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];
134 vm_page_lock_queues();
136 /* handled by vm_fault now */
137 /* vm_page_zero_invalid(m, TRUE); */
138 for (i = 0; i < npages; ++i) {
139 if (i != ap->a_reqpage)
140 vm_page_free(pages[i]);
142 vm_page_unlock_queues();
145 vm_page_unlock_queues();
149 * We use only the kva address for the buffer, but this is extremely
150 * convienient and fast.
152 bp = getpbuf(&nfs_pbuf_freecnt);
154 kva = (vm_offset_t) bp->b_data;
155 pmap_qenter(kva, pages, npages);
157 cnt.v_vnodepgsin += npages;
159 iov.iov_base = (caddr_t) kva;
163 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
164 uio.uio_resid = count;
165 uio.uio_segflg = UIO_SYSSPACE;
166 uio.uio_rw = UIO_READ;
169 error = nfs_readrpc(vp, &uio, cred);
170 pmap_qremove(kva, npages);
172 relpbuf(bp, &nfs_pbuf_freecnt);
174 if (error && (uio.uio_resid == count)) {
175 printf("nfs_getpages: error %d\n", error);
176 vm_page_lock_queues();
177 for (i = 0; i < npages; ++i) {
178 if (i != ap->a_reqpage)
179 vm_page_free(pages[i]);
181 vm_page_unlock_queues();
182 return VM_PAGER_ERROR;
186 * Calculate the number of bytes read and validate only that number
187 * of bytes. Note that due to pending writes, size may be 0. This
188 * does not mean that the remaining data is invalid!
191 size = count - uio.uio_resid;
192 vm_page_lock_queues();
193 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
195 nextoff = toff + PAGE_SIZE;
198 m->flags &= ~PG_ZERO;
200 if (nextoff <= size) {
202 * Read operation filled an entire page
204 m->valid = VM_PAGE_BITS_ALL;
206 } else if (size > toff) {
208 * Read operation filled a partial page.
211 vm_page_set_validclean(m, 0, size - toff);
212 /* handled by vm_fault now */
213 /* vm_page_zero_invalid(m, TRUE); */
216 * Read operation was short. If no error occured
217 * we may have hit a zero-fill section. We simply
218 * leave valid set to 0.
222 if (i != ap->a_reqpage) {
224 * Whether or not to leave the page activated is up in
225 * the air, but we should put the page on a page queue
226 * somewhere (it already is in the object). Result:
227 * It appears that emperical results show that
228 * deactivating pages is best.
232 * Just in case someone was asking for this page we
233 * now tell them that it is ok to use.
236 if (m->flags & PG_WANTED)
239 vm_page_deactivate(m);
246 vm_page_unlock_queues();
251 * Vnode op for VM putpages.
254 nfs_putpages(struct vop_putpages_args *ap)
260 int iomode, must_commit, i, error, npages, count;
266 struct nfsmount *nmp;
274 td = curthread; /* XXX */
275 cred = curthread->td_ucred; /* XXX */
276 nmp = VFSTONFS(vp->v_mount);
279 rtvals = ap->a_rtvals;
280 npages = btoc(count);
281 offset = IDX_TO_OFF(pages[0]->pindex);
283 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
284 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
285 (void)nfs_fsinfo(nmp, vp, cred, td);
288 for (i = 0; i < npages; i++)
289 rtvals[i] = VM_PAGER_AGAIN;
292 * When putting pages, do not extend file past EOF.
295 if (offset + count > np->n_size) {
296 count = np->n_size - offset;
302 * We use only the kva address for the buffer, but this is extremely
303 * convienient and fast.
