xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 7be9a3b45356747f9fcb6d69a722c1c95f8060bf)
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 			/*
788 			 * When doing direct I/O we do not care if the
789 			 * server's write verifier has changed, but we
790 			 * do not want to update the verifier if it has
791 			 * changed, since that hides the change from
792 			 * writes being done through the buffer cache.
793 			 * By passing must_commit in set to two, the code
794 			 * in nfsrpc_writerpc() will not update the
795 			 * verifier on the mount point.
796 			 */
797 			must_commit = 2;
798 			error = ncl_writerpc(vp, &uio, cred, &iomode,
799 			    &must_commit, 0);
800 			KASSERT((must_commit == 2),
801 			    ("ncl_directio_write: Updated write verifier"));
802 			if (error)
803 				return (error);
804 			if (iomode != NFSWRITE_FILESYNC)
805 				printf("nfs_directio_write: Broken server "
806 				    "did not reply FILE_SYNC\n");
807 			uiop->uio_offset += size;
808 			uiop->uio_resid -= size;
809 			if (uiop->uio_iov->iov_len <= size) {
810 				uiop->uio_iovcnt--;
811 				uiop->uio_iov++;
812 			} else {
813 				uiop->uio_iov->iov_base =
814 					(char *)uiop->uio_iov->iov_base + size;
815 				uiop->uio_iov->iov_len -= size;
816 			}
817 		}
818 	} else {
819 		struct uio *t_uio;
820 		struct iovec *t_iov;
821 		struct buf *bp;
822 
823 		/*
824 		 * Break up the write into blocksize chunks and hand these
825 		 * over to nfsiod's for write back.
826 		 * Unfortunately, this incurs a copy of the data. Since
827 		 * the user could modify the buffer before the write is
828 		 * initiated.
829 		 *
830 		 * The obvious optimization here is that one of the 2 copies
831 		 * in the async write path can be eliminated by copying the
832 		 * data here directly into mbufs and passing the mbuf chain
833 		 * down. But that will require a fair amount of re-working
834 		 * of the code and can be done if there's enough interest
835 		 * in NFS directio access.
836 		 */
837 		while (uiop->uio_resid > 0) {
838 			size = MIN(uiop->uio_resid, wsize);
839 			size = MIN(uiop->uio_iov->iov_len, size);
840 			bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
841 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
842 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
843 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
844 			t_iov->iov_len = size;
845 			t_uio->uio_iov = t_iov;
846 			t_uio->uio_iovcnt = 1;
847 			t_uio->uio_offset = uiop->uio_offset;
848 			t_uio->uio_resid = size;
849 			t_uio->uio_segflg = UIO_SYSSPACE;
850 			t_uio->uio_rw = UIO_WRITE;
851 			t_uio->uio_td = td;
852 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
853 			    uiop->uio_segflg == UIO_SYSSPACE,
854 			    ("nfs_directio_write: Bad uio_segflg"));
855 			if (uiop->uio_segflg == UIO_USERSPACE) {
856 				error = copyin(uiop->uio_iov->iov_base,
857 				    t_iov->iov_base, size);
858 				if (error != 0)
859 					goto err_free;
860 			} else
861 				/*
862 				 * UIO_SYSSPACE may never happen, but handle
863 				 * it just in case it does.
864 				 */
865 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
866 				    size);
867 			bp->b_flags |= B_DIRECT;
868 			bp->b_iocmd = BIO_WRITE;
869 			if (cred != NOCRED) {
870 				crhold(cred);
871 				bp->b_wcred = cred;
872 			} else
873 				bp->b_wcred = NOCRED;
874 			bp->b_caller1 = (void *)t_uio;
875 			bp->b_vp = vp;
876 			error = ncl_asyncio(nmp, bp, NOCRED, td);
877 err_free:
878 			if (error) {
879 				free(t_iov->iov_base, M_NFSDIRECTIO);
880 				free(t_iov, M_NFSDIRECTIO);
881 				free(t_uio, M_NFSDIRECTIO);
882 				bp->b_vp = NULL;
883 				uma_zfree(ncl_pbuf_zone, bp);
884 				if (error == EINTR)
885 					return (error);
886 				goto do_sync;
887 			}
888 			uiop->uio_offset += size;
889 			uiop->uio_resid -= size;
890 			if (uiop->uio_iov->iov_len <= size) {
891 				uiop->uio_iovcnt--;
892 				uiop->uio_iov++;
893 			} else {
894 				uiop->uio_iov->iov_base =
895 					(char *)uiop->uio_iov->iov_base + size;
896 				uiop->uio_iov->iov_len -= size;
897 			}
898 		}
899 	}
900 	return (0);
901 }
902 
903 /*
904  * Vnode op for write using bio
905  */
906 int
907 ncl_write(struct vop_write_args *ap)
908 {
909 	int biosize;
910 	struct uio *uio = ap->a_uio;
911 	struct thread *td = uio->uio_td;
912 	struct vnode *vp = ap->a_vp;
913 	struct nfsnode *np = VTONFS(vp);
914 	struct ucred *cred = ap->a_cred;
915 	int ioflag = ap->a_ioflag;
916 	struct buf *bp;
917 	struct vattr vattr;
918 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
919 	daddr_t lbn;
920 	int bcount, noncontig_write, obcount;
921 	int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
922 	size_t orig_resid, local_resid;
923 	off_t orig_size, tmp_off;
924 	struct timespec ts;
925 
926 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
927 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
928 	    ("ncl_write proc"));
929 	if (vp->v_type != VREG)
930 		return (EIO);
931 	NFSLOCKNODE(np);
932 	if (np->n_flag & NWRITEERR) {
933 		np->n_flag &= ~NWRITEERR;
934 		NFSUNLOCKNODE(np);
935 		return (np->n_error);
936 	} else
937 		NFSUNLOCKNODE(np);
938 	mtx_lock(&nmp->nm_mtx);
939 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
940 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
941 		mtx_unlock(&nmp->nm_mtx);
942 		(void)ncl_fsinfo(nmp, vp, cred, td);
943 		mtx_lock(&nmp->nm_mtx);
944 	}
945 	if (nmp->nm_wsize == 0)
946 		(void) newnfs_iosize(nmp);
947 	mtx_unlock(&nmp->nm_mtx);
948 
949 	/*
950 	 * Synchronously flush pending buffers if we are in synchronous
951 	 * mode or if we are appending.
