xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 6ea46cb5e46c49111a6ecf3bcbe3c7e2730fe9f6)
1 /*	$NetBSD: nfs_bio.c,v 1.17 1994/07/20 04:16:01 mycroft Exp $	*/
2 
3 /*
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Rick Macklem at The University of Guelph.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	@(#)nfs_bio.c	8.5 (Berkeley) 1/4/94
39  */
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/resourcevar.h>
44 #include <sys/proc.h>
45 #include <sys/buf.h>
46 #include <sys/vnode.h>
47 #include <sys/trace.h>
48 #include <sys/mount.h>
49 #include <sys/kernel.h>
50 
51 #include <vm/vm.h>
52 
53 #include <nfs/nfsnode.h>
54 #include <nfs/rpcv2.h>
55 #include <nfs/nfsv2.h>
56 #include <nfs/nfs.h>
57 #include <nfs/nfsmount.h>
58 #include <nfs/nqnfs.h>
59 
60 struct buf *incore(), *nfs_getcacheblk();
61 extern struct proc *nfs_iodwant[NFS_MAXASYNCDAEMON];
62 extern int nfs_numasync;
63 
64 /*
65  * Vnode op for read using bio
66  * Any similarity to readip() is purely coincidental
67  */
68 nfs_bioread(vp, uio, ioflag, cred)
69 	register struct vnode *vp;
70 	register struct uio *uio;
71 	int ioflag;
72 	struct ucred *cred;
73 {
74 	register struct nfsnode *np = VTONFS(vp);
75 	register int biosize, diff;
76 	struct buf *bp, *rabp;
77 	struct vattr vattr;
78 	struct proc *p;
79 	struct nfsmount *nmp;
80 	daddr_t lbn, bn, rabn;
81 	caddr_t baddr;
82 	int got_buf, nra, error = 0, n, on, not_readin;
83 
84 #ifdef lint
85 	ioflag = ioflag;
86 #endif /* lint */
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 && vp->v_type != VDIR)
94 		return (EINVAL);
95 	nmp = VFSTONFS(vp->v_mount);
96 	biosize = nmp->nm_rsize;
97 	p = uio->uio_procp;
98 	/*
99 	 * For nfs, cache consistency can only be maintained approximately.
100 	 * Although RFC1094 does not specify the criteria, the following is
101 	 * believed to be compatible with the reference port.
102 	 * For nqnfs, full cache consistency is maintained within the loop.
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 	 * The mount flag NFSMNT_MYWRITE says "Assume that my writes are
111 	 * the ones changing the modify time.
112 	 * NB: This implies that cache data can be read when up to
113 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
114 	 * attributes this could be forced by setting n_attrstamp to 0 before
115 	 * the VOP_GETATTR() call.
116 	 */
117 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
118 		if (np->n_flag & NMODIFIED) {
119 			if ((nmp->nm_flag & NFSMNT_MYWRITE) == 0 ||
120 			     vp->v_type != VREG) {
121 				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
122 					return (error);
123 			}
124 			np->n_attrstamp = 0;
125 			np->n_direofoffset = 0;
126 			if (error = VOP_GETATTR(vp, &vattr, cred, p))
127 				return (error);
128 			np->n_mtime = vattr.va_mtime.ts_sec;
129 		} else {
130 			if (error = VOP_GETATTR(vp, &vattr, cred, p))
131 				return (error);
132 			if (np->n_mtime != vattr.va_mtime.ts_sec) {
133 				np->n_direofoffset = 0;
134 				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
135 					return (error);
136 				np->n_mtime = vattr.va_mtime.ts_sec;
137 			}
138 		}
139 	}
140 	do {
141 
142 	    /*
143 	     * Get a valid lease. If cached data is stale, flush it.
