xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 76dfffe33547c37f8bdd446e3e4ab0f3c16cea4b)
1 /*	$NetBSD: nfs_bio.c,v 1.29 1996/10/13 01:39:04 christos 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 struct proc *nfs_iodwant[NFS_MAXASYNCDAEMON];
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)
74 	register struct vnode *vp;
75 	register struct uio *uio;
76 	int ioflag;
77 	struct ucred *cred;
78 {
79 	register struct nfsnode *np = VTONFS(vp);
80 	register int biosize, diff, i;
81 	struct buf *bp = NULL, *rabp;
82 	struct vattr vattr;
83 	struct proc *p;
84 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
85 	daddr_t lbn, bn, rabn;
86 	caddr_t baddr;
87 	int got_buf = 0, nra, error = 0, n = 0, on = 0, not_readin;
88 
89 #ifdef DIAGNOSTIC
90 	if (uio->uio_rw != UIO_READ)
91 		panic("nfs_read mode");
92 #endif
93 	if (uio->uio_resid == 0)
94 		return (0);
95 	if (uio->uio_offset < 0)
96 		return (EINVAL);
97 	p = uio->uio_procp;
98 	if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3)
99 		(void)nfs_fsinfo(nmp, vp, cred, p);
100 	biosize = nmp->nm_rsize;
101 	/*
102 	 * For nfs, cache consistency can only be maintained approximately.
103 	 * Although RFC1094 does not specify the criteria, the following is
104 	 * believed to be compatible with the reference port.
105 	 * For nqnfs, full cache consistency is maintained within the loop.
106 	 * For nfs:
107 	 * If the file's modify time on the server has changed since the
108 	 * last read rpc or you have written to the file,
109 	 * you may have lost data cache consistency with the
110 	 * server, so flush all of the file's data out of the cache.
111 	 * Then force a getattr rpc to ensure that you have up to date
112 	 * attributes.
113 	 * NB: This implies that cache data can be read when up to
114 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
115 	 * attributes this could be forced by setting n_attrstamp to 0 before
116 	 * the VOP_GETATTR() call.
117 	 */
118 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
119 		if (np->n_flag & NMODIFIED) {
120 			if (vp->v_type != VREG) {
121 				if (vp->v_type != VDIR)
122 					panic("nfs: bioread, not dir");
123 				nfs_invaldir(vp);
124 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
125 				if (error)
126 					return (error);
127 			}
128 			np->n_attrstamp = 0;
129 			error = VOP_GETATTR(vp, &vattr, cred, p);
130 			if (error)
131 				return (error);
132 			np->n_mtime = vattr.va_mtime.tv_sec;
133 		} else {
134 			error = VOP_GETATTR(vp, &vattr, cred, p);
135 			if (error)
136 				return (error);
137 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
138 				if (vp->v_type == VDIR)
139 					nfs_invaldir(vp);
140 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
141 				if (error)
142 					return (error);
143 				np->n_mtime = vattr.va_mtime.tv_sec;
144 			}
145 		}
146 	}
147 	do {
148 
149 	    /*
150 	     * Get a valid lease. If cached data is stale, flush it.
151 	     */
152 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
153 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
154 		    do {
155 			error = nqnfs_getlease(vp, ND_READ, cred, p);
156 		    } while (error == NQNFS_EXPIRED);
157 		    if (error)
158 			return (error);
159 		    if (np->n_lrev != np->n_brev ||
160 			(np->n_flag & NQNFSNONCACHE) ||
161 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
162 			if (vp->v_type == VDIR)
163 			    nfs_invaldir(vp);
164 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
165 			if (error)
166 			    return (error);
167 			np->n_brev = np->n_lrev;
168 		    }
169 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
170 		    nfs_invaldir(vp);
171 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
172 		    if (error)
173 			return (error);
174 		}
175 	    }
176 	    /*
177 	     * Don't cache symlinks.
