xref: /netbsd-src/sys/nfs/nfs_bio.c (revision d48f14661dda8638fee055ba15d35bdfb29b9fa8)
1 /*	$NetBSD: nfs_bio.c,v 1.144 2006/06/30 09:55:34 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 1.144 2006/06/30 09:55:34 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 #include <sys/kauth.h>
56 
57 #include <uvm/uvm_extern.h>
58 #include <uvm/uvm.h>
59 
60 #include <nfs/rpcv2.h>
61 #include <nfs/nfsproto.h>
62 #include <nfs/nfs.h>
63 #include <nfs/nfsmount.h>
64 #include <nfs/nqnfs.h>
65 #include <nfs/nfsnode.h>
66 #include <nfs/nfs_var.h>
67 
68 extern int nfs_numasync;
69 extern int nfs_commitsize;
70 extern struct nfsstats nfsstats;
71 
72 static int nfs_doio_read __P((struct buf *, struct uio *));
73 static int nfs_doio_write __P((struct buf *, struct uio *));
74 static int nfs_doio_phys __P((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(vp, uio, ioflag, cred, cflag)
82 	struct vnode *vp;
83 	struct uio *uio;
84 	int ioflag, cflag;
85 	kauth_cred_t cred;
86 {
87 	struct nfsnode *np = VTONFS(vp);
88 	struct buf *bp = NULL, *rabp;
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 	int advice;
97 	struct lwp *l = curlwp;
98 
99 #ifdef DIAGNOSTIC
100 	if (uio->uio_rw != UIO_READ)
101 		panic("nfs_read mode");
102 #endif
103 	if (uio->uio_resid == 0)
104 		return (0);
105 	if (vp->v_type != VDIR && uio->uio_offset < 0)
106 		return (EINVAL);
107 #ifndef NFS_V2_ONLY
108 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
109 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
110 		(void)nfs_fsinfo(nmp, vp, cred, l);
111 #endif
112 	if (vp->v_type != VDIR &&
113 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
114 		return (EFBIG);
115 
116 	/*
117 	 * For nfs, cache consistency can only be maintained approximately.
118 	 * Although RFC1094 does not specify the criteria, the following is
119 	 * believed to be compatible with the reference port.
120 	 * For nqnfs, full cache consistency is maintained within the loop.
121 	 * For nfs:
122 	 * If the file's modify time on the server has changed since the
123 	 * last read rpc or you have written to the file,
124 	 * you may have lost data cache consistency with the
125 	 * server, so flush all of the file's data out of the cache.
126 	 * Then force a getattr rpc to ensure that you have up to date
127 	 * attributes.
128 	 * NB: This implies that cache data can be read when up to
129 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
130 	 * attributes this could be forced by setting n_attrstamp to 0 before
131 	 * the VOP_GETATTR() call.
132 	 */
133 
134 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
135 		error = nfs_flushstalebuf(vp, cred, l,
136 		    NFS_FLUSHSTALEBUF_MYWRITE);
137 		if (error)
138 			return error;
139 	}
140 
141 	do {
142 #ifndef NFS_V2_ONLY
143 	    /*
144 	     * Get a valid lease. If cached data is stale, flush it.
145 	     */
146 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
147 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
148 		    do {
149 			error = nqnfs_getlease(vp, ND_READ, cred, l);
150 		    } while (error == NQNFS_EXPIRED);
151 		    if (error)
152 			return (error);
153 		    if (np->n_lrev != np->n_brev ||
154 			(np->n_flag & NQNFSNONCACHE) ||
155 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
156 			if (vp->v_type == VDIR) {
157 				nfs_invaldircache(vp, 0);
158 			}
159 			error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 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, l, 1);
167 		    if (error)
168 			return (error);
169 		}
170 	    }
171 #endif
172 	    /*
173 	     * Don't cache symlinks.
174 	     */
175 	    if (np->n_flag & NQNFSNONCACHE
176 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
177 		switch (vp->v_type) {
178 		case VREG:
179 			return (nfs_readrpc(vp, uio));
180 		case VLNK:
181 			return (nfs_readlinkrpc(vp, uio, cred));
182 		case VDIR:
183 			break;
184 		default:
185 			printf(" NQNFSNONCACHE: type %x unexpected\n",
186 			    vp->v_type);
187 		};
188 	    }
189 	    baddr = (caddr_t)0;
190 	    switch (vp->v_type) {
191 	    case VREG:
192 		nfsstats.biocache_reads++;
193 
194 		advice = IO_ADV_DECODE(ioflag);
195 		error = 0;
196 		while (uio->uio_resid > 0) {
197 			void *win;
198 			int flags;
199 			vsize_t bytelen;
200 
201 			nfs_delayedtruncate(vp);
202 			if (np->n_size <= uio->uio_offset) {
203 				break;
204 			}
205 			bytelen =
206 			    MIN(np->n_size - uio->uio_offset, uio->uio_resid);
207 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
208 					&bytelen, advice, UBC_READ);
209 			error = uiomove(win, bytelen, uio);
210 			flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
211 			ubc_release(win, flags);
212 			if (error) {
213 				/*
214 				 * XXXkludge
215 				 * the file has been truncated on the server.
216 				 * there isn't much we can do.
