xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 5944f899a2519c6321bac3c17cc076418643a088)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Rick Macklem at The University of Guelph.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/bio.h>
41 #include <sys/buf.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/rwlock.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_pager.h>
54 #include <vm/vnode_pager.h>
55 
56 #include <fs/nfs/nfsport.h>
57 #include <fs/nfsclient/nfsmount.h>
58 #include <fs/nfsclient/nfs.h>
59 #include <fs/nfsclient/nfsnode.h>
60 #include <fs/nfsclient/nfs_kdtrace.h>
61 
62 extern int newnfs_directio_allow_mmap;
63 extern struct nfsstatsv1 nfsstatsv1;
64 extern struct mtx ncl_iod_mutex;
65 extern int ncl_numasync;
66 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
67 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
68 extern int newnfs_directio_enable;
69 extern int nfs_keep_dirty_on_error;
70 
71 int ncl_pbuf_freecnt = -1;	/* start out unlimited */
72 
73 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
74     struct thread *td);
75 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
76     struct ucred *cred, int ioflag);
77 
78 /*
79  * Vnode op for VM getpages.
80  */
81 SYSCTL_DECL(_vfs_nfs);
82 static int use_buf_pager = 1;
83 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
84     &use_buf_pager, 0,
85     "Use buffer pager instead of direct readrpc call");
86 
87 static daddr_t
88 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
89 {
90 
91 	return (off / vp->v_bufobj.bo_bsize);
92 }
93 
94 static int
95 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn)
96 {
97 	struct nfsnode *np;
98 	u_quad_t nsize;
99 	int biosize, bcount;
100 
101 	np = VTONFS(vp);
102 	mtx_lock(&np->n_mtx);
103 	nsize = np->n_size;
104 	mtx_unlock(&np->n_mtx);
105 
106 	biosize = vp->v_bufobj.bo_bsize;
107 	bcount = biosize;
108 	if ((off_t)lbn * biosize >= nsize)
109 		bcount = 0;
110 	else if ((off_t)(lbn + 1) * biosize > nsize)
111 		bcount = nsize - (off_t)lbn * biosize;
112 	return (bcount);
113 }
114 
115 int
116 ncl_getpages(struct vop_getpages_args *ap)
117 {
118 	int i, error, nextoff, size, toff, count, npages;
119 	struct uio uio;
120 	struct iovec iov;
121 	vm_offset_t kva;
122 	struct buf *bp;
123 	struct vnode *vp;
124 	struct thread *td;
125 	struct ucred *cred;
126 	struct nfsmount *nmp;
127 	vm_object_t object;
128 	vm_page_t *pages;
129 	struct nfsnode *np;
130 
131 	vp = ap->a_vp;
132 	np = VTONFS(vp);
133 	td = curthread;
134 	cred = curthread->td_ucred;
135 	nmp = VFSTONFS(vp->v_mount);
136 	pages = ap->a_m;
137 	npages = ap->a_count;
138 
139 	if ((object = vp->v_object) == NULL) {
140 		printf("ncl_getpages: called with non-merged cache vnode\n");
141 		return (VM_PAGER_ERROR);
142 	}
143 
144 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
145 		mtx_lock(&np->n_mtx);
146 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
147 			mtx_unlock(&np->n_mtx);
148 			printf("ncl_getpages: called on non-cacheable vnode\n");
149 			return (VM_PAGER_ERROR);
150 		} else
151 			mtx_unlock(&np->n_mtx);
152 	}
153 
154 	mtx_lock(&nmp->nm_mtx);
155 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
156 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
157 		mtx_unlock(&nmp->nm_mtx);
158 		/* We'll never get here for v4, because we always have fsinfo */
159 		(void)ncl_fsinfo(nmp, vp, cred, td);
160 	} else
161 		mtx_unlock(&nmp->nm_mtx);
162 
163 	if (use_buf_pager)
164 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
165 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
166 
167 	/*
168 	 * If the requested page is partially valid, just return it and
169 	 * allow the pager to zero-out the blanks.  Partially valid pages
170 	 * can only occur at the file EOF.
171 	 *
172 	 * XXXGL: is that true for NFS, where short read can occur???
173 	 */
174 	VM_OBJECT_WLOCK(object);
175 	if (pages[npages - 1]->valid != 0 && --npages == 0)
176 		goto out;
177 	VM_OBJECT_WUNLOCK(object);
178 
179 	/*
180 	 * We use only the kva address for the buffer, but this is extremely
181 	 * convenient and fast.
182 	 */
183 	bp = getpbuf(&ncl_pbuf_freecnt);
184 
185 	kva = (vm_offset_t) bp->b_data;
186 	pmap_qenter(kva, pages, npages);
187 	PCPU_INC(cnt.v_vnodein);
188 	PCPU_ADD(cnt.v_vnodepgsin, npages);
189 
190 	count = npages << PAGE_SHIFT;
191 	iov.iov_base = (caddr_t) kva;
192 	iov.iov_len = count;
193 	uio.uio_iov = &iov;
194 	uio.uio_iovcnt = 1;
195 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
196 	uio.uio_resid = count;
197 	uio.uio_segflg = UIO_SYSSPACE;
198 	uio.uio_rw = UIO_READ;
199 	uio.uio_td = td;
200 
201 	error = ncl_readrpc(vp, &uio, cred);
202 	pmap_qremove(kva, npages);
203 
204 	relpbuf(bp, &ncl_pbuf_freecnt);
205 
206 	if (error && (uio.uio_resid == count)) {
207 		printf("ncl_getpages: error %d\n", error);
208 		return (VM_PAGER_ERROR);
209 	}
210 
211 	/*
212 	 * Calculate the number of bytes read and validate only that number
213 	 * of bytes.  Note that due to pending writes, size may be 0.  This
214 	 * does not mean that the remaining data is invalid!
215 	 */
216 
217 	size = count - uio.uio_resid;
218 	VM_OBJECT_WLOCK(object);
219 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
220 		vm_page_t m;
221 		nextoff = toff + PAGE_SIZE;
222 		m = pages[i];
223 
224 		if (nextoff <= size) {
225 			/*
226 			 * Read operation filled an entire page
227 			 */
228 			m->valid = VM_PAGE_BITS_ALL;
229 			KASSERT(m->dirty == 0,
230 			    ("nfs_getpages: page %p is dirty", m));
231 		} else if (size > toff) {
232 			/*
233 			 * Read operation filled a partial page.
234 			 */
235 			m->valid = 0;
236 			vm_page_set_valid_range(m, 0, size - toff);
237 			KASSERT(m->dirty == 0,
238 			    ("nfs_getpages: page %p is dirty", m));
239 		} else {
240 			/*
241 			 * Read operation was short.  If no error
242 			 * occurred we may have hit a zero-fill
243 			 * section.  We leave valid set to 0, and page
244 			 * is freed by vm_page_readahead_finish() if
245 			 * its index is not equal to requested, or
246 			 * page is zeroed and set valid by
247 			 * vm_pager_get_pages() for requested page.
248 			 */
249 			;
250 		}
251 	}
252 out:
253 	VM_OBJECT_WUNLOCK(object);
254 	if (ap->a_rbehind)
255 		*ap->a_rbehind = 0;
256 	if (ap->a_rahead)
257 		*ap->a_rahead = 0;
258 	return (VM_PAGER_OK);
259 }
260 
261 /*
262  * Vnode op for VM putpages.
