xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: nfs_bio.c,v 1.122 2004/10/26 04:34:46 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 1.122 2004/10/26 04:34:46 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 vattr vattr;
89 	struct proc *p;
90 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
91 	struct nfsdircache *ndp = NULL, *nndp = NULL;
92 	caddr_t baddr, ep, edp;
93 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
94 	int enough = 0;
95 	struct dirent *dp, *pdp;
96 	off_t curoff = 0;
97 
98 #ifdef DIAGNOSTIC
99 	if (uio->uio_rw != UIO_READ)
100 		panic("nfs_read mode");
101 #endif
102 	if (uio->uio_resid == 0)
103 		return (0);
104 	if (vp->v_type != VDIR && uio->uio_offset < 0)
105 		return (EINVAL);
106 	p = uio->uio_procp;
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, p);
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 		if (np->n_flag & NMODIFIED) {
136 			if (vp->v_type != VREG) {
137 				if (vp->v_type != VDIR)
138 					panic("nfs: bioread, not dir");
139 				nfs_invaldircache(vp, 0);
140 				np->n_direofoffset = 0;
141 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
142 				if (error)
143 					return (error);
144 			}
145 			NFS_INVALIDATE_ATTRCACHE(np);
146 			error = VOP_GETATTR(vp, &vattr, cred, p);
147 			if (error)
148 				return (error);
149 			np->n_mtime = vattr.va_mtime;
150 		} else {
151 			error = VOP_GETATTR(vp, &vattr, cred, p);
152 			if (error)
153 				return (error);
154 			if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
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_mtime = vattr.va_mtime;
163 			}
164 		}
165 	}
166 
167 	do {
168 #ifndef NFS_V2_ONLY
169 	    /*
170 	     * Get a valid lease. If cached data is stale, flush it.
171 	     */
172 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
173 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
174 		    do {
175 			error = nqnfs_getlease(vp, ND_READ, cred, p);
176 		    } while (error == NQNFS_EXPIRED);
177 		    if (error)
178 			return (error);
179 		    if (np->n_lrev != np->n_brev ||
180 			(np->n_flag & NQNFSNONCACHE) ||
181 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
182 			if (vp->v_type == VDIR) {
183 				nfs_invaldircache(vp, 0);
184 				np->n_direofoffset = 0;
185 			}
186 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
187 			if (error)
188 			    return (error);
189 			np->n_brev = np->n_lrev;
190 		    }
191 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
192 		    nfs_invaldircache(vp, 0);
193 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
194 		    np->n_direofoffset = 0;
195 		    if (error)
196 			return (error);
197 		}
198 	    }
199 #endif
200 	    /*
201 	     * Don't cache symlinks.
202 	     */
203 	    if (np->n_flag & NQNFSNONCACHE
204 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
205 		switch (vp->v_type) {
206 		case VREG:
207 			return (nfs_readrpc(vp, uio));
208 		case VLNK:
209 			return (nfs_readlinkrpc(vp, uio, cred));
210 		case VDIR:
211 			break;
212 		default:
213 			printf(" NQNFSNONCACHE: type %x unexpected\n",
214 			    vp->v_type);
215 		};
216 	    }
217 	    baddr = (caddr_t)0;
218 	    switch (vp->v_type) {
219 	    case VREG:
220 		nfsstats.biocache_reads++;
221 
222 		error = 0;
223 		if (uio->uio_offset >= np->n_size) {
224 			break;
225 		}
226 		while (uio->uio_resid > 0) {
227 			void *win;
228 			vsize_t bytelen = MIN(np->n_size - uio->uio_offset,
229 					      uio->uio_resid);
230 
231 			if (bytelen == 0)
232 				break;
233 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
234 					&bytelen, UBC_READ);
235 			error = uiomove(win, bytelen, uio);
236 			ubc_release(win, 0);
237 			if (error) {
238 				break;
239 			}
240 		}
241 		n = 0;
242 		break;
243 
244 	    case VLNK:
245 		nfsstats.biocache_readlinks++;
246 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
247 		if (!bp)
248 			return (EINTR);
249 		if ((bp->b_flags & B_DONE) == 0) {
250 			bp->b_flags |= B_READ;
251 			error = nfs_doio(bp, p);
252 			if (error) {
253 				brelse(bp);
254 				return (error);
255 			}
256 		}
257 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
258 		got_buf = 1;
259 		on = 0;
260 		break;
261 	    case VDIR:
262 diragain:
263 		nfsstats.biocache_readdirs++;
264 		ndp = nfs_searchdircache(vp, uio->uio_offset,
265 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
266 		if (!ndp) {
267 			/*
268 			 * We've been handed a cookie that is not
269 			 * in the cache. If we're not translating
270 			 * 32 <-> 64, it may be a value that was
271 			 * flushed out of the cache because it grew
272 			 * too big. Let the server judge if it's
273 			 * valid or not. In the translation case,
274 			 * we have no way of validating this value,
275 			 * so punt.
