xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 27578b9aac214cc7796ead81dcc5427e79d5f2a0)
1 /*	$NetBSD: nfs_bio.c,v 1.69 2001/09/15 20:36:39 chs 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 	int biosize;
84 	struct buf *bp = NULL, *rabp;
85 	struct vattr vattr;
86 	struct proc *p;
87 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
88 	struct nfsdircache *ndp = NULL, *nndp = NULL;
89 	caddr_t baddr, ep, edp;
90 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
91 	int enough = 0;
92 	struct dirent *dp, *pdp;
93 	off_t curoff = 0;
94 
95 #ifdef DIAGNOSTIC
96 	if (uio->uio_rw != UIO_READ)
97 		panic("nfs_read mode");
98 #endif
99 	if (uio->uio_resid == 0)
100 		return (0);
101 	if (vp->v_type != VDIR && uio->uio_offset < 0)
102 		return (EINVAL);
103 	p = uio->uio_procp;
104 #ifndef NFS_V2_ONLY
105 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
106 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
107 		(void)nfs_fsinfo(nmp, vp, cred, p);
108 #endif
109 	if (vp->v_type != VDIR &&
110 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
111 		return (EFBIG);
112 	biosize = nmp->nm_rsize;
113 
114 	/*
115 	 * For nfs, cache consistency can only be maintained approximately.
116 	 * Although RFC1094 does not specify the criteria, the following is
117 	 * believed to be compatible with the reference port.
118 	 * For nqnfs, full cache consistency is maintained within the loop.
119 	 * For nfs:
120 	 * If the file's modify time on the server has changed since the
121 	 * last read rpc or you have written to the file,
122 	 * you may have lost data cache consistency with the
123 	 * server, so flush all of the file's data out of the cache.
124 	 * Then force a getattr rpc to ensure that you have up to date
125 	 * attributes.
126 	 * NB: This implies that cache data can be read when up to
127 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
128 	 * attributes this could be forced by setting n_attrstamp to 0 before
129 	 * the VOP_GETATTR() call.
130 	 */
131 
132 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
133 		if (np->n_flag & NMODIFIED) {
134 			if (vp->v_type != VREG) {
135 				if (vp->v_type != VDIR)
136 					panic("nfs: bioread, not dir");
137 				nfs_invaldircache(vp, 0);
138 				np->n_direofoffset = 0;
139 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
140 				if (error)
141 					return (error);
142 			}
143 			np->n_attrstamp = 0;
144 			error = VOP_GETATTR(vp, &vattr, cred, p);
145 			if (error)
146 				return (error);
147 			np->n_mtime = vattr.va_mtime.tv_sec;
148 		} else {
149 			error = VOP_GETATTR(vp, &vattr, cred, p);
150 			if (error)
151 				return (error);
152 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
153 				if (vp->v_type == VDIR) {
154 					nfs_invaldircache(vp, 0);
155 					np->n_direofoffset = 0;
156 				}
157 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
158 				if (error)
159 					return (error);
160 				np->n_mtime = vattr.va_mtime.tv_sec;
161 			}
162 		}
163 	}
164 
165 	/*
166 	 * update the cached read creds for this node.
167 	 */
168 
169 	if (np->n_rcred) {
170 		crfree(np->n_rcred);
171 	}
172 	np->n_rcred = cred;
173 	crhold(cred);
174 
175 	do {
176 #ifndef NFS_V2_ONLY
177 	    /*
178 	     * Get a valid lease. If cached data is stale, flush it.
179 	     */
180 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
181 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
182 		    do {
183 			error = nqnfs_getlease(vp, ND_READ, cred, p);
184 		    } while (error == NQNFS_EXPIRED);
185 		    if (error)
186 			return (error);
187 		    if (np->n_lrev != np->n_brev ||
188 			(np->n_flag & NQNFSNONCACHE) ||
189 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
190 			if (vp->v_type == VDIR) {
191 				nfs_invaldircache(vp, 0);
192 				np->n_direofoffset = 0;
193 			}
194 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
195 			if (error)
196 			    return (error);
197 			np->n_brev = np->n_lrev;
198 		    }
199 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
200 		    nfs_invaldircache(vp, 0);
201 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
202 		    np->n_direofoffset = 0;
203 		    if (error)
204 			return (error);
205 		}
206 	    }
207 #endif
208 	    /*
209 	     * Don't cache symlinks.
