xref: /netbsd-src/sys/nfs/nfs_bio.c (revision db6316d1518382eecd2fdbe55a1205e0620a1b35)
1 /*	$NetBSD: nfs_bio.c,v 1.123 2004/12/14 09:13:13 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 1.123 2004/12/14 09:13:13 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;
89 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
90 	struct nfsdircache *ndp = NULL, *nndp = NULL;
91 	caddr_t baddr, ep, edp;
92 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
93 	int enough = 0;
94 	struct dirent *dp, *pdp;
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 	p = uio->uio_procp;
106 #ifndef NFS_V2_ONLY
107 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
108 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
109 		(void)nfs_fsinfo(nmp, vp, cred, p);
110 #endif
111 	if (vp->v_type != VDIR &&
112 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
113 		return (EFBIG);
114 
115 	/*
116 	 * For nfs, cache consistency can only be maintained approximately.
117 	 * Although RFC1094 does not specify the criteria, the following is
118 	 * believed to be compatible with the reference port.
119 	 * For nqnfs, full cache consistency is maintained within the loop.
120 	 * For nfs:
121 	 * If the file's modify time on the server has changed since the
122 	 * last read rpc or you have written to the file,
123 	 * you may have lost data cache consistency with the
124 	 * server, so flush all of the file's data out of the cache.
125 	 * Then force a getattr rpc to ensure that you have up to date
126 	 * attributes.
127 	 * NB: This implies that cache data can be read when up to
128 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
129 	 * attributes this could be forced by setting n_attrstamp to 0 before
130 	 * the VOP_GETATTR() call.
131 	 */
132 
133 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
134 		error = nfs_flushstalebuf(vp, cred, p,
135 		    NFS_FLUSHSTALEBUF_MYWRITE);
136 		if (error)
137 			return error;
138 	}
139 
140 	do {
141 #ifndef NFS_V2_ONLY
142 	    /*
143 	     * Get a valid lease. If cached data is stale, flush it.
144 	     */
145 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
146 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
147 		    do {
148 			error = nqnfs_getlease(vp, ND_READ, cred, p);
149 		    } while (error == NQNFS_EXPIRED);
150 		    if (error)
151 			return (error);
152 		    if (np->n_lrev != np->n_brev ||
153 			(np->n_flag & NQNFSNONCACHE) ||
154 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
155 			if (vp->v_type == VDIR) {
156 				nfs_invaldircache(vp, 0);
157 				np->n_direofoffset = 0;
158 			}
159 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
160 			if (error)
161 			    return (error);
162 			np->n_brev = np->n_lrev;
163 		    }
164 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
165 		    nfs_invaldircache(vp, 0);
166 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
167 		    np->n_direofoffset = 0;
168 		    if (error)
169 			return (error);
170 		}
171 	    }
172 #endif
173 	    /*
174 	     * Don't cache symlinks.
175 	     */
176 	    if (np->n_flag & NQNFSNONCACHE
177 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
178 		switch (vp->v_type) {
179 		case VREG:
180 			return (nfs_readrpc(vp, uio));
181 		case VLNK:
182 			return (nfs_readlinkrpc(vp, uio, cred));
183 		case VDIR:
184 			break;
185 		default:
186 			printf(" NQNFSNONCACHE: type %x unexpected\n",
187 			    vp->v_type);
188 		};
189 	    }
190 	    baddr = (caddr_t)0;
191 	    switch (vp->v_type) {
192 	    case VREG:
193 		nfsstats.biocache_reads++;
194 
195 		error = 0;
196 		if (uio->uio_offset >= np->n_size) {
197 			break;
198 		}
199 		while (uio->uio_resid > 0) {
200 			void *win;
201 			vsize_t bytelen = MIN(np->n_size - uio->uio_offset,
202 					      uio->uio_resid);
203 
204 			if (bytelen == 0)
205 				break;
206 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
207 					&bytelen, UBC_READ);
208 			error = uiomove(win, bytelen, uio);
209 			ubc_release(win, 0);
210 			if (error) {
211 				break;
212 			}
213 		}
214 		n = 0;
215 		break;
216 
217 	    case VLNK:
218 		nfsstats.biocache_readlinks++;
219 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
220 		if (!bp)
221 			return (EINTR);
222 		if ((bp->b_flags & B_DONE) == 0) {
223 			bp->b_flags |= B_READ;
224 			error = nfs_doio(bp, p);
225 			if (error) {
226 				brelse(bp);
227 				return (error);
228 			}
229 		}
230 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
231 		got_buf = 1;
232 		on = 0;
233 		break;
234 	    case VDIR:
235 diragain:
236 		nfsstats.biocache_readdirs++;
237 		ndp = nfs_searchdircache(vp, uio->uio_offset,
238 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
239 		if (!ndp) {
240 			/*
241 			 * We've been handed a cookie that is not
242 			 * in the cache. If we're not translating
243 			 * 32 <-> 64, it may be a value that was
244 			 * flushed out of the cache because it grew
245 			 * too big. Let the server judge if it's
246 			 * valid or not. In the translation case,
247 			 * we have no way of validating this value,
248 			 * so punt.