276 			 */
277 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
278 				return (EINVAL);
279 			ndp = nfs_enterdircache(vp, uio->uio_offset,
280 				uio->uio_offset, 0, 0);
281 		}
282 
283 		if (uio->uio_offset != 0 &&
284 		    ndp->dc_cookie == np->n_direofoffset) {
285 			nfs_putdircache(np, ndp);
286 			nfsstats.direofcache_hits++;
287 			return (0);
288 		}
289 
290 		bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, p);
291 		if (!bp)
292 		    return (EINTR);
293 		if ((bp->b_flags & B_DONE) == 0) {
294 		    bp->b_flags |= B_READ;
295 		    bp->b_dcookie = ndp->dc_blkcookie;
296 		    error = nfs_doio(bp, p);
297 		    if (error) {
298 			/*
299 			 * Yuck! The directory has been modified on the
300 			 * server. Punt and let the userland code
301 			 * deal with it.
302 			 */
303 			nfs_putdircache(np, ndp);
304 			brelse(bp);
305 			if (error == NFSERR_BAD_COOKIE) {
306 			    nfs_invaldircache(vp, 0);
307 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
308 			    error = EINVAL;
309 			}
310 			return (error);
311 		    }
312 		}
313 
314 		/*
315 		 * Just return if we hit EOF right away with this
316 		 * block. Always check here, because direofoffset
317 		 * may have been set by an nfsiod since the last
318 		 * check.
319 		 */
320 		if (np->n_direofoffset != 0 &&
321 			ndp->dc_blkcookie == np->n_direofoffset) {
322 			nfs_putdircache(np, ndp);
323 			brelse(bp);
324 			return (0);
325 		}
326 
327 		/*
328 		 * Find the entry we were looking for in the block.
329 		 */
330 
331 		en = ndp->dc_entry;
332 
333 		pdp = dp = (struct dirent *)bp->b_data;
334 		edp = bp->b_data + bp->b_bcount - bp->b_resid;
335 		enn = 0;
336 		while (enn < en && (caddr_t)dp < edp) {
337 			pdp = dp;
338 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
339 			enn++;
340 		}
341 
342 		/*
343 		 * If the entry number was bigger than the number of
344 		 * entries in the block, or the cookie of the previous
345 		 * entry doesn't match, the directory cache is
346 		 * stale. Flush it and try again (i.e. go to
347 		 * the server).
348 		 */
349 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
350 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
351 #ifdef DEBUG
352 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
353 				pdp, dp, edp,
354 				(unsigned long)uio->uio_offset,
355 				(unsigned long)NFS_GETCOOKIE(pdp));
356 #endif
357 			nfs_putdircache(np, ndp);
358 			brelse(bp);
359 			nfs_invaldircache(vp, 0);
360 			nfs_vinvalbuf(vp, 0, cred, p, 0);
361 			goto diragain;
362 		}
363 
364 		on = (caddr_t)dp - bp->b_data;
365 
366 		/*
367 		 * Cache all entries that may be exported to the
368 		 * user, as they may be thrown back at us. The
369 		 * NFSBIO_CACHECOOKIES flag indicates that all
370 		 * entries are being 'exported', so cache them all.
371 		 */
372 
373 		if (en == 0 && pdp == dp) {
374 			dp = (struct dirent *)
375 			    ((caddr_t)dp + dp->d_reclen);
376 			enn++;
377 		}
378 
379 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
380 			n = uio->uio_resid;
381 			enough = 1;
382 		} else
383 			n = bp->b_bcount - bp->b_resid - on;
384 
385 		ep = bp->b_data + on + n;
386 
387 		/*
388 		 * Find last complete entry to copy, caching entries
389 		 * (if requested) as we go.
390 		 */
391 
392 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
393 			if (cflag & NFSBIO_CACHECOOKIES) {
394 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
395 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
396 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
397 					NFS_STASHCOOKIE32(pdp,
398 					    nndp->dc_cookie32);
399 				}
400 				nfs_putdircache(np, nndp);
401 			}
402 			pdp = dp;
403 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
404 			enn++;
405 		}
406 		nfs_putdircache(np, ndp);
407 
408 		/*
409 		 * If the last requested entry was not the last in the
410 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
411 		 * cache the cookie of the last requested one, and
412 		 * set of the offset to it.
413 		 */
414 
415 		if ((on + n) < bp->b_bcount - bp->b_resid) {
416 			curoff = NFS_GETCOOKIE(pdp);
417 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
418 			    enn, bp->b_lblkno);
419 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
420 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
421 				curoff = nndp->dc_cookie32;
422 			}
423 			nfs_putdircache(np, nndp);
424 		} else
425 			curoff = bp->b_dcookie;
426 
427 		/*
428 		 * Always cache the entry for the next block,
429 		 * so that readaheads can use it.
430 		 */
431 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
432 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
433 			if (curoff == bp->b_dcookie) {
434 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
435 				curoff = nndp->dc_cookie32;
436 			}
437 		}
438 
439 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
440 
441 		/*
442 		 * If not eof and read aheads are enabled, start one.
443 		 * (You need the current block first, so that you have the
444 		 *  directory offset cookie of the next block.)
