xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 788ca347b816afd83b2885e0c79aeeb88649b2ab)
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 		    FALSE);
145 		return (VM_PAGER_OK);
146 	}
147 
148 	/*
149 	 * We use only the kva address for the buffer, but this is extremely
150 	 * convienient and fast.
151 	 */
152 	bp = getpbuf(&ncl_pbuf_freecnt);
153 
154 	kva = (vm_offset_t) bp->b_data;
155 	pmap_qenter(kva, pages, npages);
156 	PCPU_INC(cnt.v_vnodein);
157 	PCPU_ADD(cnt.v_vnodepgsin, npages);
158 
159 	iov.iov_base = (caddr_t) kva;
160 	iov.iov_len = count;
161 	uio.uio_iov = &iov;
162 	uio.uio_iovcnt = 1;
163 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
164 	uio.uio_resid = count;
165 	uio.uio_segflg = UIO_SYSSPACE;
166 	uio.uio_rw = UIO_READ;
167 	uio.uio_td = td;
168 
169 	error = ncl_readrpc(vp, &uio, cred);
170 	pmap_qremove(kva, npages);
171 
172 	relpbuf(bp, &ncl_pbuf_freecnt);
173 
174 	if (error && (uio.uio_resid == count)) {
175 		ncl_printf("nfs_getpages: error %d\n", error);
176 		vm_pager_free_nonreq(object, pages, ap->a_reqpage, npages,
177 		    FALSE);
178 		return (VM_PAGER_ERROR);
179 	}
180 
181 	/*
182 	 * Calculate the number of bytes read and validate only that number
183 	 * of bytes.  Note that due to pending writes, size may be 0.  This
184 	 * does not mean that the remaining data is invalid!
185 	 */
186 
187 	size = count - uio.uio_resid;
188 	VM_OBJECT_WLOCK(object);
189 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
190 		vm_page_t m;
191 		nextoff = toff + PAGE_SIZE;
192 		m = pages[i];
193 
194 		if (nextoff <= size) {
195 			/*
196 			 * Read operation filled an entire page
197 			 */
198 			m->valid = VM_PAGE_BITS_ALL;
199 			KASSERT(m->dirty == 0,
200 			    ("nfs_getpages: page %p is dirty", m));
201 		} else if (size > toff) {
202 			/*
203 			 * Read operation filled a partial page.
204 			 */
205 			m->valid = 0;
206 			vm_page_set_valid_range(m, 0, size - toff);
207 			KASSERT(m->dirty == 0,
208 			    ("nfs_getpages: page %p is dirty", m));
209 		} else {
210 			/*
211 			 * Read operation was short.  If no error
212 			 * occured we may have hit a zero-fill
213 			 * section.  We leave valid set to 0, and page
214 			 * is freed by vm_page_readahead_finish() if
215 			 * its index is not equal to requested, or
216 			 * page is zeroed and set valid by
217 			 * vm_pager_get_pages() for requested page.
218 			 */
219 			;
220 		}
221 		if (i != ap->a_reqpage)
222 			vm_page_readahead_finish(m);
223 	}
224 	VM_OBJECT_WUNLOCK(object);
225 	return (0);
226 }
227 
228 /*
229  * Vnode op for VM putpages.
230  */
231 int
232 ncl_putpages(struct vop_putpages_args *ap)
233 {
234 	struct uio uio;
235 	struct iovec iov;
236 	vm_offset_t kva;
237 	struct buf *bp;
238 	int iomode, must_commit, i, error, npages, count;
239 	off_t offset;
240 	int *rtvals;
241 	struct vnode *vp;
242 	struct thread *td;
243 	struct ucred *cred;
244 	struct nfsmount *nmp;
245 	struct nfsnode *np;
246 	vm_page_t *pages;
247 
248 	vp = ap->a_vp;
249 	np = VTONFS(vp);
250 	td = curthread;				/* XXX */
251 	/* Set the cred to n_writecred for the write rpcs. */
252 	if (np->n_writecred != NULL)
253 		cred = crhold(np->n_writecred);
254 	else
255 		cred = crhold(curthread->td_ucred);	/* XXX */
256 	nmp = VFSTONFS(vp->v_mount);
257 	pages = ap->a_m;
258 	count = ap->a_count;
259 	rtvals = ap->a_rtvals;
260 	npages = btoc(count);
261 	offset = IDX_TO_OFF(pages[0]->pindex);
262 
263 	mtx_lock(&nmp->nm_mtx);
264 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
265 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
266 		mtx_unlock(&nmp->nm_mtx);
267 		(void)ncl_fsinfo(nmp, vp, cred, td);
268 	} else
269 		mtx_unlock(&nmp->nm_mtx);
270 
271 	mtx_lock(&np->n_mtx);
272 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
273 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
274 		mtx_unlock(&np->n_mtx);
275 		ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
276 		mtx_lock(&np->n_mtx);
277 	}
278 
279 	for (i = 0; i < npages; i++)
280 		rtvals[i] = VM_PAGER_ERROR;
281 
282 	/*
283 	 * When putting pages, do not extend file past EOF.
284 	 */
285 	if (offset + count > np->n_size) {
286 		count = np->n_size - offset;
287 		if (count < 0)
288 			count = 0;
289 	}
290 	mtx_unlock(&np->n_mtx);
291 
292 	/*
293 	 * We use only the kva address for the buffer, but this is extremely
294 	 * convienient and fast.
295 	 */
296 	bp = getpbuf(&ncl_pbuf_freecnt);
297 
298 	kva = (vm_offset_t) bp->b_data;
299 	pmap_qenter(kva, pages, npages);
300 	PCPU_INC(cnt.v_vnodeout);
301 	PCPU_ADD(cnt.v_vnodepgsout, count);
302 
303 	iov.iov_base = (caddr_t) kva;
304 	iov.iov_len = count;
305 	uio.uio_iov = &iov;
306 	uio.uio_iovcnt = 1;
307 	uio.uio_offset = offset;
308 	uio.uio_resid = count;
309 	uio.uio_segflg = UIO_SYSSPACE;
310 	uio.uio_rw = UIO_WRITE;
311 	uio.uio_td = td;
312 
313 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
314 	    iomode = NFSWRITE_UNSTABLE;
315 	else
316 	    iomode = NFSWRITE_FILESYNC;
317 
318 	error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
319 	crfree(cred);
320 
321 	pmap_qremove(kva, npages);
322 	relpbuf(bp, &ncl_pbuf_freecnt);
323 
324 	if (error == 0 || !nfs_keep_dirty_on_error) {
325 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
326 		if (must_commit)
327 			ncl_clearcommit(vp->v_mount);
328 	}
329 	return rtvals[0];
330 }
331 
332 /*
333  * For nfs, cache consistency can only be maintained approximately.
334  * Although RFC1094 does not specify the criteria, the following is
335  * believed to be compatible with the reference port.
336  * For nfs:
337  * If the file's modify time on the server has changed since the
338  * last read rpc or you have written to the file,
339  * you may have lost data cache consistency with the
340  * server, so flush all of the file's data out of the cache.
341  * Then force a getattr rpc to ensure that you have up to date
342  * attributes.
343  * NB: This implies that cache data can be read when up to
344  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
345  * attributes this could be forced by setting n_attrstamp to 0 before
346  * the VOP_GETATTR() call.
347  */
348 static inline int
349 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
350 {
351 	int error = 0;
352 	struct vattr vattr;
353 	struct nfsnode *np = VTONFS(vp);
354 	int old_lock;
355 
356 	/*
357 	 * Grab the exclusive lock before checking whether the cache is
358 	 * consistent.
