xref: /netbsd-src/sys/nfs/nfs_bio.c (revision aa73cae19608873cc4d1f712c4a0f8f8435f1ffa)
1 /*	$NetBSD: nfs_bio.c,v 1.128 2005/02/26 22:39:50 perry 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 1.128 2005/02/26 22:39:50 perry Exp $");
39 
40 #include "opt_nfs.h"
41 #include "opt_ddb.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/proc.h>
48 #include <sys/buf.h>
49 #include <sys/vnode.h>
50 #include <sys/mount.h>
51 #include <sys/kernel.h>
52 #include <sys/namei.h>
53 #include <sys/dirent.h>
54 #include <sys/malloc.h>
55 
56 #include <uvm/uvm_extern.h>
57 #include <uvm/uvm.h>
58 
59 #include <nfs/rpcv2.h>
60 #include <nfs/nfsproto.h>
61 #include <nfs/nfs.h>
62 #include <nfs/nfsmount.h>
63 #include <nfs/nqnfs.h>
64 #include <nfs/nfsnode.h>
65 #include <nfs/nfs_var.h>
66 
67 extern int nfs_numasync;
68 extern int nfs_commitsize;
69 extern struct nfsstats nfsstats;
70 
71 static int nfs_doio_read __P((struct buf *, struct uio *));
72 static int nfs_doio_write __P((struct buf *, struct uio *));
73 static int nfs_doio_phys __P((struct buf *, struct uio *));
74 
75 /*
76  * Vnode op for read using bio
77  * Any similarity to readip() is purely coincidental
78  */
79 int
80 nfs_bioread(vp, uio, ioflag, cred, cflag)
81 	struct vnode *vp;
82 	struct uio *uio;
83 	int ioflag, cflag;
84 	struct ucred *cred;
85 {
86 	struct nfsnode *np = VTONFS(vp);
87 	struct buf *bp = NULL, *rabp;
88 	struct proc *p;
89 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
90 	struct nfsdircache *ndp = NULL, *nndp = NULL;
91 	caddr_t baddr, ep, edp;
92 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
93 	int enough = 0;
94 	struct dirent *dp, *pdp;
95 	off_t curoff = 0;
96 
97 #ifdef DIAGNOSTIC
98 	if (uio->uio_rw != UIO_READ)
99 		panic("nfs_read mode");
100 #endif
101 	if (uio->uio_resid == 0)
102 		return (0);
103 	if (vp->v_type != VDIR && uio->uio_offset < 0)
104 		return (EINVAL);
105 	p = uio->uio_procp;
106 #ifndef NFS_V2_ONLY
107 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
108 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
109 		(void)nfs_fsinfo(nmp, vp, cred, p);
110 #endif
111 	if (vp->v_type != VDIR &&
112 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
113 		return (EFBIG);
114 
115 	/*
116 	 * For nfs, cache consistency can only be maintained approximately.
117 	 * Although RFC1094 does not specify the criteria, the following is
118 	 * believed to be compatible with the reference port.
119 	 * For nqnfs, full cache consistency is maintained within the loop.
120 	 * For nfs:
121 	 * If the file's modify time on the server has changed since the
122 	 * last read rpc or you have written to the file,
123 	 * you may have lost data cache consistency with the
124 	 * server, so flush all of the file's data out of the cache.
125 	 * Then force a getattr rpc to ensure that you have up to date
126 	 * attributes.
127 	 * NB: This implies that cache data can be read when up to
128 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
129 	 * attributes this could be forced by setting n_attrstamp to 0 before
130 	 * the VOP_GETATTR() call.
131 	 */
132 
133 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
134 		error = nfs_flushstalebuf(vp, cred, p,
135 		    NFS_FLUSHSTALEBUF_MYWRITE);
136 		if (error)
137 			return error;
138 	}
139 
140 	do {
141 #ifndef NFS_V2_ONLY
142 	    /*
143 	     * Get a valid lease. If cached data is stale, flush it.
144 	     */
145 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
146 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
147 		    do {
148 			error = nqnfs_getlease(vp, ND_READ, cred, p);
149 		    } while (error == NQNFS_EXPIRED);
150 		    if (error)
151 			return (error);
152 		    if (np->n_lrev != np->n_brev ||
153 			(np->n_flag & NQNFSNONCACHE) ||
154 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
155 			if (vp->v_type == VDIR) {
156 				nfs_invaldircache(vp, 0);
157 			}
158 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
159 			if (error)
160 			    return (error);
161 			np->n_brev = np->n_lrev;
162 		    }
163 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
164 		    nfs_invaldircache(vp, 0);
165 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
166 		    if (error)
167 			return (error);
168 		}
169 	    }
170 #endif
171 	    /*
172 	     * Don't cache symlinks.
