xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision efe3b0de1438e7a8473d92f2be57072394559e3c)
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 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 	PCPU_INC(cnt.v_vnodein);
188 	PCPU_ADD(cnt.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 	vm_offset_t kva;
270 	struct buf *bp;
271 	int iomode, must_commit, i, error, npages, count;
272 	off_t offset;
273 	int *rtvals;
274 	struct vnode *vp;
275 	struct thread *td;
276 	struct ucred *cred;
277 	struct nfsmount *nmp;
278 	struct nfsnode *np;
279 	vm_page_t *pages;
280 
281 	vp = ap->a_vp;
282 	np = VTONFS(vp);
283 	td = curthread;				/* XXX */
284 	/* Set the cred to n_writecred for the write rpcs. */
285 	if (np->n_writecred != NULL)
286 		cred = crhold(np->n_writecred);
287 	else
288 		cred = crhold(curthread->td_ucred);	/* XXX */
289 	nmp = VFSTONFS(vp->v_mount);
290 	pages = ap->a_m;
291 	count = ap->a_count;
292 	rtvals = ap->a_rtvals;
293 	npages = btoc(count);
294 	offset = IDX_TO_OFF(pages[0]->pindex);
295 
296 	mtx_lock(&nmp->nm_mtx);
297 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
298 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
299 		mtx_unlock(&nmp->nm_mtx);
300 		(void)ncl_fsinfo(nmp, vp, cred, td);
301 	} else
302 		mtx_unlock(&nmp->nm_mtx);
303 
304 	mtx_lock(&np->n_mtx);
305 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
306 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
307 		mtx_unlock(&np->n_mtx);
308 		printf("ncl_putpages: called on noncache-able vnode\n");
309 		mtx_lock(&np->n_mtx);
310 	}
311 
312 	for (i = 0; i < npages; i++)
313 		rtvals[i] = VM_PAGER_ERROR;
314 
315 	/*
316 	 * When putting pages, do not extend file past EOF.
317 	 */
318 	if (offset + count > np->n_size) {
319 		count = np->n_size - offset;
320 		if (count < 0)
321 			count = 0;
322 	}
323 	mtx_unlock(&np->n_mtx);
324 
325 	/*
326 	 * We use only the kva address for the buffer, but this is extremely
327 	 * convenient and fast.
328 	 */
329 	bp = getpbuf(&ncl_pbuf_freecnt);
330 
331 	kva = (vm_offset_t) bp->b_data;
332 	pmap_qenter(kva, pages, npages);
333 	PCPU_INC(cnt.v_vnodeout);
334 	PCPU_ADD(cnt.v_vnodepgsout, count);
335 
336 	iov.iov_base = (caddr_t) kva;
337 	iov.iov_len = count;
338 	uio.uio_iov = &iov;
339 	uio.uio_iovcnt = 1;
340 	uio.uio_offset = offset;
341 	uio.uio_resid = count;
342 	uio.uio_segflg = UIO_SYSSPACE;
343 	uio.uio_rw = UIO_WRITE;
344 	uio.uio_td = td;
345 
346 	if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
347 	    iomode = NFSWRITE_UNSTABLE;
348 	else
349 	    iomode = NFSWRITE_FILESYNC;
350 
351 	error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
352 	crfree(cred);
353 
354 	pmap_qremove(kva, npages);
355 	relpbuf(bp, &ncl_pbuf_freecnt);
356 
357 	if (error == 0 || !nfs_keep_dirty_on_error) {
358 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
359 		if (must_commit)
360 			ncl_clearcommit(vp->v_mount);
361 	}
362 	return rtvals[0];
363 }
364 
365 /*
366  * For nfs, cache consistency can only be maintained approximately.
367  * Although RFC1094 does not specify the criteria, the following is
368  * believed to be compatible with the reference port.
369  * For nfs:
370  * If the file's modify time on the server has changed since the
371  * last read rpc or you have written to the file,
372  * you may have lost data cache consistency with the
373  * server, so flush all of the file's data out of the cache.
374  * Then force a getattr rpc to ensure that you have up to date
375  * attributes.
376  * NB: This implies that cache data can be read when up to
377  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
378  * attributes this could be forced by setting n_attrstamp to 0 before
379  * the VOP_GETATTR() call.
380  */
381 static inline int
382 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
383 {
384 	int error = 0;
385 	struct vattr vattr;
386 	struct nfsnode *np = VTONFS(vp);
387 	int old_lock;
388 
389 	/*
390 	 * Grab the exclusive lock before checking whether the cache is
391 	 * consistent.
392 	 * XXX - We can make this cheaper later (by acquiring cheaper locks).
393 	 * But for now, this suffices.
394 	 */
395 	old_lock = ncl_upgrade_vnlock(vp);
396 	if (vp->v_iflag & VI_DOOMED) {
397 		ncl_downgrade_vnlock(vp, old_lock);
398 		return (EBADF);
399 	}
400 
401 	mtx_lock(&np->n_mtx);
402 	if (np->n_flag & NMODIFIED) {
403 		mtx_unlock(&np->n_mtx);
404 		if (vp->v_type != VREG) {
405 			if (vp->v_type != VDIR)
406 				panic("nfs: bioread, not dir");
407 			ncl_invaldir(vp);
408 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
409 			if (error)
410 				goto out;
411 		}
412 		np->n_attrstamp = 0;
413 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
414 		error = VOP_GETATTR(vp, &vattr, cred);
415 		if (error)
416 			goto out;
417 		mtx_lock(&np->n_mtx);
418 		np->n_mtime = vattr.va_mtime;
419 		mtx_unlock(&np->n_mtx);
420 	} else {
421 		mtx_unlock(&np->n_mtx);
422 		error = VOP_GETATTR(vp, &vattr, cred);
423 		if (error)
424 			return (error);
425 		mtx_lock(&np->n_mtx);
426 		if ((np->n_flag & NSIZECHANGED)
427 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
428 			mtx_unlock(&np->n_mtx);
429 			if (vp->v_type == VDIR)
430 				ncl_invaldir(vp);
431 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
432 			if (error)
433 				goto out;
434 			mtx_lock(&np->n_mtx);
435 			np->n_mtime = vattr.va_mtime;
436 			np->n_flag &= ~NSIZECHANGED;
437 		}
438 		mtx_unlock(&np->n_mtx);
439 	}
440 out:
441 	ncl_downgrade_vnlock(vp, old_lock);
442 	return error;
443 }
444 
445 /*
446  * Vnode op for read using bio
447  */
448 int
449 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
450 {
451 	struct nfsnode *np = VTONFS(vp);
452 	int biosize, i;
453 	struct buf *bp, *rabp;
454 	struct thread *td;
455 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
456 	daddr_t lbn, rabn;
457 	int bcount;
458 	int seqcount;
459 	int nra, error = 0, n = 0, on = 0;
460 	off_t tmp_off;
461 
462 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
463 	if (uio->uio_resid == 0)
464 		return (0);
465 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
466 		return (EINVAL);
467 	td = uio->uio_td;
468 
469 	mtx_lock(&nmp->nm_mtx);
470 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
471 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
472 		mtx_unlock(&nmp->nm_mtx);
473 		(void)ncl_fsinfo(nmp, vp, cred, td);
474 		mtx_lock(&nmp->nm_mtx);
475 	}
476 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
477 		(void) newnfs_iosize(nmp);
478 
479 	tmp_off = uio->uio_offset + uio->uio_resid;
480 	if (vp->v_type != VDIR &&
481 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
482 		mtx_unlock(&nmp->nm_mtx);
483 		return (EFBIG);
484 	}
485 	mtx_unlock(&nmp->nm_mtx);
486 
487 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
488 		/* No caching/ no readaheads. Just read data into the user buffer */
489 		return ncl_readrpc(vp, uio, cred);
490 
491 	biosize = vp->v_bufobj.bo_bsize;
492 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
493 
494 	error = nfs_bioread_check_cons(vp, td, cred);
495 	if (error)
496 		return error;
497 
498 	do {
499 	    u_quad_t nsize;
500 
501 	    mtx_lock(&np->n_mtx);
502 	    nsize = np->n_size;
503 	    mtx_unlock(&np->n_mtx);
504 
505 	    switch (vp->v_type) {
506 	    case VREG:
507 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
508 		lbn = uio->uio_offset / biosize;
509 		on = uio->uio_offset - (lbn * biosize);
510 
511 		/*
512 		 * Start the read ahead(s), as required.
