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: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.28 2006/03/24 18:35:34 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>
52 #include <sys/msfbuf.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 <sys/thread2.h>
70 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset,
71 int size, struct thread *td);
73 extern int nfs_numasync;
74 extern int nfs_pbuf_freecnt;
75 extern struct nfsstats nfsstats;
78 * Vnode op for VM getpages.
80 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
81 * int a_reqpage, vm_ooffset_t a_offset)
84 nfs_getpages(struct vop_getpages_args *ap)
86 struct thread *td = curthread; /* XXX */
87 int i, error, nextoff, size, toff, count, npages;
98 nmp = VFSTONFS(vp->v_mount);
102 if (vp->v_object == NULL) {
103 printf("nfs_getpages: called with non-merged cache vnode??\n");
104 return VM_PAGER_ERROR;
107 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
108 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
109 (void)nfs_fsinfo(nmp, vp, td);
111 npages = btoc(count);
114 * NOTE that partially valid pages may occur in cases other
115 * then file EOF, such as when a file is partially written and
116 * ftruncate()-extended to a larger size. It is also possible
117 * for the valid bits to be set on garbage beyond the file EOF and
118 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
119 * occur due to vtruncbuf() and the buffer cache's handling of
120 * pages which 'straddle' buffers or when b_bufsize is not a
121 * multiple of PAGE_SIZE.... the buffer cache cannot normally
122 * clear the extra bits. This kind of situation occurs when you
123 * make a small write() (m->valid == 0x03) and then mmap() and
124 * fault in the buffer(m->valid = 0xFF). When NFS flushes the
125 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
127 * This is combined with the possibility that the pages are partially
128 * dirty or that there is a buffer backing the pages that is dirty
129 * (even if m->dirty is 0).
131 * To solve this problem several hacks have been made: (1) NFS
132 * guarentees that the IO block size is a multiple of PAGE_SIZE and
133 * (2) The buffer cache, when invalidating an NFS buffer, will
134 * disregard the buffer's fragmentory b_bufsize and invalidate
135 * the whole page rather then just the piece the buffer owns.
137 * This allows us to assume that a partially valid page found here
138 * is fully valid (vm_fault will zero'd out areas of the page not
141 m = pages[ap->a_reqpage];
143 for (i = 0; i < npages; ++i) {
144 if (i != ap->a_reqpage)
145 vnode_pager_freepage(pages[i]);
151 * Use an MSF_BUF as a medium to retrieve data from the pages.
153 msf_map_pagelist(&msf, pages, npages, 0);
155 kva = msf_buf_kva(msf);
161 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
162 uio.uio_resid = count;
163 uio.uio_segflg = UIO_SYSSPACE;
164 uio.uio_rw = UIO_READ;
167 error = nfs_readrpc(vp, &uio);
170 if (error && (uio.uio_resid == count)) {
171 printf("nfs_getpages: error %d\n", error);
172 for (i = 0; i < npages; ++i) {
173 if (i != ap->a_reqpage)
174 vnode_pager_freepage(pages[i]);
176 return VM_PAGER_ERROR;
180 * Calculate the number of bytes read and validate only that number
181 * of bytes. Note that due to pending writes, size may be 0. This
182 * does not mean that the remaining data is invalid!
185 size = count - uio.uio_resid;
187 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
188 nextoff = toff + PAGE_SIZE;
191 m->flags &= ~PG_ZERO;
193 if (nextoff <= size) {
195 * Read operation filled an entire page
197 m->valid = VM_PAGE_BITS_ALL;
199 } else if (size > toff) {
201 * Read operation filled a partial page.
204 vm_page_set_validclean(m, 0, size - toff);
205 /* handled by vm_fault now */
206 /* vm_page_zero_invalid(m, TRUE); */
209 * Read operation was short. If no error occured
210 * we may have hit a zero-fill section. We simply
211 * leave valid set to 0.
215 if (i != ap->a_reqpage) {
217 * Whether or not to leave the page activated is up in
218 * the air, but we should put the page on a page queue
219 * somewhere (it already is in the object). Result:
220 * It appears that emperical results show that
221 * deactivating pages is best.
225 * Just in case someone was asking for this page we
226 * now tell them that it is ok to use.
229 if (m->flags & PG_WANTED)
232 vm_page_deactivate(m);
235 vnode_pager_freepage(m);
243 * Vnode op for VM putpages.
