xref: /netbsd-src/sys/nfs/nfs_bio.c (revision c2f76ff004a2cb67efe5b12d97bd3ef7fe89e18d)
1 /*	$NetBSD: nfs_bio.c,v 1.185 2010/06/12 21:10:55 jakllsch 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 1.185 2010/06/12 21:10:55 jakllsch Exp $");
39 
40 #ifdef _KERNEL_OPT
41 #include "opt_nfs.h"
42 #include "opt_ddb.h"
43 #endif
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/resourcevar.h>
48 #include <sys/signalvar.h>
49 #include <sys/proc.h>
50 #include <sys/buf.h>
51 #include <sys/vnode.h>
52 #include <sys/mount.h>
53 #include <sys/kernel.h>
54 #include <sys/namei.h>
55 #include <sys/dirent.h>
56 #include <sys/kauth.h>
57 
58 #include <uvm/uvm_extern.h>
59 #include <uvm/uvm.h>
60 
61 #include <nfs/rpcv2.h>
62 #include <nfs/nfsproto.h>
63 #include <nfs/nfs.h>
64 #include <nfs/nfsmount.h>
65 #include <nfs/nfsnode.h>
66 #include <nfs/nfs_var.h>
67 
68 extern int nfs_numasync;
69 extern int nfs_commitsize;
70 extern struct nfsstats nfsstats;
71 
72 static int nfs_doio_read(struct buf *, struct uio *);
73 static int nfs_doio_write(struct buf *, struct uio *);
74 static int nfs_doio_phys(struct buf *, struct uio *);
75 
76 /*
77  * Vnode op for read using bio
78  * Any similarity to readip() is purely coincidental
79  */
80 int
81 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag,
82 	    kauth_cred_t cred, int cflag)
83 {
84 	struct nfsnode *np = VTONFS(vp);
85 	struct buf *bp = NULL, *rabp;
86 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
87 	struct nfsdircache *ndp = NULL, *nndp = NULL;
88 	void *baddr;
89 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
90 	int enough = 0;
91 	struct dirent *dp, *pdp, *edp, *ep;
92 	off_t curoff = 0;
93 	int advice;
94 	struct lwp *l = curlwp;
95 
96 #ifdef DIAGNOSTIC
97 	if (uio->uio_rw != UIO_READ)
98 		panic("nfs_read mode");
99 #endif
100 	if (uio->uio_resid == 0)
101 		return (0);
102 	if (vp->v_type != VDIR && uio->uio_offset < 0)
103 		return (EINVAL);
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, l);
108 #endif
109 	if (vp->v_type != VDIR &&
110 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
111 		return (EFBIG);
112 
113 	/*
114 	 * For nfs, cache consistency can only be maintained approximately.
115 	 * Although RFC1094 does not specify the criteria, the following is
116 	 * believed to be compatible with the reference port.
117 	 *
118 	 * If the file's modify time on the server has changed since the
119 	 * last read rpc or you have written to the file,
120 	 * you may have lost data cache consistency with the
121 	 * server, so flush all of the file's data out of the cache.
122 	 * Then force a getattr rpc to ensure that you have up to date
123 	 * attributes.
124 	 * NB: This implies that cache data can be read when up to
125 	 * nfs_attrtimeo seconds out of date. If you find that you need current
126 	 * attributes this could be forced by setting n_attrstamp to 0 before
127 	 * the VOP_GETATTR() call.
128 	 */
129 
130 	if (vp->v_type != VLNK) {
131 		error = nfs_flushstalebuf(vp, cred, l,
132 		    NFS_FLUSHSTALEBUF_MYWRITE);
133 		if (error)
134 			return error;
135 	}
136 
137 	do {
138 	    /*
139 	     * Don't cache symlinks.
140 	     */
141 	    if ((vp->v_vflag & VV_ROOT) && vp->v_type == VLNK) {
142 		return (nfs_readlinkrpc(vp, uio, cred));
143 	    }
144 	    baddr = (void *)0;
145 	    switch (vp->v_type) {
146 	    case VREG:
147 		nfsstats.biocache_reads++;
148 
149 		advice = IO_ADV_DECODE(ioflag);
150 		error = 0;
151 		while (uio->uio_resid > 0) {
152 			vsize_t bytelen;
153 
154 			nfs_delayedtruncate(vp);
155 			if (np->n_size <= uio->uio_offset) {
156 				break;
157 			}
158 			bytelen =
159 			    MIN(np->n_size - uio->uio_offset, uio->uio_resid);
160 			error = ubc_uiomove(&vp->v_uobj, uio, bytelen,
161 			    advice, UBC_READ | UBC_PARTIALOK |
162 			    (UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0));
163 			if (error) {
164 				/*
165 				 * XXXkludge
166 				 * the file has been truncated on the server.
167 				 * there isn't much we can do.
168 				 */
169 				if (uio->uio_offset >= np->n_size) {
170 					/* end of file */
171 					error = 0;
172 				} else {
173 					break;
174 				}
175 			}
176 		}
177 		break;
178 
179 	    case VLNK:
180 		nfsstats.biocache_readlinks++;
181 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, l);
182 		if (!bp)
183 			return (EINTR);
184 		if ((bp->b_oflags & BO_DONE) == 0) {
185 			bp->b_flags |= B_READ;
186 			error = nfs_doio(bp);
187 			if (error) {
188 				brelse(bp, 0);
189 				return (error);
190 			}
191 		}
192 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
193 		got_buf = 1;
194 		on = 0;
195 		break;
196 	    case VDIR:
197 diragain:
198 		nfsstats.biocache_readdirs++;
199 		ndp = nfs_searchdircache(vp, uio->uio_offset,
200 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
201 		if (!ndp) {
202 			/*
203 			 * We've been handed a cookie that is not
204 			 * in the cache. If we're not translating
205 			 * 32 <-> 64, it may be a value that was
206 			 * flushed out of the cache because it grew
207 			 * too big. Let the server judge if it's
208 			 * valid or not. In the translation case,
209 			 * we have no way of validating this value,
210 			 * so punt.
