xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 21a3d2f02241c56556f4b2305ef1b8036f268f70)
1 /*	$NetBSD: nfs_bio.c,v 1.70 2001/10/13 23:25:58 simonb 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 "opt_nfs.h"
42 #include "opt_ddb.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/proc.h>
49 #include <sys/buf.h>
50 #include <sys/vnode.h>
51 #include <sys/mount.h>
52 #include <sys/kernel.h>
53 #include <sys/namei.h>
54 #include <sys/dirent.h>
55 #include <sys/malloc.h>
56 
57 #include <uvm/uvm_extern.h>
58 #include <uvm/uvm.h>
59 
60 #include <nfs/rpcv2.h>
61 #include <nfs/nfsproto.h>
62 #include <nfs/nfs.h>
63 #include <nfs/nfsmount.h>
64 #include <nfs/nqnfs.h>
65 #include <nfs/nfsnode.h>
66 #include <nfs/nfs_var.h>
67 
68 extern int nfs_numasync;
69 extern struct nfsstats nfsstats;
70 
71 /*
72  * Vnode op for read using bio
73  * Any similarity to readip() is purely coincidental
74  */
75 int
76 nfs_bioread(vp, uio, ioflag, cred, cflag)
77 	struct vnode *vp;
78 	struct uio *uio;
79 	int ioflag, cflag;
80 	struct ucred *cred;
81 {
82 	struct nfsnode *np = VTONFS(vp);
83 	struct buf *bp = NULL, *rabp;
84 	struct vattr vattr;
85 	struct proc *p;
86 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
87 	struct nfsdircache *ndp = NULL, *nndp = NULL;
88 	caddr_t baddr, ep, edp;
89 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
90 	int enough = 0;
91 	struct dirent *dp, *pdp;
92 	off_t curoff = 0;
93 
94 #ifdef DIAGNOSTIC
95 	if (uio->uio_rw != UIO_READ)
96 		panic("nfs_read mode");
97 #endif
98 	if (uio->uio_resid == 0)
99 		return (0);
100 	if (vp->v_type != VDIR && uio->uio_offset < 0)
101 		return (EINVAL);
102 	p = uio->uio_procp;
103 #ifndef NFS_V2_ONLY
104 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
105 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
106 		(void)nfs_fsinfo(nmp, vp, cred, p);
107 #endif
108 	if (vp->v_type != VDIR &&
109 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
110 		return (EFBIG);
111 
112 	/*
113 	 * For nfs, cache consistency can only be maintained approximately.
114 	 * Although RFC1094 does not specify the criteria, the following is
115 	 * believed to be compatible with the reference port.
116 	 * For nqnfs, full cache consistency is maintained within the loop.
117 	 * For nfs:
118 	 * If the file's modify time on the server has changed since the
119 	 * last read rpc or you have written to the file,
120 	 * you may have lost data cache consistency with the
121 	 * server, so flush all of the file's data out of the cache.
122 	 * Then force a getattr rpc to ensure that you have up to date
123 	 * attributes.
124 	 * NB: This implies that cache data can be read when up to
125 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
126 	 * attributes this could be forced by setting n_attrstamp to 0 before
127 	 * the VOP_GETATTR() call.
128 	 */
129 
130 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
131 		if (np->n_flag & NMODIFIED) {
132 			if (vp->v_type != VREG) {
133 				if (vp->v_type != VDIR)
134 					panic("nfs: bioread, not dir");
135 				nfs_invaldircache(vp, 0);
136 				np->n_direofoffset = 0;
137 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
138 				if (error)
139 					return (error);
140 			}
141 			np->n_attrstamp = 0;
142 			error = VOP_GETATTR(vp, &vattr, cred, p);
143 			if (error)
144 				return (error);
145 			np->n_mtime = vattr.va_mtime.tv_sec;
146 		} else {
147 			error = VOP_GETATTR(vp, &vattr, cred, p);
148 			if (error)
149 				return (error);
150 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
151 				if (vp->v_type == VDIR) {
152 					nfs_invaldircache(vp, 0);
153 					np->n_direofoffset = 0;
154 				}
155 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
156 				if (error)
157 					return (error);
158 				np->n_mtime = vattr.va_mtime.tv_sec;
159 			}
160 		}
161 	}
162 
163 	/*
164 	 * update the cached read creds for this node.
