xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 0957b409)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Rick Macklem at The University of Guelph.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bio.h>
43 #include <sys/buf.h>
44 #include <sys/kernel.h>
45 #include <sys/mount.h>
46 #include <sys/rwlock.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
49 
50 #include <vm/vm.h>
51 #include <vm/vm_param.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
57 
58 #include <fs/nfs/nfsport.h>
59 #include <fs/nfsclient/nfsmount.h>
60 #include <fs/nfsclient/nfs.h>
61 #include <fs/nfsclient/nfsnode.h>
62 #include <fs/nfsclient/nfs_kdtrace.h>
63 
64 extern int newnfs_directio_allow_mmap;
65 extern struct nfsstatsv1 nfsstatsv1;
66 extern struct mtx ncl_iod_mutex;
67 extern int ncl_numasync;
68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
70 extern int newnfs_directio_enable;
71 extern int nfs_keep_dirty_on_error;
72 
73 uma_zone_t ncl_pbuf_zone;
74 
75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
76     struct thread *td);
77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
78     struct ucred *cred, int ioflag);
79 
80 /*
81  * Vnode op for VM getpages.
82  */
83 SYSCTL_DECL(_vfs_nfs);
84 static int use_buf_pager = 1;
85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
86     &use_buf_pager, 0,
87     "Use buffer pager instead of direct readrpc call");
88 
89 static daddr_t
90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
91 {
92 
93 	return (off / vp->v_bufobj.bo_bsize);
94 }
95 
96 static int
97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn)
98 {
99 	struct nfsnode *np;
100 	u_quad_t nsize;
101 	int biosize, bcount;
102 
103 	np = VTONFS(vp);
104 	mtx_lock(&np->n_mtx);
105 	nsize = np->n_size;
106 	mtx_unlock(&np->n_mtx);
107 
108 	biosize = vp->v_bufobj.bo_bsize;
109 	bcount = biosize;
110 	if ((off_t)lbn * biosize >= nsize)
111 		bcount = 0;
112 	else if ((off_t)(lbn + 1) * biosize > nsize)
113 		bcount = nsize - (off_t)lbn * biosize;
114 	return (bcount);
115 }
116 
117 int
118 ncl_getpages(struct vop_getpages_args *ap)
119 {
120 	int i, error, nextoff, size, toff, count, npages;
121 	struct uio uio;
122 	struct iovec iov;
123 	vm_offset_t kva;
124 	struct buf *bp;
125 	struct vnode *vp;
126 	struct thread *td;
127 	struct ucred *cred;
128 	struct nfsmount *nmp;
129 	vm_object_t object;
130 	vm_page_t *pages;
131 	struct nfsnode *np;
132 
133 	vp = ap->a_vp;
134 	np = VTONFS(vp);
135 	td = curthread;
136 	cred = curthread->td_ucred;
137 	nmp = VFSTONFS(vp->v_mount);
138 	pages = ap->a_m;
139 	npages = ap->a_count;
140 
141 	if ((object = vp->v_object) == NULL) {
142 		printf("ncl_getpages: called with non-merged cache vnode\n");
143 		return (VM_PAGER_ERROR);
144 	}
145 
146 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
147 		mtx_lock(&np->n_mtx);
148 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
149 			mtx_unlock(&np->n_mtx);
150 			printf("ncl_getpages: called on non-cacheable vnode\n");
151 			return (VM_PAGER_ERROR);
152 		} else
153 			mtx_unlock(&np->n_mtx);
154 	}
155 
156 	mtx_lock(&nmp->nm_mtx);
157 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
158 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
159 		mtx_unlock(&nmp->nm_mtx);
160 		/* We'll never get here for v4, because we always have fsinfo */
161 		(void)ncl_fsinfo(nmp, vp, cred, td);
162 	} else
163 		mtx_unlock(&nmp->nm_mtx);
164 
165 	if (use_buf_pager)
166 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
167 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
168 
169 	/*
170 	 * If the requested page is partially valid, just return it and
171 	 * allow the pager to zero-out the blanks.  Partially valid pages
172 	 * can only occur at the file EOF.
173 	 *
174 	 * XXXGL: is that true for NFS, where short read can occur???
175 	 */
176 	VM_OBJECT_WLOCK(object);
177 	if (pages[npages - 1]->valid != 0 && --npages == 0)
178 		goto out;
179 	VM_OBJECT_WUNLOCK(object);
180 
181 	/*
182 	 * We use only the kva address for the buffer, but this is extremely
183 	 * convenient and fast.
184 	 */
185 	bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
186 
187 	kva = (vm_offset_t) bp->b_data;
188 	pmap_qenter(kva, pages, npages);
189 	VM_CNT_INC(v_vnodein);
190 	VM_CNT_ADD(v_vnodepgsin, npages);
191 
192 	count = npages << PAGE_SHIFT;
193 	iov.iov_base = (caddr_t) kva;
194 	iov.iov_len = count;
195 	uio.uio_iov = &iov;
196 	uio.uio_iovcnt = 1;
197 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
198 	uio.uio_resid = count;
199 	uio.uio_segflg = UIO_SYSSPACE;
200 	uio.uio_rw = UIO_READ;
201 	uio.uio_td = td;
202 
203 	error = ncl_readrpc(vp, &uio, cred);
204 	pmap_qremove(kva, npages);
205 
206 	uma_zfree(ncl_pbuf_zone, bp);
207 
208 	if (error && (uio.uio_resid == count)) {
209 		printf("ncl_getpages: error %d\n", error);
210 		return (VM_PAGER_ERROR);
211 	}
212 
213 	/*
214 	 * Calculate the number of bytes read and validate only that number
215 	 * of bytes.  Note that due to pending writes, size may be 0.  This
216 	 * does not mean that the remaining data is invalid!
217 	 */
218 
219 	size = count - uio.uio_resid;
220 	VM_OBJECT_WLOCK(object);
221 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
222 		vm_page_t m;
223 		nextoff = toff + PAGE_SIZE;
224 		m = pages[i];
225 
226 		if (nextoff <= size) {
227 			/*
228 			 * Read operation filled an entire page
229 			 */
230 			m->valid = VM_PAGE_BITS_ALL;
231 			KASSERT(m->dirty == 0,
232 			    ("nfs_getpages: page %p is dirty", m));
233 		} else if (size > toff) {
234 			/*
235 			 * Read operation filled a partial page.
236 			 */
237 			m->valid = 0;
238 			vm_page_set_valid_range(m, 0, size - toff);
239 			KASSERT(m->dirty == 0,
240 			    ("nfs_getpages: page %p is dirty", m));
241 		} else {
242 			/*
243 			 * Read operation was short.  If no error
244 			 * occurred we may have hit a zero-fill
245 			 * section.  We leave valid set to 0, and page
246 			 * is freed by vm_page_readahead_finish() if
247 			 * its index is not equal to requested, or
248 			 * page is zeroed and set valid by
249 			 * vm_pager_get_pages() for requested page.
250 			 */
251 			;
252 		}
253 	}
254 out:
255 	VM_OBJECT_WUNLOCK(object);
256 	if (ap->a_rbehind)
257 		*ap->a_rbehind = 0;
258 	if (ap->a_rahead)
259 		*ap->a_rahead = 0;
260 	return (VM_PAGER_OK);
261 }
262 
263 /*
264  * Vnode op for VM putpages.