263  */
264 int
265 ncl_putpages(struct vop_putpages_args *ap)
266 {
267 	struct uio uio;
268 	struct iovec iov;
269 	int i, error, npages, count;
270 	off_t offset;
271 	int *rtvals;
272 	struct vnode *vp;
273 	struct thread *td;
274 	struct ucred *cred;
275 	struct nfsmount *nmp;
276 	struct nfsnode *np;
277 	vm_page_t *pages;
278 
279 	vp = ap->a_vp;
280 	np = VTONFS(vp);
281 	td = curthread;				/* XXX */
282 	/* Set the cred to n_writecred for the write rpcs. */
283 	if (np->n_writecred != NULL)
284 		cred = crhold(np->n_writecred);
285 	else
286 		cred = crhold(curthread->td_ucred);	/* XXX */
287 	nmp = VFSTONFS(vp->v_mount);
288 	pages = ap->a_m;
289 	count = ap->a_count;
290 	rtvals = ap->a_rtvals;
291 	npages = btoc(count);
292 	offset = IDX_TO_OFF(pages[0]->pindex);
293 
294 	mtx_lock(&nmp->nm_mtx);
295 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
296 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
297 		mtx_unlock(&nmp->nm_mtx);
298 		(void)ncl_fsinfo(nmp, vp, cred, td);
299 	} else
300 		mtx_unlock(&nmp->nm_mtx);
301 
302 	mtx_lock(&np->n_mtx);
303 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
304 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
305 		mtx_unlock(&np->n_mtx);
306 		printf("ncl_putpages: called on noncache-able vnode\n");
307 		mtx_lock(&np->n_mtx);
308 	}
309 
310 	for (i = 0; i < npages; i++)
311 		rtvals[i] = VM_PAGER_ERROR;
312 
313 	/*
314 	 * When putting pages, do not extend file past EOF.
315 	 */
316 	if (offset + count > np->n_size) {
317 		count = np->n_size - offset;
318 		if (count < 0)
319 			count = 0;
320 	}
321 	mtx_unlock(&np->n_mtx);
322 
323 	PCPU_INC(cnt.v_vnodeout);
324 	PCPU_ADD(cnt.v_vnodepgsout, count);
325 
326 	iov.iov_base = unmapped_buf;
327 	iov.iov_len = count;
328 	uio.uio_iov = &iov;
329 	uio.uio_iovcnt = 1;
330 	uio.uio_offset = offset;
331 	uio.uio_resid = count;
332 	uio.uio_segflg = UIO_NOCOPY;
333 	uio.uio_rw = UIO_WRITE;
334 	uio.uio_td = td;
335 
336 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
337 	    cred);
338 	crfree(cred);
339 
340 	if (error == 0 || !nfs_keep_dirty_on_error)
341 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
342 	return (rtvals[0]);
343 }
344 
345 /*
346  * For nfs, cache consistency can only be maintained approximately.
347  * Although RFC1094 does not specify the criteria, the following is
348  * believed to be compatible with the reference port.
349  * For nfs:
350  * If the file's modify time on the server has changed since the
351  * last read rpc or you have written to the file,
352  * you may have lost data cache consistency with the
353  * server, so flush all of the file's data out of the cache.
354  * Then force a getattr rpc to ensure that you have up to date
355  * attributes.
356  * NB: This implies that cache data can be read when up to
357  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
358  * attributes this could be forced by setting n_attrstamp to 0 before
359  * the VOP_GETATTR() call.
360  */
361 static inline int
362 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
363 {
364 	int error = 0;
365 	struct vattr vattr;
366 	struct nfsnode *np = VTONFS(vp);
367 	int old_lock;
368 
369 	/*
370 	 * Grab the exclusive lock before checking whether the cache is
371 	 * consistent.
372 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
373 	 * But for now, this suffices.
374 	 */
375 	old_lock = ncl_upgrade_vnlock(vp);
376 	if (vp->v_iflag & VI_DOOMED) {
377 		error = EBADF;
378 		goto out;
379 	}
380 
381 	mtx_lock(&np->n_mtx);
382 	if (np->n_flag & NMODIFIED) {
383 		mtx_unlock(&np->n_mtx);
384 		if (vp->v_type != VREG) {
385 			if (vp->v_type != VDIR)
386 				panic("nfs: bioread, not dir");
387 			ncl_invaldir(vp);
388 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
389 			if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
390 				error = EBADF;
391 			if (error != 0)
392 				goto out;
393 		}
394 		np->n_attrstamp = 0;
395 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
396 		error = VOP_GETATTR(vp, &vattr, cred);
397 		if (error)
398 			goto out;
399 		mtx_lock(&np->n_mtx);
400 		np->n_mtime = vattr.va_mtime;
401 		mtx_unlock(&np->n_mtx);
402 	} else {
403 		mtx_unlock(&np->n_mtx);
404 		error = VOP_GETATTR(vp, &vattr, cred);
405 		if (error)
406 			return (error);
407 		mtx_lock(&np->n_mtx);
408 		if ((np->n_flag & NSIZECHANGED)
409 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
410 			mtx_unlock(&np->n_mtx);
411 			if (vp->v_type == VDIR)
412 				ncl_invaldir(vp);
413 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
414 			if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
415 				error = EBADF;
416 			if (error != 0)
417 				goto out;
418 			mtx_lock(&np->n_mtx);
419 			np->n_mtime = vattr.va_mtime;
420 			np->n_flag &= ~NSIZECHANGED;
421 		}
422 		mtx_unlock(&np->n_mtx);
423 	}
424 out:
425 	ncl_downgrade_vnlock(vp, old_lock);
426 	return (error);
427 }
428 
429 /*
430  * Vnode op for read using bio
431  */
432 int
433 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
434 {
435 	struct nfsnode *np = VTONFS(vp);
436 	int biosize, i;
437 	struct buf *bp, *rabp;
438 	struct thread *td;
439 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
440 	daddr_t lbn, rabn;
441 	int bcount;
442 	int seqcount;
443 	int nra, error = 0, n = 0, on = 0;
444 	off_t tmp_off;
445 
446 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
447 	if (uio->uio_resid == 0)
448 		return (0);
449 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
450 		return (EINVAL);
451 	td = uio->uio_td;
452 
453 	mtx_lock(&nmp->nm_mtx);
454 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
455 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
456 		mtx_unlock(&nmp->nm_mtx);
457 		(void)ncl_fsinfo(nmp, vp, cred, td);
458 		mtx_lock(&nmp->nm_mtx);
459 	}
460 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
461 		(void) newnfs_iosize(nmp);
462 
463 	tmp_off = uio->uio_offset + uio->uio_resid;
464 	if (vp->v_type != VDIR &&
465 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
466 		mtx_unlock(&nmp->nm_mtx);
467 		return (EFBIG);
468 	}
469 	mtx_unlock(&nmp->nm_mtx);
470 
471 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
472 		/* No caching/ no readaheads. Just read data into the user buffer */
473 		return ncl_readrpc(vp, uio, cred);
474 
475 	biosize = vp->v_bufobj.bo_bsize;
476 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
477 
478 	error = nfs_bioread_check_cons(vp, td, cred);
479 	if (error)
480 		return error;
481 
482 	do {
483 	    u_quad_t nsize;
484 
485 	    mtx_lock(&np->n_mtx);
486 	    nsize = np->n_size;
487 	    mtx_unlock(&np->n_mtx);
488 
489 	    switch (vp->v_type) {
490 	    case VREG:
491 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
492 		lbn = uio->uio_offset / biosize;
493 		on = uio->uio_offset - (lbn * biosize);
494 
495 		/*
496 		 * Start the read ahead(s), as required.
497 		 */
498 		if (nmp->nm_readahead > 0) {
499 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
500 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
501 			rabn = lbn + 1 + nra;
502 			if (incore(&vp->v_bufobj, rabn) == NULL) {
503 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
504 			    if (!rabp) {
505 				error = newnfs_sigintr(nmp, td);
506 				return (error ? error : EINTR);
507 			    }
508 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
509 				rabp->b_flags |= B_ASYNC;
510 				rabp->b_iocmd = BIO_READ;
511 				vfs_busy_pages(rabp, 0);
512 				if (ncl_asyncio(nmp, rabp, cred, td)) {
513 				    rabp->b_flags |= B_INVAL;
514 				    rabp->b_ioflags |= BIO_ERROR;
515 				    vfs_unbusy_pages(rabp);
516 				    brelse(rabp);
517 				    break;
518 				}
519 			    } else {
520 				brelse(rabp);
521 			    }
522 			}
523 		    }
524 		}
525 
526 		/* Note that bcount is *not* DEV_BSIZE aligned. */
527 		bcount = biosize;
528 		if ((off_t)lbn * biosize >= nsize) {
529 			bcount = 0;
530 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
531 			bcount = nsize - (off_t)lbn * biosize;
532 		}
533 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
534 
535 		if (!bp) {
536 			error = newnfs_sigintr(nmp, td);
537 			return (error ? error : EINTR);
538 		}
539 
540 		/*
541 		 * If B_CACHE is not set, we must issue the read.  If this
542 		 * fails, we return an error.
