xref: /dflybsd-src/sys/vfs/nfs/nfs_bio.c (revision bfc09ba0a4d805c1860f88e64d6ae9a407d3567d)
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. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
37  * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
38  * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.45 2008/07/18 00:09:39 dillon Exp $
39  */
40 
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/resourcevar.h>
45 #include <sys/signalvar.h>
46 #include <sys/proc.h>
47 #include <sys/buf.h>
48 #include <sys/vnode.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
51 #include <sys/mbuf.h>
52 #include <sys/msfbuf.h>
53 
54 #include <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_pager.h>
59 #include <vm/vnode_pager.h>
60 
61 #include <sys/buf2.h>
62 #include <sys/thread2.h>
63 #include <vm/vm_page2.h>
64 
65 #include "rpcv2.h"
66 #include "nfsproto.h"
67 #include "nfs.h"
68 #include "nfsmount.h"
69 #include "nfsnode.h"
70 #include "xdr_subs.h"
71 #include "nfsm_subs.h"
72 
73 
74 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset,
75 				   int size, struct thread *td);
76 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen);
77 static void nfsiodone_sync(struct bio *bio);
78 static void nfs_readrpc_bio_done(nfsm_info_t info);
79 static void nfs_writerpc_bio_done(nfsm_info_t info);
80 static void nfs_commitrpc_bio_done(nfsm_info_t info);
81 
82 /*
83  * Vnode op for VM getpages.
84  *
85  * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
86  *		int a_reqpage, vm_ooffset_t a_offset)
87  */
88 int
89 nfs_getpages(struct vop_getpages_args *ap)
90 {
91 	struct thread *td = curthread;		/* XXX */
92 	int i, error, nextoff, size, toff, count, npages;
93 	struct uio uio;
94 	struct iovec iov;
95 	char *kva;
96 	struct vnode *vp;
97 	struct nfsmount *nmp;
98 	vm_page_t *pages;
99 	vm_page_t m;
100 	struct msf_buf *msf;
101 
102 	vp = ap->a_vp;
103 	nmp = VFSTONFS(vp->v_mount);
104 	pages = ap->a_m;
105 	count = ap->a_count;
106 
107 	if (vp->v_object == NULL) {
108 		kprintf("nfs_getpages: called with non-merged cache vnode??\n");
109 		return VM_PAGER_ERROR;
110 	}
111 
112 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
113 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
114 		(void)nfs_fsinfo(nmp, vp, td);
115 
116 	npages = btoc(count);
117 
118 	/*
119 	 * NOTE that partially valid pages may occur in cases other
120 	 * then file EOF, such as when a file is partially written and
121 	 * ftruncate()-extended to a larger size.   It is also possible
122 	 * for the valid bits to be set on garbage beyond the file EOF and
123 	 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
124 	 * occur due to vtruncbuf() and the buffer cache's handling of
125 	 * pages which 'straddle' buffers or when b_bufsize is not a
126 	 * multiple of PAGE_SIZE.... the buffer cache cannot normally
127 	 * clear the extra bits.  This kind of situation occurs when you
128 	 * make a small write() (m->valid == 0x03) and then mmap() and
129 	 * fault in the buffer(m->valid = 0xFF).  When NFS flushes the
130 	 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
131 	 *
132 	 * This is combined with the possibility that the pages are partially
133 	 * dirty or that there is a buffer backing the pages that is dirty
134 	 * (even if m->dirty is 0).
135 	 *
136 	 * To solve this problem several hacks have been made:  (1) NFS
137 	 * guarentees that the IO block size is a multiple of PAGE_SIZE and
138 	 * (2) The buffer cache, when invalidating an NFS buffer, will
139 	 * disregard the buffer's fragmentory b_bufsize and invalidate
140 	 * the whole page rather then just the piece the buffer owns.
141 	 *
142 	 * This allows us to assume that a partially valid page found here
143 	 * is fully valid (vm_fault will zero'd out areas of the page not
144 	 * marked as valid).
145 	 */
146 	m = pages[ap->a_reqpage];
147 	if (m->valid != 0) {
148 		for (i = 0; i < npages; ++i) {
149 			if (i != ap->a_reqpage)
150 				vnode_pager_freepage(pages[i]);
151 		}
152 		return(0);
153 	}
154 
155 	/*
156 	 * Use an MSF_BUF as a medium to retrieve data from the pages.
157 	 */
158 	msf_map_pagelist(&msf, pages, npages, 0);
159 	KKASSERT(msf);
160 	kva = msf_buf_kva(msf);
161 
162 	iov.iov_base = kva;
163 	iov.iov_len = count;
164 	uio.uio_iov = &iov;
165 	uio.uio_iovcnt = 1;
166 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
167 	uio.uio_resid = count;
168 	uio.uio_segflg = UIO_SYSSPACE;
169 	uio.uio_rw = UIO_READ;
170 	uio.uio_td = td;
171 
172 	error = nfs_readrpc_uio(vp, &uio);
173 	msf_buf_free(msf);
174 
175 	if (error && ((int)uio.uio_resid == count)) {
176 		kprintf("nfs_getpages: error %d\n", error);
177 		for (i = 0; i < npages; ++i) {
178 			if (i != ap->a_reqpage)
179 				vnode_pager_freepage(pages[i]);
180 		}
181 		return VM_PAGER_ERROR;
182 	}
183 
184 	/*
185 	 * Calculate the number of bytes read and validate only that number
186 	 * of bytes.  Note that due to pending writes, size may be 0.  This
187 	 * does not mean that the remaining data is invalid!
188 	 */
189 
190 	size = count - (int)uio.uio_resid;
191 
192 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
193 		nextoff = toff + PAGE_SIZE;
194 		m = pages[i];
195 
196 		m->flags &= ~PG_ZERO;
197 
198 		if (nextoff <= size) {
199 			/*
200 			 * Read operation filled an entire page
201 			 */
202 			m->valid = VM_PAGE_BITS_ALL;
203 			vm_page_undirty(m);
204 		} else if (size > toff) {
205 			/*
206 			 * Read operation filled a partial page.
207 			 */
208 			m->valid = 0;
209 			vm_page_set_valid(m, 0, size - toff);
210 			vm_page_clear_dirty_end_nonincl(m, 0, size - toff);
211 			/* handled by vm_fault now	  */
212 			/* vm_page_zero_invalid(m, TRUE); */
213 		} else {
214 			/*
215 			 * Read operation was short.  If no error occured
216 			 * we may have hit a zero-fill section.   We simply
217 			 * leave valid set to 0.
218 			 */
219 			;
220 		}
221 		if (i != ap->a_reqpage) {
222 			/*
223 			 * Whether or not to leave the page activated is up in
224 			 * the air, but we should put the page on a page queue
225 			 * somewhere (it already is in the object).  Result:
226 			 * It appears that emperical results show that
227 			 * deactivating pages is best.
228 			 */
229 
230 			/*
231 			 * Just in case someone was asking for this page we
232 			 * now tell them that it is ok to use.
233 			 */
234 			if (!error) {
235 				if (m->flags & PG_WANTED)
236 					vm_page_activate(m);
237 				else
238 					vm_page_deactivate(m);
239 				vm_page_wakeup(m);
240 			} else {
241 				vnode_pager_freepage(m);
242 			}
243 		}
244 	}
245 	return 0;
246 }
247 
248 /*
249  * Vnode op for VM putpages.
250  *
251  * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
252  *		int *a_rtvals, vm_ooffset_t a_offset)
253  */
254 int
255 nfs_putpages(struct vop_putpages_args *ap)
256 {
257 	struct thread *td = curthread;
258 	struct uio uio;
259 	struct iovec iov;
260 	char *kva;
261 	int iomode, must_commit, i, error, npages, count;
262 	off_t offset;
263 	int *rtvals;
264 	struct vnode *vp;
265 	struct nfsmount *nmp;
266 	struct nfsnode *np;
267 	vm_page_t *pages;
268 	struct msf_buf *msf;
269 
270 	vp = ap->a_vp;
271 	np = VTONFS(vp);
272 	nmp = VFSTONFS(vp->v_mount);
273 	pages = ap->a_m;
274 	count = ap->a_count;
275 	rtvals = ap->a_rtvals;
276 	npages = btoc(count);
277 	offset = IDX_TO_OFF(pages[0]->pindex);
278 
279 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
280 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
281 		(void)nfs_fsinfo(nmp, vp, td);
282 
283 	for (i = 0; i < npages; i++) {
284 		rtvals[i] = VM_PAGER_AGAIN;
285 	}
286 
287 	/*
288 	 * When putting pages, do not extend file past EOF.
289 	 */
290 
291 	if (offset + count > np->n_size) {
292 		count = np->n_size - offset;
293 		if (count < 0)
294 			count = 0;
295 	}
296 
297 	/*
298 	 * Use an MSF_BUF as a medium to retrieve data from the pages.
