xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision b0d29bc47dba79f6f38e67eabadfb4b32ffd9390)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Rick Macklem at The University of Guelph.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bio.h>
43 #include <sys/buf.h>
44 #include <sys/kernel.h>
45 #include <sys/mount.h>
46 #include <sys/rwlock.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
49 
50 #include <vm/vm.h>
51 #include <vm/vm_param.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
57 
58 #include <fs/nfs/nfsport.h>
59 #include <fs/nfsclient/nfsmount.h>
60 #include <fs/nfsclient/nfs.h>
61 #include <fs/nfsclient/nfsnode.h>
62 #include <fs/nfsclient/nfs_kdtrace.h>
63 
64 extern int newnfs_directio_allow_mmap;
65 extern struct nfsstatsv1 nfsstatsv1;
66 extern struct mtx ncl_iod_mutex;
67 extern int ncl_numasync;
68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
70 extern int newnfs_directio_enable;
71 extern int nfs_keep_dirty_on_error;
72 
73 uma_zone_t ncl_pbuf_zone;
74 
75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
76     struct thread *td);
77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
78     struct ucred *cred, int ioflag);
79 
80 /*
81  * Vnode op for VM getpages.
82  */
83 SYSCTL_DECL(_vfs_nfs);
84 static int use_buf_pager = 1;
85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
86     &use_buf_pager, 0,
87     "Use buffer pager instead of direct readrpc call");
88 
89 static daddr_t
90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
91 {
92 
93 	return (off / vp->v_bufobj.bo_bsize);
94 }
95 
96 static int
97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn)
98 {
99 	struct nfsnode *np;
100 	u_quad_t nsize;
101 	int biosize, bcount;
102 
103 	np = VTONFS(vp);
104 	NFSLOCKNODE(np);
105 	nsize = np->n_size;
106 	NFSUNLOCKNODE(np);
107 
108 	biosize = vp->v_bufobj.bo_bsize;
109 	bcount = biosize;
110 	if ((off_t)lbn * biosize >= nsize)
111 		bcount = 0;
112 	else if ((off_t)(lbn + 1) * biosize > nsize)
113 		bcount = nsize - (off_t)lbn * biosize;
114 	return (bcount);
115 }
116 
117 int
118 ncl_getpages(struct vop_getpages_args *ap)
119 {
120 	int i, error, nextoff, size, toff, count, npages;
121 	struct uio uio;
122 	struct iovec iov;
123 	vm_offset_t kva;
124 	struct buf *bp;
125 	struct vnode *vp;
126 	struct thread *td;
127 	struct ucred *cred;
128 	struct nfsmount *nmp;
129 	vm_object_t object;
130 	vm_page_t *pages;
131 	struct nfsnode *np;
132 
133 	vp = ap->a_vp;
134 	np = VTONFS(vp);
135 	td = curthread;
136 	cred = curthread->td_ucred;
137 	nmp = VFSTONFS(vp->v_mount);
138 	pages = ap->a_m;
139 	npages = ap->a_count;
140 
141 	if ((object = vp->v_object) == NULL) {
142 		printf("ncl_getpages: called with non-merged cache vnode\n");
143 		return (VM_PAGER_ERROR);
144 	}
145 
146 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
147 		NFSLOCKNODE(np);
148 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
149 			NFSUNLOCKNODE(np);
150 			printf("ncl_getpages: called on non-cacheable vnode\n");
151 			return (VM_PAGER_ERROR);
152 		} else
153 			NFSUNLOCKNODE(np);
154 	}
155 
156 	mtx_lock(&nmp->nm_mtx);
157 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
158 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
159 		mtx_unlock(&nmp->nm_mtx);
160 		/* We'll never get here for v4, because we always have fsinfo */
161 		(void)ncl_fsinfo(nmp, vp, cred, td);
162 	} else
163 		mtx_unlock(&nmp->nm_mtx);
164 
165 	if (use_buf_pager)
166 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
167 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
168 
169 	/*
170 	 * If the requested page is partially valid, just return it and
171 	 * allow the pager to zero-out the blanks.  Partially valid pages
172 	 * can only occur at the file EOF.
173 	 *
174 	 * XXXGL: is that true for NFS, where short read can occur???
175 	 */
176 	VM_OBJECT_WLOCK(object);
177 	if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
178 		goto out;
179 	VM_OBJECT_WUNLOCK(object);
180 
181 	/*
182 	 * We use only the kva address for the buffer, but this is extremely
183 	 * convenient and fast.
184 	 */
185 	bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
186 
187 	kva = (vm_offset_t) bp->b_data;
188 	pmap_qenter(kva, pages, npages);
189 	VM_CNT_INC(v_vnodein);
190 	VM_CNT_ADD(v_vnodepgsin, npages);
191 
192 	count = npages << PAGE_SHIFT;
193 	iov.iov_base = (caddr_t) kva;
194 	iov.iov_len = count;
195 	uio.uio_iov = &iov;
196 	uio.uio_iovcnt = 1;
197 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
198 	uio.uio_resid = count;
199 	uio.uio_segflg = UIO_SYSSPACE;
200 	uio.uio_rw = UIO_READ;
201 	uio.uio_td = td;
202 
203 	error = ncl_readrpc(vp, &uio, cred);
204 	pmap_qremove(kva, npages);
205 
206 	uma_zfree(ncl_pbuf_zone, bp);
207 
208 	if (error && (uio.uio_resid == count)) {
209 		printf("ncl_getpages: error %d\n", error);
210 		return (VM_PAGER_ERROR);
211 	}
212 
213 	/*
214 	 * Calculate the number of bytes read and validate only that number
215 	 * of bytes.  Note that due to pending writes, size may be 0.  This
216 	 * does not mean that the remaining data is invalid!
217 	 */
218 
219 	size = count - uio.uio_resid;
220 	VM_OBJECT_WLOCK(object);
221 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
222 		vm_page_t m;
223 		nextoff = toff + PAGE_SIZE;
224 		m = pages[i];
225 
226 		if (nextoff <= size) {
227 			/*
228 			 * Read operation filled an entire page
229 			 */
230 			vm_page_valid(m);
231 			KASSERT(m->dirty == 0,
232 			    ("nfs_getpages: page %p is dirty", m));
233 		} else if (size > toff) {
234 			/*
235 			 * Read operation filled a partial page.
236 			 */
237 			vm_page_invalid(m);
238 			vm_page_set_valid_range(m, 0, size - toff);
239 			KASSERT(m->dirty == 0,
240 			    ("nfs_getpages: page %p is dirty", m));
241 		} else {
242 			/*
243 			 * Read operation was short.  If no error
244 			 * occurred we may have hit a zero-fill
245 			 * section.  We leave valid set to 0, and page
246 			 * is freed by vm_page_readahead_finish() if
247 			 * its index is not equal to requested, or
248 			 * page is zeroed and set valid by
249 			 * vm_pager_get_pages() for requested page.
250 			 */
251 			;
252 		}
253 	}
254 out:
255 	VM_OBJECT_WUNLOCK(object);
256 	if (ap->a_rbehind)
257 		*ap->a_rbehind = 0;
258 	if (ap->a_rahead)
259 		*ap->a_rahead = 0;
260 	return (VM_PAGER_OK);
261 }
262 
263 /*
264  * Vnode op for VM putpages.
265  */
266 int
267 ncl_putpages(struct vop_putpages_args *ap)
268 {
269 	struct uio uio;
270 	struct iovec iov;
271 	int 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 	NFSLOCKNODE(np);
305 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
306 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
307 		NFSUNLOCKNODE(np);
308 		printf("ncl_putpages: called on noncache-able vnode\n");
309 		NFSLOCKNODE(np);
310 	}
311 	/*
312 	 * When putting pages, do not extend file past EOF.
