xref: /netbsd-src/sys/nfs/nfs_bio.c (revision f5d3fbbc6ff4a77159fb268d247bd94cb7d7e332)
1 /*	$NetBSD: nfs_bio.c,v 1.36 1997/10/19 20:27:32 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 #include <sys/dirent.h>
54 
55 #include <vm/vm.h>
56 
57 #include <nfs/rpcv2.h>
58 #include <nfs/nfsproto.h>
59 #include <nfs/nfs.h>
60 #include <nfs/nfsmount.h>
61 #include <nfs/nqnfs.h>
62 #include <nfs/nfsnode.h>
63 #include <nfs/nfs_var.h>
64 
65 extern int nfs_numasync;
66 extern struct nfsstats nfsstats;
67 
68 /*
69  * Vnode op for read using bio
70  * Any similarity to readip() is purely coincidental
71  */
72 int
73 nfs_bioread(vp, uio, ioflag, cred, cflag)
74 	register struct vnode *vp;
75 	register struct uio *uio;
76 	int ioflag, cflag;
77 	struct ucred *cred;
78 {
79 	register struct nfsnode *np = VTONFS(vp);
80 	register int biosize, diff;
81 	struct buf *bp = NULL, *rabp;
82 	struct vattr vattr;
83 	struct proc *p;
84 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
85 	struct nfsdircache *ndp = NULL, *nndp = NULL;
86 	daddr_t lbn, bn, rabn;
87 	caddr_t baddr, ep, edp;
88 	int got_buf = 0, nra, error = 0, n = 0, on = 0, not_readin, en, enn;
89 	int enough = 0;
90 	struct dirent *dp, *pdp;
91 	off_t curoff = 0;
92 
93 #ifdef DIAGNOSTIC
94 	if (uio->uio_rw != UIO_READ)
95 		panic("nfs_read mode");
96 #endif
97 	if (uio->uio_resid == 0)
98 		return (0);
99 	if (vp->v_type != VDIR && uio->uio_offset < 0)
100 		return (EINVAL);
101 	p = uio->uio_procp;
102 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
103 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
104 		(void)nfs_fsinfo(nmp, vp, cred, p);
105 	if (vp->v_type != VDIR &&
106 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
107 		return (EFBIG);
108 	biosize = nmp->nm_rsize;
109 	/*
110 	 * For nfs, cache consistency can only be maintained approximately.
111 	 * Although RFC1094 does not specify the criteria, the following is
112 	 * believed to be compatible with the reference port.
113 	 * For nqnfs, full cache consistency is maintained within the loop.
114 	 * For nfs:
115 	 * If the file's modify time on the server has changed since the
116 	 * last read rpc or you have written to the file,
117 	 * you may have lost data cache consistency with the
118 	 * server, so flush all of the file's data out of the cache.
119 	 * Then force a getattr rpc to ensure that you have up to date
120 	 * attributes.
121 	 * NB: This implies that cache data can be read when up to
122 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
123 	 * attributes this could be forced by setting n_attrstamp to 0 before
124 	 * the VOP_GETATTR() call.
125 	 */
126 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
127 		if (np->n_flag & NMODIFIED) {
128 			if (vp->v_type != VREG) {
129 				if (vp->v_type != VDIR)
130 					panic("nfs: bioread, not dir");
131 				nfs_invaldircache(vp, 0);
132 				np->n_direofoffset = 0;
133 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
134 				if (error)
135 					return (error);
136 			}
137 			np->n_attrstamp = 0;
138 			error = VOP_GETATTR(vp, &vattr, cred, p);
139 			if (error)
140 				return (error);
141 			np->n_mtime = vattr.va_mtime.tv_sec;
142 		} else {
143 			error = VOP_GETATTR(vp, &vattr, cred, p);
144 			if (error)
145 				return (error);
146 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
147 				if (vp->v_type == VDIR) {
148 					nfs_invaldircache(vp, 0);
149 					np->n_direofoffset = 0;
150 				}
151 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
152 				if (error)
153 					return (error);
154 				np->n_mtime = vattr.va_mtime.tv_sec;
155 			}
156 		}
157 	}
158 	do {
159 
160 	    /*
161 	     * Get a valid lease. If cached data is stale, flush it.
