xref: /dragonfly/sys/vfs/nfs/nfs_bio.c (revision 0de61e28)
1 /*
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Rick Macklem at The University of Guelph.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
33  * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
34  */
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/uio.h>
39 #include <sys/resourcevar.h>
40 #include <sys/signalvar.h>
41 #include <sys/proc.h>
42 #include <sys/buf.h>
43 #include <sys/vnode.h>
44 #include <sys/mount.h>
45 #include <sys/kernel.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_pager.h>
54 #include <vm/vnode_pager.h>
55 
56 #include <sys/buf2.h>
57 #include <sys/thread2.h>
58 #include <vm/vm_page2.h>
59 
60 #include "rpcv2.h"
61 #include "nfsproto.h"
62 #include "nfs.h"
63 #include "nfsmount.h"
64 #include "nfsnode.h"
65 #include "xdr_subs.h"
66 #include "nfsm_subs.h"
67 
68 
69 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset,
70 				   int size, struct thread *td);
71 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen);
72 static void nfsiodone_sync(struct bio *bio);
73 static void nfs_readrpc_bio_done(nfsm_info_t info);
74 static void nfs_writerpc_bio_done(nfsm_info_t info);
75 static void nfs_commitrpc_bio_done(nfsm_info_t info);
76 
77 static __inline
78 void
79 nfs_knote(struct vnode *vp, int flags)
80 {
81 	if (flags)
82 		KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
83 }
84 
85 /*
86  * Vnode op for read using bio
87  */
88 int
89 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
90 {
91 	struct nfsnode *np = VTONFS(vp);
92 	int biosize, i;
93 	struct buf *bp, *rabp;
94 	struct vattr vattr;
95 	struct thread *td;
96 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
97 	off_t lbn, rabn;
98 	off_t raoffset;
99 	off_t loffset;
100 	int seqcount;
101 	int nra, error = 0;
102 	int boff = 0;
103 	size_t n;
104 
105 #ifdef DIAGNOSTIC
106 	if (uio->uio_rw != UIO_READ)
107 		panic("nfs_read mode");
108 #endif
109 	if (uio->uio_resid == 0)
110 		return (0);
111 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
112 		return (EINVAL);
113 	td = uio->uio_td;
114 
115 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
116 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
117 		(void)nfs_fsinfo(nmp, vp, td);
118 	if (vp->v_type != VDIR &&
119 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
120 		return (EFBIG);
121 	biosize = vp->v_mount->mnt_stat.f_iosize;
122 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / MAXBSIZE);
123 
124 	/*
125 	 * For nfs, cache consistency can only be maintained approximately.
126 	 * Although RFC1094 does not specify the criteria, the following is
127 	 * believed to be compatible with the reference port.
128 	 *
129 	 * NFS:		If local changes have been made and this is a
130 	 *		directory, the directory must be invalidated and
131 	 *		the attribute cache must be cleared.
132 	 *
133 	 *		GETATTR is called to synchronize the file size.  To
134 	 *		avoid a deadlock again the VM system, we cannot do
135 	 *		this for UIO_NOCOPY reads.
136 	 *
137 	 *		If remote changes are detected local data is flushed
138 	 *		and the cache is invalidated.
139 	 *
140 	 *		NOTE: In the normal case the attribute cache is not
141 	 *		cleared which means GETATTR may use cached data and
142 	 *		not immediately detect changes made on the server.
143 	 */
144 	if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
145 		nfs_invaldir(vp);
146 		error = nfs_vinvalbuf(vp, V_SAVE, 1);
147 		if (error)
148 			return (error);
149 		np->n_attrstamp = 0;
150 	}
151 
152 	/*
153 	 * Synchronize the file size when possible.  We can't do this without
154 	 * risking a deadlock if this is NOCOPY read from a vm_fault->getpages
155 	 * sequence.
156 	 */
157 	if (uio->uio_segflg != UIO_NOCOPY) {
158 		error = VOP_GETATTR(vp, &vattr);
159 		if (error)
160 			return (error);
161 	}
162 
163 	/*
164 	 * This can deadlock getpages/putpages for regular
165 	 * files.  Only do it for directories.
166 	 */
167 	if (np->n_flag & NRMODIFIED) {
168 		if (vp->v_type == VDIR) {
169 			nfs_invaldir(vp);
170 			error = nfs_vinvalbuf(vp, V_SAVE, 1);
171 			if (error)
172 				return (error);
173 			np->n_flag &= ~NRMODIFIED;
174 		}
175 	}
176 
177 	/*
178 	 * Loop until uio exhausted or we hit EOF
179 	 */
180 	do {
181 	    bp = NULL;
182 
183 	    switch (vp->v_type) {
184 	    case VREG:
185 		nfsstats.biocache_reads++;
186 		lbn = uio->uio_offset / biosize;
187 		boff = uio->uio_offset & (biosize - 1);
188 		loffset = lbn * biosize;
189 
190 		/*
191 		 * Start the read ahead(s), as required.
192 		 */
193 		if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) {
194 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
195 			(off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
196 			rabn = lbn + 1 + nra;
197 			raoffset = rabn * biosize;
198 			if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
199 			    rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
200 			    if (!rabp)
201 				return (EINTR);
202 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
203 				rabp->b_cmd = BUF_CMD_READ;
204 				vfs_busy_pages(vp, rabp);
205 				nfs_asyncio(vp, &rabp->b_bio2);
206 			    } else {
207 				brelse(rabp);
208 			    }
209 			}
210 		    }
211 		}
212 
213 		/*
214 		 * Obtain the buffer cache block.  Figure out the buffer size
215 		 * when we are at EOF.  If we are modifying the size of the
216 		 * buffer based on an EOF condition we need to hold
217 		 * nfs_rslock() through obtaining the buffer to prevent
218 		 * a potential writer-appender from messing with n_size.
219 		 * Otherwise we may accidently truncate the buffer and
220 		 * lose dirty data.
221 		 *
222 		 * Note that bcount is *not* DEV_BSIZE aligned.
223 		 */
224 		if (loffset + boff >= np->n_size) {
225 			n = 0;
226 			break;
227 		}
228 		bp = nfs_getcacheblk(vp, loffset, biosize, td);
229 
230 		if (bp == NULL)
231 			return (EINTR);
232 
233 		/*
234 		 * If B_CACHE is not set, we must issue the read.  If this
235 		 * fails, we return an error.
236 		 */
237 		if ((bp->b_flags & B_CACHE) == 0) {
238 			bp->b_cmd = BUF_CMD_READ;
239 			bp->b_bio2.bio_done = nfsiodone_sync;
240 			bp->b_bio2.bio_flags |= BIO_SYNC;
241 			vfs_busy_pages(vp, bp);
242 			error = nfs_doio(vp, &bp->b_bio2, td);
243 			if (error) {
244 				brelse(bp);
245 				return (error);
246 			}
247 		}
248 
249 		/*
250 		 * on is the offset into the current bp.  Figure out how many
251 		 * bytes we can copy out of the bp.  Note that bcount is
252 		 * NOT DEV_BSIZE aligned.
253 		 *
254 		 * Then figure out how many bytes we can copy into the uio.
