xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
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 	npages = 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 	/*
130 	 * If the requested page is partially valid, just return it and
131 	 * allow the pager to zero-out the blanks.  Partially valid pages
132 	 * can only occur at the file EOF.
133 	 *
134 	 * XXXGL: is that true for NFS, where short read can occur???
135 	 */
136 	VM_OBJECT_WLOCK(object);
137 	if (pages[npages - 1]->valid != 0 && --npages == 0)
138 		goto out;
139 	VM_OBJECT_WUNLOCK(object);
140 
141 	/*
142 	 * We use only the kva address for the buffer, but this is extremely
143 	 * convienient and fast.
144 	 */
145 	bp = getpbuf(&ncl_pbuf_freecnt);
146 
147 	kva = (vm_offset_t) bp->b_data;
148 	pmap_qenter(kva, pages, npages);
149 	PCPU_INC(cnt.v_vnodein);
150 	PCPU_ADD(cnt.v_vnodepgsin, npages);
151 
152 	count = npages << PAGE_SHIFT;
153 	iov.iov_base = (caddr_t) kva;
154 	iov.iov_len = count;
155 	uio.uio_iov = &iov;
156 	uio.uio_iovcnt = 1;
157 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
158 	uio.uio_resid = count;
159 	uio.uio_segflg = UIO_SYSSPACE;
160 	uio.uio_rw = UIO_READ;
161 	uio.uio_td = td;
162 
163 	error = ncl_readrpc(vp, &uio, cred);
164 	pmap_qremove(kva, npages);
165 
166 	relpbuf(bp, &ncl_pbuf_freecnt);
167 
168 	if (error && (uio.uio_resid == count)) {
169 		ncl_printf("nfs_getpages: error %d\n", error);
170 		return (VM_PAGER_ERROR);
171 	}
172 
173 	/*
174 	 * Calculate the number of bytes read and validate only that number
175 	 * of bytes.  Note that due to pending writes, size may be 0.  This
176 	 * does not mean that the remaining data is invalid!
177 	 */
178 
179 	size = count - uio.uio_resid;
180 	VM_OBJECT_WLOCK(object);
181 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
182 		vm_page_t m;
183 		nextoff = toff + PAGE_SIZE;
184 		m = pages[i];
185 
186 		if (nextoff <= size) {
187 			/*
188 			 * Read operation filled an entire page
189 			 */
190 			m->valid = VM_PAGE_BITS_ALL;
191 			KASSERT(m->dirty == 0,
192 			    ("nfs_getpages: page %p is dirty", m));
193 		} else if (size > toff) {
194 			/*
195 			 * Read operation filled a partial page.
196 			 */
197 			m->valid = 0;
198 			vm_page_set_valid_range(m, 0, size - toff);
199 			KASSERT(m->dirty == 0,
200 			    ("nfs_getpages: page %p is dirty", m));
201 		} else {
202 			/*
203 			 * Read operation was short.  If no error
204 			 * occured we may have hit a zero-fill
205 			 * section.  We leave valid set to 0, and page
206 			 * is freed by vm_page_readahead_finish() if
207 			 * its index is not equal to requested, or
208 			 * page is zeroed and set valid by
209 			 * vm_pager_get_pages() for requested page.
210 			 */
211 			;
212 		}
213 	}
214 out:
215 	VM_OBJECT_WUNLOCK(object);
216 	if (ap->a_rbehind)
217 		*ap->a_rbehind = 0;
218 	if (ap->a_rahead)
219 		*ap->a_rahead = 0;
220 	return (VM_PAGER_OK);
221 }
222 
223 /*
224  * Vnode op for VM putpages.
225  */
226 int
227 ncl_putpages(struct vop_putpages_args *ap)
228 {
229 	struct uio uio;
230 	struct iovec iov;
231 	vm_offset_t kva;
232 	struct buf *bp;
233 	int iomode, must_commit, i, error, npages, count;
234 	off_t offset;
235 	int *rtvals;
236 	struct vnode *vp;
237 	struct thread *td;
238 	struct ucred *cred;
239 	struct nfsmount *nmp;
240 	struct nfsnode *np;
241 	vm_page_t *pages;
242 
243 	vp = ap->a_vp;
244 	np = VTONFS(vp);
245 	td = curthread;				/* XXX */
246 	/* Set the cred to n_writecred for the write rpcs. */
247 	if (np->n_writecred != NULL)
248 		cred = crhold(np->n_writecred);
249 	else
250 		cred = crhold(curthread->td_ucred);	/* XXX */
251 	nmp = VFSTONFS(vp->v_mount);
252 	pages = ap->a_m;
253 	count = ap->a_count;
254 	rtvals = ap->a_rtvals;
255 	npages = btoc(count);
256 	offset = IDX_TO_OFF(pages[0]->pindex);
257 
258 	mtx_lock(&nmp->nm_mtx);
259 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
260 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
261 		mtx_unlock(&nmp->nm_mtx);
262 		(void)ncl_fsinfo(nmp, vp, cred, td);
263 	} else
264 		mtx_unlock(&nmp->nm_mtx);
265 
266 	mtx_lock(&np->n_mtx);
267 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
268 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
269 		mtx_unlock(&np->n_mtx);
270 		ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
271 		mtx_lock(&np->n_mtx);
272 	}
273 
274 	for (i = 0; i < npages; i++)
275 		rtvals[i] = VM_PAGER_ERROR;
276 
277 	/*
278 	 * When putting pages, do not extend file past EOF.
279 	 */
280 	if (offset + count > np->n_size) {
281 		count = np->n_size - offset;
282 		if (count < 0)
283 			count = 0;
284 	}
285 	mtx_unlock(&np->n_mtx);
286 
287 	/*
288 	 * We use only the kva address for the buffer, but this is extremely
289 	 * convienient and fast.
290 	 */
291 	bp = getpbuf(&ncl_pbuf_freecnt);
292 
293 	kva = (vm_offset_t) bp->b_data;
294 	pmap_qenter(kva, pages, npages);
295 	PCPU_INC(cnt.v_vnodeout);
296 	PCPU_ADD(cnt.v_vnodepgsout, count);
297 
298 	iov.iov_base = (caddr_t) kva;
299 	iov.iov_len = count;
300 	uio.uio_iov = &iov;
301 	uio.uio_iovcnt = 1;
302 	uio.uio_offset = offset;
303 	uio.uio_resid = count;
304 	uio.uio_segflg = UIO_SYSSPACE;
305 	uio.uio_rw = UIO_WRITE;
306 	uio.uio_td = td;
307 
308 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
309 	    iomode = NFSWRITE_UNSTABLE;
310 	else
311 	    iomode = NFSWRITE_FILESYNC;
312 
313 	error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
314 	crfree(cred);
315 
316 	pmap_qremove(kva, npages);
317 	relpbuf(bp, &ncl_pbuf_freecnt);
318 
319 	if (error == 0 || !nfs_keep_dirty_on_error) {
320 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
321 		if (must_commit)
322 			ncl_clearcommit(vp->v_mount);
323 	}
324 	return rtvals[0];
325 }
326 
327 /*
328  * For nfs, cache consistency can only be maintained approximately.
329  * Although RFC1094 does not specify the criteria, the following is
330  * believed to be compatible with the reference port.
331  * For nfs:
332  * If the file's modify time on the server has changed since the
333  * last read rpc or you have written to the file,
334  * you may have lost data cache consistency with the
335  * server, so flush all of the file's data out of the cache.
336  * Then force a getattr rpc to ensure that you have up to date
337  * attributes.