217 				 */
218 				if (uio->uio_offset >= np->n_size) {
219 					/* end of file */
220 					error = 0;
221 				} else {
222 					break;
223 				}
224 			}
225 		}
226 		break;
227 
228 	    case VLNK:
229 		nfsstats.biocache_readlinks++;
230 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, l);
231 		if (!bp)
232 			return (EINTR);
233 		if ((bp->b_flags & B_DONE) == 0) {
234 			bp->b_flags |= B_READ;
235 			error = nfs_doio(bp);
236 			if (error) {
237 				brelse(bp);
238 				return (error);
239 			}
240 		}
241 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
242 		got_buf = 1;
243 		on = 0;
244 		break;
245 	    case VDIR:
246 diragain:
247 		nfsstats.biocache_readdirs++;
248 		ndp = nfs_searchdircache(vp, uio->uio_offset,
249 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
250 		if (!ndp) {
251 			/*
252 			 * We've been handed a cookie that is not
253 			 * in the cache. If we're not translating
254 			 * 32 <-> 64, it may be a value that was
255 			 * flushed out of the cache because it grew
256 			 * too big. Let the server judge if it's
257 			 * valid or not. In the translation case,
258 			 * we have no way of validating this value,
259 			 * so punt.
260 			 */
261 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
262 				return (EINVAL);
263 			ndp = nfs_enterdircache(vp, uio->uio_offset,
264 				uio->uio_offset, 0, 0);
265 		}
266 
267 		if (NFS_EOFVALID(np) &&
268 		    ndp->dc_cookie == np->n_direofoffset) {
269 			nfs_putdircache(np, ndp);
270 			nfsstats.direofcache_hits++;
271 			return (0);
272 		}
273 
274 		bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, l);
275 		if (!bp)
276 		    return (EINTR);
277 		if ((bp->b_flags & B_DONE) == 0) {
278 		    bp->b_flags |= B_READ;
279 		    bp->b_dcookie = ndp->dc_blkcookie;
280 		    error = nfs_doio(bp);
281 		    if (error) {
282 			/*
283 			 * Yuck! The directory has been modified on the
284 			 * server. Punt and let the userland code
285 			 * deal with it.
286 			 */
287 			nfs_putdircache(np, ndp);
288 			brelse(bp);
289 			/*
290 			 * nfs_request maps NFSERR_BAD_COOKIE to EINVAL.
291 			 */
292 			if (error == EINVAL) { /* NFSERR_BAD_COOKIE */
293 			    nfs_invaldircache(vp, 0);
294 			    nfs_vinvalbuf(vp, 0, cred, l, 1);
295 			}
296 			return (error);
297 		    }
298 		}
299 
300 		/*
301 		 * Just return if we hit EOF right away with this
302 		 * block. Always check here, because direofoffset
303 		 * may have been set by an nfsiod since the last
304 		 * check.
305 		 *
306 		 * also, empty block implies EOF.
307 		 */
308 
309 		if (bp->b_bcount == bp->b_resid ||
310 		    (NFS_EOFVALID(np) &&
311 		    ndp->dc_blkcookie == np->n_direofoffset)) {
312 			KASSERT(bp->b_bcount != bp->b_resid ||
313 			    ndp->dc_blkcookie == bp->b_dcookie);
314 			nfs_putdircache(np, ndp);
315 			bp->b_flags |= B_NOCACHE;
316 			brelse(bp);
317 			return 0;
318 		}
319 
320 		/*
321 		 * Find the entry we were looking for in the block.
322 		 */
323 
324 		en = ndp->dc_entry;
325 
326 		pdp = dp = (struct dirent *)bp->b_data;
327 		edp = (struct dirent *)(void *)(bp->b_data + bp->b_bcount -
328 		    bp->b_resid);
329 		enn = 0;
330 		while (enn < en && dp < edp) {
331 			pdp = dp;
332 			dp = _DIRENT_NEXT(dp);
333 			enn++;
334 		}
335 
336 		/*
337 		 * If the entry number was bigger than the number of
338 		 * entries in the block, or the cookie of the previous
339 		 * entry doesn't match, the directory cache is
340 		 * stale. Flush it and try again (i.e. go to
341 		 * the server).
342 		 */
343 		if (dp >= edp || (struct dirent *)_DIRENT_NEXT(dp) > edp ||
344 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
345 #ifdef DEBUG
346 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
347 				pdp, dp, edp,
348 				(unsigned long)uio->uio_offset,
349 				(unsigned long)NFS_GETCOOKIE(pdp));
350 #endif
351 			nfs_putdircache(np, ndp);
352 			brelse(bp);
353 			nfs_invaldircache(vp, 0);
354 			nfs_vinvalbuf(vp, 0, cred, l, 0);
355 			goto diragain;
356 		}
357 
358 		on = (caddr_t)dp - bp->b_data;
359 
360 		/*
361 		 * Cache all entries that may be exported to the
362 		 * user, as they may be thrown back at us. The
363 		 * NFSBIO_CACHECOOKIES flag indicates that all
364 		 * entries are being 'exported', so cache them all.
365 		 */
366 
367 		if (en == 0 && pdp == dp) {
368 			dp = _DIRENT_NEXT(dp);
369 			enn++;
370 		}
371 
372 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
373 			n = uio->uio_resid;
374 			enough = 1;
375 		} else
376 			n = bp->b_bcount - bp->b_resid - on;
377 
378 		ep = (struct dirent *)(void *)(bp->b_data + on + n);
379 
380 		/*
381 		 * Find last complete entry to copy, caching entries
382 		 * (if requested) as we go.
383 		 */
384 
385 		while (dp < ep && (struct dirent *)_DIRENT_NEXT(dp) <= ep) {
386 			if (cflag & NFSBIO_CACHECOOKIES) {
387 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
388 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
389 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
390 					NFS_STASHCOOKIE32(pdp,
391 					    nndp->dc_cookie32);
392 				}
393 				nfs_putdircache(np, nndp);
394 			}
395 			pdp = dp;
396 			dp = _DIRENT_NEXT(dp);
397 			enn++;
398 		}
399 		nfs_putdircache(np, ndp);
400 
401 		/*
402 		 * If the last requested entry was not the last in the
403 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
404 		 * cache the cookie of the last requested one, and
405 		 * set of the offset to it.
