xref: /openbsd-src/sys/nfs/nfs_bio.c (revision 43003dfe3ad45d1698bed8a37f2b0f5b14f20d4f)
1 /*	$OpenBSD: nfs_bio.c,v 1.67 2009/09/02 18:20:54 thib Exp $	*/
2 /*	$NetBSD: nfs_bio.c,v 1.25.4.2 1996/07/08 20:47:04 jtc Exp $	*/
3 
4 /*
5  * Copyright (c) 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * Rick Macklem at The University of Guelph.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
36  */
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/resourcevar.h>
41 #include <sys/signalvar.h>
42 #include <sys/proc.h>
43 #include <sys/buf.h>
44 #include <sys/vnode.h>
45 #include <sys/mount.h>
46 #include <sys/kernel.h>
47 #include <sys/namei.h>
48 #include <sys/queue.h>
49 #include <sys/time.h>
50 
51 #include <uvm/uvm_extern.h>
52 
53 #include <nfs/rpcv2.h>
54 #include <nfs/nfsproto.h>
55 #include <nfs/nfs.h>
56 #include <nfs/nfsmount.h>
57 #include <nfs/nfsnode.h>
58 #include <nfs/nfs_var.h>
59 
60 extern int nfs_numasync;
61 extern struct nfsstats nfsstats;
62 struct nfs_bufqhead nfs_bufq;
63 uint32_t nfs_bufqmax, nfs_bufqlen;
64 
65 /*
66  * Vnode op for read using bio
67  * Any similarity to readip() is purely coincidental
68  */
69 int
70 nfs_bioread(vp, uio, ioflag, cred)
71 	struct vnode *vp;
72 	struct uio *uio;
73 	int ioflag;
74 	struct ucred *cred;
75 {
76 	struct nfsnode *np = VTONFS(vp);
77 	int biosize, diff;
78 	struct buf *bp = NULL, *rabp;
79 	struct vattr vattr;
80 	struct proc *p;
81 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
82 	daddr64_t lbn, bn, rabn;
83 	caddr_t baddr;
84 	int got_buf = 0, nra, error = 0, n = 0, on = 0, not_readin;
85 	off_t offdiff;
86 
87 #ifdef DIAGNOSTIC
88 	if (uio->uio_rw != UIO_READ)
89 		panic("nfs_read mode");
90 #endif
91 	if (uio->uio_resid == 0)
92 		return (0);
93 	if (uio->uio_offset < 0)
94 		return (EINVAL);
95 	p = uio->uio_procp;
96 	if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3)
97 		(void)nfs_fsinfo(nmp, vp, cred, p);
98 	biosize = nmp->nm_rsize;
99 	/*
100 	 * For nfs, cache consistency can only be maintained approximately.
101 	 * Although RFC1094 does not specify the criteria, the following is
102 	 * believed to be compatible with the reference port.
103 	 * For nfs:
104 	 * If the file's modify time on the server has changed since the
105 	 * last read rpc or you have written to the file,
106 	 * you may have lost data cache consistency with the
107 	 * server, so flush all of the file's data out of the cache.
108 	 * Then force a getattr rpc to ensure that you have up to date
109 	 * attributes.
110 	 */
111 	if (np->n_flag & NMODIFIED) {
112 		NFS_INVALIDATE_ATTRCACHE(np);
113 		error = VOP_GETATTR(vp, &vattr, cred, p);
114 		if (error)
115 			return (error);
116 		np->n_mtime = vattr.va_mtime;
117 	} else {
118 		error = VOP_GETATTR(vp, &vattr, cred, p);
119 		if (error)
120 			return (error);
121 		if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
122 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p);
123 			if (error)
124 				return (error);
125 			np->n_mtime = vattr.va_mtime;
126 		}
127 	}
128 
129 	/*
130 	 * update the cache read creds for this vnode
131 	 */
132 	if (np->n_rcred)
133 		crfree(np->n_rcred);
134 	np->n_rcred = cred;
135 	crhold(cred);
136 
137 	do {
138 	    if ((vp->v_flag & VROOT) && vp->v_type == VLNK) {
139 		    return (nfs_readlinkrpc(vp, uio, cred));
140 	    }
141 	    baddr = (caddr_t)0;
142 	    switch (vp->v_type) {
143 	    case VREG:
144 		nfsstats.biocache_reads++;
145 		lbn = uio->uio_offset / biosize;
146 		on = uio->uio_offset & (biosize - 1);
147 		bn = lbn * (biosize / DEV_BSIZE);
148 		not_readin = 1;
149 
150 		/*
151 		 * Start the read ahead(s), as required.