445 		 */
446 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
447 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
448 			rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp),
449 						NFS_DIRBLKSIZ, p);
450 			if (rabp) {
451 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
452 				rabp->b_dcookie = nndp->dc_cookie;
453 				rabp->b_flags |= (B_READ | B_ASYNC);
454 				if (nfs_asyncio(rabp)) {
455 				    rabp->b_flags |= B_INVAL;
456 				    brelse(rabp);
457 				}
458 			    } else
459 				brelse(rabp);
460 			}
461 		}
462 		nfs_putdircache(np, nndp);
463 		got_buf = 1;
464 		break;
465 	    default:
466 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
467 		break;
468 	    }
469 
470 	    if (n > 0) {
471 		if (!baddr)
472 			baddr = bp->b_data;
473 		error = uiomove(baddr + on, (int)n, uio);
474 	    }
475 	    switch (vp->v_type) {
476 	    case VREG:
477 		break;
478 	    case VLNK:
479 		n = 0;
480 		break;
481 	    case VDIR:
482 		if (np->n_flag & NQNFSNONCACHE)
483 			bp->b_flags |= B_INVAL;
484 		uio->uio_offset = curoff;
485 		if (enough)
486 			n = 0;
487 		break;
488 	    default:
489 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
490 	    }
491 	    if (got_buf)
492 		brelse(bp);
493 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
494 	return (error);
495 }
496 
497 /*
498  * Vnode op for write using bio
499  */
500 int
501 nfs_write(v)
502 	void *v;
503 {
504 	struct vop_write_args /* {
505 		struct vnode *a_vp;
506 		struct uio *a_uio;
507 		int  a_ioflag;
508 		struct ucred *a_cred;
509 	} */ *ap = v;
510 	struct uio *uio = ap->a_uio;
511 	struct proc *p = uio->uio_procp;
512 	struct vnode *vp = ap->a_vp;
513 	struct nfsnode *np = VTONFS(vp);
514 	struct ucred *cred = ap->a_cred;
515 	int ioflag = ap->a_ioflag;
516 	struct vattr vattr;
517 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
518 	void *win;
519 	voff_t oldoff, origoff;
520 	vsize_t bytelen;
521 	int error = 0;
522 	int extended = 0, wrotedta = 0;
523 
524 #ifdef DIAGNOSTIC
525 	if (uio->uio_rw != UIO_WRITE)
526 		panic("nfs_write mode");
527 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
528 		panic("nfs_write proc");
529 #endif
530 	if (vp->v_type != VREG)
531 		return (EIO);
532 	if (np->n_flag & NWRITEERR) {
533 		np->n_flag &= ~NWRITEERR;
534 		return (np->n_error);
535 	}
536 #ifndef NFS_V2_ONLY
537 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
538 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
539 		(void)nfs_fsinfo(nmp, vp, cred, p);
540 #endif
541 	if (ioflag & (IO_APPEND | IO_SYNC)) {
542 		if (np->n_flag & NMODIFIED) {
543 			NFS_INVALIDATE_ATTRCACHE(np);
544 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
545 			if (error)
546 				return (error);
547 		}
548 		if (ioflag & IO_APPEND) {
549 			NFS_INVALIDATE_ATTRCACHE(np);
550 			error = VOP_GETATTR(vp, &vattr, cred, p);
551 			if (error)
552 				return (error);
553 			uio->uio_offset = np->n_size;
554 		}
555 	}
556 	if (uio->uio_offset < 0)
557 		return (EINVAL);
558 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
559 		return (EFBIG);
560 	if (uio->uio_resid == 0)
561 		return (0);
562 	/*
563 	 * Maybe this should be above the vnode op call, but so long as
564 	 * file servers have no limits, i don't think it matters
565 	 */
566 	if (p && uio->uio_offset + uio->uio_resid >
567 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
568 		psignal(p, SIGXFSZ);
569 		return (EFBIG);
570 	}
571 
572 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
573 		int iomode = NFSV3WRITE_FILESYNC;
574 		boolean_t stalewriteverf = FALSE;
575 
576 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
577 		error = nfs_writerpc(vp, uio, &iomode, FALSE, &stalewriteverf);
578 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
579 		if (stalewriteverf)
580 			nfs_clearcommit(vp->v_mount);
581 		return (error);
582 	}
583 
584 	origoff = uio->uio_offset;
585 	do {
586 		boolean_t extending; /* if we are extending whole pages */
587 		u_quad_t oldsize;
588 		oldoff = uio->uio_offset;
589 		bytelen = uio->uio_resid;
590 
591 #ifndef NFS_V2_ONLY
592 		/*
593 		 * Check for a valid write lease.
594 		 */
595 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
596 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
597 			do {
598 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
599 			} while (error == NQNFS_EXPIRED);
600 			if (error)
601 				return (error);
602 			if (np->n_lrev != np->n_brev ||
603 			    (np->n_flag & NQNFSNONCACHE)) {
604 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
605 				if (error)
606 					return (error);
607 				np->n_brev = np->n_lrev;
608 			}
609 		}
610 #endif
611 		nfsstats.biocache_writes++;
612 
613 		oldsize = np->n_size;
614 		np->n_flag |= NMODIFIED;
615 		if (np->n_size < uio->uio_offset + bytelen) {
616 			np->n_size = uio->uio_offset + bytelen;
617 		}
618 		extending = ((uio->uio_offset & PAGE_MASK) == 0 &&
619 		    (bytelen & PAGE_MASK) == 0 &&
620 		    uio->uio_offset >= vp->v_size);
621 		win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
622 			    UBC_WRITE | (extending ? UBC_FAULTBUSY : 0));
623 		error = uiomove(win, bytelen, uio);
624 		ubc_release(win, 0);
625 		if (error) {
626 			if (extending) {
627 				/*
628 				 * backout size and free pages past eof.