543 		 */
544 
545 		if ((bp->b_flags & B_CACHE) == 0) {
546 		    bp->b_iocmd = BIO_READ;
547 		    vfs_busy_pages(bp, 0);
548 		    error = ncl_doio(vp, bp, cred, td, 0);
549 		    if (error) {
550 			brelse(bp);
551 			return (error);
552 		    }
553 		}
554 
555 		/*
556 		 * on is the offset into the current bp.  Figure out how many
557 		 * bytes we can copy out of the bp.  Note that bcount is
558 		 * NOT DEV_BSIZE aligned.
559 		 *
560 		 * Then figure out how many bytes we can copy into the uio.
561 		 */
562 
563 		n = 0;
564 		if (on < bcount)
565 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
566 		break;
567 	    case VLNK:
568 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
569 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
570 		if (!bp) {
571 			error = newnfs_sigintr(nmp, td);
572 			return (error ? error : EINTR);
573 		}
574 		if ((bp->b_flags & B_CACHE) == 0) {
575 		    bp->b_iocmd = BIO_READ;
576 		    vfs_busy_pages(bp, 0);
577 		    error = ncl_doio(vp, bp, cred, td, 0);
578 		    if (error) {
579 			bp->b_ioflags |= BIO_ERROR;
580 			brelse(bp);
581 			return (error);
582 		    }
583 		}
584 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
585 		on = 0;
586 		break;
587 	    case VDIR:
588 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
589 		if (np->n_direofoffset
590 		    && uio->uio_offset >= np->n_direofoffset) {
591 		    return (0);
592 		}
593 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
594 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
595 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
596 		if (!bp) {
597 		    error = newnfs_sigintr(nmp, td);
598 		    return (error ? error : EINTR);
599 		}
600 		if ((bp->b_flags & B_CACHE) == 0) {
601 		    bp->b_iocmd = BIO_READ;
602 		    vfs_busy_pages(bp, 0);
603 		    error = ncl_doio(vp, bp, cred, td, 0);
604 		    if (error) {
605 			    brelse(bp);
606 		    }
607 		    while (error == NFSERR_BAD_COOKIE) {
608 			ncl_invaldir(vp);
609 			error = ncl_vinvalbuf(vp, 0, td, 1);
610 			if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
611 				return (EBADF);
612 
613 			/*
614 			 * Yuck! The directory has been modified on the
615 			 * server. The only way to get the block is by
616 			 * reading from the beginning to get all the
617 			 * offset cookies.
618 			 *
619 			 * Leave the last bp intact unless there is an error.
620 			 * Loop back up to the while if the error is another
621 			 * NFSERR_BAD_COOKIE (double yuch!).
622 			 */
623 			for (i = 0; i <= lbn && !error; i++) {
624 			    if (np->n_direofoffset
625 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
626 				    return (0);
627 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
628 			    if (!bp) {
629 				error = newnfs_sigintr(nmp, td);
630 				return (error ? error : EINTR);
631 			    }
632 			    if ((bp->b_flags & B_CACHE) == 0) {
633 				    bp->b_iocmd = BIO_READ;
634 				    vfs_busy_pages(bp, 0);
635 				    error = ncl_doio(vp, bp, cred, td, 0);
636 				    /*
637 				     * no error + B_INVAL == directory EOF,
638 				     * use the block.
639 				     */
640 				    if (error == 0 && (bp->b_flags & B_INVAL))
641 					    break;
642 			    }
643 			    /*
644 			     * An error will throw away the block and the
645 			     * for loop will break out.  If no error and this
646 			     * is not the block we want, we throw away the
647 			     * block and go for the next one via the for loop.
648 			     */
649 			    if (error || i < lbn)
650 				    brelse(bp);
651 			}
652 		    }
653 		    /*
654 		     * The above while is repeated if we hit another cookie
655 		     * error.  If we hit an error and it wasn't a cookie error,
656 		     * we give up.
657 		     */
658 		    if (error)
659 			    return (error);
660 		}
661 
662 		/*
663 		 * If not eof and read aheads are enabled, start one.
664 		 * (You need the current block first, so that you have the
665 		 *  directory offset cookie of the next block.)
666 		 */
667 		if (nmp->nm_readahead > 0 &&
668 		    (bp->b_flags & B_INVAL) == 0 &&
669 		    (np->n_direofoffset == 0 ||
670 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
671 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
672 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
673 			if (rabp) {
674 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
675 				rabp->b_flags |= B_ASYNC;
676 				rabp->b_iocmd = BIO_READ;
677 				vfs_busy_pages(rabp, 0);
678 				if (ncl_asyncio(nmp, rabp, cred, td)) {
679 				    rabp->b_flags |= B_INVAL;
680 				    rabp->b_ioflags |= BIO_ERROR;
681 				    vfs_unbusy_pages(rabp);
682 				    brelse(rabp);
683 				}
684 			    } else {
685 				brelse(rabp);
686 			    }
687 			}
688 		}
689 		/*
690 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
691 		 * chopped for the EOF condition, we cannot tell how large
692 		 * NFS directories are going to be until we hit EOF.  So
693 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
694 		 * it just so happens that b_resid will effectively chop it
695 		 * to EOF.  *BUT* this information is lost if the buffer goes
696 		 * away and is reconstituted into a B_CACHE state ( due to
697 		 * being VMIO ) later.  So we keep track of the directory eof
698 		 * in np->n_direofoffset and chop it off as an extra step
699 		 * right here.
700 		 */
701 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
702 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
703 			n = np->n_direofoffset - uio->uio_offset;
704 		break;
705 	    default:
706 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
707 		bp = NULL;
708 		break;
709 	    }
710 
711 	    if (n > 0) {
712 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
713 	    }
714 	    if (vp->v_type == VLNK)
715 		n = 0;
716 	    if (bp != NULL)
717 		brelse(bp);
718 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
719 	return (error);
720 }
721 
722 /*
723  * The NFS write path cannot handle iovecs with len > 1. So we need to
724  * break up iovecs accordingly (restricting them to wsize).
725  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
726  * For the ASYNC case, 2 copies are needed. The first a copy from the
727  * user buffer to a staging buffer and then a second copy from the staging
728  * buffer to mbufs. This can be optimized by copying from the user buffer
729  * directly into mbufs and passing the chain down, but that requires a
730  * fair amount of re-working of the relevant codepaths (and can be done
731  * later).