265  */
266 int
267 ncl_putpages(struct vop_putpages_args *ap)
268 {
269 	struct uio uio;
270 	struct iovec iov;
271 	int i, error, npages, count;
272 	off_t offset;
273 	int *rtvals;
274 	struct vnode *vp;
275 	struct thread *td;
276 	struct ucred *cred;
277 	struct nfsmount *nmp;
278 	struct nfsnode *np;
279 	vm_page_t *pages;
280 
281 	vp = ap->a_vp;
282 	np = VTONFS(vp);
283 	td = curthread;				/* XXX */
284 	/* Set the cred to n_writecred for the write rpcs. */
285 	if (np->n_writecred != NULL)
286 		cred = crhold(np->n_writecred);
287 	else
288 		cred = crhold(curthread->td_ucred);	/* XXX */
289 	nmp = VFSTONFS(vp->v_mount);
290 	pages = ap->a_m;
291 	count = ap->a_count;
292 	rtvals = ap->a_rtvals;
293 	npages = btoc(count);
294 	offset = IDX_TO_OFF(pages[0]->pindex);
295 
296 	mtx_lock(&nmp->nm_mtx);
297 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
298 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
299 		mtx_unlock(&nmp->nm_mtx);
300 		(void)ncl_fsinfo(nmp, vp, cred, td);
301 	} else
302 		mtx_unlock(&nmp->nm_mtx);
303 
304 	mtx_lock(&np->n_mtx);
305 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
306 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
307 		mtx_unlock(&np->n_mtx);
308 		printf("ncl_putpages: called on noncache-able vnode\n");
309 		mtx_lock(&np->n_mtx);
310 	}
311 	/*
312 	 * When putting pages, do not extend file past EOF.
313 	 */
314 	if (offset + count > np->n_size) {
315 		count = np->n_size - offset;
316 		if (count < 0)
317 			count = 0;
318 	}
319 	mtx_unlock(&np->n_mtx);
320 
321 	for (i = 0; i < npages; i++)
322 		rtvals[i] = VM_PAGER_ERROR;
323 
324 	VM_CNT_INC(v_vnodeout);
325 	VM_CNT_ADD(v_vnodepgsout, count);
326 
327 	iov.iov_base = unmapped_buf;
328 	iov.iov_len = count;
329 	uio.uio_iov = &iov;
330 	uio.uio_iovcnt = 1;
331 	uio.uio_offset = offset;
332 	uio.uio_resid = count;
333 	uio.uio_segflg = UIO_NOCOPY;
334 	uio.uio_rw = UIO_WRITE;
335 	uio.uio_td = td;
336 
337 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
338 	    cred);
339 	crfree(cred);
340 
341 	if (error == 0 || !nfs_keep_dirty_on_error) {
342 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
343 		    np->n_size - offset, npages * PAGE_SIZE);
344 	}
345 	return (rtvals[0]);
346 }
347 
348 /*
349  * For nfs, cache consistency can only be maintained approximately.
350  * Although RFC1094 does not specify the criteria, the following is
351  * believed to be compatible with the reference port.
352  * For nfs:
353  * If the file's modify time on the server has changed since the
354  * last read rpc or you have written to the file,
355  * you may have lost data cache consistency with the
356  * server, so flush all of the file's data out of the cache.
357  * Then force a getattr rpc to ensure that you have up to date
358  * attributes.
359  * NB: This implies that cache data can be read when up to
360  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
361  * attributes this could be forced by setting n_attrstamp to 0 before
362  * the VOP_GETATTR() call.
363  */
364 static inline int
365 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
366 {
367 	int error = 0;
368 	struct vattr vattr;
369 	struct nfsnode *np = VTONFS(vp);
370 	bool old_lock;
371 
372 	/*
373 	 * Ensure the exclusove access to the node before checking
374 	 * whether the cache is consistent.
375 	 */
376 	old_lock = ncl_excl_start(vp);
377 	mtx_lock(&np->n_mtx);
378 	if (np->n_flag & NMODIFIED) {
379 		mtx_unlock(&np->n_mtx);
380 		if (vp->v_type != VREG) {
381 			if (vp->v_type != VDIR)
382 				panic("nfs: bioread, not dir");
383 			ncl_invaldir(vp);
384 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
385 			if (error != 0)
386 				goto out;
387 		}
388 		np->n_attrstamp = 0;
389 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
390 		error = VOP_GETATTR(vp, &vattr, cred);
391 		if (error)
392 			goto out;
393 		mtx_lock(&np->n_mtx);
394 		np->n_mtime = vattr.va_mtime;
395 		mtx_unlock(&np->n_mtx);
396 	} else {
397 		mtx_unlock(&np->n_mtx);
398 		error = VOP_GETATTR(vp, &vattr, cred);
399 		if (error)
400 			goto out;
401 		mtx_lock(&np->n_mtx);
402 		if ((np->n_flag & NSIZECHANGED)
403 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
404 			mtx_unlock(&np->n_mtx);
405 			if (vp->v_type == VDIR)
406 				ncl_invaldir(vp);
407 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
408 			if (error != 0)
409 				goto out;
410 			mtx_lock(&np->n_mtx);
411 			np->n_mtime = vattr.va_mtime;
412 			np->n_flag &= ~NSIZECHANGED;
413 		}
414 		mtx_unlock(&np->n_mtx);
415 	}
416 out:
417 	ncl_excl_finish(vp, old_lock);
418 	return (error);
419 }
420 
421 /*
422  * Vnode op for read using bio
423  */
424 int
425 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
426 {
427 	struct nfsnode *np = VTONFS(vp);
428 	int biosize, i;
429 	struct buf *bp, *rabp;
430 	struct thread *td;
431 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
432 	daddr_t lbn, rabn;
433 	int bcount;
434 	int seqcount;
435 	int nra, error = 0, n = 0, on = 0;
436 	off_t tmp_off;
437 
438 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
439 	if (uio->uio_resid == 0)
440 		return (0);
441 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
442 		return (EINVAL);
443 	td = uio->uio_td;
444 
445 	mtx_lock(&nmp->nm_mtx);
446 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
447 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
448 		mtx_unlock(&nmp->nm_mtx);
449 		(void)ncl_fsinfo(nmp, vp, cred, td);
450 		mtx_lock(&nmp->nm_mtx);
451 	}
452 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
453 		(void) newnfs_iosize(nmp);
454 
455 	tmp_off = uio->uio_offset + uio->uio_resid;
456 	if (vp->v_type != VDIR &&
457 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
458 		mtx_unlock(&nmp->nm_mtx);
459 		return (EFBIG);
460 	}
461 	mtx_unlock(&nmp->nm_mtx);
462 
463 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
464 		/* No caching/ no readaheads. Just read data into the user buffer */
465 		return ncl_readrpc(vp, uio, cred);
466 
467 	biosize = vp->v_bufobj.bo_bsize;
468 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
469 
470 	error = nfs_bioread_check_cons(vp, td, cred);
471 	if (error)
472 		return error;
473 
474 	do {
475 	    u_quad_t nsize;
476 
477 	    mtx_lock(&np->n_mtx);
478 	    nsize = np->n_size;
479 	    mtx_unlock(&np->n_mtx);
480 
481 	    switch (vp->v_type) {
482 	    case VREG:
483 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
484 		lbn = uio->uio_offset / biosize;
485 		on = uio->uio_offset - (lbn * biosize);
486 
487 		/*
488 		 * Start the read ahead(s), as required.
489 		 */
490 		if (nmp->nm_readahead > 0) {
491 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
492 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
493 			rabn = lbn + 1 + nra;
494 			if (incore(&vp->v_bufobj, rabn) == NULL) {
495 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
496 			    if (!rabp) {
497 				error = newnfs_sigintr(nmp, td);
498 				return (error ? error : EINTR);
499 			    }
500 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
501 				rabp->b_flags |= B_ASYNC;
502 				rabp->b_iocmd = BIO_READ;
503 				vfs_busy_pages(rabp, 0);
504 				if (ncl_asyncio(nmp, rabp, cred, td)) {
505 				    rabp->b_flags |= B_INVAL;
506 				    rabp->b_ioflags |= BIO_ERROR;
507 				    vfs_unbusy_pages(rabp);
508 				    brelse(rabp);
509 				    break;
510 				}
511 			    } else {
512 				brelse(rabp);
513 			    }
514 			}
515 		    }
516 		}
517 
518 		/* Note that bcount is *not* DEV_BSIZE aligned. */
519 		bcount = biosize;
520 		if ((off_t)lbn * biosize >= nsize) {
521 			bcount = 0;
522 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
523 			bcount = nsize - (off_t)lbn * biosize;
524 		}
525 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
526 
527 		if (!bp) {
528 			error = newnfs_sigintr(nmp, td);
529 			return (error ? error : EINTR);
530 		}
531 
532 		/*
533 		 * If B_CACHE is not set, we must issue the read.  If this
534 		 * fails, we return an error.