178 	     */
179 	    if (np->n_flag & NQNFSNONCACHE
180 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
181 		switch (vp->v_type) {
182 		case VREG:
183 			return (nfs_readrpc(vp, uio, cred));
184 		case VLNK:
185 			return (nfs_readlinkrpc(vp, uio, cred));
186 		case VDIR:
187 			break;
188 		default:
189 			printf(" NQNFSNONCACHE: type %x unexpected\n",
190 			    vp->v_type);
191 		};
192 	    }
193 	    baddr = (caddr_t)0;
194 	    switch (vp->v_type) {
195 	    case VREG:
196 		nfsstats.biocache_reads++;
197 		lbn = uio->uio_offset / biosize;
198 		on = uio->uio_offset & (biosize - 1);
199 		bn = lbn * (biosize / DEV_BSIZE);
200 		not_readin = 1;
201 
202 		/*
203 		 * Start the read ahead(s), as required.
204 		 */
205 		if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
206 		    for (nra = 0; nra < nmp->nm_readahead &&
207 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
208 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
209 			if (!incore(vp, rabn)) {
210 			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
211 			    if (!rabp)
212 				return (EINTR);
213 			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
214 				rabp->b_flags |= (B_READ | B_ASYNC);
215 				if (nfs_asyncio(rabp, cred)) {
216 				    rabp->b_flags |= B_INVAL;
217 				    brelse(rabp);
218 				}
219 			    } else
220 				brelse(rabp);
221 			}
222 		    }
223 		}
224 
225 		/*
226 		 * If the block is in the cache and has the required data
227 		 * in a valid region, just copy it out.
228 		 * Otherwise, get the block and write back/read in,
229 		 * as required.
230 		 */
231 		if ((bp = incore(vp, bn)) &&
232 		    (bp->b_flags & (B_BUSY | B_WRITEINPROG)) ==
233 		    (B_BUSY | B_WRITEINPROG))
234 			got_buf = 0;
235 		else {
236 again:
237 			bp = nfs_getcacheblk(vp, bn, biosize, p);
238 			if (!bp)
239 				return (EINTR);
240 			got_buf = 1;
241 			if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
242 				bp->b_flags |= B_READ;
243 				not_readin = 0;
244 				error = nfs_doio(bp, cred, p);
245 				if (error) {
246 				    brelse(bp);
247 				    return (error);
248 				}
249 			}
250 		}
251 		n = min((unsigned)(biosize - on), uio->uio_resid);
252 		diff = np->n_size - uio->uio_offset;
253 		if (diff < n)
254 			n = diff;
255 		if (not_readin && n > 0) {
256 			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
257 				if (!got_buf) {
258 				    bp = nfs_getcacheblk(vp, bn, biosize, p);
259 				    if (!bp)
260 					return (EINTR);
261 				    got_buf = 1;
262 				}
263 				bp->b_flags |= B_INVAFTERWRITE;
264 				if (bp->b_dirtyend > 0) {
265 				    if ((bp->b_flags & B_DELWRI) == 0)
266 					panic("nfsbioread");
267 				    if (VOP_BWRITE(bp) == EINTR)
268 					return (EINTR);
269 				} else
270 				    brelse(bp);
271 				goto again;
272 			}
273 		}
274 		vp->v_lastr = lbn;
275 		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
276 		if (diff < n)
277 			n = diff;
278 		break;
279 	    case VLNK:
280 		nfsstats.biocache_readlinks++;
281 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
282 		if (!bp)
283 			return (EINTR);
284 		if ((bp->b_flags & B_DONE) == 0) {
285 			bp->b_flags |= B_READ;
286 			error = nfs_doio(bp, cred, p);
287 			if (error) {
288 				brelse(bp);
289 				return (error);
290 			}
291 		}
292 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
293 		got_buf = 1;
294 		on = 0;
295 		break;
296 	    case VDIR:
297 		if (uio->uio_resid < NFS_READDIRBLKSIZ)
298 			return (0);
299 		nfsstats.biocache_readdirs++;
300 		lbn = uio->uio_offset / NFS_DIRBLKSIZ;
301 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
302 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p);
303 		if (!bp)
304 		    return (EINTR);
305 		if ((bp->b_flags & B_DONE) == 0) {
306 		    bp->b_flags |= B_READ;
307 		    error = nfs_doio(bp, cred, p);
308 		    if (error) {
309 			brelse(bp);
310 			while (error == NFSERR_BAD_COOKIE) {
311 			    nfs_invaldir(vp);
312 			    error = nfs_vinvalbuf(vp, 0, cred, p, 1);
313 			    /*
314 			     * Yuck! The directory has been modified on the
315 			     * server. The only way to get the block is by
316 			     * reading from the beginning to get all the
317 			     * offset cookies.