305 bp = getpbuf(&nfs_pbuf_freecnt);
307 kva = (vm_offset_t) bp->b_data;
308 pmap_qenter(kva, pages, npages);
310 cnt.v_vnodepgsout += count;
312 iov.iov_base = (caddr_t) kva;
316 uio.uio_offset = offset;
317 uio.uio_resid = count;
318 uio.uio_segflg = UIO_SYSSPACE;
319 uio.uio_rw = UIO_WRITE;
322 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
323 iomode = NFSV3WRITE_UNSTABLE;
325 iomode = NFSV3WRITE_FILESYNC;
327 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit);
329 pmap_qremove(kva, npages);
330 relpbuf(bp, &nfs_pbuf_freecnt);
333 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
334 for (i = 0; i < nwritten; i++) {
335 rtvals[i] = VM_PAGER_OK;
336 vm_page_undirty(pages[i]);
339 nfs_clearcommit(vp->v_mount);
346 * Vnode op for read using bio
349 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
351 struct nfsnode *np = VTONFS(vp);
353 struct buf *bp = 0, *rabp;
356 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
360 int nra, error = 0, n = 0, on = 0;
363 if (uio->uio_rw != UIO_READ)
364 panic("nfs_read mode");
366 if (uio->uio_resid == 0)
368 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
372 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
373 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
374 (void)nfs_fsinfo(nmp, vp, cred, td);
375 if (vp->v_type != VDIR &&
376 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
378 biosize = vp->v_mount->mnt_stat.f_iosize;
379 seqcount = (int)((off_t)(ioflag >> 16) * biosize / BKVASIZE);
381 * For nfs, cache consistency can only be maintained approximately.
382 * Although RFC1094 does not specify the criteria, the following is
383 * believed to be compatible with the reference port.
385 * If the file's modify time on the server has changed since the
386 * last read rpc or you have written to the file,
387 * you may have lost data cache consistency with the
388 * server, so flush all of the file's data out of the cache.
389 * Then force a getattr rpc to ensure that you have up to date
391 * NB: This implies that cache data can be read when up to
392 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
393 * attributes this could be forced by setting n_attrstamp to 0 before
394 * the VOP_GETATTR() call.
396 if (np->n_flag & NMODIFIED) {
397 if (vp->v_type != VREG) {
398 if (vp->v_type != VDIR)
399 panic("nfs: bioread, not dir");
401 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
406 error = VOP_GETATTR(vp, &vattr, cred, td);
409 np->n_mtime = vattr.va_mtime.tv_sec;
411 error = VOP_GETATTR(vp, &vattr, cred, td);
414 if (np->n_mtime != vattr.va_mtime.tv_sec) {
415 if (vp->v_type == VDIR)
417 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
420 np->n_mtime = vattr.va_mtime.tv_sec;
424 switch (vp->v_type) {
426 nfsstats.biocache_reads++;
427 lbn = uio->uio_offset / biosize;
428 on = uio->uio_offset & (biosize - 1);
431 * Start the read ahead(s), as required.
433 if (nmp->nm_readahead > 0) {
434 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
435 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
436 rabn = lbn + 1 + nra;
437 if (incore(vp, rabn) == NULL) {
438 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
441 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
442 rabp->b_flags |= B_ASYNC;
443 rabp->b_iocmd = BIO_READ;
444 vfs_busy_pages(rabp, 0);
445 if (nfs_asyncio(rabp, cred, td)) {
446 rabp->b_flags |= B_INVAL;
447 rabp->b_ioflags |= BIO_ERROR;
448 vfs_unbusy_pages(rabp);
460 * Obtain the buffer cache block. Figure out the buffer size
461 * when we are at EOF. If we are modifying the size of the
462 * buffer based on an EOF condition we need to hold
463 * nfs_rslock() through obtaining the buffer to prevent
464 * a potential writer-appender from messing with n_size.
465 * Otherwise we may accidently truncate the buffer and
468 * Note that bcount is *not* DEV_BSIZE aligned.
473 if ((off_t)lbn * biosize >= np->n_size) {
475 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
476 bcount = np->n_size - (off_t)lbn * biosize;
478 if (bcount != biosize) {
479 switch(nfs_rslock(np, td)) {
492 bp = nfs_getcacheblk(vp, lbn, bcount, td);
494 if (bcount != biosize)
495 nfs_rsunlock(np, td);
500 * If B_CACHE is not set, we must issue the read. If this
501 * fails, we return an error.