313 	 */
314 	if (offset + count > np->n_size) {
315 		count = np->n_size - offset;
316 		if (count < 0)
317 			count = 0;
318 	}
319 	NFSUNLOCKNODE(np);
320 
321 	for (i = 0; i < npages; i++)
322 		rtvals[i] = VM_PAGER_ERROR;
323 
324 	VM_CNT_INC(v_vnodeout);
325 	VM_CNT_ADD(v_vnodepgsout, count);
326 
327 	iov.iov_base = unmapped_buf;
328 	iov.iov_len = count;
329 	uio.uio_iov = &iov;
330 	uio.uio_iovcnt = 1;
331 	uio.uio_offset = offset;
332 	uio.uio_resid = count;
333 	uio.uio_segflg = UIO_NOCOPY;
334 	uio.uio_rw = UIO_WRITE;
335 	uio.uio_td = td;
336 
337 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
338 	    cred);
339 	crfree(cred);
340 
341 	if (error == 0 || !nfs_keep_dirty_on_error) {
342 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
343 		    np->n_size - offset, npages * PAGE_SIZE);
344 	}
345 	return (rtvals[0]);
346 }
347 
348 /*
349  * For nfs, cache consistency can only be maintained approximately.
350  * Although RFC1094 does not specify the criteria, the following is
351  * believed to be compatible with the reference port.
352  * For nfs:
353  * If the file's modify time on the server has changed since the
354  * last read rpc or you have written to the file,
355  * you may have lost data cache consistency with the
356  * server, so flush all of the file's data out of the cache.
357  * Then force a getattr rpc to ensure that you have up to date
358  * attributes.
359  * NB: This implies that cache data can be read when up to
360  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
361  * attributes this could be forced by setting n_attrstamp to 0 before
362  * the VOP_GETATTR() call.
363  */
364 static inline int
365 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
366 {
367 	int error = 0;
368 	struct vattr vattr;
369 	struct nfsnode *np = VTONFS(vp);
370 	bool old_lock;
371 
372 	/*
373 	 * Ensure the exclusove access to the node before checking
374 	 * whether the cache is consistent.
375 	 */
376 	old_lock = ncl_excl_start(vp);
377 	NFSLOCKNODE(np);
378 	if (np->n_flag & NMODIFIED) {
379 		NFSUNLOCKNODE(np);
380 		if (vp->v_type != VREG) {
381 			if (vp->v_type != VDIR)
382 				panic("nfs: bioread, not dir");
383 			ncl_invaldir(vp);
384 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
385 			if (error != 0)
386 				goto out;
387 		}
388 		np->n_attrstamp = 0;
389 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
390 		error = VOP_GETATTR(vp, &vattr, cred);
391 		if (error)
392 			goto out;
393 		NFSLOCKNODE(np);
394 		np->n_mtime = vattr.va_mtime;
395 		NFSUNLOCKNODE(np);
396 	} else {
397 		NFSUNLOCKNODE(np);
398 		error = VOP_GETATTR(vp, &vattr, cred);
399 		if (error)
400 			goto out;
401 		NFSLOCKNODE(np);
402 		if ((np->n_flag & NSIZECHANGED)
403 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
404 			NFSUNLOCKNODE(np);
405 			if (vp->v_type == VDIR)
406 				ncl_invaldir(vp);
407 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
408 			if (error != 0)
409 				goto out;
410 			NFSLOCKNODE(np);
411 			np->n_mtime = vattr.va_mtime;
412 			np->n_flag &= ~NSIZECHANGED;
413 		}
414 		NFSUNLOCKNODE(np);
415 	}
416 out:
417 	ncl_excl_finish(vp, old_lock);
418 	return (error);
419 }
420 
421 /*
422  * Vnode op for read using bio
423  */
424 int
425 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
426 {
427 	struct nfsnode *np = VTONFS(vp);
428 	struct buf *bp, *rabp;
429 	struct thread *td;
430 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
431 	daddr_t lbn, rabn;
432 	int biosize, bcount, error, i, n, nra, on, save2, seqcount;
433 	off_t tmp_off;
434 
435 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
436 	if (uio->uio_resid == 0)
437 		return (0);
438 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
439 		return (EINVAL);
440 	td = uio->uio_td;
441 
442 	mtx_lock(&nmp->nm_mtx);
443 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
444 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
445 		mtx_unlock(&nmp->nm_mtx);
446 		(void)ncl_fsinfo(nmp, vp, cred, td);
447 		mtx_lock(&nmp->nm_mtx);
448 	}
449 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
450 		(void) newnfs_iosize(nmp);
451 
452 	tmp_off = uio->uio_offset + uio->uio_resid;
453 	if (vp->v_type != VDIR &&
454 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
455 		mtx_unlock(&nmp->nm_mtx);
456 		return (EFBIG);
457 	}
458 	mtx_unlock(&nmp->nm_mtx);
459 
460 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
461 		/* No caching/ no readaheads. Just read data into the user buffer */
462 		return ncl_readrpc(vp, uio, cred);
463 
464 	n = 0;
465 	on = 0;
466 	biosize = vp->v_bufobj.bo_bsize;
467 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
468 
469 	error = nfs_bioread_check_cons(vp, td, cred);
470 	if (error)
471 		return error;
472 
473 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
474 	do {
475 	    u_quad_t nsize;
476 
477 	    NFSLOCKNODE(np);
478 	    nsize = np->n_size;
479 	    NFSUNLOCKNODE(np);
480 
481 	    switch (vp->v_type) {
482 	    case VREG:
483 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
484 		lbn = uio->uio_offset / biosize;
485 		on = uio->uio_offset - (lbn * biosize);
486 
487 		/*
488 		 * Start the read ahead(s), as required.
489 		 */
490 		if (nmp->nm_readahead > 0) {
491 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
492 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
493 			rabn = lbn + 1 + nra;
494 			if (incore(&vp->v_bufobj, rabn) == NULL) {
495 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
496 			    if (!rabp) {
497 				error = newnfs_sigintr(nmp, td);
498 				if (error == 0)
499 					error = EINTR;
500 				goto out;
501 			    }
502 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
503 				rabp->b_flags |= B_ASYNC;
504 				rabp->b_iocmd = BIO_READ;
505 				vfs_busy_pages(rabp, 0);
506 				if (ncl_asyncio(nmp, rabp, cred, td)) {
507 				    rabp->b_flags |= B_INVAL;
508 				    rabp->b_ioflags |= BIO_ERROR;
509 				    vfs_unbusy_pages(rabp);
510 				    brelse(rabp);
511 				    break;
512 				}
513 			    } else {
514 				brelse(rabp);
515 			    }
516 			}
517 		    }
518 		}
519 
520 		/* Note that bcount is *not* DEV_BSIZE aligned. */
521 		bcount = biosize;
522 		if ((off_t)lbn * biosize >= nsize) {
523 			bcount = 0;
524 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
525 			bcount = nsize - (off_t)lbn * biosize;
526 		}
527 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
528 
529 		if (!bp) {
530 			error = newnfs_sigintr(nmp, td);
531 			if (error == 0)
532 				error = EINTR;
533 			goto out;
534 		}
535 
536 		/*
537 		 * If B_CACHE is not set, we must issue the read.  If this
538 		 * fails, we return an error.
539 		 */
540 
541 		if ((bp->b_flags & B_CACHE) == 0) {
542 		    bp->b_iocmd = BIO_READ;
543 		    vfs_busy_pages(bp, 0);
544 		    error = ncl_doio(vp, bp, cred, td, 0);
545 		    if (error) {
546 			brelse(bp);
547 			goto out;
548 		    }
549 		}
550 
551 		/*
552 		 * on is the offset into the current bp.  Figure out how many
553 		 * bytes we can copy out of the bp.  Note that bcount is
554 		 * NOT DEV_BSIZE aligned.