162 	     */
163 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
164 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
165 		    do {
166 			error = nqnfs_getlease(vp, ND_READ, cred, p);
167 		    } while (error == NQNFS_EXPIRED);
168 		    if (error)
169 			return (error);
170 		    if (np->n_lrev != np->n_brev ||
171 			(np->n_flag & NQNFSNONCACHE) ||
172 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
173 			if (vp->v_type == VDIR) {
174 				nfs_invaldircache(vp, 0);
175 				np->n_direofoffset = 0;
176 			}
177 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
178 			if (error)
179 			    return (error);
180 			np->n_brev = np->n_lrev;
181 		    }
182 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
183 		    nfs_invaldircache(vp, 0);
184 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
185 		    np->n_direofoffset = 0;
186 		    if (error)
187 			return (error);
188 		}
189 	    }
190 	    /*
191 	     * Don't cache symlinks.
192 	     */
193 	    if (np->n_flag & NQNFSNONCACHE
194 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
195 		switch (vp->v_type) {
196 		case VREG:
197 			return (nfs_readrpc(vp, uio, cred));
198 		case VLNK:
199 			return (nfs_readlinkrpc(vp, uio, cred));
200 		case VDIR:
201 			break;
202 		default:
203 			printf(" NQNFSNONCACHE: type %x unexpected\n",
204 			    vp->v_type);
205 		};
206 	    }
207 	    baddr = (caddr_t)0;
208 	    switch (vp->v_type) {
209 	    case VREG:
210 		nfsstats.biocache_reads++;
211 		lbn = uio->uio_offset / biosize;
212 		on = uio->uio_offset & (biosize - 1);
213 		bn = lbn * (biosize / DEV_BSIZE);
214 		not_readin = 1;
215 
216 		/*
217 		 * Start the read ahead(s), as required.
218 		 */
219 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
220 		    lbn - 1 == vp->v_lastr) {
221 		    for (nra = 0; nra < nmp->nm_readahead &&
222 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
223 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
224 			if (!incore(vp, rabn)) {
225 			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
226 			    if (!rabp)
227 				return (EINTR);
228 			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
229 				rabp->b_flags |= (B_READ | B_ASYNC);
230 				if (nfs_asyncio(rabp, cred)) {
231 				    rabp->b_flags |= B_INVAL;
232 				    brelse(rabp);
233 				}
234 			    } else
235 				brelse(rabp);
236 			}
237 		    }
238 		}
239 
240 		/*
241 		 * If the block is in the cache and has the required data
242 		 * in a valid region, just copy it out.
243 		 * Otherwise, get the block and write back/read in,
244 		 * as required.
245 		 */
246 		if ((bp = incore(vp, bn)) &&
247 		    (bp->b_flags & (B_BUSY | B_WRITEINPROG)) ==
248 		    (B_BUSY | B_WRITEINPROG))
249 			got_buf = 0;
250 		else {
251 again:
252 			bp = nfs_getcacheblk(vp, bn, biosize, p);
253 			if (!bp)
254 				return (EINTR);
255 			got_buf = 1;
256 			if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
257 				bp->b_flags |= B_READ;
258 				not_readin = 0;
259 				error = nfs_doio(bp, cred, p);
260 				if (error) {
261 				    brelse(bp);
262 				    return (error);
263 				}
264 			}
265 		}
266 		n = min((unsigned)(biosize - on), uio->uio_resid);
267 		diff = np->n_size - uio->uio_offset;
268 		if (diff < n)
269 			n = diff;
270 		if (not_readin && n > 0) {
271 			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
272 				if (!got_buf) {
273 				    bp = nfs_getcacheblk(vp, bn, biosize, p);
274 				    if (!bp)
275 					return (EINTR);
276 				    got_buf = 1;
277 				}
278 				bp->b_flags |= B_INVAFTERWRITE;
279 				if (bp->b_dirtyend > 0) {
280 				    if ((bp->b_flags & B_DELWRI) == 0)
281 					panic("nfsbioread");
282 				    if (VOP_BWRITE(bp) == EINTR)
283 					return (EINTR);
284 				} else
285 				    brelse(bp);
286 				goto again;
287 			}
288 		}
289 		vp->v_lastr = lbn;
290 		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
291 		if (diff < n)
292 			n = diff;
293 		break;
294 	    case VLNK:
295 		nfsstats.biocache_readlinks++;
296 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
297 		if (!bp)
298 			return (EINTR);
299 		if ((bp->b_flags & B_DONE) == 0) {
300 			bp->b_flags |= B_READ;
301 			error = nfs_doio(bp, cred, p);
302 			if (error) {
303 				brelse(bp);
304 				return (error);
305 			}
306 		}
307 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
308 		got_buf = 1;
309 		on = 0;
310 		break;
311 	    case VDIR:
312 diragain:
313 		nfsstats.biocache_readdirs++;
314 		ndp = nfs_searchdircache(vp, uio->uio_offset,
315 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
316 		if (!ndp) {
317 			/*
318 			 * We've been handed a cookie that is not
319 			 * in the cache. If we're not translating
320 			 * 32 <-> 64, it may be a value that was
321 			 * flushed out of the cache because it grew
322 			 * too big. Let the server judge if it's
323 			 * valid or not. In the translation case,
324 			 * we have no way of validating this value,
325 			 * so punt.
