xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 8657387683946d0c03e09fe77029edfe309eeb20)
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  * 3. 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 nfsstatsv1 nfsstatsv1;
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 SYSCTL_DECL(_vfs_nfs);
82 static int use_buf_pager = 1;
83 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
84     &use_buf_pager, 0,
85     "Use buffer pager instead of direct readrpc call");
86 
87 static daddr_t
88 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
89 {
90 
91 	return (off / vp->v_bufobj.bo_bsize);
92 }
93 
94 static int
95 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn)
96 {
97 	struct nfsnode *np;
98 	u_quad_t nsize;
99 	int biosize, bcount;
100 
101 	np = VTONFS(vp);
102 	mtx_lock(&np->n_mtx);
103 	nsize = np->n_size;
104 	mtx_unlock(&np->n_mtx);
105 
106 	biosize = vp->v_bufobj.bo_bsize;
107 	bcount = biosize;
108 	if ((off_t)lbn * biosize >= nsize)
109 		bcount = 0;
110 	else if ((off_t)(lbn + 1) * biosize > nsize)
111 		bcount = nsize - (off_t)lbn * biosize;
112 	return (bcount);
113 }
114 
115 int
116 ncl_getpages(struct vop_getpages_args *ap)
117 {
118 	int i, error, nextoff, size, toff, count, npages;
119 	struct uio uio;
120 	struct iovec iov;
121 	vm_offset_t kva;
122 	struct buf *bp;
123 	struct vnode *vp;
124 	struct thread *td;
125 	struct ucred *cred;
126 	struct nfsmount *nmp;
127 	vm_object_t object;
128 	vm_page_t *pages;
129 	struct nfsnode *np;
130 
131 	vp = ap->a_vp;
132 	np = VTONFS(vp);
133 	td = curthread;
134 	cred = curthread->td_ucred;
135 	nmp = VFSTONFS(vp->v_mount);
136 	pages = ap->a_m;
137 	npages = ap->a_count;
138 
139 	if ((object = vp->v_object) == NULL) {
140 		printf("ncl_getpages: called with non-merged cache vnode\n");
141 		return (VM_PAGER_ERROR);
142 	}
143 
144 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
145 		mtx_lock(&np->n_mtx);
146 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
147 			mtx_unlock(&np->n_mtx);
148 			printf("ncl_getpages: called on non-cacheable vnode\n");
149 			return (VM_PAGER_ERROR);
150 		} else
151 			mtx_unlock(&np->n_mtx);
152 	}
153 
154 	mtx_lock(&nmp->nm_mtx);
155 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
156 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
157 		mtx_unlock(&nmp->nm_mtx);
158 		/* We'll never get here for v4, because we always have fsinfo */
159 		(void)ncl_fsinfo(nmp, vp, cred, td);
160 	} else
161 		mtx_unlock(&nmp->nm_mtx);
162 
163 	if (use_buf_pager)
164 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
165 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
166 
167 	/*
168 	 * If the requested page is partially valid, just return it and
169 	 * allow the pager to zero-out the blanks.  Partially valid pages
170 	 * can only occur at the file EOF.
171 	 *
172 	 * XXXGL: is that true for NFS, where short read can occur???
173 	 */
174 	VM_OBJECT_WLOCK(object);
175 	if (pages[npages - 1]->valid != 0 && --npages == 0)
176 		goto out;
177 	VM_OBJECT_WUNLOCK(object);
178 
179 	/*
180 	 * We use only the kva address for the buffer, but this is extremely
181 	 * convenient and fast.
182 	 */
183 	bp = getpbuf(&ncl_pbuf_freecnt);
184 
185 	kva = (vm_offset_t) bp->b_data;
186 	pmap_qenter(kva, pages, npages);
187 	VM_CNT_INC(v_vnodein);
188 	VM_CNT_ADD(v_vnodepgsin, npages);
189 
190 	count = npages << PAGE_SHIFT;
191 	iov.iov_base = (caddr_t) kva;
192 	iov.iov_len = count;
193 	uio.uio_iov = &iov;
194 	uio.uio_iovcnt = 1;
195 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
196 	uio.uio_resid = count;
197 	uio.uio_segflg = UIO_SYSSPACE;
198 	uio.uio_rw = UIO_READ;
199 	uio.uio_td = td;
200 
201 	error = ncl_readrpc(vp, &uio, cred);
202 	pmap_qremove(kva, npages);
203 
204 	relpbuf(bp, &ncl_pbuf_freecnt);
205 
206 	if (error && (uio.uio_resid == count)) {
207 		printf("ncl_getpages: error %d\n", error);
208 		return (VM_PAGER_ERROR);
209 	}
210 
211 	/*
212 	 * Calculate the number of bytes read and validate only that number
213 	 * of bytes.  Note that due to pending writes, size may be 0.  This
214 	 * does not mean that the remaining data is invalid!
215 	 */
216 
217 	size = count - uio.uio_resid;
218 	VM_OBJECT_WLOCK(object);
219 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
220 		vm_page_t m;
221 		nextoff = toff + PAGE_SIZE;
222 		m = pages[i];
223 
224 		if (nextoff <= size) {
225 			/*
226 			 * Read operation filled an entire page
227 			 */
228 			m->valid = VM_PAGE_BITS_ALL;
229 			KASSERT(m->dirty == 0,
230 			    ("nfs_getpages: page %p is dirty", m));
231 		} else if (size > toff) {
232 			/*
233 			 * Read operation filled a partial page.
234 			 */
235 			m->valid = 0;
236 			vm_page_set_valid_range(m, 0, size - toff);
237 			KASSERT(m->dirty == 0,
238 			    ("nfs_getpages: page %p is dirty", m));
239 		} else {
240 			/*
241 			 * Read operation was short.  If no error
242 			 * occurred we may have hit a zero-fill
243 			 * section.  We leave valid set to 0, and page
244 			 * is freed by vm_page_readahead_finish() if
245 			 * its index is not equal to requested, or
246 			 * page is zeroed and set valid by
247 			 * vm_pager_get_pages() for requested page.
248 			 */
249 			;
250 		}
251 	}
252 out:
253 	VM_OBJECT_WUNLOCK(object);
254 	if (ap->a_rbehind)
255 		*ap->a_rbehind = 0;
256 	if (ap->a_rahead)
257 		*ap->a_rahead = 0;
258 	return (VM_PAGER_OK);
259 }
260 
261 /*
262  * Vnode op for VM putpages.
263  */
264 int
265 ncl_putpages(struct vop_putpages_args *ap)
266 {
267 	struct uio uio;
268 	struct iovec iov;
269 	int i, error, npages, count;
270 	off_t offset;
271 	int *rtvals;
272 	struct vnode *vp;
273 	struct thread *td;
274 	struct ucred *cred;
275 	struct nfsmount *nmp;
276 	struct nfsnode *np;
277 	vm_page_t *pages;
278 
279 	vp = ap->a_vp;
280 	np = VTONFS(vp);
281 	td = curthread;				/* XXX */
282 	/* Set the cred to n_writecred for the write rpcs. */
283 	if (np->n_writecred != NULL)
284 		cred = crhold(np->n_writecred);
285 	else
286 		cred = crhold(curthread->td_ucred);	/* XXX */
287 	nmp = VFSTONFS(vp->v_mount);
288 	pages = ap->a_m;
289 	count = ap->a_count;
290 	rtvals = ap->a_rtvals;
291 	npages = btoc(count);
292 	offset = IDX_TO_OFF(pages[0]->pindex);
293 
294 	mtx_lock(&nmp->nm_mtx);
295 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
296 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
297 		mtx_unlock(&nmp->nm_mtx);
298 		(void)ncl_fsinfo(nmp, vp, cred, td);
299 	} else
300 		mtx_unlock(&nmp->nm_mtx);
301 
302 	mtx_lock(&np->n_mtx);
303 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
304 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
305 		mtx_unlock(&np->n_mtx);
306 		printf("ncl_putpages: called on noncache-able vnode\n");
307 		mtx_lock(&np->n_mtx);
308 	}
309 	/*
310 	 * When putting pages, do not extend file past EOF.
