xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision dda5b39711dab90ae1c5624bdd6ff7453177df31)
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 	VM_OBJECT_WLOCK(object);
137 	if (pages[ap->a_reqpage]->valid != 0) {
138 		for (i = 0; i < npages; ++i) {
139 			if (i != ap->a_reqpage) {
140 				vm_page_lock(pages[i]);
141 				vm_page_free(pages[i]);
142 				vm_page_unlock(pages[i]);
143 			}
144 		}
145 		VM_OBJECT_WUNLOCK(object);
146 		return (0);
147 	}
148 	VM_OBJECT_WUNLOCK(object);
149 
150 	/*
151 	 * We use only the kva address for the buffer, but this is extremely
152 	 * convienient and fast.
153 	 */
154 	bp = getpbuf(&ncl_pbuf_freecnt);
155 
156 	kva = (vm_offset_t) bp->b_data;
157 	pmap_qenter(kva, pages, npages);
158 	PCPU_INC(cnt.v_vnodein);
159 	PCPU_ADD(cnt.v_vnodepgsin, npages);
160 
161 	iov.iov_base = (caddr_t) kva;
162 	iov.iov_len = count;
163 	uio.uio_iov = &iov;
164 	uio.uio_iovcnt = 1;
165 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
166 	uio.uio_resid = count;
167 	uio.uio_segflg = UIO_SYSSPACE;
168 	uio.uio_rw = UIO_READ;
169 	uio.uio_td = td;
170 
171 	error = ncl_readrpc(vp, &uio, cred);
172 	pmap_qremove(kva, npages);
173 
174 	relpbuf(bp, &ncl_pbuf_freecnt);
175 
176 	if (error && (uio.uio_resid == count)) {
177 		ncl_printf("nfs_getpages: error %d\n", error);
178 		VM_OBJECT_WLOCK(object);
179 		for (i = 0; i < npages; ++i) {
180 			if (i != ap->a_reqpage) {
181 				vm_page_lock(pages[i]);
182 				vm_page_free(pages[i]);
183 				vm_page_unlock(pages[i]);
184 			}
185 		}
186 		VM_OBJECT_WUNLOCK(object);
187 		return (VM_PAGER_ERROR);
188 	}
189 
190 	/*
191 	 * Calculate the number of bytes read and validate only that number
192 	 * of bytes.  Note that due to pending writes, size may be 0.  This
193 	 * does not mean that the remaining data is invalid!
194 	 */
195 
196 	size = count - uio.uio_resid;
197 	VM_OBJECT_WLOCK(object);
198 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
199 		vm_page_t m;
200 		nextoff = toff + PAGE_SIZE;
201 		m = pages[i];
202 
203 		if (nextoff <= size) {
204 			/*
205 			 * Read operation filled an entire page
206 			 */
207 			m->valid = VM_PAGE_BITS_ALL;
208 			KASSERT(m->dirty == 0,
209 			    ("nfs_getpages: page %p is dirty", m));
210 		} else if (size > toff) {
211 			/*
212 			 * Read operation filled a partial page.
213 			 */
214 			m->valid = 0;
215 			vm_page_set_valid_range(m, 0, size - toff);
216 			KASSERT(m->dirty == 0,
217 			    ("nfs_getpages: page %p is dirty", m));
218 		} else {
219 			/*
220 			 * Read operation was short.  If no error
221 			 * occured we may have hit a zero-fill
222 			 * section.  We leave valid set to 0, and page
223 			 * is freed by vm_page_readahead_finish() if
224 			 * its index is not equal to requested, or
225 			 * page is zeroed and set valid by
226 			 * vm_pager_get_pages() for requested page.
227 			 */
228 			;
229 		}
230 		if (i != ap->a_reqpage)
231 			vm_page_readahead_finish(m);
232 	}
233 	VM_OBJECT_WUNLOCK(object);
234 	return (0);
235 }
236 
237 /*
238  * Vnode op for VM putpages.
239  */
240 int
241 ncl_putpages(struct vop_putpages_args *ap)
242 {
243 	struct uio uio;
244 	struct iovec iov;
245 	vm_offset_t kva;
246 	struct buf *bp;
247 	int iomode, must_commit, i, error, npages, count;
248 	off_t offset;
249 	int *rtvals;
250 	struct vnode *vp;
251 	struct thread *td;
252 	struct ucred *cred;
253 	struct nfsmount *nmp;
254 	struct nfsnode *np;
255 	vm_page_t *pages;
256 
257 	vp = ap->a_vp;
258 	np = VTONFS(vp);
259 	td = curthread;				/* XXX */
260 	/* Set the cred to n_writecred for the write rpcs. */
261 	if (np->n_writecred != NULL)
262 		cred = crhold(np->n_writecred);
263 	else
264 		cred = crhold(curthread->td_ucred);	/* XXX */
265 	nmp = VFSTONFS(vp->v_mount);
266 	pages = ap->a_m;
267 	count = ap->a_count;
268 	rtvals = ap->a_rtvals;
269 	npages = btoc(count);
270 	offset = IDX_TO_OFF(pages[0]->pindex);
271 
272 	mtx_lock(&nmp->nm_mtx);
273 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
274 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
275 		mtx_unlock(&nmp->nm_mtx);
276 		(void)ncl_fsinfo(nmp, vp, cred, td);
277 	} else
278 		mtx_unlock(&nmp->nm_mtx);
279 
280 	mtx_lock(&np->n_mtx);
281 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
282 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
283 		mtx_unlock(&np->n_mtx);
284 		ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
285 		mtx_lock(&np->n_mtx);
286 	}
287 
288 	for (i = 0; i < npages; i++)
289 		rtvals[i] = VM_PAGER_ERROR;
290 
291 	/*
292 	 * When putting pages, do not extend file past EOF.
293 	 */
294 	if (offset + count > np->n_size) {
295 		count = np->n_size - offset;
296 		if (count < 0)
297 			count = 0;
298 	}
299 	mtx_unlock(&np->n_mtx);
300 
301 	/*
302 	 * We use only the kva address for the buffer, but this is extremely
303 	 * convienient and fast.
304 	 */
305 	bp = getpbuf(&ncl_pbuf_freecnt);
306 
307 	kva = (vm_offset_t) bp->b_data;
308 	pmap_qenter(kva, pages, npages);
309 	PCPU_INC(cnt.v_vnodeout);
310 	PCPU_ADD(cnt.v_vnodepgsout, count);
311 
312 	iov.iov_base = (caddr_t) kva;
313 	iov.iov_len = count;
314 	uio.uio_iov = &iov;
315 	uio.uio_iovcnt = 1;
316 	uio.uio_offset = offset;
317 	uio.uio_resid = count;
318 	uio.uio_segflg = UIO_SYSSPACE;
319 	uio.uio_rw = UIO_WRITE;
320 	uio.uio_td = td;
321 
322 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
323 	    iomode = NFSWRITE_UNSTABLE;
324 	else
325 	    iomode = NFSWRITE_FILESYNC;
326 
327 	error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
328 	crfree(cred);
329 
330 	pmap_qremove(kva, npages);
331 	relpbuf(bp, &ncl_pbuf_freecnt);
332 
333 	if (error == 0 || !nfs_keep_dirty_on_error) {
334 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
335 		if (must_commit)
336 			ncl_clearcommit(vp->v_mount);
337 	}
338 	return rtvals[0];
339 }
340 
341 /*
342  * For nfs, cache consistency can only be maintained approximately.
343  * Although RFC1094 does not specify the criteria, the following is
344  * believed to be compatible with the reference port.
345  * For nfs:
346  * If the file's modify time on the server has changed since the
347  * last read rpc or you have written to the file,
348  * you may have lost data cache consistency with the
349  * server, so flush all of the file's data out of the cache.