255 		 */
256 		n = biosize - boff;
257 		if (n > uio->uio_resid)
258 			n = uio->uio_resid;
259 		if (loffset + boff + n > np->n_size)
260 			n = np->n_size - loffset - boff;
261 		break;
262 	    case VLNK:
263 		biosize = min(NFS_MAXPATHLEN, np->n_size);
264 		nfsstats.biocache_readlinks++;
265 		bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
266 		if (bp == NULL)
267 			return (EINTR);
268 		if ((bp->b_flags & B_CACHE) == 0) {
269 			bp->b_cmd = BUF_CMD_READ;
270 			bp->b_bio2.bio_done = nfsiodone_sync;
271 			bp->b_bio2.bio_flags |= BIO_SYNC;
272 			vfs_busy_pages(vp, bp);
273 			error = nfs_doio(vp, &bp->b_bio2, td);
274 			if (error) {
275 				bp->b_flags |= B_ERROR | B_INVAL;
276 				brelse(bp);
277 				return (error);
278 			}
279 		}
280 		n = szmin(uio->uio_resid, (size_t)bp->b_bcount - bp->b_resid);
281 		boff = 0;
282 		break;
283 	    case VDIR:
284 		nfsstats.biocache_readdirs++;
285 		if (np->n_direofoffset &&
286 		    uio->uio_offset >= np->n_direofoffset
287 		) {
288 			return (0);
289 		}
290 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
291 		boff = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
292 		loffset = uio->uio_offset - boff;
293 		bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
294 		if (bp == NULL)
295 			return (EINTR);
296 
297 		if ((bp->b_flags & B_CACHE) == 0) {
298 		    bp->b_cmd = BUF_CMD_READ;
299 		    bp->b_bio2.bio_done = nfsiodone_sync;
300 		    bp->b_bio2.bio_flags |= BIO_SYNC;
301 		    vfs_busy_pages(vp, bp);
302 		    error = nfs_doio(vp, &bp->b_bio2, td);
303 		    if (error)
304 			    brelse(bp);
305 		    while (error == NFSERR_BAD_COOKIE) {
306 			kprintf("got bad cookie vp %p bp %p\n", vp, bp);
307 			nfs_invaldir(vp);
308 			error = nfs_vinvalbuf(vp, 0, 1);
309 			/*
310 			 * Yuck! The directory has been modified on the
311 			 * server. The only way to get the block is by
312 			 * reading from the beginning to get all the
313 			 * offset cookies.
314 			 *
315 			 * Leave the last bp intact unless there is an error.
316 			 * Loop back up to the while if the error is another
317 			 * NFSERR_BAD_COOKIE (double yuch!).
318 			 */
319 			for (i = 0; i <= lbn && !error; i++) {
320 			    if (np->n_direofoffset
321 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
322 				    return (0);
323 			    bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
324 						 NFS_DIRBLKSIZ, td);
325 			    if (!bp)
326 				return (EINTR);
327 			    if ((bp->b_flags & B_CACHE) == 0) {
328 				    bp->b_cmd = BUF_CMD_READ;
329 				    bp->b_bio2.bio_done = nfsiodone_sync;
330 				    bp->b_bio2.bio_flags |= BIO_SYNC;
331 				    vfs_busy_pages(vp, bp);
332 				    error = nfs_doio(vp, &bp->b_bio2, td);
333 				    /*
334 				     * no error + B_INVAL == directory EOF,
335 				     * use the block.
336 				     */
337 				    if (error == 0 && (bp->b_flags & B_INVAL))
338 					    break;
339 			    }
340 			    /*
341 			     * An error will throw away the block and the
342 			     * for loop will break out.  If no error and this
343 			     * is not the block we want, we throw away the
344 			     * block and go for the next one via the for loop.
345 			     */
346 			    if (error || i < lbn)
347 				    brelse(bp);
348 			}
349 		    }
350 		    /*
351 		     * The above while is repeated if we hit another cookie
352 		     * error.  If we hit an error and it wasn't a cookie error,
353 		     * we give up.
354 		     */
355 		    if (error)
356 			    return (error);
357 		}
358 
359 		/*
360 		 * If not eof and read aheads are enabled, start one.
361 		 * (You need the current block first, so that you have the
362 		 *  directory offset cookie of the next block.)
363 		 */
364 		if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) &&
365 		    (bp->b_flags & B_INVAL) == 0 &&
366 		    (np->n_direofoffset == 0 ||
367 		    loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
368 		    findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
369 		) {
370 			rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
371 					       NFS_DIRBLKSIZ, td);
372 			if (rabp) {
373 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
374 				rabp->b_cmd = BUF_CMD_READ;
375 				vfs_busy_pages(vp, rabp);
376 				nfs_asyncio(vp, &rabp->b_bio2);
377 			    } else {
378 				brelse(rabp);
379 			    }
380 			}
381 		}
382 		/*
383 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
384 		 * chopped for the EOF condition, we cannot tell how large
385 		 * NFS directories are going to be until we hit EOF.  So
386 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
387 		 * it just so happens that b_resid will effectively chop it
388 		 * to EOF.  *BUT* this information is lost if the buffer goes
389 		 * away and is reconstituted into a B_CACHE state ( due to
390 		 * being VMIO ) later.  So we keep track of the directory eof
391 		 * in np->n_direofoffset and chop it off as an extra step
392 		 * right here.
393 		 *
394 		 * NOTE: boff could already be beyond EOF.
395 		 */
396 		if ((size_t)boff > NFS_DIRBLKSIZ - bp->b_resid) {
397 			n = 0;
398 		} else {
399 			n = szmin(uio->uio_resid,
400 				  NFS_DIRBLKSIZ - bp->b_resid - (size_t)boff);
401 		}
402 		if (np->n_direofoffset &&
403 		    n > (size_t)(np->n_direofoffset - uio->uio_offset)) {
404 			n = (size_t)(np->n_direofoffset - uio->uio_offset);
405 		}
406 		break;
407 	    default:
408 		kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
409 		n = 0;
410 		break;
411 	    }
412 
413 	    switch (vp->v_type) {
414 	    case VREG:
415 		if (n > 0)
416 		    error = uiomovebp(bp, bp->b_data + boff, n, uio);
417 		break;
418 	    case VLNK:
419 		if (n > 0)
420 		    error = uiomovebp(bp, bp->b_data + boff, n, uio);
421 		n = 0;
422 		break;
423 	    case VDIR:
424 		if (n > 0) {
425 		    off_t old_off = uio->uio_offset;
426 		    caddr_t cpos, epos;
427 		    struct nfs_dirent *dp;
428 
429 		    /*
430 		     * We are casting cpos to nfs_dirent, it must be
431 		     * int-aligned.
432 		     */
433 		    if (boff & 3) {
434 			error = EINVAL;
435 			break;
436 		    }
437 
438 		    cpos = bp->b_data + boff;
439 		    epos = bp->b_data + boff + n;
440 		    while (cpos < epos && error == 0 && uio->uio_resid > 0) {
441 			    dp = (struct nfs_dirent *)cpos;
442 			    error = nfs_check_dirent(dp, (int)(epos - cpos));
443 			    if (error)
444 				    break;
445 			    if (vop_write_dirent(&error, uio, dp->nfs_ino,
446 				dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
447 				    break;
448 			    }
449 			    cpos += dp->nfs_reclen;
450 		    }
451 		    n = 0;
452 		    if (error == 0) {
453 			    uio->uio_offset = old_off + cpos -
454 					      bp->b_data - boff;
455 		    }
456 		}
457 		break;
458 	    default:
459 		kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
460 	    }
461 	    if (bp)
462 		    brelse(bp);
463 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
464 	return (error);
465 }
466 
467 /*
468  * Userland can supply any 'seek' offset when reading a NFS directory.
469  * Validate the structure so we don't panic the kernel.  Note that
470  * the element name is nul terminated and the nul is not included
471  * in nfs_namlen.