165 	 */
166 
167 	if (np->n_rcred) {
168 		crfree(np->n_rcred);
169 	}
170 	np->n_rcred = cred;
171 	crhold(cred);
172 
173 	do {
174 #ifndef NFS_V2_ONLY
175 	    /*
176 	     * Get a valid lease. If cached data is stale, flush it.
177 	     */
178 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
179 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
180 		    do {
181 			error = nqnfs_getlease(vp, ND_READ, cred, p);
182 		    } while (error == NQNFS_EXPIRED);
183 		    if (error)
184 			return (error);
185 		    if (np->n_lrev != np->n_brev ||
186 			(np->n_flag & NQNFSNONCACHE) ||
187 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
188 			if (vp->v_type == VDIR) {
189 				nfs_invaldircache(vp, 0);
190 				np->n_direofoffset = 0;
191 			}
192 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
193 			if (error)
194 			    return (error);
195 			np->n_brev = np->n_lrev;
196 		    }
197 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
198 		    nfs_invaldircache(vp, 0);
199 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
200 		    np->n_direofoffset = 0;
201 		    if (error)
202 			return (error);
203 		}
204 	    }
205 #endif
206 	    /*
207 	     * Don't cache symlinks.
208 	     */
209 	    if (np->n_flag & NQNFSNONCACHE
210 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
211 		switch (vp->v_type) {
212 		case VREG:
213 			return (nfs_readrpc(vp, uio));
214 		case VLNK:
215 			return (nfs_readlinkrpc(vp, uio, cred));
216 		case VDIR:
217 			break;
218 		default:
219 			printf(" NQNFSNONCACHE: type %x unexpected\n",
220 			    vp->v_type);
221 		};
222 	    }
223 	    baddr = (caddr_t)0;
224 	    switch (vp->v_type) {
225 	    case VREG:
226 		nfsstats.biocache_reads++;
227 
228 		error = 0;
229 		if (uio->uio_offset >= np->n_size) {
230 			break;
231 		}
232 		while (uio->uio_resid > 0) {
233 			void *win;
234 			vsize_t bytelen = MIN(np->n_size - uio->uio_offset,
235 					      uio->uio_resid);
236 
237 			if (bytelen == 0)
238 				break;
239 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
240 					&bytelen, UBC_READ);
241 			error = uiomove(win, bytelen, uio);
242 			ubc_release(win, 0);
243 			if (error) {
244 				break;
245 			}
246 		}
247 		n = 0;
248 		break;
249 
250 	    case VLNK:
251 		nfsstats.biocache_readlinks++;
252 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
253 		if (!bp)
254 			return (EINTR);
255 		if ((bp->b_flags & B_DONE) == 0) {
256 			bp->b_flags |= B_READ;
257 			error = nfs_doio(bp, p);
258 			if (error) {
259 				brelse(bp);
260 				return (error);
261 			}
262 		}
263 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
264 		got_buf = 1;
265 		on = 0;
266 		break;
267 	    case VDIR:
268 diragain:
269 		nfsstats.biocache_readdirs++;
270 		ndp = nfs_searchdircache(vp, uio->uio_offset,
271 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
272 		if (!ndp) {
273 			/*
274 			 * We've been handed a cookie that is not
275 			 * in the cache. If we're not translating
276 			 * 32 <-> 64, it may be a value that was
277 			 * flushed out of the cache because it grew
278 			 * too big. Let the server judge if it's
279 			 * valid or not. In the translation case,
280 			 * we have no way of validating this value,
281 			 * so punt.
282 			 */
283 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
284 				return (EINVAL);
285 			ndp = nfs_enterdircache(vp, uio->uio_offset,
286 				uio->uio_offset, 0, 0);
287 		}
288 
289 		if (uio->uio_offset != 0 &&
290 		    ndp->dc_cookie == np->n_direofoffset) {
291 			nfsstats.direofcache_hits++;
292 			return (0);
293 		}
294 
295 		bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p);
296 		if (!bp)
297 		    return (EINTR);
298 		if ((bp->b_flags & B_DONE) == 0) {
299 		    bp->b_flags |= B_READ;
300 		    bp->b_dcookie = ndp->dc_blkcookie;
301 		    error = nfs_doio(bp, p);
302 		    if (error) {
303 			/*
304 			 * Yuck! The directory has been modified on the
305 			 * server. Punt and let the userland code
306 			 * deal with it.