211 			 */
212 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
213 				return (EINVAL);
214 			ndp = nfs_enterdircache(vp, uio->uio_offset,
215 				uio->uio_offset, 0, 0);
216 		}
217 
218 		if (NFS_EOFVALID(np) &&
219 		    ndp->dc_cookie == np->n_direofoffset) {
220 			nfs_putdircache(np, ndp);
221 			nfsstats.direofcache_hits++;
222 			return (0);
223 		}
224 
225 		bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, l);
226 		if (!bp)
227 		    return (EINTR);
228 		if ((bp->b_oflags & BO_DONE) == 0) {
229 		    bp->b_flags |= B_READ;
230 		    bp->b_dcookie = ndp->dc_blkcookie;
231 		    error = nfs_doio(bp);
232 		    if (error) {
233 			/*
234 			 * Yuck! The directory has been modified on the
235 			 * server. Punt and let the userland code
236 			 * deal with it.
237 			 */
238 			nfs_putdircache(np, ndp);
239 			brelse(bp, 0);
240 			/*
241 			 * nfs_request maps NFSERR_BAD_COOKIE to EINVAL.
242 			 */
243 			if (error == EINVAL) { /* NFSERR_BAD_COOKIE */
244 			    nfs_invaldircache(vp, 0);
245 			    nfs_vinvalbuf(vp, 0, cred, l, 1);
246 			}
247 			return (error);
248 		    }
249 		}
250 
251 		/*
252 		 * Just return if we hit EOF right away with this
253 		 * block. Always check here, because direofoffset
254 		 * may have been set by an nfsiod since the last
255 		 * check.
256 		 *
257 		 * also, empty block implies EOF.
258 		 */
259 
260 		if (bp->b_bcount == bp->b_resid ||
261 		    (NFS_EOFVALID(np) &&
262 		    ndp->dc_blkcookie == np->n_direofoffset)) {
263 			KASSERT(bp->b_bcount != bp->b_resid ||
264 			    ndp->dc_blkcookie == bp->b_dcookie);
265 			nfs_putdircache(np, ndp);
266 			brelse(bp, BC_NOCACHE);
267 			return 0;
268 		}
269 
270 		/*
271 		 * Find the entry we were looking for in the block.
272 		 */
273 
274 		en = ndp->dc_entry;
275 
276 		pdp = dp = (struct dirent *)bp->b_data;
277 		edp = (struct dirent *)(void *)((char *)bp->b_data + bp->b_bcount -
278 		    bp->b_resid);
279 		enn = 0;
280 		while (enn < en && dp < edp) {
281 			pdp = dp;
282 			dp = _DIRENT_NEXT(dp);
283 			enn++;
284 		}
285 
286 		/*
287 		 * If the entry number was bigger than the number of
288 		 * entries in the block, or the cookie of the previous
289 		 * entry doesn't match, the directory cache is
290 		 * stale. Flush it and try again (i.e. go to
291 		 * the server).
292 		 */
293 		if (dp >= edp || (struct dirent *)_DIRENT_NEXT(dp) > edp ||
294 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
295 #ifdef DEBUG
296 		    	printf("invalid cache: %p %p %p off %jx %jx\n",
297 				pdp, dp, edp,
298 				(uintmax_t)uio->uio_offset,
299 				(uintmax_t)NFS_GETCOOKIE(pdp));
300 #endif
301 			nfs_putdircache(np, ndp);
302 			brelse(bp, 0);
303 			nfs_invaldircache(vp, 0);
304 			nfs_vinvalbuf(vp, 0, cred, l, 0);
305 			goto diragain;
306 		}
307 
308 		on = (char *)dp - (char *)bp->b_data;
309 
310 		/*
311 		 * Cache all entries that may be exported to the
312 		 * user, as they may be thrown back at us. The
313 		 * NFSBIO_CACHECOOKIES flag indicates that all
314 		 * entries are being 'exported', so cache them all.
315 		 */
316 
317 		if (en == 0 && pdp == dp) {
318 			dp = _DIRENT_NEXT(dp);
319 			enn++;
320 		}
321 
322 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
323 			n = uio->uio_resid;
324 			enough = 1;
325 		} else
326 			n = bp->b_bcount - bp->b_resid - on;
327 
328 		ep = (struct dirent *)(void *)((char *)bp->b_data + on + n);
329 
330 		/*
331 		 * Find last complete entry to copy, caching entries
332 		 * (if requested) as we go.
333 		 */
334 
335 		while (dp < ep && (struct dirent *)_DIRENT_NEXT(dp) <= ep) {
336 			if (cflag & NFSBIO_CACHECOOKIES) {
337 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
338 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
339 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
340 					NFS_STASHCOOKIE32(pdp,
341 					    nndp->dc_cookie32);
342 				}
343 				nfs_putdircache(np, nndp);
344 			}
345 			pdp = dp;
346 			dp = _DIRENT_NEXT(dp);
347 			enn++;
348 		}
349 		nfs_putdircache(np, ndp);
350 
351 		/*
352 		 * If the last requested entry was not the last in the
353 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
354 		 * cache the cookie of the last requested one, and
355 		 * set of the offset to it.
356 		 */
357 
358 		if ((on + n) < bp->b_bcount - bp->b_resid) {
359 			curoff = NFS_GETCOOKIE(pdp);
360 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
361 			    enn, bp->b_lblkno);
362 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
363 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
364 				curoff = nndp->dc_cookie32;
365 			}
366 			nfs_putdircache(np, nndp);
367 		} else
368 			curoff = bp->b_dcookie;
369 
370 		/*
371 		 * Always cache the entry for the next block,
372 		 * so that readaheads can use it.