350  * Then force a getattr rpc to ensure that you have up to date
351  * attributes.
352  * NB: This implies that cache data can be read when up to
353  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
354  * attributes this could be forced by setting n_attrstamp to 0 before
355  * the VOP_GETATTR() call.
356  */
357 static inline int
358 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
359 {
360 	int error = 0;
361 	struct vattr vattr;
362 	struct nfsnode *np = VTONFS(vp);
363 	int old_lock;
364 
365 	/*
366 	 * Grab the exclusive lock before checking whether the cache is
367 	 * consistent.
368 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
369 	 * But for now, this suffices.
370 	 */
371 	old_lock = ncl_upgrade_vnlock(vp);
372 	if (vp->v_iflag & VI_DOOMED) {
373 		ncl_downgrade_vnlock(vp, old_lock);
374 		return (EBADF);
375 	}
376 
377 	mtx_lock(&np->n_mtx);
378 	if (np->n_flag & NMODIFIED) {
379 		mtx_unlock(&np->n_mtx);
380 		if (vp->v_type != VREG) {
381 			if (vp->v_type != VDIR)
382 				panic("nfs: bioread, not dir");
383 			ncl_invaldir(vp);
384 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
385 			if (error)
386 				goto out;
387 		}
388 		np->n_attrstamp = 0;
389 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
390 		error = VOP_GETATTR(vp, &vattr, cred);
391 		if (error)
392 			goto out;
393 		mtx_lock(&np->n_mtx);
394 		np->n_mtime = vattr.va_mtime;
395 		mtx_unlock(&np->n_mtx);
396 	} else {
397 		mtx_unlock(&np->n_mtx);
398 		error = VOP_GETATTR(vp, &vattr, cred);
399 		if (error)
400 			return (error);
401 		mtx_lock(&np->n_mtx);
402 		if ((np->n_flag & NSIZECHANGED)
403 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
404 			mtx_unlock(&np->n_mtx);
405 			if (vp->v_type == VDIR)
406 				ncl_invaldir(vp);
407 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
408 			if (error)
409 				goto out;
410 			mtx_lock(&np->n_mtx);
411 			np->n_mtime = vattr.va_mtime;
412 			np->n_flag &= ~NSIZECHANGED;
413 		}
414 		mtx_unlock(&np->n_mtx);
415 	}
416 out:
417 	ncl_downgrade_vnlock(vp, old_lock);
418 	return error;
419 }
420 
421 /*
422  * Vnode op for read using bio
423  */
424 int
425 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
426 {
427 	struct nfsnode *np = VTONFS(vp);
428 	int biosize, i;
429 	struct buf *bp, *rabp;
430 	struct thread *td;
431 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
432 	daddr_t lbn, rabn;
433 	int bcount;
434 	int seqcount;
435 	int nra, error = 0, n = 0, on = 0;
436 	off_t tmp_off;
437 
438 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
439 	if (uio->uio_resid == 0)
440 		return (0);
441 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
442 		return (EINVAL);
443 	td = uio->uio_td;
444 
445 	mtx_lock(&nmp->nm_mtx);
446 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
447 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
448 		mtx_unlock(&nmp->nm_mtx);
449 		(void)ncl_fsinfo(nmp, vp, cred, td);
450 		mtx_lock(&nmp->nm_mtx);
451 	}
452 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
453 		(void) newnfs_iosize(nmp);
454 
455 	tmp_off = uio->uio_offset + uio->uio_resid;
456 	if (vp->v_type != VDIR &&
457 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
458 		mtx_unlock(&nmp->nm_mtx);
459 		return (EFBIG);
460 	}
461 	mtx_unlock(&nmp->nm_mtx);
462 
463 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
464 		/* No caching/ no readaheads. Just read data into the user buffer */
465 		return ncl_readrpc(vp, uio, cred);
466 
467 	biosize = vp->v_bufobj.bo_bsize;
468 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
469 
470 	error = nfs_bioread_check_cons(vp, td, cred);
471 	if (error)
472 		return error;
473 
474 	do {
475 	    u_quad_t nsize;
476 
477 	    mtx_lock(&np->n_mtx);
478 	    nsize = np->n_size;
479 	    mtx_unlock(&np->n_mtx);
480 
481 	    switch (vp->v_type) {
482 	    case VREG:
483 		NFSINCRGLOBAL(newnfsstats.biocache_reads);
484 		lbn = uio->uio_offset / biosize;
485 		on = uio->uio_offset - (lbn * biosize);
486 
487 		/*
488 		 * Start the read ahead(s), as required.
489 		 */
490 		if (nmp->nm_readahead > 0) {
491 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
492 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
493 			rabn = lbn + 1 + nra;
494 			if (incore(&vp->v_bufobj, rabn) == NULL) {
495 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
496 			    if (!rabp) {
497 				error = newnfs_sigintr(nmp, td);
498 				return (error ? error : EINTR);
499 			    }
500 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
501 				rabp->b_flags |= B_ASYNC;
502 				rabp->b_iocmd = BIO_READ;
503 				vfs_busy_pages(rabp, 0);
504 				if (ncl_asyncio(nmp, rabp, cred, td)) {
505 				    rabp->b_flags |= B_INVAL;
506 				    rabp->b_ioflags |= BIO_ERROR;
507 				    vfs_unbusy_pages(rabp);
508 				    brelse(rabp);
509 				    break;
510 				}
511 			    } else {
512 				brelse(rabp);
513 			    }
514 			}
515 		    }
516 		}
517 
518 		/* Note that bcount is *not* DEV_BSIZE aligned. */
519 		bcount = biosize;
520 		if ((off_t)lbn * biosize >= nsize) {
521 			bcount = 0;
522 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
523 			bcount = nsize - (off_t)lbn * biosize;
524 		}
525 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
526 
527 		if (!bp) {
528 			error = newnfs_sigintr(nmp, td);
529 			return (error ? error : EINTR);
530 		}
531 
532 		/*
533 		 * If B_CACHE is not set, we must issue the read.  If this
534 		 * fails, we return an error.
535 		 */
536 
537 		if ((bp->b_flags & B_CACHE) == 0) {
538 		    bp->b_iocmd = BIO_READ;
539 		    vfs_busy_pages(bp, 0);
540 		    error = ncl_doio(vp, bp, cred, td, 0);
541 		    if (error) {
542 			brelse(bp);
543 			return (error);
544 		    }
545 		}
546 
547 		/*
548 		 * on is the offset into the current bp.  Figure out how many
549 		 * bytes we can copy out of the bp.  Note that bcount is
550 		 * NOT DEV_BSIZE aligned.
551 		 *
552 		 * Then figure out how many bytes we can copy into the uio.
553 		 */
554 
555 		n = 0;
556 		if (on < bcount)
557 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
558 		break;
559 	    case VLNK:
560 		NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
561 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
562 		if (!bp) {
563 			error = newnfs_sigintr(nmp, td);
564 			return (error ? error : EINTR);
565 		}
566 		if ((bp->b_flags & B_CACHE) == 0) {
567 		    bp->b_iocmd = BIO_READ;
568 		    vfs_busy_pages(bp, 0);
569 		    error = ncl_doio(vp, bp, cred, td, 0);
570 		    if (error) {
571 			bp->b_ioflags |= BIO_ERROR;
572 			brelse(bp);
573 			return (error);
574 		    }
575 		}
576 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
577 		on = 0;
578 		break;
579 	    case VDIR:
580 		NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
581 		if (np->n_direofoffset
582 		    && uio->uio_offset >= np->n_direofoffset) {
583 		    return (0);
584 		}
585 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
586 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
587 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
588 		if (!bp) {
589 		    error = newnfs_sigintr(nmp, td);
590 		    return (error ? error : EINTR);
591 		}
592 		if ((bp->b_flags & B_CACHE) == 0) {
593 		    bp->b_iocmd = BIO_READ;
594 		    vfs_busy_pages(bp, 0);
595 		    error = ncl_doio(vp, bp, cred, td, 0);
596 		    if (error) {
597 			    brelse(bp);
598 		    }
599 		    while (error == NFSERR_BAD_COOKIE) {
600 			ncl_invaldir(vp);
601 			error = ncl_vinvalbuf(vp, 0, td, 1);
602 			/*
603 			 * Yuck! The directory has been modified on the
604 			 * server. The only way to get the block is by
605 			 * reading from the beginning to get all the
606 			 * offset cookies.