732  */
733 static int
734 nfs_directio_write(vp, uiop, cred, ioflag)
735 	struct vnode *vp;
736 	struct uio *uiop;
737 	struct ucred *cred;
738 	int ioflag;
739 {
740 	int error;
741 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
742 	struct thread *td = uiop->uio_td;
743 	int size;
744 	int wsize;
745 
746 	mtx_lock(&nmp->nm_mtx);
747 	wsize = nmp->nm_wsize;
748 	mtx_unlock(&nmp->nm_mtx);
749 	if (ioflag & IO_SYNC) {
750 		int iomode, must_commit;
751 		struct uio uio;
752 		struct iovec iov;
753 do_sync:
754 		while (uiop->uio_resid > 0) {
755 			size = MIN(uiop->uio_resid, wsize);
756 			size = MIN(uiop->uio_iov->iov_len, size);
757 			iov.iov_base = uiop->uio_iov->iov_base;
758 			iov.iov_len = size;
759 			uio.uio_iov = &iov;
760 			uio.uio_iovcnt = 1;
761 			uio.uio_offset = uiop->uio_offset;
762 			uio.uio_resid = size;
763 			uio.uio_segflg = UIO_USERSPACE;
764 			uio.uio_rw = UIO_WRITE;
765 			uio.uio_td = td;
766 			iomode = NFSWRITE_FILESYNC;
767 			error = ncl_writerpc(vp, &uio, cred, &iomode,
768 			    &must_commit, 0);
769 			KASSERT((must_commit == 0),
770 				("ncl_directio_write: Did not commit write"));
771 			if (error)
772 				return (error);
773 			uiop->uio_offset += size;
774 			uiop->uio_resid -= size;
775 			if (uiop->uio_iov->iov_len <= size) {
776 				uiop->uio_iovcnt--;
777 				uiop->uio_iov++;
778 			} else {
779 				uiop->uio_iov->iov_base =
780 					(char *)uiop->uio_iov->iov_base + size;
781 				uiop->uio_iov->iov_len -= size;
782 			}
783 		}
784 	} else {
785 		struct uio *t_uio;
786 		struct iovec *t_iov;
787 		struct buf *bp;
788 
789 		/*
790 		 * Break up the write into blocksize chunks and hand these
791 		 * over to nfsiod's for write back.
792 		 * Unfortunately, this incurs a copy of the data. Since
793 		 * the user could modify the buffer before the write is
794 		 * initiated.
795 		 *
796 		 * The obvious optimization here is that one of the 2 copies
797 		 * in the async write path can be eliminated by copying the
798 		 * data here directly into mbufs and passing the mbuf chain
799 		 * down. But that will require a fair amount of re-working
800 		 * of the code and can be done if there's enough interest
801 		 * in NFS directio access.
802 		 */
803 		while (uiop->uio_resid > 0) {
804 			size = MIN(uiop->uio_resid, wsize);
805 			size = MIN(uiop->uio_iov->iov_len, size);
806 			bp = getpbuf(&ncl_pbuf_freecnt);
807 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
808 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
809 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
810 			t_iov->iov_len = size;
811 			t_uio->uio_iov = t_iov;
812 			t_uio->uio_iovcnt = 1;
813 			t_uio->uio_offset = uiop->uio_offset;
814 			t_uio->uio_resid = size;
815 			t_uio->uio_segflg = UIO_SYSSPACE;
816 			t_uio->uio_rw = UIO_WRITE;
817 			t_uio->uio_td = td;
818 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
819 			    uiop->uio_segflg == UIO_SYSSPACE,
820 			    ("nfs_directio_write: Bad uio_segflg"));
821 			if (uiop->uio_segflg == UIO_USERSPACE) {
822 				error = copyin(uiop->uio_iov->iov_base,
823 				    t_iov->iov_base, size);
824 				if (error != 0)
825 					goto err_free;
826 			} else
827 				/*
828 				 * UIO_SYSSPACE may never happen, but handle
829 				 * it just in case it does.
830 				 */
831 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
832 				    size);
833 			bp->b_flags |= B_DIRECT;
834 			bp->b_iocmd = BIO_WRITE;
835 			if (cred != NOCRED) {
836 				crhold(cred);
837 				bp->b_wcred = cred;
838 			} else
839 				bp->b_wcred = NOCRED;
840 			bp->b_caller1 = (void *)t_uio;
841 			bp->b_vp = vp;
842 			error = ncl_asyncio(nmp, bp, NOCRED, td);
843 err_free:
844 			if (error) {
845 				free(t_iov->iov_base, M_NFSDIRECTIO);
846 				free(t_iov, M_NFSDIRECTIO);
847 				free(t_uio, M_NFSDIRECTIO);
848 				bp->b_vp = NULL;
849 				relpbuf(bp, &ncl_pbuf_freecnt);
850 				if (error == EINTR)
851 					return (error);
852 				goto do_sync;
853 			}
854 			uiop->uio_offset += size;
855 			uiop->uio_resid -= size;
856 			if (uiop->uio_iov->iov_len <= size) {
857 				uiop->uio_iovcnt--;
858 				uiop->uio_iov++;
859 			} else {
860 				uiop->uio_iov->iov_base =
861 					(char *)uiop->uio_iov->iov_base + size;
862 				uiop->uio_iov->iov_len -= size;
863 			}
864 		}
865 	}
866 	return (0);
867 }
868 
869 /*
870  * Vnode op for write using bio
871  */
872 int
873 ncl_write(struct vop_write_args *ap)
874 {
875 	int biosize;
876 	struct uio *uio = ap->a_uio;
877 	struct thread *td = uio->uio_td;
878 	struct vnode *vp = ap->a_vp;
879 	struct nfsnode *np = VTONFS(vp);
880 	struct ucred *cred = ap->a_cred;
881 	int ioflag = ap->a_ioflag;
882 	struct buf *bp;
883 	struct vattr vattr;
884 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
885 	daddr_t lbn;
886 	int bcount, noncontig_write, obcount;
887 	int bp_cached, n, on, error = 0, error1, wouldcommit;
888 	size_t orig_resid, local_resid;
889 	off_t orig_size, tmp_off;
890 
891 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
892 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
893 	    ("ncl_write proc"));
894 	if (vp->v_type != VREG)
895 		return (EIO);
896 	mtx_lock(&np->n_mtx);
897 	if (np->n_flag & NWRITEERR) {
898 		np->n_flag &= ~NWRITEERR;
899 		mtx_unlock(&np->n_mtx);
900 		return (np->n_error);
901 	} else
902 		mtx_unlock(&np->n_mtx);
903 	mtx_lock(&nmp->nm_mtx);
904 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
905 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
906 		mtx_unlock(&nmp->nm_mtx);
907 		(void)ncl_fsinfo(nmp, vp, cred, td);
908 		mtx_lock(&nmp->nm_mtx);
909 	}
910 	if (nmp->nm_wsize == 0)
911 		(void) newnfs_iosize(nmp);
912 	mtx_unlock(&nmp->nm_mtx);
913 
914 	/*
915 	 * Synchronously flush pending buffers if we are in synchronous
916 	 * mode or if we are appending.
917 	 */
918 	if (ioflag & (IO_APPEND | IO_SYNC)) {
919 		mtx_lock(&np->n_mtx);
920 		if (np->n_flag & NMODIFIED) {
921 			mtx_unlock(&np->n_mtx);
922 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
923 			/*
924 			 * Require non-blocking, synchronous writes to
925 			 * dirty files to inform the program it needs
926 			 * to fsync(2) explicitly.
927 			 */
928 			if (ioflag & IO_NDELAY)
929 				return (EAGAIN);
930 #endif
931 			np->n_attrstamp = 0;
932 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
933 			error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
934 			    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
935 			if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
936 				error = EBADF;
937 			if (error != 0)
938 				return (error);
939 		} else
940 			mtx_unlock(&np->n_mtx);
941 	}
942 
943 	orig_resid = uio->uio_resid;
944 	mtx_lock(&np->n_mtx);
945 	orig_size = np->n_size;
946 	mtx_unlock(&np->n_mtx);
947 
948 	/*
949 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
950 	 * get the append lock.
