xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 57718be8fa0bd5edc11ab9a72e68cc71982939a6)
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 	 * Since the caller has busied the requested page, that page's valid
133 	 * field will not be changed by other threads.
134 	 */
135 	vm_page_assert_xbusied(pages[ap->a_reqpage]);
136 
137 	/*
138 	 * If the requested page is partially valid, just return it and
139 	 * allow the pager to zero-out the blanks.  Partially valid pages
140 	 * can only occur at the file EOF.
141 	 */
142 	if (pages[ap->a_reqpage]->valid != 0) {
143 		vm_pager_free_nonreq(object, pages, ap->a_reqpage, npages);
144 		return (VM_PAGER_OK);
145 	}
146 
147 	/*
148 	 * We use only the kva address for the buffer, but this is extremely
149 	 * convienient and fast.
150 	 */
151 	bp = getpbuf(&ncl_pbuf_freecnt);
152 
153 	kva = (vm_offset_t) bp->b_data;
154 	pmap_qenter(kva, pages, npages);
155 	PCPU_INC(cnt.v_vnodein);
156 	PCPU_ADD(cnt.v_vnodepgsin, npages);
157 
158 	iov.iov_base = (caddr_t) kva;
159 	iov.iov_len = count;
160 	uio.uio_iov = &iov;
161 	uio.uio_iovcnt = 1;
162 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
163 	uio.uio_resid = count;
164 	uio.uio_segflg = UIO_SYSSPACE;
165 	uio.uio_rw = UIO_READ;
166 	uio.uio_td = td;
167 
168 	error = ncl_readrpc(vp, &uio, cred);
169 	pmap_qremove(kva, npages);
170 
171 	relpbuf(bp, &ncl_pbuf_freecnt);
172 
173 	if (error && (uio.uio_resid == count)) {
174 		ncl_printf("nfs_getpages: error %d\n", error);
175 		vm_pager_free_nonreq(object, pages, ap->a_reqpage, npages);
176 		return (VM_PAGER_ERROR);
177 	}
178 
179 	/*
180 	 * Calculate the number of bytes read and validate only that number
181 	 * of bytes.  Note that due to pending writes, size may be 0.  This
182 	 * does not mean that the remaining data is invalid!
183 	 */
184 
185 	size = count - uio.uio_resid;
186 	VM_OBJECT_WLOCK(object);
187 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
188 		vm_page_t m;
189 		nextoff = toff + PAGE_SIZE;
190 		m = pages[i];
191 
192 		if (nextoff <= size) {
193 			/*
194 			 * Read operation filled an entire page
195 			 */
196 			m->valid = VM_PAGE_BITS_ALL;
197 			KASSERT(m->dirty == 0,
198 			    ("nfs_getpages: page %p is dirty", m));
199 		} else if (size > toff) {
200 			/*
201 			 * Read operation filled a partial page.
202 			 */
203 			m->valid = 0;
204 			vm_page_set_valid_range(m, 0, size - toff);
205 			KASSERT(m->dirty == 0,
206 			    ("nfs_getpages: page %p is dirty", m));
207 		} else {
208 			/*
209 			 * Read operation was short.  If no error
210 			 * occured we may have hit a zero-fill
211 			 * section.  We leave valid set to 0, and page
212 			 * is freed by vm_page_readahead_finish() if
213 			 * its index is not equal to requested, or
214 			 * page is zeroed and set valid by
215 			 * vm_pager_get_pages() for requested page.
216 			 */
217 			;
218 		}
219 		if (i != ap->a_reqpage)
220 			vm_page_readahead_finish(m);
221 	}
222 	VM_OBJECT_WUNLOCK(object);
223 	return (0);
224 }
225 
226 /*
227  * Vnode op for VM putpages.
228  */
229 int
230 ncl_putpages(struct vop_putpages_args *ap)
231 {
232 	struct uio uio;
233 	struct iovec iov;
234 	vm_offset_t kva;
235 	struct buf *bp;
236 	int iomode, must_commit, i, error, npages, count;
237 	off_t offset;
238 	int *rtvals;
239 	struct vnode *vp;
240 	struct thread *td;
241 	struct ucred *cred;
242 	struct nfsmount *nmp;
243 	struct nfsnode *np;
244 	vm_page_t *pages;
245 
246 	vp = ap->a_vp;
247 	np = VTONFS(vp);
248 	td = curthread;				/* XXX */
249 	/* Set the cred to n_writecred for the write rpcs. */
250 	if (np->n_writecred != NULL)
251 		cred = crhold(np->n_writecred);
252 	else
253 		cred = crhold(curthread->td_ucred);	/* XXX */
254 	nmp = VFSTONFS(vp->v_mount);
255 	pages = ap->a_m;
256 	count = ap->a_count;
257 	rtvals = ap->a_rtvals;
258 	npages = btoc(count);
259 	offset = IDX_TO_OFF(pages[0]->pindex);
260 
261 	mtx_lock(&nmp->nm_mtx);
262 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
263 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
264 		mtx_unlock(&nmp->nm_mtx);
265 		(void)ncl_fsinfo(nmp, vp, cred, td);
266 	} else
267 		mtx_unlock(&nmp->nm_mtx);
268 
269 	mtx_lock(&np->n_mtx);
270 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
271 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
272 		mtx_unlock(&np->n_mtx);
273 		ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
274 		mtx_lock(&np->n_mtx);
275 	}
276 
277 	for (i = 0; i < npages; i++)
278 		rtvals[i] = VM_PAGER_ERROR;
279 
280 	/*
281 	 * When putting pages, do not extend file past EOF.
282 	 */
283 	if (offset + count > np->n_size) {
284 		count = np->n_size - offset;
285 		if (count < 0)
286 			count = 0;
287 	}
288 	mtx_unlock(&np->n_mtx);
289 
290 	/*
291 	 * We use only the kva address for the buffer, but this is extremely
292 	 * convienient and fast.
293 	 */
294 	bp = getpbuf(&ncl_pbuf_freecnt);
295 
296 	kva = (vm_offset_t) bp->b_data;
297 	pmap_qenter(kva, pages, npages);
298 	PCPU_INC(cnt.v_vnodeout);
299 	PCPU_ADD(cnt.v_vnodepgsout, count);
300 
301 	iov.iov_base = (caddr_t) kva;
302 	iov.iov_len = count;
303 	uio.uio_iov = &iov;
304 	uio.uio_iovcnt = 1;
305 	uio.uio_offset = offset;
306 	uio.uio_resid = count;
307 	uio.uio_segflg = UIO_SYSSPACE;
308 	uio.uio_rw = UIO_WRITE;
309 	uio.uio_td = td;
310 
311 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
312 	    iomode = NFSWRITE_UNSTABLE;
313 	else
314 	    iomode = NFSWRITE_FILESYNC;
315 
316 	error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
317 	crfree(cred);
318 
319 	pmap_qremove(kva, npages);
320 	relpbuf(bp, &ncl_pbuf_freecnt);
321 
322 	if (error == 0 || !nfs_keep_dirty_on_error) {
323 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
324 		if (must_commit)
325 			ncl_clearcommit(vp->v_mount);
326 	}
327 	return rtvals[0];
328 }
329 
330 /*
331  * For nfs, cache consistency can only be maintained approximately.
332  * Although RFC1094 does not specify the criteria, the following is
333  * believed to be compatible with the reference port.
334  * For nfs:
335  * If the file's modify time on the server has changed since the
336  * last read rpc or you have written to the file,
337  * you may have lost data cache consistency with the
338  * server, so flush all of the file's data out of the cache.
339  * Then force a getattr rpc to ensure that you have up to date
340  * attributes.