307 			 */
308 			brelse(bp);
309 			if (error == NFSERR_BAD_COOKIE) {
310 			    nfs_invaldircache(vp, 0);
311 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
312 			    error = EINVAL;
313 			}
314 			return (error);
315 		    }
316 		}
317 
318 		/*
319 		 * Just return if we hit EOF right away with this
320 		 * block. Always check here, because direofoffset
321 		 * may have been set by an nfsiod since the last
322 		 * check.
323 		 */
324 		if (np->n_direofoffset != 0 &&
325 			ndp->dc_blkcookie == np->n_direofoffset) {
326 			brelse(bp);
327 			return (0);
328 		}
329 
330 		/*
331 		 * Find the entry we were looking for in the block.
332 		 */
333 
334 		en = ndp->dc_entry;
335 
336 		pdp = dp = (struct dirent *)bp->b_data;
337 		edp = bp->b_data + bp->b_bcount - bp->b_resid;
338 		enn = 0;
339 		while (enn < en && (caddr_t)dp < edp) {
340 			pdp = dp;
341 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
342 			enn++;
343 		}
344 
345 		/*
346 		 * If the entry number was bigger than the number of
347 		 * entries in the block, or the cookie of the previous
348 		 * entry doesn't match, the directory cache is
349 		 * stale. Flush it and try again (i.e. go to
350 		 * the server).
351 		 */
352 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
353 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
354 #ifdef DEBUG
355 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
356 				pdp, dp, edp,
357 				(unsigned long)uio->uio_offset,
358 				(unsigned long)NFS_GETCOOKIE(pdp));
359 #endif
360 			brelse(bp);
361 			nfs_invaldircache(vp, 0);
362 			nfs_vinvalbuf(vp, 0, cred, p, 0);
363 			goto diragain;
364 		}
365 
366 		on = (caddr_t)dp - bp->b_data;
367 
368 		/*
369 		 * Cache all entries that may be exported to the
370 		 * user, as they may be thrown back at us. The
371 		 * NFSBIO_CACHECOOKIES flag indicates that all
372 		 * entries are being 'exported', so cache them all.
373 		 */
374 
375 		if (en == 0 && pdp == dp) {
376 			dp = (struct dirent *)
377 			    ((caddr_t)dp + dp->d_reclen);
378 			enn++;
379 		}
380 
381 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
382 			n = uio->uio_resid;
383 			enough = 1;
384 		} else
385 			n = bp->b_bcount - bp->b_resid - on;
386 
387 		ep = bp->b_data + on + n;
388 
389 		/*
390 		 * Find last complete entry to copy, caching entries
391 		 * (if requested) as we go.
392 		 */
393 
394 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
395 			if (cflag & NFSBIO_CACHECOOKIES) {
396 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
397 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
398 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
399 					NFS_STASHCOOKIE32(pdp,
400 					    nndp->dc_cookie32);
401 				}
402 			}
403 			pdp = dp;
404 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
405 			enn++;
406 		}
407 
408 		/*
409 		 * If the last requested entry was not the last in the
410 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
411 		 * cache the cookie of the last requested one, and
412 		 * set of the offset to it.
413 		 */
414 
415 		if ((on + n) < bp->b_bcount - bp->b_resid) {
416 			curoff = NFS_GETCOOKIE(pdp);
417 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
418 			    enn, bp->b_lblkno);
419 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
420 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
421 				curoff = nndp->dc_cookie32;
422 			}
423 		} else
424 			curoff = bp->b_dcookie;
425 
426 		/*
427 		 * Always cache the entry for the next block,
428 		 * so that readaheads can use it.
429 		 */
430 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
431 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
432 			if (curoff == bp->b_dcookie) {
433 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
434 				curoff = nndp->dc_cookie32;
435 			}
436 		}
437 
438 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
439 
440 		/*
441 		 * If not eof and read aheads are enabled, start one.
442 		 * (You need the current block first, so that you have the
443 		 *  directory offset cookie of the next block.)