359 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
360 	 * But for now, this suffices.
361 	 */
362 	old_lock = ncl_upgrade_vnlock(vp);
363 	if (vp->v_iflag & VI_DOOMED) {
364 		ncl_downgrade_vnlock(vp, old_lock);
365 		return (EBADF);
366 	}
367 
368 	mtx_lock(&np->n_mtx);
369 	if (np->n_flag & NMODIFIED) {
370 		mtx_unlock(&np->n_mtx);
371 		if (vp->v_type != VREG) {
372 			if (vp->v_type != VDIR)
373 				panic("nfs: bioread, not dir");
374 			ncl_invaldir(vp);
375 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
376 			if (error)
377 				goto out;
378 		}
379 		np->n_attrstamp = 0;
380 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
381 		error = VOP_GETATTR(vp, &vattr, cred);
382 		if (error)
383 			goto out;
384 		mtx_lock(&np->n_mtx);
385 		np->n_mtime = vattr.va_mtime;
386 		mtx_unlock(&np->n_mtx);
387 	} else {
388 		mtx_unlock(&np->n_mtx);
389 		error = VOP_GETATTR(vp, &vattr, cred);
390 		if (error)
391 			return (error);
392 		mtx_lock(&np->n_mtx);
393 		if ((np->n_flag & NSIZECHANGED)
394 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
395 			mtx_unlock(&np->n_mtx);
396 			if (vp->v_type == VDIR)
397 				ncl_invaldir(vp);
398 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
399 			if (error)
400 				goto out;
401 			mtx_lock(&np->n_mtx);
402 			np->n_mtime = vattr.va_mtime;
403 			np->n_flag &= ~NSIZECHANGED;
404 		}
405 		mtx_unlock(&np->n_mtx);
406 	}
407 out:
408 	ncl_downgrade_vnlock(vp, old_lock);
409 	return error;
410 }
411 
412 /*
413  * Vnode op for read using bio
414  */
415 int
416 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
417 {
418 	struct nfsnode *np = VTONFS(vp);
419 	int biosize, i;
420 	struct buf *bp, *rabp;
421 	struct thread *td;
422 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
423 	daddr_t lbn, rabn;
424 	int bcount;
425 	int seqcount;
426 	int nra, error = 0, n = 0, on = 0;
427 	off_t tmp_off;
428 
429 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
430 	if (uio->uio_resid == 0)
431 		return (0);
432 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
433 		return (EINVAL);
434 	td = uio->uio_td;
435 
436 	mtx_lock(&nmp->nm_mtx);
437 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
438 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
439 		mtx_unlock(&nmp->nm_mtx);
440 		(void)ncl_fsinfo(nmp, vp, cred, td);
441 		mtx_lock(&nmp->nm_mtx);
442 	}
443 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
444 		(void) newnfs_iosize(nmp);
445 
446 	tmp_off = uio->uio_offset + uio->uio_resid;
447 	if (vp->v_type != VDIR &&
448 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
449 		mtx_unlock(&nmp->nm_mtx);
450 		return (EFBIG);
451 	}
452 	mtx_unlock(&nmp->nm_mtx);
453 
454 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
455 		/* No caching/ no readaheads. Just read data into the user buffer */
456 		return ncl_readrpc(vp, uio, cred);
457 
458 	biosize = vp->v_bufobj.bo_bsize;
459 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
460 
461 	error = nfs_bioread_check_cons(vp, td, cred);
462 	if (error)
463 		return error;
464 
465 	do {
466 	    u_quad_t nsize;
467 
468 	    mtx_lock(&np->n_mtx);
469 	    nsize = np->n_size;
470 	    mtx_unlock(&np->n_mtx);
471 
472 	    switch (vp->v_type) {
473 	    case VREG:
474 		NFSINCRGLOBAL(newnfsstats.biocache_reads);
475 		lbn = uio->uio_offset / biosize;
476 		on = uio->uio_offset - (lbn * biosize);
477 
478 		/*
479 		 * Start the read ahead(s), as required.
480 		 */
481 		if (nmp->nm_readahead > 0) {
482 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
483 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
484 			rabn = lbn + 1 + nra;
485 			if (incore(&vp->v_bufobj, rabn) == NULL) {
486 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
487 			    if (!rabp) {
488 				error = newnfs_sigintr(nmp, td);
489 				return (error ? error : EINTR);
490 			    }
491 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
492 				rabp->b_flags |= B_ASYNC;
493 				rabp->b_iocmd = BIO_READ;
494 				vfs_busy_pages(rabp, 0);
495 				if (ncl_asyncio(nmp, rabp, cred, td)) {
496 				    rabp->b_flags |= B_INVAL;
497 				    rabp->b_ioflags |= BIO_ERROR;
498 				    vfs_unbusy_pages(rabp);
499 				    brelse(rabp);
500 				    break;
501 				}
502 			    } else {
503 				brelse(rabp);
504 			    }
505 			}
506 		    }
507 		}
508 
509 		/* Note that bcount is *not* DEV_BSIZE aligned. */
510 		bcount = biosize;
511 		if ((off_t)lbn * biosize >= nsize) {
512 			bcount = 0;
513 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
514 			bcount = nsize - (off_t)lbn * biosize;
515 		}
516 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
517 
518 		if (!bp) {
519 			error = newnfs_sigintr(nmp, td);
520 			return (error ? error : EINTR);
521 		}
522 
523 		/*
524 		 * If B_CACHE is not set, we must issue the read.  If this
525 		 * fails, we return an error.
526 		 */
527 
528 		if ((bp->b_flags & B_CACHE) == 0) {
529 		    bp->b_iocmd = BIO_READ;
530 		    vfs_busy_pages(bp, 0);
531 		    error = ncl_doio(vp, bp, cred, td, 0);
532 		    if (error) {
533 			brelse(bp);
534 			return (error);
535 		    }
536 		}
537 
538 		/*
539 		 * on is the offset into the current bp.  Figure out how many
540 		 * bytes we can copy out of the bp.  Note that bcount is
541 		 * NOT DEV_BSIZE aligned.
542 		 *
543 		 * Then figure out how many bytes we can copy into the uio.
544 		 */
545 
546 		n = 0;
547 		if (on < bcount)
548 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
549 		break;
550 	    case VLNK:
551 		NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
552 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
553 		if (!bp) {
554 			error = newnfs_sigintr(nmp, td);
555 			return (error ? error : EINTR);
556 		}
557 		if ((bp->b_flags & B_CACHE) == 0) {
558 		    bp->b_iocmd = BIO_READ;
559 		    vfs_busy_pages(bp, 0);
560 		    error = ncl_doio(vp, bp, cred, td, 0);
561 		    if (error) {
562 			bp->b_ioflags |= BIO_ERROR;
563 			brelse(bp);
564 			return (error);
565 		    }
566 		}
567 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
568 		on = 0;
569 		break;
570 	    case VDIR:
571 		NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
572 		if (np->n_direofoffset
573 		    && uio->uio_offset >= np->n_direofoffset) {
574 		    return (0);
575 		}
576 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
577 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
578 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
579 		if (!bp) {
580 		    error = newnfs_sigintr(nmp, td);
581 		    return (error ? error : EINTR);
582 		}
583 		if ((bp->b_flags & B_CACHE) == 0) {
584 		    bp->b_iocmd = BIO_READ;
585 		    vfs_busy_pages(bp, 0);
586 		    error = ncl_doio(vp, bp, cred, td, 0);
587 		    if (error) {
588 			    brelse(bp);
589 		    }
590 		    while (error == NFSERR_BAD_COOKIE) {
591 			ncl_invaldir(vp);
592 			error = ncl_vinvalbuf(vp, 0, td, 1);
593 			/*
594 			 * Yuck! The directory has been modified on the
595 			 * server. The only way to get the block is by
596 			 * reading from the beginning to get all the
597 			 * offset cookies.