513 		 */
514 		if (nmp->nm_readahead > 0) {
515 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
516 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
517 			rabn = lbn + 1 + nra;
518 			if (incore(&vp->v_bufobj, rabn) == NULL) {
519 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
520 			    if (!rabp) {
521 				error = newnfs_sigintr(nmp, td);
522 				return (error ? error : EINTR);
523 			    }
524 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
525 				rabp->b_flags |= B_ASYNC;
526 				rabp->b_iocmd = BIO_READ;
527 				vfs_busy_pages(rabp, 0);
528 				if (ncl_asyncio(nmp, rabp, cred, td)) {
529 				    rabp->b_flags |= B_INVAL;
530 				    rabp->b_ioflags |= BIO_ERROR;
531 				    vfs_unbusy_pages(rabp);
532 				    brelse(rabp);
533 				    break;
534 				}
535 			    } else {
536 				brelse(rabp);
537 			    }
538 			}
539 		    }
540 		}
541 
542 		/* Note that bcount is *not* DEV_BSIZE aligned. */
543 		bcount = biosize;
544 		if ((off_t)lbn * biosize >= nsize) {
545 			bcount = 0;
546 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
547 			bcount = nsize - (off_t)lbn * biosize;
548 		}
549 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
550 
551 		if (!bp) {
552 			error = newnfs_sigintr(nmp, td);
553 			return (error ? error : EINTR);
554 		}
555 
556 		/*
557 		 * If B_CACHE is not set, we must issue the read.  If this
558 		 * fails, we return an error.
559 		 */
560 
561 		if ((bp->b_flags & B_CACHE) == 0) {
562 		    bp->b_iocmd = BIO_READ;
563 		    vfs_busy_pages(bp, 0);
564 		    error = ncl_doio(vp, bp, cred, td, 0);
565 		    if (error) {
566 			brelse(bp);
567 			return (error);
568 		    }
569 		}
570 
571 		/*
572 		 * on is the offset into the current bp.  Figure out how many
573 		 * bytes we can copy out of the bp.  Note that bcount is
574 		 * NOT DEV_BSIZE aligned.
575 		 *
576 		 * Then figure out how many bytes we can copy into the uio.
577 		 */
578 
579 		n = 0;
580 		if (on < bcount)
581 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
582 		break;
583 	    case VLNK:
584 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
585 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, 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 			bp->b_ioflags |= BIO_ERROR;
596 			brelse(bp);
597 			return (error);
598 		    }
599 		}
600 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
601 		on = 0;
602 		break;
603 	    case VDIR:
604 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
605 		if (np->n_direofoffset
606 		    && uio->uio_offset >= np->n_direofoffset) {
607 		    return (0);
608 		}
609 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
610 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
611 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
612 		if (!bp) {
613 		    error = newnfs_sigintr(nmp, td);
614 		    return (error ? error : EINTR);
615 		}
616 		if ((bp->b_flags & B_CACHE) == 0) {
617 		    bp->b_iocmd = BIO_READ;
618 		    vfs_busy_pages(bp, 0);
619 		    error = ncl_doio(vp, bp, cred, td, 0);
620 		    if (error) {
621 			    brelse(bp);
622 		    }
623 		    while (error == NFSERR_BAD_COOKIE) {
624 			ncl_invaldir(vp);
625 			error = ncl_vinvalbuf(vp, 0, td, 1);
626 			/*
627 			 * Yuck! The directory has been modified on the
628 			 * server. The only way to get the block is by
629 			 * reading from the beginning to get all the
630 			 * offset cookies.
631 			 *
632 			 * Leave the last bp intact unless there is an error.
633 			 * Loop back up to the while if the error is another
634 			 * NFSERR_BAD_COOKIE (double yuch!).
635 			 */
636 			for (i = 0; i <= lbn && !error; i++) {
637 			    if (np->n_direofoffset
638 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
639 				    return (0);
640 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
641 			    if (!bp) {
642 				error = newnfs_sigintr(nmp, td);
643 				return (error ? error : EINTR);
644 			    }
645 			    if ((bp->b_flags & B_CACHE) == 0) {
646 				    bp->b_iocmd = BIO_READ;
647 				    vfs_busy_pages(bp, 0);
648 				    error = ncl_doio(vp, bp, cred, td, 0);
649 				    /*
650 				     * no error + B_INVAL == directory EOF,
651 				     * use the block.
652 				     */
653 				    if (error == 0 && (bp->b_flags & B_INVAL))
654 					    break;
655 			    }
656 			    /*
657 			     * An error will throw away the block and the
658 			     * for loop will break out.  If no error and this
659 			     * is not the block we want, we throw away the
660 			     * block and go for the next one via the for loop.
661 			     */
662 			    if (error || i < lbn)
663 				    brelse(bp);
664 			}
665 		    }
666 		    /*
667 		     * The above while is repeated if we hit another cookie
668 		     * error.  If we hit an error and it wasn't a cookie error,
669 		     * we give up.
670 		     */
671 		    if (error)
672 			    return (error);
673 		}
674 
675 		/*
676 		 * If not eof and read aheads are enabled, start one.
677 		 * (You need the current block first, so that you have the
678 		 *  directory offset cookie of the next block.)