952 	 */
953 	if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
954 	    NMODIFIED))) {
955 		/*
956 		 * For the case where IO_APPEND is being done using a
957 		 * direct output (to the NFS server) RPC and
958 		 * newnfs_directio_enable is 0, all buffer cache buffers,
959 		 * including ones not modified, must be invalidated.
960 		 * This ensures that stale data is not read out of the
961 		 * buffer cache.  The call also invalidates all mapped
962 		 * pages and, since the exclusive lock is held on the vnode,
963 		 * new pages cannot be faulted in.
964 		 *
965 		 * For the case where newnfs_directio_enable is set
966 		 * (which is not the default), it is not obvious that
967 		 * stale data should be left in the buffer cache, but
968 		 * the code has been this way for over a decade without
969 		 * complaints.  Note that, unlike doing IO_APPEND via
970 		 * a direct write RPC when newnfs_directio_enable is not set,
971 		 * when newnfs_directio_enable is set, reading is done via
972 		 * direct to NFS server RPCs as well.
973 		 */
974 		np->n_attrstamp = 0;
975 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
976 		error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
977 		    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
978 		if (error != 0)
979 			return (error);
980 	}
981 
982 	orig_resid = uio->uio_resid;
983 	NFSLOCKNODE(np);
984 	orig_size = np->n_size;
985 	NFSUNLOCKNODE(np);
986 
987 	/*
988 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
989 	 * get the append lock.
990 	 */
991 	if (ioflag & IO_APPEND) {
992 		np->n_attrstamp = 0;
993 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
994 		error = VOP_GETATTR(vp, &vattr, cred);
995 		if (error)
996 			return (error);
997 		NFSLOCKNODE(np);
998 		uio->uio_offset = np->n_size;
999 		NFSUNLOCKNODE(np);
1000 	}
1001 
1002 	if (uio->uio_offset < 0)
1003 		return (EINVAL);
1004 	tmp_off = uio->uio_offset + uio->uio_resid;
1005 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
1006 		return (EFBIG);
1007 	if (uio->uio_resid == 0)
1008 		return (0);
1009 
1010 	/*
1011 	 * Do IO_APPEND writing via a synchronous direct write.
1012 	 * This can result in a significant performance improvement.
1013 	 */
1014 	if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
1015 	    (ioflag & IO_APPEND)) {
1016 		/*
1017 		 * Direct writes to the server must be done NFSWRITE_FILESYNC,
1018 		 * because the write data is not cached and, therefore, the
1019 		 * write cannot be redone after a server reboot.
1020 		 * Set IO_SYNC to make this happen.
1021 		 */
1022 		ioflag |= IO_SYNC;
1023 		return (nfs_directio_write(vp, uio, cred, ioflag));
1024 	}
1025 
1026 	/*
1027 	 * Maybe this should be above the vnode op call, but so long as
1028 	 * file servers have no limits, i don't think it matters
1029 	 */
1030 	if (vn_rlimit_fsize(vp, uio, td))
1031 		return (EFBIG);
1032 
1033 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
1034 	biosize = vp->v_bufobj.bo_bsize;
1035 	/*
1036 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
1037 	 * would exceed the local maximum per-file write commit size when
1038 	 * combined with those, we must decide whether to flush,
1039 	 * go synchronous, or return error.  We don't bother checking
1040 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
1041 	 * no point optimizing for something that really won't ever happen.