245 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
246 * int *a_rtvals, vm_ooffset_t a_offset)
249 nfs_putpages(struct vop_putpages_args *ap)
251 struct thread *td = curthread;
255 int iomode, must_commit, i, error, npages, count;
259 struct nfsmount *nmp;
266 nmp = VFSTONFS(vp->v_mount);
269 rtvals = ap->a_rtvals;
270 npages = btoc(count);
271 offset = IDX_TO_OFF(pages[0]->pindex);
273 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
274 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
275 (void)nfs_fsinfo(nmp, vp, td);
277 for (i = 0; i < npages; i++) {
278 rtvals[i] = VM_PAGER_AGAIN;
282 * When putting pages, do not extend file past EOF.
285 if (offset + count > np->n_size) {
286 count = np->n_size - offset;
292 * Use an MSF_BUF as a medium to retrieve data from the pages.
294 msf_map_pagelist(&msf, pages, npages, 0);
296 kva = msf_buf_kva(msf);
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, &iomode, &must_commit);
318 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
319 for (i = 0; i < nwritten; i++) {
320 rtvals[i] = VM_PAGER_OK;
321 vm_page_undirty(pages[i]);
324 nfs_clearcommit(vp->v_mount);
330 * Vnode op for read using bio
333 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
335 struct nfsnode *np = VTONFS(vp);
337 struct buf *bp = 0, *rabp;
340 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
346 int nra, error = 0, n = 0, on = 0;
349 if (uio->uio_rw != UIO_READ)
350 panic("nfs_read mode");
352 if (uio->uio_resid == 0)
354 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
358 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
359 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
360 (void)nfs_fsinfo(nmp, vp, td);
361 if (vp->v_type != VDIR &&
362 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
364 biosize = vp->v_mount->mnt_stat.f_iosize;
365 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
368 * For nfs, cache consistency can only be maintained approximately.
369 * Although RFC1094 does not specify the criteria, the following is
370 * believed to be compatible with the reference port.
372 * NQNFS: Full cache coherency is maintained within the loop.
374 * NFS: If local changes have been made and this is a
375 * directory, the directory must be invalidated and
376 * the attribute cache must be cleared.
378 * GETATTR is called to synchronize the file size.
380 * If remote changes are detected local data is flushed
381 * and the cache is invalidated.
384 * NOTE: In the normal case the attribute cache is not
385 * cleared which means GETATTR may use cached data and
386 * not immediately detect changes made on the server.
388 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
389 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
391 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
396 error = VOP_GETATTR(vp, &vattr, td);
399 if (np->n_flag & NRMODIFIED) {
400 if (vp->v_type == VDIR)
402 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
405 np->n_flag &= ~NRMODIFIED;
411 * Get a valid lease. If cached data is stale, flush it.
413 if (nmp->nm_flag & NFSMNT_NQNFS) {
414 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
416 error = nqnfs_getlease(vp, ND_READ, td);
417 } while (error == NQNFS_EXPIRED);
420 if (np->n_lrev != np->n_brev ||
421 (np->n_flag & NQNFSNONCACHE) ||
422 ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR)) {
423 if (vp->v_type == VDIR)
425 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
428 np->n_brev = np->n_lrev;
430 } else if (vp->v_type == VDIR && (np->n_flag & NLMODIFIED)) {
432 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
437 if (np->n_flag & NQNFSNONCACHE) {
438 switch (vp->v_type) {
440 return (nfs_readrpc(vp, uio));
442 return (nfs_readlinkrpc(vp, uio));
446 printf(" NQNFSNONCACHE: type %x unexpected\n",
450 switch (vp->v_type) {
452 nfsstats.biocache_reads++;
453 lbn = uio->uio_offset / biosize;
454 on = uio->uio_offset & (biosize - 1);
455 loffset = (off_t)lbn * biosize;
458 * Start the read ahead(s), as required.
460 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
461 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
462 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
463 rabn = lbn + 1 + nra;
464 raoffset = (off_t)rabn * biosize;
465 if (!findblk(vp, raoffset)) {
466 rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
469 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
470 rabp->b_flags |= (B_READ | B_ASYNC);
471 vfs_busy_pages(rabp, 0);
472 if (nfs_asyncio(vp, &rabp->b_bio2, td)) {
473 rabp->b_flags |= B_INVAL|B_ERROR;
474 vfs_unbusy_pages(rabp);
486 * Obtain the buffer cache block. Figure out the buffer size
487 * when we are at EOF. If we are modifying the size of the
488 * buffer based on an EOF condition we need to hold
489 * nfs_rslock() through obtaining the buffer to prevent
490 * a potential writer-appender from messing with n_size.