338  * NB: This implies that cache data can be read when up to
339  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
340  * attributes this could be forced by setting n_attrstamp to 0 before
341  * the VOP_GETATTR() call.
342  */
343 static inline int
344 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
345 {
346 	int error = 0;
347 	struct vattr vattr;
348 	struct nfsnode *np = VTONFS(vp);
349 	int old_lock;
350 
351 	/*
352 	 * Grab the exclusive lock before checking whether the cache is
353 	 * consistent.
354 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
355 	 * But for now, this suffices.
356 	 */
357 	old_lock = ncl_upgrade_vnlock(vp);
358 	if (vp->v_iflag & VI_DOOMED) {
359 		ncl_downgrade_vnlock(vp, old_lock);
360 		return (EBADF);
361 	}
362 
363 	mtx_lock(&np->n_mtx);
364 	if (np->n_flag & NMODIFIED) {
365 		mtx_unlock(&np->n_mtx);
366 		if (vp->v_type != VREG) {
367 			if (vp->v_type != VDIR)
368 				panic("nfs: bioread, not dir");
369 			ncl_invaldir(vp);
370 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
371 			if (error)
372 				goto out;
373 		}
374 		np->n_attrstamp = 0;
375 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
376 		error = VOP_GETATTR(vp, &vattr, cred);
377 		if (error)
378 			goto out;
379 		mtx_lock(&np->n_mtx);
380 		np->n_mtime = vattr.va_mtime;
381 		mtx_unlock(&np->n_mtx);
382 	} else {
383 		mtx_unlock(&np->n_mtx);
384 		error = VOP_GETATTR(vp, &vattr, cred);
385 		if (error)
386 			return (error);
387 		mtx_lock(&np->n_mtx);
388 		if ((np->n_flag & NSIZECHANGED)
389 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
390 			mtx_unlock(&np->n_mtx);
391 			if (vp->v_type == VDIR)
392 				ncl_invaldir(vp);
393 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
394 			if (error)
395 				goto out;
396 			mtx_lock(&np->n_mtx);
397 			np->n_mtime = vattr.va_mtime;
398 			np->n_flag &= ~NSIZECHANGED;
399 		}
400 		mtx_unlock(&np->n_mtx);
401 	}
402 out:
403 	ncl_downgrade_vnlock(vp, old_lock);
404 	return error;
405 }
406 
407 /*
408  * Vnode op for read using bio
409  */
410 int
411 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
412 {
413 	struct nfsnode *np = VTONFS(vp);
414 	int biosize, i;
415 	struct buf *bp, *rabp;
416 	struct thread *td;
417 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
418 	daddr_t lbn, rabn;
419 	int bcount;
420 	int seqcount;
421 	int nra, error = 0, n = 0, on = 0;
422 	off_t tmp_off;
423 
424 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
425 	if (uio->uio_resid == 0)
426 		return (0);
427 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
428 		return (EINVAL);
429 	td = uio->uio_td;
430 
431 	mtx_lock(&nmp->nm_mtx);
432 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
433 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
434 		mtx_unlock(&nmp->nm_mtx);
435 		(void)ncl_fsinfo(nmp, vp, cred, td);
436 		mtx_lock(&nmp->nm_mtx);
437 	}
438 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
439 		(void) newnfs_iosize(nmp);
440 
441 	tmp_off = uio->uio_offset + uio->uio_resid;
442 	if (vp->v_type != VDIR &&
443 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
444 		mtx_unlock(&nmp->nm_mtx);
445 		return (EFBIG);
446 	}
447 	mtx_unlock(&nmp->nm_mtx);
448 
449 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
450 		/* No caching/ no readaheads. Just read data into the user buffer */
451 		return ncl_readrpc(vp, uio, cred);
452 
453 	biosize = vp->v_bufobj.bo_bsize;
454 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
455 
456 	error = nfs_bioread_check_cons(vp, td, cred);
457 	if (error)
458 		return error;
459 
460 	do {
461 	    u_quad_t nsize;
462 
463 	    mtx_lock(&np->n_mtx);
464 	    nsize = np->n_size;
465 	    mtx_unlock(&np->n_mtx);
466 
467 	    switch (vp->v_type) {
468 	    case VREG:
469 		NFSINCRGLOBAL(newnfsstats.biocache_reads);
470 		lbn = uio->uio_offset / biosize;
471 		on = uio->uio_offset - (lbn * biosize);
472 
473 		/*
474 		 * Start the read ahead(s), as required.
475 		 */
476 		if (nmp->nm_readahead > 0) {
477 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
478 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
479 			rabn = lbn + 1 + nra;
480 			if (incore(&vp->v_bufobj, rabn) == NULL) {
481 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
482 			    if (!rabp) {
483 				error = newnfs_sigintr(nmp, td);
484 				return (error ? error : EINTR);
485 			    }
486 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
487 				rabp->b_flags |= B_ASYNC;
488 				rabp->b_iocmd = BIO_READ;
489 				vfs_busy_pages(rabp, 0);
490 				if (ncl_asyncio(nmp, rabp, cred, td)) {
491 				    rabp->b_flags |= B_INVAL;
492 				    rabp->b_ioflags |= BIO_ERROR;
493 				    vfs_unbusy_pages(rabp);
494 				    brelse(rabp);
495 				    break;
496 				}
497 			    } else {
498 				brelse(rabp);
499 			    }
500 			}
501 		    }
502 		}
503 
504 		/* Note that bcount is *not* DEV_BSIZE aligned. */
505 		bcount = biosize;
506 		if ((off_t)lbn * biosize >= nsize) {
507 			bcount = 0;
508 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
509 			bcount = nsize - (off_t)lbn * biosize;
510 		}
511 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
512 
513 		if (!bp) {
514 			error = newnfs_sigintr(nmp, td);
515 			return (error ? error : EINTR);
516 		}
517 
518 		/*
519 		 * If B_CACHE is not set, we must issue the read.  If this
520 		 * fails, we return an error.
521 		 */
522 
523 		if ((bp->b_flags & B_CACHE) == 0) {
524 		    bp->b_iocmd = BIO_READ;
525 		    vfs_busy_pages(bp, 0);
526 		    error = ncl_doio(vp, bp, cred, td, 0);
527 		    if (error) {
528 			brelse(bp);
529 			return (error);
530 		    }
531 		}
532 
533 		/*
534 		 * on is the offset into the current bp.  Figure out how many
535 		 * bytes we can copy out of the bp.  Note that bcount is
536 		 * NOT DEV_BSIZE aligned.
537 		 *
538 		 * Then figure out how many bytes we can copy into the uio.
539 		 */
540 
541 		n = 0;
542 		if (on < bcount)
543 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
544 		break;
545 	    case VLNK:
546 		NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
547 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
548 		if (!bp) {
549 			error = newnfs_sigintr(nmp, td);
550 			return (error ? error : EINTR);
551 		}
552 		if ((bp->b_flags & B_CACHE) == 0) {
553 		    bp->b_iocmd = BIO_READ;
554 		    vfs_busy_pages(bp, 0);
555 		    error = ncl_doio(vp, bp, cred, td, 0);
556 		    if (error) {
557 			bp->b_ioflags |= BIO_ERROR;
558 			brelse(bp);
559 			return (error);
560 		    }
561 		}
562 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
563 		on = 0;
564 		break;
565 	    case VDIR:
566 		NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
567 		if (np->n_direofoffset
568 		    && uio->uio_offset >= np->n_direofoffset) {
569 		    return (0);
570 		}
571 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
572 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
573 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
574 		if (!bp) {
575 		    error = newnfs_sigintr(nmp, td);
576 		    return (error ? error : EINTR);
577 		}
578 		if ((bp->b_flags & B_CACHE) == 0) {
579 		    bp->b_iocmd = BIO_READ;
580 		    vfs_busy_pages(bp, 0);
581 		    error = ncl_doio(vp, bp, cred, td, 0);
582 		    if (error) {
583 			    brelse(bp);
584 		    }
585 		    while (error == NFSERR_BAD_COOKIE) {
586 			ncl_invaldir(vp);
587 			error = ncl_vinvalbuf(vp, 0, td, 1);
588 			/*
589 			 * Yuck! The directory has been modified on the
590 			 * server. The only way to get the block is by
591 			 * reading from the beginning to get all the
592 			 * offset cookies.