318 			     */
319 			    for (i = 0; i <= lbn && !error; i++) {
320 				bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p);
321 				if (!bp)
322 				    return (EINTR);
323 				if ((bp->b_flags & B_DONE) == 0) {
324 				    bp->b_flags |= B_READ;
325 				    error = nfs_doio(bp, cred, p);
326 				    if (error)
327 					brelse(bp);
328 				}
329 			    }
330 			}
331 			if (error)
332 			    return (error);
333 		    }
334 		}
335 
336 		/*
337 		 * If not eof and read aheads are enabled, start one.
338 		 * (You need the current block first, so that you have the
339 		 *  directory offset cookie of the next block.)
340 		 */
341 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
342 		    (np->n_direofoffset == 0 ||
343 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
344 		    !(np->n_flag & NQNFSNONCACHE) &&
345 		    !incore(vp, lbn + 1)) {
346 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p);
347 			if (rabp) {
348 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
349 				rabp->b_flags |= (B_READ | B_ASYNC);
350 				if (nfs_asyncio(rabp, cred)) {
351 				    rabp->b_flags |= B_INVAL;
352 				    brelse(rabp);
353 				}
354 			    } else
355 				brelse(rabp);
356 			}
357 		}
358 		n = min(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
359 		got_buf = 1;
360 		break;
361 	    default:
362 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
363 		break;
364 	    };
365 
366 	    if (n > 0) {
367 		if (!baddr)
368 			baddr = bp->b_data;
369 		error = uiomove(baddr + on, (int)n, uio);
370 	    }
371 	    switch (vp->v_type) {
372 	    case VREG:
373 		break;
374 	    case VLNK:
375 		n = 0;
376 		break;
377 	    case VDIR:
378 		if (np->n_flag & NQNFSNONCACHE)
379 			bp->b_flags |= B_INVAL;
380 		break;
381 	    default:
382 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
383 	    }
384 	    if (got_buf)
385 		brelse(bp);
386 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
387 	return (error);
388 }
389 
390 /*
391  * Vnode op for write using bio
392  */
393 int
394 nfs_write(v)
395 	void *v;
396 {
397 	struct vop_write_args /* {
398 		struct vnode *a_vp;
399 		struct uio *a_uio;
400 		int  a_ioflag;
401 		struct ucred *a_cred;
402 	} */ *ap = v;
403 	register int biosize;
404 	register struct uio *uio = ap->a_uio;
405 	struct proc *p = uio->uio_procp;
406 	register struct vnode *vp = ap->a_vp;
407 	struct nfsnode *np = VTONFS(vp);
408 	register struct ucred *cred = ap->a_cred;
409 	int ioflag = ap->a_ioflag;
410 	struct buf *bp;
411 	struct vattr vattr;
412 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
413 	daddr_t lbn, bn;
414 	int n, on, error = 0, iomode, must_commit;
415 
416 #ifdef DIAGNOSTIC
417 	if (uio->uio_rw != UIO_WRITE)
418 		panic("nfs_write mode");
419 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
420 		panic("nfs_write proc");
421 #endif
422 	if (vp->v_type != VREG)
423 		return (EIO);
424 	if (np->n_flag & NWRITEERR) {
425 		np->n_flag &= ~NWRITEERR;
426 		return (np->n_error);
427 	}
428 	if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3)
429 		(void)nfs_fsinfo(nmp, vp, cred, p);
430 	if (ioflag & (IO_APPEND | IO_SYNC)) {
431 		if (np->n_flag & NMODIFIED) {
432 			np->n_attrstamp = 0;
433 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
434 			if (error)
435 				return (error);
436 		}
437 		if (ioflag & IO_APPEND) {
438 			np->n_attrstamp = 0;
439 			error = VOP_GETATTR(vp, &vattr, cred, p);
440 			if (error)
441 				return (error);
442 			uio->uio_offset = np->n_size;
443 		}
444 	}
445 	if (uio->uio_offset < 0)
446 		return (EINVAL);
447 	if (uio->uio_resid == 0)
448 		return (0);
449 	/*
450 	 * Maybe this should be above the vnode op call, but so long as
451 	 * file servers have no limits, i don't think it matters
452 	 */
453 	if (p && uio->uio_offset + uio->uio_resid >
454 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
455 		psignal(p, SIGXFSZ);
456 		return (EFBIG);
457 	}
458 	/*
459 	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
460 	 * will be the same size within a filesystem. nfs_writerpc will
461 	 * still use nm_wsize when sizing the rpc's.
