xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 481fca6e59249d8ffcf24fef7cfbe7b131bfb080)
1 /*	$NetBSD: nfs_bio.c,v 1.50 2000/06/27 17:52:28 mrg 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 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/resourcevar.h>
44 #include <sys/signalvar.h>
45 #include <sys/proc.h>
46 #include <sys/buf.h>
47 #include <sys/vnode.h>
48 #include <sys/trace.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
51 #include <sys/namei.h>
52 #include <sys/dirent.h>
53 
54 #include <uvm/uvm_extern.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, cflag)
73 	struct vnode *vp;
74 	struct uio *uio;
75 	int ioflag, cflag;
76 	struct ucred *cred;
77 {
78 	struct nfsnode *np = VTONFS(vp);
79 	int biosize, diff;
80 	struct buf *bp = NULL, *rabp;
81 	struct vattr vattr;
82 	struct proc *p;
83 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
84 	struct nfsdircache *ndp = NULL, *nndp = NULL;
85 	daddr_t lbn, bn, rabn;
86 	caddr_t baddr, ep, edp;
87 	int got_buf = 0, nra, error = 0, n = 0, on = 0, not_readin, en, enn;
88 	int enough = 0;
89 	struct dirent *dp, *pdp;
90 	off_t curoff = 0, offdiff;
91 
92 #ifdef DIAGNOSTIC
93 	if (uio->uio_rw != UIO_READ)
94 		panic("nfs_read mode");
95 #endif
96 	if (uio->uio_resid == 0)
97 		return (0);
98 	if (vp->v_type != VDIR && uio->uio_offset < 0)
99 		return (EINVAL);
100 	p = uio->uio_procp;
101 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
102 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
103 		(void)nfs_fsinfo(nmp, vp, cred, p);
104 	if (vp->v_type != VDIR &&
105 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
106 		return (EFBIG);
107 	biosize = nmp->nm_rsize;
108 	/*
109 	 * For nfs, cache consistency can only be maintained approximately.
110 	 * Although RFC1094 does not specify the criteria, the following is
111 	 * believed to be compatible with the reference port.
112 	 * For nqnfs, full cache consistency is maintained within the loop.
113 	 * For nfs:
114 	 * If the file's modify time on the server has changed since the
115 	 * last read rpc or you have written to the file,
116 	 * you may have lost data cache consistency with the
117 	 * server, so flush all of the file's data out of the cache.
118 	 * Then force a getattr rpc to ensure that you have up to date
119 	 * attributes.
120 	 * NB: This implies that cache data can be read when up to
121 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
122 	 * attributes this could be forced by setting n_attrstamp to 0 before
123 	 * the VOP_GETATTR() call.
124 	 */
125 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
126 		if (np->n_flag & NMODIFIED) {
127 			if (vp->v_type != VREG) {
128 				if (vp->v_type != VDIR)
129 					panic("nfs: bioread, not dir");
130 				nfs_invaldircache(vp, 0);
131 				np->n_direofoffset = 0;
132 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
133 				if (error)
134 					return (error);
135 			}
136 			np->n_attrstamp = 0;
137 			error = VOP_GETATTR(vp, &vattr, cred, p);
138 			if (error)
139 				return (error);
140 			np->n_mtime = vattr.va_mtime.tv_sec;
141 		} else {
142 			error = VOP_GETATTR(vp, &vattr, cred, p);
143 			if (error)
144 				return (error);
145 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
146 				if (vp->v_type == VDIR) {
147 					nfs_invaldircache(vp, 0);
148 					np->n_direofoffset = 0;
149 				}
150 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
151 				if (error)
152 					return (error);
153 				np->n_mtime = vattr.va_mtime.tv_sec;
154 			}
155 		}
156 	}
157 	do {
158 
159 	    /*
160 	     * Get a valid lease. If cached data is stale, flush it.
161 	     */
162 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
163 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
164 		    do {
165 			error = nqnfs_getlease(vp, ND_READ, cred, p);
166 		    } while (error == NQNFS_EXPIRED);
167 		    if (error)
168 			return (error);
169 		    if (np->n_lrev != np->n_brev ||
170 			(np->n_flag & NQNFSNONCACHE) ||
171 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
172 			if (vp->v_type == VDIR) {
173 				nfs_invaldircache(vp, 0);
174 				np->n_direofoffset = 0;
175 			}
176 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
177 			if (error)
178 			    return (error);
179 			np->n_brev = np->n_lrev;
180 		    }
181 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
182 		    nfs_invaldircache(vp, 0);
183 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
184 		    np->n_direofoffset = 0;
185 		    if (error)
186 			return (error);
187 		}
188 	    }
189 	    /*
190 	     * Don't cache symlinks.