593 			 *
594 			 * Leave the last bp intact unless there is an error.
595 			 * Loop back up to the while if the error is another
596 			 * NFSERR_BAD_COOKIE (double yuch!).
597 			 */
598 			for (i = 0; i <= lbn && !error; i++) {
599 			    if (np->n_direofoffset
600 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
601 				    return (0);
602 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
603 			    if (!bp) {
604 				error = newnfs_sigintr(nmp, td);
605 				return (error ? error : EINTR);
606 			    }
607 			    if ((bp->b_flags & B_CACHE) == 0) {
608 				    bp->b_iocmd = BIO_READ;
609 				    vfs_busy_pages(bp, 0);
610 				    error = ncl_doio(vp, bp, cred, td, 0);
611 				    /*
612 				     * no error + B_INVAL == directory EOF,
613 				     * use the block.
614 				     */
615 				    if (error == 0 && (bp->b_flags & B_INVAL))
616 					    break;
617 			    }
618 			    /*
619 			     * An error will throw away the block and the
620 			     * for loop will break out.  If no error and this
621 			     * is not the block we want, we throw away the
622 			     * block and go for the next one via the for loop.
623 			     */
624 			    if (error || i < lbn)
625 				    brelse(bp);
626 			}
627 		    }
628 		    /*
629 		     * The above while is repeated if we hit another cookie
630 		     * error.  If we hit an error and it wasn't a cookie error,
631 		     * we give up.
632 		     */
633 		    if (error)
634 			    return (error);
635 		}
636 
637 		/*
638 		 * If not eof and read aheads are enabled, start one.
639 		 * (You need the current block first, so that you have the
640 		 *  directory offset cookie of the next block.)
641 		 */
642 		if (nmp->nm_readahead > 0 &&
643 		    (bp->b_flags & B_INVAL) == 0 &&
644 		    (np->n_direofoffset == 0 ||
645 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
646 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
647 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
648 			if (rabp) {
649 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
650 				rabp->b_flags |= B_ASYNC;
651 				rabp->b_iocmd = BIO_READ;
652 				vfs_busy_pages(rabp, 0);
653 				if (ncl_asyncio(nmp, rabp, cred, td)) {
654 				    rabp->b_flags |= B_INVAL;
655 				    rabp->b_ioflags |= BIO_ERROR;
656 				    vfs_unbusy_pages(rabp);
657 				    brelse(rabp);
658 				}
659 			    } else {
660 				brelse(rabp);
661 			    }
662 			}
663 		}
664 		/*
665 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
666 		 * chopped for the EOF condition, we cannot tell how large
667 		 * NFS directories are going to be until we hit EOF.  So
668 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
669 		 * it just so happens that b_resid will effectively chop it
670 		 * to EOF.  *BUT* this information is lost if the buffer goes
671 		 * away and is reconstituted into a B_CACHE state ( due to
672 		 * being VMIO ) later.  So we keep track of the directory eof
673 		 * in np->n_direofoffset and chop it off as an extra step
674 		 * right here.
675 		 */
676 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
677 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
678 			n = np->n_direofoffset - uio->uio_offset;
679 		break;
680 	    default:
681 		ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
682 		bp = NULL;
683 		break;
684 	    };
685 
686 	    if (n > 0) {
687 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
688 	    }
689 	    if (vp->v_type == VLNK)
690 		n = 0;
691 	    if (bp != NULL)
692 		brelse(bp);
693 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
694 	return (error);
695 }
696 
697 /*
698  * The NFS write path cannot handle iovecs with len > 1. So we need to
699  * break up iovecs accordingly (restricting them to wsize).
700  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
701  * For the ASYNC case, 2 copies are needed. The first a copy from the
702  * user buffer to a staging buffer and then a second copy from the staging
703  * buffer to mbufs. This can be optimized by copying from the user buffer
704  * directly into mbufs and passing the chain down, but that requires a
705  * fair amount of re-working of the relevant codepaths (and can be done
706  * later).
707  */
708 static int
709 nfs_directio_write(vp, uiop, cred, ioflag)
710 	struct vnode *vp;
711 	struct uio *uiop;
712 	struct ucred *cred;
713 	int ioflag;
714 {
715 	int error;
716 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
717 	struct thread *td = uiop->uio_td;
718 	int size;
719 	int wsize;
720 
721 	mtx_lock(&nmp->nm_mtx);
722 	wsize = nmp->nm_wsize;
723 	mtx_unlock(&nmp->nm_mtx);
724 	if (ioflag & IO_SYNC) {
725 		int iomode, must_commit;
726 		struct uio uio;
727 		struct iovec iov;
728 do_sync:
729 		while (uiop->uio_resid > 0) {
730 			size = MIN(uiop->uio_resid, wsize);
731 			size = MIN(uiop->uio_iov->iov_len, size);
732 			iov.iov_base = uiop->uio_iov->iov_base;
733 			iov.iov_len = size;
734 			uio.uio_iov = &iov;
735 			uio.uio_iovcnt = 1;
736 			uio.uio_offset = uiop->uio_offset;
737 			uio.uio_resid = size;
738 			uio.uio_segflg = UIO_USERSPACE;
739 			uio.uio_rw = UIO_WRITE;
740 			uio.uio_td = td;
741 			iomode = NFSWRITE_FILESYNC;
742 			error = ncl_writerpc(vp, &uio, cred, &iomode,
743 			    &must_commit, 0);
744 			KASSERT((must_commit == 0),
745 				("ncl_directio_write: Did not commit write"));
746 			if (error)
747 				return (error);
748 			uiop->uio_offset += size;
749 			uiop->uio_resid -= size;
750 			if (uiop->uio_iov->iov_len <= size) {
751 				uiop->uio_iovcnt--;
752 				uiop->uio_iov++;
753 			} else {
754 				uiop->uio_iov->iov_base =
755 					(char *)uiop->uio_iov->iov_base + size;
756 				uiop->uio_iov->iov_len -= size;
757 			}
758 		}
759 	} else {
760 		struct uio *t_uio;
761 		struct iovec *t_iov;
762 		struct buf *bp;
763 
764 		/*
765 		 * Break up the write into blocksize chunks and hand these
766 		 * over to nfsiod's for write back.
767 		 * Unfortunately, this incurs a copy of the data. Since
768 		 * the user could modify the buffer before the write is
769 		 * initiated.
770 		 *
771 		 * The obvious optimization here is that one of the 2 copies
772 		 * in the async write path can be eliminated by copying the
773 		 * data here directly into mbufs and passing the mbuf chain
774 		 * down. But that will require a fair amount of re-working
775 		 * of the code and can be done if there's enough interest
776 		 * in NFS directio access.