299 	 */
300 	msf_map_pagelist(&msf, pages, npages, 0);
301 	KKASSERT(msf);
302 	kva = msf_buf_kva(msf);
303 
304 	iov.iov_base = kva;
305 	iov.iov_len = count;
306 	uio.uio_iov = &iov;
307 	uio.uio_iovcnt = 1;
308 	uio.uio_offset = offset;
309 	uio.uio_resid = (size_t)count;
310 	uio.uio_segflg = UIO_SYSSPACE;
311 	uio.uio_rw = UIO_WRITE;
312 	uio.uio_td = td;
313 
314 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
315 	    iomode = NFSV3WRITE_UNSTABLE;
316 	else
317 	    iomode = NFSV3WRITE_FILESYNC;
318 
319 	error = nfs_writerpc_uio(vp, &uio, &iomode, &must_commit);
320 
321 	msf_buf_free(msf);
322 
323 	if (!error) {
324 		int nwritten = round_page(count - (int)uio.uio_resid) / PAGE_SIZE;
325 		for (i = 0; i < nwritten; i++) {
326 			rtvals[i] = VM_PAGER_OK;
327 			vm_page_undirty(pages[i]);
328 		}
329 		if (must_commit)
330 			nfs_clearcommit(vp->v_mount);
331 	}
332 	return rtvals[0];
333 }
334 
335 /*
336  * Vnode op for read using bio
337  */
338 int
339 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
340 {
341 	struct nfsnode *np = VTONFS(vp);
342 	int biosize, i;
343 	struct buf *bp = 0, *rabp;
344 	struct vattr vattr;
345 	struct thread *td;
346 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
347 	daddr_t lbn, rabn;
348 	off_t raoffset;
349 	off_t loffset;
350 	int bcount;
351 	int seqcount;
352 	int nra, error = 0, n = 0, on = 0;
353 
354 #ifdef DIAGNOSTIC
355 	if (uio->uio_rw != UIO_READ)
356 		panic("nfs_read mode");
357 #endif
358 	if (uio->uio_resid == 0)
359 		return (0);
360 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
361 		return (EINVAL);
362 	td = uio->uio_td;
363 
364 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
365 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
366 		(void)nfs_fsinfo(nmp, vp, td);
367 	if (vp->v_type != VDIR &&
368 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
369 		return (EFBIG);
370 	biosize = vp->v_mount->mnt_stat.f_iosize;
371 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
372 
373 	/*
374 	 * For nfs, cache consistency can only be maintained approximately.
375 	 * Although RFC1094 does not specify the criteria, the following is
376 	 * believed to be compatible with the reference port.
377 	 *
378 	 * NFS:		If local changes have been made and this is a
379 	 *		directory, the directory must be invalidated and
380 	 *		the attribute cache must be cleared.
381 	 *
382 	 *		GETATTR is called to synchronize the file size.
383 	 *
384 	 *		If remote changes are detected local data is flushed
385 	 *		and the cache is invalidated.
386 	 *
387 	 *		NOTE: In the normal case the attribute cache is not
388 	 *		cleared which means GETATTR may use cached data and
389 	 *		not immediately detect changes made on the server.
390 	 */
391 	if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
392 		nfs_invaldir(vp);
393 		error = nfs_vinvalbuf(vp, V_SAVE, 1);
394 		if (error)
395 			return (error);
396 		np->n_attrstamp = 0;
397 	}
398 	error = VOP_GETATTR(vp, &vattr);
399 	if (error)
400 		return (error);
401 	if (np->n_flag & NRMODIFIED) {
402 		if (vp->v_type == VDIR)
403 			nfs_invaldir(vp);
404 		error = nfs_vinvalbuf(vp, V_SAVE, 1);
405 		if (error)
406 			return (error);
407 		np->n_flag &= ~NRMODIFIED;
408 	}
409 	do {
410 	    if (np->n_flag & NDONTCACHE) {
411 		switch (vp->v_type) {
412 		case VREG:
413 			return (nfs_readrpc_uio(vp, uio));
414 		case VLNK:
415 			return (nfs_readlinkrpc_uio(vp, uio));
416 		case VDIR:
417 			break;
418 		default:
419 			kprintf(" NDONTCACHE: type %x unexpected\n", vp->v_type);
420 			break;
421 		};
422 	    }
423 	    switch (vp->v_type) {
424 	    case VREG:
425 		nfsstats.biocache_reads++;
426 		lbn = uio->uio_offset / biosize;
427 		on = uio->uio_offset & (biosize - 1);
428 		loffset = (off_t)lbn * biosize;
429 
430 		/*
431 		 * Start the read ahead(s), as required.
432 		 */
433 		if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) {
434 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
435 			(off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
436 			rabn = lbn + 1 + nra;
437 			raoffset = (off_t)rabn * biosize;
438 			if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
439 			    rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
440 			    if (!rabp)
441 				return (EINTR);
442 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
443 				rabp->b_cmd = BUF_CMD_READ;
444 				vfs_busy_pages(vp, rabp);
445 				nfs_asyncio(vp, &rabp->b_bio2);
446 			    } else {
447 				brelse(rabp);
448 			    }
449 			}
450 		    }
451 		}
452 
453 		/*
454 		 * Obtain the buffer cache block.  Figure out the buffer size
455 		 * when we are at EOF.  If we are modifying the size of the
456 		 * buffer based on an EOF condition we need to hold
457 		 * nfs_rslock() through obtaining the buffer to prevent
458 		 * a potential writer-appender from messing with n_size.
459 		 * Otherwise we may accidently truncate the buffer and
460 		 * lose dirty data.
461 		 *
462 		 * Note that bcount is *not* DEV_BSIZE aligned.
463 		 */
464 
465 again:
466 		bcount = biosize;
467 		if (loffset >= np->n_size) {
468 			bcount = 0;
469 		} else if (loffset + biosize > np->n_size) {
470 			bcount = np->n_size - loffset;
471 		}
472 		if (bcount != biosize) {
473 			switch(nfs_rslock(np)) {
474 			case ENOLCK:
475 				goto again;
476 				/* not reached */
477 			case EINTR:
478 			case ERESTART:
479 				return(EINTR);
480 				/* not reached */
481 			default:
482 				break;
483 			}
484 		}
485 
486 		bp = nfs_getcacheblk(vp, loffset, bcount, td);
487 
488 		if (bcount != biosize)
489 			nfs_rsunlock(np);
490 		if (!bp)
491 			return (EINTR);
492 
493 		/*
494 		 * If B_CACHE is not set, we must issue the read.  If this
495 		 * fails, we return an error.
496 		 */
497 		if ((bp->b_flags & B_CACHE) == 0) {
498 			bp->b_cmd = BUF_CMD_READ;
499 			bp->b_bio2.bio_done = nfsiodone_sync;
500 			bp->b_bio2.bio_flags |= BIO_SYNC;
501 			vfs_busy_pages(vp, bp);
502 			error = nfs_doio(vp, &bp->b_bio2, td);
503 			if (error) {
504 				brelse(bp);
505 				return (error);
506 			}
507 		}
508 
509 		/*
510 		 * on is the offset into the current bp.  Figure out how many
511 		 * bytes we can copy out of the bp.  Note that bcount is
512 		 * NOT DEV_BSIZE aligned.
513 		 *
514 		 * Then figure out how many bytes we can copy into the uio.
515 		 */
516 		n = 0;
517 		if (on < bcount)
518 			n = (int)szmin((unsigned)(bcount - on), uio->uio_resid);
519 		break;
520 	    case VLNK:
521 		biosize = min(NFS_MAXPATHLEN, np->n_size);
522 		nfsstats.biocache_readlinks++;
523 		bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
524 		if (bp == NULL)
525 			return (EINTR);
526 		if ((bp->b_flags & B_CACHE) == 0) {
527 			bp->b_cmd = BUF_CMD_READ;
528 			bp->b_bio2.bio_done = nfsiodone_sync;
529 			bp->b_bio2.bio_flags |= BIO_SYNC;
530 			vfs_busy_pages(vp, bp);
531 			error = nfs_doio(vp, &bp->b_bio2, td);
532 			if (error) {
533 				bp->b_flags |= B_ERROR | B_INVAL;
534 				brelse(bp);
535 				return (error);
536 			}
537 		}
538 		n = (int)szmin(uio->uio_resid, bp->b_bcount - bp->b_resid);
539 		on = 0;
540 		break;
541 	    case VDIR:
542 		nfsstats.biocache_readdirs++;
543 		if (np->n_direofoffset
544 		    && uio->uio_offset >= np->n_direofoffset) {
545 		    return (0);
546 		}
547 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
548 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
549 		loffset = uio->uio_offset - on;
550 		bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
551 		if (bp == NULL)
552 		    return (EINTR);
553 
554 		if ((bp->b_flags & B_CACHE) == 0) {
555 		    bp->b_cmd = BUF_CMD_READ;
556 		    bp->b_bio2.bio_done = nfsiodone_sync;
557 		    bp->b_bio2.bio_flags |= BIO_SYNC;
558 		    vfs_busy_pages(vp, bp);
559 		    error = nfs_doio(vp, &bp->b_bio2, td);
560 		    if (error)
561 			    brelse(bp);
562 		    while (error == NFSERR_BAD_COOKIE) {
563 			kprintf("got bad cookie vp %p bp %p\n", vp, bp);
564 			nfs_invaldir(vp);
565 			error = nfs_vinvalbuf(vp, 0, 1);
566 			/*
567 			 * Yuck! The directory has been modified on the
568 			 * server. The only way to get the block is by
569 			 * reading from the beginning to get all the
570 			 * offset cookies.
