xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision f7c32ed617858bcd22f8d1b03199099d50125721)
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, long *sz)
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 	*sz = bcount;
115 	return (0);
116 }
117 
118 int
119 ncl_getpages(struct vop_getpages_args *ap)
120 {
121 	int i, error, nextoff, size, toff, count, npages;
122 	struct uio uio;
123 	struct iovec iov;
124 	vm_offset_t kva;
125 	struct buf *bp;
126 	struct vnode *vp;
127 	struct thread *td;
128 	struct ucred *cred;
129 	struct nfsmount *nmp;
130 	vm_object_t object;
131 	vm_page_t *pages;
132 	struct nfsnode *np;
133 
134 	vp = ap->a_vp;
135 	np = VTONFS(vp);
136 	td = curthread;
137 	cred = curthread->td_ucred;
138 	nmp = VFSTONFS(vp->v_mount);
139 	pages = ap->a_m;
140 	npages = ap->a_count;
141 
142 	if ((object = vp->v_object) == NULL) {
143 		printf("ncl_getpages: called with non-merged cache vnode\n");
144 		return (VM_PAGER_ERROR);
145 	}
146 
147 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
148 		NFSLOCKNODE(np);
149 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
150 			NFSUNLOCKNODE(np);
151 			printf("ncl_getpages: called on non-cacheable vnode\n");
152 			return (VM_PAGER_ERROR);
153 		} else
154 			NFSUNLOCKNODE(np);
155 	}
156 
157 	mtx_lock(&nmp->nm_mtx);
158 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
159 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
160 		mtx_unlock(&nmp->nm_mtx);
161 		/* We'll never get here for v4, because we always have fsinfo */
162 		(void)ncl_fsinfo(nmp, vp, cred, td);
163 	} else
164 		mtx_unlock(&nmp->nm_mtx);
165 
166 	if (use_buf_pager)
167 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
168 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
169 
170 	/*
171 	 * If the requested page is partially valid, just return it and
172 	 * allow the pager to zero-out the blanks.  Partially valid pages
173 	 * can only occur at the file EOF.
174 	 *
175 	 * XXXGL: is that true for NFS, where short read can occur???
176 	 */
177 	VM_OBJECT_WLOCK(object);
178 	if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
179 		goto out;
180 	VM_OBJECT_WUNLOCK(object);
181 
182 	/*
183 	 * We use only the kva address for the buffer, but this is extremely
184 	 * convenient and fast.
185 	 */
186 	bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
187 
188 	kva = (vm_offset_t) bp->b_data;
189 	pmap_qenter(kva, pages, npages);
190 	VM_CNT_INC(v_vnodein);
191 	VM_CNT_ADD(v_vnodepgsin, npages);
192 
193 	count = npages << PAGE_SHIFT;
194 	iov.iov_base = (caddr_t) kva;
195 	iov.iov_len = count;
196 	uio.uio_iov = &iov;
197 	uio.uio_iovcnt = 1;
198 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
199 	uio.uio_resid = count;
200 	uio.uio_segflg = UIO_SYSSPACE;
201 	uio.uio_rw = UIO_READ;
202 	uio.uio_td = td;
203 
204 	error = ncl_readrpc(vp, &uio, cred);
205 	pmap_qremove(kva, npages);
206 
207 	uma_zfree(ncl_pbuf_zone, bp);
208 
209 	if (error && (uio.uio_resid == count)) {
210 		printf("ncl_getpages: error %d\n", error);
211 		return (VM_PAGER_ERROR);
212 	}
213 
214 	/*
215 	 * Calculate the number of bytes read and validate only that number
216 	 * of bytes.  Note that due to pending writes, size may be 0.  This
217 	 * does not mean that the remaining data is invalid!
218 	 */
219 
220 	size = count - uio.uio_resid;
221 	VM_OBJECT_WLOCK(object);
222 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
223 		vm_page_t m;
224 		nextoff = toff + PAGE_SIZE;
225 		m = pages[i];
226 
227 		if (nextoff <= size) {
228 			/*
229 			 * Read operation filled an entire page
230 			 */
231 			vm_page_valid(m);
232 			KASSERT(m->dirty == 0,
233 			    ("nfs_getpages: page %p is dirty", m));
234 		} else if (size > toff) {
235 			/*
236 			 * Read operation filled a partial page.
237 			 */
238 			vm_page_invalid(m);
239 			vm_page_set_valid_range(m, 0, size - toff);
240 			KASSERT(m->dirty == 0,
241 			    ("nfs_getpages: page %p is dirty", m));
242 		} else {
243 			/*
244 			 * Read operation was short.  If no error
245 			 * occurred we may have hit a zero-fill
246 			 * section.  We leave valid set to 0, and page
247 			 * is freed by vm_page_readahead_finish() if
248 			 * its index is not equal to requested, or
249 			 * page is zeroed and set valid by
250 			 * vm_pager_get_pages() for requested page.
251 			 */
252 			;
253 		}
254 	}
255 out:
256 	VM_OBJECT_WUNLOCK(object);
257 	if (ap->a_rbehind)
258 		*ap->a_rbehind = 0;
259 	if (ap->a_rahead)
260 		*ap->a_rahead = 0;
261 	return (VM_PAGER_OK);
262 }
263 
264 /*
265  * Vnode op for VM putpages.
266  */
267 int
268 ncl_putpages(struct vop_putpages_args *ap)
269 {
270 	struct uio uio;
271 	struct iovec iov;
272 	int i, error, npages, count;
273 	off_t offset;
274 	int *rtvals;
275 	struct vnode *vp;
276 	struct thread *td;
277 	struct ucred *cred;
278 	struct nfsmount *nmp;
279 	struct nfsnode *np;
280 	vm_page_t *pages;
281 
282 	vp = ap->a_vp;
283 	np = VTONFS(vp);
284 	td = curthread;				/* XXX */
285 	/* Set the cred to n_writecred for the write rpcs. */
286 	if (np->n_writecred != NULL)
287 		cred = crhold(np->n_writecred);
288 	else
289 		cred = crhold(curthread->td_ucred);	/* XXX */
290 	nmp = VFSTONFS(vp->v_mount);
291 	pages = ap->a_m;
292 	count = ap->a_count;
293 	rtvals = ap->a_rtvals;
294 	npages = btoc(count);
295 	offset = IDX_TO_OFF(pages[0]->pindex);
296 
297 	mtx_lock(&nmp->nm_mtx);
298 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
299 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
300 		mtx_unlock(&nmp->nm_mtx);
301 		(void)ncl_fsinfo(nmp, vp, cred, td);
302 	} else
303 		mtx_unlock(&nmp->nm_mtx);
304 
305 	NFSLOCKNODE(np);
306 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
307 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
308 		NFSUNLOCKNODE(np);
309 		printf("ncl_putpages: called on noncache-able vnode\n");
310 		NFSLOCKNODE(np);
311 	}
312 	/*
313 	 * When putting pages, do not extend file past EOF.
314 	 */
315 	if (offset + count > np->n_size) {
316 		count = np->n_size - offset;
317 		if (count < 0)
318 			count = 0;
319 	}
320 	NFSUNLOCKNODE(np);
321 
322 	for (i = 0; i < npages; i++)
323 		rtvals[i] = VM_PAGER_ERROR;
324 
325 	VM_CNT_INC(v_vnodeout);
326 	VM_CNT_ADD(v_vnodepgsout, count);
327 
328 	iov.iov_base = unmapped_buf;
329 	iov.iov_len = count;
330 	uio.uio_iov = &iov;
331 	uio.uio_iovcnt = 1;
332 	uio.uio_offset = offset;
333 	uio.uio_resid = count;
334 	uio.uio_segflg = UIO_NOCOPY;
335 	uio.uio_rw = UIO_WRITE;
336 	uio.uio_td = td;
337 
338 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
339 	    cred);
340 	crfree(cred);
341 
342 	if (error == 0 || !nfs_keep_dirty_on_error) {
343 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
344 		    np->n_size - offset, npages * PAGE_SIZE);
345 	}
346 	return (rtvals[0]);
347 }
348 
349 /*
350  * For nfs, cache consistency can only be maintained approximately.