555 		 *
556 		 * Then figure out how many bytes we can copy into the uio.
557 		 */
558 
559 		n = 0;
560 		if (on < bcount)
561 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
562 		break;
563 	    case VLNK:
564 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
565 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
566 		if (!bp) {
567 			error = newnfs_sigintr(nmp, td);
568 			if (error == 0)
569 				error = EINTR;
570 			goto out;
571 		}
572 		if ((bp->b_flags & B_CACHE) == 0) {
573 		    bp->b_iocmd = BIO_READ;
574 		    vfs_busy_pages(bp, 0);
575 		    error = ncl_doio(vp, bp, cred, td, 0);
576 		    if (error) {
577 			bp->b_ioflags |= BIO_ERROR;
578 			brelse(bp);
579 			goto out;
580 		    }
581 		}
582 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
583 		on = 0;
584 		break;
585 	    case VDIR:
586 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
587 		if (np->n_direofoffset
588 		    && uio->uio_offset >= np->n_direofoffset) {
589 			error = 0;
590 			goto out;
591 		}
592 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
593 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
594 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
595 		if (!bp) {
596 			error = newnfs_sigintr(nmp, td);
597 			if (error == 0)
598 				error = EINTR;
599 			goto out;
600 		}
601 		if ((bp->b_flags & B_CACHE) == 0) {
602 		    bp->b_iocmd = BIO_READ;
603 		    vfs_busy_pages(bp, 0);
604 		    error = ncl_doio(vp, bp, cred, td, 0);
605 		    if (error) {
606 			    brelse(bp);
607 		    }
608 		    while (error == NFSERR_BAD_COOKIE) {
609 			ncl_invaldir(vp);
610 			error = ncl_vinvalbuf(vp, 0, td, 1);
611 
612 			/*
613 			 * Yuck! The directory has been modified on the
614 			 * server. The only way to get the block is by
615 			 * reading from the beginning to get all the
616 			 * offset cookies.
617 			 *
618 			 * Leave the last bp intact unless there is an error.
619 			 * Loop back up to the while if the error is another
620 			 * NFSERR_BAD_COOKIE (double yuch!).
621 			 */
622 			for (i = 0; i <= lbn && !error; i++) {
623 			    if (np->n_direofoffset
624 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
625 				    error = 0;
626 				    goto out;
627 			    }
628 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
629 			    if (!bp) {
630 				error = newnfs_sigintr(nmp, td);
631 				if (error == 0)
632 					error = EINTR;
633 				goto out;
634 			    }
635 			    if ((bp->b_flags & B_CACHE) == 0) {
636 				    bp->b_iocmd = BIO_READ;
637 				    vfs_busy_pages(bp, 0);
638 				    error = ncl_doio(vp, bp, cred, td, 0);
639 				    /*
640 				     * no error + B_INVAL == directory EOF,
641 				     * use the block.
642 				     */
643 				    if (error == 0 && (bp->b_flags & B_INVAL))
644 					    break;
645 			    }
646 			    /*
647 			     * An error will throw away the block and the
648 			     * for loop will break out.  If no error and this
649 			     * is not the block we want, we throw away the
650 			     * block and go for the next one via the for loop.
651 			     */
652 			    if (error || i < lbn)
653 				    brelse(bp);
654 			}
655 		    }
656 		    /*
657 		     * The above while is repeated if we hit another cookie
658 		     * error.  If we hit an error and it wasn't a cookie error,
659 		     * we give up.
660 		     */
661 		    if (error)
662 			    goto out;
663 		}
664 
665 		/*
666 		 * If not eof and read aheads are enabled, start one.
667 		 * (You need the current block first, so that you have the
668 		 *  directory offset cookie of the next block.)
669 		 */
670 		if (nmp->nm_readahead > 0 &&
671 		    (bp->b_flags & B_INVAL) == 0 &&
672 		    (np->n_direofoffset == 0 ||
673 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
674 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
675 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
676 			if (rabp) {
677 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
678 				rabp->b_flags |= B_ASYNC;
679 				rabp->b_iocmd = BIO_READ;
680 				vfs_busy_pages(rabp, 0);
681 				if (ncl_asyncio(nmp, rabp, cred, td)) {
682 				    rabp->b_flags |= B_INVAL;
683 				    rabp->b_ioflags |= BIO_ERROR;
684 				    vfs_unbusy_pages(rabp);
685 				    brelse(rabp);
686 				}
687 			    } else {
688 				brelse(rabp);
689 			    }
690 			}
691 		}
692 		/*
693 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
694 		 * chopped for the EOF condition, we cannot tell how large
695 		 * NFS directories are going to be until we hit EOF.  So
696 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
697 		 * it just so happens that b_resid will effectively chop it
698 		 * to EOF.  *BUT* this information is lost if the buffer goes
699 		 * away and is reconstituted into a B_CACHE state ( due to
700 		 * being VMIO ) later.  So we keep track of the directory eof
701 		 * in np->n_direofoffset and chop it off as an extra step
702 		 * right here.
703 		 */
704 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
705 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
706 			n = np->n_direofoffset - uio->uio_offset;
707 		break;
708 	    default:
709 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
710 		bp = NULL;
711 		break;
712 	    }
713 
714 	    if (n > 0) {
715 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
716 	    }
717 	    if (vp->v_type == VLNK)
718 		n = 0;
719 	    if (bp != NULL)
720 		brelse(bp);
721 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
722 out:
723 	curthread_pflags2_restore(save2);
724 	if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
725 		NFSLOCKNODE(np);
726 		ncl_pager_setsize(vp, NULL);
727 	}
728 	return (error);
729 }
730 
731 /*
732  * The NFS write path cannot handle iovecs with len > 1. So we need to
733  * break up iovecs accordingly (restricting them to wsize).
734  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
735  * For the ASYNC case, 2 copies are needed. The first a copy from the
736  * user buffer to a staging buffer and then a second copy from the staging
737  * buffer to mbufs. This can be optimized by copying from the user buffer
738  * directly into mbufs and passing the chain down, but that requires a
739  * fair amount of re-working of the relevant codepaths (and can be done
740  * later).
741  */
742 static int
743 nfs_directio_write(vp, uiop, cred, ioflag)
744 	struct vnode *vp;
745 	struct uio *uiop;
746 	struct ucred *cred;
747 	int ioflag;
748 {
749 	int error;
750 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
751 	struct thread *td = uiop->uio_td;
752 	int size;
753 	int wsize;
754 
755 	mtx_lock(&nmp->nm_mtx);
756 	wsize = nmp->nm_wsize;
757 	mtx_unlock(&nmp->nm_mtx);
758 	if (ioflag & IO_SYNC) {
759 		int iomode, must_commit;
760 		struct uio uio;
761 		struct iovec iov;
762 do_sync:
763 		while (uiop->uio_resid > 0) {
764 			size = MIN(uiop->uio_resid, wsize);
765 			size = MIN(uiop->uio_iov->iov_len, size);
766 			iov.iov_base = uiop->uio_iov->iov_base;
767 			iov.iov_len = size;
768 			uio.uio_iov = &iov;
769 			uio.uio_iovcnt = 1;
770 			uio.uio_offset = uiop->uio_offset;
771 			uio.uio_resid = size;
772 			uio.uio_segflg = UIO_USERSPACE;
773 			uio.uio_rw = UIO_WRITE;
774 			uio.uio_td = td;
775 			iomode = NFSWRITE_FILESYNC;
776 			error = ncl_writerpc(vp, &uio, cred, &iomode,
777 			    &must_commit, 0);
778 			KASSERT((must_commit == 0),
779 				("ncl_directio_write: Did not commit write"));
780 			if (error)
781 				return (error);
782 			uiop->uio_offset += size;
783 			uiop->uio_resid -= size;
784 			if (uiop->uio_iov->iov_len <= size) {
785 				uiop->uio_iovcnt--;
786 				uiop->uio_iov++;
787 			} else {
788 				uiop->uio_iov->iov_base =
789 					(char *)uiop->uio_iov->iov_base + size;
790 				uiop->uio_iov->iov_len -= size;
791 			}
792 		}
793 	} else {
794 		struct uio *t_uio;
795 		struct iovec *t_iov;
796 		struct buf *bp;
797 
798 		/*
799 		 * Break up the write into blocksize chunks and hand these
800 		 * over to nfsiod's for write back.