777 		 */
778 		while (uiop->uio_resid > 0) {
779 			size = MIN(uiop->uio_resid, wsize);
780 			size = MIN(uiop->uio_iov->iov_len, size);
781 			bp = getpbuf(&ncl_pbuf_freecnt);
782 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
783 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
784 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
785 			t_iov->iov_len = size;
786 			t_uio->uio_iov = t_iov;
787 			t_uio->uio_iovcnt = 1;
788 			t_uio->uio_offset = uiop->uio_offset;
789 			t_uio->uio_resid = size;
790 			t_uio->uio_segflg = UIO_SYSSPACE;
791 			t_uio->uio_rw = UIO_WRITE;
792 			t_uio->uio_td = td;
793 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
794 			    uiop->uio_segflg == UIO_SYSSPACE,
795 			    ("nfs_directio_write: Bad uio_segflg"));
796 			if (uiop->uio_segflg == UIO_USERSPACE) {
797 				error = copyin(uiop->uio_iov->iov_base,
798 				    t_iov->iov_base, size);
799 				if (error != 0)
800 					goto err_free;
801 			} else
802 				/*
803 				 * UIO_SYSSPACE may never happen, but handle
804 				 * it just in case it does.
805 				 */
806 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
807 				    size);
808 			bp->b_flags |= B_DIRECT;
809 			bp->b_iocmd = BIO_WRITE;
810 			if (cred != NOCRED) {
811 				crhold(cred);
812 				bp->b_wcred = cred;
813 			} else
814 				bp->b_wcred = NOCRED;
815 			bp->b_caller1 = (void *)t_uio;
816 			bp->b_vp = vp;
817 			error = ncl_asyncio(nmp, bp, NOCRED, td);
818 err_free:
819 			if (error) {
820 				free(t_iov->iov_base, M_NFSDIRECTIO);
821 				free(t_iov, M_NFSDIRECTIO);
822 				free(t_uio, M_NFSDIRECTIO);
823 				bp->b_vp = NULL;
824 				relpbuf(bp, &ncl_pbuf_freecnt);
825 				if (error == EINTR)
826 					return (error);
827 				goto do_sync;
828 			}
829 			uiop->uio_offset += size;
830 			uiop->uio_resid -= size;
831 			if (uiop->uio_iov->iov_len <= size) {
832 				uiop->uio_iovcnt--;
833 				uiop->uio_iov++;
834 			} else {
835 				uiop->uio_iov->iov_base =
836 					(char *)uiop->uio_iov->iov_base + size;
837 				uiop->uio_iov->iov_len -= size;
838 			}
839 		}
840 	}
841 	return (0);
842 }
843 
844 /*
845  * Vnode op for write using bio
846  */
847 int
848 ncl_write(struct vop_write_args *ap)
849 {
850 	int biosize;
851 	struct uio *uio = ap->a_uio;
852 	struct thread *td = uio->uio_td;
853 	struct vnode *vp = ap->a_vp;
854 	struct nfsnode *np = VTONFS(vp);
855 	struct ucred *cred = ap->a_cred;
856 	int ioflag = ap->a_ioflag;
857 	struct buf *bp;
858 	struct vattr vattr;
859 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
860 	daddr_t lbn;
861 	int bcount, noncontig_write, obcount;
862 	int bp_cached, n, on, error = 0, error1, wouldcommit;
863 	size_t orig_resid, local_resid;
864 	off_t orig_size, tmp_off;
865 
866 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
867 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
868 	    ("ncl_write proc"));
869 	if (vp->v_type != VREG)
870 		return (EIO);
871 	mtx_lock(&np->n_mtx);
872 	if (np->n_flag & NWRITEERR) {
873 		np->n_flag &= ~NWRITEERR;
874 		mtx_unlock(&np->n_mtx);
875 		return (np->n_error);
876 	} else
877 		mtx_unlock(&np->n_mtx);
878 	mtx_lock(&nmp->nm_mtx);
879 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
880 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
881 		mtx_unlock(&nmp->nm_mtx);
882 		(void)ncl_fsinfo(nmp, vp, cred, td);
883 		mtx_lock(&nmp->nm_mtx);
884 	}
885 	if (nmp->nm_wsize == 0)
886 		(void) newnfs_iosize(nmp);
887 	mtx_unlock(&nmp->nm_mtx);
888 
889 	/*
890 	 * Synchronously flush pending buffers if we are in synchronous
891 	 * mode or if we are appending.
892 	 */
893 	if (ioflag & (IO_APPEND | IO_SYNC)) {
894 		mtx_lock(&np->n_mtx);
895 		if (np->n_flag & NMODIFIED) {
896 			mtx_unlock(&np->n_mtx);
897 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
898 			/*
899 			 * Require non-blocking, synchronous writes to
900 			 * dirty files to inform the program it needs
901 			 * to fsync(2) explicitly.
902 			 */
903 			if (ioflag & IO_NDELAY)
904 				return (EAGAIN);
905 #endif
906 			np->n_attrstamp = 0;
907 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
908 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
909 			if (error)
910 				return (error);
911 		} else
912 			mtx_unlock(&np->n_mtx);
913 	}
914 
915 	orig_resid = uio->uio_resid;
916 	mtx_lock(&np->n_mtx);
917 	orig_size = np->n_size;
918 	mtx_unlock(&np->n_mtx);
919 
920 	/*
921 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
922 	 * get the append lock.
923 	 */
924 	if (ioflag & IO_APPEND) {
925 		np->n_attrstamp = 0;
926 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
927 		error = VOP_GETATTR(vp, &vattr, cred);
928 		if (error)
929 			return (error);
930 		mtx_lock(&np->n_mtx);
931 		uio->uio_offset = np->n_size;
932 		mtx_unlock(&np->n_mtx);
933 	}
934 
935 	if (uio->uio_offset < 0)
936 		return (EINVAL);
937 	tmp_off = uio->uio_offset + uio->uio_resid;
938 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
939 		return (EFBIG);
940 	if (uio->uio_resid == 0)
941 		return (0);
942 
943 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
944 		return nfs_directio_write(vp, uio, cred, ioflag);
945 
946 	/*
947 	 * Maybe this should be above the vnode op call, but so long as
948 	 * file servers have no limits, i don't think it matters
949 	 */
950 	if (vn_rlimit_fsize(vp, uio, td))
951 		return (EFBIG);
952 
953 	biosize = vp->v_bufobj.bo_bsize;
954 	/*
955 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
956 	 * would exceed the local maximum per-file write commit size when
957 	 * combined with those, we must decide whether to flush,
958 	 * go synchronous, or return error.  We don't bother checking
959 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
960 	 * no point optimizing for something that really won't ever happen.
961 	 */
962 	wouldcommit = 0;
963 	if (!(ioflag & IO_SYNC)) {
964 		int nflag;
965 
966 		mtx_lock(&np->n_mtx);
967 		nflag = np->n_flag;
968 		mtx_unlock(&np->n_mtx);
969 		if (nflag & NMODIFIED) {
970 			BO_LOCK(&vp->v_bufobj);
971 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
972 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
973 				    b_bobufs) {
974 					if (bp->b_flags & B_NEEDCOMMIT)
975 						wouldcommit += bp->b_bcount;
976 				}
977 			}
978 			BO_UNLOCK(&vp->v_bufobj);
979 		}
980 	}
981 
982 	do {
983 		if (!(ioflag & IO_SYNC)) {
984 			wouldcommit += biosize;
985 			if (wouldcommit > nmp->nm_wcommitsize) {
986 				np->n_attrstamp = 0;
987 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
988 				error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
989 				if (error)
990 					return (error);
991 				wouldcommit = biosize;
992 			}
993 		}
994 
995 		NFSINCRGLOBAL(newnfsstats.biocache_writes);
996 		lbn = uio->uio_offset / biosize;
997 		on = uio->uio_offset - (lbn * biosize);
998 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
999 again:
1000 		/*
1001 		 * Handle direct append and file extension cases, calculate
1002 		 * unaligned buffer size.
