xref: /netbsd-src/sys/nfs/nfs_bio.c (revision d0fed6c87ddc40a8bffa6f99e7433ddfc864dd83)
1 /*	$NetBSD: nfs_bio.c,v 1.31 1997/04/20 16:24:44 fvdl 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	8.9 (Berkeley) 3/30/95
39  */
40 
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/resourcevar.h>
45 #include <sys/signalvar.h>
46 #include <sys/proc.h>
47 #include <sys/buf.h>
48 #include <sys/vnode.h>
49 #include <sys/trace.h>
50 #include <sys/mount.h>
51 #include <sys/kernel.h>
52 #include <sys/namei.h>
53 
54 #include <vm/vm.h>
55 
56 #include <nfs/rpcv2.h>
57 #include <nfs/nfsproto.h>
58 #include <nfs/nfs.h>
59 #include <nfs/nfsmount.h>
60 #include <nfs/nqnfs.h>
61 #include <nfs/nfsnode.h>
62 #include <nfs/nfs_var.h>
63 
64 extern int nfs_numasync;
65 extern struct nfsstats nfsstats;
66 
67 /*
68  * Vnode op for read using bio
69  * Any similarity to readip() is purely coincidental
70  */
71 int
72 nfs_bioread(vp, uio, ioflag, cred)
73 	register struct vnode *vp;
74 	register struct uio *uio;
75 	int ioflag;
76 	struct ucred *cred;
77 {
78 	register struct nfsnode *np = VTONFS(vp);
79 	register int biosize, diff, i;
80 	struct buf *bp = NULL, *rabp;
81 	struct vattr vattr;
82 	struct proc *p;
83 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
84 	daddr_t lbn, bn, rabn;
85 	caddr_t baddr;
86 	int got_buf = 0, nra, error = 0, n = 0, on = 0, not_readin;
87 
88 #ifdef DIAGNOSTIC
89 	if (uio->uio_rw != UIO_READ)
90 		panic("nfs_read mode");
91 #endif
92 	if (uio->uio_resid == 0)
93 		return (0);
94 	if (uio->uio_offset < 0)
95 		return (EINVAL);
96 	p = uio->uio_procp;
97 	if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3)
98 		(void)nfs_fsinfo(nmp, vp, cred, p);
99 	biosize = nmp->nm_rsize;
100 	/*
101 	 * For nfs, cache consistency can only be maintained approximately.
102 	 * Although RFC1094 does not specify the criteria, the following is
103 	 * believed to be compatible with the reference port.
104 	 * For nqnfs, full cache consistency is maintained within the loop.
105 	 * For nfs:
106 	 * If the file's modify time on the server has changed since the
107 	 * last read rpc or you have written to the file,
108 	 * you may have lost data cache consistency with the
109 	 * server, so flush all of the file's data out of the cache.
110 	 * Then force a getattr rpc to ensure that you have up to date
111 	 * attributes.
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 (vp->v_type != VREG) {
120 				if (vp->v_type != VDIR)
121 					panic("nfs: bioread, not dir");
122 				nfs_invaldir(vp);
123 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
124 				if (error)
125 					return (error);
126 			}
127 			np->n_attrstamp = 0;
128 			error = VOP_GETATTR(vp, &vattr, cred, p);
129 			if (error)
130 				return (error);
131 			np->n_mtime = vattr.va_mtime.tv_sec;
132 		} else {
133 			error = VOP_GETATTR(vp, &vattr, cred, p);
134 			if (error)
135 				return (error);
136 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
137 				if (vp->v_type == VDIR)
138 					nfs_invaldir(vp);
139 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
140 				if (error)
141 					return (error);
142 				np->n_mtime = vattr.va_mtime.tv_sec;
143 			}
144 		}
145 	}
146 	do {
147 
148 	    /*
149 	     * Get a valid lease. If cached data is stale, flush it.
150 	     */
151 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
152 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
153 		    do {
154 			error = nqnfs_getlease(vp, ND_READ, cred, p);
155 		    } while (error == NQNFS_EXPIRED);
156 		    if (error)
157 			return (error);
158 		    if (np->n_lrev != np->n_brev ||
159 			(np->n_flag & NQNFSNONCACHE) ||
160 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
161 			if (vp->v_type == VDIR)
162 			    nfs_invaldir(vp);
163 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
164 			if (error)
165 			    return (error);
166 			np->n_brev = np->n_lrev;
167 		    }
168 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
169 		    nfs_invaldir(vp);
170 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
171 		    if (error)
172 			return (error);
173 		}
174 	    }
175 	    /*
176 	     * Don't cache symlinks.