191 	     */
192 	    if (np->n_flag & NQNFSNONCACHE
193 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
194 		switch (vp->v_type) {
195 		case VREG:
196 			return (nfs_readrpc(vp, uio, cred));
197 		case VLNK:
198 			return (nfs_readlinkrpc(vp, uio, cred));
199 		case VDIR:
200 			break;
201 		default:
202 			printf(" NQNFSNONCACHE: type %x unexpected\n",
203 			    vp->v_type);
204 		};
205 	    }
206 	    baddr = (caddr_t)0;
207 	    switch (vp->v_type) {
208 	    case VREG:
209 		nfsstats.biocache_reads++;
210 		lbn = uio->uio_offset / biosize;
211 		on = uio->uio_offset & (biosize - 1);
212 		bn = lbn * (biosize / DEV_BSIZE);
213 		not_readin = 1;
214 
215 		/*
216 		 * Start the read ahead(s), as required.
217 		 */
218 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
219 		    lbn - 1 == vp->v_lastr) {
220 		    for (nra = 0; nra < nmp->nm_readahead &&
221 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
222 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
223 			if (!incore(vp, rabn)) {
224 			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
225 			    if (!rabp)
226 				return (EINTR);
227 			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
228 				rabp->b_flags |= (B_READ | B_ASYNC);
229 				if (nfs_asyncio(rabp, cred)) {
230 				    rabp->b_flags |= B_INVAL;
231 				    brelse(rabp);
232 				}
233 			    } else
234 				brelse(rabp);
235 			}
236 		    }
237 		}
238 
239 		/*
240 		 * If the block is in the cache and has the required data
241 		 * in a valid region, just copy it out.
242 		 * Otherwise, get the block and write back/read in,
243 		 * as required.
244 		 */
245 		if ((bp = incore(vp, bn)) &&
246 		    (bp->b_flags & (B_BUSY | B_WRITEINPROG)) ==
247 		    (B_BUSY | B_WRITEINPROG))
248 			got_buf = 0;
249 		else {
250 again:
251 			bp = nfs_getcacheblk(vp, bn, biosize, p);
252 			if (!bp)
253 				return (EINTR);
254 			got_buf = 1;
255 			if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
256 				bp->b_flags |= B_READ;
257 				not_readin = 0;
258 				error = nfs_doio(bp, cred, p);
259 				if (error) {
260 				    brelse(bp);
261 				    return (error);
262 				}
263 			}
264 		}
265 		n = min((unsigned)(biosize - on), uio->uio_resid);
266 		offdiff = np->n_size - uio->uio_offset;
267 		if (offdiff < (off_t)n)
268 			n = (int)offdiff;
269 		if (not_readin && n > 0) {
270 			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
271 				if (!got_buf) {
272 				    bp = nfs_getcacheblk(vp, bn, biosize, p);
273 				    if (!bp)
274 					return (EINTR);
275 				    got_buf = 1;
276 				}
277 				bp->b_flags |= B_INVAFTERWRITE;
278 				if (bp->b_dirtyend > 0) {
279 				    if ((bp->b_flags & B_DELWRI) == 0)
280 					panic("nfsbioread");
281 				    if (VOP_BWRITE(bp) == EINTR)
282 					return (EINTR);
283 				} else
284 				    brelse(bp);
285 				goto again;
286 			}
287 		}
288 		vp->v_lastr = lbn;
289 		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
290 		if (diff < n)
291 			n = diff;
292 		break;
293 	    case VLNK:
294 		nfsstats.biocache_readlinks++;
295 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
296 		if (!bp)
297 			return (EINTR);
298 		if ((bp->b_flags & B_DONE) == 0) {
299 			bp->b_flags |= B_READ;
300 			error = nfs_doio(bp, cred, p);
301 			if (error) {
302 				brelse(bp);
303 				return (error);
304 			}
305 		}
306 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
307 		got_buf = 1;
308 		on = 0;
309 		break;
310 	    case VDIR:
311 diragain:
312 		nfsstats.biocache_readdirs++;
313 		ndp = nfs_searchdircache(vp, uio->uio_offset,
314 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
315 		if (!ndp) {
316 			/*
317 			 * We've been handed a cookie that is not
318 			 * in the cache. If we're not translating
319 			 * 32 <-> 64, it may be a value that was
320 			 * flushed out of the cache because it grew
321 			 * too big. Let the server judge if it's
322 			 * valid or not. In the translation case,
323 			 * we have no way of validating this value,
324 			 * so punt.