598 			 *
599 			 * Leave the last bp intact unless there is an error.
600 			 * Loop back up to the while if the error is another
601 			 * NFSERR_BAD_COOKIE (double yuch!).
602 			 */
603 			for (i = 0; i <= lbn && !error; i++) {
604 			    if (np->n_direofoffset
605 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
606 				    return (0);
607 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
608 			    if (!bp) {
609 				error = newnfs_sigintr(nmp, td);
610 				return (error ? error : EINTR);
611 			    }
612 			    if ((bp->b_flags & B_CACHE) == 0) {
613 				    bp->b_iocmd = BIO_READ;
614 				    vfs_busy_pages(bp, 0);
615 				    error = ncl_doio(vp, bp, cred, td, 0);
616 				    /*
617 				     * no error + B_INVAL == directory EOF,
618 				     * use the block.
619 				     */
620 				    if (error == 0 && (bp->b_flags & B_INVAL))
621 					    break;
622 			    }
623 			    /*
624 			     * An error will throw away the block and the
625 			     * for loop will break out.  If no error and this
626 			     * is not the block we want, we throw away the
627 			     * block and go for the next one via the for loop.
628 			     */
629 			    if (error || i < lbn)
630 				    brelse(bp);
631 			}
632 		    }
633 		    /*
634 		     * The above while is repeated if we hit another cookie
635 		     * error.  If we hit an error and it wasn't a cookie error,
636 		     * we give up.
637 		     */
638 		    if (error)
639 			    return (error);
640 		}
641 
642 		/*
643 		 * If not eof and read aheads are enabled, start one.
644 		 * (You need the current block first, so that you have the
645 		 *  directory offset cookie of the next block.)
646 		 */
647 		if (nmp->nm_readahead > 0 &&
648 		    (bp->b_flags & B_INVAL) == 0 &&
649 		    (np->n_direofoffset == 0 ||
650 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
651 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
652 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
653 			if (rabp) {
654 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
655 				rabp->b_flags |= B_ASYNC;
656 				rabp->b_iocmd = BIO_READ;
657 				vfs_busy_pages(rabp, 0);
658 				if (ncl_asyncio(nmp, rabp, cred, td)) {
659 				    rabp->b_flags |= B_INVAL;
660 				    rabp->b_ioflags |= BIO_ERROR;
661 				    vfs_unbusy_pages(rabp);
662 				    brelse(rabp);
663 				}
664 			    } else {
665 				brelse(rabp);
666 			    }
667 			}
668 		}
669 		/*
670 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
671 		 * chopped for the EOF condition, we cannot tell how large
672 		 * NFS directories are going to be until we hit EOF.  So
673 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
674 		 * it just so happens that b_resid will effectively chop it
675 		 * to EOF.  *BUT* this information is lost if the buffer goes
676 		 * away and is reconstituted into a B_CACHE state ( due to
677 		 * being VMIO ) later.  So we keep track of the directory eof
678 		 * in np->n_direofoffset and chop it off as an extra step
679 		 * right here.
680 		 */
681 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
682 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
683 			n = np->n_direofoffset - uio->uio_offset;
684 		break;
685 	    default:
686 		ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
687 		bp = NULL;
688 		break;
689 	    };
690 
691 	    if (n > 0) {
692 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
693 	    }
694 	    if (vp->v_type == VLNK)
695 		n = 0;
696 	    if (bp != NULL)
697 		brelse(bp);
698 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
699 	return (error);
700 }
701 
702 /*
703  * The NFS write path cannot handle iovecs with len > 1. So we need to
704  * break up iovecs accordingly (restricting them to wsize).
705  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
706  * For the ASYNC case, 2 copies are needed. The first a copy from the
707  * user buffer to a staging buffer and then a second copy from the staging
708  * buffer to mbufs. This can be optimized by copying from the user buffer
709  * directly into mbufs and passing the chain down, but that requires a
710  * fair amount of re-working of the relevant codepaths (and can be done
711  * later).
712  */
713 static int
714 nfs_directio_write(vp, uiop, cred, ioflag)
715 	struct vnode *vp;
716 	struct uio *uiop;
717 	struct ucred *cred;
718 	int ioflag;
719 {
720 	int error;
721 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
722 	struct thread *td = uiop->uio_td;
723 	int size;
724 	int wsize;
725 
726 	mtx_lock(&nmp->nm_mtx);
727 	wsize = nmp->nm_wsize;
728 	mtx_unlock(&nmp->nm_mtx);
729 	if (ioflag & IO_SYNC) {
730 		int iomode, must_commit;
731 		struct uio uio;
732 		struct iovec iov;
733 do_sync:
734 		while (uiop->uio_resid > 0) {
735 			size = MIN(uiop->uio_resid, wsize);
736 			size = MIN(uiop->uio_iov->iov_len, size);
737 			iov.iov_base = uiop->uio_iov->iov_base;
738 			iov.iov_len = size;
739 			uio.uio_iov = &iov;
740 			uio.uio_iovcnt = 1;
741 			uio.uio_offset = uiop->uio_offset;
742 			uio.uio_resid = size;
743 			uio.uio_segflg = UIO_USERSPACE;
744 			uio.uio_rw = UIO_WRITE;
745 			uio.uio_td = td;
746 			iomode = NFSWRITE_FILESYNC;
747 			error = ncl_writerpc(vp, &uio, cred, &iomode,
748 			    &must_commit, 0);
749 			KASSERT((must_commit == 0),
750 				("ncl_directio_write: Did not commit write"));
751 			if (error)
752 				return (error);
753 			uiop->uio_offset += size;
754 			uiop->uio_resid -= size;
755 			if (uiop->uio_iov->iov_len <= size) {
756 				uiop->uio_iovcnt--;
757 				uiop->uio_iov++;
758 			} else {
759 				uiop->uio_iov->iov_base =
760 					(char *)uiop->uio_iov->iov_base + size;
761 				uiop->uio_iov->iov_len -= size;
762 			}
763 		}
764 	} else {
765 		struct uio *t_uio;
766 		struct iovec *t_iov;
767 		struct buf *bp;
768 
769 		/*
770 		 * Break up the write into blocksize chunks and hand these
771 		 * over to nfsiod's for write back.
772 		 * Unfortunately, this incurs a copy of the data. Since
773 		 * the user could modify the buffer before the write is
774 		 * initiated.
775 		 *
776 		 * The obvious optimization here is that one of the 2 copies
777 		 * in the async write path can be eliminated by copying the
778 		 * data here directly into mbufs and passing the mbuf chain
779 		 * down. But that will require a fair amount of re-working
780 		 * of the code and can be done if there's enough interest
781 		 * in NFS directio access.
782 		 */
783 		while (uiop->uio_resid > 0) {
784 			size = MIN(uiop->uio_resid, wsize);
785 			size = MIN(uiop->uio_iov->iov_len, size);
786 			bp = getpbuf(&ncl_pbuf_freecnt);
787 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
788 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
789 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
790 			t_iov->iov_len = size;
791 			t_uio->uio_iov = t_iov;
792 			t_uio->uio_iovcnt = 1;
793 			t_uio->uio_offset = uiop->uio_offset;
794 			t_uio->uio_resid = size;
795 			t_uio->uio_segflg = UIO_SYSSPACE;
796 			t_uio->uio_rw = UIO_WRITE;
797 			t_uio->uio_td = td;
798 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
799 			    uiop->uio_segflg == UIO_SYSSPACE,
800 			    ("nfs_directio_write: Bad uio_segflg"));
801 			if (uiop->uio_segflg == UIO_USERSPACE) {
802 				error = copyin(uiop->uio_iov->iov_base,
803 				    t_iov->iov_base, size);
804 				if (error != 0)
805 					goto err_free;
806 			} else
807 				/*
808 				 * UIO_SYSSPACE may never happen, but handle
809 				 * it just in case it does.