311 	 */
312 	if (offset + count > np->n_size) {
313 		count = np->n_size - offset;
314 		if (count < 0)
315 			count = 0;
316 	}
317 	mtx_unlock(&np->n_mtx);
318 
319 	for (i = 0; i < npages; i++)
320 		rtvals[i] = VM_PAGER_ERROR;
321 
322 	VM_CNT_INC(v_vnodeout);
323 	VM_CNT_ADD(v_vnodepgsout, count);
324 
325 	iov.iov_base = unmapped_buf;
326 	iov.iov_len = count;
327 	uio.uio_iov = &iov;
328 	uio.uio_iovcnt = 1;
329 	uio.uio_offset = offset;
330 	uio.uio_resid = count;
331 	uio.uio_segflg = UIO_NOCOPY;
332 	uio.uio_rw = UIO_WRITE;
333 	uio.uio_td = td;
334 
335 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
336 	    cred);
337 	crfree(cred);
338 
339 	if (error == 0 || !nfs_keep_dirty_on_error) {
340 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
341 		    np->n_size - offset, npages * PAGE_SIZE);
342 	}
343 	return (rtvals[0]);
344 }
345 
346 /*
347  * For nfs, cache consistency can only be maintained approximately.
348  * Although RFC1094 does not specify the criteria, the following is
349  * believed to be compatible with the reference port.
350  * For nfs:
351  * If the file's modify time on the server has changed since the
352  * last read rpc or you have written to the file,
353  * you may have lost data cache consistency with the
354  * server, so flush all of the file's data out of the cache.
355  * Then force a getattr rpc to ensure that you have up to date
356  * attributes.
357  * NB: This implies that cache data can be read when up to
358  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
359  * attributes this could be forced by setting n_attrstamp to 0 before
360  * the VOP_GETATTR() call.
361  */
362 static inline int
363 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
364 {
365 	int error = 0;
366 	struct vattr vattr;
367 	struct nfsnode *np = VTONFS(vp);
368 	bool old_lock;
369 
370 	/*
371 	 * Ensure the exclusove access to the node before checking
372 	 * whether the cache is consistent.
373 	 */
374 	old_lock = ncl_excl_start(vp);
375 	mtx_lock(&np->n_mtx);
376 	if (np->n_flag & NMODIFIED) {
377 		mtx_unlock(&np->n_mtx);
378 		if (vp->v_type != VREG) {
379 			if (vp->v_type != VDIR)
380 				panic("nfs: bioread, not dir");
381 			ncl_invaldir(vp);
382 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
383 			if (error != 0)
384 				goto out;
385 		}
386 		np->n_attrstamp = 0;
387 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
388 		error = VOP_GETATTR(vp, &vattr, cred);
389 		if (error)
390 			goto out;
391 		mtx_lock(&np->n_mtx);
392 		np->n_mtime = vattr.va_mtime;
393 		mtx_unlock(&np->n_mtx);
394 	} else {
395 		mtx_unlock(&np->n_mtx);
396 		error = VOP_GETATTR(vp, &vattr, cred);
397 		if (error)
398 			goto out;
399 		mtx_lock(&np->n_mtx);
400 		if ((np->n_flag & NSIZECHANGED)
401 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
402 			mtx_unlock(&np->n_mtx);
403 			if (vp->v_type == VDIR)
404 				ncl_invaldir(vp);
405 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
406 			if (error != 0)
407 				goto out;
408 			mtx_lock(&np->n_mtx);
409 			np->n_mtime = vattr.va_mtime;
410 			np->n_flag &= ~NSIZECHANGED;
411 		}
412 		mtx_unlock(&np->n_mtx);
413 	}
414 out:
415 	ncl_excl_finish(vp, old_lock);
416 	return (error);
417 }
418 
419 /*
420  * Vnode op for read using bio
421  */
422 int
423 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
424 {
425 	struct nfsnode *np = VTONFS(vp);
426 	int biosize, i;
427 	struct buf *bp, *rabp;
428 	struct thread *td;
429 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
430 	daddr_t lbn, rabn;
431 	int bcount;
432 	int seqcount;
433 	int nra, error = 0, n = 0, on = 0;
434 	off_t tmp_off;
435 
436 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
437 	if (uio->uio_resid == 0)
438 		return (0);
439 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
440 		return (EINVAL);
441 	td = uio->uio_td;
442 
443 	mtx_lock(&nmp->nm_mtx);
444 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
445 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
446 		mtx_unlock(&nmp->nm_mtx);
447 		(void)ncl_fsinfo(nmp, vp, cred, td);
448 		mtx_lock(&nmp->nm_mtx);
449 	}
450 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
451 		(void) newnfs_iosize(nmp);
452 
453 	tmp_off = uio->uio_offset + uio->uio_resid;
454 	if (vp->v_type != VDIR &&
455 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
456 		mtx_unlock(&nmp->nm_mtx);
457 		return (EFBIG);
458 	}
459 	mtx_unlock(&nmp->nm_mtx);
460 
461 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
462 		/* No caching/ no readaheads. Just read data into the user buffer */
463 		return ncl_readrpc(vp, uio, cred);
464 
465 	biosize = vp->v_bufobj.bo_bsize;
466 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
467 
468 	error = nfs_bioread_check_cons(vp, td, cred);
469 	if (error)
470 		return error;
471 
472 	do {
473 	    u_quad_t nsize;
474 
475 	    mtx_lock(&np->n_mtx);
476 	    nsize = np->n_size;
477 	    mtx_unlock(&np->n_mtx);
478 
479 	    switch (vp->v_type) {
480 	    case VREG:
481 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
482 		lbn = uio->uio_offset / biosize;
483 		on = uio->uio_offset - (lbn * biosize);
484 
485 		/*
486 		 * Start the read ahead(s), as required.
487 		 */
488 		if (nmp->nm_readahead > 0) {
489 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
490 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
491 			rabn = lbn + 1 + nra;
492 			if (incore(&vp->v_bufobj, rabn) == NULL) {
493 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
494 			    if (!rabp) {
495 				error = newnfs_sigintr(nmp, td);
496 				return (error ? error : EINTR);
497 			    }
498 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
499 				rabp->b_flags |= B_ASYNC;
500 				rabp->b_iocmd = BIO_READ;
501 				vfs_busy_pages(rabp, 0);
502 				if (ncl_asyncio(nmp, rabp, cred, td)) {
503 				    rabp->b_flags |= B_INVAL;
504 				    rabp->b_ioflags |= BIO_ERROR;
505 				    vfs_unbusy_pages(rabp);
506 				    brelse(rabp);
507 				    break;
508 				}
509 			    } else {
510 				brelse(rabp);
511 			    }
512 			}
513 		    }
514 		}
515 
516 		/* Note that bcount is *not* DEV_BSIZE aligned. */
517 		bcount = biosize;
518 		if ((off_t)lbn * biosize >= nsize) {
519 			bcount = 0;
520 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
521 			bcount = nsize - (off_t)lbn * biosize;
522 		}
523 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
524 
525 		if (!bp) {
526 			error = newnfs_sigintr(nmp, td);
527 			return (error ? error : EINTR);
528 		}
529 
530 		/*
531 		 * If B_CACHE is not set, we must issue the read.  If this
532 		 * fails, we return an error.