472  */
473 static
474 int
475 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
476 {
477 	int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
478 
479 	if (nfs_name_off >= maxlen)
480 		return (EINVAL);
481 	if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
482 		return (EINVAL);
483 	if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
484 		return (EINVAL);
485 	if (dp->nfs_reclen & 3)
486 		return (EINVAL);
487 	return (0);
488 }
489 
490 /*
491  * Vnode op for write using bio
492  *
493  * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
494  *	     struct ucred *a_cred)
495  */
496 int
497 nfs_write(struct vop_write_args *ap)
498 {
499 	struct uio *uio = ap->a_uio;
500 	struct thread *td = uio->uio_td;
501 	struct vnode *vp = ap->a_vp;
502 	struct nfsnode *np = VTONFS(vp);
503 	int ioflag = ap->a_ioflag;
504 	struct buf *bp;
505 	struct vattr vattr;
506 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
507 	off_t loffset;
508 	int boff, bytes;
509 	int error = 0;
510 	int haverslock = 0;
511 	int bcount;
512 	int biosize;
513 	int trivial;
514 	int kflags = 0;
515 
516 #ifdef DIAGNOSTIC
517 	if (uio->uio_rw != UIO_WRITE)
518 		panic("nfs_write mode");
519 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
520 		panic("nfs_write proc");
521 #endif
522 	if (vp->v_type != VREG)
523 		return (EIO);
524 
525 	lwkt_gettoken(&nmp->nm_token);
526 
527 	if (np->n_flag & NWRITEERR) {
528 		np->n_flag &= ~NWRITEERR;
529 		lwkt_reltoken(&nmp->nm_token);
530 		return (np->n_error);
531 	}
532 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
533 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
534 		(void)nfs_fsinfo(nmp, vp, td);
535 	}
536 
537 	/*
538 	 * Synchronously flush pending buffers if we are in synchronous
539 	 * mode or if we are appending.
540 	 */
541 	if (ioflag & (IO_APPEND | IO_SYNC)) {
542 		if (np->n_flag & NLMODIFIED) {
543 			np->n_attrstamp = 0;
544 			error = nfs_flush(vp, MNT_WAIT, td, 0);
545 			/* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
546 			if (error)
547 				goto  done;
548 		}
549 	}
550 
551 	/*
552 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
553 	 * get the append lock.
554 	 */
555 restart:
556 	if (ioflag & IO_APPEND) {
557 		np->n_attrstamp = 0;
558 		error = VOP_GETATTR(vp, &vattr);
559 		if (error)
560 			goto done;
561 		uio->uio_offset = np->n_size;
562 	}
563 
564 	if (uio->uio_offset < 0) {
565 		error = EINVAL;
566 		goto done;
567 	}
568 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) {
569 		error = EFBIG;
570 		goto done;
571 	}
572 	if (uio->uio_resid == 0) {
573 		error = 0;
574 		goto done;
575 	}
576 
577 	/*
578 	 * We need to obtain the rslock if we intend to modify np->n_size
579 	 * in order to guarentee the append point with multiple contending
580 	 * writers, to guarentee that no other appenders modify n_size
581 	 * while we are trying to obtain a truncated buffer (i.e. to avoid
582 	 * accidently truncating data written by another appender due to
583 	 * the race), and to ensure that the buffer is populated prior to
584 	 * our extending of the file.  We hold rslock through the entire
585 	 * operation.
586 	 *
587 	 * Note that we do not synchronize the case where someone truncates
588 	 * the file while we are appending to it because attempting to lock
589 	 * this case may deadlock other parts of the system unexpectedly.
590 	 */
591 	if ((ioflag & IO_APPEND) ||
592 	    uio->uio_offset + uio->uio_resid > np->n_size) {
593 		switch(nfs_rslock(np)) {
594 		case ENOLCK:
595 			goto restart;
596 			/* not reached */
597 		case EINTR:
598 		case ERESTART:
599 			error = EINTR;
600 			goto done;
601 			/* not reached */
602 		default:
603 			break;
604 		}
605 		haverslock = 1;
606 	}
607 
608 	/*
609 	 * Maybe this should be above the vnode op call, but so long as
610 	 * file servers have no limits, i don't think it matters
611 	 */
612 	if (td && td->td_proc && uio->uio_offset + uio->uio_resid >
613 	      td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
614 		lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
615 		if (haverslock)
616 			nfs_rsunlock(np);
617 		error = EFBIG;
618 		goto done;
619 	}
620 
621 	biosize = vp->v_mount->mnt_stat.f_iosize;
622 
623 	do {
624 		nfsstats.biocache_writes++;
625 		boff = uio->uio_offset & (biosize-1);
626 		loffset = uio->uio_offset - boff;
627 		bytes = (int)szmin((unsigned)(biosize - boff), uio->uio_resid);
628 again:
629 		/*
630 		 * Handle direct append and file extension cases, calculate
631 		 * unaligned buffer size.  When extending B_CACHE will be
632 		 * set if possible.  See UIO_NOCOPY note below.
633 		 */
634 		if (uio->uio_offset + bytes > np->n_size) {
635 			np->n_flag |= NLMODIFIED;
636 			trivial = (uio->uio_segflg != UIO_NOCOPY &&
637 				   uio->uio_offset <= np->n_size) ?
638 				  NVEXTF_TRIVIAL : 0;
639 			nfs_meta_setsize(vp, td, uio->uio_offset + bytes,
640 					 trivial);
641 			kflags |= NOTE_EXTEND;
642 		}
643 		bp = nfs_getcacheblk(vp, loffset, biosize, td);
644 		if (bp == NULL) {
645 			error = EINTR;
646 			break;
647 		}
648 
649 		/*
650 		 * Actual bytes in buffer which we care about
651 		 */
652 		if (loffset + biosize < np->n_size)
653 			bcount = biosize;
654 		else
655 			bcount = (int)(np->n_size - loffset);
656 
657 		/*
658 		 * Avoid a read by setting B_CACHE where the data we
659 		 * intend to write covers the entire buffer.  Note
660 		 * that the buffer may have been set to B_CACHE by
661 		 * nfs_meta_setsize() above or otherwise inherited the
662 		 * flag, but if B_CACHE isn't set the buffer may be
663 		 * uninitialized and must be zero'd to accomodate
664 		 * future seek+write's.
665 		 *
666 		 * See the comments in kern/vfs_bio.c's getblk() for
667 		 * more information.
668 		 *
669 		 * When doing a UIO_NOCOPY write the buffer is not
670 		 * overwritten and we cannot just set B_CACHE unconditionally
671 		 * for full-block writes.
672 		 */
673 		if (boff == 0 && bytes == biosize &&
674 		    uio->uio_segflg != UIO_NOCOPY) {
675 			bp->b_flags |= B_CACHE;
676 			bp->b_flags &= ~(B_ERROR | B_INVAL);
677 		}
678 
679 		/*
680 		 * b_resid may be set due to file EOF if we extended out.
681 		 * The NFS bio code will zero the difference anyway so
682 		 * just acknowledged the fact and set b_resid to 0.
683 		 */
684 		if ((bp->b_flags & B_CACHE) == 0) {
685 			bp->b_cmd = BUF_CMD_READ;
686 			bp->b_bio2.bio_done = nfsiodone_sync;
687 			bp->b_bio2.bio_flags |= BIO_SYNC;
688 			vfs_busy_pages(vp, bp);
689 			error = nfs_doio(vp, &bp->b_bio2, td);
690 			if (error) {
691 				brelse(bp);
692 				break;
693 			}
694 			bp->b_resid = 0;
695 		}
696 		np->n_flag |= NLMODIFIED;
697 		kflags |= NOTE_WRITE;
698 
699 		/*
700 		 * If dirtyend exceeds file size, chop it down.  This should
701 		 * not normally occur but there is an append race where it
702 		 * might occur XXX, so we log it.
703 		 *
704 		 * If the chopping creates a reverse-indexed or degenerate
705 		 * situation with dirtyoff/end, we 0 both of them.
706 		 */
707 		if (bp->b_dirtyend > bcount) {
708 			kprintf("NFS append race @%08llx:%d\n",
709 			    (long long)bp->b_bio2.bio_offset,
710 			    bp->b_dirtyend - bcount);
711 			bp->b_dirtyend = bcount;
712 		}
713 
714 		if (bp->b_dirtyoff >= bp->b_dirtyend)
715 			bp->b_dirtyoff = bp->b_dirtyend = 0;
716 
717 		/*
718 		 * If the new write will leave a contiguous dirty
719 		 * area, just update the b_dirtyoff and b_dirtyend,
720 		 * otherwise force a write rpc of the old dirty area.