210 	     */
211 	    if (np->n_flag & NQNFSNONCACHE
212 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
213 		switch (vp->v_type) {
214 		case VREG:
215 			return (nfs_readrpc(vp, uio));
216 		case VLNK:
217 			return (nfs_readlinkrpc(vp, uio, cred));
218 		case VDIR:
219 			break;
220 		default:
221 			printf(" NQNFSNONCACHE: type %x unexpected\n",
222 			    vp->v_type);
223 		};
224 	    }
225 	    baddr = (caddr_t)0;
226 	    switch (vp->v_type) {
227 	    case VREG:
228 		nfsstats.biocache_reads++;
229 
230 		error = 0;
231 		if (uio->uio_offset >= np->n_size) {
232 			break;
233 		}
234 		while (uio->uio_resid > 0) {
235 			void *win;
236 			vsize_t bytelen = MIN(np->n_size - uio->uio_offset,
237 					      uio->uio_resid);
238 
239 			if (bytelen == 0)
240 				break;
241 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
242 					&bytelen, UBC_READ);
243 			error = uiomove(win, bytelen, uio);
244 			ubc_release(win, 0);
245 			if (error) {
246 				break;
247 			}
248 		}
249 		n = 0;
250 		break;
251 
252 	    case VLNK:
253 		nfsstats.biocache_readlinks++;
254 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
255 		if (!bp)
256 			return (EINTR);
257 		if ((bp->b_flags & B_DONE) == 0) {
258 			bp->b_flags |= B_READ;
259 			error = nfs_doio(bp, p);
260 			if (error) {
261 				brelse(bp);
262 				return (error);
263 			}
264 		}
265 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
266 		got_buf = 1;
267 		on = 0;
268 		break;
269 	    case VDIR:
270 diragain:
271 		nfsstats.biocache_readdirs++;
272 		ndp = nfs_searchdircache(vp, uio->uio_offset,
273 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
274 		if (!ndp) {
275 			/*
276 			 * We've been handed a cookie that is not
277 			 * in the cache. If we're not translating
278 			 * 32 <-> 64, it may be a value that was
279 			 * flushed out of the cache because it grew
280 			 * too big. Let the server judge if it's
281 			 * valid or not. In the translation case,
282 			 * we have no way of validating this value,
283 			 * so punt.
284 			 */
285 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
286 				return (EINVAL);
287 			ndp = nfs_enterdircache(vp, uio->uio_offset,
288 				uio->uio_offset, 0, 0);
289 		}
290 
291 		if (uio->uio_offset != 0 &&
292 		    ndp->dc_cookie == np->n_direofoffset) {
293 			nfsstats.direofcache_hits++;
294 			return (0);
295 		}
296 
297 		bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p);
298 		if (!bp)
299 		    return (EINTR);
300 		if ((bp->b_flags & B_DONE) == 0) {
301 		    bp->b_flags |= B_READ;
302 		    bp->b_dcookie = ndp->dc_blkcookie;
303 		    error = nfs_doio(bp, p);
304 		    if (error) {
305 			/*
306 			 * Yuck! The directory has been modified on the
307 			 * server. Punt and let the userland code
308 			 * deal with it.
309 			 */
310 			brelse(bp);
311 			if (error == NFSERR_BAD_COOKIE) {
312 			    nfs_invaldircache(vp, 0);
313 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
314 			    error = EINVAL;
315 			}
316 			return (error);
317 		    }
318 		}
319 
320 		/*
321 		 * Just return if we hit EOF right away with this
322 		 * block. Always check here, because direofoffset
323 		 * may have been set by an nfsiod since the last
324 		 * check.
325 		 */
326 		if (np->n_direofoffset != 0 &&
327 			ndp->dc_blkcookie == np->n_direofoffset) {
328 			brelse(bp);
329 			return (0);
330 		}
331 
332 		/*
333 		 * Find the entry we were looking for in the block.