629 				 */
630 				np->n_size = oldsize;
631 				simple_lock(&vp->v_interlock);
632 				(void)VOP_PUTPAGES(vp, round_page(vp->v_size),
633 				    0, PGO_SYNCIO | PGO_FREE);
634 			}
635 			break;
636 		}
637 		wrotedta = 1;
638 
639 		/*
640 		 * update UVM's notion of the size now that we've
641 		 * copied the data into the vnode's pages.
642 		 */
643 
644 		if (vp->v_size < uio->uio_offset) {
645 			uvm_vnp_setsize(vp, uio->uio_offset);
646 			extended = 1;
647 		}
648 
649 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
650 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
651 			simple_lock(&vp->v_interlock);
652 			error = VOP_PUTPAGES(vp,
653 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
654 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
655 				       ~(nmp->nm_wsize - 1)), PGO_CLEANIT);
656 		}
657 	} while (uio->uio_resid > 0);
658 	if (wrotedta)
659 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0));
660 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
661 		simple_lock(&vp->v_interlock);
662 		error = VOP_PUTPAGES(vp,
663 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
664 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
665 			       ~(nmp->nm_wsize - 1)),
666 		    PGO_CLEANIT | PGO_SYNCIO);
667 	}
668 	return error;
669 }
670 
671 /*
672  * Get an nfs cache block.
673  * Allocate a new one if the block isn't currently in the cache
674  * and return the block marked busy. If the calling process is
675  * interrupted by a signal for an interruptible mount point, return
676  * NULL.
677  */
678 struct buf *
679 nfs_getcacheblk(vp, bn, size, p)
680 	struct vnode *vp;
681 	daddr_t bn;
682 	int size;
683 	struct proc *p;
684 {
685 	struct buf *bp;
686 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
687 
688 	if (nmp->nm_flag & NFSMNT_INT) {
689 		bp = getblk(vp, bn, size, PCATCH, 0);
690 		while (bp == NULL) {
691 			if (nfs_sigintr(nmp, NULL, p))
692 				return (NULL);
693 			bp = getblk(vp, bn, size, 0, 2 * hz);
694 		}
695 	} else
696 		bp = getblk(vp, bn, size, 0, 0);
697 	return (bp);
698 }
699 
700 /*
701  * Flush and invalidate all dirty buffers. If another process is already
702  * doing the flush, just wait for completion.
703  */
704 int
705 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
706 	struct vnode *vp;
707 	int flags;
708 	struct ucred *cred;
709 	struct proc *p;
710 	int intrflg;
711 {
712 	struct nfsnode *np = VTONFS(vp);
713 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
714 	int error = 0, slpflag, slptimeo;
715 
716 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
717 		intrflg = 0;
718 	if (intrflg) {
719 		slpflag = PCATCH;
720 		slptimeo = 2 * hz;
721 	} else {
722 		slpflag = 0;
723 		slptimeo = 0;
724 	}
725 	/*
726 	 * First wait for any other process doing a flush to complete.
727 	 */
728 	simple_lock(&vp->v_interlock);
729 	while (np->n_flag & NFLUSHINPROG) {
730 		np->n_flag |= NFLUSHWANT;
731 		error = ltsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
732 			slptimeo, &vp->v_interlock);
733 		if (error && intrflg && nfs_sigintr(nmp, NULL, p)) {
734 			simple_unlock(&vp->v_interlock);
735 			return EINTR;
736 		}
737 	}
738 
739 	/*
740 	 * Now, flush as required.
741 	 */
742 	np->n_flag |= NFLUSHINPROG;
743 	simple_unlock(&vp->v_interlock);
744 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
745 	while (error) {
746 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
747 			error = EINTR;
748 			break;
749 		}
750 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
751 	}
752 	simple_lock(&vp->v_interlock);
753 	if (error == 0)
754 		np->n_flag &= ~NMODIFIED;
755 	np->n_flag &= ~NFLUSHINPROG;
756 	if (np->n_flag & NFLUSHWANT) {
757 		np->n_flag &= ~NFLUSHWANT;
758 		wakeup(&np->n_flag);
759 	}
760 	simple_unlock(&vp->v_interlock);
761 	return error;
762 }
763 
764 /*
765  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
766  * This is mainly to avoid queueing async I/O requests when the nfsiods
767  * are all hung on a dead server.
768  */
769 
770 int
771 nfs_asyncio(bp)
772 	struct buf *bp;
773 {
774 	int i;
775 	struct nfsmount *nmp;
776 	int gotiod, slpflag = 0, slptimeo = 0, error;
777 
778 	if (nfs_numasync == 0)
779 		return (EIO);
780 
781 	nmp = VFSTONFS(bp->b_vp->v_mount);
782 again:
783 	if (nmp->nm_flag & NFSMNT_INT)
784 		slpflag = PCATCH;
785 	gotiod = FALSE;
786 
787 	/*
788 	 * Find a free iod to process this request.