351  * Although RFC1094 does not specify the criteria, the following is
352  * believed to be compatible with the reference port.
353  * For nfs:
354  * If the file's modify time on the server has changed since the
355  * last read rpc or you have written to the file,
356  * you may have lost data cache consistency with the
357  * server, so flush all of the file's data out of the cache.
358  * Then force a getattr rpc to ensure that you have up to date
359  * attributes.
360  * NB: This implies that cache data can be read when up to
361  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
362  * attributes this could be forced by setting n_attrstamp to 0 before
363  * the VOP_GETATTR() call.
364  */
365 static inline int
366 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
367 {
368 	int error = 0;
369 	struct vattr vattr;
370 	struct nfsnode *np = VTONFS(vp);
371 	bool old_lock;
372 
373 	/*
374 	 * Ensure the exclusove access to the node before checking
375 	 * whether the cache is consistent.
376 	 */
377 	old_lock = ncl_excl_start(vp);
378 	NFSLOCKNODE(np);
379 	if (np->n_flag & NMODIFIED) {
380 		NFSUNLOCKNODE(np);
381 		if (vp->v_type != VREG) {
382 			if (vp->v_type != VDIR)
383 				panic("nfs: bioread, not dir");
384 			ncl_invaldir(vp);
385 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
386 			if (error != 0)
387 				goto out;
388 		}
389 		np->n_attrstamp = 0;
390 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
391 		error = VOP_GETATTR(vp, &vattr, cred);
392 		if (error)
393 			goto out;
394 		NFSLOCKNODE(np);
395 		np->n_mtime = vattr.va_mtime;
396 		NFSUNLOCKNODE(np);
397 	} else {
398 		NFSUNLOCKNODE(np);
399 		error = VOP_GETATTR(vp, &vattr, cred);
400 		if (error)
401 			goto out;
402 		NFSLOCKNODE(np);
403 		if ((np->n_flag & NSIZECHANGED)
404 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
405 			NFSUNLOCKNODE(np);
406 			if (vp->v_type == VDIR)
407 				ncl_invaldir(vp);
408 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
409 			if (error != 0)
410 				goto out;
411 			NFSLOCKNODE(np);
412 			np->n_mtime = vattr.va_mtime;
413 			np->n_flag &= ~NSIZECHANGED;
414 		}
415 		NFSUNLOCKNODE(np);
416 	}
417 out:
418 	ncl_excl_finish(vp, old_lock);
419 	return (error);
420 }
421 
422 /*
423  * Vnode op for read using bio
424  */
425 int
426 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
427 {
428 	struct nfsnode *np = VTONFS(vp);
429 	struct buf *bp, *rabp;
430 	struct thread *td;
431 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
432 	daddr_t lbn, rabn;
433 	int biosize, bcount, error, i, n, nra, on, save2, seqcount;
434 	off_t tmp_off;
435 
436 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
437 	if (uio->uio_resid == 0)
438 		return (0);
439 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
440 		return (EINVAL);
441 	td = uio->uio_td;
442 
443 	mtx_lock(&nmp->nm_mtx);
444 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
445 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
446 		mtx_unlock(&nmp->nm_mtx);
447 		(void)ncl_fsinfo(nmp, vp, cred, td);
448 		mtx_lock(&nmp->nm_mtx);
449 	}
450 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
451 		(void) newnfs_iosize(nmp);
452 
453 	tmp_off = uio->uio_offset + uio->uio_resid;
454 	if (vp->v_type != VDIR &&
455 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
456 		mtx_unlock(&nmp->nm_mtx);
457 		return (EFBIG);
458 	}
459 	mtx_unlock(&nmp->nm_mtx);
460 
461 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
462 		/* No caching/ no readaheads. Just read data into the user buffer */
463 		return ncl_readrpc(vp, uio, cred);
464 
465 	n = 0;
466 	on = 0;
467 	biosize = vp->v_bufobj.bo_bsize;
468 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
469 
470 	error = nfs_bioread_check_cons(vp, td, cred);
471 	if (error)
472 		return error;
473 
474 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
475 	do {
476 	    u_quad_t nsize;
477 
478 	    NFSLOCKNODE(np);
479 	    nsize = np->n_size;
480 	    NFSUNLOCKNODE(np);
481 
482 	    switch (vp->v_type) {
483 	    case VREG:
484 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
485 		lbn = uio->uio_offset / biosize;
486 		on = uio->uio_offset - (lbn * biosize);
487 
488 		/*
489 		 * Start the read ahead(s), as required.
490 		 */
491 		if (nmp->nm_readahead > 0) {
492 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
493 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
494 			rabn = lbn + 1 + nra;
495 			if (incore(&vp->v_bufobj, rabn) == NULL) {
496 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
497 			    if (!rabp) {
498 				error = newnfs_sigintr(nmp, td);
499 				if (error == 0)
500 					error = EINTR;
501 				goto out;
502 			    }
503 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
504 				rabp->b_flags |= B_ASYNC;
505 				rabp->b_iocmd = BIO_READ;
506 				vfs_busy_pages(rabp, 0);
507 				if (ncl_asyncio(nmp, rabp, cred, td)) {
508 				    rabp->b_flags |= B_INVAL;
509 				    rabp->b_ioflags |= BIO_ERROR;
510 				    vfs_unbusy_pages(rabp);
511 				    brelse(rabp);
512 				    break;
513 				}
514 			    } else {
515 				brelse(rabp);
516 			    }
517 			}
518 		    }
519 		}
520 
521 		/* Note that bcount is *not* DEV_BSIZE aligned. */
522 		bcount = biosize;
523 		if ((off_t)lbn * biosize >= nsize) {
524 			bcount = 0;
525 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
526 			bcount = nsize - (off_t)lbn * biosize;
527 		}
528 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
529 
530 		if (!bp) {
531 			error = newnfs_sigintr(nmp, td);
532 			if (error == 0)
533 				error = EINTR;
534 			goto out;
535 		}
536 
537 		/*
538 		 * If B_CACHE is not set, we must issue the read.  If this
539 		 * fails, we return an error.
540 		 */
541 
542 		if ((bp->b_flags & B_CACHE) == 0) {
543 		    bp->b_iocmd = BIO_READ;
544 		    vfs_busy_pages(bp, 0);
545 		    error = ncl_doio(vp, bp, cred, td, 0);
546 		    if (error) {
547 			brelse(bp);
548 			goto out;
549 		    }
550 		}
551 
552 		/*
553 		 * on is the offset into the current bp.  Figure out how many
554 		 * bytes we can copy out of the bp.  Note that bcount is
555 		 * NOT DEV_BSIZE aligned.
556 		 *
557 		 * Then figure out how many bytes we can copy into the uio.