373 		 */
374 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
375 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
376 			if (curoff == bp->b_dcookie) {
377 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
378 				curoff = nndp->dc_cookie32;
379 			}
380 		}
381 
382 		n = (char *)_DIRENT_NEXT(pdp) - ((char *)bp->b_data + on);
383 
384 		/*
385 		 * If not eof and read aheads are enabled, start one.
386 		 * (You need the current block first, so that you have the
387 		 *  directory offset cookie of the next block.)
388 		 */
389 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
390 		    !NFS_EOFVALID(np)) {
391 			rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp),
392 						NFS_DIRBLKSIZ, l);
393 			if (rabp) {
394 			    if ((rabp->b_oflags & (BO_DONE | BO_DELWRI)) == 0) {
395 				rabp->b_dcookie = nndp->dc_cookie;
396 				rabp->b_flags |= (B_READ | B_ASYNC);
397 				if (nfs_asyncio(rabp)) {
398 				    brelse(rabp, BC_INVAL);
399 				}
400 			    } else
401 				brelse(rabp, 0);
402 			}
403 		}
404 		nfs_putdircache(np, nndp);
405 		got_buf = 1;
406 		break;
407 	    default:
408 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
409 		break;
410 	    }
411 
412 	    if (n > 0) {
413 		if (!baddr)
414 			baddr = bp->b_data;
415 		error = uiomove((char *)baddr + on, (int)n, uio);
416 	    }
417 	    switch (vp->v_type) {
418 	    case VREG:
419 		break;
420 	    case VLNK:
421 		n = 0;
422 		break;
423 	    case VDIR:
424 		uio->uio_offset = curoff;
425 		if (enough)
426 			n = 0;
427 		break;
428 	    default:
429 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
430 	    }
431 	    if (got_buf)
432 		brelse(bp, 0);
433 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
434 	return (error);
435 }
436 
437 /*
438  * Vnode op for write using bio
439  */
440 int
441 nfs_write(void *v)
442 {
443 	struct vop_write_args /* {
444 		struct vnode *a_vp;
445 		struct uio *a_uio;
446 		int  a_ioflag;
447 		kauth_cred_t a_cred;
448 	} */ *ap = v;
449 	struct uio *uio = ap->a_uio;
450 	struct lwp *l = curlwp;
451 	struct vnode *vp = ap->a_vp;
452 	struct nfsnode *np = VTONFS(vp);
453 	kauth_cred_t cred = ap->a_cred;
454 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
455 	voff_t oldoff, origoff;
456 	vsize_t bytelen;
457 	int error = 0;
458 	int ioflag = ap->a_ioflag;
459 	int extended = 0, wrotedata = 0;
460 
461 #ifdef DIAGNOSTIC
462 	if (uio->uio_rw != UIO_WRITE)
463 		panic("nfs_write mode");
464 #endif
465 	if (vp->v_type != VREG)
466 		return (EIO);
467 	if (np->n_flag & NWRITEERR) {
468 		np->n_flag &= ~NWRITEERR;
469 		return (np->n_error);
470 	}
471 #ifndef NFS_V2_ONLY
472 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
473 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
474 		(void)nfs_fsinfo(nmp, vp, cred, l);
475 #endif
476 	if (ioflag & IO_APPEND) {
477 		NFS_INVALIDATE_ATTRCACHE(np);
478 		error = nfs_flushstalebuf(vp, cred, l,
479 		    NFS_FLUSHSTALEBUF_MYWRITE);
480 		if (error)
481 			return (error);
482 		uio->uio_offset = np->n_size;
483 
484 		/*
485 		 * This is already checked above VOP_WRITE, but recheck
486 		 * the append case here to make sure our idea of the
487 		 * file size is as fresh as possible.
488 		 */
489 		if (uio->uio_offset + uio->uio_resid >
490 		      l->l_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
491 			mutex_enter(proc_lock);
492 			psignal(l->l_proc, SIGXFSZ);
493 			mutex_exit(proc_lock);
494 			return (EFBIG);
495 		}
496 	}
497 	if (uio->uio_offset < 0)
498 		return (EINVAL);
499 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
500 		return (EFBIG);
501 	if (uio->uio_resid == 0)
502 		return (0);
503 
504 	origoff = uio->uio_offset;
505 	do {
506 		bool overwrite; /* if we are overwriting whole pages */
507 		u_quad_t oldsize;
508 		oldoff = uio->uio_offset;
509 		bytelen = uio->uio_resid;
510 
511 		nfsstats.biocache_writes++;
512 
513 		oldsize = np->n_size;
514 		np->n_flag |= NMODIFIED;
515 		if (np->n_size < uio->uio_offset + bytelen) {
516 			np->n_size = uio->uio_offset + bytelen;
517 		}
518 		overwrite = false;
519 		if ((uio->uio_offset & PAGE_MASK) == 0) {
520 			if ((vp->v_vflag & VV_MAPPED) == 0 &&
521 			    bytelen > PAGE_SIZE) {
522 				bytelen = trunc_page(bytelen);
523 				overwrite = true;
524 			} else if ((bytelen & PAGE_MASK) == 0 &&
525 			    uio->uio_offset >= vp->v_size) {
526 				overwrite = true;
527 			}
528 		}
529 		if (vp->v_size < uio->uio_offset + bytelen) {
530 			uvm_vnp_setwritesize(vp, uio->uio_offset + bytelen);
531 		}
532 		error = ubc_uiomove(&vp->v_uobj, uio, bytelen,
533 		    UVM_ADV_RANDOM, UBC_WRITE | UBC_PARTIALOK |
534 		    (overwrite ? UBC_FAULTBUSY : 0) |
535 		    (UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0));
536 		if (error) {
537 			uvm_vnp_setwritesize(vp, vp->v_size);
538 			if (overwrite && np->n_size != oldsize) {
539 				/*
540 				 * backout size and free pages past eof.