173 	     */
174 	    if (np->n_flag & NQNFSNONCACHE
175 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
176 		switch (vp->v_type) {
177 		case VREG:
178 			return (nfs_readrpc(vp, uio));
179 		case VLNK:
180 			return (nfs_readlinkrpc(vp, uio, cred));
181 		case VDIR:
182 			break;
183 		default:
184 			printf(" NQNFSNONCACHE: type %x unexpected\n",
185 			    vp->v_type);
186 		};
187 	    }
188 	    baddr = (caddr_t)0;
189 	    switch (vp->v_type) {
190 	    case VREG:
191 		nfsstats.biocache_reads++;
192 
193 		error = 0;
194 		if (uio->uio_offset >= np->n_size) {
195 			break;
196 		}
197 		while (uio->uio_resid > 0) {
198 			void *win;
199 			int flags;
200 			vsize_t bytelen = MIN(np->n_size - uio->uio_offset,
201 					      uio->uio_resid);
202 
203 			if (bytelen == 0)
204 				break;
205 			win = ubc_alloc(&vp->v_uobj, uio->uio_offset,
206 					&bytelen, UBC_READ);
207 			error = uiomove(win, bytelen, uio);
208 			flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
209 			ubc_release(win, flags);
210 			if (error) {
211 				break;
212 			}
213 		}
214 		n = 0;
215 		break;
216 
217 	    case VLNK:
218 		nfsstats.biocache_readlinks++;
219 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
220 		if (!bp)
221 			return (EINTR);
222 		if ((bp->b_flags & B_DONE) == 0) {
223 			bp->b_flags |= B_READ;
224 			error = nfs_doio(bp, p);
225 			if (error) {
226 				brelse(bp);
227 				return (error);
228 			}
229 		}
230 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
231 		got_buf = 1;
232 		on = 0;
233 		break;
234 	    case VDIR:
235 diragain:
236 		nfsstats.biocache_readdirs++;
237 		ndp = nfs_searchdircache(vp, uio->uio_offset,
238 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
239 		if (!ndp) {
240 			/*
241 			 * We've been handed a cookie that is not
242 			 * in the cache. If we're not translating
243 			 * 32 <-> 64, it may be a value that was
244 			 * flushed out of the cache because it grew
245 			 * too big. Let the server judge if it's
246 			 * valid or not. In the translation case,
247 			 * we have no way of validating this value,
248 			 * so punt.
249 			 */
250 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
251 				return (EINVAL);
252 			ndp = nfs_enterdircache(vp, uio->uio_offset,
253 				uio->uio_offset, 0, 0);
254 		}
255 
256 		if (NFS_EOFVALID(np) &&
257 		    ndp->dc_cookie == np->n_direofoffset) {
258 			nfs_putdircache(np, ndp);
259 			nfsstats.direofcache_hits++;
260 			return (0);
261 		}
262 
263 		bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, p);
264 		if (!bp)
265 		    return (EINTR);
266 		if ((bp->b_flags & B_DONE) == 0) {
267 		    bp->b_flags |= B_READ;
268 		    bp->b_dcookie = ndp->dc_blkcookie;
269 		    error = nfs_doio(bp, p);
270 		    if (error) {
271 			/*
272 			 * Yuck! The directory has been modified on the
273 			 * server. Punt and let the userland code
274 			 * deal with it.
275 			 */
276 			nfs_putdircache(np, ndp);
277 			brelse(bp);
278 			if (error == NFSERR_BAD_COOKIE) {
279 			    nfs_invaldircache(vp, 0);
280 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
281 			    error = EINVAL;
282 			}
283 			return (error);
284 		    }
285 		}
286 
287 		/*
288 		 * Just return if we hit EOF right away with this
289 		 * block. Always check here, because direofoffset
290 		 * may have been set by an nfsiod since the last
291 		 * check.
292 		 *
293 		 * also, empty block implies EOF.
294 		 */
295 
296 		if (bp->b_bcount == bp->b_resid ||
297 		    (NFS_EOFVALID(np) &&
298 		    ndp->dc_blkcookie == np->n_direofoffset)) {
299 			KASSERT(bp->b_bcount != bp->b_resid ||
300 			    ndp->dc_blkcookie == bp->b_dcookie);
301 			nfs_putdircache(np, ndp);
302 			bp->b_flags |= B_NOCACHE;
303 			brelse(bp);
304 			return 0;
305 		}
306 
307 		/*
308 		 * Find the entry we were looking for in the block.
309 		 */
310 
311 		en = ndp->dc_entry;
312 
313 		pdp = dp = (struct dirent *)bp->b_data;
314 		edp = bp->b_data + bp->b_bcount - bp->b_resid;
315 		enn = 0;
316 		while (enn < en && (caddr_t)dp < edp) {
317 			pdp = dp;
318 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
319 			enn++;
320 		}
321 
322 		/*
323 		 * If the entry number was bigger than the number of
324 		 * entries in the block, or the cookie of the previous
325 		 * entry doesn't match, the directory cache is
326 		 * stale. Flush it and try again (i.e. go to
327 		 * the server).
328 		 */
329 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
330 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
331 #ifdef DEBUG
332 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
333 				pdp, dp, edp,
334 				(unsigned long)uio->uio_offset,
335 				(unsigned long)NFS_GETCOOKIE(pdp));
336 #endif
337 			nfs_putdircache(np, ndp);
338 			brelse(bp);
339 			nfs_invaldircache(vp, 0);
340 			nfs_vinvalbuf(vp, 0, cred, p, 0);
341 			goto diragain;
342 		}
343 
344 		on = (caddr_t)dp - bp->b_data;
345 
346 		/*
347 		 * Cache all entries that may be exported to the
348 		 * user, as they may be thrown back at us. The
349 		 * NFSBIO_CACHECOOKIES flag indicates that all
350 		 * entries are being 'exported', so cache them all.
351 		 */
352 
353 		if (en == 0 && pdp == dp) {
354 			dp = (struct dirent *)
355 			    ((caddr_t)dp + dp->d_reclen);
356 			enn++;
357 		}
358 
359 		if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) {
360 			n = uio->uio_resid;
361 			enough = 1;
362 		} else
363 			n = bp->b_bcount - bp->b_resid - on;
364 
365 		ep = bp->b_data + on + n;
366 
367 		/*
368 		 * Find last complete entry to copy, caching entries
369 		 * (if requested) as we go.
370 		 */
371 
372 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
373 			if (cflag & NFSBIO_CACHECOOKIES) {
374 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
375 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
376 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
377 					NFS_STASHCOOKIE32(pdp,
378 					    nndp->dc_cookie32);
379 				}
380 				nfs_putdircache(np, nndp);
381 			}
382 			pdp = dp;
383 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
384 			enn++;
385 		}
386 		nfs_putdircache(np, ndp);
387 
388 		/*
389 		 * If the last requested entry was not the last in the
390 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
391 		 * cache the cookie of the last requested one, and
392 		 * set of the offset to it.
