xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision f4b37ed0f8b307b1f3f0f630ca725d68f1dff30d)
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  * 4. 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 nfsstats newnfsstats;
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 int
82 ncl_getpages(struct vop_getpages_args *ap)
83 {
84 	int i, error, nextoff, size, toff, count, npages;
85 	struct uio uio;
86 	struct iovec iov;
87 	vm_offset_t kva;
88 	struct buf *bp;
89 	struct vnode *vp;
90 	struct thread *td;
91 	struct ucred *cred;
92 	struct nfsmount *nmp;
93 	vm_object_t object;
94 	vm_page_t *pages;
95 	struct nfsnode *np;
96 
97 	vp = ap->a_vp;
98 	np = VTONFS(vp);
99 	td = curthread;				/* XXX */
100 	cred = curthread->td_ucred;		/* XXX */
101 	nmp = VFSTONFS(vp->v_mount);
102 	pages = ap->a_m;
103 	count = ap->a_count;
104 
105 	if ((object = vp->v_object) == NULL) {
106 		ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
107 		return (VM_PAGER_ERROR);
108 	}
109 
110 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
111 		mtx_lock(&np->n_mtx);
112 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
113 			mtx_unlock(&np->n_mtx);
114 			ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
115 			return (VM_PAGER_ERROR);
116 		} else
117 			mtx_unlock(&np->n_mtx);
118 	}
119 
120 	mtx_lock(&nmp->nm_mtx);
121 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
122 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
123 		mtx_unlock(&nmp->nm_mtx);
124 		/* We'll never get here for v4, because we always have fsinfo */
125 		(void)ncl_fsinfo(nmp, vp, cred, td);
126 	} else
127 		mtx_unlock(&nmp->nm_mtx);
128 
129 	npages = btoc(count);
130 
131 	/*
132 	 * If the requested page is partially valid, just return it and
133 	 * allow the pager to zero-out the blanks.  Partially valid pages
134 	 * can only occur at the file EOF.
135 	 */
136 	if (pages[ap->a_reqpage]->valid != 0) {
137 		vm_pager_free_nonreq(object, pages, ap->a_reqpage, npages,
138 		    FALSE);
139 		return (VM_PAGER_OK);
140 	}
141 
142 	/*
143 	 * We use only the kva address for the buffer, but this is extremely
144 	 * convienient and fast.
145 	 */
146 	bp = getpbuf(&ncl_pbuf_freecnt);
147 
148 	kva = (vm_offset_t) bp->b_data;
149 	pmap_qenter(kva, pages, npages);
150 	PCPU_INC(cnt.v_vnodein);
151 	PCPU_ADD(cnt.v_vnodepgsin, npages);
152 
153 	iov.iov_base = (caddr_t) kva;
154 	iov.iov_len = count;
155 	uio.uio_iov = &iov;
156 	uio.uio_iovcnt = 1;
157 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
158 	uio.uio_resid = count;
159 	uio.uio_segflg = UIO_SYSSPACE;
160 	uio.uio_rw = UIO_READ;
161 	uio.uio_td = td;
162 
163 	error = ncl_readrpc(vp, &uio, cred);
164 	pmap_qremove(kva, npages);
165 
166 	relpbuf(bp, &ncl_pbuf_freecnt);
167 
168 	if (error && (uio.uio_resid == count)) {
169 		ncl_printf("nfs_getpages: error %d\n", error);
170 		vm_pager_free_nonreq(object, pages, ap->a_reqpage, npages,
171 		    FALSE);
172 		return (VM_PAGER_ERROR);
173 	}
174 
175 	/*
176 	 * Calculate the number of bytes read and validate only that number
177 	 * of bytes.  Note that due to pending writes, size may be 0.  This
178 	 * does not mean that the remaining data is invalid!
179 	 */
180 
181 	size = count - uio.uio_resid;
182 	VM_OBJECT_WLOCK(object);
183 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
184 		vm_page_t m;
185 		nextoff = toff + PAGE_SIZE;
186 		m = pages[i];
187 
188 		if (nextoff <= size) {
189 			/*
190 			 * Read operation filled an entire page
191 			 */
192 			m->valid = VM_PAGE_BITS_ALL;
193 			KASSERT(m->dirty == 0,
194 			    ("nfs_getpages: page %p is dirty", m));
195 		} else if (size > toff) {
196 			/*
197 			 * Read operation filled a partial page.
198 			 */
199 			m->valid = 0;
200 			vm_page_set_valid_range(m, 0, size - toff);
201 			KASSERT(m->dirty == 0,
202 			    ("nfs_getpages: page %p is dirty", m));
203 		} else {
204 			/*
205 			 * Read operation was short.  If no error
206 			 * occured we may have hit a zero-fill
207 			 * section.  We leave valid set to 0, and page
208 			 * is freed by vm_page_readahead_finish() if
209 			 * its index is not equal to requested, or
210 			 * page is zeroed and set valid by
211 			 * vm_pager_get_pages() for requested page.
212 			 */
213 			;
214 		}
215 		if (i != ap->a_reqpage)
216 			vm_page_readahead_finish(m);
217 	}
218 	VM_OBJECT_WUNLOCK(object);
219 	return (0);
220 }
221 
222 /*
223  * Vnode op for VM putpages.
224  */
225 int
226 ncl_putpages(struct vop_putpages_args *ap)
227 {
228 	struct uio uio;
229 	struct iovec iov;
230 	vm_offset_t kva;
231 	struct buf *bp;
232 	int iomode, must_commit, i, error, npages, count;
233 	off_t offset;
234 	int *rtvals;
235 	struct vnode *vp;
236 	struct thread *td;
237 	struct ucred *cred;
238 	struct nfsmount *nmp;
239 	struct nfsnode *np;
240 	vm_page_t *pages;
241 
242 	vp = ap->a_vp;
243 	np = VTONFS(vp);
244 	td = curthread;				/* XXX */
245 	/* Set the cred to n_writecred for the write rpcs. */
246 	if (np->n_writecred != NULL)
247 		cred = crhold(np->n_writecred);
248 	else
249 		cred = crhold(curthread->td_ucred);	/* XXX */
250 	nmp = VFSTONFS(vp->v_mount);
251 	pages = ap->a_m;
252 	count = ap->a_count;
253 	rtvals = ap->a_rtvals;
254 	npages = btoc(count);
255 	offset = IDX_TO_OFF(pages[0]->pindex);
256 
257 	mtx_lock(&nmp->nm_mtx);
258 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
259 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
260 		mtx_unlock(&nmp->nm_mtx);
261 		(void)ncl_fsinfo(nmp, vp, cred, td);
262 	} else
263 		mtx_unlock(&nmp->nm_mtx);
264 
265 	mtx_lock(&np->n_mtx);
266 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
267 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
268 		mtx_unlock(&np->n_mtx);
269 		ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
270 		mtx_lock(&np->n_mtx);
271 	}
272 
273 	for (i = 0; i < npages; i++)
274 		rtvals[i] = VM_PAGER_ERROR;
275 
276 	/*
277 	 * When putting pages, do not extend file past EOF.
278 	 */
279 	if (offset + count > np->n_size) {
280 		count = np->n_size - offset;
281 		if (count < 0)
282 			count = 0;
283 	}
284 	mtx_unlock(&np->n_mtx);
285 
286 	/*
287 	 * We use only the kva address for the buffer, but this is extremely
288 	 * convienient and fast.