535 		 */
536 
537 		if ((bp->b_flags & B_CACHE) == 0) {
538 		    bp->b_iocmd = BIO_READ;
539 		    vfs_busy_pages(bp, 0);
540 		    error = ncl_doio(vp, bp, cred, td, 0);
541 		    if (error) {
542 			brelse(bp);
543 			return (error);
544 		    }
545 		}
546 
547 		/*
548 		 * on is the offset into the current bp.  Figure out how many
549 		 * bytes we can copy out of the bp.  Note that bcount is
550 		 * NOT DEV_BSIZE aligned.
551 		 *
552 		 * Then figure out how many bytes we can copy into the uio.
553 		 */
554 
555 		n = 0;
556 		if (on < bcount)
557 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
558 		break;
559 	    case VLNK:
560 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
561 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
562 		if (!bp) {
563 			error = newnfs_sigintr(nmp, td);
564 			return (error ? error : EINTR);
565 		}
566 		if ((bp->b_flags & B_CACHE) == 0) {
567 		    bp->b_iocmd = BIO_READ;
568 		    vfs_busy_pages(bp, 0);
569 		    error = ncl_doio(vp, bp, cred, td, 0);
570 		    if (error) {
571 			bp->b_ioflags |= BIO_ERROR;
572 			brelse(bp);
573 			return (error);
574 		    }
575 		}
576 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
577 		on = 0;
578 		break;
579 	    case VDIR:
580 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
581 		if (np->n_direofoffset
582 		    && uio->uio_offset >= np->n_direofoffset) {
583 		    return (0);
584 		}
585 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
586 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
587 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
588 		if (!bp) {
589 		    error = newnfs_sigintr(nmp, td);
590 		    return (error ? error : EINTR);
591 		}
592 		if ((bp->b_flags & B_CACHE) == 0) {
593 		    bp->b_iocmd = BIO_READ;
594 		    vfs_busy_pages(bp, 0);
595 		    error = ncl_doio(vp, bp, cred, td, 0);
596 		    if (error) {
597 			    brelse(bp);
598 		    }
599 		    while (error == NFSERR_BAD_COOKIE) {
600 			ncl_invaldir(vp);
601 			error = ncl_vinvalbuf(vp, 0, td, 1);
602 
603 			/*
604 			 * Yuck! The directory has been modified on the
605 			 * server. The only way to get the block is by
606 			 * reading from the beginning to get all the
607 			 * offset cookies.
608 			 *
609 			 * Leave the last bp intact unless there is an error.
610 			 * Loop back up to the while if the error is another
611 			 * NFSERR_BAD_COOKIE (double yuch!).
612 			 */
613 			for (i = 0; i <= lbn && !error; i++) {
614 			    if (np->n_direofoffset
615 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
616 				    return (0);
617 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
618 			    if (!bp) {
619 				error = newnfs_sigintr(nmp, td);
620 				return (error ? error : EINTR);
621 			    }
622 			    if ((bp->b_flags & B_CACHE) == 0) {
623 				    bp->b_iocmd = BIO_READ;
624 				    vfs_busy_pages(bp, 0);
625 				    error = ncl_doio(vp, bp, cred, td, 0);
626 				    /*
627 				     * no error + B_INVAL == directory EOF,
628 				     * use the block.
629 				     */
630 				    if (error == 0 && (bp->b_flags & B_INVAL))
631 					    break;
632 			    }
633 			    /*
634 			     * An error will throw away the block and the
635 			     * for loop will break out.  If no error and this
636 			     * is not the block we want, we throw away the
637 			     * block and go for the next one via the for loop.
638 			     */
639 			    if (error || i < lbn)
640 				    brelse(bp);
641 			}
642 		    }
643 		    /*
644 		     * The above while is repeated if we hit another cookie
645 		     * error.  If we hit an error and it wasn't a cookie error,
646 		     * we give up.
647 		     */
648 		    if (error)
649 			    return (error);
650 		}
651 
652 		/*
653 		 * If not eof and read aheads are enabled, start one.
654 		 * (You need the current block first, so that you have the
655 		 *  directory offset cookie of the next block.)
656 		 */
657 		if (nmp->nm_readahead > 0 &&
658 		    (bp->b_flags & B_INVAL) == 0 &&
659 		    (np->n_direofoffset == 0 ||
660 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
661 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
662 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
663 			if (rabp) {
664 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
665 				rabp->b_flags |= B_ASYNC;
666 				rabp->b_iocmd = BIO_READ;
667 				vfs_busy_pages(rabp, 0);
668 				if (ncl_asyncio(nmp, rabp, cred, td)) {
669 				    rabp->b_flags |= B_INVAL;
670 				    rabp->b_ioflags |= BIO_ERROR;
671 				    vfs_unbusy_pages(rabp);
672 				    brelse(rabp);
673 				}
674 			    } else {
675 				brelse(rabp);
676 			    }
677 			}
678 		}
679 		/*
680 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
681 		 * chopped for the EOF condition, we cannot tell how large
682 		 * NFS directories are going to be until we hit EOF.  So
683 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
684 		 * it just so happens that b_resid will effectively chop it
685 		 * to EOF.  *BUT* this information is lost if the buffer goes
686 		 * away and is reconstituted into a B_CACHE state ( due to
687 		 * being VMIO ) later.  So we keep track of the directory eof
688 		 * in np->n_direofoffset and chop it off as an extra step
689 		 * right here.
690 		 */
691 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
692 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
693 			n = np->n_direofoffset - uio->uio_offset;
694 		break;
695 	    default:
696 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
697 		bp = NULL;
698 		break;
699 	    }
700 
701 	    if (n > 0) {
702 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
703 	    }
704 	    if (vp->v_type == VLNK)
705 		n = 0;
706 	    if (bp != NULL)
707 		brelse(bp);
708 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
709 	return (error);
710 }
711 
712 /*
713  * The NFS write path cannot handle iovecs with len > 1. So we need to
714  * break up iovecs accordingly (restricting them to wsize).
715  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
716  * For the ASYNC case, 2 copies are needed. The first a copy from the
717  * user buffer to a staging buffer and then a second copy from the staging
718  * buffer to mbufs. This can be optimized by copying from the user buffer
719  * directly into mbufs and passing the chain down, but that requires a
720  * fair amount of re-working of the relevant codepaths (and can be done
721  * later).