679 		 */
680 		if (nmp->nm_readahead > 0 &&
681 		    (bp->b_flags & B_INVAL) == 0 &&
682 		    (np->n_direofoffset == 0 ||
683 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
684 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
685 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
686 			if (rabp) {
687 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
688 				rabp->b_flags |= B_ASYNC;
689 				rabp->b_iocmd = BIO_READ;
690 				vfs_busy_pages(rabp, 0);
691 				if (ncl_asyncio(nmp, rabp, cred, td)) {
692 				    rabp->b_flags |= B_INVAL;
693 				    rabp->b_ioflags |= BIO_ERROR;
694 				    vfs_unbusy_pages(rabp);
695 				    brelse(rabp);
696 				}
697 			    } else {
698 				brelse(rabp);
699 			    }
700 			}
701 		}
702 		/*
703 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
704 		 * chopped for the EOF condition, we cannot tell how large
705 		 * NFS directories are going to be until we hit EOF.  So
706 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
707 		 * it just so happens that b_resid will effectively chop it
708 		 * to EOF.  *BUT* this information is lost if the buffer goes
709 		 * away and is reconstituted into a B_CACHE state ( due to
710 		 * being VMIO ) later.  So we keep track of the directory eof
711 		 * in np->n_direofoffset and chop it off as an extra step
712 		 * right here.
713 		 */
714 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
715 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
716 			n = np->n_direofoffset - uio->uio_offset;
717 		break;
718 	    default:
719 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
720 		bp = NULL;
721 		break;
722 	    }
723 
724 	    if (n > 0) {
725 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
726 	    }
727 	    if (vp->v_type == VLNK)
728 		n = 0;
729 	    if (bp != NULL)
730 		brelse(bp);
731 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
732 	return (error);
733 }
734 
735 /*
736  * The NFS write path cannot handle iovecs with len > 1. So we need to
737  * break up iovecs accordingly (restricting them to wsize).
738  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
739  * For the ASYNC case, 2 copies are needed. The first a copy from the
740  * user buffer to a staging buffer and then a second copy from the staging
741  * buffer to mbufs. This can be optimized by copying from the user buffer
742  * directly into mbufs and passing the chain down, but that requires a
743  * fair amount of re-working of the relevant codepaths (and can be done
744  * later).
745  */
746 static int
747 nfs_directio_write(vp, uiop, cred, ioflag)
748 	struct vnode *vp;
749 	struct uio *uiop;
750 	struct ucred *cred;
751 	int ioflag;
752 {
753 	int error;
754 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
755 	struct thread *td = uiop->uio_td;
756 	int size;
757 	int wsize;
758 
759 	mtx_lock(&nmp->nm_mtx);
760 	wsize = nmp->nm_wsize;
761 	mtx_unlock(&nmp->nm_mtx);
762 	if (ioflag & IO_SYNC) {
763 		int iomode, must_commit;
764 		struct uio uio;
765 		struct iovec iov;
766 do_sync:
767 		while (uiop->uio_resid > 0) {
768 			size = MIN(uiop->uio_resid, wsize);
769 			size = MIN(uiop->uio_iov->iov_len, size);
770 			iov.iov_base = uiop->uio_iov->iov_base;
771 			iov.iov_len = size;
772 			uio.uio_iov = &iov;
773 			uio.uio_iovcnt = 1;
774 			uio.uio_offset = uiop->uio_offset;
775 			uio.uio_resid = size;
776 			uio.uio_segflg = UIO_USERSPACE;
777 			uio.uio_rw = UIO_WRITE;
778 			uio.uio_td = td;
779 			iomode = NFSWRITE_FILESYNC;
780 			error = ncl_writerpc(vp, &uio, cred, &iomode,
781 			    &must_commit, 0);
782 			KASSERT((must_commit == 0),
783 				("ncl_directio_write: Did not commit write"));
784 			if (error)
785 				return (error);
786 			uiop->uio_offset += size;
787 			uiop->uio_resid -= size;
788 			if (uiop->uio_iov->iov_len <= size) {
789 				uiop->uio_iovcnt--;
790 				uiop->uio_iov++;
791 			} else {
792 				uiop->uio_iov->iov_base =
793 					(char *)uiop->uio_iov->iov_base + size;
794 				uiop->uio_iov->iov_len -= size;
795 			}
796 		}
797 	} else {
798 		struct uio *t_uio;
799 		struct iovec *t_iov;
800 		struct buf *bp;
801 
802 		/*
803 		 * Break up the write into blocksize chunks and hand these
804 		 * over to nfsiod's for write back.
805 		 * Unfortunately, this incurs a copy of the data. Since
806 		 * the user could modify the buffer before the write is
807 		 * initiated.
808 		 *
809 		 * The obvious optimization here is that one of the 2 copies
810 		 * in the async write path can be eliminated by copying the
811 		 * data here directly into mbufs and passing the mbuf chain
812 		 * down. But that will require a fair amount of re-working
813 		 * of the code and can be done if there's enough interest
814 		 * in NFS directio access.
815 		 */
816 		while (uiop->uio_resid > 0) {
817 			size = MIN(uiop->uio_resid, wsize);
818 			size = MIN(uiop->uio_iov->iov_len, size);
819 			bp = getpbuf(&ncl_pbuf_freecnt);
820 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
821 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
822 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
823 			t_iov->iov_len = size;
824 			t_uio->uio_iov = t_iov;
825 			t_uio->uio_iovcnt = 1;
826 			t_uio->uio_offset = uiop->uio_offset;
827 			t_uio->uio_resid = size;
828 			t_uio->uio_segflg = UIO_SYSSPACE;
829 			t_uio->uio_rw = UIO_WRITE;
830 			t_uio->uio_td = td;
831 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
832 			    uiop->uio_segflg == UIO_SYSSPACE,
833 			    ("nfs_directio_write: Bad uio_segflg"));
834 			if (uiop->uio_segflg == UIO_USERSPACE) {
835 				error = copyin(uiop->uio_iov->iov_base,
836 				    t_iov->iov_base, size);
837 				if (error != 0)
838 					goto err_free;
839 			} else
840 				/*
841 				 * UIO_SYSSPACE may never happen, but handle
842 				 * it just in case it does.