491 * Otherwise we may accidently truncate the buffer and
494 * Note that bcount is *not* DEV_BSIZE aligned.
499 if (loffset >= np->n_size) {
501 } else if (loffset + biosize > np->n_size) {
502 bcount = np->n_size - loffset;
504 if (bcount != biosize) {
505 switch(nfs_rslock(np, td)) {
518 bp = nfs_getcacheblk(vp, loffset, bcount, td);
520 if (bcount != biosize)
521 nfs_rsunlock(np, td);
526 * If B_CACHE is not set, we must issue the read. If this
527 * fails, we return an error.
530 if ((bp->b_flags & B_CACHE) == 0) {
531 bp->b_flags |= B_READ;
532 vfs_busy_pages(bp, 0);
533 error = nfs_doio(vp, &bp->b_bio2, td);
541 * on is the offset into the current bp. Figure out how many
542 * bytes we can copy out of the bp. Note that bcount is
543 * NOT DEV_BSIZE aligned.
545 * Then figure out how many bytes we can copy into the uio.
550 n = min((unsigned)(bcount - on), uio->uio_resid);
553 nfsstats.biocache_readlinks++;
554 bp = nfs_getcacheblk(vp, (off_t)0, NFS_MAXPATHLEN, td);
557 if ((bp->b_flags & B_CACHE) == 0) {
558 bp->b_flags |= B_READ;
559 vfs_busy_pages(bp, 0);
560 error = nfs_doio(vp, &bp->b_bio2, td);
562 bp->b_flags |= B_ERROR;
567 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
571 nfsstats.biocache_readdirs++;
572 if (np->n_direofoffset
573 && uio->uio_offset >= np->n_direofoffset) {
576 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
577 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
578 loffset = uio->uio_offset - on;
579 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
582 if ((bp->b_flags & B_CACHE) == 0) {
583 bp->b_flags |= B_READ;
584 vfs_busy_pages(bp, 0);
585 error = nfs_doio(vp, &bp->b_bio2, td);
589 while (error == NFSERR_BAD_COOKIE) {
590 printf("got bad cookie vp %p bp %p\n", vp, bp);
592 error = nfs_vinvalbuf(vp, 0, td, 1);
594 * Yuck! The directory has been modified on the
595 * server. The only way to get the block is by
596 * reading from the beginning to get all the
599 * Leave the last bp intact unless there is an error.
600 * Loop back up to the while if the error is another
601 * NFSERR_BAD_COOKIE (double yuch!).
603 for (i = 0; i <= lbn && !error; i++) {
604 if (np->n_direofoffset
605 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
607 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
611 if ((bp->b_flags & B_CACHE) == 0) {
612 bp->b_flags |= B_READ;
613 vfs_busy_pages(bp, 0);
614 error = nfs_doio(vp, &bp->b_bio2, td);
616 * no error + B_INVAL == directory EOF,
619 if (error == 0 && (bp->b_flags & B_INVAL))
623 * An error will throw away the block and the
624 * for loop will break out. If no error and this
625 * is not the block we want, we throw away the
626 * block and go for the next one via the for loop.
628 if (error || i < lbn)
633 * The above while is repeated if we hit another cookie
634 * error. If we hit an error and it wasn't a cookie error,
642 * If not eof and read aheads are enabled, start one.
643 * (You need the current block first, so that you have the
644 * directory offset cookie of the next block.)