1003 		 */
1004 		mtx_lock(&np->n_mtx);
1005 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1006 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1007 			noncontig_write = 1;
1008 		else
1009 			noncontig_write = 0;
1010 		if ((uio->uio_offset == np->n_size ||
1011 		    (noncontig_write != 0 &&
1012 		    lbn == (np->n_size / biosize) &&
1013 		    uio->uio_offset + n > np->n_size)) && n) {
1014 			mtx_unlock(&np->n_mtx);
1015 			/*
1016 			 * Get the buffer (in its pre-append state to maintain
1017 			 * B_CACHE if it was previously set).  Resize the
1018 			 * nfsnode after we have locked the buffer to prevent
1019 			 * readers from reading garbage.
1020 			 */
1021 			obcount = np->n_size - (lbn * biosize);
1022 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1023 
1024 			if (bp != NULL) {
1025 				long save;
1026 
1027 				mtx_lock(&np->n_mtx);
1028 				np->n_size = uio->uio_offset + n;
1029 				np->n_flag |= NMODIFIED;
1030 				vnode_pager_setsize(vp, np->n_size);
1031 				mtx_unlock(&np->n_mtx);
1032 
1033 				save = bp->b_flags & B_CACHE;
1034 				bcount = on + n;
1035 				allocbuf(bp, bcount);
1036 				bp->b_flags |= save;
1037 				if (noncontig_write != 0 && on > obcount)
1038 					vfs_bio_bzero_buf(bp, obcount, on -
1039 					    obcount);
1040 			}
1041 		} else {
1042 			/*
1043 			 * Obtain the locked cache block first, and then
1044 			 * adjust the file's size as appropriate.
1045 			 */
1046 			bcount = on + n;
1047 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1048 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1049 					bcount = biosize;
1050 				else
1051 					bcount = np->n_size - (off_t)lbn * biosize;
1052 			}
1053 			mtx_unlock(&np->n_mtx);
1054 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1055 			mtx_lock(&np->n_mtx);
1056 			if (uio->uio_offset + n > np->n_size) {
1057 				np->n_size = uio->uio_offset + n;
1058 				np->n_flag |= NMODIFIED;
1059 				vnode_pager_setsize(vp, np->n_size);
1060 			}
1061 			mtx_unlock(&np->n_mtx);
1062 		}
1063 
1064 		if (!bp) {
1065 			error = newnfs_sigintr(nmp, td);
1066 			if (!error)
1067 				error = EINTR;
1068 			break;
1069 		}
1070 
1071 		/*
1072 		 * Issue a READ if B_CACHE is not set.  In special-append
1073 		 * mode, B_CACHE is based on the buffer prior to the write
1074 		 * op and is typically set, avoiding the read.  If a read
1075 		 * is required in special append mode, the server will
1076 		 * probably send us a short-read since we extended the file
1077 		 * on our end, resulting in b_resid == 0 and, thusly,
1078 		 * B_CACHE getting set.
1079 		 *
1080 		 * We can also avoid issuing the read if the write covers
1081 		 * the entire buffer.  We have to make sure the buffer state
1082 		 * is reasonable in this case since we will not be initiating
1083 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1084 		 * more information.
1085 		 *
1086 		 * B_CACHE may also be set due to the buffer being cached
1087 		 * normally.
1088 		 */
1089 
1090 		bp_cached = 1;
1091 		if (on == 0 && n == bcount) {
1092 			if ((bp->b_flags & B_CACHE) == 0)
1093 				bp_cached = 0;
1094 			bp->b_flags |= B_CACHE;
1095 			bp->b_flags &= ~B_INVAL;
1096 			bp->b_ioflags &= ~BIO_ERROR;
1097 		}
1098 
1099 		if ((bp->b_flags & B_CACHE) == 0) {
1100 			bp->b_iocmd = BIO_READ;
1101 			vfs_busy_pages(bp, 0);
1102 			error = ncl_doio(vp, bp, cred, td, 0);
1103 			if (error) {
1104 				brelse(bp);
1105 				break;
1106 			}
1107 		}
1108 		if (bp->b_wcred == NOCRED)
1109 			bp->b_wcred = crhold(cred);
1110 		mtx_lock(&np->n_mtx);
1111 		np->n_flag |= NMODIFIED;
1112 		mtx_unlock(&np->n_mtx);
1113 
1114 		/*
1115 		 * If dirtyend exceeds file size, chop it down.  This should
1116 		 * not normally occur but there is an append race where it
1117 		 * might occur XXX, so we log it.
1118 		 *
1119 		 * If the chopping creates a reverse-indexed or degenerate
1120 		 * situation with dirtyoff/end, we 0 both of them.
1121 		 */
1122 
1123 		if (bp->b_dirtyend > bcount) {
1124 			ncl_printf("NFS append race @%lx:%d\n",
1125 			    (long)bp->b_blkno * DEV_BSIZE,
1126 			    bp->b_dirtyend - bcount);
1127 			bp->b_dirtyend = bcount;
1128 		}
1129 
1130 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1131 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1132 
1133 		/*
1134 		 * If the new write will leave a contiguous dirty
1135 		 * area, just update the b_dirtyoff and b_dirtyend,
1136 		 * otherwise force a write rpc of the old dirty area.
1137 		 *
1138 		 * If there has been a file lock applied to this file
1139 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1140 		 * While it is possible to merge discontiguous writes due to
1141 		 * our having a B_CACHE buffer ( and thus valid read data
1142 		 * for the hole), we don't because it could lead to
1143 		 * significant cache coherency problems with multiple clients,
1144 		 * especially if locking is implemented later on.
1145 		 *
1146 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1147 		 * not been file locking done on this file:
1148 		 * Relax coherency a bit for the sake of performance and
1149 		 * expand the current dirty region to contain the new
1150 		 * write even if it means we mark some non-dirty data as
1151 		 * dirty.
1152 		 */
1153 
1154 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1155 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1156 			if (bwrite(bp) == EINTR) {
1157 				error = EINTR;
1158 				break;
1159 			}
1160 			goto again;
1161 		}
1162 
1163 		local_resid = uio->uio_resid;
1164 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1165 
1166 		if (error != 0 && !bp_cached) {
1167 			/*
1168 			 * This block has no other content then what
1169 			 * possibly was written by the faulty uiomove.
1170 			 * Release it, forgetting the data pages, to
1171 			 * prevent the leak of uninitialized data to
1172 			 * usermode.
1173 			 */
1174 			bp->b_ioflags |= BIO_ERROR;
1175 			brelse(bp);
1176 			uio->uio_offset -= local_resid - uio->uio_resid;
1177 			uio->uio_resid = local_resid;
1178 			break;
1179 		}
1180 
1181 		/*
1182 		 * Since this block is being modified, it must be written
1183 		 * again and not just committed.  Since write clustering does
1184 		 * not work for the stage 1 data write, only the stage 2
1185 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1186 		 */
1187 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1188 
1189 		/*
1190 		 * Get the partial update on the progress made from
1191 		 * uiomove, if an error occured.
1192 		 */
1193 		if (error != 0)
1194 			n = local_resid - uio->uio_resid;
1195 
1196 		/*
1197 		 * Only update dirtyoff/dirtyend if not a degenerate
1198 		 * condition.