607 			 *
608 			 * Leave the last bp intact unless there is an error.
609 			 * Loop back up to the while if the error is another
610 			 * NFSERR_BAD_COOKIE (double yuch!).
611 			 */
612 			for (i = 0; i <= lbn && !error; i++) {
613 			    if (np->n_direofoffset
614 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
615 				    return (0);
616 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
617 			    if (!bp) {
618 				error = newnfs_sigintr(nmp, td);
619 				return (error ? error : EINTR);
620 			    }
621 			    if ((bp->b_flags & B_CACHE) == 0) {
622 				    bp->b_iocmd = BIO_READ;
623 				    vfs_busy_pages(bp, 0);
624 				    error = ncl_doio(vp, bp, cred, td, 0);
625 				    /*
626 				     * no error + B_INVAL == directory EOF,
627 				     * use the block.
628 				     */
629 				    if (error == 0 && (bp->b_flags & B_INVAL))
630 					    break;
631 			    }
632 			    /*
633 			     * An error will throw away the block and the
634 			     * for loop will break out.  If no error and this
635 			     * is not the block we want, we throw away the
636 			     * block and go for the next one via the for loop.
637 			     */
638 			    if (error || i < lbn)
639 				    brelse(bp);
640 			}
641 		    }
642 		    /*
643 		     * The above while is repeated if we hit another cookie
644 		     * error.  If we hit an error and it wasn't a cookie error,
645 		     * we give up.
646 		     */
647 		    if (error)
648 			    return (error);
649 		}
650 
651 		/*
652 		 * If not eof and read aheads are enabled, start one.
653 		 * (You need the current block first, so that you have the
654 		 *  directory offset cookie of the next block.)
655 		 */
656 		if (nmp->nm_readahead > 0 &&
657 		    (bp->b_flags & B_INVAL) == 0 &&
658 		    (np->n_direofoffset == 0 ||
659 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
660 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
661 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
662 			if (rabp) {
663 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
664 				rabp->b_flags |= B_ASYNC;
665 				rabp->b_iocmd = BIO_READ;
666 				vfs_busy_pages(rabp, 0);
667 				if (ncl_asyncio(nmp, rabp, cred, td)) {
668 				    rabp->b_flags |= B_INVAL;
669 				    rabp->b_ioflags |= BIO_ERROR;
670 				    vfs_unbusy_pages(rabp);
671 				    brelse(rabp);
672 				}
673 			    } else {
674 				brelse(rabp);
675 			    }
676 			}
677 		}
678 		/*
679 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
680 		 * chopped for the EOF condition, we cannot tell how large
681 		 * NFS directories are going to be until we hit EOF.  So
682 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
683 		 * it just so happens that b_resid will effectively chop it
684 		 * to EOF.  *BUT* this information is lost if the buffer goes
685 		 * away and is reconstituted into a B_CACHE state ( due to
686 		 * being VMIO ) later.  So we keep track of the directory eof
687 		 * in np->n_direofoffset and chop it off as an extra step
688 		 * right here.
689 		 */
690 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
691 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
692 			n = np->n_direofoffset - uio->uio_offset;
693 		break;
694 	    default:
695 		ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
696 		bp = NULL;
697 		break;
698 	    };
699 
700 	    if (n > 0) {
701 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
702 	    }
703 	    if (vp->v_type == VLNK)
704 		n = 0;
705 	    if (bp != NULL)
706 		brelse(bp);
707 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
708 	return (error);
709 }
710 
711 /*
712  * The NFS write path cannot handle iovecs with len > 1. So we need to
713  * break up iovecs accordingly (restricting them to wsize).
714  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
715  * For the ASYNC case, 2 copies are needed. The first a copy from the
716  * user buffer to a staging buffer and then a second copy from the staging
717  * buffer to mbufs. This can be optimized by copying from the user buffer
718  * directly into mbufs and passing the chain down, but that requires a
719  * fair amount of re-working of the relevant codepaths (and can be done
720  * later).
721  */
722 static int
723 nfs_directio_write(vp, uiop, cred, ioflag)
724 	struct vnode *vp;
725 	struct uio *uiop;
726 	struct ucred *cred;
727 	int ioflag;
728 {
729 	int error;
730 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
731 	struct thread *td = uiop->uio_td;
732 	int size;
733 	int wsize;
734 
735 	mtx_lock(&nmp->nm_mtx);
736 	wsize = nmp->nm_wsize;
737 	mtx_unlock(&nmp->nm_mtx);
738 	if (ioflag & IO_SYNC) {
739 		int iomode, must_commit;
740 		struct uio uio;
741 		struct iovec iov;
742 do_sync:
743 		while (uiop->uio_resid > 0) {
744 			size = MIN(uiop->uio_resid, wsize);
745 			size = MIN(uiop->uio_iov->iov_len, size);
746 			iov.iov_base = uiop->uio_iov->iov_base;
747 			iov.iov_len = size;
748 			uio.uio_iov = &iov;
749 			uio.uio_iovcnt = 1;
750 			uio.uio_offset = uiop->uio_offset;
751 			uio.uio_resid = size;
752 			uio.uio_segflg = UIO_USERSPACE;
753 			uio.uio_rw = UIO_WRITE;
754 			uio.uio_td = td;
755 			iomode = NFSWRITE_FILESYNC;
756 			error = ncl_writerpc(vp, &uio, cred, &iomode,
757 			    &must_commit, 0);
758 			KASSERT((must_commit == 0),
759 				("ncl_directio_write: Did not commit write"));
760 			if (error)
761 				return (error);
762 			uiop->uio_offset += size;
763 			uiop->uio_resid -= size;
764 			if (uiop->uio_iov->iov_len <= size) {
765 				uiop->uio_iovcnt--;
766 				uiop->uio_iov++;
767 			} else {
768 				uiop->uio_iov->iov_base =
769 					(char *)uiop->uio_iov->iov_base + size;
770 				uiop->uio_iov->iov_len -= size;
771 			}
772 		}
773 	} else {
774 		struct uio *t_uio;
775 		struct iovec *t_iov;
776 		struct buf *bp;
777 
778 		/*
779 		 * Break up the write into blocksize chunks and hand these
780 		 * over to nfsiod's for write back.
781 		 * Unfortunately, this incurs a copy of the data. Since
782 		 * the user could modify the buffer before the write is
783 		 * initiated.
784 		 *
785 		 * The obvious optimization here is that one of the 2 copies
786 		 * in the async write path can be eliminated by copying the
787 		 * data here directly into mbufs and passing the mbuf chain
788 		 * down. But that will require a fair amount of re-working
789 		 * of the code and can be done if there's enough interest
790 		 * in NFS directio access.
