xref: /dragonfly/sys/vfs/nfs/nfs_bio.c (revision ae24b5e0)
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 
37 #include <sys/param.h>
38 #include <sys/systm.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/mbuf.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_extern.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_object.h>
52 #include <vm/vm_pager.h>
53 #include <vm/vnode_pager.h>
54 
55 #include <sys/buf2.h>
56 #include <sys/thread2.h>
57 #include <vm/vm_page2.h>
58 
59 #include "rpcv2.h"
60 #include "nfsproto.h"
61 #include "nfs.h"
62 #include "nfsmount.h"
63 #include "nfsnode.h"
64 #include "xdr_subs.h"
65 #include "nfsm_subs.h"
66 
67 
68 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset,
69 				   int size, struct thread *td);
70 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen);
71 static void nfsiodone_sync(struct bio *bio);
72 static void nfs_readrpc_bio_done(nfsm_info_t info);
73 static void nfs_writerpc_bio_done(nfsm_info_t info);
74 static void nfs_commitrpc_bio_done(nfsm_info_t info);
75 
76 static __inline
77 void
78 nfs_knote(struct vnode *vp, int flags)
79 {
80 	if (flags)
81 		KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
82 }
83 
84 /*
85  * Vnode op for read using bio
86  */
87 int
88 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
89 {
90 	struct nfsnode *np = VTONFS(vp);
91 	int biosize, i;
92 	struct buf *bp, *rabp;
93 	struct vattr vattr;
94 	struct thread *td;
95 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
96 	off_t lbn, rabn;
97 	off_t raoffset;
98 	off_t loffset;
99 	int seqcount;
100 	int nra, error = 0;
101 	int boff = 0;
102 	size_t n;
103 
104 #ifdef DIAGNOSTIC
105 	if (uio->uio_rw != UIO_READ)
106 		panic("nfs_read mode");
107 #endif
108 	if (uio->uio_resid == 0)
109 		return (0);
110 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
111 		return (EINVAL);
112 	td = uio->uio_td;
113 
114 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
115 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
116 		(void)nfs_fsinfo(nmp, vp, td);
117 	if (vp->v_type != VDIR &&
118 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
119 		return (EFBIG);
120 	biosize = vp->v_mount->mnt_stat.f_iosize;
121 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / MAXBSIZE);
122 
123 	/*
124 	 * For nfs, cache consistency can only be maintained approximately.
125 	 * Although RFC1094 does not specify the criteria, the following is
126 	 * believed to be compatible with the reference port.
127 	 *
128 	 * NFS:		If local changes have been made and this is a
129 	 *		directory, the directory must be invalidated and
130 	 *		the attribute cache must be cleared.
131 	 *
132 	 *		GETATTR is called to synchronize the file size.
133 	 *
134 	 *		If remote changes are detected local data is flushed
135 	 *		and the cache is invalidated.
136 	 *
137 	 *		NOTE: In the normal case the attribute cache is not
138 	 *		cleared which means GETATTR may use cached data and
139 	 *		not immediately detect changes made on the server.
140 	 */
141 	if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
142 		nfs_invaldir(vp);
143 		error = nfs_vinvalbuf(vp, V_SAVE, 1);
144 		if (error)
145 			return (error);
146 		np->n_attrstamp = 0;
147 	}
148 	error = VOP_GETATTR(vp, &vattr);
149 	if (error)
150 		return (error);
151 
152 	/*
153 	 * This can deadlock getpages/putpages for regular
154 	 * files.  Only do it for directories.
155 	 */
156 	if (np->n_flag & NRMODIFIED) {
157 		if (vp->v_type == VDIR) {
158 			nfs_invaldir(vp);
159 			error = nfs_vinvalbuf(vp, V_SAVE, 1);
160 			if (error)
161 				return (error);
162 			np->n_flag &= ~NRMODIFIED;
163 		}
164 	}
165 
166 	/*
167 	 * Loop until uio exhausted or we hit EOF
168 	 */
169 	do {
170 	    bp = NULL;
171 
172 	    switch (vp->v_type) {
173 	    case VREG:
174 		nfsstats.biocache_reads++;
175 		lbn = uio->uio_offset / biosize;
176 		boff = uio->uio_offset & (biosize - 1);
177 		loffset = lbn * biosize;
178 
179 		/*
180 		 * Start the read ahead(s), as required.
181 		 */
182 		if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) {
183 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
184 			(off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
185 			rabn = lbn + 1 + nra;
186 			raoffset = rabn * biosize;
187 			if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
188 			    rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
189 			    if (!rabp)
190 				return (EINTR);
191 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
192 				rabp->b_cmd = BUF_CMD_READ;
193 				vfs_busy_pages(vp, rabp);
194 				nfs_asyncio(vp, &rabp->b_bio2);
195 			    } else {
196 				brelse(rabp);
197 			    }
198 			}
199 		    }
200 		}
201 
202 		/*
203 		 * Obtain the buffer cache block.  Figure out the buffer size
204 		 * when we are at EOF.  If we are modifying the size of the
205 		 * buffer based on an EOF condition we need to hold
206 		 * nfs_rslock() through obtaining the buffer to prevent
207 		 * a potential writer-appender from messing with n_size.
208 		 * Otherwise we may accidently truncate the buffer and
209 		 * lose dirty data.
210 		 *
211 		 * Note that bcount is *not* DEV_BSIZE aligned.
212 		 */
213 		if (loffset + boff >= np->n_size) {
214 			n = 0;
215 			break;
216 		}
217 		bp = nfs_getcacheblk(vp, loffset, biosize, td);
218 
219 		if (bp == NULL)
220 			return (EINTR);
221 
222 		/*
223 		 * If B_CACHE is not set, we must issue the read.  If this
224 		 * fails, we return an error.
225 		 */
226 		if ((bp->b_flags & B_CACHE) == 0) {
227 			bp->b_cmd = BUF_CMD_READ;
228 			bp->b_bio2.bio_done = nfsiodone_sync;
229 			bp->b_bio2.bio_flags |= BIO_SYNC;
230 			vfs_busy_pages(vp, bp);
231 			error = nfs_doio(vp, &bp->b_bio2, td);
232 			if (error) {
233 				brelse(bp);
234 				return (error);
235 			}
236 		}
237 
238 		/*
239 		 * on is the offset into the current bp.  Figure out how many
240 		 * bytes we can copy out of the bp.  Note that bcount is
241 		 * NOT DEV_BSIZE aligned.
242 		 *
243 		 * Then figure out how many bytes we can copy into the uio.
244 		 */
245 		n = biosize - boff;
246 		if (n > uio->uio_resid)
247 			n = uio->uio_resid;
248 		if (loffset + boff + n > np->n_size)
249 			n = np->n_size - loffset - boff;
250 		break;
251 	    case VLNK:
252 		biosize = min(NFS_MAXPATHLEN, np->n_size);
253 		nfsstats.biocache_readlinks++;
254 		bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
255 		if (bp == NULL)
256 			return (EINTR);
257 		if ((bp->b_flags & B_CACHE) == 0) {
258 			bp->b_cmd = BUF_CMD_READ;
259 			bp->b_bio2.bio_done = nfsiodone_sync;
260 			bp->b_bio2.bio_flags |= BIO_SYNC;
261 			vfs_busy_pages(vp, bp);
262 			error = nfs_doio(vp, &bp->b_bio2, td);
263 			if (error) {
264 				bp->b_flags |= B_ERROR | B_INVAL;
265 				brelse(bp);
266 				return (error);
267 			}
268 		}
269 		n = szmin(uio->uio_resid, (size_t)bp->b_bcount - bp->b_resid);
270 		boff = 0;
271 		break;
272 	    case VDIR:
273 		nfsstats.biocache_readdirs++;
274 		if (np->n_direofoffset &&
275 		    uio->uio_offset >= np->n_direofoffset
276 		) {
277 			return (0);
278 		}
279 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
280 		boff = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
281 		loffset = uio->uio_offset - boff;
282 		bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
283 		if (bp == NULL)
284 			return (EINTR);
285 
286 		if ((bp->b_flags & B_CACHE) == 0) {
287 		    bp->b_cmd = BUF_CMD_READ;
288 		    bp->b_bio2.bio_done = nfsiodone_sync;
289 		    bp->b_bio2.bio_flags |= BIO_SYNC;
290 		    vfs_busy_pages(vp, bp);
291 		    error = nfs_doio(vp, &bp->b_bio2, td);
292 		    if (error)
293 			    brelse(bp);
294 		    while (error == NFSERR_BAD_COOKIE) {
295 			kprintf("got bad cookie vp %p bp %p\n", vp, bp);
296 			nfs_invaldir(vp);
297 			error = nfs_vinvalbuf(vp, 0, 1);
298 			/*
299 			 * Yuck! The directory has been modified on the
300 			 * server. The only way to get the block is by
301 			 * reading from the beginning to get all the
302 			 * offset cookies.