289 	 */
290 	bp = getpbuf(&ncl_pbuf_freecnt);
291 
292 	kva = (vm_offset_t) bp->b_data;
293 	pmap_qenter(kva, pages, npages);
294 	PCPU_INC(cnt.v_vnodeout);
295 	PCPU_ADD(cnt.v_vnodepgsout, count);
296 
297 	iov.iov_base = (caddr_t) kva;
298 	iov.iov_len = count;
299 	uio.uio_iov = &iov;
300 	uio.uio_iovcnt = 1;
301 	uio.uio_offset = offset;
302 	uio.uio_resid = count;
303 	uio.uio_segflg = UIO_SYSSPACE;
304 	uio.uio_rw = UIO_WRITE;
305 	uio.uio_td = td;
306 
307 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
308 	    iomode = NFSWRITE_UNSTABLE;
309 	else
310 	    iomode = NFSWRITE_FILESYNC;
311 
312 	error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
313 	crfree(cred);
314 
315 	pmap_qremove(kva, npages);
316 	relpbuf(bp, &ncl_pbuf_freecnt);
317 
318 	if (error == 0 || !nfs_keep_dirty_on_error) {
319 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
320 		if (must_commit)
321 			ncl_clearcommit(vp->v_mount);
322 	}
323 	return rtvals[0];
324 }
325 
326 /*
327  * For nfs, cache consistency can only be maintained approximately.
328  * Although RFC1094 does not specify the criteria, the following is
329  * believed to be compatible with the reference port.
330  * For nfs:
331  * If the file's modify time on the server has changed since the
332  * last read rpc or you have written to the file,
333  * you may have lost data cache consistency with the
334  * server, so flush all of the file's data out of the cache.
335  * Then force a getattr rpc to ensure that you have up to date
336  * attributes.
337  * NB: This implies that cache data can be read when up to
338  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
339  * attributes this could be forced by setting n_attrstamp to 0 before
340  * the VOP_GETATTR() call.
341  */
342 static inline int
343 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
344 {
345 	int error = 0;
346 	struct vattr vattr;
347 	struct nfsnode *np = VTONFS(vp);
348 	int old_lock;
349 
350 	/*
351 	 * Grab the exclusive lock before checking whether the cache is
352 	 * consistent.
353 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
354 	 * But for now, this suffices.
355 	 */
356 	old_lock = ncl_upgrade_vnlock(vp);
357 	if (vp->v_iflag & VI_DOOMED) {
358 		ncl_downgrade_vnlock(vp, old_lock);
359 		return (EBADF);
360 	}
361 
362 	mtx_lock(&np->n_mtx);
363 	if (np->n_flag & NMODIFIED) {
364 		mtx_unlock(&np->n_mtx);
365 		if (vp->v_type != VREG) {
366 			if (vp->v_type != VDIR)
367 				panic("nfs: bioread, not dir");
368 			ncl_invaldir(vp);
369 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
370 			if (error)
371 				goto out;
372 		}
373 		np->n_attrstamp = 0;
374 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
375 		error = VOP_GETATTR(vp, &vattr, cred);
376 		if (error)
377 			goto out;
378 		mtx_lock(&np->n_mtx);
379 		np->n_mtime = vattr.va_mtime;
380 		mtx_unlock(&np->n_mtx);
381 	} else {
382 		mtx_unlock(&np->n_mtx);
383 		error = VOP_GETATTR(vp, &vattr, cred);
384 		if (error)
385 			return (error);
386 		mtx_lock(&np->n_mtx);
387 		if ((np->n_flag & NSIZECHANGED)
388 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
389 			mtx_unlock(&np->n_mtx);
390 			if (vp->v_type == VDIR)
391 				ncl_invaldir(vp);
392 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
393 			if (error)
394 				goto out;
395 			mtx_lock(&np->n_mtx);
396 			np->n_mtime = vattr.va_mtime;
397 			np->n_flag &= ~NSIZECHANGED;
398 		}
399 		mtx_unlock(&np->n_mtx);
400 	}
401 out:
402 	ncl_downgrade_vnlock(vp, old_lock);
403 	return error;
404 }
405 
406 /*
407  * Vnode op for read using bio
408  */
409 int
410 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
411 {
412 	struct nfsnode *np = VTONFS(vp);
413 	int biosize, i;
414 	struct buf *bp, *rabp;
415 	struct thread *td;
416 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
417 	daddr_t lbn, rabn;
418 	int bcount;
419 	int seqcount;
420 	int nra, error = 0, n = 0, on = 0;
421 	off_t tmp_off;
422 
423 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
424 	if (uio->uio_resid == 0)
425 		return (0);
426 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
427 		return (EINVAL);
428 	td = uio->uio_td;
429 
430 	mtx_lock(&nmp->nm_mtx);
431 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
432 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
433 		mtx_unlock(&nmp->nm_mtx);
434 		(void)ncl_fsinfo(nmp, vp, cred, td);
435 		mtx_lock(&nmp->nm_mtx);
436 	}
437 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
438 		(void) newnfs_iosize(nmp);
439 
440 	tmp_off = uio->uio_offset + uio->uio_resid;
441 	if (vp->v_type != VDIR &&
442 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
443 		mtx_unlock(&nmp->nm_mtx);
444 		return (EFBIG);
445 	}
446 	mtx_unlock(&nmp->nm_mtx);
447 
448 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
449 		/* No caching/ no readaheads. Just read data into the user buffer */
450 		return ncl_readrpc(vp, uio, cred);
451 
452 	biosize = vp->v_bufobj.bo_bsize;
453 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
454 
455 	error = nfs_bioread_check_cons(vp, td, cred);
456 	if (error)
457 		return error;
458 
459 	do {
460 	    u_quad_t nsize;
461 
462 	    mtx_lock(&np->n_mtx);
463 	    nsize = np->n_size;
464 	    mtx_unlock(&np->n_mtx);
465 
466 	    switch (vp->v_type) {
467 	    case VREG:
468 		NFSINCRGLOBAL(newnfsstats.biocache_reads);
469 		lbn = uio->uio_offset / biosize;
470 		on = uio->uio_offset - (lbn * biosize);
471 
472 		/*
473 		 * Start the read ahead(s), as required.
474 		 */
475 		if (nmp->nm_readahead > 0) {
476 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
477 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
478 			rabn = lbn + 1 + nra;
479 			if (incore(&vp->v_bufobj, rabn) == NULL) {
480 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
481 			    if (!rabp) {
482 				error = newnfs_sigintr(nmp, td);
483 				return (error ? error : EINTR);
484 			    }
485 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
486 				rabp->b_flags |= B_ASYNC;
487 				rabp->b_iocmd = BIO_READ;
488 				vfs_busy_pages(rabp, 0);
489 				if (ncl_asyncio(nmp, rabp, cred, td)) {
490 				    rabp->b_flags |= B_INVAL;
491 				    rabp->b_ioflags |= BIO_ERROR;
492 				    vfs_unbusy_pages(rabp);
493 				    brelse(rabp);
494 				    break;
495 				}
496 			    } else {
497 				brelse(rabp);
498 			    }
499 			}
500 		    }
501 		}
502 
503 		/* Note that bcount is *not* DEV_BSIZE aligned. */
504 		bcount = biosize;
505 		if ((off_t)lbn * biosize >= nsize) {
506 			bcount = 0;
507 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
508 			bcount = nsize - (off_t)lbn * biosize;
509 		}
510 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
511 
512 		if (!bp) {
513 			error = newnfs_sigintr(nmp, td);
514 			return (error ? error : EINTR);
515 		}
516 
517 		/*
518 		 * If B_CACHE is not set, we must issue the read.  If this
519 		 * fails, we return an error.
520 		 */
521 
522 		if ((bp->b_flags & B_CACHE) == 0) {
523 		    bp->b_iocmd = BIO_READ;
524 		    vfs_busy_pages(bp, 0);
525 		    error = ncl_doio(vp, bp, cred, td, 0);
526 		    if (error) {
527 			brelse(bp);
528 			return (error);
529 		    }
530 		}
531 
532 		/*
533 		 * on is the offset into the current bp.  Figure out how many
534 		 * bytes we can copy out of the bp.  Note that bcount is
535 		 * NOT DEV_BSIZE aligned.