326 			 */
327 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
328 				return (EINVAL);
329 			ndp = nfs_enterdircache(vp, uio->uio_offset,
330 				uio->uio_offset, 0, 0);
331 		}
332 
333 		if (uio->uio_offset != 0 &&
334 		    ndp->dc_cookie == np->n_direofoffset) {
335 			nfsstats.direofcache_hits++;
336 			return (0);
337 		}
338 
339 		bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p);
340 		if (!bp)
341 		    return (EINTR);
342 		if ((bp->b_flags & B_DONE) == 0) {
343 		    bp->b_flags |= B_READ;
344 		    bp->b_dcookie = ndp->dc_blkcookie;
345 		    error = nfs_doio(bp, cred, p);
346 		    if (error) {
347 			/*
348 			 * Yuck! The directory has been modified on the
349 			 * server. Punt and let the userland code
350 			 * deal with it.
351 			 */
352 			brelse(bp);
353 			if (error == NFSERR_BAD_COOKIE) {
354 			    nfs_invaldircache(vp, 0);
355 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
356 			    error = EINVAL;
357 			}
358 			return (error);
359 		    }
360 		}
361 
362 		/*
363 		 * Find the entry we were looking for in the block.
364 		 */
365 
366 		en = ndp->dc_entry;
367 
368 		pdp = dp = (struct dirent *)bp->b_data;
369 		edp = bp->b_data + bp->b_validend;
370 		enn = 0;
371 		while (enn < en && (caddr_t)dp < edp) {
372 			pdp = dp;
373 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
374 			enn++;
375 		}
376 
377 		/*
378 		 * If the entry number was bigger than the number of
379 		 * entries in the block, or the cookie of the previous
380 		 * entry doesn't match, the directory cache is
381 		 * stale. Flush it and try again (i.e. go to
382 		 * the server).
383 		 */
384 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
385 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
386 #ifdef DEBUG
387 		    	printf("invalid cache: %p %p %p len %u off %lx %lx\n",
388 				pdp, dp, edp, dp->d_reclen,
389 				(unsigned long)uio->uio_offset,
390 				(unsigned long)NFS_GETCOOKIE(pdp));
391 #endif
392 			brelse(bp);
393 			nfs_invaldircache(vp, 0);
394 			nfs_vinvalbuf(vp, 0, cred, p, 0);
395 			goto diragain;
396 		}
397 
398 		on = (caddr_t)dp - bp->b_data;
399 
400 		/*
401 		 * Cache all entries that may be exported to the
402 		 * user, as they may be thrown back at us. The
403 		 * NFSBIO_CACHECOOKIES flag indicates that all
404 		 * entries are being 'exported', so cache them all.
405 		 */
406 
407 		if (en == 0 && pdp == dp) {
408 			dp = (struct dirent *)
409 			    ((caddr_t)dp + dp->d_reclen);
410 			enn++;
411 		}
412 
413 		if (uio->uio_resid < (bp->b_validend - on)) {
414 			n = uio->uio_resid;
415 			enough = 1;
416 		} else
417 			n = bp->b_validend - on;
418 
419 		ep = bp->b_data + on + n;
420 
421 		/*
422 		 * Find last complete entry to copy, caching entries
423 		 * (if requested) as we go.
424 		 */
425 
426 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
427 			if (cflag & NFSBIO_CACHECOOKIES) {
428 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
429 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
430 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
431 					NFS_STASHCOOKIE32(pdp,
432 					    nndp->dc_cookie32);
433 				}
434 			}
435 			pdp = dp;
436 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
437 			enn++;
438 		}
439 
440 		/*
441 		 * If the last requested entry was not the last in the
442 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
443 		 * cache the cookie of the last requested one, and
444 		 * set of the offset to it.