325 			 */
326 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
327 				return (EINVAL);
328 			ndp = nfs_enterdircache(vp, uio->uio_offset,
329 				uio->uio_offset, 0, 0);
330 		}
331 
332 		if (uio->uio_offset != 0 &&
333 		    ndp->dc_cookie == np->n_direofoffset) {
334 			nfsstats.direofcache_hits++;
335 			return (0);
336 		}
337 
338 		bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p);
339 		if (!bp)
340 		    return (EINTR);
341 		if ((bp->b_flags & B_DONE) == 0) {
342 		    bp->b_flags |= B_READ;
343 		    bp->b_dcookie = ndp->dc_blkcookie;
344 		    error = nfs_doio(bp, cred, p);
345 		    if (error) {
346 			/*
347 			 * Yuck! The directory has been modified on the
348 			 * server. Punt and let the userland code
349 			 * deal with it.
350 			 */
351 			brelse(bp);
352 			if (error == NFSERR_BAD_COOKIE) {
353 			    nfs_invaldircache(vp, 0);
354 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
355 			    error = EINVAL;
356 			}
357 			return (error);
358 		    }
359 		}
360 
361 		/*
362 		 * Just return if we hit EOF right away with this
363 		 * block. Always check here, because direofoffset
364 		 * may have been set by an nfsiod since the last
365 		 * check.
366 		 */
367 		if (np->n_direofoffset != 0 &&
368 			ndp->dc_blkcookie == np->n_direofoffset) {
369 			brelse(bp);
370 			return (0);
371 		}
372 
373 		/*
374 		 * Find the entry we were looking for in the block.
375 		 */
376 
377 		en = ndp->dc_entry;
378 
379 		pdp = dp = (struct dirent *)bp->b_data;
380 		edp = bp->b_data + bp->b_validend;
381 		enn = 0;
382 		while (enn < en && (caddr_t)dp < edp) {
383 			pdp = dp;
384 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
385 			enn++;
386 		}
387 
388 		/*
389 		 * If the entry number was bigger than the number of
390 		 * entries in the block, or the cookie of the previous
391 		 * entry doesn't match, the directory cache is
392 		 * stale. Flush it and try again (i.e. go to
393 		 * the server).
394 		 */
395 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
396 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
397 #ifdef DEBUG
398 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
399 				pdp, dp, edp,
400 				(unsigned long)uio->uio_offset,
401 				(unsigned long)NFS_GETCOOKIE(pdp));
402 #endif
403 			brelse(bp);
404 			nfs_invaldircache(vp, 0);
405 			nfs_vinvalbuf(vp, 0, cred, p, 0);
406 			goto diragain;
407 		}
408 
409 		on = (caddr_t)dp - bp->b_data;
410 
411 		/*
412 		 * Cache all entries that may be exported to the
413 		 * user, as they may be thrown back at us. The
414 		 * NFSBIO_CACHECOOKIES flag indicates that all
415 		 * entries are being 'exported', so cache them all.
416 		 */
417 
418 		if (en == 0 && pdp == dp) {
419 			dp = (struct dirent *)
420 			    ((caddr_t)dp + dp->d_reclen);
421 			enn++;
422 		}
423 
424 		if (uio->uio_resid < (bp->b_validend - on)) {
425 			n = uio->uio_resid;
426 			enough = 1;
427 		} else
428 			n = bp->b_validend - on;
429 
430 		ep = bp->b_data + on + n;
431 
432 		/*
433 		 * Find last complete entry to copy, caching entries
434 		 * (if requested) as we go.
435 		 */
436 
437 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
438 			if (cflag & NFSBIO_CACHECOOKIES) {
439 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
440 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
441 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
442 					NFS_STASHCOOKIE32(pdp,
443 					    nndp->dc_cookie32);
444 				}
445 			}
446 			pdp = dp;
447 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
448 			enn++;
449 		}
450 
451 		/*
452 		 * If the last requested entry was not the last in the
453 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
454 		 * cache the cookie of the last requested one, and
455 		 * set of the offset to it.