571 			 *
572 			 * Leave the last bp intact unless there is an error.
573 			 * Loop back up to the while if the error is another
574 			 * NFSERR_BAD_COOKIE (double yuch!).
575 			 */
576 			for (i = 0; i <= lbn && !error; i++) {
577 			    if (np->n_direofoffset
578 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
579 				    return (0);
580 			    bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
581 						 NFS_DIRBLKSIZ, td);
582 			    if (!bp)
583 				return (EINTR);
584 			    if ((bp->b_flags & B_CACHE) == 0) {
585 				    bp->b_cmd = BUF_CMD_READ;
586 				    bp->b_bio2.bio_done = nfsiodone_sync;
587 				    bp->b_bio2.bio_flags |= BIO_SYNC;
588 				    vfs_busy_pages(vp, bp);
589 				    error = nfs_doio(vp, &bp->b_bio2, td);
590 				    /*
591 				     * no error + B_INVAL == directory EOF,
592 				     * use the block.
593 				     */
594 				    if (error == 0 && (bp->b_flags & B_INVAL))
595 					    break;
596 			    }
597 			    /*
598 			     * An error will throw away the block and the
599 			     * for loop will break out.  If no error and this
600 			     * is not the block we want, we throw away the
601 			     * block and go for the next one via the for loop.
602 			     */
603 			    if (error || i < lbn)
604 				    brelse(bp);
605 			}
606 		    }
607 		    /*
608 		     * The above while is repeated if we hit another cookie
609 		     * error.  If we hit an error and it wasn't a cookie error,
610 		     * we give up.
611 		     */
612 		    if (error)
613 			    return (error);
614 		}
615 
616 		/*
617 		 * If not eof and read aheads are enabled, start one.
618 		 * (You need the current block first, so that you have the
619 		 *  directory offset cookie of the next block.)
620 		 */
621 		if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) &&
622 		    (bp->b_flags & B_INVAL) == 0 &&
623 		    (np->n_direofoffset == 0 ||
624 		    loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
625 		    (np->n_flag & NDONTCACHE) == 0 &&
626 		    findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
627 		) {
628 			rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
629 					       NFS_DIRBLKSIZ, td);
630 			if (rabp) {
631 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
632 				rabp->b_cmd = BUF_CMD_READ;
633 				vfs_busy_pages(vp, rabp);
634 				nfs_asyncio(vp, &rabp->b_bio2);
635 			    } else {
636 				brelse(rabp);
637 			    }
638 			}
639 		}
640 		/*
641 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
642 		 * chopped for the EOF condition, we cannot tell how large
643 		 * NFS directories are going to be until we hit EOF.  So
644 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
645 		 * it just so happens that b_resid will effectively chop it
646 		 * to EOF.  *BUT* this information is lost if the buffer goes
647 		 * away and is reconstituted into a B_CACHE state ( due to
648 		 * being VMIO ) later.  So we keep track of the directory eof
649 		 * in np->n_direofoffset and chop it off as an extra step
650 		 * right here.
651 		 */
652 		n = (int)szmin(uio->uio_resid,
653 			       NFS_DIRBLKSIZ - bp->b_resid - on);
654 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
655 			n = np->n_direofoffset - uio->uio_offset;
656 		break;
657 	    default:
658 		kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
659 		break;
660 	    };
661 
662 	    switch (vp->v_type) {
663 	    case VREG:
664 		if (n > 0)
665 		    error = uiomove(bp->b_data + on, (int)n, uio);
666 		break;
667 	    case VLNK:
668 		if (n > 0)
669 		    error = uiomove(bp->b_data + on, (int)n, uio);
670 		n = 0;
671 		break;
672 	    case VDIR:
673 		if (n > 0) {
674 		    off_t old_off = uio->uio_offset;
675 		    caddr_t cpos, epos;
676 		    struct nfs_dirent *dp;
677 
678 		    /*
679 		     * We are casting cpos to nfs_dirent, it must be
680 		     * int-aligned.
681 		     */
682 		    if (on & 3) {
683 			error = EINVAL;
684 			break;
685 		    }
686 
687 		    cpos = bp->b_data + on;
688 		    epos = bp->b_data + on + n;
689 		    while (cpos < epos && error == 0 && uio->uio_resid > 0) {
690 			    dp = (struct nfs_dirent *)cpos;
691 			    error = nfs_check_dirent(dp, (int)(epos - cpos));
692 			    if (error)
693 				    break;
694 			    if (vop_write_dirent(&error, uio, dp->nfs_ino,
695 				dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
696 				    break;
697 			    }
698 			    cpos += dp->nfs_reclen;
699 		    }
700 		    n = 0;
701 		    if (error == 0)
702 			    uio->uio_offset = old_off + cpos - bp->b_data - on;
703 		}
704 		/*
705 		 * Invalidate buffer if caching is disabled, forcing a
706 		 * re-read from the remote later.
707 		 */
708 		if (np->n_flag & NDONTCACHE)
709 			bp->b_flags |= B_INVAL;
710 		break;
711 	    default:
712 		kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
713 	    }
714 	    brelse(bp);
715 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
716 	return (error);
717 }
718 
719 /*
720  * Userland can supply any 'seek' offset when reading a NFS directory.
721  * Validate the structure so we don't panic the kernel.  Note that
722  * the element name is nul terminated and the nul is not included
723  * in nfs_namlen.
724  */
725 static
726 int
727 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
728 {
729 	int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
730 
731 	if (nfs_name_off >= maxlen)
732 		return (EINVAL);
733 	if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
734 		return (EINVAL);
735 	if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
736 		return (EINVAL);
737 	if (dp->nfs_reclen & 3)
738 		return (EINVAL);
739 	return (0);
740 }
741 
742 /*
743  * Vnode op for write using bio
744  *
745  * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
746  *	     struct ucred *a_cred)
747  */
748 int
749 nfs_write(struct vop_write_args *ap)
750 {
751 	struct uio *uio = ap->a_uio;
752 	struct thread *td = uio->uio_td;
753 	struct vnode *vp = ap->a_vp;
754 	struct nfsnode *np = VTONFS(vp);
755 	int ioflag = ap->a_ioflag;
756 	struct buf *bp;
757 	struct vattr vattr;
758 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
759 	daddr_t lbn;
760 	off_t loffset;
761 	int n, on, error = 0, iomode, must_commit;
762 	int haverslock = 0;
763 	int bcount;
764 	int biosize;
765 
766 #ifdef DIAGNOSTIC
767 	if (uio->uio_rw != UIO_WRITE)
768 		panic("nfs_write mode");
769 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
770 		panic("nfs_write proc");
771 #endif
772 	if (vp->v_type != VREG)
773 		return (EIO);
774 	if (np->n_flag & NWRITEERR) {
775 		np->n_flag &= ~NWRITEERR;
776 		return (np->n_error);
777 	}
778 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
779 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
780 		(void)nfs_fsinfo(nmp, vp, td);
781 
782 	/*
783 	 * Synchronously flush pending buffers if we are in synchronous
784 	 * mode or if we are appending.
785 	 */
786 	if (ioflag & (IO_APPEND | IO_SYNC)) {
787 		if (np->n_flag & NLMODIFIED) {
788 			np->n_attrstamp = 0;
789 			error = nfs_flush(vp, MNT_WAIT, td, 0);
790 			/* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
791 			if (error)
792 				return (error);
793 		}
794 	}
795 
796 	/*
797 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
798 	 * get the append lock.
799 	 */
800 restart:
801 	if (ioflag & IO_APPEND) {
802 		np->n_attrstamp = 0;
803 		error = VOP_GETATTR(vp, &vattr);
804 		if (error)
805 			return (error);
806 		uio->uio_offset = np->n_size;
807 	}
808 
809 	if (uio->uio_offset < 0)
810 		return (EINVAL);
811 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
812 		return (EFBIG);
813 	if (uio->uio_resid == 0)
814 		return (0);
815 
816 	/*
817 	 * We need to obtain the rslock if we intend to modify np->n_size
818 	 * in order to guarentee the append point with multiple contending
819 	 * writers, to guarentee that no other appenders modify n_size
820 	 * while we are trying to obtain a truncated buffer (i.e. to avoid
821 	 * accidently truncating data written by another appender due to
822 	 * the race), and to ensure that the buffer is populated prior to
823 	 * our extending of the file.  We hold rslock through the entire
824 	 * operation.