504 if ((bp->b_flags & B_CACHE) == 0) {
505 bp->b_iocmd = BIO_READ;
506 vfs_busy_pages(bp, 0);
507 error = nfs_doio(bp, cred, td);
515 * on is the offset into the current bp. Figure out how many
516 * bytes we can copy out of the bp. Note that bcount is
517 * NOT DEV_BSIZE aligned.
519 * Then figure out how many bytes we can copy into the uio.
524 n = min((unsigned)(bcount - on), uio->uio_resid);
527 nfsstats.biocache_readlinks++;
528 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
531 if ((bp->b_flags & B_CACHE) == 0) {
532 bp->b_iocmd = BIO_READ;
533 vfs_busy_pages(bp, 0);
534 error = nfs_doio(bp, cred, td);
536 bp->b_ioflags |= BIO_ERROR;
541 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
545 nfsstats.biocache_readdirs++;
546 if (np->n_direofoffset
547 && uio->uio_offset >= np->n_direofoffset) {
550 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
551 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
552 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
555 if ((bp->b_flags & B_CACHE) == 0) {
556 bp->b_iocmd = BIO_READ;
557 vfs_busy_pages(bp, 0);
558 error = nfs_doio(bp, cred, td);
562 while (error == NFSERR_BAD_COOKIE) {
563 printf("got bad cookie vp %p bp %p\n", vp, bp);
565 error = nfs_vinvalbuf(vp, 0, cred, td, 1);
567 * Yuck! The directory has been modified on the
568 * server. The only way to get the block is by
569 * reading from the beginning to get all the
572 * Leave the last bp intact unless there is an error.
573 * Loop back up to the while if the error is another
574 * NFSERR_BAD_COOKIE (double yuch!).
576 for (i = 0; i <= lbn && !error; i++) {
577 if (np->n_direofoffset
578 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
580 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
583 if ((bp->b_flags & B_CACHE) == 0) {
584 bp->b_iocmd = BIO_READ;
585 vfs_busy_pages(bp, 0);
586 error = nfs_doio(bp, cred, td);
588 * no error + B_INVAL == directory EOF,
591 if (error == 0 && (bp->b_flags & B_INVAL))
595 * An error will throw away the block and the
596 * for loop will break out. If no error and this
597 * is not the block we want, we throw away the
598 * block and go for the next one via the for loop.
600 if (error || i < lbn)
605 * The above while is repeated if we hit another cookie
606 * error. If we hit an error and it wasn't a cookie error,
614 * If not eof and read aheads are enabled, start one.
615 * (You need the current block first, so that you have the
616 * directory offset cookie of the next block.)
618 if (nmp->nm_readahead > 0 &&
619 (bp->b_flags & B_INVAL) == 0 &&
620 (np->n_direofoffset == 0 ||
621 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
622 incore(vp, lbn + 1) == NULL) {
623 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
625 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
626 rabp->b_flags |= B_ASYNC;
627 rabp->b_iocmd = BIO_READ;
628 vfs_busy_pages(rabp, 0);
629 if (nfs_asyncio(rabp, cred, td)) {
630 rabp->b_flags |= B_INVAL;
631 rabp->b_ioflags |= BIO_ERROR;
632 vfs_unbusy_pages(rabp);
641 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
642 * chopped for the EOF condition, we cannot tell how large
643 * NFS directories are going to be until we hit EOF. So
644 * an NFS directory buffer is *not* chopped to its EOF. Now,
645 * it just so happens that b_resid will effectively chop it
646 * to EOF. *BUT* this information is lost if the buffer goes
647 * away and is reconstituted into a B_CACHE state ( due to
648 * being VMIO ) later. So we keep track of the directory eof
649 * in np->n_direofoffset and chop it off as an extra step
652 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
653 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
654 n = np->n_direofoffset - uio->uio_offset;
657 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
662 error = uiomove(bp->b_data + on, (int)n, uio);
664 switch (vp->v_type) {
673 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
676 } while (error == 0 && uio->uio_resid > 0 && n > 0);
681 * Vnode op for write using bio
684 nfs_write(struct vop_write_args *ap)
687 struct uio *uio = ap->a_uio;
688 struct thread *td = uio->uio_td;
689 struct vnode *vp = ap->a_vp;
690 struct nfsnode *np = VTONFS(vp);
691 struct ucred *cred = ap->a_cred;
692 int ioflag = ap->a_ioflag;
695 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
698 int n, on, error = 0;
700 struct proc *p = td?td->td_proc:NULL;
705 if (uio->uio_rw != UIO_WRITE)
706 panic("nfs_write mode");
707 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
708 panic("nfs_write proc");
710 if (vp->v_type != VREG)
712 if (np->n_flag & NWRITEERR) {
713 np->n_flag &= ~NWRITEERR;
714 return (np->n_error);
716 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
717 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
718 (void)nfs_fsinfo(nmp, vp, cred, td);
721 * Synchronously flush pending buffers if we are in synchronous
722 * mode or if we are appending.