558 		 */
559 
560 		n = 0;
561 		if (on < bcount)
562 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
563 		break;
564 	    case VLNK:
565 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
566 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
567 		if (!bp) {
568 			error = newnfs_sigintr(nmp, td);
569 			if (error == 0)
570 				error = EINTR;
571 			goto out;
572 		}
573 		if ((bp->b_flags & B_CACHE) == 0) {
574 		    bp->b_iocmd = BIO_READ;
575 		    vfs_busy_pages(bp, 0);
576 		    error = ncl_doio(vp, bp, cred, td, 0);
577 		    if (error) {
578 			bp->b_ioflags |= BIO_ERROR;
579 			brelse(bp);
580 			goto out;
581 		    }
582 		}
583 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
584 		on = 0;
585 		break;
586 	    case VDIR:
587 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
588 		NFSLOCKNODE(np);
589 		if (np->n_direofoffset
590 		    && uio->uio_offset >= np->n_direofoffset) {
591 			NFSUNLOCKNODE(np);
592 			error = 0;
593 			goto out;
594 		}
595 		NFSUNLOCKNODE(np);
596 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
597 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
598 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
599 		if (!bp) {
600 			error = newnfs_sigintr(nmp, td);
601 			if (error == 0)
602 				error = EINTR;
603 			goto out;
604 		}
605 		if ((bp->b_flags & B_CACHE) == 0) {
606 		    bp->b_iocmd = BIO_READ;
607 		    vfs_busy_pages(bp, 0);
608 		    error = ncl_doio(vp, bp, cred, td, 0);
609 		    if (error) {
610 			    brelse(bp);
611 		    }
612 		    while (error == NFSERR_BAD_COOKIE) {
613 			ncl_invaldir(vp);
614 			error = ncl_vinvalbuf(vp, 0, td, 1);
615 
616 			/*
617 			 * Yuck! The directory has been modified on the
618 			 * server. The only way to get the block is by
619 			 * reading from the beginning to get all the
620 			 * offset cookies.
621 			 *
622 			 * Leave the last bp intact unless there is an error.
623 			 * Loop back up to the while if the error is another
624 			 * NFSERR_BAD_COOKIE (double yuch!).
625 			 */
626 			for (i = 0; i <= lbn && !error; i++) {
627 			    NFSLOCKNODE(np);
628 			    if (np->n_direofoffset
629 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
630 				    NFSUNLOCKNODE(np);
631 				    error = 0;
632 				    goto out;
633 			    }
634 			    NFSUNLOCKNODE(np);
635 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
636 			    if (!bp) {
637 				error = newnfs_sigintr(nmp, td);
638 				if (error == 0)
639 					error = EINTR;
640 				goto out;
641 			    }
642 			    if ((bp->b_flags & B_CACHE) == 0) {
643 				    bp->b_iocmd = BIO_READ;
644 				    vfs_busy_pages(bp, 0);
645 				    error = ncl_doio(vp, bp, cred, td, 0);
646 				    /*
647 				     * no error + B_INVAL == directory EOF,
648 				     * use the block.
649 				     */
650 				    if (error == 0 && (bp->b_flags & B_INVAL))
651 					    break;
652 			    }
653 			    /*
654 			     * An error will throw away the block and the
655 			     * for loop will break out.  If no error and this
656 			     * is not the block we want, we throw away the
657 			     * block and go for the next one via the for loop.
658 			     */
659 			    if (error || i < lbn)
660 				    brelse(bp);
661 			}
662 		    }
663 		    /*
664 		     * The above while is repeated if we hit another cookie
665 		     * error.  If we hit an error and it wasn't a cookie error,
666 		     * we give up.
667 		     */
668 		    if (error)
669 			    goto out;
670 		}
671 
672 		/*
673 		 * If not eof and read aheads are enabled, start one.
674 		 * (You need the current block first, so that you have the
675 		 *  directory offset cookie of the next block.)
676 		 */
677 		NFSLOCKNODE(np);
678 		if (nmp->nm_readahead > 0 &&
679 		    (bp->b_flags & B_INVAL) == 0 &&
680 		    (np->n_direofoffset == 0 ||
681 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
682 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
683 			NFSUNLOCKNODE(np);
684 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
685 			if (rabp) {
686 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
687 				rabp->b_flags |= B_ASYNC;
688 				rabp->b_iocmd = BIO_READ;
689 				vfs_busy_pages(rabp, 0);
690 				if (ncl_asyncio(nmp, rabp, cred, td)) {
691 				    rabp->b_flags |= B_INVAL;
692 				    rabp->b_ioflags |= BIO_ERROR;
693 				    vfs_unbusy_pages(rabp);
694 				    brelse(rabp);
695 				}
696 			    } else {
697 				brelse(rabp);
698 			    }
699 			}
700 			NFSLOCKNODE(np);
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 		NFSUNLOCKNODE(np);
718 		break;
719 	    default:
720 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
721 		bp = NULL;
722 		break;
723 	    }
724 
725 	    if (n > 0) {
726 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
727 	    }
728 	    if (vp->v_type == VLNK)
729 		n = 0;
730 	    if (bp != NULL)
731 		brelse(bp);
732 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
733 out:
734 	curthread_pflags2_restore(save2);
735 	if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
736 		NFSLOCKNODE(np);
737 		ncl_pager_setsize(vp, NULL);
738 	}
739 	return (error);
740 }
741 
742 /*
743  * The NFS write path cannot handle iovecs with len > 1. So we need to
744  * break up iovecs accordingly (restricting them to wsize).
745  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
746  * For the ASYNC case, 2 copies are needed. The first a copy from the
747  * user buffer to a staging buffer and then a second copy from the staging
748  * buffer to mbufs. This can be optimized by copying from the user buffer
749  * directly into mbufs and passing the chain down, but that requires a
750  * fair amount of re-working of the relevant codepaths (and can be done
751  * later).
752  */
753 static int
754 nfs_directio_write(vp, uiop, cred, ioflag)
755 	struct vnode *vp;
756 	struct uio *uiop;
757 	struct ucred *cred;
758 	int ioflag;
759 {
760 	int error;
761 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
762 	struct thread *td = uiop->uio_td;
763 	int size;
764 	int wsize;
765 
766 	mtx_lock(&nmp->nm_mtx);
767 	wsize = nmp->nm_wsize;
768 	mtx_unlock(&nmp->nm_mtx);
769 	if (ioflag & IO_SYNC) {
770 		int iomode, must_commit;
771 		struct uio uio;
772 		struct iovec iov;
773 do_sync:
774 		while (uiop->uio_resid > 0) {
775 			size = MIN(uiop->uio_resid, wsize);
776 			size = MIN(uiop->uio_iov->iov_len, size);
777 			iov.iov_base = uiop->uio_iov->iov_base;
778 			iov.iov_len = size;
779 			uio.uio_iov = &iov;
780 			uio.uio_iovcnt = 1;
781 			uio.uio_offset = uiop->uio_offset;
782 			uio.uio_resid = size;
783 			uio.uio_segflg = UIO_USERSPACE;
784 			uio.uio_rw = UIO_WRITE;
785 			uio.uio_td = td;
786 			iomode = NFSWRITE_FILESYNC;
787 			error = ncl_writerpc(vp, &uio, cred, &iomode,
788 			    &must_commit, 0);
789 			KASSERT((must_commit == 0),
790 				("ncl_directio_write: Did not commit write"));
791 			if (error)
792 				return (error);
793 			uiop->uio_offset += size;
794 			uiop->uio_resid -= size;
795 			if (uiop->uio_iov->iov_len <= size) {
796 				uiop->uio_iovcnt--;
797 				uiop->uio_iov++;
798 			} else {
799 				uiop->uio_iov->iov_base =
800 					(char *)uiop->uio_iov->iov_base + size;
801 				uiop->uio_iov->iov_len -= size;
802 			}
803 		}
804 	} else {
805 		struct uio *t_uio;
806 		struct iovec *t_iov;
807 		struct buf *bp;
808 
809 		/*
810 		 * Break up the write into blocksize chunks and hand these
811 		 * over to nfsiod's for write back.