801 		 * Unfortunately, this incurs a copy of the data. Since
802 		 * the user could modify the buffer before the write is
803 		 * initiated.
804 		 *
805 		 * The obvious optimization here is that one of the 2 copies
806 		 * in the async write path can be eliminated by copying the
807 		 * data here directly into mbufs and passing the mbuf chain
808 		 * down. But that will require a fair amount of re-working
809 		 * of the code and can be done if there's enough interest
810 		 * in NFS directio access.
811 		 */
812 		while (uiop->uio_resid > 0) {
813 			size = MIN(uiop->uio_resid, wsize);
814 			size = MIN(uiop->uio_iov->iov_len, size);
815 			bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
816 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
817 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
818 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
819 			t_iov->iov_len = size;
820 			t_uio->uio_iov = t_iov;
821 			t_uio->uio_iovcnt = 1;
822 			t_uio->uio_offset = uiop->uio_offset;
823 			t_uio->uio_resid = size;
824 			t_uio->uio_segflg = UIO_SYSSPACE;
825 			t_uio->uio_rw = UIO_WRITE;
826 			t_uio->uio_td = td;
827 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
828 			    uiop->uio_segflg == UIO_SYSSPACE,
829 			    ("nfs_directio_write: Bad uio_segflg"));
830 			if (uiop->uio_segflg == UIO_USERSPACE) {
831 				error = copyin(uiop->uio_iov->iov_base,
832 				    t_iov->iov_base, size);
833 				if (error != 0)
834 					goto err_free;
835 			} else
836 				/*
837 				 * UIO_SYSSPACE may never happen, but handle
838 				 * it just in case it does.
839 				 */
840 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
841 				    size);
842 			bp->b_flags |= B_DIRECT;
843 			bp->b_iocmd = BIO_WRITE;
844 			if (cred != NOCRED) {
845 				crhold(cred);
846 				bp->b_wcred = cred;
847 			} else
848 				bp->b_wcred = NOCRED;
849 			bp->b_caller1 = (void *)t_uio;
850 			bp->b_vp = vp;
851 			error = ncl_asyncio(nmp, bp, NOCRED, td);
852 err_free:
853 			if (error) {
854 				free(t_iov->iov_base, M_NFSDIRECTIO);
855 				free(t_iov, M_NFSDIRECTIO);
856 				free(t_uio, M_NFSDIRECTIO);
857 				bp->b_vp = NULL;
858 				uma_zfree(ncl_pbuf_zone, bp);
859 				if (error == EINTR)
860 					return (error);
861 				goto do_sync;
862 			}
863 			uiop->uio_offset += size;
864 			uiop->uio_resid -= size;
865 			if (uiop->uio_iov->iov_len <= size) {
866 				uiop->uio_iovcnt--;
867 				uiop->uio_iov++;
868 			} else {
869 				uiop->uio_iov->iov_base =
870 					(char *)uiop->uio_iov->iov_base + size;
871 				uiop->uio_iov->iov_len -= size;
872 			}
873 		}
874 	}
875 	return (0);
876 }
877 
878 /*
879  * Vnode op for write using bio
880  */
881 int
882 ncl_write(struct vop_write_args *ap)
883 {
884 	int biosize;
885 	struct uio *uio = ap->a_uio;
886 	struct thread *td = uio->uio_td;
887 	struct vnode *vp = ap->a_vp;
888 	struct nfsnode *np = VTONFS(vp);
889 	struct ucred *cred = ap->a_cred;
890 	int ioflag = ap->a_ioflag;
891 	struct buf *bp;
892 	struct vattr vattr;
893 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
894 	daddr_t lbn;
895 	int bcount, noncontig_write, obcount;
896 	int bp_cached, n, on, error = 0, error1, wouldcommit;
897 	size_t orig_resid, local_resid;
898 	off_t orig_size, tmp_off;
899 
900 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
901 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
902 	    ("ncl_write proc"));
903 	if (vp->v_type != VREG)
904 		return (EIO);
905 	NFSLOCKNODE(np);
906 	if (np->n_flag & NWRITEERR) {
907 		np->n_flag &= ~NWRITEERR;
908 		NFSUNLOCKNODE(np);
909 		return (np->n_error);
910 	} else
911 		NFSUNLOCKNODE(np);
912 	mtx_lock(&nmp->nm_mtx);
913 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
914 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
915 		mtx_unlock(&nmp->nm_mtx);
916 		(void)ncl_fsinfo(nmp, vp, cred, td);
917 		mtx_lock(&nmp->nm_mtx);
918 	}
919 	if (nmp->nm_wsize == 0)
920 		(void) newnfs_iosize(nmp);
921 	mtx_unlock(&nmp->nm_mtx);
922 
923 	/*
924 	 * Synchronously flush pending buffers if we are in synchronous
925 	 * mode or if we are appending.
926 	 */
927 	if (ioflag & (IO_APPEND | IO_SYNC)) {
928 		NFSLOCKNODE(np);
929 		if (np->n_flag & NMODIFIED) {
930 			NFSUNLOCKNODE(np);
931 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
932 			/*
933 			 * Require non-blocking, synchronous writes to
934 			 * dirty files to inform the program it needs
935 			 * to fsync(2) explicitly.
936 			 */
937 			if (ioflag & IO_NDELAY)
938 				return (EAGAIN);
939 #endif
940 			np->n_attrstamp = 0;
941 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
942 			error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
943 			    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
944 			if (error != 0)
945 				return (error);
946 		} else
947 			NFSUNLOCKNODE(np);
948 	}
949 
950 	orig_resid = uio->uio_resid;
951 	NFSLOCKNODE(np);
952 	orig_size = np->n_size;
953 	NFSUNLOCKNODE(np);
954 
955 	/*
956 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
957 	 * get the append lock.
958 	 */
959 	if (ioflag & IO_APPEND) {
960 		np->n_attrstamp = 0;
961 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
962 		error = VOP_GETATTR(vp, &vattr, cred);
963 		if (error)
964 			return (error);
965 		NFSLOCKNODE(np);
966 		uio->uio_offset = np->n_size;
967 		NFSUNLOCKNODE(np);
968 	}
969 
970 	if (uio->uio_offset < 0)
971 		return (EINVAL);
972 	tmp_off = uio->uio_offset + uio->uio_resid;
973 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
974 		return (EFBIG);
975 	if (uio->uio_resid == 0)
976 		return (0);
977 
978 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
979 		return nfs_directio_write(vp, uio, cred, ioflag);
980 
981 	/*
982 	 * Maybe this should be above the vnode op call, but so long as
983 	 * file servers have no limits, i don't think it matters
984 	 */
985 	if (vn_rlimit_fsize(vp, uio, td))
986 		return (EFBIG);
987 
988 	biosize = vp->v_bufobj.bo_bsize;
989 	/*
990 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
991 	 * would exceed the local maximum per-file write commit size when
992 	 * combined with those, we must decide whether to flush,
993 	 * go synchronous, or return error.  We don't bother checking
994 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
995 	 * no point optimizing for something that really won't ever happen.
