xref: /dflybsd-src/sys/vfs/nfs/nfs_bio.c (revision 53e987cee557d989dbf172d8a3c2ade9ea6fc46f)
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 1.22 2005/04/15 19:08:21 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/buf2.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 "rpcv2.h"
62 #include "nfsproto.h"
63 #include "nfs.h"
64 #include "nfsmount.h"
65 #include "nqnfs.h"
66 #include "nfsnode.h"
67 
68 static struct buf *nfs_getcacheblk (struct vnode *vp, daddr_t bn, int size,
69 					struct thread *td);
70 
71 extern int nfs_numasync;
72 extern int nfs_pbuf_freecnt;
73 extern struct nfsstats nfsstats;
74 
75 /*
76  * Vnode op for VM getpages.
77  *
78  * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
79  *		int a_reqpage, vm_ooffset_t a_offset)
80  */
81 int
82 nfs_getpages(struct vop_getpages_args *ap)
83 {
84 	struct thread *td = curthread;		/* XXX */
85 	int i, error, nextoff, size, toff, count, npages;
86 	struct uio uio;
87 	struct iovec iov;
88 	char *kva;
89 	struct vnode *vp;
90 	struct nfsmount *nmp;
91 	vm_page_t *pages;
92 	vm_page_t m;
93 	struct msf_buf *msf;
94 
95 	vp = ap->a_vp;
96 	nmp = VFSTONFS(vp->v_mount);
97 	pages = ap->a_m;
98 	count = ap->a_count;
99 
100 	if (vp->v_object == NULL) {
101 		printf("nfs_getpages: called with non-merged cache vnode??\n");
102 		return VM_PAGER_ERROR;
103 	}
104 
105 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
106 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
107 		(void)nfs_fsinfo(nmp, vp, td);
108 
109 	npages = btoc(count);
110 
111 	/*
112 	 * NOTE that partially valid pages may occur in cases other
113 	 * then file EOF, such as when a file is partially written and
114 	 * ftruncate()-extended to a larger size.   It is also possible
115 	 * for the valid bits to be set on garbage beyond the file EOF and
116 	 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
117 	 * occur due to vtruncbuf() and the buffer cache's handling of
118 	 * pages which 'straddle' buffers or when b_bufsize is not a
119 	 * multiple of PAGE_SIZE.... the buffer cache cannot normally
120 	 * clear the extra bits.  This kind of situation occurs when you
121 	 * make a small write() (m->valid == 0x03) and then mmap() and
122 	 * fault in the buffer(m->valid = 0xFF).  When NFS flushes the
123 	 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
124 	 *
125 	 * This is combined with the possibility that the pages are partially
126 	 * dirty or that there is a buffer backing the pages that is dirty
127 	 * (even if m->dirty is 0).
128 	 *
129 	 * To solve this problem several hacks have been made:  (1) NFS
130 	 * guarentees that the IO block size is a multiple of PAGE_SIZE and
131 	 * (2) The buffer cache, when invalidating an NFS buffer, will
132 	 * disregard the buffer's fragmentory b_bufsize and invalidate
133 	 * the whole page rather then just the piece the buffer owns.
134 	 *
135 	 * This allows us to assume that a partially valid page found here
136 	 * is fully valid (vm_fault will zero'd out areas of the page not
137 	 * marked as valid).
138 	 */
139 	m = pages[ap->a_reqpage];
140 	if (m->valid != 0) {
141 		for (i = 0; i < npages; ++i) {
142 			if (i != ap->a_reqpage)
143 				vnode_pager_freepage(pages[i]);
144 		}
145 		return(0);
146 	}
147 
148 	/*
149 	 * Use an MSF_BUF as a medium to retrieve data from the pages.
150 	 */
151 	msf_map_pagelist(&msf, pages, npages, 0);
152 	KKASSERT(msf);
153 	kva = msf_buf_kva(msf);
154 
155 	iov.iov_base = kva;
156 	iov.iov_len = count;
157 	uio.uio_iov = &iov;
158 	uio.uio_iovcnt = 1;
159 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
160 	uio.uio_resid = count;
161 	uio.uio_segflg = UIO_SYSSPACE;
162 	uio.uio_rw = UIO_READ;
163 	uio.uio_td = td;
164 
165 	error = nfs_readrpc(vp, &uio);
166 	msf_buf_free(msf);
167 
168 	if (error && (uio.uio_resid == count)) {
169 		printf("nfs_getpages: error %d\n", error);
170 		for (i = 0; i < npages; ++i) {
171 			if (i != ap->a_reqpage)
172 				vnode_pager_freepage(pages[i]);
173 		}
174 		return VM_PAGER_ERROR;
175 	}
176 
177 	/*
178 	 * Calculate the number of bytes read and validate only that number
179 	 * of bytes.  Note that due to pending writes, size may be 0.  This
180 	 * does not mean that the remaining data is invalid!
181 	 */
182 
183 	size = count - uio.uio_resid;
184 
185 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
186 		nextoff = toff + PAGE_SIZE;
187 		m = pages[i];
188 
189 		m->flags &= ~PG_ZERO;
190 
191 		if (nextoff <= size) {
192 			/*
193 			 * Read operation filled an entire page
194 			 */
195 			m->valid = VM_PAGE_BITS_ALL;
196 			vm_page_undirty(m);
197 		} else if (size > toff) {
198 			/*
199 			 * Read operation filled a partial page.
200 			 */
201 			m->valid = 0;
202 			vm_page_set_validclean(m, 0, size - toff);
203 			/* handled by vm_fault now	  */
204 			/* vm_page_zero_invalid(m, TRUE); */
205 		} else {
206 			/*
207 			 * Read operation was short.  If no error occured
208 			 * we may have hit a zero-fill section.   We simply
209 			 * leave valid set to 0.
210 			 */
211 			;
212 		}
213 		if (i != ap->a_reqpage) {
214 			/*
215 			 * Whether or not to leave the page activated is up in
216 			 * the air, but we should put the page on a page queue
217 			 * somewhere (it already is in the object).  Result:
218 			 * It appears that emperical results show that
219 			 * deactivating pages is best.
220 			 */
221 
222 			/*
223 			 * Just in case someone was asking for this page we
224 			 * now tell them that it is ok to use.