462 	 */
463 	biosize = nmp->nm_rsize;
464 	do {
465 
466 		/*
467 		 * XXX make sure we aren't cached in the VM page cache
468 		 */
469 		(void)vnode_pager_uncache(vp);
470 
471 		/*
472 		 * Check for a valid write lease.
473 		 */
474 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
475 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
476 			do {
477 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
478 			} while (error == NQNFS_EXPIRED);
479 			if (error)
480 				return (error);
481 			if (np->n_lrev != np->n_brev ||
482 			    (np->n_flag & NQNFSNONCACHE)) {
483 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
484 				if (error)
485 					return (error);
486 				np->n_brev = np->n_lrev;
487 			}
488 		}
489 		if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
490 		    iomode = NFSV3WRITE_FILESYNC;
491 		    error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
492 		    if (must_commit)
493 			nfs_clearcommit(vp->v_mount);
494 		    return (error);
495 		}
496 		nfsstats.biocache_writes++;
497 		lbn = uio->uio_offset / biosize;
498 		on = uio->uio_offset & (biosize-1);
499 		n = min((unsigned)(biosize - on), uio->uio_resid);
500 		bn = lbn * (biosize / DEV_BSIZE);
501 again:
502 		bp = nfs_getcacheblk(vp, bn, biosize, p);
503 		if (!bp)
504 			return (EINTR);
505 		if (bp->b_wcred == NOCRED) {
506 			crhold(cred);
507 			bp->b_wcred = cred;
508 		}
509 		np->n_flag |= NMODIFIED;
510 		if (uio->uio_offset + n > np->n_size) {
511 			np->n_size = uio->uio_offset + n;
512 			vnode_pager_setsize(vp, (u_long)np->n_size);
513 		}
514 
515 		/*
516 		 * If the new write will leave a contiguous dirty
517 		 * area, just update the b_dirtyoff and b_dirtyend,
518 		 * otherwise force a write rpc of the old dirty area.
519 		 */
520 		if (bp->b_dirtyend > 0 &&
521 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
522 			bp->b_proc = p;
523 			if (VOP_BWRITE(bp) == EINTR)
524 				return (EINTR);
525 			goto again;
526 		}
527 
528 		/*
529 		 * Check for valid write lease and get one as required.
530 		 * In case getblk() and/or bwrite() delayed us.
531 		 */
532 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
533 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
534 			do {
535 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
536 			} while (error == NQNFS_EXPIRED);
537 			if (error) {
538 				brelse(bp);
539 				return (error);
540 			}
541 			if (np->n_lrev != np->n_brev ||
542 			    (np->n_flag & NQNFSNONCACHE)) {
543 				brelse(bp);
544 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
545 				if (error)
546 					return (error);
547 				np->n_brev = np->n_lrev;
548 				goto again;
549 			}
550 		}
551 		error = uiomove((char *)bp->b_data + on, n, uio);
552 		if (error) {
553 			bp->b_flags |= B_ERROR;
554 			brelse(bp);
555 			return (error);
556 		}
557 		if (bp->b_dirtyend > 0) {
558 			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
559 			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
560 		} else {
561 			bp->b_dirtyoff = on;
562 			bp->b_dirtyend = on + n;
563 		}
564 		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
565 		    bp->b_validoff > bp->b_dirtyend) {
566 			bp->b_validoff = bp->b_dirtyoff;
567 			bp->b_validend = bp->b_dirtyend;
568 		} else {
569 			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
570 			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
571 		}
572 
573 		/*
574 		 * Since this block is being modified, it must be written
575 		 * again and not just committed.