791 		 */
792 		while (uiop->uio_resid > 0) {
793 			size = MIN(uiop->uio_resid, wsize);
794 			size = MIN(uiop->uio_iov->iov_len, size);
795 			bp = getpbuf(&ncl_pbuf_freecnt);
796 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
797 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
798 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
799 			t_iov->iov_len = size;
800 			t_uio->uio_iov = t_iov;
801 			t_uio->uio_iovcnt = 1;
802 			t_uio->uio_offset = uiop->uio_offset;
803 			t_uio->uio_resid = size;
804 			t_uio->uio_segflg = UIO_SYSSPACE;
805 			t_uio->uio_rw = UIO_WRITE;
806 			t_uio->uio_td = td;
807 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
808 			    uiop->uio_segflg == UIO_SYSSPACE,
809 			    ("nfs_directio_write: Bad uio_segflg"));
810 			if (uiop->uio_segflg == UIO_USERSPACE) {
811 				error = copyin(uiop->uio_iov->iov_base,
812 				    t_iov->iov_base, size);
813 				if (error != 0)
814 					goto err_free;
815 			} else
816 				/*
817 				 * UIO_SYSSPACE may never happen, but handle
818 				 * it just in case it does.
819 				 */
820 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
821 				    size);
822 			bp->b_flags |= B_DIRECT;
823 			bp->b_iocmd = BIO_WRITE;
824 			if (cred != NOCRED) {
825 				crhold(cred);
826 				bp->b_wcred = cred;
827 			} else
828 				bp->b_wcred = NOCRED;
829 			bp->b_caller1 = (void *)t_uio;
830 			bp->b_vp = vp;
831 			error = ncl_asyncio(nmp, bp, NOCRED, td);
832 err_free:
833 			if (error) {
834 				free(t_iov->iov_base, M_NFSDIRECTIO);
835 				free(t_iov, M_NFSDIRECTIO);
836 				free(t_uio, M_NFSDIRECTIO);
837 				bp->b_vp = NULL;
838 				relpbuf(bp, &ncl_pbuf_freecnt);
839 				if (error == EINTR)
840 					return (error);
841 				goto do_sync;
842 			}
843 			uiop->uio_offset += size;
844 			uiop->uio_resid -= size;
845 			if (uiop->uio_iov->iov_len <= size) {
846 				uiop->uio_iovcnt--;
847 				uiop->uio_iov++;
848 			} else {
849 				uiop->uio_iov->iov_base =
850 					(char *)uiop->uio_iov->iov_base + size;
851 				uiop->uio_iov->iov_len -= size;
852 			}
853 		}
854 	}
855 	return (0);
856 }
857 
858 /*
859  * Vnode op for write using bio
860  */
861 int
862 ncl_write(struct vop_write_args *ap)
863 {
864 	int biosize;
865 	struct uio *uio = ap->a_uio;
866 	struct thread *td = uio->uio_td;
867 	struct vnode *vp = ap->a_vp;
868 	struct nfsnode *np = VTONFS(vp);
869 	struct ucred *cred = ap->a_cred;
870 	int ioflag = ap->a_ioflag;
871 	struct buf *bp;
872 	struct vattr vattr;
873 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
874 	daddr_t lbn;
875 	int bcount, noncontig_write, obcount;
876 	int bp_cached, n, on, error = 0, error1;
877 	size_t orig_resid, local_resid;
878 	off_t orig_size, tmp_off;
879 
880 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
881 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
882 	    ("ncl_write proc"));
883 	if (vp->v_type != VREG)
884 		return (EIO);
885 	mtx_lock(&np->n_mtx);
886 	if (np->n_flag & NWRITEERR) {
887 		np->n_flag &= ~NWRITEERR;
888 		mtx_unlock(&np->n_mtx);
889 		return (np->n_error);
890 	} else
891 		mtx_unlock(&np->n_mtx);
892 	mtx_lock(&nmp->nm_mtx);
893 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
894 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
895 		mtx_unlock(&nmp->nm_mtx);
896 		(void)ncl_fsinfo(nmp, vp, cred, td);
897 		mtx_lock(&nmp->nm_mtx);
898 	}
899 	if (nmp->nm_wsize == 0)
900 		(void) newnfs_iosize(nmp);
901 	mtx_unlock(&nmp->nm_mtx);
902 
903 	/*
904 	 * Synchronously flush pending buffers if we are in synchronous
905 	 * mode or if we are appending.
906 	 */
907 	if (ioflag & (IO_APPEND | IO_SYNC)) {
908 		mtx_lock(&np->n_mtx);
909 		if (np->n_flag & NMODIFIED) {
910 			mtx_unlock(&np->n_mtx);
911 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
912 			/*
913 			 * Require non-blocking, synchronous writes to
914 			 * dirty files to inform the program it needs
915 			 * to fsync(2) explicitly.
916 			 */
917 			if (ioflag & IO_NDELAY)
918 				return (EAGAIN);
919 #endif
920 flush_and_restart:
921 			np->n_attrstamp = 0;
922 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
923 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
924 			if (error)
925 				return (error);
926 		} else
927 			mtx_unlock(&np->n_mtx);
928 	}
929 
930 	orig_resid = uio->uio_resid;
931 	mtx_lock(&np->n_mtx);
932 	orig_size = np->n_size;
933 	mtx_unlock(&np->n_mtx);
934 
935 	/*
936 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
937 	 * get the append lock.
938 	 */
939 	if (ioflag & IO_APPEND) {
940 		np->n_attrstamp = 0;
941 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
942 		error = VOP_GETATTR(vp, &vattr, cred);
943 		if (error)
944 			return (error);
945 		mtx_lock(&np->n_mtx);
946 		uio->uio_offset = np->n_size;
947 		mtx_unlock(&np->n_mtx);
948 	}
949 
950 	if (uio->uio_offset < 0)
951 		return (EINVAL);
952 	tmp_off = uio->uio_offset + uio->uio_resid;
953 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
954 		return (EFBIG);
955 	if (uio->uio_resid == 0)
956 		return (0);
957 
958 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
959 		return nfs_directio_write(vp, uio, cred, ioflag);
960 
961 	/*
962 	 * Maybe this should be above the vnode op call, but so long as
963 	 * file servers have no limits, i don't think it matters
964 	 */
965 	if (vn_rlimit_fsize(vp, uio, td))
966 		return (EFBIG);
967 
968 	biosize = vp->v_bufobj.bo_bsize;
969 	/*
970 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
971 	 * would exceed the local maximum per-file write commit size when
972 	 * combined with those, we must decide whether to flush,
973 	 * go synchronous, or return error.  We don't bother checking
974 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
975 	 * no point optimizing for something that really won't ever happen.
976 	 */
977 	if (!(ioflag & IO_SYNC)) {
978 		int nflag;
979 
980 		mtx_lock(&np->n_mtx);
981 		nflag = np->n_flag;
982 		mtx_unlock(&np->n_mtx);
983 		int needrestart = 0;
984 		if (nmp->nm_wcommitsize < uio->uio_resid) {
985 			/*
986 			 * If this request could not possibly be completed
987 			 * without exceeding the maximum outstanding write
988 			 * commit size, see if we can convert it into a
989 			 * synchronous write operation.
990 			 */
991 			if (ioflag & IO_NDELAY)
992 				return (EAGAIN);
993 			ioflag |= IO_SYNC;
994 			if (nflag & NMODIFIED)
995 				needrestart = 1;
996 		} else if (nflag & NMODIFIED) {
997 			int wouldcommit = 0;
998 			BO_LOCK(&vp->v_bufobj);
999 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1000 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1001 				    b_bobufs) {
1002 					if (bp->b_flags & B_NEEDCOMMIT)
1003 						wouldcommit += bp->b_bcount;
1004 				}
1005 			}
1006 			BO_UNLOCK(&vp->v_bufobj);
1007 			/*
1008 			 * Since we're not operating synchronously and
1009 			 * bypassing the buffer cache, we are in a commit
1010 			 * and holding all of these buffers whether
1011 			 * transmitted or not.  If not limited, this
1012 			 * will lead to the buffer cache deadlocking,
1013 			 * as no one else can flush our uncommitted buffers.