225 			 */
226 			if (!error) {
227 				if (m->flags & PG_WANTED)
228 					vm_page_activate(m);
229 				else
230 					vm_page_deactivate(m);
231 				vm_page_wakeup(m);
232 			} else {
233 				vnode_pager_freepage(m);
234 			}
235 		}
236 	}
237 	return 0;
238 }
239 
240 /*
241  * Vnode op for VM putpages.
242  *
243  * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
244  *		int *a_rtvals, vm_ooffset_t a_offset)
245  */
246 int
247 nfs_putpages(struct vop_putpages_args *ap)
248 {
249 	struct thread *td = curthread;
250 	struct uio uio;
251 	struct iovec iov;
252 	char *kva;
253 	int iomode, must_commit, i, error, npages, count;
254 	off_t offset;
255 	int *rtvals;
256 	struct vnode *vp;
257 	struct nfsmount *nmp;
258 	struct nfsnode *np;
259 	vm_page_t *pages;
260 	struct msf_buf *msf;
261 
262 	vp = ap->a_vp;
263 	np = VTONFS(vp);
264 	nmp = VFSTONFS(vp->v_mount);
265 	pages = ap->a_m;
266 	count = ap->a_count;
267 	rtvals = ap->a_rtvals;
268 	npages = btoc(count);
269 	offset = IDX_TO_OFF(pages[0]->pindex);
270 
271 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
272 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
273 		(void)nfs_fsinfo(nmp, vp, td);
274 
275 	for (i = 0; i < npages; i++) {
276 		rtvals[i] = VM_PAGER_AGAIN;
277 	}
278 
279 	/*
280 	 * When putting pages, do not extend file past EOF.
281 	 */
282 
283 	if (offset + count > np->n_size) {
284 		count = np->n_size - offset;
285 		if (count < 0)
286 			count = 0;
287 	}
288 
289 	/*
290 	 * Use an MSF_BUF as a medium to retrieve data from the pages.
291 	 */
292 	msf_map_pagelist(&msf, pages, npages, 0);
293 	KKASSERT(msf);
294 	kva = msf_buf_kva(msf);
295 
296 	iov.iov_base = kva;
297 	iov.iov_len = count;
298 	uio.uio_iov = &iov;
299 	uio.uio_iovcnt = 1;
300 	uio.uio_offset = offset;
301 	uio.uio_resid = count;
302 	uio.uio_segflg = UIO_SYSSPACE;
303 	uio.uio_rw = UIO_WRITE;
304 	uio.uio_td = td;
305 
306 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
307 	    iomode = NFSV3WRITE_UNSTABLE;
308 	else
309 	    iomode = NFSV3WRITE_FILESYNC;
310 
311 	error = nfs_writerpc(vp, &uio, &iomode, &must_commit);
312 
313 	msf_buf_free(msf);
314 
315 	if (!error) {
316 		int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
317 		for (i = 0; i < nwritten; i++) {
318 			rtvals[i] = VM_PAGER_OK;
319 			vm_page_undirty(pages[i]);
320 		}
321 		if (must_commit)
322 			nfs_clearcommit(vp->v_mount);
323 	}
324 	return rtvals[0];
325 }
326 
327 /*
328  * Vnode op for read using bio
329  */
330 int
331 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
332 {
333 	struct nfsnode *np = VTONFS(vp);
334 	int biosize, i;
335 	struct buf *bp = 0, *rabp;
336 	struct vattr vattr;
337 	struct thread *td;
338 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
339 	daddr_t lbn, rabn;
340 	int bcount;
341 	int seqcount;
342 	int nra, error = 0, n = 0, on = 0;
343 
344 #ifdef DIAGNOSTIC
345 	if (uio->uio_rw != UIO_READ)
346 		panic("nfs_read mode");
347 #endif
348 	if (uio->uio_resid == 0)
349 		return (0);
350 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
351 		return (EINVAL);
352 	td = uio->uio_td;
353 
354 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
355 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
356 		(void)nfs_fsinfo(nmp, vp, td);
357 	if (vp->v_type != VDIR &&
358 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
359 		return (EFBIG);
360 	biosize = vp->v_mount->mnt_stat.f_iosize;
361 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
362 
363 	/*
364 	 * For nfs, cache consistency can only be maintained approximately.
365 	 * Although RFC1094 does not specify the criteria, the following is
366 	 * believed to be compatible with the reference port.
367 	 *
368 	 * NQNFS:	Full cache coherency is maintained within the loop.
369 	 *
370 	 * NFS:		If local changes have been made and this is a
371 	 *		directory, the directory must be invalidated and
372 	 *		the attribute cache must be cleared.
373 	 *
374 	 *		GETATTR is called to synchronize the file size.
375 	 *
376 	 *		If remote changes are detected local data is flushed
377 	 *		and the cache is invalidated.
378 	 *
379 	 *
380 	 *		NOTE: In the normal case the attribute cache is not
381 	 *		cleared which means GETATTR may use cached data and
382 	 *		not immediately detect changes made on the server.
383 	 */
384 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
385 		if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
386 			nfs_invaldir(vp);
387 			error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
388 			if (error)
389 				return (error);
390 			np->n_attrstamp = 0;
391 		}
392 		error = VOP_GETATTR(vp, &vattr, td);
393 		if (error)
394 			return (error);
395 		if (np->n_flag & NRMODIFIED) {
396 			if (vp->v_type == VDIR)
397 				nfs_invaldir(vp);
398 			error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
399 			if (error)
400 				return (error);
401 			np->n_flag &= ~NRMODIFIED;
402 		}
403 	}
404 	do {
405 
406 	    /*
407 	     * Get a valid lease. If cached data is stale, flush it.