144 	     */
145 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
146 		if (NQNFS_CKINVALID(vp, np, NQL_READ)) {
147 		    do {
148 			error = nqnfs_getlease(vp, NQL_READ, cred, p);
149 		    } while (error == NQNFS_EXPIRED);
150 		    if (error)
151 			return (error);
152 		    if (np->n_lrev != np->n_brev ||
153 			(np->n_flag & NQNFSNONCACHE) ||
154 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
155 			if (vp->v_type == VDIR) {
156 			    np->n_direofoffset = 0;
157 			    cache_purge(vp);
158 			}
159 			if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
160 			    return (error);
161 			np->n_brev = np->n_lrev;
162 		    }
163 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
164 		    np->n_direofoffset = 0;
165 		    cache_purge(vp);
166 		    if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
167 			return (error);
168 		}
169 	    }
170 	    if (np->n_flag & NQNFSNONCACHE) {
171 		switch (vp->v_type) {
172 		case VREG:
173 			error = nfs_readrpc(vp, uio, cred);
174 			break;
175 		case VLNK:
176 			error = nfs_readlinkrpc(vp, uio, cred);
177 			break;
178 		case VDIR:
179 			error = nfs_readdirrpc(vp, uio, cred);
180 			break;
181 		};
182 		return (error);
183 	    }
184 	    baddr = (caddr_t)0;
185 	    switch (vp->v_type) {
186 	    case VREG:
187 		nfsstats.biocache_reads++;
188 		lbn = uio->uio_offset / biosize;
189 		on = uio->uio_offset & (biosize-1);
190 		bn = lbn * (biosize / DEV_BSIZE);
191 		not_readin = 1;
192 
193 		/*
194 		 * Start the read ahead(s), as required.
195 		 */
196 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
197 		    lbn == vp->v_lastr + 1) {
198 		    for (nra = 0; nra < nmp->nm_readahead &&
199 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
200 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
201 			if (!incore(vp, rabn)) {
202 			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
203 			    if (!rabp)
204 				return (EINTR);
205 			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
206 				rabp->b_flags |= (B_READ | B_ASYNC);
207 				if (nfs_asyncio(rabp, cred)) {
208 				    rabp->b_flags |= B_INVAL;
209 				    brelse(rabp);
210 				}
211 			    }
212 			}
213 		    }
214 		}
215 
216 		/*
217 		 * If the block is in the cache and has the required data
218 		 * in a valid region, just copy it out.
219 		 * Otherwise, get the block and write back/read in,
220 		 * as required.
221 		 */
222 		if ((bp = incore(vp, bn)) &&
223 		    (bp->b_flags & (B_BUSY | B_WRITEINPROG)) ==
224 		    (B_BUSY | B_WRITEINPROG))
225 			got_buf = 0;
226 		else {
227 again:
228 			bp = nfs_getcacheblk(vp, bn, biosize, p);
229 			if (!bp)
230 				return (EINTR);
231 			got_buf = 1;
232 			if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
233 				bp->b_flags |= B_READ;
234 				not_readin = 0;
235 				if (error = nfs_doio(bp, cred, p)) {
236 				    brelse(bp);
237 				    return (error);
238 				}
239 			}
240 		}
241 		n = min((unsigned)(biosize - on), uio->uio_resid);
242 		diff = np->n_size - uio->uio_offset;
243 		if (diff < n)
244 			n = diff;
245 		if (not_readin && n > 0) {
246 			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
247 				if (!got_buf) {
248 				    bp = nfs_getcacheblk(vp, bn, biosize, p);
249 				    if (!bp)
250 					return (EINTR);
251 				    got_buf = 1;
252 				}
253 				bp->b_flags |= B_INVAL;
254 				if (bp->b_dirtyend > 0) {
255 				    if ((bp->b_flags & B_DELWRI) == 0)
256 					panic("nfsbioread");
257 				    if (VOP_BWRITE(bp) == EINTR)
258 					return (EINTR);
259 				} else
260 				    brelse(bp);
261 				goto again;
262 			}
263 		}
264 		vp->v_lastr = lbn;
265 		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
266 		if (diff < n)
267 			n = diff;
268 		break;
269 	    case VLNK:
270 		nfsstats.biocache_readlinks++;
271 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
272 		if (!bp)
273 			return (EINTR);
274 		if ((bp->b_flags & B_DONE) == 0) {
275 			bp->b_flags |= B_READ;
276 			if (error = nfs_doio(bp, cred, p)) {
277 				brelse(bp);
278 				return (error);
279 			}
280 		}
281 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
282 		got_buf = 1;
283 		on = 0;
284 		break;
285 	    case VDIR:
286 		nfsstats.biocache_readdirs++;
287 		bn = (daddr_t)uio->uio_offset;
288 		bp = nfs_getcacheblk(vp, bn, NFS_DIRBLKSIZ, p);
289 		if (!bp)
290 			return (EINTR);
291 		if ((bp->b_flags & B_DONE) == 0) {
292 			bp->b_flags |= B_READ;
293 			if (error = nfs_doio(bp, cred, p)) {
294 				brelse(bp);
295 				return (error);
296 			}
297 		}
298 
299 		/*
300 		 * If not eof and read aheads are enabled, start one.