456 		 */
457 
458 		if ((on + n) < bp->b_validend) {
459 			curoff = NFS_GETCOOKIE(pdp);
460 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
461 			    enn, bp->b_lblkno);
462 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
463 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
464 				curoff = nndp->dc_cookie32;
465 			}
466 		} else
467 			curoff = bp->b_dcookie;
468 
469 		/*
470 		 * Always cache the entry for the next block,
471 		 * so that readaheads can use it.
472 		 */
473 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
474 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
475 			if (curoff == bp->b_dcookie) {
476 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
477 				curoff = nndp->dc_cookie32;
478 			}
479 		}
480 
481 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
482 
483 		/*
484 		 * If not eof and read aheads are enabled, start one.
485 		 * (You need the current block first, so that you have the
486 		 *  directory offset cookie of the next block.)
487 		 */
488 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
489 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
490 			rabp = nfs_getcacheblk(vp, nndp->dc_blkno,
491 						NFS_DIRBLKSIZ, p);
492 			if (rabp) {
493 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
494 				rabp->b_dcookie = nndp->dc_cookie;
495 				rabp->b_flags |= (B_READ | B_ASYNC);
496 				if (nfs_asyncio(rabp, cred)) {
497 				    rabp->b_flags |= B_INVAL;
498 				    brelse(rabp);
499 				}
500 			    } else
501 				brelse(rabp);
502 			}
503 		}
504 		got_buf = 1;
505 		break;
506 	    default:
507 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
508 		break;
509 	    };
510 
511 	    if (n > 0) {
512 		if (!baddr)
513 			baddr = bp->b_data;
514 		error = uiomove(baddr + on, (int)n, uio);
515 	    }
516 	    switch (vp->v_type) {
517 	    case VREG:
518 		break;
519 	    case VLNK:
520 		n = 0;
521 		break;
522 	    case VDIR:
523 		if (np->n_flag & NQNFSNONCACHE)
524 			bp->b_flags |= B_INVAL;
525 		uio->uio_offset = curoff;
526 		if (enough)
527 			n = 0;
528 		break;
529 	    default:
530 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
531 	    }
532 	    if (got_buf)
533 		brelse(bp);
534 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
535 	return (error);
536 }
537 
538 /*
539  * Vnode op for write using bio
540  */
541 int
542 nfs_write(v)
543 	void *v;
544 {
545 	struct vop_write_args /* {
546 		struct vnode *a_vp;
547 		struct uio *a_uio;
548 		int  a_ioflag;
549 		struct ucred *a_cred;
550 	} */ *ap = v;
551 	int biosize;
552 	struct uio *uio = ap->a_uio;
553 	struct proc *p = uio->uio_procp;
554 	struct vnode *vp = ap->a_vp;
555 	struct nfsnode *np = VTONFS(vp);
556 	struct ucred *cred = ap->a_cred;
557 	int ioflag = ap->a_ioflag;
558 	struct buf *bp;
559 	struct vattr vattr;
560 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
561 	daddr_t lbn, bn;
562 	int n, on, error = 0, iomode, must_commit;
563 
564 #ifdef DIAGNOSTIC
565 	if (uio->uio_rw != UIO_WRITE)
566 		panic("nfs_write mode");
567 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
568 		panic("nfs_write proc");
569 #endif
570 	if (vp->v_type != VREG)
571 		return (EIO);
572 	if (np->n_flag & NWRITEERR) {
573 		np->n_flag &= ~NWRITEERR;
574 		return (np->n_error);
575 	}
576 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
577 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
578 		(void)nfs_fsinfo(nmp, vp, cred, p);
579 	if (ioflag & (IO_APPEND | IO_SYNC)) {
580 		if (np->n_flag & NMODIFIED) {
581 			np->n_attrstamp = 0;
582 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
583 			if (error)
584 				return (error);
585 		}
586 		if (ioflag & IO_APPEND) {
587 			np->n_attrstamp = 0;
588 			error = VOP_GETATTR(vp, &vattr, cred, p);
589 			if (error)
590 				return (error);
591 			uio->uio_offset = np->n_size;
592 		}
593 	}
594 	if (uio->uio_offset < 0)
595 		return (EINVAL);
596 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
597 		return (EFBIG);
598 	if (uio->uio_resid == 0)
599 		return (0);
600 	/*
601 	 * Maybe this should be above the vnode op call, but so long as
602 	 * file servers have no limits, i don't think it matters
603 	 */
604 	if (p && uio->uio_offset + uio->uio_resid >
605 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
606 		psignal(p, SIGXFSZ);
607 		return (EFBIG);
608 	}
609 	/*
610 	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
611 	 * will be the same size within a filesystem. nfs_writerpc will
612 	 * still use nm_wsize when sizing the rpc's.