408 	     */
409 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
410 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
411 		    do {
412 			error = nqnfs_getlease(vp, ND_READ, td);
413 		    } while (error == NQNFS_EXPIRED);
414 		    if (error)
415 			return (error);
416 		    if (np->n_lrev != np->n_brev ||
417 			(np->n_flag & NQNFSNONCACHE) ||
418 			((np->n_flag & NLMODIFIED) && vp->v_type == VDIR)) {
419 			if (vp->v_type == VDIR)
420 			    nfs_invaldir(vp);
421 			error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
422 			if (error)
423 			    return (error);
424 			np->n_brev = np->n_lrev;
425 		    }
426 		} else if (vp->v_type == VDIR && (np->n_flag & NLMODIFIED)) {
427 		    nfs_invaldir(vp);
428 		    error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
429 		    if (error)
430 			return (error);
431 		}
432 	    }
433 	    if (np->n_flag & NQNFSNONCACHE) {
434 		switch (vp->v_type) {
435 		case VREG:
436 			return (nfs_readrpc(vp, uio));
437 		case VLNK:
438 			return (nfs_readlinkrpc(vp, uio));
439 		case VDIR:
440 			break;
441 		default:
442 			printf(" NQNFSNONCACHE: type %x unexpected\n",
443 				vp->v_type);
444 		};
445 	    }
446 	    switch (vp->v_type) {
447 	    case VREG:
448 		nfsstats.biocache_reads++;
449 		lbn = uio->uio_offset / biosize;
450 		on = uio->uio_offset & (biosize - 1);
451 
452 		/*
453 		 * Start the read ahead(s), as required.
454 		 */
455 		if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
456 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
457 			(off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
458 			rabn = lbn + 1 + nra;
459 			if (!incore(vp, rabn)) {
460 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
461 			    if (!rabp)
462 				return (EINTR);
463 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
464 				rabp->b_flags |= (B_READ | B_ASYNC);
465 				vfs_busy_pages(rabp, 0);
466 				if (nfs_asyncio(rabp, td)) {
467 				    rabp->b_flags |= B_INVAL|B_ERROR;
468 				    vfs_unbusy_pages(rabp);
469 				    brelse(rabp);
470 				    break;
471 				}
472 			    } else {
473 				brelse(rabp);
474 			    }
475 			}
476 		    }
477 		}
478 
479 		/*
480 		 * Obtain the buffer cache block.  Figure out the buffer size
481 		 * when we are at EOF.  If we are modifying the size of the
482 		 * buffer based on an EOF condition we need to hold
483 		 * nfs_rslock() through obtaining the buffer to prevent
484 		 * a potential writer-appender from messing with n_size.
485 		 * Otherwise we may accidently truncate the buffer and
486 		 * lose dirty data.
487 		 *
488 		 * Note that bcount is *not* DEV_BSIZE aligned.
489 		 */
490 
491 again:
492 		bcount = biosize;
493 		if ((off_t)lbn * biosize >= np->n_size) {
494 			bcount = 0;
495 		} else if ((off_t)(lbn + 1) * biosize > np->n_size) {
496 			bcount = np->n_size - (off_t)lbn * biosize;
497 		}
498 		if (bcount != biosize) {
499 			switch(nfs_rslock(np, td)) {
500 			case ENOLCK:
501 				goto again;
502 				/* not reached */
503 			case EINTR:
504 			case ERESTART:
505 				return(EINTR);
506 				/* not reached */
507 			default:
508 				break;
509 			}
510 		}
511 
512 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
513 
514 		if (bcount != biosize)
515 			nfs_rsunlock(np, td);
516 		if (!bp)
517 			return (EINTR);
518 
519 		/*
520 		 * If B_CACHE is not set, we must issue the read.  If this
521 		 * fails, we return an error.
522 		 */
523 
524 		if ((bp->b_flags & B_CACHE) == 0) {
525 		    bp->b_flags |= B_READ;
526 		    vfs_busy_pages(bp, 0);
527 		    error = nfs_doio(bp, td);
528 		    if (error) {
529 			brelse(bp);
530 			return (error);
531 		    }
532 		}
533 
534 		/*
535 		 * on is the offset into the current bp.  Figure out how many
536 		 * bytes we can copy out of the bp.  Note that bcount is
537 		 * NOT DEV_BSIZE aligned.
538 		 *
539 		 * Then figure out how many bytes we can copy into the uio.
540 		 */
541 
542 		n = 0;
543 		if (on < bcount)
544 			n = min((unsigned)(bcount - on), uio->uio_resid);
545 		break;
546 	    case VLNK:
547 		nfsstats.biocache_readlinks++;
548 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
549 		if (!bp)
550 			return (EINTR);
551 		if ((bp->b_flags & B_CACHE) == 0) {
552 		    bp->b_flags |= B_READ;
553 		    vfs_busy_pages(bp, 0);
554 		    error = nfs_doio(bp, td);
555 		    if (error) {
556 			bp->b_flags |= B_ERROR;
557 			brelse(bp);
558 			return (error);
559 		    }
560 		}
561 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
562 		on = 0;
563 		break;
564 	    case VDIR:
565 		nfsstats.biocache_readdirs++;
566 		if (np->n_direofoffset
567 		    && uio->uio_offset >= np->n_direofoffset) {
568 		    return (0);
569 		}
570 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
571 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
572 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
573 		if (!bp)
574 		    return (EINTR);
575 		if ((bp->b_flags & B_CACHE) == 0) {
576 		    bp->b_flags |= B_READ;
577 		    vfs_busy_pages(bp, 0);
578 		    error = nfs_doio(bp, td);
579 		    if (error) {
580 			    brelse(bp);
581 		    }
582 		    while (error == NFSERR_BAD_COOKIE) {
583 			printf("got bad cookie vp %p bp %p\n", vp, bp);
584 			nfs_invaldir(vp);
585 			error = nfs_vinvalbuf(vp, 0, td, 1);
586 			/*
587 			 * Yuck! The directory has been modified on the
588 			 * server. The only way to get the block is by
589 			 * reading from the beginning to get all the
590 			 * offset cookies.
591 			 *
592 			 * Leave the last bp intact unless there is an error.
593 			 * Loop back up to the while if the error is another
594 			 * NFSERR_BAD_COOKIE (double yuch!).
595 			 */
596 			for (i = 0; i <= lbn && !error; i++) {
597 			    if (np->n_direofoffset
598 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
599 				    return (0);
600 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
601 			    if (!bp)
602 				return (EINTR);
603 			    if ((bp->b_flags & B_CACHE) == 0) {
604 				    bp->b_flags |= B_READ;
605 				    vfs_busy_pages(bp, 0);
606 				    error = nfs_doio(bp, td);
607 				    /*
608 				     * no error + B_INVAL == directory EOF,
609 				     * use the block.
610 				     */
611 				    if (error == 0 && (bp->b_flags & B_INVAL))
612 					    break;
613 			    }
614 			    /*
615 			     * An error will throw away the block and the
616 			     * for loop will break out.  If no error and this
617 			     * is not the block we want, we throw away the
618 			     * block and go for the next one via the for loop.