1042 	 */
1043 	wouldcommit = 0;
1044 	if (!(ioflag & IO_SYNC)) {
1045 		int nflag;
1046 
1047 		NFSLOCKNODE(np);
1048 		nflag = np->n_flag;
1049 		NFSUNLOCKNODE(np);
1050 		if (nflag & NMODIFIED) {
1051 			BO_LOCK(&vp->v_bufobj);
1052 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1053 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1054 				    b_bobufs) {
1055 					if (bp->b_flags & B_NEEDCOMMIT)
1056 						wouldcommit += bp->b_bcount;
1057 				}
1058 			}
1059 			BO_UNLOCK(&vp->v_bufobj);
1060 		}
1061 	}
1062 
1063 	do {
1064 		if (!(ioflag & IO_SYNC)) {
1065 			wouldcommit += biosize;
1066 			if (wouldcommit > nmp->nm_wcommitsize) {
1067 				np->n_attrstamp = 0;
1068 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1069 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1070 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1071 				if (error != 0)
1072 					goto out;
1073 				wouldcommit = biosize;
1074 			}
1075 		}
1076 
1077 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1078 		lbn = uio->uio_offset / biosize;
1079 		on = uio->uio_offset - (lbn * biosize);
1080 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1081 again:
1082 		/*
1083 		 * Handle direct append and file extension cases, calculate
1084 		 * unaligned buffer size.
1085 		 */
1086 		NFSLOCKNODE(np);
1087 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1088 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1089 			noncontig_write = 1;
1090 		else
1091 			noncontig_write = 0;
1092 		if ((uio->uio_offset == np->n_size ||
1093 		    (noncontig_write != 0 &&
1094 		    lbn == (np->n_size / biosize) &&
1095 		    uio->uio_offset + n > np->n_size)) && n) {
1096 			NFSUNLOCKNODE(np);
1097 			/*
1098 			 * Get the buffer (in its pre-append state to maintain
1099 			 * B_CACHE if it was previously set).  Resize the
1100 			 * nfsnode after we have locked the buffer to prevent
1101 			 * readers from reading garbage.
1102 			 */
1103 			obcount = np->n_size - (lbn * biosize);
1104 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1105 
1106 			if (bp != NULL) {
1107 				long save;
1108 
1109 				NFSLOCKNODE(np);
1110 				np->n_size = uio->uio_offset + n;
1111 				np->n_flag |= NMODIFIED;
1112 				np->n_flag &= ~NVNSETSZSKIP;
1113 				vnode_pager_setsize(vp, np->n_size);
1114 				NFSUNLOCKNODE(np);
1115 
1116 				save = bp->b_flags & B_CACHE;
1117 				bcount = on + n;
1118 				allocbuf(bp, bcount);
1119 				bp->b_flags |= save;
1120 				if (noncontig_write != 0 && on > obcount)
1121 					vfs_bio_bzero_buf(bp, obcount, on -
1122 					    obcount);
1123 			}
1124 		} else {
1125 			/*
1126 			 * Obtain the locked cache block first, and then
1127 			 * adjust the file's size as appropriate.
1128 			 */
1129 			bcount = on + n;
1130 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1131 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1132 					bcount = biosize;
1133 				else
1134 					bcount = np->n_size - (off_t)lbn * biosize;
1135 			}
1136 			NFSUNLOCKNODE(np);
1137 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1138 			NFSLOCKNODE(np);
1139 			if (uio->uio_offset + n > np->n_size) {
1140 				np->n_size = uio->uio_offset + n;
1141 				np->n_flag |= NMODIFIED;
1142 				np->n_flag &= ~NVNSETSZSKIP;
1143 				vnode_pager_setsize(vp, np->n_size);
1144 			}
1145 			NFSUNLOCKNODE(np);
1146 		}
1147 
1148 		if (!bp) {
1149 			error = newnfs_sigintr(nmp, td);
1150 			if (!error)
1151 				error = EINTR;
1152 			break;
1153 		}
1154 
1155 		/*
1156 		 * Issue a READ if B_CACHE is not set.  In special-append
1157 		 * mode, B_CACHE is based on the buffer prior to the write
1158 		 * op and is typically set, avoiding the read.  If a read
1159 		 * is required in special append mode, the server will
1160 		 * probably send us a short-read since we extended the file
1161 		 * on our end, resulting in b_resid == 0 and, thusly,
1162 		 * B_CACHE getting set.
1163 		 *
1164 		 * We can also avoid issuing the read if the write covers
1165 		 * the entire buffer.  We have to make sure the buffer state
1166 		 * is reasonable in this case since we will not be initiating
1167 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1168 		 * more information.
1169 		 *
1170 		 * B_CACHE may also be set due to the buffer being cached
1171 		 * normally.
1172 		 */
1173 
1174 		bp_cached = 1;
1175 		if (on == 0 && n == bcount) {
1176 			if ((bp->b_flags & B_CACHE) == 0)
1177 				bp_cached = 0;
1178 			bp->b_flags |= B_CACHE;
1179 			bp->b_flags &= ~B_INVAL;
1180 			bp->b_ioflags &= ~BIO_ERROR;
1181 		}
1182 
1183 		if ((bp->b_flags & B_CACHE) == 0) {
1184 			bp->b_iocmd = BIO_READ;
1185 			vfs_busy_pages(bp, 0);
1186 			error = ncl_doio(vp, bp, cred, td, 0);
1187 			if (error) {
1188 				brelse(bp);
1189 				break;
1190 			}
1191 		}
1192 		if (bp->b_wcred == NOCRED)
1193 			bp->b_wcred = crhold(cred);
1194 		NFSLOCKNODE(np);
1195 		np->n_flag |= NMODIFIED;
1196 		NFSUNLOCKNODE(np);
1197 
1198 		/*
1199 		 * If dirtyend exceeds file size, chop it down.  This should
1200 		 * not normally occur but there is an append race where it
1201 		 * might occur XXX, so we log it.