613 	 */
614 	biosize = nmp->nm_rsize;
615 	do {
616 
617 		/*
618 		 * XXX make sure we aren't cached in the VM page cache
619 		 */
620 		(void)uvm_vnp_uncache(vp);
621 
622 		/*
623 		 * Check for a valid write lease.
624 		 */
625 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
626 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
627 			do {
628 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
629 			} while (error == NQNFS_EXPIRED);
630 			if (error)
631 				return (error);
632 			if (np->n_lrev != np->n_brev ||
633 			    (np->n_flag & NQNFSNONCACHE)) {
634 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
635 				if (error)
636 					return (error);
637 				np->n_brev = np->n_lrev;
638 			}
639 		}
640 		if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
641 		    iomode = NFSV3WRITE_FILESYNC;
642 		    error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
643 		    if (must_commit)
644 			nfs_clearcommit(vp->v_mount);
645 		    return (error);
646 		}
647 		nfsstats.biocache_writes++;
648 		lbn = uio->uio_offset / biosize;
649 		on = uio->uio_offset & (biosize-1);
650 		n = min((unsigned)(biosize - on), uio->uio_resid);
651 		bn = lbn * (biosize / DEV_BSIZE);
652 again:
653 		bp = nfs_getcacheblk(vp, bn, biosize, p);
654 		if (!bp)
655 			return (EINTR);
656 		if (bp->b_wcred == NOCRED) {
657 			crhold(cred);
658 			bp->b_wcred = cred;
659 		}
660 		np->n_flag |= NMODIFIED;
661 		if (uio->uio_offset + n > np->n_size) {
662 			np->n_size = uio->uio_offset + n;
663 			uvm_vnp_setsize(vp, np->n_size);
664 		}
665 
666 		/*
667 		 * If the new write will leave a contiguous dirty
668 		 * area, just update the b_dirtyoff and b_dirtyend,
669 		 * otherwise force a write rpc of the old dirty area.
670 		 */
671 		if (bp->b_dirtyend > 0 &&
672 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
673 			bp->b_proc = p;
674 			if (VOP_BWRITE(bp) == EINTR)
675 				return (EINTR);
676 			goto again;
677 		}
678 
679 		/*
680 		 * Check for valid write lease and get one as required.
681 		 * In case getblk() and/or bwrite() delayed us.
682 		 */
683 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
684 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
685 			do {
686 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
687 			} while (error == NQNFS_EXPIRED);
688 			if (error) {
689 				brelse(bp);
690 				return (error);
691 			}
692 			if (np->n_lrev != np->n_brev ||
693 			    (np->n_flag & NQNFSNONCACHE)) {
694 				brelse(bp);
695 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
696 				if (error)
697 					return (error);
698 				np->n_brev = np->n_lrev;
699 				goto again;
700 			}
701 		}
702 		error = uiomove((char *)bp->b_data + on, n, uio);
703 		if (error) {
704 			bp->b_flags |= B_ERROR;
705 			brelse(bp);
706 			return (error);
707 		}
708 		if (bp->b_dirtyend > 0) {
709 			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
710 			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
711 		} else {
712 			bp->b_dirtyoff = on;
713 			bp->b_dirtyend = on + n;
714 		}
715 		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
716 		    bp->b_validoff > bp->b_dirtyend) {
717 			bp->b_validoff = bp->b_dirtyoff;
718 			bp->b_validend = bp->b_dirtyend;
719 		} else {
720 			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
721 			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
722 		}
723 
724 		/*
725 		 * Since this block is being modified, it must be written
726 		 * again and not just committed.
727 		 */
728 		bp->b_flags &= ~B_NEEDCOMMIT;
729 
730 		/*
731 		 * If the lease is non-cachable or IO_SYNC do bwrite().