619 			     */
620 			    if (error || i < lbn)
621 				    brelse(bp);
622 			}
623 		    }
624 		    /*
625 		     * The above while is repeated if we hit another cookie
626 		     * error.  If we hit an error and it wasn't a cookie error,
627 		     * we give up.
628 		     */
629 		    if (error)
630 			    return (error);
631 		}
632 
633 		/*
634 		 * If not eof and read aheads are enabled, start one.
635 		 * (You need the current block first, so that you have the
636 		 *  directory offset cookie of the next block.)
637 		 */
638 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
639 		    (bp->b_flags & B_INVAL) == 0 &&
640 		    (np->n_direofoffset == 0 ||
641 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
642 		    !(np->n_flag & NQNFSNONCACHE) &&
643 		    !incore(vp, lbn + 1)) {
644 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
645 			if (rabp) {
646 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
647 				rabp->b_flags |= (B_READ | B_ASYNC);
648 				vfs_busy_pages(rabp, 0);
649 				if (nfs_asyncio(rabp, td)) {
650 				    rabp->b_flags |= B_INVAL|B_ERROR;
651 				    vfs_unbusy_pages(rabp);
652 				    brelse(rabp);
653 				}
654 			    } else {
655 				brelse(rabp);
656 			    }
657 			}
658 		}
659 		/*
660 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
661 		 * chopped for the EOF condition, we cannot tell how large
662 		 * NFS directories are going to be until we hit EOF.  So
663 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
664 		 * it just so happens that b_resid will effectively chop it
665 		 * to EOF.  *BUT* this information is lost if the buffer goes
666 		 * away and is reconstituted into a B_CACHE state ( due to
667 		 * being VMIO ) later.  So we keep track of the directory eof
668 		 * in np->n_direofoffset and chop it off as an extra step
669 		 * right here.
670 		 */
671 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
672 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
673 			n = np->n_direofoffset - uio->uio_offset;
674 		break;
675 	    default:
676 		printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
677 		break;
678 	    };
679 
680 	    if (n > 0) {
681 		    error = uiomove(bp->b_data + on, (int)n, uio);
682 	    }
683 	    switch (vp->v_type) {
684 	    case VREG:
685 		break;
686 	    case VLNK:
687 		n = 0;
688 		break;
689 	    case VDIR:
690 		/*
691 		 * Invalidate buffer if caching is disabled, forcing a
692 		 * re-read from the remote later.
693 		 */
694 		if (np->n_flag & NQNFSNONCACHE)
695 			bp->b_flags |= B_INVAL;
696 		break;
697 	    default:
698 		printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
699 	    }
700 	    brelse(bp);
701 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
702 	return (error);
703 }
704 
705 /*
706  * Vnode op for write using bio
707  *
708  * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
709  *	     struct ucred *a_cred)
710  */
711 int
712 nfs_write(struct vop_write_args *ap)
713 {
714 	int biosize;
715 	struct uio *uio = ap->a_uio;
716 	struct thread *td = uio->uio_td;
717 	struct vnode *vp = ap->a_vp;
718 	struct nfsnode *np = VTONFS(vp);
719 	int ioflag = ap->a_ioflag;
720 	struct buf *bp;
721 	struct vattr vattr;
722 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
723 	daddr_t lbn;
724 	int bcount;
725 	int n, on, error = 0, iomode, must_commit;
726 	int haverslock = 0;
727 
728 #ifdef DIAGNOSTIC
729 	if (uio->uio_rw != UIO_WRITE)
730 		panic("nfs_write mode");
731 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
732 		panic("nfs_write proc");
733 #endif
734 	if (vp->v_type != VREG)
735 		return (EIO);
736 	if (np->n_flag & NWRITEERR) {
737 		np->n_flag &= ~NWRITEERR;
738 		return (np->n_error);
739 	}
740 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
741 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
742 		(void)nfs_fsinfo(nmp, vp, td);
743 
744 	/*
745 	 * Synchronously flush pending buffers if we are in synchronous
746 	 * mode or if we are appending.
747 	 */
748 	if (ioflag & (IO_APPEND | IO_SYNC)) {
749 		if (np->n_flag & NLMODIFIED) {
750 			np->n_attrstamp = 0;
751 			error = nfs_flush(vp, MNT_WAIT, td, 0);
752 			/* error = nfs_vinvalbuf(vp, V_SAVE, td, 1); */
753 			if (error)
754 				return (error);
755 		}
756 	}
757 
758 	/*
759 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
760 	 * get the append lock.
761 	 */
762 restart:
763 	if (ioflag & IO_APPEND) {
764 		np->n_attrstamp = 0;
765 		error = VOP_GETATTR(vp, &vattr, td);
766 		if (error)
767 			return (error);
768 		uio->uio_offset = np->n_size;
769 	}
770 
771 	if (uio->uio_offset < 0)
772 		return (EINVAL);
773 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
774 		return (EFBIG);
775 	if (uio->uio_resid == 0)
776 		return (0);
777 
778 	/*
779 	 * We need to obtain the rslock if we intend to modify np->n_size
780 	 * in order to guarentee the append point with multiple contending
781 	 * writers, to guarentee that no other appenders modify n_size
782 	 * while we are trying to obtain a truncated buffer (i.e. to avoid
783 	 * accidently truncating data written by another appender due to
784 	 * the race), and to ensure that the buffer is populated prior to
785 	 * our extending of the file.  We hold rslock through the entire
786 	 * operation.
787 	 *
788 	 * Note that we do not synchronize the case where someone truncates
789 	 * the file while we are appending to it because attempting to lock
790 	 * this case may deadlock other parts of the system unexpectedly.
791 	 */
792 	if ((ioflag & IO_APPEND) ||
793 	    uio->uio_offset + uio->uio_resid > np->n_size) {
794 		switch(nfs_rslock(np, td)) {
795 		case ENOLCK:
796 			goto restart;
797 			/* not reached */
798 		case EINTR:
799 		case ERESTART:
800 			return(EINTR);
801 			/* not reached */
802 		default:
803 			break;
804 		}
805 		haverslock = 1;
806 	}
807 
808 	/*
809 	 * Maybe this should be above the vnode op call, but so long as
810 	 * file servers have no limits, i don't think it matters
811 	 */
812 	if (td->td_proc && uio->uio_offset + uio->uio_resid >
813 	      td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
814 		psignal(td->td_proc, SIGXFSZ);
815 		if (haverslock)
816 			nfs_rsunlock(np, td);
817 		return (EFBIG);
818 	}
819 
820 	biosize = vp->v_mount->mnt_stat.f_iosize;
821 
822 	do {
823 		/*
824 		 * Check for a valid write lease.