724 if (ioflag & (IO_APPEND | IO_SYNC)) {
725 if (np->n_flag & NMODIFIED) {
727 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
734 * If IO_APPEND then load uio_offset. We restart here if we cannot
735 * get the append lock.
738 if (ioflag & IO_APPEND) {
740 error = VOP_GETATTR(vp, &vattr, cred, td);
743 uio->uio_offset = np->n_size;
746 if (uio->uio_offset < 0)
748 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
750 if (uio->uio_resid == 0)
754 * We need to obtain the rslock if we intend to modify np->n_size
755 * in order to guarentee the append point with multiple contending
756 * writers, to guarentee that no other appenders modify n_size
757 * while we are trying to obtain a truncated buffer (i.e. to avoid
758 * accidently truncating data written by another appender due to
759 * the race), and to ensure that the buffer is populated prior to
760 * our extending of the file. We hold rslock through the entire
763 * Note that we do not synchronize the case where someone truncates
764 * the file while we are appending to it because attempting to lock
765 * this case may deadlock other parts of the system unexpectedly.
767 if ((ioflag & IO_APPEND) ||
768 uio->uio_offset + uio->uio_resid > np->n_size) {
769 switch(nfs_rslock(np, td)) {
784 * Maybe this should be above the vnode op call, but so long as
785 * file servers have no limits, i don't think it matters
787 if (p && uio->uio_offset + uio->uio_resid >
788 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
793 nfs_rsunlock(np, td);
797 biosize = vp->v_mount->mnt_stat.f_iosize;
800 nfsstats.biocache_writes++;
801 lbn = uio->uio_offset / biosize;
802 on = uio->uio_offset & (biosize-1);
803 n = min((unsigned)(biosize - on), uio->uio_resid);
806 * Handle direct append and file extension cases, calculate
807 * unaligned buffer size.
810 if (uio->uio_offset == np->n_size && n) {
812 * Get the buffer (in its pre-append state to maintain
813 * B_CACHE if it was previously set). Resize the
814 * nfsnode after we have locked the buffer to prevent
815 * readers from reading garbage.
818 bp = nfs_getcacheblk(vp, lbn, bcount, td);
823 np->n_size = uio->uio_offset + n;
824 np->n_flag |= NMODIFIED;
825 vnode_pager_setsize(vp, np->n_size);
827 save = bp->b_flags & B_CACHE;
829 allocbuf(bp, bcount);
831 bp->b_magic = B_MAGIC_NFS;
832 bp->b_op = &buf_ops_nfs;
836 * Obtain the locked cache block first, and then
837 * adjust the file's size as appropriate.
840 if ((off_t)lbn * biosize + bcount < np->n_size) {
841 if ((off_t)(lbn + 1) * biosize < np->n_size)
844 bcount = np->n_size - (off_t)lbn * biosize;
846 bp = nfs_getcacheblk(vp, lbn, bcount, td);
847 if (uio->uio_offset + n > np->n_size) {
848 np->n_size = uio->uio_offset + n;
849 np->n_flag |= NMODIFIED;
850 vnode_pager_setsize(vp, np->n_size);
860 * Issue a READ if B_CACHE is not set. In special-append
861 * mode, B_CACHE is based on the buffer prior to the write
862 * op and is typically set, avoiding the read. If a read
863 * is required in special append mode, the server will
864 * probably send us a short-read since we extended the file
865 * on our end, resulting in b_resid == 0 and, thusly,
866 * B_CACHE getting set.