721 		 *
722 		 * While it is possible to merge discontiguous writes due to
723 		 * our having a B_CACHE buffer ( and thus valid read data
724 		 * for the hole), we don't because it could lead to
725 		 * significant cache coherency problems with multiple clients,
726 		 * especially if locking is implemented later on.
727 		 *
728 		 * as an optimization we could theoretically maintain
729 		 * a linked list of discontinuous areas, but we would still
730 		 * have to commit them separately so there isn't much
731 		 * advantage to it except perhaps a bit of asynchronization.
732 		 */
733 		if (bp->b_dirtyend > 0 &&
734 		    (boff > bp->b_dirtyend ||
735 		     (boff + bytes) < bp->b_dirtyoff)
736 		) {
737 			if (bwrite(bp) == EINTR) {
738 				error = EINTR;
739 				break;
740 			}
741 			goto again;
742 		}
743 
744 		error = uiomovebp(bp, bp->b_data + boff, bytes, uio);
745 
746 		/*
747 		 * Since this block is being modified, it must be written
748 		 * again and not just committed.  Since write clustering does
749 		 * not work for the stage 1 data write, only the stage 2
750 		 * commit rpc, we have to clear B_CLUSTEROK as well.
751 		 */
752 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
753 
754 		if (error) {
755 			brelse(bp);
756 			break;
757 		}
758 
759 		/*
760 		 * Only update dirtyoff/dirtyend if not a degenerate
761 		 * condition.
762 		 *
763 		 * The underlying VM pages have been marked valid by
764 		 * virtue of acquiring the bp.  Because the entire buffer
765 		 * is marked dirty we do not have to worry about cleaning
766 		 * out the related dirty bits (and wouldn't really know
767 		 * how to deal with byte ranges anyway)
768 		 */
769 		if (bytes) {
770 			if (bp->b_dirtyend > 0) {
771 				bp->b_dirtyoff = imin(boff, bp->b_dirtyoff);
772 				bp->b_dirtyend = imax(boff + bytes,
773 						      bp->b_dirtyend);
774 			} else {
775 				bp->b_dirtyoff = boff;
776 				bp->b_dirtyend = boff + bytes;
777 			}
778 		}
779 
780 		/*
781 		 * If the lease is non-cachable or IO_SYNC do bwrite().
782 		 *
783 		 * IO_INVAL appears to be unused.  The idea appears to be
784 		 * to turn off caching in this case.  Very odd.  XXX
785 		 *
786 		 * If nfs_async is set bawrite() will use an unstable write
787 		 * (build dirty bufs on the server), so we might as well
788 		 * push it out with bawrite().  If nfs_async is not set we
789 		 * use bdwrite() to cache dirty bufs on the client.
790 		 */
791 		if (ioflag & IO_SYNC) {
792 			if (ioflag & IO_INVAL)
793 				bp->b_flags |= B_NOCACHE;
794 			error = bwrite(bp);
795 			if (error)
796 				break;
797 		} else if (boff + bytes == biosize && nfs_async) {
798 			bawrite(bp);
799 		} else {
800 			bdwrite(bp);
801 		}
802 	} while (uio->uio_resid > 0 && bytes > 0);
803 
804 	if (haverslock)
805 		nfs_rsunlock(np);
806 
807 done:
808 	nfs_knote(vp, kflags);
809 	lwkt_reltoken(&nmp->nm_token);
810 	return (error);
811 }
812 
813 /*
814  * Get an nfs cache block.
815  *
816  * Allocate a new one if the block isn't currently in the cache
817  * and return the block marked busy. If the calling process is
818  * interrupted by a signal for an interruptible mount point, return
819  * NULL.
820  *
821  * The caller must carefully deal with the possible B_INVAL state of
822  * the buffer.  nfs_startio() clears B_INVAL (and nfs_asyncio() clears it
823  * indirectly), so synchronous reads can be issued without worrying about
824  * the B_INVAL state.  We have to be a little more careful when dealing
825  * with writes (see comments in nfs_write()) when extending a file past
826  * its EOF.
827  */
828 static struct buf *
829 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
830 {
831 	struct buf *bp;
832 	struct mount *mp;
833 	struct nfsmount *nmp;
834 
835 	mp = vp->v_mount;
836 	nmp = VFSTONFS(mp);
837 
838 	if (nmp->nm_flag & NFSMNT_INT) {
839 		bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
840 		while (bp == NULL) {
841 			if (nfs_sigintr(nmp, NULL, td))
842 				return (NULL);
843 			bp = getblk(vp, loffset, size, 0, 2 * hz);
844 		}
845 	} else {
846 		bp = getblk(vp, loffset, size, 0, 0);
847 	}
848 
849 	/*
850 	 * bio2, the 'device' layer.  Since BIOs use 64 bit byte offsets
851 	 * now, no translation is necessary.
852 	 */
853 	bp->b_bio2.bio_offset = loffset;
854 	return (bp);
855 }
856 
857 /*
858  * Flush and invalidate all dirty buffers. If another process is already
859  * doing the flush, just wait for completion.
860  */
861 int
862 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
863 {
864 	struct nfsnode *np = VTONFS(vp);
865 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
866 	int error = 0, slpflag, slptimeo;
867 	thread_t td = curthread;
868 
869 	if (vp->v_flag & VRECLAIMED)
870 		return (0);
871 
872 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
873 		intrflg = 0;
874 	if (intrflg) {
875 		slpflag = PCATCH;
876 		slptimeo = 2 * hz;
877 	} else {
878 		slpflag = 0;
879 		slptimeo = 0;
880 	}
881 	/*
882 	 * First wait for any other process doing a flush to complete.
883 	 */
884 	while (np->n_flag & NFLUSHINPROG) {
885 		np->n_flag |= NFLUSHWANT;
886 		error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
887 		if (error && intrflg && nfs_sigintr(nmp, NULL, td))
888 			return (EINTR);
889 	}
890 
891 	/*
892 	 * Now, flush as required.
893 	 */
894 	np->n_flag |= NFLUSHINPROG;
895 	error = vinvalbuf(vp, flags, slpflag, 0);
896 	while (error) {
897 		if (intrflg && nfs_sigintr(nmp, NULL, td)) {
898 			np->n_flag &= ~NFLUSHINPROG;
899 			if (np->n_flag & NFLUSHWANT) {
900 				np->n_flag &= ~NFLUSHWANT;
901 				wakeup((caddr_t)&np->n_flag);
902 			}
903 			return (EINTR);
904 		}
905 		error = vinvalbuf(vp, flags, 0, slptimeo);
906 	}
907 	np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
908 	if (np->n_flag & NFLUSHWANT) {
909 		np->n_flag &= ~NFLUSHWANT;
910 		wakeup((caddr_t)&np->n_flag);
911 	}
912 	return (0);
913 }
914 
915 /*
916  * Return true (non-zero) if the txthread and rxthread are operational
917  * and we do not already have too many not-yet-started BIO's built up.
918  */
919 int
920 nfs_asyncok(struct nfsmount *nmp)
921 {
922 	return (nmp->nm_bioqlen < nfs_maxasyncbio &&
923 		nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE &&
924 		nmp->nm_rxstate <= NFSSVC_PENDING &&
925 		nmp->nm_txstate <= NFSSVC_PENDING);
926 }
927 
928 /*
929  * The read-ahead code calls this to queue a bio to the txthread.
930  *
931  * We don't touch the bio otherwise... that is, we do not even
932  * construct or send the initial rpc.  The txthread will do it
933  * for us.
934  *
935  * NOTE!  nm_bioqlen is not decremented until the request completes,
936  *	  so it does not reflect the number of bio's on bioq.