406 		 */
407 
408 		if ((on + n) < bp->b_bcount - bp->b_resid) {
409 			curoff = NFS_GETCOOKIE(pdp);
410 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
411 			    enn, bp->b_lblkno);
412 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
413 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
414 				curoff = nndp->dc_cookie32;
415 			}
416 			nfs_putdircache(np, nndp);
417 		} else
418 			curoff = bp->b_dcookie;
419 
420 		/*
421 		 * Always cache the entry for the next block,
422 		 * so that readaheads can use it.
423 		 */
424 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
425 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
426 			if (curoff == bp->b_dcookie) {
427 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
428 				curoff = nndp->dc_cookie32;
429 			}
430 		}
431 
432 		n = (char *)_DIRENT_NEXT(pdp) - (bp->b_data + on);
433 
434 		/*
435 		 * If not eof and read aheads are enabled, start one.
436 		 * (You need the current block first, so that you have the
437 		 *  directory offset cookie of the next block.)
438 		 */
439 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
440 		    !NFS_EOFVALID(np) && !(np->n_flag & NQNFSNONCACHE)) {
441 			rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp),
442 						NFS_DIRBLKSIZ, l);
443 			if (rabp) {
444 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
445 				rabp->b_dcookie = nndp->dc_cookie;
446 				rabp->b_flags |= (B_READ | B_ASYNC);
447 				if (nfs_asyncio(rabp)) {
448 				    rabp->b_flags |= B_INVAL;
449 				    brelse(rabp);
450 				}
451 			    } else
452 				brelse(rabp);
453 			}
454 		}
455 		nfs_putdircache(np, nndp);
456 		got_buf = 1;
457 		break;
458 	    default:
459 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
460 		break;
461 	    }
462 
463 	    if (n > 0) {
464 		if (!baddr)
465 			baddr = bp->b_data;
466 		error = uiomove(baddr + on, (int)n, uio);
467 	    }
468 	    switch (vp->v_type) {
469 	    case VREG:
470 		break;
471 	    case VLNK:
472 		n = 0;
473 		break;
474 	    case VDIR:
475 		if (np->n_flag & NQNFSNONCACHE)
476 			bp->b_flags |= B_INVAL;
477 		uio->uio_offset = curoff;
478 		if (enough)
479 			n = 0;
480 		break;
481 	    default:
482 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
483 	    }
484 	    if (got_buf)
485 		brelse(bp);
486 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
487 	return (error);
488 }
489 
490 /*
491  * Vnode op for write using bio
492  */
493 int
494 nfs_write(v)
495 	void *v;
496 {
497 	struct vop_write_args /* {
498 		struct vnode *a_vp;
499 		struct uio *a_uio;
500 		int  a_ioflag;
501 		kauth_cred_t a_cred;
502 	} */ *ap = v;
503 	struct uio *uio = ap->a_uio;
504 	struct lwp *l = curlwp;
505 	struct vnode *vp = ap->a_vp;
506 	struct nfsnode *np = VTONFS(vp);
507 	kauth_cred_t cred = ap->a_cred;
508 	struct vattr vattr;
509 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
510 	void *win;
511 	voff_t oldoff, origoff;
512 	vsize_t bytelen;
513 	int flags, error = 0;
514 	int ioflag = ap->a_ioflag;
515 	int extended = 0, wrotedata = 0;
516 
517 #ifdef DIAGNOSTIC
518 	if (uio->uio_rw != UIO_WRITE)
519 		panic("nfs_write mode");
520 #endif
521 	if (vp->v_type != VREG)
522 		return (EIO);
523 	if (np->n_flag & NWRITEERR) {
524 		np->n_flag &= ~NWRITEERR;
525 		return (np->n_error);
526 	}
527 #ifndef NFS_V2_ONLY
528 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
529 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
530 		(void)nfs_fsinfo(nmp, vp, cred, l);
531 #endif
532 	if (ioflag & (IO_APPEND | IO_SYNC)) {
533 		if (np->n_flag & NMODIFIED) {
534 			NFS_INVALIDATE_ATTRCACHE(np);
535 			error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1);
536 			if (error)
537 				return (error);
538 		}
539 		if (ioflag & IO_APPEND) {
540 			NFS_INVALIDATE_ATTRCACHE(np);
541 			error = VOP_GETATTR(vp, &vattr, cred, l);
542 			if (error)
543 				return (error);
544 			uio->uio_offset = np->n_size;
545 		}
546 	}
547 	if (uio->uio_offset < 0)
548 		return (EINVAL);
549 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
550 		return (EFBIG);
551 	if (uio->uio_resid == 0)
552 		return (0);
553 	/*
554 	 * Maybe this should be above the vnode op call, but so long as
555 	 * file servers have no limits, i don't think it matters
556 	 */
557 	if (l && l->l_proc && uio->uio_offset + uio->uio_resid >
558 	      l->l_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
559 		psignal(l->l_proc, SIGXFSZ);
560 		return (EFBIG);
561 	}
562 
563 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
564 		int iomode = NFSV3WRITE_FILESYNC;
565 		boolean_t stalewriteverf = FALSE;
566 
567 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
568 		error = nfs_writerpc(vp, uio, &iomode, FALSE, &stalewriteverf);
569 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
570 		if (stalewriteverf)
571 			nfs_clearcommit(vp->v_mount);
572 		return (error);
573 	}
574 
575 	origoff = uio->uio_offset;
576 	do {
577 		boolean_t extending; /* if we are extending whole pages */
578 		u_quad_t oldsize;
579 		oldoff = uio->uio_offset;
580 		bytelen = uio->uio_resid;
581 
582 #ifndef NFS_V2_ONLY
583 		/*
584 		 * Check for a valid write lease.