177 	     */
178 	    if (np->n_flag & NQNFSNONCACHE
179 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
180 		switch (vp->v_type) {
181 		case VREG:
182 			return (nfs_readrpc(vp, uio, cred));
183 		case VLNK:
184 			return (nfs_readlinkrpc(vp, uio, cred));
185 		case VDIR:
186 			break;
187 		default:
188 			printf(" NQNFSNONCACHE: type %x unexpected\n",
189 			    vp->v_type);
190 		};
191 	    }
192 	    baddr = (caddr_t)0;
193 	    switch (vp->v_type) {
194 	    case VREG:
195 		nfsstats.biocache_reads++;
196 		lbn = uio->uio_offset / biosize;
197 		on = uio->uio_offset & (biosize - 1);
198 		bn = lbn * (biosize / DEV_BSIZE);
199 		not_readin = 1;
200 
201 		/*
202 		 * Start the read ahead(s), as required.
203 		 */
204 		if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
205 		    for (nra = 0; nra < nmp->nm_readahead &&
206 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
207 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
208 			if (!incore(vp, rabn)) {
209 			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
210 			    if (!rabp)
211 				return (EINTR);
212 			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
213 				rabp->b_flags |= (B_READ | B_ASYNC);
214 				if (nfs_asyncio(rabp, cred)) {
215 				    rabp->b_flags |= B_INVAL;
216 				    brelse(rabp);
217 				}
218 			    } else
219 				brelse(rabp);
220 			}
221 		    }
222 		}
223 
224 		/*
225 		 * If the block is in the cache and has the required data
226 		 * in a valid region, just copy it out.
227 		 * Otherwise, get the block and write back/read in,
228 		 * as required.
229 		 */
230 		if ((bp = incore(vp, bn)) &&
231 		    (bp->b_flags & (B_BUSY | B_WRITEINPROG)) ==
232 		    (B_BUSY | B_WRITEINPROG))
233 			got_buf = 0;
234 		else {
235 again:
236 			bp = nfs_getcacheblk(vp, bn, biosize, p);
237 			if (!bp)
238 				return (EINTR);
239 			got_buf = 1;
240 			if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
241 				bp->b_flags |= B_READ;
242 				not_readin = 0;
243 				error = nfs_doio(bp, cred, p);
244 				if (error) {
245 				    brelse(bp);
246 				    return (error);
247 				}
248 			}
249 		}
250 		n = min((unsigned)(biosize - on), uio->uio_resid);
251 		diff = np->n_size - uio->uio_offset;
252 		if (diff < n)
253 			n = diff;
254 		if (not_readin && n > 0) {
255 			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
256 				if (!got_buf) {
257 				    bp = nfs_getcacheblk(vp, bn, biosize, p);
258 				    if (!bp)
259 					return (EINTR);
260 				    got_buf = 1;
261 				}
262 				bp->b_flags |= B_INVAFTERWRITE;
263 				if (bp->b_dirtyend > 0) {
264 				    if ((bp->b_flags & B_DELWRI) == 0)
265 					panic("nfsbioread");
266 				    if (VOP_BWRITE(bp) == EINTR)
267 					return (EINTR);
268 				} else
269 				    brelse(bp);
270 				goto again;
271 			}
272 		}
273 		vp->v_lastr = lbn;
274 		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
275 		if (diff < n)
276 			n = diff;
277 		break;
278 	    case VLNK:
279 		nfsstats.biocache_readlinks++;
280 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
281 		if (!bp)
282 			return (EINTR);
283 		if ((bp->b_flags & B_DONE) == 0) {
284 			bp->b_flags |= B_READ;
285 			error = nfs_doio(bp, cred, p);
286 			if (error) {
287 				brelse(bp);
288 				return (error);
289 			}
290 		}
291 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
292 		got_buf = 1;
293 		on = 0;
294 		break;
295 	    case VDIR:
296 		if (uio->uio_resid < NFS_READDIRBLKSIZ)
297 			return (0);
298 		nfsstats.biocache_readdirs++;
299 		lbn = uio->uio_offset / NFS_DIRBLKSIZ;
300 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
301 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p);
302 		if (!bp)
303 		    return (EINTR);
304 		if ((bp->b_flags & B_DONE) == 0) {
305 		    bp->b_flags |= B_READ;
306 		    error = nfs_doio(bp, cred, p);
307 		    if (error) {
308 			brelse(bp);
309 			while (error == NFSERR_BAD_COOKIE) {
310 			    nfs_invaldir(vp);
311 			    error = nfs_vinvalbuf(vp, 0, cred, p, 1);
312 			    /*
313 			     * Yuck! The directory has been modified on the
314 			     * server. The only way to get the block is by
315 			     * reading from the beginning to get all the
316 			     * offset cookies.