1202 		 *
1203 		 * If the chopping creates a reverse-indexed or degenerate
1204 		 * situation with dirtyoff/end, we 0 both of them.
1205 		 */
1206 
1207 		if (bp->b_dirtyend > bcount) {
1208 			printf("NFS append race @%lx:%d\n",
1209 			    (long)bp->b_blkno * DEV_BSIZE,
1210 			    bp->b_dirtyend - bcount);
1211 			bp->b_dirtyend = bcount;
1212 		}
1213 
1214 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1215 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1216 
1217 		/*
1218 		 * If the new write will leave a contiguous dirty
1219 		 * area, just update the b_dirtyoff and b_dirtyend,
1220 		 * otherwise force a write rpc of the old dirty area.
1221 		 *
1222 		 * If there has been a file lock applied to this file
1223 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1224 		 * While it is possible to merge discontiguous writes due to
1225 		 * our having a B_CACHE buffer ( and thus valid read data
1226 		 * for the hole), we don't because it could lead to
1227 		 * significant cache coherency problems with multiple clients,
1228 		 * especially if locking is implemented later on.
1229 		 *
1230 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1231 		 * not been file locking done on this file:
1232 		 * Relax coherency a bit for the sake of performance and
1233 		 * expand the current dirty region to contain the new
1234 		 * write even if it means we mark some non-dirty data as
1235 		 * dirty.
1236 		 */
1237 
1238 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1239 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1240 			if (bwrite(bp) == EINTR) {
1241 				error = EINTR;
1242 				break;
1243 			}
1244 			goto again;
1245 		}
1246 
1247 		local_resid = uio->uio_resid;
1248 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1249 
1250 		if (error != 0 && !bp_cached) {
1251 			/*
1252 			 * This block has no other content then what
1253 			 * possibly was written by the faulty uiomove.
1254 			 * Release it, forgetting the data pages, to
1255 			 * prevent the leak of uninitialized data to
1256 			 * usermode.
1257 			 */
1258 			bp->b_ioflags |= BIO_ERROR;
1259 			brelse(bp);
1260 			uio->uio_offset -= local_resid - uio->uio_resid;
1261 			uio->uio_resid = local_resid;
1262 			break;
1263 		}
1264 
1265 		/*
1266 		 * Since this block is being modified, it must be written
1267 		 * again and not just committed.  Since write clustering does
1268 		 * not work for the stage 1 data write, only the stage 2
1269 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1270 		 */
1271 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1272 
1273 		/*
1274 		 * Get the partial update on the progress made from
1275 		 * uiomove, if an error occurred.
1276 		 */
1277 		if (error != 0)
1278 			n = local_resid - uio->uio_resid;
1279 
1280 		/*
1281 		 * Only update dirtyoff/dirtyend if not a degenerate
1282 		 * condition.
1283 		 */
1284 		if (n > 0) {
1285 			if (bp->b_dirtyend > 0) {
1286 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1287 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1288 			} else {
1289 				bp->b_dirtyoff = on;
1290 				bp->b_dirtyend = on + n;
1291 			}
1292 			vfs_bio_set_valid(bp, on, n);
1293 		}
1294 
1295 		/*
1296 		 * If IO_SYNC do bwrite().
1297 		 *
1298 		 * IO_INVAL appears to be unused.  The idea appears to be
1299 		 * to turn off caching in this case.  Very odd.  XXX
1300 		 */
1301 		if ((ioflag & IO_SYNC)) {
1302 			if (ioflag & IO_INVAL)
1303 				bp->b_flags |= B_NOCACHE;
1304 			error1 = bwrite(bp);
1305 			if (error1 != 0) {
1306 				if (error == 0)
1307 					error = error1;
1308 				break;
1309 			}
1310 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1311 			bp->b_flags |= B_ASYNC;
1312 			(void) ncl_writebp(bp, 0, NULL);
1313 		} else {
1314 			bdwrite(bp);
1315 		}
1316 
1317 		if (error != 0)
1318 			break;
1319 	} while (uio->uio_resid > 0 && n > 0);
1320 
1321 	if (error == 0) {
1322 		nanouptime(&ts);
1323 		NFSLOCKNODE(np);
1324 		np->n_localmodtime = ts;
1325 		NFSUNLOCKNODE(np);
1326 	} else {
1327 		if (ioflag & IO_UNIT) {
1328 			VATTR_NULL(&vattr);
1329 			vattr.va_size = orig_size;
1330 			/* IO_SYNC is handled implicitely */
1331 			(void)VOP_SETATTR(vp, &vattr, cred);
1332 			uio->uio_offset -= orig_resid - uio->uio_resid;
1333 			uio->uio_resid = orig_resid;
1334 		}
1335 	}
1336 
1337 out:
1338 	curthread_pflags2_restore(save2);
1339 	return (error);
1340 }
1341 
1342 /*
1343  * Get an nfs cache block.
1344  *
1345  * Allocate a new one if the block isn't currently in the cache
1346  * and return the block marked busy. If the calling process is
1347  * interrupted by a signal for an interruptible mount point, return
1348  * NULL.