585 		 */
586 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
587 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
588 			do {
589 				error = nqnfs_getlease(vp, ND_WRITE, cred, l);
590 			} while (error == NQNFS_EXPIRED);
591 			if (error)
592 				return (error);
593 			if (np->n_lrev != np->n_brev ||
594 			    (np->n_flag & NQNFSNONCACHE)) {
595 				error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1);
596 				if (error)
597 					return (error);
598 				np->n_brev = np->n_lrev;
599 			}
600 		}
601 #endif
602 		nfsstats.biocache_writes++;
603 
604 		oldsize = np->n_size;
605 		np->n_flag |= NMODIFIED;
606 		if (np->n_size < uio->uio_offset + bytelen) {
607 			np->n_size = uio->uio_offset + bytelen;
608 		}
609 		extending = ((uio->uio_offset & PAGE_MASK) == 0 &&
610 		    (bytelen & PAGE_MASK) == 0 &&
611 		    uio->uio_offset >= vp->v_size);
612 		win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
613 		    UVM_ADV_NORMAL,
614 		    UBC_WRITE | (extending ? UBC_FAULTBUSY : 0));
615 		error = uiomove(win, bytelen, uio);
616 		flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
617 		ubc_release(win, flags);
618 		if (error) {
619 			if (extending) {
620 				/*
621 				 * backout size and free pages past eof.
622 				 */
623 				np->n_size = oldsize;
624 				simple_lock(&vp->v_interlock);
625 				(void)VOP_PUTPAGES(vp, round_page(vp->v_size),
626 				    0, PGO_SYNCIO | PGO_FREE);
627 			}
628 			break;
629 		}
630 		wrotedata = 1;
631 
632 		/*
633 		 * update UVM's notion of the size now that we've
634 		 * copied the data into the vnode's pages.
635 		 */
636 
637 		if (vp->v_size < uio->uio_offset) {
638 			uvm_vnp_setsize(vp, uio->uio_offset);
639 			extended = 1;
640 		}
641 
642 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
643 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
644 			simple_lock(&vp->v_interlock);
645 			error = VOP_PUTPAGES(vp,
646 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
647 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
648 				       ~(nmp->nm_wsize - 1)), PGO_CLEANIT);
649 		}
650 	} while (uio->uio_resid > 0);
651 	if (wrotedata)
652 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0));
653 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
654 		simple_lock(&vp->v_interlock);
655 		error = VOP_PUTPAGES(vp,
656 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
657 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
658 			       ~(nmp->nm_wsize - 1)),
659 		    PGO_CLEANIT | PGO_SYNCIO);
660 	}
661 	return error;
662 }
663 
664 /*
665  * Get an nfs cache block.
666  * Allocate a new one if the block isn't currently in the cache
667  * and return the block marked busy. If the calling process is
668  * interrupted by a signal for an interruptible mount point, return
669  * NULL.
670  */
671 struct buf *
672 nfs_getcacheblk(vp, bn, size, l)
673 	struct vnode *vp;
674 	daddr_t bn;
675 	int size;
676 	struct lwp *l;
677 {
678 	struct buf *bp;
679 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
680 
681 	if (nmp->nm_flag & NFSMNT_INT) {
682 		bp = getblk(vp, bn, size, PCATCH, 0);
683 		while (bp == NULL) {
684 			if (nfs_sigintr(nmp, NULL, l))
685 				return (NULL);
686 			bp = getblk(vp, bn, size, 0, 2 * hz);
687 		}
688 	} else
689 		bp = getblk(vp, bn, size, 0, 0);
690 	return (bp);
691 }
692 
693 /*
694  * Flush and invalidate all dirty buffers. If another process is already
695  * doing the flush, just wait for completion.
696  */
697 int
698 nfs_vinvalbuf(vp, flags, cred, l, intrflg)
699 	struct vnode *vp;
700 	int flags;
701 	kauth_cred_t cred;
702 	struct lwp *l;
703 	int intrflg;
704 {
705 	struct nfsnode *np = VTONFS(vp);
706 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
707 	int error = 0, slpflag, slptimeo;
708 
709 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
710 		intrflg = 0;
711 	if (intrflg) {
712 		slpflag = PCATCH;
713 		slptimeo = 2 * hz;
714 	} else {
715 		slpflag = 0;
716 		slptimeo = 0;
717 	}
718 	/*
719 	 * First wait for any other process doing a flush to complete.
720 	 */
721 	simple_lock(&vp->v_interlock);
722 	while (np->n_flag & NFLUSHINPROG) {
723 		np->n_flag |= NFLUSHWANT;
724 		error = ltsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
725 			slptimeo, &vp->v_interlock);
726 		if (error && intrflg && nfs_sigintr(nmp, NULL, l)) {
727 			simple_unlock(&vp->v_interlock);
728 			return EINTR;
729 		}
730 	}
731 
732 	/*
733 	 * Now, flush as required.
734 	 */
735 	np->n_flag |= NFLUSHINPROG;
736 	simple_unlock(&vp->v_interlock);
737 	error = vinvalbuf(vp, flags, cred, l, slpflag, 0);
738 	while (error) {
739 		if (intrflg && nfs_sigintr(nmp, NULL, l)) {
740 			error = EINTR;
741 			break;
742 		}
743 		error = vinvalbuf(vp, flags, cred, l, 0, slptimeo);
744 	}
745 	simple_lock(&vp->v_interlock);
746 	if (error == 0)
747 		np->n_flag &= ~NMODIFIED;
748 	np->n_flag &= ~NFLUSHINPROG;
749 	if (np->n_flag & NFLUSHWANT) {
750 		np->n_flag &= ~NFLUSHWANT;
751 		wakeup(&np->n_flag);
752 	}
753 	simple_unlock(&vp->v_interlock);
754 	return error;
755 }
756 
757 /*
758  * nfs_flushstalebuf: flush cache if it's stale.