536 		 *
537 		 * Then figure out how many bytes we can copy into the uio.
538 		 */
539 
540 		n = 0;
541 		if (on < bcount)
542 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
543 		break;
544 	    case VLNK:
545 		NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
546 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
547 		if (!bp) {
548 			error = newnfs_sigintr(nmp, td);
549 			return (error ? error : EINTR);
550 		}
551 		if ((bp->b_flags & B_CACHE) == 0) {
552 		    bp->b_iocmd = BIO_READ;
553 		    vfs_busy_pages(bp, 0);
554 		    error = ncl_doio(vp, bp, cred, td, 0);
555 		    if (error) {
556 			bp->b_ioflags |= BIO_ERROR;
557 			brelse(bp);
558 			return (error);
559 		    }
560 		}
561 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
562 		on = 0;
563 		break;
564 	    case VDIR:
565 		NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
566 		if (np->n_direofoffset
567 		    && uio->uio_offset >= np->n_direofoffset) {
568 		    return (0);
569 		}
570 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
571 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
572 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
573 		if (!bp) {
574 		    error = newnfs_sigintr(nmp, td);
575 		    return (error ? error : EINTR);
576 		}
577 		if ((bp->b_flags & B_CACHE) == 0) {
578 		    bp->b_iocmd = BIO_READ;
579 		    vfs_busy_pages(bp, 0);
580 		    error = ncl_doio(vp, bp, cred, td, 0);
581 		    if (error) {
582 			    brelse(bp);
583 		    }
584 		    while (error == NFSERR_BAD_COOKIE) {
585 			ncl_invaldir(vp);
586 			error = ncl_vinvalbuf(vp, 0, td, 1);
587 			/*
588 			 * Yuck! The directory has been modified on the
589 			 * server. The only way to get the block is by
590 			 * reading from the beginning to get all the
591 			 * offset cookies.
592 			 *
593 			 * Leave the last bp intact unless there is an error.
594 			 * Loop back up to the while if the error is another
595 			 * NFSERR_BAD_COOKIE (double yuch!).
596 			 */
597 			for (i = 0; i <= lbn && !error; i++) {
598 			    if (np->n_direofoffset
599 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
600 				    return (0);
601 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
602 			    if (!bp) {
603 				error = newnfs_sigintr(nmp, td);
604 				return (error ? error : EINTR);
605 			    }
606 			    if ((bp->b_flags & B_CACHE) == 0) {
607 				    bp->b_iocmd = BIO_READ;
608 				    vfs_busy_pages(bp, 0);
609 				    error = ncl_doio(vp, bp, cred, td, 0);
610 				    /*
611 				     * no error + B_INVAL == directory EOF,
612 				     * use the block.
613 				     */
614 				    if (error == 0 && (bp->b_flags & B_INVAL))
615 					    break;
616 			    }
617 			    /*
618 			     * An error will throw away the block and the
619 			     * for loop will break out.  If no error and this
620 			     * is not the block we want, we throw away the
621 			     * block and go for the next one via the for loop.
622 			     */
623 			    if (error || i < lbn)
624 				    brelse(bp);
625 			}
626 		    }
627 		    /*
628 		     * The above while is repeated if we hit another cookie
629 		     * error.  If we hit an error and it wasn't a cookie error,
630 		     * we give up.
631 		     */
632 		    if (error)
633 			    return (error);
634 		}
635 
636 		/*
637 		 * If not eof and read aheads are enabled, start one.
638 		 * (You need the current block first, so that you have the
639 		 *  directory offset cookie of the next block.)
640 		 */
641 		if (nmp->nm_readahead > 0 &&
642 		    (bp->b_flags & B_INVAL) == 0 &&
643 		    (np->n_direofoffset == 0 ||
644 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
645 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
646 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
647 			if (rabp) {
648 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
649 				rabp->b_flags |= B_ASYNC;
650 				rabp->b_iocmd = BIO_READ;
651 				vfs_busy_pages(rabp, 0);
652 				if (ncl_asyncio(nmp, rabp, cred, td)) {
653 				    rabp->b_flags |= B_INVAL;
654 				    rabp->b_ioflags |= BIO_ERROR;
655 				    vfs_unbusy_pages(rabp);
656 				    brelse(rabp);
657 				}
658 			    } else {
659 				brelse(rabp);
660 			    }
661 			}
662 		}
663 		/*
664 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
665 		 * chopped for the EOF condition, we cannot tell how large
666 		 * NFS directories are going to be until we hit EOF.  So
667 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
668 		 * it just so happens that b_resid will effectively chop it
669 		 * to EOF.  *BUT* this information is lost if the buffer goes
670 		 * away and is reconstituted into a B_CACHE state ( due to
671 		 * being VMIO ) later.  So we keep track of the directory eof
672 		 * in np->n_direofoffset and chop it off as an extra step
673 		 * right here.
674 		 */
675 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
676 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
677 			n = np->n_direofoffset - uio->uio_offset;
678 		break;
679 	    default:
680 		ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
681 		bp = NULL;
682 		break;
683 	    };
684 
685 	    if (n > 0) {
686 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
687 	    }
688 	    if (vp->v_type == VLNK)
689 		n = 0;
690 	    if (bp != NULL)
691 		brelse(bp);
692 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
693 	return (error);
694 }
695 
696 /*
697  * The NFS write path cannot handle iovecs with len > 1. So we need to
698  * break up iovecs accordingly (restricting them to wsize).
699  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
700  * For the ASYNC case, 2 copies are needed. The first a copy from the
701  * user buffer to a staging buffer and then a second copy from the staging
702  * buffer to mbufs. This can be optimized by copying from the user buffer
703  * directly into mbufs and passing the chain down, but that requires a
704  * fair amount of re-working of the relevant codepaths (and can be done
705  * later).
706  */
707 static int
708 nfs_directio_write(vp, uiop, cred, ioflag)
709 	struct vnode *vp;
710 	struct uio *uiop;
711 	struct ucred *cred;
712 	int ioflag;
713 {
714 	int error;
715 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
716 	struct thread *td = uiop->uio_td;
717 	int size;
718 	int wsize;
719 
720 	mtx_lock(&nmp->nm_mtx);
721 	wsize = nmp->nm_wsize;
722 	mtx_unlock(&nmp->nm_mtx);
723 	if (ioflag & IO_SYNC) {
724 		int iomode, must_commit;
725 		struct uio uio;
726 		struct iovec iov;
727 do_sync:
728 		while (uiop->uio_resid > 0) {
729 			size = MIN(uiop->uio_resid, wsize);
730 			size = MIN(uiop->uio_iov->iov_len, size);
731 			iov.iov_base = uiop->uio_iov->iov_base;
732 			iov.iov_len = size;
733 			uio.uio_iov = &iov;
734 			uio.uio_iovcnt = 1;
735 			uio.uio_offset = uiop->uio_offset;
736 			uio.uio_resid = size;
737 			uio.uio_segflg = UIO_USERSPACE;
738 			uio.uio_rw = UIO_WRITE;
739 			uio.uio_td = td;
740 			iomode = NFSWRITE_FILESYNC;
741 			error = ncl_writerpc(vp, &uio, cred, &iomode,
742 			    &must_commit, 0);
743 			KASSERT((must_commit == 0),
744 				("ncl_directio_write: Did not commit write"));
745 			if (error)
746 				return (error);
747 			uiop->uio_offset += size;
748 			uiop->uio_resid -= size;
749 			if (uiop->uio_iov->iov_len <= size) {
750 				uiop->uio_iovcnt--;
751 				uiop->uio_iov++;
752 			} else {
753 				uiop->uio_iov->iov_base =
754 					(char *)uiop->uio_iov->iov_base + size;
755 				uiop->uio_iov->iov_len -= size;
756 			}
757 		}
758 	} else {
759 		struct uio *t_uio;
760 		struct iovec *t_iov;
761 		struct buf *bp;
762 
763 		/*
764 		 * Break up the write into blocksize chunks and hand these
765 		 * over to nfsiod's for write back.