812 		 * Unfortunately, this incurs a copy of the data. Since
813 		 * the user could modify the buffer before the write is
814 		 * initiated.
815 		 *
816 		 * The obvious optimization here is that one of the 2 copies
817 		 * in the async write path can be eliminated by copying the
818 		 * data here directly into mbufs and passing the mbuf chain
819 		 * down. But that will require a fair amount of re-working
820 		 * of the code and can be done if there's enough interest
821 		 * in NFS directio access.
822 		 */
823 		while (uiop->uio_resid > 0) {
824 			size = MIN(uiop->uio_resid, wsize);
825 			size = MIN(uiop->uio_iov->iov_len, size);
826 			bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
827 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
828 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
829 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
830 			t_iov->iov_len = size;
831 			t_uio->uio_iov = t_iov;
832 			t_uio->uio_iovcnt = 1;
833 			t_uio->uio_offset = uiop->uio_offset;
834 			t_uio->uio_resid = size;
835 			t_uio->uio_segflg = UIO_SYSSPACE;
836 			t_uio->uio_rw = UIO_WRITE;
837 			t_uio->uio_td = td;
838 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
839 			    uiop->uio_segflg == UIO_SYSSPACE,
840 			    ("nfs_directio_write: Bad uio_segflg"));
841 			if (uiop->uio_segflg == UIO_USERSPACE) {
842 				error = copyin(uiop->uio_iov->iov_base,
843 				    t_iov->iov_base, size);
844 				if (error != 0)
845 					goto err_free;
846 			} else
847 				/*
848 				 * UIO_SYSSPACE may never happen, but handle
849 				 * it just in case it does.
850 				 */
851 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
852 				    size);
853 			bp->b_flags |= B_DIRECT;
854 			bp->b_iocmd = BIO_WRITE;
855 			if (cred != NOCRED) {
856 				crhold(cred);
857 				bp->b_wcred = cred;
858 			} else
859 				bp->b_wcred = NOCRED;
860 			bp->b_caller1 = (void *)t_uio;
861 			bp->b_vp = vp;
862 			error = ncl_asyncio(nmp, bp, NOCRED, td);
863 err_free:
864 			if (error) {
865 				free(t_iov->iov_base, M_NFSDIRECTIO);
866 				free(t_iov, M_NFSDIRECTIO);
867 				free(t_uio, M_NFSDIRECTIO);
868 				bp->b_vp = NULL;
869 				uma_zfree(ncl_pbuf_zone, bp);
870 				if (error == EINTR)
871 					return (error);
872 				goto do_sync;
873 			}
874 			uiop->uio_offset += size;
875 			uiop->uio_resid -= size;
876 			if (uiop->uio_iov->iov_len <= size) {
877 				uiop->uio_iovcnt--;
878 				uiop->uio_iov++;
879 			} else {
880 				uiop->uio_iov->iov_base =
881 					(char *)uiop->uio_iov->iov_base + size;
882 				uiop->uio_iov->iov_len -= size;
883 			}
884 		}
885 	}
886 	return (0);
887 }
888 
889 /*
890  * Vnode op for write using bio
891  */
892 int
893 ncl_write(struct vop_write_args *ap)
894 {
895 	int biosize;
896 	struct uio *uio = ap->a_uio;
897 	struct thread *td = uio->uio_td;
898 	struct vnode *vp = ap->a_vp;
899 	struct nfsnode *np = VTONFS(vp);
900 	struct ucred *cred = ap->a_cred;
901 	int ioflag = ap->a_ioflag;
902 	struct buf *bp;
903 	struct vattr vattr;
904 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
905 	daddr_t lbn;
906 	int bcount, noncontig_write, obcount;
907 	int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
908 	size_t orig_resid, local_resid;
909 	off_t orig_size, tmp_off;
910 	struct timespec ts;
911 
912 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
913 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
914 	    ("ncl_write proc"));
915 	if (vp->v_type != VREG)
916 		return (EIO);
917 	NFSLOCKNODE(np);
918 	if (np->n_flag & NWRITEERR) {
919 		np->n_flag &= ~NWRITEERR;
920 		NFSUNLOCKNODE(np);
921 		return (np->n_error);
922 	} else
923 		NFSUNLOCKNODE(np);
924 	mtx_lock(&nmp->nm_mtx);
925 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
926 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
927 		mtx_unlock(&nmp->nm_mtx);
928 		(void)ncl_fsinfo(nmp, vp, cred, td);
929 		mtx_lock(&nmp->nm_mtx);
930 	}
931 	if (nmp->nm_wsize == 0)
932 		(void) newnfs_iosize(nmp);
933 	mtx_unlock(&nmp->nm_mtx);
934 
935 	/*
936 	 * Synchronously flush pending buffers if we are in synchronous
937 	 * mode or if we are appending.
938 	 */
939 	if (ioflag & (IO_APPEND | IO_SYNC)) {
940 		NFSLOCKNODE(np);
941 		if (np->n_flag & NMODIFIED) {
942 			NFSUNLOCKNODE(np);
943 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
944 			/*
945 			 * Require non-blocking, synchronous writes to
946 			 * dirty files to inform the program it needs
947 			 * to fsync(2) explicitly.
948 			 */
949 			if (ioflag & IO_NDELAY)
950 				return (EAGAIN);
951 #endif
952 			np->n_attrstamp = 0;
953 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
954 			error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
955 			    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
956 			if (error != 0)
957 				return (error);
958 		} else
959 			NFSUNLOCKNODE(np);
960 	}
961 
962 	orig_resid = uio->uio_resid;
963 	NFSLOCKNODE(np);
964 	orig_size = np->n_size;
965 	NFSUNLOCKNODE(np);
966 
967 	/*
968 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
969 	 * get the append lock.
970 	 */
971 	if (ioflag & IO_APPEND) {
972 		np->n_attrstamp = 0;
973 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
974 		error = VOP_GETATTR(vp, &vattr, cred);
975 		if (error)
976 			return (error);
977 		NFSLOCKNODE(np);
978 		uio->uio_offset = np->n_size;
979 		NFSUNLOCKNODE(np);
980 	}
981 
982 	if (uio->uio_offset < 0)
983 		return (EINVAL);
984 	tmp_off = uio->uio_offset + uio->uio_resid;
985 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
986 		return (EFBIG);
987 	if (uio->uio_resid == 0)
988 		return (0);
989 
990 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
991 		return nfs_directio_write(vp, uio, cred, ioflag);
992 
993 	/*
994 	 * Maybe this should be above the vnode op call, but so long as
995 	 * file servers have no limits, i don't think it matters
996 	 */
997 	if (vn_rlimit_fsize(vp, uio, td))
998 		return (EFBIG);
999 
1000 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
1001 	biosize = vp->v_bufobj.bo_bsize;
1002 	/*
1003 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
1004 	 * would exceed the local maximum per-file write commit size when
1005 	 * combined with those, we must decide whether to flush,
1006 	 * go synchronous, or return error.  We don't bother checking
1007 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
1008 	 * no point optimizing for something that really won't ever happen.