393 		 */
394 
395 		if ((on + n) < bp->b_bcount - bp->b_resid) {
396 			curoff = NFS_GETCOOKIE(pdp);
397 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
398 			    enn, bp->b_lblkno);
399 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
400 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
401 				curoff = nndp->dc_cookie32;
402 			}
403 			nfs_putdircache(np, nndp);
404 		} else
405 			curoff = bp->b_dcookie;
406 
407 		/*
408 		 * Always cache the entry for the next block,
409 		 * so that readaheads can use it.
410 		 */
411 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
412 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
413 			if (curoff == bp->b_dcookie) {
414 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
415 				curoff = nndp->dc_cookie32;
416 			}
417 		}
418 
419 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
420 
421 		/*
422 		 * If not eof and read aheads are enabled, start one.
423 		 * (You need the current block first, so that you have the
424 		 *  directory offset cookie of the next block.)
425 		 */
426 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
427 		    !NFS_EOFVALID(np) && !(np->n_flag & NQNFSNONCACHE)) {
428 			rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp),
429 						NFS_DIRBLKSIZ, p);
430 			if (rabp) {
431 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
432 				rabp->b_dcookie = nndp->dc_cookie;
433 				rabp->b_flags |= (B_READ | B_ASYNC);
434 				if (nfs_asyncio(rabp)) {
435 				    rabp->b_flags |= B_INVAL;
436 				    brelse(rabp);
437 				}
438 			    } else
439 				brelse(rabp);
440 			}
441 		}
442 		nfs_putdircache(np, nndp);
443 		got_buf = 1;
444 		break;
445 	    default:
446 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
447 		break;
448 	    }
449 
450 	    if (n > 0) {
451 		if (!baddr)
452 			baddr = bp->b_data;
453 		error = uiomove(baddr + on, (int)n, uio);
454 	    }
455 	    switch (vp->v_type) {
456 	    case VREG:
457 		break;
458 	    case VLNK:
459 		n = 0;
460 		break;
461 	    case VDIR:
462 		if (np->n_flag & NQNFSNONCACHE)
463 			bp->b_flags |= B_INVAL;
464 		uio->uio_offset = curoff;
465 		if (enough)
466 			n = 0;
467 		break;
468 	    default:
469 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
470 	    }
471 	    if (got_buf)
472 		brelse(bp);
473 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
474 	return (error);
475 }
476 
477 /*
478  * Vnode op for write using bio
479  */
480 int
481 nfs_write(v)
482 	void *v;
483 {
484 	struct vop_write_args /* {
485 		struct vnode *a_vp;
486 		struct uio *a_uio;
487 		int  a_ioflag;
488 		struct ucred *a_cred;
489 	} */ *ap = v;
490 	struct uio *uio = ap->a_uio;
491 	struct proc *p = uio->uio_procp;
492 	struct vnode *vp = ap->a_vp;
493 	struct nfsnode *np = VTONFS(vp);
494 	struct ucred *cred = ap->a_cred;
495 	struct vattr vattr;
496 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
497 	void *win;
498 	voff_t oldoff, origoff;
499 	vsize_t bytelen;
500 	int flags, error = 0;
501 	int ioflag = ap->a_ioflag;
502 	int extended = 0, wrotedata = 0;
503 
504 #ifdef DIAGNOSTIC
505 	if (uio->uio_rw != UIO_WRITE)
506 		panic("nfs_write mode");
507 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
508 		panic("nfs_write proc");
509 #endif
510 	if (vp->v_type != VREG)
511 		return (EIO);
512 	if (np->n_flag & NWRITEERR) {
513 		np->n_flag &= ~NWRITEERR;
514 		return (np->n_error);
515 	}
516 #ifndef NFS_V2_ONLY
517 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
518 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
519 		(void)nfs_fsinfo(nmp, vp, cred, p);
520 #endif
521 	if (ioflag & (IO_APPEND | IO_SYNC)) {
522 		if (np->n_flag & NMODIFIED) {
523 			NFS_INVALIDATE_ATTRCACHE(np);
524 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
525 			if (error)
526 				return (error);
527 		}
528 		if (ioflag & IO_APPEND) {
529 			NFS_INVALIDATE_ATTRCACHE(np);
530 			error = VOP_GETATTR(vp, &vattr, cred, p);
531 			if (error)
532 				return (error);
533 			uio->uio_offset = np->n_size;
534 		}
535 	}
536 	if (uio->uio_offset < 0)
537 		return (EINVAL);
538 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
539 		return (EFBIG);
540 	if (uio->uio_resid == 0)
541 		return (0);
542 	/*
543 	 * Maybe this should be above the vnode op call, but so long as
544 	 * file servers have no limits, i don't think it matters
545 	 */
546 	if (p && uio->uio_offset + uio->uio_resid >
547 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
548 		psignal(p, SIGXFSZ);
549 		return (EFBIG);
550 	}
551 
552 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
553 		int iomode = NFSV3WRITE_FILESYNC;
554 		boolean_t stalewriteverf = FALSE;
555 
556 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
557 		error = nfs_writerpc(vp, uio, &iomode, FALSE, &stalewriteverf);
558 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
559 		if (stalewriteverf)
560 			nfs_clearcommit(vp->v_mount);
561 		return (error);
562 	}
563 
564 	origoff = uio->uio_offset;
565 	do {
566 		boolean_t extending; /* if we are extending whole pages */
567 		u_quad_t oldsize;
568 		oldoff = uio->uio_offset;
569 		bytelen = uio->uio_resid;
570 
571 #ifndef NFS_V2_ONLY
572 		/*
573 		 * Check for a valid write lease.