646 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
647 (bp->b_flags & B_INVAL) == 0 &&
648 (np->n_direofoffset == 0 ||
649 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
650 !(np->n_flag & NQNFSNONCACHE) &&
651 !findblk(vp, loffset + NFS_DIRBLKSIZ)) {
652 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
655 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
656 rabp->b_flags |= (B_READ | B_ASYNC);
657 vfs_busy_pages(rabp, 0);
658 if (nfs_asyncio(vp, &rabp->b_bio2, td)) {
659 rabp->b_flags |= B_INVAL|B_ERROR;
660 vfs_unbusy_pages(rabp);
669 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
670 * chopped for the EOF condition, we cannot tell how large
671 * NFS directories are going to be until we hit EOF. So
672 * an NFS directory buffer is *not* chopped to its EOF. Now,
673 * it just so happens that b_resid will effectively chop it
674 * to EOF. *BUT* this information is lost if the buffer goes
675 * away and is reconstituted into a B_CACHE state ( due to
676 * being VMIO ) later. So we keep track of the directory eof
677 * in np->n_direofoffset and chop it off as an extra step
680 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
681 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
682 n = np->n_direofoffset - uio->uio_offset;
685 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
689 switch (vp->v_type) {
692 error = uiomove(bp->b_data + on, (int)n, uio);
696 error = uiomove(bp->b_data + on, (int)n, uio);
701 off_t old_off = uio->uio_offset;
703 struct nfs_dirent *dp;
705 cpos = bp->b_data + on;
706 epos = bp->b_data + on + n;
707 while (cpos < epos && error == 0 && uio->uio_resid > 0) {
708 dp = (struct nfs_dirent *)cpos;
709 if (vop_write_dirent(&error, uio, dp->nfs_ino,
710 dp->nfs_type, dp->nfs_namlen, dp->nfs_name))
712 cpos += dp->nfs_reclen;
716 uio->uio_offset = old_off + cpos - bp->b_data - on;
719 * Invalidate buffer if caching is disabled, forcing a
720 * re-read from the remote later.
722 if (np->n_flag & NQNFSNONCACHE)
723 bp->b_flags |= B_INVAL;
726 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
729 } while (error == 0 && uio->uio_resid > 0 && n > 0);
734 * Vnode op for write using bio
736 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
737 * struct ucred *a_cred)
740 nfs_write(struct vop_write_args *ap)
742 struct uio *uio = ap->a_uio;
743 struct thread *td = uio->uio_td;
744 struct vnode *vp = ap->a_vp;
745 struct nfsnode *np = VTONFS(vp);
746 int ioflag = ap->a_ioflag;
749 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
752 int n, on, error = 0, iomode, must_commit;
758 if (uio->uio_rw != UIO_WRITE)
759 panic("nfs_write mode");
760 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
761 panic("nfs_write proc");
763 if (vp->v_type != VREG)
765 if (np->n_flag & NWRITEERR) {
766 np->n_flag &= ~NWRITEERR;
767 return (np->n_error);
769 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
770 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
771 (void)nfs_fsinfo(nmp, vp, td);
774 * Synchronously flush pending buffers if we are in synchronous
775 * mode or if we are appending.
777 if (ioflag & (IO_APPEND | IO_SYNC)) {
778 if (np->n_flag & NLMODIFIED) {
780 error = nfs_flush(vp, MNT_WAIT, td, 0);
781 /* error = nfs_vinvalbuf(vp, V_SAVE, td, 1); */
788 * If IO_APPEND then load uio_offset. We restart here if we cannot
789 * get the append lock.
792 if (ioflag & IO_APPEND) {
794 error = VOP_GETATTR(vp, &vattr, td);
797 uio->uio_offset = np->n_size;
800 if (uio->uio_offset < 0)
802 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
804 if (uio->uio_resid == 0)
808 * We need to obtain the rslock if we intend to modify np->n_size
809 * in order to guarentee the append point with multiple contending
810 * writers, to guarentee that no other appenders modify n_size
811 * while we are trying to obtain a truncated buffer (i.e. to avoid
812 * accidently truncating data written by another appender due to
813 * the race), and to ensure that the buffer is populated prior to
814 * our extending of the file. We hold rslock through the entire
817 * Note that we do not synchronize the case where someone truncates
818 * the file while we are appending to it because attempting to lock
819 * this case may deadlock other parts of the system unexpectedly.
821 if ((ioflag & IO_APPEND) ||
822 uio->uio_offset + uio->uio_resid > np->n_size) {
823 switch(nfs_rslock(np, td)) {
838 * Maybe this should be above the vnode op call, but so long as
839 * file servers have no limits, i don't think it matters
841 if (td->td_proc && uio->uio_offset + uio->uio_resid >
842 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
843 psignal(td->td_proc, SIGXFSZ);
845 nfs_rsunlock(np, td);
849 biosize = vp->v_mount->mnt_stat.f_iosize;
853 * Check for a valid write lease.
855 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
856 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
858 error = nqnfs_getlease(vp, ND_WRITE, td);
859 } while (error == NQNFS_EXPIRED);
862 if (np->n_lrev != np->n_brev ||
863 (np->n_flag & NQNFSNONCACHE)) {
864 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
867 np->n_brev = np->n_lrev;
870 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
871 iomode = NFSV3WRITE_FILESYNC;
872 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
874 nfs_clearcommit(vp->v_mount);
877 nfsstats.biocache_writes++;
878 lbn = uio->uio_offset / biosize;
879 on = uio->uio_offset & (biosize-1);
880 loffset = uio->uio_offset - on;
881 n = min((unsigned)(biosize - on), uio->uio_resid);
884 * Handle direct append and file extension cases, calculate
885 * unaligned buffer size.
