xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 5b84b3983f71fd20a534cfa5d1556623a8aaa717)
1 /*	$NetBSD: nfs_bio.c,v 1.134 2005/08/19 10:08:48 yamt 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. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
35  */
36 
37 #include <sys/cdefs.h>
38 __KERNEL_RCSID(0, "$NetBSD: nfs_bio.c,v 1.134 2005/08/19 10:08:48 yamt Exp $");
39 
40 #include "opt_nfs.h"
41 #include "opt_ddb.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/proc.h>
48 #include <sys/buf.h>
49 #include <sys/vnode.h>
50 #include <sys/mount.h>
51 #include <sys/kernel.h>
52 #include <sys/namei.h>
53 #include <sys/dirent.h>
54 #include <sys/malloc.h>
55 
56 #include <uvm/uvm_extern.h>
57 #include <uvm/uvm.h>
58 
59 #include <nfs/rpcv2.h>
60 #include <nfs/nfsproto.h>
61 #include <nfs/nfs.h>
62 #include <nfs/nfsmount.h>
63 #include <nfs/nqnfs.h>
64 #include <nfs/nfsnode.h>
65 #include <nfs/nfs_var.h>
66 
67 extern int nfs_numasync;
68 extern int nfs_commitsize;
69 extern struct nfsstats nfsstats;
70 
71 static int nfs_doio_read __P((struct buf *, struct uio *));
72 static int nfs_doio_write __P((struct buf *, struct uio *));
73 static int nfs_doio_phys __P((struct buf *, struct uio *));
74 
75 /*
76  * Vnode op for read using bio
77  * Any similarity to readip() is purely coincidental
78  */
79 int
80 nfs_bioread(vp, uio, ioflag, cred, cflag)
81 	struct vnode *vp;
82 	struct uio *uio;
83 	int ioflag, cflag;
84 	struct ucred *cred;
85 {
86 	struct nfsnode *np = VTONFS(vp);
87 	struct buf *bp = NULL, *rabp;
88 	struct proc *p = uio->uio_procp;
89 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
90 	struct nfsdircache *ndp = NULL, *nndp = NULL;
91 	caddr_t baddr;
92 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
93 	int enough = 0;
94 	struct dirent *dp, *pdp, *edp, *ep;
95 	off_t curoff = 0;
96 
97 #ifdef DIAGNOSTIC
98 	if (uio->uio_rw != UIO_READ)
99 		panic("nfs_read mode");
100 #endif
101 	if (uio->uio_resid == 0)
102 		return (0);
103 	if (vp->v_type != VDIR && uio->uio_offset < 0)
104 		return (EINVAL);
105 #ifndef NFS_V2_ONLY
106 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
107 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
108 		(void)nfs_fsinfo(nmp, vp, cred, p);
109 #endif
110 	if (vp->v_type != VDIR &&
111 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
112 		return (EFBIG);
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 		error = nfs_flushstalebuf(vp, cred, p,
134 		    NFS_FLUSHSTALEBUF_MYWRITE);
135 		if (error)
136 			return error;
137 	}
138 
139 	do {
140 #ifndef NFS_V2_ONLY
141 	    /*
142 	     * Get a valid lease. If cached data is stale, flush it.
143 	     */
144 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
145 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
146 		    do {
147 			error = nqnfs_getlease(vp, ND_READ, cred, p);
148 		    } while (error == NQNFS_EXPIRED);
149 		    if (error)
150 			return (error);
151 		    if (np->n_lrev != np->n_brev ||
152 			(np->n_flag & NQNFSNONCACHE) ||
153 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
154 			if (vp->v_type == VDIR) {
155 				nfs_invaldircache(vp, 0);
156 			}
157 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
158 			if (error)
159 			    return (error);
160 			np->n_brev = np->n_lrev;
161 		    }
162 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
163 		    nfs_invaldircache(vp, 0);
164 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
165 		    if (error)
166 			return (error);
167 		}
168 	    }
169 #endif
170 	    /*
171 	     * Don't cache symlinks.
172 	     */
173 	    if (np->n_flag & NQNFSNONCACHE
174 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
175 		switch (vp->v_type) {
176 		case VREG:
177 			return (nfs_readrpc(vp, uio));
178 		case VLNK:
179 			return (nfs_readlinkrpc(vp, uio, cred));
180 		case VDIR:
181 			break;
182 		default:
183 			printf(" NQNFSNONCACHE: type %x unexpected\n",
184 			    vp->v_type);
185 		};
186 	    }
187 	    baddr = (caddr_t)0;
188 	    switch (vp->v_type) {
189 	    case VREG:
190 		nfsstats.biocache_reads++;
191 
192 		error = 0;
193 		if (uio->uio_offset >= np->n_size) {
194 			break;
195 		}
196 		while (uio->uio_resid > 0) {
197 			void *win;
198 			int flags;
199 			vsize_t bytelen = MIN(np->n_size - uio->uio_offset,
200 					      uio->uio_resid);
201 
202 			if (bytelen == 0)
203 				break;
204 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
205 					&bytelen, UBC_READ);
206 			error = uiomove(win, bytelen, uio);
207 			flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
208 			ubc_release(win, flags);
209 			if (error) {
210 				break;
211 			}
212 		}
213 		n = 0;
214 		break;
215 
216 	    case VLNK:
217 		nfsstats.biocache_readlinks++;
218 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
219 		if (!bp)
220 			return (EINTR);
221 		if ((bp->b_flags & B_DONE) == 0) {
222 			bp->b_flags |= B_READ;
223 			error = nfs_doio(bp);
224 			if (error) {
225 				brelse(bp);
226 				return (error);
227 			}
228 		}
229 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
230 		got_buf = 1;
231 		on = 0;
232 		break;
233 	    case VDIR:
234 diragain:
235 		nfsstats.biocache_readdirs++;
236 		ndp = nfs_searchdircache(vp, uio->uio_offset,
237 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
238 		if (!ndp) {
239 			/*
240 			 * We've been handed a cookie that is not
241 			 * in the cache. If we're not translating
242 			 * 32 <-> 64, it may be a value that was
243 			 * flushed out of the cache because it grew
244 			 * too big. Let the server judge if it's
245 			 * valid or not. In the translation case,
246 			 * we have no way of validating this value,
247 			 * so punt.