1009 	 */
1010 	wouldcommit = 0;
1011 	if (!(ioflag & IO_SYNC)) {
1012 		int nflag;
1013 
1014 		NFSLOCKNODE(np);
1015 		nflag = np->n_flag;
1016 		NFSUNLOCKNODE(np);
1017 		if (nflag & NMODIFIED) {
1018 			BO_LOCK(&vp->v_bufobj);
1019 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1020 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1021 				    b_bobufs) {
1022 					if (bp->b_flags & B_NEEDCOMMIT)
1023 						wouldcommit += bp->b_bcount;
1024 				}
1025 			}
1026 			BO_UNLOCK(&vp->v_bufobj);
1027 		}
1028 	}
1029 
1030 	do {
1031 		if (!(ioflag & IO_SYNC)) {
1032 			wouldcommit += biosize;
1033 			if (wouldcommit > nmp->nm_wcommitsize) {
1034 				np->n_attrstamp = 0;
1035 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1036 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1037 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1038 				if (error != 0)
1039 					goto out;
1040 				wouldcommit = biosize;
1041 			}
1042 		}
1043 
1044 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1045 		lbn = uio->uio_offset / biosize;
1046 		on = uio->uio_offset - (lbn * biosize);
1047 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1048 again:
1049 		/*
1050 		 * Handle direct append and file extension cases, calculate
1051 		 * unaligned buffer size.
1052 		 */
1053 		NFSLOCKNODE(np);
1054 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1055 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1056 			noncontig_write = 1;
1057 		else
1058 			noncontig_write = 0;
1059 		if ((uio->uio_offset == np->n_size ||
1060 		    (noncontig_write != 0 &&
1061 		    lbn == (np->n_size / biosize) &&
1062 		    uio->uio_offset + n > np->n_size)) && n) {
1063 			NFSUNLOCKNODE(np);
1064 			/*
1065 			 * Get the buffer (in its pre-append state to maintain
1066 			 * B_CACHE if it was previously set).  Resize the
1067 			 * nfsnode after we have locked the buffer to prevent
1068 			 * readers from reading garbage.
1069 			 */
1070 			obcount = np->n_size - (lbn * biosize);
1071 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1072 
1073 			if (bp != NULL) {
1074 				long save;
1075 
1076 				NFSLOCKNODE(np);
1077 				np->n_size = uio->uio_offset + n;
1078 				np->n_flag |= NMODIFIED;
1079 				np->n_flag &= ~NVNSETSZSKIP;
1080 				vnode_pager_setsize(vp, np->n_size);
1081 				NFSUNLOCKNODE(np);
1082 
1083 				save = bp->b_flags & B_CACHE;
1084 				bcount = on + n;
1085 				allocbuf(bp, bcount);
1086 				bp->b_flags |= save;
1087 				if (noncontig_write != 0 && on > obcount)
1088 					vfs_bio_bzero_buf(bp, obcount, on -
1089 					    obcount);
1090 			}
1091 		} else {
1092 			/*
1093 			 * Obtain the locked cache block first, and then
1094 			 * adjust the file's size as appropriate.
1095 			 */
1096 			bcount = on + n;
1097 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1098 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1099 					bcount = biosize;
1100 				else
1101 					bcount = np->n_size - (off_t)lbn * biosize;
1102 			}
1103 			NFSUNLOCKNODE(np);
1104 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1105 			NFSLOCKNODE(np);
1106 			if (uio->uio_offset + n > np->n_size) {
1107 				np->n_size = uio->uio_offset + n;
1108 				np->n_flag |= NMODIFIED;
1109 				np->n_flag &= ~NVNSETSZSKIP;
1110 				vnode_pager_setsize(vp, np->n_size);
1111 			}
1112 			NFSUNLOCKNODE(np);
1113 		}
1114 
1115 		if (!bp) {
1116 			error = newnfs_sigintr(nmp, td);
1117 			if (!error)
1118 				error = EINTR;
1119 			break;
1120 		}
1121 
1122 		/*
1123 		 * Issue a READ if B_CACHE is not set.  In special-append
1124 		 * mode, B_CACHE is based on the buffer prior to the write
1125 		 * op and is typically set, avoiding the read.  If a read
1126 		 * is required in special append mode, the server will
1127 		 * probably send us a short-read since we extended the file
1128 		 * on our end, resulting in b_resid == 0 and, thusly,
1129 		 * B_CACHE getting set.
1130 		 *
1131 		 * We can also avoid issuing the read if the write covers
1132 		 * the entire buffer.  We have to make sure the buffer state
1133 		 * is reasonable in this case since we will not be initiating
1134 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1135 		 * more information.
1136 		 *
1137 		 * B_CACHE may also be set due to the buffer being cached
1138 		 * normally.
1139 		 */
1140 
1141 		bp_cached = 1;
1142 		if (on == 0 && n == bcount) {
1143 			if ((bp->b_flags & B_CACHE) == 0)
1144 				bp_cached = 0;
1145 			bp->b_flags |= B_CACHE;
1146 			bp->b_flags &= ~B_INVAL;
1147 			bp->b_ioflags &= ~BIO_ERROR;
1148 		}
1149 
1150 		if ((bp->b_flags & B_CACHE) == 0) {
1151 			bp->b_iocmd = BIO_READ;
1152 			vfs_busy_pages(bp, 0);
1153 			error = ncl_doio(vp, bp, cred, td, 0);
1154 			if (error) {
1155 				brelse(bp);
1156 				break;
1157 			}
1158 		}
1159 		if (bp->b_wcred == NOCRED)
1160 			bp->b_wcred = crhold(cred);
1161 		NFSLOCKNODE(np);
1162 		np->n_flag |= NMODIFIED;
1163 		NFSUNLOCKNODE(np);
1164 
1165 		/*
1166 		 * If dirtyend exceeds file size, chop it down.  This should
1167 		 * not normally occur but there is an append race where it
1168 		 * might occur XXX, so we log it.
1169 		 *
1170 		 * If the chopping creates a reverse-indexed or degenerate
1171 		 * situation with dirtyoff/end, we 0 both of them.
1172 		 */
1173 
1174 		if (bp->b_dirtyend > bcount) {
1175 			printf("NFS append race @%lx:%d\n",
1176 			    (long)bp->b_blkno * DEV_BSIZE,
1177 			    bp->b_dirtyend - bcount);
1178 			bp->b_dirtyend = bcount;
1179 		}
1180 
1181 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1182 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1183 
1184 		/*
1185 		 * If the new write will leave a contiguous dirty
1186 		 * area, just update the b_dirtyoff and b_dirtyend,
1187 		 * otherwise force a write rpc of the old dirty area.
1188 		 *
1189 		 * If there has been a file lock applied to this file
1190 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1191 		 * While it is possible to merge discontiguous writes due to
1192 		 * our having a B_CACHE buffer ( and thus valid read data
1193 		 * for the hole), we don't because it could lead to
1194 		 * significant cache coherency problems with multiple clients,
1195 		 * especially if locking is implemented later on.
1196 		 *
1197 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1198 		 * not been file locking done on this file:
1199 		 * Relax coherency a bit for the sake of performance and
1200 		 * expand the current dirty region to contain the new
1201 		 * write even if it means we mark some non-dirty data as
1202 		 * dirty.
1203 		 */
1204 
1205 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1206 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1207 			if (bwrite(bp) == EINTR) {
1208 				error = EINTR;
1209 				break;
1210 			}
1211 			goto again;
1212 		}
1213 
1214 		local_resid = uio->uio_resid;
1215 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1216 
1217 		if (error != 0 && !bp_cached) {
1218 			/*
1219 			 * This block has no other content then what
1220 			 * possibly was written by the faulty uiomove.
1221 			 * Release it, forgetting the data pages, to
1222 			 * prevent the leak of uninitialized data to
1223 			 * usermode.