732 		 */
733 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
734 			bp->b_proc = p;
735 			error = VOP_BWRITE(bp);
736 			if (error)
737 				return (error);
738 			if (np->n_flag & NQNFSNONCACHE) {
739 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
740 				if (error)
741 					return (error);
742 			}
743 		} else if ((n + on) == biosize &&
744 			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
745 			bp->b_proc = (struct proc *)0;
746 			bp->b_flags |= B_ASYNC;
747 			(void)nfs_writebp(bp, 0);
748 		} else {
749 			bdwrite(bp);
750 		}
751 	} while (uio->uio_resid > 0 && n > 0);
752 	return (0);
753 }
754 
755 /*
756  * Get an nfs cache block.
757  * Allocate a new one if the block isn't currently in the cache
758  * and return the block marked busy. If the calling process is
759  * interrupted by a signal for an interruptible mount point, return
760  * NULL.
761  */
762 struct buf *
763 nfs_getcacheblk(vp, bn, size, p)
764 	struct vnode *vp;
765 	daddr_t bn;
766 	int size;
767 	struct proc *p;
768 {
769 	struct buf *bp;
770 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
771 
772 	if (nmp->nm_flag & NFSMNT_INT) {
773 		bp = getblk(vp, bn, size, PCATCH, 0);
774 		while (bp == (struct buf *)0) {
775 			if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
776 				return ((struct buf *)0);
777 			bp = getblk(vp, bn, size, 0, 2 * hz);
778 		}
779 	} else
780 		bp = getblk(vp, bn, size, 0, 0);
781 	return (bp);
782 }
783 
784 /*
785  * Flush and invalidate all dirty buffers. If another process is already
786  * doing the flush, just wait for completion.
787  */
788 int
789 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
790 	struct vnode *vp;
791 	int flags;
792 	struct ucred *cred;
793 	struct proc *p;
794 	int intrflg;
795 {
796 	struct nfsnode *np = VTONFS(vp);
797 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
798 	int error = 0, slpflag, slptimeo;
799 
800 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
801 		intrflg = 0;
802 	if (intrflg) {
803 		slpflag = PCATCH;
804 		slptimeo = 2 * hz;
805 	} else {
806 		slpflag = 0;
807 		slptimeo = 0;
808 	}
809 	/*
810 	 * First wait for any other process doing a flush to complete.
811 	 */
812 	while (np->n_flag & NFLUSHINPROG) {
813 		np->n_flag |= NFLUSHWANT;
814 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
815 			slptimeo);
816 		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
817 			return (EINTR);
818 	}
819 
820 	/*
821 	 * Now, flush as required.
822 	 */
823 	np->n_flag |= NFLUSHINPROG;
824 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
825 	while (error) {
826 		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
827 			np->n_flag &= ~NFLUSHINPROG;
828 			if (np->n_flag & NFLUSHWANT) {
829 				np->n_flag &= ~NFLUSHWANT;
830 				wakeup((caddr_t)&np->n_flag);
831 			}
832 			return (EINTR);
833 		}
834 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
835 	}
836 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
837 	if (np->n_flag & NFLUSHWANT) {
838 		np->n_flag &= ~NFLUSHWANT;
839 		wakeup((caddr_t)&np->n_flag);
840 	}
841 	return (0);
842 }
843 
844 /*
845  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
846  * This is mainly to avoid queueing async I/O requests when the nfsiods
847  * are all hung on a dead server.
848  */
849 int
850 nfs_asyncio(bp, cred)
851 	struct buf *bp;
852 	struct ucred *cred;
853 {
854 	int i;
855 	struct nfsmount *nmp;
856 	int gotiod, slpflag = 0, slptimeo = 0, error;
857 
858 	if (nfs_numasync == 0)
859 		return (EIO);
860 
861 
862 	nmp = VFSTONFS(bp->b_vp->v_mount);
863 again:
864 	if (nmp->nm_flag & NFSMNT_INT)
865 		slpflag = PCATCH;
866 	gotiod = FALSE;
867 
868 	/*
869 	 * Find a free iod to process this request.
870 	 */
871 
872 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
873 		if (nfs_iodwant[i]) {
874 			/*
875 			 * Found one, so wake it up and tell it which
876 			 * mount to process.
877 			 */
878 			nfs_iodwant[i] = (struct proc *)0;
879 			nfs_iodmount[i] = nmp;
880 			nmp->nm_bufqiods++;
881 			wakeup((caddr_t)&nfs_iodwant[i]);
882 			gotiod = TRUE;
883 			break;
884 		}
885 	/*
886 	 * If none are free, we may already have an iod working on this mount
887 	 * point.  If so, it will process our request.