303 			 *
304 			 * Leave the last bp intact unless there is an error.
305 			 * Loop back up to the while if the error is another
306 			 * NFSERR_BAD_COOKIE (double yuch!).
307 			 */
308 			for (i = 0; i <= lbn && !error; i++) {
309 			    if (np->n_direofoffset
310 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
311 				    return (0);
312 			    bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
313 						 NFS_DIRBLKSIZ, td);
314 			    if (!bp)
315 				return (EINTR);
316 			    if ((bp->b_flags & B_CACHE) == 0) {
317 				    bp->b_cmd = BUF_CMD_READ;
318 				    bp->b_bio2.bio_done = nfsiodone_sync;
319 				    bp->b_bio2.bio_flags |= BIO_SYNC;
320 				    vfs_busy_pages(vp, bp);
321 				    error = nfs_doio(vp, &bp->b_bio2, td);
322 				    /*
323 				     * no error + B_INVAL == directory EOF,
324 				     * use the block.
325 				     */
326 				    if (error == 0 && (bp->b_flags & B_INVAL))
327 					    break;
328 			    }
329 			    /*
330 			     * An error will throw away the block and the
331 			     * for loop will break out.  If no error and this
332 			     * is not the block we want, we throw away the
333 			     * block and go for the next one via the for loop.
334 			     */
335 			    if (error || i < lbn)
336 				    brelse(bp);
337 			}
338 		    }
339 		    /*
340 		     * The above while is repeated if we hit another cookie
341 		     * error.  If we hit an error and it wasn't a cookie error,
342 		     * we give up.
343 		     */
344 		    if (error)
345 			    return (error);
346 		}
347 
348 		/*
349 		 * If not eof and read aheads are enabled, start one.
350 		 * (You need the current block first, so that you have the
351 		 *  directory offset cookie of the next block.)
352 		 */
353 		if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) &&
354 		    (bp->b_flags & B_INVAL) == 0 &&
355 		    (np->n_direofoffset == 0 ||
356 		    loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
357 		    findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
358 		) {
359 			rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
360 					       NFS_DIRBLKSIZ, td);
361 			if (rabp) {
362 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
363 				rabp->b_cmd = BUF_CMD_READ;
364 				vfs_busy_pages(vp, rabp);
365 				nfs_asyncio(vp, &rabp->b_bio2);
366 			    } else {
367 				brelse(rabp);
368 			    }
369 			}
370 		}
371 		/*
372 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
373 		 * chopped for the EOF condition, we cannot tell how large
374 		 * NFS directories are going to be until we hit EOF.  So
375 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
376 		 * it just so happens that b_resid will effectively chop it
377 		 * to EOF.  *BUT* this information is lost if the buffer goes
378 		 * away and is reconstituted into a B_CACHE state ( due to
379 		 * being VMIO ) later.  So we keep track of the directory eof
380 		 * in np->n_direofoffset and chop it off as an extra step
381 		 * right here.
382 		 *
383 		 * NOTE: boff could already be beyond EOF.
384 		 */
385 		if ((size_t)boff > NFS_DIRBLKSIZ - bp->b_resid) {
386 			n = 0;
387 		} else {
388 			n = szmin(uio->uio_resid,
389 				  NFS_DIRBLKSIZ - bp->b_resid - (size_t)boff);
390 		}
391 		if (np->n_direofoffset &&
392 		    n > (size_t)(np->n_direofoffset - uio->uio_offset)) {
393 			n = (size_t)(np->n_direofoffset - uio->uio_offset);
394 		}
395 		break;
396 	    default:
397 		kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
398 		n = 0;
399 		break;
400 	    }
401 
402 	    switch (vp->v_type) {
403 	    case VREG:
404 		if (n > 0)
405 		    error = uiomovebp(bp, bp->b_data + boff, n, uio);
406 		break;
407 	    case VLNK:
408 		if (n > 0)
409 		    error = uiomovebp(bp, bp->b_data + boff, n, uio);
410 		n = 0;
411 		break;
412 	    case VDIR:
413 		if (n > 0) {
414 		    off_t old_off = uio->uio_offset;
415 		    caddr_t cpos, epos;
416 		    struct nfs_dirent *dp;
417 
418 		    /*
419 		     * We are casting cpos to nfs_dirent, it must be
420 		     * int-aligned.
421 		     */
422 		    if (boff & 3) {
423 			error = EINVAL;
424 			break;
425 		    }
426 
427 		    cpos = bp->b_data + boff;
428 		    epos = bp->b_data + boff + n;
429 		    while (cpos < epos && error == 0 && uio->uio_resid > 0) {
430 			    dp = (struct nfs_dirent *)cpos;
431 			    error = nfs_check_dirent(dp, (int)(epos - cpos));
432 			    if (error)
433 				    break;
434 			    if (vop_write_dirent(&error, uio, dp->nfs_ino,
435 				dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
436 				    break;
437 			    }
438 			    cpos += dp->nfs_reclen;
439 		    }
440 		    n = 0;
441 		    if (error == 0) {
442 			    uio->uio_offset = old_off + cpos -
443 					      bp->b_data - boff;
444 		    }
445 		}
446 		break;
447 	    default:
448 		kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
449 	    }
450 	    if (bp)
451 		    brelse(bp);
452 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
453 	return (error);
454 }
455 
456 /*
457  * Userland can supply any 'seek' offset when reading a NFS directory.
458  * Validate the structure so we don't panic the kernel.  Note that
459  * the element name is nul terminated and the nul is not included
460  * in nfs_namlen.
461  */
462 static
463 int
464 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
465 {
466 	int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
467 
468 	if (nfs_name_off >= maxlen)
469 		return (EINVAL);
470 	if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
471 		return (EINVAL);
472 	if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
473 		return (EINVAL);
474 	if (dp->nfs_reclen & 3)
475 		return (EINVAL);
476 	return (0);
477 }
478 
479 /*
480  * Vnode op for write using bio
481  *
482  * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
483  *	     struct ucred *a_cred)
484  */
485 int
486 nfs_write(struct vop_write_args *ap)
487 {
488 	struct uio *uio = ap->a_uio;
489 	struct thread *td = uio->uio_td;
490 	struct vnode *vp = ap->a_vp;
491 	struct nfsnode *np = VTONFS(vp);
492 	int ioflag = ap->a_ioflag;
493 	struct buf *bp;
494 	struct vattr vattr;
495 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
496 	off_t loffset;
497 	int boff, bytes;
498 	int error = 0;
499 	int haverslock = 0;
500 	int bcount;
501 	int biosize;
502 	int trivial;
503 	int kflags = 0;
504 
505 #ifdef DIAGNOSTIC
506 	if (uio->uio_rw != UIO_WRITE)
507 		panic("nfs_write mode");
508 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
509 		panic("nfs_write proc");
510 #endif
511 	if (vp->v_type != VREG)
512 		return (EIO);
513 
514 	lwkt_gettoken(&nmp->nm_token);
515 
516 	if (np->n_flag & NWRITEERR) {
517 		np->n_flag &= ~NWRITEERR;
518 		lwkt_reltoken(&nmp->nm_token);
519 		return (np->n_error);
520 	}
521 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
522 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
523 		(void)nfs_fsinfo(nmp, vp, td);
524 	}
525 
526 	/*
527 	 * Synchronously flush pending buffers if we are in synchronous
528 	 * mode or if we are appending.