759  *
760  * => caller shouldn't own any pages or buffers which belong to the vnode.
761  */
762 
763 int
764 nfs_flushstalebuf(struct vnode *vp, kauth_cred_t cred, struct lwp *l,
765     int flags)
766 {
767 	struct nfsnode *np = VTONFS(vp);
768 	struct vattr vattr;
769 	int error;
770 
771 	if (np->n_flag & NMODIFIED) {
772 		if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0
773 		    || vp->v_type != VREG) {
774 			error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1);
775 			if (error)
776 				return error;
777 			if (vp->v_type == VDIR) {
778 				nfs_invaldircache(vp, 0);
779 			}
780 		} else {
781 			/*
782 			 * XXX assuming writes are ours.
783 			 */
784 		}
785 		NFS_INVALIDATE_ATTRCACHE(np);
786 		error = VOP_GETATTR(vp, &vattr, cred, l);
787 		if (error)
788 			return error;
789 		np->n_mtime = vattr.va_mtime;
790 	} else {
791 		error = VOP_GETATTR(vp, &vattr, cred, l);
792 		if (error)
793 			return error;
794 		if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
795 			if (vp->v_type == VDIR) {
796 				nfs_invaldircache(vp, 0);
797 			}
798 			error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1);
799 			if (error)
800 				return error;
801 			np->n_mtime = vattr.va_mtime;
802 		}
803 	}
804 
805 	return error;
806 }
807 
808 /*
809  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
810  * This is mainly to avoid queueing async I/O requests when the nfsiods
811  * are all hung on a dead server.
812  */
813 
814 int
815 nfs_asyncio(bp)
816 	struct buf *bp;
817 {
818 	int i;
819 	struct nfsmount *nmp;
820 	int gotiod, slpflag = 0, slptimeo = 0, error;
821 
822 	if (nfs_numasync == 0)
823 		return (EIO);
824 
825 	nmp = VFSTONFS(bp->b_vp->v_mount);
826 again:
827 	if (nmp->nm_flag & NFSMNT_INT)
828 		slpflag = PCATCH;
829 	gotiod = FALSE;
830 
831 	/*
832 	 * Find a free iod to process this request.
833 	 */
834 
835 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++) {
836 		struct nfs_iod *iod = &nfs_asyncdaemon[i];
837 
838 		simple_lock(&iod->nid_slock);
839 		if (iod->nid_want) {
840 			/*
841 			 * Found one, so wake it up and tell it which
842 			 * mount to process.
843 			 */
844 			iod->nid_want = NULL;
845 			iod->nid_mount = nmp;
846 			wakeup(&iod->nid_want);
847 			simple_lock(&nmp->nm_slock);
848 			simple_unlock(&iod->nid_slock);
849 			nmp->nm_bufqiods++;
850 			gotiod = TRUE;
851 			break;
852 		}
853 		simple_unlock(&iod->nid_slock);
854 	}
855 
856 	/*
857 	 * If none are free, we may already have an iod working on this mount
858 	 * point.  If so, it will process our request.
859 	 */
860 
861 	if (!gotiod) {
862 		simple_lock(&nmp->nm_slock);
863 		if (nmp->nm_bufqiods > 0)
864 			gotiod = TRUE;
865 	}
866 
867 	LOCK_ASSERT(simple_lock_held(&nmp->nm_slock));
868 
869 	/*
870 	 * If we have an iod which can process the request, then queue
871 	 * the buffer.  However, even if we have an iod, do not initiate
872 	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
873 	 * is via local loopback, we may put curproc (pagedaemon) to sleep
874 	 * waiting for the writes to complete. But the server (ourself)
875 	 * may block the write, waiting for its (ie., our) pagedaemon
876 	 * to produce clean pages to handle the write: deadlock.
877 	 * XXX: start non-loopback mounts straight away?  If "lots free",
878 	 * let pagedaemon start loopback writes anyway?
879 	 */
880 	if (gotiod) {
881 
882 		/*
883 		 * Ensure that the queue never grows too large.
884 		 */
885 		if (curproc == uvm.pagedaemon_proc) {
886 	  		/* Enque for later, to avoid free-page deadlock */
887 			  (void) 0;
888 		} else while (nmp->nm_bufqlen >= 2*nfs_numasync) {
889 			nmp->nm_bufqwant = TRUE;
890 			error = ltsleep(&nmp->nm_bufq,
891 			    slpflag | PRIBIO | PNORELOCK,
892 			    "nfsaio", slptimeo, &nmp->nm_slock);
893 			if (error) {
894 				if (nfs_sigintr(nmp, NULL, curlwp))
895 					return (EINTR);
896 				if (slpflag == PCATCH) {
897 					slpflag = 0;
898 					slptimeo = 2 * hz;
899 				}
900 			}
901 
902 			/*
903 			 * We might have lost our iod while sleeping,
904 			 * so check and loop if nescessary.
905 			 */
906 
907 			if (nmp->nm_bufqiods == 0)
908 				goto again;
909 
910 			simple_lock(&nmp->nm_slock);
911 		}
912 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
913 		nmp->nm_bufqlen++;
914 		simple_unlock(&nmp->nm_slock);
915 		return (0);
916 	}
917 	simple_unlock(&nmp->nm_slock);
918 
919 	/*
920 	 * All the iods are busy on other mounts, so return EIO to
921 	 * force the caller to process the i/o synchronously.
922 	 */
923 
924 	return (EIO);
925 }
926 
927 /*
928  * nfs_doio for read.