766 		 * Unfortunately, this incurs a copy of the data. Since
767 		 * the user could modify the buffer before the write is
768 		 * initiated.
769 		 *
770 		 * The obvious optimization here is that one of the 2 copies
771 		 * in the async write path can be eliminated by copying the
772 		 * data here directly into mbufs and passing the mbuf chain
773 		 * down. But that will require a fair amount of re-working
774 		 * of the code and can be done if there's enough interest
775 		 * in NFS directio access.
776 		 */
777 		while (uiop->uio_resid > 0) {
778 			size = MIN(uiop->uio_resid, wsize);
779 			size = MIN(uiop->uio_iov->iov_len, size);
780 			bp = getpbuf(&ncl_pbuf_freecnt);
781 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
782 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
783 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
784 			t_iov->iov_len = size;
785 			t_uio->uio_iov = t_iov;
786 			t_uio->uio_iovcnt = 1;
787 			t_uio->uio_offset = uiop->uio_offset;
788 			t_uio->uio_resid = size;
789 			t_uio->uio_segflg = UIO_SYSSPACE;
790 			t_uio->uio_rw = UIO_WRITE;
791 			t_uio->uio_td = td;
792 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
793 			    uiop->uio_segflg == UIO_SYSSPACE,
794 			    ("nfs_directio_write: Bad uio_segflg"));
795 			if (uiop->uio_segflg == UIO_USERSPACE) {
796 				error = copyin(uiop->uio_iov->iov_base,
797 				    t_iov->iov_base, size);
798 				if (error != 0)
799 					goto err_free;
800 			} else
801 				/*
802 				 * UIO_SYSSPACE may never happen, but handle
803 				 * it just in case it does.
804 				 */
805 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
806 				    size);
807 			bp->b_flags |= B_DIRECT;
808 			bp->b_iocmd = BIO_WRITE;
809 			if (cred != NOCRED) {
810 				crhold(cred);
811 				bp->b_wcred = cred;
812 			} else
813 				bp->b_wcred = NOCRED;
814 			bp->b_caller1 = (void *)t_uio;
815 			bp->b_vp = vp;
816 			error = ncl_asyncio(nmp, bp, NOCRED, td);
817 err_free:
818 			if (error) {
819 				free(t_iov->iov_base, M_NFSDIRECTIO);
820 				free(t_iov, M_NFSDIRECTIO);
821 				free(t_uio, M_NFSDIRECTIO);
822 				bp->b_vp = NULL;
823 				relpbuf(bp, &ncl_pbuf_freecnt);
824 				if (error == EINTR)
825 					return (error);
826 				goto do_sync;
827 			}
828 			uiop->uio_offset += size;
829 			uiop->uio_resid -= size;
830 			if (uiop->uio_iov->iov_len <= size) {
831 				uiop->uio_iovcnt--;
832 				uiop->uio_iov++;
833 			} else {
834 				uiop->uio_iov->iov_base =
835 					(char *)uiop->uio_iov->iov_base + size;
836 				uiop->uio_iov->iov_len -= size;
837 			}
838 		}
839 	}
840 	return (0);
841 }
842 
843 /*
844  * Vnode op for write using bio
845  */
846 int
847 ncl_write(struct vop_write_args *ap)
848 {
849 	int biosize;
850 	struct uio *uio = ap->a_uio;
851 	struct thread *td = uio->uio_td;
852 	struct vnode *vp = ap->a_vp;
853 	struct nfsnode *np = VTONFS(vp);
854 	struct ucred *cred = ap->a_cred;
855 	int ioflag = ap->a_ioflag;
856 	struct buf *bp;
857 	struct vattr vattr;
858 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
859 	daddr_t lbn;
860 	int bcount, noncontig_write, obcount;
861 	int bp_cached, n, on, error = 0, error1, wouldcommit;
862 	size_t orig_resid, local_resid;
863 	off_t orig_size, tmp_off;
864 
865 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
866 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
867 	    ("ncl_write proc"));
868 	if (vp->v_type != VREG)
869 		return (EIO);
870 	mtx_lock(&np->n_mtx);
871 	if (np->n_flag & NWRITEERR) {
872 		np->n_flag &= ~NWRITEERR;
873 		mtx_unlock(&np->n_mtx);
874 		return (np->n_error);
875 	} else
876 		mtx_unlock(&np->n_mtx);
877 	mtx_lock(&nmp->nm_mtx);
878 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
879 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
880 		mtx_unlock(&nmp->nm_mtx);
881 		(void)ncl_fsinfo(nmp, vp, cred, td);
882 		mtx_lock(&nmp->nm_mtx);
883 	}
884 	if (nmp->nm_wsize == 0)
885 		(void) newnfs_iosize(nmp);
886 	mtx_unlock(&nmp->nm_mtx);
887 
888 	/*
889 	 * Synchronously flush pending buffers if we are in synchronous
890 	 * mode or if we are appending.
891 	 */
892 	if (ioflag & (IO_APPEND | IO_SYNC)) {
893 		mtx_lock(&np->n_mtx);
894 		if (np->n_flag & NMODIFIED) {
895 			mtx_unlock(&np->n_mtx);
896 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
897 			/*
898 			 * Require non-blocking, synchronous writes to
899 			 * dirty files to inform the program it needs
900 			 * to fsync(2) explicitly.
901 			 */
902 			if (ioflag & IO_NDELAY)
903 				return (EAGAIN);
904 #endif
905 			np->n_attrstamp = 0;
906 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
907 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
908 			if (error)
909 				return (error);
910 		} else
911 			mtx_unlock(&np->n_mtx);
912 	}
913 
914 	orig_resid = uio->uio_resid;
915 	mtx_lock(&np->n_mtx);
916 	orig_size = np->n_size;
917 	mtx_unlock(&np->n_mtx);
918 
919 	/*
920 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
921 	 * get the append lock.
922 	 */
923 	if (ioflag & IO_APPEND) {
924 		np->n_attrstamp = 0;
925 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
926 		error = VOP_GETATTR(vp, &vattr, cred);
927 		if (error)
928 			return (error);
929 		mtx_lock(&np->n_mtx);
930 		uio->uio_offset = np->n_size;
931 		mtx_unlock(&np->n_mtx);
932 	}
933 
934 	if (uio->uio_offset < 0)
935 		return (EINVAL);
936 	tmp_off = uio->uio_offset + uio->uio_resid;
937 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
938 		return (EFBIG);
939 	if (uio->uio_resid == 0)
940 		return (0);
941 
942 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
943 		return nfs_directio_write(vp, uio, cred, ioflag);
944 
945 	/*
946 	 * Maybe this should be above the vnode op call, but so long as
947 	 * file servers have no limits, i don't think it matters
948 	 */
949 	if (vn_rlimit_fsize(vp, uio, td))
950 		return (EFBIG);
951 
952 	biosize = vp->v_bufobj.bo_bsize;
953 	/*
954 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
955 	 * would exceed the local maximum per-file write commit size when
956 	 * combined with those, we must decide whether to flush,
957 	 * go synchronous, or return error.  We don't bother checking
958 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
959 	 * no point optimizing for something that really won't ever happen.