529 	 */
530 	if (ioflag & (IO_APPEND | IO_SYNC)) {
531 		if (np->n_flag & NLMODIFIED) {
532 			np->n_attrstamp = 0;
533 			error = nfs_flush(vp, MNT_WAIT, td, 0);
534 			/* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
535 			if (error)
536 				goto  done;
537 		}
538 	}
539 
540 	/*
541 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
542 	 * get the append lock.
543 	 */
544 restart:
545 	if (ioflag & IO_APPEND) {
546 		np->n_attrstamp = 0;
547 		error = VOP_GETATTR(vp, &vattr);
548 		if (error)
549 			goto done;
550 		uio->uio_offset = np->n_size;
551 	}
552 
553 	if (uio->uio_offset < 0) {
554 		error = EINVAL;
555 		goto done;
556 	}
557 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) {
558 		error = EFBIG;
559 		goto done;
560 	}
561 	if (uio->uio_resid == 0) {
562 		error = 0;
563 		goto done;
564 	}
565 
566 	/*
567 	 * We need to obtain the rslock if we intend to modify np->n_size
568 	 * in order to guarentee the append point with multiple contending
569 	 * writers, to guarentee that no other appenders modify n_size
570 	 * while we are trying to obtain a truncated buffer (i.e. to avoid
571 	 * accidently truncating data written by another appender due to
572 	 * the race), and to ensure that the buffer is populated prior to
573 	 * our extending of the file.  We hold rslock through the entire
574 	 * operation.
575 	 *
576 	 * Note that we do not synchronize the case where someone truncates
577 	 * the file while we are appending to it because attempting to lock
578 	 * this case may deadlock other parts of the system unexpectedly.
579 	 */
580 	if ((ioflag & IO_APPEND) ||
581 	    uio->uio_offset + uio->uio_resid > np->n_size) {
582 		switch(nfs_rslock(np)) {
583 		case ENOLCK:
584 			goto restart;
585 			/* not reached */
586 		case EINTR:
587 		case ERESTART:
588 			error = EINTR;
589 			goto done;
590 			/* not reached */
591 		default:
592 			break;
593 		}
594 		haverslock = 1;
595 	}
596 
597 	/*
598 	 * Maybe this should be above the vnode op call, but so long as
599 	 * file servers have no limits, i don't think it matters
600 	 */
601 	if (td && td->td_proc && uio->uio_offset + uio->uio_resid >
602 	      td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
603 		lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
604 		if (haverslock)
605 			nfs_rsunlock(np);
606 		error = EFBIG;
607 		goto done;
608 	}
609 
610 	biosize = vp->v_mount->mnt_stat.f_iosize;
611 
612 	do {
613 		nfsstats.biocache_writes++;
614 		boff = uio->uio_offset & (biosize-1);
615 		loffset = uio->uio_offset - boff;
616 		bytes = (int)szmin((unsigned)(biosize - boff), uio->uio_resid);
617 again:
618 		/*
619 		 * Handle direct append and file extension cases, calculate
620 		 * unaligned buffer size.  When extending B_CACHE will be
621 		 * set if possible.  See UIO_NOCOPY note below.
622 		 */
623 		if (uio->uio_offset + bytes > np->n_size) {
624 			np->n_flag |= NLMODIFIED;
625 			trivial = (uio->uio_segflg != UIO_NOCOPY &&
626 				   uio->uio_offset <= np->n_size);
627 			nfs_meta_setsize(vp, td, uio->uio_offset + bytes,
628 					 trivial);
629 			kflags |= NOTE_EXTEND;
630 		}
631 		bp = nfs_getcacheblk(vp, loffset, biosize, td);
632 		if (bp == NULL) {
633 			error = EINTR;
634 			break;
635 		}
636 
637 		/*
638 		 * Actual bytes in buffer which we care about
639 		 */
640 		if (loffset + biosize < np->n_size)
641 			bcount = biosize;
642 		else
643 			bcount = (int)(np->n_size - loffset);
644 
645 		/*
646 		 * Avoid a read by setting B_CACHE where the data we
647 		 * intend to write covers the entire buffer.  Note
648 		 * that the buffer may have been set to B_CACHE by
649 		 * nfs_meta_setsize() above or otherwise inherited the
650 		 * flag, but if B_CACHE isn't set the buffer may be
651 		 * uninitialized and must be zero'd to accomodate
652 		 * future seek+write's.
653 		 *
654 		 * See the comments in kern/vfs_bio.c's getblk() for
655 		 * more information.
656 		 *
657 		 * When doing a UIO_NOCOPY write the buffer is not
658 		 * overwritten and we cannot just set B_CACHE unconditionally
659 		 * for full-block writes.
660 		 */
661 		if (boff == 0 && bytes == biosize &&
662 		    uio->uio_segflg != UIO_NOCOPY) {
663 			bp->b_flags |= B_CACHE;
664 			bp->b_flags &= ~(B_ERROR | B_INVAL);
665 		}
666 
667 		/*
668 		 * b_resid may be set due to file EOF if we extended out.
669 		 * The NFS bio code will zero the difference anyway so
670 		 * just acknowledged the fact and set b_resid to 0.
671 		 */
672 		if ((bp->b_flags & B_CACHE) == 0) {
673 			bp->b_cmd = BUF_CMD_READ;
674 			bp->b_bio2.bio_done = nfsiodone_sync;
675 			bp->b_bio2.bio_flags |= BIO_SYNC;
676 			vfs_busy_pages(vp, bp);
677 			error = nfs_doio(vp, &bp->b_bio2, td);
678 			if (error) {
679 				brelse(bp);
680 				break;
681 			}
682 			bp->b_resid = 0;
683 		}
684 		np->n_flag |= NLMODIFIED;
685 		kflags |= NOTE_WRITE;
686 
687 		/*
688 		 * If dirtyend exceeds file size, chop it down.  This should
689 		 * not normally occur but there is an append race where it
690 		 * might occur XXX, so we log it.
691 		 *
692 		 * If the chopping creates a reverse-indexed or degenerate
693 		 * situation with dirtyoff/end, we 0 both of them.
694 		 */
695 		if (bp->b_dirtyend > bcount) {
696 			kprintf("NFS append race @%08llx:%d\n",
697 			    (long long)bp->b_bio2.bio_offset,
698 			    bp->b_dirtyend - bcount);
699 			bp->b_dirtyend = bcount;
700 		}
701 
702 		if (bp->b_dirtyoff >= bp->b_dirtyend)
703 			bp->b_dirtyoff = bp->b_dirtyend = 0;
704 
705 		/*
706 		 * If the new write will leave a contiguous dirty
707 		 * area, just update the b_dirtyoff and b_dirtyend,
708 		 * otherwise force a write rpc of the old dirty area.
709 		 *
710 		 * While it is possible to merge discontiguous writes due to
711 		 * our having a B_CACHE buffer ( and thus valid read data
712 		 * for the hole), we don't because it could lead to
713 		 * significant cache coherency problems with multiple clients,
714 		 * especially if locking is implemented later on.