152 		 */
153 		if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
154 		    for (nra = 0; nra < nmp->nm_readahead &&
155 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
156 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
157 			if (!incore(vp, rabn)) {
158 			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
159 			    if (!rabp)
160 				return (EINTR);
161 			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
162 				rabp->b_flags |= (B_READ | B_ASYNC);
163 				if (nfs_asyncio(rabp)) {
164 				    rabp->b_flags |= B_INVAL;
165 				    brelse(rabp);
166 				}
167 			    } else
168 				brelse(rabp);
169 			}
170 		    }
171 		}
172 
173 again:
174 		bp = nfs_getcacheblk(vp, bn, biosize, p);
175 		if (!bp)
176 			return (EINTR);
177 		got_buf = 1;
178 		if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
179 			bp->b_flags |= B_READ;
180 			not_readin = 0;
181 			error = nfs_doio(bp, p);
182 			if (error) {
183 			    brelse(bp);
184 			    return (error);
185 			}
186 		}
187 		n = min((unsigned)(biosize - on), uio->uio_resid);
188 		offdiff = np->n_size - uio->uio_offset;
189 		if (offdiff < (off_t)n)
190 			n = (int)offdiff;
191 		if (not_readin && n > 0) {
192 			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
193 				bp->b_flags |= B_INVAFTERWRITE;
194 				if (bp->b_dirtyend > 0) {
195 				    if ((bp->b_flags & B_DELWRI) == 0)
196 					panic("nfsbioread");
197 				    if (VOP_BWRITE(bp) == EINTR)
198 					return (EINTR);
199 				} else
200 				    brelse(bp);
201 				goto again;
202 			}
203 		}
204 		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
205 		if (diff < n)
206 			n = diff;
207 		break;
208 	    case VLNK:
209 		nfsstats.biocache_readlinks++;
210 		bp = nfs_getcacheblk(vp, 0, NFS_MAXPATHLEN, p);
211 		if (!bp)
212 			return (EINTR);
213 		if ((bp->b_flags & B_DONE) == 0) {
214 			bp->b_flags |= B_READ;
215 			error = nfs_doio(bp, p);
216 			if (error) {
217 				brelse(bp);
218 				return (error);
219 			}
220 		}
221 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
222 		got_buf = 1;
223 		on = 0;
224 		break;
225 	    default:
226 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
227 		break;
228 	    }
229 
230 	    if (n > 0) {
231 		if (!baddr)
232 			baddr = bp->b_data;
233 		error = uiomove(baddr + on, (int)n, uio);
234 	    }
235 	    switch (vp->v_type) {
236 	    case VREG:
237 		break;
238 	    case VLNK:
239 		n = 0;
240 		break;
241 	    default:
242 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
243 	    }
244 	    if (got_buf)
245 		brelse(bp);
246 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
247 	return (error);
248 }
249 
250 /*
251  * Vnode op for write using bio
252  */
253 int
254 nfs_write(v)
255 	void *v;
256 {
257 	struct vop_write_args *ap = v;
258 	int biosize;
259 	struct uio *uio = ap->a_uio;
260 	struct proc *p = uio->uio_procp;