929  */
930 static int
931 nfs_doio_read(bp, uiop)
932 	struct buf *bp;
933 	struct uio *uiop;
934 {
935 	struct vnode *vp = bp->b_vp;
936 	struct nfsnode *np = VTONFS(vp);
937 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
938 	int error = 0;
939 
940 	uiop->uio_rw = UIO_READ;
941 	switch (vp->v_type) {
942 	case VREG:
943 		nfsstats.read_bios++;
944 		error = nfs_readrpc(vp, uiop);
945 		if (!error && uiop->uio_resid) {
946 			int diff, len;
947 
948 			/*
949 			 * If uio_resid > 0, there is a hole in the file and
950 			 * no writes after the hole have been pushed to
951 			 * the server yet or the file has been truncated
952 			 * on the server.
953 			 * Just zero fill the rest of the valid area.
954 			 */
955 
956 			KASSERT(vp->v_size >=
957 			    uiop->uio_offset + uiop->uio_resid);
958 			diff = bp->b_bcount - uiop->uio_resid;
959 			len = uiop->uio_resid;
960 			memset((char *)bp->b_data + diff, 0, len);
961 			uiop->uio_resid = 0;
962 		}
963 #if 0
964 		if (uiop->uio_lwp && (vp->v_flag & VTEXT) &&
965 		    (((nmp->nm_flag & NFSMNT_NQNFS) &&
966 		      NQNFS_CKINVALID(vp, np, ND_READ) &&
967 		      np->n_lrev != np->n_brev) ||
968 		     (!(nmp->nm_flag & NFSMNT_NQNFS) &&
969 		      timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)))) {
970 			killproc(uiop->uio_lwp->l_proc, "process text file was modified");
971 #if 0 /* XXX NJWLWP */
972 			uiop->uio_lwp->l_proc->p_holdcnt++;
973 #endif
974 		}
975 #endif
976 		break;
977 	case VLNK:
978 		KASSERT(uiop->uio_offset == (off_t)0);
979 		nfsstats.readlink_bios++;
980 		error = nfs_readlinkrpc(vp, uiop, np->n_rcred);
981 		break;
982 	case VDIR:
983 		nfsstats.readdir_bios++;
984 		uiop->uio_offset = bp->b_dcookie;
985 #ifndef NFS_V2_ONLY
986 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
987 			error = nfs_readdirplusrpc(vp, uiop,
988 			    curlwp->l_proc->p_cred);
989 			/*
990 			 * nfs_request maps NFSERR_NOTSUPP to ENOTSUP.
991 			 */
992 			if (error == ENOTSUP)
993 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
994 		}
995 #else
996 		nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
997 #endif
998 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
999 			error = nfs_readdirrpc(vp, uiop,
1000 			    curlwp->l_proc->p_cred);
1001 		if (!error) {
1002 			bp->b_dcookie = uiop->uio_offset;
1003 		}
1004 		break;
1005 	default:
1006 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
1007 		break;
1008 	}
1009 	if (error) {
1010 		bp->b_flags |= B_ERROR;
1011 		bp->b_error = error;
1012 	}
1013 	return error;
1014 }
1015 
1016 /*
1017  * nfs_doio for write.
1018  */
1019 static int
1020 nfs_doio_write(bp, uiop)
1021 	struct buf *bp;
1022 	struct uio *uiop;
1023 {
1024 	struct vnode *vp = bp->b_vp;
1025 	struct nfsnode *np = VTONFS(vp);
1026 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1027 	int iomode;
1028 	boolean_t stalewriteverf = FALSE;
1029 	int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT;
1030 	struct vm_page *pgs[npages];
1031 #ifndef NFS_V2_ONLY
1032 	boolean_t needcommit = TRUE; /* need only COMMIT RPC */
1033 #else
1034 	boolean_t needcommit = FALSE; /* need only COMMIT RPC */
1035 #endif
1036 	boolean_t pageprotected;
1037 	struct uvm_object *uobj = &vp->v_uobj;
1038 	int error;
1039 	off_t off, cnt;
1040 
1041 	if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) {
1042 		iomode = NFSV3WRITE_UNSTABLE;
1043 	} else {
1044 		iomode = NFSV3WRITE_FILESYNC;
1045 	}
1046 
1047 #ifndef NFS_V2_ONLY
1048 again:
1049 #endif
1050 	lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1051 
1052 	for (i = 0; i < npages; i++) {
1053 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
1054 		if (pgs[i]->uobject == uobj &&
1055 		    pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) {
1056 			KASSERT(pgs[i]->flags & PG_BUSY);
1057 			/*
1058 			 * this page belongs to our object.
1059 			 */
1060 			simple_lock(&uobj->vmobjlock);
1061 			/*
1062 			 * write out the page stably if it's about to
1063 			 * be released because we can't resend it
1064 			 * on the server crash.
1065 			 *
1066 			 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be
1067 			 * changed until unbusy the page.
1068 			 */
1069 			if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT))
1070 				iomode = NFSV3WRITE_FILESYNC;
1071 			/*
1072 			 * if we met a page which hasn't been sent yet,
1073 			 * we need do WRITE RPC.
1074 			 */
1075 			if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0)
1076 				needcommit = FALSE;
1077 			simple_unlock(&uobj->vmobjlock);
1078 		} else {
1079 			iomode = NFSV3WRITE_FILESYNC;
1080 			needcommit = FALSE;
1081 		}
1082 	}
1083 	if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) {
1084 		simple_lock(&uobj->vmobjlock);
1085 		for (i = 0; i < npages; i++) {
1086 			pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY;
1087 			pmap_page_protect(pgs[i], VM_PROT_READ);
1088 		}
1089 		simple_unlock(&uobj->vmobjlock);
1090 		pageprotected = TRUE; /* pages can't be modified during i/o. */
1091 	} else
1092 		pageprotected = FALSE;
1093 
1094 	/*
1095 	 * Send the data to the server if necessary,
1096 	 * otherwise just send a commit rpc.