444 		 */
445 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
446 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
447 			rabp = nfs_getcacheblk(vp, nndp->dc_blkno,
448 						NFS_DIRBLKSIZ, p);
449 			if (rabp) {
450 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
451 				rabp->b_dcookie = nndp->dc_cookie;
452 				rabp->b_flags |= (B_READ | B_ASYNC);
453 				if (nfs_asyncio(rabp)) {
454 				    rabp->b_flags |= B_INVAL;
455 				    brelse(rabp);
456 				}
457 			    } else
458 				brelse(rabp);
459 			}
460 		}
461 		got_buf = 1;
462 		break;
463 	    default:
464 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
465 		break;
466 	    }
467 
468 	    if (n > 0) {
469 		if (!baddr)
470 			baddr = bp->b_data;
471 		error = uiomove(baddr + on, (int)n, uio);
472 	    }
473 	    switch (vp->v_type) {
474 	    case VREG:
475 		break;
476 	    case VLNK:
477 		n = 0;
478 		break;
479 	    case VDIR:
480 		if (np->n_flag & NQNFSNONCACHE)
481 			bp->b_flags |= B_INVAL;
482 		uio->uio_offset = curoff;
483 		if (enough)
484 			n = 0;
485 		break;
486 	    default:
487 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
488 	    }
489 	    if (got_buf)
490 		brelse(bp);
491 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
492 	return (error);
493 }
494 
495 /*
496  * Vnode op for write using bio
497  */
498 int
499 nfs_write(v)
500 	void *v;
501 {
502 	struct vop_write_args /* {
503 		struct vnode *a_vp;
504 		struct uio *a_uio;
505 		int  a_ioflag;
506 		struct ucred *a_cred;
507 	} */ *ap = v;
508 	struct uio *uio = ap->a_uio;
509 	struct proc *p = uio->uio_procp;
510 	struct vnode *vp = ap->a_vp;
511 	struct nfsnode *np = VTONFS(vp);
512 	struct ucred *cred = ap->a_cred;
513 	int ioflag = ap->a_ioflag;
514 	struct vattr vattr;
515 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
516 	void *win;
517 	voff_t oldoff, origoff;
518 	vsize_t bytelen;
519 	int error = 0, iomode, must_commit;
520 
521 #ifdef DIAGNOSTIC
522 	if (uio->uio_rw != UIO_WRITE)
523 		panic("nfs_write mode");
524 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
525 		panic("nfs_write proc");
526 #endif
527 	if (vp->v_type != VREG)
528 		return (EIO);
529 	if (np->n_flag & NWRITEERR) {
530 		np->n_flag &= ~NWRITEERR;
531 		return (np->n_error);
532 	}
533 #ifndef NFS_V2_ONLY
534 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
535 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
536 		(void)nfs_fsinfo(nmp, vp, cred, p);
537 #endif
538 	if (ioflag & (IO_APPEND | IO_SYNC)) {
539 		if (np->n_flag & NMODIFIED) {
540 			np->n_attrstamp = 0;
541 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
542 			if (error)
543 				return (error);
544 		}
545 		if (ioflag & IO_APPEND) {
546 			np->n_attrstamp = 0;
547 			error = VOP_GETATTR(vp, &vattr, cred, p);
548 			if (error)
549 				return (error);
550 			uio->uio_offset = np->n_size;
551 		}
552 	}
553 	if (uio->uio_offset < 0)
554 		return (EINVAL);
555 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
556 		return (EFBIG);
557 	if (uio->uio_resid == 0)
558 		return (0);
559 	/*
560 	 * Maybe this should be above the vnode op call, but so long as
561 	 * file servers have no limits, i don't think it matters
562 	 */
563 	if (p && uio->uio_offset + uio->uio_resid >
564 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
565 		psignal(p, SIGXFSZ);
566 		return (EFBIG);
567 	}
568 
569 	/*
570 	 * update the cached write creds for this node.
571 	 */
572 
573 	if (np->n_wcred) {
574 		crfree(np->n_wcred);
575 	}
576 	np->n_wcred = cred;
577 	crhold(cred);
578 
579 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
580 		iomode = NFSV3WRITE_FILESYNC;
581 		error = nfs_writerpc(vp, uio, &iomode, &must_commit);
582 		if (must_commit)
583 			nfs_clearcommit(vp->v_mount);
584 		return (error);
585 	}
586 
587 	origoff = uio->uio_offset;
588 	do {
589 		oldoff = uio->uio_offset;
590 		bytelen = uio->uio_resid;
591 
592 #ifndef NFS_V2_ONLY
593 		/*
594 		 * Check for a valid write lease.