715 		 *
716 		 * as an optimization we could theoretically maintain
717 		 * a linked list of discontinuous areas, but we would still
718 		 * have to commit them separately so there isn't much
719 		 * advantage to it except perhaps a bit of asynchronization.
720 		 */
721 		if (bp->b_dirtyend > 0 &&
722 		    (boff > bp->b_dirtyend ||
723 		     (boff + bytes) < bp->b_dirtyoff)
724 		) {
725 			if (bwrite(bp) == EINTR) {
726 				error = EINTR;
727 				break;
728 			}
729 			goto again;
730 		}
731 
732 		error = uiomovebp(bp, bp->b_data + boff, bytes, uio);
733 
734 		/*
735 		 * Since this block is being modified, it must be written
736 		 * again and not just committed.  Since write clustering does
737 		 * not work for the stage 1 data write, only the stage 2
738 		 * commit rpc, we have to clear B_CLUSTEROK as well.
739 		 */
740 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
741 
742 		if (error) {
743 			brelse(bp);
744 			break;
745 		}
746 
747 		/*
748 		 * Only update dirtyoff/dirtyend if not a degenerate
749 		 * condition.
750 		 *
751 		 * The underlying VM pages have been marked valid by
752 		 * virtue of acquiring the bp.  Because the entire buffer
753 		 * is marked dirty we do not have to worry about cleaning
754 		 * out the related dirty bits (and wouldn't really know
755 		 * how to deal with byte ranges anyway)
756 		 */
757 		if (bytes) {
758 			if (bp->b_dirtyend > 0) {
759 				bp->b_dirtyoff = imin(boff, bp->b_dirtyoff);
760 				bp->b_dirtyend = imax(boff + bytes,
761 						      bp->b_dirtyend);
762 			} else {
763 				bp->b_dirtyoff = boff;
764 				bp->b_dirtyend = boff + bytes;
765 			}
766 		}
767 
768 		/*
769 		 * If the lease is non-cachable or IO_SYNC do bwrite().
770 		 *
771 		 * IO_INVAL appears to be unused.  The idea appears to be
772 		 * to turn off caching in this case.  Very odd.  XXX
773 		 *
774 		 * If nfs_async is set bawrite() will use an unstable write
775 		 * (build dirty bufs on the server), so we might as well
776 		 * push it out with bawrite().  If nfs_async is not set we
777 		 * use bdwrite() to cache dirty bufs on the client.
778 		 */
779 		if (ioflag & IO_SYNC) {
780 			if (ioflag & IO_INVAL)
781 				bp->b_flags |= B_NOCACHE;
782 			error = bwrite(bp);
783 			if (error)
784 				break;
785 		} else if (boff + bytes == biosize && nfs_async) {
786 			bawrite(bp);
787 		} else {
788 			bdwrite(bp);
789 		}
790 	} while (uio->uio_resid > 0 && bytes > 0);
791 
792 	if (haverslock)
793 		nfs_rsunlock(np);
794 
795 done:
796 	nfs_knote(vp, kflags);
797 	lwkt_reltoken(&nmp->nm_token);
798 	return (error);
799 }
800 
801 /*
802  * Get an nfs cache block.
803  *
804  * Allocate a new one if the block isn't currently in the cache
805  * and return the block marked busy. If the calling process is
806  * interrupted by a signal for an interruptible mount point, return
807  * NULL.
808  *
809  * The caller must carefully deal with the possible B_INVAL state of
810  * the buffer.  nfs_startio() clears B_INVAL (and nfs_asyncio() clears it
811  * indirectly), so synchronous reads can be issued without worrying about
812  * the B_INVAL state.  We have to be a little more careful when dealing
813  * with writes (see comments in nfs_write()) when extending a file past
814  * its EOF.
815  */
816 static struct buf *
817 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
818 {
819 	struct buf *bp;
820 	struct mount *mp;
821 	struct nfsmount *nmp;
822 
823 	mp = vp->v_mount;
824 	nmp = VFSTONFS(mp);
825 
826 	if (nmp->nm_flag & NFSMNT_INT) {
827 		bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
828 		while (bp == NULL) {
829 			if (nfs_sigintr(nmp, NULL, td))
830 				return (NULL);
831 			bp = getblk(vp, loffset, size, 0, 2 * hz);
832 		}
833 	} else {
834 		bp = getblk(vp, loffset, size, 0, 0);
835 	}
836 
837 	/*
838 	 * bio2, the 'device' layer.  Since BIOs use 64 bit byte offsets
839 	 * now, no translation is necessary.
840 	 */
841 	bp->b_bio2.bio_offset = loffset;
842 	return (bp);
843 }
844 
845 /*
846  * Flush and invalidate all dirty buffers. If another process is already
847  * doing the flush, just wait for completion.
848  */
849 int
850 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
851 {
852 	struct nfsnode *np = VTONFS(vp);
853 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
854 	int error = 0, slpflag, slptimeo;
855 	thread_t td = curthread;
856 
857 	if (vp->v_flag & VRECLAIMED)
858 		return (0);
859 
860 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
861 		intrflg = 0;
862 	if (intrflg) {
863 		slpflag = PCATCH;
864 		slptimeo = 2 * hz;
865 	} else {
866 		slpflag = 0;
867 		slptimeo = 0;
868 	}
869 	/*
870 	 * First wait for any other process doing a flush to complete.
871 	 */
872 	while (np->n_flag & NFLUSHINPROG) {
873 		np->n_flag |= NFLUSHWANT;
874 		error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
875 		if (error && intrflg && nfs_sigintr(nmp, NULL, td))
876 			return (EINTR);
877 	}
878 
879 	/*
880 	 * Now, flush as required.
881 	 */
882 	np->n_flag |= NFLUSHINPROG;
883 	error = vinvalbuf(vp, flags, slpflag, 0);
884 	while (error) {
885 		if (intrflg && nfs_sigintr(nmp, NULL, td)) {
886 			np->n_flag &= ~NFLUSHINPROG;
887 			if (np->n_flag & NFLUSHWANT) {
888 				np->n_flag &= ~NFLUSHWANT;
889 				wakeup((caddr_t)&np->n_flag);
890 			}
891 			return (EINTR);
892 		}
893 		error = vinvalbuf(vp, flags, 0, slptimeo);
894 	}
895 	np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
896 	if (np->n_flag & NFLUSHWANT) {
897 		np->n_flag &= ~NFLUSHWANT;
898 		wakeup((caddr_t)&np->n_flag);
899 	}
900 	return (0);
901 }
902 
903 /*
904  * Return true (non-zero) if the txthread and rxthread are operational
905  * and we do not already have too many not-yet-started BIO's built up.
906  */
907 int
908 nfs_asyncok(struct nfsmount *nmp)
909 {
910 	return (nmp->nm_bioqlen < nfs_maxasyncbio &&
911 		nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE &&
912 		nmp->nm_rxstate <= NFSSVC_PENDING &&
913 		nmp->nm_txstate <= NFSSVC_PENDING);
914 }
915 
916 /*
917  * The read-ahead code calls this to queue a bio to the txthread.
918  *
919  * We don't touch the bio otherwise... that is, we do not even
920  * construct or send the initial rpc.  The txthread will do it
921  * for us.
922  *
923  * NOTE!  nm_bioqlen is not decremented until the request completes,
924  *	  so it does not reflect the number of bio's on bioq.
925  */
926 void
927 nfs_asyncio(struct vnode *vp, struct bio *bio)
928 {
929 	struct buf *bp = bio->bio_buf;
930 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
931 
932 	KKASSERT(vp->v_tag == VT_NFS);
933 	BUF_KERNPROC(bp);
934 
935 	/*
936 	 * Shortcut swap cache (not done automatically because we are not
937 	 * using bread()).