1014 			 */
1015 			wouldcommit += uio->uio_resid;
1016 			/*
1017 			 * If we would initially exceed the maximum
1018 			 * outstanding write commit size, flush and restart.
1019 			 */
1020 			if (wouldcommit > nmp->nm_wcommitsize)
1021 				needrestart = 1;
1022 		}
1023 		if (needrestart)
1024 			goto flush_and_restart;
1025 	}
1026 
1027 	do {
1028 		NFSINCRGLOBAL(newnfsstats.biocache_writes);
1029 		lbn = uio->uio_offset / biosize;
1030 		on = uio->uio_offset - (lbn * biosize);
1031 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1032 again:
1033 		/*
1034 		 * Handle direct append and file extension cases, calculate
1035 		 * unaligned buffer size.
1036 		 */
1037 		mtx_lock(&np->n_mtx);
1038 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1039 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1040 			noncontig_write = 1;
1041 		else
1042 			noncontig_write = 0;
1043 		if ((uio->uio_offset == np->n_size ||
1044 		    (noncontig_write != 0 &&
1045 		    lbn == (np->n_size / biosize) &&
1046 		    uio->uio_offset + n > np->n_size)) && n) {
1047 			mtx_unlock(&np->n_mtx);
1048 			/*
1049 			 * Get the buffer (in its pre-append state to maintain
1050 			 * B_CACHE if it was previously set).  Resize the
1051 			 * nfsnode after we have locked the buffer to prevent
1052 			 * readers from reading garbage.
1053 			 */
1054 			obcount = np->n_size - (lbn * biosize);
1055 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1056 
1057 			if (bp != NULL) {
1058 				long save;
1059 
1060 				mtx_lock(&np->n_mtx);
1061 				np->n_size = uio->uio_offset + n;
1062 				np->n_flag |= NMODIFIED;
1063 				vnode_pager_setsize(vp, np->n_size);
1064 				mtx_unlock(&np->n_mtx);
1065 
1066 				save = bp->b_flags & B_CACHE;
1067 				bcount = on + n;
1068 				allocbuf(bp, bcount);
1069 				bp->b_flags |= save;
1070 				if (noncontig_write != 0 && on > obcount)
1071 					vfs_bio_bzero_buf(bp, obcount, on -
1072 					    obcount);
1073 			}
1074 		} else {
1075 			/*
1076 			 * Obtain the locked cache block first, and then
1077 			 * adjust the file's size as appropriate.
1078 			 */
1079 			bcount = on + n;
1080 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1081 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1082 					bcount = biosize;
1083 				else
1084 					bcount = np->n_size - (off_t)lbn * biosize;
1085 			}
1086 			mtx_unlock(&np->n_mtx);
1087 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1088 			mtx_lock(&np->n_mtx);
1089 			if (uio->uio_offset + n > np->n_size) {
1090 				np->n_size = uio->uio_offset + n;
1091 				np->n_flag |= NMODIFIED;
1092 				vnode_pager_setsize(vp, np->n_size);
1093 			}
1094 			mtx_unlock(&np->n_mtx);
1095 		}
1096 
1097 		if (!bp) {
1098 			error = newnfs_sigintr(nmp, td);
1099 			if (!error)
1100 				error = EINTR;
1101 			break;
1102 		}
1103 
1104 		/*
1105 		 * Issue a READ if B_CACHE is not set.  In special-append
1106 		 * mode, B_CACHE is based on the buffer prior to the write
1107 		 * op and is typically set, avoiding the read.  If a read
1108 		 * is required in special append mode, the server will
1109 		 * probably send us a short-read since we extended the file
1110 		 * on our end, resulting in b_resid == 0 and, thusly,
1111 		 * B_CACHE getting set.
1112 		 *
1113 		 * We can also avoid issuing the read if the write covers
1114 		 * the entire buffer.  We have to make sure the buffer state
1115 		 * is reasonable in this case since we will not be initiating
1116 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1117 		 * more information.
1118 		 *
1119 		 * B_CACHE may also be set due to the buffer being cached
1120 		 * normally.
1121 		 */
1122 
1123 		bp_cached = 1;
1124 		if (on == 0 && n == bcount) {
1125 			if ((bp->b_flags & B_CACHE) == 0)
1126 				bp_cached = 0;
1127 			bp->b_flags |= B_CACHE;
1128 			bp->b_flags &= ~B_INVAL;
1129 			bp->b_ioflags &= ~BIO_ERROR;
1130 		}
1131 
1132 		if ((bp->b_flags & B_CACHE) == 0) {
1133 			bp->b_iocmd = BIO_READ;
1134 			vfs_busy_pages(bp, 0);
1135 			error = ncl_doio(vp, bp, cred, td, 0);
1136 			if (error) {
1137 				brelse(bp);
1138 				break;
1139 			}
1140 		}
1141 		if (bp->b_wcred == NOCRED)
1142 			bp->b_wcred = crhold(cred);
1143 		mtx_lock(&np->n_mtx);
1144 		np->n_flag |= NMODIFIED;
1145 		mtx_unlock(&np->n_mtx);
1146 
1147 		/*
1148 		 * If dirtyend exceeds file size, chop it down.  This should
1149 		 * not normally occur but there is an append race where it
1150 		 * might occur XXX, so we log it.
1151 		 *
1152 		 * If the chopping creates a reverse-indexed or degenerate
1153 		 * situation with dirtyoff/end, we 0 both of them.
1154 		 */
1155 
1156 		if (bp->b_dirtyend > bcount) {
1157 			ncl_printf("NFS append race @%lx:%d\n",
1158 			    (long)bp->b_blkno * DEV_BSIZE,
1159 			    bp->b_dirtyend - bcount);
1160 			bp->b_dirtyend = bcount;
1161 		}
1162 
1163 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1164 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1165 
1166 		/*
1167 		 * If the new write will leave a contiguous dirty
1168 		 * area, just update the b_dirtyoff and b_dirtyend,
1169 		 * otherwise force a write rpc of the old dirty area.
1170 		 *
1171 		 * If there has been a file lock applied to this file
1172 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1173 		 * While it is possible to merge discontiguous writes due to
1174 		 * our having a B_CACHE buffer ( and thus valid read data
1175 		 * for the hole), we don't because it could lead to
1176 		 * significant cache coherency problems with multiple clients,
1177 		 * especially if locking is implemented later on.
1178 		 *
1179 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1180 		 * not been file locking done on this file:
1181 		 * Relax coherency a bit for the sake of performance and
1182 		 * expand the current dirty region to contain the new
1183 		 * write even if it means we mark some non-dirty data as
1184 		 * dirty.
1185 		 */
1186 
1187 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1188 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1189 			if (bwrite(bp) == EINTR) {
1190 				error = EINTR;
1191 				break;
1192 			}
1193 			goto again;
1194 		}
1195 
1196 		local_resid = uio->uio_resid;
1197 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1198 
1199 		if (error != 0 && !bp_cached) {
1200 			/*
1201 			 * This block has no other content then what
1202 			 * possibly was written by the faulty uiomove.
1203 			 * Release it, forgetting the data pages, to
1204 			 * prevent the leak of uninitialized data to
1205 			 * usermode.