445 		 */
446 
447 		if ((on + n) < bp->b_validend) {
448 			curoff = NFS_GETCOOKIE(pdp);
449 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
450 			    enn, bp->b_lblkno);
451 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
452 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
453 				curoff = nndp->dc_cookie32;
454 			}
455 		} else
456 			curoff = bp->b_dcookie;
457 
458 		/*
459 		 * Always cache the entry for the next block,
460 		 * so that readaheads can use it.
461 		 */
462 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
463 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
464 			if (curoff == bp->b_dcookie) {
465 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
466 				curoff = nndp->dc_cookie32;
467 			}
468 		}
469 
470 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
471 
472 		/*
473 		 * If not eof and read aheads are enabled, start one.
474 		 * (You need the current block first, so that you have the
475 		 *  directory offset cookie of the next block.)
476 		 */
477 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
478 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
479 			rabp = nfs_getcacheblk(vp, nndp->dc_blkno,
480 						NFS_DIRBLKSIZ, p);
481 			if (rabp) {
482 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
483 				rabp->b_dcookie = nndp->dc_cookie;
484 				rabp->b_flags |= (B_READ | B_ASYNC);
485 				if (nfs_asyncio(rabp, cred)) {
486 				    rabp->b_flags |= B_INVAL;
487 				    brelse(rabp);
488 				}
489 			    } else
490 				brelse(rabp);
491 			}
492 		}
493 		got_buf = 1;
494 		break;
495 	    default:
496 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
497 		break;
498 	    };
499 
500 	    if (n > 0) {
501 		if (!baddr)
502 			baddr = bp->b_data;
503 		error = uiomove(baddr + on, (int)n, uio);
504 	    }
505 	    switch (vp->v_type) {
506 	    case VREG:
507 		break;
508 	    case VLNK:
509 		n = 0;
510 		break;
511 	    case VDIR:
512 		if (np->n_flag & NQNFSNONCACHE)
513 			bp->b_flags |= B_INVAL;
514 		uio->uio_offset = curoff;
515 		if (enough)
516 			n = 0;
517 		break;
518 	    default:
519 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
520 	    }
521 	    if (got_buf)
522 		brelse(bp);
523 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
524 	return (error);
525 }
526 
527 /*
528  * Vnode op for write using bio
529  */
530 int
531 nfs_write(v)
532 	void *v;
533 {
534 	struct vop_write_args /* {
535 		struct vnode *a_vp;
536 		struct uio *a_uio;
537 		int  a_ioflag;
538 		struct ucred *a_cred;
539 	} */ *ap = v;
540 	register int biosize;
541 	register struct uio *uio = ap->a_uio;
542 	struct proc *p = uio->uio_procp;
543 	register struct vnode *vp = ap->a_vp;
544 	struct nfsnode *np = VTONFS(vp);
545 	register struct ucred *cred = ap->a_cred;
546 	int ioflag = ap->a_ioflag;
547 	struct buf *bp;
548 	struct vattr vattr;
549 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
550 	daddr_t lbn, bn;
551 	int n, on, error = 0, iomode, must_commit;
552 
553 #ifdef DIAGNOSTIC
554 	if (uio->uio_rw != UIO_WRITE)
555 		panic("nfs_write mode");
556 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
557 		panic("nfs_write proc");
558 #endif
559 	if (vp->v_type != VREG)
560 		return (EIO);
561 	if (np->n_flag & NWRITEERR) {
562 		np->n_flag &= ~NWRITEERR;
563 		return (np->n_error);
564 	}
565 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
566 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
567 		(void)nfs_fsinfo(nmp, vp, cred, p);
568 	if (ioflag & (IO_APPEND | IO_SYNC)) {
569 		if (np->n_flag & NMODIFIED) {
570 			np->n_attrstamp = 0;
571 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
572 			if (error)
573 				return (error);
574 		}
575 		if (ioflag & IO_APPEND) {
576 			np->n_attrstamp = 0;
577 			error = VOP_GETATTR(vp, &vattr, cred, p);
578 			if (error)
579 				return (error);
580 			uio->uio_offset = np->n_size;
581 		}
582 	}
583 	if (uio->uio_offset < 0)
584 		return (EINVAL);
585 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
586 		return (EFBIG);
587 	if (uio->uio_resid == 0)
588 		return (0);
589 	/*
590 	 * Maybe this should be above the vnode op call, but so long as
591 	 * file servers have no limits, i don't think it matters
592 	 */
593 	if (p && uio->uio_offset + uio->uio_resid >
594 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
595 		psignal(p, SIGXFSZ);
596 		return (EFBIG);
597 	}
598 	/*
599 	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
600 	 * will be the same size within a filesystem. nfs_writerpc will
601 	 * still use nm_wsize when sizing the rpc's.