810 				 */
811 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
812 				    size);
813 			bp->b_flags |= B_DIRECT;
814 			bp->b_iocmd = BIO_WRITE;
815 			if (cred != NOCRED) {
816 				crhold(cred);
817 				bp->b_wcred = cred;
818 			} else
819 				bp->b_wcred = NOCRED;
820 			bp->b_caller1 = (void *)t_uio;
821 			bp->b_vp = vp;
822 			error = ncl_asyncio(nmp, bp, NOCRED, td);
823 err_free:
824 			if (error) {
825 				free(t_iov->iov_base, M_NFSDIRECTIO);
826 				free(t_iov, M_NFSDIRECTIO);
827 				free(t_uio, M_NFSDIRECTIO);
828 				bp->b_vp = NULL;
829 				relpbuf(bp, &ncl_pbuf_freecnt);
830 				if (error == EINTR)
831 					return (error);
832 				goto do_sync;
833 			}
834 			uiop->uio_offset += size;
835 			uiop->uio_resid -= size;
836 			if (uiop->uio_iov->iov_len <= size) {
837 				uiop->uio_iovcnt--;
838 				uiop->uio_iov++;
839 			} else {
840 				uiop->uio_iov->iov_base =
841 					(char *)uiop->uio_iov->iov_base + size;
842 				uiop->uio_iov->iov_len -= size;
843 			}
844 		}
845 	}
846 	return (0);
847 }
848 
849 /*
850  * Vnode op for write using bio
851  */
852 int
853 ncl_write(struct vop_write_args *ap)
854 {
855 	int biosize;
856 	struct uio *uio = ap->a_uio;
857 	struct thread *td = uio->uio_td;
858 	struct vnode *vp = ap->a_vp;
859 	struct nfsnode *np = VTONFS(vp);
860 	struct ucred *cred = ap->a_cred;
861 	int ioflag = ap->a_ioflag;
862 	struct buf *bp;
863 	struct vattr vattr;
864 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
865 	daddr_t lbn;
866 	int bcount, noncontig_write, obcount;
867 	int bp_cached, n, on, error = 0, error1;
868 	size_t orig_resid, local_resid;
869 	off_t orig_size, tmp_off;
870 
871 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
872 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
873 	    ("ncl_write proc"));
874 	if (vp->v_type != VREG)
875 		return (EIO);
876 	mtx_lock(&np->n_mtx);
877 	if (np->n_flag & NWRITEERR) {
878 		np->n_flag &= ~NWRITEERR;
879 		mtx_unlock(&np->n_mtx);
880 		return (np->n_error);
881 	} else
882 		mtx_unlock(&np->n_mtx);
883 	mtx_lock(&nmp->nm_mtx);
884 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
885 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
886 		mtx_unlock(&nmp->nm_mtx);
887 		(void)ncl_fsinfo(nmp, vp, cred, td);
888 		mtx_lock(&nmp->nm_mtx);
889 	}
890 	if (nmp->nm_wsize == 0)
891 		(void) newnfs_iosize(nmp);
892 	mtx_unlock(&nmp->nm_mtx);
893 
894 	/*
895 	 * Synchronously flush pending buffers if we are in synchronous
896 	 * mode or if we are appending.
897 	 */
898 	if (ioflag & (IO_APPEND | IO_SYNC)) {
899 		mtx_lock(&np->n_mtx);
900 		if (np->n_flag & NMODIFIED) {
901 			mtx_unlock(&np->n_mtx);
902 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
903 			/*
904 			 * Require non-blocking, synchronous writes to
905 			 * dirty files to inform the program it needs
906 			 * to fsync(2) explicitly.
907 			 */
908 			if (ioflag & IO_NDELAY)
909 				return (EAGAIN);
910 #endif
911 flush_and_restart:
912 			np->n_attrstamp = 0;
913 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
914 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
915 			if (error)
916 				return (error);
917 		} else
918 			mtx_unlock(&np->n_mtx);
919 	}
920 
921 	orig_resid = uio->uio_resid;
922 	mtx_lock(&np->n_mtx);
923 	orig_size = np->n_size;
924 	mtx_unlock(&np->n_mtx);
925 
926 	/*
927 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
928 	 * get the append lock.
929 	 */
930 	if (ioflag & IO_APPEND) {
931 		np->n_attrstamp = 0;
932 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
933 		error = VOP_GETATTR(vp, &vattr, cred);
934 		if (error)
935 			return (error);
936 		mtx_lock(&np->n_mtx);
937 		uio->uio_offset = np->n_size;
938 		mtx_unlock(&np->n_mtx);
939 	}
940 
941 	if (uio->uio_offset < 0)
942 		return (EINVAL);
943 	tmp_off = uio->uio_offset + uio->uio_resid;
944 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
945 		return (EFBIG);
946 	if (uio->uio_resid == 0)
947 		return (0);
948 
949 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
950 		return nfs_directio_write(vp, uio, cred, ioflag);
951 
952 	/*
953 	 * Maybe this should be above the vnode op call, but so long as
954 	 * file servers have no limits, i don't think it matters
955 	 */
956 	if (vn_rlimit_fsize(vp, uio, td))
957 		return (EFBIG);
958 
959 	biosize = vp->v_bufobj.bo_bsize;
960 	/*
961 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
962 	 * would exceed the local maximum per-file write commit size when
963 	 * combined with those, we must decide whether to flush,
964 	 * go synchronous, or return error.  We don't bother checking
965 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
966 	 * no point optimizing for something that really won't ever happen.
967 	 */
968 	if (!(ioflag & IO_SYNC)) {
969 		int nflag;
970 
971 		mtx_lock(&np->n_mtx);
972 		nflag = np->n_flag;
973 		mtx_unlock(&np->n_mtx);
974 		int needrestart = 0;
975 		if (nmp->nm_wcommitsize < uio->uio_resid) {
976 			/*
977 			 * If this request could not possibly be completed
978 			 * without exceeding the maximum outstanding write
979 			 * commit size, see if we can convert it into a
980 			 * synchronous write operation.
981 			 */
982 			if (ioflag & IO_NDELAY)
983 				return (EAGAIN);
984 			ioflag |= IO_SYNC;
985 			if (nflag & NMODIFIED)
986 				needrestart = 1;
987 		} else if (nflag & NMODIFIED) {
988 			int wouldcommit = 0;
989 			BO_LOCK(&vp->v_bufobj);
990 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
991 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
992 				    b_bobufs) {
993 					if (bp->b_flags & B_NEEDCOMMIT)
994 						wouldcommit += bp->b_bcount;
995 				}
996 			}
997 			BO_UNLOCK(&vp->v_bufobj);
998 			/*
999 			 * Since we're not operating synchronously and
1000 			 * bypassing the buffer cache, we are in a commit
1001 			 * and holding all of these buffers whether
1002 			 * transmitted or not.  If not limited, this
1003 			 * will lead to the buffer cache deadlocking,
1004 			 * as no one else can flush our uncommitted buffers.
1005 			 */
1006 			wouldcommit += uio->uio_resid;
1007 			/*
1008 			 * If we would initially exceed the maximum
1009 			 * outstanding write commit size, flush and restart.