868 * We can also avoid issuing the read if the write covers
869 * the entire buffer. We have to make sure the buffer state
870 * is reasonable in this case since we will not be initiating
871 * I/O. See the comments in kern/vfs_bio.c's getblk() for
874 * B_CACHE may also be set due to the buffer being cached
878 if (on == 0 && n == bcount) {
879 bp->b_flags |= B_CACHE;
880 bp->b_flags &= ~B_INVAL;
881 bp->b_ioflags &= ~BIO_ERROR;
884 if ((bp->b_flags & B_CACHE) == 0) {
885 bp->b_iocmd = BIO_READ;
886 vfs_busy_pages(bp, 0);
887 error = nfs_doio(bp, cred, td);
897 if (bp->b_wcred == NOCRED)
898 bp->b_wcred = crhold(cred);
899 np->n_flag |= NMODIFIED;
902 * If dirtyend exceeds file size, chop it down. This should
903 * not normally occur but there is an append race where it
904 * might occur XXX, so we log it.
906 * If the chopping creates a reverse-indexed or degenerate
907 * situation with dirtyoff/end, we 0 both of them.
910 if (bp->b_dirtyend > bcount) {
911 printf("NFS append race @%lx:%d\n",
912 (long)bp->b_blkno * DEV_BSIZE,
913 bp->b_dirtyend - bcount);
914 bp->b_dirtyend = bcount;
917 if (bp->b_dirtyoff >= bp->b_dirtyend)
918 bp->b_dirtyoff = bp->b_dirtyend = 0;
921 * If the new write will leave a contiguous dirty
922 * area, just update the b_dirtyoff and b_dirtyend,
923 * otherwise force a write rpc of the old dirty area.
925 * While it is possible to merge discontiguous writes due to
926 * our having a B_CACHE buffer ( and thus valid read data
927 * for the hole), we don't because it could lead to
928 * significant cache coherency problems with multiple clients,
929 * especially if locking is implemented later on.
931 * as an optimization we could theoretically maintain
932 * a linked list of discontinuous areas, but we would still
933 * have to commit them separately so there isn't much
934 * advantage to it except perhaps a bit of asynchronization.
937 if (bp->b_dirtyend > 0 &&
938 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
939 if (BUF_WRITE(bp) == EINTR) {
946 error = uiomove((char *)bp->b_data + on, n, uio);
949 * Since this block is being modified, it must be written
950 * again and not just committed. Since write clustering does
951 * not work for the stage 1 data write, only the stage 2
952 * commit rpc, we have to clear B_CLUSTEROK as well.
954 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
957 bp->b_ioflags |= BIO_ERROR;
963 * Only update dirtyoff/dirtyend if not a degenerate
967 if (bp->b_dirtyend > 0) {
968 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
969 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
972 bp->b_dirtyend = on + n;
974 vfs_bio_set_validclean(bp, on, n);
977 * If IO_NOWDRAIN then set B_NOWDRAIN (nfs-backed MD
980 if (ioflag & IO_NOWDRAIN)
981 bp->b_flags |= B_NOWDRAIN;
984 * If IO_SYNC do bwrite().
986 * IO_INVAL appears to be unused. The idea appears to be
987 * to turn off caching in this case. Very odd. XXX
989 if ((ioflag & IO_SYNC)) {
990 if (ioflag & IO_INVAL)
991 bp->b_flags |= B_NOCACHE;
992 error = BUF_WRITE(bp);
995 } else if ((n + on) == biosize) {
996 bp->b_flags |= B_ASYNC;
997 (void)nfs_writebp(bp, 0, 0);
1001 } while (uio->uio_resid > 0 && n > 0);
1004 nfs_rsunlock(np, td);
1010 * Get an nfs cache block.