825 	 *
826 	 * Note that we do not synchronize the case where someone truncates
827 	 * the file while we are appending to it because attempting to lock
828 	 * this case may deadlock other parts of the system unexpectedly.
829 	 */
830 	if ((ioflag & IO_APPEND) ||
831 	    uio->uio_offset + uio->uio_resid > np->n_size) {
832 		switch(nfs_rslock(np)) {
833 		case ENOLCK:
834 			goto restart;
835 			/* not reached */
836 		case EINTR:
837 		case ERESTART:
838 			return(EINTR);
839 			/* not reached */
840 		default:
841 			break;
842 		}
843 		haverslock = 1;
844 	}
845 
846 	/*
847 	 * Maybe this should be above the vnode op call, but so long as
848 	 * file servers have no limits, i don't think it matters
849 	 */
850 	if (td->td_proc && uio->uio_offset + uio->uio_resid >
851 	      td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
852 		lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
853 		if (haverslock)
854 			nfs_rsunlock(np);
855 		return (EFBIG);
856 	}
857 
858 	biosize = vp->v_mount->mnt_stat.f_iosize;
859 
860 	do {
861 		if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) {
862 		    iomode = NFSV3WRITE_FILESYNC;
863 		    error = nfs_writerpc_uio(vp, uio, &iomode, &must_commit);
864 		    if (must_commit)
865 			    nfs_clearcommit(vp->v_mount);
866 		    break;
867 		}
868 		nfsstats.biocache_writes++;
869 		lbn = uio->uio_offset / biosize;
870 		on = uio->uio_offset & (biosize-1);
871 		loffset = uio->uio_offset - on;
872 		n = (int)szmin((unsigned)(biosize - on), uio->uio_resid);
873 again:
874 		/*
875 		 * Handle direct append and file extension cases, calculate
876 		 * unaligned buffer size.
877 		 */
878 
879 		if (uio->uio_offset == np->n_size && n) {
880 			/*
881 			 * Get the buffer (in its pre-append state to maintain
882 			 * B_CACHE if it was previously set).  Resize the
883 			 * nfsnode after we have locked the buffer to prevent
884 			 * readers from reading garbage.
885 			 */
886 			bcount = on;
887 			bp = nfs_getcacheblk(vp, loffset, bcount, td);
888 
889 			if (bp != NULL) {
890 				long save;
891 
892 				np->n_size = uio->uio_offset + n;
893 				np->n_flag |= NLMODIFIED;
894 				vnode_pager_setsize(vp, np->n_size);
895 
896 				save = bp->b_flags & B_CACHE;
897 				bcount += n;
898 				allocbuf(bp, bcount);
899 				bp->b_flags |= save;
900 			}
901 		} else {
902 			/*
903 			 * Obtain the locked cache block first, and then
904 			 * adjust the file's size as appropriate.
905 			 */
906 			bcount = on + n;
907 			if (loffset + bcount < np->n_size) {
908 				if (loffset + biosize < np->n_size)
909 					bcount = biosize;
910 				else
911 					bcount = np->n_size - loffset;
912 			}
913 			bp = nfs_getcacheblk(vp, loffset, bcount, td);
914 			if (uio->uio_offset + n > np->n_size) {
915 				np->n_size = uio->uio_offset + n;
916 				np->n_flag |= NLMODIFIED;
917 				vnode_pager_setsize(vp, np->n_size);
918 			}
919 		}
920 
921 		if (bp == NULL) {
922 			error = EINTR;
923 			break;
924 		}
925 
926 		/*
927 		 * Avoid a read by setting B_CACHE where the data we
928 		 * intend to write covers the entire buffer.  This also
929 		 * handles the normal append case as bcount will have
930 		 * byte resolution.  The buffer state must also be adjusted.
931 		 *
932 		 * See the comments in kern/vfs_bio.c's getblk() for
933 		 * more information.
934 		 *
935 		 * When doing a UIO_NOCOPY write the buffer is not
936 		 * overwritten and we cannot just set B_CACHE unconditionally
937 		 * for full-block writes.
938 		 */
939 		if (on == 0 && n == bcount && uio->uio_segflg != UIO_NOCOPY) {
940 			bp->b_flags |= B_CACHE;
941 			bp->b_flags &= ~(B_ERROR | B_INVAL);
942 		}
943 
944 		/*
945 		 * b_resid may be set due to file EOF if we extended out.
946 		 * The NFS bio code will zero the difference anyway so
947 		 * just acknowledged the fact and set b_resid to 0.
948 		 */
949 		if ((bp->b_flags & B_CACHE) == 0) {
950 			bp->b_cmd = BUF_CMD_READ;
951 			bp->b_bio2.bio_done = nfsiodone_sync;
952 			bp->b_bio2.bio_flags |= BIO_SYNC;
953 			vfs_busy_pages(vp, bp);
954 			error = nfs_doio(vp, &bp->b_bio2, td);
955 			if (error) {
956 				brelse(bp);
957 				break;
958 			}
959 			bp->b_resid = 0;
960 		}
961 		if (!bp) {
962 			error = EINTR;
963 			break;
964 		}
965 		np->n_flag |= NLMODIFIED;
966 
967 		/*
968 		 * If dirtyend exceeds file size, chop it down.  This should
969 		 * not normally occur but there is an append race where it
970 		 * might occur XXX, so we log it.
971 		 *
972 		 * If the chopping creates a reverse-indexed or degenerate
973 		 * situation with dirtyoff/end, we 0 both of them.
974 		 */
975 
976 		if (bp->b_dirtyend > bcount) {
977 			kprintf("NFS append race @%08llx:%d\n",
978 			    (long long)bp->b_bio2.bio_offset,
979 			    bp->b_dirtyend - bcount);
980 			bp->b_dirtyend = bcount;
981 		}
982 
983 		if (bp->b_dirtyoff >= bp->b_dirtyend)
984 			bp->b_dirtyoff = bp->b_dirtyend = 0;
985 
986 		/*
987 		 * If the new write will leave a contiguous dirty
988 		 * area, just update the b_dirtyoff and b_dirtyend,
989 		 * otherwise force a write rpc of the old dirty area.
990 		 *
991 		 * While it is possible to merge discontiguous writes due to
992 		 * our having a B_CACHE buffer ( and thus valid read data
993 		 * for the hole), we don't because it could lead to
994 		 * significant cache coherency problems with multiple clients,
995 		 * especially if locking is implemented later on.
996 		 *
997 		 * as an optimization we could theoretically maintain
998 		 * a linked list of discontinuous areas, but we would still
999 		 * have to commit them separately so there isn't much
1000 		 * advantage to it except perhaps a bit of asynchronization.
1001 		 */
1002 		if (bp->b_dirtyend > 0 &&
1003 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1004 			if (bwrite(bp) == EINTR) {
1005 				error = EINTR;
1006 				break;
1007 			}
1008 			goto again;
1009 		}
1010 
1011 		error = uiomove((char *)bp->b_data + on, n, uio);
1012 
1013 		/*
1014 		 * Since this block is being modified, it must be written
1015 		 * again and not just committed.  Since write clustering does
1016 		 * not work for the stage 1 data write, only the stage 2
1017 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1018 		 */
1019 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1020 
1021 		if (error) {
1022 			bp->b_flags |= B_ERROR;
1023 			brelse(bp);
1024 			break;
1025 		}
1026 
1027 		/*
1028 		 * Only update dirtyoff/dirtyend if not a degenerate
1029 		 * condition.
1030 		 *
1031 		 * The underlying VM pages have been marked valid by
1032 		 * virtue of acquiring the bp.  Because the entire buffer
1033 		 * is marked dirty we do not have to worry about cleaning
1034 		 * out the related dirty bits (and wouldn't really know
1035 		 * how to deal with byte ranges anyway)
1036 		 */
1037 		if (n) {
1038 			if (bp->b_dirtyend > 0) {
1039 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1040 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1041 			} else {
1042 				bp->b_dirtyoff = on;
1043 				bp->b_dirtyend = on + n;
1044 			}
1045 		}
1046 
1047 		/*
1048 		 * If the lease is non-cachable or IO_SYNC do bwrite().
1049 		 *
1050 		 * IO_INVAL appears to be unused.  The idea appears to be
1051 		 * to turn off caching in this case.  Very odd.  XXX
1052 		 *
1053 		 * If nfs_async is set bawrite() will use an unstable write
1054 		 * (build dirty bufs on the server), so we might as well
1055 		 * push it out with bawrite().  If nfs_async is not set we
1056 		 * use bdwrite() to cache dirty bufs on the client.