317 			     */
318 			    for (i = 0; i <= lbn && !error; i++) {
319 				bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p);
320 				if (!bp)
321 				    return (EINTR);
322 				if ((bp->b_flags & B_DONE) == 0) {
323 				    bp->b_flags |= B_READ;
324 				    error = nfs_doio(bp, cred, p);
325 				    if (error)
326 					brelse(bp);
327 				}
328 			    }
329 			}
330 			if (error)
331 			    return (error);
332 		    }
333 		}
334 
335 		/*
336 		 * If not eof and read aheads are enabled, start one.
337 		 * (You need the current block first, so that you have the
338 		 *  directory offset cookie of the next block.)
339 		 */
340 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
341 		    (np->n_direofoffset == 0 ||
342 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
343 		    !(np->n_flag & NQNFSNONCACHE) &&
344 		    !incore(vp, lbn + 1)) {
345 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p);
346 			if (rabp) {
347 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
348 				rabp->b_flags |= (B_READ | B_ASYNC);
349 				if (nfs_asyncio(rabp, cred)) {
350 				    rabp->b_flags |= B_INVAL;
351 				    brelse(rabp);
352 				}
353 			    } else
354 				brelse(rabp);
355 			}
356 		}
357 		n = min(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
358 		got_buf = 1;
359 		break;
360 	    default:
361 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
362 		break;
363 	    };
364 
365 	    if (n > 0) {
366 		if (!baddr)
367 			baddr = bp->b_data;
368 		error = uiomove(baddr + on, (int)n, uio);
369 	    }
370 	    switch (vp->v_type) {
371 	    case VREG:
372 		break;
373 	    case VLNK:
374 		n = 0;
375 		break;
376 	    case VDIR:
377 		if (np->n_flag & NQNFSNONCACHE)
378 			bp->b_flags |= B_INVAL;
379 		break;
380 	    default:
381 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
382 	    }
383 	    if (got_buf)
384 		brelse(bp);
385 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
386 	return (error);
387 }
388 
389 /*
390  * Vnode op for write using bio
391  */
392 int
393 nfs_write(v)
394 	void *v;
395 {
396 	struct vop_write_args /* {
397 		struct vnode *a_vp;
398 		struct uio *a_uio;
399 		int  a_ioflag;
400 		struct ucred *a_cred;
401 	} */ *ap = v;
402 	register int biosize;
403 	register struct uio *uio = ap->a_uio;
404 	struct proc *p = uio->uio_procp;
405 	register struct vnode *vp = ap->a_vp;
406 	struct nfsnode *np = VTONFS(vp);
407 	register struct ucred *cred = ap->a_cred;
408 	int ioflag = ap->a_ioflag;
409 	struct buf *bp;
410 	struct vattr vattr;
411 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
412 	daddr_t lbn, bn;
413 	int n, on, error = 0, iomode, must_commit;
414 
415 #ifdef DIAGNOSTIC
416 	if (uio->uio_rw != UIO_WRITE)
417 		panic("nfs_write mode");
418 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
419 		panic("nfs_write proc");
420 #endif
421 	if (vp->v_type != VREG)
422 		return (EIO);
423 	if (np->n_flag & NWRITEERR) {
424 		np->n_flag &= ~NWRITEERR;
425 		return (np->n_error);
426 	}
427 	if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3)
428 		(void)nfs_fsinfo(nmp, vp, cred, p);
429 	if (ioflag & (IO_APPEND | IO_SYNC)) {
430 		if (np->n_flag & NMODIFIED) {
431 			np->n_attrstamp = 0;
432 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
433 			if (error)
434 				return (error);
435 		}
436 		if (ioflag & IO_APPEND) {
437 			np->n_attrstamp = 0;
438 			error = VOP_GETATTR(vp, &vattr, cred, p);
439 			if (error)
440 				return (error);
441 			uio->uio_offset = np->n_size;
442 		}
443 	}
444 	if (uio->uio_offset < 0)
445 		return (EINVAL);
446 	if (uio->uio_resid == 0)
447 		return (0);
448 	/*
449 	 * Maybe this should be above the vnode op call, but so long as
450 	 * file servers have no limits, i don't think it matters
451 	 */
452 	if (p && uio->uio_offset + uio->uio_resid >
453 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
454 		psignal(p, SIGXFSZ);
455 		return (EFBIG);
456 	}
457 	/*
458 	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
459 	 * will be the same size within a filesystem. nfs_writerpc will
460 	 * still use nm_wsize when sizing the rpc's.