825 		 */
826 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
827 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
828 			do {
829 				error = nqnfs_getlease(vp, ND_WRITE, td);
830 			} while (error == NQNFS_EXPIRED);
831 			if (error)
832 				break;
833 			if (np->n_lrev != np->n_brev ||
834 			    (np->n_flag & NQNFSNONCACHE)) {
835 				error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
836 				if (error)
837 					break;
838 				np->n_brev = np->n_lrev;
839 			}
840 		}
841 		if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
842 		    iomode = NFSV3WRITE_FILESYNC;
843 		    error = nfs_writerpc(vp, uio, &iomode, &must_commit);
844 		    if (must_commit)
845 			    nfs_clearcommit(vp->v_mount);
846 		    break;
847 		}
848 		nfsstats.biocache_writes++;
849 		lbn = uio->uio_offset / biosize;
850 		on = uio->uio_offset & (biosize-1);
851 		n = min((unsigned)(biosize - on), uio->uio_resid);
852 again:
853 		/*
854 		 * Handle direct append and file extension cases, calculate
855 		 * unaligned buffer size.
856 		 */
857 
858 		if (uio->uio_offset == np->n_size && n) {
859 			/*
860 			 * Get the buffer (in its pre-append state to maintain
861 			 * B_CACHE if it was previously set).  Resize the
862 			 * nfsnode after we have locked the buffer to prevent
863 			 * readers from reading garbage.
864 			 */
865 			bcount = on;
866 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
867 
868 			if (bp != NULL) {
869 				long save;
870 
871 				np->n_size = uio->uio_offset + n;
872 				np->n_flag |= NLMODIFIED;
873 				vnode_pager_setsize(vp, np->n_size);
874 
875 				save = bp->b_flags & B_CACHE;
876 				bcount += n;
877 				allocbuf(bp, bcount);
878 				bp->b_flags |= save;
879 			}
880 		} else {
881 			/*
882 			 * Obtain the locked cache block first, and then
883 			 * adjust the file's size as appropriate.
884 			 */
885 			bcount = on + n;
886 			if ((off_t)lbn * biosize + bcount < np->n_size) {
887 				if ((off_t)(lbn + 1) * biosize < np->n_size)
888 					bcount = biosize;
889 				else
890 					bcount = np->n_size - (off_t)lbn * biosize;
891 			}
892 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
893 			if (uio->uio_offset + n > np->n_size) {
894 				np->n_size = uio->uio_offset + n;
895 				np->n_flag |= NLMODIFIED;
896 				vnode_pager_setsize(vp, np->n_size);
897 			}
898 		}
899 
900 		if (!bp) {
901 			error = EINTR;
902 			break;
903 		}
904 
905 		/*
906 		 * Issue a READ if B_CACHE is not set.  In special-append
907 		 * mode, B_CACHE is based on the buffer prior to the write
908 		 * op and is typically set, avoiding the read.  If a read
909 		 * is required in special append mode, the server will
910 		 * probably send us a short-read since we extended the file
911 		 * on our end, resulting in b_resid == 0 and, thusly,
912 		 * B_CACHE getting set.
913 		 *
914 		 * We can also avoid issuing the read if the write covers
915 		 * the entire buffer.  We have to make sure the buffer state
916 		 * is reasonable in this case since we will not be initiating
917 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
918 		 * more information.
919 		 *
920 		 * B_CACHE may also be set due to the buffer being cached
921 		 * normally.
922 		 */
923 
924 		if (on == 0 && n == bcount) {
925 			bp->b_flags |= B_CACHE;
926 			bp->b_flags &= ~(B_ERROR | B_INVAL);
927 		}
928 
929 		if ((bp->b_flags & B_CACHE) == 0) {
930 			bp->b_flags |= B_READ;
931 			vfs_busy_pages(bp, 0);
932 			error = nfs_doio(bp, td);
933 			if (error) {
934 				brelse(bp);
935 				break;
936 			}
937 		}
938 		if (!bp) {
939 			error = EINTR;
940 			break;
941 		}
942 		np->n_flag |= NLMODIFIED;
943 
944 		/*
945 		 * If dirtyend exceeds file size, chop it down.  This should
946 		 * not normally occur but there is an append race where it
947 		 * might occur XXX, so we log it.
948 		 *
949 		 * If the chopping creates a reverse-indexed or degenerate
950 		 * situation with dirtyoff/end, we 0 both of them.
951 		 */
952 
953 		if (bp->b_dirtyend > bcount) {
954 			printf("NFS append race @%lx:%d\n",
955 			    (long)bp->b_blkno * DEV_BSIZE,
956 			    bp->b_dirtyend - bcount);
957 			bp->b_dirtyend = bcount;
958 		}
959 
960 		if (bp->b_dirtyoff >= bp->b_dirtyend)
961 			bp->b_dirtyoff = bp->b_dirtyend = 0;
962 
963 		/*
964 		 * If the new write will leave a contiguous dirty
965 		 * area, just update the b_dirtyoff and b_dirtyend,
966 		 * otherwise force a write rpc of the old dirty area.
967 		 *
968 		 * While it is possible to merge discontiguous writes due to
969 		 * our having a B_CACHE buffer ( and thus valid read data
970 		 * for the hole), we don't because it could lead to
971 		 * significant cache coherency problems with multiple clients,
972 		 * especially if locking is implemented later on.
973 		 *
974 		 * as an optimization we could theoretically maintain
975 		 * a linked list of discontinuous areas, but we would still
976 		 * have to commit them separately so there isn't much
977 		 * advantage to it except perhaps a bit of asynchronization.
978 		 */
979 
980 		if (bp->b_dirtyend > 0 &&
981 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
982 			if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
983 				error = EINTR;
984 				break;
985 			}
986 			goto again;
987 		}
988 
989 		/*
990 		 * Check for valid write lease and get one as required.
991 		 * In case getblk() and/or bwrite() delayed us.