960 	 */
961 	wouldcommit = 0;
962 	if (!(ioflag & IO_SYNC)) {
963 		int nflag;
964 
965 		mtx_lock(&np->n_mtx);
966 		nflag = np->n_flag;
967 		mtx_unlock(&np->n_mtx);
968 		if (nflag & NMODIFIED) {
969 			BO_LOCK(&vp->v_bufobj);
970 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
971 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
972 				    b_bobufs) {
973 					if (bp->b_flags & B_NEEDCOMMIT)
974 						wouldcommit += bp->b_bcount;
975 				}
976 			}
977 			BO_UNLOCK(&vp->v_bufobj);
978 		}
979 	}
980 
981 	do {
982 		if (!(ioflag & IO_SYNC)) {
983 			wouldcommit += biosize;
984 			if (wouldcommit > nmp->nm_wcommitsize) {
985 				np->n_attrstamp = 0;
986 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
987 				error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
988 				if (error)
989 					return (error);
990 				wouldcommit = biosize;
991 			}
992 		}
993 
994 		NFSINCRGLOBAL(newnfsstats.biocache_writes);
995 		lbn = uio->uio_offset / biosize;
996 		on = uio->uio_offset - (lbn * biosize);
997 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
998 again:
999 		/*
1000 		 * Handle direct append and file extension cases, calculate
1001 		 * unaligned buffer size.
1002 		 */
1003 		mtx_lock(&np->n_mtx);
1004 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1005 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1006 			noncontig_write = 1;
1007 		else
1008 			noncontig_write = 0;
1009 		if ((uio->uio_offset == np->n_size ||
1010 		    (noncontig_write != 0 &&
1011 		    lbn == (np->n_size / biosize) &&
1012 		    uio->uio_offset + n > np->n_size)) && n) {
1013 			mtx_unlock(&np->n_mtx);
1014 			/*
1015 			 * Get the buffer (in its pre-append state to maintain
1016 			 * B_CACHE if it was previously set).  Resize the
1017 			 * nfsnode after we have locked the buffer to prevent
1018 			 * readers from reading garbage.
1019 			 */
1020 			obcount = np->n_size - (lbn * biosize);
1021 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1022 
1023 			if (bp != NULL) {
1024 				long save;
1025 
1026 				mtx_lock(&np->n_mtx);
1027 				np->n_size = uio->uio_offset + n;
1028 				np->n_flag |= NMODIFIED;
1029 				vnode_pager_setsize(vp, np->n_size);
1030 				mtx_unlock(&np->n_mtx);
1031 
1032 				save = bp->b_flags & B_CACHE;
1033 				bcount = on + n;
1034 				allocbuf(bp, bcount);
1035 				bp->b_flags |= save;
1036 				if (noncontig_write != 0 && on > obcount)
1037 					vfs_bio_bzero_buf(bp, obcount, on -
1038 					    obcount);
1039 			}
1040 		} else {
1041 			/*
1042 			 * Obtain the locked cache block first, and then
1043 			 * adjust the file's size as appropriate.
1044 			 */
1045 			bcount = on + n;
1046 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1047 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1048 					bcount = biosize;
1049 				else
1050 					bcount = np->n_size - (off_t)lbn * biosize;
1051 			}
1052 			mtx_unlock(&np->n_mtx);
1053 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1054 			mtx_lock(&np->n_mtx);
1055 			if (uio->uio_offset + n > np->n_size) {
1056 				np->n_size = uio->uio_offset + n;
1057 				np->n_flag |= NMODIFIED;
1058 				vnode_pager_setsize(vp, np->n_size);
1059 			}
1060 			mtx_unlock(&np->n_mtx);
1061 		}
1062 
1063 		if (!bp) {
1064 			error = newnfs_sigintr(nmp, td);
1065 			if (!error)
1066 				error = EINTR;
1067 			break;
1068 		}
1069 
1070 		/*
1071 		 * Issue a READ if B_CACHE is not set.  In special-append
1072 		 * mode, B_CACHE is based on the buffer prior to the write
1073 		 * op and is typically set, avoiding the read.  If a read
1074 		 * is required in special append mode, the server will
1075 		 * probably send us a short-read since we extended the file
1076 		 * on our end, resulting in b_resid == 0 and, thusly,
1077 		 * B_CACHE getting set.
1078 		 *
1079 		 * We can also avoid issuing the read if the write covers
1080 		 * the entire buffer.  We have to make sure the buffer state
1081 		 * is reasonable in this case since we will not be initiating
1082 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1083 		 * more information.
1084 		 *
1085 		 * B_CACHE may also be set due to the buffer being cached
1086 		 * normally.
1087 		 */
1088 
1089 		bp_cached = 1;
1090 		if (on == 0 && n == bcount) {
1091 			if ((bp->b_flags & B_CACHE) == 0)
1092 				bp_cached = 0;
1093 			bp->b_flags |= B_CACHE;
1094 			bp->b_flags &= ~B_INVAL;
1095 			bp->b_ioflags &= ~BIO_ERROR;
1096 		}
1097 
1098 		if ((bp->b_flags & B_CACHE) == 0) {
1099 			bp->b_iocmd = BIO_READ;
1100 			vfs_busy_pages(bp, 0);
1101 			error = ncl_doio(vp, bp, cred, td, 0);
1102 			if (error) {
1103 				brelse(bp);
1104 				break;
1105 			}
1106 		}
1107 		if (bp->b_wcred == NOCRED)
1108 			bp->b_wcred = crhold(cred);
1109 		mtx_lock(&np->n_mtx);
1110 		np->n_flag |= NMODIFIED;
1111 		mtx_unlock(&np->n_mtx);
1112 
1113 		/*
1114 		 * If dirtyend exceeds file size, chop it down.  This should
1115 		 * not normally occur but there is an append race where it
1116 		 * might occur XXX, so we log it.
1117 		 *
1118 		 * If the chopping creates a reverse-indexed or degenerate
1119 		 * situation with dirtyoff/end, we 0 both of them.
1120 		 */
1121 
1122 		if (bp->b_dirtyend > bcount) {
1123 			ncl_printf("NFS append race @%lx:%d\n",
1124 			    (long)bp->b_blkno * DEV_BSIZE,
1125 			    bp->b_dirtyend - bcount);
1126 			bp->b_dirtyend = bcount;
1127 		}
1128 
1129 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1130 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1131 
1132 		/*
1133 		 * If the new write will leave a contiguous dirty
1134 		 * area, just update the b_dirtyoff and b_dirtyend,
1135 		 * otherwise force a write rpc of the old dirty area.
1136 		 *
1137 		 * If there has been a file lock applied to this file
1138 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1139 		 * While it is possible to merge discontiguous writes due to
1140 		 * our having a B_CACHE buffer ( and thus valid read data
1141 		 * for the hole), we don't because it could lead to
1142 		 * significant cache coherency problems with multiple clients,
1143 		 * especially if locking is implemented later on.
1144 		 *
1145 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1146 		 * not been file locking done on this file:
1147 		 * Relax coherency a bit for the sake of performance and
1148 		 * expand the current dirty region to contain the new
1149 		 * write even if it means we mark some non-dirty data as
1150 		 * dirty.
1151 		 */
1152 
1153 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1154 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1155 			if (bwrite(bp) == EINTR) {
1156 				error = EINTR;
1157 				break;
1158 			}
1159 			goto again;
1160 		}
1161 
1162 		local_resid = uio->uio_resid;
1163 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1164 
1165 		if (error != 0 && !bp_cached) {
1166 			/*
1167 			 * This block has no other content then what
1168 			 * possibly was written by the faulty uiomove.
1169 			 * Release it, forgetting the data pages, to
1170 			 * prevent the leak of uninitialized data to
1171 			 * usermode.