301 		 * (You need the current block first, so that you have the
302 		 *  directory offset cookie of the next block.
303 		 */
304 		rabn = bp->b_blkno;
305 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
306 		    rabn != 0 && rabn != np->n_direofoffset &&
307 		    !incore(vp, rabn)) {
308 			rabp = nfs_getcacheblk(vp, rabn, NFS_DIRBLKSIZ, p);
309 			if (rabp) {
310 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
311 				rabp->b_flags |= (B_READ | B_ASYNC);
312 				if (nfs_asyncio(rabp, cred)) {
313 				    rabp->b_flags |= B_INVAL;
314 				    brelse(rabp);
315 				}
316 			    }
317 			}
318 		}
319 		on = 0;
320 		n = min(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid);
321 		got_buf = 1;
322 		break;
323 	    };
324 
325 	    if (n > 0) {
326 		if (!baddr)
327 			baddr = bp->b_data;
328 		error = uiomove(baddr + on, (int)n, uio);
329 	    }
330 	    switch (vp->v_type) {
331 	    case VREG:
332 		if (n + on == biosize || uio->uio_offset == np->n_size)
333 			bp->b_flags |= B_AGE;
334 		break;
335 	    case VLNK:
336 		n = 0;
337 		break;
338 	    case VDIR:
339 		uio->uio_offset = bp->b_blkno;
340 		break;
341 	    };
342 	    if (got_buf)
343 		brelse(bp);
344 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
345 	return (error);
346 }
347 
348 /*
349  * Vnode op for write using bio
350  */
351 nfs_write(ap)
352 	struct vop_write_args /* {
353 		struct vnode *a_vp;
354 		struct uio *a_uio;
355 		int  a_ioflag;
356 		struct ucred *a_cred;
357 	} */ *ap;
358 {
359 	register int biosize;
360 	register struct uio *uio = ap->a_uio;
361 	struct proc *p = uio->uio_procp;
362 	register struct vnode *vp = ap->a_vp;
363 	struct nfsnode *np = VTONFS(vp);
364 	register struct ucred *cred = ap->a_cred;
365 	int ioflag = ap->a_ioflag;
366 	struct buf *bp;
367 	struct vattr vattr;
368 	struct nfsmount *nmp;
369 	daddr_t lbn, bn;
370 	int n, on, error = 0;
371 
372 #ifdef DIAGNOSTIC
373 	if (uio->uio_rw != UIO_WRITE)
374 		panic("nfs_write mode");
375 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
376 		panic("nfs_write proc");
377 #endif
378 	if (vp->v_type != VREG)
379 		return (EIO);
380 	if (np->n_flag & NWRITEERR) {
381 		np->n_flag &= ~NWRITEERR;
382 		return (np->n_error);
383 	}
384 	if (ioflag & (IO_APPEND | IO_SYNC)) {
385 		if (np->n_flag & NMODIFIED) {
386 			np->n_attrstamp = 0;
387 			if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
388 				return (error);
389 		}
390 		if (ioflag & IO_APPEND) {
391 			np->n_attrstamp = 0;
392 			if (error = VOP_GETATTR(vp, &vattr, cred, p))
393 				return (error);
394 			uio->uio_offset = np->n_size;
395 		}
396 	}
397 	nmp = VFSTONFS(vp->v_mount);
398 	if (uio->uio_offset < 0)
399 		return (EINVAL);
400 	if (uio->uio_resid == 0)
401 		return (0);
402 	/*
403 	 * Maybe this should be above the vnode op call, but so long as
404 	 * file servers have no limits, i don't think it matters
405 	 */
406 	if (p && uio->uio_offset + uio->uio_resid >
407 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
408 		psignal(p, SIGXFSZ);
409 		return (EFBIG);
410 	}
411 	/*
412 	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
413 	 * will be the same size within a filesystem. nfs_writerpc will
414 	 * still use nm_wsize when sizing the rpc's.