533 		 */
534 
535 		if ((bp->b_flags & B_CACHE) == 0) {
536 		    bp->b_iocmd = BIO_READ;
537 		    vfs_busy_pages(bp, 0);
538 		    error = ncl_doio(vp, bp, cred, td, 0);
539 		    if (error) {
540 			brelse(bp);
541 			return (error);
542 		    }
543 		}
544 
545 		/*
546 		 * on is the offset into the current bp.  Figure out how many
547 		 * bytes we can copy out of the bp.  Note that bcount is
548 		 * NOT DEV_BSIZE aligned.
549 		 *
550 		 * Then figure out how many bytes we can copy into the uio.
551 		 */
552 
553 		n = 0;
554 		if (on < bcount)
555 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
556 		break;
557 	    case VLNK:
558 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
559 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
560 		if (!bp) {
561 			error = newnfs_sigintr(nmp, td);
562 			return (error ? error : EINTR);
563 		}
564 		if ((bp->b_flags & B_CACHE) == 0) {
565 		    bp->b_iocmd = BIO_READ;
566 		    vfs_busy_pages(bp, 0);
567 		    error = ncl_doio(vp, bp, cred, td, 0);
568 		    if (error) {
569 			bp->b_ioflags |= BIO_ERROR;
570 			brelse(bp);
571 			return (error);
572 		    }
573 		}
574 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
575 		on = 0;
576 		break;
577 	    case VDIR:
578 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
579 		if (np->n_direofoffset
580 		    && uio->uio_offset >= np->n_direofoffset) {
581 		    return (0);
582 		}
583 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
584 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
585 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
586 		if (!bp) {
587 		    error = newnfs_sigintr(nmp, td);
588 		    return (error ? error : EINTR);
589 		}
590 		if ((bp->b_flags & B_CACHE) == 0) {
591 		    bp->b_iocmd = BIO_READ;
592 		    vfs_busy_pages(bp, 0);
593 		    error = ncl_doio(vp, bp, cred, td, 0);
594 		    if (error) {
595 			    brelse(bp);
596 		    }
597 		    while (error == NFSERR_BAD_COOKIE) {
598 			ncl_invaldir(vp);
599 			error = ncl_vinvalbuf(vp, 0, td, 1);
600 
601 			/*
602 			 * Yuck! The directory has been modified on the
603 			 * server. The only way to get the block is by
604 			 * reading from the beginning to get all the
605 			 * offset cookies.
606 			 *
607 			 * Leave the last bp intact unless there is an error.
608 			 * Loop back up to the while if the error is another
609 			 * NFSERR_BAD_COOKIE (double yuch!).
610 			 */
611 			for (i = 0; i <= lbn && !error; i++) {
612 			    if (np->n_direofoffset
613 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
614 				    return (0);
615 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
616 			    if (!bp) {
617 				error = newnfs_sigintr(nmp, td);
618 				return (error ? error : EINTR);
619 			    }
620 			    if ((bp->b_flags & B_CACHE) == 0) {
621 				    bp->b_iocmd = BIO_READ;
622 				    vfs_busy_pages(bp, 0);
623 				    error = ncl_doio(vp, bp, cred, td, 0);
624 				    /*
625 				     * no error + B_INVAL == directory EOF,
626 				     * use the block.
627 				     */
628 				    if (error == 0 && (bp->b_flags & B_INVAL))
629 					    break;
630 			    }
631 			    /*
632 			     * An error will throw away the block and the
633 			     * for loop will break out.  If no error and this
634 			     * is not the block we want, we throw away the
635 			     * block and go for the next one via the for loop.
636 			     */
637 			    if (error || i < lbn)
638 				    brelse(bp);
639 			}
640 		    }
641 		    /*
642 		     * The above while is repeated if we hit another cookie
643 		     * error.  If we hit an error and it wasn't a cookie error,
644 		     * we give up.
645 		     */
646 		    if (error)
647 			    return (error);
648 		}
649 
650 		/*
651 		 * If not eof and read aheads are enabled, start one.
652 		 * (You need the current block first, so that you have the
653 		 *  directory offset cookie of the next block.)
654 		 */
655 		if (nmp->nm_readahead > 0 &&
656 		    (bp->b_flags & B_INVAL) == 0 &&
657 		    (np->n_direofoffset == 0 ||
658 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
659 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
660 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
661 			if (rabp) {
662 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
663 				rabp->b_flags |= B_ASYNC;
664 				rabp->b_iocmd = BIO_READ;
665 				vfs_busy_pages(rabp, 0);
666 				if (ncl_asyncio(nmp, rabp, cred, td)) {
667 				    rabp->b_flags |= B_INVAL;
668 				    rabp->b_ioflags |= BIO_ERROR;
669 				    vfs_unbusy_pages(rabp);
670 				    brelse(rabp);
671 				}
672 			    } else {
673 				brelse(rabp);
674 			    }
675 			}
676 		}
677 		/*
678 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
679 		 * chopped for the EOF condition, we cannot tell how large
680 		 * NFS directories are going to be until we hit EOF.  So
681 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
682 		 * it just so happens that b_resid will effectively chop it
683 		 * to EOF.  *BUT* this information is lost if the buffer goes
684 		 * away and is reconstituted into a B_CACHE state ( due to
685 		 * being VMIO ) later.  So we keep track of the directory eof
686 		 * in np->n_direofoffset and chop it off as an extra step
687 		 * right here.
688 		 */
689 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
690 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
691 			n = np->n_direofoffset - uio->uio_offset;
692 		break;
693 	    default:
694 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
695 		bp = NULL;
696 		break;
697 	    }
698 
699 	    if (n > 0) {
700 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
701 	    }
702 	    if (vp->v_type == VLNK)
703 		n = 0;
704 	    if (bp != NULL)
705 		brelse(bp);
706 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
707 	return (error);
708 }
709 
710 /*
711  * The NFS write path cannot handle iovecs with len > 1. So we need to
712  * break up iovecs accordingly (restricting them to wsize).
713  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
714  * For the ASYNC case, 2 copies are needed. The first a copy from the
715  * user buffer to a staging buffer and then a second copy from the staging
716  * buffer to mbufs. This can be optimized by copying from the user buffer
717  * directly into mbufs and passing the chain down, but that requires a
718  * fair amount of re-working of the relevant codepaths (and can be done
719  * later).
720  */
721 static int
722 nfs_directio_write(vp, uiop, cred, ioflag)
723 	struct vnode *vp;
724 	struct uio *uiop;
725 	struct ucred *cred;
726 	int ioflag;
727 {
728 	int error;
729 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
730 	struct thread *td = uiop->uio_td;
731 	int size;
732 	int wsize;
733 
734 	mtx_lock(&nmp->nm_mtx);
735 	wsize = nmp->nm_wsize;
736 	mtx_unlock(&nmp->nm_mtx);
737 	if (ioflag & IO_SYNC) {
738 		int iomode, must_commit;
739 		struct uio uio;
740 		struct iovec iov;
741 do_sync:
742 		while (uiop->uio_resid > 0) {
743 			size = MIN(uiop->uio_resid, wsize);
744 			size = MIN(uiop->uio_iov->iov_len, size);
745 			iov.iov_base = uiop->uio_iov->iov_base;
746 			iov.iov_len = size;
747 			uio.uio_iov = &iov;
748 			uio.uio_iovcnt = 1;
749 			uio.uio_offset = uiop->uio_offset;
750 			uio.uio_resid = size;
751 			uio.uio_segflg = UIO_USERSPACE;
752 			uio.uio_rw = UIO_WRITE;
753 			uio.uio_td = td;
754 			iomode = NFSWRITE_FILESYNC;
755 			error = ncl_writerpc(vp, &uio, cred, &iomode,
756 			    &must_commit, 0);
757 			KASSERT((must_commit == 0),
758 				("ncl_directio_write: Did not commit write"));
759 			if (error)
760 				return (error);
761 			uiop->uio_offset += size;
762 			uiop->uio_resid -= size;
763 			if (uiop->uio_iov->iov_len <= size) {
764 				uiop->uio_iovcnt--;
765 				uiop->uio_iov++;
766 			} else {
767 				uiop->uio_iov->iov_base =
768 					(char *)uiop->uio_iov->iov_base + size;
769 				uiop->uio_iov->iov_len -= size;
770 			}
771 		}
772 	} else {
773 		struct uio *t_uio;
774 		struct iovec *t_iov;
775 		struct buf *bp;
776 
777 		/*
778 		 * Break up the write into blocksize chunks and hand these
779 		 * over to nfsiod's for write back.