1057 		 */
1058 		if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) {
1059 			if (ioflag & IO_INVAL)
1060 				bp->b_flags |= B_NOCACHE;
1061 			error = bwrite(bp);
1062 			if (error)
1063 				break;
1064 			if (np->n_flag & NDONTCACHE) {
1065 				error = nfs_vinvalbuf(vp, V_SAVE, 1);
1066 				if (error)
1067 					break;
1068 			}
1069 		} else if ((n + on) == biosize && nfs_async) {
1070 			bawrite(bp);
1071 		} else {
1072 			bdwrite(bp);
1073 		}
1074 	} while (uio->uio_resid > 0 && n > 0);
1075 
1076 	if (haverslock)
1077 		nfs_rsunlock(np);
1078 
1079 	return (error);
1080 }
1081 
1082 /*
1083  * Get an nfs cache block.
1084  *
1085  * Allocate a new one if the block isn't currently in the cache
1086  * and return the block marked busy. If the calling process is
1087  * interrupted by a signal for an interruptible mount point, return
1088  * NULL.
1089  *
1090  * The caller must carefully deal with the possible B_INVAL state of
1091  * the buffer.  nfs_startio() clears B_INVAL (and nfs_asyncio() clears it
1092  * indirectly), so synchronous reads can be issued without worrying about
1093  * the B_INVAL state.  We have to be a little more careful when dealing
1094  * with writes (see comments in nfs_write()) when extending a file past
1095  * its EOF.
1096  */
1097 static struct buf *
1098 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
1099 {
1100 	struct buf *bp;
1101 	struct mount *mp;
1102 	struct nfsmount *nmp;
1103 
1104 	mp = vp->v_mount;
1105 	nmp = VFSTONFS(mp);
1106 
1107 	if (nmp->nm_flag & NFSMNT_INT) {
1108 		bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
1109 		while (bp == NULL) {
1110 			if (nfs_sigintr(nmp, NULL, td))
1111 				return (NULL);
1112 			bp = getblk(vp, loffset, size, 0, 2 * hz);
1113 		}
1114 	} else {
1115 		bp = getblk(vp, loffset, size, 0, 0);
1116 	}
1117 
1118 	/*
1119 	 * bio2, the 'device' layer.  Since BIOs use 64 bit byte offsets
1120 	 * now, no translation is necessary.
1121 	 */
1122 	bp->b_bio2.bio_offset = loffset;
1123 	return (bp);
1124 }
1125 
1126 /*
1127  * Flush and invalidate all dirty buffers. If another process is already
1128  * doing the flush, just wait for completion.
1129  */
1130 int
1131 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
1132 {
1133 	struct nfsnode *np = VTONFS(vp);
1134 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1135 	int error = 0, slpflag, slptimeo;
1136 	thread_t td = curthread;
1137 
1138 	if (vp->v_flag & VRECLAIMED)
1139 		return (0);
1140 
1141 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1142 		intrflg = 0;
1143 	if (intrflg) {
1144 		slpflag = PCATCH;
1145 		slptimeo = 2 * hz;
1146 	} else {
1147 		slpflag = 0;
1148 		slptimeo = 0;
1149 	}
1150 	/*
1151 	 * First wait for any other process doing a flush to complete.
1152 	 */
1153 	while (np->n_flag & NFLUSHINPROG) {
1154 		np->n_flag |= NFLUSHWANT;
1155 		error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1156 		if (error && intrflg && nfs_sigintr(nmp, NULL, td))
1157 			return (EINTR);
1158 	}
1159 
1160 	/*
1161 	 * Now, flush as required.
1162 	 */
1163 	np->n_flag |= NFLUSHINPROG;
1164 	error = vinvalbuf(vp, flags, slpflag, 0);
1165 	while (error) {
1166 		if (intrflg && nfs_sigintr(nmp, NULL, td)) {
1167 			np->n_flag &= ~NFLUSHINPROG;
1168 			if (np->n_flag & NFLUSHWANT) {
1169 				np->n_flag &= ~NFLUSHWANT;
1170 				wakeup((caddr_t)&np->n_flag);
1171 			}
1172 			return (EINTR);
1173 		}
1174 		error = vinvalbuf(vp, flags, 0, slptimeo);
1175 	}
1176 	np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
1177 	if (np->n_flag & NFLUSHWANT) {
1178 		np->n_flag &= ~NFLUSHWANT;
1179 		wakeup((caddr_t)&np->n_flag);
1180 	}
1181 	return (0);
1182 }
1183 
1184 /*
1185  * Return true (non-zero) if the txthread and rxthread are operational
1186  * and we do not already have too many not-yet-started BIO's built up.
1187  */
1188 int
1189 nfs_asyncok(struct nfsmount *nmp)
1190 {
1191 	return (nmp->nm_bioqlen < nfs_maxasyncbio &&
1192 		nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE &&
1193 		nmp->nm_rxstate <= NFSSVC_PENDING &&
1194 		nmp->nm_txstate <= NFSSVC_PENDING);
1195 }
1196 
1197 /*
1198  * The read-ahead code calls this to queue a bio to the txthread.
1199  *
1200  * We don't touch the bio otherwise... that is, we do not even
1201  * construct or send the initial rpc.  The txthread will do it
1202  * for us.
1203  *
1204  * NOTE!  nm_bioqlen is not decremented until the request completes,
1205  *	  so it does not reflect the number of bio's on bioq.
1206  */
1207 void
1208 nfs_asyncio(struct vnode *vp, struct bio *bio)
1209 {
1210 	struct buf *bp = bio->bio_buf;
1211 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1212 
1213 	KKASSERT(vp->v_tag == VT_NFS);
1214 	BUF_KERNPROC(bp);
1215 	bio->bio_driver_info = vp;
1216 	crit_enter();
1217 	TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
1218 	atomic_add_int(&nmp->nm_bioqlen, 1);
1219 	crit_exit();
1220 	nfssvc_iod_writer_wakeup(nmp);
1221 }
1222 
1223 /*
1224  * nfs_dio()	- Execute a BIO operation synchronously.  The BIO will be
1225  *		  completed and its error returned.  The caller is responsible
1226  *		  for brelse()ing it.  ONLY USE FOR BIO_SYNC IOs!  Otherwise
1227  *		  our error probe will be against an invalid pointer.
1228  *
1229  * nfs_startio()- Execute a BIO operation assynchronously.
1230  *
1231  * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation,
1232  *	 which basically just queues it to the txthread.  nfs_startio()
1233  *	 actually initiates the I/O AFTER it has gotten to the txthread.
1234  *
1235  * NOTE: td might be NULL.
1236  */
1237 void
1238 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td)
1239 {
1240 	struct buf *bp = bio->bio_buf;
1241 	struct nfsnode *np;
1242 	struct nfsmount *nmp;
1243 
1244 	KKASSERT(vp->v_tag == VT_NFS);
1245 	np = VTONFS(vp);
1246 	nmp = VFSTONFS(vp->v_mount);
1247 
1248 	/*
1249 	 * clear B_ERROR and B_INVAL state prior to initiating the I/O.  We
1250 	 * do this here so we do not have to do it in all the code that
1251 	 * calls us.
1252 	 */
1253 	bp->b_flags &= ~(B_ERROR | B_INVAL);
1254 
1255 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
1256 		("nfs_doio: bp %p already marked done!", bp));
1257 
1258 	if (bp->b_cmd == BUF_CMD_READ) {
1259 	    switch (vp->v_type) {
1260 	    case VREG:
1261 		nfsstats.read_bios++;
1262 		nfs_readrpc_bio(vp, bio);
1263 		break;
1264 	    case VLNK:
1265 #if 0
1266 		bio->bio_offset = 0;
1267 		nfsstats.readlink_bios++;
1268 		nfs_readlinkrpc_bio(vp, bio);
1269 #else
1270 		nfs_doio(vp, bio, td);
1271 #endif
1272 		break;
1273 	    case VDIR:
1274 		/*
1275 		 * NOTE: If nfs_readdirplusrpc_bio() is requested but
1276 		 *	 not supported, it will chain to
1277 		 *	 nfs_readdirrpc_bio().
1278 		 */
1279 #if 0
1280 		nfsstats.readdir_bios++;
1281 		uiop->uio_offset = bio->bio_offset;
1282 		if (nmp->nm_flag & NFSMNT_RDIRPLUS)
1283 			nfs_readdirplusrpc_bio(vp, bio);
1284 		else
1285 			nfs_readdirrpc_bio(vp, bio);
1286 #else
1287 		nfs_doio(vp, bio, td);
1288 #endif
1289 		break;
1290 	    default:
1291 		kprintf("nfs_doio:  type %x unexpected\n",vp->v_type);
1292 		bp->b_flags |= B_ERROR;
1293 		bp->b_error = EINVAL;
1294 		biodone(bio);
1295 		break;
1296 	    }
1297 	} else {
1298 	    /*
1299 	     * If we only need to commit, try to commit.  If this fails
1300 	     * it will chain through to the write.  Basically all the logic
1301 	     * in nfs_doio() is replicated.