1224 			 */
1225 			bp->b_ioflags |= BIO_ERROR;
1226 			brelse(bp);
1227 			uio->uio_offset -= local_resid - uio->uio_resid;
1228 			uio->uio_resid = local_resid;
1229 			break;
1230 		}
1231 
1232 		/*
1233 		 * Since this block is being modified, it must be written
1234 		 * again and not just committed.  Since write clustering does
1235 		 * not work for the stage 1 data write, only the stage 2
1236 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1237 		 */
1238 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1239 
1240 		/*
1241 		 * Get the partial update on the progress made from
1242 		 * uiomove, if an error occurred.
1243 		 */
1244 		if (error != 0)
1245 			n = local_resid - uio->uio_resid;
1246 
1247 		/*
1248 		 * Only update dirtyoff/dirtyend if not a degenerate
1249 		 * condition.
1250 		 */
1251 		if (n > 0) {
1252 			if (bp->b_dirtyend > 0) {
1253 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1254 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1255 			} else {
1256 				bp->b_dirtyoff = on;
1257 				bp->b_dirtyend = on + n;
1258 			}
1259 			vfs_bio_set_valid(bp, on, n);
1260 		}
1261 
1262 		/*
1263 		 * If IO_SYNC do bwrite().
1264 		 *
1265 		 * IO_INVAL appears to be unused.  The idea appears to be
1266 		 * to turn off caching in this case.  Very odd.  XXX
1267 		 */
1268 		if ((ioflag & IO_SYNC)) {
1269 			if (ioflag & IO_INVAL)
1270 				bp->b_flags |= B_NOCACHE;
1271 			error1 = bwrite(bp);
1272 			if (error1 != 0) {
1273 				if (error == 0)
1274 					error = error1;
1275 				break;
1276 			}
1277 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1278 			bp->b_flags |= B_ASYNC;
1279 			(void) ncl_writebp(bp, 0, NULL);
1280 		} else {
1281 			bdwrite(bp);
1282 		}
1283 
1284 		if (error != 0)
1285 			break;
1286 	} while (uio->uio_resid > 0 && n > 0);
1287 
1288 	if (error == 0) {
1289 		nanouptime(&ts);
1290 		NFSLOCKNODE(np);
1291 		np->n_localmodtime = ts;
1292 		NFSUNLOCKNODE(np);
1293 	} else {
1294 		if (ioflag & IO_UNIT) {
1295 			VATTR_NULL(&vattr);
1296 			vattr.va_size = orig_size;
1297 			/* IO_SYNC is handled implicitely */
1298 			(void)VOP_SETATTR(vp, &vattr, cred);
1299 			uio->uio_offset -= orig_resid - uio->uio_resid;
1300 			uio->uio_resid = orig_resid;
1301 		}
1302 	}
1303 
1304 out:
1305 	curthread_pflags2_restore(save2);
1306 	return (error);
1307 }
1308 
1309 /*
1310  * Get an nfs cache block.
1311  *
1312  * Allocate a new one if the block isn't currently in the cache
1313  * and return the block marked busy. If the calling process is
1314  * interrupted by a signal for an interruptible mount point, return
1315  * NULL.
1316  *
1317  * The caller must carefully deal with the possible B_INVAL state of
1318  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1319  * indirectly), so synchronous reads can be issued without worrying about
1320  * the B_INVAL state.  We have to be a little more careful when dealing
1321  * with writes (see comments in nfs_write()) when extending a file past
1322  * its EOF.
1323  */
1324 static struct buf *
1325 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1326 {
1327 	struct buf *bp;
1328 	struct mount *mp;
1329 	struct nfsmount *nmp;
1330 
1331 	mp = vp->v_mount;
1332 	nmp = VFSTONFS(mp);
1333 
1334 	if (nmp->nm_flag & NFSMNT_INT) {
1335 		sigset_t oldset;
1336 
1337 		newnfs_set_sigmask(td, &oldset);
1338 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1339 		newnfs_restore_sigmask(td, &oldset);
1340 		while (bp == NULL) {
1341 			if (newnfs_sigintr(nmp, td))
1342 				return (NULL);
1343 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1344 		}
1345 	} else {
1346 		bp = getblk(vp, bn, size, 0, 0, 0);
1347 	}
1348 
1349 	if (vp->v_type == VREG)
1350 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1351 	return (bp);
1352 }
1353 
1354 /*
1355  * Flush and invalidate all dirty buffers. If another process is already
1356  * doing the flush, just wait for completion.
1357  */
1358 int
1359 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1360 {
1361 	struct nfsnode *np = VTONFS(vp);
1362 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1363 	int error = 0, slpflag, slptimeo;
1364 	bool old_lock;
1365 	struct timespec ts;
1366 
1367 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1368 
1369 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1370 		intrflg = 0;
1371 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1372 		intrflg = 1;
1373 	if (intrflg) {
1374 		slpflag = PCATCH;
1375 		slptimeo = 2 * hz;
1376 	} else {
1377 		slpflag = 0;
1378 		slptimeo = 0;
1379 	}
1380 
1381 	old_lock = ncl_excl_start(vp);
1382 	if (old_lock)
1383 		flags |= V_ALLOWCLEAN;
1384 
1385 	/*
1386 	 * Now, flush as required.
1387 	 */
1388 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1389 	     vp->v_bufobj.bo_object != NULL) {
1390 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1391 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1392 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1393 		/*
1394 		 * If the page clean was interrupted, fail the invalidation.
1395 		 * Not doing so, we run the risk of losing dirty pages in the
1396 		 * vinvalbuf() call below.
1397 		 */
1398 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1399 			goto out;
1400 	}
1401 
1402 	error = vinvalbuf(vp, flags, slpflag, 0);
1403 	while (error) {
1404 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1405 			goto out;
1406 		error = vinvalbuf(vp, flags, 0, slptimeo);
1407 	}
1408 	if (NFSHASPNFS(nmp)) {
1409 		nfscl_layoutcommit(vp, td);
1410 		nanouptime(&ts);
1411 		/*
1412 		 * Invalidate the attribute cache, since writes to a DS
1413 		 * won't update the size attribute.
1414 		 */
1415 		NFSLOCKNODE(np);
1416 		np->n_attrstamp = 0;
1417 	} else {
1418 		nanouptime(&ts);
1419 		NFSLOCKNODE(np);
1420 	}
1421 	if (np->n_directio_asyncwr == 0 && (np->n_flag & NMODIFIED) != 0) {
1422 		np->n_localmodtime = ts;
1423 		np->n_flag &= ~NMODIFIED;
1424 	}
1425 	NFSUNLOCKNODE(np);
1426 out:
1427 	ncl_excl_finish(vp, old_lock);
1428 	return error;
1429 }
1430 
1431 /*
1432  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1433  * This is mainly to avoid queueing async I/O requests when the nfsiods
1434  * are all hung on a dead server.
1435  *
1436  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1437  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1438  */
1439 int
1440 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1441 {
1442 	int iod;
1443 	int gotiod;
1444 	int slpflag = 0;
1445 	int slptimeo = 0;
1446 	int error, error2;
1447 
1448 	/*
1449 	 * Commits are usually short and sweet so lets save some cpu and
1450 	 * leave the async daemons for more important rpc's (such as reads
1451 	 * and writes).
1452 	 *
1453 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1454 	 * in the directory in order to update attributes. This can deadlock
1455 	 * with another thread that is waiting for async I/O to be done by
1456 	 * an nfsiod thread while holding a lock on one of these vnodes.
1457 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1458 	 * perform Readdirplus RPCs.
1459 	 */
1460 	NFSLOCKIOD();
1461 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1462 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1463 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1464 		NFSUNLOCKIOD();
1465 		return(EIO);
1466 	}
1467 again:
1468 	if (nmp->nm_flag & NFSMNT_INT)
1469 		slpflag = PCATCH;
1470 	gotiod = FALSE;
1471 
1472 	/*
1473 	 * Find a free iod to process this request.