1010 			 */
1011 			if (wouldcommit > nmp->nm_wcommitsize)
1012 				needrestart = 1;
1013 		}
1014 		if (needrestart)
1015 			goto flush_and_restart;
1016 	}
1017 
1018 	do {
1019 		NFSINCRGLOBAL(newnfsstats.biocache_writes);
1020 		lbn = uio->uio_offset / biosize;
1021 		on = uio->uio_offset - (lbn * biosize);
1022 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1023 again:
1024 		/*
1025 		 * Handle direct append and file extension cases, calculate
1026 		 * unaligned buffer size.
1027 		 */
1028 		mtx_lock(&np->n_mtx);
1029 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1030 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1031 			noncontig_write = 1;
1032 		else
1033 			noncontig_write = 0;
1034 		if ((uio->uio_offset == np->n_size ||
1035 		    (noncontig_write != 0 &&
1036 		    lbn == (np->n_size / biosize) &&
1037 		    uio->uio_offset + n > np->n_size)) && n) {
1038 			mtx_unlock(&np->n_mtx);
1039 			/*
1040 			 * Get the buffer (in its pre-append state to maintain
1041 			 * B_CACHE if it was previously set).  Resize the
1042 			 * nfsnode after we have locked the buffer to prevent
1043 			 * readers from reading garbage.
1044 			 */
1045 			obcount = np->n_size - (lbn * biosize);
1046 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1047 
1048 			if (bp != NULL) {
1049 				long save;
1050 
1051 				mtx_lock(&np->n_mtx);
1052 				np->n_size = uio->uio_offset + n;
1053 				np->n_flag |= NMODIFIED;
1054 				vnode_pager_setsize(vp, np->n_size);
1055 				mtx_unlock(&np->n_mtx);
1056 
1057 				save = bp->b_flags & B_CACHE;
1058 				bcount = on + n;
1059 				allocbuf(bp, bcount);
1060 				bp->b_flags |= save;
1061 				if (noncontig_write != 0 && on > obcount)
1062 					vfs_bio_bzero_buf(bp, obcount, on -
1063 					    obcount);
1064 			}
1065 		} else {
1066 			/*
1067 			 * Obtain the locked cache block first, and then
1068 			 * adjust the file's size as appropriate.
1069 			 */
1070 			bcount = on + n;
1071 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1072 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1073 					bcount = biosize;
1074 				else
1075 					bcount = np->n_size - (off_t)lbn * biosize;
1076 			}
1077 			mtx_unlock(&np->n_mtx);
1078 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1079 			mtx_lock(&np->n_mtx);
1080 			if (uio->uio_offset + n > np->n_size) {
1081 				np->n_size = uio->uio_offset + n;
1082 				np->n_flag |= NMODIFIED;
1083 				vnode_pager_setsize(vp, np->n_size);
1084 			}
1085 			mtx_unlock(&np->n_mtx);
1086 		}
1087 
1088 		if (!bp) {
1089 			error = newnfs_sigintr(nmp, td);
1090 			if (!error)
1091 				error = EINTR;
1092 			break;
1093 		}
1094 
1095 		/*
1096 		 * Issue a READ if B_CACHE is not set.  In special-append
1097 		 * mode, B_CACHE is based on the buffer prior to the write
1098 		 * op and is typically set, avoiding the read.  If a read
1099 		 * is required in special append mode, the server will
1100 		 * probably send us a short-read since we extended the file
1101 		 * on our end, resulting in b_resid == 0 and, thusly,
1102 		 * B_CACHE getting set.
1103 		 *
1104 		 * We can also avoid issuing the read if the write covers
1105 		 * the entire buffer.  We have to make sure the buffer state
1106 		 * is reasonable in this case since we will not be initiating
1107 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1108 		 * more information.
1109 		 *
1110 		 * B_CACHE may also be set due to the buffer being cached
1111 		 * normally.
1112 		 */
1113 
1114 		bp_cached = 1;
1115 		if (on == 0 && n == bcount) {
1116 			if ((bp->b_flags & B_CACHE) == 0)
1117 				bp_cached = 0;
1118 			bp->b_flags |= B_CACHE;
1119 			bp->b_flags &= ~B_INVAL;
1120 			bp->b_ioflags &= ~BIO_ERROR;
1121 		}
1122 
1123 		if ((bp->b_flags & B_CACHE) == 0) {
1124 			bp->b_iocmd = BIO_READ;
1125 			vfs_busy_pages(bp, 0);
1126 			error = ncl_doio(vp, bp, cred, td, 0);
1127 			if (error) {
1128 				brelse(bp);
1129 				break;
1130 			}
1131 		}
1132 		if (bp->b_wcred == NOCRED)
1133 			bp->b_wcred = crhold(cred);
1134 		mtx_lock(&np->n_mtx);
1135 		np->n_flag |= NMODIFIED;
1136 		mtx_unlock(&np->n_mtx);
1137 
1138 		/*
1139 		 * If dirtyend exceeds file size, chop it down.  This should
1140 		 * not normally occur but there is an append race where it
1141 		 * might occur XXX, so we log it.
1142 		 *
1143 		 * If the chopping creates a reverse-indexed or degenerate
1144 		 * situation with dirtyoff/end, we 0 both of them.
1145 		 */
1146 
1147 		if (bp->b_dirtyend > bcount) {
1148 			ncl_printf("NFS append race @%lx:%d\n",
1149 			    (long)bp->b_blkno * DEV_BSIZE,
1150 			    bp->b_dirtyend - bcount);
1151 			bp->b_dirtyend = bcount;
1152 		}
1153 
1154 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1155 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1156 
1157 		/*
1158 		 * If the new write will leave a contiguous dirty
1159 		 * area, just update the b_dirtyoff and b_dirtyend,
1160 		 * otherwise force a write rpc of the old dirty area.
1161 		 *
1162 		 * If there has been a file lock applied to this file
1163 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1164 		 * While it is possible to merge discontiguous writes due to
1165 		 * our having a B_CACHE buffer ( and thus valid read data
1166 		 * for the hole), we don't because it could lead to
1167 		 * significant cache coherency problems with multiple clients,
1168 		 * especially if locking is implemented later on.
1169 		 *
1170 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1171 		 * not been file locking done on this file:
1172 		 * Relax coherency a bit for the sake of performance and
1173 		 * expand the current dirty region to contain the new
1174 		 * write even if it means we mark some non-dirty data as
1175 		 * dirty.
1176 		 */
1177 
1178 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1179 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1180 			if (bwrite(bp) == EINTR) {
1181 				error = EINTR;
1182 				break;
1183 			}
1184 			goto again;
1185 		}
1186 
1187 		local_resid = uio->uio_resid;
1188 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1189 
1190 		if (error != 0 && !bp_cached) {
1191 			/*
1192 			 * This block has no other content then what
1193 			 * possibly was written by the faulty uiomove.
1194 			 * Release it, forgetting the data pages, to
1195 			 * prevent the leak of uninitialized data to
1196 			 * usermode.
1197 			 */
1198 			bp->b_ioflags |= BIO_ERROR;
1199 			brelse(bp);
1200 			uio->uio_offset -= local_resid - uio->uio_resid;
1201 			uio->uio_resid = local_resid;
1202 			break;
1203 		}
1204 
1205 		/*
1206 		 * Since this block is being modified, it must be written
1207 		 * again and not just committed.  Since write clustering does
1208 		 * not work for the stage 1 data write, only the stage 2
1209 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1210 		 */
1211 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1212 
1213 		/*
1214 		 * Get the partial update on the progress made from
1215 		 * uiomove, if an error occured.