1172 			 */
1173 			bp->b_ioflags |= BIO_ERROR;
1174 			brelse(bp);
1175 			uio->uio_offset -= local_resid - uio->uio_resid;
1176 			uio->uio_resid = local_resid;
1177 			break;
1178 		}
1179 
1180 		/*
1181 		 * Since this block is being modified, it must be written
1182 		 * again and not just committed.  Since write clustering does
1183 		 * not work for the stage 1 data write, only the stage 2
1184 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1185 		 */
1186 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1187 
1188 		/*
1189 		 * Get the partial update on the progress made from
1190 		 * uiomove, if an error occured.
1191 		 */
1192 		if (error != 0)
1193 			n = local_resid - uio->uio_resid;
1194 
1195 		/*
1196 		 * Only update dirtyoff/dirtyend if not a degenerate
1197 		 * condition.
1198 		 */
1199 		if (n > 0) {
1200 			if (bp->b_dirtyend > 0) {
1201 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1202 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1203 			} else {
1204 				bp->b_dirtyoff = on;
1205 				bp->b_dirtyend = on + n;
1206 			}
1207 			vfs_bio_set_valid(bp, on, n);
1208 		}
1209 
1210 		/*
1211 		 * If IO_SYNC do bwrite().
1212 		 *
1213 		 * IO_INVAL appears to be unused.  The idea appears to be
1214 		 * to turn off caching in this case.  Very odd.  XXX
1215 		 */
1216 		if ((ioflag & IO_SYNC)) {
1217 			if (ioflag & IO_INVAL)
1218 				bp->b_flags |= B_NOCACHE;
1219 			error1 = bwrite(bp);
1220 			if (error1 != 0) {
1221 				if (error == 0)
1222 					error = error1;
1223 				break;
1224 			}
1225 		} else if ((n + on) == biosize) {
1226 			bp->b_flags |= B_ASYNC;
1227 			(void) ncl_writebp(bp, 0, NULL);
1228 		} else {
1229 			bdwrite(bp);
1230 		}
1231 
1232 		if (error != 0)
1233 			break;
1234 	} while (uio->uio_resid > 0 && n > 0);
1235 
1236 	if (error != 0) {
1237 		if (ioflag & IO_UNIT) {
1238 			VATTR_NULL(&vattr);
1239 			vattr.va_size = orig_size;
1240 			/* IO_SYNC is handled implicitely */
1241 			(void)VOP_SETATTR(vp, &vattr, cred);
1242 			uio->uio_offset -= orig_resid - uio->uio_resid;
1243 			uio->uio_resid = orig_resid;
1244 		}
1245 	}
1246 
1247 	return (error);
1248 }
1249 
1250 /*
1251  * Get an nfs cache block.
1252  *
1253  * Allocate a new one if the block isn't currently in the cache
1254  * and return the block marked busy. If the calling process is
1255  * interrupted by a signal for an interruptible mount point, return
1256  * NULL.
1257  *
1258  * The caller must carefully deal with the possible B_INVAL state of
1259  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1260  * indirectly), so synchronous reads can be issued without worrying about
1261  * the B_INVAL state.  We have to be a little more careful when dealing
1262  * with writes (see comments in nfs_write()) when extending a file past
1263  * its EOF.
1264  */
1265 static struct buf *
1266 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1267 {
1268 	struct buf *bp;
1269 	struct mount *mp;
1270 	struct nfsmount *nmp;
1271 
1272 	mp = vp->v_mount;
1273 	nmp = VFSTONFS(mp);
1274 
1275 	if (nmp->nm_flag & NFSMNT_INT) {
1276 		sigset_t oldset;
1277 
1278 		newnfs_set_sigmask(td, &oldset);
1279 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1280 		newnfs_restore_sigmask(td, &oldset);
1281 		while (bp == NULL) {
1282 			if (newnfs_sigintr(nmp, td))
1283 				return (NULL);
1284 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1285 		}
1286 	} else {
1287 		bp = getblk(vp, bn, size, 0, 0, 0);
1288 	}
1289 
1290 	if (vp->v_type == VREG)
1291 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1292 	return (bp);
1293 }
1294 
1295 /*
1296  * Flush and invalidate all dirty buffers. If another process is already
1297  * doing the flush, just wait for completion.
1298  */
1299 int
1300 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1301 {
1302 	struct nfsnode *np = VTONFS(vp);
1303 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1304 	int error = 0, slpflag, slptimeo;
1305 	int old_lock = 0;
1306 
1307 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1308 
1309 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1310 		intrflg = 0;
1311 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1312 		intrflg = 1;
1313 	if (intrflg) {
1314 		slpflag = PCATCH;
1315 		slptimeo = 2 * hz;
1316 	} else {
1317 		slpflag = 0;
1318 		slptimeo = 0;
1319 	}
1320 
1321 	old_lock = ncl_upgrade_vnlock(vp);
1322 	if (vp->v_iflag & VI_DOOMED) {
1323 		/*
1324 		 * Since vgonel() uses the generic vinvalbuf() to flush
1325 		 * dirty buffers and it does not call this function, it
1326 		 * is safe to just return OK when VI_DOOMED is set.
1327 		 */
1328 		ncl_downgrade_vnlock(vp, old_lock);
1329 		return (0);
1330 	}
1331 
1332 	/*
1333 	 * Now, flush as required.
1334 	 */
1335 	if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1336 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1337 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1338 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1339 		/*
1340 		 * If the page clean was interrupted, fail the invalidation.
1341 		 * Not doing so, we run the risk of losing dirty pages in the
1342 		 * vinvalbuf() call below.
1343 		 */
1344 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1345 			goto out;
1346 	}
1347 
1348 	error = vinvalbuf(vp, flags, slpflag, 0);
1349 	while (error) {
1350 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1351 			goto out;
1352 		error = vinvalbuf(vp, flags, 0, slptimeo);
1353 	}
1354 	if (NFSHASPNFS(nmp)) {
1355 		nfscl_layoutcommit(vp, td);
1356 		/*
1357 		 * Invalidate the attribute cache, since writes to a DS
1358 		 * won't update the size attribute.
1359 		 */
1360 		mtx_lock(&np->n_mtx);
1361 		np->n_attrstamp = 0;
1362 	} else
1363 		mtx_lock(&np->n_mtx);
1364 	if (np->n_directio_asyncwr == 0)
1365 		np->n_flag &= ~NMODIFIED;
1366 	mtx_unlock(&np->n_mtx);
1367 out:
1368 	ncl_downgrade_vnlock(vp, old_lock);
1369 	return error;
1370 }
1371 
1372 /*
1373  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1374  * This is mainly to avoid queueing async I/O requests when the nfsiods
1375  * are all hung on a dead server.
1376  *
1377  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1378  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1379  */
1380 int
1381 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1382 {
1383 	int iod;
1384 	int gotiod;
1385 	int slpflag = 0;
1386 	int slptimeo = 0;
1387 	int error, error2;
1388 
1389 	/*
1390 	 * Commits are usually short and sweet so lets save some cpu and
1391 	 * leave the async daemons for more important rpc's (such as reads
1392 	 * and writes).
1393 	 *
1394 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1395 	 * in the directory in order to update attributes. This can deadlock
1396 	 * with another thread that is waiting for async I/O to be done by
1397 	 * an nfsiod thread while holding a lock on one of these vnodes.
1398 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1399 	 * perform Readdirplus RPCs.
1400 	 */
1401 	mtx_lock(&ncl_iod_mutex);
1402 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1403 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1404 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1405 		mtx_unlock(&ncl_iod_mutex);
1406 		return(EIO);
1407 	}
1408 again:
1409 	if (nmp->nm_flag & NFSMNT_INT)
1410 		slpflag = PCATCH;
1411 	gotiod = FALSE;
1412 
1413 	/*
1414 	 * Find a free iod to process this request.