1474 	 */
1475 	for (iod = 0; iod < ncl_numasync; iod++)
1476 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1477 			gotiod = TRUE;
1478 			break;
1479 		}
1480 
1481 	/*
1482 	 * Try to create one if none are free.
1483 	 */
1484 	if (!gotiod)
1485 		ncl_nfsiodnew();
1486 	else {
1487 		/*
1488 		 * Found one, so wake it up and tell it which
1489 		 * mount to process.
1490 		 */
1491 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1492 		    iod, nmp));
1493 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1494 		ncl_iodmount[iod] = nmp;
1495 		nmp->nm_bufqiods++;
1496 		wakeup(&ncl_iodwant[iod]);
1497 	}
1498 
1499 	/*
1500 	 * If none are free, we may already have an iod working on this mount
1501 	 * point.  If so, it will process our request.
1502 	 */
1503 	if (!gotiod) {
1504 		if (nmp->nm_bufqiods > 0) {
1505 			NFS_DPF(ASYNCIO,
1506 				("ncl_asyncio: %d iods are already processing mount %p\n",
1507 				 nmp->nm_bufqiods, nmp));
1508 			gotiod = TRUE;
1509 		}
1510 	}
1511 
1512 	/*
1513 	 * If we have an iod which can process the request, then queue
1514 	 * the buffer.
1515 	 */
1516 	if (gotiod) {
1517 		/*
1518 		 * Ensure that the queue never grows too large.  We still want
1519 		 * to asynchronize so we block rather then return EIO.
1520 		 */
1521 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1522 			NFS_DPF(ASYNCIO,
1523 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1524 			nmp->nm_bufqwant = TRUE;
1525 			error = newnfs_msleep(td, &nmp->nm_bufq,
1526 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1527 			   slptimeo);
1528 			if (error) {
1529 				error2 = newnfs_sigintr(nmp, td);
1530 				if (error2) {
1531 					NFSUNLOCKIOD();
1532 					return (error2);
1533 				}
1534 				if (slpflag == PCATCH) {
1535 					slpflag = 0;
1536 					slptimeo = 2 * hz;
1537 				}
1538 			}
1539 			/*
1540 			 * We might have lost our iod while sleeping,
1541 			 * so check and loop if necessary.
1542 			 */
1543 			goto again;
1544 		}
1545 
1546 		/* We might have lost our nfsiod */
1547 		if (nmp->nm_bufqiods == 0) {
1548 			NFS_DPF(ASYNCIO,
1549 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1550 			goto again;
1551 		}
1552 
1553 		if (bp->b_iocmd == BIO_READ) {
1554 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1555 				bp->b_rcred = crhold(cred);
1556 		} else {
1557 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1558 				bp->b_wcred = crhold(cred);
1559 		}
1560 
1561 		if (bp->b_flags & B_REMFREE)
1562 			bremfreef(bp);
1563 		BUF_KERNPROC(bp);
1564 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1565 		nmp->nm_bufqlen++;
1566 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1567 			NFSLOCKNODE(VTONFS(bp->b_vp));
1568 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1569 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1570 			NFSUNLOCKNODE(VTONFS(bp->b_vp));
1571 		}
1572 		NFSUNLOCKIOD();
1573 		return (0);
1574 	}
1575 
1576 	NFSUNLOCKIOD();
1577 
1578 	/*
1579 	 * All the iods are busy on other mounts, so return EIO to
1580 	 * force the caller to process the i/o synchronously.
1581 	 */
1582 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1583 	return (EIO);
1584 }
1585 
1586 void
1587 ncl_doio_directwrite(struct buf *bp)
1588 {
1589 	int iomode, must_commit;
1590 	struct uio *uiop = (struct uio *)bp->b_caller1;
1591 	char *iov_base = uiop->uio_iov->iov_base;
1592 
1593 	iomode = NFSWRITE_FILESYNC;
1594 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1595 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1596 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1597 	free(iov_base, M_NFSDIRECTIO);
1598 	free(uiop->uio_iov, M_NFSDIRECTIO);
1599 	free(uiop, M_NFSDIRECTIO);
1600 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1601 		struct nfsnode *np = VTONFS(bp->b_vp);
1602 		NFSLOCKNODE(np);
1603 		if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1604 			/*
1605 			 * Invalidate the attribute cache, since writes to a DS
1606 			 * won't update the size attribute.
1607 			 */
1608 			np->n_attrstamp = 0;
1609 		}
1610 		np->n_directio_asyncwr--;
1611 		if (np->n_directio_asyncwr == 0) {
1612 			np->n_flag &= ~NMODIFIED;
1613 			if ((np->n_flag & NFSYNCWAIT)) {
1614 				np->n_flag &= ~NFSYNCWAIT;
1615 				wakeup((caddr_t)&np->n_directio_asyncwr);
1616 			}
1617 		}
1618 		NFSUNLOCKNODE(np);
1619 	}
1620 	bp->b_vp = NULL;
1621 	uma_zfree(ncl_pbuf_zone, bp);
1622 }
1623 
1624 /*
1625  * Do an I/O operation to/from a cache block. This may be called
1626  * synchronously or from an nfsiod.
1627  */
1628 int
1629 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1630     int called_from_strategy)
1631 {
1632 	struct uio *uiop;
1633 	struct nfsnode *np;
1634 	struct nfsmount *nmp;
1635 	int error = 0, iomode, must_commit = 0;
1636 	struct uio uio;
1637 	struct iovec io;
1638 	struct proc *p = td ? td->td_proc : NULL;
1639 	uint8_t	iocmd;
1640 
1641 	np = VTONFS(vp);
1642 	nmp = VFSTONFS(vp->v_mount);
1643 	uiop = &uio;
1644 	uiop->uio_iov = &io;
1645 	uiop->uio_iovcnt = 1;
1646 	uiop->uio_segflg = UIO_SYSSPACE;
1647 	uiop->uio_td = td;
1648 
1649 	/*
1650 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1651 	 * do this here so we do not have to do it in all the code that
1652 	 * calls us.
1653 	 */
1654 	bp->b_flags &= ~B_INVAL;
1655 	bp->b_ioflags &= ~BIO_ERROR;
1656 
1657 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1658 	iocmd = bp->b_iocmd;
1659 	if (iocmd == BIO_READ) {
1660 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1661 	    io.iov_base = bp->b_data;
1662 	    uiop->uio_rw = UIO_READ;
1663 
1664 	    switch (vp->v_type) {
1665 	    case VREG:
1666 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1667 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1668 		error = ncl_readrpc(vp, uiop, cr);
1669 
1670 		if (!error) {
1671 		    if (uiop->uio_resid) {
1672 			/*
1673 			 * If we had a short read with no error, we must have
1674 			 * hit a file hole.  We should zero-fill the remainder.
1675 			 * This can also occur if the server hits the file EOF.
1676 			 *
1677 			 * Holes used to be able to occur due to pending
1678 			 * writes, but that is not possible any longer.