996 	 */
997 	wouldcommit = 0;
998 	if (!(ioflag & IO_SYNC)) {
999 		int nflag;
1000 
1001 		NFSLOCKNODE(np);
1002 		nflag = np->n_flag;
1003 		NFSUNLOCKNODE(np);
1004 		if (nflag & NMODIFIED) {
1005 			BO_LOCK(&vp->v_bufobj);
1006 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1007 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1008 				    b_bobufs) {
1009 					if (bp->b_flags & B_NEEDCOMMIT)
1010 						wouldcommit += bp->b_bcount;
1011 				}
1012 			}
1013 			BO_UNLOCK(&vp->v_bufobj);
1014 		}
1015 	}
1016 
1017 	do {
1018 		if (!(ioflag & IO_SYNC)) {
1019 			wouldcommit += biosize;
1020 			if (wouldcommit > nmp->nm_wcommitsize) {
1021 				np->n_attrstamp = 0;
1022 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1023 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1024 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1025 				if (error != 0)
1026 					return (error);
1027 				wouldcommit = biosize;
1028 			}
1029 		}
1030 
1031 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1032 		lbn = uio->uio_offset / biosize;
1033 		on = uio->uio_offset - (lbn * biosize);
1034 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1035 again:
1036 		/*
1037 		 * Handle direct append and file extension cases, calculate
1038 		 * unaligned buffer size.
1039 		 */
1040 		NFSLOCKNODE(np);
1041 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1042 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1043 			noncontig_write = 1;
1044 		else
1045 			noncontig_write = 0;
1046 		if ((uio->uio_offset == np->n_size ||
1047 		    (noncontig_write != 0 &&
1048 		    lbn == (np->n_size / biosize) &&
1049 		    uio->uio_offset + n > np->n_size)) && n) {
1050 			NFSUNLOCKNODE(np);
1051 			/*
1052 			 * Get the buffer (in its pre-append state to maintain
1053 			 * B_CACHE if it was previously set).  Resize the
1054 			 * nfsnode after we have locked the buffer to prevent
1055 			 * readers from reading garbage.
1056 			 */
1057 			obcount = np->n_size - (lbn * biosize);
1058 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1059 
1060 			if (bp != NULL) {
1061 				long save;
1062 
1063 				NFSLOCKNODE(np);
1064 				np->n_size = uio->uio_offset + n;
1065 				np->n_flag |= NMODIFIED;
1066 				vnode_pager_setsize(vp, np->n_size);
1067 				NFSUNLOCKNODE(np);
1068 
1069 				save = bp->b_flags & B_CACHE;
1070 				bcount = on + n;
1071 				allocbuf(bp, bcount);
1072 				bp->b_flags |= save;
1073 				if (noncontig_write != 0 && on > obcount)
1074 					vfs_bio_bzero_buf(bp, obcount, on -
1075 					    obcount);
1076 			}
1077 		} else {
1078 			/*
1079 			 * Obtain the locked cache block first, and then
1080 			 * adjust the file's size as appropriate.
1081 			 */
1082 			bcount = on + n;
1083 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1084 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1085 					bcount = biosize;
1086 				else
1087 					bcount = np->n_size - (off_t)lbn * biosize;
1088 			}
1089 			NFSUNLOCKNODE(np);
1090 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1091 			NFSLOCKNODE(np);
1092 			if (uio->uio_offset + n > np->n_size) {
1093 				np->n_size = uio->uio_offset + n;
1094 				np->n_flag |= NMODIFIED;
1095 				vnode_pager_setsize(vp, np->n_size);
1096 			}
1097 			NFSUNLOCKNODE(np);
1098 		}
1099 
1100 		if (!bp) {
1101 			error = newnfs_sigintr(nmp, td);
1102 			if (!error)
1103 				error = EINTR;
1104 			break;
1105 		}
1106 
1107 		/*
1108 		 * Issue a READ if B_CACHE is not set.  In special-append
1109 		 * mode, B_CACHE is based on the buffer prior to the write
1110 		 * op and is typically set, avoiding the read.  If a read
1111 		 * is required in special append mode, the server will
1112 		 * probably send us a short-read since we extended the file
1113 		 * on our end, resulting in b_resid == 0 and, thusly,
1114 		 * B_CACHE getting set.
1115 		 *
1116 		 * We can also avoid issuing the read if the write covers
1117 		 * the entire buffer.  We have to make sure the buffer state
1118 		 * is reasonable in this case since we will not be initiating
1119 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1120 		 * more information.
1121 		 *
1122 		 * B_CACHE may also be set due to the buffer being cached
1123 		 * normally.
1124 		 */
1125 
1126 		bp_cached = 1;
1127 		if (on == 0 && n == bcount) {
1128 			if ((bp->b_flags & B_CACHE) == 0)
1129 				bp_cached = 0;
1130 			bp->b_flags |= B_CACHE;
1131 			bp->b_flags &= ~B_INVAL;
1132 			bp->b_ioflags &= ~BIO_ERROR;
1133 		}
1134 
1135 		if ((bp->b_flags & B_CACHE) == 0) {
1136 			bp->b_iocmd = BIO_READ;
1137 			vfs_busy_pages(bp, 0);
1138 			error = ncl_doio(vp, bp, cred, td, 0);
1139 			if (error) {
1140 				brelse(bp);
1141 				break;
1142 			}
1143 		}
1144 		if (bp->b_wcred == NOCRED)
1145 			bp->b_wcred = crhold(cred);
1146 		NFSLOCKNODE(np);
1147 		np->n_flag |= NMODIFIED;
1148 		NFSUNLOCKNODE(np);
1149 
1150 		/*
1151 		 * If dirtyend exceeds file size, chop it down.  This should
1152 		 * not normally occur but there is an append race where it
1153 		 * might occur XXX, so we log it.
1154 		 *
1155 		 * If the chopping creates a reverse-indexed or degenerate
1156 		 * situation with dirtyoff/end, we 0 both of them.
1157 		 */
1158 
1159 		if (bp->b_dirtyend > bcount) {
1160 			printf("NFS append race @%lx:%d\n",
1161 			    (long)bp->b_blkno * DEV_BSIZE,
1162 			    bp->b_dirtyend - bcount);
1163 			bp->b_dirtyend = bcount;
1164 		}
1165 
1166 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1167 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1168 
1169 		/*
1170 		 * If the new write will leave a contiguous dirty
1171 		 * area, just update the b_dirtyoff and b_dirtyend,
1172 		 * otherwise force a write rpc of the old dirty area.
1173 		 *
1174 		 * If there has been a file lock applied to this file
1175 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1176 		 * While it is possible to merge discontiguous writes due to
1177 		 * our having a B_CACHE buffer ( and thus valid read data
1178 		 * for the hole), we don't because it could lead to
1179 		 * significant cache coherency problems with multiple clients,
1180 		 * especially if locking is implemented later on.
1181 		 *
1182 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1183 		 * not been file locking done on this file:
1184 		 * Relax coherency a bit for the sake of performance and
1185 		 * expand the current dirty region to contain the new
1186 		 * write even if it means we mark some non-dirty data as
1187 		 * dirty.