261 	struct vnode *vp = ap->a_vp;
262 	struct nfsnode *np = VTONFS(vp);
263 	struct ucred *cred = ap->a_cred;
264 	int ioflag = ap->a_ioflag;
265 	struct buf *bp;
266 	struct vattr vattr;
267 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
268 	daddr64_t lbn, bn;
269 	int n, on, error = 0, extended = 0, wrotedta = 0, truncated = 0;
270 
271 #ifdef DIAGNOSTIC
272 	if (uio->uio_rw != UIO_WRITE)
273 		panic("nfs_write mode");
274 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
275 		panic("nfs_write proc");
276 #endif
277 	if (vp->v_type != VREG)
278 		return (EIO);
279 	if (np->n_flag & NWRITEERR) {
280 		np->n_flag &= ~NWRITEERR;
281 		return (np->n_error);
282 	}
283 	if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3)
284 		(void)nfs_fsinfo(nmp, vp, cred, p);
285 	if (ioflag & (IO_APPEND | IO_SYNC)) {
286 		if (np->n_flag & NMODIFIED) {
287 			NFS_INVALIDATE_ATTRCACHE(np);
288 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p);
289 			if (error)
290 				return (error);
291 		}
292 		if (ioflag & IO_APPEND) {
293 			NFS_INVALIDATE_ATTRCACHE(np);
294 			error = VOP_GETATTR(vp, &vattr, cred, p);
295 			if (error)
296 				return (error);
297 			uio->uio_offset = np->n_size;
298 		}
299 	}
300 	if (uio->uio_offset < 0)
301 		return (EINVAL);
302 	if (uio->uio_resid == 0)
303 		return (0);
304 	/*
305 	 * Maybe this should be above the vnode op call, but so long as
306 	 * file servers have no limits, i don't think it matters
307 	 */
308 	if (p && uio->uio_offset + uio->uio_resid >
309 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
310 		psignal(p, SIGXFSZ);
311 		return (EFBIG);
312 	}
313 
314 	/*
315 	 * update the cache write creds for this node.
316 	 */
317 	if (np->n_wcred)
318 		crfree(np->n_wcred);
319 	np->n_wcred = cred;
320 	crhold(cred);
321 
322 	/*
323 	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
324 	 * will be the same size within a filesystem. nfs_writerpc will
325 	 * still use nm_wsize when sizing the rpc's.
326 	 */
327 	biosize = nmp->nm_rsize;
328 	do {
329 
330 		/*
331 		 * XXX make sure we aren't cached in the VM page cache
332 		 */
333 		uvm_vnp_uncache(vp);
334 
335 		nfsstats.biocache_writes++;
336 		lbn = uio->uio_offset / biosize;
337 		on = uio->uio_offset & (biosize-1);
338 		n = min((unsigned)(biosize - on), uio->uio_resid);
339 		bn = lbn * (biosize / DEV_BSIZE);
340 again:
341 		bp = nfs_getcacheblk(vp, bn, biosize, p);
342 		if (!bp)
343 			return (EINTR);
344 		np->n_flag |= NMODIFIED;
345 		if (uio->uio_offset + n > np->n_size) {
346 			np->n_size = uio->uio_offset + n;
347 			uvm_vnp_setsize(vp, (u_long)np->n_size);
348 			extended = 1;
349 		} else if (uio->uio_offset + n < np->n_size)
350 			truncated = 1;
351 
352 		/*
353 		 * If the new write will leave a contiguous dirty
354 		 * area, just update the b_dirtyoff and b_dirtyend,
355 		 * otherwise force a write rpc of the old dirty area.