1349  *
1350  * The caller must carefully deal with the possible B_INVAL state of
1351  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1352  * indirectly), so synchronous reads can be issued without worrying about
1353  * the B_INVAL state.  We have to be a little more careful when dealing
1354  * with writes (see comments in nfs_write()) when extending a file past
1355  * its EOF.
1356  */
1357 static struct buf *
1358 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1359 {
1360 	struct buf *bp;
1361 	struct mount *mp;
1362 	struct nfsmount *nmp;
1363 
1364 	mp = vp->v_mount;
1365 	nmp = VFSTONFS(mp);
1366 
1367 	if (nmp->nm_flag & NFSMNT_INT) {
1368 		sigset_t oldset;
1369 
1370 		newnfs_set_sigmask(td, &oldset);
1371 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1372 		newnfs_restore_sigmask(td, &oldset);
1373 		while (bp == NULL) {
1374 			if (newnfs_sigintr(nmp, td))
1375 				return (NULL);
1376 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1377 		}
1378 	} else {
1379 		bp = getblk(vp, bn, size, 0, 0, 0);
1380 	}
1381 
1382 	if (vp->v_type == VREG)
1383 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1384 	return (bp);
1385 }
1386 
1387 /*
1388  * Flush and invalidate all dirty buffers. If another process is already
1389  * doing the flush, just wait for completion.
1390  */
1391 int
1392 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1393 {
1394 	struct nfsnode *np = VTONFS(vp);
1395 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1396 	int error = 0, slpflag, slptimeo;
1397 	bool old_lock;
1398 	struct timespec ts;
1399 
1400 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1401 
1402 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1403 		intrflg = 0;
1404 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1405 		intrflg = 1;
1406 	if (intrflg) {
1407 		slpflag = PCATCH;
1408 		slptimeo = 2 * hz;
1409 	} else {
1410 		slpflag = 0;
1411 		slptimeo = 0;
1412 	}
1413 
1414 	old_lock = ncl_excl_start(vp);
1415 	if (old_lock)
1416 		flags |= V_ALLOWCLEAN;
1417 
1418 	/*
1419 	 * Now, flush as required.
1420 	 */
1421 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1422 	     vp->v_bufobj.bo_object != NULL) {
1423 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1424 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1425 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1426 		/*
1427 		 * If the page clean was interrupted, fail the invalidation.
1428 		 * Not doing so, we run the risk of losing dirty pages in the
1429 		 * vinvalbuf() call below.
1430 		 */
1431 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1432 			goto out;
1433 	}
1434 
1435 	error = vinvalbuf(vp, flags, slpflag, 0);
1436 	while (error) {
1437 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1438 			goto out;
1439 		error = vinvalbuf(vp, flags, 0, slptimeo);
1440 	}
1441 	if (NFSHASPNFS(nmp)) {
1442 		nfscl_layoutcommit(vp, td);
1443 		nanouptime(&ts);
1444 		/*
1445 		 * Invalidate the attribute cache, since writes to a DS
1446 		 * won't update the size attribute.
1447 		 */
1448 		NFSLOCKNODE(np);
1449 		np->n_attrstamp = 0;
1450 	} else {
1451 		nanouptime(&ts);
1452 		NFSLOCKNODE(np);
1453 	}
1454 	if (np->n_directio_asyncwr == 0 && (np->n_flag & NMODIFIED) != 0) {
1455 		np->n_localmodtime = ts;
1456 		np->n_flag &= ~NMODIFIED;
1457 	}
1458 	NFSUNLOCKNODE(np);
1459 out:
1460 	ncl_excl_finish(vp, old_lock);
1461 	return error;
1462 }
1463 
1464 /*
1465  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1466  * This is mainly to avoid queueing async I/O requests when the nfsiods
1467  * are all hung on a dead server.
1468  *
1469  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1470  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1471  */
1472 int
1473 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1474 {
1475 	int iod;
1476 	int gotiod;
1477 	int slpflag = 0;
1478 	int slptimeo = 0;
1479 	int error, error2;
1480 
1481 	/*
1482 	 * Commits are usually short and sweet so lets save some cpu and
1483 	 * leave the async daemons for more important rpc's (such as reads
1484 	 * and writes).
1485 	 *
1486 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1487 	 * in the directory in order to update attributes. This can deadlock
1488 	 * with another thread that is waiting for async I/O to be done by
1489 	 * an nfsiod thread while holding a lock on one of these vnodes.
1490 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1491 	 * perform Readdirplus RPCs.
1492 	 */
1493 	NFSLOCKIOD();
1494 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1495 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1496 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1497 		NFSUNLOCKIOD();
1498 		return(EIO);
1499 	}
1500 again:
1501 	if (nmp->nm_flag & NFSMNT_INT)
1502 		slpflag = PCATCH;
1503 	gotiod = FALSE;
1504 
1505 	/*
1506 	 * Find a free iod to process this request.
1507 	 */
1508 	for (iod = 0; iod < ncl_numasync; iod++)
1509 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1510 			gotiod = TRUE;
1511 			break;
1512 		}
1513 
1514 	/*
1515 	 * Try to create one if none are free.
1516 	 */
1517 	if (!gotiod)
1518 		ncl_nfsiodnew();
1519 	else {
1520 		/*
1521 		 * Found one, so wake it up and tell it which
1522 		 * mount to process.