1302 	     */
1303 	    KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1304 	    if (bp->b_flags & B_NEEDCOMMIT)
1305 		nfs_commitrpc_bio(vp, bio);
1306 	    else
1307 		nfs_writerpc_bio(vp, bio);
1308 	}
1309 }
1310 
1311 int
1312 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1313 {
1314 	struct buf *bp = bio->bio_buf;
1315 	struct uio *uiop;
1316 	struct nfsnode *np;
1317 	struct nfsmount *nmp;
1318 	int error = 0;
1319 	int iomode, must_commit;
1320 	size_t n;
1321 	struct uio uio;
1322 	struct iovec io;
1323 
1324 	KKASSERT(vp->v_tag == VT_NFS);
1325 	np = VTONFS(vp);
1326 	nmp = VFSTONFS(vp->v_mount);
1327 	uiop = &uio;
1328 	uiop->uio_iov = &io;
1329 	uiop->uio_iovcnt = 1;
1330 	uiop->uio_segflg = UIO_SYSSPACE;
1331 	uiop->uio_td = td;
1332 
1333 	/*
1334 	 * clear B_ERROR and B_INVAL state prior to initiating the I/O.  We
1335 	 * do this here so we do not have to do it in all the code that
1336 	 * calls us.
1337 	 */
1338 	bp->b_flags &= ~(B_ERROR | B_INVAL);
1339 
1340 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
1341 		("nfs_doio: bp %p already marked done!", bp));
1342 
1343 	if (bp->b_cmd == BUF_CMD_READ) {
1344 	    io.iov_len = uiop->uio_resid = (size_t)bp->b_bcount;
1345 	    io.iov_base = bp->b_data;
1346 	    uiop->uio_rw = UIO_READ;
1347 
1348 	    switch (vp->v_type) {
1349 	    case VREG:
1350 		/*
1351 		 * When reading from a regular file zero-fill any residual.
1352 		 * Note that this residual has nothing to do with NFS short
1353 		 * reads, which nfs_readrpc_uio() will handle for us.
1354 		 *
1355 		 * We have to do this because when we are write extending
1356 		 * a file the server may not have the same notion of
1357 		 * filesize as we do.  Our BIOs should already be sized
1358 		 * (b_bcount) to account for the file EOF.
1359 		 */
1360 		nfsstats.read_bios++;
1361 		uiop->uio_offset = bio->bio_offset;
1362 		error = nfs_readrpc_uio(vp, uiop);
1363 		if (error == 0 && uiop->uio_resid) {
1364 			n = (size_t)bp->b_bcount - uiop->uio_resid;
1365 			bzero(bp->b_data + n, bp->b_bcount - n);
1366 			uiop->uio_resid = 0;
1367 		}
1368 		if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1369 		    np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1370 			uprintf("Process killed due to text file modification\n");
1371 			ksignal(td->td_proc, SIGKILL);
1372 		}
1373 		break;
1374 	    case VLNK:
1375 		uiop->uio_offset = 0;
1376 		nfsstats.readlink_bios++;
1377 		error = nfs_readlinkrpc_uio(vp, uiop);
1378 		break;
1379 	    case VDIR:
1380 		nfsstats.readdir_bios++;
1381 		uiop->uio_offset = bio->bio_offset;
1382 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1383 			error = nfs_readdirplusrpc_uio(vp, uiop);
1384 			if (error == NFSERR_NOTSUPP)
1385 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1386 		}
1387 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1388 			error = nfs_readdirrpc_uio(vp, uiop);
1389 		/*
1390 		 * end-of-directory sets B_INVAL but does not generate an
1391 		 * error.
1392 		 */
1393 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1394 			bp->b_flags |= B_INVAL;
1395 		break;
1396 	    default:
1397 		kprintf("nfs_doio:  type %x unexpected\n",vp->v_type);
1398 		break;
1399 	    };
1400 	    if (error) {
1401 		bp->b_flags |= B_ERROR;
1402 		bp->b_error = error;
1403 	    }
1404 	    bp->b_resid = uiop->uio_resid;
1405 	} else {
1406 	    /*
1407 	     * If we only need to commit, try to commit
1408 	     */
1409 	    KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1410 	    if (bp->b_flags & B_NEEDCOMMIT) {
1411 		    int retv;
1412 		    off_t off;
1413 
1414 		    off = bio->bio_offset + bp->b_dirtyoff;
1415 		    retv = nfs_commitrpc_uio(vp, off,
1416 					     bp->b_dirtyend - bp->b_dirtyoff,
1417 					     td);
1418 		    if (retv == 0) {
1419 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1420 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1421 			    bp->b_resid = 0;
1422 			    biodone(bio);
1423 			    return(0);
1424 		    }
1425 		    if (retv == NFSERR_STALEWRITEVERF) {
1426 			    nfs_clearcommit(vp->v_mount);
1427 		    }
1428 	    }
1429 
1430 	    /*
1431 	     * Setup for actual write
1432 	     */
1433 	    if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1434 		bp->b_dirtyend = np->n_size - bio->bio_offset;
1435 
1436 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1437 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1438 		    - bp->b_dirtyoff;
1439 		uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1440 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1441 		uiop->uio_rw = UIO_WRITE;
1442 		nfsstats.write_bios++;
1443 
1444 		if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1445 		    iomode = NFSV3WRITE_UNSTABLE;
1446 		else
1447 		    iomode = NFSV3WRITE_FILESYNC;
1448 
1449 		must_commit = 0;
1450 		error = nfs_writerpc_uio(vp, uiop, &iomode, &must_commit);
1451 
1452 		/*
1453 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1454 		 * to cluster the buffers needing commit.  This will allow
1455 		 * the system to submit a single commit rpc for the whole
1456 		 * cluster.  We can do this even if the buffer is not 100%
1457 		 * dirty (relative to the NFS blocksize), so we optimize the
1458 		 * append-to-file-case.
1459 		 *
1460 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1461 		 * cleared because write clustering only works for commit
1462 		 * rpc's, not for the data portion of the write).
1463 		 */
1464 
1465 		if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1466 		    bp->b_flags |= B_NEEDCOMMIT;
1467 		    if (bp->b_dirtyoff == 0
1468 			&& bp->b_dirtyend == bp->b_bcount)
1469 			bp->b_flags |= B_CLUSTEROK;
1470 		} else {
1471 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1472 		}
1473 
1474 		/*
1475 		 * For an interrupted write, the buffer is still valid
1476 		 * and the write hasn't been pushed to the server yet,
1477 		 * so we can't set B_ERROR and report the interruption
1478 		 * by setting B_EINTR. For the async case, B_EINTR
1479 		 * is not relevant, so the rpc attempt is essentially
1480 		 * a noop.  For the case of a V3 write rpc not being
1481 		 * committed to stable storage, the block is still
1482 		 * dirty and requires either a commit rpc or another
1483 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1484 		 * the block is reused. This is indicated by setting
1485 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1486 		 *
1487 		 * If the buffer is marked B_PAGING, it does not reside on
1488 		 * the vp's paging queues so we cannot call bdirty().  The
1489 		 * bp in this case is not an NFS cache block so we should
1490 		 * be safe. XXX
1491 		 */
1492     		if (error == EINTR
1493 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1494 			crit_enter();
1495 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1496 			if ((bp->b_flags & B_PAGING) == 0)
1497 			    bdirty(bp);
1498 			if (error)
1499 			    bp->b_flags |= B_EINTR;
1500 			crit_exit();
1501 	    	} else {
1502 		    if (error) {
1503 			bp->b_flags |= B_ERROR;
1504 			bp->b_error = np->n_error = error;
1505 			np->n_flag |= NWRITEERR;
1506 		    }
1507 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1508 		}
1509 		if (must_commit)
1510 		    nfs_clearcommit(vp->v_mount);
1511 		bp->b_resid = uiop->uio_resid;
1512 	    } else {
1513 		bp->b_resid = 0;
1514 	    }
1515 	}
1516 
1517 	/*
1518 	 * I/O was run synchronously, biodone() it and calculate the
1519 	 * error to return.
1520 	 */
1521 	biodone(bio);
1522 	KKASSERT(bp->b_cmd == BUF_CMD_DONE);
1523 	if (bp->b_flags & B_EINTR)
1524 		return (EINTR);
1525 	if (bp->b_flags & B_ERROR)
1526 		return (bp->b_error ? bp->b_error : EIO);
1527 	return (0);
1528 }
1529 
1530 /*
1531  * Used to aid in handling ftruncate() operations on the NFS client side.
1532  * Truncation creates a number of special problems for NFS.  We have to
1533  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1534  * we have to properly handle VM pages or (potentially dirty) buffers
1535  * that straddle the truncation point.