1206 			 */
1207 			bp->b_ioflags |= BIO_ERROR;
1208 			brelse(bp);
1209 			uio->uio_offset -= local_resid - uio->uio_resid;
1210 			uio->uio_resid = local_resid;
1211 			break;
1212 		}
1213 
1214 		/*
1215 		 * Since this block is being modified, it must be written
1216 		 * again and not just committed.  Since write clustering does
1217 		 * not work for the stage 1 data write, only the stage 2
1218 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1219 		 */
1220 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1221 
1222 		/*
1223 		 * Get the partial update on the progress made from
1224 		 * uiomove, if an error occured.
1225 		 */
1226 		if (error != 0)
1227 			n = local_resid - uio->uio_resid;
1228 
1229 		/*
1230 		 * Only update dirtyoff/dirtyend if not a degenerate
1231 		 * condition.
1232 		 */
1233 		if (n > 0) {
1234 			if (bp->b_dirtyend > 0) {
1235 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1236 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1237 			} else {
1238 				bp->b_dirtyoff = on;
1239 				bp->b_dirtyend = on + n;
1240 			}
1241 			vfs_bio_set_valid(bp, on, n);
1242 		}
1243 
1244 		/*
1245 		 * If IO_SYNC do bwrite().
1246 		 *
1247 		 * IO_INVAL appears to be unused.  The idea appears to be
1248 		 * to turn off caching in this case.  Very odd.  XXX
1249 		 */
1250 		if ((ioflag & IO_SYNC)) {
1251 			if (ioflag & IO_INVAL)
1252 				bp->b_flags |= B_NOCACHE;
1253 			error1 = bwrite(bp);
1254 			if (error1 != 0) {
1255 				if (error == 0)
1256 					error = error1;
1257 				break;
1258 			}
1259 		} else if ((n + on) == biosize) {
1260 			bp->b_flags |= B_ASYNC;
1261 			(void) ncl_writebp(bp, 0, NULL);
1262 		} else {
1263 			bdwrite(bp);
1264 		}
1265 
1266 		if (error != 0)
1267 			break;
1268 	} while (uio->uio_resid > 0 && n > 0);
1269 
1270 	if (error != 0) {
1271 		if (ioflag & IO_UNIT) {
1272 			VATTR_NULL(&vattr);
1273 			vattr.va_size = orig_size;
1274 			/* IO_SYNC is handled implicitely */
1275 			(void)VOP_SETATTR(vp, &vattr, cred);
1276 			uio->uio_offset -= orig_resid - uio->uio_resid;
1277 			uio->uio_resid = orig_resid;
1278 		}
1279 	}
1280 
1281 	return (error);
1282 }
1283 
1284 /*
1285  * Get an nfs cache block.
1286  *
1287  * Allocate a new one if the block isn't currently in the cache
1288  * and return the block marked busy. If the calling process is
1289  * interrupted by a signal for an interruptible mount point, return
1290  * NULL.
1291  *
1292  * The caller must carefully deal with the possible B_INVAL state of
1293  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1294  * indirectly), so synchronous reads can be issued without worrying about
1295  * the B_INVAL state.  We have to be a little more careful when dealing
1296  * with writes (see comments in nfs_write()) when extending a file past
1297  * its EOF.
1298  */
1299 static struct buf *
1300 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1301 {
1302 	struct buf *bp;
1303 	struct mount *mp;
1304 	struct nfsmount *nmp;
1305 
1306 	mp = vp->v_mount;
1307 	nmp = VFSTONFS(mp);
1308 
1309 	if (nmp->nm_flag & NFSMNT_INT) {
1310 		sigset_t oldset;
1311 
1312 		newnfs_set_sigmask(td, &oldset);
1313 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1314 		newnfs_restore_sigmask(td, &oldset);
1315 		while (bp == NULL) {
1316 			if (newnfs_sigintr(nmp, td))
1317 				return (NULL);
1318 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1319 		}
1320 	} else {
1321 		bp = getblk(vp, bn, size, 0, 0, 0);
1322 	}
1323 
1324 	if (vp->v_type == VREG)
1325 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1326 	return (bp);
1327 }
1328 
1329 /*
1330  * Flush and invalidate all dirty buffers. If another process is already
1331  * doing the flush, just wait for completion.
1332  */
1333 int
1334 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1335 {
1336 	struct nfsnode *np = VTONFS(vp);
1337 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1338 	int error = 0, slpflag, slptimeo;
1339 	int old_lock = 0;
1340 
1341 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1342 
1343 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1344 		intrflg = 0;
1345 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1346 		intrflg = 1;
1347 	if (intrflg) {
1348 		slpflag = PCATCH;
1349 		slptimeo = 2 * hz;
1350 	} else {
1351 		slpflag = 0;
1352 		slptimeo = 0;
1353 	}
1354 
1355 	old_lock = ncl_upgrade_vnlock(vp);
1356 	if (vp->v_iflag & VI_DOOMED) {
1357 		/*
1358 		 * Since vgonel() uses the generic vinvalbuf() to flush
1359 		 * dirty buffers and it does not call this function, it
1360 		 * is safe to just return OK when VI_DOOMED is set.
1361 		 */
1362 		ncl_downgrade_vnlock(vp, old_lock);
1363 		return (0);
1364 	}
1365 
1366 	/*
1367 	 * Now, flush as required.
1368 	 */
1369 	if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1370 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1371 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1372 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1373 		/*
1374 		 * If the page clean was interrupted, fail the invalidation.
1375 		 * Not doing so, we run the risk of losing dirty pages in the
1376 		 * vinvalbuf() call below.
1377 		 */
1378 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1379 			goto out;
1380 	}
1381 
1382 	error = vinvalbuf(vp, flags, slpflag, 0);
1383 	while (error) {
1384 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1385 			goto out;
1386 		error = vinvalbuf(vp, flags, 0, slptimeo);
1387 	}
1388 	if (NFSHASPNFS(nmp)) {
1389 		nfscl_layoutcommit(vp, td);
1390 		/*
1391 		 * Invalidate the attribute cache, since writes to a DS
1392 		 * won't update the size attribute.
1393 		 */
1394 		mtx_lock(&np->n_mtx);
1395 		np->n_attrstamp = 0;
1396 	} else
1397 		mtx_lock(&np->n_mtx);
1398 	if (np->n_directio_asyncwr == 0)
1399 		np->n_flag &= ~NMODIFIED;
1400 	mtx_unlock(&np->n_mtx);
1401 out:
1402 	ncl_downgrade_vnlock(vp, old_lock);
1403 	return error;
1404 }
1405 
1406 /*
1407  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1408  * This is mainly to avoid queueing async I/O requests when the nfsiods
1409  * are all hung on a dead server.
1410  *
1411  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1412  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1413  */
1414 int
1415 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1416 {
1417 	int iod;
1418 	int gotiod;
1419 	int slpflag = 0;
1420 	int slptimeo = 0;
1421 	int error, error2;
1422 
1423 	/*
1424 	 * Commits are usually short and sweet so lets save some cpu and
1425 	 * leave the async daemons for more important rpc's (such as reads
1426 	 * and writes).
1427 	 *
1428 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1429 	 * in the directory in order to update attributes. This can deadlock
1430 	 * with another thread that is waiting for async I/O to be done by
1431 	 * an nfsiod thread while holding a lock on one of these vnodes.
1432 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1433 	 * perform Readdirplus RPCs.
1434 	 */
1435 	mtx_lock(&ncl_iod_mutex);
1436 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1437 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1438 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1439 		mtx_unlock(&ncl_iod_mutex);
1440 		return(EIO);
1441 	}
1442 again:
1443 	if (nmp->nm_flag & NFSMNT_INT)
1444 		slpflag = PCATCH;
1445 	gotiod = FALSE;
1446 
1447 	/*
1448 	 * Find a free iod to process this request.
1449 	 */
1450 	for (iod = 0; iod < ncl_numasync; iod++)
1451 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1452 			gotiod = TRUE;
1453 			break;
1454 		}
1455 
1456 	/*
1457 	 * Try to create one if none are free.