1523 		 */
1524 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1525 		    iod, nmp));
1526 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1527 		ncl_iodmount[iod] = nmp;
1528 		nmp->nm_bufqiods++;
1529 		wakeup(&ncl_iodwant[iod]);
1530 	}
1531 
1532 	/*
1533 	 * If none are free, we may already have an iod working on this mount
1534 	 * point.  If so, it will process our request.
1535 	 */
1536 	if (!gotiod) {
1537 		if (nmp->nm_bufqiods > 0) {
1538 			NFS_DPF(ASYNCIO,
1539 				("ncl_asyncio: %d iods are already processing mount %p\n",
1540 				 nmp->nm_bufqiods, nmp));
1541 			gotiod = TRUE;
1542 		}
1543 	}
1544 
1545 	/*
1546 	 * If we have an iod which can process the request, then queue
1547 	 * the buffer.
1548 	 */
1549 	if (gotiod) {
1550 		/*
1551 		 * Ensure that the queue never grows too large.  We still want
1552 		 * to asynchronize so we block rather then return EIO.
1553 		 */
1554 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1555 			NFS_DPF(ASYNCIO,
1556 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1557 			nmp->nm_bufqwant = TRUE;
1558 			error = newnfs_msleep(td, &nmp->nm_bufq,
1559 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1560 			   slptimeo);
1561 			if (error) {
1562 				error2 = newnfs_sigintr(nmp, td);
1563 				if (error2) {
1564 					NFSUNLOCKIOD();
1565 					return (error2);
1566 				}
1567 				if (slpflag == PCATCH) {
1568 					slpflag = 0;
1569 					slptimeo = 2 * hz;
1570 				}
1571 			}
1572 			/*
1573 			 * We might have lost our iod while sleeping,
1574 			 * so check and loop if necessary.
1575 			 */
1576 			goto again;
1577 		}
1578 
1579 		/* We might have lost our nfsiod */
1580 		if (nmp->nm_bufqiods == 0) {
1581 			NFS_DPF(ASYNCIO,
1582 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1583 			goto again;
1584 		}
1585 
1586 		if (bp->b_iocmd == BIO_READ) {
1587 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1588 				bp->b_rcred = crhold(cred);
1589 		} else {
1590 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1591 				bp->b_wcred = crhold(cred);
1592 		}
1593 
1594 		if (bp->b_flags & B_REMFREE)
1595 			bremfreef(bp);
1596 		BUF_KERNPROC(bp);
1597 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1598 		nmp->nm_bufqlen++;
1599 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1600 			NFSLOCKNODE(VTONFS(bp->b_vp));
1601 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1602 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1603 			NFSUNLOCKNODE(VTONFS(bp->b_vp));
1604 		}
1605 		NFSUNLOCKIOD();
1606 		return (0);
1607 	}
1608 
1609 	NFSUNLOCKIOD();
1610 
1611 	/*
1612 	 * All the iods are busy on other mounts, so return EIO to
1613 	 * force the caller to process the i/o synchronously.
1614 	 */
1615 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1616 	return (EIO);
1617 }
1618 
1619 void
1620 ncl_doio_directwrite(struct buf *bp)
1621 {
1622 	int iomode, must_commit;
1623 	struct uio *uiop = (struct uio *)bp->b_caller1;
1624 	char *iov_base = uiop->uio_iov->iov_base;
1625 
1626 	iomode = NFSWRITE_FILESYNC;
1627 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1628 	/*
1629 	 * When doing direct I/O we do not care if the
1630 	 * server's write verifier has changed, but we
1631 	 * do not want to update the verifier if it has
1632 	 * changed, since that hides the change from
1633 	 * writes being done through the buffer cache.
1634 	 * By passing must_commit in set to two, the code
1635 	 * in nfsrpc_writerpc() will not update the
1636 	 * verifier on the mount point.
1637 	 */
1638 	must_commit = 2;
1639 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1640 	KASSERT((must_commit == 2), ("ncl_doio_directwrite: Updated write"
1641 	    " verifier"));
1642 	if (iomode != NFSWRITE_FILESYNC)
1643 		printf("ncl_doio_directwrite: Broken server "
1644 		    "did not reply FILE_SYNC\n");
1645 	free(iov_base, M_NFSDIRECTIO);
1646 	free(uiop->uio_iov, M_NFSDIRECTIO);
1647 	free(uiop, M_NFSDIRECTIO);
1648 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1649 		struct nfsnode *np = VTONFS(bp->b_vp);
1650 		NFSLOCKNODE(np);
1651 		if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1652 			/*
1653 			 * Invalidate the attribute cache, since writes to a DS
1654 			 * won't update the size attribute.
1655 			 */
1656 			np->n_attrstamp = 0;
1657 		}
1658 		np->n_directio_asyncwr--;
1659 		if (np->n_directio_asyncwr == 0) {
1660 			np->n_flag &= ~NMODIFIED;
1661 			if ((np->n_flag & NFSYNCWAIT)) {
1662 				np->n_flag &= ~NFSYNCWAIT;
1663 				wakeup((caddr_t)&np->n_directio_asyncwr);
1664 			}
1665 		}
1666 		NFSUNLOCKNODE(np);
1667 	}
1668 	bp->b_vp = NULL;
1669 	uma_zfree(ncl_pbuf_zone, bp);
1670 }
1671 
1672 /*
1673  * Do an I/O operation to/from a cache block. This may be called
1674  * synchronously or from an nfsiod.