938 	 */
939 	if (vn_cache_strategy(vp, bio))
940 		return;
941 
942 	bio->bio_driver_info = vp;
943 	crit_enter();
944 	TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
945 	atomic_add_int(&nmp->nm_bioqlen, 1);
946 	crit_exit();
947 	nfssvc_iod_writer_wakeup(nmp);
948 }
949 
950 /*
951  * nfs_doio()	- Execute a BIO operation synchronously.  The BIO will be
952  *		  completed and its error returned.  The caller is responsible
953  *		  for brelse()ing it.  ONLY USE FOR BIO_SYNC IOs!  Otherwise
954  *		  our error probe will be against an invalid pointer.
955  *
956  * nfs_startio()- Execute a BIO operation assynchronously.
957  *
958  * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation,
959  *	 which basically just queues it to the txthread.  nfs_startio()
960  *	 actually initiates the I/O AFTER it has gotten to the txthread.
961  *
962  * NOTE: td might be NULL.
963  *
964  * NOTE: Caller has already busied the I/O.
965  */
966 void
967 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td)
968 {
969 	struct buf *bp = bio->bio_buf;
970 
971 	KKASSERT(vp->v_tag == VT_NFS);
972 
973 	/*
974 	 * clear B_ERROR and B_INVAL state prior to initiating the I/O.  We
975 	 * do this here so we do not have to do it in all the code that
976 	 * calls us.
977 	 */
978 	bp->b_flags &= ~(B_ERROR | B_INVAL);
979 
980 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
981 		("nfs_doio: bp %p already marked done!", bp));
982 
983 	if (bp->b_cmd == BUF_CMD_READ) {
984 	    switch (vp->v_type) {
985 	    case VREG:
986 		nfsstats.read_bios++;
987 		nfs_readrpc_bio(vp, bio);
988 		break;
989 	    case VLNK:
990 #if 0
991 		bio->bio_offset = 0;
992 		nfsstats.readlink_bios++;
993 		nfs_readlinkrpc_bio(vp, bio);
994 #else
995 		nfs_doio(vp, bio, td);
996 #endif
997 		break;
998 	    case VDIR:
999 		/*
1000 		 * NOTE: If nfs_readdirplusrpc_bio() is requested but
1001 		 *	 not supported, it will chain to
1002 		 *	 nfs_readdirrpc_bio().
1003 		 */
1004 #if 0
1005 		nfsstats.readdir_bios++;
1006 		uiop->uio_offset = bio->bio_offset;
1007 		if (nmp->nm_flag & NFSMNT_RDIRPLUS)
1008 			nfs_readdirplusrpc_bio(vp, bio);
1009 		else
1010 			nfs_readdirrpc_bio(vp, bio);
1011 #else
1012 		nfs_doio(vp, bio, td);
1013 #endif
1014 		break;
1015 	    default:
1016 		kprintf("nfs_doio:  type %x unexpected\n",vp->v_type);
1017 		bp->b_flags |= B_ERROR;
1018 		bp->b_error = EINVAL;
1019 		biodone(bio);
1020 		break;
1021 	    }
1022 	} else {
1023 	    /*
1024 	     * If we only need to commit, try to commit.  If this fails
1025 	     * it will chain through to the write.  Basically all the logic
1026 	     * in nfs_doio() is replicated.
1027 	     */
1028 	    KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1029 	    if (bp->b_flags & B_NEEDCOMMIT)
1030 		nfs_commitrpc_bio(vp, bio);
1031 	    else
1032 		nfs_writerpc_bio(vp, bio);
1033 	}
1034 }
1035 
1036 int
1037 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1038 {
1039 	struct buf *bp = bio->bio_buf;
1040 	struct uio *uiop;
1041 	struct nfsnode *np;
1042 	struct nfsmount *nmp;
1043 	int error = 0;
1044 	int iomode, must_commit;
1045 	size_t n;
1046 	struct uio uio;
1047 	struct iovec io;
1048 
1049 #if 0
1050 	/*
1051 	 * Shortcut swap cache (not done automatically because we are not
1052 	 * using bread()).
1053 	 *
1054 	 * XXX The biowait is a hack until we can figure out how to stop a
1055 	 * biodone chain when a middle element is BIO_SYNC.  BIO_SYNC is
1056 	 * set so the bp shouldn't get ripped out from under us.  The only
1057 	 * use-cases are fully synchronous I/O cases.
1058 	 *
1059 	 * XXX This is having problems, give up for now.
1060 	 */
1061 	if (vn_cache_strategy(vp, bio)) {
1062 		error = biowait(&bio->bio_buf->b_bio1, "nfsrsw");
1063 		return (error);
1064 	}
1065 #endif
1066 
1067 	KKASSERT(vp->v_tag == VT_NFS);
1068 	np = VTONFS(vp);
1069 	nmp = VFSTONFS(vp->v_mount);
1070 	uiop = &uio;
1071 	uiop->uio_iov = &io;
1072 	uiop->uio_iovcnt = 1;
1073 	uiop->uio_segflg = UIO_SYSSPACE;
1074 	uiop->uio_td = td;
1075 
1076 	/*
1077 	 * clear B_ERROR and B_INVAL state prior to initiating the I/O.  We
1078 	 * do this here so we do not have to do it in all the code that
1079 	 * calls us.
1080 	 */
1081 	bp->b_flags &= ~(B_ERROR | B_INVAL);
1082 
1083 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
1084 		("nfs_doio: bp %p already marked done!", bp));
1085 
1086 	if (bp->b_cmd == BUF_CMD_READ) {
1087 	    io.iov_len = uiop->uio_resid = (size_t)bp->b_bcount;
1088 	    io.iov_base = bp->b_data;
1089 	    uiop->uio_rw = UIO_READ;
1090 
1091 	    switch (vp->v_type) {
1092 	    case VREG:
1093 		/*
1094 		 * When reading from a regular file zero-fill any residual.
1095 		 * Note that this residual has nothing to do with NFS short
1096 		 * reads, which nfs_readrpc_uio() will handle for us.
1097 		 *
1098 		 * We have to do this because when we are write extending
1099 		 * a file the server may not have the same notion of
1100 		 * filesize as we do.  Our BIOs should already be sized
1101 		 * (b_bcount) to account for the file EOF.
1102 		 */
1103 		nfsstats.read_bios++;
1104 		uiop->uio_offset = bio->bio_offset;
1105 		error = nfs_readrpc_uio(vp, uiop);
1106 		if (error == 0 && uiop->uio_resid) {
1107 			n = (size_t)bp->b_bcount - uiop->uio_resid;
1108 			bzero(bp->b_data + n, bp->b_bcount - n);
1109 			uiop->uio_resid = 0;
1110 		}
1111 		if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1112 		    np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1113 			uprintf("Process killed due to text file modification\n");
1114 			ksignal(td->td_proc, SIGKILL);
1115 		}
1116 		break;
1117 	    case VLNK:
1118 		uiop->uio_offset = 0;
1119 		nfsstats.readlink_bios++;
1120 		error = nfs_readlinkrpc_uio(vp, uiop);
1121 		break;
1122 	    case VDIR:
1123 		nfsstats.readdir_bios++;
1124 		uiop->uio_offset = bio->bio_offset;
1125 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1126 			error = nfs_readdirplusrpc_uio(vp, uiop);
1127 			if (error == NFSERR_NOTSUPP)
1128 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1129 		}
1130 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1131 			error = nfs_readdirrpc_uio(vp, uiop);
1132 		/*
1133 		 * end-of-directory sets B_INVAL but does not generate an
1134 		 * error.