1679 			 */
1680 			int nread = bp->b_bcount - uiop->uio_resid;
1681 			ssize_t left = uiop->uio_resid;
1682 
1683 			if (left > 0)
1684 				bzero((char *)bp->b_data + nread, left);
1685 			uiop->uio_resid = 0;
1686 		    }
1687 		}
1688 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1689 		if (p && vp->v_writecount <= -1) {
1690 			NFSLOCKNODE(np);
1691 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1692 				NFSUNLOCKNODE(np);
1693 				PROC_LOCK(p);
1694 				killproc(p, "text file modification");
1695 				PROC_UNLOCK(p);
1696 			} else
1697 				NFSUNLOCKNODE(np);
1698 		}
1699 		break;
1700 	    case VLNK:
1701 		uiop->uio_offset = (off_t)0;
1702 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1703 		error = ncl_readlinkrpc(vp, uiop, cr);
1704 		break;
1705 	    case VDIR:
1706 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1707 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1708 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1709 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1710 			if (error == NFSERR_NOTSUPP)
1711 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1712 		}
1713 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1714 			error = ncl_readdirrpc(vp, uiop, cr, td);
1715 		/*
1716 		 * end-of-directory sets B_INVAL but does not generate an
1717 		 * error.
1718 		 */
1719 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1720 			bp->b_flags |= B_INVAL;
1721 		break;
1722 	    default:
1723 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1724 		break;
1725 	    }
1726 	    if (error) {
1727 		bp->b_ioflags |= BIO_ERROR;
1728 		bp->b_error = error;
1729 	    }
1730 	} else {
1731 	    /*
1732 	     * If we only need to commit, try to commit
1733 	     */
1734 	    if (bp->b_flags & B_NEEDCOMMIT) {
1735 		    int retv;
1736 		    off_t off;
1737 
1738 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1739 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1740 			bp->b_wcred, td);
1741 		    if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1742 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1743 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1744 			    bp->b_resid = 0;
1745 			    bufdone(bp);
1746 			    return (0);
1747 		    }
1748 		    if (retv == NFSERR_STALEWRITEVERF) {
1749 			    ncl_clearcommit(vp->v_mount);
1750 		    }
1751 	    }
1752 
1753 	    /*
1754 	     * Setup for actual write
1755 	     */
1756 	    NFSLOCKNODE(np);
1757 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1758 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1759 	    NFSUNLOCKNODE(np);
1760 
1761 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1762 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1763 		    - bp->b_dirtyoff;
1764 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1765 		    + bp->b_dirtyoff;
1766 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1767 		uiop->uio_rw = UIO_WRITE;
1768 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1769 
1770 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1771 		    iomode = NFSWRITE_UNSTABLE;
1772 		else
1773 		    iomode = NFSWRITE_FILESYNC;
1774 
1775 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1776 		    called_from_strategy);
1777 
1778 		/*
1779 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1780 		 * to cluster the buffers needing commit.  This will allow
1781 		 * the system to submit a single commit rpc for the whole
1782 		 * cluster.  We can do this even if the buffer is not 100%
1783 		 * dirty (relative to the NFS blocksize), so we optimize the
1784 		 * append-to-file-case.
1785 		 *
1786 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1787 		 * cleared because write clustering only works for commit
1788 		 * rpc's, not for the data portion of the write).
1789 		 */
1790 
1791 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1792 		    bp->b_flags |= B_NEEDCOMMIT;
1793 		    if (bp->b_dirtyoff == 0
1794 			&& bp->b_dirtyend == bp->b_bcount)
1795 			bp->b_flags |= B_CLUSTEROK;
1796 		} else {
1797 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1798 		}
1799 
1800 		/*
1801 		 * For an interrupted write, the buffer is still valid
1802 		 * and the write hasn't been pushed to the server yet,
1803 		 * so we can't set BIO_ERROR and report the interruption
1804 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1805 		 * is not relevant, so the rpc attempt is essentially
1806 		 * a noop.  For the case of a V3 write rpc not being
1807 		 * committed to stable storage, the block is still
1808 		 * dirty and requires either a commit rpc or another
1809 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1810 		 * the block is reused. This is indicated by setting
1811 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1812 		 *
1813 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1814 		 * write error and is handled as above, except that
1815 		 * B_EINTR isn't set. One cause of this is a stale stateid
1816 		 * error for the RPC that indicates recovery is required,
1817 		 * when called with called_from_strategy != 0.
1818 		 *
1819 		 * If the buffer is marked B_PAGING, it does not reside on
1820 		 * the vp's paging queues so we cannot call bdirty().  The
1821 		 * bp in this case is not an NFS cache block so we should
1822 		 * be safe. XXX
1823 		 *
1824 		 * The logic below breaks up errors into recoverable and
1825 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1826 		 * and keep the buffer around for potential write retries.
1827 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1828 		 * and save the error in the nfsnode. This is less than ideal
1829 		 * but necessary. Keeping such buffers around could potentially
1830 		 * cause buffer exhaustion eventually (they can never be written
1831 		 * out, so will get constantly be re-dirtied). It also causes
1832 		 * all sorts of vfs panics. For non-recoverable write errors,
1833 		 * also invalidate the attrcache, so we'll be forced to go over
1834 		 * the wire for this object, returning an error to user on next
1835 		 * call (most of the time).
1836 		 */
1837 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1838 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1839 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1840 			if ((bp->b_flags & B_PAGING) == 0) {
1841 			    bdirty(bp);
1842 			    bp->b_flags &= ~B_DONE;
1843 			}
1844 			if ((error == EINTR || error == ETIMEDOUT) &&
1845 			    (bp->b_flags & B_ASYNC) == 0)
1846 			    bp->b_flags |= B_EINTR;
1847 		} else {
1848 		    if (error) {
1849 			bp->b_ioflags |= BIO_ERROR;
1850 			bp->b_flags |= B_INVAL;
1851 			bp->b_error = np->n_error = error;
1852 			NFSLOCKNODE(np);
1853 			np->n_flag |= NWRITEERR;
1854 			np->n_attrstamp = 0;
1855 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1856 			NFSUNLOCKNODE(np);
1857 		    }
1858 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1859 		}
1860 	    } else {
1861 		bp->b_resid = 0;
1862 		bufdone(bp);
1863 		return (0);
1864 	    }
1865 	}
1866 	bp->b_resid = uiop->uio_resid;
1867 	if (must_commit)
1868 	    ncl_clearcommit(vp->v_mount);
1869 	bufdone(bp);
1870 	return (error);
1871 }
1872 
1873 /*
1874  * Used to aid in handling ftruncate() operations on the NFS client side.
1875  * Truncation creates a number of special problems for NFS.  We have to
1876  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1877  * we have to properly handle VM pages or (potentially dirty) buffers
1878  * that straddle the truncation point.
1879  */
1880 
1881 int
1882 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1883 {
1884 	struct nfsnode *np = VTONFS(vp);
1885 	u_quad_t tsize;
1886 	int biosize = vp->v_bufobj.bo_bsize;
1887 	int error = 0;
1888 
1889 	NFSLOCKNODE(np);
1890 	tsize = np->n_size;
1891 	np->n_size = nsize;
1892 	NFSUNLOCKNODE(np);
1893 
1894 	if (nsize < tsize) {
1895 		struct buf *bp;
1896 		daddr_t lbn;
1897 		int bufsize;
1898 
1899 		/*
1900 		 * vtruncbuf() doesn't get the buffer overlapping the
1901 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1902 		 * buffer that now needs to be truncated.
1903 		 */
1904 		error = vtruncbuf(vp, nsize, biosize);
1905 		lbn = nsize / biosize;
1906 		bufsize = nsize - (lbn * biosize);
1907 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1908 		if (!bp)
1909 			return EINTR;
1910 		if (bp->b_dirtyoff > bp->b_bcount)
1911 			bp->b_dirtyoff = bp->b_bcount;
1912 		if (bp->b_dirtyend > bp->b_bcount)
1913 			bp->b_dirtyend = bp->b_bcount;
1914 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1915 		brelse(bp);
1916 	} else {
1917 		vnode_pager_setsize(vp, nsize);
1918 	}
1919 	return(error);
1920 }
1921