249 			 */
250 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
251 				return (EINVAL);
252 			ndp = nfs_enterdircache(vp, uio->uio_offset,
253 				uio->uio_offset, 0, 0);
254 		}
255 
256 		if (uio->uio_offset != 0 &&
257 		    ndp->dc_cookie == np->n_direofoffset) {
258 			nfs_putdircache(np, ndp);
259 			nfsstats.direofcache_hits++;
260 			return (0);
261 		}
262 
263 		bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, p);
264 		if (!bp)
265 		    return (EINTR);
266 		if ((bp->b_flags & B_DONE) == 0) {
267 		    bp->b_flags |= B_READ;
268 		    bp->b_dcookie = ndp->dc_blkcookie;
269 		    error = nfs_doio(bp, p);
270 		    if (error) {
271 			/*
272 			 * Yuck! The directory has been modified on the
273 			 * server. Punt and let the userland code
274 			 * deal with it.
275 			 */
276 			nfs_putdircache(np, ndp);
277 			brelse(bp);
278 			if (error == NFSERR_BAD_COOKIE) {
279 			    nfs_invaldircache(vp, 0);
280 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
281 			    error = EINVAL;
282 			}
283 			return (error);
284 		    }
285 		}
286 
287 		/*
288 		 * Just return if we hit EOF right away with this
289 		 * block. Always check here, because direofoffset
290 		 * may have been set by an nfsiod since the last
291 		 * check.
292 		 */
293 		if (np->n_direofoffset != 0 &&
294 			ndp->dc_blkcookie == np->n_direofoffset) {
295 			nfs_putdircache(np, ndp);
296 			brelse(bp);
297 			return (0);
298 		}
299 
300 		/*
301 		 * Find the entry we were looking for in the block.
302 		 */
303 
304 		en = ndp->dc_entry;
305 
306 		pdp = dp = (struct dirent *)bp->b_data;
307 		edp = bp->b_data + bp->b_bcount - bp->b_resid;
308 		enn = 0;
309 		while (enn < en && (caddr_t)dp < edp) {
310 			pdp = dp;
311 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
312 			enn++;
313 		}
314 
315 		/*
316 		 * If the entry number was bigger than the number of
317 		 * entries in the block, or the cookie of the previous
318 		 * entry doesn't match, the directory cache is
319 		 * stale. Flush it and try again (i.e. go to
320 		 * the server).
321 		 */
322 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
323 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
324 #ifdef DEBUG
325 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
326 				pdp, dp, edp,
327 				(unsigned long)uio->uio_offset,
328 				(unsigned long)NFS_GETCOOKIE(pdp));
329 #endif
330 			nfs_putdircache(np, ndp);
331 			brelse(bp);
332 			nfs_invaldircache(vp, 0);
333 			nfs_vinvalbuf(vp, 0, cred, p, 0);
334 			goto diragain;
335 		}
336 
337 		on = (caddr_t)dp - bp->b_data;
338 
339 		/*
340 		 * Cache all entries that may be exported to the
341 		 * user, as they may be thrown back at us. The
342 		 * NFSBIO_CACHECOOKIES flag indicates that all
343 		 * entries are being 'exported', so cache them all.
344 		 */
345 
346 		if (en == 0 && pdp == dp) {
347 			dp = (struct dirent *)
348 			    ((caddr_t)dp + dp->d_reclen);
349 			enn++;
350 		}
351 
352 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
353 			n = uio->uio_resid;
354 			enough = 1;
355 		} else
356 			n = bp->b_bcount - bp->b_resid - on;
357 
358 		ep = bp->b_data + on + n;
359 
360 		/*
361 		 * Find last complete entry to copy, caching entries
362 		 * (if requested) as we go.
363 		 */
364 
365 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
366 			if (cflag & NFSBIO_CACHECOOKIES) {
367 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
368 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
369 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
370 					NFS_STASHCOOKIE32(pdp,
371 					    nndp->dc_cookie32);
372 				}
373 				nfs_putdircache(np, nndp);
374 			}
375 			pdp = dp;
376 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
377 			enn++;
378 		}
379 		nfs_putdircache(np, ndp);
380 
381 		/*
382 		 * If the last requested entry was not the last in the
383 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
384 		 * cache the cookie of the last requested one, and
385 		 * set of the offset to it.
386 		 */
387 
388 		if ((on + n) < bp->b_bcount - bp->b_resid) {
389 			curoff = NFS_GETCOOKIE(pdp);
390 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
391 			    enn, bp->b_lblkno);
392 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
393 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
394 				curoff = nndp->dc_cookie32;
395 			}
396 			nfs_putdircache(np, nndp);
397 		} else
398 			curoff = bp->b_dcookie;
399 
400 		/*
401 		 * Always cache the entry for the next block,
402 		 * so that readaheads can use it.
403 		 */
404 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
405 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
406 			if (curoff == bp->b_dcookie) {
407 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
408 				curoff = nndp->dc_cookie32;
409 			}
410 		}
411 
412 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
413 
414 		/*
415 		 * If not eof and read aheads are enabled, start one.
416 		 * (You need the current block first, so that you have the
417 		 *  directory offset cookie of the next block.)