951 	 */
952 	if (ioflag & IO_APPEND) {
953 		np->n_attrstamp = 0;
954 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
955 		error = VOP_GETATTR(vp, &vattr, cred);
956 		if (error)
957 			return (error);
958 		mtx_lock(&np->n_mtx);
959 		uio->uio_offset = np->n_size;
960 		mtx_unlock(&np->n_mtx);
961 	}
962 
963 	if (uio->uio_offset < 0)
964 		return (EINVAL);
965 	tmp_off = uio->uio_offset + uio->uio_resid;
966 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
967 		return (EFBIG);
968 	if (uio->uio_resid == 0)
969 		return (0);
970 
971 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
972 		return nfs_directio_write(vp, uio, cred, ioflag);
973 
974 	/*
975 	 * Maybe this should be above the vnode op call, but so long as
976 	 * file servers have no limits, i don't think it matters
977 	 */
978 	if (vn_rlimit_fsize(vp, uio, td))
979 		return (EFBIG);
980 
981 	biosize = vp->v_bufobj.bo_bsize;
982 	/*
983 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
984 	 * would exceed the local maximum per-file write commit size when
985 	 * combined with those, we must decide whether to flush,
986 	 * go synchronous, or return error.  We don't bother checking
987 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
988 	 * no point optimizing for something that really won't ever happen.
989 	 */
990 	wouldcommit = 0;
991 	if (!(ioflag & IO_SYNC)) {
992 		int nflag;
993 
994 		mtx_lock(&np->n_mtx);
995 		nflag = np->n_flag;
996 		mtx_unlock(&np->n_mtx);
997 		if (nflag & NMODIFIED) {
998 			BO_LOCK(&vp->v_bufobj);
999 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1000 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1001 				    b_bobufs) {
1002 					if (bp->b_flags & B_NEEDCOMMIT)
1003 						wouldcommit += bp->b_bcount;
1004 				}
1005 			}
1006 			BO_UNLOCK(&vp->v_bufobj);
1007 		}
1008 	}
1009 
1010 	do {
1011 		if (!(ioflag & IO_SYNC)) {
1012 			wouldcommit += biosize;
1013 			if (wouldcommit > nmp->nm_wcommitsize) {
1014 				np->n_attrstamp = 0;
1015 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1016 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1017 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1018 				if (error == 0 &&
1019 				    (vp->v_iflag & VI_DOOMED) != 0)
1020 					error = EBADF;
1021 				if (error != 0)
1022 					return (error);
1023 				wouldcommit = biosize;
1024 			}
1025 		}
1026 
1027 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1028 		lbn = uio->uio_offset / biosize;
1029 		on = uio->uio_offset - (lbn * biosize);
1030 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1031 again:
1032 		/*
1033 		 * Handle direct append and file extension cases, calculate
1034 		 * unaligned buffer size.
1035 		 */
1036 		mtx_lock(&np->n_mtx);
1037 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1038 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1039 			noncontig_write = 1;
1040 		else
1041 			noncontig_write = 0;
1042 		if ((uio->uio_offset == np->n_size ||
1043 		    (noncontig_write != 0 &&
1044 		    lbn == (np->n_size / biosize) &&
1045 		    uio->uio_offset + n > np->n_size)) && n) {
1046 			mtx_unlock(&np->n_mtx);
1047 			/*
1048 			 * Get the buffer (in its pre-append state to maintain
1049 			 * B_CACHE if it was previously set).  Resize the
1050 			 * nfsnode after we have locked the buffer to prevent
1051 			 * readers from reading garbage.
1052 			 */
1053 			obcount = np->n_size - (lbn * biosize);
1054 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1055 
1056 			if (bp != NULL) {
1057 				long save;
1058 
1059 				mtx_lock(&np->n_mtx);
1060 				np->n_size = uio->uio_offset + n;
1061 				np->n_flag |= NMODIFIED;
1062 				vnode_pager_setsize(vp, np->n_size);
1063 				mtx_unlock(&np->n_mtx);
1064 
1065 				save = bp->b_flags & B_CACHE;
1066 				bcount = on + n;
1067 				allocbuf(bp, bcount);
1068 				bp->b_flags |= save;
1069 				if (noncontig_write != 0 && on > obcount)
1070 					vfs_bio_bzero_buf(bp, obcount, on -
1071 					    obcount);
1072 			}
1073 		} else {
1074 			/*
1075 			 * Obtain the locked cache block first, and then
1076 			 * adjust the file's size as appropriate.
1077 			 */
1078 			bcount = on + n;
1079 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1080 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1081 					bcount = biosize;
1082 				else
1083 					bcount = np->n_size - (off_t)lbn * biosize;
1084 			}
1085 			mtx_unlock(&np->n_mtx);
1086 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1087 			mtx_lock(&np->n_mtx);
1088 			if (uio->uio_offset + n > np->n_size) {
1089 				np->n_size = uio->uio_offset + n;
1090 				np->n_flag |= NMODIFIED;
1091 				vnode_pager_setsize(vp, np->n_size);
1092 			}
1093 			mtx_unlock(&np->n_mtx);
1094 		}
1095 
1096 		if (!bp) {
1097 			error = newnfs_sigintr(nmp, td);
1098 			if (!error)
1099 				error = EINTR;
1100 			break;
1101 		}
1102 
1103 		/*
1104 		 * Issue a READ if B_CACHE is not set.  In special-append
1105 		 * mode, B_CACHE is based on the buffer prior to the write
1106 		 * op and is typically set, avoiding the read.  If a read
1107 		 * is required in special append mode, the server will
1108 		 * probably send us a short-read since we extended the file
1109 		 * on our end, resulting in b_resid == 0 and, thusly,
1110 		 * B_CACHE getting set.
1111 		 *
1112 		 * We can also avoid issuing the read if the write covers
1113 		 * the entire buffer.  We have to make sure the buffer state
1114 		 * is reasonable in this case since we will not be initiating
1115 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1116 		 * more information.
1117 		 *
1118 		 * B_CACHE may also be set due to the buffer being cached
1119 		 * normally.
1120 		 */
1121 
1122 		bp_cached = 1;
1123 		if (on == 0 && n == bcount) {
1124 			if ((bp->b_flags & B_CACHE) == 0)
1125 				bp_cached = 0;
1126 			bp->b_flags |= B_CACHE;
1127 			bp->b_flags &= ~B_INVAL;
1128 			bp->b_ioflags &= ~BIO_ERROR;
1129 		}
1130 
1131 		if ((bp->b_flags & B_CACHE) == 0) {
1132 			bp->b_iocmd = BIO_READ;
1133 			vfs_busy_pages(bp, 0);
1134 			error = ncl_doio(vp, bp, cred, td, 0);
1135 			if (error) {
1136 				brelse(bp);
1137 				break;
1138 			}
1139 		}
1140 		if (bp->b_wcred == NOCRED)
1141 			bp->b_wcred = crhold(cred);
1142 		mtx_lock(&np->n_mtx);
1143 		np->n_flag |= NMODIFIED;
1144 		mtx_unlock(&np->n_mtx);
1145 
1146 		/*
1147 		 * If dirtyend exceeds file size, chop it down.  This should
1148 		 * not normally occur but there is an append race where it
1149 		 * might occur XXX, so we log it.
1150 		 *
1151 		 * If the chopping creates a reverse-indexed or degenerate
1152 		 * situation with dirtyoff/end, we 0 both of them.
1153 		 */
1154 
1155 		if (bp->b_dirtyend > bcount) {
1156 			printf("NFS append race @%lx:%d\n",
1157 			    (long)bp->b_blkno * DEV_BSIZE,
1158 			    bp->b_dirtyend - bcount);
1159 			bp->b_dirtyend = bcount;
1160 		}
1161 
1162 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1163 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1164 
1165 		/*
1166 		 * If the new write will leave a contiguous dirty
1167 		 * area, just update the b_dirtyoff and b_dirtyend,
1168 		 * otherwise force a write rpc of the old dirty area.
1169 		 *
1170 		 * If there has been a file lock applied to this file
1171 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1172 		 * While it is possible to merge discontiguous writes due to
1173 		 * our having a B_CACHE buffer ( and thus valid read data
1174 		 * for the hole), we don't because it could lead to
1175 		 * significant cache coherency problems with multiple clients,
1176 		 * especially if locking is implemented later on.
1177 		 *
1178 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1179 		 * not been file locking done on this file:
1180 		 * Relax coherency a bit for the sake of performance and
1181 		 * expand the current dirty region to contain the new
1182 		 * write even if it means we mark some non-dirty data as
1183 		 * dirty.