992 		 */
993 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
994 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
995 			do {
996 				error = nqnfs_getlease(vp, ND_WRITE, td);
997 			} while (error == NQNFS_EXPIRED);
998 			if (error) {
999 				brelse(bp);
1000 				break;
1001 			}
1002 			if (np->n_lrev != np->n_brev ||
1003 			    (np->n_flag & NQNFSNONCACHE)) {
1004 				brelse(bp);
1005 				error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1006 				if (error)
1007 					break;
1008 				np->n_brev = np->n_lrev;
1009 				goto again;
1010 			}
1011 		}
1012 
1013 		error = uiomove((char *)bp->b_data + on, n, uio);
1014 
1015 		/*
1016 		 * Since this block is being modified, it must be written
1017 		 * again and not just committed.  Since write clustering does
1018 		 * not work for the stage 1 data write, only the stage 2
1019 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1020 		 */
1021 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1022 
1023 		if (error) {
1024 			bp->b_flags |= B_ERROR;
1025 			brelse(bp);
1026 			break;
1027 		}
1028 
1029 		/*
1030 		 * Only update dirtyoff/dirtyend if not a degenerate
1031 		 * condition.
1032 		 */
1033 		if (n) {
1034 			if (bp->b_dirtyend > 0) {
1035 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1036 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1037 			} else {
1038 				bp->b_dirtyoff = on;
1039 				bp->b_dirtyend = on + n;
1040 			}
1041 			vfs_bio_set_validclean(bp, on, n);
1042 		}
1043 		/*
1044 		 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1045 		 * filesystem).  XXX also use for loopback NFS mounts.
1046 		 */
1047 		if (ioflag & IO_NOWDRAIN)
1048 			bp->b_flags |= B_NOWDRAIN;
1049 
1050 		/*
1051 		 * If the lease is non-cachable or IO_SYNC do bwrite().
1052 		 *
1053 		 * IO_INVAL appears to be unused.  The idea appears to be
1054 		 * to turn off caching in this case.  Very odd.  XXX
1055 		 */
1056 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1057 			if (ioflag & IO_INVAL)
1058 				bp->b_flags |= B_NOCACHE;
1059 			error = VOP_BWRITE(bp->b_vp, bp);
1060 			if (error)
1061 				break;
1062 			if (np->n_flag & NQNFSNONCACHE) {
1063 				error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1064 				if (error)
1065 					break;
1066 			}
1067 		} else if ((n + on) == biosize &&
1068 			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1069 			bp->b_flags |= B_ASYNC;
1070 			(void)nfs_writebp(bp, 0, 0);
1071 		} else {
1072 			bdwrite(bp);
1073 		}
1074 	} while (uio->uio_resid > 0 && n > 0);
1075 
1076 	if (haverslock)
1077 		nfs_rsunlock(np, td);
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_doio() 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, daddr_t bn, 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, bn, size, PCATCH, 0);
1109 		while (bp == (struct buf *)0) {
1110 			if (nfs_sigintr(nmp, (struct nfsreq *)0, td))
1111 				return ((struct buf *)0);
1112 			bp = getblk(vp, bn, size, 0, 2 * hz);
1113 		}
1114 	} else {
1115 		bp = getblk(vp, bn, size, 0, 0);
1116 	}
1117 
1118 	if (vp->v_type == VREG) {
1119 		int biosize;
1120 
1121 		biosize = mp->mnt_stat.f_iosize;
1122 		bp->b_blkno = bn * (biosize / DEV_BSIZE);
1123 	}
1124 	return (bp);
1125 }
1126 
1127 /*
1128  * Flush and invalidate all dirty buffers. If another process is already
1129  * doing the flush, just wait for completion.
1130  */
1131 int
1132 nfs_vinvalbuf(struct vnode *vp, int flags,
1133 	      struct thread *td, int intrflg)
1134 {
1135 	struct nfsnode *np = VTONFS(vp);
1136 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1137 	int error = 0, slpflag, slptimeo;
1138 
1139 	if (vp->v_flag & VRECLAIMED)
1140 		return (0);
1141 
1142 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1143 		intrflg = 0;
1144 	if (intrflg) {
1145 		slpflag = PCATCH;
1146 		slptimeo = 2 * hz;
1147 	} else {
1148 		slpflag = 0;
1149 		slptimeo = 0;
1150 	}
1151 	/*
1152 	 * First wait for any other process doing a flush to complete.
1153 	 */
1154 	while (np->n_flag & NFLUSHINPROG) {
1155 		np->n_flag |= NFLUSHWANT;
1156 		error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1157 		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td))
1158 			return (EINTR);
1159 	}
1160 
1161 	/*
1162 	 * Now, flush as required.
1163 	 */
1164 	np->n_flag |= NFLUSHINPROG;
1165 	error = vinvalbuf(vp, flags, td, slpflag, 0);
1166 	while (error) {
1167 		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) {
1168 			np->n_flag &= ~NFLUSHINPROG;
1169 			if (np->n_flag & NFLUSHWANT) {
1170 				np->n_flag &= ~NFLUSHWANT;
1171 				wakeup((caddr_t)&np->n_flag);
1172 			}
1173 			return (EINTR);
1174 		}
1175 		error = vinvalbuf(vp, flags, td, 0, slptimeo);
1176 	}
1177 	np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
1178 	if (np->n_flag & NFLUSHWANT) {
1179 		np->n_flag &= ~NFLUSHWANT;
1180 		wakeup((caddr_t)&np->n_flag);
1181 	}
1182 	return (0);
1183 }
1184 
1185 /*
1186  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1187  * This is mainly to avoid queueing async I/O requests when the nfsiods
1188  * are all hung on a dead server.
1189  *
1190  * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1191  * is eventually dequeued by the async daemon, nfs_doio() *will*.
1192  */
1193 int
1194 nfs_asyncio(struct buf *bp, struct thread *td)
1195 {
1196 	struct nfsmount *nmp;
1197 	int i;
1198 	int gotiod;
1199 	int slpflag = 0;
1200 	int slptimeo = 0;
1201 	int error;
1202 
1203 	/*
1204 	 * If no async daemons then return EIO to force caller to run the rpc
1205 	 * synchronously.
1206 	 */
1207 	if (nfs_numasync == 0)
1208 		return (EIO);
1209 
1210 	nmp = VFSTONFS(bp->b_vp->v_mount);
1211 
1212 	/*
1213 	 * Commits are usually short and sweet so lets save some cpu and
1214 	 * leave the async daemons for more important rpc's (such as reads
1215 	 * and writes).