418 		 */
419 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
420 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
421 			rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp),
422 						NFS_DIRBLKSIZ, p);
423 			if (rabp) {
424 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
425 				rabp->b_dcookie = nndp->dc_cookie;
426 				rabp->b_flags |= (B_READ | B_ASYNC);
427 				if (nfs_asyncio(rabp)) {
428 				    rabp->b_flags |= B_INVAL;
429 				    brelse(rabp);
430 				}
431 			    } else
432 				brelse(rabp);
433 			}
434 		}
435 		nfs_putdircache(np, nndp);
436 		got_buf = 1;
437 		break;
438 	    default:
439 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
440 		break;
441 	    }
442 
443 	    if (n > 0) {
444 		if (!baddr)
445 			baddr = bp->b_data;
446 		error = uiomove(baddr + on, (int)n, uio);
447 	    }
448 	    switch (vp->v_type) {
449 	    case VREG:
450 		break;
451 	    case VLNK:
452 		n = 0;
453 		break;
454 	    case VDIR:
455 		if (np->n_flag & NQNFSNONCACHE)
456 			bp->b_flags |= B_INVAL;
457 		uio->uio_offset = curoff;
458 		if (enough)
459 			n = 0;
460 		break;
461 	    default:
462 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
463 	    }
464 	    if (got_buf)
465 		brelse(bp);
466 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
467 	return (error);
468 }
469 
470 /*
471  * Vnode op for write using bio
472  */
473 int
474 nfs_write(v)
475 	void *v;
476 {
477 	struct vop_write_args /* {
478 		struct vnode *a_vp;
479 		struct uio *a_uio;
480 		int  a_ioflag;
481 		struct ucred *a_cred;
482 	} */ *ap = v;
483 	struct uio *uio = ap->a_uio;
484 	struct proc *p = uio->uio_procp;
485 	struct vnode *vp = ap->a_vp;
486 	struct nfsnode *np = VTONFS(vp);
487 	struct ucred *cred = ap->a_cred;
488 	int ioflag = ap->a_ioflag;
489 	struct vattr vattr;
490 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
491 	void *win;
492 	voff_t oldoff, origoff;
493 	vsize_t bytelen;
494 	int error = 0;
495 	int extended = 0, wrotedta = 0;
496 
497 #ifdef DIAGNOSTIC
498 	if (uio->uio_rw != UIO_WRITE)
499 		panic("nfs_write mode");
500 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
501 		panic("nfs_write proc");
502 #endif
503 	if (vp->v_type != VREG)
504 		return (EIO);
505 	if (np->n_flag & NWRITEERR) {
506 		np->n_flag &= ~NWRITEERR;
507 		return (np->n_error);
508 	}
509 #ifndef NFS_V2_ONLY
510 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
511 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
512 		(void)nfs_fsinfo(nmp, vp, cred, p);
513 #endif
514 	if (ioflag & (IO_APPEND | IO_SYNC)) {
515 		if (np->n_flag & NMODIFIED) {
516 			NFS_INVALIDATE_ATTRCACHE(np);
517 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
518 			if (error)
519 				return (error);
520 		}
521 		if (ioflag & IO_APPEND) {
522 			NFS_INVALIDATE_ATTRCACHE(np);
523 			error = VOP_GETATTR(vp, &vattr, cred, p);
524 			if (error)
525 				return (error);
526 			uio->uio_offset = np->n_size;
527 		}
528 	}
529 	if (uio->uio_offset < 0)
530 		return (EINVAL);
531 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
532 		return (EFBIG);
533 	if (uio->uio_resid == 0)
534 		return (0);
535 	/*
536 	 * Maybe this should be above the vnode op call, but so long as
537 	 * file servers have no limits, i don't think it matters
538 	 */
539 	if (p && uio->uio_offset + uio->uio_resid >
540 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
541 		psignal(p, SIGXFSZ);
542 		return (EFBIG);
543 	}
544 
545 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
546 		int iomode = NFSV3WRITE_FILESYNC;
547 		boolean_t stalewriteverf = FALSE;
548 
549 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
550 		error = nfs_writerpc(vp, uio, &iomode, FALSE, &stalewriteverf);
551 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
552 		if (stalewriteverf)
553 			nfs_clearcommit(vp->v_mount);
554 		return (error);
555 	}
556 
557 	origoff = uio->uio_offset;
558 	do {
559 		boolean_t extending; /* if we are extending whole pages */
560 		u_quad_t oldsize;
561 		oldoff = uio->uio_offset;
562 		bytelen = uio->uio_resid;
563 
564 #ifndef NFS_V2_ONLY
565 		/*
566 		 * Check for a valid write lease.
567 		 */
568 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
569 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
570 			do {
571 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
572 			} while (error == NQNFS_EXPIRED);
573 			if (error)
574 				return (error);
575 			if (np->n_lrev != np->n_brev ||
576 			    (np->n_flag & NQNFSNONCACHE)) {
577 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
578 				if (error)
579 					return (error);
580 				np->n_brev = np->n_lrev;
581 			}
582 		}
583 #endif
584 		nfsstats.biocache_writes++;
585 
586 		oldsize = np->n_size;
587 		np->n_flag |= NMODIFIED;
588 		if (np->n_size < uio->uio_offset + bytelen) {
589 			np->n_size = uio->uio_offset + bytelen;
590 		}
591 		extending = ((uio->uio_offset & PAGE_MASK) == 0 &&
592 		    (bytelen & PAGE_MASK) == 0 &&
593 		    uio->uio_offset >= vp->v_size);
594 		win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
595 			    UBC_WRITE | (extending ? UBC_FAULTBUSY : 0));
596 		error = uiomove(win, bytelen, uio);
597 		ubc_release(win, 0);
598 		if (error) {
599 			if (extending) {
600 				/*
601 				 * backout size and free pages past eof.