541 				 */
542 				np->n_size = oldsize;
543 				mutex_enter(&vp->v_interlock);
544 				(void)VOP_PUTPAGES(vp, round_page(vp->v_size),
545 				    0, PGO_SYNCIO | PGO_FREE);
546 			}
547 			break;
548 		}
549 		wrotedata = 1;
550 
551 		/*
552 		 * update UVM's notion of the size now that we've
553 		 * copied the data into the vnode's pages.
554 		 */
555 
556 		if (vp->v_size < uio->uio_offset) {
557 			uvm_vnp_setsize(vp, uio->uio_offset);
558 			extended = 1;
559 		}
560 
561 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
562 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
563 			mutex_enter(&vp->v_interlock);
564 			error = VOP_PUTPAGES(vp,
565 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
566 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
567 				       ~(nmp->nm_wsize - 1)), PGO_CLEANIT);
568 		}
569 	} while (uio->uio_resid > 0);
570 	if (wrotedata)
571 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0));
572 	if (error == 0 && (ioflag & IO_SYNC) != 0) {
573 		mutex_enter(&vp->v_interlock);
574 		error = VOP_PUTPAGES(vp,
575 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
576 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
577 			       ~(nmp->nm_wsize - 1)),
578 		    PGO_CLEANIT | PGO_SYNCIO);
579 	}
580 	return error;
581 }
582 
583 /*
584  * Get an nfs cache block.
585  * Allocate a new one if the block isn't currently in the cache
586  * and return the block marked busy. If the calling process is
587  * interrupted by a signal for an interruptible mount point, return
588  * NULL.
589  */
590 struct buf *
591 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct lwp *l)
592 {
593 	struct buf *bp;
594 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
595 
596 	if (nmp->nm_flag & NFSMNT_INT) {
597 		bp = getblk(vp, bn, size, PCATCH, 0);
598 		while (bp == NULL) {
599 			if (nfs_sigintr(nmp, NULL, l))
600 				return (NULL);
601 			bp = getblk(vp, bn, size, 0, 2 * hz);
602 		}
603 	} else
604 		bp = getblk(vp, bn, size, 0, 0);
605 	return (bp);
606 }
607 
608 /*
609  * Flush and invalidate all dirty buffers. If another process is already
610  * doing the flush, just wait for completion.
611  */
612 int
613 nfs_vinvalbuf(struct vnode *vp, int flags, kauth_cred_t cred,
614 		struct lwp *l, int intrflg)
615 {
616 	struct nfsnode *np = VTONFS(vp);
617 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
618 	int error = 0, slptimeo;
619 	bool catch;
620 
621 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
622 		intrflg = 0;
623 	if (intrflg) {
624 		catch = true;
625 		slptimeo = 2 * hz;
626 	} else {
627 		catch = false;
628 		slptimeo = 0;
629 	}
630 	/*
631 	 * First wait for any other process doing a flush to complete.
632 	 */
633 	mutex_enter(&vp->v_interlock);
634 	while (np->n_flag & NFLUSHINPROG) {
635 		np->n_flag |= NFLUSHWANT;
636 		error = mtsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
637 			slptimeo, &vp->v_interlock);
638 		if (error && intrflg && nfs_sigintr(nmp, NULL, l)) {
639 			mutex_exit(&vp->v_interlock);
640 			return EINTR;
641 		}
642 	}
643 
644 	/*
645 	 * Now, flush as required.
646 	 */
647 	np->n_flag |= NFLUSHINPROG;
648 	mutex_exit(&vp->v_interlock);
649 	error = vinvalbuf(vp, flags, cred, l, catch, 0);
650 	while (error) {
651 		if (intrflg && nfs_sigintr(nmp, NULL, l)) {
652 			error = EINTR;
653 			break;
654 		}
655 		error = vinvalbuf(vp, flags, cred, l, 0, slptimeo);
656 	}
657 	mutex_enter(&vp->v_interlock);
658 	if (error == 0)
659 		np->n_flag &= ~NMODIFIED;
660 	np->n_flag &= ~NFLUSHINPROG;
661 	if (np->n_flag & NFLUSHWANT) {
662 		np->n_flag &= ~NFLUSHWANT;
663 		wakeup(&np->n_flag);
664 	}
665 	mutex_exit(&vp->v_interlock);
666 	return error;
667 }
668 
669 /*
670  * nfs_flushstalebuf: flush cache if it's stale.
671  *
672  * => caller shouldn't own any pages or buffers which belong to the vnode.
673  */
674 
675 int
676 nfs_flushstalebuf(struct vnode *vp, kauth_cred_t cred, struct lwp *l,
677     int flags)
678 {
679 	struct nfsnode *np = VTONFS(vp);
680 	struct vattr vattr;
681 	int error;
682 
683 	if (np->n_flag & NMODIFIED) {
684 		if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0
685 		    || vp->v_type != VREG) {
686 			error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1);
687 			if (error)
688 				return error;
689 			if (vp->v_type == VDIR) {
690 				nfs_invaldircache(vp, 0);
691 			}
692 		} else {
693 			/*
694 			 * XXX assuming writes are ours.