937  */
938 void
939 nfs_asyncio(struct vnode *vp, struct bio *bio)
940 {
941 	struct buf *bp = bio->bio_buf;
942 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
943 
944 	KKASSERT(vp->v_tag == VT_NFS);
945 	BUF_KERNPROC(bp);
946 
947 	/*
948 	 * Shortcut swap cache (not done automatically because we are not
949 	 * using bread()).
950 	 */
951 	if (vn_cache_strategy(vp, bio))
952 		return;
953 
954 	bio->bio_driver_info = vp;
955 	crit_enter();
956 	TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
957 	atomic_add_int(&nmp->nm_bioqlen, 1);
958 	crit_exit();
959 	nfssvc_iod_writer_wakeup(nmp);
960 }
961 
962 /*
963  * nfs_doio()	- Execute a BIO operation synchronously.  The BIO will be
964  *		  completed and its error returned.  The caller is responsible
965  *		  for brelse()ing it.  ONLY USE FOR BIO_SYNC IOs!  Otherwise
966  *		  our error probe will be against an invalid pointer.
967  *
968  * nfs_startio()- Execute a BIO operation assynchronously.
969  *
970  * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation,
971  *	 which basically just queues it to the txthread.  nfs_startio()
972  *	 actually initiates the I/O AFTER it has gotten to the txthread.
973  *
974  * NOTE: td might be NULL.
975  *
976  * NOTE: Caller has already busied the I/O.
977  */
978 void
979 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td)
980 {
981 	struct buf *bp = bio->bio_buf;
982 
983 	KKASSERT(vp->v_tag == VT_NFS);
984 
985 	/*
986 	 * clear B_ERROR and B_INVAL state prior to initiating the I/O.  We
987 	 * do this here so we do not have to do it in all the code that
988 	 * calls us.
989 	 */
990 	bp->b_flags &= ~(B_ERROR | B_INVAL);
991 
992 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
993 		("nfs_doio: bp %p already marked done!", bp));
994 
995 	if (bp->b_cmd == BUF_CMD_READ) {
996 	    switch (vp->v_type) {
997 	    case VREG:
998 		nfsstats.read_bios++;
999 		nfs_readrpc_bio(vp, bio);
1000 		break;
1001 	    case VLNK:
1002 #if 0
1003 		bio->bio_offset = 0;
1004 		nfsstats.readlink_bios++;
1005 		nfs_readlinkrpc_bio(vp, bio);
1006 #else
1007 		nfs_doio(vp, bio, td);
1008 #endif
1009 		break;
1010 	    case VDIR:
1011 		/*
1012 		 * NOTE: If nfs_readdirplusrpc_bio() is requested but
1013 		 *	 not supported, it will chain to
1014 		 *	 nfs_readdirrpc_bio().
1015 		 */
1016 #if 0
1017 		nfsstats.readdir_bios++;
1018 		uiop->uio_offset = bio->bio_offset;
1019 		if (nmp->nm_flag & NFSMNT_RDIRPLUS)
1020 			nfs_readdirplusrpc_bio(vp, bio);
1021 		else
1022 			nfs_readdirrpc_bio(vp, bio);
1023 #else
1024 		nfs_doio(vp, bio, td);
1025 #endif
1026 		break;
1027 	    default:
1028 		kprintf("nfs_doio:  type %x unexpected\n",vp->v_type);
1029 		bp->b_flags |= B_ERROR;
1030 		bp->b_error = EINVAL;
1031 		biodone(bio);
1032 		break;
1033 	    }
1034 	} else {
1035 	    /*
1036 	     * If we only need to commit, try to commit.  If this fails
1037 	     * it will chain through to the write.  Basically all the logic
1038 	     * in nfs_doio() is replicated.
1039 	     */
1040 	    KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1041 	    if (bp->b_flags & B_NEEDCOMMIT)
1042 		nfs_commitrpc_bio(vp, bio);
1043 	    else
1044 		nfs_writerpc_bio(vp, bio);
1045 	}
1046 }
1047 
1048 int
1049 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1050 {
1051 	struct buf *bp = bio->bio_buf;
1052 	struct uio *uiop;
1053 	struct nfsnode *np;
1054 	struct nfsmount *nmp;
1055 	int error = 0;
1056 	int iomode, must_commit;
1057 	size_t n;
1058 	struct uio uio;
1059 	struct iovec io;
1060 
1061 #if 0
1062 	/*
1063 	 * Shortcut swap cache (not done automatically because we are not
1064 	 * using bread()).
1065 	 *
1066 	 * XXX The biowait is a hack until we can figure out how to stop a
1067 	 * biodone chain when a middle element is BIO_SYNC.  BIO_SYNC is
1068 	 * set so the bp shouldn't get ripped out from under us.  The only
1069 	 * use-cases are fully synchronous I/O cases.
1070 	 *
1071 	 * XXX This is having problems, give up for now.
1072 	 */
1073 	if (vn_cache_strategy(vp, bio)) {
1074 		error = biowait(&bio->bio_buf->b_bio1, "nfsrsw");
1075 		return (error);
1076 	}
1077 #endif
1078 
1079 	KKASSERT(vp->v_tag == VT_NFS);
1080 	np = VTONFS(vp);
1081 	nmp = VFSTONFS(vp->v_mount);
1082 	uiop = &uio;
1083 	uiop->uio_iov = &io;
1084 	uiop->uio_iovcnt = 1;
1085 	uiop->uio_segflg = UIO_SYSSPACE;
1086 	uiop->uio_td = td;
1087 
1088 	/*
1089 	 * clear B_ERROR and B_INVAL state prior to initiating the I/O.  We
1090 	 * do this here so we do not have to do it in all the code that
1091 	 * calls us.
1092 	 */
1093 	bp->b_flags &= ~(B_ERROR | B_INVAL);
1094 
1095 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
1096 		("nfs_doio: bp %p already marked done!", bp));
1097 
1098 	if (bp->b_cmd == BUF_CMD_READ) {
1099 	    io.iov_len = uiop->uio_resid = (size_t)bp->b_bcount;
1100 	    io.iov_base = bp->b_data;
1101 	    uiop->uio_rw = UIO_READ;
1102 
1103 	    switch (vp->v_type) {
1104 	    case VREG:
1105 		/*
1106 		 * When reading from a regular file zero-fill any residual.
1107 		 * Note that this residual has nothing to do with NFS short
1108 		 * reads, which nfs_readrpc_uio() will handle for us.
1109 		 *
1110 		 * We have to do this because when we are write extending
1111 		 * a file the server may not have the same notion of
1112 		 * filesize as we do.  Our BIOs should already be sized
1113 		 * (b_bcount) to account for the file EOF.
1114 		 */
1115 		nfsstats.read_bios++;
1116 		uiop->uio_offset = bio->bio_offset;
1117 		error = nfs_readrpc_uio(vp, uiop);
1118 		if (error == 0 && uiop->uio_resid) {
1119 			n = (size_t)bp->b_bcount - uiop->uio_resid;
1120 			bzero(bp->b_data + n, bp->b_bcount - n);
1121 			uiop->uio_resid = 0;
1122 		}
1123 		if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1124 		    np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1125 			uprintf("Process killed due to text file modification\n");
1126 			ksignal(td->td_proc, SIGKILL);
1127 		}
1128 		break;
1129 	    case VLNK:
1130 		uiop->uio_offset = 0;
1131 		nfsstats.readlink_bios++;
1132 		error = nfs_readlinkrpc_uio(vp, uiop);
1133 		break;
1134 	    case VDIR:
1135 		nfsstats.readdir_bios++;
1136 		uiop->uio_offset = bio->bio_offset;
1137 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1138 			error = nfs_readdirplusrpc_uio(vp, uiop);
1139 			if (error == NFSERR_NOTSUPP)
1140 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1141 		}
1142 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1143 			error = nfs_readdirrpc_uio(vp, uiop);
1144 		/*
1145 		 * end-of-directory sets B_INVAL but does not generate an
1146 		 * error.