602 				 */
603 				np->n_size = oldsize;
604 				simple_lock(&vp->v_interlock);
605 				(void)VOP_PUTPAGES(vp, round_page(vp->v_size),
606 				    0, PGO_SYNCIO | PGO_FREE);
607 			}
608 			break;
609 		}
610 		wrotedta = 1;
611 
612 		/*
613 		 * update UVM's notion of the size now that we've
614 		 * copied the data into the vnode's pages.
615 		 */
616 
617 		if (vp->v_size < uio->uio_offset) {
618 			uvm_vnp_setsize(vp, uio->uio_offset);
619 			extended = 1;
620 		}
621 
622 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
623 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
624 			simple_lock(&vp->v_interlock);
625 			error = VOP_PUTPAGES(vp,
626 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
627 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
628 				       ~(nmp->nm_wsize - 1)), PGO_CLEANIT);
629 		}
630 	} while (uio->uio_resid > 0);
631 	if (wrotedta)
632 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0));
633 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
634 		simple_lock(&vp->v_interlock);
635 		error = VOP_PUTPAGES(vp,
636 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
637 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
638 			       ~(nmp->nm_wsize - 1)),
639 		    PGO_CLEANIT | PGO_SYNCIO);
640 	}
641 	return error;
642 }
643 
644 /*
645  * Get an nfs cache block.
646  * Allocate a new one if the block isn't currently in the cache
647  * and return the block marked busy. If the calling process is
648  * interrupted by a signal for an interruptible mount point, return
649  * NULL.
650  */
651 struct buf *
652 nfs_getcacheblk(vp, bn, size, p)
653 	struct vnode *vp;
654 	daddr_t bn;
655 	int size;
656 	struct proc *p;
657 {
658 	struct buf *bp;
659 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
660 
661 	if (nmp->nm_flag & NFSMNT_INT) {
662 		bp = getblk(vp, bn, size, PCATCH, 0);
663 		while (bp == NULL) {
664 			if (nfs_sigintr(nmp, NULL, p))
665 				return (NULL);
666 			bp = getblk(vp, bn, size, 0, 2 * hz);
667 		}
668 	} else
669 		bp = getblk(vp, bn, size, 0, 0);
670 	return (bp);
671 }
672 
673 /*
674  * Flush and invalidate all dirty buffers. If another process is already
675  * doing the flush, just wait for completion.
676  */
677 int
678 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
679 	struct vnode *vp;
680 	int flags;
681 	struct ucred *cred;
682 	struct proc *p;
683 	int intrflg;
684 {
685 	struct nfsnode *np = VTONFS(vp);
686 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
687 	int error = 0, slpflag, slptimeo;
688 
689 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
690 		intrflg = 0;
691 	if (intrflg) {
692 		slpflag = PCATCH;
693 		slptimeo = 2 * hz;
694 	} else {
695 		slpflag = 0;
696 		slptimeo = 0;
697 	}
698 	/*
699 	 * First wait for any other process doing a flush to complete.
700 	 */
701 	simple_lock(&vp->v_interlock);
702 	while (np->n_flag & NFLUSHINPROG) {
703 		np->n_flag |= NFLUSHWANT;
704 		error = ltsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
705 			slptimeo, &vp->v_interlock);
706 		if (error && intrflg && nfs_sigintr(nmp, NULL, p)) {
707 			simple_unlock(&vp->v_interlock);
708 			return EINTR;
709 		}
710 	}
711 
712 	/*
713 	 * Now, flush as required.
714 	 */
715 	np->n_flag |= NFLUSHINPROG;
716 	simple_unlock(&vp->v_interlock);
717 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
718 	while (error) {
719 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
720 			error = EINTR;
721 			break;
722 		}
723 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
724 	}
725 	simple_lock(&vp->v_interlock);
726 	if (error == 0)
727 		np->n_flag &= ~NMODIFIED;
728 	np->n_flag &= ~NFLUSHINPROG;
729 	if (np->n_flag & NFLUSHWANT) {
730 		np->n_flag &= ~NFLUSHWANT;
731 		wakeup(&np->n_flag);
732 	}
733 	simple_unlock(&vp->v_interlock);
734 	return error;
735 }
736 
737 /*
738  * nfs_flushstalebuf: flush cache if it's stale.
739  *
740  * => caller shouldn't own any pages or buffers which belong to the vnode.
741  */
742 
743 int
744 nfs_flushstalebuf(struct vnode *vp, struct ucred *cred, struct proc *p,
745     int flags)
746 {
747 	struct nfsnode *np = VTONFS(vp);
748 	struct vattr vattr;
749 	int error;
750 
751 	if (np->n_flag & NMODIFIED) {
752 		if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0
753 		    || vp->v_type != VREG) {
754 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
755 			if (error)
756 				return error;
757 			if (vp->v_type == VDIR) {
758 				nfs_invaldircache(vp, 0);
759 				np->n_direofoffset = 0;
760 			}
761 		} else {
762 			/*
763 			 * XXX assuming writes are ours.
764 			 */
765 		}
766 		NFS_INVALIDATE_ATTRCACHE(np);
767 		error = VOP_GETATTR(vp, &vattr, cred, p);
768 		if (error)
769 			return error;
770 		np->n_mtime = vattr.va_mtime;
771 	} else {
772 		error = VOP_GETATTR(vp, &vattr, cred, p);
773 		if (error)
774 			return error;
775 		if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
776 			if (vp->v_type == VDIR) {
777 				nfs_invaldircache(vp, 0);
778 				np->n_direofoffset = 0;
779 			}
780 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
781 			if (error)
782 				return error;
783 			np->n_mtime = vattr.va_mtime;
784 		}
785 	}
786 
787 	return error;
788 }
789 
790 /*
791  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
792  * This is mainly to avoid queueing async I/O requests when the nfsiods
793  * are all hung on a dead server.