722  */
723 static int
724 nfs_directio_write(vp, uiop, cred, ioflag)
725 	struct vnode *vp;
726 	struct uio *uiop;
727 	struct ucred *cred;
728 	int ioflag;
729 {
730 	int error;
731 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
732 	struct thread *td = uiop->uio_td;
733 	int size;
734 	int wsize;
735 
736 	mtx_lock(&nmp->nm_mtx);
737 	wsize = nmp->nm_wsize;
738 	mtx_unlock(&nmp->nm_mtx);
739 	if (ioflag & IO_SYNC) {
740 		int iomode, must_commit;
741 		struct uio uio;
742 		struct iovec iov;
743 do_sync:
744 		while (uiop->uio_resid > 0) {
745 			size = MIN(uiop->uio_resid, wsize);
746 			size = MIN(uiop->uio_iov->iov_len, size);
747 			iov.iov_base = uiop->uio_iov->iov_base;
748 			iov.iov_len = size;
749 			uio.uio_iov = &iov;
750 			uio.uio_iovcnt = 1;
751 			uio.uio_offset = uiop->uio_offset;
752 			uio.uio_resid = size;
753 			uio.uio_segflg = UIO_USERSPACE;
754 			uio.uio_rw = UIO_WRITE;
755 			uio.uio_td = td;
756 			iomode = NFSWRITE_FILESYNC;
757 			error = ncl_writerpc(vp, &uio, cred, &iomode,
758 			    &must_commit, 0);
759 			KASSERT((must_commit == 0),
760 				("ncl_directio_write: Did not commit write"));
761 			if (error)
762 				return (error);
763 			uiop->uio_offset += size;
764 			uiop->uio_resid -= size;
765 			if (uiop->uio_iov->iov_len <= size) {
766 				uiop->uio_iovcnt--;
767 				uiop->uio_iov++;
768 			} else {
769 				uiop->uio_iov->iov_base =
770 					(char *)uiop->uio_iov->iov_base + size;
771 				uiop->uio_iov->iov_len -= size;
772 			}
773 		}
774 	} else {
775 		struct uio *t_uio;
776 		struct iovec *t_iov;
777 		struct buf *bp;
778 
779 		/*
780 		 * Break up the write into blocksize chunks and hand these
781 		 * over to nfsiod's for write back.
782 		 * Unfortunately, this incurs a copy of the data. Since
783 		 * the user could modify the buffer before the write is
784 		 * initiated.
785 		 *
786 		 * The obvious optimization here is that one of the 2 copies
787 		 * in the async write path can be eliminated by copying the
788 		 * data here directly into mbufs and passing the mbuf chain
789 		 * down. But that will require a fair amount of re-working
790 		 * of the code and can be done if there's enough interest
791 		 * in NFS directio access.
792 		 */
793 		while (uiop->uio_resid > 0) {
794 			size = MIN(uiop->uio_resid, wsize);
795 			size = MIN(uiop->uio_iov->iov_len, size);
796 			bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
797 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
798 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
799 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
800 			t_iov->iov_len = size;
801 			t_uio->uio_iov = t_iov;
802 			t_uio->uio_iovcnt = 1;
803 			t_uio->uio_offset = uiop->uio_offset;
804 			t_uio->uio_resid = size;
805 			t_uio->uio_segflg = UIO_SYSSPACE;
806 			t_uio->uio_rw = UIO_WRITE;
807 			t_uio->uio_td = td;
808 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
809 			    uiop->uio_segflg == UIO_SYSSPACE,
810 			    ("nfs_directio_write: Bad uio_segflg"));
811 			if (uiop->uio_segflg == UIO_USERSPACE) {
812 				error = copyin(uiop->uio_iov->iov_base,
813 				    t_iov->iov_base, size);
814 				if (error != 0)
815 					goto err_free;
816 			} else
817 				/*
818 				 * UIO_SYSSPACE may never happen, but handle
819 				 * it just in case it does.
820 				 */
821 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
822 				    size);
823 			bp->b_flags |= B_DIRECT;
824 			bp->b_iocmd = BIO_WRITE;
825 			if (cred != NOCRED) {
826 				crhold(cred);
827 				bp->b_wcred = cred;
828 			} else
829 				bp->b_wcred = NOCRED;
830 			bp->b_caller1 = (void *)t_uio;
831 			bp->b_vp = vp;
832 			error = ncl_asyncio(nmp, bp, NOCRED, td);
833 err_free:
834 			if (error) {
835 				free(t_iov->iov_base, M_NFSDIRECTIO);
836 				free(t_iov, M_NFSDIRECTIO);
837 				free(t_uio, M_NFSDIRECTIO);
838 				bp->b_vp = NULL;
839 				uma_zfree(ncl_pbuf_zone, bp);
840 				if (error == EINTR)
841 					return (error);
842 				goto do_sync;
843 			}
844 			uiop->uio_offset += size;
845 			uiop->uio_resid -= size;
846 			if (uiop->uio_iov->iov_len <= size) {
847 				uiop->uio_iovcnt--;
848 				uiop->uio_iov++;
849 			} else {
850 				uiop->uio_iov->iov_base =
851 					(char *)uiop->uio_iov->iov_base + size;
852 				uiop->uio_iov->iov_len -= size;
853 			}
854 		}
855 	}
856 	return (0);
857 }
858 
859 /*
860  * Vnode op for write using bio
861  */
862 int
863 ncl_write(struct vop_write_args *ap)
864 {
865 	int biosize;
866 	struct uio *uio = ap->a_uio;
867 	struct thread *td = uio->uio_td;
868 	struct vnode *vp = ap->a_vp;
869 	struct nfsnode *np = VTONFS(vp);
870 	struct ucred *cred = ap->a_cred;
871 	int ioflag = ap->a_ioflag;
872 	struct buf *bp;
873 	struct vattr vattr;
874 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
875 	daddr_t lbn;
876 	int bcount, noncontig_write, obcount;
877 	int bp_cached, n, on, error = 0, error1, wouldcommit;
878 	size_t orig_resid, local_resid;
879 	off_t orig_size, tmp_off;
880 
881 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
882 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
883 	    ("ncl_write proc"));
884 	if (vp->v_type != VREG)
885 		return (EIO);
886 	mtx_lock(&np->n_mtx);
887 	if (np->n_flag & NWRITEERR) {
888 		np->n_flag &= ~NWRITEERR;
889 		mtx_unlock(&np->n_mtx);
890 		return (np->n_error);
891 	} else
892 		mtx_unlock(&np->n_mtx);
893 	mtx_lock(&nmp->nm_mtx);
894 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
895 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
896 		mtx_unlock(&nmp->nm_mtx);
897 		(void)ncl_fsinfo(nmp, vp, cred, td);
898 		mtx_lock(&nmp->nm_mtx);
899 	}
900 	if (nmp->nm_wsize == 0)
901 		(void) newnfs_iosize(nmp);
902 	mtx_unlock(&nmp->nm_mtx);
903 
904 	/*
905 	 * Synchronously flush pending buffers if we are in synchronous
906 	 * mode or if we are appending.
907 	 */
908 	if (ioflag & (IO_APPEND | IO_SYNC)) {
909 		mtx_lock(&np->n_mtx);
910 		if (np->n_flag & NMODIFIED) {
911 			mtx_unlock(&np->n_mtx);
912 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
913 			/*
914 			 * Require non-blocking, synchronous writes to
915 			 * dirty files to inform the program it needs
916 			 * to fsync(2) explicitly.
917 			 */
918 			if (ioflag & IO_NDELAY)
919 				return (EAGAIN);
920 #endif
921 			np->n_attrstamp = 0;
922 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
923 			error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
924 			    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
925 			if (error != 0)
926 				return (error);
927 		} else
928 			mtx_unlock(&np->n_mtx);
929 	}
930 
931 	orig_resid = uio->uio_resid;
932 	mtx_lock(&np->n_mtx);
933 	orig_size = np->n_size;
934 	mtx_unlock(&np->n_mtx);
935 
936 	/*
937 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
938 	 * get the append lock.