248 			 */
249 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
250 				return (EINVAL);
251 			ndp = nfs_enterdircache(vp, uio->uio_offset,
252 				uio->uio_offset, 0, 0);
253 		}
254 
255 		if (NFS_EOFVALID(np) &&
256 		    ndp->dc_cookie == np->n_direofoffset) {
257 			nfs_putdircache(np, ndp);
258 			nfsstats.direofcache_hits++;
259 			return (0);
260 		}
261 
262 		bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, p);
263 		if (!bp)
264 		    return (EINTR);
265 		if ((bp->b_flags & B_DONE) == 0) {
266 		    bp->b_flags |= B_READ;
267 		    bp->b_dcookie = ndp->dc_blkcookie;
268 		    error = nfs_doio(bp);
269 		    if (error) {
270 			/*
271 			 * Yuck! The directory has been modified on the
272 			 * server. Punt and let the userland code
273 			 * deal with it.
274 			 */
275 			nfs_putdircache(np, ndp);
276 			brelse(bp);
277 			if (error == NFSERR_BAD_COOKIE) {
278 			    nfs_invaldircache(vp, 0);
279 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
280 			    error = EINVAL;
281 			}
282 			return (error);
283 		    }
284 		}
285 
286 		/*
287 		 * Just return if we hit EOF right away with this
288 		 * block. Always check here, because direofoffset
289 		 * may have been set by an nfsiod since the last
290 		 * check.
291 		 *
292 		 * also, empty block implies EOF.
293 		 */
294 
295 		if (bp->b_bcount == bp->b_resid ||
296 		    (NFS_EOFVALID(np) &&
297 		    ndp->dc_blkcookie == np->n_direofoffset)) {
298 			KASSERT(bp->b_bcount != bp->b_resid ||
299 			    ndp->dc_blkcookie == bp->b_dcookie);
300 			nfs_putdircache(np, ndp);
301 			bp->b_flags |= B_NOCACHE;
302 			brelse(bp);
303 			return 0;
304 		}
305 
306 		/*
307 		 * Find the entry we were looking for in the block.
308 		 */
309 
310 		en = ndp->dc_entry;
311 
312 		pdp = dp = (struct dirent *)bp->b_data;
313 		edp = (struct dirent *)(void *)(bp->b_data + bp->b_bcount -
314 		    bp->b_resid);
315 		enn = 0;
316 		while (enn < en && dp < edp) {
317 			pdp = dp;
318 			dp = _DIRENT_NEXT(dp);
319 			enn++;
320 		}
321 
322 		/*
323 		 * If the entry number was bigger than the number of
324 		 * entries in the block, or the cookie of the previous
325 		 * entry doesn't match, the directory cache is
326 		 * stale. Flush it and try again (i.e. go to
327 		 * the server).
328 		 */
329 		if (dp >= edp || (struct dirent *)_DIRENT_NEXT(dp) > edp ||
330 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
331 #ifdef DEBUG
332 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
333 				pdp, dp, edp,
334 				(unsigned long)uio->uio_offset,
335 				(unsigned long)NFS_GETCOOKIE(pdp));
336 #endif
337 			nfs_putdircache(np, ndp);
338 			brelse(bp);
339 			nfs_invaldircache(vp, 0);
340 			nfs_vinvalbuf(vp, 0, cred, p, 0);
341 			goto diragain;
342 		}
343 
344 		on = (caddr_t)dp - bp->b_data;
345 
346 		/*
347 		 * Cache all entries that may be exported to the
348 		 * user, as they may be thrown back at us. The
349 		 * NFSBIO_CACHECOOKIES flag indicates that all
350 		 * entries are being 'exported', so cache them all.
351 		 */
352 
353 		if (en == 0 && pdp == dp) {
354 			dp = _DIRENT_NEXT(dp);
355 			enn++;
356 		}
357 
358 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
359 			n = uio->uio_resid;
360 			enough = 1;
361 		} else
362 			n = bp->b_bcount - bp->b_resid - on;
363 
364 		ep = (struct dirent *)(void *)(bp->b_data + on + n);
365 
366 		/*
367 		 * Find last complete entry to copy, caching entries
368 		 * (if requested) as we go.
369 		 */
370 
371 		while (dp < ep && (struct dirent *)_DIRENT_NEXT(dp) <= ep) {
372 			if (cflag & NFSBIO_CACHECOOKIES) {
373 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
374 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
375 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
376 					NFS_STASHCOOKIE32(pdp,
377 					    nndp->dc_cookie32);
378 				}
379 				nfs_putdircache(np, nndp);
380 			}
381 			pdp = dp;
382 			dp = _DIRENT_NEXT(dp);
383 			enn++;
384 		}
385 		nfs_putdircache(np, ndp);
386 
387 		/*
388 		 * If the last requested entry was not the last in the
389 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
390 		 * cache the cookie of the last requested one, and
391 		 * set of the offset to it.