794  */
795 
796 int
797 nfs_asyncio(bp)
798 	struct buf *bp;
799 {
800 	int i;
801 	struct nfsmount *nmp;
802 	int gotiod, slpflag = 0, slptimeo = 0, error;
803 
804 	if (nfs_numasync == 0)
805 		return (EIO);
806 
807 	nmp = VFSTONFS(bp->b_vp->v_mount);
808 again:
809 	if (nmp->nm_flag & NFSMNT_INT)
810 		slpflag = PCATCH;
811 	gotiod = FALSE;
812 
813 	/*
814 	 * Find a free iod to process this request.
815 	 */
816 
817 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++) {
818 		struct nfs_iod *iod = &nfs_asyncdaemon[i];
819 
820 		simple_lock(&iod->nid_slock);
821 		if (iod->nid_want) {
822 			/*
823 			 * Found one, so wake it up and tell it which
824 			 * mount to process.
825 			 */
826 			iod->nid_want = NULL;
827 			iod->nid_mount = nmp;
828 			wakeup(&iod->nid_want);
829 			simple_lock(&nmp->nm_slock);
830 			simple_unlock(&iod->nid_slock);
831 			nmp->nm_bufqiods++;
832 			gotiod = TRUE;
833 			break;
834 		}
835 		simple_unlock(&iod->nid_slock);
836 	}
837 
838 	/*
839 	 * If none are free, we may already have an iod working on this mount
840 	 * point.  If so, it will process our request.
841 	 */
842 
843 	if (!gotiod) {
844 		simple_lock(&nmp->nm_slock);
845 		if (nmp->nm_bufqiods > 0)
846 			gotiod = TRUE;
847 	}
848 
849 	LOCK_ASSERT(simple_lock_held(&nmp->nm_slock));
850 
851 	/*
852 	 * If we have an iod which can process the request, then queue
853 	 * the buffer.  However, even if we have an iod, do not initiate
854 	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
855 	 * is via local loopback, we may put curproc (pagedaemon) to sleep
856 	 * waiting for the writes to complete. But the server (ourself)
857 	 * may block the write, waiting for its (ie., our) pagedaemon
858 	 * to produce clean pages to handle the write: deadlock.
859 	 * XXX: start non-loopback mounts straight away?  If "lots free",
860 	 * let pagedaemon start loopback writes anyway?
861 	 */
862 	if (gotiod) {
863 
864 		/*
865 		 * Ensure that the queue never grows too large.
866 		 */
867 		if (curproc == uvm.pagedaemon_proc) {
868 	  		/* Enque for later, to avoid free-page deadlock */
869 			  (void) 0;
870 		} else while (nmp->nm_bufqlen >= 2*nfs_numasync) {
871 			nmp->nm_bufqwant = TRUE;
872 			error = ltsleep(&nmp->nm_bufq,
873 			    slpflag | PRIBIO | PNORELOCK,
874 			    "nfsaio", slptimeo, &nmp->nm_slock);
875 			if (error) {
876 				if (nfs_sigintr(nmp, NULL, curproc))
877 					return (EINTR);
878 				if (slpflag == PCATCH) {
879 					slpflag = 0;
880 					slptimeo = 2 * hz;
881 				}
882 			}
883 
884 			/*
885 			 * We might have lost our iod while sleeping,
886 			 * so check and loop if nescessary.
887 			 */
888 
889 			if (nmp->nm_bufqiods == 0)
890 				goto again;
891 
892 			simple_lock(&nmp->nm_slock);
893 		}
894 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
895 		nmp->nm_bufqlen++;
896 		simple_unlock(&nmp->nm_slock);
897 		return (0);
898 	}
899 	simple_unlock(&nmp->nm_slock);
900 
901 	/*
902 	 * All the iods are busy on other mounts, so return EIO to
903 	 * force the caller to process the i/o synchronously.
904 	 */
905 
906 	return (EIO);
907 }
908 
909 /*
910  * nfs_doio for read.
911  */
912 static int
913 nfs_doio_read(bp, uiop)
914 	struct buf *bp;
915 	struct uio *uiop;
916 {
917 	struct vnode *vp = bp->b_vp;
918 	struct nfsnode *np = VTONFS(vp);
919 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
920 	int error = 0;
921 
922 	uiop->uio_rw = UIO_READ;
923 	switch (vp->v_type) {
924 	case VREG:
925 		nfsstats.read_bios++;
926 		error = nfs_readrpc(vp, uiop);
927 		if (!error && uiop->uio_resid) {
928 			int diff, len;
929 
930 			/*
931 			 * If uio_resid > 0, there is a hole in the file and
932 			 * no writes after the hole have been pushed to
933 			 * the server yet or the file has been truncated
934 			 * on the server.
935 			 * Just zero fill the rest of the valid area.