1675  */
1676 int
1677 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1678     int called_from_strategy)
1679 {
1680 	struct uio *uiop;
1681 	struct nfsnode *np;
1682 	struct nfsmount *nmp;
1683 	int error = 0, iomode, must_commit = 0;
1684 	struct uio uio;
1685 	struct iovec io;
1686 	struct proc *p = td ? td->td_proc : NULL;
1687 	uint8_t	iocmd;
1688 
1689 	np = VTONFS(vp);
1690 	nmp = VFSTONFS(vp->v_mount);
1691 	uiop = &uio;
1692 	uiop->uio_iov = &io;
1693 	uiop->uio_iovcnt = 1;
1694 	uiop->uio_segflg = UIO_SYSSPACE;
1695 	uiop->uio_td = td;
1696 
1697 	/*
1698 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1699 	 * do this here so we do not have to do it in all the code that
1700 	 * calls us.
1701 	 */
1702 	bp->b_flags &= ~B_INVAL;
1703 	bp->b_ioflags &= ~BIO_ERROR;
1704 
1705 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1706 	iocmd = bp->b_iocmd;
1707 	if (iocmd == BIO_READ) {
1708 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1709 	    io.iov_base = bp->b_data;
1710 	    uiop->uio_rw = UIO_READ;
1711 
1712 	    switch (vp->v_type) {
1713 	    case VREG:
1714 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1715 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1716 		error = ncl_readrpc(vp, uiop, cr);
1717 
1718 		if (!error) {
1719 		    if (uiop->uio_resid) {
1720 			/*
1721 			 * If we had a short read with no error, we must have
1722 			 * hit a file hole.  We should zero-fill the remainder.
1723 			 * This can also occur if the server hits the file EOF.
1724 			 *
1725 			 * Holes used to be able to occur due to pending
1726 			 * writes, but that is not possible any longer.
1727 			 */
1728 			int nread = bp->b_bcount - uiop->uio_resid;
1729 			ssize_t left = uiop->uio_resid;
1730 
1731 			if (left > 0)
1732 				bzero((char *)bp->b_data + nread, left);
1733 			uiop->uio_resid = 0;
1734 		    }
1735 		}
1736 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1737 		if (p && vp->v_writecount <= -1) {
1738 			NFSLOCKNODE(np);
1739 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1740 				NFSUNLOCKNODE(np);
1741 				PROC_LOCK(p);
1742 				killproc(p, "text file modification");
1743 				PROC_UNLOCK(p);
1744 			} else
1745 				NFSUNLOCKNODE(np);
1746 		}
1747 		break;
1748 	    case VLNK:
1749 		uiop->uio_offset = (off_t)0;
1750 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1751 		error = ncl_readlinkrpc(vp, uiop, cr);
1752 		break;
1753 	    case VDIR:
1754 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1755 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1756 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1757 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1758 			if (error == NFSERR_NOTSUPP)
1759 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1760 		}
1761 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1762 			error = ncl_readdirrpc(vp, uiop, cr, td);
1763 		/*
1764 		 * end-of-directory sets B_INVAL but does not generate an
1765 		 * error.
1766 		 */
1767 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1768 			bp->b_flags |= B_INVAL;
1769 		break;
1770 	    default:
1771 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1772 		break;
1773 	    }
1774 	    if (error) {
1775 		bp->b_ioflags |= BIO_ERROR;
1776 		bp->b_error = error;
1777 	    }
1778 	} else {
1779 	    /*
1780 	     * If we only need to commit, try to commit
1781 	     */
1782 	    if (bp->b_flags & B_NEEDCOMMIT) {
1783 		    int retv;
1784 		    off_t off;
1785 
1786 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1787 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1788 			bp->b_wcred, td);
1789 		    if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1790 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1791 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1792 			    bp->b_resid = 0;
1793 			    bufdone(bp);
1794 			    return (0);
1795 		    }
1796 		    if (retv == NFSERR_STALEWRITEVERF) {
1797 			    ncl_clearcommit(vp->v_mount);
1798 		    }
1799 	    }
1800 
1801 	    /*
1802 	     * Setup for actual write
1803 	     */
1804 	    NFSLOCKNODE(np);
1805 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1806 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1807 	    NFSUNLOCKNODE(np);
1808 
1809 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1810 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1811 		    - bp->b_dirtyoff;
1812 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1813 		    + bp->b_dirtyoff;
1814 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1815 		uiop->uio_rw = UIO_WRITE;
1816 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1817 
1818 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1819 		    iomode = NFSWRITE_UNSTABLE;
1820 		else
1821 		    iomode = NFSWRITE_FILESYNC;
1822 
1823 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1824 		    called_from_strategy);
1825 
1826 		/*
1827 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1828 		 * to cluster the buffers needing commit.  This will allow
1829 		 * the system to submit a single commit rpc for the whole
1830 		 * cluster.  We can do this even if the buffer is not 100%
1831 		 * dirty (relative to the NFS blocksize), so we optimize the
1832 		 * append-to-file-case.