576 		 */
577 		bp->b_flags &= ~B_NEEDCOMMIT;
578 
579 		/*
580 		 * If the lease is non-cachable or IO_SYNC do bwrite().
581 		 */
582 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
583 			bp->b_proc = p;
584 			error = VOP_BWRITE(bp);
585 			if (error)
586 				return (error);
587 			if (np->n_flag & NQNFSNONCACHE) {
588 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
589 				if (error)
590 					return (error);
591 			}
592 		} else if ((n + on) == biosize &&
593 			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
594 			bp->b_proc = (struct proc *)0;
595 			bp->b_flags |= B_ASYNC;
596 			(void)nfs_writebp(bp, 0);
597 		} else {
598 			bdwrite(bp);
599 		}
600 	} while (uio->uio_resid > 0 && n > 0);
601 	return (0);
602 }
603 
604 /*
605  * Get an nfs cache block.
606  * Allocate a new one if the block isn't currently in the cache
607  * and return the block marked busy. If the calling process is
608  * interrupted by a signal for an interruptible mount point, return
609  * NULL.
610  */
611 struct buf *
612 nfs_getcacheblk(vp, bn, size, p)
613 	struct vnode *vp;
614 	daddr_t bn;
615 	int size;
616 	struct proc *p;
617 {
618 	register struct buf *bp;
619 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
620 
621 	if (nmp->nm_flag & NFSMNT_INT) {
622 		bp = getblk(vp, bn, size, PCATCH, 0);
623 		while (bp == (struct buf *)0) {
624 			if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
625 				return ((struct buf *)0);
626 			bp = getblk(vp, bn, size, 0, 2 * hz);
627 		}
628 	} else
629 		bp = getblk(vp, bn, size, 0, 0);
630 	return (bp);
631 }
632 
633 /*
634  * Flush and invalidate all dirty buffers. If another process is already
635  * doing the flush, just wait for completion.
636  */
637 int
638 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
639 	struct vnode *vp;
640 	int flags;
641 	struct ucred *cred;
642 	struct proc *p;
643 	int intrflg;
644 {
645 	register struct nfsnode *np = VTONFS(vp);
646 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
647 	int error = 0, slpflag, slptimeo;
648 
649 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
650 		intrflg = 0;
651 	if (intrflg) {
652 		slpflag = PCATCH;
653 		slptimeo = 2 * hz;
654 	} else {
655 		slpflag = 0;
656 		slptimeo = 0;
657 	}
658 	/*
659 	 * First wait for any other process doing a flush to complete.
660 	 */
661 	while (np->n_flag & NFLUSHINPROG) {
662 		np->n_flag |= NFLUSHWANT;
663 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
664 			slptimeo);
665 		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
666 			return (EINTR);
667 	}
668 
669 	/*
670 	 * Now, flush as required.
671 	 */
672 	np->n_flag |= NFLUSHINPROG;
673 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
674 	while (error) {
675 		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
676 			np->n_flag &= ~NFLUSHINPROG;
677 			if (np->n_flag & NFLUSHWANT) {
678 				np->n_flag &= ~NFLUSHWANT;
679 				wakeup((caddr_t)&np->n_flag);
680 			}
681 			return (EINTR);
682 		}
683 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
684 	}
685 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
686 	if (np->n_flag & NFLUSHWANT) {
687 		np->n_flag &= ~NFLUSHWANT;
688 		wakeup((caddr_t)&np->n_flag);
689 	}
690 	return (0);
691 }
692 
693 /*
694  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
695  * This is mainly to avoid queueing async I/O requests when the nfsiods
696  * are all hung on a dead server.