843 				 */
844 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
845 				    size);
846 			bp->b_flags |= B_DIRECT;
847 			bp->b_iocmd = BIO_WRITE;
848 			if (cred != NOCRED) {
849 				crhold(cred);
850 				bp->b_wcred = cred;
851 			} else
852 				bp->b_wcred = NOCRED;
853 			bp->b_caller1 = (void *)t_uio;
854 			bp->b_vp = vp;
855 			error = ncl_asyncio(nmp, bp, NOCRED, td);
856 err_free:
857 			if (error) {
858 				free(t_iov->iov_base, M_NFSDIRECTIO);
859 				free(t_iov, M_NFSDIRECTIO);
860 				free(t_uio, M_NFSDIRECTIO);
861 				bp->b_vp = NULL;
862 				relpbuf(bp, &ncl_pbuf_freecnt);
863 				if (error == EINTR)
864 					return (error);
865 				goto do_sync;
866 			}
867 			uiop->uio_offset += size;
868 			uiop->uio_resid -= size;
869 			if (uiop->uio_iov->iov_len <= size) {
870 				uiop->uio_iovcnt--;
871 				uiop->uio_iov++;
872 			} else {
873 				uiop->uio_iov->iov_base =
874 					(char *)uiop->uio_iov->iov_base + size;
875 				uiop->uio_iov->iov_len -= size;
876 			}
877 		}
878 	}
879 	return (0);
880 }
881 
882 /*
883  * Vnode op for write using bio
884  */
885 int
886 ncl_write(struct vop_write_args *ap)
887 {
888 	int biosize;
889 	struct uio *uio = ap->a_uio;
890 	struct thread *td = uio->uio_td;
891 	struct vnode *vp = ap->a_vp;
892 	struct nfsnode *np = VTONFS(vp);
893 	struct ucred *cred = ap->a_cred;
894 	int ioflag = ap->a_ioflag;
895 	struct buf *bp;
896 	struct vattr vattr;
897 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
898 	daddr_t lbn;
899 	int bcount, noncontig_write, obcount;
900 	int bp_cached, n, on, error = 0, error1, wouldcommit;
901 	size_t orig_resid, local_resid;
902 	off_t orig_size, tmp_off;
903 
904 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
905 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
906 	    ("ncl_write proc"));
907 	if (vp->v_type != VREG)
908 		return (EIO);
909 	mtx_lock(&np->n_mtx);
910 	if (np->n_flag & NWRITEERR) {
911 		np->n_flag &= ~NWRITEERR;
912 		mtx_unlock(&np->n_mtx);
913 		return (np->n_error);
914 	} else
915 		mtx_unlock(&np->n_mtx);
916 	mtx_lock(&nmp->nm_mtx);
917 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
918 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
919 		mtx_unlock(&nmp->nm_mtx);
920 		(void)ncl_fsinfo(nmp, vp, cred, td);
921 		mtx_lock(&nmp->nm_mtx);
922 	}
923 	if (nmp->nm_wsize == 0)
924 		(void) newnfs_iosize(nmp);
925 	mtx_unlock(&nmp->nm_mtx);
926 
927 	/*
928 	 * Synchronously flush pending buffers if we are in synchronous
929 	 * mode or if we are appending.
930 	 */
931 	if (ioflag & (IO_APPEND | IO_SYNC)) {
932 		mtx_lock(&np->n_mtx);
933 		if (np->n_flag & NMODIFIED) {
934 			mtx_unlock(&np->n_mtx);
935 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
936 			/*
937 			 * Require non-blocking, synchronous writes to
938 			 * dirty files to inform the program it needs
939 			 * to fsync(2) explicitly.
940 			 */
941 			if (ioflag & IO_NDELAY)
942 				return (EAGAIN);
943 #endif
944 			np->n_attrstamp = 0;
945 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
946 			error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
947 			if (error)
948 				return (error);
949 		} else
950 			mtx_unlock(&np->n_mtx);
951 	}
952 
953 	orig_resid = uio->uio_resid;
954 	mtx_lock(&np->n_mtx);
955 	orig_size = np->n_size;
956 	mtx_unlock(&np->n_mtx);
957 
958 	/*
959 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
960 	 * get the append lock.
961 	 */
962 	if (ioflag & IO_APPEND) {
963 		np->n_attrstamp = 0;
964 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
965 		error = VOP_GETATTR(vp, &vattr, cred);
966 		if (error)
967 			return (error);
968 		mtx_lock(&np->n_mtx);
969 		uio->uio_offset = np->n_size;
970 		mtx_unlock(&np->n_mtx);
971 	}
972 
973 	if (uio->uio_offset < 0)
974 		return (EINVAL);
975 	tmp_off = uio->uio_offset + uio->uio_resid;
976 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
977 		return (EFBIG);
978 	if (uio->uio_resid == 0)
979 		return (0);
980 
981 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
982 		return nfs_directio_write(vp, uio, cred, ioflag);
983 
984 	/*
985 	 * Maybe this should be above the vnode op call, but so long as
986 	 * file servers have no limits, i don't think it matters
987 	 */
988 	if (vn_rlimit_fsize(vp, uio, td))
989 		return (EFBIG);
990 
991 	biosize = vp->v_bufobj.bo_bsize;
992 	/*
993 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
994 	 * would exceed the local maximum per-file write commit size when
995 	 * combined with those, we must decide whether to flush,
996 	 * go synchronous, or return error.  We don't bother checking
997 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
998 	 * no point optimizing for something that really won't ever happen.
999 	 */
1000 	wouldcommit = 0;
1001 	if (!(ioflag & IO_SYNC)) {
1002 		int nflag;
1003 
1004 		mtx_lock(&np->n_mtx);
1005 		nflag = np->n_flag;
1006 		mtx_unlock(&np->n_mtx);
1007 		if (nflag & NMODIFIED) {
1008 			BO_LOCK(&vp->v_bufobj);
1009 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1010 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1011 				    b_bobufs) {
1012 					if (bp->b_flags & B_NEEDCOMMIT)
1013 						wouldcommit += bp->b_bcount;
1014 				}
1015 			}
1016 			BO_UNLOCK(&vp->v_bufobj);
1017 		}
1018 	}
1019 
1020 	do {
1021 		if (!(ioflag & IO_SYNC)) {
1022 			wouldcommit += biosize;
1023 			if (wouldcommit > nmp->nm_wcommitsize) {
1024 				np->n_attrstamp = 0;
1025 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1026 				error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
1027 				if (error)
1028 					return (error);
1029 				wouldcommit = biosize;
1030 			}
1031 		}
1032 
1033 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1034 		lbn = uio->uio_offset / biosize;
1035 		on = uio->uio_offset - (lbn * biosize);
1036 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1037 again:
1038 		/*
1039 		 * Handle direct append and file extension cases, calculate
1040 		 * unaligned buffer size.
1041 		 */
1042 		mtx_lock(&np->n_mtx);
1043 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1044 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1045 			noncontig_write = 1;
1046 		else
1047 			noncontig_write = 0;
1048 		if ((uio->uio_offset == np->n_size ||
1049 		    (noncontig_write != 0 &&
1050 		    lbn == (np->n_size / biosize) &&
1051 		    uio->uio_offset + n > np->n_size)) && n) {
1052 			mtx_unlock(&np->n_mtx);
1053 			/*
1054 			 * Get the buffer (in its pre-append state to maintain
1055 			 * B_CACHE if it was previously set).  Resize the
1056 			 * nfsnode after we have locked the buffer to prevent
1057 			 * readers from reading garbage.
1058 			 */
1059 			obcount = np->n_size - (lbn * biosize);
1060 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1061 
1062 			if (bp != NULL) {
1063 				long save;
1064 
1065 				mtx_lock(&np->n_mtx);
1066 				np->n_size = uio->uio_offset + n;
1067 				np->n_flag |= NMODIFIED;
1068 				vnode_pager_setsize(vp, np->n_size);
1069 				mtx_unlock(&np->n_mtx);
1070 
1071 				save = bp->b_flags & B_CACHE;
1072 				bcount = on + n;
1073 				allocbuf(bp, bcount);
1074 				bp->b_flags |= save;
1075 				if (noncontig_write != 0 && on > obcount)
1076 					vfs_bio_bzero_buf(bp, obcount, on -
1077 					    obcount);
1078 			}
1079 		} else {
1080 			/*
1081 			 * Obtain the locked cache block first, and then
1082 			 * adjust the file's size as appropriate.