1415 	 */
1416 	for (iod = 0; iod < ncl_numasync; iod++)
1417 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1418 			gotiod = TRUE;
1419 			break;
1420 		}
1421 
1422 	/*
1423 	 * Try to create one if none are free.
1424 	 */
1425 	if (!gotiod)
1426 		ncl_nfsiodnew();
1427 	else {
1428 		/*
1429 		 * Found one, so wake it up and tell it which
1430 		 * mount to process.
1431 		 */
1432 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1433 		    iod, nmp));
1434 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1435 		ncl_iodmount[iod] = nmp;
1436 		nmp->nm_bufqiods++;
1437 		wakeup(&ncl_iodwant[iod]);
1438 	}
1439 
1440 	/*
1441 	 * If none are free, we may already have an iod working on this mount
1442 	 * point.  If so, it will process our request.
1443 	 */
1444 	if (!gotiod) {
1445 		if (nmp->nm_bufqiods > 0) {
1446 			NFS_DPF(ASYNCIO,
1447 				("ncl_asyncio: %d iods are already processing mount %p\n",
1448 				 nmp->nm_bufqiods, nmp));
1449 			gotiod = TRUE;
1450 		}
1451 	}
1452 
1453 	/*
1454 	 * If we have an iod which can process the request, then queue
1455 	 * the buffer.
1456 	 */
1457 	if (gotiod) {
1458 		/*
1459 		 * Ensure that the queue never grows too large.  We still want
1460 		 * to asynchronize so we block rather then return EIO.
1461 		 */
1462 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1463 			NFS_DPF(ASYNCIO,
1464 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1465 			nmp->nm_bufqwant = TRUE;
1466 			error = newnfs_msleep(td, &nmp->nm_bufq,
1467 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1468 			   slptimeo);
1469 			if (error) {
1470 				error2 = newnfs_sigintr(nmp, td);
1471 				if (error2) {
1472 					mtx_unlock(&ncl_iod_mutex);
1473 					return (error2);
1474 				}
1475 				if (slpflag == PCATCH) {
1476 					slpflag = 0;
1477 					slptimeo = 2 * hz;
1478 				}
1479 			}
1480 			/*
1481 			 * We might have lost our iod while sleeping,
1482 			 * so check and loop if nescessary.
1483 			 */
1484 			goto again;
1485 		}
1486 
1487 		/* We might have lost our nfsiod */
1488 		if (nmp->nm_bufqiods == 0) {
1489 			NFS_DPF(ASYNCIO,
1490 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1491 			goto again;
1492 		}
1493 
1494 		if (bp->b_iocmd == BIO_READ) {
1495 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1496 				bp->b_rcred = crhold(cred);
1497 		} else {
1498 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1499 				bp->b_wcred = crhold(cred);
1500 		}
1501 
1502 		if (bp->b_flags & B_REMFREE)
1503 			bremfreef(bp);
1504 		BUF_KERNPROC(bp);
1505 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1506 		nmp->nm_bufqlen++;
1507 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1508 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1509 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1510 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1511 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1512 		}
1513 		mtx_unlock(&ncl_iod_mutex);
1514 		return (0);
1515 	}
1516 
1517 	mtx_unlock(&ncl_iod_mutex);
1518 
1519 	/*
1520 	 * All the iods are busy on other mounts, so return EIO to
1521 	 * force the caller to process the i/o synchronously.
1522 	 */
1523 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1524 	return (EIO);
1525 }
1526 
1527 void
1528 ncl_doio_directwrite(struct buf *bp)
1529 {
1530 	int iomode, must_commit;
1531 	struct uio *uiop = (struct uio *)bp->b_caller1;
1532 	char *iov_base = uiop->uio_iov->iov_base;
1533 
1534 	iomode = NFSWRITE_FILESYNC;
1535 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1536 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1537 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1538 	free(iov_base, M_NFSDIRECTIO);
1539 	free(uiop->uio_iov, M_NFSDIRECTIO);
1540 	free(uiop, M_NFSDIRECTIO);
1541 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1542 		struct nfsnode *np = VTONFS(bp->b_vp);
1543 		mtx_lock(&np->n_mtx);
1544 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1545 			/*
1546 			 * Invalidate the attribute cache, since writes to a DS
1547 			 * won't update the size attribute.
1548 			 */
1549 			np->n_attrstamp = 0;
1550 		}
1551 		np->n_directio_asyncwr--;
1552 		if (np->n_directio_asyncwr == 0) {
1553 			np->n_flag &= ~NMODIFIED;
1554 			if ((np->n_flag & NFSYNCWAIT)) {
1555 				np->n_flag &= ~NFSYNCWAIT;
1556 				wakeup((caddr_t)&np->n_directio_asyncwr);
1557 			}
1558 		}
1559 		mtx_unlock(&np->n_mtx);
1560 	}
1561 	bp->b_vp = NULL;
1562 	relpbuf(bp, &ncl_pbuf_freecnt);
1563 }
1564 
1565 /*
1566  * Do an I/O operation to/from a cache block. This may be called
1567  * synchronously or from an nfsiod.
1568  */
1569 int
1570 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1571     int called_from_strategy)
1572 {
1573 	struct uio *uiop;
1574 	struct nfsnode *np;
1575 	struct nfsmount *nmp;
1576 	int error = 0, iomode, must_commit = 0;
1577 	struct uio uio;
1578 	struct iovec io;
1579 	struct proc *p = td ? td->td_proc : NULL;
1580 	uint8_t	iocmd;
1581 
1582 	np = VTONFS(vp);
1583 	nmp = VFSTONFS(vp->v_mount);
1584 	uiop = &uio;
1585 	uiop->uio_iov = &io;
1586 	uiop->uio_iovcnt = 1;
1587 	uiop->uio_segflg = UIO_SYSSPACE;
1588 	uiop->uio_td = td;
1589 
1590 	/*
1591 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1592 	 * do this here so we do not have to do it in all the code that
1593 	 * calls us.
1594 	 */
1595 	bp->b_flags &= ~B_INVAL;
1596 	bp->b_ioflags &= ~BIO_ERROR;
1597 
1598 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1599 	iocmd = bp->b_iocmd;
1600 	if (iocmd == BIO_READ) {
1601 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1602 	    io.iov_base = bp->b_data;
1603 	    uiop->uio_rw = UIO_READ;
1604 
1605 	    switch (vp->v_type) {
1606 	    case VREG:
1607 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1608 		NFSINCRGLOBAL(newnfsstats.read_bios);
1609 		error = ncl_readrpc(vp, uiop, cr);
1610 
1611 		if (!error) {
1612 		    if (uiop->uio_resid) {
1613 			/*
1614 			 * If we had a short read with no error, we must have
1615 			 * hit a file hole.  We should zero-fill the remainder.
1616 			 * This can also occur if the server hits the file EOF.
1617 			 *
1618 			 * Holes used to be able to occur due to pending
1619 			 * writes, but that is not possible any longer.
1620 			 */
1621 			int nread = bp->b_bcount - uiop->uio_resid;
1622 			ssize_t left = uiop->uio_resid;
1623 
1624 			if (left > 0)
1625 				bzero((char *)bp->b_data + nread, left);
1626 			uiop->uio_resid = 0;
1627 		    }
1628 		}
1629 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1630 		if (p && (vp->v_vflag & VV_TEXT)) {
1631 			mtx_lock(&np->n_mtx);
1632 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1633 				mtx_unlock(&np->n_mtx);
1634 				PROC_LOCK(p);
1635 				killproc(p, "text file modification");
1636 				PROC_UNLOCK(p);
1637 			} else
1638 				mtx_unlock(&np->n_mtx);
1639 		}
1640 		break;
1641 	    case VLNK:
1642 		uiop->uio_offset = (off_t)0;
1643 		NFSINCRGLOBAL(newnfsstats.readlink_bios);
1644 		error = ncl_readlinkrpc(vp, uiop, cr);
1645 		break;
1646 	    case VDIR:
1647 		NFSINCRGLOBAL(newnfsstats.readdir_bios);
1648 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1649 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1650 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1651 			if (error == NFSERR_NOTSUPP)
1652 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1653 		}
1654 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1655 			error = ncl_readdirrpc(vp, uiop, cr, td);
1656 		/*
1657 		 * end-of-directory sets B_INVAL but does not generate an
1658 		 * error.