780 		 * Unfortunately, this incurs a copy of the data. Since
781 		 * the user could modify the buffer before the write is
782 		 * initiated.
783 		 *
784 		 * The obvious optimization here is that one of the 2 copies
785 		 * in the async write path can be eliminated by copying the
786 		 * data here directly into mbufs and passing the mbuf chain
787 		 * down. But that will require a fair amount of re-working
788 		 * of the code and can be done if there's enough interest
789 		 * in NFS directio access.
790 		 */
791 		while (uiop->uio_resid > 0) {
792 			size = MIN(uiop->uio_resid, wsize);
793 			size = MIN(uiop->uio_iov->iov_len, size);
794 			bp = getpbuf(&ncl_pbuf_freecnt);
795 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
796 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
797 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
798 			t_iov->iov_len = size;
799 			t_uio->uio_iov = t_iov;
800 			t_uio->uio_iovcnt = 1;
801 			t_uio->uio_offset = uiop->uio_offset;
802 			t_uio->uio_resid = size;
803 			t_uio->uio_segflg = UIO_SYSSPACE;
804 			t_uio->uio_rw = UIO_WRITE;
805 			t_uio->uio_td = td;
806 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
807 			    uiop->uio_segflg == UIO_SYSSPACE,
808 			    ("nfs_directio_write: Bad uio_segflg"));
809 			if (uiop->uio_segflg == UIO_USERSPACE) {
810 				error = copyin(uiop->uio_iov->iov_base,
811 				    t_iov->iov_base, size);
812 				if (error != 0)
813 					goto err_free;
814 			} else
815 				/*
816 				 * UIO_SYSSPACE may never happen, but handle
817 				 * it just in case it does.
818 				 */
819 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
820 				    size);
821 			bp->b_flags |= B_DIRECT;
822 			bp->b_iocmd = BIO_WRITE;
823 			if (cred != NOCRED) {
824 				crhold(cred);
825 				bp->b_wcred = cred;
826 			} else
827 				bp->b_wcred = NOCRED;
828 			bp->b_caller1 = (void *)t_uio;
829 			bp->b_vp = vp;
830 			error = ncl_asyncio(nmp, bp, NOCRED, td);
831 err_free:
832 			if (error) {
833 				free(t_iov->iov_base, M_NFSDIRECTIO);
834 				free(t_iov, M_NFSDIRECTIO);
835 				free(t_uio, M_NFSDIRECTIO);
836 				bp->b_vp = NULL;
837 				relpbuf(bp, &ncl_pbuf_freecnt);
838 				if (error == EINTR)
839 					return (error);
840 				goto do_sync;
841 			}
842 			uiop->uio_offset += size;
843 			uiop->uio_resid -= size;
844 			if (uiop->uio_iov->iov_len <= size) {
845 				uiop->uio_iovcnt--;
846 				uiop->uio_iov++;
847 			} else {
848 				uiop->uio_iov->iov_base =
849 					(char *)uiop->uio_iov->iov_base + size;
850 				uiop->uio_iov->iov_len -= size;
851 			}
852 		}
853 	}
854 	return (0);
855 }
856 
857 /*
858  * Vnode op for write using bio
859  */
860 int
861 ncl_write(struct vop_write_args *ap)
862 {
863 	int biosize;
864 	struct uio *uio = ap->a_uio;
865 	struct thread *td = uio->uio_td;
866 	struct vnode *vp = ap->a_vp;
867 	struct nfsnode *np = VTONFS(vp);
868 	struct ucred *cred = ap->a_cred;
869 	int ioflag = ap->a_ioflag;
870 	struct buf *bp;
871 	struct vattr vattr;
872 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
873 	daddr_t lbn;
874 	int bcount, noncontig_write, obcount;
875 	int bp_cached, n, on, error = 0, error1, wouldcommit;
876 	size_t orig_resid, local_resid;
877 	off_t orig_size, tmp_off;
878 
879 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
880 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
881 	    ("ncl_write proc"));
882 	if (vp->v_type != VREG)
883 		return (EIO);
884 	mtx_lock(&np->n_mtx);
885 	if (np->n_flag & NWRITEERR) {
886 		np->n_flag &= ~NWRITEERR;
887 		mtx_unlock(&np->n_mtx);
888 		return (np->n_error);
889 	} else
890 		mtx_unlock(&np->n_mtx);
891 	mtx_lock(&nmp->nm_mtx);
892 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
893 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
894 		mtx_unlock(&nmp->nm_mtx);
895 		(void)ncl_fsinfo(nmp, vp, cred, td);
896 		mtx_lock(&nmp->nm_mtx);
897 	}
898 	if (nmp->nm_wsize == 0)
899 		(void) newnfs_iosize(nmp);
900 	mtx_unlock(&nmp->nm_mtx);
901 
902 	/*
903 	 * Synchronously flush pending buffers if we are in synchronous
904 	 * mode or if we are appending.
905 	 */
906 	if (ioflag & (IO_APPEND | IO_SYNC)) {
907 		mtx_lock(&np->n_mtx);
908 		if (np->n_flag & NMODIFIED) {
909 			mtx_unlock(&np->n_mtx);
910 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
911 			/*
912 			 * Require non-blocking, synchronous writes to
913 			 * dirty files to inform the program it needs
914 			 * to fsync(2) explicitly.
915 			 */
916 			if (ioflag & IO_NDELAY)
917 				return (EAGAIN);
918 #endif
919 			np->n_attrstamp = 0;
920 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
921 			error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
922 			    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
923 			if (error != 0)
924 				return (error);
925 		} else
926 			mtx_unlock(&np->n_mtx);
927 	}
928 
929 	orig_resid = uio->uio_resid;
930 	mtx_lock(&np->n_mtx);
931 	orig_size = np->n_size;
932 	mtx_unlock(&np->n_mtx);
933 
934 	/*
935 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
936 	 * get the append lock.
937 	 */
938 	if (ioflag & IO_APPEND) {
939 		np->n_attrstamp = 0;
940 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
941 		error = VOP_GETATTR(vp, &vattr, cred);
942 		if (error)
943 			return (error);
944 		mtx_lock(&np->n_mtx);
945 		uio->uio_offset = np->n_size;
946 		mtx_unlock(&np->n_mtx);
947 	}
948 
949 	if (uio->uio_offset < 0)
950 		return (EINVAL);
951 	tmp_off = uio->uio_offset + uio->uio_resid;
952 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
953 		return (EFBIG);
954 	if (uio->uio_resid == 0)
955 		return (0);
956 
957 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
958 		return nfs_directio_write(vp, uio, cred, ioflag);
959 
960 	/*
961 	 * Maybe this should be above the vnode op call, but so long as
962 	 * file servers have no limits, i don't think it matters
963 	 */
964 	if (vn_rlimit_fsize(vp, uio, td))
965 		return (EFBIG);
966 
967 	biosize = vp->v_bufobj.bo_bsize;
968 	/*
969 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
970 	 * would exceed the local maximum per-file write commit size when
971 	 * combined with those, we must decide whether to flush,
972 	 * go synchronous, or return error.  We don't bother checking
973 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
974 	 * no point optimizing for something that really won't ever happen.