939 	 */
940 	if (ioflag & IO_APPEND) {
941 		np->n_attrstamp = 0;
942 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
943 		error = VOP_GETATTR(vp, &vattr, cred);
944 		if (error)
945 			return (error);
946 		mtx_lock(&np->n_mtx);
947 		uio->uio_offset = np->n_size;
948 		mtx_unlock(&np->n_mtx);
949 	}
950 
951 	if (uio->uio_offset < 0)
952 		return (EINVAL);
953 	tmp_off = uio->uio_offset + uio->uio_resid;
954 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
955 		return (EFBIG);
956 	if (uio->uio_resid == 0)
957 		return (0);
958 
959 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
960 		return nfs_directio_write(vp, uio, cred, ioflag);
961 
962 	/*
963 	 * Maybe this should be above the vnode op call, but so long as
964 	 * file servers have no limits, i don't think it matters
965 	 */
966 	if (vn_rlimit_fsize(vp, uio, td))
967 		return (EFBIG);
968 
969 	biosize = vp->v_bufobj.bo_bsize;
970 	/*
971 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
972 	 * would exceed the local maximum per-file write commit size when
973 	 * combined with those, we must decide whether to flush,
974 	 * go synchronous, or return error.  We don't bother checking
975 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
976 	 * no point optimizing for something that really won't ever happen.
977 	 */
978 	wouldcommit = 0;
979 	if (!(ioflag & IO_SYNC)) {
980 		int nflag;
981 
982 		mtx_lock(&np->n_mtx);
983 		nflag = np->n_flag;
984 		mtx_unlock(&np->n_mtx);
985 		if (nflag & NMODIFIED) {
986 			BO_LOCK(&vp->v_bufobj);
987 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
988 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
989 				    b_bobufs) {
990 					if (bp->b_flags & B_NEEDCOMMIT)
991 						wouldcommit += bp->b_bcount;
992 				}
993 			}
994 			BO_UNLOCK(&vp->v_bufobj);
995 		}
996 	}
997 
998 	do {
999 		if (!(ioflag & IO_SYNC)) {
1000 			wouldcommit += biosize;
1001 			if (wouldcommit > nmp->nm_wcommitsize) {
1002 				np->n_attrstamp = 0;
1003 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1004 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1005 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1006 				if (error != 0)
1007 					return (error);
1008 				wouldcommit = biosize;
1009 			}
1010 		}
1011 
1012 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1013 		lbn = uio->uio_offset / biosize;
1014 		on = uio->uio_offset - (lbn * biosize);
1015 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1016 again:
1017 		/*
1018 		 * Handle direct append and file extension cases, calculate
1019 		 * unaligned buffer size.
1020 		 */
1021 		mtx_lock(&np->n_mtx);
1022 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1023 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1024 			noncontig_write = 1;
1025 		else
1026 			noncontig_write = 0;
1027 		if ((uio->uio_offset == np->n_size ||
1028 		    (noncontig_write != 0 &&
1029 		    lbn == (np->n_size / biosize) &&
1030 		    uio->uio_offset + n > np->n_size)) && n) {
1031 			mtx_unlock(&np->n_mtx);
1032 			/*
1033 			 * Get the buffer (in its pre-append state to maintain
1034 			 * B_CACHE if it was previously set).  Resize the
1035 			 * nfsnode after we have locked the buffer to prevent
1036 			 * readers from reading garbage.
1037 			 */
1038 			obcount = np->n_size - (lbn * biosize);
1039 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1040 
1041 			if (bp != NULL) {
1042 				long save;
1043 
1044 				mtx_lock(&np->n_mtx);
1045 				np->n_size = uio->uio_offset + n;
1046 				np->n_flag |= NMODIFIED;
1047 				vnode_pager_setsize(vp, np->n_size);
1048 				mtx_unlock(&np->n_mtx);
1049 
1050 				save = bp->b_flags & B_CACHE;
1051 				bcount = on + n;
1052 				allocbuf(bp, bcount);
1053 				bp->b_flags |= save;
1054 				if (noncontig_write != 0 && on > obcount)
1055 					vfs_bio_bzero_buf(bp, obcount, on -
1056 					    obcount);
1057 			}
1058 		} else {
1059 			/*
1060 			 * Obtain the locked cache block first, and then
1061 			 * adjust the file's size as appropriate.
1062 			 */
1063 			bcount = on + n;
1064 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1065 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1066 					bcount = biosize;
1067 				else
1068 					bcount = np->n_size - (off_t)lbn * biosize;
1069 			}
1070 			mtx_unlock(&np->n_mtx);
1071 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1072 			mtx_lock(&np->n_mtx);
1073 			if (uio->uio_offset + n > np->n_size) {
1074 				np->n_size = uio->uio_offset + n;
1075 				np->n_flag |= NMODIFIED;
1076 				vnode_pager_setsize(vp, np->n_size);
1077 			}
1078 			mtx_unlock(&np->n_mtx);
1079 		}
1080 
1081 		if (!bp) {
1082 			error = newnfs_sigintr(nmp, td);
1083 			if (!error)
1084 				error = EINTR;
1085 			break;
1086 		}
1087 
1088 		/*
1089 		 * Issue a READ if B_CACHE is not set.  In special-append
1090 		 * mode, B_CACHE is based on the buffer prior to the write
1091 		 * op and is typically set, avoiding the read.  If a read
1092 		 * is required in special append mode, the server will
1093 		 * probably send us a short-read since we extended the file
1094 		 * on our end, resulting in b_resid == 0 and, thusly,
1095 		 * B_CACHE getting set.
1096 		 *
1097 		 * We can also avoid issuing the read if the write covers
1098 		 * the entire buffer.  We have to make sure the buffer state
1099 		 * is reasonable in this case since we will not be initiating
1100 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1101 		 * more information.
1102 		 *
1103 		 * B_CACHE may also be set due to the buffer being cached
1104 		 * normally.
1105 		 */
1106 
1107 		bp_cached = 1;
1108 		if (on == 0 && n == bcount) {
1109 			if ((bp->b_flags & B_CACHE) == 0)
1110 				bp_cached = 0;
1111 			bp->b_flags |= B_CACHE;
1112 			bp->b_flags &= ~B_INVAL;
1113 			bp->b_ioflags &= ~BIO_ERROR;
1114 		}
1115 
1116 		if ((bp->b_flags & B_CACHE) == 0) {
1117 			bp->b_iocmd = BIO_READ;
1118 			vfs_busy_pages(bp, 0);
1119 			error = ncl_doio(vp, bp, cred, td, 0);
1120 			if (error) {
1121 				brelse(bp);
1122 				break;
1123 			}
1124 		}
1125 		if (bp->b_wcred == NOCRED)
1126 			bp->b_wcred = crhold(cred);
1127 		mtx_lock(&np->n_mtx);
1128 		np->n_flag |= NMODIFIED;
1129 		mtx_unlock(&np->n_mtx);
1130 
1131 		/*
1132 		 * If dirtyend exceeds file size, chop it down.  This should
1133 		 * not normally occur but there is an append race where it
1134 		 * might occur XXX, so we log it.
1135 		 *
1136 		 * If the chopping creates a reverse-indexed or degenerate
1137 		 * situation with dirtyoff/end, we 0 both of them.
1138 		 */
1139 
1140 		if (bp->b_dirtyend > bcount) {
1141 			printf("NFS append race @%lx:%d\n",
1142 			    (long)bp->b_blkno * DEV_BSIZE,
1143 			    bp->b_dirtyend - bcount);
1144 			bp->b_dirtyend = bcount;
1145 		}
1146 
1147 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1148 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1149 
1150 		/*
1151 		 * If the new write will leave a contiguous dirty
1152 		 * area, just update the b_dirtyoff and b_dirtyend,
1153 		 * otherwise force a write rpc of the old dirty area.
1154 		 *
1155 		 * If there has been a file lock applied to this file
1156 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1157 		 * While it is possible to merge discontiguous writes due to
1158 		 * our having a B_CACHE buffer ( and thus valid read data
1159 		 * for the hole), we don't because it could lead to
1160 		 * significant cache coherency problems with multiple clients,
1161 		 * especially if locking is implemented later on.