334 		 */
335 
336 		en = ndp->dc_entry;
337 
338 		pdp = dp = (struct dirent *)bp->b_data;
339 		edp = bp->b_data + bp->b_bcount - bp->b_resid;
340 		enn = 0;
341 		while (enn < en && (caddr_t)dp < edp) {
342 			pdp = dp;
343 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
344 			enn++;
345 		}
346 
347 		/*
348 		 * If the entry number was bigger than the number of
349 		 * entries in the block, or the cookie of the previous
350 		 * entry doesn't match, the directory cache is
351 		 * stale. Flush it and try again (i.e. go to
352 		 * the server).
353 		 */
354 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
355 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
356 #ifdef DEBUG
357 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
358 				pdp, dp, edp,
359 				(unsigned long)uio->uio_offset,
360 				(unsigned long)NFS_GETCOOKIE(pdp));
361 #endif
362 			brelse(bp);
363 			nfs_invaldircache(vp, 0);
364 			nfs_vinvalbuf(vp, 0, cred, p, 0);
365 			goto diragain;
366 		}
367 
368 		on = (caddr_t)dp - bp->b_data;
369 
370 		/*
371 		 * Cache all entries that may be exported to the
372 		 * user, as they may be thrown back at us. The
373 		 * NFSBIO_CACHECOOKIES flag indicates that all
374 		 * entries are being 'exported', so cache them all.
375 		 */
376 
377 		if (en == 0 && pdp == dp) {
378 			dp = (struct dirent *)
379 			    ((caddr_t)dp + dp->d_reclen);
380 			enn++;
381 		}
382 
383 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
384 			n = uio->uio_resid;
385 			enough = 1;
386 		} else
387 			n = bp->b_bcount - bp->b_resid - on;
388 
389 		ep = bp->b_data + on + n;
390 
391 		/*
392 		 * Find last complete entry to copy, caching entries
393 		 * (if requested) as we go.
394 		 */
395 
396 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
397 			if (cflag & NFSBIO_CACHECOOKIES) {
398 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
399 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
400 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
401 					NFS_STASHCOOKIE32(pdp,
402 					    nndp->dc_cookie32);
403 				}
404 			}
405 			pdp = dp;
406 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
407 			enn++;
408 		}
409 
410 		/*
411 		 * If the last requested entry was not the last in the
412 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
413 		 * cache the cookie of the last requested one, and
414 		 * set of the offset to it.
415 		 */
416 
417 		if ((on + n) < bp->b_bcount - bp->b_resid) {
418 			curoff = NFS_GETCOOKIE(pdp);
419 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
420 			    enn, bp->b_lblkno);
421 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
422 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
423 				curoff = nndp->dc_cookie32;
424 			}
425 		} else
426 			curoff = bp->b_dcookie;
427 
428 		/*
429 		 * Always cache the entry for the next block,
430 		 * so that readaheads can use it.
431 		 */
432 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
433 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
434 			if (curoff == bp->b_dcookie) {
435 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
436 				curoff = nndp->dc_cookie32;
437 			}
438 		}
439 
440 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
441 
442 		/*
443 		 * If not eof and read aheads are enabled, start one.
444 		 * (You need the current block first, so that you have the
445 		 *  directory offset cookie of the next block.)