975 	 */
976 	wouldcommit = 0;
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 		if (nflag & NMODIFIED) {
984 			BO_LOCK(&vp->v_bufobj);
985 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
986 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
987 				    b_bobufs) {
988 					if (bp->b_flags & B_NEEDCOMMIT)
989 						wouldcommit += bp->b_bcount;
990 				}
991 			}
992 			BO_UNLOCK(&vp->v_bufobj);
993 		}
994 	}
995 
996 	do {
997 		if (!(ioflag & IO_SYNC)) {
998 			wouldcommit += biosize;
999 			if (wouldcommit > nmp->nm_wcommitsize) {
1000 				np->n_attrstamp = 0;
1001 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1002 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1003 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1004 				if (error != 0)
1005 					return (error);
1006 				wouldcommit = biosize;
1007 			}
1008 		}
1009 
1010 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1011 		lbn = uio->uio_offset / biosize;
1012 		on = uio->uio_offset - (lbn * biosize);
1013 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1014 again:
1015 		/*
1016 		 * Handle direct append and file extension cases, calculate
1017 		 * unaligned buffer size.
1018 		 */
1019 		mtx_lock(&np->n_mtx);
1020 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1021 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1022 			noncontig_write = 1;
1023 		else
1024 			noncontig_write = 0;
1025 		if ((uio->uio_offset == np->n_size ||
1026 		    (noncontig_write != 0 &&
1027 		    lbn == (np->n_size / biosize) &&
1028 		    uio->uio_offset + n > np->n_size)) && n) {
1029 			mtx_unlock(&np->n_mtx);
1030 			/*
1031 			 * Get the buffer (in its pre-append state to maintain
1032 			 * B_CACHE if it was previously set).  Resize the
1033 			 * nfsnode after we have locked the buffer to prevent
1034 			 * readers from reading garbage.
1035 			 */
1036 			obcount = np->n_size - (lbn * biosize);
1037 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1038 
1039 			if (bp != NULL) {
1040 				long save;
1041 
1042 				mtx_lock(&np->n_mtx);
1043 				np->n_size = uio->uio_offset + n;
1044 				np->n_flag |= NMODIFIED;
1045 				vnode_pager_setsize(vp, np->n_size);
1046 				mtx_unlock(&np->n_mtx);
1047 
1048 				save = bp->b_flags & B_CACHE;
1049 				bcount = on + n;
1050 				allocbuf(bp, bcount);
1051 				bp->b_flags |= save;
1052 				if (noncontig_write != 0 && on > obcount)
1053 					vfs_bio_bzero_buf(bp, obcount, on -
1054 					    obcount);
1055 			}
1056 		} else {
1057 			/*
1058 			 * Obtain the locked cache block first, and then
1059 			 * adjust the file's size as appropriate.
1060 			 */
1061 			bcount = on + n;
1062 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1063 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1064 					bcount = biosize;
1065 				else
1066 					bcount = np->n_size - (off_t)lbn * biosize;
1067 			}
1068 			mtx_unlock(&np->n_mtx);
1069 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1070 			mtx_lock(&np->n_mtx);
1071 			if (uio->uio_offset + n > np->n_size) {
1072 				np->n_size = uio->uio_offset + n;
1073 				np->n_flag |= NMODIFIED;
1074 				vnode_pager_setsize(vp, np->n_size);
1075 			}
1076 			mtx_unlock(&np->n_mtx);
1077 		}
1078 
1079 		if (!bp) {
1080 			error = newnfs_sigintr(nmp, td);
1081 			if (!error)
1082 				error = EINTR;
1083 			break;
1084 		}
1085 
1086 		/*
1087 		 * Issue a READ if B_CACHE is not set.  In special-append
1088 		 * mode, B_CACHE is based on the buffer prior to the write
1089 		 * op and is typically set, avoiding the read.  If a read
1090 		 * is required in special append mode, the server will
1091 		 * probably send us a short-read since we extended the file
1092 		 * on our end, resulting in b_resid == 0 and, thusly,
1093 		 * B_CACHE getting set.
1094 		 *
1095 		 * We can also avoid issuing the read if the write covers
1096 		 * the entire buffer.  We have to make sure the buffer state
1097 		 * is reasonable in this case since we will not be initiating
1098 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1099 		 * more information.
1100 		 *
1101 		 * B_CACHE may also be set due to the buffer being cached
1102 		 * normally.
1103 		 */
1104 
1105 		bp_cached = 1;
1106 		if (on == 0 && n == bcount) {
1107 			if ((bp->b_flags & B_CACHE) == 0)
1108 				bp_cached = 0;
1109 			bp->b_flags |= B_CACHE;
1110 			bp->b_flags &= ~B_INVAL;
1111 			bp->b_ioflags &= ~BIO_ERROR;
1112 		}
1113 
1114 		if ((bp->b_flags & B_CACHE) == 0) {
1115 			bp->b_iocmd = BIO_READ;
1116 			vfs_busy_pages(bp, 0);
1117 			error = ncl_doio(vp, bp, cred, td, 0);
1118 			if (error) {
1119 				brelse(bp);
1120 				break;
1121 			}
1122 		}
1123 		if (bp->b_wcred == NOCRED)
1124 			bp->b_wcred = crhold(cred);
1125 		mtx_lock(&np->n_mtx);
1126 		np->n_flag |= NMODIFIED;
1127 		mtx_unlock(&np->n_mtx);
1128 
1129 		/*
1130 		 * If dirtyend exceeds file size, chop it down.  This should
1131 		 * not normally occur but there is an append race where it
1132 		 * might occur XXX, so we log it.
1133 		 *
1134 		 * If the chopping creates a reverse-indexed or degenerate
1135 		 * situation with dirtyoff/end, we 0 both of them.
1136 		 */
1137 
1138 		if (bp->b_dirtyend > bcount) {
1139 			printf("NFS append race @%lx:%d\n",
1140 			    (long)bp->b_blkno * DEV_BSIZE,
1141 			    bp->b_dirtyend - bcount);
1142 			bp->b_dirtyend = bcount;
1143 		}
1144 
1145 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1146 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1147 
1148 		/*
1149 		 * If the new write will leave a contiguous dirty
1150 		 * area, just update the b_dirtyoff and b_dirtyend,
1151 		 * otherwise force a write rpc of the old dirty area.
1152 		 *
1153 		 * If there has been a file lock applied to this file
1154 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1155 		 * While it is possible to merge discontiguous writes due to
1156 		 * our having a B_CACHE buffer ( and thus valid read data
1157 		 * for the hole), we don't because it could lead to
1158 		 * significant cache coherency problems with multiple clients,
1159 		 * especially if locking is implemented later on.
1160 		 *
1161 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1162 		 * not been file locking done on this file:
1163 		 * Relax coherency a bit for the sake of performance and
1164 		 * expand the current dirty region to contain the new
1165 		 * write even if it means we mark some non-dirty data as
1166 		 * dirty.
1167 		 */
1168 
1169 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1170 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1171 			if (bwrite(bp) == EINTR) {
1172 				error = EINTR;
1173 				break;
1174 			}
1175 			goto again;
1176 		}
1177 
1178 		local_resid = uio->uio_resid;
1179 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1180 
1181 		if (error != 0 && !bp_cached) {
1182 			/*
1183 			 * This block has no other content then what
1184 			 * possibly was written by the faulty uiomove.