415 	 */
416 	biosize = nmp->nm_rsize;
417 	do {
418 
419 		/*
420 		 * XXX make sure we aren't cached in the VM page cache
421 		 */
422 		(void)vnode_pager_uncache(vp);
423 
424 		/*
425 		 * Check for a valid write lease.
426 		 * If non-cachable, just do the rpc
427 		 */
428 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
429 		    NQNFS_CKINVALID(vp, np, NQL_WRITE)) {
430 			do {
431 				error = nqnfs_getlease(vp, NQL_WRITE, cred, p);
432 			} while (error == NQNFS_EXPIRED);
433 			if (error)
434 				return (error);
435 			if (np->n_lrev != np->n_brev ||
436 			    (np->n_flag & NQNFSNONCACHE)) {
437 				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
438 					return (error);
439 				np->n_brev = np->n_lrev;
440 			}
441 		}
442 		if (np->n_flag & NQNFSNONCACHE)
443 			return (nfs_writerpc(vp, uio, cred, ioflag));
444 		nfsstats.biocache_writes++;
445 		lbn = uio->uio_offset / biosize;
446 		on = uio->uio_offset & (biosize-1);
447 		n = min((unsigned)(biosize - on), uio->uio_resid);
448 		bn = lbn * (biosize / DEV_BSIZE);
449 again:
450 		bp = nfs_getcacheblk(vp, bn, biosize, p);
451 		if (!bp)
452 			return (EINTR);
453 		if (bp->b_wcred == NOCRED) {
454 			crhold(cred);
455 			bp->b_wcred = cred;
456 		}
457 		np->n_flag |= NMODIFIED;
458 		if (uio->uio_offset + n > np->n_size) {
459 			np->n_size = uio->uio_offset + n;
460 			vnode_pager_setsize(vp, (u_long)np->n_size);
461 		}
462 
463 		/*
464 		 * If the new write will leave a contiguous dirty
465 		 * area, just update the b_dirtyoff and b_dirtyend,
466 		 * otherwise force a write rpc of the old dirty area.
467 		 */
468 		if (bp->b_dirtyend > 0 &&
469 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
470 			bp->b_proc = p;
471 			if (VOP_BWRITE(bp) == EINTR)
472 				return (EINTR);
473 			goto again;
474 		}
475 
476 		/*
477 		 * Check for valid write lease and get one as required.
478 		 * In case getblk() and/or bwrite() delayed us.