1162 		 *
1163 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1164 		 * not been file locking done on this file:
1165 		 * Relax coherency a bit for the sake of performance and
1166 		 * expand the current dirty region to contain the new
1167 		 * write even if it means we mark some non-dirty data as
1168 		 * dirty.
1169 		 */
1170 
1171 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1172 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1173 			if (bwrite(bp) == EINTR) {
1174 				error = EINTR;
1175 				break;
1176 			}
1177 			goto again;
1178 		}
1179 
1180 		local_resid = uio->uio_resid;
1181 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1182 
1183 		if (error != 0 && !bp_cached) {
1184 			/*
1185 			 * This block has no other content then what
1186 			 * possibly was written by the faulty uiomove.
1187 			 * Release it, forgetting the data pages, to
1188 			 * prevent the leak of uninitialized data to
1189 			 * usermode.
1190 			 */
1191 			bp->b_ioflags |= BIO_ERROR;
1192 			brelse(bp);
1193 			uio->uio_offset -= local_resid - uio->uio_resid;
1194 			uio->uio_resid = local_resid;
1195 			break;
1196 		}
1197 
1198 		/*
1199 		 * Since this block is being modified, it must be written
1200 		 * again and not just committed.  Since write clustering does
1201 		 * not work for the stage 1 data write, only the stage 2
1202 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1203 		 */
1204 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1205 
1206 		/*
1207 		 * Get the partial update on the progress made from
1208 		 * uiomove, if an error occurred.
1209 		 */
1210 		if (error != 0)
1211 			n = local_resid - uio->uio_resid;
1212 
1213 		/*
1214 		 * Only update dirtyoff/dirtyend if not a degenerate
1215 		 * condition.
1216 		 */
1217 		if (n > 0) {
1218 			if (bp->b_dirtyend > 0) {
1219 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1220 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1221 			} else {
1222 				bp->b_dirtyoff = on;
1223 				bp->b_dirtyend = on + n;
1224 			}
1225 			vfs_bio_set_valid(bp, on, n);
1226 		}
1227 
1228 		/*
1229 		 * If IO_SYNC do bwrite().
1230 		 *
1231 		 * IO_INVAL appears to be unused.  The idea appears to be
1232 		 * to turn off caching in this case.  Very odd.  XXX
1233 		 */
1234 		if ((ioflag & IO_SYNC)) {
1235 			if (ioflag & IO_INVAL)
1236 				bp->b_flags |= B_NOCACHE;
1237 			error1 = bwrite(bp);
1238 			if (error1 != 0) {
1239 				if (error == 0)
1240 					error = error1;
1241 				break;
1242 			}
1243 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1244 			bp->b_flags |= B_ASYNC;
1245 			(void) ncl_writebp(bp, 0, NULL);
1246 		} else {
1247 			bdwrite(bp);
1248 		}
1249 
1250 		if (error != 0)
1251 			break;
1252 	} while (uio->uio_resid > 0 && n > 0);
1253 
1254 	if (error != 0) {
1255 		if (ioflag & IO_UNIT) {
1256 			VATTR_NULL(&vattr);
1257 			vattr.va_size = orig_size;
1258 			/* IO_SYNC is handled implicitely */
1259 			(void)VOP_SETATTR(vp, &vattr, cred);
1260 			uio->uio_offset -= orig_resid - uio->uio_resid;
1261 			uio->uio_resid = orig_resid;
1262 		}
1263 	}
1264 
1265 	return (error);
1266 }
1267 
1268 /*
1269  * Get an nfs cache block.
1270  *
1271  * Allocate a new one if the block isn't currently in the cache
1272  * and return the block marked busy. If the calling process is
1273  * interrupted by a signal for an interruptible mount point, return
1274  * NULL.
1275  *
1276  * The caller must carefully deal with the possible B_INVAL state of
1277  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1278  * indirectly), so synchronous reads can be issued without worrying about
1279  * the B_INVAL state.  We have to be a little more careful when dealing
1280  * with writes (see comments in nfs_write()) when extending a file past
1281  * its EOF.
1282  */
1283 static struct buf *
1284 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1285 {
1286 	struct buf *bp;
1287 	struct mount *mp;
1288 	struct nfsmount *nmp;
1289 
1290 	mp = vp->v_mount;
1291 	nmp = VFSTONFS(mp);
1292 
1293 	if (nmp->nm_flag & NFSMNT_INT) {
1294 		sigset_t oldset;
1295 
1296 		newnfs_set_sigmask(td, &oldset);
1297 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1298 		newnfs_restore_sigmask(td, &oldset);
1299 		while (bp == NULL) {
1300 			if (newnfs_sigintr(nmp, td))
1301 				return (NULL);
1302 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1303 		}
1304 	} else {
1305 		bp = getblk(vp, bn, size, 0, 0, 0);
1306 	}
1307 
1308 	if (vp->v_type == VREG)
1309 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1310 	return (bp);
1311 }
1312 
1313 /*
1314  * Flush and invalidate all dirty buffers. If another process is already
1315  * doing the flush, just wait for completion.
1316  */
1317 int
1318 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1319 {
1320 	struct nfsnode *np = VTONFS(vp);
1321 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1322 	int error = 0, slpflag, slptimeo;
1323 	bool old_lock;
1324 
1325 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1326 
1327 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1328 		intrflg = 0;
1329 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1330 		intrflg = 1;
1331 	if (intrflg) {
1332 		slpflag = PCATCH;
1333 		slptimeo = 2 * hz;
1334 	} else {
1335 		slpflag = 0;
1336 		slptimeo = 0;
1337 	}
1338 
1339 	old_lock = ncl_excl_start(vp);
1340 	if (old_lock)
1341 		flags |= V_ALLOWCLEAN;
1342 
1343 	/*
1344 	 * Now, flush as required.
1345 	 */
1346 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1347 	     vp->v_bufobj.bo_object != NULL) {
1348 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1349 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1350 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1351 		/*
1352 		 * If the page clean was interrupted, fail the invalidation.
1353 		 * Not doing so, we run the risk of losing dirty pages in the
1354 		 * vinvalbuf() call below.
1355 		 */
1356 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1357 			goto out;
1358 	}
1359 
1360 	error = vinvalbuf(vp, flags, slpflag, 0);
1361 	while (error) {
1362 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1363 			goto out;
1364 		error = vinvalbuf(vp, flags, 0, slptimeo);
1365 	}
1366 	if (NFSHASPNFS(nmp)) {
1367 		nfscl_layoutcommit(vp, td);
1368 		/*
1369 		 * Invalidate the attribute cache, since writes to a DS
1370 		 * won't update the size attribute.
1371 		 */
1372 		mtx_lock(&np->n_mtx);
1373 		np->n_attrstamp = 0;
1374 	} else
1375 		mtx_lock(&np->n_mtx);
1376 	if (np->n_directio_asyncwr == 0)
1377 		np->n_flag &= ~NMODIFIED;
1378 	mtx_unlock(&np->n_mtx);
1379 out:
1380 	ncl_excl_finish(vp, old_lock);
1381 	return error;
1382 }
1383 
1384 /*
1385  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1386  * This is mainly to avoid queueing async I/O requests when the nfsiods
1387  * are all hung on a dead server.
1388  *
1389  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1390  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1391  */
1392 int
1393 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1394 {
1395 	int iod;
1396 	int gotiod;
1397 	int slpflag = 0;
1398 	int slptimeo = 0;
1399 	int error, error2;
1400 
1401 	/*
1402 	 * Commits are usually short and sweet so lets save some cpu and
1403 	 * leave the async daemons for more important rpc's (such as reads
1404 	 * and writes).
1405 	 *
1406 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1407 	 * in the directory in order to update attributes. This can deadlock
1408 	 * with another thread that is waiting for async I/O to be done by
1409 	 * an nfsiod thread while holding a lock on one of these vnodes.
1410 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1411 	 * perform Readdirplus RPCs.