695 			 */
696 		}
697 		NFS_INVALIDATE_ATTRCACHE(np);
698 		error = VOP_GETATTR(vp, &vattr, cred);
699 		if (error)
700 			return error;
701 		np->n_mtime = vattr.va_mtime;
702 	} else {
703 		error = VOP_GETATTR(vp, &vattr, cred);
704 		if (error)
705 			return error;
706 		if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
707 			if (vp->v_type == VDIR) {
708 				nfs_invaldircache(vp, 0);
709 			}
710 			error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1);
711 			if (error)
712 				return error;
713 			np->n_mtime = vattr.va_mtime;
714 		}
715 	}
716 
717 	return error;
718 }
719 
720 /*
721  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
722  * This is mainly to avoid queueing async I/O requests when the nfsiods
723  * are all hung on a dead server.
724  */
725 
726 int
727 nfs_asyncio(struct buf *bp)
728 {
729 	struct nfs_iod *iod;
730 	struct nfsmount *nmp;
731 	int slptimeo = 0, error;
732 	bool catch = false;
733 
734 	if (nfs_numasync == 0)
735 		return (EIO);
736 
737 	nmp = VFSTONFS(bp->b_vp->v_mount);
738 again:
739 	if (nmp->nm_flag & NFSMNT_INT)
740 		catch = true;
741 
742 	/*
743 	 * Find a free iod to process this request.
744 	 */
745 
746 	mutex_enter(&nfs_iodlist_lock);
747 	iod = LIST_FIRST(&nfs_iodlist_idle);
748 	if (iod) {
749 		/*
750 		 * Found one, so wake it up and tell it which
751 		 * mount to process.
752 		 */
753 		LIST_REMOVE(iod, nid_idle);
754 		mutex_enter(&iod->nid_lock);
755 		mutex_exit(&nfs_iodlist_lock);
756 		KASSERT(iod->nid_mount == NULL);
757 		iod->nid_mount = nmp;
758 		cv_signal(&iod->nid_cv);
759 		mutex_enter(&nmp->nm_lock);
760 		mutex_exit(&iod->nid_lock);
761 		nmp->nm_bufqiods++;
762 		if (nmp->nm_bufqlen < 2 * nmp->nm_bufqiods) {
763 			cv_broadcast(&nmp->nm_aiocv);
764 		}
765 	} else {
766 		mutex_exit(&nfs_iodlist_lock);
767 		mutex_enter(&nmp->nm_lock);
768 	}
769 
770 	KASSERT(mutex_owned(&nmp->nm_lock));
771 
772 	/*
773 	 * If we have an iod which can process the request, then queue
774 	 * the buffer.  However, even if we have an iod, do not initiate
775 	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
776 	 * is via local loopback, we may put curproc (pagedaemon) to sleep
777 	 * waiting for the writes to complete. But the server (ourself)
778 	 * may block the write, waiting for its (ie., our) pagedaemon
779 	 * to produce clean pages to handle the write: deadlock.
780 	 * XXX: start non-loopback mounts straight away?  If "lots free",
781 	 * let pagedaemon start loopback writes anyway?
782 	 */
783 	if (nmp->nm_bufqiods > 0) {
784 
785 		/*
786 		 * Ensure that the queue never grows too large.
787 		 */
788 		if (curlwp == uvm.pagedaemon_lwp) {
789 	  		/* Enque for later, to avoid free-page deadlock */
790 		} else while (nmp->nm_bufqlen >= 2 * nmp->nm_bufqiods) {
791 			if (catch) {
792 				error = cv_timedwait_sig(&nmp->nm_aiocv,
793 				    &nmp->nm_lock, slptimeo);
794 			} else {
795 				error = cv_timedwait(&nmp->nm_aiocv,
796 				    &nmp->nm_lock, slptimeo);
797 			}
798 			if (error) {
799 				if (nfs_sigintr(nmp, NULL, curlwp)) {
800 					mutex_exit(&nmp->nm_lock);
801 					return (EINTR);
802 				}
803 				if (catch) {
804 					catch = false;
805 					slptimeo = 2 * hz;
806 				}
807 			}
808 
809 			/*
810 			 * We might have lost our iod while sleeping,
811 			 * so check and loop if necessary.
812 			 */
813 
814 			if (nmp->nm_bufqiods == 0) {
815 				mutex_exit(&nmp->nm_lock);
816 				goto again;
817 			}
818 		}
819 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
820 		nmp->nm_bufqlen++;
821 		mutex_exit(&nmp->nm_lock);
822 		return (0);
823 	}
824 	mutex_exit(&nmp->nm_lock);
825 
826 	/*
827 	 * All the iods are busy on other mounts, so return EIO to
828 	 * force the caller to process the i/o synchronously.
829 	 */
830 
831 	return (EIO);
832 }
833 
834 /*
835  * nfs_doio for read.
836  */
837 static int
838 nfs_doio_read(struct buf *bp, struct uio *uiop)
839 {
840 	struct vnode *vp = bp->b_vp;
841 	struct nfsnode *np = VTONFS(vp);
842 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
843 	int error = 0;
844 
845 	uiop->uio_rw = UIO_READ;
846 	switch (vp->v_type) {
847 	case VREG:
848 		nfsstats.read_bios++;
849 		error = nfs_readrpc(vp, uiop);
850 		if (!error && uiop->uio_resid) {
851 			int diff, len;
852 
853 			/*
854 			 * If uio_resid > 0, there is a hole in the file and
855 			 * no writes after the hole have been pushed to
856 			 * the server yet or the file has been truncated
857 			 * on the server.
858 			 * Just zero fill the rest of the valid area.