341  * NB: This implies that cache data can be read when up to
342  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
343  * attributes this could be forced by setting n_attrstamp to 0 before
344  * the VOP_GETATTR() call.
345  */
346 static inline int
347 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
348 {
349 	int error = 0;
350 	struct vattr vattr;
351 	struct nfsnode *np = VTONFS(vp);
352 	int old_lock;
353 
354 	/*
355 	 * Grab the exclusive lock before checking whether the cache is
356 	 * consistent.
357 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
358 	 * But for now, this suffices.
359 	 */
360 	old_lock = ncl_upgrade_vnlock(vp);
361 	if (vp->v_iflag & VI_DOOMED) {
362 		ncl_downgrade_vnlock(vp, old_lock);
363 		return (EBADF);
364 	}
365 
366 	mtx_lock(&np->n_mtx);
367 	if (np->n_flag & NMODIFIED) {
368 		mtx_unlock(&np->n_mtx);
369 		if (vp->v_type != VREG) {
370 			if (vp->v_type != VDIR)
371 				panic("nfs: bioread, not dir");
372 			ncl_invaldir(vp);
373 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
374 			if (error)
375 				goto out;
376 		}
377 		np->n_attrstamp = 0;
378 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
379 		error = VOP_GETATTR(vp, &vattr, cred);
380 		if (error)
381 			goto out;
382 		mtx_lock(&np->n_mtx);
383 		np->n_mtime = vattr.va_mtime;
384 		mtx_unlock(&np->n_mtx);
385 	} else {
386 		mtx_unlock(&np->n_mtx);
387 		error = VOP_GETATTR(vp, &vattr, cred);
388 		if (error)
389 			return (error);
390 		mtx_lock(&np->n_mtx);
391 		if ((np->n_flag & NSIZECHANGED)
392 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
393 			mtx_unlock(&np->n_mtx);
394 			if (vp->v_type == VDIR)
395 				ncl_invaldir(vp);
396 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
397 			if (error)
398 				goto out;
399 			mtx_lock(&np->n_mtx);
400 			np->n_mtime = vattr.va_mtime;
401 			np->n_flag &= ~NSIZECHANGED;
402 		}
403 		mtx_unlock(&np->n_mtx);
404 	}
405 out:
406 	ncl_downgrade_vnlock(vp, old_lock);
407 	return error;
408 }
409 
410 /*
411  * Vnode op for read using bio
412  */
413 int
414 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
415 {
416 	struct nfsnode *np = VTONFS(vp);
417 	int biosize, i;
418 	struct buf *bp, *rabp;
419 	struct thread *td;
420 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
421 	daddr_t lbn, rabn;
422 	int bcount;
423 	int seqcount;
424 	int nra, error = 0, n = 0, on = 0;
425 	off_t tmp_off;
426 
427 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
428 	if (uio->uio_resid == 0)
429 		return (0);
430 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
431 		return (EINVAL);
432 	td = uio->uio_td;
433 
434 	mtx_lock(&nmp->nm_mtx);
435 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
436 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
437 		mtx_unlock(&nmp->nm_mtx);
438 		(void)ncl_fsinfo(nmp, vp, cred, td);
439 		mtx_lock(&nmp->nm_mtx);
440 	}
441 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
442 		(void) newnfs_iosize(nmp);
443 
444 	tmp_off = uio->uio_offset + uio->uio_resid;
445 	if (vp->v_type != VDIR &&
446 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
447 		mtx_unlock(&nmp->nm_mtx);
448 		return (EFBIG);
449 	}
450 	mtx_unlock(&nmp->nm_mtx);
451 
452 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
453 		/* No caching/ no readaheads. Just read data into the user buffer */
454 		return ncl_readrpc(vp, uio, cred);
455 
456 	biosize = vp->v_bufobj.bo_bsize;
457 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
458 
459 	error = nfs_bioread_check_cons(vp, td, cred);
460 	if (error)
461 		return error;
462 
463 	do {
464 	    u_quad_t nsize;
465 
466 	    mtx_lock(&np->n_mtx);
467 	    nsize = np->n_size;
468 	    mtx_unlock(&np->n_mtx);
469 
470 	    switch (vp->v_type) {
471 	    case VREG:
472 		NFSINCRGLOBAL(newnfsstats.biocache_reads);
473 		lbn = uio->uio_offset / biosize;
474 		on = uio->uio_offset - (lbn * biosize);
475 
476 		/*
477 		 * Start the read ahead(s), as required.
478 		 */
479 		if (nmp->nm_readahead > 0) {
480 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
481 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
482 			rabn = lbn + 1 + nra;
483 			if (incore(&vp->v_bufobj, rabn) == NULL) {
484 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
485 			    if (!rabp) {
486 				error = newnfs_sigintr(nmp, td);
487 				return (error ? error : EINTR);
488 			    }
489 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
490 				rabp->b_flags |= B_ASYNC;
491 				rabp->b_iocmd = BIO_READ;
492 				vfs_busy_pages(rabp, 0);
493 				if (ncl_asyncio(nmp, rabp, cred, td)) {
494 				    rabp->b_flags |= B_INVAL;
495 				    rabp->b_ioflags |= BIO_ERROR;
496 				    vfs_unbusy_pages(rabp);
497 				    brelse(rabp);
498 				    break;
499 				}
500 			    } else {
501 				brelse(rabp);
502 			    }
503 			}
504 		    }
505 		}
506 
507 		/* Note that bcount is *not* DEV_BSIZE aligned. */
508 		bcount = biosize;
509 		if ((off_t)lbn * biosize >= nsize) {
510 			bcount = 0;
511 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
512 			bcount = nsize - (off_t)lbn * biosize;
513 		}
514 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
515 
516 		if (!bp) {
517 			error = newnfs_sigintr(nmp, td);
518 			return (error ? error : EINTR);
519 		}
520 
521 		/*
522 		 * If B_CACHE is not set, we must issue the read.  If this
523 		 * fails, we return an error.
524 		 */
525 
526 		if ((bp->b_flags & B_CACHE) == 0) {
527 		    bp->b_iocmd = BIO_READ;
528 		    vfs_busy_pages(bp, 0);
529 		    error = ncl_doio(vp, bp, cred, td, 0);
530 		    if (error) {
531 			brelse(bp);
532 			return (error);
533 		    }
534 		}
535 
536 		/*
537 		 * on is the offset into the current bp.  Figure out how many
538 		 * bytes we can copy out of the bp.  Note that bcount is
539 		 * NOT DEV_BSIZE aligned.
540 		 *
541 		 * Then figure out how many bytes we can copy into the uio.
542 		 */
543 
544 		n = 0;
545 		if (on < bcount)
546 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
547 		break;
548 	    case VLNK:
549 		NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
550 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
551 		if (!bp) {
552 			error = newnfs_sigintr(nmp, td);
553 			return (error ? error : EINTR);
554 		}
555 		if ((bp->b_flags & B_CACHE) == 0) {
556 		    bp->b_iocmd = BIO_READ;
557 		    vfs_busy_pages(bp, 0);
558 		    error = ncl_doio(vp, bp, cred, td, 0);
559 		    if (error) {
560 			bp->b_ioflags |= BIO_ERROR;
561 			brelse(bp);
562 			return (error);
563 		    }
564 		}
565 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
566 		on = 0;
567 		break;
568 	    case VDIR:
569 		NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
570 		if (np->n_direofoffset
571 		    && uio->uio_offset >= np->n_direofoffset) {
572 		    return (0);
573 		}
574 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
575 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
576 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
577 		if (!bp) {
578 		    error = newnfs_sigintr(nmp, td);
579 		    return (error ? error : EINTR);
580 		}
581 		if ((bp->b_flags & B_CACHE) == 0) {
582 		    bp->b_iocmd = BIO_READ;
583 		    vfs_busy_pages(bp, 0);
584 		    error = ncl_doio(vp, bp, cred, td, 0);
585 		    if (error) {
586 			    brelse(bp);
587 		    }
588 		    while (error == NFSERR_BAD_COOKIE) {
589 			ncl_invaldir(vp);
590 			error = ncl_vinvalbuf(vp, 0, td, 1);
591 			/*
592 			 * Yuck! The directory has been modified on the
593 			 * server. The only way to get the block is by
594 			 * reading from the beginning to get all the
595 			 * offset cookies.