1097 	 */
1098 #ifndef NFS_V2_ONLY
1099 	if (needcommit) {
1100 
1101 		/*
1102 		 * If the buffer is in the range that we already committed,
1103 		 * there's nothing to do.
1104 		 *
1105 		 * If it's in the range that we need to commit, push the
1106 		 * whole range at once, otherwise only push the buffer.
1107 		 * In both these cases, acquire the commit lock to avoid
1108 		 * other processes modifying the range.
1109 		 */
1110 
1111 		off = uiop->uio_offset;
1112 		cnt = bp->b_bcount;
1113 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1114 		if (!nfs_in_committed_range(vp, off, bp->b_bcount)) {
1115 			boolean_t pushedrange;
1116 			if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) {
1117 				pushedrange = TRUE;
1118 				off = np->n_pushlo;
1119 				cnt = np->n_pushhi - np->n_pushlo;
1120 			} else {
1121 				pushedrange = FALSE;
1122 			}
1123 			error = nfs_commit(vp, off, cnt, curlwp);
1124 			if (error == 0) {
1125 				if (pushedrange) {
1126 					nfs_merge_commit_ranges(vp);
1127 				} else {
1128 					nfs_add_committed_range(vp, off, cnt);
1129 				}
1130 			}
1131 		} else {
1132 			error = 0;
1133 		}
1134 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1135 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1136 		if (!error) {
1137 			/*
1138 			 * pages are now on stable storage.
1139 			 */
1140 			uiop->uio_resid = 0;
1141 			simple_lock(&uobj->vmobjlock);
1142 			for (i = 0; i < npages; i++) {
1143 				pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1144 			}
1145 			simple_unlock(&uobj->vmobjlock);
1146 			return 0;
1147 		} else if (error == NFSERR_STALEWRITEVERF) {
1148 			nfs_clearcommit(vp->v_mount);
1149 			goto again;
1150 		}
1151 		if (error) {
1152 			bp->b_flags |= B_ERROR;
1153 			bp->b_error = np->n_error = error;
1154 			np->n_flag |= NWRITEERR;
1155 		}
1156 		return error;
1157 	}
1158 #endif
1159 	off = uiop->uio_offset;
1160 	cnt = bp->b_bcount;
1161 	uiop->uio_rw = UIO_WRITE;
1162 	nfsstats.write_bios++;
1163 	error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf);
1164 #ifndef NFS_V2_ONLY
1165 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1166 		/*
1167 		 * we need to commit pages later.
1168 		 */
1169 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1170 		nfs_add_tobecommitted_range(vp, off, cnt);
1171 		/*
1172 		 * if there can be too many uncommitted pages, commit them now.
1173 		 */
1174 		if (np->n_pushhi - np->n_pushlo > nfs_commitsize) {
1175 			off = np->n_pushlo;
1176 			cnt = nfs_commitsize >> 1;
1177 			error = nfs_commit(vp, off, cnt, curlwp);
1178 			if (!error) {
1179 				nfs_add_committed_range(vp, off, cnt);
1180 				nfs_del_tobecommitted_range(vp, off, cnt);
1181 			}
1182 			if (error == NFSERR_STALEWRITEVERF) {
1183 				stalewriteverf = TRUE;
1184 				error = 0; /* it isn't a real error */
1185 			}
1186 		} else {
1187 			/*
1188 			 * re-dirty pages so that they will be passed
1189 			 * to us later again.
1190 			 */
1191 			simple_lock(&uobj->vmobjlock);
1192 			for (i = 0; i < npages; i++) {
1193 				pgs[i]->flags &= ~PG_CLEAN;
1194 			}
1195 			simple_unlock(&uobj->vmobjlock);
1196 		}
1197 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1198 	} else
1199 #endif
1200 	if (!error) {
1201 		/*
1202 		 * pages are now on stable storage.
1203 		 */
1204 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1205 		nfs_del_committed_range(vp, off, cnt);
1206 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1207 		simple_lock(&uobj->vmobjlock);
1208 		for (i = 0; i < npages; i++) {
1209 			pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1210 		}
1211 		simple_unlock(&uobj->vmobjlock);
1212 	} else {
1213 		/*
1214 		 * we got an error.
1215 		 */
1216 		bp->b_flags |= B_ERROR;
1217 		bp->b_error = np->n_error = error;
1218 		np->n_flag |= NWRITEERR;
1219 	}
1220 
1221 	lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1222 
1223 	if (stalewriteverf) {
1224 		nfs_clearcommit(vp->v_mount);
1225 	}
1226 	return error;
1227 }
1228 
1229 /*
1230  * nfs_doio for B_PHYS.
1231  */
1232 static int
1233 nfs_doio_phys(bp, uiop)
1234 	struct buf *bp;
1235 	struct uio *uiop;
1236 {
1237 	struct vnode *vp = bp->b_vp;
1238 	int error;
1239 
1240 	uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
1241 	if (bp->b_flags & B_READ) {
1242 		uiop->uio_rw = UIO_READ;
1243 		nfsstats.read_physios++;
1244 		error = nfs_readrpc(vp, uiop);
1245 	} else {
1246 		int iomode = NFSV3WRITE_DATASYNC;
1247 		boolean_t stalewriteverf;
1248 		struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1249 
1250 		uiop->uio_rw = UIO_WRITE;
1251 		nfsstats.write_physios++;
1252 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1253 		error = nfs_writerpc(vp, uiop, &iomode, FALSE, &stalewriteverf);
1254 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1255 		if (stalewriteverf) {
1256 			nfs_clearcommit(bp->b_vp->v_mount);
1257 		}
1258 	}
1259 	if (error) {
1260 		bp->b_flags |= B_ERROR;
1261 		bp->b_error = error;
1262 	}
1263 	return error;
1264 }
1265 
1266 /*
1267  * Do an I/O operation to/from a cache block. This may be called
1268  * synchronously or from an nfsiod.