595 		 */
596 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
597 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
598 			do {
599 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
600 			} while (error == NQNFS_EXPIRED);
601 			if (error)
602 				return (error);
603 			if (np->n_lrev != np->n_brev ||
604 			    (np->n_flag & NQNFSNONCACHE)) {
605 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
606 				if (error)
607 					return (error);
608 				np->n_brev = np->n_lrev;
609 			}
610 		}
611 #endif
612 		nfsstats.biocache_writes++;
613 
614 		np->n_flag |= NMODIFIED;
615 		if (np->n_size < uio->uio_offset + bytelen) {
616 			np->n_size = uio->uio_offset + bytelen;
617 		}
618 		if ((uio->uio_offset & PAGE_MASK) == 0 &&
619 		    ((uio->uio_offset + bytelen) & PAGE_MASK) == 0) {
620 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
621 			    UBC_WRITE | UBC_FAULTBUSY);
622 		} else {
623 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
624 			    UBC_WRITE);
625 		}
626 		error = uiomove(win, bytelen, uio);
627 		ubc_release(win, 0);
628 		if (error) {
629 			break;
630 		}
631 
632 		/*
633 		 * update UVM's notion of the size now that we've
634 		 * copied the data into the vnode's pages.
635 		 */
636 
637 		if (vp->v_size < uio->uio_offset) {
638 			uvm_vnp_setsize(vp, uio->uio_offset);
639 		}
640 
641 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
642 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
643 			simple_lock(&vp->v_uobj.vmobjlock);
644 			error = (vp->v_uobj.pgops->pgo_put)(&vp->v_uobj,
645 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
646 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
647 				       ~(nmp->nm_wsize - 1)),
648 			    PGO_CLEANIT|PGO_WEAK);
649 		}
650 	} while (uio->uio_resid > 0);
651 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
652 		simple_lock(&vp->v_uobj.vmobjlock);
653 		error = (vp->v_uobj.pgops->pgo_put)(&vp->v_uobj,
654 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
655 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
656 			       ~(nmp->nm_wsize - 1)),
657 		    PGO_CLEANIT|PGO_SYNCIO);
658 	}
659 	return error;
660 }
661 
662 /*
663  * Get an nfs cache block.
664  * Allocate a new one if the block isn't currently in the cache
665  * and return the block marked busy. If the calling process is
666  * interrupted by a signal for an interruptible mount point, return
667  * NULL.
668  */
669 struct buf *
670 nfs_getcacheblk(vp, bn, size, p)
671 	struct vnode *vp;
672 	daddr_t bn;
673 	int size;
674 	struct proc *p;
675 {
676 	struct buf *bp;
677 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
678 
679 	if (nmp->nm_flag & NFSMNT_INT) {
680 		bp = getblk(vp, bn, size, PCATCH, 0);
681 		while (bp == NULL) {
682 			if (nfs_sigintr(nmp, NULL, p))
683 				return (NULL);
684 			bp = getblk(vp, bn, size, 0, 2 * hz);
685 		}
686 	} else
687 		bp = getblk(vp, bn, size, 0, 0);
688 	return (bp);
689 }
690 
691 /*
692  * Flush and invalidate all dirty buffers. If another process is already
693  * doing the flush, just wait for completion.
694  */
695 int
696 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
697 	struct vnode *vp;
698 	int flags;
699 	struct ucred *cred;
700 	struct proc *p;
701 	int intrflg;
702 {
703 	struct nfsnode *np = VTONFS(vp);
704 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
705 	int error = 0, slpflag, slptimeo;
706 
707 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
708 		intrflg = 0;
709 	if (intrflg) {
710 		slpflag = PCATCH;
711 		slptimeo = 2 * hz;
712 	} else {
713 		slpflag = 0;
714 		slptimeo = 0;
715 	}
716 	/*
717 	 * First wait for any other process doing a flush to complete.
718 	 */
719 	while (np->n_flag & NFLUSHINPROG) {
720 		np->n_flag |= NFLUSHWANT;
721 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
722 			slptimeo);
723 		if (error && intrflg && nfs_sigintr(nmp, NULL, p))
724 			return (EINTR);
725 	}
726 
727 	/*
728 	 * Now, flush as required.