392 		 */
393 
394 		if ((on + n) < bp->b_bcount - bp->b_resid) {
395 			curoff = NFS_GETCOOKIE(pdp);
396 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
397 			    enn, bp->b_lblkno);
398 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
399 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
400 				curoff = nndp->dc_cookie32;
401 			}
402 			nfs_putdircache(np, nndp);
403 		} else
404 			curoff = bp->b_dcookie;
405 
406 		/*
407 		 * Always cache the entry for the next block,
408 		 * so that readaheads can use it.
409 		 */
410 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
411 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
412 			if (curoff == bp->b_dcookie) {
413 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
414 				curoff = nndp->dc_cookie32;
415 			}
416 		}
417 
418 		n = (char *)_DIRENT_NEXT(pdp) - (bp->b_data + on);
419 
420 		/*
421 		 * If not eof and read aheads are enabled, start one.
422 		 * (You need the current block first, so that you have the
423 		 *  directory offset cookie of the next block.)
424 		 */
425 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
426 		    !NFS_EOFVALID(np) && !(np->n_flag & NQNFSNONCACHE)) {
427 			rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp),
428 						NFS_DIRBLKSIZ, p);
429 			if (rabp) {
430 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
431 				rabp->b_dcookie = nndp->dc_cookie;
432 				rabp->b_flags |= (B_READ | B_ASYNC);
433 				if (nfs_asyncio(rabp)) {
434 				    rabp->b_flags |= B_INVAL;
435 				    brelse(rabp);
436 				}
437 			    } else
438 				brelse(rabp);
439 			}
440 		}
441 		nfs_putdircache(np, nndp);
442 		got_buf = 1;
443 		break;
444 	    default:
445 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
446 		break;
447 	    }
448 
449 	    if (n > 0) {
450 		if (!baddr)
451 			baddr = bp->b_data;
452 		error = uiomove(baddr + on, (int)n, uio);
453 	    }
454 	    switch (vp->v_type) {
455 	    case VREG:
456 		break;
457 	    case VLNK:
458 		n = 0;
459 		break;
460 	    case VDIR:
461 		if (np->n_flag & NQNFSNONCACHE)
462 			bp->b_flags |= B_INVAL;
463 		uio->uio_offset = curoff;
464 		if (enough)
465 			n = 0;
466 		break;
467 	    default:
468 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
469 	    }
470 	    if (got_buf)
471 		brelse(bp);
472 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
473 	return (error);
474 }
475 
476 /*
477  * Vnode op for write using bio
478  */
479 int
480 nfs_write(v)
481 	void *v;
482 {
483 	struct vop_write_args /* {
484 		struct vnode *a_vp;
485 		struct uio *a_uio;
486 		int  a_ioflag;
487 		struct ucred *a_cred;
488 	} */ *ap = v;
489 	struct uio *uio = ap->a_uio;
490 	struct proc *p = uio->uio_procp;
491 	struct vnode *vp = ap->a_vp;
492 	struct nfsnode *np = VTONFS(vp);
493 	struct ucred *cred = ap->a_cred;
494 	struct vattr vattr;
495 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
496 	void *win;
497 	voff_t oldoff, origoff;
498 	vsize_t bytelen;
499 	int flags, error = 0;
500 	int ioflag = ap->a_ioflag;
501 	int extended = 0, wrotedata = 0;
502 
503 #ifdef DIAGNOSTIC
504 	if (uio->uio_rw != UIO_WRITE)
505 		panic("nfs_write mode");
506 #endif
507 	if (vp->v_type != VREG)
508 		return (EIO);
509 	if (np->n_flag & NWRITEERR) {
510 		np->n_flag &= ~NWRITEERR;
511 		return (np->n_error);
512 	}
513 #ifndef NFS_V2_ONLY
514 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
515 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
516 		(void)nfs_fsinfo(nmp, vp, cred, p);
517 #endif
518 	if (ioflag & (IO_APPEND | IO_SYNC)) {
519 		if (np->n_flag & NMODIFIED) {
520 			NFS_INVALIDATE_ATTRCACHE(np);
521 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
522 			if (error)
523 				return (error);
524 		}
525 		if (ioflag & IO_APPEND) {
526 			NFS_INVALIDATE_ATTRCACHE(np);
527 			error = VOP_GETATTR(vp, &vattr, cred, p);
528 			if (error)
529 				return (error);
530 			uio->uio_offset = np->n_size;
531 		}
532 	}
533 	if (uio->uio_offset < 0)
534 		return (EINVAL);
535 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
536 		return (EFBIG);
537 	if (uio->uio_resid == 0)
538 		return (0);
539 	/*
540 	 * Maybe this should be above the vnode op call, but so long as
541 	 * file servers have no limits, i don't think it matters
542 	 */
543 	if (p && uio->uio_offset + uio->uio_resid >
544 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
545 		psignal(p, SIGXFSZ);
546 		return (EFBIG);
547 	}
548 
549 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
550 		int iomode = NFSV3WRITE_FILESYNC;
551 		boolean_t stalewriteverf = FALSE;
552 
553 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
554 		error = nfs_writerpc(vp, uio, &iomode, FALSE, &stalewriteverf);
555 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
556 		if (stalewriteverf)
557 			nfs_clearcommit(vp->v_mount);
558 		return (error);
559 	}
560 
561 	origoff = uio->uio_offset;
562 	do {
563 		boolean_t extending; /* if we are extending whole pages */
564 		u_quad_t oldsize;
565 		oldoff = uio->uio_offset;
566 		bytelen = uio->uio_resid;
567 
568 #ifndef NFS_V2_ONLY
569 		/*
570 		 * Check for a valid write lease.