356 		 */
357 		if (bp->b_dirtyend > 0 &&
358 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
359 			bp->b_proc = p;
360 			if (VOP_BWRITE(bp) == EINTR)
361 				return (EINTR);
362 			goto again;
363 		}
364 
365 		error = uiomove((char *)bp->b_data + on, n, uio);
366 		if (error) {
367 			bp->b_flags |= B_ERROR;
368 			brelse(bp);
369 			return (error);
370 		}
371 		if (bp->b_dirtyend > 0) {
372 			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
373 			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
374 		} else {
375 			bp->b_dirtyoff = on;
376 			bp->b_dirtyend = on + n;
377 		}
378 		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
379 		    bp->b_validoff > bp->b_dirtyend) {
380 			bp->b_validoff = bp->b_dirtyoff;
381 			bp->b_validend = bp->b_dirtyend;
382 		} else {
383 			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
384 			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
385 		}
386 
387 		wrotedta = 1;
388 
389 		/*
390 		 * Since this block is being modified, it must be written
391 		 * again and not just committed.
392 		 */
393 
394 		if (NFS_ISV3(vp)) {
395 			rw_enter_write(&np->n_commitlock);
396 			if (bp->b_flags & B_NEEDCOMMIT) {
397 				bp->b_flags &= ~B_NEEDCOMMIT;
398 				nfs_del_tobecommitted_range(vp, bp);
399 			}
400 			nfs_del_committed_range(vp, bp);
401 			rw_exit_write(&np->n_commitlock);
402 		} else
403 			bp->b_flags &= ~B_NEEDCOMMIT;
404 
405 		if (ioflag & IO_SYNC) {
406 			bp->b_proc = p;
407 			error = VOP_BWRITE(bp);
408 			if (error)
409 				return (error);
410 		} else if ((n + on) == biosize) {
411 			bp->b_proc = NULL;
412 			bp->b_flags |= B_ASYNC;
413 			(void)nfs_writebp(bp, 0);
414 		} else {
415 			bdwrite(bp);
416 		}
417 	} while (uio->uio_resid > 0 && n > 0);
418 
419 	if (wrotedta)
420 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0) |
421 		    (truncated ? NOTE_TRUNCATE : 0));
422 
423 	return (0);
424 }
425 
426 /*
427  * Get an nfs cache block.
428  * Allocate a new one if the block isn't currently in the cache
429  * and return the block marked busy. If the calling process is
430  * interrupted by a signal for an interruptible mount point, return
431  * NULL.
432  */
433 struct buf *
434 nfs_getcacheblk(vp, bn, size, p)
435 	struct vnode *vp;
436 	daddr64_t bn;
437 	int size;
438 	struct proc *p;
439 {
440 	struct buf *bp;
441 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
442 
443 	if (nmp->nm_flag & NFSMNT_INT) {
444 		bp = getblk(vp, bn, size, PCATCH, 0);
445 		while (bp == NULL) {
446 			if (nfs_sigintr(nmp, NULL, p))
447 				return (NULL);
448 			bp = getblk(vp, bn, size, 0, 2 * hz);
449 		}
450 	} else
451 		bp = getblk(vp, bn, size, 0, 0);
452 	return (bp);
453 }
454 
455 /*
456  * Flush and invalidate all dirty buffers. If another process is already
457  * doing the flush, just wait for completion.
458  */
459 int
460 nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred, struct proc *p)
461 {
462 	struct nfsmount		*nmp= VFSTONFS(vp->v_mount);
463 	struct nfsnode		*np = VTONFS(vp);
464 	int			 error, sintr, stimeo;
465 
466 	error = sintr = stimeo = 0;
467 
468 	if (ISSET(nmp->nm_flag, NFSMNT_INT)) {
469 		sintr = PCATCH;
470 		stimeo = 2 * hz;
471 	}
472 
473 	/* First wait for any other process doing a flush to complete. */
474 	while (np->n_flag & NFLUSHINPROG) {
475 		np->n_flag |= NFLUSHWANT;
476 		error = tsleep(&np->n_flag, PRIBIO|sintr, "nfsvinval", stimeo);
477 		if (error && sintr && nfs_sigintr(nmp, NULL, p))
478 			return (EINTR);
479 	}
480 
481 	/* Now, flush as required. */
482 	np->n_flag |= NFLUSHINPROG;
483 	error = vinvalbuf(vp, flags, cred, p, sintr, 0);
484 	while (error) {
485 		if (sintr && nfs_sigintr(nmp, NULL, p)) {
486 			np->n_flag &= ~NFLUSHINPROG;
487 			if (np->n_flag & NFLUSHWANT) {
488 				np->n_flag &= ~NFLUSHWANT;
489 				wakeup(&np->n_flag);
490 			}
491 			return (EINTR);
492 		}
493 		error = vinvalbuf(vp, flags, cred, p, 0, stimeo);
494 	}
495 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
496 	if (np->n_flag & NFLUSHWANT) {
497 		np->n_flag &= ~NFLUSHWANT;
498 		wakeup(&np->n_flag);
499 	}
500 	return (0);
501 }
502 
503 /*
504  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
505  * This is mainly to avoid queueing async I/O requests when the nfsiods
506  * are all hung on a dead server.