602 	 */
603 	biosize = nmp->nm_rsize;
604 	do {
605 
606 		/*
607 		 * XXX make sure we aren't cached in the VM page cache
608 		 */
609 		(void)vnode_pager_uncache(vp);
610 
611 		/*
612 		 * Check for a valid write lease.
613 		 */
614 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
615 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
616 			do {
617 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
618 			} while (error == NQNFS_EXPIRED);
619 			if (error)
620 				return (error);
621 			if (np->n_lrev != np->n_brev ||
622 			    (np->n_flag & NQNFSNONCACHE)) {
623 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
624 				if (error)
625 					return (error);
626 				np->n_brev = np->n_lrev;
627 			}
628 		}
629 		if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
630 		    iomode = NFSV3WRITE_FILESYNC;
631 		    error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
632 		    if (must_commit)
633 			nfs_clearcommit(vp->v_mount);
634 		    return (error);
635 		}
636 		nfsstats.biocache_writes++;
637 		lbn = uio->uio_offset / biosize;
638 		on = uio->uio_offset & (biosize-1);
639 		n = min((unsigned)(biosize - on), uio->uio_resid);
640 		bn = lbn * (biosize / DEV_BSIZE);
641 again:
642 		bp = nfs_getcacheblk(vp, bn, biosize, p);
643 		if (!bp)
644 			return (EINTR);
645 		if (bp->b_wcred == NOCRED) {
646 			crhold(cred);
647 			bp->b_wcred = cred;
648 		}
649 		np->n_flag |= NMODIFIED;
650 		if (uio->uio_offset + n > np->n_size) {
651 			np->n_size = uio->uio_offset + n;
652 			vnode_pager_setsize(vp, np->n_size);
653 		}
654 
655 		/*
656 		 * If the new write will leave a contiguous dirty
657 		 * area, just update the b_dirtyoff and b_dirtyend,
658 		 * otherwise force a write rpc of the old dirty area.
659 		 */
660 		if (bp->b_dirtyend > 0 &&
661 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
662 			bp->b_proc = p;
663 			if (VOP_BWRITE(bp) == EINTR)
664 				return (EINTR);
665 			goto again;
666 		}
667 
668 		/*
669 		 * Check for valid write lease and get one as required.
670 		 * In case getblk() and/or bwrite() delayed us.
671 		 */
672 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
673 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
674 			do {
675 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
676 			} while (error == NQNFS_EXPIRED);
677 			if (error) {
678 				brelse(bp);
679 				return (error);
680 			}
681 			if (np->n_lrev != np->n_brev ||
682 			    (np->n_flag & NQNFSNONCACHE)) {
683 				brelse(bp);
684 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
685 				if (error)
686 					return (error);
687 				np->n_brev = np->n_lrev;
688 				goto again;
689 			}
690 		}
691 		error = uiomove((char *)bp->b_data + on, n, uio);
692 		if (error) {
693 			bp->b_flags |= B_ERROR;
694 			brelse(bp);
695 			return (error);
696 		}
697 		if (bp->b_dirtyend > 0) {
698 			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
699 			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
700 		} else {
701 			bp->b_dirtyoff = on;
702 			bp->b_dirtyend = on + n;
703 		}
704 		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
705 		    bp->b_validoff > bp->b_dirtyend) {
706 			bp->b_validoff = bp->b_dirtyoff;
707 			bp->b_validend = bp->b_dirtyend;
708 		} else {
709 			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
710 			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
711 		}
712 
713 		/*
714 		 * Since this block is being modified, it must be written
715 		 * again and not just committed.
716 		 */
717 		bp->b_flags &= ~B_NEEDCOMMIT;
718 
719 		/*
720 		 * If the lease is non-cachable or IO_SYNC do bwrite().