1083 			 */
1084 			bcount = on + n;
1085 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1086 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1087 					bcount = biosize;
1088 				else
1089 					bcount = np->n_size - (off_t)lbn * biosize;
1090 			}
1091 			mtx_unlock(&np->n_mtx);
1092 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1093 			mtx_lock(&np->n_mtx);
1094 			if (uio->uio_offset + n > np->n_size) {
1095 				np->n_size = uio->uio_offset + n;
1096 				np->n_flag |= NMODIFIED;
1097 				vnode_pager_setsize(vp, np->n_size);
1098 			}
1099 			mtx_unlock(&np->n_mtx);
1100 		}
1101 
1102 		if (!bp) {
1103 			error = newnfs_sigintr(nmp, td);
1104 			if (!error)
1105 				error = EINTR;
1106 			break;
1107 		}
1108 
1109 		/*
1110 		 * Issue a READ if B_CACHE is not set.  In special-append
1111 		 * mode, B_CACHE is based on the buffer prior to the write
1112 		 * op and is typically set, avoiding the read.  If a read
1113 		 * is required in special append mode, the server will
1114 		 * probably send us a short-read since we extended the file
1115 		 * on our end, resulting in b_resid == 0 and, thusly,
1116 		 * B_CACHE getting set.
1117 		 *
1118 		 * We can also avoid issuing the read if the write covers
1119 		 * the entire buffer.  We have to make sure the buffer state
1120 		 * is reasonable in this case since we will not be initiating
1121 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1122 		 * more information.
1123 		 *
1124 		 * B_CACHE may also be set due to the buffer being cached
1125 		 * normally.
1126 		 */
1127 
1128 		bp_cached = 1;
1129 		if (on == 0 && n == bcount) {
1130 			if ((bp->b_flags & B_CACHE) == 0)
1131 				bp_cached = 0;
1132 			bp->b_flags |= B_CACHE;
1133 			bp->b_flags &= ~B_INVAL;
1134 			bp->b_ioflags &= ~BIO_ERROR;
1135 		}
1136 
1137 		if ((bp->b_flags & B_CACHE) == 0) {
1138 			bp->b_iocmd = BIO_READ;
1139 			vfs_busy_pages(bp, 0);
1140 			error = ncl_doio(vp, bp, cred, td, 0);
1141 			if (error) {
1142 				brelse(bp);
1143 				break;
1144 			}
1145 		}
1146 		if (bp->b_wcred == NOCRED)
1147 			bp->b_wcred = crhold(cred);
1148 		mtx_lock(&np->n_mtx);
1149 		np->n_flag |= NMODIFIED;
1150 		mtx_unlock(&np->n_mtx);
1151 
1152 		/*
1153 		 * If dirtyend exceeds file size, chop it down.  This should
1154 		 * not normally occur but there is an append race where it
1155 		 * might occur XXX, so we log it.
1156 		 *
1157 		 * If the chopping creates a reverse-indexed or degenerate
1158 		 * situation with dirtyoff/end, we 0 both of them.
1159 		 */
1160 
1161 		if (bp->b_dirtyend > bcount) {
1162 			printf("NFS append race @%lx:%d\n",
1163 			    (long)bp->b_blkno * DEV_BSIZE,
1164 			    bp->b_dirtyend - bcount);
1165 			bp->b_dirtyend = bcount;
1166 		}
1167 
1168 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1169 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1170 
1171 		/*
1172 		 * If the new write will leave a contiguous dirty
1173 		 * area, just update the b_dirtyoff and b_dirtyend,
1174 		 * otherwise force a write rpc of the old dirty area.
1175 		 *
1176 		 * If there has been a file lock applied to this file
1177 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1178 		 * While it is possible to merge discontiguous writes due to
1179 		 * our having a B_CACHE buffer ( and thus valid read data
1180 		 * for the hole), we don't because it could lead to
1181 		 * significant cache coherency problems with multiple clients,
1182 		 * especially if locking is implemented later on.
1183 		 *
1184 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1185 		 * not been file locking done on this file:
1186 		 * Relax coherency a bit for the sake of performance and
1187 		 * expand the current dirty region to contain the new
1188 		 * write even if it means we mark some non-dirty data as
1189 		 * dirty.
1190 		 */
1191 
1192 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1193 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1194 			if (bwrite(bp) == EINTR) {
1195 				error = EINTR;
1196 				break;
1197 			}
1198 			goto again;
1199 		}
1200 
1201 		local_resid = uio->uio_resid;
1202 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1203 
1204 		if (error != 0 && !bp_cached) {
1205 			/*
1206 			 * This block has no other content then what
1207 			 * possibly was written by the faulty uiomove.
1208 			 * Release it, forgetting the data pages, to
1209 			 * prevent the leak of uninitialized data to
1210 			 * usermode.
1211 			 */
1212 			bp->b_ioflags |= BIO_ERROR;
1213 			brelse(bp);
1214 			uio->uio_offset -= local_resid - uio->uio_resid;
1215 			uio->uio_resid = local_resid;
1216 			break;
1217 		}
1218 
1219 		/*
1220 		 * Since this block is being modified, it must be written
1221 		 * again and not just committed.  Since write clustering does
1222 		 * not work for the stage 1 data write, only the stage 2
1223 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1224 		 */
1225 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1226 
1227 		/*
1228 		 * Get the partial update on the progress made from
1229 		 * uiomove, if an error occurred.
1230 		 */
1231 		if (error != 0)
1232 			n = local_resid - uio->uio_resid;
1233 
1234 		/*
1235 		 * Only update dirtyoff/dirtyend if not a degenerate
1236 		 * condition.
1237 		 */
1238 		if (n > 0) {
1239 			if (bp->b_dirtyend > 0) {
1240 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1241 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1242 			} else {
1243 				bp->b_dirtyoff = on;
1244 				bp->b_dirtyend = on + n;
1245 			}
1246 			vfs_bio_set_valid(bp, on, n);
1247 		}
1248 
1249 		/*
1250 		 * If IO_SYNC do bwrite().
1251 		 *
1252 		 * IO_INVAL appears to be unused.  The idea appears to be
1253 		 * to turn off caching in this case.  Very odd.  XXX
1254 		 */
1255 		if ((ioflag & IO_SYNC)) {
1256 			if (ioflag & IO_INVAL)
1257 				bp->b_flags |= B_NOCACHE;
1258 			error1 = bwrite(bp);
1259 			if (error1 != 0) {
1260 				if (error == 0)
1261 					error = error1;
1262 				break;
1263 			}
1264 		} else if ((n + on) == biosize) {
1265 			bp->b_flags |= B_ASYNC;
1266 			(void) ncl_writebp(bp, 0, NULL);
1267 		} else {
1268 			bdwrite(bp);
1269 		}
1270 
1271 		if (error != 0)
1272 			break;
1273 	} while (uio->uio_resid > 0 && n > 0);
1274 
1275 	if (error != 0) {
1276 		if (ioflag & IO_UNIT) {
1277 			VATTR_NULL(&vattr);
1278 			vattr.va_size = orig_size;
1279 			/* IO_SYNC is handled implicitely */
1280 			(void)VOP_SETATTR(vp, &vattr, cred);
1281 			uio->uio_offset -= orig_resid - uio->uio_resid;
1282 			uio->uio_resid = orig_resid;
1283 		}
1284 	}
1285 
1286 	return (error);
1287 }
1288 
1289 /*
1290  * Get an nfs cache block.