446 		 */
447 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
448 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
449 			rabp = nfs_getcacheblk(vp, nndp->dc_blkno,
450 						NFS_DIRBLKSIZ, p);
451 			if (rabp) {
452 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
453 				rabp->b_dcookie = nndp->dc_cookie;
454 				rabp->b_flags |= (B_READ | B_ASYNC);
455 				if (nfs_asyncio(rabp)) {
456 				    rabp->b_flags |= B_INVAL;
457 				    brelse(rabp);
458 				}
459 			    } else
460 				brelse(rabp);
461 			}
462 		}
463 		got_buf = 1;
464 		break;
465 	    default:
466 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
467 		break;
468 	    }
469 
470 	    if (n > 0) {
471 		if (!baddr)
472 			baddr = bp->b_data;
473 		error = uiomove(baddr + on, (int)n, uio);
474 	    }
475 	    switch (vp->v_type) {
476 	    case VREG:
477 		break;
478 	    case VLNK:
479 		n = 0;
480 		break;
481 	    case VDIR:
482 		if (np->n_flag & NQNFSNONCACHE)
483 			bp->b_flags |= B_INVAL;
484 		uio->uio_offset = curoff;
485 		if (enough)
486 			n = 0;
487 		break;
488 	    default:
489 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
490 	    }
491 	    if (got_buf)
492 		brelse(bp);
493 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
494 	return (error);
495 }
496 
497 /*
498  * Vnode op for write using bio
499  */
500 int
501 nfs_write(v)
502 	void *v;
503 {
504 	struct vop_write_args /* {
505 		struct vnode *a_vp;
506 		struct uio *a_uio;
507 		int  a_ioflag;
508 		struct ucred *a_cred;
509 	} */ *ap = v;
510 	struct uio *uio = ap->a_uio;
511 	struct proc *p = uio->uio_procp;
512 	struct vnode *vp = ap->a_vp;
513 	struct nfsnode *np = VTONFS(vp);
514 	struct ucred *cred = ap->a_cred;
515 	int ioflag = ap->a_ioflag;
516 	struct vattr vattr;
517 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
518 	void *win;
519 	voff_t oldoff, origoff;
520 	vsize_t bytelen;
521 	int error = 0, iomode, must_commit;
522 
523 #ifdef DIAGNOSTIC
524 	if (uio->uio_rw != UIO_WRITE)
525 		panic("nfs_write mode");
526 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
527 		panic("nfs_write proc");
528 #endif
529 	if (vp->v_type != VREG)
530 		return (EIO);
531 	if (np->n_flag & NWRITEERR) {
532 		np->n_flag &= ~NWRITEERR;
533 		return (np->n_error);
534 	}
535 #ifndef NFS_V2_ONLY
536 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
537 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
538 		(void)nfs_fsinfo(nmp, vp, cred, p);
539 #endif
540 	if (ioflag & (IO_APPEND | IO_SYNC)) {
541 		if (np->n_flag & NMODIFIED) {
542 			np->n_attrstamp = 0;
543 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
544 			if (error)
545 				return (error);
546 		}
547 		if (ioflag & IO_APPEND) {
548 			np->n_attrstamp = 0;
549 			error = VOP_GETATTR(vp, &vattr, cred, p);
550 			if (error)
551 				return (error);
552 			uio->uio_offset = np->n_size;
553 		}
554 	}
555 	if (uio->uio_offset < 0)
556 		return (EINVAL);
557 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
558 		return (EFBIG);
559 	if (uio->uio_resid == 0)
560 		return (0);
561 	/*
562 	 * Maybe this should be above the vnode op call, but so long as
563 	 * file servers have no limits, i don't think it matters
564 	 */
565 	if (p && uio->uio_offset + uio->uio_resid >
566 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
567 		psignal(p, SIGXFSZ);
568 		return (EFBIG);
569 	}
570 
571 	/*
572 	 * update the cached write creds for this node.
573 	 */
574 
575 	if (np->n_wcred) {
576 		crfree(np->n_wcred);
577 	}
578 	np->n_wcred = cred;
579 	crhold(cred);
580 
581 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
582 		iomode = NFSV3WRITE_FILESYNC;
583 		error = nfs_writerpc(vp, uio, &iomode, &must_commit);
584 		if (must_commit)
585 			nfs_clearcommit(vp->v_mount);
586 		return (error);
587 	}
588 
589 	origoff = uio->uio_offset;
590 	do {
591 		oldoff = uio->uio_offset;
592 		bytelen = uio->uio_resid;
593 
594 #ifndef NFS_V2_ONLY
595 		/*
596 		 * Check for a valid write lease.
597 		 */
598 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
599 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
600 			do {
601 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
602 			} while (error == NQNFS_EXPIRED);
603 			if (error)
604 				return (error);
605 			if (np->n_lrev != np->n_brev ||
606 			    (np->n_flag & NQNFSNONCACHE)) {
607 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
608 				if (error)
609 					return (error);
610 				np->n_brev = np->n_lrev;
611 			}
612 		}
613 #endif
614 		nfsstats.biocache_writes++;
615 
616 		np->n_flag |= NMODIFIED;
617 		if (np->n_size < uio->uio_offset + bytelen) {
618 			np->n_size = uio->uio_offset + bytelen;
619 		}
620 		if ((uio->uio_offset & PAGE_MASK) == 0 &&
621 		    ((uio->uio_offset + bytelen) & PAGE_MASK) == 0) {
622 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
623 			    UBC_WRITE | UBC_FAULTBUSY);
624 		} else {
625 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
626 			    UBC_WRITE);
627 		}
628 		error = uiomove(win, bytelen, uio);
629 		ubc_release(win, 0);
630 		if (error) {
631 			break;
632 		}
633 
634 		/*
635 		 * update UVM's notion of the size now that we've
636 		 * copied the data into the vnode's pages.