1147 		 */
1148 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1149 			bp->b_flags |= B_INVAL;
1150 		break;
1151 	    default:
1152 		kprintf("nfs_doio:  type %x unexpected\n",vp->v_type);
1153 		break;
1154 	    }
1155 	    if (error) {
1156 		bp->b_flags |= B_ERROR;
1157 		bp->b_error = error;
1158 	    }
1159 	    bp->b_resid = uiop->uio_resid;
1160 	} else {
1161 	    /*
1162 	     * If we only need to commit, try to commit.
1163 	     *
1164 	     * NOTE: The I/O has already been staged for the write and
1165 	     *	     its pages busied, so b_dirtyoff/end is valid.
1166 	     */
1167 	    KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1168 	    if (bp->b_flags & B_NEEDCOMMIT) {
1169 		    int retv;
1170 		    off_t off;
1171 
1172 		    off = bio->bio_offset + bp->b_dirtyoff;
1173 		    retv = nfs_commitrpc_uio(vp, off,
1174 					     bp->b_dirtyend - bp->b_dirtyoff,
1175 					     td);
1176 		    if (retv == 0) {
1177 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1178 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1179 			    bp->b_resid = 0;
1180 			    biodone(bio);
1181 			    return(0);
1182 		    }
1183 		    if (retv == NFSERR_STALEWRITEVERF) {
1184 			    nfs_clearcommit(vp->v_mount);
1185 		    }
1186 	    }
1187 
1188 	    /*
1189 	     * Setup for actual write
1190 	     */
1191 	    if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1192 		bp->b_dirtyend = np->n_size - bio->bio_offset;
1193 
1194 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1195 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1196 		    - bp->b_dirtyoff;
1197 		uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1198 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1199 		uiop->uio_rw = UIO_WRITE;
1200 		nfsstats.write_bios++;
1201 
1202 		if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1203 		    iomode = NFSV3WRITE_UNSTABLE;
1204 		else
1205 		    iomode = NFSV3WRITE_FILESYNC;
1206 
1207 		must_commit = 0;
1208 		error = nfs_writerpc_uio(vp, uiop, &iomode, &must_commit);
1209 
1210 		/*
1211 		 * We no longer try to use kern/vfs_bio's cluster code to
1212 		 * cluster commits, so B_CLUSTEROK is no longer set with
1213 		 * B_NEEDCOMMIT.  The problem is that a vfs_busy_pages()
1214 		 * may have to clear B_NEEDCOMMIT if it finds underlying
1215 		 * pages have been redirtied through a memory mapping
1216 		 * and doing this on a clustered bp will probably cause
1217 		 * a panic, plus the flag in the underlying NFS bufs
1218 		 * making up the cluster bp will not be properly cleared.
1219 		 */
1220 		if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1221 		    bp->b_flags |= B_NEEDCOMMIT;
1222 #if 0
1223 		    /* XXX do not enable commit clustering */
1224 		    if (bp->b_dirtyoff == 0
1225 			&& bp->b_dirtyend == bp->b_bcount)
1226 			bp->b_flags |= B_CLUSTEROK;
1227 #endif
1228 		} else {
1229 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1230 		}
1231 
1232 		/*
1233 		 * For an interrupted write, the buffer is still valid
1234 		 * and the write hasn't been pushed to the server yet,
1235 		 * so we can't set B_ERROR and report the interruption
1236 		 * by setting B_EINTR. For the async case, B_EINTR
1237 		 * is not relevant, so the rpc attempt is essentially
1238 		 * a noop.  For the case of a V3 write rpc not being
1239 		 * committed to stable storage, the block is still
1240 		 * dirty and requires either a commit rpc or another
1241 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1242 		 * the block is reused. This is indicated by setting
1243 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1244 		 *
1245 		 * If the buffer is marked B_PAGING, it does not reside on
1246 		 * the vp's paging queues so we cannot call bdirty().  The
1247 		 * bp in this case is not an NFS cache block so we should
1248 		 * be safe. XXX
1249 		 */
1250     		if (error == EINTR
1251 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1252 			crit_enter();
1253 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1254 			if ((bp->b_flags & B_PAGING) == 0)
1255 			    bdirty(bp);
1256 			if (error)
1257 			    bp->b_flags |= B_EINTR;
1258 			crit_exit();
1259 	    	} else {
1260 		    if (error) {
1261 			bp->b_flags |= B_ERROR;
1262 			bp->b_error = np->n_error = error;
1263 			np->n_flag |= NWRITEERR;
1264 		    }
1265 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1266 		}
1267 		if (must_commit)
1268 		    nfs_clearcommit(vp->v_mount);
1269 		bp->b_resid = uiop->uio_resid;
1270 	    } else {
1271 		bp->b_resid = 0;
1272 	    }
1273 	}
1274 
1275 	/*
1276 	 * I/O was run synchronously, biodone() it and calculate the
1277 	 * error to return.
1278 	 */
1279 	biodone(bio);
1280 	KKASSERT(bp->b_cmd == BUF_CMD_DONE);
1281 	if (bp->b_flags & B_EINTR)
1282 		return (EINTR);
1283 	if (bp->b_flags & B_ERROR)
1284 		return (bp->b_error ? bp->b_error : EIO);
1285 	return (0);
1286 }
1287 
1288 /*
1289  * Handle all truncation, write-extend, and ftruncate()-extend operations
1290  * on the NFS lcient side.
1291  *
1292  * We use the new API in kern/vfs_vm.c to perform these operations in a
1293  * VM-friendly way.  With this API VM pages are properly zerod and pages
1294  * still mapped into the buffer straddling EOF are not invalidated.
1295  */
1296 int
1297 nfs_meta_setsize(struct vnode *vp, struct thread *td, off_t nsize, int flags)
1298 {
1299 	struct nfsnode *np = VTONFS(vp);
1300 	off_t osize;
1301 	int biosize = vp->v_mount->mnt_stat.f_iosize;
1302 	int error;
1303 
1304 	osize = np->n_size;
1305 	np->n_size = nsize;
1306 
1307 	if (nsize < osize) {
1308 		error = nvtruncbuf(vp, nsize, biosize, -1, flags);
1309 	} else {
1310 		error = nvextendbuf(vp, osize, nsize,
1311 				    biosize, biosize, -1, -1, flags);
1312 	}
1313 	return(error);
1314 }
1315 
1316 /*
1317  * Synchronous completion for nfs_doio.  Call bpdone() with elseit=FALSE.
1318  * Caller is responsible for brelse()'ing the bp.