461 	 */
462 	biosize = nmp->nm_rsize;
463 	do {
464 
465 		/*
466 		 * XXX make sure we aren't cached in the VM page cache
467 		 */
468 		(void)vnode_pager_uncache(vp);
469 
470 		/*
471 		 * Check for a valid write lease.
472 		 */
473 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
474 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
475 			do {
476 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
477 			} while (error == NQNFS_EXPIRED);
478 			if (error)
479 				return (error);
480 			if (np->n_lrev != np->n_brev ||
481 			    (np->n_flag & NQNFSNONCACHE)) {
482 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
483 				if (error)
484 					return (error);
485 				np->n_brev = np->n_lrev;
486 			}
487 		}
488 		if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
489 		    iomode = NFSV3WRITE_FILESYNC;
490 		    error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
491 		    if (must_commit)
492 			nfs_clearcommit(vp->v_mount);
493 		    return (error);
494 		}
495 		nfsstats.biocache_writes++;
496 		lbn = uio->uio_offset / biosize;
497 		on = uio->uio_offset & (biosize-1);
498 		n = min((unsigned)(biosize - on), uio->uio_resid);
499 		bn = lbn * (biosize / DEV_BSIZE);
500 again:
501 		bp = nfs_getcacheblk(vp, bn, biosize, p);
502 		if (!bp)
503 			return (EINTR);
504 		if (bp->b_wcred == NOCRED) {
505 			crhold(cred);
506 			bp->b_wcred = cred;
507 		}
508 		np->n_flag |= NMODIFIED;
509 		if (uio->uio_offset + n > np->n_size) {
510 			np->n_size = uio->uio_offset + n;
511 			vnode_pager_setsize(vp, (u_long)np->n_size);
512 		}
513 
514 		/*
515 		 * If the new write will leave a contiguous dirty
516 		 * area, just update the b_dirtyoff and b_dirtyend,
517 		 * otherwise force a write rpc of the old dirty area.
518 		 */
519 		if (bp->b_dirtyend > 0 &&
520 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
521 			bp->b_proc = p;
522 			if (VOP_BWRITE(bp) == EINTR)
523 				return (EINTR);
524 			goto again;
525 		}
526 
527 		/*
528 		 * Check for valid write lease and get one as required.
529 		 * In case getblk() and/or bwrite() delayed us.
530 		 */
531 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
532 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
533 			do {
534 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
535 			} while (error == NQNFS_EXPIRED);
536 			if (error) {
537 				brelse(bp);
538 				return (error);
539 			}
540 			if (np->n_lrev != np->n_brev ||
541 			    (np->n_flag & NQNFSNONCACHE)) {
542 				brelse(bp);
543 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
544 				if (error)
545 					return (error);
546 				np->n_brev = np->n_lrev;
547 				goto again;
548 			}
549 		}
550 		error = uiomove((char *)bp->b_data + on, n, uio);
551 		if (error) {
552 			bp->b_flags |= B_ERROR;
553 			brelse(bp);
554 			return (error);
555 		}
556 		if (bp->b_dirtyend > 0) {
557 			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
558 			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
559 		} else {
560 			bp->b_dirtyoff = on;
561 			bp->b_dirtyend = on + n;
562 		}
563 		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
564 		    bp->b_validoff > bp->b_dirtyend) {
565 			bp->b_validoff = bp->b_dirtyoff;
566 			bp->b_validend = bp->b_dirtyend;
567 		} else {
568 			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
569 			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
570 		}
571 
572 		/*
573 		 * Since this block is being modified, it must be written
574 		 * again and not just committed.