1269  */
1270 int
1271 nfs_doio(bp)
1272 	struct buf *bp;
1273 {
1274 	int error;
1275 	struct uio uio;
1276 	struct uio *uiop = &uio;
1277 	struct iovec io;
1278 	UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist);
1279 
1280 	uiop->uio_iov = &io;
1281 	uiop->uio_iovcnt = 1;
1282 	uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
1283 	UIO_SETUP_SYSSPACE(uiop);
1284 	io.iov_base = bp->b_data;
1285 	io.iov_len = uiop->uio_resid = bp->b_bcount;
1286 
1287 	/*
1288 	 * Historically, paging was done with physio, but no more...
1289 	 */
1290 	if (bp->b_flags & B_PHYS) {
1291 		/*
1292 		 * ...though reading /dev/drum still gets us here.
1293 		 */
1294 		error = nfs_doio_phys(bp, uiop);
1295 	} else if (bp->b_flags & B_READ) {
1296 		error = nfs_doio_read(bp, uiop);
1297 	} else {
1298 		error = nfs_doio_write(bp, uiop);
1299 	}
1300 	bp->b_resid = uiop->uio_resid;
1301 	biodone(bp);
1302 	return (error);
1303 }
1304 
1305 /*
1306  * Vnode op for VM getpages.
1307  */
1308 
1309 int
1310 nfs_getpages(v)
1311 	void *v;
1312 {
1313 	struct vop_getpages_args /* {
1314 		struct vnode *a_vp;
1315 		voff_t a_offset;
1316 		struct vm_page **a_m;
1317 		int *a_count;
1318 		int a_centeridx;
1319 		vm_prot_t a_access_type;
1320 		int a_advice;
1321 		int a_flags;
1322 	} */ *ap = v;
1323 
1324 	struct vnode *vp = ap->a_vp;
1325 	struct uvm_object *uobj = &vp->v_uobj;
1326 	struct nfsnode *np = VTONFS(vp);
1327 	const int npages = *ap->a_count;
1328 	struct vm_page *pg, **pgs, *opgs[npages];
1329 	off_t origoffset, len;
1330 	int i, error;
1331 	boolean_t v3 = NFS_ISV3(vp);
1332 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1333 	boolean_t locked = (ap->a_flags & PGO_LOCKED) != 0;
1334 
1335 	/*
1336 	 * call the genfs code to get the pages.  `pgs' may be NULL
1337 	 * when doing read-ahead.
1338 	 */
1339 
1340 	pgs = ap->a_m;
1341 	if (write && locked && v3) {
1342 		KASSERT(pgs != NULL);
1343 #ifdef DEBUG
1344 
1345 		/*
1346 		 * If PGO_LOCKED is set, real pages shouldn't exists
1347 		 * in the array.
1348 		 */
1349 
1350 		for (i = 0; i < npages; i++)
1351 			KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE);
1352 #endif
1353 		memcpy(opgs, pgs, npages * sizeof(struct vm_pages *));
1354 	}
1355 	error = genfs_getpages(v);
1356 	if (error) {
1357 		return (error);
1358 	}
1359 
1360 	/*
1361 	 * for read faults where the nfs node is not yet marked NMODIFIED,
1362 	 * set PG_RDONLY on the pages so that we come back here if someone
1363 	 * tries to modify later via the mapping that will be entered for
1364 	 * this fault.
1365 	 */
1366 
1367 	if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) {
1368 		if (!locked) {
1369 			simple_lock(&uobj->vmobjlock);
1370 		}
1371 		for (i = 0; i < npages; i++) {
1372 			pg = pgs[i];
1373 			if (pg == NULL || pg == PGO_DONTCARE) {
1374 				continue;
1375 			}
1376 			pg->flags |= PG_RDONLY;
1377 		}
1378 		if (!locked) {
1379 			simple_unlock(&uobj->vmobjlock);
1380 		}
1381 	}
1382 	if (!write) {
1383 		return (0);
1384 	}
1385 
1386 	/*
1387 	 * this is a write fault, update the commit info.
1388 	 */
1389 
1390 	origoffset = ap->a_offset;
1391 	len = npages << PAGE_SHIFT;
1392 
1393 	if (v3) {
1394 		error = lockmgr(&np->n_commitlock,
1395 		    LK_EXCLUSIVE | (locked ? LK_NOWAIT : 0), NULL);
1396 		if (error) {
1397 			KASSERT(locked != 0);
1398 
1399 			/*
1400 			 * Since PGO_LOCKED is set, we need to unbusy
1401 			 * all pages fetched by genfs_getpages() above,
1402 			 * tell the caller that there are no pages
1403 			 * available and put back original pgs array.
1404 			 */
1405 
1406 			uvm_lock_pageq();
1407 			uvm_page_unbusy(pgs, npages);
1408 			uvm_unlock_pageq();
1409 			*ap->a_count = 0;
1410 			memcpy(pgs, opgs,
1411 			    npages * sizeof(struct vm_pages *));
1412 			return (error);
1413 		}
1414 		nfs_del_committed_range(vp, origoffset, len);
1415 		nfs_del_tobecommitted_range(vp, origoffset, len);
1416 	}
1417 	np->n_flag |= NMODIFIED;
1418 	if (!locked) {
1419 		simple_lock(&uobj->vmobjlock);
1420 	}
1421 	for (i = 0; i < npages; i++) {
1422 		pg = pgs[i];
1423 		if (pg == NULL || pg == PGO_DONTCARE) {
1424 			continue;
1425 		}
1426 		pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1427 	}
1428 	if (!locked) {
1429 		simple_unlock(&uobj->vmobjlock);
1430 	}
1431 	if (v3) {
1432 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1433 	}
1434 	return (0);
1435 }
1436