1412 	 */
1413 	mtx_lock(&ncl_iod_mutex);
1414 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1415 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1416 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1417 		mtx_unlock(&ncl_iod_mutex);
1418 		return(EIO);
1419 	}
1420 again:
1421 	if (nmp->nm_flag & NFSMNT_INT)
1422 		slpflag = PCATCH;
1423 	gotiod = FALSE;
1424 
1425 	/*
1426 	 * Find a free iod to process this request.
1427 	 */
1428 	for (iod = 0; iod < ncl_numasync; iod++)
1429 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1430 			gotiod = TRUE;
1431 			break;
1432 		}
1433 
1434 	/*
1435 	 * Try to create one if none are free.
1436 	 */
1437 	if (!gotiod)
1438 		ncl_nfsiodnew();
1439 	else {
1440 		/*
1441 		 * Found one, so wake it up and tell it which
1442 		 * mount to process.
1443 		 */
1444 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1445 		    iod, nmp));
1446 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1447 		ncl_iodmount[iod] = nmp;
1448 		nmp->nm_bufqiods++;
1449 		wakeup(&ncl_iodwant[iod]);
1450 	}
1451 
1452 	/*
1453 	 * If none are free, we may already have an iod working on this mount
1454 	 * point.  If so, it will process our request.
1455 	 */
1456 	if (!gotiod) {
1457 		if (nmp->nm_bufqiods > 0) {
1458 			NFS_DPF(ASYNCIO,
1459 				("ncl_asyncio: %d iods are already processing mount %p\n",
1460 				 nmp->nm_bufqiods, nmp));
1461 			gotiod = TRUE;
1462 		}
1463 	}
1464 
1465 	/*
1466 	 * If we have an iod which can process the request, then queue
1467 	 * the buffer.
1468 	 */
1469 	if (gotiod) {
1470 		/*
1471 		 * Ensure that the queue never grows too large.  We still want
1472 		 * to asynchronize so we block rather then return EIO.
1473 		 */
1474 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1475 			NFS_DPF(ASYNCIO,
1476 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1477 			nmp->nm_bufqwant = TRUE;
1478 			error = newnfs_msleep(td, &nmp->nm_bufq,
1479 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1480 			   slptimeo);
1481 			if (error) {
1482 				error2 = newnfs_sigintr(nmp, td);
1483 				if (error2) {
1484 					mtx_unlock(&ncl_iod_mutex);
1485 					return (error2);
1486 				}
1487 				if (slpflag == PCATCH) {
1488 					slpflag = 0;
1489 					slptimeo = 2 * hz;
1490 				}
1491 			}
1492 			/*
1493 			 * We might have lost our iod while sleeping,
1494 			 * so check and loop if necessary.
1495 			 */
1496 			goto again;
1497 		}
1498 
1499 		/* We might have lost our nfsiod */
1500 		if (nmp->nm_bufqiods == 0) {
1501 			NFS_DPF(ASYNCIO,
1502 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1503 			goto again;
1504 		}
1505 
1506 		if (bp->b_iocmd == BIO_READ) {
1507 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1508 				bp->b_rcred = crhold(cred);
1509 		} else {
1510 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1511 				bp->b_wcred = crhold(cred);
1512 		}
1513 
1514 		if (bp->b_flags & B_REMFREE)
1515 			bremfreef(bp);
1516 		BUF_KERNPROC(bp);
1517 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1518 		nmp->nm_bufqlen++;
1519 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1520 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1521 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1522 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1523 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1524 		}
1525 		mtx_unlock(&ncl_iod_mutex);
1526 		return (0);
1527 	}
1528 
1529 	mtx_unlock(&ncl_iod_mutex);
1530 
1531 	/*
1532 	 * All the iods are busy on other mounts, so return EIO to
1533 	 * force the caller to process the i/o synchronously.
1534 	 */
1535 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1536 	return (EIO);
1537 }
1538 
1539 void
1540 ncl_doio_directwrite(struct buf *bp)
1541 {
1542 	int iomode, must_commit;
1543 	struct uio *uiop = (struct uio *)bp->b_caller1;
1544 	char *iov_base = uiop->uio_iov->iov_base;
1545 
1546 	iomode = NFSWRITE_FILESYNC;
1547 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1548 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1549 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1550 	free(iov_base, M_NFSDIRECTIO);
1551 	free(uiop->uio_iov, M_NFSDIRECTIO);
1552 	free(uiop, M_NFSDIRECTIO);
1553 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1554 		struct nfsnode *np = VTONFS(bp->b_vp);
1555 		mtx_lock(&np->n_mtx);
1556 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1557 			/*
1558 			 * Invalidate the attribute cache, since writes to a DS
1559 			 * won't update the size attribute.
1560 			 */
1561 			np->n_attrstamp = 0;
1562 		}
1563 		np->n_directio_asyncwr--;
1564 		if (np->n_directio_asyncwr == 0) {
1565 			np->n_flag &= ~NMODIFIED;
1566 			if ((np->n_flag & NFSYNCWAIT)) {
1567 				np->n_flag &= ~NFSYNCWAIT;
1568 				wakeup((caddr_t)&np->n_directio_asyncwr);
1569 			}
1570 		}
1571 		mtx_unlock(&np->n_mtx);
1572 	}
1573 	bp->b_vp = NULL;
1574 	uma_zfree(ncl_pbuf_zone, bp);
1575 }
1576 
1577 /*
1578  * Do an I/O operation to/from a cache block. This may be called
1579  * synchronously or from an nfsiod.
1580  */
1581 int
1582 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1583     int called_from_strategy)
1584 {
1585 	struct uio *uiop;
1586 	struct nfsnode *np;
1587 	struct nfsmount *nmp;
1588 	int error = 0, iomode, must_commit = 0;
1589 	struct uio uio;
1590 	struct iovec io;
1591 	struct proc *p = td ? td->td_proc : NULL;
1592 	uint8_t	iocmd;
1593 
1594 	np = VTONFS(vp);
1595 	nmp = VFSTONFS(vp->v_mount);
1596 	uiop = &uio;
1597 	uiop->uio_iov = &io;
1598 	uiop->uio_iovcnt = 1;
1599 	uiop->uio_segflg = UIO_SYSSPACE;
1600 	uiop->uio_td = td;
1601 
1602 	/*
1603 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1604 	 * do this here so we do not have to do it in all the code that
1605 	 * calls us.
1606 	 */
1607 	bp->b_flags &= ~B_INVAL;
1608 	bp->b_ioflags &= ~BIO_ERROR;
1609 
1610 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1611 	iocmd = bp->b_iocmd;
1612 	if (iocmd == BIO_READ) {
1613 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1614 	    io.iov_base = bp->b_data;
1615 	    uiop->uio_rw = UIO_READ;
1616 
1617 	    switch (vp->v_type) {
1618 	    case VREG:
1619 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1620 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1621 		error = ncl_readrpc(vp, uiop, cr);
1622 
1623 		if (!error) {
1624 		    if (uiop->uio_resid) {
1625 			/*
1626 			 * If we had a short read with no error, we must have
1627 			 * hit a file hole.  We should zero-fill the remainder.
1628 			 * This can also occur if the server hits the file EOF.
1629 			 *
1630 			 * Holes used to be able to occur due to pending
1631 			 * writes, but that is not possible any longer.