575 		 */
576 		bp->b_flags &= ~B_NEEDCOMMIT;
577 
578 		/*
579 		 * If the lease is non-cachable or IO_SYNC do bwrite().
580 		 */
581 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
582 			bp->b_proc = p;
583 			error = VOP_BWRITE(bp);
584 			if (error)
585 				return (error);
586 			if (np->n_flag & NQNFSNONCACHE) {
587 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
588 				if (error)
589 					return (error);
590 			}
591 		} else if ((n + on) == biosize &&
592 			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
593 			bp->b_proc = (struct proc *)0;
594 			bp->b_flags |= B_ASYNC;
595 			(void)nfs_writebp(bp, 0);
596 		} else {
597 			bdwrite(bp);
598 		}
599 	} while (uio->uio_resid > 0 && n > 0);
600 	return (0);
601 }
602 
603 /*
604  * Get an nfs cache block.
605  * Allocate a new one if the block isn't currently in the cache
606  * and return the block marked busy. If the calling process is
607  * interrupted by a signal for an interruptible mount point, return
608  * NULL.
609  */
610 struct buf *
611 nfs_getcacheblk(vp, bn, size, p)
612 	struct vnode *vp;
613 	daddr_t bn;
614 	int size;
615 	struct proc *p;
616 {
617 	register struct buf *bp;
618 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
619 
620 	if (nmp->nm_flag & NFSMNT_INT) {
621 		bp = getblk(vp, bn, size, PCATCH, 0);
622 		while (bp == (struct buf *)0) {
623 			if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
624 				return ((struct buf *)0);
625 			bp = getblk(vp, bn, size, 0, 2 * hz);
626 		}
627 	} else
628 		bp = getblk(vp, bn, size, 0, 0);
629 	return (bp);
630 }
631 
632 /*
633  * Flush and invalidate all dirty buffers. If another process is already
634  * doing the flush, just wait for completion.
635  */
636 int
637 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
638 	struct vnode *vp;
639 	int flags;
640 	struct ucred *cred;
641 	struct proc *p;
642 	int intrflg;
643 {
644 	register struct nfsnode *np = VTONFS(vp);
645 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
646 	int error = 0, slpflag, slptimeo;
647 
648 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
649 		intrflg = 0;
650 	if (intrflg) {
651 		slpflag = PCATCH;
652 		slptimeo = 2 * hz;
653 	} else {
654 		slpflag = 0;
655 		slptimeo = 0;
656 	}
657 	/*
658 	 * First wait for any other process doing a flush to complete.
659 	 */
660 	while (np->n_flag & NFLUSHINPROG) {
661 		np->n_flag |= NFLUSHWANT;
662 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
663 			slptimeo);
664 		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
665 			return (EINTR);
666 	}
667 
668 	/*
669 	 * Now, flush as required.
670 	 */
671 	np->n_flag |= NFLUSHINPROG;
672 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
673 	while (error) {
674 		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
675 			np->n_flag &= ~NFLUSHINPROG;
676 			if (np->n_flag & NFLUSHWANT) {
677 				np->n_flag &= ~NFLUSHWANT;
678 				wakeup((caddr_t)&np->n_flag);
679 			}
680 			return (EINTR);
681 		}
682 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
683 	}
684 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
685 	if (np->n_flag & NFLUSHWANT) {
686 		np->n_flag &= ~NFLUSHWANT;
687 		wakeup((caddr_t)&np->n_flag);
688 	}
689 	return (0);
690 }
691 
692 /*
693  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
694  * This is mainly to avoid queueing async I/O requests when the nfsiods
695  * are all hung on a dead server.
696  */
697 int
698 nfs_asyncio(bp, cred)
699 	register struct buf *bp;
700 	struct ucred *cred;
701 {
702 	register int i;
703 	register struct nfsmount *nmp;
704 	int gotiod, slpflag = 0, slptimeo = 0, error;
705 
706 	if (nfs_numasync == 0)
707 		return (EIO);
708 
709 
710 	nmp = VFSTONFS(bp->b_vp->v_mount);
711 again:
712 	if (nmp->nm_flag & NFSMNT_INT)
713 		slpflag = PCATCH;
714 	gotiod = FALSE;
715 
716 	/*
717 	 * Find a free iod to process this request.