574 		 */
575 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
576 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
577 			do {
578 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
579 			} while (error == NQNFS_EXPIRED);
580 			if (error)
581 				return (error);
582 			if (np->n_lrev != np->n_brev ||
583 			    (np->n_flag & NQNFSNONCACHE)) {
584 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
585 				if (error)
586 					return (error);
587 				np->n_brev = np->n_lrev;
588 			}
589 		}
590 #endif
591 		nfsstats.biocache_writes++;
592 
593 		oldsize = np->n_size;
594 		np->n_flag |= NMODIFIED;
595 		if (np->n_size < uio->uio_offset + bytelen) {
596 			np->n_size = uio->uio_offset + bytelen;
597 		}
598 		extending = ((uio->uio_offset & PAGE_MASK) == 0 &&
599 		    (bytelen & PAGE_MASK) == 0 &&
600 		    uio->uio_offset >= vp->v_size);
601 		win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen,
602 			    UBC_WRITE | (extending ? UBC_FAULTBUSY : 0));
603 		error = uiomove(win, bytelen, uio);
604 		flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
605 		ubc_release(win, flags);
606 		if (error) {
607 			if (extending) {
608 				/*
609 				 * backout size and free pages past eof.
610 				 */
611 				np->n_size = oldsize;
612 				simple_lock(&vp->v_interlock);
613 				(void)VOP_PUTPAGES(vp, round_page(vp->v_size),
614 				    0, PGO_SYNCIO | PGO_FREE);
615 			}
616 			break;
617 		}
618 		wrotedata = 1;
619 
620 		/*
621 		 * update UVM's notion of the size now that we've
622 		 * copied the data into the vnode's pages.
623 		 */
624 
625 		if (vp->v_size < uio->uio_offset) {
626 			uvm_vnp_setsize(vp, uio->uio_offset);
627 			extended = 1;
628 		}
629 
630 		if ((oldoff & ~(nmp->nm_wsize - 1)) !=
631 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
632 			simple_lock(&vp->v_interlock);
633 			error = VOP_PUTPAGES(vp,
634 			    trunc_page(oldoff & ~(nmp->nm_wsize - 1)),
635 			    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
636 				       ~(nmp->nm_wsize - 1)), PGO_CLEANIT);
637 		}
638 	} while (uio->uio_resid > 0);
639 	if (wrotedata)
640 		VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0));
641 	if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
642 		simple_lock(&vp->v_interlock);
643 		error = VOP_PUTPAGES(vp,
644 		    trunc_page(origoff & ~(nmp->nm_wsize - 1)),
645 		    round_page((uio->uio_offset + nmp->nm_wsize - 1) &
646 			       ~(nmp->nm_wsize - 1)),
647 		    PGO_CLEANIT | PGO_SYNCIO);
648 	}
649 	return error;
650 }
651 
652 /*
653  * Get an nfs cache block.
654  * Allocate a new one if the block isn't currently in the cache
655  * and return the block marked busy. If the calling process is
656  * interrupted by a signal for an interruptible mount point, return
657  * NULL.
658  */
659 struct buf *
660 nfs_getcacheblk(vp, bn, size, p)
661 	struct vnode *vp;
662 	daddr_t bn;
663 	int size;
664 	struct proc *p;
665 {
666 	struct buf *bp;
667 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
668 
669 	if (nmp->nm_flag & NFSMNT_INT) {
670 		bp = getblk(vp, bn, size, PCATCH, 0);
671 		while (bp == NULL) {
672 			if (nfs_sigintr(nmp, NULL, p))
673 				return (NULL);
674 			bp = getblk(vp, bn, size, 0, 2 * hz);
675 		}
676 	} else
677 		bp = getblk(vp, bn, size, 0, 0);
678 	return (bp);
679 }
680 
681 /*
682  * Flush and invalidate all dirty buffers. If another process is already
683  * doing the flush, just wait for completion.
684  */
685 int
686 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
687 	struct vnode *vp;
688 	int flags;
689 	struct ucred *cred;
690 	struct proc *p;
691 	int intrflg;
692 {
693 	struct nfsnode *np = VTONFS(vp);
694 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
695 	int error = 0, slpflag, slptimeo;
696 
697 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
698 		intrflg = 0;
699 	if (intrflg) {
700 		slpflag = PCATCH;
701 		slptimeo = 2 * hz;
702 	} else {
703 		slpflag = 0;
704 		slptimeo = 0;
705 	}
706 	/*
707 	 * First wait for any other process doing a flush to complete.
708 	 */
709 	simple_lock(&vp->v_interlock);
710 	while (np->n_flag & NFLUSHINPROG) {
711 		np->n_flag |= NFLUSHWANT;
712 		error = ltsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
713 			slptimeo, &vp->v_interlock);
714 		if (error && intrflg && nfs_sigintr(nmp, NULL, p)) {
715 			simple_unlock(&vp->v_interlock);
716 			return EINTR;
717 		}
718 	}
719 
720 	/*
721 	 * Now, flush as required.
722 	 */
723 	np->n_flag |= NFLUSHINPROG;
724 	simple_unlock(&vp->v_interlock);
725 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
726 	while (error) {
727 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
728 			error = EINTR;
729 			break;
730 		}
731 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
732 	}
733 	simple_lock(&vp->v_interlock);
734 	if (error == 0)
735 		np->n_flag &= ~NMODIFIED;
736 	np->n_flag &= ~NFLUSHINPROG;
737 	if (np->n_flag & NFLUSHWANT) {
738 		np->n_flag &= ~NFLUSHWANT;
739 		wakeup(&np->n_flag);
740 	}
741 	simple_unlock(&vp->v_interlock);
742 	return error;
743 }
744 
745 /*
746  * nfs_flushstalebuf: flush cache if it's stale.