1135 		 */
1136 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1137 			bp->b_flags |= B_INVAL;
1138 		break;
1139 	    default:
1140 		kprintf("nfs_doio:  type %x unexpected\n",vp->v_type);
1141 		break;
1142 	    }
1143 	    if (error) {
1144 		bp->b_flags |= B_ERROR;
1145 		bp->b_error = error;
1146 	    }
1147 	    bp->b_resid = uiop->uio_resid;
1148 	} else {
1149 	    /*
1150 	     * If we only need to commit, try to commit.
1151 	     *
1152 	     * NOTE: The I/O has already been staged for the write and
1153 	     *	     its pages busied, so b_dirtyoff/end is valid.
1154 	     */
1155 	    KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1156 	    if (bp->b_flags & B_NEEDCOMMIT) {
1157 		    int retv;
1158 		    off_t off;
1159 
1160 		    off = bio->bio_offset + bp->b_dirtyoff;
1161 		    retv = nfs_commitrpc_uio(vp, off,
1162 					     bp->b_dirtyend - bp->b_dirtyoff,
1163 					     td);
1164 		    if (retv == 0) {
1165 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1166 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1167 			    bp->b_resid = 0;
1168 			    biodone(bio);
1169 			    return(0);
1170 		    }
1171 		    if (retv == NFSERR_STALEWRITEVERF) {
1172 			    nfs_clearcommit(vp->v_mount);
1173 		    }
1174 	    }
1175 
1176 	    /*
1177 	     * Setup for actual write
1178 	     */
1179 	    if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1180 		bp->b_dirtyend = np->n_size - bio->bio_offset;
1181 
1182 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1183 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1184 		    - bp->b_dirtyoff;
1185 		uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1186 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1187 		uiop->uio_rw = UIO_WRITE;
1188 		nfsstats.write_bios++;
1189 
1190 		if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1191 		    iomode = NFSV3WRITE_UNSTABLE;
1192 		else
1193 		    iomode = NFSV3WRITE_FILESYNC;
1194 
1195 		must_commit = 0;
1196 		error = nfs_writerpc_uio(vp, uiop, &iomode, &must_commit);
1197 
1198 		/*
1199 		 * We no longer try to use kern/vfs_bio's cluster code to
1200 		 * cluster commits, so B_CLUSTEROK is no longer set with
1201 		 * B_NEEDCOMMIT.  The problem is that a vfs_busy_pages()
1202 		 * may have to clear B_NEEDCOMMIT if it finds underlying
1203 		 * pages have been redirtied through a memory mapping
1204 		 * and doing this on a clustered bp will probably cause
1205 		 * a panic, plus the flag in the underlying NFS bufs
1206 		 * making up the cluster bp will not be properly cleared.
1207 		 */
1208 		if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1209 		    bp->b_flags |= B_NEEDCOMMIT;
1210 #if 0
1211 		    /* XXX do not enable commit clustering */
1212 		    if (bp->b_dirtyoff == 0
1213 			&& bp->b_dirtyend == bp->b_bcount)
1214 			bp->b_flags |= B_CLUSTEROK;
1215 #endif
1216 		} else {
1217 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1218 		}
1219 
1220 		/*
1221 		 * For an interrupted write, the buffer is still valid
1222 		 * and the write hasn't been pushed to the server yet,
1223 		 * so we can't set B_ERROR and report the interruption
1224 		 * by setting B_EINTR. For the async case, B_EINTR
1225 		 * is not relevant, so the rpc attempt is essentially
1226 		 * a noop.  For the case of a V3 write rpc not being
1227 		 * committed to stable storage, the block is still
1228 		 * dirty and requires either a commit rpc or another
1229 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1230 		 * the block is reused. This is indicated by setting
1231 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1232 		 *
1233 		 * If the buffer is marked B_PAGING, it does not reside on
1234 		 * the vp's paging queues so we cannot call bdirty().  The
1235 		 * bp in this case is not an NFS cache block so we should
1236 		 * be safe. XXX
1237 		 */
1238     		if (error == EINTR
1239 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1240 			crit_enter();
1241 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1242 			if ((bp->b_flags & B_PAGING) == 0)
1243 			    bdirty(bp);
1244 			if (error)
1245 			    bp->b_flags |= B_EINTR;
1246 			crit_exit();
1247 	    	} else {
1248 		    if (error) {
1249 			bp->b_flags |= B_ERROR;
1250 			bp->b_error = np->n_error = error;
1251 			np->n_flag |= NWRITEERR;
1252 		    }
1253 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1254 		}
1255 		if (must_commit)
1256 		    nfs_clearcommit(vp->v_mount);
1257 		bp->b_resid = uiop->uio_resid;
1258 	    } else {
1259 		bp->b_resid = 0;
1260 	    }
1261 	}
1262 
1263 	/*
1264 	 * I/O was run synchronously, biodone() it and calculate the
1265 	 * error to return.
1266 	 */
1267 	biodone(bio);
1268 	KKASSERT(bp->b_cmd == BUF_CMD_DONE);
1269 	if (bp->b_flags & B_EINTR)
1270 		return (EINTR);
1271 	if (bp->b_flags & B_ERROR)
1272 		return (bp->b_error ? bp->b_error : EIO);
1273 	return (0);
1274 }
1275 
1276 /*
1277  * Handle all truncation, write-extend, and ftruncate()-extend operations
1278  * on the NFS lcient side.
1279  *
1280  * We use the new API in kern/vfs_vm.c to perform these operations in a
1281  * VM-friendly way.  With this API VM pages are properly zerod and pages
1282  * still mapped into the buffer straddling EOF are not invalidated.
1283  */
1284 int
1285 nfs_meta_setsize(struct vnode *vp, struct thread *td, off_t nsize, int trivial)
1286 {
1287 	struct nfsnode *np = VTONFS(vp);
1288 	off_t osize;
1289 	int biosize = vp->v_mount->mnt_stat.f_iosize;
1290 	int error;
1291 
1292 	osize = np->n_size;
1293 	np->n_size = nsize;
1294 
1295 	if (nsize < osize) {
1296 		error = nvtruncbuf(vp, nsize, biosize, -1, 0);
1297 	} else {
1298 		error = nvextendbuf(vp, osize, nsize,
1299 				    biosize, biosize, -1, -1,
1300 				    trivial);
1301 	}
1302 	return(error);
1303 }
1304 
1305 /*
1306  * Synchronous completion for nfs_doio.  Call bpdone() with elseit=FALSE.
1307  * Caller is responsible for brelse()'ing the bp.