888 	 */
889 	if (!gotiod && nmp->nm_bufqiods > 0)
890 		gotiod = TRUE;
891 
892 	/*
893 	 * If we have an iod which can process the request, then queue
894 	 * the buffer.
895 	 */
896 	if (gotiod) {
897 		/*
898 		 * Ensure that the queue never grows too large.
899 		 */
900 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
901 			nmp->nm_bufqwant = TRUE;
902 			error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
903 				"nfsaio", slptimeo);
904 			if (error) {
905 				if (nfs_sigintr(nmp, NULL, bp->b_proc))
906 					return (EINTR);
907 				if (slpflag == PCATCH) {
908 					slpflag = 0;
909 					slptimeo = 2 * hz;
910 				}
911 			}
912 			/*
913 			 * We might have lost our iod while sleeping,
914 			 * so check and loop if nescessary.
915 			 */
916 			if (nmp->nm_bufqiods == 0)
917 				goto again;
918 		}
919 
920 		if (bp->b_flags & B_READ) {
921 			if (bp->b_rcred == NOCRED && cred != NOCRED) {
922 				crhold(cred);
923 				bp->b_rcred = cred;
924 			}
925 		} else {
926 			bp->b_flags |= B_WRITEINPROG;
927 			if (bp->b_wcred == NOCRED && cred != NOCRED) {
928 				crhold(cred);
929 				bp->b_wcred = cred;
930 			}
931 		}
932 
933 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
934 		nmp->nm_bufqlen++;
935 		return (0);
936 	    }
937 
938 	/*
939 	 * All the iods are busy on other mounts, so return EIO to
940 	 * force the caller to process the i/o synchronously.
941 	 */
942 	return (EIO);
943 }
944 
945 /*
946  * Do an I/O operation to/from a cache block. This may be called
947  * synchronously or from an nfsiod.
948  */
949 int
950 nfs_doio(bp, cr, p)
951 	struct buf *bp;
952 	struct ucred *cr;
953 	struct proc *p;
954 {
955 	struct uio *uiop;
956 	struct vnode *vp;
957 	struct nfsnode *np;
958 	struct nfsmount *nmp;
959 	int error = 0, diff, len, iomode, must_commit = 0, s;
960 	struct uio uio;
961 	struct iovec io;
962 
963 	vp = bp->b_vp;
964 	np = VTONFS(vp);
965 	nmp = VFSTONFS(vp->v_mount);
966 	uiop = &uio;
967 	uiop->uio_iov = &io;
968 	uiop->uio_iovcnt = 1;
969 	uiop->uio_segflg = UIO_SYSSPACE;
970 	uiop->uio_procp = p;
971 
972 	/*
973 	 * Historically, paging was done with physio, but no more...
974 	 */
975 	if (bp->b_flags & B_PHYS) {
976 	    /*
977 	     * ...though reading /dev/drum still gets us here.
978 	     */
979 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
980 	    /* mapping was done by vmapbuf() */
981 	    io.iov_base = bp->b_data;
982 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
983 	    if (bp->b_flags & B_READ) {
984 		uiop->uio_rw = UIO_READ;
985 		nfsstats.read_physios++;
986 		error = nfs_readrpc(vp, uiop, cr);
987 	    } else {
988 		iomode = NFSV3WRITE_DATASYNC;
989 		uiop->uio_rw = UIO_WRITE;
990 		nfsstats.write_physios++;
991 		error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
992 	    }
993 	    if (error) {
994 		bp->b_flags |= B_ERROR;
995 		bp->b_error = error;
996 	    }
997 	} else if (bp->b_flags & B_READ) {
998 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
999 	    io.iov_base = bp->b_data;
1000 	    uiop->uio_rw = UIO_READ;
1001 	    switch (vp->v_type) {
1002 	    case VREG:
1003 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1004 		nfsstats.read_bios++;
1005 		error = nfs_readrpc(vp, uiop, cr);
1006 		if (!error) {
1007 		    bp->b_validoff = 0;
1008 		    if (uiop->uio_resid) {
1009 			/*
1010 			 * If len > 0, there is a hole in the file and
1011 			 * no writes after the hole have been pushed to
1012 			 * the server yet.
1013 			 * Just zero fill the rest of the valid area.