596 			 *
597 			 * Leave the last bp intact unless there is an error.
598 			 * Loop back up to the while if the error is another
599 			 * NFSERR_BAD_COOKIE (double yuch!).
600 			 */
601 			for (i = 0; i <= lbn && !error; i++) {
602 			    if (np->n_direofoffset
603 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
604 				    return (0);
605 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
606 			    if (!bp) {
607 				error = newnfs_sigintr(nmp, td);
608 				return (error ? error : EINTR);
609 			    }
610 			    if ((bp->b_flags & B_CACHE) == 0) {
611 				    bp->b_iocmd = BIO_READ;
612 				    vfs_busy_pages(bp, 0);
613 				    error = ncl_doio(vp, bp, cred, td, 0);
614 				    /*
615 				     * no error + B_INVAL == directory EOF,
616 				     * use the block.
617 				     */
618 				    if (error == 0 && (bp->b_flags & B_INVAL))
619 					    break;
620 			    }
621 			    /*
622 			     * An error will throw away the block and the
623 			     * for loop will break out.  If no error and this
624 			     * is not the block we want, we throw away the
625 			     * block and go for the next one via the for loop.
626 			     */
627 			    if (error || i < lbn)
628 				    brelse(bp);
629 			}
630 		    }
631 		    /*
632 		     * The above while is repeated if we hit another cookie
633 		     * error.  If we hit an error and it wasn't a cookie error,
634 		     * we give up.
635 		     */
636 		    if (error)
637 			    return (error);
638 		}
639 
640 		/*
641 		 * If not eof and read aheads are enabled, start one.
642 		 * (You need the current block first, so that you have the
643 		 *  directory offset cookie of the next block.)
644 		 */
645 		if (nmp->nm_readahead > 0 &&
646 		    (bp->b_flags & B_INVAL) == 0 &&
647 		    (np->n_direofoffset == 0 ||
648 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
649 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
650 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
651 			if (rabp) {
652 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
653 				rabp->b_flags |= B_ASYNC;
654 				rabp->b_iocmd = BIO_READ;
655 				vfs_busy_pages(rabp, 0);
656 				if (ncl_asyncio(nmp, rabp, cred, td)) {
657 				    rabp->b_flags |= B_INVAL;
658 				    rabp->b_ioflags |= BIO_ERROR;
659 				    vfs_unbusy_pages(rabp);
660 				    brelse(rabp);
661 				}
662 			    } else {
663 				brelse(rabp);
664 			    }
665 			}
666 		}
667 		/*
668 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
669 		 * chopped for the EOF condition, we cannot tell how large
670 		 * NFS directories are going to be until we hit EOF.  So
671 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
672 		 * it just so happens that b_resid will effectively chop it
673 		 * to EOF.  *BUT* this information is lost if the buffer goes
674 		 * away and is reconstituted into a B_CACHE state ( due to
675 		 * being VMIO ) later.  So we keep track of the directory eof
676 		 * in np->n_direofoffset and chop it off as an extra step
677 		 * right here.
678 		 */
679 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
680 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
681 			n = np->n_direofoffset - uio->uio_offset;
682 		break;
683 	    default:
684 		ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
685 		bp = NULL;
686 		break;
687 	    };
688 
689 	    if (n > 0) {
690 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
691 	    }
692 	    if (vp->v_type == VLNK)
693 		n = 0;
694 	    if (bp != NULL)
695 		brelse(bp);
696 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
697 	return (error);
698 }
699 
700 /*
701  * The NFS write path cannot handle iovecs with len > 1. So we need to
702  * break up iovecs accordingly (restricting them to wsize).
703  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
704  * For the ASYNC case, 2 copies are needed. The first a copy from the
705  * user buffer to a staging buffer and then a second copy from the staging
706  * buffer to mbufs. This can be optimized by copying from the user buffer
707  * directly into mbufs and passing the chain down, but that requires a
708  * fair amount of re-working of the relevant codepaths (and can be done
709  * later).
710  */
711 static int
712 nfs_directio_write(vp, uiop, cred, ioflag)
713 	struct vnode *vp;
714 	struct uio *uiop;
715 	struct ucred *cred;
716 	int ioflag;
717 {
718 	int error;
719 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
720 	struct thread *td = uiop->uio_td;
721 	int size;
722 	int wsize;
723 
724 	mtx_lock(&nmp->nm_mtx);
725 	wsize = nmp->nm_wsize;
726 	mtx_unlock(&nmp->nm_mtx);
727 	if (ioflag & IO_SYNC) {
728 		int iomode, must_commit;
729 		struct uio uio;
730 		struct iovec iov;
731 do_sync:
732 		while (uiop->uio_resid > 0) {
733 			size = MIN(uiop->uio_resid, wsize);
734 			size = MIN(uiop->uio_iov->iov_len, size);
735 			iov.iov_base = uiop->uio_iov->iov_base;
736 			iov.iov_len = size;
737 			uio.uio_iov = &iov;
738 			uio.uio_iovcnt = 1;
739 			uio.uio_offset = uiop->uio_offset;
740 			uio.uio_resid = size;
741 			uio.uio_segflg = UIO_USERSPACE;
742 			uio.uio_rw = UIO_WRITE;
743 			uio.uio_td = td;
744 			iomode = NFSWRITE_FILESYNC;
745 			error = ncl_writerpc(vp, &uio, cred, &iomode,
746 			    &must_commit, 0);
747 			KASSERT((must_commit == 0),
748 				("ncl_directio_write: Did not commit write"));
749 			if (error)
750 				return (error);
751 			uiop->uio_offset += size;
752 			uiop->uio_resid -= size;
753 			if (uiop->uio_iov->iov_len <= size) {
754 				uiop->uio_iovcnt--;
755 				uiop->uio_iov++;
756 			} else {
757 				uiop->uio_iov->iov_base =
758 					(char *)uiop->uio_iov->iov_base + size;
759 				uiop->uio_iov->iov_len -= size;
760 			}
761 		}
762 	} else {
763 		struct uio *t_uio;
764 		struct iovec *t_iov;
765 		struct buf *bp;
766 
767 		/*
768 		 * Break up the write into blocksize chunks and hand these
769 		 * over to nfsiod's for write back.
770 		 * Unfortunately, this incurs a copy of the data. Since
771 		 * the user could modify the buffer before the write is
772 		 * initiated.
773 		 *
774 		 * The obvious optimization here is that one of the 2 copies
775 		 * in the async write path can be eliminated by copying the
776 		 * data here directly into mbufs and passing the mbuf chain
777 		 * down. But that will require a fair amount of re-working
778 		 * of the code and can be done if there's enough interest
779 		 * in NFS directio access.