637 		 */
638 
639 		if (vp->v_size < uio->uio_offset) {
640 			uvm_vnp_setsize(vp, uio->uio_offset);
641 		}
642 
643 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
644 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
645 			simple_lock(&vp->v_uobj.vmobjlock);
646 			error = (vp->v_uobj.pgops->pgo_put)(&vp->v_uobj,
647 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
648 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
649 				       ~(nmp->nm_wsize - 1)),
650 			    PGO_CLEANIT|PGO_WEAK);
651 		}
652 	} while (uio->uio_resid > 0);
653 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
654 		simple_lock(&vp->v_uobj.vmobjlock);
655 		error = (vp->v_uobj.pgops->pgo_put)(&vp->v_uobj,
656 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
657 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
658 			       ~(nmp->nm_wsize - 1)),
659 		    PGO_CLEANIT|PGO_SYNCIO);
660 	}
661 	return error;
662 }
663 
664 /*
665  * Get an nfs cache block.
666  * Allocate a new one if the block isn't currently in the cache
667  * and return the block marked busy. If the calling process is
668  * interrupted by a signal for an interruptible mount point, return
669  * NULL.
670  */
671 struct buf *
672 nfs_getcacheblk(vp, bn, size, p)
673 	struct vnode *vp;
674 	daddr_t bn;
675 	int size;
676 	struct proc *p;
677 {
678 	struct buf *bp;
679 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
680 
681 	if (nmp->nm_flag & NFSMNT_INT) {
682 		bp = getblk(vp, bn, size, PCATCH, 0);
683 		while (bp == NULL) {
684 			if (nfs_sigintr(nmp, NULL, p))
685 				return (NULL);
686 			bp = getblk(vp, bn, size, 0, 2 * hz);
687 		}
688 	} else
689 		bp = getblk(vp, bn, size, 0, 0);
690 	return (bp);
691 }
692 
693 /*
694  * Flush and invalidate all dirty buffers. If another process is already
695  * doing the flush, just wait for completion.
696  */
697 int
698 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
699 	struct vnode *vp;
700 	int flags;
701 	struct ucred *cred;
702 	struct proc *p;
703 	int intrflg;
704 {
705 	struct nfsnode *np = VTONFS(vp);
706 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
707 	int error = 0, slpflag, slptimeo;
708 
709 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
710 		intrflg = 0;
711 	if (intrflg) {
712 		slpflag = PCATCH;
713 		slptimeo = 2 * hz;
714 	} else {
715 		slpflag = 0;
716 		slptimeo = 0;
717 	}
718 	/*
719 	 * First wait for any other process doing a flush to complete.
720 	 */
721 	while (np->n_flag & NFLUSHINPROG) {
722 		np->n_flag |= NFLUSHWANT;
723 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
724 			slptimeo);
725 		if (error && intrflg && nfs_sigintr(nmp, NULL, p))
726 			return (EINTR);
727 	}
728 
729 	/*
730 	 * Now, flush as required.
731 	 */
732 	np->n_flag |= NFLUSHINPROG;
733 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
734 	while (error) {
735 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
736 			np->n_flag &= ~NFLUSHINPROG;
737 			if (np->n_flag & NFLUSHWANT) {
738 				np->n_flag &= ~NFLUSHWANT;
739 				wakeup((caddr_t)&np->n_flag);
740 			}
741 			return (EINTR);
742 		}
743 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
744 	}
745 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
746 	if (np->n_flag & NFLUSHWANT) {
747 		np->n_flag &= ~NFLUSHWANT;
748 		wakeup((caddr_t)&np->n_flag);
749 	}
750 	return (0);
751 }
752 
753 /*
754  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
755  * This is mainly to avoid queueing async I/O requests when the nfsiods
756  * are all hung on a dead server.