747  *
748  * => caller shouldn't own any pages or buffers which belong to the vnode.
749  */
750 
751 int
752 nfs_flushstalebuf(struct vnode *vp, struct ucred *cred, struct proc *p,
753     int flags)
754 {
755 	struct nfsnode *np = VTONFS(vp);
756 	struct vattr vattr;
757 	int error;
758 
759 	if (np->n_flag & NMODIFIED) {
760 		if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0
761 		    || vp->v_type != VREG) {
762 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
763 			if (error)
764 				return error;
765 			if (vp->v_type == VDIR) {
766 				nfs_invaldircache(vp, 0);
767 			}
768 		} else {
769 			/*
770 			 * XXX assuming writes are ours.
771 			 */
772 		}
773 		NFS_INVALIDATE_ATTRCACHE(np);
774 		error = VOP_GETATTR(vp, &vattr, cred, p);
775 		if (error)
776 			return error;
777 		np->n_mtime = vattr.va_mtime;
778 	} else {
779 		error = VOP_GETATTR(vp, &vattr, cred, p);
780 		if (error)
781 			return error;
782 		if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) {
783 			if (vp->v_type == VDIR) {
784 				nfs_invaldircache(vp, 0);
785 			}
786 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
787 			if (error)
788 				return error;
789 			np->n_mtime = vattr.va_mtime;
790 		}
791 	}
792 
793 	return error;
794 }
795 
796 /*
797  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
798  * This is mainly to avoid queueing async I/O requests when the nfsiods
799  * are all hung on a dead server.
800  */
801 
802 int
803 nfs_asyncio(bp)
804 	struct buf *bp;
805 {
806 	int i;
807 	struct nfsmount *nmp;
808 	int gotiod, slpflag = 0, slptimeo = 0, error;
809 
810 	if (nfs_numasync == 0)
811 		return (EIO);
812 
813 	nmp = VFSTONFS(bp->b_vp->v_mount);
814 again:
815 	if (nmp->nm_flag & NFSMNT_INT)
816 		slpflag = PCATCH;
817 	gotiod = FALSE;
818 
819 	/*
820 	 * Find a free iod to process this request.
821 	 */
822 
823 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++) {
824 		struct nfs_iod *iod = &nfs_asyncdaemon[i];
825 
826 		simple_lock(&iod->nid_slock);
827 		if (iod->nid_want) {
828 			/*
829 			 * Found one, so wake it up and tell it which
830 			 * mount to process.
831 			 */
832 			iod->nid_want = NULL;
833 			iod->nid_mount = nmp;
834 			wakeup(&iod->nid_want);
835 			simple_lock(&nmp->nm_slock);
836 			simple_unlock(&iod->nid_slock);
837 			nmp->nm_bufqiods++;
838 			gotiod = TRUE;
839 			break;
840 		}
841 		simple_unlock(&iod->nid_slock);
842 	}
843 
844 	/*
845 	 * If none are free, we may already have an iod working on this mount
846 	 * point.  If so, it will process our request.
847 	 */
848 
849 	if (!gotiod) {
850 		simple_lock(&nmp->nm_slock);
851 		if (nmp->nm_bufqiods > 0)
852 			gotiod = TRUE;
853 	}
854 
855 	LOCK_ASSERT(simple_lock_held(&nmp->nm_slock));
856 
857 	/*
858 	 * If we have an iod which can process the request, then queue
859 	 * the buffer.  However, even if we have an iod, do not initiate
860 	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
861 	 * is via local loopback, we may put curproc (pagedaemon) to sleep
862 	 * waiting for the writes to complete. But the server (ourself)
863 	 * may block the write, waiting for its (ie., our) pagedaemon
864 	 * to produce clean pages to handle the write: deadlock.
865 	 * XXX: start non-loopback mounts straight away?  If "lots free",
866 	 * let pagedaemon start loopback writes anyway?
867 	 */
868 	if (gotiod) {
869 
870 		/*
871 		 * Ensure that the queue never grows too large.
872 		 */
873 		if (curproc == uvm.pagedaemon_proc) {
874 	  		/* Enque for later, to avoid free-page deadlock */
875 			  (void) 0;
876 		} else while (nmp->nm_bufqlen >= 2*nfs_numasync) {
877 			nmp->nm_bufqwant = TRUE;
878 			error = ltsleep(&nmp->nm_bufq,
879 			    slpflag | PRIBIO | PNORELOCK,
880 			    "nfsaio", slptimeo, &nmp->nm_slock);
881 			if (error) {
882 				if (nfs_sigintr(nmp, NULL, curproc))
883 					return (EINTR);
884 				if (slpflag == PCATCH) {
885 					slpflag = 0;
886 					slptimeo = 2 * hz;
887 				}
888 			}
889 
890 			/*
891 			 * We might have lost our iod while sleeping,
892 			 * so check and loop if nescessary.
893 			 */
894 
895 			if (nmp->nm_bufqiods == 0)
896 				goto again;
897 
898 			simple_lock(&nmp->nm_slock);
899 		}
900 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
901 		nmp->nm_bufqlen++;
902 		simple_unlock(&nmp->nm_slock);
903 		return (0);
904 	}
905 	simple_unlock(&nmp->nm_slock);
906 
907 	/*
908 	 * All the iods are busy on other mounts, so return EIO to
909 	 * force the caller to process the i/o synchronously.
910 	 */
911 
912 	return (EIO);
913 }
914 
915 /*
916  * nfs_doio for read.