859 			 */
860 
861 			KASSERT(vp->v_size >=
862 			    uiop->uio_offset + uiop->uio_resid);
863 			diff = bp->b_bcount - uiop->uio_resid;
864 			len = uiop->uio_resid;
865 			memset((char *)bp->b_data + diff, 0, len);
866 			uiop->uio_resid = 0;
867 		}
868 #if 0
869 		if (uiop->uio_lwp && (vp->v_iflag & VI_TEXT) &&
870 		    timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)) {
871 		    	mutex_enter(proc_lock);
872 			killproc(uiop->uio_lwp->l_proc, "process text file was modified");
873 		    	mutex_exit(proc_lock);
874 #if 0 /* XXX NJWLWP */
875 			uiop->uio_lwp->l_proc->p_holdcnt++;
876 #endif
877 		}
878 #endif
879 		break;
880 	case VLNK:
881 		KASSERT(uiop->uio_offset == (off_t)0);
882 		nfsstats.readlink_bios++;
883 		error = nfs_readlinkrpc(vp, uiop, np->n_rcred);
884 		break;
885 	case VDIR:
886 		nfsstats.readdir_bios++;
887 		uiop->uio_offset = bp->b_dcookie;
888 #ifndef NFS_V2_ONLY
889 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
890 			error = nfs_readdirplusrpc(vp, uiop,
891 			    curlwp->l_cred);
892 			/*
893 			 * nfs_request maps NFSERR_NOTSUPP to ENOTSUP.
894 			 */
895 			if (error == ENOTSUP)
896 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
897 		}
898 #else
899 		nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
900 #endif
901 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
902 			error = nfs_readdirrpc(vp, uiop,
903 			    curlwp->l_cred);
904 		if (!error) {
905 			bp->b_dcookie = uiop->uio_offset;
906 		}
907 		break;
908 	default:
909 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
910 		break;
911 	}
912 	bp->b_error = error;
913 	return error;
914 }
915 
916 /*
917  * nfs_doio for write.
918  */
919 static int
920 nfs_doio_write(struct buf *bp, struct uio *uiop)
921 {
922 	struct vnode *vp = bp->b_vp;
923 	struct nfsnode *np = VTONFS(vp);
924 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
925 	int iomode;
926 	bool stalewriteverf = false;
927 	int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT;
928 	struct vm_page **pgs, *spgs[UBC_MAX_PAGES];
929 #ifndef NFS_V2_ONLY
930 	bool needcommit = true; /* need only COMMIT RPC */
931 #else
932 	bool needcommit = false; /* need only COMMIT RPC */
933 #endif
934 	bool pageprotected;
935 	struct uvm_object *uobj = &vp->v_uobj;
936 	int error;
937 	off_t off, cnt;
938 
939 	if (npages < __arraycount(spgs))
940 		pgs = spgs;
941 	else {
942 		if ((pgs = kmem_alloc(sizeof(*pgs) * npages, KM_NOSLEEP)) ==
943 		    NULL)
944 			return ENOMEM;
945 	}
946 
947 	if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) {
948 		iomode = NFSV3WRITE_UNSTABLE;
949 	} else {
950 		iomode = NFSV3WRITE_FILESYNC;
951 	}
952 
953 #ifndef NFS_V2_ONLY
954 again:
955 #endif
956 	rw_enter(&nmp->nm_writeverflock, RW_READER);
957 
958 	for (i = 0; i < npages; i++) {
959 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
960 		if (pgs[i]->uobject == uobj &&
961 		    pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) {
962 			KASSERT(pgs[i]->flags & PG_BUSY);
963 			/*
964 			 * this page belongs to our object.
965 			 */
966 			mutex_enter(&uobj->vmobjlock);
967 			/*
968 			 * write out the page stably if it's about to
969 			 * be released because we can't resend it
970 			 * on the server crash.
971 			 *
972 			 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be
973 			 * changed until unbusy the page.
974 			 */
975 			if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT))
976 				iomode = NFSV3WRITE_FILESYNC;
977 			/*
978 			 * if we met a page which hasn't been sent yet,
979 			 * we need do WRITE RPC.
980 			 */
981 			if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0)
982 				needcommit = false;
983 			mutex_exit(&uobj->vmobjlock);
984 		} else {
985 			iomode = NFSV3WRITE_FILESYNC;
986 			needcommit = false;
987 		}
988 	}
989 	if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) {
990 		mutex_enter(&uobj->vmobjlock);
991 		for (i = 0; i < npages; i++) {
992 			pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY;
993 			pmap_page_protect(pgs[i], VM_PROT_READ);
994 		}
995 		mutex_exit(&uobj->vmobjlock);
996 		pageprotected = true; /* pages can't be modified during i/o. */
997 	} else
998 		pageprotected = false;
999 
1000 	/*
1001 	 * Send the data to the server if necessary,
1002 	 * otherwise just send a commit rpc.
1003 	 */
1004 #ifndef NFS_V2_ONLY
1005 	if (needcommit) {
1006 
1007 		/*
1008 		 * If the buffer is in the range that we already committed,
1009 		 * there's nothing to do.
1010 		 *
1011 		 * If it's in the range that we need to commit, push the
1012 		 * whole range at once, otherwise only push the buffer.
1013 		 * In both these cases, acquire the commit lock to avoid
1014 		 * other processes modifying the range.
1015 		 */
1016 
1017 		off = uiop->uio_offset;
1018 		cnt = bp->b_bcount;
1019 		mutex_enter(&np->n_commitlock);
1020 		if (!nfs_in_committed_range(vp, off, bp->b_bcount)) {
1021 			bool pushedrange;
1022 			if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) {
1023 				pushedrange = true;
1024 				off = np->n_pushlo;
1025 				cnt = np->n_pushhi - np->n_pushlo;
1026 			} else {
1027 				pushedrange = false;
1028 			}
1029 			error = nfs_commit(vp, off, cnt, curlwp);
1030 			if (error == 0) {
1031 				if (pushedrange) {
1032 					nfs_merge_commit_ranges(vp);
1033 				} else {
1034 					nfs_add_committed_range(vp, off, cnt);
1035 				}
1036 			}
1037 		} else {
1038 			error = 0;
1039 		}
1040 		mutex_exit(&np->n_commitlock);
1041 		rw_exit(&nmp->nm_writeverflock);
1042 		if (!error) {
1043 			/*
1044 			 * pages are now on stable storage.