1536  */
1537 
1538 int
1539 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1540 {
1541 	struct nfsnode *np = VTONFS(vp);
1542 	u_quad_t tsize = np->n_size;
1543 	int biosize = vp->v_mount->mnt_stat.f_iosize;
1544 	int error = 0;
1545 
1546 	np->n_size = nsize;
1547 
1548 	if (nsize < tsize) {
1549 		struct buf *bp;
1550 		off_t loffset;
1551 		int bufsize;
1552 
1553 		/*
1554 		 * vtruncbuf() doesn't get the buffer overlapping the
1555 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1556 		 * buffer that now needs to be truncated.
1557 		 */
1558 		error = vtruncbuf(vp, nsize, biosize);
1559 		bufsize = nsize & (biosize - 1);
1560 		loffset = nsize - bufsize;
1561 		bp = nfs_getcacheblk(vp, loffset, bufsize, td);
1562 		if (bp->b_dirtyoff > bp->b_bcount)
1563 			bp->b_dirtyoff = bp->b_bcount;
1564 		if (bp->b_dirtyend > bp->b_bcount)
1565 			bp->b_dirtyend = bp->b_bcount;
1566 		bp->b_flags |= B_RELBUF;    /* don't leave garbage around */
1567 		brelse(bp);
1568 	} else {
1569 		vnode_pager_setsize(vp, nsize);
1570 	}
1571 	return(error);
1572 }
1573 
1574 /*
1575  * Synchronous completion for nfs_doio.  Call bpdone() with elseit=FALSE.
1576  * Caller is responsible for brelse()'ing the bp.
1577  */
1578 static void
1579 nfsiodone_sync(struct bio *bio)
1580 {
1581 	bio->bio_flags = 0;
1582 	bpdone(bio->bio_buf, 0);
1583 }
1584 
1585 /*
1586  * nfs read rpc - BIO version
1587  */
1588 void
1589 nfs_readrpc_bio(struct vnode *vp, struct bio *bio)
1590 {
1591 	struct buf *bp = bio->bio_buf;
1592 	u_int32_t *tl;
1593 	struct nfsmount *nmp;
1594 	int error = 0, len, tsiz;
1595 	struct nfsm_info *info;
1596 
1597 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1598 	info->mrep = NULL;
1599 	info->v3 = NFS_ISV3(vp);
1600 
1601 	nmp = VFSTONFS(vp->v_mount);
1602 	tsiz = bp->b_bcount;
1603 	KKASSERT(tsiz <= nmp->nm_rsize);
1604 	if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) {
1605 		error = EFBIG;
1606 		goto nfsmout;
1607 	}
1608 	nfsstats.rpccnt[NFSPROC_READ]++;
1609 	len = tsiz;
1610 	nfsm_reqhead(info, vp, NFSPROC_READ,
1611 		     NFSX_FH(info->v3) + NFSX_UNSIGNED * 3);
1612 	ERROROUT(nfsm_fhtom(info, vp));
1613 	tl = nfsm_build(info, NFSX_UNSIGNED * 3);
1614 	if (info->v3) {
1615 		txdr_hyper(bio->bio_offset, tl);
1616 		*(tl + 2) = txdr_unsigned(len);
1617 	} else {
1618 		*tl++ = txdr_unsigned(bio->bio_offset);
1619 		*tl++ = txdr_unsigned(len);
1620 		*tl = 0;
1621 	}
1622 	info->bio = bio;
1623 	info->done = nfs_readrpc_bio_done;
1624 	nfsm_request_bio(info, vp, NFSPROC_READ, NULL,
1625 			 nfs_vpcred(vp, ND_READ));
1626 	return;
1627 nfsmout:
1628 	kfree(info, M_NFSREQ);
1629 	bp->b_error = error;
1630 	bp->b_flags |= B_ERROR;
1631 	biodone(bio);
1632 }
1633 
1634 static void
1635 nfs_readrpc_bio_done(nfsm_info_t info)
1636 {
1637 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1638 	struct bio *bio = info->bio;
1639 	struct buf *bp = bio->bio_buf;
1640 	u_int32_t *tl;
1641 	int attrflag;
1642 	int retlen;
1643 	int eof;
1644 	int error = 0;
1645 
1646 	KKASSERT(info->state == NFSM_STATE_DONE);
1647 
1648 	if (info->v3) {
1649 		ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag,
1650 					 NFS_LATTR_NOSHRINK));
1651 		NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED));
1652 		eof = fxdr_unsigned(int, *(tl + 1));
1653 	} else {
1654 		ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1655 		eof = 0;
1656 	}
1657 	NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize));
1658 	ERROROUT(nfsm_mtobio(info, bio, retlen));
1659 	m_freem(info->mrep);
1660 	info->mrep = NULL;
1661 
1662 	/*
1663 	 * No error occured, if retlen is less then bcount and no EOF
1664 	 * and NFSv3 a zero-fill short read occured.
1665 	 *
1666 	 * For NFSv2 a short-read indicates EOF.
1667 	 */
1668 	if (retlen < bp->b_bcount && info->v3 && eof == 0) {
1669 		bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1670 		retlen = bp->b_bcount;
1671 	}
1672 
1673 	/*
1674 	 * If we hit an EOF we still zero-fill, but return the expected
1675 	 * b_resid anyway.  This should normally not occur since async
1676 	 * BIOs are not used for read-before-write case.  Races against
1677 	 * the server can cause it though and we don't want to leave
1678 	 * garbage in the buffer.
1679 	 */
1680 	if (retlen < bp->b_bcount) {
1681 		bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1682 	}
1683 	bp->b_resid = 0;
1684 	/* bp->b_resid = bp->b_bcount - retlen; */
1685 nfsmout:
1686 	kfree(info, M_NFSREQ);
1687 	if (error) {
1688 		bp->b_error = error;
1689 		bp->b_flags |= B_ERROR;
1690 	}
1691 	biodone(bio);
1692 }
1693 
1694 /*
1695  * nfs write call - BIO version
1696  */
1697 void
1698 nfs_writerpc_bio(struct vnode *vp, struct bio *bio)
1699 {
1700 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1701 	struct nfsnode *np = VTONFS(vp);
1702 	struct buf *bp = bio->bio_buf;
1703 	u_int32_t *tl;
1704 	int len;
1705 	int iomode;
1706 	int error = 0;
1707 	struct nfsm_info *info;
1708 	off_t offset;
1709 
1710 	/*
1711 	 * Setup for actual write.  Just clean up the bio if there
1712 	 * is nothing to do.
1713 	 */
1714 	if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1715 		bp->b_dirtyend = np->n_size - bio->bio_offset;
1716 
1717 	if (bp->b_dirtyend <= bp->b_dirtyoff) {
1718 		bp->b_resid = 0;
1719 		biodone(bio);
1720 		return;
1721 	}
1722 	len = bp->b_dirtyend - bp->b_dirtyoff;
1723 	offset = bio->bio_offset + bp->b_dirtyoff;
1724 	if (offset + len > nmp->nm_maxfilesize) {
1725 		bp->b_flags |= B_ERROR;
1726 		bp->b_error = EFBIG;
1727 		biodone(bio);
1728 		return;
1729 	}
1730 	bp->b_resid = len;
1731 	nfsstats.write_bios++;
1732 
1733 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1734 	info->mrep = NULL;
1735 	info->v3 = NFS_ISV3(vp);
1736 	info->info_writerpc.must_commit = 0;
1737 	if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1738 		iomode = NFSV3WRITE_UNSTABLE;
1739 	else
1740 		iomode = NFSV3WRITE_FILESYNC;
1741 
1742 	KKASSERT(len <= nmp->nm_wsize);
1743 
1744 	nfsstats.rpccnt[NFSPROC_WRITE]++;
1745 	nfsm_reqhead(info, vp, NFSPROC_WRITE,
1746 		     NFSX_FH(info->v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len));
1747 	ERROROUT(nfsm_fhtom(info, vp));
1748 	if (info->v3) {
1749 		tl = nfsm_build(info, 5 * NFSX_UNSIGNED);
1750 		txdr_hyper(offset, tl);
1751 		tl += 2;
1752 		*tl++ = txdr_unsigned(len);
1753 		*tl++ = txdr_unsigned(iomode);
1754 		*tl = txdr_unsigned(len);
1755 	} else {
1756 		u_int32_t x;
1757 
1758 		tl = nfsm_build(info, 4 * NFSX_UNSIGNED);
1759 		/* Set both "begin" and "current" to non-garbage. */
1760 		x = txdr_unsigned((u_int32_t)offset);
1761 		*tl++ = x;	/* "begin offset" */
1762 		*tl++ = x;	/* "current offset" */
1763 		x = txdr_unsigned(len);
1764 		*tl++ = x;	/* total to this offset */
1765 		*tl = x;	/* size of this write */
1766 	}
1767 	ERROROUT(nfsm_biotom(info, bio, bp->b_dirtyoff, len));
1768 	info->bio = bio;
1769 	info->done = nfs_writerpc_bio_done;
1770 	nfsm_request_bio(info, vp, NFSPROC_WRITE, NULL,
1771 			 nfs_vpcred(vp, ND_WRITE));
1772 	return;
1773 nfsmout:
1774 	kfree(info, M_NFSREQ);
1775 	bp->b_error = error;
1776 	bp->b_flags |= B_ERROR;
1777 	biodone(bio);
1778 }
1779 
1780 static void
1781 nfs_writerpc_bio_done(nfsm_info_t info)
1782 {
1783 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1784 	struct nfsnode *np = VTONFS(info->vp);
1785 	struct bio *bio = info->bio;
1786 	struct buf *bp = bio->bio_buf;
1787 	int wccflag = NFSV3_WCCRATTR;
1788 	int iomode = NFSV3WRITE_FILESYNC;
1789 	int commit;
1790 	int rlen;
1791 	int error;
1792 	int len = bp->b_resid;	/* b_resid was set to shortened length */
1793 	u_int32_t *tl;
1794 
1795 	if (info->v3) {
1796 		/*
1797 		 * The write RPC returns a before and after mtime.  The
1798 		 * nfsm_wcc_data() macro checks the before n_mtime
1799 		 * against the before time and stores the after time
1800 		 * in the nfsnode's cached vattr and n_mtime field.