721 		 */
722 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
723 			bp->b_proc = p;
724 			error = VOP_BWRITE(bp);
725 			if (error)
726 				return (error);
727 			if (np->n_flag & NQNFSNONCACHE) {
728 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
729 				if (error)
730 					return (error);
731 			}
732 		} else if ((n + on) == biosize &&
733 			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
734 			bp->b_proc = (struct proc *)0;
735 			bp->b_flags |= B_ASYNC;
736 			(void)nfs_writebp(bp, 0);
737 		} else {
738 			bdwrite(bp);
739 		}
740 	} while (uio->uio_resid > 0 && n > 0);
741 	return (0);
742 }
743 
744 /*
745  * Get an nfs cache block.
746  * Allocate a new one if the block isn't currently in the cache
747  * and return the block marked busy. If the calling process is
748  * interrupted by a signal for an interruptible mount point, return
749  * NULL.
750  */
751 struct buf *
752 nfs_getcacheblk(vp, bn, size, p)
753 	struct vnode *vp;
754 	daddr_t bn;
755 	int size;
756 	struct proc *p;
757 {
758 	register struct buf *bp;
759 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
760 
761 	if (nmp->nm_flag & NFSMNT_INT) {
762 		bp = getblk(vp, bn, size, PCATCH, 0);
763 		while (bp == (struct buf *)0) {
764 			if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
765 				return ((struct buf *)0);
766 			bp = getblk(vp, bn, size, 0, 2 * hz);
767 		}
768 	} else
769 		bp = getblk(vp, bn, size, 0, 0);
770 	return (bp);
771 }
772 
773 /*
774  * Flush and invalidate all dirty buffers. If another process is already
775  * doing the flush, just wait for completion.
776  */
777 int
778 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
779 	struct vnode *vp;
780 	int flags;
781 	struct ucred *cred;
782 	struct proc *p;
783 	int intrflg;
784 {
785 	register struct nfsnode *np = VTONFS(vp);
786 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
787 	int error = 0, slpflag, slptimeo;
788 
789 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
790 		intrflg = 0;
791 	if (intrflg) {
792 		slpflag = PCATCH;
793 		slptimeo = 2 * hz;
794 	} else {
795 		slpflag = 0;
796 		slptimeo = 0;
797 	}
798 	/*
799 	 * First wait for any other process doing a flush to complete.
800 	 */
801 	while (np->n_flag & NFLUSHINPROG) {
802 		np->n_flag |= NFLUSHWANT;
803 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
804 			slptimeo);
805 		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
806 			return (EINTR);
807 	}
808 
809 	/*
810 	 * Now, flush as required.
811 	 */
812 	np->n_flag |= NFLUSHINPROG;
813 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
814 	while (error) {
815 		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
816 			np->n_flag &= ~NFLUSHINPROG;
817 			if (np->n_flag & NFLUSHWANT) {
818 				np->n_flag &= ~NFLUSHWANT;
819 				wakeup((caddr_t)&np->n_flag);
820 			}
821 			return (EINTR);
822 		}
823 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
824 	}
825 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
826 	if (np->n_flag & NFLUSHWANT) {
827 		np->n_flag &= ~NFLUSHWANT;
828 		wakeup((caddr_t)&np->n_flag);
829 	}
830 	return (0);
831 }
832 
833 /*
834  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
835  * This is mainly to avoid queueing async I/O requests when the nfsiods
836  * are all hung on a dead server.
837  */
838 int
839 nfs_asyncio(bp, cred)
840 	register struct buf *bp;
841 	struct ucred *cred;
842 {
843 	register int i;
844 	register struct nfsmount *nmp;
845 	int gotiod, slpflag = 0, slptimeo = 0, error;
846 
847 	if (nfs_numasync == 0)
848 		return (EIO);
849 
850 
851 	nmp = VFSTONFS(bp->b_vp->v_mount);
852 again:
853 	if (nmp->nm_flag & NFSMNT_INT)
854 		slpflag = PCATCH;
855 	gotiod = FALSE;
856 
857 	/*
858 	 * Find a free iod to process this request.
859 	 */
860 
861 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
862 		if (nfs_iodwant[i]) {
863 			/*
864 			 * Found one, so wake it up and tell it which
865 			 * mount to process.
866 			 */
867 			nfs_iodwant[i] = (struct proc *)0;
868 			nfs_iodmount[i] = nmp;
869 			nmp->nm_bufqiods++;
870 			wakeup((caddr_t)&nfs_iodwant[i]);
871 			gotiod = TRUE;
872 			break;
873 		}
874 	/*
875 	 * If none are free, we may already have an iod working on this mount
876 	 * point.  If so, it will process our request.