1291  *
1292  * Allocate a new one if the block isn't currently in the cache
1293  * and return the block marked busy. If the calling process is
1294  * interrupted by a signal for an interruptible mount point, return
1295  * NULL.
1296  *
1297  * The caller must carefully deal with the possible B_INVAL state of
1298  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1299  * indirectly), so synchronous reads can be issued without worrying about
1300  * the B_INVAL state.  We have to be a little more careful when dealing
1301  * with writes (see comments in nfs_write()) when extending a file past
1302  * its EOF.
1303  */
1304 static struct buf *
1305 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1306 {
1307 	struct buf *bp;
1308 	struct mount *mp;
1309 	struct nfsmount *nmp;
1310 
1311 	mp = vp->v_mount;
1312 	nmp = VFSTONFS(mp);
1313 
1314 	if (nmp->nm_flag & NFSMNT_INT) {
1315 		sigset_t oldset;
1316 
1317 		newnfs_set_sigmask(td, &oldset);
1318 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1319 		newnfs_restore_sigmask(td, &oldset);
1320 		while (bp == NULL) {
1321 			if (newnfs_sigintr(nmp, td))
1322 				return (NULL);
1323 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1324 		}
1325 	} else {
1326 		bp = getblk(vp, bn, size, 0, 0, 0);
1327 	}
1328 
1329 	if (vp->v_type == VREG)
1330 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1331 	return (bp);
1332 }
1333 
1334 /*
1335  * Flush and invalidate all dirty buffers. If another process is already
1336  * doing the flush, just wait for completion.
1337  */
1338 int
1339 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1340 {
1341 	struct nfsnode *np = VTONFS(vp);
1342 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1343 	int error = 0, slpflag, slptimeo;
1344 	int old_lock = 0;
1345 
1346 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1347 
1348 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1349 		intrflg = 0;
1350 	if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1351 		intrflg = 1;
1352 	if (intrflg) {
1353 		slpflag = PCATCH;
1354 		slptimeo = 2 * hz;
1355 	} else {
1356 		slpflag = 0;
1357 		slptimeo = 0;
1358 	}
1359 
1360 	old_lock = ncl_upgrade_vnlock(vp);
1361 	if (vp->v_iflag & VI_DOOMED) {
1362 		/*
1363 		 * Since vgonel() uses the generic vinvalbuf() to flush
1364 		 * dirty buffers and it does not call this function, it
1365 		 * is safe to just return OK when VI_DOOMED is set.
1366 		 */
1367 		ncl_downgrade_vnlock(vp, old_lock);
1368 		return (0);
1369 	}
1370 
1371 	/*
1372 	 * Now, flush as required.
1373 	 */
1374 	if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1375 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1376 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1377 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1378 		/*
1379 		 * If the page clean was interrupted, fail the invalidation.
1380 		 * Not doing so, we run the risk of losing dirty pages in the
1381 		 * vinvalbuf() call below.
1382 		 */
1383 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1384 			goto out;
1385 	}
1386 
1387 	error = vinvalbuf(vp, flags, slpflag, 0);
1388 	while (error) {
1389 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1390 			goto out;
1391 		error = vinvalbuf(vp, flags, 0, slptimeo);
1392 	}
1393 	if (NFSHASPNFS(nmp)) {
1394 		nfscl_layoutcommit(vp, td);
1395 		/*
1396 		 * Invalidate the attribute cache, since writes to a DS
1397 		 * won't update the size attribute.
1398 		 */
1399 		mtx_lock(&np->n_mtx);
1400 		np->n_attrstamp = 0;
1401 	} else
1402 		mtx_lock(&np->n_mtx);
1403 	if (np->n_directio_asyncwr == 0)
1404 		np->n_flag &= ~NMODIFIED;
1405 	mtx_unlock(&np->n_mtx);
1406 out:
1407 	ncl_downgrade_vnlock(vp, old_lock);
1408 	return error;
1409 }
1410 
1411 /*
1412  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1413  * This is mainly to avoid queueing async I/O requests when the nfsiods
1414  * are all hung on a dead server.
1415  *
1416  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1417  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1418  */
1419 int
1420 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1421 {
1422 	int iod;
1423 	int gotiod;
1424 	int slpflag = 0;
1425 	int slptimeo = 0;
1426 	int error, error2;
1427 
1428 	/*
1429 	 * Commits are usually short and sweet so lets save some cpu and
1430 	 * leave the async daemons for more important rpc's (such as reads
1431 	 * and writes).
1432 	 *
1433 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1434 	 * in the directory in order to update attributes. This can deadlock
1435 	 * with another thread that is waiting for async I/O to be done by
1436 	 * an nfsiod thread while holding a lock on one of these vnodes.
1437 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1438 	 * perform Readdirplus RPCs.
1439 	 */
1440 	mtx_lock(&ncl_iod_mutex);
1441 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1442 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1443 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1444 		mtx_unlock(&ncl_iod_mutex);
1445 		return(EIO);
1446 	}
1447 again:
1448 	if (nmp->nm_flag & NFSMNT_INT)
1449 		slpflag = PCATCH;
1450 	gotiod = FALSE;
1451 
1452 	/*
1453 	 * Find a free iod to process this request.
1454 	 */
1455 	for (iod = 0; iod < ncl_numasync; iod++)
1456 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1457 			gotiod = TRUE;
1458 			break;
1459 		}
1460 
1461 	/*
1462 	 * Try to create one if none are free.
1463 	 */
1464 	if (!gotiod)
1465 		ncl_nfsiodnew();
1466 	else {
1467 		/*
1468 		 * Found one, so wake it up and tell it which
1469 		 * mount to process.
1470 		 */
1471 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1472 		    iod, nmp));
1473 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1474 		ncl_iodmount[iod] = nmp;
1475 		nmp->nm_bufqiods++;
1476 		wakeup(&ncl_iodwant[iod]);
1477 	}
1478 
1479 	/*
1480 	 * If none are free, we may already have an iod working on this mount
1481 	 * point.  If so, it will process our request.
1482 	 */
1483 	if (!gotiod) {
1484 		if (nmp->nm_bufqiods > 0) {
1485 			NFS_DPF(ASYNCIO,
1486 				("ncl_asyncio: %d iods are already processing mount %p\n",
1487 				 nmp->nm_bufqiods, nmp));
1488 			gotiod = TRUE;
1489 		}
1490 	}
1491 
1492 	/*
1493 	 * If we have an iod which can process the request, then queue
1494 	 * the buffer.