729 	 */
730 	np->n_flag |= NFLUSHINPROG;
731 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
732 	while (error) {
733 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
734 			np->n_flag &= ~NFLUSHINPROG;
735 			if (np->n_flag & NFLUSHWANT) {
736 				np->n_flag &= ~NFLUSHWANT;
737 				wakeup((caddr_t)&np->n_flag);
738 			}
739 			return (EINTR);
740 		}
741 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
742 	}
743 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
744 	if (np->n_flag & NFLUSHWANT) {
745 		np->n_flag &= ~NFLUSHWANT;
746 		wakeup((caddr_t)&np->n_flag);
747 	}
748 	return (0);
749 }
750 
751 /*
752  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
753  * This is mainly to avoid queueing async I/O requests when the nfsiods
754  * are all hung on a dead server.
755  */
756 
757 int
758 nfs_asyncio(bp)
759 	struct buf *bp;
760 {
761 	int i;
762 	struct nfsmount *nmp;
763 	int gotiod, slpflag = 0, slptimeo = 0, error;
764 
765 	if (nfs_numasync == 0)
766 		return (EIO);
767 
768 
769 	nmp = VFSTONFS(bp->b_vp->v_mount);
770 again:
771 	if (nmp->nm_flag & NFSMNT_INT)
772 		slpflag = PCATCH;
773 	gotiod = FALSE;
774 
775 	/*
776 	 * Find a free iod to process this request.
777 	 */
778 
779 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
780 		if (nfs_iodwant[i]) {
781 			/*
782 			 * Found one, so wake it up and tell it which
783 			 * mount to process.
784 			 */
785 			nfs_iodwant[i] = NULL;
786 			nfs_iodmount[i] = nmp;
787 			nmp->nm_bufqiods++;
788 			wakeup((caddr_t)&nfs_iodwant[i]);
789 			gotiod = TRUE;
790 			break;
791 		}
792 	/*
793 	 * If none are free, we may already have an iod working on this mount
794 	 * point.  If so, it will process our request.
795 	 */
796 	if (!gotiod && nmp->nm_bufqiods > 0)
797 		gotiod = TRUE;
798 
799 	/*
800 	 * If we have an iod which can process the request, then queue
801 	 * the buffer.
802 	 */
803 	if (gotiod) {
804 		/*
805 		 * Ensure that the queue never grows too large.
806 		 */
807 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
808 			nmp->nm_bufqwant = TRUE;
809 			error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
810 				"nfsaio", slptimeo);
811 			if (error) {
812 				if (nfs_sigintr(nmp, NULL, bp->b_proc))
813 					return (EINTR);
814 				if (slpflag == PCATCH) {
815 					slpflag = 0;
816 					slptimeo = 2 * hz;
817 				}
818 			}
819 			/*
820 			 * We might have lost our iod while sleeping,
821 			 * so check and loop if nescessary.
822 			 */
823 			if (nmp->nm_bufqiods == 0)
824 				goto again;
825 		}
826 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
827 		nmp->nm_bufqlen++;
828 		return (0);
829 	    }
830 
831 	/*
832 	 * All the iods are busy on other mounts, so return EIO to
833 	 * force the caller to process the i/o synchronously.
834 	 */
835 	return (EIO);
836 }
837 
838 /*
839  * Do an I/O operation to/from a cache block. This may be called
840  * synchronously or from an nfsiod.
841  */
842 int
843 nfs_doio(bp, p)
844 	struct buf *bp;
845 	struct proc *p;
846 {
847 	struct uio *uiop;
848 	struct vnode *vp;
849 	struct nfsnode *np;
850 	struct nfsmount *nmp;
851 	int error = 0, diff, len, iomode, must_commit = 0;
852 	struct uio uio;
853 	struct iovec io;
854 
855 	vp = bp->b_vp;
856 	np = VTONFS(vp);
857 	nmp = VFSTONFS(vp->v_mount);
858 	uiop = &uio;
859 	uiop->uio_iov = &io;
860 	uiop->uio_iovcnt = 1;
861 	uiop->uio_segflg = UIO_SYSSPACE;
862 	uiop->uio_procp = p;
863 
864 	/*
865 	 * Historically, paging was done with physio, but no more...
866 	 */
867 	if (bp->b_flags & B_PHYS) {
868 	    /*
869 	     * ...though reading /dev/drum still gets us here.