936 			 */
937 
938 			KASSERT(vp->v_size >=
939 			    uiop->uio_offset + uiop->uio_resid);
940 			diff = bp->b_bcount - uiop->uio_resid;
941 			len = uiop->uio_resid;
942 			memset((char *)bp->b_data + diff, 0, len);
943 		}
944 		if (uiop->uio_procp && (vp->v_flag & VTEXT) &&
945 		    (((nmp->nm_flag & NFSMNT_NQNFS) &&
946 		      NQNFS_CKINVALID(vp, np, ND_READ) &&
947 		      np->n_lrev != np->n_brev) ||
948 		     (!(nmp->nm_flag & NFSMNT_NQNFS) &&
949 		      timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)))) {
950 			uprintf("Process killed due to "
951 				"text file modification\n");
952 			psignal(uiop->uio_procp, SIGKILL);
953 #if 0 /* XXX NJWLWP */
954 			uiop->uio_procp->p_holdcnt++;
955 #endif
956 		}
957 		break;
958 	case VLNK:
959 		KASSERT(uiop->uio_offset == (off_t)0);
960 		nfsstats.readlink_bios++;
961 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
962 		break;
963 	case VDIR:
964 		nfsstats.readdir_bios++;
965 		uiop->uio_offset = bp->b_dcookie;
966 #ifndef NFS_V2_ONLY
967 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
968 			error = nfs_readdirplusrpc(vp, uiop, np->n_rcred);
969 			if (error == NFSERR_NOTSUPP)
970 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
971 		}
972 #else
973 		nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
974 #endif
975 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
976 			error = nfs_readdirrpc(vp, uiop, np->n_rcred);
977 		if (!error) {
978 			bp->b_dcookie = uiop->uio_offset;
979 		}
980 		break;
981 	default:
982 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
983 		break;
984 	}
985 	if (error) {
986 		bp->b_flags |= B_ERROR;
987 		bp->b_error = error;
988 	}
989 	return error;
990 }
991 
992 /*
993  * nfs_doio for write.
994  */
995 static int
996 nfs_doio_write(bp, uiop)
997 	struct buf *bp;
998 	struct uio *uiop;
999 {
1000 	struct vnode *vp = bp->b_vp;
1001 	struct nfsnode *np = VTONFS(vp);
1002 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1003 	int iomode;
1004 	boolean_t stalewriteverf = FALSE;
1005 	int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT;
1006 	struct vm_page *pgs[npages];
1007 #ifndef NFS_V2_ONLY
1008 	boolean_t needcommit = TRUE; /* need only COMMIT RPC */
1009 #else
1010 	boolean_t needcommit = FALSE; /* need only COMMIT RPC */
1011 #endif
1012 	boolean_t pageprotected;
1013 	struct uvm_object *uobj = &vp->v_uobj;
1014 	int error;
1015 	off_t off, cnt;
1016 
1017 	if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) {
1018 		iomode = NFSV3WRITE_UNSTABLE;
1019 	} else {
1020 		iomode = NFSV3WRITE_FILESYNC;
1021 	}
1022 
1023 #ifndef NFS_V2_ONLY
1024 again:
1025 #endif
1026 	lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1027 
1028 	for (i = 0; i < npages; i++) {
1029 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
1030 		if (pgs[i]->uobject == uobj &&
1031 		    pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) {
1032 			KASSERT(pgs[i]->flags & PG_BUSY);
1033 			/*
1034 			 * this page belongs to our object.
1035 			 */
1036 			simple_lock(&uobj->vmobjlock);
1037 			/*
1038 			 * write out the page stably if it's about to
1039 			 * be released because we can't resend it
1040 			 * on the server crash.
1041 			 *
1042 			 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be
1043 			 * changed until unbusy the page.
1044 			 */
1045 			if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT))
1046 				iomode = NFSV3WRITE_FILESYNC;
1047 			/*
1048 			 * if we met a page which hasn't been sent yet,
1049 			 * we need do WRITE RPC.
1050 			 */
1051 			if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0)
1052 				needcommit = FALSE;
1053 			simple_unlock(&uobj->vmobjlock);
1054 		} else {
1055 			iomode = NFSV3WRITE_FILESYNC;
1056 			needcommit = FALSE;
1057 		}
1058 	}
1059 	if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) {
1060 		simple_lock(&uobj->vmobjlock);
1061 		for (i = 0; i < npages; i++) {
1062 			pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY;
1063 			pmap_page_protect(pgs[i], VM_PROT_READ);
1064 		}
1065 		simple_unlock(&uobj->vmobjlock);
1066 		pageprotected = TRUE; /* pages can't be modified during i/o. */
1067 	} else
1068 		pageprotected = FALSE;
1069 
1070 	/*
1071 	 * Send the data to the server if necessary,
1072 	 * otherwise just send a commit rpc.
1073 	 */
1074 #ifndef NFS_V2_ONLY
1075 	if (needcommit) {
1076 
1077 		/*
1078 		 * If the buffer is in the range that we already committed,
1079 		 * there's nothing to do.
1080 		 *
1081 		 * If it's in the range that we need to commit, push the
1082 		 * whole range at once, otherwise only push the buffer.
1083 		 * In both these cases, acquire the commit lock to avoid
1084 		 * other processes modifying the range.