571 		 */
572 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
573 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
574 			do {
575 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
576 			} while (error == NQNFS_EXPIRED);
577 			if (error)
578 				return (error);
579 			if (np->n_lrev != np->n_brev ||
580 			    (np->n_flag & NQNFSNONCACHE)) {
581 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
582 				if (error)
583 					return (error);
584 				np->n_brev = np->n_lrev;
585 			}
586 		}
587 #endif
588 		nfsstats.biocache_writes++;
589 
590 		oldsize = np->n_size;
591 		np->n_flag |= NMODIFIED;
592 		if (np->n_size < uio->uio_offset + bytelen) {
593 			np->n_size = uio->uio_offset + bytelen;
594 		}
595 		extending = ((uio->uio_offset & PAGE_MASK) == 0 &&
596 		    (bytelen & PAGE_MASK) == 0 &&
597 		    uio->uio_offset >= vp->v_size);
598 		win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
599 			    UBC_WRITE | (extending ? UBC_FAULTBUSY : 0));
600 		error = uiomove(win, bytelen, uio);
601 		flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
602 		ubc_release(win, flags);
603 		if (error) {
604 			if (extending) {
605 				/*
606 				 * backout size and free pages past eof.
607 				 */
608 				np->n_size = oldsize;
609 				simple_lock(&vp->v_interlock);
610 				(void)VOP_PUTPAGES(vp, round_page(vp->v_size),
611 				    0, PGO_SYNCIO | PGO_FREE);
612 			}
613 			break;
614 		}
615 		wrotedata = 1;
616 
617 		/*
618 		 * update UVM's notion of the size now that we've
619 		 * copied the data into the vnode's pages.
620 		 */
621 
622 		if (vp->v_size < uio->uio_offset) {
623 			uvm_vnp_setsize(vp, uio->uio_offset);
624 			extended = 1;
625 		}
626 
627 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
628 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
629 			simple_lock(&vp->v_interlock);
630 			error = VOP_PUTPAGES(vp,
631 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
632 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
633 				       ~(nmp->nm_wsize - 1)), PGO_CLEANIT);
634 		}
635 	} while (uio->uio_resid > 0);
636 	if (wrotedata)
637 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0));
638 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
639 		simple_lock(&vp->v_interlock);
640 		error = VOP_PUTPAGES(vp,
641 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
642 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
643 			       ~(nmp->nm_wsize - 1)),
644 		    PGO_CLEANIT | PGO_SYNCIO);
645 	}
646 	return error;
647 }
648 
649 /*
650  * Get an nfs cache block.
651  * Allocate a new one if the block isn't currently in the cache
652  * and return the block marked busy. If the calling process is
653  * interrupted by a signal for an interruptible mount point, return
654  * NULL.
655  */
656 struct buf *
657 nfs_getcacheblk(vp, bn, size, p)
658 	struct vnode *vp;
659 	daddr_t bn;
660 	int size;
661 	struct proc *p;
662 {
663 	struct buf *bp;
664 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
665 
666 	if (nmp->nm_flag & NFSMNT_INT) {
667 		bp = getblk(vp, bn, size, PCATCH, 0);
668 		while (bp == NULL) {
669 			if (nfs_sigintr(nmp, NULL, p))
670 				return (NULL);
671 			bp = getblk(vp, bn, size, 0, 2 * hz);
672 		}
673 	} else
674 		bp = getblk(vp, bn, size, 0, 0);
675 	return (bp);
676 }
677 
678 /*
679  * Flush and invalidate all dirty buffers. If another process is already
680  * doing the flush, just wait for completion.
681  */
682 int
683 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
684 	struct vnode *vp;
685 	int flags;
686 	struct ucred *cred;
687 	struct proc *p;
688 	int intrflg;
689 {
690 	struct nfsnode *np = VTONFS(vp);
691 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
692 	int error = 0, slpflag, slptimeo;
693 
694 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
695 		intrflg = 0;
696 	if (intrflg) {
697 		slpflag = PCATCH;
698 		slptimeo = 2 * hz;
699 	} else {
700 		slpflag = 0;
701 		slptimeo = 0;
702 	}
703 	/*
704 	 * First wait for any other process doing a flush to complete.
705 	 */
706 	simple_lock(&vp->v_interlock);
707 	while (np->n_flag & NFLUSHINPROG) {
708 		np->n_flag |= NFLUSHWANT;
709 		error = ltsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
710 			slptimeo, &vp->v_interlock);
711 		if (error && intrflg && nfs_sigintr(nmp, NULL, p)) {
712 			simple_unlock(&vp->v_interlock);
713 			return EINTR;
714 		}
715 	}
716 
717 	/*
718 	 * Now, flush as required.
719 	 */
720 	np->n_flag |= NFLUSHINPROG;
721 	simple_unlock(&vp->v_interlock);
722 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
723 	while (error) {
724 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
725 			error = EINTR;
726 			break;
727 		}
728 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
729 	}
730 	simple_lock(&vp->v_interlock);
731 	if (error == 0)
732 		np->n_flag &= ~NMODIFIED;
733 	np->n_flag &= ~NFLUSHINPROG;
734 	if (np->n_flag & NFLUSHWANT) {
735 		np->n_flag &= ~NFLUSHWANT;
736 		wakeup(&np->n_flag);
737 	}
738 	simple_unlock(&vp->v_interlock);
739 	return error;
740 }
741 
742 /*
743  * nfs_flushstalebuf: flush cache if it's stale.
744  *
745  * => caller shouldn't own any pages or buffers which belong to the vnode.