888 if (uio->uio_offset == np->n_size && n) {
890 * Get the buffer (in its pre-append state to maintain
891 * B_CACHE if it was previously set). Resize the
892 * nfsnode after we have locked the buffer to prevent
893 * readers from reading garbage.
896 bp = nfs_getcacheblk(vp, loffset, bcount, td);
901 np->n_size = uio->uio_offset + n;
902 np->n_flag |= NLMODIFIED;
903 vnode_pager_setsize(vp, np->n_size);
905 save = bp->b_flags & B_CACHE;
907 allocbuf(bp, bcount);
912 * Obtain the locked cache block first, and then
913 * adjust the file's size as appropriate.
916 if (loffset + bcount < np->n_size) {
917 if (loffset + biosize < np->n_size)
920 bcount = np->n_size - loffset;
922 bp = nfs_getcacheblk(vp, loffset, bcount, td);
923 if (uio->uio_offset + n > np->n_size) {
924 np->n_size = uio->uio_offset + n;
925 np->n_flag |= NLMODIFIED;
926 vnode_pager_setsize(vp, np->n_size);
936 * Issue a READ if B_CACHE is not set. In special-append
937 * mode, B_CACHE is based on the buffer prior to the write
938 * op and is typically set, avoiding the read. If a read
939 * is required in special append mode, the server will
940 * probably send us a short-read since we extended the file
941 * on our end, resulting in b_resid == 0 and, thusly,
942 * B_CACHE getting set.
944 * We can also avoid issuing the read if the write covers
945 * the entire buffer. We have to make sure the buffer state
946 * is reasonable in this case since we will not be initiating
947 * I/O. See the comments in kern/vfs_bio.c's getblk() for
950 * B_CACHE may also be set due to the buffer being cached
954 if (on == 0 && n == bcount) {
955 bp->b_flags |= B_CACHE;
956 bp->b_flags &= ~(B_ERROR | B_INVAL);
959 if ((bp->b_flags & B_CACHE) == 0) {
960 bp->b_flags |= B_READ;
961 vfs_busy_pages(bp, 0);
962 error = nfs_doio(vp, &bp->b_bio2, td);
972 np->n_flag |= NLMODIFIED;
975 * If dirtyend exceeds file size, chop it down. This should
976 * not normally occur but there is an append race where it
977 * might occur XXX, so we log it.
979 * If the chopping creates a reverse-indexed or degenerate
980 * situation with dirtyoff/end, we 0 both of them.
983 if (bp->b_dirtyend > bcount) {
984 printf("NFS append race @%08llx:%d\n",
985 bp->b_bio2.bio_offset,
986 bp->b_dirtyend - bcount);
987 bp->b_dirtyend = bcount;
990 if (bp->b_dirtyoff >= bp->b_dirtyend)
991 bp->b_dirtyoff = bp->b_dirtyend = 0;
994 * If the new write will leave a contiguous dirty
995 * area, just update the b_dirtyoff and b_dirtyend,
996 * otherwise force a write rpc of the old dirty area.
998 * While it is possible to merge discontiguous writes due to
999 * our having a B_CACHE buffer ( and thus valid read data
1000 * for the hole), we don't because it could lead to
1001 * significant cache coherency problems with multiple clients,
1002 * especially if locking is implemented later on.
1004 * as an optimization we could theoretically maintain
1005 * a linked list of discontinuous areas, but we would still
1006 * have to commit them separately so there isn't much
1007 * advantage to it except perhaps a bit of asynchronization.
1010 if (bp->b_dirtyend > 0 &&
1011 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1012 if (VOP_BWRITE(vp, bp) == EINTR) {
1020 * Check for valid write lease and get one as required.
1021 * In case getblk() and/or bwrite() delayed us.
1023 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
1024 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
1026 error = nqnfs_getlease(vp, ND_WRITE, td);
1027 } while (error == NQNFS_EXPIRED);
1032 if (np->n_lrev != np->n_brev ||
1033 (np->n_flag & NQNFSNONCACHE)) {
1035 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1038 np->n_brev = np->n_lrev;
1043 error = uiomove((char *)bp->b_data + on, n, uio);
1046 * Since this block is being modified, it must be written
1047 * again and not just committed. Since write clustering does
1048 * not work for the stage 1 data write, only the stage 2
1049 * commit rpc, we have to clear B_CLUSTEROK as well.