1216 		 */
1217 		if (error != 0)
1218 			n = local_resid - uio->uio_resid;
1219 
1220 		/*
1221 		 * Only update dirtyoff/dirtyend if not a degenerate
1222 		 * condition.
1223 		 */
1224 		if (n > 0) {
1225 			if (bp->b_dirtyend > 0) {
1226 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1227 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1228 			} else {
1229 				bp->b_dirtyoff = on;
1230 				bp->b_dirtyend = on + n;
1231 			}
1232 			vfs_bio_set_valid(bp, on, n);
1233 		}
1234 
1235 		/*
1236 		 * If IO_SYNC do bwrite().
1237 		 *
1238 		 * IO_INVAL appears to be unused.  The idea appears to be
1239 		 * to turn off caching in this case.  Very odd.  XXX
1240 		 */
1241 		if ((ioflag & IO_SYNC)) {
1242 			if (ioflag & IO_INVAL)
1243 				bp->b_flags |= B_NOCACHE;
1244 			error1 = bwrite(bp);
1245 			if (error1 != 0) {
1246 				if (error == 0)
1247 					error = error1;
1248 				break;
1249 			}
1250 		} else if ((n + on) == biosize) {
1251 			bp->b_flags |= B_ASYNC;
1252 			(void) ncl_writebp(bp, 0, NULL);
1253 		} else {
1254 			bdwrite(bp);
1255 		}
1256 
1257 		if (error != 0)
1258 			break;
1259 	} while (uio->uio_resid > 0 && n > 0);
1260 
1261 	if (error != 0) {
1262 		if (ioflag & IO_UNIT) {
1263 			VATTR_NULL(&vattr);
1264 			vattr.va_size = orig_size;
1265 			/* IO_SYNC is handled implicitely */
1266 			(void)VOP_SETATTR(vp, &vattr, cred);
1267 			uio->uio_offset -= orig_resid - uio->uio_resid;
1268 			uio->uio_resid = orig_resid;
1269 		}
1270 	}
1271 
1272 	return (error);
1273 }
1274 
1275 /*
1276  * Get an nfs cache block.
1277  *
1278  * Allocate a new one if the block isn't currently in the cache
1279  * and return the block marked busy. If the calling process is
1280  * interrupted by a signal for an interruptible mount point, return
1281  * NULL.
1282  *
1283  * The caller must carefully deal with the possible B_INVAL state of
1284  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1285  * indirectly), so synchronous reads can be issued without worrying about
1286  * the B_INVAL state.  We have to be a little more careful when dealing
1287  * with writes (see comments in nfs_write()) when extending a file past
1288  * its EOF.
1289  */
1290 static struct buf *
1291 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1292 {
1293 	struct buf *bp;
1294 	struct mount *mp;
1295 	struct nfsmount *nmp;
1296 
1297 	mp = vp->v_mount;
1298 	nmp = VFSTONFS(mp);
1299 
1300 	if (nmp->nm_flag & NFSMNT_INT) {
1301 		sigset_t oldset;
1302 
1303 		newnfs_set_sigmask(td, &oldset);
1304 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1305 		newnfs_restore_sigmask(td, &oldset);
1306 		while (bp == NULL) {
1307 			if (newnfs_sigintr(nmp, td))
1308 				return (NULL);
1309 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1310 		}
1311 	} else {
1312 		bp = getblk(vp, bn, size, 0, 0, 0);
1313 	}
1314 
1315 	if (vp->v_type == VREG)
1316 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1317 	return (bp);
1318 }
1319 
1320 /*
1321  * Flush and invalidate all dirty buffers. If another process is already
1322  * doing the flush, just wait for completion.
1323  */
1324 int
1325 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1326 {
1327 	struct nfsnode *np = VTONFS(vp);
1328 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1329 	int error = 0, slpflag, slptimeo;
1330 	int old_lock = 0;
1331 
1332 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1333 
1334 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1335 		intrflg = 0;
1336 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1337 		intrflg = 1;
1338 	if (intrflg) {
1339 		slpflag = PCATCH;
1340 		slptimeo = 2 * hz;
1341 	} else {
1342 		slpflag = 0;
1343 		slptimeo = 0;
1344 	}
1345 
1346 	old_lock = ncl_upgrade_vnlock(vp);
1347 	if (vp->v_iflag & VI_DOOMED) {
1348 		/*
1349 		 * Since vgonel() uses the generic vinvalbuf() to flush
1350 		 * dirty buffers and it does not call this function, it
1351 		 * is safe to just return OK when VI_DOOMED is set.
1352 		 */
1353 		ncl_downgrade_vnlock(vp, old_lock);
1354 		return (0);
1355 	}
1356 
1357 	/*
1358 	 * Now, flush as required.
1359 	 */
1360 	if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1361 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1362 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1363 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1364 		/*
1365 		 * If the page clean was interrupted, fail the invalidation.
1366 		 * Not doing so, we run the risk of losing dirty pages in the
1367 		 * vinvalbuf() call below.
1368 		 */
1369 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1370 			goto out;
1371 	}
1372 
1373 	error = vinvalbuf(vp, flags, slpflag, 0);
1374 	while (error) {
1375 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1376 			goto out;
1377 		error = vinvalbuf(vp, flags, 0, slptimeo);
1378 	}
1379 	if (NFSHASPNFS(nmp)) {
1380 		nfscl_layoutcommit(vp, td);
1381 		/*
1382 		 * Invalidate the attribute cache, since writes to a DS
1383 		 * won't update the size attribute.
1384 		 */
1385 		mtx_lock(&np->n_mtx);
1386 		np->n_attrstamp = 0;
1387 	} else
1388 		mtx_lock(&np->n_mtx);
1389 	if (np->n_directio_asyncwr == 0)
1390 		np->n_flag &= ~NMODIFIED;
1391 	mtx_unlock(&np->n_mtx);
1392 out:
1393 	ncl_downgrade_vnlock(vp, old_lock);
1394 	return error;
1395 }
1396 
1397 /*
1398  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1399  * This is mainly to avoid queueing async I/O requests when the nfsiods
1400  * are all hung on a dead server.
1401  *
1402  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1403  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1404  */
1405 int
1406 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1407 {
1408 	int iod;
1409 	int gotiod;
1410 	int slpflag = 0;
1411 	int slptimeo = 0;
1412 	int error, error2;
1413 
1414 	/*
1415 	 * Commits are usually short and sweet so lets save some cpu and
1416 	 * leave the async daemons for more important rpc's (such as reads
1417 	 * and writes).
1418 	 *
1419 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1420 	 * in the directory in order to update attributes. This can deadlock
1421 	 * with another thread that is waiting for async I/O to be done by
1422 	 * an nfsiod thread while holding a lock on one of these vnodes.
1423 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1424 	 * perform Readdirplus RPCs.
1425 	 */
1426 	mtx_lock(&ncl_iod_mutex);
1427 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1428 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1429 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1430 		mtx_unlock(&ncl_iod_mutex);
1431 		return(EIO);
1432 	}
1433 again:
1434 	if (nmp->nm_flag & NFSMNT_INT)
1435 		slpflag = PCATCH;
1436 	gotiod = FALSE;
1437 
1438 	/*
1439 	 * Find a free iod to process this request.
1440 	 */
1441 	for (iod = 0; iod < ncl_numasync; iod++)
1442 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1443 			gotiod = TRUE;
1444 			break;
1445 		}
1446 
1447 	/*
1448 	 * Try to create one if none are free.