1185 			 * Release it, forgetting the data pages, to
1186 			 * prevent the leak of uninitialized data to
1187 			 * usermode.
1188 			 */
1189 			bp->b_ioflags |= BIO_ERROR;
1190 			brelse(bp);
1191 			uio->uio_offset -= local_resid - uio->uio_resid;
1192 			uio->uio_resid = local_resid;
1193 			break;
1194 		}
1195 
1196 		/*
1197 		 * Since this block is being modified, it must be written
1198 		 * again and not just committed.  Since write clustering does
1199 		 * not work for the stage 1 data write, only the stage 2
1200 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1201 		 */
1202 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1203 
1204 		/*
1205 		 * Get the partial update on the progress made from
1206 		 * uiomove, if an error occurred.
1207 		 */
1208 		if (error != 0)
1209 			n = local_resid - uio->uio_resid;
1210 
1211 		/*
1212 		 * Only update dirtyoff/dirtyend if not a degenerate
1213 		 * condition.
1214 		 */
1215 		if (n > 0) {
1216 			if (bp->b_dirtyend > 0) {
1217 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1218 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1219 			} else {
1220 				bp->b_dirtyoff = on;
1221 				bp->b_dirtyend = on + n;
1222 			}
1223 			vfs_bio_set_valid(bp, on, n);
1224 		}
1225 
1226 		/*
1227 		 * If IO_SYNC do bwrite().
1228 		 *
1229 		 * IO_INVAL appears to be unused.  The idea appears to be
1230 		 * to turn off caching in this case.  Very odd.  XXX
1231 		 */
1232 		if ((ioflag & IO_SYNC)) {
1233 			if (ioflag & IO_INVAL)
1234 				bp->b_flags |= B_NOCACHE;
1235 			error1 = bwrite(bp);
1236 			if (error1 != 0) {
1237 				if (error == 0)
1238 					error = error1;
1239 				break;
1240 			}
1241 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1242 			bp->b_flags |= B_ASYNC;
1243 			(void) ncl_writebp(bp, 0, NULL);
1244 		} else {
1245 			bdwrite(bp);
1246 		}
1247 
1248 		if (error != 0)
1249 			break;
1250 	} while (uio->uio_resid > 0 && n > 0);
1251 
1252 	if (error != 0) {
1253 		if (ioflag & IO_UNIT) {
1254 			VATTR_NULL(&vattr);
1255 			vattr.va_size = orig_size;
1256 			/* IO_SYNC is handled implicitely */
1257 			(void)VOP_SETATTR(vp, &vattr, cred);
1258 			uio->uio_offset -= orig_resid - uio->uio_resid;
1259 			uio->uio_resid = orig_resid;
1260 		}
1261 	}
1262 
1263 	return (error);
1264 }
1265 
1266 /*
1267  * Get an nfs cache block.
1268  *
1269  * Allocate a new one if the block isn't currently in the cache
1270  * and return the block marked busy. If the calling process is
1271  * interrupted by a signal for an interruptible mount point, return
1272  * NULL.
1273  *
1274  * The caller must carefully deal with the possible B_INVAL state of
1275  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1276  * indirectly), so synchronous reads can be issued without worrying about
1277  * the B_INVAL state.  We have to be a little more careful when dealing
1278  * with writes (see comments in nfs_write()) when extending a file past
1279  * its EOF.
1280  */
1281 static struct buf *
1282 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1283 {
1284 	struct buf *bp;
1285 	struct mount *mp;
1286 	struct nfsmount *nmp;
1287 
1288 	mp = vp->v_mount;
1289 	nmp = VFSTONFS(mp);
1290 
1291 	if (nmp->nm_flag & NFSMNT_INT) {
1292 		sigset_t oldset;
1293 
1294 		newnfs_set_sigmask(td, &oldset);
1295 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1296 		newnfs_restore_sigmask(td, &oldset);
1297 		while (bp == NULL) {
1298 			if (newnfs_sigintr(nmp, td))
1299 				return (NULL);
1300 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1301 		}
1302 	} else {
1303 		bp = getblk(vp, bn, size, 0, 0, 0);
1304 	}
1305 
1306 	if (vp->v_type == VREG)
1307 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1308 	return (bp);
1309 }
1310 
1311 /*
1312  * Flush and invalidate all dirty buffers. If another process is already
1313  * doing the flush, just wait for completion.
1314  */
1315 int
1316 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1317 {
1318 	struct nfsnode *np = VTONFS(vp);
1319 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1320 	int error = 0, slpflag, slptimeo;
1321 	bool old_lock;
1322 
1323 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1324 
1325 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1326 		intrflg = 0;
1327 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1328 		intrflg = 1;
1329 	if (intrflg) {
1330 		slpflag = PCATCH;
1331 		slptimeo = 2 * hz;
1332 	} else {
1333 		slpflag = 0;
1334 		slptimeo = 0;
1335 	}
1336 
1337 	old_lock = ncl_excl_start(vp);
1338 	if (old_lock)
1339 		flags |= V_ALLOWCLEAN;
1340 
1341 	/*
1342 	 * Now, flush as required.
1343 	 */
1344 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1345 	     vp->v_bufobj.bo_object != NULL) {
1346 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1347 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1348 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1349 		/*
1350 		 * If the page clean was interrupted, fail the invalidation.
1351 		 * Not doing so, we run the risk of losing dirty pages in the
1352 		 * vinvalbuf() call below.
1353 		 */
1354 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1355 			goto out;
1356 	}
1357 
1358 	error = vinvalbuf(vp, flags, slpflag, 0);
1359 	while (error) {
1360 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1361 			goto out;
1362 		error = vinvalbuf(vp, flags, 0, slptimeo);
1363 	}
1364 	if (NFSHASPNFS(nmp)) {
1365 		nfscl_layoutcommit(vp, td);
1366 		/*
1367 		 * Invalidate the attribute cache, since writes to a DS
1368 		 * won't update the size attribute.
1369 		 */
1370 		mtx_lock(&np->n_mtx);
1371 		np->n_attrstamp = 0;
1372 	} else
1373 		mtx_lock(&np->n_mtx);
1374 	if (np->n_directio_asyncwr == 0)
1375 		np->n_flag &= ~NMODIFIED;
1376 	mtx_unlock(&np->n_mtx);
1377 out:
1378 	ncl_excl_finish(vp, old_lock);
1379 	return error;
1380 }
1381 
1382 /*
1383  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1384  * This is mainly to avoid queueing async I/O requests when the nfsiods
1385  * are all hung on a dead server.
1386  *
1387  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1388  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1389  */
1390 int
1391 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1392 {
1393 	int iod;
1394 	int gotiod;
1395 	int slpflag = 0;
1396 	int slptimeo = 0;
1397 	int error, error2;
1398 
1399 	/*
1400 	 * Commits are usually short and sweet so lets save some cpu and
1401 	 * leave the async daemons for more important rpc's (such as reads
1402 	 * and writes).
1403 	 *
1404 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1405 	 * in the directory in order to update attributes. This can deadlock
1406 	 * with another thread that is waiting for async I/O to be done by
1407 	 * an nfsiod thread while holding a lock on one of these vnodes.
1408 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1409 	 * perform Readdirplus RPCs.
1410 	 */
1411 	mtx_lock(&ncl_iod_mutex);
1412 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1413 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1414 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1415 		mtx_unlock(&ncl_iod_mutex);
1416 		return(EIO);
1417 	}
1418 again:
1419 	if (nmp->nm_flag & NFSMNT_INT)
1420 		slpflag = PCATCH;
1421 	gotiod = FALSE;
1422 
1423 	/*
1424 	 * Find a free iod to process this request.