1495 	 */
1496 	if (gotiod) {
1497 		/*
1498 		 * Ensure that the queue never grows too large.  We still want
1499 		 * to asynchronize so we block rather then return EIO.
1500 		 */
1501 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1502 			NFS_DPF(ASYNCIO,
1503 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1504 			nmp->nm_bufqwant = TRUE;
1505 			error = newnfs_msleep(td, &nmp->nm_bufq,
1506 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1507 			   slptimeo);
1508 			if (error) {
1509 				error2 = newnfs_sigintr(nmp, td);
1510 				if (error2) {
1511 					mtx_unlock(&ncl_iod_mutex);
1512 					return (error2);
1513 				}
1514 				if (slpflag == PCATCH) {
1515 					slpflag = 0;
1516 					slptimeo = 2 * hz;
1517 				}
1518 			}
1519 			/*
1520 			 * We might have lost our iod while sleeping,
1521 			 * so check and loop if necessary.
1522 			 */
1523 			goto again;
1524 		}
1525 
1526 		/* We might have lost our nfsiod */
1527 		if (nmp->nm_bufqiods == 0) {
1528 			NFS_DPF(ASYNCIO,
1529 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1530 			goto again;
1531 		}
1532 
1533 		if (bp->b_iocmd == BIO_READ) {
1534 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1535 				bp->b_rcred = crhold(cred);
1536 		} else {
1537 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1538 				bp->b_wcred = crhold(cred);
1539 		}
1540 
1541 		if (bp->b_flags & B_REMFREE)
1542 			bremfreef(bp);
1543 		BUF_KERNPROC(bp);
1544 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1545 		nmp->nm_bufqlen++;
1546 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1547 			mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1548 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1549 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1550 			mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1551 		}
1552 		mtx_unlock(&ncl_iod_mutex);
1553 		return (0);
1554 	}
1555 
1556 	mtx_unlock(&ncl_iod_mutex);
1557 
1558 	/*
1559 	 * All the iods are busy on other mounts, so return EIO to
1560 	 * force the caller to process the i/o synchronously.
1561 	 */
1562 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1563 	return (EIO);
1564 }
1565 
1566 void
1567 ncl_doio_directwrite(struct buf *bp)
1568 {
1569 	int iomode, must_commit;
1570 	struct uio *uiop = (struct uio *)bp->b_caller1;
1571 	char *iov_base = uiop->uio_iov->iov_base;
1572 
1573 	iomode = NFSWRITE_FILESYNC;
1574 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1575 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1576 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1577 	free(iov_base, M_NFSDIRECTIO);
1578 	free(uiop->uio_iov, M_NFSDIRECTIO);
1579 	free(uiop, M_NFSDIRECTIO);
1580 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1581 		struct nfsnode *np = VTONFS(bp->b_vp);
1582 		mtx_lock(&np->n_mtx);
1583 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1584 			/*
1585 			 * Invalidate the attribute cache, since writes to a DS
1586 			 * won't update the size attribute.
1587 			 */
1588 			np->n_attrstamp = 0;
1589 		}
1590 		np->n_directio_asyncwr--;
1591 		if (np->n_directio_asyncwr == 0) {
1592 			np->n_flag &= ~NMODIFIED;
1593 			if ((np->n_flag & NFSYNCWAIT)) {
1594 				np->n_flag &= ~NFSYNCWAIT;
1595 				wakeup((caddr_t)&np->n_directio_asyncwr);
1596 			}
1597 		}
1598 		mtx_unlock(&np->n_mtx);
1599 	}
1600 	bp->b_vp = NULL;
1601 	relpbuf(bp, &ncl_pbuf_freecnt);
1602 }
1603 
1604 /*
1605  * Do an I/O operation to/from a cache block. This may be called
1606  * synchronously or from an nfsiod.
1607  */
1608 int
1609 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1610     int called_from_strategy)
1611 {
1612 	struct uio *uiop;
1613 	struct nfsnode *np;
1614 	struct nfsmount *nmp;
1615 	int error = 0, iomode, must_commit = 0;
1616 	struct uio uio;
1617 	struct iovec io;
1618 	struct proc *p = td ? td->td_proc : NULL;
1619 	uint8_t	iocmd;
1620 
1621 	np = VTONFS(vp);
1622 	nmp = VFSTONFS(vp->v_mount);
1623 	uiop = &uio;
1624 	uiop->uio_iov = &io;
1625 	uiop->uio_iovcnt = 1;
1626 	uiop->uio_segflg = UIO_SYSSPACE;
1627 	uiop->uio_td = td;
1628 
1629 	/*
1630 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1631 	 * do this here so we do not have to do it in all the code that
1632 	 * calls us.
1633 	 */
1634 	bp->b_flags &= ~B_INVAL;
1635 	bp->b_ioflags &= ~BIO_ERROR;
1636 
1637 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1638 	iocmd = bp->b_iocmd;
1639 	if (iocmd == BIO_READ) {
1640 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1641 	    io.iov_base = bp->b_data;
1642 	    uiop->uio_rw = UIO_READ;
1643 
1644 	    switch (vp->v_type) {
1645 	    case VREG:
1646 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1647 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1648 		error = ncl_readrpc(vp, uiop, cr);
1649 
1650 		if (!error) {
1651 		    if (uiop->uio_resid) {
1652 			/*
1653 			 * If we had a short read with no error, we must have
1654 			 * hit a file hole.  We should zero-fill the remainder.
1655 			 * This can also occur if the server hits the file EOF.
1656 			 *
1657 			 * Holes used to be able to occur due to pending
1658 			 * writes, but that is not possible any longer.
1659 			 */
1660 			int nread = bp->b_bcount - uiop->uio_resid;
1661 			ssize_t left = uiop->uio_resid;
1662 
1663 			if (left > 0)
1664 				bzero((char *)bp->b_data + nread, left);
1665 			uiop->uio_resid = 0;
1666 		    }
1667 		}
1668 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1669 		if (p && (vp->v_vflag & VV_TEXT)) {
1670 			mtx_lock(&np->n_mtx);
1671 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1672 				mtx_unlock(&np->n_mtx);
1673 				PROC_LOCK(p);
1674 				killproc(p, "text file modification");
1675 				PROC_UNLOCK(p);
1676 			} else
1677 				mtx_unlock(&np->n_mtx);
1678 		}
1679 		break;
1680 	    case VLNK:
1681 		uiop->uio_offset = (off_t)0;
1682 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1683 		error = ncl_readlinkrpc(vp, uiop, cr);
1684 		break;
1685 	    case VDIR:
1686 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1687 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1688 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1689 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1690 			if (error == NFSERR_NOTSUPP)
1691 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1692 		}
1693 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1694 			error = ncl_readdirrpc(vp, uiop, cr, td);
1695 		/*
1696 		 * end-of-directory sets B_INVAL but does not generate an
1697 		 * error.