1199 		 */
1200 		if (n > 0) {
1201 			if (bp->b_dirtyend > 0) {
1202 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1203 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1204 			} else {
1205 				bp->b_dirtyoff = on;
1206 				bp->b_dirtyend = on + n;
1207 			}
1208 			vfs_bio_set_valid(bp, on, n);
1209 		}
1210 
1211 		/*
1212 		 * If IO_SYNC do bwrite().
1213 		 *
1214 		 * IO_INVAL appears to be unused.  The idea appears to be
1215 		 * to turn off caching in this case.  Very odd.  XXX
1216 		 */
1217 		if ((ioflag & IO_SYNC)) {
1218 			if (ioflag & IO_INVAL)
1219 				bp->b_flags |= B_NOCACHE;
1220 			error1 = bwrite(bp);
1221 			if (error1 != 0) {
1222 				if (error == 0)
1223 					error = error1;
1224 				break;
1225 			}
1226 		} else if ((n + on) == biosize) {
1227 			bp->b_flags |= B_ASYNC;
1228 			(void) ncl_writebp(bp, 0, NULL);
1229 		} else {
1230 			bdwrite(bp);
1231 		}
1232 
1233 		if (error != 0)
1234 			break;
1235 	} while (uio->uio_resid > 0 && n > 0);
1236 
1237 	if (error != 0) {
1238 		if (ioflag & IO_UNIT) {
1239 			VATTR_NULL(&vattr);
1240 			vattr.va_size = orig_size;
1241 			/* IO_SYNC is handled implicitely */
1242 			(void)VOP_SETATTR(vp, &vattr, cred);
1243 			uio->uio_offset -= orig_resid - uio->uio_resid;
1244 			uio->uio_resid = orig_resid;
1245 		}
1246 	}
1247 
1248 	return (error);
1249 }
1250 
1251 /*
1252  * Get an nfs cache block.
1253  *
1254  * Allocate a new one if the block isn't currently in the cache
1255  * and return the block marked busy. If the calling process is
1256  * interrupted by a signal for an interruptible mount point, return
1257  * NULL.
1258  *
1259  * The caller must carefully deal with the possible B_INVAL state of
1260  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1261  * indirectly), so synchronous reads can be issued without worrying about
1262  * the B_INVAL state.  We have to be a little more careful when dealing
1263  * with writes (see comments in nfs_write()) when extending a file past
1264  * its EOF.
1265  */
1266 static struct buf *
1267 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1268 {
1269 	struct buf *bp;
1270 	struct mount *mp;
1271 	struct nfsmount *nmp;
1272 
1273 	mp = vp->v_mount;
1274 	nmp = VFSTONFS(mp);
1275 
1276 	if (nmp->nm_flag & NFSMNT_INT) {
1277 		sigset_t oldset;
1278 
1279 		newnfs_set_sigmask(td, &oldset);
1280 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1281 		newnfs_restore_sigmask(td, &oldset);
1282 		while (bp == NULL) {
1283 			if (newnfs_sigintr(nmp, td))
1284 				return (NULL);
1285 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1286 		}
1287 	} else {
1288 		bp = getblk(vp, bn, size, 0, 0, 0);
1289 	}
1290 
1291 	if (vp->v_type == VREG)
1292 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1293 	return (bp);
1294 }
1295 
1296 /*
1297  * Flush and invalidate all dirty buffers. If another process is already
1298  * doing the flush, just wait for completion.
1299  */
1300 int
1301 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1302 {
1303 	struct nfsnode *np = VTONFS(vp);
1304 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1305 	int error = 0, slpflag, slptimeo;
1306 	int old_lock = 0;
1307 
1308 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1309 
1310 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1311 		intrflg = 0;
1312 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1313 		intrflg = 1;
1314 	if (intrflg) {
1315 		slpflag = PCATCH;
1316 		slptimeo = 2 * hz;
1317 	} else {
1318 		slpflag = 0;
1319 		slptimeo = 0;
1320 	}
1321 
1322 	old_lock = ncl_upgrade_vnlock(vp);
1323 	if (vp->v_iflag & VI_DOOMED) {
1324 		/*
1325 		 * Since vgonel() uses the generic vinvalbuf() to flush
1326 		 * dirty buffers and it does not call this function, it
1327 		 * is safe to just return OK when VI_DOOMED is set.
1328 		 */
1329 		ncl_downgrade_vnlock(vp, old_lock);
1330 		return (0);
1331 	}
1332 
1333 	/*
1334 	 * Now, flush as required.
1335 	 */
1336 	if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1337 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1338 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1339 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1340 		/*
1341 		 * If the page clean was interrupted, fail the invalidation.
1342 		 * Not doing so, we run the risk of losing dirty pages in the
1343 		 * vinvalbuf() call below.
1344 		 */
1345 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1346 			goto out;
1347 	}
1348 
1349 	error = vinvalbuf(vp, flags, slpflag, 0);
1350 	while (error) {
1351 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1352 			goto out;
1353 		error = vinvalbuf(vp, flags, 0, slptimeo);
1354 	}
1355 	if (NFSHASPNFS(nmp)) {
1356 		nfscl_layoutcommit(vp, td);
1357 		/*
1358 		 * Invalidate the attribute cache, since writes to a DS
1359 		 * won't update the size attribute.
1360 		 */
1361 		mtx_lock(&np->n_mtx);
1362 		np->n_attrstamp = 0;
1363 	} else
1364 		mtx_lock(&np->n_mtx);
1365 	if (np->n_directio_asyncwr == 0)
1366 		np->n_flag &= ~NMODIFIED;
1367 	mtx_unlock(&np->n_mtx);
1368 out:
1369 	ncl_downgrade_vnlock(vp, old_lock);
1370 	return error;
1371 }
1372 
1373 /*
1374  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1375  * This is mainly to avoid queueing async I/O requests when the nfsiods
1376  * are all hung on a dead server.
1377  *
1378  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1379  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1380  */
1381 int
1382 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1383 {
1384 	int iod;
1385 	int gotiod;
1386 	int slpflag = 0;
1387 	int slptimeo = 0;
1388 	int error, error2;
1389 
1390 	/*
1391 	 * Commits are usually short and sweet so lets save some cpu and
1392 	 * leave the async daemons for more important rpc's (such as reads
1393 	 * and writes).
1394 	 *
1395 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1396 	 * in the directory in order to update attributes. This can deadlock
1397 	 * with another thread that is waiting for async I/O to be done by
1398 	 * an nfsiod thread while holding a lock on one of these vnodes.
1399 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1400 	 * perform Readdirplus RPCs.
1401 	 */
1402 	mtx_lock(&ncl_iod_mutex);
1403 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1404 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1405 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1406 		mtx_unlock(&ncl_iod_mutex);
1407 		return(EIO);
1408 	}
1409 again:
1410 	if (nmp->nm_flag & NFSMNT_INT)
1411 		slpflag = PCATCH;
1412 	gotiod = FALSE;
1413 
1414 	/*
1415 	 * Find a free iod to process this request.
1416 	 */
1417 	for (iod = 0; iod < ncl_numasync; iod++)
1418 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1419 			gotiod = TRUE;
1420 			break;
1421 		}
1422 
1423 	/*
1424 	 * Try to create one if none are free.
1425 	 */
1426 	if (!gotiod)
1427 		ncl_nfsiodnew();
1428 	else {
1429 		/*
1430 		 * Found one, so wake it up and tell it which
1431 		 * mount to process.