507  */
508 int
509 nfs_asyncio(bp)
510 	struct buf *bp;
511 {
512 	if (nfs_numasync == 0)
513 		goto out;
514 
515 	if (nfs_bufqlen > nfs_bufqmax)
516 		goto out; /* too many bufs in use, force sync */
517 
518 	if ((bp->b_flags & B_READ) == 0) {
519 		bp->b_flags |= B_WRITEINPROG;
520 	}
521 
522 	TAILQ_INSERT_TAIL(&nfs_bufq, bp, b_freelist);
523 	nfs_bufqlen++;
524 
525 	wakeup_one(&nfs_bufq);
526 	return (0);
527 
528 out:
529 	nfsstats.forcedsync++;
530 	return (EIO);
531 }
532 
533 /*
534  * Do an I/O operation to/from a cache block. This may be called
535  * synchronously or from an nfsiod.
536  */
537 int
538 nfs_doio(bp, p)
539 	struct buf *bp;
540 	struct proc *p;
541 {
542 	struct uio *uiop;
543 	struct vnode *vp;
544 	struct nfsnode *np;
545 	struct nfsmount *nmp;
546 	int s, error = 0, diff, len, iomode, must_commit = 0;
547 	struct uio uio;
548 	struct iovec io;
549 
550 	vp = bp->b_vp;
551 	np = VTONFS(vp);
552 	nmp = VFSTONFS(vp->v_mount);
553 	uiop = &uio;
554 	uiop->uio_iov = &io;
555 	uiop->uio_iovcnt = 1;
556 	uiop->uio_segflg = UIO_SYSSPACE;
557 	uiop->uio_procp = p;
558 
559 	/*
560 	 * Historically, paging was done with physio, but no more.
561 	 */
562 	if (bp->b_flags & B_PHYS) {
563 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
564 	    /* mapping was done by vmapbuf() */
565 	    io.iov_base = bp->b_data;
566 	    uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
567 	    if (bp->b_flags & B_READ) {
568 		uiop->uio_rw = UIO_READ;
569 		nfsstats.read_physios++;
570 		error = nfs_readrpc(vp, uiop);
571 	    } else {
572 		iomode = NFSV3WRITE_DATASYNC;
573 		uiop->uio_rw = UIO_WRITE;
574 		nfsstats.write_physios++;
575 		error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
576 	    }
577 	    if (error) {
578 		bp->b_flags |= B_ERROR;
579 		bp->b_error = error;
580 	    }
581 	} else if (bp->b_flags & B_READ) {
582 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
583 	    io.iov_base = bp->b_data;
584 	    uiop->uio_rw = UIO_READ;
585 	    switch (vp->v_type) {
586 	    case VREG:
587 		uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
588 		nfsstats.read_bios++;
589 		bcstats.pendingreads++;
590 		bcstats.numreads++;
591 		error = nfs_readrpc(vp, uiop);
592 		if (!error) {
593 		    bp->b_validoff = 0;
594 		    if (uiop->uio_resid) {
595 			/*
596 			 * If len > 0, there is a hole in the file and
597 			 * no writes after the hole have been pushed to
598 			 * the server yet.