1216 	 */
1217 	if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1218 	    (nmp->nm_bufqiods > nfs_numasync / 2)) {
1219 		return(EIO);
1220 	}
1221 
1222 again:
1223 	if (nmp->nm_flag & NFSMNT_INT)
1224 		slpflag = PCATCH;
1225 	gotiod = FALSE;
1226 
1227 	/*
1228 	 * Find a free iod to process this request.
1229 	 */
1230 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1231 		if (nfs_iodwant[i]) {
1232 			/*
1233 			 * Found one, so wake it up and tell it which
1234 			 * mount to process.
1235 			 */
1236 			NFS_DPF(ASYNCIO,
1237 				("nfs_asyncio: waking iod %d for mount %p\n",
1238 				 i, nmp));
1239 			nfs_iodwant[i] = NULL;
1240 			nfs_iodmount[i] = nmp;
1241 			nmp->nm_bufqiods++;
1242 			wakeup((caddr_t)&nfs_iodwant[i]);
1243 			gotiod = TRUE;
1244 			break;
1245 		}
1246 
1247 	/*
1248 	 * If none are free, we may already have an iod working on this mount
1249 	 * point.  If so, it will process our request.
1250 	 */
1251 	if (!gotiod) {
1252 		if (nmp->nm_bufqiods > 0) {
1253 			NFS_DPF(ASYNCIO,
1254 				("nfs_asyncio: %d iods are already processing mount %p\n",
1255 				 nmp->nm_bufqiods, nmp));
1256 			gotiod = TRUE;
1257 		}
1258 	}
1259 
1260 	/*
1261 	 * If we have an iod which can process the request, then queue
1262 	 * the buffer.
1263 	 */
1264 	if (gotiod) {
1265 		/*
1266 		 * Ensure that the queue never grows too large.  We still want
1267 		 * to asynchronize so we block rather then return EIO.
1268 		 */
1269 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1270 			NFS_DPF(ASYNCIO,
1271 				("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1272 			nmp->nm_bufqwant = TRUE;
1273 			error = tsleep(&nmp->nm_bufq, slpflag,
1274 				       "nfsaio", slptimeo);
1275 			if (error) {
1276 				if (nfs_sigintr(nmp, NULL, td))
1277 					return (EINTR);
1278 				if (slpflag == PCATCH) {
1279 					slpflag = 0;
1280 					slptimeo = 2 * hz;
1281 				}
1282 			}
1283 			/*
1284 			 * We might have lost our iod while sleeping,
1285 			 * so check and loop if nescessary.
1286 			 */
1287 			if (nmp->nm_bufqiods == 0) {
1288 				NFS_DPF(ASYNCIO,
1289 					("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1290 				goto again;
1291 			}
1292 		}
1293 		BUF_KERNPROC(bp);
1294 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1295 		nmp->nm_bufqlen++;
1296 		return (0);
1297 	}
1298 
1299 	/*
1300 	 * All the iods are busy on other mounts, so return EIO to
1301 	 * force the caller to process the i/o synchronously.
1302 	 */
1303 	NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1304 	return (EIO);
1305 }
1306 
1307 /*
1308  * Do an I/O operation to/from a cache block. This may be called
1309  * synchronously or from an nfsiod.
1310  *
1311  * NOTE! TD MIGHT BE NULL
1312  */
1313 int
1314 nfs_doio(struct buf *bp, struct thread *td)
1315 {
1316 	struct uio *uiop;
1317 	struct vnode *vp;
1318 	struct nfsnode *np;
1319 	struct nfsmount *nmp;
1320 	int error = 0, iomode, must_commit = 0;
1321 	struct uio uio;
1322 	struct iovec io;
1323 
1324 	vp = bp->b_vp;
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_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1341 
1342 	/*
1343 	 * Historically, paging was done with physio, but no more.
1344 	 */
1345 	if (bp->b_flags & B_PHYS) {
1346 	    /*
1347 	     * ...though reading /dev/drum still gets us here.
1348 	     */
1349 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1350 	    /* mapping was done by vmapbuf() */
1351 	    io.iov_base = bp->b_data;
1352 	    uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1353 	    if (bp->b_flags & B_READ) {
1354 		uiop->uio_rw = UIO_READ;
1355 		nfsstats.read_physios++;
1356 		error = nfs_readrpc(vp, uiop);
1357 	    } else {
1358 		int com;
1359 
1360 		iomode = NFSV3WRITE_DATASYNC;
1361 		uiop->uio_rw = UIO_WRITE;
1362 		nfsstats.write_physios++;
1363 		error = nfs_writerpc(vp, uiop, &iomode, &com);
1364 	    }
1365 	    if (error) {
1366 		bp->b_flags |= B_ERROR;
1367 		bp->b_error = error;
1368 	    }
1369 	} else if (bp->b_flags & B_READ) {
1370 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1371 	    io.iov_base = bp->b_data;
1372 	    uiop->uio_rw = UIO_READ;
1373 
1374 	    switch (vp->v_type) {
1375 	    case VREG:
1376 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1377 		nfsstats.read_bios++;
1378 		error = nfs_readrpc(vp, uiop);
1379 
1380 		if (!error) {
1381 		    if (uiop->uio_resid) {
1382 			/*
1383 			 * If we had a short read with no error, we must have
1384 			 * hit a file hole.  We should zero-fill the remainder.
1385 			 * This can also occur if the server hits the file EOF.
1386 			 *
1387 			 * Holes used to be able to occur due to pending
1388 			 * writes, but that is not possible any longer.
1389 			 */
1390 			int nread = bp->b_bcount - uiop->uio_resid;
1391 			int left  = uiop->uio_resid;
1392 
1393 			if (left > 0)
1394 				bzero((char *)bp->b_data + nread, left);
1395 			uiop->uio_resid = 0;
1396 		    }
1397 		}
1398 		if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1399 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
1400 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
1401 			  np->n_lrev != np->n_brev) ||
1402 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1403 			  np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1404 			uprintf("Process killed due to text file modification\n");
1405 			psignal(td->td_proc, SIGKILL);
1406 			PHOLD(td->td_proc);
1407 		}
1408 		break;
1409 	    case VLNK:
1410 		uiop->uio_offset = (off_t)0;
1411 		nfsstats.readlink_bios++;
1412 		error = nfs_readlinkrpc(vp, uiop);
1413 		break;
1414 	    case VDIR:
1415 		nfsstats.readdir_bios++;
1416 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1417 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1418 			error = nfs_readdirplusrpc(vp, uiop);
1419 			if (error == NFSERR_NOTSUPP)
1420 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1421 		}
1422 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1423 			error = nfs_readdirrpc(vp, uiop);
1424 		/*
1425 		 * end-of-directory sets B_INVAL but does not generate an
1426 		 * error.