1184 		 */
1185 
1186 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1187 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1188 			if (bwrite(bp) == EINTR) {
1189 				error = EINTR;
1190 				break;
1191 			}
1192 			goto again;
1193 		}
1194 
1195 		local_resid = uio->uio_resid;
1196 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1197 
1198 		if (error != 0 && !bp_cached) {
1199 			/*
1200 			 * This block has no other content then what
1201 			 * possibly was written by the faulty uiomove.
1202 			 * Release it, forgetting the data pages, to
1203 			 * prevent the leak of uninitialized data to
1204 			 * usermode.
1205 			 */
1206 			bp->b_ioflags |= BIO_ERROR;
1207 			brelse(bp);
1208 			uio->uio_offset -= local_resid - uio->uio_resid;
1209 			uio->uio_resid = local_resid;
1210 			break;
1211 		}
1212 
1213 		/*
1214 		 * Since this block is being modified, it must be written
1215 		 * again and not just committed.  Since write clustering does
1216 		 * not work for the stage 1 data write, only the stage 2
1217 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1218 		 */
1219 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1220 
1221 		/*
1222 		 * Get the partial update on the progress made from
1223 		 * uiomove, if an error occurred.
1224 		 */
1225 		if (error != 0)
1226 			n = local_resid - uio->uio_resid;
1227 
1228 		/*
1229 		 * Only update dirtyoff/dirtyend if not a degenerate
1230 		 * condition.
1231 		 */
1232 		if (n > 0) {
1233 			if (bp->b_dirtyend > 0) {
1234 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1235 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1236 			} else {
1237 				bp->b_dirtyoff = on;
1238 				bp->b_dirtyend = on + n;
1239 			}
1240 			vfs_bio_set_valid(bp, on, n);
1241 		}
1242 
1243 		/*
1244 		 * If IO_SYNC do bwrite().
1245 		 *
1246 		 * IO_INVAL appears to be unused.  The idea appears to be
1247 		 * to turn off caching in this case.  Very odd.  XXX
1248 		 */
1249 		if ((ioflag & IO_SYNC)) {
1250 			if (ioflag & IO_INVAL)
1251 				bp->b_flags |= B_NOCACHE;
1252 			error1 = bwrite(bp);
1253 			if (error1 != 0) {
1254 				if (error == 0)
1255 					error = error1;
1256 				break;
1257 			}
1258 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1259 			bp->b_flags |= B_ASYNC;
1260 			(void) ncl_writebp(bp, 0, NULL);
1261 		} else {
1262 			bdwrite(bp);
1263 		}
1264 
1265 		if (error != 0)
1266 			break;
1267 	} while (uio->uio_resid > 0 && n > 0);
1268 
1269 	if (error != 0) {
1270 		if (ioflag & IO_UNIT) {
1271 			VATTR_NULL(&vattr);
1272 			vattr.va_size = orig_size;
1273 			/* IO_SYNC is handled implicitely */
1274 			(void)VOP_SETATTR(vp, &vattr, cred);
1275 			uio->uio_offset -= orig_resid - uio->uio_resid;
1276 			uio->uio_resid = orig_resid;
1277 		}
1278 	}
1279 
1280 	return (error);
1281 }
1282 
1283 /*
1284  * Get an nfs cache block.
1285  *
1286  * Allocate a new one if the block isn't currently in the cache
1287  * and return the block marked busy. If the calling process is
1288  * interrupted by a signal for an interruptible mount point, return
1289  * NULL.
1290  *
1291  * The caller must carefully deal with the possible B_INVAL state of
1292  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1293  * indirectly), so synchronous reads can be issued without worrying about
1294  * the B_INVAL state.  We have to be a little more careful when dealing
1295  * with writes (see comments in nfs_write()) when extending a file past
1296  * its EOF.
1297  */
1298 static struct buf *
1299 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1300 {
1301 	struct buf *bp;
1302 	struct mount *mp;
1303 	struct nfsmount *nmp;
1304 
1305 	mp = vp->v_mount;
1306 	nmp = VFSTONFS(mp);
1307 
1308 	if (nmp->nm_flag & NFSMNT_INT) {
1309 		sigset_t oldset;
1310 
1311 		newnfs_set_sigmask(td, &oldset);
1312 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1313 		newnfs_restore_sigmask(td, &oldset);
1314 		while (bp == NULL) {
1315 			if (newnfs_sigintr(nmp, td))
1316 				return (NULL);
1317 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1318 		}
1319 	} else {
1320 		bp = getblk(vp, bn, size, 0, 0, 0);
1321 	}
1322 
1323 	if (vp->v_type == VREG)
1324 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1325 	return (bp);
1326 }
1327 
1328 /*
1329  * Flush and invalidate all dirty buffers. If another process is already
1330  * doing the flush, just wait for completion.
1331  */
1332 int
1333 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1334 {
1335 	struct nfsnode *np = VTONFS(vp);
1336 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1337 	int error = 0, slpflag, slptimeo;
1338 	int old_lock = 0;
1339 
1340 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1341 
1342 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1343 		intrflg = 0;
1344 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1345 		intrflg = 1;
1346 	if (intrflg) {
1347 		slpflag = PCATCH;
1348 		slptimeo = 2 * hz;
1349 	} else {
1350 		slpflag = 0;
1351 		slptimeo = 0;
1352 	}
1353 
1354 	old_lock = ncl_upgrade_vnlock(vp);
1355 	if (vp->v_iflag & VI_DOOMED) {
1356 		/*
1357 		 * Since vgonel() uses the generic vinvalbuf() to flush
1358 		 * dirty buffers and it does not call this function, it
1359 		 * is safe to just return OK when VI_DOOMED is set.
1360 		 */
1361 		ncl_downgrade_vnlock(vp, old_lock);
1362 		return (0);
1363 	}
1364 
1365 	/*
1366 	 * Now, flush as required.
1367 	 */
1368 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1369 	     vp->v_bufobj.bo_object != NULL) {
1370 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1371 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1372 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1373 		/*
1374 		 * If the page clean was interrupted, fail the invalidation.
1375 		 * Not doing so, we run the risk of losing dirty pages in the
1376 		 * vinvalbuf() call below.
1377 		 */
1378 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1379 			goto out;
1380 	}
1381 
1382 	error = vinvalbuf(vp, flags, slpflag, 0);
1383 	while (error) {
1384 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1385 			goto out;
1386 		error = vinvalbuf(vp, flags, 0, slptimeo);
1387 	}
1388 	if (NFSHASPNFS(nmp)) {
1389 		nfscl_layoutcommit(vp, td);
1390 		/*
1391 		 * Invalidate the attribute cache, since writes to a DS
1392 		 * won't update the size attribute.
1393 		 */
1394 		mtx_lock(&np->n_mtx);
1395 		np->n_attrstamp = 0;
1396 	} else
1397 		mtx_lock(&np->n_mtx);
1398 	if (np->n_directio_asyncwr == 0)
1399 		np->n_flag &= ~NMODIFIED;
1400 	mtx_unlock(&np->n_mtx);
1401 out:
1402 	ncl_downgrade_vnlock(vp, old_lock);
1403 	return error;
1404 }
1405 
1406 /*
1407  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1408  * This is mainly to avoid queueing async I/O requests when the nfsiods
1409  * are all hung on a dead server.
1410  *
1411  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1412  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1413  */
1414 int
1415 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1416 {
1417 	int iod;
1418 	int gotiod;
1419 	int slpflag = 0;
1420 	int slptimeo = 0;
1421 	int error, error2;
1422 
1423 	/*
1424 	 * Commits are usually short and sweet so lets save some cpu and
1425 	 * leave the async daemons for more important rpc's (such as reads
1426 	 * and writes).