757  */
758 
759 int
760 nfs_asyncio(bp)
761 	struct buf *bp;
762 {
763 	int i;
764 	struct nfsmount *nmp;
765 	int gotiod, slpflag = 0, slptimeo = 0, error;
766 
767 	if (nfs_numasync == 0)
768 		return (EIO);
769 
770 
771 	nmp = VFSTONFS(bp->b_vp->v_mount);
772 again:
773 	if (nmp->nm_flag & NFSMNT_INT)
774 		slpflag = PCATCH;
775 	gotiod = FALSE;
776 
777 	/*
778 	 * Find a free iod to process this request.
779 	 */
780 
781 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
782 		if (nfs_iodwant[i]) {
783 			/*
784 			 * Found one, so wake it up and tell it which
785 			 * mount to process.
786 			 */
787 			nfs_iodwant[i] = NULL;
788 			nfs_iodmount[i] = nmp;
789 			nmp->nm_bufqiods++;
790 			wakeup((caddr_t)&nfs_iodwant[i]);
791 			gotiod = TRUE;
792 			break;
793 		}
794 	/*
795 	 * If none are free, we may already have an iod working on this mount
796 	 * point.  If so, it will process our request.
797 	 */
798 	if (!gotiod && nmp->nm_bufqiods > 0)
799 		gotiod = TRUE;
800 
801 	/*
802 	 * If we have an iod which can process the request, then queue
803 	 * the buffer.
804 	 */
805 	if (gotiod) {
806 		/*
807 		 * Ensure that the queue never grows too large.
808 		 */
809 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
810 			nmp->nm_bufqwant = TRUE;
811 			error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
812 				"nfsaio", slptimeo);
813 			if (error) {
814 				if (nfs_sigintr(nmp, NULL, bp->b_proc))
815 					return (EINTR);
816 				if (slpflag == PCATCH) {
817 					slpflag = 0;
818 					slptimeo = 2 * hz;
819 				}
820 			}
821 			/*
822 			 * We might have lost our iod while sleeping,
823 			 * so check and loop if nescessary.
824 			 */
825 			if (nmp->nm_bufqiods == 0)
826 				goto again;
827 		}
828 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
829 		nmp->nm_bufqlen++;
830 		return (0);
831 	    }
832 
833 	/*
834 	 * All the iods are busy on other mounts, so return EIO to
835 	 * force the caller to process the i/o synchronously.
836 	 */
837 	return (EIO);
838 }
839 
840 /*
841  * Do an I/O operation to/from a cache block. This may be called
842  * synchronously or from an nfsiod.
843  */
844 int
845 nfs_doio(bp, p)
846 	struct buf *bp;
847 	struct proc *p;
848 {
849 	struct uio *uiop;
850 	struct vnode *vp;
851 	struct nfsnode *np;
852 	struct nfsmount *nmp;
853 	int error = 0, diff, len, iomode, must_commit = 0;
854 	struct uio uio;
855 	struct iovec io;
856 
857 	vp = bp->b_vp;
858 	np = VTONFS(vp);
859 	nmp = VFSTONFS(vp->v_mount);
860 	uiop = &uio;
861 	uiop->uio_iov = &io;
862 	uiop->uio_iovcnt = 1;
863 	uiop->uio_segflg = UIO_SYSSPACE;
864 	uiop->uio_procp = p;
865 
866 	/*
867 	 * Historically, paging was done with physio, but no more...
868 	 */
869 	if (bp->b_flags & B_PHYS) {
870 	    /*
871 	     * ...though reading /dev/drum still gets us here.