870 	     */
871 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
872 	    /* mapping was done by vmapbuf() */
873 	    io.iov_base = bp->b_data;
874 	    uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
875 	    if (bp->b_flags & B_READ) {
876 		uiop->uio_rw = UIO_READ;
877 		nfsstats.read_physios++;
878 		error = nfs_readrpc(vp, uiop);
879 	    } else {
880 		iomode = NFSV3WRITE_DATASYNC;
881 		uiop->uio_rw = UIO_WRITE;
882 		nfsstats.write_physios++;
883 		error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
884 	    }
885 	    if (error) {
886 		bp->b_flags |= B_ERROR;
887 		bp->b_error = error;
888 	    }
889 	} else if (bp->b_flags & B_READ) {
890 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
891 	    io.iov_base = bp->b_data;
892 	    uiop->uio_rw = UIO_READ;
893 	    switch (vp->v_type) {
894 	    case VREG:
895 		uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
896 		nfsstats.read_bios++;
897 		error = nfs_readrpc(vp, uiop);
898 		if (!error && uiop->uio_resid) {
899 
900 			/*
901 			 * If len > 0, there is a hole in the file and
902 			 * no writes after the hole have been pushed to
903 			 * the server yet.
904 			 * Just zero fill the rest of the valid area.
905 			 */
906 
907 			diff = bp->b_bcount - uiop->uio_resid;
908 			len = np->n_size - ((((off_t)bp->b_blkno) << DEV_BSHIFT)
909 				+ diff);
910 			if (len > 0) {
911 				len = MIN(len, uiop->uio_resid);
912 				memset((char *)bp->b_data + diff, 0, len);
913 			}
914 		}
915 		if (p && (vp->v_flag & VTEXT) &&
916 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
917 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
918 			  np->n_lrev != np->n_brev) ||
919 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
920 			  np->n_mtime != np->n_vattr->va_mtime.tv_sec))) {
921 			uprintf("Process killed due to "
922 				"text file modification\n");
923 			psignal(p, SIGKILL);
924 			p->p_holdcnt++;
925 		}
926 		break;
927 	    case VLNK:
928 		uiop->uio_offset = (off_t)0;
929 		nfsstats.readlink_bios++;
930 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
931 		break;
932 	    case VDIR:
933 		nfsstats.readdir_bios++;
934 		uiop->uio_offset = bp->b_dcookie;
935 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
936 			error = nfs_readdirplusrpc(vp, uiop, curproc->p_ucred);
937 			if (error == NFSERR_NOTSUPP)
938 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
939 		}
940 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
941 			error = nfs_readdirrpc(vp, uiop, curproc->p_ucred);
942 		if (!error) {
943 			bp->b_dcookie = uiop->uio_offset;
944 		}
945 		break;
946 	    default:
947 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
948 		break;
949 	    }
950 	    if (error) {
951 		bp->b_flags |= B_ERROR;
952 		bp->b_error = error;
953 	    }
954 	} else {
955 	    /*
956 	     * If B_NEEDCOMMIT is set, a commit rpc may do the trick. If not
957 	     * an actual write will have to be scheduled.
958 	     */
959 
960 	    io.iov_base = bp->b_data;
961 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
962 	    uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
963 	    uiop->uio_rw = UIO_WRITE;
964 	    nfsstats.write_bios++;
965 	    iomode = NFSV3WRITE_UNSTABLE;
966 	    error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
967 	}
968 	bp->b_resid = uiop->uio_resid;
969 	if (must_commit)
970 		nfs_clearcommit(vp->v_mount);
971 	biodone(bp);
972 	return (error);
973 }
974 
975 /*
976  * Vnode op for VM getpages.
977  */
978 
979 int
980 nfs_getpages(v)
981 	void *v;
982 {
983 	struct vop_getpages_args /* {
984 		struct vnode *a_vp;
985 		voff_t a_offset;
986 		struct vm_page **a_m;
987 		int *a_count;
988 		int a_centeridx;
989 		vm_prot_t a_access_type;
990 		int a_advice;
991 		int a_flags;
992 	} */ *ap = v;
993 
994 	struct vnode *vp = ap->a_vp;
995 	struct uvm_object *uobj = &vp->v_uobj;
996 	struct nfsnode *np = VTONFS(vp);
997 	struct vm_page *pg, **pgs;
998 	off_t origoffset;
999 	int i, error, npages;
1000 	boolean_t v3 = NFS_ISV3(vp);
1001 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1002 	UVMHIST_FUNC("nfs_getpages"); UVMHIST_CALLED(ubchist);
1003 
1004 	/*
1005 	 * update the cached read creds for this node.