1319  */
1320 static void
1321 nfsiodone_sync(struct bio *bio)
1322 {
1323 	bio->bio_flags = 0;
1324 	bpdone(bio->bio_buf, 0);
1325 }
1326 
1327 /*
1328  * nfs read rpc - BIO version
1329  */
1330 void
1331 nfs_readrpc_bio(struct vnode *vp, struct bio *bio)
1332 {
1333 	struct buf *bp = bio->bio_buf;
1334 	u_int32_t *tl;
1335 	struct nfsmount *nmp;
1336 	int error = 0, len, tsiz;
1337 	struct nfsm_info *info;
1338 
1339 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1340 	info->mrep = NULL;
1341 	info->v3 = NFS_ISV3(vp);
1342 
1343 	nmp = VFSTONFS(vp->v_mount);
1344 	tsiz = bp->b_bcount;
1345 	KKASSERT(tsiz <= nmp->nm_rsize);
1346 	if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) {
1347 		error = EFBIG;
1348 		goto nfsmout;
1349 	}
1350 	nfsstats.rpccnt[NFSPROC_READ]++;
1351 	len = tsiz;
1352 	nfsm_reqhead(info, vp, NFSPROC_READ,
1353 		     NFSX_FH(info->v3) + NFSX_UNSIGNED * 3);
1354 	ERROROUT(nfsm_fhtom(info, vp));
1355 	tl = nfsm_build(info, NFSX_UNSIGNED * 3);
1356 	if (info->v3) {
1357 		txdr_hyper(bio->bio_offset, tl);
1358 		*(tl + 2) = txdr_unsigned(len);
1359 	} else {
1360 		*tl++ = txdr_unsigned(bio->bio_offset);
1361 		*tl++ = txdr_unsigned(len);
1362 		*tl = 0;
1363 	}
1364 	info->bio = bio;
1365 	info->done = nfs_readrpc_bio_done;
1366 	nfsm_request_bio(info, vp, NFSPROC_READ, NULL,
1367 			 nfs_vpcred(vp, ND_READ));
1368 	return;
1369 nfsmout:
1370 	kfree(info, M_NFSREQ);
1371 	bp->b_error = error;
1372 	bp->b_flags |= B_ERROR;
1373 	biodone(bio);
1374 }
1375 
1376 static void
1377 nfs_readrpc_bio_done(nfsm_info_t info)
1378 {
1379 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1380 	struct bio *bio = info->bio;
1381 	struct buf *bp = bio->bio_buf;
1382 	u_int32_t *tl;
1383 	int attrflag;
1384 	int retlen;
1385 	int eof;
1386 	int error = 0;
1387 
1388 	KKASSERT(info->state == NFSM_STATE_DONE);
1389 
1390 	lwkt_gettoken(&nmp->nm_token);
1391 
1392 	ERROROUT(info->error);
1393 	if (info->v3) {
1394 		ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag,
1395 					 NFS_LATTR_NOSHRINK));
1396 		NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED));
1397 		eof = fxdr_unsigned(int, *(tl + 1));
1398 	} else {
1399 		ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1400 		eof = 0;
1401 	}
1402 	NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize));
1403 	ERROROUT(nfsm_mtobio(info, bio, retlen));
1404 	m_freem(info->mrep);
1405 	info->mrep = NULL;
1406 
1407 	/*
1408 	 * No error occured, if retlen is less then bcount and no EOF
1409 	 * and NFSv3 a zero-fill short read occured.
1410 	 *
1411 	 * For NFSv2 a short-read indicates EOF.
1412 	 */
1413 	if (retlen < bp->b_bcount && info->v3 && eof == 0) {
1414 		bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1415 		retlen = bp->b_bcount;
1416 	}
1417 
1418 	/*
1419 	 * If we hit an EOF we still zero-fill, but return the expected
1420 	 * b_resid anyway.  This should normally not occur since async
1421 	 * BIOs are not used for read-before-write case.  Races against
1422 	 * the server can cause it though and we don't want to leave
1423 	 * garbage in the buffer.
1424 	 */
1425 	if (retlen < bp->b_bcount) {
1426 		bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1427 	}
1428 	bp->b_resid = 0;
1429 	/* bp->b_resid = bp->b_bcount - retlen; */
1430 nfsmout:
1431 	lwkt_reltoken(&nmp->nm_token);
1432 	kfree(info, M_NFSREQ);
1433 	if (error) {
1434 		bp->b_error = error;
1435 		bp->b_flags |= B_ERROR;
1436 	}
1437 	biodone(bio);
1438 }
1439 
1440 /*
1441  * nfs write call - BIO version
1442  *
1443  * NOTE: Caller has already busied the I/O.
1444  */
1445 void
1446 nfs_writerpc_bio(struct vnode *vp, struct bio *bio)
1447 {
1448 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1449 	struct nfsnode *np = VTONFS(vp);
1450 	struct buf *bp = bio->bio_buf;
1451 	u_int32_t *tl;
1452 	int len;
1453 	int iomode;
1454 	int error = 0;
1455 	struct nfsm_info *info;
1456 	off_t offset;
1457 
1458 	/*
1459 	 * Setup for actual write.  Just clean up the bio if there
1460 	 * is nothing to do.  b_dirtyoff/end have already been staged
1461 	 * by the bp's pages getting busied.
1462 	 */
1463 	if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1464 		bp->b_dirtyend = np->n_size - bio->bio_offset;
1465 
1466 	if (bp->b_dirtyend <= bp->b_dirtyoff) {
1467 		bp->b_resid = 0;
1468 		biodone(bio);
1469 		return;
1470 	}
1471 	len = bp->b_dirtyend - bp->b_dirtyoff;
1472 	offset = bio->bio_offset + bp->b_dirtyoff;
1473 	if (offset + len > nmp->nm_maxfilesize) {
1474 		bp->b_flags |= B_ERROR;
1475 		bp->b_error = EFBIG;
1476 		biodone(bio);
1477 		return;
1478 	}
1479 	bp->b_resid = len;
1480 	nfsstats.write_bios++;
1481 
1482 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1483 	info->mrep = NULL;
1484 	info->v3 = NFS_ISV3(vp);
1485 	info->info_writerpc.must_commit = 0;
1486 	if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1487 		iomode = NFSV3WRITE_UNSTABLE;
1488 	else
1489 		iomode = NFSV3WRITE_FILESYNC;
1490 
1491 	KKASSERT(len <= nmp->nm_wsize);
1492 
1493 	nfsstats.rpccnt[NFSPROC_WRITE]++;
1494 	nfsm_reqhead(info, vp, NFSPROC_WRITE,
1495 		     NFSX_FH(info->v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len));
1496 	ERROROUT(nfsm_fhtom(info, vp));
1497 	if (info->v3) {
1498 		tl = nfsm_build(info, 5 * NFSX_UNSIGNED);
1499 		txdr_hyper(offset, tl);
1500 		tl += 2;
1501 		*tl++ = txdr_unsigned(len);
1502 		*tl++ = txdr_unsigned(iomode);
1503 		*tl = txdr_unsigned(len);
1504 	} else {
1505 		u_int32_t x;
1506 
1507 		tl = nfsm_build(info, 4 * NFSX_UNSIGNED);
1508 		/* Set both "begin" and "current" to non-garbage. */
1509 		x = txdr_unsigned((u_int32_t)offset);
1510 		*tl++ = x;	/* "begin offset" */
1511 		*tl++ = x;	/* "current offset" */
1512 		x = txdr_unsigned(len);
1513 		*tl++ = x;	/* total to this offset */
1514 		*tl = x;	/* size of this write */
1515 	}
1516 	ERROROUT(nfsm_biotom(info, bio, bp->b_dirtyoff, len));
1517 	info->bio = bio;
1518 	info->done = nfs_writerpc_bio_done;
1519 	nfsm_request_bio(info, vp, NFSPROC_WRITE, NULL,
1520 			 nfs_vpcred(vp, ND_WRITE));
1521 	return;
1522 nfsmout:
1523 	kfree(info, M_NFSREQ);
1524 	bp->b_error = error;
1525 	bp->b_flags |= B_ERROR;
1526 	biodone(bio);
1527 }
1528 
1529 static void
1530 nfs_writerpc_bio_done(nfsm_info_t info)
1531 {
1532 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1533 	struct nfsnode *np = VTONFS(info->vp);
1534 	struct bio *bio = info->bio;
1535 	struct buf *bp = bio->bio_buf;
1536 	int wccflag = NFSV3_WCCRATTR;
1537 	int iomode = NFSV3WRITE_FILESYNC;
1538 	int commit;
1539 	int rlen;
1540 	int error;
1541 	int len = bp->b_resid;	/* b_resid was set to shortened length */
1542 	u_int32_t *tl;
1543 
1544 	lwkt_gettoken(&nmp->nm_token);
1545 
1546 	ERROROUT(info->error);
1547 	if (info->v3) {
1548 		/*
1549 		 * The write RPC returns a before and after mtime.  The
1550 		 * nfsm_wcc_data() macro checks the before n_mtime
1551 		 * against the before time and stores the after time
1552 		 * in the nfsnode's cached vattr and n_mtime field.