1449 	 */
1450 	if (!gotiod)
1451 		ncl_nfsiodnew();
1452 	else {
1453 		/*
1454 		 * Found one, so wake it up and tell it which
1455 		 * mount to process.
1456 		 */
1457 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1458 		    iod, nmp));
1459 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1460 		ncl_iodmount[iod] = nmp;
1461 		nmp->nm_bufqiods++;
1462 		wakeup(&ncl_iodwant[iod]);
1463 	}
1464 
1465 	/*
1466 	 * If none are free, we may already have an iod working on this mount
1467 	 * point.  If so, it will process our request.
1468 	 */
1469 	if (!gotiod) {
1470 		if (nmp->nm_bufqiods > 0) {
1471 			NFS_DPF(ASYNCIO,
1472 				("ncl_asyncio: %d iods are already processing mount %p\n",
1473 				 nmp->nm_bufqiods, nmp));
1474 			gotiod = TRUE;
1475 		}
1476 	}
1477 
1478 	/*
1479 	 * If we have an iod which can process the request, then queue
1480 	 * the buffer.
1481 	 */
1482 	if (gotiod) {
1483 		/*
1484 		 * Ensure that the queue never grows too large.  We still want
1485 		 * to asynchronize so we block rather then return EIO.
1486 		 */
1487 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1488 			NFS_DPF(ASYNCIO,
1489 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1490 			nmp->nm_bufqwant = TRUE;
1491 			error = newnfs_msleep(td, &nmp->nm_bufq,
1492 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1493 			   slptimeo);
1494 			if (error) {
1495 				error2 = newnfs_sigintr(nmp, td);
1496 				if (error2) {
1497 					mtx_unlock(&ncl_iod_mutex);
1498 					return (error2);
1499 				}
1500 				if (slpflag == PCATCH) {
1501 					slpflag = 0;
1502 					slptimeo = 2 * hz;
1503 				}
1504 			}
1505 			/*
1506 			 * We might have lost our iod while sleeping,
1507 			 * so check and loop if nescessary.
1508 			 */
1509 			goto again;
1510 		}
1511 
1512 		/* We might have lost our nfsiod */
1513 		if (nmp->nm_bufqiods == 0) {
1514 			NFS_DPF(ASYNCIO,
1515 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1516 			goto again;
1517 		}
1518 
1519 		if (bp->b_iocmd == BIO_READ) {
1520 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1521 				bp->b_rcred = crhold(cred);
1522 		} else {
1523 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1524 				bp->b_wcred = crhold(cred);
1525 		}
1526 
1527 		if (bp->b_flags & B_REMFREE)
1528 			bremfreef(bp);
1529 		BUF_KERNPROC(bp);
1530 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1531 		nmp->nm_bufqlen++;
1532 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1533 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1534 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1535 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1536 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1537 		}
1538 		mtx_unlock(&ncl_iod_mutex);
1539 		return (0);
1540 	}
1541 
1542 	mtx_unlock(&ncl_iod_mutex);
1543 
1544 	/*
1545 	 * All the iods are busy on other mounts, so return EIO to
1546 	 * force the caller to process the i/o synchronously.
1547 	 */
1548 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1549 	return (EIO);
1550 }
1551 
1552 void
1553 ncl_doio_directwrite(struct buf *bp)
1554 {
1555 	int iomode, must_commit;
1556 	struct uio *uiop = (struct uio *)bp->b_caller1;
1557 	char *iov_base = uiop->uio_iov->iov_base;
1558 
1559 	iomode = NFSWRITE_FILESYNC;
1560 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1561 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1562 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1563 	free(iov_base, M_NFSDIRECTIO);
1564 	free(uiop->uio_iov, M_NFSDIRECTIO);
1565 	free(uiop, M_NFSDIRECTIO);
1566 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1567 		struct nfsnode *np = VTONFS(bp->b_vp);
1568 		mtx_lock(&np->n_mtx);
1569 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1570 			/*
1571 			 * Invalidate the attribute cache, since writes to a DS
1572 			 * won't update the size attribute.
1573 			 */
1574 			np->n_attrstamp = 0;
1575 		}
1576 		np->n_directio_asyncwr--;
1577 		if (np->n_directio_asyncwr == 0) {
1578 			np->n_flag &= ~NMODIFIED;
1579 			if ((np->n_flag & NFSYNCWAIT)) {
1580 				np->n_flag &= ~NFSYNCWAIT;
1581 				wakeup((caddr_t)&np->n_directio_asyncwr);
1582 			}
1583 		}
1584 		mtx_unlock(&np->n_mtx);
1585 	}
1586 	bp->b_vp = NULL;
1587 	relpbuf(bp, &ncl_pbuf_freecnt);
1588 }
1589 
1590 /*
1591  * Do an I/O operation to/from a cache block. This may be called
1592  * synchronously or from an nfsiod.
1593  */
1594 int
1595 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1596     int called_from_strategy)
1597 {
1598 	struct uio *uiop;
1599 	struct nfsnode *np;
1600 	struct nfsmount *nmp;
1601 	int error = 0, iomode, must_commit = 0;
1602 	struct uio uio;
1603 	struct iovec io;
1604 	struct proc *p = td ? td->td_proc : NULL;
1605 	uint8_t	iocmd;
1606 
1607 	np = VTONFS(vp);
1608 	nmp = VFSTONFS(vp->v_mount);
1609 	uiop = &uio;
1610 	uiop->uio_iov = &io;
1611 	uiop->uio_iovcnt = 1;
1612 	uiop->uio_segflg = UIO_SYSSPACE;
1613 	uiop->uio_td = td;
1614 
1615 	/*
1616 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1617 	 * do this here so we do not have to do it in all the code that
1618 	 * calls us.
1619 	 */
1620 	bp->b_flags &= ~B_INVAL;
1621 	bp->b_ioflags &= ~BIO_ERROR;
1622 
1623 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1624 	iocmd = bp->b_iocmd;
1625 	if (iocmd == BIO_READ) {
1626 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1627 	    io.iov_base = bp->b_data;
1628 	    uiop->uio_rw = UIO_READ;
1629 
1630 	    switch (vp->v_type) {
1631 	    case VREG:
1632 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1633 		NFSINCRGLOBAL(newnfsstats.read_bios);
1634 		error = ncl_readrpc(vp, uiop, cr);
1635 
1636 		if (!error) {
1637 		    if (uiop->uio_resid) {
1638 			/*
1639 			 * If we had a short read with no error, we must have
1640 			 * hit a file hole.  We should zero-fill the remainder.
1641 			 * This can also occur if the server hits the file EOF.
1642 			 *
1643 			 * Holes used to be able to occur due to pending
1644 			 * writes, but that is not possible any longer.
1645 			 */
1646 			int nread = bp->b_bcount - uiop->uio_resid;
1647 			ssize_t left = uiop->uio_resid;
1648 
1649 			if (left > 0)
1650 				bzero((char *)bp->b_data + nread, left);
1651 			uiop->uio_resid = 0;
1652 		    }
1653 		}
1654 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1655 		if (p && (vp->v_vflag & VV_TEXT)) {
1656 			mtx_lock(&np->n_mtx);
1657 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1658 				mtx_unlock(&np->n_mtx);
1659 				PROC_LOCK(p);
1660 				killproc(p, "text file modification");
1661 				PROC_UNLOCK(p);
1662 			} else
1663 				mtx_unlock(&np->n_mtx);
1664 		}
1665 		break;
1666 	    case VLNK:
1667 		uiop->uio_offset = (off_t)0;
1668 		NFSINCRGLOBAL(newnfsstats.readlink_bios);
1669 		error = ncl_readlinkrpc(vp, uiop, cr);
1670 		break;
1671 	    case VDIR:
1672 		NFSINCRGLOBAL(newnfsstats.readdir_bios);
1673 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1674 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1675 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1676 			if (error == NFSERR_NOTSUPP)
1677 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1678 		}
1679 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1680 			error = ncl_readdirrpc(vp, uiop, cr, td);
1681 		/*
1682 		 * end-of-directory sets B_INVAL but does not generate an
1683 		 * error.