718 	 */
719 
720 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
721 		if (nfs_iodwant[i]) {
722 			/*
723 			 * Found one, so wake it up and tell it which
724 			 * mount to process.
725 			 */
726 			nfs_iodwant[i] = (struct proc *)0;
727 			nfs_iodmount[i] = nmp;
728 			nmp->nm_bufqiods++;
729 			wakeup((caddr_t)&nfs_iodwant[i]);
730 			gotiod = TRUE;
731 			break;
732 		}
733 	/*
734 	 * If none are free, we may already have an iod working on this mount
735 	 * point.  If so, it will process our request.
736 	 */
737 	if (!gotiod && nmp->nm_bufqiods > 0)
738 		gotiod = TRUE;
739 
740 	/*
741 	 * If we have an iod which can process the request, then queue
742 	 * the buffer.
743 	 */
744 	if (gotiod) {
745 		/*
746 		 * Ensure that the queue never grows too large.
747 		 */
748 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
749 			nmp->nm_bufqwant = TRUE;
750 			error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
751 				"nfsaio", slptimeo);
752 			if (error) {
753 				if (nfs_sigintr(nmp, NULL, bp->b_proc))
754 					return (EINTR);
755 				if (slpflag == PCATCH) {
756 					slpflag = 0;
757 					slptimeo = 2 * hz;
758 				}
759 			}
760 			/*
761 			 * We might have lost our iod while sleeping,
762 			 * so check and loop if nescessary.
763 			 */
764 			if (nmp->nm_bufqiods == 0)
765 				goto again;
766 		}
767 
768 		if (bp->b_flags & B_READ) {
769 			if (bp->b_rcred == NOCRED && cred != NOCRED) {
770 				crhold(cred);
771 				bp->b_rcred = cred;
772 			}
773 		} else {
774 			bp->b_flags |= B_WRITEINPROG;
775 			if (bp->b_wcred == NOCRED && cred != NOCRED) {
776 				crhold(cred);
777 				bp->b_wcred = cred;
778 			}
779 		}
780 
781 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
782 		nmp->nm_bufqlen++;
783 		return (0);
784 	    }
785 
786 	/*
787 	 * All the iods are busy on other mounts, so return EIO to
788 	 * force the caller to process the i/o synchronously.
789 	 */
790 	return (EIO);
791 }
792 
793 /*
794  * Do an I/O operation to/from a cache block. This may be called
795  * synchronously or from an nfsiod.
796  */
797 int
798 nfs_doio(bp, cr, p)
799 	register struct buf *bp;
800 	struct ucred *cr;
801 	struct proc *p;
802 {
803 	register struct uio *uiop;
804 	register struct vnode *vp;
805 	struct nfsnode *np;
806 	struct nfsmount *nmp;
807 	int error = 0, diff, len, iomode, must_commit = 0;
808 	struct uio uio;
809 	struct iovec io;
810 
811 	vp = bp->b_vp;
812 	np = VTONFS(vp);
813 	nmp = VFSTONFS(vp->v_mount);
814 	uiop = &uio;
815 	uiop->uio_iov = &io;
816 	uiop->uio_iovcnt = 1;
817 	uiop->uio_segflg = UIO_SYSSPACE;
818 	uiop->uio_procp = p;
819 
820 	/*
821 	 * Historically, paging was done with physio, but no more...
822 	 */
823 	if (bp->b_flags & B_PHYS) {
824 	    /*
825 	     * ...though reading /dev/drum still gets us here.
826 	     */
827 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
828 	    /* mapping was done by vmapbuf() */
829 	    io.iov_base = bp->b_data;
830 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
831 	    if (bp->b_flags & B_READ) {
832 		uiop->uio_rw = UIO_READ;
833 		nfsstats.read_physios++;
834 		error = nfs_readrpc(vp, uiop, cr);
835 	    } else {
836 		iomode = NFSV3WRITE_DATASYNC;
837 		uiop->uio_rw = UIO_WRITE;
838 		nfsstats.write_physios++;
839 		error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
840 	    }
841 	    if (error) {
842 		bp->b_flags |= B_ERROR;
843 		bp->b_error = error;
844 	    }
845 	} else if (bp->b_flags & B_READ) {
846 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
847 	    io.iov_base = bp->b_data;
848 	    uiop->uio_rw = UIO_READ;
849 	    switch (vp->v_type) {
850 	    case VREG:
851 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
852 		nfsstats.read_bios++;
853 		error = nfs_readrpc(vp, uiop, cr);
854 		if (!error) {
855 		    bp->b_validoff = 0;
856 		    if (uiop->uio_resid) {
857 			/*
858 			 * If len > 0, there is a hole in the file and
859 			 * no writes after the hole have been pushed to
860 			 * the server yet.