1012 * Allocate a new one if the block isn't currently in the cache
1013 * and return the block marked busy. If the calling process is
1014 * interrupted by a signal for an interruptible mount point, return
1017 * The caller must carefully deal with the possible B_INVAL state of
1018 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1019 * indirectly), so synchronous reads can be issued without worrying about
1020 * the B_INVAL state. We have to be a little more careful when dealing
1021 * with writes (see comments in nfs_write()) when extending a file past
1025 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1029 struct nfsmount *nmp;
1034 if (nmp->nm_flag & NFSMNT_INT) {
1035 bp = getblk(vp, bn, size, PCATCH, 0);
1036 while (bp == NULL) {
1037 if (nfs_sigintr(nmp, NULL, td))
1039 bp = getblk(vp, bn, size, 0, 2 * hz);
1042 bp = getblk(vp, bn, size, 0, 0);
1045 if (vp->v_type == VREG) {
1048 biosize = mp->mnt_stat.f_iosize;
1049 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1055 * Flush and invalidate all dirty buffers. If another process is already
1056 * doing the flush, just wait for completion.
1059 nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred,
1060 struct thread *td, int intrflg)
1062 struct nfsnode *np = VTONFS(vp);
1063 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1064 int error = 0, slpflag, slptimeo;
1066 if (vp->v_flag & VXLOCK) {
1070 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1080 * First wait for any other process doing a flush to complete.
1082 while (np->n_flag & NFLUSHINPROG) {
1083 np->n_flag |= NFLUSHWANT;
1084 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
1086 if (error && intrflg &&
1087 nfs_sigintr(nmp, NULL, td))
1092 * Now, flush as required.
1094 np->n_flag |= NFLUSHINPROG;
1095 error = vinvalbuf(vp, flags, cred, td, slpflag, 0);
1098 nfs_sigintr(nmp, NULL, td)) {
1099 np->n_flag &= ~NFLUSHINPROG;
1100 if (np->n_flag & NFLUSHWANT) {
1101 np->n_flag &= ~NFLUSHWANT;
1102 wakeup((caddr_t)&np->n_flag);
1106 error = vinvalbuf(vp, flags, cred, td, 0, slptimeo);
1108 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1109 if (np->n_flag & NFLUSHWANT) {
1110 np->n_flag &= ~NFLUSHWANT;
1111 wakeup((caddr_t)&np->n_flag);
1117 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1118 * This is mainly to avoid queueing async I/O requests when the nfsiods
1119 * are all hung on a dead server.
1121 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1122 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1125 nfs_asyncio(struct buf *bp, struct ucred *cred, struct thread *td)
1127 struct nfsmount *nmp;
1134 nmp = VFSTONFS(bp->b_vp->v_mount);
1137 * Commits are usually short and sweet so lets save some cpu and
1138 * leave the async daemons for more important rpc's (such as reads
1141 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1142 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1147 if (nmp->nm_flag & NFSMNT_INT)
1152 * Find a free iod to process this request.
1154 for (iod = 0; iod < nfs_numasync; iod++)
1155 if (nfs_iodwant[iod]) {
1161 * Try to create one if none are free.
1164 iod = nfs_nfsiodnew();
1171 * Found one, so wake it up and tell it which
1174 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1176 nfs_iodwant[iod] = NULL;
1177 nfs_iodmount[iod] = nmp;
1179 wakeup((caddr_t)&nfs_iodwant[iod]);
1183 * If none are free, we may already have an iod working on this mount
1184 * point. If so, it will process our request.
1187 if (nmp->nm_bufqiods > 0) {
1189 ("nfs_asyncio: %d iods are already processing mount %p\n",
1190 nmp->nm_bufqiods, nmp));
1196 * If we have an iod which can process the request, then queue
1201 * Ensure that the queue never grows too large. We still want
1202 * to asynchronize so we block rather then return EIO.
1204 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1206 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1207 nmp->nm_bufqwant = TRUE;
1208 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
1209 "nfsaio", slptimeo);
1211 if (nfs_sigintr(nmp, NULL, td))
1213 if (slpflag == PCATCH) {
1219 * We might have lost our iod while sleeping,
1220 * so check and loop if nescessary.