1051 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1054 bp->b_flags |= B_ERROR;
1060 * Only update dirtyoff/dirtyend if not a degenerate
1064 if (bp->b_dirtyend > 0) {
1065 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1066 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1068 bp->b_dirtyoff = on;
1069 bp->b_dirtyend = on + n;
1071 vfs_bio_set_validclean(bp, on, n);
1074 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1075 * filesystem). XXX also use for loopback NFS mounts.
1077 if (ioflag & IO_NOWDRAIN)
1078 bp->b_flags |= B_NOWDRAIN;
1081 * If the lease is non-cachable or IO_SYNC do bwrite().
1083 * IO_INVAL appears to be unused. The idea appears to be
1084 * to turn off caching in this case. Very odd. XXX
1086 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1087 if (ioflag & IO_INVAL)
1088 bp->b_flags |= B_NOCACHE;
1089 error = VOP_BWRITE(vp, bp);
1092 if (np->n_flag & NQNFSNONCACHE) {
1093 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1097 } else if ((n + on) == biosize &&
1098 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1099 bp->b_flags |= B_ASYNC;
1100 nfs_writebp(bp, 0, 0);
1104 } while (uio->uio_resid > 0 && n > 0);
1107 nfs_rsunlock(np, td);
1113 * Get an nfs cache block.
1115 * Allocate a new one if the block isn't currently in the cache
1116 * and return the block marked busy. If the calling process is
1117 * interrupted by a signal for an interruptible mount point, return
1120 * The caller must carefully deal with the possible B_INVAL state of
1121 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1122 * indirectly), so synchronous reads can be issued without worrying about
1123 * the B_INVAL state. We have to be a little more careful when dealing
1124 * with writes (see comments in nfs_write()) when extending a file past
1128 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
1132 struct nfsmount *nmp;
1137 if (nmp->nm_flag & NFSMNT_INT) {
1138 bp = getblk(vp, loffset, size, PCATCH, 0);
1139 while (bp == NULL) {
1140 if (nfs_sigintr(nmp, (struct nfsreq *)0, td))
1142 bp = getblk(vp, loffset, size, 0, 2 * hz);
1145 bp = getblk(vp, loffset, size, 0, 0);
1149 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets
1150 * now, no translation is necessary.
1152 bp->b_bio2.bio_offset = loffset;
1157 * Flush and invalidate all dirty buffers. If another process is already
1158 * doing the flush, just wait for completion.
1161 nfs_vinvalbuf(struct vnode *vp, int flags,
1162 struct thread *td, int intrflg)
1164 struct nfsnode *np = VTONFS(vp);
1165 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1166 int error = 0, slpflag, slptimeo;
1168 if (vp->v_flag & VRECLAIMED)
1171 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1181 * First wait for any other process doing a flush to complete.
1183 while (np->n_flag & NFLUSHINPROG) {
1184 np->n_flag |= NFLUSHWANT;
1185 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1186 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td))
1191 * Now, flush as required.
1193 np->n_flag |= NFLUSHINPROG;
1194 error = vinvalbuf(vp, flags, td, slpflag, 0);
1196 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) {
1197 np->n_flag &= ~NFLUSHINPROG;
1198 if (np->n_flag & NFLUSHWANT) {
1199 np->n_flag &= ~NFLUSHWANT;
1200 wakeup((caddr_t)&np->n_flag);
1204 error = vinvalbuf(vp, flags, td, 0, slptimeo);
1206 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
1207 if (np->n_flag & NFLUSHWANT) {
1208 np->n_flag &= ~NFLUSHWANT;
1209 wakeup((caddr_t)&np->n_flag);
1215 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1216 * This is mainly to avoid queueing async I/O requests when the nfsiods
1217 * are all hung on a dead server.