1188 		 */
1189 
1190 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1191 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1192 			if (bwrite(bp) == EINTR) {
1193 				error = EINTR;
1194 				break;
1195 			}
1196 			goto again;
1197 		}
1198 
1199 		local_resid = uio->uio_resid;
1200 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1201 
1202 		if (error != 0 && !bp_cached) {
1203 			/*
1204 			 * This block has no other content then what
1205 			 * possibly was written by the faulty uiomove.
1206 			 * Release it, forgetting the data pages, to
1207 			 * prevent the leak of uninitialized data to
1208 			 * usermode.
1209 			 */
1210 			bp->b_ioflags |= BIO_ERROR;
1211 			brelse(bp);
1212 			uio->uio_offset -= local_resid - uio->uio_resid;
1213 			uio->uio_resid = local_resid;
1214 			break;
1215 		}
1216 
1217 		/*
1218 		 * Since this block is being modified, it must be written
1219 		 * again and not just committed.  Since write clustering does
1220 		 * not work for the stage 1 data write, only the stage 2
1221 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1222 		 */
1223 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1224 
1225 		/*
1226 		 * Get the partial update on the progress made from
1227 		 * uiomove, if an error occurred.
1228 		 */
1229 		if (error != 0)
1230 			n = local_resid - uio->uio_resid;
1231 
1232 		/*
1233 		 * Only update dirtyoff/dirtyend if not a degenerate
1234 		 * condition.
1235 		 */
1236 		if (n > 0) {
1237 			if (bp->b_dirtyend > 0) {
1238 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1239 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1240 			} else {
1241 				bp->b_dirtyoff = on;
1242 				bp->b_dirtyend = on + n;
1243 			}
1244 			vfs_bio_set_valid(bp, on, n);
1245 		}
1246 
1247 		/*
1248 		 * If IO_SYNC do bwrite().
1249 		 *
1250 		 * IO_INVAL appears to be unused.  The idea appears to be
1251 		 * to turn off caching in this case.  Very odd.  XXX
1252 		 */
1253 		if ((ioflag & IO_SYNC)) {
1254 			if (ioflag & IO_INVAL)
1255 				bp->b_flags |= B_NOCACHE;
1256 			error1 = bwrite(bp);
1257 			if (error1 != 0) {
1258 				if (error == 0)
1259 					error = error1;
1260 				break;
1261 			}
1262 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1263 			bp->b_flags |= B_ASYNC;
1264 			(void) ncl_writebp(bp, 0, NULL);
1265 		} else {
1266 			bdwrite(bp);
1267 		}
1268 
1269 		if (error != 0)
1270 			break;
1271 	} while (uio->uio_resid > 0 && n > 0);
1272 
1273 	if (error != 0) {
1274 		if (ioflag & IO_UNIT) {
1275 			VATTR_NULL(&vattr);
1276 			vattr.va_size = orig_size;
1277 			/* IO_SYNC is handled implicitely */
1278 			(void)VOP_SETATTR(vp, &vattr, cred);
1279 			uio->uio_offset -= orig_resid - uio->uio_resid;
1280 			uio->uio_resid = orig_resid;
1281 		}
1282 	}
1283 
1284 	return (error);
1285 }
1286 
1287 /*
1288  * Get an nfs cache block.
1289  *
1290  * Allocate a new one if the block isn't currently in the cache
1291  * and return the block marked busy. If the calling process is
1292  * interrupted by a signal for an interruptible mount point, return
1293  * NULL.
1294  *
1295  * The caller must carefully deal with the possible B_INVAL state of
1296  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1297  * indirectly), so synchronous reads can be issued without worrying about
1298  * the B_INVAL state.  We have to be a little more careful when dealing
1299  * with writes (see comments in nfs_write()) when extending a file past
1300  * its EOF.
1301  */
1302 static struct buf *
1303 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1304 {
1305 	struct buf *bp;
1306 	struct mount *mp;
1307 	struct nfsmount *nmp;
1308 
1309 	mp = vp->v_mount;
1310 	nmp = VFSTONFS(mp);
1311 
1312 	if (nmp->nm_flag & NFSMNT_INT) {
1313 		sigset_t oldset;
1314 
1315 		newnfs_set_sigmask(td, &oldset);
1316 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1317 		newnfs_restore_sigmask(td, &oldset);
1318 		while (bp == NULL) {
1319 			if (newnfs_sigintr(nmp, td))
1320 				return (NULL);
1321 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1322 		}
1323 	} else {
1324 		bp = getblk(vp, bn, size, 0, 0, 0);
1325 	}
1326 
1327 	if (vp->v_type == VREG)
1328 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1329 	return (bp);
1330 }
1331 
1332 /*
1333  * Flush and invalidate all dirty buffers. If another process is already
1334  * doing the flush, just wait for completion.
1335  */
1336 int
1337 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1338 {
1339 	struct nfsnode *np = VTONFS(vp);
1340 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1341 	int error = 0, slpflag, slptimeo;
1342 	bool old_lock;
1343 
1344 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1345 
1346 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1347 		intrflg = 0;
1348 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1349 		intrflg = 1;
1350 	if (intrflg) {
1351 		slpflag = PCATCH;
1352 		slptimeo = 2 * hz;
1353 	} else {
1354 		slpflag = 0;
1355 		slptimeo = 0;
1356 	}
1357 
1358 	old_lock = ncl_excl_start(vp);
1359 	if (old_lock)
1360 		flags |= V_ALLOWCLEAN;
1361 
1362 	/*
1363 	 * Now, flush as required.
1364 	 */
1365 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1366 	     vp->v_bufobj.bo_object != NULL) {
1367 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1368 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1369 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1370 		/*
1371 		 * If the page clean was interrupted, fail the invalidation.
1372 		 * Not doing so, we run the risk of losing dirty pages in the
1373 		 * vinvalbuf() call below.
1374 		 */
1375 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1376 			goto out;
1377 	}
1378 
1379 	error = vinvalbuf(vp, flags, slpflag, 0);
1380 	while (error) {
1381 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1382 			goto out;
1383 		error = vinvalbuf(vp, flags, 0, slptimeo);
1384 	}
1385 	if (NFSHASPNFS(nmp)) {
1386 		nfscl_layoutcommit(vp, td);
1387 		/*
1388 		 * Invalidate the attribute cache, since writes to a DS
1389 		 * won't update the size attribute.
1390 		 */
1391 		NFSLOCKNODE(np);
1392 		np->n_attrstamp = 0;
1393 	} else
1394 		NFSLOCKNODE(np);
1395 	if (np->n_directio_asyncwr == 0)
1396 		np->n_flag &= ~NMODIFIED;
1397 	NFSUNLOCKNODE(np);
1398 out:
1399 	ncl_excl_finish(vp, old_lock);
1400 	return error;
1401 }
1402 
1403 /*
1404  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1405  * This is mainly to avoid queueing async I/O requests when the nfsiods
1406  * are all hung on a dead server.
1407  *
1408  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1409  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1410  */
1411 int
1412 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1413 {
1414 	int iod;
1415 	int gotiod;
1416 	int slpflag = 0;
1417 	int slptimeo = 0;
1418 	int error, error2;
1419 
1420 	/*
1421 	 * Commits are usually short and sweet so lets save some cpu and
1422 	 * leave the async daemons for more important rpc's (such as reads
1423 	 * and writes).
1424 	 *
1425 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1426 	 * in the directory in order to update attributes. This can deadlock
1427 	 * with another thread that is waiting for async I/O to be done by
1428 	 * an nfsiod thread while holding a lock on one of these vnodes.
1429 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1430 	 * perform Readdirplus RPCs.
1431 	 */
1432 	NFSLOCKIOD();
1433 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1434 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1435 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1436 		NFSUNLOCKIOD();
1437 		return(EIO);
1438 	}
1439 again:
1440 	if (nmp->nm_flag & NFSMNT_INT)
1441 		slpflag = PCATCH;
1442 	gotiod = FALSE;
1443 
1444 	/*
1445 	 * Find a free iod to process this request.