479 		 */
480 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
481 		    NQNFS_CKINVALID(vp, np, NQL_WRITE)) {
482 			do {
483 				error = nqnfs_getlease(vp, NQL_WRITE, cred, p);
484 			} while (error == NQNFS_EXPIRED);
485 			if (error) {
486 				brelse(bp);
487 				return (error);
488 			}
489 			if (np->n_lrev != np->n_brev ||
490 			    (np->n_flag & NQNFSNONCACHE)) {
491 				brelse(bp);
492 				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
493 					return (error);
494 				np->n_brev = np->n_lrev;
495 				goto again;
496 			}
497 		}
498 		if (error = uiomove((char *)bp->b_data + on, n, uio)) {
499 			bp->b_flags |= B_ERROR;
500 			brelse(bp);
501 			return (error);
502 		}
503 		if (bp->b_dirtyend > 0) {
504 			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
505 			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
506 		} else {
507 			bp->b_dirtyoff = on;
508 			bp->b_dirtyend = on + n;
509 		}
510 #ifndef notdef
511 		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
512 		    bp->b_validoff > bp->b_dirtyend) {
513 			bp->b_validoff = bp->b_dirtyoff;
514 			bp->b_validend = bp->b_dirtyend;
515 		} else {
516 			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
517 			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
518 		}
519 #else
520 		bp->b_validoff = bp->b_dirtyoff;
521 		bp->b_validend = bp->b_dirtyend;
522 #endif
523 		if (ioflag & IO_APPEND)
524 			bp->b_flags |= B_APPENDWRITE;
525 
526 		/*
527 		 * If the lease is non-cachable or IO_SYNC do bwrite().
528 		 */
529 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
530 			bp->b_proc = p;
531 			if (error = VOP_BWRITE(bp))
532 				return (error);
533 		} else if ((n + on) == biosize &&
534 			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
535 			bp->b_proc = (struct proc *)0;
536 			bawrite(bp);
537 		} else
538 			bdwrite(bp);
539 	} while (uio->uio_resid > 0 && n > 0);
540 	return (0);
541 }
542 
543 /*
544  * Get an nfs cache block.
545  * Allocate a new one if the block isn't currently in the cache
546  * and return the block marked busy. If the calling process is
547  * interrupted by a signal for an interruptible mount point, return
548  * NULL.
549  */
550 struct buf *
551 nfs_getcacheblk(vp, bn, size, p)
552 	struct vnode *vp;
553 	daddr_t bn;
554 	int size;
555 	struct proc *p;
556 {
557 	register struct buf *bp;
558 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
559 
560 	if (nmp->nm_flag & NFSMNT_INT) {
561 		bp = getblk(vp, bn, size, PCATCH, 0);
562 		while (bp == (struct buf *)0) {
563 			if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
564 				return ((struct buf *)0);
565 			bp = getblk(vp, bn, size, 0, 2 * hz);
566 		}
567 	} else
568 		bp = getblk(vp, bn, size, 0, 0);
569 	return (bp);
570 }
571 
572 /*
573  * Flush and invalidate all dirty buffers. If another process is already
574  * doing the flush, just wait for completion.
575  */
576 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
577 	struct vnode *vp;
578 	int flags;
579 	struct ucred *cred;
580 	struct proc *p;
581 	int intrflg;
582 {
583 	register struct nfsnode *np = VTONFS(vp);
584 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
585 	int error = 0, slpflag, slptimeo;
586 
587 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
588 		intrflg = 0;
589 	if (intrflg) {
590 		slpflag = PCATCH;
591 		slptimeo = 2 * hz;
592 	} else {
593 		slpflag = 0;
594 		slptimeo = 0;
595 	}
596 	/*
597 	 * First wait for any other process doing a flush to complete.
598 	 */
599 	while (np->n_flag & NFLUSHINPROG) {
600 		np->n_flag |= NFLUSHWANT;
601 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
602 			slptimeo);
603 		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
604 			return (EINTR);
605 	}
606 
607 	/*
608 	 * Now, flush as required.