1833 		 *
1834 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1835 		 * cleared because write clustering only works for commit
1836 		 * rpc's, not for the data portion of the write).
1837 		 */
1838 
1839 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1840 		    bp->b_flags |= B_NEEDCOMMIT;
1841 		    if (bp->b_dirtyoff == 0
1842 			&& bp->b_dirtyend == bp->b_bcount)
1843 			bp->b_flags |= B_CLUSTEROK;
1844 		} else {
1845 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1846 		}
1847 
1848 		/*
1849 		 * For an interrupted write, the buffer is still valid
1850 		 * and the write hasn't been pushed to the server yet,
1851 		 * so we can't set BIO_ERROR and report the interruption
1852 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1853 		 * is not relevant, so the rpc attempt is essentially
1854 		 * a noop.  For the case of a V3 write rpc not being
1855 		 * committed to stable storage, the block is still
1856 		 * dirty and requires either a commit rpc or another
1857 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1858 		 * the block is reused. This is indicated by setting
1859 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1860 		 *
1861 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1862 		 * write error and is handled as above, except that
1863 		 * B_EINTR isn't set. One cause of this is a stale stateid
1864 		 * error for the RPC that indicates recovery is required,
1865 		 * when called with called_from_strategy != 0.
1866 		 *
1867 		 * If the buffer is marked B_PAGING, it does not reside on
1868 		 * the vp's paging queues so we cannot call bdirty().  The
1869 		 * bp in this case is not an NFS cache block so we should
1870 		 * be safe. XXX
1871 		 *
1872 		 * The logic below breaks up errors into recoverable and
1873 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1874 		 * and keep the buffer around for potential write retries.
1875 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1876 		 * and save the error in the nfsnode. This is less than ideal
1877 		 * but necessary. Keeping such buffers around could potentially
1878 		 * cause buffer exhaustion eventually (they can never be written
1879 		 * out, so will get constantly be re-dirtied). It also causes
1880 		 * all sorts of vfs panics. For non-recoverable write errors,
1881 		 * also invalidate the attrcache, so we'll be forced to go over
1882 		 * the wire for this object, returning an error to user on next
1883 		 * call (most of the time).
1884 		 */
1885 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1886 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1887 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1888 			if ((bp->b_flags & B_PAGING) == 0) {
1889 			    bdirty(bp);
1890 			    bp->b_flags &= ~B_DONE;
1891 			}
1892 			if ((error == EINTR || error == ETIMEDOUT) &&
1893 			    (bp->b_flags & B_ASYNC) == 0)
1894 			    bp->b_flags |= B_EINTR;
1895 		} else {
1896 		    if (error) {
1897 			bp->b_ioflags |= BIO_ERROR;
1898 			bp->b_flags |= B_INVAL;
1899 			bp->b_error = np->n_error = error;
1900 			NFSLOCKNODE(np);
1901 			np->n_flag |= NWRITEERR;
1902 			np->n_attrstamp = 0;
1903 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1904 			NFSUNLOCKNODE(np);
1905 		    }
1906 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1907 		}
1908 	    } else {
1909 		bp->b_resid = 0;
1910 		bufdone(bp);
1911 		return (0);
1912 	    }
1913 	}
1914 	bp->b_resid = uiop->uio_resid;
1915 	if (must_commit == 1)
1916 	    ncl_clearcommit(vp->v_mount);
1917 	bufdone(bp);
1918 	return (error);
1919 }
1920 
1921 /*
1922  * Used to aid in handling ftruncate() operations on the NFS client side.
1923  * Truncation creates a number of special problems for NFS.  We have to
1924  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1925  * we have to properly handle VM pages or (potentially dirty) buffers
1926  * that straddle the truncation point.
1927  */
1928 
1929 int
1930 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1931 {
1932 	struct nfsnode *np = VTONFS(vp);
1933 	u_quad_t tsize;
1934 	int biosize = vp->v_bufobj.bo_bsize;
1935 	int error = 0;
1936 
1937 	NFSLOCKNODE(np);
1938 	tsize = np->n_size;
1939 	np->n_size = nsize;
1940 	NFSUNLOCKNODE(np);
1941 
1942 	if (nsize < tsize) {
1943 		struct buf *bp;
1944 		daddr_t lbn;
1945 		int bufsize;
1946 
1947 		/*
1948 		 * vtruncbuf() doesn't get the buffer overlapping the
1949 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1950 		 * buffer that now needs to be truncated.
1951 		 */
1952 		error = vtruncbuf(vp, nsize, biosize);
1953 		lbn = nsize / biosize;
1954 		bufsize = nsize - (lbn * biosize);
1955 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1956 		if (!bp)
1957 			return EINTR;
1958 		if (bp->b_dirtyoff > bp->b_bcount)
1959 			bp->b_dirtyoff = bp->b_bcount;
1960 		if (bp->b_dirtyend > bp->b_bcount)
1961 			bp->b_dirtyend = bp->b_bcount;
1962 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1963 		brelse(bp);
1964 	} else {
1965 		vnode_pager_setsize(vp, nsize);
1966 	}
1967 	return(error);
1968 }
1969