1446 	 */
1447 	for (iod = 0; iod < ncl_numasync; iod++)
1448 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1449 			gotiod = TRUE;
1450 			break;
1451 		}
1452 
1453 	/*
1454 	 * Try to create one if none are free.
1455 	 */
1456 	if (!gotiod)
1457 		ncl_nfsiodnew();
1458 	else {
1459 		/*
1460 		 * Found one, so wake it up and tell it which
1461 		 * mount to process.
1462 		 */
1463 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1464 		    iod, nmp));
1465 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1466 		ncl_iodmount[iod] = nmp;
1467 		nmp->nm_bufqiods++;
1468 		wakeup(&ncl_iodwant[iod]);
1469 	}
1470 
1471 	/*
1472 	 * If none are free, we may already have an iod working on this mount
1473 	 * point.  If so, it will process our request.
1474 	 */
1475 	if (!gotiod) {
1476 		if (nmp->nm_bufqiods > 0) {
1477 			NFS_DPF(ASYNCIO,
1478 				("ncl_asyncio: %d iods are already processing mount %p\n",
1479 				 nmp->nm_bufqiods, nmp));
1480 			gotiod = TRUE;
1481 		}
1482 	}
1483 
1484 	/*
1485 	 * If we have an iod which can process the request, then queue
1486 	 * the buffer.
1487 	 */
1488 	if (gotiod) {
1489 		/*
1490 		 * Ensure that the queue never grows too large.  We still want
1491 		 * to asynchronize so we block rather then return EIO.
1492 		 */
1493 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1494 			NFS_DPF(ASYNCIO,
1495 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1496 			nmp->nm_bufqwant = TRUE;
1497 			error = newnfs_msleep(td, &nmp->nm_bufq,
1498 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1499 			   slptimeo);
1500 			if (error) {
1501 				error2 = newnfs_sigintr(nmp, td);
1502 				if (error2) {
1503 					NFSUNLOCKIOD();
1504 					return (error2);
1505 				}
1506 				if (slpflag == PCATCH) {
1507 					slpflag = 0;
1508 					slptimeo = 2 * hz;
1509 				}
1510 			}
1511 			/*
1512 			 * We might have lost our iod while sleeping,
1513 			 * so check and loop if necessary.
1514 			 */
1515 			goto again;
1516 		}
1517 
1518 		/* We might have lost our nfsiod */
1519 		if (nmp->nm_bufqiods == 0) {
1520 			NFS_DPF(ASYNCIO,
1521 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1522 			goto again;
1523 		}
1524 
1525 		if (bp->b_iocmd == BIO_READ) {
1526 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1527 				bp->b_rcred = crhold(cred);
1528 		} else {
1529 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1530 				bp->b_wcred = crhold(cred);
1531 		}
1532 
1533 		if (bp->b_flags & B_REMFREE)
1534 			bremfreef(bp);
1535 		BUF_KERNPROC(bp);
1536 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1537 		nmp->nm_bufqlen++;
1538 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1539 			NFSLOCKNODE(VTONFS(bp->b_vp));
1540 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1541 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1542 			NFSUNLOCKNODE(VTONFS(bp->b_vp));
1543 		}
1544 		NFSUNLOCKIOD();
1545 		return (0);
1546 	}
1547 
1548 	NFSUNLOCKIOD();
1549 
1550 	/*
1551 	 * All the iods are busy on other mounts, so return EIO to
1552 	 * force the caller to process the i/o synchronously.
1553 	 */
1554 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1555 	return (EIO);
1556 }
1557 
1558 void
1559 ncl_doio_directwrite(struct buf *bp)
1560 {
1561 	int iomode, must_commit;
1562 	struct uio *uiop = (struct uio *)bp->b_caller1;
1563 	char *iov_base = uiop->uio_iov->iov_base;
1564 
1565 	iomode = NFSWRITE_FILESYNC;
1566 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1567 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1568 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1569 	free(iov_base, M_NFSDIRECTIO);
1570 	free(uiop->uio_iov, M_NFSDIRECTIO);
1571 	free(uiop, M_NFSDIRECTIO);
1572 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1573 		struct nfsnode *np = VTONFS(bp->b_vp);
1574 		NFSLOCKNODE(np);
1575 		if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1576 			/*
1577 			 * Invalidate the attribute cache, since writes to a DS
1578 			 * won't update the size attribute.
1579 			 */
1580 			np->n_attrstamp = 0;
1581 		}
1582 		np->n_directio_asyncwr--;
1583 		if (np->n_directio_asyncwr == 0) {
1584 			np->n_flag &= ~NMODIFIED;
1585 			if ((np->n_flag & NFSYNCWAIT)) {
1586 				np->n_flag &= ~NFSYNCWAIT;
1587 				wakeup((caddr_t)&np->n_directio_asyncwr);
1588 			}
1589 		}
1590 		NFSUNLOCKNODE(np);
1591 	}
1592 	bp->b_vp = NULL;
1593 	uma_zfree(ncl_pbuf_zone, bp);
1594 }
1595 
1596 /*
1597  * Do an I/O operation to/from a cache block. This may be called
1598  * synchronously or from an nfsiod.
1599  */
1600 int
1601 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1602     int called_from_strategy)
1603 {
1604 	struct uio *uiop;
1605 	struct nfsnode *np;
1606 	struct nfsmount *nmp;
1607 	int error = 0, iomode, must_commit = 0;
1608 	struct uio uio;
1609 	struct iovec io;
1610 	struct proc *p = td ? td->td_proc : NULL;
1611 	uint8_t	iocmd;
1612 
1613 	np = VTONFS(vp);
1614 	nmp = VFSTONFS(vp->v_mount);
1615 	uiop = &uio;
1616 	uiop->uio_iov = &io;
1617 	uiop->uio_iovcnt = 1;
1618 	uiop->uio_segflg = UIO_SYSSPACE;
1619 	uiop->uio_td = td;
1620 
1621 	/*
1622 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1623 	 * do this here so we do not have to do it in all the code that
1624 	 * calls us.
1625 	 */
1626 	bp->b_flags &= ~B_INVAL;
1627 	bp->b_ioflags &= ~BIO_ERROR;
1628 
1629 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1630 	iocmd = bp->b_iocmd;
1631 	if (iocmd == BIO_READ) {
1632 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1633 	    io.iov_base = bp->b_data;
1634 	    uiop->uio_rw = UIO_READ;
1635 
1636 	    switch (vp->v_type) {
1637 	    case VREG:
1638 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1639 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1640 		error = ncl_readrpc(vp, uiop, cr);
1641 
1642 		if (!error) {
1643 		    if (uiop->uio_resid) {
1644 			/*
1645 			 * If we had a short read with no error, we must have
1646 			 * hit a file hole.  We should zero-fill the remainder.
1647 			 * This can also occur if the server hits the file EOF.
1648 			 *
1649 			 * Holes used to be able to occur due to pending
1650 			 * writes, but that is not possible any longer.