1308  */
1309 static void
1310 nfsiodone_sync(struct bio *bio)
1311 {
1312 	bio->bio_flags = 0;
1313 	bpdone(bio->bio_buf, 0);
1314 }
1315 
1316 /*
1317  * nfs read rpc - BIO version
1318  */
1319 void
1320 nfs_readrpc_bio(struct vnode *vp, struct bio *bio)
1321 {
1322 	struct buf *bp = bio->bio_buf;
1323 	u_int32_t *tl;
1324 	struct nfsmount *nmp;
1325 	int error = 0, len, tsiz;
1326 	struct nfsm_info *info;
1327 
1328 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1329 	info->mrep = NULL;
1330 	info->v3 = NFS_ISV3(vp);
1331 
1332 	nmp = VFSTONFS(vp->v_mount);
1333 	tsiz = bp->b_bcount;
1334 	KKASSERT(tsiz <= nmp->nm_rsize);
1335 	if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) {
1336 		error = EFBIG;
1337 		goto nfsmout;
1338 	}
1339 	nfsstats.rpccnt[NFSPROC_READ]++;
1340 	len = tsiz;
1341 	nfsm_reqhead(info, vp, NFSPROC_READ,
1342 		     NFSX_FH(info->v3) + NFSX_UNSIGNED * 3);
1343 	ERROROUT(nfsm_fhtom(info, vp));
1344 	tl = nfsm_build(info, NFSX_UNSIGNED * 3);
1345 	if (info->v3) {
1346 		txdr_hyper(bio->bio_offset, tl);
1347 		*(tl + 2) = txdr_unsigned(len);
1348 	} else {
1349 		*tl++ = txdr_unsigned(bio->bio_offset);
1350 		*tl++ = txdr_unsigned(len);
1351 		*tl = 0;
1352 	}
1353 	info->bio = bio;
1354 	info->done = nfs_readrpc_bio_done;
1355 	nfsm_request_bio(info, vp, NFSPROC_READ, NULL,
1356 			 nfs_vpcred(vp, ND_READ));
1357 	return;
1358 nfsmout:
1359 	kfree(info, M_NFSREQ);
1360 	bp->b_error = error;
1361 	bp->b_flags |= B_ERROR;
1362 	biodone(bio);
1363 }
1364 
1365 static void
1366 nfs_readrpc_bio_done(nfsm_info_t info)
1367 {
1368 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1369 	struct bio *bio = info->bio;
1370 	struct buf *bp = bio->bio_buf;
1371 	u_int32_t *tl;
1372 	int attrflag;
1373 	int retlen;
1374 	int eof;
1375 	int error = 0;
1376 
1377 	KKASSERT(info->state == NFSM_STATE_DONE);
1378 
1379 	lwkt_gettoken(&nmp->nm_token);
1380 
1381 	ERROROUT(info->error);
1382 	if (info->v3) {
1383 		ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag,
1384 					 NFS_LATTR_NOSHRINK));
1385 		NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED));
1386 		eof = fxdr_unsigned(int, *(tl + 1));
1387 	} else {
1388 		ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1389 		eof = 0;
1390 	}
1391 	NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize));
1392 	ERROROUT(nfsm_mtobio(info, bio, retlen));
1393 	m_freem(info->mrep);
1394 	info->mrep = NULL;
1395 
1396 	/*
1397 	 * No error occured, if retlen is less then bcount and no EOF
1398 	 * and NFSv3 a zero-fill short read occured.
1399 	 *
1400 	 * For NFSv2 a short-read indicates EOF.
1401 	 */
1402 	if (retlen < bp->b_bcount && info->v3 && eof == 0) {
1403 		bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1404 		retlen = bp->b_bcount;
1405 	}
1406 
1407 	/*
1408 	 * If we hit an EOF we still zero-fill, but return the expected
1409 	 * b_resid anyway.  This should normally not occur since async
1410 	 * BIOs are not used for read-before-write case.  Races against
1411 	 * the server can cause it though and we don't want to leave
1412 	 * garbage in the buffer.
1413 	 */
1414 	if (retlen < bp->b_bcount) {
1415 		bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1416 	}
1417 	bp->b_resid = 0;
1418 	/* bp->b_resid = bp->b_bcount - retlen; */
1419 nfsmout:
1420 	lwkt_reltoken(&nmp->nm_token);
1421 	kfree(info, M_NFSREQ);
1422 	if (error) {
1423 		bp->b_error = error;
1424 		bp->b_flags |= B_ERROR;
1425 	}
1426 	biodone(bio);
1427 }
1428 
1429 /*
1430  * nfs write call - BIO version
1431  *
1432  * NOTE: Caller has already busied the I/O.
1433  */
1434 void
1435 nfs_writerpc_bio(struct vnode *vp, struct bio *bio)
1436 {
1437 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1438 	struct nfsnode *np = VTONFS(vp);
1439 	struct buf *bp = bio->bio_buf;
1440 	u_int32_t *tl;
1441 	int len;
1442 	int iomode;
1443 	int error = 0;
1444 	struct nfsm_info *info;
1445 	off_t offset;
1446 
1447 	/*
1448 	 * Setup for actual write.  Just clean up the bio if there
1449 	 * is nothing to do.  b_dirtyoff/end have already been staged
1450 	 * by the bp's pages getting busied.
1451 	 */
1452 	if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1453 		bp->b_dirtyend = np->n_size - bio->bio_offset;
1454 
1455 	if (bp->b_dirtyend <= bp->b_dirtyoff) {
1456 		bp->b_resid = 0;
1457 		biodone(bio);
1458 		return;
1459 	}
1460 	len = bp->b_dirtyend - bp->b_dirtyoff;
1461 	offset = bio->bio_offset + bp->b_dirtyoff;
1462 	if (offset + len > nmp->nm_maxfilesize) {
1463 		bp->b_flags |= B_ERROR;
1464 		bp->b_error = EFBIG;
1465 		biodone(bio);
1466 		return;
1467 	}
1468 	bp->b_resid = len;
1469 	nfsstats.write_bios++;
1470 
1471 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1472 	info->mrep = NULL;
1473 	info->v3 = NFS_ISV3(vp);
1474 	info->info_writerpc.must_commit = 0;
1475 	if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1476 		iomode = NFSV3WRITE_UNSTABLE;
1477 	else
1478 		iomode = NFSV3WRITE_FILESYNC;
1479 
1480 	KKASSERT(len <= nmp->nm_wsize);
1481 
1482 	nfsstats.rpccnt[NFSPROC_WRITE]++;
1483 	nfsm_reqhead(info, vp, NFSPROC_WRITE,
1484 		     NFSX_FH(info->v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len));
1485 	ERROROUT(nfsm_fhtom(info, vp));
1486 	if (info->v3) {
1487 		tl = nfsm_build(info, 5 * NFSX_UNSIGNED);
1488 		txdr_hyper(offset, tl);
1489 		tl += 2;
1490 		*tl++ = txdr_unsigned(len);
1491 		*tl++ = txdr_unsigned(iomode);
1492 		*tl = txdr_unsigned(len);
1493 	} else {
1494 		u_int32_t x;
1495 
1496 		tl = nfsm_build(info, 4 * NFSX_UNSIGNED);
1497 		/* Set both "begin" and "current" to non-garbage. */
1498 		x = txdr_unsigned((u_int32_t)offset);
1499 		*tl++ = x;	/* "begin offset" */
1500 		*tl++ = x;	/* "current offset" */
1501 		x = txdr_unsigned(len);
1502 		*tl++ = x;	/* total to this offset */
1503 		*tl = x;	/* size of this write */
1504 	}
1505 	ERROROUT(nfsm_biotom(info, bio, bp->b_dirtyoff, len));
1506 	info->bio = bio;
1507 	info->done = nfs_writerpc_bio_done;
1508 	nfsm_request_bio(info, vp, NFSPROC_WRITE, NULL,
1509 			 nfs_vpcred(vp, ND_WRITE));
1510 	return;
1511 nfsmout:
1512 	kfree(info, M_NFSREQ);
1513 	bp->b_error = error;
1514 	bp->b_flags |= B_ERROR;
1515 	biodone(bio);
1516 }
1517 
1518 static void
1519 nfs_writerpc_bio_done(nfsm_info_t info)
1520 {
1521 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1522 	struct nfsnode *np = VTONFS(info->vp);
1523 	struct bio *bio = info->bio;
1524 	struct buf *bp = bio->bio_buf;
1525 	int wccflag = NFSV3_WCCRATTR;
1526 	int iomode = NFSV3WRITE_FILESYNC;
1527 	int commit;
1528 	int rlen;
1529 	int error;
1530 	int len = bp->b_resid;	/* b_resid was set to shortened length */
1531 	u_int32_t *tl;
1532 
1533 	lwkt_gettoken(&nmp->nm_token);
1534 
1535 	ERROROUT(info->error);
1536 	if (info->v3) {
1537 		/*
1538 		 * The write RPC returns a before and after mtime.  The
1539 		 * nfsm_wcc_data() macro checks the before n_mtime
1540 		 * against the before time and stores the after time
1541 		 * in the nfsnode's cached vattr and n_mtime field.