780 		 */
781 		while (uiop->uio_resid > 0) {
782 			size = MIN(uiop->uio_resid, wsize);
783 			size = MIN(uiop->uio_iov->iov_len, size);
784 			bp = getpbuf(&ncl_pbuf_freecnt);
785 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
786 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
787 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
788 			t_iov->iov_len = size;
789 			t_uio->uio_iov = t_iov;
790 			t_uio->uio_iovcnt = 1;
791 			t_uio->uio_offset = uiop->uio_offset;
792 			t_uio->uio_resid = size;
793 			t_uio->uio_segflg = UIO_SYSSPACE;
794 			t_uio->uio_rw = UIO_WRITE;
795 			t_uio->uio_td = td;
796 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
797 			    uiop->uio_segflg == UIO_SYSSPACE,
798 			    ("nfs_directio_write: Bad uio_segflg"));
799 			if (uiop->uio_segflg == UIO_USERSPACE) {
800 				error = copyin(uiop->uio_iov->iov_base,
801 				    t_iov->iov_base, size);
802 				if (error != 0)
803 					goto err_free;
804 			} else
805 				/*
806 				 * UIO_SYSSPACE may never happen, but handle
807 				 * it just in case it does.
808 				 */
809 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
810 				    size);
811 			bp->b_flags |= B_DIRECT;
812 			bp->b_iocmd = BIO_WRITE;
813 			if (cred != NOCRED) {
814 				crhold(cred);
815 				bp->b_wcred = cred;
816 			} else
817 				bp->b_wcred = NOCRED;
818 			bp->b_caller1 = (void *)t_uio;
819 			bp->b_vp = vp;
820 			error = ncl_asyncio(nmp, bp, NOCRED, td);
821 err_free:
822 			if (error) {
823 				free(t_iov->iov_base, M_NFSDIRECTIO);
824 				free(t_iov, M_NFSDIRECTIO);
825 				free(t_uio, M_NFSDIRECTIO);
826 				bp->b_vp = NULL;
827 				relpbuf(bp, &ncl_pbuf_freecnt);
828 				if (error == EINTR)
829 					return (error);
830 				goto do_sync;
831 			}
832 			uiop->uio_offset += size;
833 			uiop->uio_resid -= size;
834 			if (uiop->uio_iov->iov_len <= size) {
835 				uiop->uio_iovcnt--;
836 				uiop->uio_iov++;
837 			} else {
838 				uiop->uio_iov->iov_base =
839 					(char *)uiop->uio_iov->iov_base + size;
840 				uiop->uio_iov->iov_len -= size;
841 			}
842 		}
843 	}
844 	return (0);
845 }
846 
847 /*
848  * Vnode op for write using bio
849  */
850 int
851 ncl_write(struct vop_write_args *ap)
852 {
853 	int biosize;
854 	struct uio *uio = ap->a_uio;
855 	struct thread *td = uio->uio_td;
856 	struct vnode *vp = ap->a_vp;
857 	struct nfsnode *np = VTONFS(vp);
858 	struct ucred *cred = ap->a_cred;
859 	int ioflag = ap->a_ioflag;
860 	struct buf *bp;
861 	struct vattr vattr;
862 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
863 	daddr_t lbn;
864 	int bcount, noncontig_write, obcount;
865 	int bp_cached, n, on, error = 0, error1;
866 	size_t orig_resid, local_resid;
867 	off_t orig_size, tmp_off;
868 
869 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
870 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
871 	    ("ncl_write proc"));
872 	if (vp->v_type != VREG)
873 		return (EIO);
874 	mtx_lock(&np->n_mtx);
875 	if (np->n_flag & NWRITEERR) {
876 		np->n_flag &= ~NWRITEERR;
877 		mtx_unlock(&np->n_mtx);
878 		return (np->n_error);
879 	} else
880 		mtx_unlock(&np->n_mtx);
881 	mtx_lock(&nmp->nm_mtx);
882 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
883 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
884 		mtx_unlock(&nmp->nm_mtx);
885 		(void)ncl_fsinfo(nmp, vp, cred, td);
886 		mtx_lock(&nmp->nm_mtx);
887 	}
888 	if (nmp->nm_wsize == 0)
889 		(void) newnfs_iosize(nmp);
890 	mtx_unlock(&nmp->nm_mtx);
891 
892 	/*
893 	 * Synchronously flush pending buffers if we are in synchronous
894 	 * mode or if we are appending.
895 	 */
896 	if (ioflag & (IO_APPEND | IO_SYNC)) {
897 		mtx_lock(&np->n_mtx);
898 		if (np->n_flag & NMODIFIED) {
899 			mtx_unlock(&np->n_mtx);
900 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
901 			/*
902 			 * Require non-blocking, synchronous writes to
903 			 * dirty files to inform the program it needs
904 			 * to fsync(2) explicitly.
905 			 */
906 			if (ioflag & IO_NDELAY)
907 				return (EAGAIN);
908 #endif
909 flush_and_restart:
910 			np->n_attrstamp = 0;
911 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
912 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
913 			if (error)
914 				return (error);
915 		} else
916 			mtx_unlock(&np->n_mtx);
917 	}
918 
919 	orig_resid = uio->uio_resid;
920 	mtx_lock(&np->n_mtx);
921 	orig_size = np->n_size;
922 	mtx_unlock(&np->n_mtx);
923 
924 	/*
925 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
926 	 * get the append lock.
927 	 */
928 	if (ioflag & IO_APPEND) {
929 		np->n_attrstamp = 0;
930 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
931 		error = VOP_GETATTR(vp, &vattr, cred);
932 		if (error)
933 			return (error);
934 		mtx_lock(&np->n_mtx);
935 		uio->uio_offset = np->n_size;
936 		mtx_unlock(&np->n_mtx);
937 	}
938 
939 	if (uio->uio_offset < 0)
940 		return (EINVAL);
941 	tmp_off = uio->uio_offset + uio->uio_resid;
942 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
943 		return (EFBIG);
944 	if (uio->uio_resid == 0)
945 		return (0);
946 
947 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
948 		return nfs_directio_write(vp, uio, cred, ioflag);
949 
950 	/*
951 	 * Maybe this should be above the vnode op call, but so long as
952 	 * file servers have no limits, i don't think it matters
953 	 */
954 	if (vn_rlimit_fsize(vp, uio, td))
955 		return (EFBIG);
956 
957 	biosize = vp->v_bufobj.bo_bsize;
958 	/*
959 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
960 	 * would exceed the local maximum per-file write commit size when
961 	 * combined with those, we must decide whether to flush,
962 	 * go synchronous, or return error.  We don't bother checking
963 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
964 	 * no point optimizing for something that really won't ever happen.
965 	 */
966 	if (!(ioflag & IO_SYNC)) {
967 		int nflag;
968 
969 		mtx_lock(&np->n_mtx);
970 		nflag = np->n_flag;
971 		mtx_unlock(&np->n_mtx);
972 		int needrestart = 0;
973 		if (nmp->nm_wcommitsize < uio->uio_resid) {
974 			/*
975 			 * If this request could not possibly be completed
976 			 * without exceeding the maximum outstanding write
977 			 * commit size, see if we can convert it into a
978 			 * synchronous write operation.
979 			 */
980 			if (ioflag & IO_NDELAY)
981 				return (EAGAIN);
982 			ioflag |= IO_SYNC;
983 			if (nflag & NMODIFIED)
984 				needrestart = 1;
985 		} else if (nflag & NMODIFIED) {
986 			int wouldcommit = 0;
987 			BO_LOCK(&vp->v_bufobj);
988 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
989 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
990 				    b_bobufs) {
991 					if (bp->b_flags & B_NEEDCOMMIT)
992 						wouldcommit += bp->b_bcount;
993 				}
994 			}
995 			BO_UNLOCK(&vp->v_bufobj);
996 			/*
997 			 * Since we're not operating synchronously and
998 			 * bypassing the buffer cache, we are in a commit
999 			 * and holding all of these buffers whether
1000 			 * transmitted or not.  If not limited, this
1001 			 * will lead to the buffer cache deadlocking,
1002 			 * as no one else can flush our uncommitted buffers.