1458 	 */
1459 	if (!gotiod)
1460 		ncl_nfsiodnew();
1461 	else {
1462 		/*
1463 		 * Found one, so wake it up and tell it which
1464 		 * mount to process.
1465 		 */
1466 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1467 		    iod, nmp));
1468 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1469 		ncl_iodmount[iod] = nmp;
1470 		nmp->nm_bufqiods++;
1471 		wakeup(&ncl_iodwant[iod]);
1472 	}
1473 
1474 	/*
1475 	 * If none are free, we may already have an iod working on this mount
1476 	 * point.  If so, it will process our request.
1477 	 */
1478 	if (!gotiod) {
1479 		if (nmp->nm_bufqiods > 0) {
1480 			NFS_DPF(ASYNCIO,
1481 				("ncl_asyncio: %d iods are already processing mount %p\n",
1482 				 nmp->nm_bufqiods, nmp));
1483 			gotiod = TRUE;
1484 		}
1485 	}
1486 
1487 	/*
1488 	 * If we have an iod which can process the request, then queue
1489 	 * the buffer.
1490 	 */
1491 	if (gotiod) {
1492 		/*
1493 		 * Ensure that the queue never grows too large.  We still want
1494 		 * to asynchronize so we block rather then return EIO.
1495 		 */
1496 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1497 			NFS_DPF(ASYNCIO,
1498 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1499 			nmp->nm_bufqwant = TRUE;
1500 			error = newnfs_msleep(td, &nmp->nm_bufq,
1501 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1502 			   slptimeo);
1503 			if (error) {
1504 				error2 = newnfs_sigintr(nmp, td);
1505 				if (error2) {
1506 					mtx_unlock(&ncl_iod_mutex);
1507 					return (error2);
1508 				}
1509 				if (slpflag == PCATCH) {
1510 					slpflag = 0;
1511 					slptimeo = 2 * hz;
1512 				}
1513 			}
1514 			/*
1515 			 * We might have lost our iod while sleeping,
1516 			 * so check and loop if nescessary.
1517 			 */
1518 			goto again;
1519 		}
1520 
1521 		/* We might have lost our nfsiod */
1522 		if (nmp->nm_bufqiods == 0) {
1523 			NFS_DPF(ASYNCIO,
1524 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1525 			goto again;
1526 		}
1527 
1528 		if (bp->b_iocmd == BIO_READ) {
1529 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1530 				bp->b_rcred = crhold(cred);
1531 		} else {
1532 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1533 				bp->b_wcred = crhold(cred);
1534 		}
1535 
1536 		if (bp->b_flags & B_REMFREE)
1537 			bremfreef(bp);
1538 		BUF_KERNPROC(bp);
1539 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1540 		nmp->nm_bufqlen++;
1541 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1542 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1543 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1544 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1545 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1546 		}
1547 		mtx_unlock(&ncl_iod_mutex);
1548 		return (0);
1549 	}
1550 
1551 	mtx_unlock(&ncl_iod_mutex);
1552 
1553 	/*
1554 	 * All the iods are busy on other mounts, so return EIO to
1555 	 * force the caller to process the i/o synchronously.
1556 	 */
1557 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1558 	return (EIO);
1559 }
1560 
1561 void
1562 ncl_doio_directwrite(struct buf *bp)
1563 {
1564 	int iomode, must_commit;
1565 	struct uio *uiop = (struct uio *)bp->b_caller1;
1566 	char *iov_base = uiop->uio_iov->iov_base;
1567 
1568 	iomode = NFSWRITE_FILESYNC;
1569 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1570 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1571 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1572 	free(iov_base, M_NFSDIRECTIO);
1573 	free(uiop->uio_iov, M_NFSDIRECTIO);
1574 	free(uiop, M_NFSDIRECTIO);
1575 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1576 		struct nfsnode *np = VTONFS(bp->b_vp);
1577 		mtx_lock(&np->n_mtx);
1578 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1579 			/*
1580 			 * Invalidate the attribute cache, since writes to a DS
1581 			 * won't update the size attribute.
1582 			 */
1583 			np->n_attrstamp = 0;
1584 		}
1585 		np->n_directio_asyncwr--;
1586 		if (np->n_directio_asyncwr == 0) {
1587 			np->n_flag &= ~NMODIFIED;
1588 			if ((np->n_flag & NFSYNCWAIT)) {
1589 				np->n_flag &= ~NFSYNCWAIT;
1590 				wakeup((caddr_t)&np->n_directio_asyncwr);
1591 			}
1592 		}
1593 		mtx_unlock(&np->n_mtx);
1594 	}
1595 	bp->b_vp = NULL;
1596 	relpbuf(bp, &ncl_pbuf_freecnt);
1597 }
1598 
1599 /*
1600  * Do an I/O operation to/from a cache block. This may be called
1601  * synchronously or from an nfsiod.
1602  */
1603 int
1604 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1605     int called_from_strategy)
1606 {
1607 	struct uio *uiop;
1608 	struct nfsnode *np;
1609 	struct nfsmount *nmp;
1610 	int error = 0, iomode, must_commit = 0;
1611 	struct uio uio;
1612 	struct iovec io;
1613 	struct proc *p = td ? td->td_proc : NULL;
1614 	uint8_t	iocmd;
1615 
1616 	np = VTONFS(vp);
1617 	nmp = VFSTONFS(vp->v_mount);
1618 	uiop = &uio;
1619 	uiop->uio_iov = &io;
1620 	uiop->uio_iovcnt = 1;
1621 	uiop->uio_segflg = UIO_SYSSPACE;
1622 	uiop->uio_td = td;
1623 
1624 	/*
1625 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1626 	 * do this here so we do not have to do it in all the code that
1627 	 * calls us.
1628 	 */
1629 	bp->b_flags &= ~B_INVAL;
1630 	bp->b_ioflags &= ~BIO_ERROR;
1631 
1632 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1633 	iocmd = bp->b_iocmd;
1634 	if (iocmd == BIO_READ) {
1635 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1636 	    io.iov_base = bp->b_data;
1637 	    uiop->uio_rw = UIO_READ;
1638 
1639 	    switch (vp->v_type) {
1640 	    case VREG:
1641 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1642 		NFSINCRGLOBAL(newnfsstats.read_bios);
1643 		error = ncl_readrpc(vp, uiop, cr);
1644 
1645 		if (!error) {
1646 		    if (uiop->uio_resid) {
1647 			/*
1648 			 * If we had a short read with no error, we must have
1649 			 * hit a file hole.  We should zero-fill the remainder.
1650 			 * This can also occur if the server hits the file EOF.
1651 			 *
1652 			 * Holes used to be able to occur due to pending
1653 			 * writes, but that is not possible any longer.
1654 			 */
1655 			int nread = bp->b_bcount - uiop->uio_resid;
1656 			ssize_t left = uiop->uio_resid;
1657 
1658 			if (left > 0)
1659 				bzero((char *)bp->b_data + nread, left);
1660 			uiop->uio_resid = 0;
1661 		    }
1662 		}
1663 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1664 		if (p && (vp->v_vflag & VV_TEXT)) {
1665 			mtx_lock(&np->n_mtx);
1666 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1667 				mtx_unlock(&np->n_mtx);
1668 				PROC_LOCK(p);
1669 				killproc(p, "text file modification");
1670 				PROC_UNLOCK(p);
1671 			} else
1672 				mtx_unlock(&np->n_mtx);
1673 		}
1674 		break;
1675 	    case VLNK:
1676 		uiop->uio_offset = (off_t)0;
1677 		NFSINCRGLOBAL(newnfsstats.readlink_bios);
1678 		error = ncl_readlinkrpc(vp, uiop, cr);
1679 		break;
1680 	    case VDIR:
1681 		NFSINCRGLOBAL(newnfsstats.readdir_bios);
1682 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1683 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1684 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1685 			if (error == NFSERR_NOTSUPP)
1686 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1687 		}
1688 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1689 			error = ncl_readdirrpc(vp, uiop, cr, td);
1690 		/*
1691 		 * end-of-directory sets B_INVAL but does not generate an
1692 		 * error.