917  */
918 static int
919 nfs_doio_read(bp, uiop)
920 	struct buf *bp;
921 	struct uio *uiop;
922 {
923 	struct vnode *vp = bp->b_vp;
924 	struct nfsnode *np = VTONFS(vp);
925 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
926 	int error = 0;
927 
928 	uiop->uio_rw = UIO_READ;
929 	switch (vp->v_type) {
930 	case VREG:
931 		nfsstats.read_bios++;
932 		error = nfs_readrpc(vp, uiop);
933 		if (!error && uiop->uio_resid) {
934 			int diff, len;
935 
936 			/*
937 			 * If uio_resid > 0, there is a hole in the file and
938 			 * no writes after the hole have been pushed to
939 			 * the server yet or the file has been truncated
940 			 * on the server.
941 			 * Just zero fill the rest of the valid area.
942 			 */
943 
944 			KASSERT(vp->v_size >=
945 			    uiop->uio_offset + uiop->uio_resid);
946 			diff = bp->b_bcount - uiop->uio_resid;
947 			len = uiop->uio_resid;
948 			memset((char *)bp->b_data + diff, 0, len);
949 		}
950 		if (uiop->uio_procp && (vp->v_flag & VTEXT) &&
951 		    (((nmp->nm_flag & NFSMNT_NQNFS) &&
952 		      NQNFS_CKINVALID(vp, np, ND_READ) &&
953 		      np->n_lrev != np->n_brev) ||
954 		     (!(nmp->nm_flag & NFSMNT_NQNFS) &&
955 		      timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)))) {
956 			uprintf("Process killed due to "
957 				"text file modification\n");
958 			psignal(uiop->uio_procp, SIGKILL);
959 #if 0 /* XXX NJWLWP */
960 			uiop->uio_procp->p_holdcnt++;
961 #endif
962 		}
963 		break;
964 	case VLNK:
965 		KASSERT(uiop->uio_offset == (off_t)0);
966 		nfsstats.readlink_bios++;
967 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
968 		break;
969 	case VDIR:
970 		nfsstats.readdir_bios++;
971 		uiop->uio_offset = bp->b_dcookie;
972 #ifndef NFS_V2_ONLY
973 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
974 			error = nfs_readdirplusrpc(vp, uiop, np->n_rcred);
975 			if (error == NFSERR_NOTSUPP)
976 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
977 		}
978 #else
979 		nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
980 #endif
981 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
982 			error = nfs_readdirrpc(vp, uiop, np->n_rcred);
983 		if (!error) {
984 			bp->b_dcookie = uiop->uio_offset;
985 		}
986 		break;
987 	default:
988 		printf("nfs_doio:  type %x unexpected\n", vp->v_type);
989 		break;
990 	}
991 	if (error) {
992 		bp->b_flags |= B_ERROR;
993 		bp->b_error = error;
994 	}
995 	return error;
996 }
997 
998 /*
999  * nfs_doio for write.
1000  */
1001 static int
1002 nfs_doio_write(bp, uiop)
1003 	struct buf *bp;
1004 	struct uio *uiop;
1005 {
1006 	struct vnode *vp = bp->b_vp;
1007 	struct nfsnode *np = VTONFS(vp);
1008 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1009 	int iomode;
1010 	boolean_t stalewriteverf = FALSE;
1011 	int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT;
1012 	struct vm_page *pgs[npages];
1013 #ifndef NFS_V2_ONLY
1014 	boolean_t needcommit = TRUE; /* need only COMMIT RPC */
1015 #else
1016 	boolean_t needcommit = FALSE; /* need only COMMIT RPC */
1017 #endif
1018 	boolean_t pageprotected;
1019 	struct uvm_object *uobj = &vp->v_uobj;
1020 	int error;
1021 	off_t off, cnt;
1022 
1023 	if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) {
1024 		iomode = NFSV3WRITE_UNSTABLE;
1025 	} else {
1026 		iomode = NFSV3WRITE_FILESYNC;
1027 	}
1028 
1029 #ifndef NFS_V2_ONLY
1030 again:
1031 #endif
1032 	lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1033 
1034 	for (i = 0; i < npages; i++) {
1035 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
1036 		if (pgs[i]->uobject == uobj &&
1037 		    pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) {
1038 			KASSERT(pgs[i]->flags & PG_BUSY);
1039 			/*
1040 			 * this page belongs to our object.
1041 			 */
1042 			simple_lock(&uobj->vmobjlock);
1043 			/*
1044 			 * write out the page stably if it's about to
1045 			 * be released because we can't resend it
1046 			 * on the server crash.
1047 			 *
1048 			 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be
1049 			 * changed until unbusy the page.
1050 			 */
1051 			if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT))
1052 				iomode = NFSV3WRITE_FILESYNC;
1053 			/*
1054 			 * if we met a page which hasn't been sent yet,
1055 			 * we need do WRITE RPC.
1056 			 */
1057 			if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0)
1058 				needcommit = FALSE;
1059 			simple_unlock(&uobj->vmobjlock);
1060 		} else {
1061 			iomode = NFSV3WRITE_FILESYNC;
1062 			needcommit = FALSE;
1063 		}
1064 	}
1065 	if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) {
1066 		simple_lock(&uobj->vmobjlock);
1067 		for (i = 0; i < npages; i++) {
1068 			pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY;
1069 			pmap_page_protect(pgs[i], VM_PROT_READ);
1070 		}
1071 		simple_unlock(&uobj->vmobjlock);
1072 		pageprotected = TRUE; /* pages can't be modified during i/o. */
1073 	} else
1074 		pageprotected = FALSE;
1075 
1076 	/*
1077 	 * Send the data to the server if necessary,
1078 	 * otherwise just send a commit rpc.
1079 	 */
1080 #ifndef NFS_V2_ONLY
1081 	if (needcommit) {
1082 
1083 		/*
1084 		 * If the buffer is in the range that we already committed,
1085 		 * there's nothing to do.
1086 		 *
1087 		 * If it's in the range that we need to commit, push the
1088 		 * whole range at once, otherwise only push the buffer.