1427 		 */
1428 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1429 			bp->b_flags |= B_INVAL;
1430 		break;
1431 	    default:
1432 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
1433 		break;
1434 	    };
1435 	    if (error) {
1436 		bp->b_flags |= B_ERROR;
1437 		bp->b_error = error;
1438 	    }
1439 	} else {
1440 	    /*
1441 	     * If we only need to commit, try to commit
1442 	     */
1443 	    if (bp->b_flags & B_NEEDCOMMIT) {
1444 		    int retv;
1445 		    off_t off;
1446 
1447 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1448 		    retv = nfs_commit(bp->b_vp, off,
1449 				bp->b_dirtyend - bp->b_dirtyoff, td);
1450 		    if (retv == 0) {
1451 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1452 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1453 			    bp->b_resid = 0;
1454 			    biodone(bp);
1455 			    return (0);
1456 		    }
1457 		    if (retv == NFSERR_STALEWRITEVERF) {
1458 			    nfs_clearcommit(bp->b_vp->v_mount);
1459 		    }
1460 	    }
1461 
1462 	    /*
1463 	     * Setup for actual write
1464 	     */
1465 
1466 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1467 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1468 
1469 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1470 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1471 		    - bp->b_dirtyoff;
1472 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1473 		    + bp->b_dirtyoff;
1474 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1475 		uiop->uio_rw = UIO_WRITE;
1476 		nfsstats.write_bios++;
1477 
1478 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1479 		    iomode = NFSV3WRITE_UNSTABLE;
1480 		else
1481 		    iomode = NFSV3WRITE_FILESYNC;
1482 
1483 		error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1484 
1485 		/*
1486 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1487 		 * to cluster the buffers needing commit.  This will allow
1488 		 * the system to submit a single commit rpc for the whole
1489 		 * cluster.  We can do this even if the buffer is not 100%
1490 		 * dirty (relative to the NFS blocksize), so we optimize the
1491 		 * append-to-file-case.
1492 		 *
1493 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1494 		 * cleared because write clustering only works for commit
1495 		 * rpc's, not for the data portion of the write).
1496 		 */
1497 
1498 		if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1499 		    bp->b_flags |= B_NEEDCOMMIT;
1500 		    if (bp->b_dirtyoff == 0
1501 			&& bp->b_dirtyend == bp->b_bcount)
1502 			bp->b_flags |= B_CLUSTEROK;
1503 		} else {
1504 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1505 		}
1506 
1507 		/*
1508 		 * For an interrupted write, the buffer is still valid
1509 		 * and the write hasn't been pushed to the server yet,
1510 		 * so we can't set B_ERROR and report the interruption
1511 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1512 		 * is not relevant, so the rpc attempt is essentially
1513 		 * a noop.  For the case of a V3 write rpc not being
1514 		 * committed to stable storage, the block is still
1515 		 * dirty and requires either a commit rpc or another
1516 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1517 		 * the block is reused. This is indicated by setting
1518 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1519 		 *
1520 		 * If the buffer is marked B_PAGING, it does not reside on
1521 		 * the vp's paging queues so we cannot call bdirty().  The
1522 		 * bp in this case is not an NFS cache block so we should
1523 		 * be safe. XXX
1524 		 */
1525     		if (error == EINTR
1526 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1527 			int s;
1528 
1529 			s = splbio();
1530 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1531 			if ((bp->b_flags & B_PAGING) == 0) {
1532 			    bdirty(bp);
1533 			    bp->b_flags &= ~B_DONE;
1534 			}
1535 			if (error && (bp->b_flags & B_ASYNC) == 0)
1536 			    bp->b_flags |= B_EINTR;
1537 			splx(s);
1538 	    	} else {
1539 		    if (error) {
1540 			bp->b_flags |= B_ERROR;
1541 			bp->b_error = np->n_error = error;
1542 			np->n_flag |= NWRITEERR;
1543 		    }
1544 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1545 		}
1546 	    } else {
1547 		bp->b_resid = 0;
1548 		biodone(bp);
1549 		return (0);
1550 	    }
1551 	}
1552 	bp->b_resid = uiop->uio_resid;
1553 	if (must_commit)
1554 	    nfs_clearcommit(vp->v_mount);
1555 	biodone(bp);
1556 	return (error);
1557 }
1558 
1559 /*
1560  * Used to aid in handling ftruncate() operations on the NFS client side.
1561  * Truncation creates a number of special problems for NFS.  We have to
1562  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1563  * we have to properly handle VM pages or (potentially dirty) buffers
1564  * that straddle the truncation point.
1565  */
1566 
1567 int
1568 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1569 {
1570 	struct nfsnode *np = VTONFS(vp);
1571 	u_quad_t tsize = np->n_size;
1572 	int biosize = vp->v_mount->mnt_stat.f_iosize;
1573 	int error = 0;
1574 
1575 	np->n_size = nsize;
1576 
1577 	if (np->n_size < tsize) {
1578 		struct buf *bp;
1579 		daddr_t lbn;
1580 		int bufsize;
1581 
1582 		/*
1583 		 * vtruncbuf() doesn't get the buffer overlapping the
1584 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1585 		 * buffer that now needs to be truncated.
1586 		 */
1587 		error = vtruncbuf(vp, td, nsize, biosize);
1588 		lbn = nsize / biosize;
1589 		bufsize = nsize & (biosize - 1);
1590 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1591 		if (bp->b_dirtyoff > bp->b_bcount)
1592 			bp->b_dirtyoff = bp->b_bcount;
1593 		if (bp->b_dirtyend > bp->b_bcount)
1594 			bp->b_dirtyend = bp->b_bcount;
1595 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1596 		brelse(bp);
1597 	} else {
1598 		vnode_pager_setsize(vp, nsize);
1599 	}
1600 	return(error);
1601 }
1602 
1603