1222 if (nmp->nm_bufqiods == 0) {
1224 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1229 if (bp->b_iocmd == BIO_READ) {
1230 if (bp->b_rcred == NOCRED && cred != NOCRED)
1231 bp->b_rcred = crhold(cred);
1233 bp->b_flags |= B_WRITEINPROG;
1234 if (bp->b_wcred == NOCRED && cred != NOCRED)
1235 bp->b_wcred = crhold(cred);
1239 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1245 * All the iods are busy on other mounts, so return EIO to
1246 * force the caller to process the i/o synchronously.
1248 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1253 * Do an I/O operation to/from a cache block. This may be called
1254 * synchronously or from an nfsiod.
1257 nfs_doio(struct buf *bp, struct ucred *cr, struct thread *td)
1262 struct nfsmount *nmp;
1263 int error = 0, iomode, must_commit = 0;
1266 struct proc *p = td ? td->td_proc : NULL;
1270 nmp = VFSTONFS(vp->v_mount);
1272 uiop->uio_iov = &io;
1273 uiop->uio_iovcnt = 1;
1274 uiop->uio_segflg = UIO_SYSSPACE;
1278 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1279 * do this here so we do not have to do it in all the code that
1282 bp->b_flags &= ~B_INVAL;
1283 bp->b_ioflags &= ~BIO_ERROR;
1285 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1288 * Historically, paging was done with physio, but no more.
1290 if (bp->b_flags & B_PHYS) {
1292 * ...though reading /dev/drum still gets us here.
1294 io.iov_len = uiop->uio_resid = bp->b_bcount;
1295 /* mapping was done by vmapbuf() */
1296 io.iov_base = bp->b_data;
1297 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1298 if (bp->b_iocmd == BIO_READ) {
1299 uiop->uio_rw = UIO_READ;
1300 nfsstats.read_physios++;
1301 error = nfs_readrpc(vp, uiop, cr);
1305 iomode = NFSV3WRITE_DATASYNC;
1306 uiop->uio_rw = UIO_WRITE;
1307 nfsstats.write_physios++;
1308 error = nfs_writerpc(vp, uiop, cr, &iomode, &com);
1311 bp->b_ioflags |= BIO_ERROR;
1312 bp->b_error = error;
1314 } else if (bp->b_iocmd == BIO_READ) {
1315 io.iov_len = uiop->uio_resid = bp->b_bcount;
1316 io.iov_base = bp->b_data;
1317 uiop->uio_rw = UIO_READ;
1319 switch (vp->v_type) {
1321 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1322 nfsstats.read_bios++;
1323 error = nfs_readrpc(vp, uiop, cr);
1326 if (uiop->uio_resid) {
1328 * If we had a short read with no error, we must have
1329 * hit a file hole. We should zero-fill the remainder.
1330 * This can also occur if the server hits the file EOF.
1332 * Holes used to be able to occur due to pending
1333 * writes, but that is not possible any longer.
1335 int nread = bp->b_bcount - uiop->uio_resid;
1336 int left = uiop->uio_resid;
1339 bzero((char *)bp->b_data + nread, left);
1340 uiop->uio_resid = 0;
1343 if (p && (vp->v_flag & VTEXT) &&
1344 (np->n_mtime != np->n_vattr.va_mtime.tv_sec)) {
1345 uprintf("Process killed due to text file modification\n");
1347 psignal(p, SIGKILL);
1353 uiop->uio_offset = (off_t)0;
1354 nfsstats.readlink_bios++;
1355 error = nfs_readlinkrpc(vp, uiop, cr);
1358 nfsstats.readdir_bios++;
1359 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1360 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1361 error = nfs_readdirplusrpc(vp, uiop, cr);
1362 if (error == NFSERR_NOTSUPP)
1363 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1365 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1366 error = nfs_readdirrpc(vp, uiop, cr);
1368 * end-of-directory sets B_INVAL but does not generate an
1371 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1372 bp->b_flags |= B_INVAL;
1375 printf("nfs_doio: type %x unexpected\n", vp->v_type);
1379 bp->b_ioflags |= BIO_ERROR;
1380 bp->b_error = error;
1384 * If we only need to commit, try to commit
1386 if (bp->b_flags & B_NEEDCOMMIT) {
1390 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1391 bp->b_flags |= B_WRITEINPROG;
1393 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1395 bp->b_flags &= ~B_WRITEINPROG;
1397 bp->b_dirtyoff = bp->b_dirtyend = 0;
1398 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1403 if (retv == NFSERR_STALEWRITEVERF) {
1404 nfs_clearcommit(bp->b_vp->v_mount);
1409 * Setup for actual write
1412 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1413 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1415 if (bp->b_dirtyend > bp->b_dirtyoff) {
1416 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1418 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1420 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1421 uiop->uio_rw = UIO_WRITE;
1422 nfsstats.write_bios++;
1424 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1425 iomode = NFSV3WRITE_UNSTABLE;
1427 iomode = NFSV3WRITE_FILESYNC;
1429 bp->b_flags |= B_WRITEINPROG;
1430 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1433 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1434 * to cluster the buffers needing commit. This will allow
1435 * the system to submit a single commit rpc for the whole
1436 * cluster. We can do this even if the buffer is not 100%
1437 * dirty (relative to the NFS blocksize), so we optimize the
1438 * append-to-file-case.