1651 			 */
1652 			int nread = bp->b_bcount - uiop->uio_resid;
1653 			ssize_t left = uiop->uio_resid;
1654 
1655 			if (left > 0)
1656 				bzero((char *)bp->b_data + nread, left);
1657 			uiop->uio_resid = 0;
1658 		    }
1659 		}
1660 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1661 		if (p && vp->v_writecount <= -1) {
1662 			NFSLOCKNODE(np);
1663 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1664 				NFSUNLOCKNODE(np);
1665 				PROC_LOCK(p);
1666 				killproc(p, "text file modification");
1667 				PROC_UNLOCK(p);
1668 			} else
1669 				NFSUNLOCKNODE(np);
1670 		}
1671 		break;
1672 	    case VLNK:
1673 		uiop->uio_offset = (off_t)0;
1674 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1675 		error = ncl_readlinkrpc(vp, uiop, cr);
1676 		break;
1677 	    case VDIR:
1678 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1679 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1680 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1681 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1682 			if (error == NFSERR_NOTSUPP)
1683 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1684 		}
1685 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1686 			error = ncl_readdirrpc(vp, uiop, cr, td);
1687 		/*
1688 		 * end-of-directory sets B_INVAL but does not generate an
1689 		 * error.
1690 		 */
1691 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1692 			bp->b_flags |= B_INVAL;
1693 		break;
1694 	    default:
1695 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1696 		break;
1697 	    }
1698 	    if (error) {
1699 		bp->b_ioflags |= BIO_ERROR;
1700 		bp->b_error = error;
1701 	    }
1702 	} else {
1703 	    /*
1704 	     * If we only need to commit, try to commit
1705 	     */
1706 	    if (bp->b_flags & B_NEEDCOMMIT) {
1707 		    int retv;
1708 		    off_t off;
1709 
1710 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1711 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1712 			bp->b_wcred, td);
1713 		    if (retv == 0) {
1714 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1715 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1716 			    bp->b_resid = 0;
1717 			    bufdone(bp);
1718 			    return (0);
1719 		    }
1720 		    if (retv == NFSERR_STALEWRITEVERF) {
1721 			    ncl_clearcommit(vp->v_mount);
1722 		    }
1723 	    }
1724 
1725 	    /*
1726 	     * Setup for actual write
1727 	     */
1728 	    NFSLOCKNODE(np);
1729 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1730 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1731 	    NFSUNLOCKNODE(np);
1732 
1733 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1734 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1735 		    - bp->b_dirtyoff;
1736 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1737 		    + bp->b_dirtyoff;
1738 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1739 		uiop->uio_rw = UIO_WRITE;
1740 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1741 
1742 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1743 		    iomode = NFSWRITE_UNSTABLE;
1744 		else
1745 		    iomode = NFSWRITE_FILESYNC;
1746 
1747 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1748 		    called_from_strategy);
1749 
1750 		/*
1751 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1752 		 * to cluster the buffers needing commit.  This will allow
1753 		 * the system to submit a single commit rpc for the whole
1754 		 * cluster.  We can do this even if the buffer is not 100%
1755 		 * dirty (relative to the NFS blocksize), so we optimize the
1756 		 * append-to-file-case.
1757 		 *
1758 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1759 		 * cleared because write clustering only works for commit
1760 		 * rpc's, not for the data portion of the write).
1761 		 */
1762 
1763 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1764 		    bp->b_flags |= B_NEEDCOMMIT;
1765 		    if (bp->b_dirtyoff == 0
1766 			&& bp->b_dirtyend == bp->b_bcount)
1767 			bp->b_flags |= B_CLUSTEROK;
1768 		} else {
1769 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1770 		}
1771 
1772 		/*
1773 		 * For an interrupted write, the buffer is still valid
1774 		 * and the write hasn't been pushed to the server yet,
1775 		 * so we can't set BIO_ERROR and report the interruption
1776 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1777 		 * is not relevant, so the rpc attempt is essentially
1778 		 * a noop.  For the case of a V3 write rpc not being
1779 		 * committed to stable storage, the block is still
1780 		 * dirty and requires either a commit rpc or another
1781 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1782 		 * the block is reused. This is indicated by setting
1783 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1784 		 *
1785 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1786 		 * write error and is handled as above, except that
1787 		 * B_EINTR isn't set. One cause of this is a stale stateid
1788 		 * error for the RPC that indicates recovery is required,
1789 		 * when called with called_from_strategy != 0.
1790 		 *
1791 		 * If the buffer is marked B_PAGING, it does not reside on
1792 		 * the vp's paging queues so we cannot call bdirty().  The
1793 		 * bp in this case is not an NFS cache block so we should
1794 		 * be safe. XXX
1795 		 *
1796 		 * The logic below breaks up errors into recoverable and
1797 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1798 		 * and keep the buffer around for potential write retries.
1799 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1800 		 * and save the error in the nfsnode. This is less than ideal
1801 		 * but necessary. Keeping such buffers around could potentially
1802 		 * cause buffer exhaustion eventually (they can never be written
1803 		 * out, so will get constantly be re-dirtied). It also causes
1804 		 * all sorts of vfs panics. For non-recoverable write errors,
1805 		 * also invalidate the attrcache, so we'll be forced to go over
1806 		 * the wire for this object, returning an error to user on next
1807 		 * call (most of the time).
1808 		 */
1809 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1810 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1811 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1812 			if ((bp->b_flags & B_PAGING) == 0) {
1813 			    bdirty(bp);
1814 			    bp->b_flags &= ~B_DONE;
1815 			}
1816 			if ((error == EINTR || error == ETIMEDOUT) &&
1817 			    (bp->b_flags & B_ASYNC) == 0)
1818 			    bp->b_flags |= B_EINTR;
1819 		} else {
1820 		    if (error) {
1821 			bp->b_ioflags |= BIO_ERROR;
1822 			bp->b_flags |= B_INVAL;
1823 			bp->b_error = np->n_error = error;
1824 			NFSLOCKNODE(np);
1825 			np->n_flag |= NWRITEERR;
1826 			np->n_attrstamp = 0;
1827 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1828 			NFSUNLOCKNODE(np);
1829 		    }
1830 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1831 		}
1832 	    } else {
1833 		bp->b_resid = 0;
1834 		bufdone(bp);
1835 		return (0);
1836 	    }
1837 	}
1838 	bp->b_resid = uiop->uio_resid;
1839 	if (must_commit)
1840 	    ncl_clearcommit(vp->v_mount);
1841 	bufdone(bp);
1842 	return (error);
1843 }
1844 
1845 /*
1846  * Used to aid in handling ftruncate() operations on the NFS client side.
1847  * Truncation creates a number of special problems for NFS.  We have to
1848  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1849  * we have to properly handle VM pages or (potentially dirty) buffers
1850  * that straddle the truncation point.
1851  */
1852 
1853 int
1854 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1855 {
1856 	struct nfsnode *np = VTONFS(vp);
1857 	u_quad_t tsize;
1858 	int biosize = vp->v_bufobj.bo_bsize;
1859 	int error = 0;
1860 
1861 	NFSLOCKNODE(np);
1862 	tsize = np->n_size;
1863 	np->n_size = nsize;
1864 	NFSUNLOCKNODE(np);
1865 
1866 	if (nsize < tsize) {
1867 		struct buf *bp;
1868 		daddr_t lbn;
1869 		int bufsize;
1870 
1871 		/*
1872 		 * vtruncbuf() doesn't get the buffer overlapping the
1873 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1874 		 * buffer that now needs to be truncated.
1875 		 */
1876 		error = vtruncbuf(vp, nsize, biosize);
1877 		lbn = nsize / biosize;
1878 		bufsize = nsize - (lbn * biosize);
1879 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1880 		if (!bp)
1881 			return EINTR;
1882 		if (bp->b_dirtyoff > bp->b_bcount)
1883 			bp->b_dirtyoff = bp->b_bcount;
1884 		if (bp->b_dirtyend > bp->b_bcount)
1885 			bp->b_dirtyend = bp->b_bcount;
1886 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1887 		brelse(bp);
1888 	} else {
1889 		vnode_pager_setsize(vp, nsize);
1890 	}
1891 	return(error);
1892 }
1893 
1894