1219 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1220 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1223 nfs_asyncio(struct vnode *vp, struct bio *bio, struct thread *td)
1225 struct buf *bp = bio->bio_buf;
1226 struct nfsmount *nmp;
1234 * If no async daemons then return EIO to force caller to run the rpc
1237 if (nfs_numasync == 0)
1240 KKASSERT(vp->v_tag == VT_NFS);
1241 nmp = VFSTONFS(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_bioqiods > 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] = NULL;
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_bioqiods > 0) {
1285 ("nfs_asyncio: %d iods are already processing mount %p\n",
1286 nmp->nm_bioqiods, 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_bioqlen >= 2*nfs_numasync) {
1302 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1303 nmp->nm_bioqwant = TRUE;
1304 error = tsleep(&nmp->nm_bioq, slpflag,
1305 "nfsaio", slptimeo);
1307 if (nfs_sigintr(nmp, NULL, td))
1309 if (slpflag == PCATCH) {
1315 * We might have lost our iod while sleeping,
1316 * so check and loop if nescessary.
1318 if (nmp->nm_bioqiods == 0) {
1320 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1327 * The passed bio's buffer is not necessary associated with
1328 * the NFS vnode it is being written to. Store the NFS vnode
1329 * in the BIO driver info.
1331 bio->bio_driver_info = vp;
1332 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
1338 * All the iods are busy on other mounts, so return EIO to
1339 * force the caller to process the i/o synchronously.
1341 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1346 * Do an I/O operation to/from a cache block. This may be called
1347 * synchronously or from an nfsiod. The BIO is normalized for DEV_BSIZE.
1349 * NOTE! TD MIGHT BE NULL
1352 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1354 struct buf *bp = bio->bio_buf;
1357 struct nfsmount *nmp;
1358 int error = 0, iomode, must_commit = 0;
1362 KKASSERT(vp->v_tag == VT_NFS);
1364 nmp = VFSTONFS(vp->v_mount);
1366 uiop->uio_iov = &io;
1367 uiop->uio_iovcnt = 1;
1368 uiop->uio_segflg = UIO_SYSSPACE;
1372 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1373 * do this here so we do not have to do it in all the code that
1376 bp->b_flags &= ~(B_ERROR | B_INVAL);
1378 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1381 * Historically, paging was done with physio, but no more.
1383 if (bp->b_flags & B_PHYS) {
1385 * ...though reading /dev/drum still gets us here.
1387 io.iov_len = uiop->uio_resid = bp->b_bcount;
1388 /* mapping was done by vmapbuf() */
1389 io.iov_base = bp->b_data;
1390 uiop->uio_offset = bio->bio_offset;
1391 if (bp->b_flags & B_READ) {
1392 uiop->uio_rw = UIO_READ;
1393 nfsstats.read_physios++;
1394 error = nfs_readrpc(vp, uiop);
1398 iomode = NFSV3WRITE_DATASYNC;
1399 uiop->uio_rw = UIO_WRITE;
1400 nfsstats.write_physios++;
1401 error = nfs_writerpc(vp, uiop, &iomode, &com);
1404 bp->b_flags |= B_ERROR;
1405 bp->b_error = error;
1407 } else if (bp->b_flags & B_READ) {
1408 io.iov_len = uiop->uio_resid = bp->b_bcount;
1409 io.iov_base = bp->b_data;
1410 uiop->uio_rw = UIO_READ;
1412 switch (vp->v_type) {
1414 uiop->uio_offset = bio->bio_offset;
1415 nfsstats.read_bios++;
1416 error = nfs_readrpc(vp, uiop);
1419 if (uiop->uio_resid) {
1421 * If we had a short read with no error, we must have
1422 * hit a file hole. We should zero-fill the remainder.
1423 * This can also occur if the server hits the file EOF.
1425 * Holes used to be able to occur due to pending
1426 * writes, but that is not possible any longer.