789 	 */
790 
791 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++) {
792 		struct nfs_iod *iod = &nfs_asyncdaemon[i];
793 
794 		simple_lock(&iod->nid_slock);
795 		if (iod->nid_want) {
796 			/*
797 			 * Found one, so wake it up and tell it which
798 			 * mount to process.
799 			 */
800 			iod->nid_want = NULL;
801 			iod->nid_mount = nmp;
802 			wakeup(&iod->nid_want);
803 			simple_lock(&nmp->nm_slock);
804 			simple_unlock(&iod->nid_slock);
805 			nmp->nm_bufqiods++;
806 			gotiod = TRUE;
807 			break;
808 		}
809 		simple_unlock(&iod->nid_slock);
810 	}
811 
812 	/*
813 	 * If none are free, we may already have an iod working on this mount
814 	 * point.  If so, it will process our request.
815 	 */
816 
817 	if (!gotiod) {
818 		simple_lock(&nmp->nm_slock);
819 		if (nmp->nm_bufqiods > 0)
820 			gotiod = TRUE;
821 	}
822 
823 	LOCK_ASSERT(simple_lock_held(&nmp->nm_slock));
824 
825 	/*
826 	 * If we have an iod which can process the request, then queue
827 	 * the buffer.  However, even if we have an iod, do not initiate
828 	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
829 	 * is via local loopback, we may put curproc (pagedaemon) to sleep
830 	 * waiting for the writes to complete. But the server (ourself)
831 	 * may block the write, waiting for its (ie., our) pagedaemon
832 	 * to produce clean pages to handle the write: deadlock.
833 	 * XXX: start non-loopback mounts straight away?  If "lots free",
834 	 * let pagedaemon start loopback writes anyway?
835 	 */
836 	if (gotiod) {
837 
838 		/*
839 		 * Ensure that the queue never grows too large.
840 		 */
841 		if (curproc == uvm.pagedaemon_proc) {
842 	  		/* Enque for later, to avoid free-page deadlock */
843 			  (void) 0;
844 		} else while (nmp->nm_bufqlen >= 2*nfs_numasync) {
845 			nmp->nm_bufqwant = TRUE;
846 			error = ltsleep(&nmp->nm_bufq,
847 			    slpflag | PRIBIO | PNORELOCK,
848 			    "nfsaio", slptimeo, &nmp->nm_slock);
849 			if (error) {
850 				if (nfs_sigintr(nmp, NULL, curproc))
851 					return (EINTR);
852 				if (slpflag == PCATCH) {
853 					slpflag = 0;
854 					slptimeo = 2 * hz;
855 				}
856 			}
857 
858 			/*
859 			 * We might have lost our iod while sleeping,
860 			 * so check and loop if nescessary.
861 			 */
862 
863 			if (nmp->nm_bufqiods == 0)
864 				goto again;
865 
866 			simple_lock(&nmp->nm_slock);
867 		}
868 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
869 		nmp->nm_bufqlen++;
870 		simple_unlock(&nmp->nm_slock);
871 		return (0);
872 	}
873 	simple_unlock(&nmp->nm_slock);
874 
875 	/*
876 	 * All the iods are busy on other mounts, so return EIO to
877 	 * force the caller to process the i/o synchronously.
878 	 */
879 
880 	return (EIO);
881 }
882 
883 /*
884  * nfs_doio for read.
885  */
886 static int
887 nfs_doio_read(bp, uiop)
888 	struct buf *bp;
889 	struct uio *uiop;
890 {
891 	struct vnode *vp = bp->b_vp;
892 	struct nfsnode *np = VTONFS(vp);
893 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
894 	int error = 0;
895 
896 	uiop->uio_rw = UIO_READ;
897 	switch (vp->v_type) {
898 	case VREG:
899 		nfsstats.read_bios++;
900 		error = nfs_readrpc(vp, uiop);
901 		if (!error && uiop->uio_resid) {
902 			int diff, len;
903 
904 			/*
905 			 * If uio_resid > 0, there is a hole in the file and
906 			 * no writes after the hole have been pushed to
907 			 * the server yet or the file has been truncated
908 			 * on the server.
909 			 * Just zero fill the rest of the valid area.