1085 		 */
1086 
1087 		off = uiop->uio_offset;
1088 		cnt = bp->b_bcount;
1089 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1090 		if (!nfs_in_committed_range(vp, off, bp->b_bcount)) {
1091 			boolean_t pushedrange;
1092 			if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) {
1093 				pushedrange = TRUE;
1094 				off = np->n_pushlo;
1095 				cnt = np->n_pushhi - np->n_pushlo;
1096 			} else {
1097 				pushedrange = FALSE;
1098 			}
1099 			error = nfs_commit(vp, off, cnt, curproc);
1100 			if (error == 0) {
1101 				if (pushedrange) {
1102 					nfs_merge_commit_ranges(vp);
1103 				} else {
1104 					nfs_add_committed_range(vp, off, cnt);
1105 				}
1106 			}
1107 		} else {
1108 			error = 0;
1109 		}
1110 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1111 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1112 		if (!error) {
1113 			/*
1114 			 * pages are now on stable storage.
1115 			 */
1116 			uiop->uio_resid = 0;
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 			return 0;
1123 		} else if (error == NFSERR_STALEWRITEVERF) {
1124 			nfs_clearcommit(vp->v_mount);
1125 			goto again;
1126 		}
1127 		if (error) {
1128 			bp->b_flags |= B_ERROR;
1129 			bp->b_error = np->n_error = error;
1130 			np->n_flag |= NWRITEERR;
1131 		}
1132 		return error;
1133 	}
1134 #endif
1135 	off = uiop->uio_offset;
1136 	cnt = bp->b_bcount;
1137 	uiop->uio_rw = UIO_WRITE;
1138 	nfsstats.write_bios++;
1139 	error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf);
1140 #ifndef NFS_V2_ONLY
1141 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1142 		/*
1143 		 * we need to commit pages later.
1144 		 */
1145 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1146 		nfs_add_tobecommitted_range(vp, off, cnt);
1147 		/*
1148 		 * if there can be too many uncommitted pages, commit them now.
1149 		 */
1150 		if (np->n_pushhi - np->n_pushlo > nfs_commitsize) {
1151 			off = np->n_pushlo;
1152 			cnt = nfs_commitsize >> 1;
1153 			error = nfs_commit(vp, off, cnt, curproc);
1154 			if (!error) {
1155 				nfs_add_committed_range(vp, off, cnt);
1156 				nfs_del_tobecommitted_range(vp, off, cnt);
1157 			}
1158 			if (error == NFSERR_STALEWRITEVERF) {
1159 				stalewriteverf = TRUE;
1160 				error = 0; /* it isn't a real error */
1161 			}
1162 		} else {
1163 			/*
1164 			 * re-dirty pages so that they will be passed
1165 			 * to us later again.
1166 			 */
1167 			simple_lock(&uobj->vmobjlock);
1168 			for (i = 0; i < npages; i++) {
1169 				pgs[i]->flags &= ~PG_CLEAN;
1170 			}
1171 			simple_unlock(&uobj->vmobjlock);
1172 		}
1173 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1174 	} else
1175 #endif
1176 	if (!error) {
1177 		/*
1178 		 * pages are now on stable storage.
1179 		 */
1180 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1181 		nfs_del_committed_range(vp, off, cnt);
1182 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1183 		simple_lock(&uobj->vmobjlock);
1184 		for (i = 0; i < npages; i++) {
1185 			pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1186 		}
1187 		simple_unlock(&uobj->vmobjlock);
1188 	} else {
1189 		/*
1190 		 * we got an error.
1191 		 */
1192 		bp->b_flags |= B_ERROR;
1193 		bp->b_error = np->n_error = error;
1194 		np->n_flag |= NWRITEERR;
1195 	}
1196 
1197 	lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1198 
1199 	if (stalewriteverf) {
1200 		nfs_clearcommit(vp->v_mount);
1201 	}
1202 	return error;
1203 }
1204 
1205 /*
1206  * nfs_doio for B_PHYS.
1207  */
1208 static int
1209 nfs_doio_phys(bp, uiop)
1210 	struct buf *bp;
1211 	struct uio *uiop;
1212 {
1213 	struct vnode *vp = bp->b_vp;
1214 	int error;
1215 
1216 	uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
1217 	if (bp->b_flags & B_READ) {
1218 		uiop->uio_rw = UIO_READ;
1219 		nfsstats.read_physios++;
1220 		error = nfs_readrpc(vp, uiop);
1221 	} else {
1222 		int iomode = NFSV3WRITE_DATASYNC;
1223 		boolean_t stalewriteverf;
1224 		struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1225 
1226 		uiop->uio_rw = UIO_WRITE;
1227 		nfsstats.write_physios++;
1228 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1229 		error = nfs_writerpc(vp, uiop, &iomode, FALSE, &stalewriteverf);
1230 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1231 		if (stalewriteverf) {
1232 			nfs_clearcommit(bp->b_vp->v_mount);
1233 		}
1234 	}
1235 	if (error) {
1236 		bp->b_flags |= B_ERROR;
1237 		bp->b_error = error;
1238 	}
1239 	return error;
1240 }
1241 
1242 /*
1243  * Do an I/O operation to/from a cache block. This may be called
1244  * synchronously or from an nfsiod.
1245  */
1246 int
1247 nfs_doio(bp, p)
1248 	struct buf *bp;
1249 	struct proc *p;
1250 {
1251 	int error;
1252 	struct uio uio;
1253 	struct uio *uiop = &uio;
1254 	struct iovec io;
1255 	UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist);
1256 
1257 	uiop->uio_iov = &io;
1258 	uiop->uio_iovcnt = 1;
1259 	uiop->uio_segflg = UIO_SYSSPACE;
1260 	uiop->uio_procp = NULL;
1261 	uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
1262 	io.iov_base = bp->b_data;
1263 	io.iov_len = uiop->uio_resid = bp->b_bcount;
1264 
1265 	/*
1266 	 * Historically, paging was done with physio, but no more...