1432 		 */
1433 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1434 		    iod, nmp));
1435 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1436 		ncl_iodmount[iod] = nmp;
1437 		nmp->nm_bufqiods++;
1438 		wakeup(&ncl_iodwant[iod]);
1439 	}
1440 
1441 	/*
1442 	 * If none are free, we may already have an iod working on this mount
1443 	 * point.  If so, it will process our request.
1444 	 */
1445 	if (!gotiod) {
1446 		if (nmp->nm_bufqiods > 0) {
1447 			NFS_DPF(ASYNCIO,
1448 				("ncl_asyncio: %d iods are already processing mount %p\n",
1449 				 nmp->nm_bufqiods, nmp));
1450 			gotiod = TRUE;
1451 		}
1452 	}
1453 
1454 	/*
1455 	 * If we have an iod which can process the request, then queue
1456 	 * the buffer.
1457 	 */
1458 	if (gotiod) {
1459 		/*
1460 		 * Ensure that the queue never grows too large.  We still want
1461 		 * to asynchronize so we block rather then return EIO.
1462 		 */
1463 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1464 			NFS_DPF(ASYNCIO,
1465 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1466 			nmp->nm_bufqwant = TRUE;
1467 			error = newnfs_msleep(td, &nmp->nm_bufq,
1468 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1469 			   slptimeo);
1470 			if (error) {
1471 				error2 = newnfs_sigintr(nmp, td);
1472 				if (error2) {
1473 					mtx_unlock(&ncl_iod_mutex);
1474 					return (error2);
1475 				}
1476 				if (slpflag == PCATCH) {
1477 					slpflag = 0;
1478 					slptimeo = 2 * hz;
1479 				}
1480 			}
1481 			/*
1482 			 * We might have lost our iod while sleeping,
1483 			 * so check and loop if nescessary.
1484 			 */
1485 			goto again;
1486 		}
1487 
1488 		/* We might have lost our nfsiod */
1489 		if (nmp->nm_bufqiods == 0) {
1490 			NFS_DPF(ASYNCIO,
1491 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1492 			goto again;
1493 		}
1494 
1495 		if (bp->b_iocmd == BIO_READ) {
1496 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1497 				bp->b_rcred = crhold(cred);
1498 		} else {
1499 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1500 				bp->b_wcred = crhold(cred);
1501 		}
1502 
1503 		if (bp->b_flags & B_REMFREE)
1504 			bremfreef(bp);
1505 		BUF_KERNPROC(bp);
1506 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1507 		nmp->nm_bufqlen++;
1508 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1509 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1510 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1511 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1512 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1513 		}
1514 		mtx_unlock(&ncl_iod_mutex);
1515 		return (0);
1516 	}
1517 
1518 	mtx_unlock(&ncl_iod_mutex);
1519 
1520 	/*
1521 	 * All the iods are busy on other mounts, so return EIO to
1522 	 * force the caller to process the i/o synchronously.
1523 	 */
1524 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1525 	return (EIO);
1526 }
1527 
1528 void
1529 ncl_doio_directwrite(struct buf *bp)
1530 {
1531 	int iomode, must_commit;
1532 	struct uio *uiop = (struct uio *)bp->b_caller1;
1533 	char *iov_base = uiop->uio_iov->iov_base;
1534 
1535 	iomode = NFSWRITE_FILESYNC;
1536 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1537 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1538 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1539 	free(iov_base, M_NFSDIRECTIO);
1540 	free(uiop->uio_iov, M_NFSDIRECTIO);
1541 	free(uiop, M_NFSDIRECTIO);
1542 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1543 		struct nfsnode *np = VTONFS(bp->b_vp);
1544 		mtx_lock(&np->n_mtx);
1545 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1546 			/*
1547 			 * Invalidate the attribute cache, since writes to a DS
1548 			 * won't update the size attribute.
1549 			 */
1550 			np->n_attrstamp = 0;
1551 		}
1552 		np->n_directio_asyncwr--;
1553 		if (np->n_directio_asyncwr == 0) {
1554 			np->n_flag &= ~NMODIFIED;
1555 			if ((np->n_flag & NFSYNCWAIT)) {
1556 				np->n_flag &= ~NFSYNCWAIT;
1557 				wakeup((caddr_t)&np->n_directio_asyncwr);
1558 			}
1559 		}
1560 		mtx_unlock(&np->n_mtx);
1561 	}
1562 	bp->b_vp = NULL;
1563 	relpbuf(bp, &ncl_pbuf_freecnt);
1564 }
1565 
1566 /*
1567  * Do an I/O operation to/from a cache block. This may be called
1568  * synchronously or from an nfsiod.
1569  */
1570 int
1571 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1572     int called_from_strategy)
1573 {
1574 	struct uio *uiop;
1575 	struct nfsnode *np;
1576 	struct nfsmount *nmp;
1577 	int error = 0, iomode, must_commit = 0;
1578 	struct uio uio;
1579 	struct iovec io;
1580 	struct proc *p = td ? td->td_proc : NULL;
1581 	uint8_t	iocmd;
1582 
1583 	np = VTONFS(vp);
1584 	nmp = VFSTONFS(vp->v_mount);
1585 	uiop = &uio;
1586 	uiop->uio_iov = &io;
1587 	uiop->uio_iovcnt = 1;
1588 	uiop->uio_segflg = UIO_SYSSPACE;
1589 	uiop->uio_td = td;
1590 
1591 	/*
1592 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1593 	 * do this here so we do not have to do it in all the code that
1594 	 * calls us.
1595 	 */
1596 	bp->b_flags &= ~B_INVAL;
1597 	bp->b_ioflags &= ~BIO_ERROR;
1598 
1599 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1600 	iocmd = bp->b_iocmd;
1601 	if (iocmd == BIO_READ) {
1602 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1603 	    io.iov_base = bp->b_data;
1604 	    uiop->uio_rw = UIO_READ;
1605 
1606 	    switch (vp->v_type) {
1607 	    case VREG:
1608 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1609 		NFSINCRGLOBAL(newnfsstats.read_bios);
1610 		error = ncl_readrpc(vp, uiop, cr);
1611 
1612 		if (!error) {
1613 		    if (uiop->uio_resid) {
1614 			/*
1615 			 * If we had a short read with no error, we must have
1616 			 * hit a file hole.  We should zero-fill the remainder.
1617 			 * This can also occur if the server hits the file EOF.
1618 			 *
1619 			 * Holes used to be able to occur due to pending
1620 			 * writes, but that is not possible any longer.
1621 			 */
1622 			int nread = bp->b_bcount - uiop->uio_resid;
1623 			ssize_t left = uiop->uio_resid;
1624 
1625 			if (left > 0)
1626 				bzero((char *)bp->b_data + nread, left);
1627 			uiop->uio_resid = 0;
1628 		    }
1629 		}
1630 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1631 		if (p && (vp->v_vflag & VV_TEXT)) {
1632 			mtx_lock(&np->n_mtx);
1633 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1634 				mtx_unlock(&np->n_mtx);
1635 				PROC_LOCK(p);
1636 				killproc(p, "text file modification");
1637 				PROC_UNLOCK(p);
1638 			} else
1639 				mtx_unlock(&np->n_mtx);
1640 		}
1641 		break;
1642 	    case VLNK:
1643 		uiop->uio_offset = (off_t)0;
1644 		NFSINCRGLOBAL(newnfsstats.readlink_bios);
1645 		error = ncl_readlinkrpc(vp, uiop, cr);
1646 		break;
1647 	    case VDIR:
1648 		NFSINCRGLOBAL(newnfsstats.readdir_bios);
1649 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1650 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1651 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1652 			if (error == NFSERR_NOTSUPP)
1653 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1654 		}
1655 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1656 			error = ncl_readdirrpc(vp, uiop, cr, td);
1657 		/*
1658 		 * end-of-directory sets B_INVAL but does not generate an
1659 		 * error.