1801 		 * The NRMODIFIED bit will be set if the before
1802 		 * time did not match the original mtime.
1803 		 */
1804 		wccflag = NFSV3_WCCCHK;
1805 		ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1806 		if (error == 0) {
1807 			NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF));
1808 			rlen = fxdr_unsigned(int, *tl++);
1809 			if (rlen == 0) {
1810 				error = NFSERR_IO;
1811 				m_freem(info->mrep);
1812 				info->mrep = NULL;
1813 				goto nfsmout;
1814 			} else if (rlen < len) {
1815 #if 0
1816 				/*
1817 				 * XXX what do we do here?
1818 				 */
1819 				backup = len - rlen;
1820 				uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup;
1821 				uiop->uio_iov->iov_len += backup;
1822 				uiop->uio_offset -= backup;
1823 				uiop->uio_resid += backup;
1824 				len = rlen;
1825 #endif
1826 			}
1827 			commit = fxdr_unsigned(int, *tl++);
1828 
1829 			/*
1830 			 * Return the lowest committment level
1831 			 * obtained by any of the RPCs.
1832 			 */
1833 			if (iomode == NFSV3WRITE_FILESYNC)
1834 				iomode = commit;
1835 			else if (iomode == NFSV3WRITE_DATASYNC &&
1836 				commit == NFSV3WRITE_UNSTABLE)
1837 				iomode = commit;
1838 			if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){
1839 			    bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1840 			    nmp->nm_state |= NFSSTA_HASWRITEVERF;
1841 			} else if (bcmp(tl, nmp->nm_verf, NFSX_V3WRITEVERF)) {
1842 			    info->info_writerpc.must_commit = 1;
1843 			    bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1844 			}
1845 		}
1846 	} else {
1847 		ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1848 	}
1849 	m_freem(info->mrep);
1850 	info->mrep = NULL;
1851 	len = 0;
1852 nfsmout:
1853 	if (info->vp->v_mount->mnt_flag & MNT_ASYNC)
1854 		iomode = NFSV3WRITE_FILESYNC;
1855 	bp->b_resid = len;
1856 
1857 	/*
1858 	 * End of RPC.  Now clean up the bp.
1859 	 *
1860 	 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1861 	 * to cluster the buffers needing commit.  This will allow
1862 	 * the system to submit a single commit rpc for the whole
1863 	 * cluster.  We can do this even if the buffer is not 100%
1864 	 * dirty (relative to the NFS blocksize), so we optimize the
1865 	 * append-to-file-case.
1866 	 *
1867 	 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1868 	 * cleared because write clustering only works for commit
1869 	 * rpc's, not for the data portion of the write).
1870 	 */
1871 	if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1872 		bp->b_flags |= B_NEEDCOMMIT;
1873 		if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount)
1874 			bp->b_flags |= B_CLUSTEROK;
1875 	} else {
1876 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1877 	}
1878 
1879 	/*
1880 	 * For an interrupted write, the buffer is still valid
1881 	 * and the write hasn't been pushed to the server yet,
1882 	 * so we can't set B_ERROR and report the interruption
1883 	 * by setting B_EINTR. For the async case, B_EINTR
1884 	 * is not relevant, so the rpc attempt is essentially
1885 	 * a noop.  For the case of a V3 write rpc not being
1886 	 * committed to stable storage, the block is still
1887 	 * dirty and requires either a commit rpc or another
1888 	 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1889 	 * the block is reused. This is indicated by setting
1890 	 * the B_DELWRI and B_NEEDCOMMIT flags.
1891 	 *
1892 	 * If the buffer is marked B_PAGING, it does not reside on
1893 	 * the vp's paging queues so we cannot call bdirty().  The
1894 	 * bp in this case is not an NFS cache block so we should
1895 	 * be safe. XXX
1896 	 */
1897 	if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1898 		crit_enter();
1899 		bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1900 		if ((bp->b_flags & B_PAGING) == 0)
1901 			bdirty(bp);
1902 		if (error)
1903 			bp->b_flags |= B_EINTR;
1904 		crit_exit();
1905 	} else {
1906 		if (error) {
1907 			bp->b_flags |= B_ERROR;
1908 			bp->b_error = np->n_error = error;
1909 			np->n_flag |= NWRITEERR;
1910 		}
1911 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1912 	}
1913 	if (info->info_writerpc.must_commit)
1914 		nfs_clearcommit(info->vp->v_mount);
1915 	kfree(info, M_NFSREQ);
1916 	if (error) {
1917 		bp->b_flags |= B_ERROR;
1918 		bp->b_error = error;
1919 	}
1920 	biodone(bio);
1921 }
1922 
1923 /*
1924  * Nfs Version 3 commit rpc - BIO version
1925  *
1926  * This function issues the commit rpc and will chain to a write
1927  * rpc if necessary.
1928  */
1929 void
1930 nfs_commitrpc_bio(struct vnode *vp, struct bio *bio)
1931 {
1932 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1933 	struct buf *bp = bio->bio_buf;
1934 	struct nfsm_info *info;
1935 	int error = 0;
1936 	u_int32_t *tl;
1937 
1938 	if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) {
1939 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1940 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1941 		bp->b_resid = 0;
1942 		biodone(bio);
1943 		return;
1944 	}
1945 
1946 	info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1947 	info->mrep = NULL;
1948 	info->v3 = 1;
1949 
1950 	nfsstats.rpccnt[NFSPROC_COMMIT]++;
1951 	nfsm_reqhead(info, vp, NFSPROC_COMMIT, NFSX_FH(1));
1952 	ERROROUT(nfsm_fhtom(info, vp));
1953 	tl = nfsm_build(info, 3 * NFSX_UNSIGNED);
1954 	txdr_hyper(bio->bio_offset + bp->b_dirtyoff, tl);
1955 	tl += 2;
1956 	*tl = txdr_unsigned(bp->b_dirtyend - bp->b_dirtyoff);
1957 	info->bio = bio;
1958 	info->done = nfs_commitrpc_bio_done;
1959 	nfsm_request_bio(info, vp, NFSPROC_COMMIT, NULL,
1960 			 nfs_vpcred(vp, ND_WRITE));
1961 	return;
1962 nfsmout:
1963 	/*
1964 	 * Chain to write RPC on (early) error
1965 	 */
1966 	kfree(info, M_NFSREQ);
1967 	nfs_writerpc_bio(vp, bio);
1968 }
1969 
1970 static void
1971 nfs_commitrpc_bio_done(nfsm_info_t info)
1972 {
1973 	struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1974 	struct bio *bio = info->bio;
1975 	struct buf *bp = bio->bio_buf;
1976 	u_int32_t *tl;
1977 	int wccflag = NFSV3_WCCRATTR;
1978 	int error = 0;
1979 
1980 	ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1981 	if (error == 0) {
1982 		NULLOUT(tl = nfsm_dissect(info, NFSX_V3WRITEVERF));
1983 		if (bcmp(nmp->nm_verf, tl, NFSX_V3WRITEVERF)) {
1984 			bcopy(tl, nmp->nm_verf, NFSX_V3WRITEVERF);
1985 			error = NFSERR_STALEWRITEVERF;
1986 		}
1987 	}
1988 	m_freem(info->mrep);
1989 	info->mrep = NULL;
1990 
1991 	/*
1992 	 * On completion we must chain to a write bio if an
1993 	 * error occurred.
1994 	 */
1995 nfsmout:
1996 	kfree(info, M_NFSREQ);
1997 	if (error == 0) {
1998 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1999 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
2000 		bp->b_resid = 0;
2001 		biodone(bio);
2002 	} else {
2003 		nfs_writerpc_bio(info->vp, bio);
2004 	}
2005 }
2006 
2007