746  */
747 
748 int
749 nfs_flushstalebuf(struct vnode *vp, struct ucred *cred, struct proc *p,
750     int flags)
751 {
752 	struct nfsnode *np = VTONFS(vp);
753 	struct vattr vattr;
754 	int error;
755 
756 	if (np->n_flag & NMODIFIED) {
757 		if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0
758 		    || vp->v_type != VREG) {
759 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
760 			if (error)
761 				return error;
762 			if (vp->v_type == VDIR) {
763 				nfs_invaldircache(vp, 0);
764 			}
765 		} else {
766 			/*
767 			 * XXX assuming writes are ours.
768 			 */
769 		}
770 		NFS_INVALIDATE_ATTRCACHE(np);
771 		error = VOP_GETATTR(vp, &vattr, cred, p);
772 		if (error)
773 			return error;
774 		np->n_mtime = vattr.va_mtime;
775 	} else {
776 		error = VOP_GETATTR(vp, &vattr, cred, p);
777 		if (error)
778 			return error;
779 		if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
780 			if (vp->v_type == VDIR) {
781 				nfs_invaldircache(vp, 0);
782 			}
783 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
784 			if (error)
785 				return error;
786 			np->n_mtime = vattr.va_mtime;
787 		}
788 	}
789 
790 	return error;
791 }
792 
793 /*
794  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
795  * This is mainly to avoid queueing async I/O requests when the nfsiods
796  * are all hung on a dead server.
797  */
798 
799 int
800 nfs_asyncio(bp)
801 	struct buf *bp;
802 {
803 	int i;
804 	struct nfsmount *nmp;
805 	int gotiod, slpflag = 0, slptimeo = 0, error;
806 
807 	if (nfs_numasync == 0)
808 		return (EIO);
809 
810 	nmp = VFSTONFS(bp->b_vp->v_mount);
811 again:
812 	if (nmp->nm_flag & NFSMNT_INT)
813 		slpflag = PCATCH;
814 	gotiod = FALSE;
815 
816 	/*
817 	 * Find a free iod to process this request.
818 	 */
819 
820 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++) {
821 		struct nfs_iod *iod = &nfs_asyncdaemon[i];
822 
823 		simple_lock(&iod->nid_slock);
824 		if (iod->nid_want) {
825 			/*
826 			 * Found one, so wake it up and tell it which
827 			 * mount to process.
828 			 */
829 			iod->nid_want = NULL;
830 			iod->nid_mount = nmp;
831 			wakeup(&iod->nid_want);
832 			simple_lock(&nmp->nm_slock);
833 			simple_unlock(&iod->nid_slock);
834 			nmp->nm_bufqiods++;
835 			gotiod = TRUE;
836 			break;
837 		}
838 		simple_unlock(&iod->nid_slock);
839 	}
840 
841 	/*
842 	 * If none are free, we may already have an iod working on this mount
843 	 * point.  If so, it will process our request.
844 	 */
845 
846 	if (!gotiod) {
847 		simple_lock(&nmp->nm_slock);
848 		if (nmp->nm_bufqiods > 0)
849 			gotiod = TRUE;
850 	}
851 
852 	LOCK_ASSERT(simple_lock_held(&nmp->nm_slock));
853 
854 	/*
855 	 * If we have an iod which can process the request, then queue
856 	 * the buffer.  However, even if we have an iod, do not initiate
857 	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
858 	 * is via local loopback, we may put curproc (pagedaemon) to sleep
859 	 * waiting for the writes to complete. But the server (ourself)
860 	 * may block the write, waiting for its (ie., our) pagedaemon
861 	 * to produce clean pages to handle the write: deadlock.
862 	 * XXX: start non-loopback mounts straight away?  If "lots free",
863 	 * let pagedaemon start loopback writes anyway?
864 	 */
865 	if (gotiod) {
866 
867 		/*
868 		 * Ensure that the queue never grows too large.
869 		 */
870 		if (curproc == uvm.pagedaemon_proc) {
871 	  		/* Enque for later, to avoid free-page deadlock */
872 			  (void) 0;
873 		} else while (nmp->nm_bufqlen >= 2*nfs_numasync) {
874 			nmp->nm_bufqwant = TRUE;
875 			error = ltsleep(&nmp->nm_bufq,
876 			    slpflag | PRIBIO | PNORELOCK,
877 			    "nfsaio", slptimeo, &nmp->nm_slock);
878 			if (error) {
879 				if (nfs_sigintr(nmp, NULL, curproc))
880 					return (EINTR);
881 				if (slpflag == PCATCH) {
882 					slpflag = 0;
883 					slptimeo = 2 * hz;
884 				}
885 			}
886 
887 			/*
888 			 * We might have lost our iod while sleeping,
889 			 * so check and loop if nescessary.
890 			 */
891 
892 			if (nmp->nm_bufqiods == 0)
893 				goto again;
894 
895 			simple_lock(&nmp->nm_slock);
896 		}
897 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
898 		nmp->nm_bufqlen++;
899 		simple_unlock(&nmp->nm_slock);
900 		return (0);
901 	}
902 	simple_unlock(&nmp->nm_slock);
903 
904 	/*
905 	 * All the iods are busy on other mounts, so return EIO to
906 	 * force the caller to process the i/o synchronously.
907 	 */
908 
909 	return (EIO);
910 }
911 
912 /*
913  * nfs_doio for read.
914  */
915 static int
916 nfs_doio_read(bp, uiop)
917 	struct buf *bp;
918 	struct uio *uiop;
919 {
920 	struct vnode *vp = bp->b_vp;
921 	struct nfsnode *np = VTONFS(vp);
922 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
923 	struct proc *p = uiop->uio_procp;
924 	int error = 0;
925 
926 	uiop->uio_rw = UIO_READ;
927 	switch (vp->v_type) {
928 	case VREG:
929 		nfsstats.read_bios++;
930 		error = nfs_readrpc(vp, uiop);
931 		if (!error && uiop->uio_resid) {
932 			int diff, len;
933 
934 			/*
935 			 * If uio_resid > 0, there is a hole in the file and
936 			 * no writes after the hole have been pushed to
937 			 * the server yet or the file has been truncated
938 			 * on the server.