1440 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1441 * cleared because write clustering only works for commit
1442 * rpc's, not for the data portion of the write).
1445 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1446 bp->b_flags |= B_NEEDCOMMIT;
1447 if (bp->b_dirtyoff == 0
1448 && bp->b_dirtyend == bp->b_bcount)
1449 bp->b_flags |= B_CLUSTEROK;
1451 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1453 bp->b_flags &= ~B_WRITEINPROG;
1456 * For an interrupted write, the buffer is still valid
1457 * and the write hasn't been pushed to the server yet,
1458 * so we can't set BIO_ERROR and report the interruption
1459 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1460 * is not relevant, so the rpc attempt is essentially
1461 * a noop. For the case of a V3 write rpc not being
1462 * committed to stable storage, the block is still
1463 * dirty and requires either a commit rpc or another
1464 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1465 * the block is reused. This is indicated by setting
1466 * the B_DELWRI and B_NEEDCOMMIT flags.
1468 * If the buffer is marked B_PAGING, it does not reside on
1469 * the vp's paging queues so we cannot call bdirty(). The
1470 * bp in this case is not an NFS cache block so we should
1474 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1478 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1479 if ((bp->b_flags & B_PAGING) == 0) {
1481 bp->b_flags &= ~B_DONE;
1483 if (error && (bp->b_flags & B_ASYNC) == 0)
1484 bp->b_flags |= B_EINTR;
1488 bp->b_ioflags |= BIO_ERROR;
1489 bp->b_error = np->n_error = error;
1490 np->n_flag |= NWRITEERR;
1492 bp->b_dirtyoff = bp->b_dirtyend = 0;
1500 bp->b_resid = uiop->uio_resid;
1502 nfs_clearcommit(vp->v_mount);
1508 * Used to aid in handling ftruncate() operations on the NFS client side.
1509 * Truncation creates a number of special problems for NFS. We have to
1510 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1511 * we have to properly handle VM pages or (potentially dirty) buffers
1512 * that straddle the truncation point.
1516 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1518 struct nfsnode *np = VTONFS(vp);
1519 u_quad_t tsize = np->n_size;
1520 int biosize = vp->v_mount->mnt_stat.f_iosize;
1525 if (np->n_size < tsize) {
1531 * vtruncbuf() doesn't get the buffer overlapping the
1532 * truncation point. We may have a B_DELWRI and/or B_CACHE
1533 * buffer that now needs to be truncated.
1535 error = vtruncbuf(vp, cred, td, nsize, biosize);
1536 lbn = nsize / biosize;
1537 bufsize = nsize & (biosize - 1);
1538 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1539 if (bp->b_dirtyoff > bp->b_bcount)
1540 bp->b_dirtyoff = bp->b_bcount;
1541 if (bp->b_dirtyend > bp->b_bcount)
1542 bp->b_dirtyend = bp->b_bcount;
1543 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1546 vnode_pager_setsize(vp, nsize);