1427 	 *
1428 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1429 	 * in the directory in order to update attributes. This can deadlock
1430 	 * with another thread that is waiting for async I/O to be done by
1431 	 * an nfsiod thread while holding a lock on one of these vnodes.
1432 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1433 	 * perform Readdirplus RPCs.
1434 	 */
1435 	mtx_lock(&ncl_iod_mutex);
1436 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1437 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1438 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1439 		mtx_unlock(&ncl_iod_mutex);
1440 		return(EIO);
1441 	}
1442 again:
1443 	if (nmp->nm_flag & NFSMNT_INT)
1444 		slpflag = PCATCH;
1445 	gotiod = FALSE;
1446 
1447 	/*
1448 	 * Find a free iod to process this request.
1449 	 */
1450 	for (iod = 0; iod < ncl_numasync; iod++)
1451 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1452 			gotiod = TRUE;
1453 			break;
1454 		}
1455 
1456 	/*
1457 	 * Try to create one if none are free.
1458 	 */
1459 	if (!gotiod)
1460 		ncl_nfsiodnew();
1461 	else {
1462 		/*
1463 		 * Found one, so wake it up and tell it which
1464 		 * mount to process.
1465 		 */
1466 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1467 		    iod, nmp));
1468 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1469 		ncl_iodmount[iod] = nmp;
1470 		nmp->nm_bufqiods++;
1471 		wakeup(&ncl_iodwant[iod]);
1472 	}
1473 
1474 	/*
1475 	 * If none are free, we may already have an iod working on this mount
1476 	 * point.  If so, it will process our request.
1477 	 */
1478 	if (!gotiod) {
1479 		if (nmp->nm_bufqiods > 0) {
1480 			NFS_DPF(ASYNCIO,
1481 				("ncl_asyncio: %d iods are already processing mount %p\n",
1482 				 nmp->nm_bufqiods, nmp));
1483 			gotiod = TRUE;
1484 		}
1485 	}
1486 
1487 	/*
1488 	 * If we have an iod which can process the request, then queue
1489 	 * the buffer.
1490 	 */
1491 	if (gotiod) {
1492 		/*
1493 		 * Ensure that the queue never grows too large.  We still want
1494 		 * to asynchronize so we block rather then return EIO.
1495 		 */
1496 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1497 			NFS_DPF(ASYNCIO,
1498 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1499 			nmp->nm_bufqwant = TRUE;
1500 			error = newnfs_msleep(td, &nmp->nm_bufq,
1501 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1502 			   slptimeo);
1503 			if (error) {
1504 				error2 = newnfs_sigintr(nmp, td);
1505 				if (error2) {
1506 					mtx_unlock(&ncl_iod_mutex);
1507 					return (error2);
1508 				}
1509 				if (slpflag == PCATCH) {
1510 					slpflag = 0;
1511 					slptimeo = 2 * hz;
1512 				}
1513 			}
1514 			/*
1515 			 * We might have lost our iod while sleeping,
1516 			 * so check and loop if necessary.
1517 			 */
1518 			goto again;
1519 		}
1520 
1521 		/* We might have lost our nfsiod */
1522 		if (nmp->nm_bufqiods == 0) {
1523 			NFS_DPF(ASYNCIO,
1524 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1525 			goto again;
1526 		}
1527 
1528 		if (bp->b_iocmd == BIO_READ) {
1529 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1530 				bp->b_rcred = crhold(cred);
1531 		} else {
1532 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1533 				bp->b_wcred = crhold(cred);
1534 		}
1535 
1536 		if (bp->b_flags & B_REMFREE)
1537 			bremfreef(bp);
1538 		BUF_KERNPROC(bp);
1539 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1540 		nmp->nm_bufqlen++;
1541 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1542 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1543 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1544 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1545 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1546 		}
1547 		mtx_unlock(&ncl_iod_mutex);
1548 		return (0);
1549 	}
1550 
1551 	mtx_unlock(&ncl_iod_mutex);
1552 
1553 	/*
1554 	 * All the iods are busy on other mounts, so return EIO to
1555 	 * force the caller to process the i/o synchronously.
1556 	 */
1557 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1558 	return (EIO);
1559 }
1560 
1561 void
1562 ncl_doio_directwrite(struct buf *bp)
1563 {
1564 	int iomode, must_commit;
1565 	struct uio *uiop = (struct uio *)bp->b_caller1;
1566 	char *iov_base = uiop->uio_iov->iov_base;
1567 
1568 	iomode = NFSWRITE_FILESYNC;
1569 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1570 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1571 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1572 	free(iov_base, M_NFSDIRECTIO);
1573 	free(uiop->uio_iov, M_NFSDIRECTIO);
1574 	free(uiop, M_NFSDIRECTIO);
1575 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1576 		struct nfsnode *np = VTONFS(bp->b_vp);
1577 		mtx_lock(&np->n_mtx);
1578 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1579 			/*
1580 			 * Invalidate the attribute cache, since writes to a DS
1581 			 * won't update the size attribute.
1582 			 */
1583 			np->n_attrstamp = 0;
1584 		}
1585 		np->n_directio_asyncwr--;
1586 		if (np->n_directio_asyncwr == 0) {
1587 			np->n_flag &= ~NMODIFIED;
1588 			if ((np->n_flag & NFSYNCWAIT)) {
1589 				np->n_flag &= ~NFSYNCWAIT;
1590 				wakeup((caddr_t)&np->n_directio_asyncwr);
1591 			}
1592 		}
1593 		mtx_unlock(&np->n_mtx);
1594 	}
1595 	bp->b_vp = NULL;
1596 	relpbuf(bp, &ncl_pbuf_freecnt);
1597 }
1598 
1599 /*
1600  * Do an I/O operation to/from a cache block. This may be called
1601  * synchronously or from an nfsiod.
1602  */
1603 int
1604 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1605     int called_from_strategy)
1606 {
1607 	struct uio *uiop;
1608 	struct nfsnode *np;
1609 	struct nfsmount *nmp;
1610 	int error = 0, iomode, must_commit = 0;
1611 	struct uio uio;
1612 	struct iovec io;
1613 	struct proc *p = td ? td->td_proc : NULL;
1614 	uint8_t	iocmd;
1615 
1616 	np = VTONFS(vp);
1617 	nmp = VFSTONFS(vp->v_mount);
1618 	uiop = &uio;
1619 	uiop->uio_iov = &io;
1620 	uiop->uio_iovcnt = 1;
1621 	uiop->uio_segflg = UIO_SYSSPACE;
1622 	uiop->uio_td = td;
1623 
1624 	/*
1625 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1626 	 * do this here so we do not have to do it in all the code that
1627 	 * calls us.
1628 	 */
1629 	bp->b_flags &= ~B_INVAL;
1630 	bp->b_ioflags &= ~BIO_ERROR;
1631 
1632 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1633 	iocmd = bp->b_iocmd;
1634 	if (iocmd == BIO_READ) {
1635 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1636 	    io.iov_base = bp->b_data;
1637 	    uiop->uio_rw = UIO_READ;
1638 
1639 	    switch (vp->v_type) {
1640 	    case VREG:
1641 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1642 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1643 		error = ncl_readrpc(vp, uiop, cr);
1644 
1645 		if (!error) {
1646 		    if (uiop->uio_resid) {
1647 			/*
1648 			 * If we had a short read with no error, we must have
1649 			 * hit a file hole.  We should zero-fill the remainder.
1650 			 * This can also occur if the server hits the file EOF.
1651 			 *
1652 			 * Holes used to be able to occur due to pending
1653 			 * writes, but that is not possible any longer.