939 			 * Just zero fill the rest of the valid area.
940 			 */
941 
942 			KASSERT(vp->v_size >=
943 			    uiop->uio_offset + uiop->uio_resid);
944 			diff = bp->b_bcount - uiop->uio_resid;
945 			len = uiop->uio_resid;
946 			memset((char *)bp->b_data + diff, 0, len);
947 		}
948 		if (p && (vp->v_flag & VTEXT) &&
949 		    (((nmp->nm_flag & NFSMNT_NQNFS) &&
950 		      NQNFS_CKINVALID(vp, np, ND_READ) &&
951 		      np->n_lrev != np->n_brev) ||
952 		     (!(nmp->nm_flag & NFSMNT_NQNFS) &&
953 		      timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)))) {
954 			uprintf("Process killed due to "
955 				"text file modification\n");
956 			psignal(p, SIGKILL);
957 #if 0 /* XXX NJWLWP */
958 			p->p_holdcnt++;
959 #endif
960 		}
961 		break;
962 	case VLNK:
963 		KASSERT(uiop->uio_offset == (off_t)0);
964 		nfsstats.readlink_bios++;
965 		error = nfs_readlinkrpc(vp, uiop, np->n_rcred);
966 		break;
967 	case VDIR:
968 		nfsstats.readdir_bios++;
969 		uiop->uio_offset = bp->b_dcookie;
970 #ifndef NFS_V2_ONLY
971 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
972 			error = nfs_readdirplusrpc(vp, uiop, np->n_rcred);
973 			if (error == NFSERR_NOTSUPP)
974 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
975 		}
976 #else
977 		nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
978 #endif
979 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
980 			error = nfs_readdirrpc(vp, uiop, np->n_rcred);
981 		if (!error) {
982 			bp->b_dcookie = uiop->uio_offset;
983 		}
984 		break;
985 	default:
986 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
987 		break;
988 	}
989 	if (error) {
990 		bp->b_flags |= B_ERROR;
991 		bp->b_error = error;
992 	}
993 	return error;
994 }
995 
996 /*
997  * nfs_doio for write.
998  */
999 static int
1000 nfs_doio_write(bp, uiop)
1001 	struct buf *bp;
1002 	struct uio *uiop;
1003 {
1004 	struct vnode *vp = bp->b_vp;
1005 	struct nfsnode *np = VTONFS(vp);
1006 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1007 	int iomode;
1008 	boolean_t stalewriteverf = FALSE;
1009 	int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT;
1010 	struct vm_page *pgs[npages];
1011 #ifndef NFS_V2_ONLY
1012 	boolean_t needcommit = TRUE; /* need only COMMIT RPC */
1013 #else
1014 	boolean_t needcommit = FALSE; /* need only COMMIT RPC */
1015 #endif
1016 	boolean_t pageprotected;
1017 	struct uvm_object *uobj = &vp->v_uobj;
1018 	int error;
1019 	off_t off, cnt;
1020 
1021 	if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) {
1022 		iomode = NFSV3WRITE_UNSTABLE;
1023 	} else {
1024 		iomode = NFSV3WRITE_FILESYNC;
1025 	}
1026 
1027 #ifndef NFS_V2_ONLY
1028 again:
1029 #endif
1030 	lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1031 
1032 	for (i = 0; i < npages; i++) {
1033 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
1034 		if (pgs[i]->uobject == uobj &&
1035 		    pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) {
1036 			KASSERT(pgs[i]->flags & PG_BUSY);
1037 			/*
1038 			 * this page belongs to our object.
1039 			 */
1040 			simple_lock(&uobj->vmobjlock);
1041 			/*
1042 			 * write out the page stably if it's about to
1043 			 * be released because we can't resend it
1044 			 * on the server crash.
1045 			 *
1046 			 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be
1047 			 * changed until unbusy the page.
1048 			 */
1049 			if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT))
1050 				iomode = NFSV3WRITE_FILESYNC;
1051 			/*
1052 			 * if we met a page which hasn't been sent yet,
1053 			 * we need do WRITE RPC.
1054 			 */
1055 			if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0)
1056 				needcommit = FALSE;
1057 			simple_unlock(&uobj->vmobjlock);
1058 		} else {
1059 			iomode = NFSV3WRITE_FILESYNC;
1060 			needcommit = FALSE;
1061 		}
1062 	}
1063 	if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) {
1064 		simple_lock(&uobj->vmobjlock);
1065 		for (i = 0; i < npages; i++) {
1066 			pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY;
1067 			pmap_page_protect(pgs[i], VM_PROT_READ);
1068 		}
1069 		simple_unlock(&uobj->vmobjlock);
1070 		pageprotected = TRUE; /* pages can't be modified during i/o. */
1071 	} else
1072 		pageprotected = FALSE;
1073 
1074 	/*
1075 	 * Send the data to the server if necessary,
1076 	 * otherwise just send a commit rpc.
1077 	 */
1078 #ifndef NFS_V2_ONLY
1079 	if (needcommit) {
1080 
1081 		/*
1082 		 * If the buffer is in the range that we already committed,
1083 		 * there's nothing to do.
1084 		 *
1085 		 * If it's in the range that we need to commit, push the
1086 		 * whole range at once, otherwise only push the buffer.