910 			 */
911 
912 			KASSERT(vp->v_size >=
913 			    uiop->uio_offset + uiop->uio_resid);
914 			diff = bp->b_bcount - uiop->uio_resid;
915 			len = uiop->uio_resid;
916 			memset((char *)bp->b_data + diff, 0, len);
917 		}
918 		if (uiop->uio_procp && (vp->v_flag & VTEXT) &&
919 		    (((nmp->nm_flag & NFSMNT_NQNFS) &&
920 		      NQNFS_CKINVALID(vp, np, ND_READ) &&
921 		      np->n_lrev != np->n_brev) ||
922 		     (!(nmp->nm_flag & NFSMNT_NQNFS) &&
923 		      timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)))) {
924 			uprintf("Process killed due to "
925 				"text file modification\n");
926 			psignal(uiop->uio_procp, SIGKILL);
927 #if 0 /* XXX NJWLWP */
928 			uiop->uio_procp->p_holdcnt++;
929 #endif
930 		}
931 		break;
932 	case VLNK:
933 		KASSERT(uiop->uio_offset == (off_t)0);
934 		nfsstats.readlink_bios++;
935 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
936 		break;
937 	case VDIR:
938 		nfsstats.readdir_bios++;
939 		uiop->uio_offset = bp->b_dcookie;
940 #ifndef NFS_V2_ONLY
941 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
942 			error = nfs_readdirplusrpc(vp, uiop, np->n_rcred);
943 			if (error == NFSERR_NOTSUPP)
944 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
945 		}
946 #else
947 		nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
948 #endif
949 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
950 			error = nfs_readdirrpc(vp, uiop, np->n_rcred);
951 		if (!error) {
952 			bp->b_dcookie = uiop->uio_offset;
953 		}
954 		break;
955 	default:
956 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
957 		break;
958 	}
959 	if (error) {
960 		bp->b_flags |= B_ERROR;
961 		bp->b_error = error;
962 	}
963 	return error;
964 }
965 
966 /*
967  * nfs_doio for write.
968  */
969 static int
970 nfs_doio_write(bp, uiop)
971 	struct buf *bp;
972 	struct uio *uiop;
973 {
974 	struct vnode *vp = bp->b_vp;
975 	struct nfsnode *np = VTONFS(vp);
976 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
977 	int iomode;
978 	boolean_t stalewriteverf = FALSE;
979 	int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT;
980 	struct vm_page *pgs[npages];
981 #ifndef NFS_V2_ONLY
982 	boolean_t needcommit = TRUE; /* need only COMMIT RPC */
983 #else
984 	boolean_t needcommit = FALSE; /* need only COMMIT RPC */
985 #endif
986 	boolean_t pageprotected;
987 	struct uvm_object *uobj = &vp->v_uobj;
988 	int error;
989 	off_t off, cnt;
990 
991 	if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) {
992 		iomode = NFSV3WRITE_UNSTABLE;
993 	} else {
994 		iomode = NFSV3WRITE_FILESYNC;
995 	}
996 
997 #ifndef NFS_V2_ONLY
998 again:
999 #endif
1000 	lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1001 
1002 	for (i = 0; i < npages; i++) {
1003 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
1004 		if (pgs[i]->uobject == uobj &&
1005 		    pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) {
1006 			KASSERT(pgs[i]->flags & PG_BUSY);
1007 			/*
1008 			 * this page belongs to our object.
1009 			 */
1010 			simple_lock(&uobj->vmobjlock);
1011 			/*
1012 			 * write out the page stably if it's about to
1013 			 * be released because we can't resend it
1014 			 * on the server crash.
1015 			 *
1016 			 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be
1017 			 * changed until unbusy the page.
1018 			 */
1019 			if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT))
1020 				iomode = NFSV3WRITE_FILESYNC;
1021 			/*
1022 			 * if we met a page which hasn't been sent yet,
1023 			 * we need do WRITE RPC.
1024 			 */
1025 			if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0)
1026 				needcommit = FALSE;
1027 			simple_unlock(&uobj->vmobjlock);
1028 		} else {
1029 			iomode = NFSV3WRITE_FILESYNC;
1030 			needcommit = FALSE;
1031 		}
1032 	}
1033 	if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) {
1034 		simple_lock(&uobj->vmobjlock);
1035 		for (i = 0; i < npages; i++) {
1036 			pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY;
1037 			pmap_page_protect(pgs[i], VM_PROT_READ);
1038 		}
1039 		simple_unlock(&uobj->vmobjlock);
1040 		pageprotected = TRUE; /* pages can't be modified during i/o. */
1041 	} else
1042 		pageprotected = FALSE;
1043 
1044 	/*
1045 	 * Send the data to the server if necessary,
1046 	 * otherwise just send a commit rpc.
1047 	 */
1048 #ifndef NFS_V2_ONLY
1049 	if (needcommit) {
1050 
1051 		/*
1052 		 * If the buffer is in the range that we already committed,
1053 		 * there's nothing to do.
1054 		 *
1055 		 * If it's in the range that we need to commit, push the
1056 		 * whole range at once, otherwise only push the buffer.
1057 		 * In both these cases, acquire the commit lock to avoid
1058 		 * other processes modifying the range.
1059 		 */
1060 
1061 		off = uiop->uio_offset;
1062 		cnt = bp->b_bcount;
1063 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1064 		if (!nfs_in_committed_range(vp, off, bp->b_bcount)) {
1065 			boolean_t pushedrange;
1066 			if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) {
1067 				pushedrange = TRUE;
1068 				off = np->n_pushlo;
1069 				cnt = np->n_pushhi - np->n_pushlo;
1070 			} else {
1071 				pushedrange = FALSE;
1072 			}
1073 			error = nfs_commit(vp, off, cnt, curproc);
1074 			if (error == 0) {
1075 				if (pushedrange) {
1076 					nfs_merge_commit_ranges(vp);
1077 				} else {
1078 					nfs_add_committed_range(vp, off, cnt);
1079 				}
1080 			}
1081 		} else {
1082 			error = 0;
1083 		}
1084 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1085 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1086 		if (!error) {
1087 			/*
1088 			 * pages are now on stable storage.