1428 int nread = bp->b_bcount - uiop->uio_resid;
1429 int left = uiop->uio_resid;
1432 bzero((char *)bp->b_data + nread, left);
1433 uiop->uio_resid = 0;
1436 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1437 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1438 NQNFS_CKINVALID(vp, np, ND_READ) &&
1439 np->n_lrev != np->n_brev) ||
1440 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1441 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1442 uprintf("Process killed due to text file modification\n");
1443 psignal(td->td_proc, SIGKILL);
1447 uiop->uio_offset = 0;
1448 nfsstats.readlink_bios++;
1449 error = nfs_readlinkrpc(vp, uiop);
1452 nfsstats.readdir_bios++;
1453 uiop->uio_offset = bio->bio_offset;
1454 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1455 error = nfs_readdirplusrpc(vp, uiop);
1456 if (error == NFSERR_NOTSUPP)
1457 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1459 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1460 error = nfs_readdirrpc(vp, uiop);
1462 * end-of-directory sets B_INVAL but does not generate an
1465 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1466 bp->b_flags |= B_INVAL;
1469 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1473 bp->b_flags |= B_ERROR;
1474 bp->b_error = error;
1478 * If we only need to commit, try to commit
1480 if (bp->b_flags & B_NEEDCOMMIT) {
1484 off = bio->bio_offset + bp->b_dirtyoff;
1485 retv = nfs_commit(vp, off,
1486 bp->b_dirtyend - bp->b_dirtyoff, td);
1488 bp->b_dirtyoff = bp->b_dirtyend = 0;
1489 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1494 if (retv == NFSERR_STALEWRITEVERF) {
1495 nfs_clearcommit(vp->v_mount);
1500 * Setup for actual write
1503 if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1504 bp->b_dirtyend = np->n_size - bio->bio_offset;
1506 if (bp->b_dirtyend > bp->b_dirtyoff) {
1507 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1509 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1510 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1511 uiop->uio_rw = UIO_WRITE;
1512 nfsstats.write_bios++;
1514 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1515 iomode = NFSV3WRITE_UNSTABLE;
1517 iomode = NFSV3WRITE_FILESYNC;
1519 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1522 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1523 * to cluster the buffers needing commit. This will allow
1524 * the system to submit a single commit rpc for the whole
1525 * cluster. We can do this even if the buffer is not 100%
1526 * dirty (relative to the NFS blocksize), so we optimize the
1527 * append-to-file-case.
1529 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1530 * cleared because write clustering only works for commit
1531 * rpc's, not for the data portion of the write).
1534 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1535 bp->b_flags |= B_NEEDCOMMIT;
1536 if (bp->b_dirtyoff == 0
1537 && bp->b_dirtyend == bp->b_bcount)
1538 bp->b_flags |= B_CLUSTEROK;
1540 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1544 * For an interrupted write, the buffer is still valid
1545 * and the write hasn't been pushed to the server yet,
1546 * so we can't set B_ERROR and report the interruption
1547 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1548 * is not relevant, so the rpc attempt is essentially
1549 * a noop. For the case of a V3 write rpc not being
1550 * committed to stable storage, the block is still
1551 * dirty and requires either a commit rpc or another
1552 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1553 * the block is reused. This is indicated by setting
1554 * the B_DELWRI and B_NEEDCOMMIT flags.
1556 * If the buffer is marked B_PAGING, it does not reside on
1557 * the vp's paging queues so we cannot call bdirty(). The
1558 * bp in this case is not an NFS cache block so we should
1562 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1564 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1565 if ((bp->b_flags & B_PAGING) == 0) {
1567 bp->b_flags &= ~B_DONE;
1569 if (error && (bp->b_flags & B_ASYNC) == 0)
1570 bp->b_flags |= B_EINTR;
1574 bp->b_flags |= B_ERROR;
1575 bp->b_error = np->n_error = error;
1576 np->n_flag |= NWRITEERR;
1578 bp->b_dirtyoff = bp->b_dirtyend = 0;
1586 bp->b_resid = uiop->uio_resid;
1588 nfs_clearcommit(vp->v_mount);
1594 * Used to aid in handling ftruncate() operations on the NFS client side.
1595 * Truncation creates a number of special problems for NFS. We have to
1596 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1597 * we have to properly handle VM pages or (potentially dirty) buffers
1598 * that straddle the truncation point.
1602 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1604 struct nfsnode *np = VTONFS(vp);
1605 u_quad_t tsize = np->n_size;
1606 int biosize = vp->v_mount->mnt_stat.f_iosize;
1611 if (np->n_size < tsize) {
1618 * vtruncbuf() doesn't get the buffer overlapping the
1619 * truncation point. We may have a B_DELWRI and/or B_CACHE
1620 * buffer that now needs to be truncated.
1622 error = vtruncbuf(vp, td, nsize, biosize);
1623 lbn = nsize / biosize;
1624 bufsize = nsize & (biosize - 1);
1625 loffset = nsize - bufsize;
1626 bp = nfs_getcacheblk(vp, loffset, bufsize, td);
1627 if (bp->b_dirtyoff > bp->b_bcount)
1628 bp->b_dirtyoff = bp->b_bcount;
1629 if (bp->b_dirtyend > bp->b_bcount)
1630 bp->b_dirtyend = bp->b_bcount;
1631 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1634 vnode_pager_setsize(vp, nsize);