1659 		 */
1660 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1661 			bp->b_flags |= B_INVAL;
1662 		break;
1663 	    default:
1664 		ncl_printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1665 		break;
1666 	    };
1667 	    if (error) {
1668 		bp->b_ioflags |= BIO_ERROR;
1669 		bp->b_error = error;
1670 	    }
1671 	} else {
1672 	    /*
1673 	     * If we only need to commit, try to commit
1674 	     */
1675 	    if (bp->b_flags & B_NEEDCOMMIT) {
1676 		    int retv;
1677 		    off_t off;
1678 
1679 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1680 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1681 			bp->b_wcred, td);
1682 		    if (retv == 0) {
1683 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1684 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1685 			    bp->b_resid = 0;
1686 			    bufdone(bp);
1687 			    return (0);
1688 		    }
1689 		    if (retv == NFSERR_STALEWRITEVERF) {
1690 			    ncl_clearcommit(vp->v_mount);
1691 		    }
1692 	    }
1693 
1694 	    /*
1695 	     * Setup for actual write
1696 	     */
1697 	    mtx_lock(&np->n_mtx);
1698 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1699 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1700 	    mtx_unlock(&np->n_mtx);
1701 
1702 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1703 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1704 		    - bp->b_dirtyoff;
1705 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1706 		    + bp->b_dirtyoff;
1707 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1708 		uiop->uio_rw = UIO_WRITE;
1709 		NFSINCRGLOBAL(newnfsstats.write_bios);
1710 
1711 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1712 		    iomode = NFSWRITE_UNSTABLE;
1713 		else
1714 		    iomode = NFSWRITE_FILESYNC;
1715 
1716 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1717 		    called_from_strategy);
1718 
1719 		/*
1720 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1721 		 * to cluster the buffers needing commit.  This will allow
1722 		 * the system to submit a single commit rpc for the whole
1723 		 * cluster.  We can do this even if the buffer is not 100%
1724 		 * dirty (relative to the NFS blocksize), so we optimize the
1725 		 * append-to-file-case.
1726 		 *
1727 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1728 		 * cleared because write clustering only works for commit
1729 		 * rpc's, not for the data portion of the write).
1730 		 */
1731 
1732 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1733 		    bp->b_flags |= B_NEEDCOMMIT;
1734 		    if (bp->b_dirtyoff == 0
1735 			&& bp->b_dirtyend == bp->b_bcount)
1736 			bp->b_flags |= B_CLUSTEROK;
1737 		} else {
1738 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1739 		}
1740 
1741 		/*
1742 		 * For an interrupted write, the buffer is still valid
1743 		 * and the write hasn't been pushed to the server yet,
1744 		 * so we can't set BIO_ERROR and report the interruption
1745 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1746 		 * is not relevant, so the rpc attempt is essentially
1747 		 * a noop.  For the case of a V3 write rpc not being
1748 		 * committed to stable storage, the block is still
1749 		 * dirty and requires either a commit rpc or another
1750 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1751 		 * the block is reused. This is indicated by setting
1752 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1753 		 *
1754 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1755 		 * write error and is handled as above, except that
1756 		 * B_EINTR isn't set. One cause of this is a stale stateid
1757 		 * error for the RPC that indicates recovery is required,
1758 		 * when called with called_from_strategy != 0.
1759 		 *
1760 		 * If the buffer is marked B_PAGING, it does not reside on
1761 		 * the vp's paging queues so we cannot call bdirty().  The
1762 		 * bp in this case is not an NFS cache block so we should
1763 		 * be safe. XXX
1764 		 *
1765 		 * The logic below breaks up errors into recoverable and
1766 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1767 		 * and keep the buffer around for potential write retries.
1768 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1769 		 * and save the error in the nfsnode. This is less than ideal
1770 		 * but necessary. Keeping such buffers around could potentially
1771 		 * cause buffer exhaustion eventually (they can never be written
1772 		 * out, so will get constantly be re-dirtied). It also causes
1773 		 * all sorts of vfs panics. For non-recoverable write errors,
1774 		 * also invalidate the attrcache, so we'll be forced to go over
1775 		 * the wire for this object, returning an error to user on next
1776 		 * call (most of the time).
1777 		 */
1778 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1779 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1780 			int s;
1781 
1782 			s = splbio();
1783 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1784 			if ((bp->b_flags & B_PAGING) == 0) {
1785 			    bdirty(bp);
1786 			    bp->b_flags &= ~B_DONE;
1787 			}
1788 			if ((error == EINTR || error == ETIMEDOUT) &&
1789 			    (bp->b_flags & B_ASYNC) == 0)
1790 			    bp->b_flags |= B_EINTR;
1791 			splx(s);
1792 		} else {
1793 		    if (error) {
1794 			bp->b_ioflags |= BIO_ERROR;
1795 			bp->b_flags |= B_INVAL;
1796 			bp->b_error = np->n_error = error;
1797 			mtx_lock(&np->n_mtx);
1798 			np->n_flag |= NWRITEERR;
1799 			np->n_attrstamp = 0;
1800 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1801 			mtx_unlock(&np->n_mtx);
1802 		    }
1803 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1804 		}
1805 	    } else {
1806 		bp->b_resid = 0;
1807 		bufdone(bp);
1808 		return (0);
1809 	    }
1810 	}
1811 	bp->b_resid = uiop->uio_resid;
1812 	if (must_commit)
1813 	    ncl_clearcommit(vp->v_mount);
1814 	bufdone(bp);
1815 	return (error);
1816 }
1817 
1818 /*
1819  * Used to aid in handling ftruncate() operations on the NFS client side.
1820  * Truncation creates a number of special problems for NFS.  We have to
1821  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1822  * we have to properly handle VM pages or (potentially dirty) buffers
1823  * that straddle the truncation point.
1824  */
1825 
1826 int
1827 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1828 {
1829 	struct nfsnode *np = VTONFS(vp);
1830 	u_quad_t tsize;
1831 	int biosize = vp->v_bufobj.bo_bsize;
1832 	int error = 0;
1833 
1834 	mtx_lock(&np->n_mtx);
1835 	tsize = np->n_size;
1836 	np->n_size = nsize;
1837 	mtx_unlock(&np->n_mtx);
1838 
1839 	if (nsize < tsize) {
1840 		struct buf *bp;
1841 		daddr_t lbn;
1842 		int bufsize;
1843 
1844 		/*
1845 		 * vtruncbuf() doesn't get the buffer overlapping the
1846 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1847 		 * buffer that now needs to be truncated.
1848 		 */
1849 		error = vtruncbuf(vp, cred, nsize, biosize);
1850 		lbn = nsize / biosize;
1851 		bufsize = nsize - (lbn * biosize);
1852 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1853 		if (!bp)
1854 			return EINTR;
1855 		if (bp->b_dirtyoff > bp->b_bcount)
1856 			bp->b_dirtyoff = bp->b_bcount;
1857 		if (bp->b_dirtyend > bp->b_bcount)
1858 			bp->b_dirtyend = bp->b_bcount;
1859 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1860 		brelse(bp);
1861 	} else {
1862 		vnode_pager_setsize(vp, nsize);
1863 	}
1864 	return(error);
1865 }
1866 
1867