1003 			 */
1004 			wouldcommit += uio->uio_resid;
1005 			/*
1006 			 * If we would initially exceed the maximum
1007 			 * outstanding write commit size, flush and restart.
1008 			 */
1009 			if (wouldcommit > nmp->nm_wcommitsize)
1010 				needrestart = 1;
1011 		}
1012 		if (needrestart)
1013 			goto flush_and_restart;
1014 	}
1015 
1016 	do {
1017 		NFSINCRGLOBAL(newnfsstats.biocache_writes);
1018 		lbn = uio->uio_offset / biosize;
1019 		on = uio->uio_offset - (lbn * biosize);
1020 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1021 again:
1022 		/*
1023 		 * Handle direct append and file extension cases, calculate
1024 		 * unaligned buffer size.
1025 		 */
1026 		mtx_lock(&np->n_mtx);
1027 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1028 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1029 			noncontig_write = 1;
1030 		else
1031 			noncontig_write = 0;
1032 		if ((uio->uio_offset == np->n_size ||
1033 		    (noncontig_write != 0 &&
1034 		    lbn == (np->n_size / biosize) &&
1035 		    uio->uio_offset + n > np->n_size)) && n) {
1036 			mtx_unlock(&np->n_mtx);
1037 			/*
1038 			 * Get the buffer (in its pre-append state to maintain
1039 			 * B_CACHE if it was previously set).  Resize the
1040 			 * nfsnode after we have locked the buffer to prevent
1041 			 * readers from reading garbage.
1042 			 */
1043 			obcount = np->n_size - (lbn * biosize);
1044 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1045 
1046 			if (bp != NULL) {
1047 				long save;
1048 
1049 				mtx_lock(&np->n_mtx);
1050 				np->n_size = uio->uio_offset + n;
1051 				np->n_flag |= NMODIFIED;
1052 				vnode_pager_setsize(vp, np->n_size);
1053 				mtx_unlock(&np->n_mtx);
1054 
1055 				save = bp->b_flags & B_CACHE;
1056 				bcount = on + n;
1057 				allocbuf(bp, bcount);
1058 				bp->b_flags |= save;
1059 				if (noncontig_write != 0 && on > obcount)
1060 					vfs_bio_bzero_buf(bp, obcount, on -
1061 					    obcount);
1062 			}
1063 		} else {
1064 			/*
1065 			 * Obtain the locked cache block first, and then
1066 			 * adjust the file's size as appropriate.
1067 			 */
1068 			bcount = on + n;
1069 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1070 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1071 					bcount = biosize;
1072 				else
1073 					bcount = np->n_size - (off_t)lbn * biosize;
1074 			}
1075 			mtx_unlock(&np->n_mtx);
1076 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1077 			mtx_lock(&np->n_mtx);
1078 			if (uio->uio_offset + n > np->n_size) {
1079 				np->n_size = uio->uio_offset + n;
1080 				np->n_flag |= NMODIFIED;
1081 				vnode_pager_setsize(vp, np->n_size);
1082 			}
1083 			mtx_unlock(&np->n_mtx);
1084 		}
1085 
1086 		if (!bp) {
1087 			error = newnfs_sigintr(nmp, td);
1088 			if (!error)
1089 				error = EINTR;
1090 			break;
1091 		}
1092 
1093 		/*
1094 		 * Issue a READ if B_CACHE is not set.  In special-append
1095 		 * mode, B_CACHE is based on the buffer prior to the write
1096 		 * op and is typically set, avoiding the read.  If a read
1097 		 * is required in special append mode, the server will
1098 		 * probably send us a short-read since we extended the file
1099 		 * on our end, resulting in b_resid == 0 and, thusly,
1100 		 * B_CACHE getting set.
1101 		 *
1102 		 * We can also avoid issuing the read if the write covers
1103 		 * the entire buffer.  We have to make sure the buffer state
1104 		 * is reasonable in this case since we will not be initiating
1105 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1106 		 * more information.
1107 		 *
1108 		 * B_CACHE may also be set due to the buffer being cached
1109 		 * normally.
1110 		 */
1111 
1112 		bp_cached = 1;
1113 		if (on == 0 && n == bcount) {
1114 			if ((bp->b_flags & B_CACHE) == 0)
1115 				bp_cached = 0;
1116 			bp->b_flags |= B_CACHE;
1117 			bp->b_flags &= ~B_INVAL;
1118 			bp->b_ioflags &= ~BIO_ERROR;
1119 		}
1120 
1121 		if ((bp->b_flags & B_CACHE) == 0) {
1122 			bp->b_iocmd = BIO_READ;
1123 			vfs_busy_pages(bp, 0);
1124 			error = ncl_doio(vp, bp, cred, td, 0);
1125 			if (error) {
1126 				brelse(bp);
1127 				break;
1128 			}
1129 		}
1130 		if (bp->b_wcred == NOCRED)
1131 			bp->b_wcred = crhold(cred);
1132 		mtx_lock(&np->n_mtx);
1133 		np->n_flag |= NMODIFIED;
1134 		mtx_unlock(&np->n_mtx);
1135 
1136 		/*
1137 		 * If dirtyend exceeds file size, chop it down.  This should
1138 		 * not normally occur but there is an append race where it
1139 		 * might occur XXX, so we log it.
1140 		 *
1141 		 * If the chopping creates a reverse-indexed or degenerate
1142 		 * situation with dirtyoff/end, we 0 both of them.
1143 		 */
1144 
1145 		if (bp->b_dirtyend > bcount) {
1146 			ncl_printf("NFS append race @%lx:%d\n",
1147 			    (long)bp->b_blkno * DEV_BSIZE,
1148 			    bp->b_dirtyend - bcount);
1149 			bp->b_dirtyend = bcount;
1150 		}
1151 
1152 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1153 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1154 
1155 		/*
1156 		 * If the new write will leave a contiguous dirty
1157 		 * area, just update the b_dirtyoff and b_dirtyend,
1158 		 * otherwise force a write rpc of the old dirty area.
1159 		 *
1160 		 * If there has been a file lock applied to this file
1161 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1162 		 * While it is possible to merge discontiguous writes due to
1163 		 * our having a B_CACHE buffer ( and thus valid read data
1164 		 * for the hole), we don't because it could lead to
1165 		 * significant cache coherency problems with multiple clients,
1166 		 * especially if locking is implemented later on.
1167 		 *
1168 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1169 		 * not been file locking done on this file:
1170 		 * Relax coherency a bit for the sake of performance and
1171 		 * expand the current dirty region to contain the new
1172 		 * write even if it means we mark some non-dirty data as
1173 		 * dirty.
1174 		 */
1175 
1176 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1177 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1178 			if (bwrite(bp) == EINTR) {
1179 				error = EINTR;
1180 				break;
1181 			}
1182 			goto again;
1183 		}
1184 
1185 		local_resid = uio->uio_resid;
1186 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1187 
1188 		if (error != 0 && !bp_cached) {
1189 			/*
1190 			 * This block has no other content then what
1191 			 * possibly was written by the faulty uiomove.
1192 			 * Release it, forgetting the data pages, to
1193 			 * prevent the leak of uninitialized data to
1194 			 * usermode.
1195 			 */
1196 			bp->b_ioflags |= BIO_ERROR;
1197 			brelse(bp);
1198 			uio->uio_offset -= local_resid - uio->uio_resid;
1199 			uio->uio_resid = local_resid;
1200 			break;
1201 		}
1202 
1203 		/*
1204 		 * Since this block is being modified, it must be written
1205 		 * again and not just committed.  Since write clustering does
1206 		 * not work for the stage 1 data write, only the stage 2
1207 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1208 		 */
1209 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1210 
1211 		/*
1212 		 * Get the partial update on the progress made from
1213 		 * uiomove, if an error occured.