877 	 */
878 	if (!gotiod && nmp->nm_bufqiods > 0)
879 		gotiod = TRUE;
880 
881 	/*
882 	 * If we have an iod which can process the request, then queue
883 	 * the buffer.
884 	 */
885 	if (gotiod) {
886 		/*
887 		 * Ensure that the queue never grows too large.
888 		 */
889 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
890 			nmp->nm_bufqwant = TRUE;
891 			error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
892 				"nfsaio", slptimeo);
893 			if (error) {
894 				if (nfs_sigintr(nmp, NULL, bp->b_proc))
895 					return (EINTR);
896 				if (slpflag == PCATCH) {
897 					slpflag = 0;
898 					slptimeo = 2 * hz;
899 				}
900 			}
901 			/*
902 			 * We might have lost our iod while sleeping,
903 			 * so check and loop if nescessary.
904 			 */
905 			if (nmp->nm_bufqiods == 0)
906 				goto again;
907 		}
908 
909 		if (bp->b_flags & B_READ) {
910 			if (bp->b_rcred == NOCRED && cred != NOCRED) {
911 				crhold(cred);
912 				bp->b_rcred = cred;
913 			}
914 		} else {
915 			bp->b_flags |= B_WRITEINPROG;
916 			if (bp->b_wcred == NOCRED && cred != NOCRED) {
917 				crhold(cred);
918 				bp->b_wcred = cred;
919 			}
920 		}
921 
922 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
923 		nmp->nm_bufqlen++;
924 		return (0);
925 	    }
926 
927 	/*
928 	 * All the iods are busy on other mounts, so return EIO to
929 	 * force the caller to process the i/o synchronously.
930 	 */
931 	return (EIO);
932 }
933 
934 /*
935  * Do an I/O operation to/from a cache block. This may be called
936  * synchronously or from an nfsiod.
937  */
938 int
939 nfs_doio(bp, cr, p)
940 	register struct buf *bp;
941 	struct ucred *cr;
942 	struct proc *p;
943 {
944 	register struct uio *uiop;
945 	register struct vnode *vp;
946 	struct nfsnode *np;
947 	struct nfsmount *nmp;
948 	int error = 0, diff, len, iomode, must_commit = 0;
949 	struct uio uio;
950 	struct iovec io;
951 
952 	vp = bp->b_vp;
953 	np = VTONFS(vp);
954 	nmp = VFSTONFS(vp->v_mount);
955 	uiop = &uio;
956 	uiop->uio_iov = &io;
957 	uiop->uio_iovcnt = 1;
958 	uiop->uio_segflg = UIO_SYSSPACE;
959 	uiop->uio_procp = p;
960 
961 	/*
962 	 * Historically, paging was done with physio, but no more...
963 	 */
964 	if (bp->b_flags & B_PHYS) {
965 	    /*
966 	     * ...though reading /dev/drum still gets us here.
967 	     */
968 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
969 	    /* mapping was done by vmapbuf() */
970 	    io.iov_base = bp->b_data;
971 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
972 	    if (bp->b_flags & B_READ) {
973 		uiop->uio_rw = UIO_READ;
974 		nfsstats.read_physios++;
975 		error = nfs_readrpc(vp, uiop, cr);
976 	    } else {
977 		iomode = NFSV3WRITE_DATASYNC;
978 		uiop->uio_rw = UIO_WRITE;
979 		nfsstats.write_physios++;
980 		error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
981 	    }
982 	    if (error) {
983 		bp->b_flags |= B_ERROR;
984 		bp->b_error = error;
985 	    }
986 	} else if (bp->b_flags & B_READ) {
987 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
988 	    io.iov_base = bp->b_data;
989 	    uiop->uio_rw = UIO_READ;
990 	    switch (vp->v_type) {
991 	    case VREG:
992 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
993 		nfsstats.read_bios++;
994 		error = nfs_readrpc(vp, uiop, cr);
995 		if (!error) {
996 		    bp->b_validoff = 0;
997 		    if (uiop->uio_resid) {
998 			/*
999 			 * If len > 0, there is a hole in the file and
1000 			 * no writes after the hole have been pushed to
1001 			 * the server yet.
1002 			 * Just zero fill the rest of the valid area.