872 	     */
873 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
874 	    /* mapping was done by vmapbuf() */
875 	    io.iov_base = bp->b_data;
876 	    uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
877 	    if (bp->b_flags & B_READ) {
878 		uiop->uio_rw = UIO_READ;
879 		nfsstats.read_physios++;
880 		error = nfs_readrpc(vp, uiop);
881 	    } else {
882 		iomode = NFSV3WRITE_DATASYNC;
883 		uiop->uio_rw = UIO_WRITE;
884 		nfsstats.write_physios++;
885 		error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
886 	    }
887 	    if (error) {
888 		bp->b_flags |= B_ERROR;
889 		bp->b_error = error;
890 	    }
891 	} else if (bp->b_flags & B_READ) {
892 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
893 	    io.iov_base = bp->b_data;
894 	    uiop->uio_rw = UIO_READ;
895 	    switch (vp->v_type) {
896 	    case VREG:
897 		uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
898 		nfsstats.read_bios++;
899 		error = nfs_readrpc(vp, uiop);
900 		if (!error && uiop->uio_resid) {
901 
902 			/*
903 			 * If len > 0, there is a hole in the file and
904 			 * no writes after the hole have been pushed to
905 			 * the server yet.
906 			 * Just zero fill the rest of the valid area.
907 			 */
908 
909 			diff = bp->b_bcount - uiop->uio_resid;
910 			len = np->n_size - ((((off_t)bp->b_blkno) << DEV_BSHIFT)
911 				+ diff);
912 			if (len > 0) {
913 				len = MIN(len, uiop->uio_resid);
914 				memset((char *)bp->b_data + diff, 0, len);
915 			}
916 		}
917 		if (p && (vp->v_flag & VTEXT) &&
918 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
919 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
920 			  np->n_lrev != np->n_brev) ||
921 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
922 			  np->n_mtime != np->n_vattr->va_mtime.tv_sec))) {
923 			uprintf("Process killed due to "
924 				"text file modification\n");
925 			psignal(p, SIGKILL);
926 			p->p_holdcnt++;
927 		}
928 		break;
929 	    case VLNK:
930 		uiop->uio_offset = (off_t)0;
931 		nfsstats.readlink_bios++;
932 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
933 		break;
934 	    case VDIR:
935 		nfsstats.readdir_bios++;
936 		uiop->uio_offset = bp->b_dcookie;
937 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
938 			error = nfs_readdirplusrpc(vp, uiop, curproc->p_ucred);
939 			if (error == NFSERR_NOTSUPP)
940 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
941 		}
942 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
943 			error = nfs_readdirrpc(vp, uiop, curproc->p_ucred);
944 		if (!error) {
945 			bp->b_dcookie = uiop->uio_offset;
946 		}
947 		break;
948 	    default:
949 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
950 		break;
951 	    }
952 	    if (error) {
953 		bp->b_flags |= B_ERROR;
954 		bp->b_error = error;
955 	    }
956 	} else {
957 	    /*
958 	     * If B_NEEDCOMMIT is set, a commit rpc may do the trick. If not
959 	     * an actual write will have to be scheduled.
960 	     */
961 
962 	    io.iov_base = bp->b_data;
963 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
964 	    uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
965 	    uiop->uio_rw = UIO_WRITE;
966 	    nfsstats.write_bios++;
967 	    iomode = NFSV3WRITE_UNSTABLE;
968 	    error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
969 	}
970 	bp->b_resid = uiop->uio_resid;
971 	if (must_commit)
972 		nfs_clearcommit(vp->v_mount);
973 	biodone(bp);
974 	return (error);
975 }
976 
977 /*
978  * Vnode op for VM getpages.
979  */
980 
981 int
982 nfs_getpages(v)
983 	void *v;
984 {
985 	struct vop_getpages_args /* {
986 		struct vnode *a_vp;
987 		voff_t a_offset;
988 		struct vm_page **a_m;
989 		int *a_count;
990 		int a_centeridx;
991 		vm_prot_t a_access_type;
992 		int a_advice;
993 		int a_flags;
994 	} */ *ap = v;
995 
996 	struct vnode *vp = ap->a_vp;
997 	struct uvm_object *uobj = &vp->v_uobj;
998 	struct nfsnode *np = VTONFS(vp);
999 	struct vm_page *pg, **pgs;
1000 	off_t origoffset;
1001 	int i, error, npages;
1002 	boolean_t v3 = NFS_ISV3(vp);
1003 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1004 	UVMHIST_FUNC("nfs_getpages"); UVMHIST_CALLED(ubchist);
1005 
1006 	/*
1007 	 * update the cached read creds for this node.