1045 			 */
1046 			uiop->uio_resid = 0;
1047 			mutex_enter(&uobj->vmobjlock);
1048 			for (i = 0; i < npages; i++) {
1049 				pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1050 			}
1051 			mutex_exit(&uobj->vmobjlock);
1052 			goto out;
1053 		} else if (error == NFSERR_STALEWRITEVERF) {
1054 			nfs_clearcommit(vp->v_mount);
1055 			goto again;
1056 		}
1057 		if (error) {
1058 			bp->b_error = np->n_error = error;
1059 			np->n_flag |= NWRITEERR;
1060 		}
1061 		goto out;
1062 	}
1063 #endif
1064 	off = uiop->uio_offset;
1065 	cnt = bp->b_bcount;
1066 	uiop->uio_rw = UIO_WRITE;
1067 	nfsstats.write_bios++;
1068 	error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf);
1069 #ifndef NFS_V2_ONLY
1070 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1071 		/*
1072 		 * we need to commit pages later.
1073 		 */
1074 		mutex_enter(&np->n_commitlock);
1075 		nfs_add_tobecommitted_range(vp, off, cnt);
1076 		/*
1077 		 * if there can be too many uncommitted pages, commit them now.
1078 		 */
1079 		if (np->n_pushhi - np->n_pushlo > nfs_commitsize) {
1080 			off = np->n_pushlo;
1081 			cnt = nfs_commitsize >> 1;
1082 			error = nfs_commit(vp, off, cnt, curlwp);
1083 			if (!error) {
1084 				nfs_add_committed_range(vp, off, cnt);
1085 				nfs_del_tobecommitted_range(vp, off, cnt);
1086 			}
1087 			if (error == NFSERR_STALEWRITEVERF) {
1088 				stalewriteverf = true;
1089 				error = 0; /* it isn't a real error */
1090 			}
1091 		} else {
1092 			/*
1093 			 * re-dirty pages so that they will be passed
1094 			 * to us later again.
1095 			 */
1096 			mutex_enter(&uobj->vmobjlock);
1097 			for (i = 0; i < npages; i++) {
1098 				pgs[i]->flags &= ~PG_CLEAN;
1099 			}
1100 			mutex_exit(&uobj->vmobjlock);
1101 		}
1102 		mutex_exit(&np->n_commitlock);
1103 	} else
1104 #endif
1105 	if (!error) {
1106 		/*
1107 		 * pages are now on stable storage.
1108 		 */
1109 		mutex_enter(&np->n_commitlock);
1110 		nfs_del_committed_range(vp, off, cnt);
1111 		mutex_exit(&np->n_commitlock);
1112 		mutex_enter(&uobj->vmobjlock);
1113 		for (i = 0; i < npages; i++) {
1114 			pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1115 		}
1116 		mutex_exit(&uobj->vmobjlock);
1117 	} else {
1118 		/*
1119 		 * we got an error.
1120 		 */
1121 		bp->b_error = np->n_error = error;
1122 		np->n_flag |= NWRITEERR;
1123 	}
1124 
1125 	rw_exit(&nmp->nm_writeverflock);
1126 
1127 
1128 	if (stalewriteverf) {
1129 		nfs_clearcommit(vp->v_mount);
1130 	}
1131 #ifndef NFS_V2_ONLY
1132 out:
1133 #endif
1134 	if (pgs != spgs)
1135 		kmem_free(pgs, sizeof(*pgs) * npages);
1136 	return error;
1137 }
1138 
1139 /*
1140  * nfs_doio for B_PHYS.
1141  */
1142 static int
1143 nfs_doio_phys(struct buf *bp, struct uio *uiop)
1144 {
1145 	struct vnode *vp = bp->b_vp;
1146 	int error;
1147 
1148 	uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
1149 	if (bp->b_flags & B_READ) {
1150 		uiop->uio_rw = UIO_READ;
1151 		nfsstats.read_physios++;
1152 		error = nfs_readrpc(vp, uiop);
1153 	} else {
1154 		int iomode = NFSV3WRITE_DATASYNC;
1155 		bool stalewriteverf;
1156 		struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1157 
1158 		uiop->uio_rw = UIO_WRITE;
1159 		nfsstats.write_physios++;
1160 		rw_enter(&nmp->nm_writeverflock, RW_READER);
1161 		error = nfs_writerpc(vp, uiop, &iomode, false, &stalewriteverf);
1162 		rw_exit(&nmp->nm_writeverflock);
1163 		if (stalewriteverf) {
1164 			nfs_clearcommit(bp->b_vp->v_mount);
1165 		}
1166 	}
1167 	bp->b_error = error;
1168 	return error;
1169 }
1170 
1171 /*
1172  * Do an I/O operation to/from a cache block. This may be called
1173  * synchronously or from an nfsiod.
1174  */
1175 int
1176 nfs_doio(struct buf *bp)
1177 {
1178 	int error;
1179 	struct uio uio;
1180 	struct uio *uiop = &uio;
1181 	struct iovec io;
1182 	UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist);
1183 
1184 	uiop->uio_iov = &io;
1185 	uiop->uio_iovcnt = 1;
1186 	uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
1187 	UIO_SETUP_SYSSPACE(uiop);
1188 	io.iov_base = bp->b_data;
1189 	io.iov_len = uiop->uio_resid = bp->b_bcount;
1190 
1191 	/*
1192 	 * Historically, paging was done with physio, but no more...