609 	 */
610 	np->n_flag |= NFLUSHINPROG;
611 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
612 	while (error) {
613 		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
614 			np->n_flag &= ~NFLUSHINPROG;
615 			if (np->n_flag & NFLUSHWANT) {
616 				np->n_flag &= ~NFLUSHWANT;
617 				wakeup((caddr_t)&np->n_flag);
618 			}
619 			return (EINTR);
620 		}
621 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
622 	}
623 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
624 	if (np->n_flag & NFLUSHWANT) {
625 		np->n_flag &= ~NFLUSHWANT;
626 		wakeup((caddr_t)&np->n_flag);
627 	}
628 	return (0);
629 }
630 
631 /*
632  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
633  * This is mainly to avoid queueing async I/O requests when the nfsiods
634  * are all hung on a dead server.
635  */
636 nfs_asyncio(bp, cred)
637 	register struct buf *bp;
638 	struct ucred *cred;
639 {
640 	register int i;
641 
642 	if (nfs_numasync == 0)
643 		return (EIO);
644 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
645 	    if (nfs_iodwant[i]) {
646 		if (bp->b_flags & B_READ) {
647 			if (bp->b_rcred == NOCRED && cred != NOCRED) {
648 				crhold(cred);
649 				bp->b_rcred = cred;
650 			}
651 		} else {
652 			if (bp->b_wcred == NOCRED && cred != NOCRED) {
653 				crhold(cred);
654 				bp->b_wcred = cred;
655 			}
656 		}
657 
658 		TAILQ_INSERT_TAIL(&nfs_bufq, bp, b_freelist);
659 		nfs_iodwant[i] = (struct proc *)0;
660 		wakeup((caddr_t)&nfs_iodwant[i]);
661 		return (0);
662 	    }
663 	return (EIO);
664 }
665 
666 /*
667  * Do an I/O operation to/from a cache block. This may be called
668  * synchronously or from an nfsiod.
669  */
670 int
671 nfs_doio(bp, cr, p)
672 	register struct buf *bp;
673 	struct cred *cr;
674 	struct proc *p;
675 {
676 	register struct uio *uiop;
677 	register struct vnode *vp;
678 	struct nfsnode *np;
679 	struct nfsmount *nmp;
680 	int error, diff, len;
681 	struct uio uio;
682 	struct iovec io;
683 
684 	vp = bp->b_vp;
685 	np = VTONFS(vp);
686 	nmp = VFSTONFS(vp->v_mount);
687 	uiop = &uio;
688 	uiop->uio_iov = &io;
689 	uiop->uio_iovcnt = 1;
690 	uiop->uio_segflg = UIO_SYSSPACE;
691 	uiop->uio_procp = p;
692 
693 	/*
694 	 * Historically, paging was done with physio, but no more...
695 	 */
696 	if (bp->b_flags & B_PHYS) {
697 	    /*
698 	     * ...though reading /dev/drum still gets us here.
699 	     */
700 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
701 	    /* mapping was done by vmapbuf() */
702 	    io.iov_base = bp->b_data;
703 	    uiop->uio_offset = bp->b_blkno * DEV_BSIZE;
704 	    if (bp->b_flags & B_READ) {
705 		uiop->uio_rw = UIO_READ;
706 		nfsstats.read_physios++;
707 		error = nfs_readrpc(vp, uiop, cr);
708 	    } else {
709 		uiop->uio_rw = UIO_WRITE;
710 		nfsstats.write_physios++;
711 		error = nfs_writerpc(vp, uiop, cr);
712 	    }
713 	    if (error) {
714 		bp->b_flags |= B_ERROR;
715 		bp->b_error = error;
716 	    }
717 	} else if (bp->b_flags & B_READ) {
718 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
719 	    io.iov_base = bp->b_data;
720 	    uiop->uio_rw = UIO_READ;
721 	    switch (vp->v_type) {
722 	    case VREG:
723 		uiop->uio_offset = bp->b_blkno * DEV_BSIZE;
724 		nfsstats.read_bios++;
725 		error = nfs_readrpc(vp, uiop, cr);
726 		if (!error) {
727 		    bp->b_validoff = 0;
728 		    if (uiop->uio_resid) {
729 			/*
730 			 * If len > 0, there is a hole in the file and
731 			 * no writes after the hole have been pushed to
732 			 * the server yet.