1425 	 */
1426 	for (iod = 0; iod < ncl_numasync; iod++)
1427 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1428 			gotiod = TRUE;
1429 			break;
1430 		}
1431 
1432 	/*
1433 	 * Try to create one if none are free.
1434 	 */
1435 	if (!gotiod)
1436 		ncl_nfsiodnew();
1437 	else {
1438 		/*
1439 		 * Found one, so wake it up and tell it which
1440 		 * mount to process.
1441 		 */
1442 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1443 		    iod, nmp));
1444 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1445 		ncl_iodmount[iod] = nmp;
1446 		nmp->nm_bufqiods++;
1447 		wakeup(&ncl_iodwant[iod]);
1448 	}
1449 
1450 	/*
1451 	 * If none are free, we may already have an iod working on this mount
1452 	 * point.  If so, it will process our request.
1453 	 */
1454 	if (!gotiod) {
1455 		if (nmp->nm_bufqiods > 0) {
1456 			NFS_DPF(ASYNCIO,
1457 				("ncl_asyncio: %d iods are already processing mount %p\n",
1458 				 nmp->nm_bufqiods, nmp));
1459 			gotiod = TRUE;
1460 		}
1461 	}
1462 
1463 	/*
1464 	 * If we have an iod which can process the request, then queue
1465 	 * the buffer.
1466 	 */
1467 	if (gotiod) {
1468 		/*
1469 		 * Ensure that the queue never grows too large.  We still want
1470 		 * to asynchronize so we block rather then return EIO.
1471 		 */
1472 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1473 			NFS_DPF(ASYNCIO,
1474 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1475 			nmp->nm_bufqwant = TRUE;
1476 			error = newnfs_msleep(td, &nmp->nm_bufq,
1477 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1478 			   slptimeo);
1479 			if (error) {
1480 				error2 = newnfs_sigintr(nmp, td);
1481 				if (error2) {
1482 					mtx_unlock(&ncl_iod_mutex);
1483 					return (error2);
1484 				}
1485 				if (slpflag == PCATCH) {
1486 					slpflag = 0;
1487 					slptimeo = 2 * hz;
1488 				}
1489 			}
1490 			/*
1491 			 * We might have lost our iod while sleeping,
1492 			 * so check and loop if necessary.
1493 			 */
1494 			goto again;
1495 		}
1496 
1497 		/* We might have lost our nfsiod */
1498 		if (nmp->nm_bufqiods == 0) {
1499 			NFS_DPF(ASYNCIO,
1500 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1501 			goto again;
1502 		}
1503 
1504 		if (bp->b_iocmd == BIO_READ) {
1505 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1506 				bp->b_rcred = crhold(cred);
1507 		} else {
1508 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1509 				bp->b_wcred = crhold(cred);
1510 		}
1511 
1512 		if (bp->b_flags & B_REMFREE)
1513 			bremfreef(bp);
1514 		BUF_KERNPROC(bp);
1515 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1516 		nmp->nm_bufqlen++;
1517 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1518 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1519 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1520 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1521 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1522 		}
1523 		mtx_unlock(&ncl_iod_mutex);
1524 		return (0);
1525 	}
1526 
1527 	mtx_unlock(&ncl_iod_mutex);
1528 
1529 	/*
1530 	 * All the iods are busy on other mounts, so return EIO to
1531 	 * force the caller to process the i/o synchronously.
1532 	 */
1533 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1534 	return (EIO);
1535 }
1536 
1537 void
1538 ncl_doio_directwrite(struct buf *bp)
1539 {
1540 	int iomode, must_commit;
1541 	struct uio *uiop = (struct uio *)bp->b_caller1;
1542 	char *iov_base = uiop->uio_iov->iov_base;
1543 
1544 	iomode = NFSWRITE_FILESYNC;
1545 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1546 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1547 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1548 	free(iov_base, M_NFSDIRECTIO);
1549 	free(uiop->uio_iov, M_NFSDIRECTIO);
1550 	free(uiop, M_NFSDIRECTIO);
1551 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1552 		struct nfsnode *np = VTONFS(bp->b_vp);
1553 		mtx_lock(&np->n_mtx);
1554 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1555 			/*
1556 			 * Invalidate the attribute cache, since writes to a DS
1557 			 * won't update the size attribute.
1558 			 */
1559 			np->n_attrstamp = 0;
1560 		}
1561 		np->n_directio_asyncwr--;
1562 		if (np->n_directio_asyncwr == 0) {
1563 			np->n_flag &= ~NMODIFIED;
1564 			if ((np->n_flag & NFSYNCWAIT)) {
1565 				np->n_flag &= ~NFSYNCWAIT;
1566 				wakeup((caddr_t)&np->n_directio_asyncwr);
1567 			}
1568 		}
1569 		mtx_unlock(&np->n_mtx);
1570 	}
1571 	bp->b_vp = NULL;
1572 	relpbuf(bp, &ncl_pbuf_freecnt);
1573 }
1574 
1575 /*
1576  * Do an I/O operation to/from a cache block. This may be called
1577  * synchronously or from an nfsiod.
1578  */
1579 int
1580 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1581     int called_from_strategy)
1582 {
1583 	struct uio *uiop;
1584 	struct nfsnode *np;
1585 	struct nfsmount *nmp;
1586 	int error = 0, iomode, must_commit = 0;
1587 	struct uio uio;
1588 	struct iovec io;
1589 	struct proc *p = td ? td->td_proc : NULL;
1590 	uint8_t	iocmd;
1591 
1592 	np = VTONFS(vp);
1593 	nmp = VFSTONFS(vp->v_mount);
1594 	uiop = &uio;
1595 	uiop->uio_iov = &io;
1596 	uiop->uio_iovcnt = 1;
1597 	uiop->uio_segflg = UIO_SYSSPACE;
1598 	uiop->uio_td = td;
1599 
1600 	/*
1601 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1602 	 * do this here so we do not have to do it in all the code that
1603 	 * calls us.
1604 	 */
1605 	bp->b_flags &= ~B_INVAL;
1606 	bp->b_ioflags &= ~BIO_ERROR;
1607 
1608 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1609 	iocmd = bp->b_iocmd;
1610 	if (iocmd == BIO_READ) {
1611 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1612 	    io.iov_base = bp->b_data;
1613 	    uiop->uio_rw = UIO_READ;
1614 
1615 	    switch (vp->v_type) {
1616 	    case VREG:
1617 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1618 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1619 		error = ncl_readrpc(vp, uiop, cr);
1620 
1621 		if (!error) {
1622 		    if (uiop->uio_resid) {
1623 			/*
1624 			 * If we had a short read with no error, we must have
1625 			 * hit a file hole.  We should zero-fill the remainder.
1626 			 * This can also occur if the server hits the file EOF.
1627 			 *
1628 			 * Holes used to be able to occur due to pending
1629 			 * writes, but that is not possible any longer.
1630 			 */
1631 			int nread = bp->b_bcount - uiop->uio_resid;
1632 			ssize_t left = uiop->uio_resid;
1633 
1634 			if (left > 0)
1635 				bzero((char *)bp->b_data + nread, left);
1636 			uiop->uio_resid = 0;
1637 		    }
1638 		}
1639 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1640 		if (p && (vp->v_vflag & VV_TEXT)) {
1641 			mtx_lock(&np->n_mtx);
1642 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1643 				mtx_unlock(&np->n_mtx);
1644 				PROC_LOCK(p);
1645 				killproc(p, "text file modification");
1646 				PROC_UNLOCK(p);
1647 			} else
1648 				mtx_unlock(&np->n_mtx);
1649 		}
1650 		break;
1651 	    case VLNK:
1652 		uiop->uio_offset = (off_t)0;
1653 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1654 		error = ncl_readlinkrpc(vp, uiop, cr);
1655 		break;
1656 	    case VDIR:
1657 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1658 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1659 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1660 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1661 			if (error == NFSERR_NOTSUPP)
1662 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1663 		}
1664 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1665 			error = ncl_readdirrpc(vp, uiop, cr, td);
1666 		/*
1667 		 * end-of-directory sets B_INVAL but does not generate an
1668 		 * error.