1193 	 */
1194 	if (bp->b_flags & B_PHYS) {
1195 		/*
1196 		 * ...though reading /dev/drum still gets us here.
1197 		 */
1198 		error = nfs_doio_phys(bp, uiop);
1199 	} else if (bp->b_flags & B_READ) {
1200 		error = nfs_doio_read(bp, uiop);
1201 	} else {
1202 		error = nfs_doio_write(bp, uiop);
1203 	}
1204 	bp->b_resid = uiop->uio_resid;
1205 	biodone(bp);
1206 	return (error);
1207 }
1208 
1209 /*
1210  * Vnode op for VM getpages.
1211  */
1212 
1213 int
1214 nfs_getpages(void *v)
1215 {
1216 	struct vop_getpages_args /* {
1217 		struct vnode *a_vp;
1218 		voff_t a_offset;
1219 		struct vm_page **a_m;
1220 		int *a_count;
1221 		int a_centeridx;
1222 		vm_prot_t a_access_type;
1223 		int a_advice;
1224 		int a_flags;
1225 	} */ *ap = v;
1226 
1227 	struct vnode *vp = ap->a_vp;
1228 	struct uvm_object *uobj = &vp->v_uobj;
1229 	struct nfsnode *np = VTONFS(vp);
1230 	const int npages = *ap->a_count;
1231 	struct vm_page *pg, **pgs, **opgs, *spgs[UBC_MAX_PAGES];
1232 	off_t origoffset, len;
1233 	int i, error;
1234 	bool v3 = NFS_ISV3(vp);
1235 	bool write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1236 	bool locked = (ap->a_flags & PGO_LOCKED) != 0;
1237 
1238 	/*
1239 	 * If we are not locked we are not really using opgs,
1240 	 * so just initialize it
1241 	 */
1242 	if (!locked || npages < __arraycount(spgs))
1243 		opgs = spgs;
1244 	else {
1245 		if ((opgs = kmem_alloc(npages * sizeof(*opgs), KM_NOSLEEP)) ==
1246 		    NULL)
1247 			return ENOMEM;
1248 	}
1249 
1250 	/*
1251 	 * call the genfs code to get the pages.  `pgs' may be NULL
1252 	 * when doing read-ahead.
1253 	 */
1254 	pgs = ap->a_m;
1255 	if (write && locked && v3) {
1256 		KASSERT(pgs != NULL);
1257 #ifdef DEBUG
1258 
1259 		/*
1260 		 * If PGO_LOCKED is set, real pages shouldn't exists
1261 		 * in the array.
1262 		 */
1263 
1264 		for (i = 0; i < npages; i++)
1265 			KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE);
1266 #endif
1267 		memcpy(opgs, pgs, npages * sizeof(struct vm_pages *));
1268 	}
1269 	error = genfs_getpages(v);
1270 	if (error)
1271 		goto out;
1272 
1273 	/*
1274 	 * for read faults where the nfs node is not yet marked NMODIFIED,
1275 	 * set PG_RDONLY on the pages so that we come back here if someone
1276 	 * tries to modify later via the mapping that will be entered for
1277 	 * this fault.
1278 	 */
1279 
1280 	if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) {
1281 		if (!locked) {
1282 			mutex_enter(&uobj->vmobjlock);
1283 		}
1284 		for (i = 0; i < npages; i++) {
1285 			pg = pgs[i];
1286 			if (pg == NULL || pg == PGO_DONTCARE) {
1287 				continue;
1288 			}
1289 			pg->flags |= PG_RDONLY;
1290 		}
1291 		if (!locked) {
1292 			mutex_exit(&uobj->vmobjlock);
1293 		}
1294 	}
1295 	if (!write)
1296 		goto out;
1297 
1298 	/*
1299 	 * this is a write fault, update the commit info.
1300 	 */
1301 
1302 	origoffset = ap->a_offset;
1303 	len = npages << PAGE_SHIFT;
1304 
1305 	if (v3) {
1306 		if (!locked) {
1307 			mutex_enter(&np->n_commitlock);
1308 		} else {
1309 			if (!mutex_tryenter(&np->n_commitlock)) {
1310 
1311 				/*
1312 				 * Since PGO_LOCKED is set, we need to unbusy
1313 				 * all pages fetched by genfs_getpages() above,
1314 				 * tell the caller that there are no pages
1315 				 * available and put back original pgs array.
1316 				 */
1317 
1318 				mutex_enter(&uvm_pageqlock);
1319 				uvm_page_unbusy(pgs, npages);
1320 				mutex_exit(&uvm_pageqlock);
1321 				*ap->a_count = 0;
1322 				memcpy(pgs, opgs,
1323 				    npages * sizeof(struct vm_pages *));
1324 				error = EBUSY;
1325 				goto out;
1326 			}
1327 		}
1328 		nfs_del_committed_range(vp, origoffset, len);
1329 		nfs_del_tobecommitted_range(vp, origoffset, len);
1330 	}
1331 	np->n_flag |= NMODIFIED;
1332 	if (!locked) {
1333 		mutex_enter(&uobj->vmobjlock);
1334 	}
1335 	for (i = 0; i < npages; i++) {
1336 		pg = pgs[i];
1337 		if (pg == NULL || pg == PGO_DONTCARE) {
1338 			continue;
1339 		}
1340 		pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1341 	}
1342 	if (!locked) {
1343 		mutex_exit(&uobj->vmobjlock);
1344 	}
1345 	if (v3) {
1346 		mutex_exit(&np->n_commitlock);
1347 	}
1348 out:
1349 	if (opgs != spgs)
1350 		kmem_free(opgs, sizeof(*opgs) * npages);
1351 	return error;
1352 }
1353