1214 		 */
1215 		if (error != 0)
1216 			n = local_resid - uio->uio_resid;
1217 
1218 		/*
1219 		 * Only update dirtyoff/dirtyend if not a degenerate
1220 		 * condition.
1221 		 */
1222 		if (n > 0) {
1223 			if (bp->b_dirtyend > 0) {
1224 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1225 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1226 			} else {
1227 				bp->b_dirtyoff = on;
1228 				bp->b_dirtyend = on + n;
1229 			}
1230 			vfs_bio_set_valid(bp, on, n);
1231 		}
1232 
1233 		/*
1234 		 * If IO_SYNC do bwrite().
1235 		 *
1236 		 * IO_INVAL appears to be unused.  The idea appears to be
1237 		 * to turn off caching in this case.  Very odd.  XXX
1238 		 */
1239 		if ((ioflag & IO_SYNC)) {
1240 			if (ioflag & IO_INVAL)
1241 				bp->b_flags |= B_NOCACHE;
1242 			error1 = bwrite(bp);
1243 			if (error1 != 0) {
1244 				if (error == 0)
1245 					error = error1;
1246 				break;
1247 			}
1248 		} else if ((n + on) == biosize) {
1249 			bp->b_flags |= B_ASYNC;
1250 			(void) ncl_writebp(bp, 0, NULL);
1251 		} else {
1252 			bdwrite(bp);
1253 		}
1254 
1255 		if (error != 0)
1256 			break;
1257 	} while (uio->uio_resid > 0 && n > 0);
1258 
1259 	if (error != 0) {
1260 		if (ioflag & IO_UNIT) {
1261 			VATTR_NULL(&vattr);
1262 			vattr.va_size = orig_size;
1263 			/* IO_SYNC is handled implicitely */
1264 			(void)VOP_SETATTR(vp, &vattr, cred);
1265 			uio->uio_offset -= orig_resid - uio->uio_resid;
1266 			uio->uio_resid = orig_resid;
1267 		}
1268 	}
1269 
1270 	return (error);
1271 }
1272 
1273 /*
1274  * Get an nfs cache block.
1275  *
1276  * Allocate a new one if the block isn't currently in the cache
1277  * and return the block marked busy. If the calling process is
1278  * interrupted by a signal for an interruptible mount point, return
1279  * NULL.
1280  *
1281  * The caller must carefully deal with the possible B_INVAL state of
1282  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1283  * indirectly), so synchronous reads can be issued without worrying about
1284  * the B_INVAL state.  We have to be a little more careful when dealing
1285  * with writes (see comments in nfs_write()) when extending a file past
1286  * its EOF.
1287  */
1288 static struct buf *
1289 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1290 {
1291 	struct buf *bp;
1292 	struct mount *mp;
1293 	struct nfsmount *nmp;
1294 
1295 	mp = vp->v_mount;
1296 	nmp = VFSTONFS(mp);
1297 
1298 	if (nmp->nm_flag & NFSMNT_INT) {
1299 		sigset_t oldset;
1300 
1301 		newnfs_set_sigmask(td, &oldset);
1302 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1303 		newnfs_restore_sigmask(td, &oldset);
1304 		while (bp == NULL) {
1305 			if (newnfs_sigintr(nmp, td))
1306 				return (NULL);
1307 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1308 		}
1309 	} else {
1310 		bp = getblk(vp, bn, size, 0, 0, 0);
1311 	}
1312 
1313 	if (vp->v_type == VREG)
1314 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1315 	return (bp);
1316 }
1317 
1318 /*
1319  * Flush and invalidate all dirty buffers. If another process is already
1320  * doing the flush, just wait for completion.
1321  */
1322 int
1323 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1324 {
1325 	struct nfsnode *np = VTONFS(vp);
1326 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1327 	int error = 0, slpflag, slptimeo;
1328 	int old_lock = 0;
1329 
1330 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1331 
1332 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1333 		intrflg = 0;
1334 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1335 		intrflg = 1;
1336 	if (intrflg) {
1337 		slpflag = PCATCH;
1338 		slptimeo = 2 * hz;
1339 	} else {
1340 		slpflag = 0;
1341 		slptimeo = 0;
1342 	}
1343 
1344 	old_lock = ncl_upgrade_vnlock(vp);
1345 	if (vp->v_iflag & VI_DOOMED) {
1346 		/*
1347 		 * Since vgonel() uses the generic vinvalbuf() to flush
1348 		 * dirty buffers and it does not call this function, it
1349 		 * is safe to just return OK when VI_DOOMED is set.
1350 		 */
1351 		ncl_downgrade_vnlock(vp, old_lock);
1352 		return (0);
1353 	}
1354 
1355 	/*
1356 	 * Now, flush as required.
1357 	 */
1358 	if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1359 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1360 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1361 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1362 		/*
1363 		 * If the page clean was interrupted, fail the invalidation.
1364 		 * Not doing so, we run the risk of losing dirty pages in the
1365 		 * vinvalbuf() call below.
1366 		 */
1367 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1368 			goto out;
1369 	}
1370 
1371 	error = vinvalbuf(vp, flags, slpflag, 0);
1372 	while (error) {
1373 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1374 			goto out;
1375 		error = vinvalbuf(vp, flags, 0, slptimeo);
1376 	}
1377 	if (NFSHASPNFS(nmp)) {
1378 		nfscl_layoutcommit(vp, td);
1379 		/*
1380 		 * Invalidate the attribute cache, since writes to a DS
1381 		 * won't update the size attribute.
1382 		 */
1383 		mtx_lock(&np->n_mtx);
1384 		np->n_attrstamp = 0;
1385 	} else
1386 		mtx_lock(&np->n_mtx);
1387 	if (np->n_directio_asyncwr == 0)
1388 		np->n_flag &= ~NMODIFIED;
1389 	mtx_unlock(&np->n_mtx);
1390 out:
1391 	ncl_downgrade_vnlock(vp, old_lock);
1392 	return error;
1393 }
1394 
1395 /*
1396  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1397  * This is mainly to avoid queueing async I/O requests when the nfsiods
1398  * are all hung on a dead server.
1399  *
1400  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1401  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1402  */
1403 int
1404 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1405 {
1406 	int iod;
1407 	int gotiod;
1408 	int slpflag = 0;
1409 	int slptimeo = 0;
1410 	int error, error2;
1411 
1412 	/*
1413 	 * Commits are usually short and sweet so lets save some cpu and
1414 	 * leave the async daemons for more important rpc's (such as reads
1415 	 * and writes).
1416 	 *
1417 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1418 	 * in the directory in order to update attributes. This can deadlock
1419 	 * with another thread that is waiting for async I/O to be done by
1420 	 * an nfsiod thread while holding a lock on one of these vnodes.
1421 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1422 	 * perform Readdirplus RPCs.
1423 	 */
1424 	mtx_lock(&ncl_iod_mutex);
1425 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1426 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1427 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1428 		mtx_unlock(&ncl_iod_mutex);
1429 		return(EIO);
1430 	}
1431 again:
1432 	if (nmp->nm_flag & NFSMNT_INT)
1433 		slpflag = PCATCH;
1434 	gotiod = FALSE;
1435 
1436 	/*
1437 	 * Find a free iod to process this request.
1438 	 */
1439 	for (iod = 0; iod < ncl_numasync; iod++)
1440 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1441 			gotiod = TRUE;
1442 			break;
1443 		}
1444 
1445 	/*
1446 	 * Try to create one if none are free.