1632 			 */
1633 			int nread = bp->b_bcount - uiop->uio_resid;
1634 			ssize_t left = uiop->uio_resid;
1635 
1636 			if (left > 0)
1637 				bzero((char *)bp->b_data + nread, left);
1638 			uiop->uio_resid = 0;
1639 		    }
1640 		}
1641 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1642 		if (p && (vp->v_vflag & VV_TEXT)) {
1643 			mtx_lock(&np->n_mtx);
1644 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1645 				mtx_unlock(&np->n_mtx);
1646 				PROC_LOCK(p);
1647 				killproc(p, "text file modification");
1648 				PROC_UNLOCK(p);
1649 			} else
1650 				mtx_unlock(&np->n_mtx);
1651 		}
1652 		break;
1653 	    case VLNK:
1654 		uiop->uio_offset = (off_t)0;
1655 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1656 		error = ncl_readlinkrpc(vp, uiop, cr);
1657 		break;
1658 	    case VDIR:
1659 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1660 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1661 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1662 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1663 			if (error == NFSERR_NOTSUPP)
1664 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1665 		}
1666 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1667 			error = ncl_readdirrpc(vp, uiop, cr, td);
1668 		/*
1669 		 * end-of-directory sets B_INVAL but does not generate an
1670 		 * error.
1671 		 */
1672 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1673 			bp->b_flags |= B_INVAL;
1674 		break;
1675 	    default:
1676 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1677 		break;
1678 	    }
1679 	    if (error) {
1680 		bp->b_ioflags |= BIO_ERROR;
1681 		bp->b_error = error;
1682 	    }
1683 	} else {
1684 	    /*
1685 	     * If we only need to commit, try to commit
1686 	     */
1687 	    if (bp->b_flags & B_NEEDCOMMIT) {
1688 		    int retv;
1689 		    off_t off;
1690 
1691 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1692 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1693 			bp->b_wcred, td);
1694 		    if (retv == 0) {
1695 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1696 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1697 			    bp->b_resid = 0;
1698 			    bufdone(bp);
1699 			    return (0);
1700 		    }
1701 		    if (retv == NFSERR_STALEWRITEVERF) {
1702 			    ncl_clearcommit(vp->v_mount);
1703 		    }
1704 	    }
1705 
1706 	    /*
1707 	     * Setup for actual write
1708 	     */
1709 	    mtx_lock(&np->n_mtx);
1710 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1711 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1712 	    mtx_unlock(&np->n_mtx);
1713 
1714 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1715 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1716 		    - bp->b_dirtyoff;
1717 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1718 		    + bp->b_dirtyoff;
1719 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1720 		uiop->uio_rw = UIO_WRITE;
1721 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1722 
1723 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1724 		    iomode = NFSWRITE_UNSTABLE;
1725 		else
1726 		    iomode = NFSWRITE_FILESYNC;
1727 
1728 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1729 		    called_from_strategy);
1730 
1731 		/*
1732 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1733 		 * to cluster the buffers needing commit.  This will allow
1734 		 * the system to submit a single commit rpc for the whole
1735 		 * cluster.  We can do this even if the buffer is not 100%
1736 		 * dirty (relative to the NFS blocksize), so we optimize the
1737 		 * append-to-file-case.
1738 		 *
1739 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1740 		 * cleared because write clustering only works for commit
1741 		 * rpc's, not for the data portion of the write).
1742 		 */
1743 
1744 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1745 		    bp->b_flags |= B_NEEDCOMMIT;
1746 		    if (bp->b_dirtyoff == 0
1747 			&& bp->b_dirtyend == bp->b_bcount)
1748 			bp->b_flags |= B_CLUSTEROK;
1749 		} else {
1750 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1751 		}
1752 
1753 		/*
1754 		 * For an interrupted write, the buffer is still valid
1755 		 * and the write hasn't been pushed to the server yet,
1756 		 * so we can't set BIO_ERROR and report the interruption
1757 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1758 		 * is not relevant, so the rpc attempt is essentially
1759 		 * a noop.  For the case of a V3 write rpc not being
1760 		 * committed to stable storage, the block is still
1761 		 * dirty and requires either a commit rpc or another
1762 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1763 		 * the block is reused. This is indicated by setting
1764 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1765 		 *
1766 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1767 		 * write error and is handled as above, except that
1768 		 * B_EINTR isn't set. One cause of this is a stale stateid
1769 		 * error for the RPC that indicates recovery is required,
1770 		 * when called with called_from_strategy != 0.
1771 		 *
1772 		 * If the buffer is marked B_PAGING, it does not reside on
1773 		 * the vp's paging queues so we cannot call bdirty().  The
1774 		 * bp in this case is not an NFS cache block so we should
1775 		 * be safe. XXX
1776 		 *
1777 		 * The logic below breaks up errors into recoverable and
1778 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1779 		 * and keep the buffer around for potential write retries.
1780 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1781 		 * and save the error in the nfsnode. This is less than ideal
1782 		 * but necessary. Keeping such buffers around could potentially
1783 		 * cause buffer exhaustion eventually (they can never be written
1784 		 * out, so will get constantly be re-dirtied). It also causes
1785 		 * all sorts of vfs panics. For non-recoverable write errors,
1786 		 * also invalidate the attrcache, so we'll be forced to go over
1787 		 * the wire for this object, returning an error to user on next
1788 		 * call (most of the time).
1789 		 */
1790 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1791 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1792 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1793 			if ((bp->b_flags & B_PAGING) == 0) {
1794 			    bdirty(bp);
1795 			    bp->b_flags &= ~B_DONE;
1796 			}
1797 			if ((error == EINTR || error == ETIMEDOUT) &&
1798 			    (bp->b_flags & B_ASYNC) == 0)
1799 			    bp->b_flags |= B_EINTR;
1800 		} else {
1801 		    if (error) {
1802 			bp->b_ioflags |= BIO_ERROR;
1803 			bp->b_flags |= B_INVAL;
1804 			bp->b_error = np->n_error = error;
1805 			mtx_lock(&np->n_mtx);
1806 			np->n_flag |= NWRITEERR;
1807 			np->n_attrstamp = 0;
1808 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1809 			mtx_unlock(&np->n_mtx);
1810 		    }
1811 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1812 		}
1813 	    } else {
1814 		bp->b_resid = 0;
1815 		bufdone(bp);
1816 		return (0);
1817 	    }
1818 	}
1819 	bp->b_resid = uiop->uio_resid;
1820 	if (must_commit)
1821 	    ncl_clearcommit(vp->v_mount);
1822 	bufdone(bp);
1823 	return (error);
1824 }
1825 
1826 /*
1827  * Used to aid in handling ftruncate() operations on the NFS client side.
1828  * Truncation creates a number of special problems for NFS.  We have to
1829  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1830  * we have to properly handle VM pages or (potentially dirty) buffers
1831  * that straddle the truncation point.
1832  */
1833 
1834 int
1835 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1836 {
1837 	struct nfsnode *np = VTONFS(vp);
1838 	u_quad_t tsize;
1839 	int biosize = vp->v_bufobj.bo_bsize;
1840 	int error = 0;
1841 
1842 	mtx_lock(&np->n_mtx);
1843 	tsize = np->n_size;
1844 	np->n_size = nsize;
1845 	mtx_unlock(&np->n_mtx);
1846 
1847 	if (nsize < tsize) {
1848 		struct buf *bp;
1849 		daddr_t lbn;
1850 		int bufsize;
1851 
1852 		/*
1853 		 * vtruncbuf() doesn't get the buffer overlapping the
1854 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1855 		 * buffer that now needs to be truncated.
1856 		 */
1857 		error = vtruncbuf(vp, cred, nsize, biosize);
1858 		lbn = nsize / biosize;
1859 		bufsize = nsize - (lbn * biosize);
1860 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1861 		if (!bp)
1862 			return EINTR;
1863 		if (bp->b_dirtyoff > bp->b_bcount)
1864 			bp->b_dirtyoff = bp->b_bcount;
1865 		if (bp->b_dirtyend > bp->b_bcount)
1866 			bp->b_dirtyend = bp->b_bcount;
1867 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1868 		brelse(bp);
1869 	} else {
1870 		vnode_pager_setsize(vp, nsize);
1871 	}
1872 	return(error);
1873 }
1874 
1875