861 			 * Just zero fill the rest of the valid area.
862 			 */
863 			diff = bp->b_bcount - uiop->uio_resid;
864 			len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE
865 				+ diff);
866 			if (len > 0) {
867 			    len = min(len, uiop->uio_resid);
868 			    bzero((char *)bp->b_data + diff, len);
869 			    bp->b_validend = diff + len;
870 			} else
871 			    bp->b_validend = diff;
872 		    } else
873 			bp->b_validend = bp->b_bcount;
874 		}
875 		if (p && (vp->v_flag & VTEXT) &&
876 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
877 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
878 			  np->n_lrev != np->n_brev) ||
879 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
880 			  np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
881 			uprintf("Process killed due to text file modification\n");
882 			psignal(p, SIGKILL);
883 			p->p_holdcnt++;
884 		}
885 		break;
886 	    case VLNK:
887 		uiop->uio_offset = (off_t)0;
888 		nfsstats.readlink_bios++;
889 		error = nfs_readlinkrpc(vp, uiop, cr);
890 		break;
891 	    case VDIR:
892 		nfsstats.readdir_bios++;
893 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
894 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
895 			error = nfs_readdirplusrpc(vp, uiop, cr);
896 			if (error == NFSERR_NOTSUPP)
897 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
898 		}
899 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
900 			error = nfs_readdirrpc(vp, uiop, cr);
901 		break;
902 	    default:
903 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
904 		break;
905 	    };
906 	    if (error) {
907 		bp->b_flags |= B_ERROR;
908 		bp->b_error = error;
909 	    }
910 	} else {
911 	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
912 		- bp->b_dirtyoff;
913 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE
914 		+ bp->b_dirtyoff;
915 	    io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
916 	    uiop->uio_rw = UIO_WRITE;
917 	    nfsstats.write_bios++;
918 	    if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC)
919 		iomode = NFSV3WRITE_UNSTABLE;
920 	    else
921 		iomode = NFSV3WRITE_FILESYNC;
922 	    bp->b_flags |= B_WRITEINPROG;
923 #ifdef fvdl_debug
924 	    printf("nfs_doio(%x): bp %x doff %d dend %d\n",
925 		vp, bp, bp->b_dirtyoff, bp->b_dirtyend);
926 #endif
927 	    error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
928 	    if (!error && iomode == NFSV3WRITE_UNSTABLE)
929 		bp->b_flags |= B_NEEDCOMMIT;
930 	    else
931 		bp->b_flags &= ~B_NEEDCOMMIT;
932 	    bp->b_flags &= ~B_WRITEINPROG;
933 
934 	    /*
935 	     * For an interrupted write, the buffer is still valid and the
936 	     * write hasn't been pushed to the server yet, so we can't set
937 	     * B_ERROR and report the interruption by setting B_EINTR. For
938 	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
939 	     * is essentially a noop.
940 	     * For the case of a V3 write rpc not being committed to stable
941 	     * storage, the block is still dirty and requires either a commit
942 	     * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC
943 	     * before the block is reused. This is indicated by setting the
944 	     * B_DELWRI and B_NEEDCOMMIT flags.
945 	     */
946 	    if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
947 		bp->b_flags |= B_DELWRI;
948 
949 		/*
950 		 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the
951 		 * buffer to the clean list, we have to reassign it back to the
952 		 * dirty one. Ugh.
953 		 */
954 		if (bp->b_flags & B_ASYNC)
955 		    reassignbuf(bp, vp);
956 		else if (error)
957 		    bp->b_flags |= B_EINTR;
958 	    } else {
959 		if (error) {
960 		    bp->b_flags |= B_ERROR;
961 		    bp->b_error = np->n_error = error;
962 		    np->n_flag |= NWRITEERR;
963 		}
964 		bp->b_dirtyoff = bp->b_dirtyend = 0;
965 	    }
966 	}
967 	bp->b_resid = uiop->uio_resid;
968 	if (must_commit)
969 		nfs_clearcommit(vp->v_mount);
970 	biodone(bp);
971 	return (error);
972 }
973