1542 		 * The NRMODIFIED bit will be set if the before
1543 		 * time did not match the original mtime.
1544 		 */
1545 		wccflag = NFSV3_WCCCHK;
1546 		ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1547 		if (error == 0) {
1548 			NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF));
1549 			rlen = fxdr_unsigned(int, *tl++);
1550 			if (rlen == 0) {
1551 				error = NFSERR_IO;
1552 				m_freem(info->mrep);
1553 				info->mrep = NULL;
1554 				goto nfsmout;
1555 			} else if (rlen < len) {
1556 #if 0
1557 				/*
1558 				 * XXX what do we do here?
1559 				 */
1560 				backup = len - rlen;
1561 				uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup;
1562 				uiop->uio_iov->iov_len += backup;
1563 				uiop->uio_offset -= backup;
1564 				uiop->uio_resid += backup;
1565 				len = rlen;
1566 #endif
1567 			}
1568 			commit = fxdr_unsigned(int, *tl++);
1569 
1570 			/*
1571 			 * Return the lowest committment level
1572 			 * obtained by any of the RPCs.
1573 			 */
1574 			if (iomode == NFSV3WRITE_FILESYNC)
1575 				iomode = commit;
1576 			else if (iomode == NFSV3WRITE_DATASYNC &&
1577 				commit == NFSV3WRITE_UNSTABLE)
1578 				iomode = commit;
1579 			if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){
1580 			    bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1581 			    nmp->nm_state |= NFSSTA_HASWRITEVERF;
1582 			} else if (bcmp(tl, nmp->nm_verf, NFSX_V3WRITEVERF)) {
1583 			    info->info_writerpc.must_commit = 1;
1584 			    bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1585 			}
1586 		}
1587 	} else {
1588 		ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1589 	}
1590 	m_freem(info->mrep);
1591 	info->mrep = NULL;
1592 	len = 0;
1593 nfsmout:
1594 	if (info->vp->v_mount->mnt_flag & MNT_ASYNC)
1595 		iomode = NFSV3WRITE_FILESYNC;
1596 	bp->b_resid = len;
1597 
1598 	/*
1599 	 * End of RPC.  Now clean up the bp.
1600 	 *
1601 	 * We no longer enable write clustering for commit operations,
1602 	 * See around line 1157 for a more detailed comment.
1603 	 */
1604 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1605 		bp->b_flags |= B_NEEDCOMMIT;
1606 #if 0
1607 		/* XXX do not enable commit clustering */
1608 		if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount)
1609 			bp->b_flags |= B_CLUSTEROK;
1610 #endif
1611 	} else {
1612 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1613 	}
1614 
1615 	/*
1616 	 * For an interrupted write, the buffer is still valid
1617 	 * and the write hasn't been pushed to the server yet,
1618 	 * so we can't set B_ERROR and report the interruption
1619 	 * by setting B_EINTR. For the async case, B_EINTR
1620 	 * is not relevant, so the rpc attempt is essentially
1621 	 * a noop.  For the case of a V3 write rpc not being
1622 	 * committed to stable storage, the block is still
1623 	 * dirty and requires either a commit rpc or another
1624 	 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1625 	 * the block is reused. This is indicated by setting
1626 	 * the B_DELWRI and B_NEEDCOMMIT flags.
1627 	 *
1628 	 * If the buffer is marked B_PAGING, it does not reside on
1629 	 * the vp's paging queues so we cannot call bdirty().  The
1630 	 * bp in this case is not an NFS cache block so we should
1631 	 * be safe. XXX
1632 	 */
1633 	if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1634 		crit_enter();
1635 		bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1636 		if ((bp->b_flags & B_PAGING) == 0)
1637 			bdirty(bp);
1638 		if (error)
1639 			bp->b_flags |= B_EINTR;
1640 		crit_exit();
1641 	} else {
1642 		if (error) {
1643 			bp->b_flags |= B_ERROR;
1644 			bp->b_error = np->n_error = error;
1645 			np->n_flag |= NWRITEERR;
1646 		}
1647 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1648 	}
1649 	if (info->info_writerpc.must_commit)
1650 		nfs_clearcommit(info->vp->v_mount);
1651 	lwkt_reltoken(&nmp->nm_token);
1652 
1653 	kfree(info, M_NFSREQ);
1654 	if (error) {
1655 		bp->b_flags |= B_ERROR;
1656 		bp->b_error = error;
1657 	}
1658 	biodone(bio);
1659 }
1660 
1661 /*
1662  * Nfs Version 3 commit rpc - BIO version
1663  *
1664  * This function issues the commit rpc and will chain to a write
1665  * rpc if necessary.
1666  */
1667 void
1668 nfs_commitrpc_bio(struct vnode *vp, struct bio *bio)
1669 {
1670 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1671 	struct buf *bp = bio->bio_buf;
1672 	struct nfsm_info *info;
1673 	int error = 0;
1674 	u_int32_t *tl;
1675 
1676 	if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) {
1677 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1678 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1679 		bp->b_resid = 0;
1680 		biodone(bio);
1681 		return;
1682 	}
1683 
1684 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1685 	info->mrep = NULL;
1686 	info->v3 = 1;
1687 
1688 	nfsstats.rpccnt[NFSPROC_COMMIT]++;
1689 	nfsm_reqhead(info, vp, NFSPROC_COMMIT, NFSX_FH(1));
1690 	ERROROUT(nfsm_fhtom(info, vp));
1691 	tl = nfsm_build(info, 3 * NFSX_UNSIGNED);
1692 	txdr_hyper(bio->bio_offset + bp->b_dirtyoff, tl);
1693 	tl += 2;
1694 	*tl = txdr_unsigned(bp->b_dirtyend - bp->b_dirtyoff);
1695 	info->bio = bio;
1696 	info->done = nfs_commitrpc_bio_done;
1697 	nfsm_request_bio(info, vp, NFSPROC_COMMIT, NULL,
1698 			 nfs_vpcred(vp, ND_WRITE));
1699 	return;
1700 nfsmout:
1701 	/*
1702 	 * Chain to write RPC on (early) error
1703 	 */
1704 	kfree(info, M_NFSREQ);
1705 	nfs_writerpc_bio(vp, bio);
1706 }
1707 
1708 static void
1709 nfs_commitrpc_bio_done(nfsm_info_t info)
1710 {
1711 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1712 	struct bio *bio = info->bio;
1713 	struct buf *bp = bio->bio_buf;
1714 	u_int32_t *tl;
1715 	int wccflag = NFSV3_WCCRATTR;
1716 	int error = 0;
1717 
1718 	lwkt_gettoken(&nmp->nm_token);
1719 
1720 	ERROROUT(info->error);
1721 	ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1722 	if (error == 0) {
1723 		NULLOUT(tl = nfsm_dissect(info, NFSX_V3WRITEVERF));
1724 		if (bcmp(nmp->nm_verf, tl, NFSX_V3WRITEVERF)) {
1725 			bcopy(tl, nmp->nm_verf, NFSX_V3WRITEVERF);
1726 			error = NFSERR_STALEWRITEVERF;
1727 		}
1728 	}
1729 	m_freem(info->mrep);
1730 	info->mrep = NULL;
1731 
1732 	/*
1733 	 * On completion we must chain to a write bio if an
1734 	 * error occurred.
1735 	 */
1736 nfsmout:
1737 	if (error == 0) {
1738 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1739 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1740 		bp->b_resid = 0;
1741 		biodone(bio);
1742 	} else {
1743 		nfs_writerpc_bio(info->vp, bio);
1744 	}
1745 	kfree(info, M_NFSREQ);
1746 	lwkt_reltoken(&nmp->nm_token);
1747 }
1748 
1749