1447 	 */
1448 	if (!gotiod)
1449 		ncl_nfsiodnew();
1450 	else {
1451 		/*
1452 		 * Found one, so wake it up and tell it which
1453 		 * mount to process.
1454 		 */
1455 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1456 		    iod, nmp));
1457 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1458 		ncl_iodmount[iod] = nmp;
1459 		nmp->nm_bufqiods++;
1460 		wakeup(&ncl_iodwant[iod]);
1461 	}
1462 
1463 	/*
1464 	 * If none are free, we may already have an iod working on this mount
1465 	 * point.  If so, it will process our request.
1466 	 */
1467 	if (!gotiod) {
1468 		if (nmp->nm_bufqiods > 0) {
1469 			NFS_DPF(ASYNCIO,
1470 				("ncl_asyncio: %d iods are already processing mount %p\n",
1471 				 nmp->nm_bufqiods, nmp));
1472 			gotiod = TRUE;
1473 		}
1474 	}
1475 
1476 	/*
1477 	 * If we have an iod which can process the request, then queue
1478 	 * the buffer.
1479 	 */
1480 	if (gotiod) {
1481 		/*
1482 		 * Ensure that the queue never grows too large.  We still want
1483 		 * to asynchronize so we block rather then return EIO.
1484 		 */
1485 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1486 			NFS_DPF(ASYNCIO,
1487 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1488 			nmp->nm_bufqwant = TRUE;
1489 			error = newnfs_msleep(td, &nmp->nm_bufq,
1490 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1491 			   slptimeo);
1492 			if (error) {
1493 				error2 = newnfs_sigintr(nmp, td);
1494 				if (error2) {
1495 					mtx_unlock(&ncl_iod_mutex);
1496 					return (error2);
1497 				}
1498 				if (slpflag == PCATCH) {
1499 					slpflag = 0;
1500 					slptimeo = 2 * hz;
1501 				}
1502 			}
1503 			/*
1504 			 * We might have lost our iod while sleeping,
1505 			 * so check and loop if nescessary.
1506 			 */
1507 			goto again;
1508 		}
1509 
1510 		/* We might have lost our nfsiod */
1511 		if (nmp->nm_bufqiods == 0) {
1512 			NFS_DPF(ASYNCIO,
1513 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1514 			goto again;
1515 		}
1516 
1517 		if (bp->b_iocmd == BIO_READ) {
1518 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1519 				bp->b_rcred = crhold(cred);
1520 		} else {
1521 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1522 				bp->b_wcred = crhold(cred);
1523 		}
1524 
1525 		if (bp->b_flags & B_REMFREE)
1526 			bremfreef(bp);
1527 		BUF_KERNPROC(bp);
1528 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1529 		nmp->nm_bufqlen++;
1530 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1531 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1532 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1533 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1534 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1535 		}
1536 		mtx_unlock(&ncl_iod_mutex);
1537 		return (0);
1538 	}
1539 
1540 	mtx_unlock(&ncl_iod_mutex);
1541 
1542 	/*
1543 	 * All the iods are busy on other mounts, so return EIO to
1544 	 * force the caller to process the i/o synchronously.
1545 	 */
1546 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1547 	return (EIO);
1548 }
1549 
1550 void
1551 ncl_doio_directwrite(struct buf *bp)
1552 {
1553 	int iomode, must_commit;
1554 	struct uio *uiop = (struct uio *)bp->b_caller1;
1555 	char *iov_base = uiop->uio_iov->iov_base;
1556 
1557 	iomode = NFSWRITE_FILESYNC;
1558 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1559 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1560 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1561 	free(iov_base, M_NFSDIRECTIO);
1562 	free(uiop->uio_iov, M_NFSDIRECTIO);
1563 	free(uiop, M_NFSDIRECTIO);
1564 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1565 		struct nfsnode *np = VTONFS(bp->b_vp);
1566 		mtx_lock(&np->n_mtx);
1567 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1568 			/*
1569 			 * Invalidate the attribute cache, since writes to a DS
1570 			 * won't update the size attribute.
1571 			 */
1572 			np->n_attrstamp = 0;
1573 		}
1574 		np->n_directio_asyncwr--;
1575 		if (np->n_directio_asyncwr == 0) {
1576 			np->n_flag &= ~NMODIFIED;
1577 			if ((np->n_flag & NFSYNCWAIT)) {
1578 				np->n_flag &= ~NFSYNCWAIT;
1579 				wakeup((caddr_t)&np->n_directio_asyncwr);
1580 			}
1581 		}
1582 		mtx_unlock(&np->n_mtx);
1583 	}
1584 	bp->b_vp = NULL;
1585 	relpbuf(bp, &ncl_pbuf_freecnt);
1586 }
1587 
1588 /*
1589  * Do an I/O operation to/from a cache block. This may be called
1590  * synchronously or from an nfsiod.
1591  */
1592 int
1593 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1594     int called_from_strategy)
1595 {
1596 	struct uio *uiop;
1597 	struct nfsnode *np;
1598 	struct nfsmount *nmp;
1599 	int error = 0, iomode, must_commit = 0;
1600 	struct uio uio;
1601 	struct iovec io;
1602 	struct proc *p = td ? td->td_proc : NULL;
1603 	uint8_t	iocmd;
1604 
1605 	np = VTONFS(vp);
1606 	nmp = VFSTONFS(vp->v_mount);
1607 	uiop = &uio;
1608 	uiop->uio_iov = &io;
1609 	uiop->uio_iovcnt = 1;
1610 	uiop->uio_segflg = UIO_SYSSPACE;
1611 	uiop->uio_td = td;
1612 
1613 	/*
1614 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1615 	 * do this here so we do not have to do it in all the code that
1616 	 * calls us.
1617 	 */
1618 	bp->b_flags &= ~B_INVAL;
1619 	bp->b_ioflags &= ~BIO_ERROR;
1620 
1621 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1622 	iocmd = bp->b_iocmd;
1623 	if (iocmd == BIO_READ) {
1624 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1625 	    io.iov_base = bp->b_data;
1626 	    uiop->uio_rw = UIO_READ;
1627 
1628 	    switch (vp->v_type) {
1629 	    case VREG:
1630 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1631 		NFSINCRGLOBAL(newnfsstats.read_bios);
1632 		error = ncl_readrpc(vp, uiop, cr);
1633 
1634 		if (!error) {
1635 		    if (uiop->uio_resid) {
1636 			/*
1637 			 * If we had a short read with no error, we must have
1638 			 * hit a file hole.  We should zero-fill the remainder.
1639 			 * This can also occur if the server hits the file EOF.
1640 			 *
1641 			 * Holes used to be able to occur due to pending
1642 			 * writes, but that is not possible any longer.
1643 			 */
1644 			int nread = bp->b_bcount - uiop->uio_resid;
1645 			ssize_t left = uiop->uio_resid;
1646 
1647 			if (left > 0)
1648 				bzero((char *)bp->b_data + nread, left);
1649 			uiop->uio_resid = 0;
1650 		    }
1651 		}
1652 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1653 		if (p && (vp->v_vflag & VV_TEXT)) {
1654 			mtx_lock(&np->n_mtx);
1655 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1656 				mtx_unlock(&np->n_mtx);
1657 				PROC_LOCK(p);
1658 				killproc(p, "text file modification");
1659 				PROC_UNLOCK(p);
1660 			} else
1661 				mtx_unlock(&np->n_mtx);
1662 		}
1663 		break;
1664 	    case VLNK:
1665 		uiop->uio_offset = (off_t)0;
1666 		NFSINCRGLOBAL(newnfsstats.readlink_bios);
1667 		error = ncl_readlinkrpc(vp, uiop, cr);
1668 		break;
1669 	    case VDIR:
1670 		NFSINCRGLOBAL(newnfsstats.readdir_bios);
1671 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1672 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1673 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1674 			if (error == NFSERR_NOTSUPP)
1675 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1676 		}
1677 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1678 			error = ncl_readdirrpc(vp, uiop, cr, td);
1679 		/*
1680 		 * end-of-directory sets B_INVAL but does not generate an
1681 		 * error.