697  */
698 int
699 nfs_asyncio(bp, cred)
700 	register struct buf *bp;
701 	struct ucred *cred;
702 {
703 	register int i;
704 
705 	if (nfs_numasync == 0)
706 		return (EIO);
707 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
708 	    if (nfs_iodwant[i]) {
709 		if (bp->b_flags & B_READ) {
710 			if (bp->b_rcred == NOCRED && cred != NOCRED) {
711 				crhold(cred);
712 				bp->b_rcred = cred;
713 			}
714 		} else {
715 			bp->b_flags |= B_WRITEINPROG;
716 			if (bp->b_wcred == NOCRED && cred != NOCRED) {
717 				crhold(cred);
718 				bp->b_wcred = cred;
719 			}
720 		}
721 
722 		TAILQ_INSERT_TAIL(&nfs_bufq, bp, b_freelist);
723 		nfs_iodwant[i] = (struct proc *)0;
724 		wakeup((caddr_t)&nfs_iodwant[i]);
725 		return (0);
726 	    }
727 
728 	/*
729 	 * If it is a read or a write already marked B_WRITEINPROG or B_NOCACHE
730 	 * return EIO so the process will call nfs_doio() and do it
731 	 * synchronously.
732 	 */
733 	if (bp->b_flags & (B_READ | B_WRITEINPROG | B_NOCACHE))
734 		return (EIO);
735 
736 	/*
737 	 * Just turn the async write into a delayed write, instead of
738 	 * doing in synchronously. Hopefully, at least one of the nfsiods
739 	 * is currently doing a write for this file and will pick up the
740 	 * delayed writes before going back to sleep.
741 	 */
742 	bp->b_flags |= B_DELWRI;
743 	reassignbuf(bp, bp->b_vp);
744 	biodone(bp);
745 	return (0);
746 }
747 
748 /*
749  * Do an I/O operation to/from a cache block. This may be called
750  * synchronously or from an nfsiod.
751  */
752 int
753 nfs_doio(bp, cr, p)
754 	register struct buf *bp;
755 	struct ucred *cr;
756 	struct proc *p;
757 {
758 	register struct uio *uiop;
759 	register struct vnode *vp;
760 	struct nfsnode *np;
761 	struct nfsmount *nmp;
762 	int error = 0, diff, len, iomode, must_commit = 0;
763 	struct uio uio;
764 	struct iovec io;
765 
766 	vp = bp->b_vp;
767 	np = VTONFS(vp);
768 	nmp = VFSTONFS(vp->v_mount);
769 	uiop = &uio;
770 	uiop->uio_iov = &io;
771 	uiop->uio_iovcnt = 1;
772 	uiop->uio_segflg = UIO_SYSSPACE;
773 	uiop->uio_procp = p;
774 
775 	/*
776 	 * Historically, paging was done with physio, but no more...
777 	 */
778 	if (bp->b_flags & B_PHYS) {
779 	    /*
780 	     * ...though reading /dev/drum still gets us here.
781 	     */
782 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
783 	    /* mapping was done by vmapbuf() */
784 	    io.iov_base = bp->b_data;
785 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
786 	    if (bp->b_flags & B_READ) {
787 		uiop->uio_rw = UIO_READ;
788 		nfsstats.read_physios++;
789 		error = nfs_readrpc(vp, uiop, cr);
790 	    } else {
791 		iomode = NFSV3WRITE_DATASYNC;
792 		uiop->uio_rw = UIO_WRITE;
793 		nfsstats.write_physios++;
794 		error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
795 	    }
796 	    if (error) {
797 		bp->b_flags |= B_ERROR;
798 		bp->b_error = error;
799 	    }
800 	} else if (bp->b_flags & B_READ) {
801 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
802 	    io.iov_base = bp->b_data;
803 	    uiop->uio_rw = UIO_READ;
804 	    switch (vp->v_type) {
805 	    case VREG:
806 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
807 		nfsstats.read_bios++;
808 		error = nfs_readrpc(vp, uiop, cr);
809 		if (!error) {
810 		    bp->b_validoff = 0;
811 		    if (uiop->uio_resid) {
812 			/*
813 			 * If len > 0, there is a hole in the file and
814 			 * no writes after the hole have been pushed to
815 			 * the server yet.