1087 		 * In both these cases, acquire the commit lock to avoid
1088 		 * other processes modifying the range.
1089 		 */
1090 
1091 		off = uiop->uio_offset;
1092 		cnt = bp->b_bcount;
1093 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1094 		if (!nfs_in_committed_range(vp, off, bp->b_bcount)) {
1095 			boolean_t pushedrange;
1096 			if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) {
1097 				pushedrange = TRUE;
1098 				off = np->n_pushlo;
1099 				cnt = np->n_pushhi - np->n_pushlo;
1100 			} else {
1101 				pushedrange = FALSE;
1102 			}
1103 			error = nfs_commit(vp, off, cnt, curproc);
1104 			if (error == 0) {
1105 				if (pushedrange) {
1106 					nfs_merge_commit_ranges(vp);
1107 				} else {
1108 					nfs_add_committed_range(vp, off, cnt);
1109 				}
1110 			}
1111 		} else {
1112 			error = 0;
1113 		}
1114 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1115 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1116 		if (!error) {
1117 			/*
1118 			 * pages are now on stable storage.
1119 			 */
1120 			uiop->uio_resid = 0;
1121 			simple_lock(&uobj->vmobjlock);
1122 			for (i = 0; i < npages; i++) {
1123 				pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1124 			}
1125 			simple_unlock(&uobj->vmobjlock);
1126 			return 0;
1127 		} else if (error == NFSERR_STALEWRITEVERF) {
1128 			nfs_clearcommit(vp->v_mount);
1129 			goto again;
1130 		}
1131 		if (error) {
1132 			bp->b_flags |= B_ERROR;
1133 			bp->b_error = np->n_error = error;
1134 			np->n_flag |= NWRITEERR;
1135 		}
1136 		return error;
1137 	}
1138 #endif
1139 	off = uiop->uio_offset;
1140 	cnt = bp->b_bcount;
1141 	uiop->uio_rw = UIO_WRITE;
1142 	nfsstats.write_bios++;
1143 	error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf);
1144 #ifndef NFS_V2_ONLY
1145 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1146 		/*
1147 		 * we need to commit pages later.
1148 		 */
1149 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1150 		nfs_add_tobecommitted_range(vp, off, cnt);
1151 		/*
1152 		 * if there can be too many uncommitted pages, commit them now.
1153 		 */
1154 		if (np->n_pushhi - np->n_pushlo > nfs_commitsize) {
1155 			off = np->n_pushlo;
1156 			cnt = nfs_commitsize >> 1;
1157 			error = nfs_commit(vp, off, cnt, curproc);
1158 			if (!error) {
1159 				nfs_add_committed_range(vp, off, cnt);
1160 				nfs_del_tobecommitted_range(vp, off, cnt);
1161 			}
1162 			if (error == NFSERR_STALEWRITEVERF) {
1163 				stalewriteverf = TRUE;
1164 				error = 0; /* it isn't a real error */
1165 			}
1166 		} else {
1167 			/*
1168 			 * re-dirty pages so that they will be passed
1169 			 * to us later again.
1170 			 */
1171 			simple_lock(&uobj->vmobjlock);
1172 			for (i = 0; i < npages; i++) {
1173 				pgs[i]->flags &= ~PG_CLEAN;
1174 			}
1175 			simple_unlock(&uobj->vmobjlock);
1176 		}
1177 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1178 	} else
1179 #endif
1180 	if (!error) {
1181 		/*
1182 		 * pages are now on stable storage.
1183 		 */
1184 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1185 		nfs_del_committed_range(vp, off, cnt);
1186 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1187 		simple_lock(&uobj->vmobjlock);
1188 		for (i = 0; i < npages; i++) {
1189 			pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1190 		}
1191 		simple_unlock(&uobj->vmobjlock);
1192 	} else {
1193 		/*
1194 		 * we got an error.
1195 		 */
1196 		bp->b_flags |= B_ERROR;
1197 		bp->b_error = np->n_error = error;
1198 		np->n_flag |= NWRITEERR;
1199 	}
1200 
1201 	lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1202 
1203 	if (stalewriteverf) {
1204 		nfs_clearcommit(vp->v_mount);
1205 	}
1206 	return error;
1207 }
1208 
1209 /*
1210  * nfs_doio for B_PHYS.
1211  */
1212 static int
1213 nfs_doio_phys(bp, uiop)
1214 	struct buf *bp;
1215 	struct uio *uiop;
1216 {
1217 	struct vnode *vp = bp->b_vp;
1218 	int error;
1219 
1220 	uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
1221 	if (bp->b_flags & B_READ) {
1222 		uiop->uio_rw = UIO_READ;
1223 		nfsstats.read_physios++;
1224 		error = nfs_readrpc(vp, uiop);
1225 	} else {
1226 		int iomode = NFSV3WRITE_DATASYNC;
1227 		boolean_t stalewriteverf;
1228 		struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1229 
1230 		uiop->uio_rw = UIO_WRITE;
1231 		nfsstats.write_physios++;
1232 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1233 		error = nfs_writerpc(vp, uiop, &iomode, FALSE, &stalewriteverf);
1234 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1235 		if (stalewriteverf) {
1236 			nfs_clearcommit(bp->b_vp->v_mount);
1237 		}
1238 	}
1239 	if (error) {
1240 		bp->b_flags |= B_ERROR;
1241 		bp->b_error = error;
1242 	}
1243 	return error;
1244 }
1245 
1246 /*
1247  * Do an I/O operation to/from a cache block. This may be called
1248  * synchronously or from an nfsiod.