1008 	 */
1009 
1010 	if (np->n_rcred) {
1011 		crfree(np->n_rcred);
1012 	}
1013 	np->n_rcred = curproc->p_ucred;
1014 	crhold(np->n_rcred);
1015 
1016 	/*
1017 	 * call the genfs code to get the pages.
1018 	 */
1019 
1020 	npages = *ap->a_count;
1021 	error = genfs_getpages(v);
1022 	if (error || !write || !v3) {
1023 		return error;
1024 	}
1025 
1026 	/*
1027 	 * this is a write fault, update the commit info.
1028 	 */
1029 
1030 	origoffset = ap->a_offset;
1031 	pgs = ap->a_m;
1032 
1033 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1034 	nfs_del_committed_range(vp, origoffset, npages);
1035 	nfs_del_tobecommitted_range(vp, origoffset, npages);
1036 	simple_lock(&uobj->vmobjlock);
1037 	for (i = 0; i < npages; i++) {
1038 		pg = pgs[i];
1039 		if (pg == NULL || pg == PGO_DONTCARE) {
1040 			continue;
1041 		}
1042 		pg->flags &= ~(PG_NEEDCOMMIT|PG_RDONLY);
1043 	}
1044 	simple_unlock(&uobj->vmobjlock);
1045 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1046 	return 0;
1047 }
1048 
1049 int
1050 nfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1051 {
1052 	struct uvm_object *uobj = &vp->v_uobj;
1053 	struct nfsnode *np = VTONFS(vp);
1054 	off_t origoffset, commitoff;
1055 	uint32_t commitbytes;
1056 	int error, i;
1057 	int bytes;
1058 	boolean_t v3 = NFS_ISV3(vp);
1059 	boolean_t weak = flags & PGO_WEAK;
1060 	UVMHIST_FUNC("nfs_gop_write"); UVMHIST_CALLED(ubchist);
1061 
1062 	/* XXX for now, skip the v3 stuff. */
1063 	v3 = FALSE;
1064 
1065 	/*
1066 	 * for NFSv2, just write normally.
1067 	 */
1068 
1069 	if (!v3) {
1070 		return genfs_gop_write(vp, pgs, npages, flags);
1071 	}
1072 
1073 	/*
1074 	 * for NFSv3, use delayed writes and the "commit" operation
1075 	 * to avoid sync writes.
1076 	 */
1077 
1078 	origoffset = pgs[0]->offset;
1079 	bytes = npages << PAGE_SHIFT;
1080 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1081 	if (nfs_in_committed_range(vp, origoffset, bytes)) {
1082 		goto committed;
1083 	}
1084 	if (nfs_in_tobecommitted_range(vp, origoffset, bytes)) {
1085 		if (weak) {
1086 			lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1087 			return 0;
1088 		} else {
1089 			commitoff = np->n_pushlo;
1090 			commitbytes = (uint32_t)(np->n_pushhi - np->n_pushlo);
1091 			goto commit;
1092 		}
1093 	} else {
1094 		commitoff = origoffset;
1095 		commitbytes = npages << PAGE_SHIFT;
1096 	}
1097 	simple_lock(&uobj->vmobjlock);
1098 	for (i = 0; i < npages; i++) {
1099 		pgs[i]->flags |= PG_NEEDCOMMIT|PG_RDONLY;
1100 		pgs[i]->flags &= ~PG_CLEAN;
1101 	}
1102 	simple_unlock(&uobj->vmobjlock);
1103 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1104 	error = genfs_gop_write(vp, pgs, npages, flags);
1105 	if (error) {
1106 		return error;
1107 	}
1108 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1109 	if (weak) {
1110 		nfs_add_tobecommitted_range(vp, origoffset,
1111 		    npages << PAGE_SHIFT);
1112 	} else {
1113 commit:
1114 		error = nfs_commit(vp, commitoff, commitbytes, curproc);
1115 		nfs_del_tobecommitted_range(vp, commitoff, commitbytes);
1116 committed:
1117 		simple_lock(&uobj->vmobjlock);
1118 		for (i = 0; i < npages; i++) {
1119 			pgs[i]->flags &= ~(PG_NEEDCOMMIT|PG_RDONLY);
1120 		}
1121 		simple_unlock(&uobj->vmobjlock);
1122 	}
1123 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1124 	return error;
1125 }
1126