816 			 * Just zero fill the rest of the valid area.
817 			 */
818 			diff = bp->b_bcount - uiop->uio_resid;
819 			len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE
820 				+ diff);
821 			if (len > 0) {
822 			    len = min(len, uiop->uio_resid);
823 			    bzero((char *)bp->b_data + diff, len);
824 			    bp->b_validend = diff + len;
825 			} else
826 			    bp->b_validend = diff;
827 		    } else
828 			bp->b_validend = bp->b_bcount;
829 		}
830 		if (p && (vp->v_flag & VTEXT) &&
831 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
832 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
833 			  np->n_lrev != np->n_brev) ||
834 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
835 			  np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
836 			uprintf("Process killed due to text file modification\n");
837 			psignal(p, SIGKILL);
838 			p->p_holdcnt++;
839 		}
840 		break;
841 	    case VLNK:
842 		uiop->uio_offset = (off_t)0;
843 		nfsstats.readlink_bios++;
844 		error = nfs_readlinkrpc(vp, uiop, cr);
845 		break;
846 	    case VDIR:
847 		nfsstats.readdir_bios++;
848 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
849 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
850 			error = nfs_readdirplusrpc(vp, uiop, cr);
851 			if (error == NFSERR_NOTSUPP)
852 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
853 		}
854 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
855 			error = nfs_readdirrpc(vp, uiop, cr);
856 		break;
857 	    default:
858 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
859 		break;
860 	    };
861 	    if (error) {
862 		bp->b_flags |= B_ERROR;
863 		bp->b_error = error;
864 	    }
865 	} else {
866 	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
867 		- bp->b_dirtyoff;
868 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE
869 		+ bp->b_dirtyoff;
870 	    io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
871 	    uiop->uio_rw = UIO_WRITE;
872 	    nfsstats.write_bios++;
873 	    if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC)
874 		iomode = NFSV3WRITE_UNSTABLE;
875 	    else
876 		iomode = NFSV3WRITE_FILESYNC;
877 	    bp->b_flags |= B_WRITEINPROG;
878 #ifdef fvdl_debug
879 	    printf("nfs_doio(%x): bp %x doff %d dend %d\n",
880 		vp, bp, bp->b_dirtyoff, bp->b_dirtyend);
881 #endif
882 	    error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
883 	    if (!error && iomode == NFSV3WRITE_UNSTABLE)
884 		bp->b_flags |= B_NEEDCOMMIT;
885 	    else
886 		bp->b_flags &= ~B_NEEDCOMMIT;
887 	    bp->b_flags &= ~B_WRITEINPROG;
888 
889 	    /*
890 	     * For an interrupted write, the buffer is still valid and the
891 	     * write hasn't been pushed to the server yet, so we can't set
892 	     * B_ERROR and report the interruption by setting B_EINTR. For
893 	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
894 	     * is essentially a noop.
895 	     * For the case of a V3 write rpc not being committed to stable
896 	     * storage, the block is still dirty and requires either a commit
897 	     * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC
898 	     * before the block is reused. This is indicated by setting the
899 	     * B_DELWRI and B_NEEDCOMMIT flags.
900 	     */
901 	    if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
902 		bp->b_flags |= B_DELWRI;
903 
904 		/*
905 		 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the
906 		 * buffer to the clean list, we have to reassign it back to the
907 		 * dirty one. Ugh.
908 		 */
909 		if (bp->b_flags & B_ASYNC)
910 		    reassignbuf(bp, vp);
911 		else if (error)
912 		    bp->b_flags |= B_EINTR;
913 	    } else {
914 		if (error) {
915 		    bp->b_flags |= B_ERROR;
916 		    bp->b_error = np->n_error = error;
917 		    np->n_flag |= NWRITEERR;
918 		}
919 		bp->b_dirtyoff = bp->b_dirtyend = 0;
920 	    }
921 	}
922 	bp->b_resid = uiop->uio_resid;
923 	if (must_commit)
924 		nfs_clearcommit(vp->v_mount);
925 	biodone(bp);
926 	return (error);
927 }
928