599 			 * Just zero fill the rest of the valid area.
600 			 */
601 			diff = bp->b_bcount - uiop->uio_resid;
602 			len = np->n_size - ((((off_t)bp->b_blkno) << DEV_BSHIFT)
603 				+ diff);
604 			if (len > 0) {
605 			    len = min(len, uiop->uio_resid);
606 			    bzero((char *)bp->b_data + diff, len);
607 			    bp->b_validend = diff + len;
608 			} else
609 			    bp->b_validend = diff;
610 		    } else
611 			bp->b_validend = bp->b_bcount;
612 		}
613 		if (p && (vp->v_flag & VTEXT) &&
614 		    (timespeccmp(&np->n_mtime, &np->n_vattr.va_mtime, !=))) {
615 			uprintf("Process killed due to text file modification\n");
616 			psignal(p, SIGKILL);
617 		}
618 		break;
619 	    case VLNK:
620 		uiop->uio_offset = (off_t)0;
621 		nfsstats.readlink_bios++;
622 		bcstats.pendingreads++;
623 		bcstats.numreads++;
624 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
625 		break;
626 	    default:
627 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
628 		break;
629 	    };
630 	    if (error) {
631 		bp->b_flags |= B_ERROR;
632 		bp->b_error = error;
633 	    }
634 	} else {
635 	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
636 		- bp->b_dirtyoff;
637 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE
638 		+ bp->b_dirtyoff;
639 	    io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
640 	    uiop->uio_rw = UIO_WRITE;
641 	    nfsstats.write_bios++;
642 	    bcstats.pendingwrites++;
643 	    bcstats.numwrites++;
644 	    if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC)
645 		iomode = NFSV3WRITE_UNSTABLE;
646 	    else
647 		iomode = NFSV3WRITE_FILESYNC;
648 	    bp->b_flags |= B_WRITEINPROG;
649 	    error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
650 
651 	    rw_enter_write(&np->n_commitlock);
652 	    if (!error && iomode == NFSV3WRITE_UNSTABLE) {
653 		bp->b_flags |= B_NEEDCOMMIT;
654 		nfs_add_tobecommitted_range(vp, bp);
655 	    } else {
656 		bp->b_flags &= ~B_NEEDCOMMIT;
657 		nfs_del_committed_range(vp, bp);
658 	    }
659 	    rw_exit_write(&np->n_commitlock);
660 
661 	    bp->b_flags &= ~B_WRITEINPROG;
662 
663 	    /*
664 	     * For an interrupted write, the buffer is still valid and the
665 	     * write hasn't been pushed to the server yet, so we can't set
666 	     * B_ERROR and report the interruption by setting B_EINTR. For
667 	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
668 	     * is essentially a noop.
669 	     * For the case of a V3 write rpc not being committed to stable
670 	     * storage, the block is still dirty and requires either a commit
671 	     * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC
672 	     * before the block is reused. This is indicated by setting the
673 	     * B_DELWRI and B_NEEDCOMMIT flags.
674 	     */
675 	    if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
676 		    s = splbio();
677 		    buf_dirty(bp);
678 		    splx(s);
679 
680 		    if (!(bp->b_flags & B_ASYNC) && error)
681 			    bp->b_flags |= B_EINTR;
682 	    } else {
683 		if (error) {
684 		    bp->b_flags |= B_ERROR;
685 		    bp->b_error = np->n_error = error;
686 		    np->n_flag |= NWRITEERR;
687 		}
688 		bp->b_dirtyoff = bp->b_dirtyend = 0;
689 	    }
690 	}
691 	bp->b_resid = uiop->uio_resid;
692 	if (must_commit)
693 		nfs_clearcommit(vp->v_mount);
694 	s = splbio();
695 	biodone(bp);
696 	splx(s);
697 	return (error);
698 }
699