1660 		 */
1661 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1662 			bp->b_flags |= B_INVAL;
1663 		break;
1664 	    default:
1665 		ncl_printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1666 		break;
1667 	    };
1668 	    if (error) {
1669 		bp->b_ioflags |= BIO_ERROR;
1670 		bp->b_error = error;
1671 	    }
1672 	} else {
1673 	    /*
1674 	     * If we only need to commit, try to commit
1675 	     */
1676 	    if (bp->b_flags & B_NEEDCOMMIT) {
1677 		    int retv;
1678 		    off_t off;
1679 
1680 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1681 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1682 			bp->b_wcred, td);
1683 		    if (retv == 0) {
1684 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1685 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1686 			    bp->b_resid = 0;
1687 			    bufdone(bp);
1688 			    return (0);
1689 		    }
1690 		    if (retv == NFSERR_STALEWRITEVERF) {
1691 			    ncl_clearcommit(vp->v_mount);
1692 		    }
1693 	    }
1694 
1695 	    /*
1696 	     * Setup for actual write
1697 	     */
1698 	    mtx_lock(&np->n_mtx);
1699 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1700 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1701 	    mtx_unlock(&np->n_mtx);
1702 
1703 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1704 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1705 		    - bp->b_dirtyoff;
1706 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1707 		    + bp->b_dirtyoff;
1708 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1709 		uiop->uio_rw = UIO_WRITE;
1710 		NFSINCRGLOBAL(newnfsstats.write_bios);
1711 
1712 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1713 		    iomode = NFSWRITE_UNSTABLE;
1714 		else
1715 		    iomode = NFSWRITE_FILESYNC;
1716 
1717 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1718 		    called_from_strategy);
1719 
1720 		/*
1721 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1722 		 * to cluster the buffers needing commit.  This will allow
1723 		 * the system to submit a single commit rpc for the whole
1724 		 * cluster.  We can do this even if the buffer is not 100%
1725 		 * dirty (relative to the NFS blocksize), so we optimize the
1726 		 * append-to-file-case.
1727 		 *
1728 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1729 		 * cleared because write clustering only works for commit
1730 		 * rpc's, not for the data portion of the write).
1731 		 */
1732 
1733 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1734 		    bp->b_flags |= B_NEEDCOMMIT;
1735 		    if (bp->b_dirtyoff == 0
1736 			&& bp->b_dirtyend == bp->b_bcount)
1737 			bp->b_flags |= B_CLUSTEROK;
1738 		} else {
1739 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1740 		}
1741 
1742 		/*
1743 		 * For an interrupted write, the buffer is still valid
1744 		 * and the write hasn't been pushed to the server yet,
1745 		 * so we can't set BIO_ERROR and report the interruption
1746 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1747 		 * is not relevant, so the rpc attempt is essentially
1748 		 * a noop.  For the case of a V3 write rpc not being
1749 		 * committed to stable storage, the block is still
1750 		 * dirty and requires either a commit rpc or another
1751 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1752 		 * the block is reused. This is indicated by setting
1753 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1754 		 *
1755 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1756 		 * write error and is handled as above, except that
1757 		 * B_EINTR isn't set. One cause of this is a stale stateid
1758 		 * error for the RPC that indicates recovery is required,
1759 		 * when called with called_from_strategy != 0.
1760 		 *
1761 		 * If the buffer is marked B_PAGING, it does not reside on
1762 		 * the vp's paging queues so we cannot call bdirty().  The
1763 		 * bp in this case is not an NFS cache block so we should
1764 		 * be safe. XXX
1765 		 *
1766 		 * The logic below breaks up errors into recoverable and
1767 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1768 		 * and keep the buffer around for potential write retries.
1769 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1770 		 * and save the error in the nfsnode. This is less than ideal
1771 		 * but necessary. Keeping such buffers around could potentially
1772 		 * cause buffer exhaustion eventually (they can never be written
1773 		 * out, so will get constantly be re-dirtied). It also causes
1774 		 * all sorts of vfs panics. For non-recoverable write errors,
1775 		 * also invalidate the attrcache, so we'll be forced to go over
1776 		 * the wire for this object, returning an error to user on next
1777 		 * call (most of the time).
1778 		 */
1779 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1780 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1781 			int s;
1782 
1783 			s = splbio();
1784 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1785 			if ((bp->b_flags & B_PAGING) == 0) {
1786 			    bdirty(bp);
1787 			    bp->b_flags &= ~B_DONE;
1788 			}
1789 			if ((error == EINTR || error == ETIMEDOUT) &&
1790 			    (bp->b_flags & B_ASYNC) == 0)
1791 			    bp->b_flags |= B_EINTR;
1792 			splx(s);
1793 		} else {
1794 		    if (error) {
1795 			bp->b_ioflags |= BIO_ERROR;
1796 			bp->b_flags |= B_INVAL;
1797 			bp->b_error = np->n_error = error;
1798 			mtx_lock(&np->n_mtx);
1799 			np->n_flag |= NWRITEERR;
1800 			np->n_attrstamp = 0;
1801 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1802 			mtx_unlock(&np->n_mtx);
1803 		    }
1804 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1805 		}
1806 	    } else {
1807 		bp->b_resid = 0;
1808 		bufdone(bp);
1809 		return (0);
1810 	    }
1811 	}
1812 	bp->b_resid = uiop->uio_resid;
1813 	if (must_commit)
1814 	    ncl_clearcommit(vp->v_mount);
1815 	bufdone(bp);
1816 	return (error);
1817 }
1818 
1819 /*
1820  * Used to aid in handling ftruncate() operations on the NFS client side.
1821  * Truncation creates a number of special problems for NFS.  We have to
1822  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1823  * we have to properly handle VM pages or (potentially dirty) buffers
1824  * that straddle the truncation point.
1825  */
1826 
1827 int
1828 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1829 {
1830 	struct nfsnode *np = VTONFS(vp);
1831 	u_quad_t tsize;
1832 	int biosize = vp->v_bufobj.bo_bsize;
1833 	int error = 0;
1834 
1835 	mtx_lock(&np->n_mtx);
1836 	tsize = np->n_size;
1837 	np->n_size = nsize;
1838 	mtx_unlock(&np->n_mtx);
1839 
1840 	if (nsize < tsize) {
1841 		struct buf *bp;
1842 		daddr_t lbn;
1843 		int bufsize;
1844 
1845 		/*
1846 		 * vtruncbuf() doesn't get the buffer overlapping the
1847 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1848 		 * buffer that now needs to be truncated.
1849 		 */
1850 		error = vtruncbuf(vp, cred, nsize, biosize);
1851 		lbn = nsize / biosize;
1852 		bufsize = nsize - (lbn * biosize);
1853 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1854 		if (!bp)
1855 			return EINTR;
1856 		if (bp->b_dirtyoff > bp->b_bcount)
1857 			bp->b_dirtyoff = bp->b_bcount;
1858 		if (bp->b_dirtyend > bp->b_bcount)
1859 			bp->b_dirtyend = bp->b_bcount;
1860 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1861 		brelse(bp);
1862 	} else {
1863 		vnode_pager_setsize(vp, nsize);
1864 	}
1865 	return(error);
1866 }
1867 
1868