1249  */
1250 int
1251 nfs_doio(bp)
1252 	struct buf *bp;
1253 {
1254 	int error;
1255 	struct uio uio;
1256 	struct uio *uiop = &uio;
1257 	struct iovec io;
1258 	UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist);
1259 
1260 	uiop->uio_iov = &io;
1261 	uiop->uio_iovcnt = 1;
1262 	uiop->uio_segflg = UIO_SYSSPACE;
1263 	uiop->uio_procp = NULL;
1264 	uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
1265 	io.iov_base = bp->b_data;
1266 	io.iov_len = uiop->uio_resid = bp->b_bcount;
1267 
1268 	/*
1269 	 * Historically, paging was done with physio, but no more...
1270 	 */
1271 	if (bp->b_flags & B_PHYS) {
1272 		/*
1273 		 * ...though reading /dev/drum still gets us here.
1274 		 */
1275 		error = nfs_doio_phys(bp, uiop);
1276 	} else if (bp->b_flags & B_READ) {
1277 		error = nfs_doio_read(bp, uiop);
1278 	} else {
1279 		error = nfs_doio_write(bp, uiop);
1280 	}
1281 	bp->b_resid = uiop->uio_resid;
1282 	biodone(bp);
1283 	return (error);
1284 }
1285 
1286 /*
1287  * Vnode op for VM getpages.
1288  */
1289 
1290 int
1291 nfs_getpages(v)
1292 	void *v;
1293 {
1294 	struct vop_getpages_args /* {
1295 		struct vnode *a_vp;
1296 		voff_t a_offset;
1297 		struct vm_page **a_m;
1298 		int *a_count;
1299 		int a_centeridx;
1300 		vm_prot_t a_access_type;
1301 		int a_advice;
1302 		int a_flags;
1303 	} */ *ap = v;
1304 
1305 	struct vnode *vp = ap->a_vp;
1306 	struct uvm_object *uobj = &vp->v_uobj;
1307 	struct nfsnode *np = VTONFS(vp);
1308 	const int npages = *ap->a_count;
1309 	struct vm_page *pg, **pgs, *opgs[npages];
1310 	off_t origoffset, len;
1311 	int i, error;
1312 	boolean_t v3 = NFS_ISV3(vp);
1313 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1314 	boolean_t locked = (ap->a_flags & PGO_LOCKED) != 0;
1315 
1316 	/*
1317 	 * call the genfs code to get the pages.  `pgs' may be NULL
1318 	 * when doing read-ahead.
1319 	 */
1320 
1321 	pgs = ap->a_m;
1322 	if (write && locked && v3) {
1323 		KASSERT(pgs != NULL);
1324 #ifdef DEBUG
1325 
1326 		/*
1327 		 * If PGO_LOCKED is set, real pages shouldn't exists
1328 		 * in the array.
1329 		 */
1330 
1331 		for (i = 0; i < npages; i++)
1332 			KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE);
1333 #endif
1334 		memcpy(opgs, pgs, npages * sizeof(struct vm_pages *));
1335 	}
1336 	error = genfs_getpages(v);
1337 	if (error) {
1338 		return (error);
1339 	}
1340 
1341 	/*
1342 	 * for read faults where the nfs node is not yet marked NMODIFIED,
1343 	 * set PG_RDONLY on the pages so that we come back here if someone
1344 	 * tries to modify later via the mapping that will be entered for
1345 	 * this fault.
1346 	 */
1347 
1348 	if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) {
1349 		if (!locked) {
1350 			simple_lock(&uobj->vmobjlock);
1351 		}
1352 		for (i = 0; i < npages; i++) {
1353 			pg = pgs[i];
1354 			if (pg == NULL || pg == PGO_DONTCARE) {
1355 				continue;
1356 			}
1357 			pg->flags |= PG_RDONLY;
1358 		}
1359 		if (!locked) {
1360 			simple_unlock(&uobj->vmobjlock);
1361 		}
1362 	}
1363 	if (!write) {
1364 		return (0);
1365 	}
1366 
1367 	/*
1368 	 * this is a write fault, update the commit info.
1369 	 */
1370 
1371 	origoffset = ap->a_offset;
1372 	len = npages << PAGE_SHIFT;
1373 
1374 	if (v3) {
1375 		error = lockmgr(&np->n_commitlock,
1376 		    LK_EXCLUSIVE | (locked ? LK_NOWAIT : 0), NULL);
1377 		if (error) {
1378 			KASSERT(locked != 0);
1379 
1380 			/*
1381 			 * Since PGO_LOCKED is set, we need to unbusy
1382 			 * all pages fetched by genfs_getpages() above,
1383 			 * tell the caller that there are no pages
1384 			 * available and put back original pgs array.
1385 			 */
1386 
1387 			uvm_lock_pageq();
1388 			uvm_page_unbusy(pgs, npages);
1389 			uvm_unlock_pageq();
1390 			*ap->a_count = 0;
1391 			memcpy(pgs, opgs,
1392 			    npages * sizeof(struct vm_pages *));
1393 			return (error);
1394 		}
1395 		nfs_del_committed_range(vp, origoffset, len);
1396 		nfs_del_tobecommitted_range(vp, origoffset, len);
1397 	}
1398 	np->n_flag |= NMODIFIED;
1399 	if (!locked) {
1400 		simple_lock(&uobj->vmobjlock);
1401 	}
1402 	for (i = 0; i < npages; i++) {
1403 		pg = pgs[i];
1404 		if (pg == NULL || pg == PGO_DONTCARE) {
1405 			continue;
1406 		}
1407 		pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1408 	}
1409 	if (!locked) {
1410 		simple_unlock(&uobj->vmobjlock);
1411 	}
1412 	if (v3) {
1413 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1414 	}
1415 	return (0);
1416 }
1417