1553 		 * The NRMODIFIED bit will be set if the before
1554 		 * time did not match the original mtime.
1555 		 */
1556 		wccflag = NFSV3_WCCCHK;
1557 		ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1558 		if (error == 0) {
1559 			NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF));
1560 			rlen = fxdr_unsigned(int, *tl++);
1561 			if (rlen == 0) {
1562 				error = NFSERR_IO;
1563 				m_freem(info->mrep);
1564 				info->mrep = NULL;
1565 				goto nfsmout;
1566 			} else if (rlen < len) {
1567 #if 0
1568 				/*
1569 				 * XXX what do we do here?
1570 				 */
1571 				backup = len - rlen;
1572 				uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup;
1573 				uiop->uio_iov->iov_len += backup;
1574 				uiop->uio_offset -= backup;
1575 				uiop->uio_resid += backup;
1576 				len = rlen;
1577 #endif
1578 			}
1579 			commit = fxdr_unsigned(int, *tl++);
1580 
1581 			/*
1582 			 * Return the lowest committment level
1583 			 * obtained by any of the RPCs.
1584 			 */
1585 			if (iomode == NFSV3WRITE_FILESYNC)
1586 				iomode = commit;
1587 			else if (iomode == NFSV3WRITE_DATASYNC &&
1588 				commit == NFSV3WRITE_UNSTABLE)
1589 				iomode = commit;
1590 			if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){
1591 			    bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1592 			    nmp->nm_state |= NFSSTA_HASWRITEVERF;
1593 			} else if (bcmp(tl, nmp->nm_verf, NFSX_V3WRITEVERF)) {
1594 			    info->info_writerpc.must_commit = 1;
1595 			    bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1596 			}
1597 		}
1598 	} else {
1599 		ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1600 	}
1601 	m_freem(info->mrep);
1602 	info->mrep = NULL;
1603 	len = 0;
1604 nfsmout:
1605 	if (info->vp->v_mount->mnt_flag & MNT_ASYNC)
1606 		iomode = NFSV3WRITE_FILESYNC;
1607 	bp->b_resid = len;
1608 
1609 	/*
1610 	 * End of RPC.  Now clean up the bp.
1611 	 *
1612 	 * We no longer enable write clustering for commit operations,
1613 	 * See around line 1157 for a more detailed comment.
1614 	 */
1615 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1616 		bp->b_flags |= B_NEEDCOMMIT;
1617 #if 0
1618 		/* XXX do not enable commit clustering */
1619 		if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount)
1620 			bp->b_flags |= B_CLUSTEROK;
1621 #endif
1622 	} else {
1623 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1624 	}
1625 
1626 	/*
1627 	 * For an interrupted write, the buffer is still valid
1628 	 * and the write hasn't been pushed to the server yet,
1629 	 * so we can't set B_ERROR and report the interruption
1630 	 * by setting B_EINTR. For the async case, B_EINTR
1631 	 * is not relevant, so the rpc attempt is essentially
1632 	 * a noop.  For the case of a V3 write rpc not being
1633 	 * committed to stable storage, the block is still
1634 	 * dirty and requires either a commit rpc or another
1635 	 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1636 	 * the block is reused. This is indicated by setting
1637 	 * the B_DELWRI and B_NEEDCOMMIT flags.
1638 	 *
1639 	 * If the buffer is marked B_PAGING, it does not reside on
1640 	 * the vp's paging queues so we cannot call bdirty().  The
1641 	 * bp in this case is not an NFS cache block so we should
1642 	 * be safe. XXX
1643 	 */
1644 	if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1645 		crit_enter();
1646 		bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1647 		if ((bp->b_flags & B_PAGING) == 0)
1648 			bdirty(bp);
1649 		if (error)
1650 			bp->b_flags |= B_EINTR;
1651 		crit_exit();
1652 	} else {
1653 		if (error) {
1654 			bp->b_flags |= B_ERROR;
1655 			bp->b_error = np->n_error = error;
1656 			np->n_flag |= NWRITEERR;
1657 		}
1658 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1659 	}
1660 	if (info->info_writerpc.must_commit)
1661 		nfs_clearcommit(info->vp->v_mount);
1662 	lwkt_reltoken(&nmp->nm_token);
1663 
1664 	kfree(info, M_NFSREQ);
1665 	if (error) {
1666 		bp->b_flags |= B_ERROR;
1667 		bp->b_error = error;
1668 	}
1669 	biodone(bio);
1670 }
1671 
1672 /*
1673  * Nfs Version 3 commit rpc - BIO version
1674  *
1675  * This function issues the commit rpc and will chain to a write
1676  * rpc if necessary.
1677  */
1678 void
1679 nfs_commitrpc_bio(struct vnode *vp, struct bio *bio)
1680 {
1681 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1682 	struct buf *bp = bio->bio_buf;
1683 	struct nfsm_info *info;
1684 	int error = 0;
1685 	u_int32_t *tl;
1686 
1687 	if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) {
1688 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1689 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1690 		bp->b_resid = 0;
1691 		biodone(bio);
1692 		return;
1693 	}
1694 
1695 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1696 	info->mrep = NULL;
1697 	info->v3 = 1;
1698 
1699 	nfsstats.rpccnt[NFSPROC_COMMIT]++;
1700 	nfsm_reqhead(info, vp, NFSPROC_COMMIT, NFSX_FH(1));
1701 	ERROROUT(nfsm_fhtom(info, vp));
1702 	tl = nfsm_build(info, 3 * NFSX_UNSIGNED);
1703 	txdr_hyper(bio->bio_offset + bp->b_dirtyoff, tl);
1704 	tl += 2;
1705 	*tl = txdr_unsigned(bp->b_dirtyend - bp->b_dirtyoff);
1706 	info->bio = bio;
1707 	info->done = nfs_commitrpc_bio_done;
1708 	nfsm_request_bio(info, vp, NFSPROC_COMMIT, NULL,
1709 			 nfs_vpcred(vp, ND_WRITE));
1710 	return;
1711 nfsmout:
1712 	/*
1713 	 * Chain to write RPC on (early) error
1714 	 */
1715 	kfree(info, M_NFSREQ);
1716 	nfs_writerpc_bio(vp, bio);
1717 }
1718 
1719 static void
1720 nfs_commitrpc_bio_done(nfsm_info_t info)
1721 {
1722 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1723 	struct bio *bio = info->bio;
1724 	struct buf *bp = bio->bio_buf;
1725 	u_int32_t *tl;
1726 	int wccflag = NFSV3_WCCRATTR;
1727 	int error = 0;
1728 
1729 	lwkt_gettoken(&nmp->nm_token);
1730 
1731 	ERROROUT(info->error);
1732 	ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1733 	if (error == 0) {
1734 		NULLOUT(tl = nfsm_dissect(info, NFSX_V3WRITEVERF));
1735 		if (bcmp(nmp->nm_verf, tl, NFSX_V3WRITEVERF)) {
1736 			bcopy(tl, nmp->nm_verf, NFSX_V3WRITEVERF);
1737 			error = NFSERR_STALEWRITEVERF;
1738 		}
1739 	}
1740 	m_freem(info->mrep);
1741 	info->mrep = NULL;
1742 
1743 	/*
1744 	 * On completion we must chain to a write bio if an
1745 	 * error occurred.
1746 	 */
1747 nfsmout:
1748 	if (error == 0) {
1749 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1750 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1751 		bp->b_resid = 0;
1752 		biodone(bio);
1753 	} else {
1754 		nfs_writerpc_bio(info->vp, bio);
1755 	}
1756 	kfree(info, M_NFSREQ);
1757 	lwkt_reltoken(&nmp->nm_token);
1758 }
1759 
1760