1089 		 * In both these cases, acquire the commit lock to avoid
1090 		 * other processes modifying the range.
1091 		 */
1092 
1093 		off = uiop->uio_offset;
1094 		cnt = bp->b_bcount;
1095 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1096 		if (!nfs_in_committed_range(vp, off, bp->b_bcount)) {
1097 			boolean_t pushedrange;
1098 			if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) {
1099 				pushedrange = TRUE;
1100 				off = np->n_pushlo;
1101 				cnt = np->n_pushhi - np->n_pushlo;
1102 			} else {
1103 				pushedrange = FALSE;
1104 			}
1105 			error = nfs_commit(vp, off, cnt, curproc);
1106 			if (error == 0) {
1107 				if (pushedrange) {
1108 					nfs_merge_commit_ranges(vp);
1109 				} else {
1110 					nfs_add_committed_range(vp, off, cnt);
1111 				}
1112 			}
1113 		} else {
1114 			error = 0;
1115 		}
1116 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1117 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1118 		if (!error) {
1119 			/*
1120 			 * pages are now on stable storage.
1121 			 */
1122 			uiop->uio_resid = 0;
1123 			simple_lock(&uobj->vmobjlock);
1124 			for (i = 0; i < npages; i++) {
1125 				pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1126 			}
1127 			simple_unlock(&uobj->vmobjlock);
1128 			return 0;
1129 		} else if (error == NFSERR_STALEWRITEVERF) {
1130 			nfs_clearcommit(vp->v_mount);
1131 			goto again;
1132 		}
1133 		if (error) {
1134 			bp->b_flags |= B_ERROR;
1135 			bp->b_error = np->n_error = error;
1136 			np->n_flag |= NWRITEERR;
1137 		}
1138 		return error;
1139 	}
1140 #endif
1141 	off = uiop->uio_offset;
1142 	cnt = bp->b_bcount;
1143 	uiop->uio_rw = UIO_WRITE;
1144 	nfsstats.write_bios++;
1145 	error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf);
1146 #ifndef NFS_V2_ONLY
1147 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1148 		/*
1149 		 * we need to commit pages later.
1150 		 */
1151 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1152 		nfs_add_tobecommitted_range(vp, off, cnt);
1153 		/*
1154 		 * if there can be too many uncommitted pages, commit them now.
1155 		 */
1156 		if (np->n_pushhi - np->n_pushlo > nfs_commitsize) {
1157 			off = np->n_pushlo;
1158 			cnt = nfs_commitsize >> 1;
1159 			error = nfs_commit(vp, off, cnt, curproc);
1160 			if (!error) {
1161 				nfs_add_committed_range(vp, off, cnt);
1162 				nfs_del_tobecommitted_range(vp, off, cnt);
1163 			}
1164 			if (error == NFSERR_STALEWRITEVERF) {
1165 				stalewriteverf = TRUE;
1166 				error = 0; /* it isn't a real error */
1167 			}
1168 		} else {
1169 			/*
1170 			 * re-dirty pages so that they will be passed
1171 			 * to us later again.
1172 			 */
1173 			simple_lock(&uobj->vmobjlock);
1174 			for (i = 0; i < npages; i++) {
1175 				pgs[i]->flags &= ~PG_CLEAN;
1176 			}
1177 			simple_unlock(&uobj->vmobjlock);
1178 		}
1179 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1180 	} else
1181 #endif
1182 	if (!error) {
1183 		/*
1184 		 * pages are now on stable storage.
1185 		 */
1186 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1187 		nfs_del_committed_range(vp, off, cnt);
1188 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1189 		simple_lock(&uobj->vmobjlock);
1190 		for (i = 0; i < npages; i++) {
1191 			pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1192 		}
1193 		simple_unlock(&uobj->vmobjlock);
1194 	} else {
1195 		/*
1196 		 * we got an error.
1197 		 */
1198 		bp->b_flags |= B_ERROR;
1199 		bp->b_error = np->n_error = error;
1200 		np->n_flag |= NWRITEERR;
1201 	}
1202 
1203 	lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1204 
1205 	if (stalewriteverf) {
1206 		nfs_clearcommit(vp->v_mount);
1207 	}
1208 	return error;
1209 }
1210 
1211 /*
1212  * nfs_doio for B_PHYS.
1213  */
1214 static int
1215 nfs_doio_phys(bp, uiop)
1216 	struct buf *bp;
1217 	struct uio *uiop;
1218 {
1219 	struct vnode *vp = bp->b_vp;
1220 	int error;
1221 
1222 	uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
1223 	if (bp->b_flags & B_READ) {
1224 		uiop->uio_rw = UIO_READ;
1225 		nfsstats.read_physios++;
1226 		error = nfs_readrpc(vp, uiop);
1227 	} else {
1228 		int iomode = NFSV3WRITE_DATASYNC;
1229 		boolean_t stalewriteverf;
1230 		struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1231 
1232 		uiop->uio_rw = UIO_WRITE;
1233 		nfsstats.write_physios++;
1234 		lockmgr(&nmp->nm_writeverflock, LK_SHARED, NULL);
1235 		error = nfs_writerpc(vp, uiop, &iomode, FALSE, &stalewriteverf);
1236 		lockmgr(&nmp->nm_writeverflock, LK_RELEASE, NULL);
1237 		if (stalewriteverf) {
1238 			nfs_clearcommit(bp->b_vp->v_mount);
1239 		}
1240 	}
1241 	if (error) {
1242 		bp->b_flags |= B_ERROR;
1243 		bp->b_error = error;
1244 	}
1245 	return error;
1246 }
1247 
1248 /*
1249  * Do an I/O operation to/from a cache block. This may be called
1250  * synchronously or from an nfsiod.