1669 		 */
1670 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1671 			bp->b_flags |= B_INVAL;
1672 		break;
1673 	    default:
1674 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1675 		break;
1676 	    }
1677 	    if (error) {
1678 		bp->b_ioflags |= BIO_ERROR;
1679 		bp->b_error = error;
1680 	    }
1681 	} else {
1682 	    /*
1683 	     * If we only need to commit, try to commit
1684 	     */
1685 	    if (bp->b_flags & B_NEEDCOMMIT) {
1686 		    int retv;
1687 		    off_t off;
1688 
1689 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1690 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1691 			bp->b_wcred, td);
1692 		    if (retv == 0) {
1693 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1694 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1695 			    bp->b_resid = 0;
1696 			    bufdone(bp);
1697 			    return (0);
1698 		    }
1699 		    if (retv == NFSERR_STALEWRITEVERF) {
1700 			    ncl_clearcommit(vp->v_mount);
1701 		    }
1702 	    }
1703 
1704 	    /*
1705 	     * Setup for actual write
1706 	     */
1707 	    mtx_lock(&np->n_mtx);
1708 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1709 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1710 	    mtx_unlock(&np->n_mtx);
1711 
1712 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1713 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1714 		    - bp->b_dirtyoff;
1715 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1716 		    + bp->b_dirtyoff;
1717 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1718 		uiop->uio_rw = UIO_WRITE;
1719 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1720 
1721 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1722 		    iomode = NFSWRITE_UNSTABLE;
1723 		else
1724 		    iomode = NFSWRITE_FILESYNC;
1725 
1726 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1727 		    called_from_strategy);
1728 
1729 		/*
1730 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1731 		 * to cluster the buffers needing commit.  This will allow
1732 		 * the system to submit a single commit rpc for the whole
1733 		 * cluster.  We can do this even if the buffer is not 100%
1734 		 * dirty (relative to the NFS blocksize), so we optimize the
1735 		 * append-to-file-case.
1736 		 *
1737 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1738 		 * cleared because write clustering only works for commit
1739 		 * rpc's, not for the data portion of the write).
1740 		 */
1741 
1742 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1743 		    bp->b_flags |= B_NEEDCOMMIT;
1744 		    if (bp->b_dirtyoff == 0
1745 			&& bp->b_dirtyend == bp->b_bcount)
1746 			bp->b_flags |= B_CLUSTEROK;
1747 		} else {
1748 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1749 		}
1750 
1751 		/*
1752 		 * For an interrupted write, the buffer is still valid
1753 		 * and the write hasn't been pushed to the server yet,
1754 		 * so we can't set BIO_ERROR and report the interruption
1755 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1756 		 * is not relevant, so the rpc attempt is essentially
1757 		 * a noop.  For the case of a V3 write rpc not being
1758 		 * committed to stable storage, the block is still
1759 		 * dirty and requires either a commit rpc or another
1760 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1761 		 * the block is reused. This is indicated by setting
1762 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1763 		 *
1764 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1765 		 * write error and is handled as above, except that
1766 		 * B_EINTR isn't set. One cause of this is a stale stateid
1767 		 * error for the RPC that indicates recovery is required,
1768 		 * when called with called_from_strategy != 0.
1769 		 *
1770 		 * If the buffer is marked B_PAGING, it does not reside on
1771 		 * the vp's paging queues so we cannot call bdirty().  The
1772 		 * bp in this case is not an NFS cache block so we should
1773 		 * be safe. XXX
1774 		 *
1775 		 * The logic below breaks up errors into recoverable and
1776 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1777 		 * and keep the buffer around for potential write retries.
1778 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1779 		 * and save the error in the nfsnode. This is less than ideal
1780 		 * but necessary. Keeping such buffers around could potentially
1781 		 * cause buffer exhaustion eventually (they can never be written
1782 		 * out, so will get constantly be re-dirtied). It also causes
1783 		 * all sorts of vfs panics. For non-recoverable write errors,
1784 		 * also invalidate the attrcache, so we'll be forced to go over
1785 		 * the wire for this object, returning an error to user on next
1786 		 * call (most of the time).
1787 		 */
1788 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1789 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1790 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1791 			if ((bp->b_flags & B_PAGING) == 0) {
1792 			    bdirty(bp);
1793 			    bp->b_flags &= ~B_DONE;
1794 			}
1795 			if ((error == EINTR || error == ETIMEDOUT) &&
1796 			    (bp->b_flags & B_ASYNC) == 0)
1797 			    bp->b_flags |= B_EINTR;
1798 		} else {
1799 		    if (error) {
1800 			bp->b_ioflags |= BIO_ERROR;
1801 			bp->b_flags |= B_INVAL;
1802 			bp->b_error = np->n_error = error;
1803 			mtx_lock(&np->n_mtx);
1804 			np->n_flag |= NWRITEERR;
1805 			np->n_attrstamp = 0;
1806 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1807 			mtx_unlock(&np->n_mtx);
1808 		    }
1809 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1810 		}
1811 	    } else {
1812 		bp->b_resid = 0;
1813 		bufdone(bp);
1814 		return (0);
1815 	    }
1816 	}
1817 	bp->b_resid = uiop->uio_resid;
1818 	if (must_commit)
1819 	    ncl_clearcommit(vp->v_mount);
1820 	bufdone(bp);
1821 	return (error);
1822 }
1823 
1824 /*
1825  * Used to aid in handling ftruncate() operations on the NFS client side.
1826  * Truncation creates a number of special problems for NFS.  We have to
1827  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1828  * we have to properly handle VM pages or (potentially dirty) buffers
1829  * that straddle the truncation point.
1830  */
1831 
1832 int
1833 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1834 {
1835 	struct nfsnode *np = VTONFS(vp);
1836 	u_quad_t tsize;
1837 	int biosize = vp->v_bufobj.bo_bsize;
1838 	int error = 0;
1839 
1840 	mtx_lock(&np->n_mtx);
1841 	tsize = np->n_size;
1842 	np->n_size = nsize;
1843 	mtx_unlock(&np->n_mtx);
1844 
1845 	if (nsize < tsize) {
1846 		struct buf *bp;
1847 		daddr_t lbn;
1848 		int bufsize;
1849 
1850 		/*
1851 		 * vtruncbuf() doesn't get the buffer overlapping the
1852 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1853 		 * buffer that now needs to be truncated.
1854 		 */
1855 		error = vtruncbuf(vp, cred, nsize, biosize);
1856 		lbn = nsize / biosize;
1857 		bufsize = nsize - (lbn * biosize);
1858 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1859 		if (!bp)
1860 			return EINTR;
1861 		if (bp->b_dirtyoff > bp->b_bcount)
1862 			bp->b_dirtyoff = bp->b_bcount;
1863 		if (bp->b_dirtyend > bp->b_bcount)
1864 			bp->b_dirtyend = bp->b_bcount;
1865 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1866 		brelse(bp);
1867 	} else {
1868 		vnode_pager_setsize(vp, nsize);
1869 	}
1870 	return(error);
1871 }
1872 
1873