1693 		 */
1694 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1695 			bp->b_flags |= B_INVAL;
1696 		break;
1697 	    default:
1698 		ncl_printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1699 		break;
1700 	    };
1701 	    if (error) {
1702 		bp->b_ioflags |= BIO_ERROR;
1703 		bp->b_error = error;
1704 	    }
1705 	} else {
1706 	    /*
1707 	     * If we only need to commit, try to commit
1708 	     */
1709 	    if (bp->b_flags & B_NEEDCOMMIT) {
1710 		    int retv;
1711 		    off_t off;
1712 
1713 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1714 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1715 			bp->b_wcred, td);
1716 		    if (retv == 0) {
1717 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1718 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1719 			    bp->b_resid = 0;
1720 			    bufdone(bp);
1721 			    return (0);
1722 		    }
1723 		    if (retv == NFSERR_STALEWRITEVERF) {
1724 			    ncl_clearcommit(vp->v_mount);
1725 		    }
1726 	    }
1727 
1728 	    /*
1729 	     * Setup for actual write
1730 	     */
1731 	    mtx_lock(&np->n_mtx);
1732 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1733 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1734 	    mtx_unlock(&np->n_mtx);
1735 
1736 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1737 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1738 		    - bp->b_dirtyoff;
1739 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1740 		    + bp->b_dirtyoff;
1741 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1742 		uiop->uio_rw = UIO_WRITE;
1743 		NFSINCRGLOBAL(newnfsstats.write_bios);
1744 
1745 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1746 		    iomode = NFSWRITE_UNSTABLE;
1747 		else
1748 		    iomode = NFSWRITE_FILESYNC;
1749 
1750 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1751 		    called_from_strategy);
1752 
1753 		/*
1754 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1755 		 * to cluster the buffers needing commit.  This will allow
1756 		 * the system to submit a single commit rpc for the whole
1757 		 * cluster.  We can do this even if the buffer is not 100%
1758 		 * dirty (relative to the NFS blocksize), so we optimize the
1759 		 * append-to-file-case.
1760 		 *
1761 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1762 		 * cleared because write clustering only works for commit
1763 		 * rpc's, not for the data portion of the write).
1764 		 */
1765 
1766 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1767 		    bp->b_flags |= B_NEEDCOMMIT;
1768 		    if (bp->b_dirtyoff == 0
1769 			&& bp->b_dirtyend == bp->b_bcount)
1770 			bp->b_flags |= B_CLUSTEROK;
1771 		} else {
1772 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1773 		}
1774 
1775 		/*
1776 		 * For an interrupted write, the buffer is still valid
1777 		 * and the write hasn't been pushed to the server yet,
1778 		 * so we can't set BIO_ERROR and report the interruption
1779 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1780 		 * is not relevant, so the rpc attempt is essentially
1781 		 * a noop.  For the case of a V3 write rpc not being
1782 		 * committed to stable storage, the block is still
1783 		 * dirty and requires either a commit rpc or another
1784 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1785 		 * the block is reused. This is indicated by setting
1786 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1787 		 *
1788 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1789 		 * write error and is handled as above, except that
1790 		 * B_EINTR isn't set. One cause of this is a stale stateid
1791 		 * error for the RPC that indicates recovery is required,
1792 		 * when called with called_from_strategy != 0.
1793 		 *
1794 		 * If the buffer is marked B_PAGING, it does not reside on
1795 		 * the vp's paging queues so we cannot call bdirty().  The
1796 		 * bp in this case is not an NFS cache block so we should
1797 		 * be safe. XXX
1798 		 *
1799 		 * The logic below breaks up errors into recoverable and
1800 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1801 		 * and keep the buffer around for potential write retries.
1802 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1803 		 * and save the error in the nfsnode. This is less than ideal
1804 		 * but necessary. Keeping such buffers around could potentially
1805 		 * cause buffer exhaustion eventually (they can never be written
1806 		 * out, so will get constantly be re-dirtied). It also causes
1807 		 * all sorts of vfs panics. For non-recoverable write errors,
1808 		 * also invalidate the attrcache, so we'll be forced to go over
1809 		 * the wire for this object, returning an error to user on next
1810 		 * call (most of the time).
1811 		 */
1812 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1813 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1814 			int s;
1815 
1816 			s = splbio();
1817 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1818 			if ((bp->b_flags & B_PAGING) == 0) {
1819 			    bdirty(bp);
1820 			    bp->b_flags &= ~B_DONE;
1821 			}
1822 			if ((error == EINTR || error == ETIMEDOUT) &&
1823 			    (bp->b_flags & B_ASYNC) == 0)
1824 			    bp->b_flags |= B_EINTR;
1825 			splx(s);
1826 		} else {
1827 		    if (error) {
1828 			bp->b_ioflags |= BIO_ERROR;
1829 			bp->b_flags |= B_INVAL;
1830 			bp->b_error = np->n_error = error;
1831 			mtx_lock(&np->n_mtx);
1832 			np->n_flag |= NWRITEERR;
1833 			np->n_attrstamp = 0;
1834 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1835 			mtx_unlock(&np->n_mtx);
1836 		    }
1837 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1838 		}
1839 	    } else {
1840 		bp->b_resid = 0;
1841 		bufdone(bp);
1842 		return (0);
1843 	    }
1844 	}
1845 	bp->b_resid = uiop->uio_resid;
1846 	if (must_commit)
1847 	    ncl_clearcommit(vp->v_mount);
1848 	bufdone(bp);
1849 	return (error);
1850 }
1851 
1852 /*
1853  * Used to aid in handling ftruncate() operations on the NFS client side.
1854  * Truncation creates a number of special problems for NFS.  We have to
1855  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1856  * we have to properly handle VM pages or (potentially dirty) buffers
1857  * that straddle the truncation point.
1858  */
1859 
1860 int
1861 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1862 {
1863 	struct nfsnode *np = VTONFS(vp);
1864 	u_quad_t tsize;
1865 	int biosize = vp->v_bufobj.bo_bsize;
1866 	int error = 0;
1867 
1868 	mtx_lock(&np->n_mtx);
1869 	tsize = np->n_size;
1870 	np->n_size = nsize;
1871 	mtx_unlock(&np->n_mtx);
1872 
1873 	if (nsize < tsize) {
1874 		struct buf *bp;
1875 		daddr_t lbn;
1876 		int bufsize;
1877 
1878 		/*
1879 		 * vtruncbuf() doesn't get the buffer overlapping the
1880 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1881 		 * buffer that now needs to be truncated.
1882 		 */
1883 		error = vtruncbuf(vp, cred, nsize, biosize);
1884 		lbn = nsize / biosize;
1885 		bufsize = nsize - (lbn * biosize);
1886 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1887 		if (!bp)
1888 			return EINTR;
1889 		if (bp->b_dirtyoff > bp->b_bcount)
1890 			bp->b_dirtyoff = bp->b_bcount;
1891 		if (bp->b_dirtyend > bp->b_bcount)
1892 			bp->b_dirtyend = bp->b_bcount;
1893 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1894 		brelse(bp);
1895 	} else {
1896 		vnode_pager_setsize(vp, nsize);
1897 	}
1898 	return(error);
1899 }
1900 
1901