1003 			 */
1004 			diff = bp->b_bcount - uiop->uio_resid;
1005 			len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE
1006 				+ diff);
1007 			if (len > 0) {
1008 			    len = min(len, uiop->uio_resid);
1009 			    bzero((char *)bp->b_data + diff, len);
1010 			    bp->b_validend = diff + len;
1011 			} else
1012 			    bp->b_validend = diff;
1013 		    } else
1014 			bp->b_validend = bp->b_bcount;
1015 		}
1016 		if (p && (vp->v_flag & VTEXT) &&
1017 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
1018 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
1019 			  np->n_lrev != np->n_brev) ||
1020 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1021 			  np->n_mtime != np->n_vattr->va_mtime.tv_sec))) {
1022 			uprintf("Process killed due to text file modification\n");
1023 			psignal(p, SIGKILL);
1024 			p->p_holdcnt++;
1025 		}
1026 		break;
1027 	    case VLNK:
1028 		uiop->uio_offset = (off_t)0;
1029 		nfsstats.readlink_bios++;
1030 		error = nfs_readlinkrpc(vp, uiop, cr);
1031 		break;
1032 	    case VDIR:
1033 		nfsstats.readdir_bios++;
1034 		uiop->uio_offset = bp->b_dcookie;
1035 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1036 			error = nfs_readdirplusrpc(vp, uiop, cr);
1037 			if (error == NFSERR_NOTSUPP)
1038 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1039 		}
1040 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1041 			error = nfs_readdirrpc(vp, uiop, cr);
1042 		if (!error) {
1043 			bp->b_dcookie = uiop->uio_offset;
1044 			bp->b_validoff = 0;
1045 			bp->b_validend = bp->b_bcount - uiop->uio_resid;
1046 		}
1047 		break;
1048 	    default:
1049 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
1050 		break;
1051 	    };
1052 	    if (error) {
1053 		bp->b_flags |= B_ERROR;
1054 		bp->b_error = error;
1055 	    }
1056 	} else {
1057 	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
1058 		- bp->b_dirtyoff;
1059 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE
1060 		+ bp->b_dirtyoff;
1061 	    io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1062 	    uiop->uio_rw = UIO_WRITE;
1063 	    nfsstats.write_bios++;
1064 	    if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC)
1065 		iomode = NFSV3WRITE_UNSTABLE;
1066 	    else
1067 		iomode = NFSV3WRITE_FILESYNC;
1068 	    bp->b_flags |= B_WRITEINPROG;
1069 #ifdef fvdl_debug
1070 	    printf("nfs_doio(%x): bp %x doff %d dend %d\n",
1071 		vp, bp, bp->b_dirtyoff, bp->b_dirtyend);
1072 #endif
1073 	    error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1074 	    if (!error && iomode == NFSV3WRITE_UNSTABLE)
1075 		bp->b_flags |= B_NEEDCOMMIT;
1076 	    else
1077 		bp->b_flags &= ~B_NEEDCOMMIT;
1078 	    bp->b_flags &= ~B_WRITEINPROG;
1079 
1080 	    /*
1081 	     * For an interrupted write, the buffer is still valid and the
1082 	     * write hasn't been pushed to the server yet, so we can't set
1083 	     * B_ERROR and report the interruption by setting B_EINTR. For
1084 	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
1085 	     * is essentially a noop.
1086 	     * For the case of a V3 write rpc not being committed to stable
1087 	     * storage, the block is still dirty and requires either a commit
1088 	     * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC
1089 	     * before the block is reused. This is indicated by setting the
1090 	     * B_DELWRI and B_NEEDCOMMIT flags.
1091 	     */
1092 	    if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1093 		bp->b_flags |= B_DELWRI;
1094 
1095 		/*
1096 		 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the
1097 		 * buffer to the clean list, we have to reassign it back to the
1098 		 * dirty one. Ugh.
1099 		 */
1100 		if (bp->b_flags & B_ASYNC)
1101 		    reassignbuf(bp, vp);
1102 		else if (error)
1103 		    bp->b_flags |= B_EINTR;
1104 	    } else {
1105 		if (error) {
1106 		    bp->b_flags |= B_ERROR;
1107 		    bp->b_error = np->n_error = error;
1108 		    np->n_flag |= NWRITEERR;
1109 		}
1110 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1111 	    }
1112 	}
1113 	bp->b_resid = uiop->uio_resid;
1114 	if (must_commit)
1115 		nfs_clearcommit(vp->v_mount);
1116 	biodone(bp);
1117 	return (error);
1118 }
1119