1089 			 */
1090 			uiop->uio_resid = 0;
1091 			simple_lock(&uobj->vmobjlock);
1092 			for (i = 0; i < npages; i++) {
1093 				pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1094 			}
1095 			simple_unlock(&uobj->vmobjlock);
1096 			return 0;
1097 		} else if (error == NFSERR_STALEWRITEVERF) {
1098 			nfs_clearcommit(vp->v_mount);
1099 			goto again;
1100 		}
1101 		if (error) {
1102 			bp->b_flags |= B_ERROR;
1103 			bp->b_error = np->n_error = error;
1104 			np->n_flag |= NWRITEERR;
1105 		}
1106 		return error;
1107 	}
1108 #endif
1109 	off = uiop->uio_offset;
1110 	cnt = bp->b_bcount;
1111 	uiop->uio_rw = UIO_WRITE;
1112 	nfsstats.write_bios++;
1113 	error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf);
1114 #ifndef NFS_V2_ONLY
1115 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1116 		/*
1117 		 * we need to commit pages later.
1118 		 */
1119 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1120 		nfs_add_tobecommitted_range(vp, off, cnt);
1121 		/*
1122 		 * if there can be too many uncommitted pages, commit them now.
1123 		 */
1124 		if (np->n_pushhi - np->n_pushlo > nfs_commitsize) {
1125 			off = np->n_pushlo;
1126 			cnt = nfs_commitsize >> 1;
1127 			error = nfs_commit(vp, off, cnt, curproc);
1128 			if (!error) {
1129 				nfs_add_committed_range(vp, off, cnt);
1130 				nfs_del_tobecommitted_range(vp, off, cnt);
1131 			}
1132 			if (error == NFSERR_STALEWRITEVERF) {
1133 				stalewriteverf = TRUE;
1134 				error = 0; /* it isn't a real error */
1135 			}
1136 		} else {
1137 			/*
1138 			 * re-dirty pages so that they will be passed
1139 			 * to us later again.
1140 			 */
1141 			simple_lock(&uobj->vmobjlock);
1142 			for (i = 0; i < npages; i++) {
1143 				pgs[i]->flags &= ~PG_CLEAN;
1144 			}
1145 			simple_unlock(&uobj->vmobjlock);
1146 		}
1147 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1148 	} else
1149 #endif
1150 	if (!error) {
1151 		/*
1152 		 * pages are now on stable storage.
1153 		 */
1154 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1155 		nfs_del_committed_range(vp, off, cnt);
1156 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1157 		simple_lock(&uobj->vmobjlock);
1158 		for (i = 0; i < npages; i++) {
1159 			pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1160 		}
1161 		simple_unlock(&uobj->vmobjlock);
1162 	} else {
1163 		/*
1164 		 * we got an error.
1165 		 */
1166 		bp->b_flags |= B_ERROR;
1167 		bp->b_error = np->n_error = error;
1168 		np->n_flag |= NWRITEERR;
1169 	}
1170 
1171 	lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1172 
1173 	if (stalewriteverf) {
1174 		nfs_clearcommit(vp->v_mount);
1175 	}
1176 	return error;
1177 }
1178 
1179 /*
1180  * nfs_doio for B_PHYS.
1181  */
1182 static int
1183 nfs_doio_phys(bp, uiop)
1184 	struct buf *bp;
1185 	struct uio *uiop;
1186 {
1187 	struct vnode *vp = bp->b_vp;
1188 	int error;
1189 
1190 	uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
1191 	if (bp->b_flags & B_READ) {
1192 		uiop->uio_rw = UIO_READ;
1193 		nfsstats.read_physios++;
1194 		error = nfs_readrpc(vp, uiop);
1195 	} else {
1196 		int iomode = NFSV3WRITE_DATASYNC;
1197 		boolean_t stalewriteverf;
1198 		struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1199 
1200 		uiop->uio_rw = UIO_WRITE;
1201 		nfsstats.write_physios++;
1202 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1203 		error = nfs_writerpc(vp, uiop, &iomode, FALSE, &stalewriteverf);
1204 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1205 		if (stalewriteverf) {
1206 			nfs_clearcommit(bp->b_vp->v_mount);
1207 		}
1208 	}
1209 	if (error) {
1210 		bp->b_flags |= B_ERROR;
1211 		bp->b_error = error;
1212 	}
1213 	return error;
1214 }
1215 
1216 /*
1217  * Do an I/O operation to/from a cache block. This may be called
1218  * synchronously or from an nfsiod.
1219  */
1220 int
1221 nfs_doio(bp, p)
1222 	struct buf *bp;
1223 	struct proc *p;
1224 {
1225 	int error;
1226 	struct uio uio;
1227 	struct uio *uiop = &uio;
1228 	struct iovec io;
1229 	UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist);
1230 
1231 	uiop->uio_iov = &io;
1232 	uiop->uio_iovcnt = 1;
1233 	uiop->uio_segflg = UIO_SYSSPACE;
1234 	uiop->uio_procp = NULL;
1235 	uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
1236 	io.iov_base = bp->b_data;
1237 	io.iov_len = uiop->uio_resid = bp->b_bcount;
1238 
1239 	/*
1240 	 * Historically, paging was done with physio, but no more...