1251  */
1252 int
1253 nfs_doio(bp, p)
1254 	struct buf *bp;
1255 	struct proc *p;
1256 {
1257 	int error;
1258 	struct uio uio;
1259 	struct uio *uiop = &uio;
1260 	struct iovec io;
1261 	UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist);
1262 
1263 	uiop->uio_iov = &io;
1264 	uiop->uio_iovcnt = 1;
1265 	uiop->uio_segflg = UIO_SYSSPACE;
1266 	uiop->uio_procp = NULL;
1267 	uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
1268 	io.iov_base = bp->b_data;
1269 	io.iov_len = uiop->uio_resid = bp->b_bcount;
1270 
1271 	/*
1272 	 * Historically, paging was done with physio, but no more...
1273 	 */
1274 	if (bp->b_flags & B_PHYS) {
1275 		/*
1276 		 * ...though reading /dev/drum still gets us here.
1277 		 */
1278 		error = nfs_doio_phys(bp, uiop);
1279 	} else if (bp->b_flags & B_READ) {
1280 		error = nfs_doio_read(bp, uiop);
1281 	} else {
1282 		error = nfs_doio_write(bp, uiop);
1283 	}
1284 	bp->b_resid = uiop->uio_resid;
1285 	biodone(bp);
1286 	return (error);
1287 }
1288 
1289 /*
1290  * Vnode op for VM getpages.
1291  */
1292 
1293 int
1294 nfs_getpages(v)
1295 	void *v;
1296 {
1297 	struct vop_getpages_args /* {
1298 		struct vnode *a_vp;
1299 		voff_t a_offset;
1300 		struct vm_page **a_m;
1301 		int *a_count;
1302 		int a_centeridx;
1303 		vm_prot_t a_access_type;
1304 		int a_advice;
1305 		int a_flags;
1306 	} */ *ap = v;
1307 
1308 	struct vnode *vp = ap->a_vp;
1309 	struct uvm_object *uobj = &vp->v_uobj;
1310 	struct nfsnode *np = VTONFS(vp);
1311 	const int npages = *ap->a_count;
1312 	struct vm_page *pg, **pgs, *opgs[npages];
1313 	off_t origoffset, len;
1314 	int i, error;
1315 	boolean_t v3 = NFS_ISV3(vp);
1316 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1317 	boolean_t locked = (ap->a_flags & PGO_LOCKED) != 0;
1318 
1319 	/*
1320 	 * call the genfs code to get the pages.  `pgs' may be NULL
1321 	 * when doing read-ahead.
1322 	 */
1323 
1324 	pgs = ap->a_m;
1325 	if (write && locked && v3) {
1326 		KASSERT(pgs != NULL);
1327 #ifdef DEBUG
1328 
1329 		/*
1330 		 * If PGO_LOCKED is set, real pages shouldn't exists
1331 		 * in the array.
1332 		 */
1333 
1334 		for (i = 0; i < npages; i++)
1335 			KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE);
1336 #endif
1337 		memcpy(opgs, pgs, npages * sizeof(struct vm_pages *));
1338 	}
1339 	error = genfs_getpages(v);
1340 	if (error) {
1341 		return (error);
1342 	}
1343 
1344 	/*
1345 	 * for read faults where the nfs node is not yet marked NMODIFIED,
1346 	 * set PG_RDONLY on the pages so that we come back here if someone
1347 	 * tries to modify later via the mapping that will be entered for
1348 	 * this fault.
1349 	 */
1350 
1351 	if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) {
1352 		if (!locked) {
1353 			simple_lock(&uobj->vmobjlock);
1354 		}
1355 		for (i = 0; i < npages; i++) {
1356 			pg = pgs[i];
1357 			if (pg == NULL || pg == PGO_DONTCARE) {
1358 				continue;
1359 			}
1360 			pg->flags |= PG_RDONLY;
1361 		}
1362 		if (!locked) {
1363 			simple_unlock(&uobj->vmobjlock);
1364 		}
1365 	}
1366 	if (!write) {
1367 		return (0);
1368 	}
1369 
1370 	/*
1371 	 * this is a write fault, update the commit info.
1372 	 */
1373 
1374 	origoffset = ap->a_offset;
1375 	len = npages << PAGE_SHIFT;
1376 
1377 	if (v3) {
1378 		error = lockmgr(&np->n_commitlock,
1379 		    LK_EXCLUSIVE | (locked ? LK_NOWAIT : 0), NULL);
1380 		if (error) {
1381 			KASSERT(locked != 0);
1382 
1383 			/*
1384 			 * Since PGO_LOCKED is set, we need to unbusy
1385 			 * all pages fetched by genfs_getpages() above,
1386 			 * tell the caller that there are no pages
1387 			 * available and put back original pgs array.
1388 			 */
1389 
1390 			uvm_lock_pageq();
1391 			uvm_page_unbusy(pgs, npages);
1392 			uvm_unlock_pageq();
1393 			*ap->a_count = 0;
1394 			memcpy(pgs, opgs,
1395 			    npages * sizeof(struct vm_pages *));
1396 			return (error);
1397 		}
1398 		nfs_del_committed_range(vp, origoffset, len);
1399 		nfs_del_tobecommitted_range(vp, origoffset, len);
1400 	}
1401 	np->n_flag |= NMODIFIED;
1402 	if (!locked) {
1403 		simple_lock(&uobj->vmobjlock);
1404 	}
1405 	for (i = 0; i < npages; i++) {
1406 		pg = pgs[i];
1407 		if (pg == NULL || pg == PGO_DONTCARE) {
1408 			continue;
1409 		}
1410 		pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY);
1411 	}
1412 	if (!locked) {
1413 		simple_unlock(&uobj->vmobjlock);
1414 	}
1415 	if (v3) {
1416 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1417 	}
1418 	return (0);
1419 }
1420