733 			 * Just zero fill the rest of the valid area.
734 			 */
735 			diff = bp->b_bcount - uiop->uio_resid;
736 			len = np->n_size - (bp->b_blkno * DEV_BSIZE
737 				+ diff);
738 			if (len > 0) {
739 			    len = min(len, uiop->uio_resid);
740 			    bzero((char *)bp->b_data + diff, len);
741 			    bp->b_validend = diff + len;
742 			} else
743 			    bp->b_validend = diff;
744 		    } else
745 			bp->b_validend = bp->b_bcount;
746 		}
747 		if (p && (vp->v_flag & VTEXT) &&
748 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
749 			  NQNFS_CKINVALID(vp, np, NQL_READ) &&
750 			  np->n_lrev != np->n_brev) ||
751 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
752 			  np->n_mtime != np->n_vattr.va_mtime.ts_sec))) {
753 			uprintf("Process killed due to text file modification\n");
754 			psignal(p, SIGKILL);
755 			p->p_holdcnt++;
756 		}
757 		break;
758 	    case VLNK:
759 		uiop->uio_offset = 0;
760 		nfsstats.readlink_bios++;
761 		error = nfs_readlinkrpc(vp, uiop, cr);
762 		break;
763 	    case VDIR:
764 		uiop->uio_offset = bp->b_lblkno;
765 		nfsstats.readdir_bios++;
766 		if (VFSTONFS(vp->v_mount)->nm_flag & NFSMNT_NQNFS)
767 		    error = nfs_readdirlookrpc(vp, uiop, cr);
768 		else
769 		    error = nfs_readdirrpc(vp, uiop, cr);
770 		/*
771 		 * Save offset cookie in b_blkno.
772 		 */
773 		bp->b_blkno = uiop->uio_offset;
774 		break;
775 	    };
776 	    if (error) {
777 		bp->b_flags |= B_ERROR;
778 		bp->b_error = error;
779 	    }
780 	} else {
781 	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
782 		- bp->b_dirtyoff;
783 	    uiop->uio_offset = (bp->b_blkno * DEV_BSIZE)
784 		+ bp->b_dirtyoff;
785 	    io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
786 	    uiop->uio_rw = UIO_WRITE;
787 	    nfsstats.write_bios++;
788 	    if (bp->b_flags & B_APPENDWRITE)
789 		error = nfs_writerpc(vp, uiop, cr, IO_APPEND);
790 	    else
791 		error = nfs_writerpc(vp, uiop, cr, 0);
792 	    bp->b_flags &= ~(B_WRITEINPROG | B_APPENDWRITE);
793 
794 	    /*
795 	     * For an interrupted write, the buffer is still valid and the
796 	     * write hasn't been pushed to the server yet, so we can't set
797 	     * B_ERROR and report the interruption by setting B_EINTR. For
798 	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
799 	     * is essentially a noop.
800 	     */
801 	    if (error == EINTR) {
802 		bp->b_flags &= ~B_INVAL;
803 		bp->b_flags |= B_DELWRI;
804 
805 		/*
806 		 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the
807 		 * buffer to the clean list, we have to reassign it back to the
808 		 * dirty one. Ugh.
809 		 */
810 		if (bp->b_flags & B_ASYNC)
811 		    reassignbuf(bp, vp);
812 		else
813 		    bp->b_flags |= B_EINTR;
814 	    } else {
815 		if (error) {
816 		    bp->b_flags |= B_ERROR;
817 		    bp->b_error = np->n_error = error;
818 		    np->n_flag |= NWRITEERR;
819 		}
820 		bp->b_dirtyoff = bp->b_dirtyend = 0;
821 	    }
822 	}
823 	bp->b_resid = uiop->uio_resid;
824 	biodone(bp);
825 	return (error);
826 }
827