1684 		 */
1685 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1686 			bp->b_flags |= B_INVAL;
1687 		break;
1688 	    default:
1689 		ncl_printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1690 		break;
1691 	    };
1692 	    if (error) {
1693 		bp->b_ioflags |= BIO_ERROR;
1694 		bp->b_error = error;
1695 	    }
1696 	} else {
1697 	    /*
1698 	     * If we only need to commit, try to commit
1699 	     */
1700 	    if (bp->b_flags & B_NEEDCOMMIT) {
1701 		    int retv;
1702 		    off_t off;
1703 
1704 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1705 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1706 			bp->b_wcred, td);
1707 		    if (retv == 0) {
1708 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1709 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1710 			    bp->b_resid = 0;
1711 			    bufdone(bp);
1712 			    return (0);
1713 		    }
1714 		    if (retv == NFSERR_STALEWRITEVERF) {
1715 			    ncl_clearcommit(vp->v_mount);
1716 		    }
1717 	    }
1718 
1719 	    /*
1720 	     * Setup for actual write
1721 	     */
1722 	    mtx_lock(&np->n_mtx);
1723 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1724 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1725 	    mtx_unlock(&np->n_mtx);
1726 
1727 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1728 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1729 		    - bp->b_dirtyoff;
1730 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1731 		    + bp->b_dirtyoff;
1732 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1733 		uiop->uio_rw = UIO_WRITE;
1734 		NFSINCRGLOBAL(newnfsstats.write_bios);
1735 
1736 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1737 		    iomode = NFSWRITE_UNSTABLE;
1738 		else
1739 		    iomode = NFSWRITE_FILESYNC;
1740 
1741 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1742 		    called_from_strategy);
1743 
1744 		/*
1745 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1746 		 * to cluster the buffers needing commit.  This will allow
1747 		 * the system to submit a single commit rpc for the whole
1748 		 * cluster.  We can do this even if the buffer is not 100%
1749 		 * dirty (relative to the NFS blocksize), so we optimize the
1750 		 * append-to-file-case.
1751 		 *
1752 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1753 		 * cleared because write clustering only works for commit
1754 		 * rpc's, not for the data portion of the write).
1755 		 */
1756 
1757 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1758 		    bp->b_flags |= B_NEEDCOMMIT;
1759 		    if (bp->b_dirtyoff == 0
1760 			&& bp->b_dirtyend == bp->b_bcount)
1761 			bp->b_flags |= B_CLUSTEROK;
1762 		} else {
1763 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1764 		}
1765 
1766 		/*
1767 		 * For an interrupted write, the buffer is still valid
1768 		 * and the write hasn't been pushed to the server yet,
1769 		 * so we can't set BIO_ERROR and report the interruption
1770 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1771 		 * is not relevant, so the rpc attempt is essentially
1772 		 * a noop.  For the case of a V3 write rpc not being
1773 		 * committed to stable storage, the block is still
1774 		 * dirty and requires either a commit rpc or another
1775 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1776 		 * the block is reused. This is indicated by setting
1777 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1778 		 *
1779 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1780 		 * write error and is handled as above, except that
1781 		 * B_EINTR isn't set. One cause of this is a stale stateid
1782 		 * error for the RPC that indicates recovery is required,
1783 		 * when called with called_from_strategy != 0.
1784 		 *
1785 		 * If the buffer is marked B_PAGING, it does not reside on
1786 		 * the vp's paging queues so we cannot call bdirty().  The
1787 		 * bp in this case is not an NFS cache block so we should
1788 		 * be safe. XXX
1789 		 *
1790 		 * The logic below breaks up errors into recoverable and
1791 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1792 		 * and keep the buffer around for potential write retries.
1793 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1794 		 * and save the error in the nfsnode. This is less than ideal
1795 		 * but necessary. Keeping such buffers around could potentially
1796 		 * cause buffer exhaustion eventually (they can never be written
1797 		 * out, so will get constantly be re-dirtied). It also causes
1798 		 * all sorts of vfs panics. For non-recoverable write errors,
1799 		 * also invalidate the attrcache, so we'll be forced to go over
1800 		 * the wire for this object, returning an error to user on next
1801 		 * call (most of the time).
1802 		 */
1803 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1804 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1805 			int s;
1806 
1807 			s = splbio();
1808 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1809 			if ((bp->b_flags & B_PAGING) == 0) {
1810 			    bdirty(bp);
1811 			    bp->b_flags &= ~B_DONE;
1812 			}
1813 			if ((error == EINTR || error == ETIMEDOUT) &&
1814 			    (bp->b_flags & B_ASYNC) == 0)
1815 			    bp->b_flags |= B_EINTR;
1816 			splx(s);
1817 		} else {
1818 		    if (error) {
1819 			bp->b_ioflags |= BIO_ERROR;
1820 			bp->b_flags |= B_INVAL;
1821 			bp->b_error = np->n_error = error;
1822 			mtx_lock(&np->n_mtx);
1823 			np->n_flag |= NWRITEERR;
1824 			np->n_attrstamp = 0;
1825 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1826 			mtx_unlock(&np->n_mtx);
1827 		    }
1828 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1829 		}
1830 	    } else {
1831 		bp->b_resid = 0;
1832 		bufdone(bp);
1833 		return (0);
1834 	    }
1835 	}
1836 	bp->b_resid = uiop->uio_resid;
1837 	if (must_commit)
1838 	    ncl_clearcommit(vp->v_mount);
1839 	bufdone(bp);
1840 	return (error);
1841 }
1842 
1843 /*
1844  * Used to aid in handling ftruncate() operations on the NFS client side.
1845  * Truncation creates a number of special problems for NFS.  We have to
1846  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1847  * we have to properly handle VM pages or (potentially dirty) buffers
1848  * that straddle the truncation point.
1849  */
1850 
1851 int
1852 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1853 {
1854 	struct nfsnode *np = VTONFS(vp);
1855 	u_quad_t tsize;
1856 	int biosize = vp->v_bufobj.bo_bsize;
1857 	int error = 0;
1858 
1859 	mtx_lock(&np->n_mtx);
1860 	tsize = np->n_size;
1861 	np->n_size = nsize;
1862 	mtx_unlock(&np->n_mtx);
1863 
1864 	if (nsize < tsize) {
1865 		struct buf *bp;
1866 		daddr_t lbn;
1867 		int bufsize;
1868 
1869 		/*
1870 		 * vtruncbuf() doesn't get the buffer overlapping the
1871 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1872 		 * buffer that now needs to be truncated.
1873 		 */
1874 		error = vtruncbuf(vp, cred, nsize, biosize);
1875 		lbn = nsize / biosize;
1876 		bufsize = nsize - (lbn * biosize);
1877 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1878 		if (!bp)
1879 			return EINTR;
1880 		if (bp->b_dirtyoff > bp->b_bcount)
1881 			bp->b_dirtyoff = bp->b_bcount;
1882 		if (bp->b_dirtyend > bp->b_bcount)
1883 			bp->b_dirtyend = bp->b_bcount;
1884 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1885 		brelse(bp);
1886 	} else {
1887 		vnode_pager_setsize(vp, nsize);
1888 	}
1889 	return(error);
1890 }
1891 
1892