1682 		 */
1683 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1684 			bp->b_flags |= B_INVAL;
1685 		break;
1686 	    default:
1687 		ncl_printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1688 		break;
1689 	    };
1690 	    if (error) {
1691 		bp->b_ioflags |= BIO_ERROR;
1692 		bp->b_error = error;
1693 	    }
1694 	} else {
1695 	    /*
1696 	     * If we only need to commit, try to commit
1697 	     */
1698 	    if (bp->b_flags & B_NEEDCOMMIT) {
1699 		    int retv;
1700 		    off_t off;
1701 
1702 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1703 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1704 			bp->b_wcred, td);
1705 		    if (retv == 0) {
1706 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1707 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1708 			    bp->b_resid = 0;
1709 			    bufdone(bp);
1710 			    return (0);
1711 		    }
1712 		    if (retv == NFSERR_STALEWRITEVERF) {
1713 			    ncl_clearcommit(vp->v_mount);
1714 		    }
1715 	    }
1716 
1717 	    /*
1718 	     * Setup for actual write
1719 	     */
1720 	    mtx_lock(&np->n_mtx);
1721 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1722 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1723 	    mtx_unlock(&np->n_mtx);
1724 
1725 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1726 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1727 		    - bp->b_dirtyoff;
1728 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1729 		    + bp->b_dirtyoff;
1730 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1731 		uiop->uio_rw = UIO_WRITE;
1732 		NFSINCRGLOBAL(newnfsstats.write_bios);
1733 
1734 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1735 		    iomode = NFSWRITE_UNSTABLE;
1736 		else
1737 		    iomode = NFSWRITE_FILESYNC;
1738 
1739 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1740 		    called_from_strategy);
1741 
1742 		/*
1743 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1744 		 * to cluster the buffers needing commit.  This will allow
1745 		 * the system to submit a single commit rpc for the whole
1746 		 * cluster.  We can do this even if the buffer is not 100%
1747 		 * dirty (relative to the NFS blocksize), so we optimize the
1748 		 * append-to-file-case.
1749 		 *
1750 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1751 		 * cleared because write clustering only works for commit
1752 		 * rpc's, not for the data portion of the write).
1753 		 */
1754 
1755 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1756 		    bp->b_flags |= B_NEEDCOMMIT;
1757 		    if (bp->b_dirtyoff == 0
1758 			&& bp->b_dirtyend == bp->b_bcount)
1759 			bp->b_flags |= B_CLUSTEROK;
1760 		} else {
1761 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1762 		}
1763 
1764 		/*
1765 		 * For an interrupted write, the buffer is still valid
1766 		 * and the write hasn't been pushed to the server yet,
1767 		 * so we can't set BIO_ERROR and report the interruption
1768 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1769 		 * is not relevant, so the rpc attempt is essentially
1770 		 * a noop.  For the case of a V3 write rpc not being
1771 		 * committed to stable storage, the block is still
1772 		 * dirty and requires either a commit rpc or another
1773 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1774 		 * the block is reused. This is indicated by setting
1775 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1776 		 *
1777 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1778 		 * write error and is handled as above, except that
1779 		 * B_EINTR isn't set. One cause of this is a stale stateid
1780 		 * error for the RPC that indicates recovery is required,
1781 		 * when called with called_from_strategy != 0.
1782 		 *
1783 		 * If the buffer is marked B_PAGING, it does not reside on
1784 		 * the vp's paging queues so we cannot call bdirty().  The
1785 		 * bp in this case is not an NFS cache block so we should
1786 		 * be safe. XXX
1787 		 *
1788 		 * The logic below breaks up errors into recoverable and
1789 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1790 		 * and keep the buffer around for potential write retries.
1791 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1792 		 * and save the error in the nfsnode. This is less than ideal
1793 		 * but necessary. Keeping such buffers around could potentially
1794 		 * cause buffer exhaustion eventually (they can never be written
1795 		 * out, so will get constantly be re-dirtied). It also causes
1796 		 * all sorts of vfs panics. For non-recoverable write errors,
1797 		 * also invalidate the attrcache, so we'll be forced to go over
1798 		 * the wire for this object, returning an error to user on next
1799 		 * call (most of the time).
1800 		 */
1801 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1802 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1803 			int s;
1804 
1805 			s = splbio();
1806 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1807 			if ((bp->b_flags & B_PAGING) == 0) {
1808 			    bdirty(bp);
1809 			    bp->b_flags &= ~B_DONE;
1810 			}
1811 			if ((error == EINTR || error == ETIMEDOUT) &&
1812 			    (bp->b_flags & B_ASYNC) == 0)
1813 			    bp->b_flags |= B_EINTR;
1814 			splx(s);
1815 		} else {
1816 		    if (error) {
1817 			bp->b_ioflags |= BIO_ERROR;
1818 			bp->b_flags |= B_INVAL;
1819 			bp->b_error = np->n_error = error;
1820 			mtx_lock(&np->n_mtx);
1821 			np->n_flag |= NWRITEERR;
1822 			np->n_attrstamp = 0;
1823 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1824 			mtx_unlock(&np->n_mtx);
1825 		    }
1826 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1827 		}
1828 	    } else {
1829 		bp->b_resid = 0;
1830 		bufdone(bp);
1831 		return (0);
1832 	    }
1833 	}
1834 	bp->b_resid = uiop->uio_resid;
1835 	if (must_commit)
1836 	    ncl_clearcommit(vp->v_mount);
1837 	bufdone(bp);
1838 	return (error);
1839 }
1840 
1841 /*
1842  * Used to aid in handling ftruncate() operations on the NFS client side.
1843  * Truncation creates a number of special problems for NFS.  We have to
1844  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1845  * we have to properly handle VM pages or (potentially dirty) buffers
1846  * that straddle the truncation point.
1847  */
1848 
1849 int
1850 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1851 {
1852 	struct nfsnode *np = VTONFS(vp);
1853 	u_quad_t tsize;
1854 	int biosize = vp->v_bufobj.bo_bsize;
1855 	int error = 0;
1856 
1857 	mtx_lock(&np->n_mtx);
1858 	tsize = np->n_size;
1859 	np->n_size = nsize;
1860 	mtx_unlock(&np->n_mtx);
1861 
1862 	if (nsize < tsize) {
1863 		struct buf *bp;
1864 		daddr_t lbn;
1865 		int bufsize;
1866 
1867 		/*
1868 		 * vtruncbuf() doesn't get the buffer overlapping the
1869 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1870 		 * buffer that now needs to be truncated.
1871 		 */
1872 		error = vtruncbuf(vp, cred, nsize, biosize);
1873 		lbn = nsize / biosize;
1874 		bufsize = nsize - (lbn * biosize);
1875 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1876 		if (!bp)
1877 			return EINTR;
1878 		if (bp->b_dirtyoff > bp->b_bcount)
1879 			bp->b_dirtyoff = bp->b_bcount;
1880 		if (bp->b_dirtyend > bp->b_bcount)
1881 			bp->b_dirtyend = bp->b_bcount;
1882 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1883 		brelse(bp);
1884 	} else {
1885 		vnode_pager_setsize(vp, nsize);
1886 	}
1887 	return(error);
1888 }
1889 
1890