1654 			 */
1655 			int nread = bp->b_bcount - uiop->uio_resid;
1656 			ssize_t left = uiop->uio_resid;
1657 
1658 			if (left > 0)
1659 				bzero((char *)bp->b_data + nread, left);
1660 			uiop->uio_resid = 0;
1661 		    }
1662 		}
1663 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1664 		if (p && (vp->v_vflag & VV_TEXT)) {
1665 			mtx_lock(&np->n_mtx);
1666 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1667 				mtx_unlock(&np->n_mtx);
1668 				PROC_LOCK(p);
1669 				killproc(p, "text file modification");
1670 				PROC_UNLOCK(p);
1671 			} else
1672 				mtx_unlock(&np->n_mtx);
1673 		}
1674 		break;
1675 	    case VLNK:
1676 		uiop->uio_offset = (off_t)0;
1677 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1678 		error = ncl_readlinkrpc(vp, uiop, cr);
1679 		break;
1680 	    case VDIR:
1681 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1682 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1683 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1684 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1685 			if (error == NFSERR_NOTSUPP)
1686 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1687 		}
1688 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1689 			error = ncl_readdirrpc(vp, uiop, cr, td);
1690 		/*
1691 		 * end-of-directory sets B_INVAL but does not generate an
1692 		 * error.
1693 		 */
1694 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1695 			bp->b_flags |= B_INVAL;
1696 		break;
1697 	    default:
1698 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1699 		break;
1700 	    }
1701 	    if (error) {
1702 		bp->b_ioflags |= BIO_ERROR;
1703 		bp->b_error = error;
1704 	    }
1705 	} else {
1706 	    /*
1707 	     * If we only need to commit, try to commit
1708 	     */
1709 	    if (bp->b_flags & B_NEEDCOMMIT) {
1710 		    int retv;
1711 		    off_t off;
1712 
1713 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1714 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1715 			bp->b_wcred, td);
1716 		    if (retv == 0) {
1717 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1718 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1719 			    bp->b_resid = 0;
1720 			    bufdone(bp);
1721 			    return (0);
1722 		    }
1723 		    if (retv == NFSERR_STALEWRITEVERF) {
1724 			    ncl_clearcommit(vp->v_mount);
1725 		    }
1726 	    }
1727 
1728 	    /*
1729 	     * Setup for actual write
1730 	     */
1731 	    mtx_lock(&np->n_mtx);
1732 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1733 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1734 	    mtx_unlock(&np->n_mtx);
1735 
1736 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1737 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1738 		    - bp->b_dirtyoff;
1739 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1740 		    + bp->b_dirtyoff;
1741 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1742 		uiop->uio_rw = UIO_WRITE;
1743 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1744 
1745 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1746 		    iomode = NFSWRITE_UNSTABLE;
1747 		else
1748 		    iomode = NFSWRITE_FILESYNC;
1749 
1750 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1751 		    called_from_strategy);
1752 
1753 		/*
1754 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1755 		 * to cluster the buffers needing commit.  This will allow
1756 		 * the system to submit a single commit rpc for the whole
1757 		 * cluster.  We can do this even if the buffer is not 100%
1758 		 * dirty (relative to the NFS blocksize), so we optimize the
1759 		 * append-to-file-case.
1760 		 *
1761 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1762 		 * cleared because write clustering only works for commit
1763 		 * rpc's, not for the data portion of the write).
1764 		 */
1765 
1766 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1767 		    bp->b_flags |= B_NEEDCOMMIT;
1768 		    if (bp->b_dirtyoff == 0
1769 			&& bp->b_dirtyend == bp->b_bcount)
1770 			bp->b_flags |= B_CLUSTEROK;
1771 		} else {
1772 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1773 		}
1774 
1775 		/*
1776 		 * For an interrupted write, the buffer is still valid
1777 		 * and the write hasn't been pushed to the server yet,
1778 		 * so we can't set BIO_ERROR and report the interruption
1779 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1780 		 * is not relevant, so the rpc attempt is essentially
1781 		 * a noop.  For the case of a V3 write rpc not being
1782 		 * committed to stable storage, the block is still
1783 		 * dirty and requires either a commit rpc or another
1784 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1785 		 * the block is reused. This is indicated by setting
1786 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1787 		 *
1788 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1789 		 * write error and is handled as above, except that
1790 		 * B_EINTR isn't set. One cause of this is a stale stateid
1791 		 * error for the RPC that indicates recovery is required,
1792 		 * when called with called_from_strategy != 0.
1793 		 *
1794 		 * If the buffer is marked B_PAGING, it does not reside on
1795 		 * the vp's paging queues so we cannot call bdirty().  The
1796 		 * bp in this case is not an NFS cache block so we should
1797 		 * be safe. XXX
1798 		 *
1799 		 * The logic below breaks up errors into recoverable and
1800 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1801 		 * and keep the buffer around for potential write retries.
1802 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1803 		 * and save the error in the nfsnode. This is less than ideal
1804 		 * but necessary. Keeping such buffers around could potentially
1805 		 * cause buffer exhaustion eventually (they can never be written
1806 		 * out, so will get constantly be re-dirtied). It also causes
1807 		 * all sorts of vfs panics. For non-recoverable write errors,
1808 		 * also invalidate the attrcache, so we'll be forced to go over
1809 		 * the wire for this object, returning an error to user on next
1810 		 * call (most of the time).
1811 		 */
1812 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1813 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1814 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1815 			if ((bp->b_flags & B_PAGING) == 0) {
1816 			    bdirty(bp);
1817 			    bp->b_flags &= ~B_DONE;
1818 			}
1819 			if ((error == EINTR || error == ETIMEDOUT) &&
1820 			    (bp->b_flags & B_ASYNC) == 0)
1821 			    bp->b_flags |= B_EINTR;
1822 		} else {
1823 		    if (error) {
1824 			bp->b_ioflags |= BIO_ERROR;
1825 			bp->b_flags |= B_INVAL;
1826 			bp->b_error = np->n_error = error;
1827 			mtx_lock(&np->n_mtx);
1828 			np->n_flag |= NWRITEERR;
1829 			np->n_attrstamp = 0;
1830 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1831 			mtx_unlock(&np->n_mtx);
1832 		    }
1833 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1834 		}
1835 	    } else {
1836 		bp->b_resid = 0;
1837 		bufdone(bp);
1838 		return (0);
1839 	    }
1840 	}
1841 	bp->b_resid = uiop->uio_resid;
1842 	if (must_commit)
1843 	    ncl_clearcommit(vp->v_mount);
1844 	bufdone(bp);
1845 	return (error);
1846 }
1847 
1848 /*
1849  * Used to aid in handling ftruncate() operations on the NFS client side.
1850  * Truncation creates a number of special problems for NFS.  We have to
1851  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1852  * we have to properly handle VM pages or (potentially dirty) buffers
1853  * that straddle the truncation point.
1854  */
1855 
1856 int
1857 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1858 {
1859 	struct nfsnode *np = VTONFS(vp);
1860 	u_quad_t tsize;
1861 	int biosize = vp->v_bufobj.bo_bsize;
1862 	int error = 0;
1863 
1864 	mtx_lock(&np->n_mtx);
1865 	tsize = np->n_size;
1866 	np->n_size = nsize;
1867 	mtx_unlock(&np->n_mtx);
1868 
1869 	if (nsize < tsize) {
1870 		struct buf *bp;
1871 		daddr_t lbn;
1872 		int bufsize;
1873 
1874 		/*
1875 		 * vtruncbuf() doesn't get the buffer overlapping the
1876 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1877 		 * buffer that now needs to be truncated.
1878 		 */
1879 		error = vtruncbuf(vp, cred, nsize, biosize);
1880 		lbn = nsize / biosize;
1881 		bufsize = nsize - (lbn * biosize);
1882 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1883 		if (!bp)
1884 			return EINTR;
1885 		if (bp->b_dirtyoff > bp->b_bcount)
1886 			bp->b_dirtyoff = bp->b_bcount;
1887 		if (bp->b_dirtyend > bp->b_bcount)
1888 			bp->b_dirtyend = bp->b_bcount;
1889 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1890 		brelse(bp);
1891 	} else {
1892 		vnode_pager_setsize(vp, nsize);
1893 	}
1894 	return(error);
1895 }
1896 
1897