1267 	 */
1268 	if (bp->b_flags & B_PHYS) {
1269 		/*
1270 		 * ...though reading /dev/drum still gets us here.
1271 		 */
1272 		error = nfs_doio_phys(bp, uiop);
1273 	} else if (bp->b_flags & B_READ) {
1274 		error = nfs_doio_read(bp, uiop);
1275 	} else {
1276 		error = nfs_doio_write(bp, uiop);
1277 	}
1278 	bp->b_resid = uiop->uio_resid;
1279 	biodone(bp);
1280 	return (error);
1281 }
1282 
1283 /*
1284  * Vnode op for VM getpages.
1285  */
1286 
1287 int
1288 nfs_getpages(v)
1289 	void *v;
1290 {
1291 	struct vop_getpages_args /* {
1292 		struct vnode *a_vp;
1293 		voff_t a_offset;
1294 		struct vm_page **a_m;
1295 		int *a_count;
1296 		int a_centeridx;
1297 		vm_prot_t a_access_type;
1298 		int a_advice;
1299 		int a_flags;
1300 	} */ *ap = v;
1301 
1302 	struct vnode *vp = ap->a_vp;
1303 	struct uvm_object *uobj = &vp->v_uobj;
1304 	struct nfsnode *np = VTONFS(vp);
1305 	const int npages = *ap->a_count;
1306 	struct vm_page *pg, **pgs, *opgs[npages];
1307 	off_t origoffset, len;
1308 	int i, error;
1309 	boolean_t v3 = NFS_ISV3(vp);
1310 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1311 	boolean_t locked = (ap->a_flags & PGO_LOCKED) != 0;
1312 
1313 	/*
1314 	 * call the genfs code to get the pages.  `pgs' may be NULL
1315 	 * when doing read-ahead.
1316 	 */
1317 
1318 	pgs = ap->a_m;
1319 	if (write && locked && v3) {
1320 		KASSERT(pgs != NULL);
1321 #ifdef DEBUG
1322 
1323 		/*
1324 		 * If PGO_LOCKED is set, real pages shouldn't exists
1325 		 * in the array.
1326 		 */
1327 
1328 		for (i = 0; i < npages; i++)
1329 			KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE);
1330 #endif
1331 		memcpy(opgs, pgs, npages * sizeof(struct vm_pages *));
1332 	}
1333 	error = genfs_getpages(v);
1334 	if (error) {
1335 		return (error);
1336 	}
1337 
1338 	/*
1339 	 * for read faults where the nfs node is not yet marked NMODIFIED,
1340 	 * set PG_RDONLY on the pages so that we come back here if someone
1341 	 * tries to modify later via the mapping that will be entered for
1342 	 * this fault.
1343 	 */
1344 
1345 	if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) {
1346 		if (!locked) {
1347 			simple_lock(&uobj->vmobjlock);
1348 		}
1349 		for (i = 0; i < npages; i++) {
1350 			pg = pgs[i];
1351 			if (pg == NULL || pg == PGO_DONTCARE) {
1352 				continue;
1353 			}
1354 			pg->flags |= PG_RDONLY;
1355 		}
1356 		if (!locked) {
1357 			simple_unlock(&uobj->vmobjlock);
1358 		}
1359 	}
1360 	if (!write) {
1361 		return (0);
1362 	}
1363 
1364 	/*
1365 	 * this is a write fault, update the commit info.
1366 	 */
1367 
1368 	origoffset = ap->a_offset;
1369 	len = npages << PAGE_SHIFT;
1370 
1371 	if (v3) {
1372 		error = lockmgr(&np->n_commitlock,
1373 		    LK_EXCLUSIVE | (locked ? LK_NOWAIT : 0), NULL);
1374 		if (error) {
1375 			KASSERT(locked != 0);
1376 
1377 			/*
1378 			 * Since PGO_LOCKED is set, we need to unbusy
1379 			 * all pages fetched by genfs_getpages() above,
1380 			 * tell the caller that there are no pages
1381 			 * available and put back original pgs array.
1382 			 */
1383 
1384 			uvm_lock_pageq();
1385 			uvm_page_unbusy(pgs, npages);
1386 			uvm_unlock_pageq();
1387 			*ap->a_count = 0;
1388 			memcpy(pgs, opgs,
1389 			    npages * sizeof(struct vm_pages *));
1390 			return (error);
1391 		}
1392 		nfs_del_committed_range(vp, origoffset, len);
1393 		nfs_del_tobecommitted_range(vp, origoffset, len);
1394 	}
1395 	np->n_flag |= NMODIFIED;
1396 	if (!locked) {
1397 		simple_lock(&uobj->vmobjlock);
1398 	}
1399 	for (i = 0; i < npages; i++) {
1400 		pg = pgs[i];
1401 		if (pg == NULL || pg == PGO_DONTCARE) {
1402 			continue;
1403 		}
1404 		pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1405 	}
1406 	if (!locked) {
1407 		simple_unlock(&uobj->vmobjlock);
1408 	}
1409 	if (v3) {
1410 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1411 	}
1412 	return (0);
1413 }
1414