1698 		 */
1699 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1700 			bp->b_flags |= B_INVAL;
1701 		break;
1702 	    default:
1703 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1704 		break;
1705 	    }
1706 	    if (error) {
1707 		bp->b_ioflags |= BIO_ERROR;
1708 		bp->b_error = error;
1709 	    }
1710 	} else {
1711 	    /*
1712 	     * If we only need to commit, try to commit
1713 	     */
1714 	    if (bp->b_flags & B_NEEDCOMMIT) {
1715 		    int retv;
1716 		    off_t off;
1717 
1718 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1719 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1720 			bp->b_wcred, td);
1721 		    if (retv == 0) {
1722 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1723 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1724 			    bp->b_resid = 0;
1725 			    bufdone(bp);
1726 			    return (0);
1727 		    }
1728 		    if (retv == NFSERR_STALEWRITEVERF) {
1729 			    ncl_clearcommit(vp->v_mount);
1730 		    }
1731 	    }
1732 
1733 	    /*
1734 	     * Setup for actual write
1735 	     */
1736 	    mtx_lock(&np->n_mtx);
1737 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1738 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1739 	    mtx_unlock(&np->n_mtx);
1740 
1741 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1742 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1743 		    - bp->b_dirtyoff;
1744 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1745 		    + bp->b_dirtyoff;
1746 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1747 		uiop->uio_rw = UIO_WRITE;
1748 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1749 
1750 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1751 		    iomode = NFSWRITE_UNSTABLE;
1752 		else
1753 		    iomode = NFSWRITE_FILESYNC;
1754 
1755 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1756 		    called_from_strategy);
1757 
1758 		/*
1759 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1760 		 * to cluster the buffers needing commit.  This will allow
1761 		 * the system to submit a single commit rpc for the whole
1762 		 * cluster.  We can do this even if the buffer is not 100%
1763 		 * dirty (relative to the NFS blocksize), so we optimize the
1764 		 * append-to-file-case.
1765 		 *
1766 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1767 		 * cleared because write clustering only works for commit
1768 		 * rpc's, not for the data portion of the write).
1769 		 */
1770 
1771 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1772 		    bp->b_flags |= B_NEEDCOMMIT;
1773 		    if (bp->b_dirtyoff == 0
1774 			&& bp->b_dirtyend == bp->b_bcount)
1775 			bp->b_flags |= B_CLUSTEROK;
1776 		} else {
1777 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1778 		}
1779 
1780 		/*
1781 		 * For an interrupted write, the buffer is still valid
1782 		 * and the write hasn't been pushed to the server yet,
1783 		 * so we can't set BIO_ERROR and report the interruption
1784 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1785 		 * is not relevant, so the rpc attempt is essentially
1786 		 * a noop.  For the case of a V3 write rpc not being
1787 		 * committed to stable storage, the block is still
1788 		 * dirty and requires either a commit rpc or another
1789 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1790 		 * the block is reused. This is indicated by setting
1791 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1792 		 *
1793 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1794 		 * write error and is handled as above, except that
1795 		 * B_EINTR isn't set. One cause of this is a stale stateid
1796 		 * error for the RPC that indicates recovery is required,
1797 		 * when called with called_from_strategy != 0.
1798 		 *
1799 		 * If the buffer is marked B_PAGING, it does not reside on
1800 		 * the vp's paging queues so we cannot call bdirty().  The
1801 		 * bp in this case is not an NFS cache block so we should
1802 		 * be safe. XXX
1803 		 *
1804 		 * The logic below breaks up errors into recoverable and
1805 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1806 		 * and keep the buffer around for potential write retries.
1807 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1808 		 * and save the error in the nfsnode. This is less than ideal
1809 		 * but necessary. Keeping such buffers around could potentially
1810 		 * cause buffer exhaustion eventually (they can never be written
1811 		 * out, so will get constantly be re-dirtied). It also causes
1812 		 * all sorts of vfs panics. For non-recoverable write errors,
1813 		 * also invalidate the attrcache, so we'll be forced to go over
1814 		 * the wire for this object, returning an error to user on next
1815 		 * call (most of the time).
1816 		 */
1817 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1818 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1819 			int s;
1820 
1821 			s = splbio();
1822 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1823 			if ((bp->b_flags & B_PAGING) == 0) {
1824 			    bdirty(bp);
1825 			    bp->b_flags &= ~B_DONE;
1826 			}
1827 			if ((error == EINTR || error == ETIMEDOUT) &&
1828 			    (bp->b_flags & B_ASYNC) == 0)
1829 			    bp->b_flags |= B_EINTR;
1830 			splx(s);
1831 		} else {
1832 		    if (error) {
1833 			bp->b_ioflags |= BIO_ERROR;
1834 			bp->b_flags |= B_INVAL;
1835 			bp->b_error = np->n_error = error;
1836 			mtx_lock(&np->n_mtx);
1837 			np->n_flag |= NWRITEERR;
1838 			np->n_attrstamp = 0;
1839 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1840 			mtx_unlock(&np->n_mtx);
1841 		    }
1842 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1843 		}
1844 	    } else {
1845 		bp->b_resid = 0;
1846 		bufdone(bp);
1847 		return (0);
1848 	    }
1849 	}
1850 	bp->b_resid = uiop->uio_resid;
1851 	if (must_commit)
1852 	    ncl_clearcommit(vp->v_mount);
1853 	bufdone(bp);
1854 	return (error);
1855 }
1856 
1857 /*
1858  * Used to aid in handling ftruncate() operations on the NFS client side.
1859  * Truncation creates a number of special problems for NFS.  We have to
1860  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1861  * we have to properly handle VM pages or (potentially dirty) buffers
1862  * that straddle the truncation point.
1863  */
1864 
1865 int
1866 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1867 {
1868 	struct nfsnode *np = VTONFS(vp);
1869 	u_quad_t tsize;
1870 	int biosize = vp->v_bufobj.bo_bsize;
1871 	int error = 0;
1872 
1873 	mtx_lock(&np->n_mtx);
1874 	tsize = np->n_size;
1875 	np->n_size = nsize;
1876 	mtx_unlock(&np->n_mtx);
1877 
1878 	if (nsize < tsize) {
1879 		struct buf *bp;
1880 		daddr_t lbn;
1881 		int bufsize;
1882 
1883 		/*
1884 		 * vtruncbuf() doesn't get the buffer overlapping the
1885 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1886 		 * buffer that now needs to be truncated.
1887 		 */
1888 		error = vtruncbuf(vp, cred, nsize, biosize);
1889 		lbn = nsize / biosize;
1890 		bufsize = nsize - (lbn * biosize);
1891 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1892 		if (!bp)
1893 			return EINTR;
1894 		if (bp->b_dirtyoff > bp->b_bcount)
1895 			bp->b_dirtyoff = bp->b_bcount;
1896 		if (bp->b_dirtyend > bp->b_bcount)
1897 			bp->b_dirtyend = bp->b_bcount;
1898 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1899 		brelse(bp);
1900 	} else {
1901 		vnode_pager_setsize(vp, nsize);
1902 	}
1903 	return(error);
1904 }
1905 
1906