1241 	 */
1242 	if (bp->b_flags & B_PHYS) {
1243 		/*
1244 		 * ...though reading /dev/drum still gets us here.
1245 		 */
1246 		error = nfs_doio_phys(bp, uiop);
1247 	} else if (bp->b_flags & B_READ) {
1248 		error = nfs_doio_read(bp, uiop);
1249 	} else {
1250 		error = nfs_doio_write(bp, uiop);
1251 	}
1252 	bp->b_resid = uiop->uio_resid;
1253 	biodone(bp);
1254 	return (error);
1255 }
1256 
1257 /*
1258  * Vnode op for VM getpages.
1259  */
1260 
1261 int
1262 nfs_getpages(v)
1263 	void *v;
1264 {
1265 	struct vop_getpages_args /* {
1266 		struct vnode *a_vp;
1267 		voff_t a_offset;
1268 		struct vm_page **a_m;
1269 		int *a_count;
1270 		int a_centeridx;
1271 		vm_prot_t a_access_type;
1272 		int a_advice;
1273 		int a_flags;
1274 	} */ *ap = v;
1275 
1276 	struct vnode *vp = ap->a_vp;
1277 	struct uvm_object *uobj = &vp->v_uobj;
1278 	struct nfsnode *np = VTONFS(vp);
1279 	const int npages = *ap->a_count;
1280 	struct vm_page *pg, **pgs, *opgs[npages];
1281 	off_t origoffset, len;
1282 	int i, error;
1283 	boolean_t v3 = NFS_ISV3(vp);
1284 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1285 	boolean_t locked = (ap->a_flags & PGO_LOCKED) != 0;
1286 
1287 	/*
1288 	 * call the genfs code to get the pages.  `pgs' may be NULL
1289 	 * when doing read-ahead.
1290 	 */
1291 
1292 	pgs = ap->a_m;
1293 	if (write && locked && v3) {
1294 		KASSERT(pgs != NULL);
1295 #ifdef DEBUG
1296 
1297 		/*
1298 		 * If PGO_LOCKED is set, real pages shouldn't exists
1299 		 * in the array.
1300 		 */
1301 
1302 		for (i = 0; i < npages; i++)
1303 			KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE);
1304 #endif
1305 		memcpy(opgs, pgs, npages * sizeof(struct vm_pages *));
1306 	}
1307 	error = genfs_getpages(v);
1308 	if (error) {
1309 		return (error);
1310 	}
1311 
1312 	/*
1313 	 * for read faults where the nfs node is not yet marked NMODIFIED,
1314 	 * set PG_RDONLY on the pages so that we come back here if someone
1315 	 * tries to modify later via the mapping that will be entered for
1316 	 * this fault.
1317 	 */
1318 
1319 	if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) {
1320 		if (!locked) {
1321 			simple_lock(&uobj->vmobjlock);
1322 		}
1323 		for (i = 0; i < npages; i++) {
1324 			pg = pgs[i];
1325 			if (pg == NULL || pg == PGO_DONTCARE) {
1326 				continue;
1327 			}
1328 			pg->flags |= PG_RDONLY;
1329 		}
1330 		if (!locked) {
1331 			simple_unlock(&uobj->vmobjlock);
1332 		}
1333 	}
1334 	if (!write) {
1335 		return (0);
1336 	}
1337 
1338 	/*
1339 	 * this is a write fault, update the commit info.
1340 	 */
1341 
1342 	origoffset = ap->a_offset;
1343 	len = npages << PAGE_SHIFT;
1344 
1345 	if (v3) {
1346 		error = lockmgr(&np->n_commitlock,
1347 		    LK_EXCLUSIVE | (locked ? LK_NOWAIT : 0), NULL);
1348 		if (error) {
1349 			KASSERT(locked != 0);
1350 
1351 			/*
1352 			 * Since PGO_LOCKED is set, we need to unbusy
1353 			 * all pages fetched by genfs_getpages() above,
1354 			 * tell the caller that there are no pages
1355 			 * available and put back original pgs array.
1356 			 */
1357 
1358 			uvm_lock_pageq();
1359 			uvm_page_unbusy(pgs, npages);
1360 			uvm_unlock_pageq();
1361 			*ap->a_count = 0;
1362 			memcpy(pgs, opgs,
1363 			    npages * sizeof(struct vm_pages *));
1364 			return (error);
1365 		}
1366 		nfs_del_committed_range(vp, origoffset, len);
1367 		nfs_del_tobecommitted_range(vp, origoffset, len);
1368 	}
1369 	np->n_flag |= NMODIFIED;
1370 	if (!locked) {
1371 		simple_lock(&uobj->vmobjlock);
1372 	}
1373 	for (i = 0; i < npages; i++) {
1374 		pg = pgs[i];
1375 		if (pg == NULL || pg == PGO_DONTCARE) {
1376 			continue;
1377 		}
1378 		pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1379 	}
1380 	if (!locked) {
1381 		simple_unlock(&uobj->vmobjlock);
1382 	}
1383 	if (v3) {
1384 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1385 	}
1386 	return (0);
1387 }
1388