1014 			 */
1015 			diff = bp->b_bcount - uiop->uio_resid;
1016 			len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE
1017 				+ diff);
1018 			if (len > 0) {
1019 			    len = min(len, uiop->uio_resid);
1020 			    memset((char *)bp->b_data + diff, 0, len);
1021 			    bp->b_validend = diff + len;
1022 			} else
1023 			    bp->b_validend = diff;
1024 		    } else
1025 			bp->b_validend = bp->b_bcount;
1026 		}
1027 		if (p && (vp->v_flag & VTEXT) &&
1028 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
1029 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
1030 			  np->n_lrev != np->n_brev) ||
1031 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1032 			  np->n_mtime != np->n_vattr->va_mtime.tv_sec))) {
1033 			uprintf("Process killed due to text file modification\n");
1034 			psignal(p, SIGKILL);
1035 			p->p_holdcnt++;
1036 		}
1037 		break;
1038 	    case VLNK:
1039 		uiop->uio_offset = (off_t)0;
1040 		nfsstats.readlink_bios++;
1041 		error = nfs_readlinkrpc(vp, uiop, cr);
1042 		break;
1043 	    case VDIR:
1044 		nfsstats.readdir_bios++;
1045 		uiop->uio_offset = bp->b_dcookie;
1046 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1047 			error = nfs_readdirplusrpc(vp, uiop, cr);
1048 			if (error == NFSERR_NOTSUPP)
1049 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1050 		}
1051 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1052 			error = nfs_readdirrpc(vp, uiop, cr);
1053 		if (!error) {
1054 			bp->b_dcookie = uiop->uio_offset;
1055 			bp->b_validoff = 0;
1056 			bp->b_validend = bp->b_bcount - uiop->uio_resid;
1057 		}
1058 		break;
1059 	    default:
1060 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
1061 		break;
1062 	    };
1063 	    if (error) {
1064 		bp->b_flags |= B_ERROR;
1065 		bp->b_error = error;
1066 	    }
1067 	} else {
1068 	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
1069 		- bp->b_dirtyoff;
1070 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE
1071 		+ bp->b_dirtyoff;
1072 	    io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1073 	    uiop->uio_rw = UIO_WRITE;
1074 	    nfsstats.write_bios++;
1075 	    if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC)
1076 		iomode = NFSV3WRITE_UNSTABLE;
1077 	    else
1078 		iomode = NFSV3WRITE_FILESYNC;
1079 	    bp->b_flags |= B_WRITEINPROG;
1080 #ifdef fvdl_debug
1081 	    printf("nfs_doio(%p): bp %p doff %d dend %d\n",
1082 		vp, bp, bp->b_dirtyoff, bp->b_dirtyend);
1083 #endif
1084 	    error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1085 	    s = splbio();
1086 	    if (!error && iomode == NFSV3WRITE_UNSTABLE)
1087 		bp->b_flags |= B_NEEDCOMMIT;
1088 	    else
1089 		bp->b_flags &= ~B_NEEDCOMMIT;
1090 	    bp->b_flags &= ~B_WRITEINPROG;
1091 
1092 	    /*
1093 	     * For an interrupted write, the buffer is still valid and the
1094 	     * write hasn't been pushed to the server yet, so we can't set
1095 	     * B_ERROR and report the interruption by setting B_EINTR. For
1096 	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
1097 	     * is essentially a noop.
1098 	     * For the case of a V3 write rpc not being committed to stable
1099 	     * storage, the block is still dirty and requires either a commit
1100 	     * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC
1101 	     * before the block is reused. This is indicated by setting the
1102 	     * B_DELWRI and B_NEEDCOMMIT flags.
1103 	     */
1104 	    if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1105 		bp->b_flags |= B_DELWRI;
1106 
1107 		/*
1108 		 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the
1109 		 * buffer to the clean list, we have to reassign it back to the
1110 		 * dirty one. Ugh.
1111 		 */
1112 		if (bp->b_flags & B_ASYNC) {
1113 		    reassignbuf(bp, vp);
1114 		} else if (error)
1115 		    bp->b_flags |= B_EINTR;
1116 	    } else {
1117 		if (error) {
1118 		    bp->b_flags |= B_ERROR;
1119 		    bp->b_error = np->n_error = error;
1120 		    np->n_flag |= NWRITEERR;
1121 		}
1122 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1123 	    }
1124 	    splx(s);
1125 	}
1126 	bp->b_resid = uiop->uio_resid;
1127 	if (must_commit)
1128 		nfs_clearcommit(vp->v_mount);
1129 	biodone(bp);
1130 	return (error);
1131 }
1132