1006 	 */
1007 
1008 	if (np->n_rcred) {
1009 		crfree(np->n_rcred);
1010 	}
1011 	np->n_rcred = curproc->p_ucred;
1012 	crhold(np->n_rcred);
1013 
1014 	/*
1015 	 * call the genfs code to get the pages.
1016 	 */
1017 
1018 	npages = *ap->a_count;
1019 	error = genfs_getpages(v);
1020 	if (error || !write || !v3) {
1021 		return error;
1022 	}
1023 
1024 	/*
1025 	 * this is a write fault, update the commit info.
1026 	 */
1027 
1028 	origoffset = ap->a_offset;
1029 	pgs = ap->a_m;
1030 
1031 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1032 	nfs_del_committed_range(vp, origoffset, npages);
1033 	nfs_del_tobecommitted_range(vp, origoffset, npages);
1034 	simple_lock(&uobj->vmobjlock);
1035 	for (i = 0; i < npages; i++) {
1036 		pg = pgs[i];
1037 		if (pg == NULL || pg == PGO_DONTCARE) {
1038 			continue;
1039 		}
1040 		pg->flags &= ~(PG_NEEDCOMMIT|PG_RDONLY);
1041 	}
1042 	simple_unlock(&uobj->vmobjlock);
1043 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1044 	return 0;
1045 }
1046 
1047 int
1048 nfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1049 {
1050 	struct uvm_object *uobj = &vp->v_uobj;
1051 	struct nfsnode *np = VTONFS(vp);
1052 	off_t origoffset, commitoff;
1053 	uint32_t commitbytes;
1054 	int error, i;
1055 	int bytes;
1056 	boolean_t v3 = NFS_ISV3(vp);
1057 	boolean_t weak = flags & PGO_WEAK;
1058 	UVMHIST_FUNC("nfs_gop_write"); UVMHIST_CALLED(ubchist);
1059 
1060 	/* XXX for now, skip the v3 stuff. */
1061 	v3 = FALSE;
1062 
1063 	/*
1064 	 * for NFSv2, just write normally.
1065 	 */
1066 
1067 	if (!v3) {
1068 		return genfs_gop_write(vp, pgs, npages, flags);
1069 	}
1070 
1071 	/*
1072 	 * for NFSv3, use delayed writes and the "commit" operation
1073 	 * to avoid sync writes.
1074 	 */
1075 
1076 	origoffset = pgs[0]->offset;
1077 	bytes = npages << PAGE_SHIFT;
1078 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1079 	if (nfs_in_committed_range(vp, origoffset, bytes)) {
1080 		goto committed;
1081 	}
1082 	if (nfs_in_tobecommitted_range(vp, origoffset, bytes)) {
1083 		if (weak) {
1084 			lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1085 			return 0;
1086 		} else {
1087 			commitoff = np->n_pushlo;
1088 			commitbytes = (uint32_t)(np->n_pushhi - np->n_pushlo);
1089 			goto commit;
1090 		}
1091 	} else {
1092 		commitoff = origoffset;
1093 		commitbytes = npages << PAGE_SHIFT;
1094 	}
1095 	simple_lock(&uobj->vmobjlock);
1096 	for (i = 0; i < npages; i++) {
1097 		pgs[i]->flags |= PG_NEEDCOMMIT|PG_RDONLY;
1098 		pgs[i]->flags &= ~PG_CLEAN;
1099 	}
1100 	simple_unlock(&uobj->vmobjlock);
1101 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1102 	error = genfs_gop_write(vp, pgs, npages, flags);
1103 	if (error) {
1104 		return error;
1105 	}
1106 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1107 	if (weak) {
1108 		nfs_add_tobecommitted_range(vp, origoffset,
1109 		    npages << PAGE_SHIFT);
1110 	} else {
1111 commit:
1112 		error = nfs_commit(vp, commitoff, commitbytes, curproc);
1113 		nfs_del_tobecommitted_range(vp, commitoff, commitbytes);
1114 committed:
1115 		simple_lock(&uobj->vmobjlock);
1116 		for (i = 0; i < npages; i++) {
1117 			pgs[i]->flags &= ~(PG_NEEDCOMMIT|PG_RDONLY);
1118 		}
1119 		simple_unlock(&uobj->vmobjlock);
1120 	}
1121 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1122 	return error;
1123 }
1124