xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Rick Macklem at The University of Guelph.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bio.h>
43 #include <sys/buf.h>
44 #include <sys/kernel.h>
45 #include <sys/mount.h>
46 #include <sys/rwlock.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
49 
50 #include <vm/vm.h>
51 #include <vm/vm_param.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
57 
58 #include <fs/nfs/nfsport.h>
59 #include <fs/nfsclient/nfsmount.h>
60 #include <fs/nfsclient/nfs.h>
61 #include <fs/nfsclient/nfsnode.h>
62 #include <fs/nfsclient/nfs_kdtrace.h>
63 
64 extern int newnfs_directio_allow_mmap;
65 extern struct nfsstatsv1 nfsstatsv1;
66 extern struct mtx ncl_iod_mutex;
67 extern int ncl_numasync;
68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
70 extern int newnfs_directio_enable;
71 extern int nfs_keep_dirty_on_error;
72 
73 uma_zone_t ncl_pbuf_zone;
74 
75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
76     struct thread *td);
77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
78     struct ucred *cred, int ioflag);
79 
80 /*
81  * Vnode op for VM getpages.
82  */
83 SYSCTL_DECL(_vfs_nfs);
84 static int use_buf_pager = 1;
85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
86     &use_buf_pager, 0,
87     "Use buffer pager instead of direct readrpc call");
88 
89 static daddr_t
90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
91 {
92 
93 	return (off / vp->v_bufobj.bo_bsize);
94 }
95 
96 static int
97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn)
98 {
99 	struct nfsnode *np;
100 	u_quad_t nsize;
101 	int biosize, bcount;
102 
103 	np = VTONFS(vp);
104 	NFSLOCKNODE(np);
105 	nsize = np->n_size;
106 	NFSUNLOCKNODE(np);
107 
108 	biosize = vp->v_bufobj.bo_bsize;
109 	bcount = biosize;
110 	if ((off_t)lbn * biosize >= nsize)
111 		bcount = 0;
112 	else if ((off_t)(lbn + 1) * biosize > nsize)
113 		bcount = nsize - (off_t)lbn * biosize;
114 	return (bcount);
115 }
116 
117 int
118 ncl_getpages(struct vop_getpages_args *ap)
119 {
120 	int i, error, nextoff, size, toff, count, npages;
121 	struct uio uio;
122 	struct iovec iov;
123 	vm_offset_t kva;
124 	struct buf *bp;
125 	struct vnode *vp;
126 	struct thread *td;
127 	struct ucred *cred;
128 	struct nfsmount *nmp;
129 	vm_object_t object;
130 	vm_page_t *pages;
131 	struct nfsnode *np;
132 
133 	vp = ap->a_vp;
134 	np = VTONFS(vp);
135 	td = curthread;
136 	cred = curthread->td_ucred;
137 	nmp = VFSTONFS(vp->v_mount);
138 	pages = ap->a_m;
139 	npages = ap->a_count;
140 
141 	if ((object = vp->v_object) == NULL) {
142 		printf("ncl_getpages: called with non-merged cache vnode\n");
143 		return (VM_PAGER_ERROR);
144 	}
145 
146 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
147 		NFSLOCKNODE(np);
148 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
149 			NFSUNLOCKNODE(np);
150 			printf("ncl_getpages: called on non-cacheable vnode\n");
151 			return (VM_PAGER_ERROR);
152 		} else
153 			NFSUNLOCKNODE(np);
154 	}
155 
156 	mtx_lock(&nmp->nm_mtx);
157 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
158 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
159 		mtx_unlock(&nmp->nm_mtx);
160 		/* We'll never get here for v4, because we always have fsinfo */
161 		(void)ncl_fsinfo(nmp, vp, cred, td);
162 	} else
163 		mtx_unlock(&nmp->nm_mtx);
164 
165 	if (use_buf_pager)
166 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
167 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
168 
169 	/*
170 	 * If the requested page is partially valid, just return it and
171 	 * allow the pager to zero-out the blanks.  Partially valid pages
172 	 * can only occur at the file EOF.
173 	 *
174 	 * XXXGL: is that true for NFS, where short read can occur???
175 	 */
176 	VM_OBJECT_WLOCK(object);
177 	if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
178 		goto out;
179 	VM_OBJECT_WUNLOCK(object);
180 
181 	/*
182 	 * We use only the kva address for the buffer, but this is extremely
183 	 * convenient and fast.
184 	 */
185 	bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
186 
187 	kva = (vm_offset_t) bp->b_data;
188 	pmap_qenter(kva, pages, npages);
189 	VM_CNT_INC(v_vnodein);
190 	VM_CNT_ADD(v_vnodepgsin, npages);
191 
192 	count = npages << PAGE_SHIFT;
193 	iov.iov_base = (caddr_t) kva;
194 	iov.iov_len = count;
195 	uio.uio_iov = &iov;
196 	uio.uio_iovcnt = 1;
197 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
198 	uio.uio_resid = count;
199 	uio.uio_segflg = UIO_SYSSPACE;
200 	uio.uio_rw = UIO_READ;
201 	uio.uio_td = td;
202 
203 	error = ncl_readrpc(vp, &uio, cred);
204 	pmap_qremove(kva, npages);
205 
206 	uma_zfree(ncl_pbuf_zone, bp);
207 
208 	if (error && (uio.uio_resid == count)) {
209 		printf("ncl_getpages: error %d\n", error);
210 		return (VM_PAGER_ERROR);
211 	}
212 
213 	/*
214 	 * Calculate the number of bytes read and validate only that number
215 	 * of bytes.  Note that due to pending writes, size may be 0.  This
216 	 * does not mean that the remaining data is invalid!
217 	 */
218 
219 	size = count - uio.uio_resid;
220 	VM_OBJECT_WLOCK(object);
221 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
222 		vm_page_t m;
223 		nextoff = toff + PAGE_SIZE;
224 		m = pages[i];
225 
226 		if (nextoff <= size) {
227 			/*
228 			 * Read operation filled an entire page
229 			 */
230 			vm_page_valid(m);
231 			KASSERT(m->dirty == 0,
232 			    ("nfs_getpages: page %p is dirty", m));
233 		} else if (size > toff) {
234 			/*
235 			 * Read operation filled a partial page.
236 			 */
237 			vm_page_invalid(m);
238 			vm_page_set_valid_range(m, 0, size - toff);
239 			KASSERT(m->dirty == 0,
240 			    ("nfs_getpages: page %p is dirty", m));
241 		} else {
242 			/*
243 			 * Read operation was short.  If no error
244 			 * occurred we may have hit a zero-fill
245 			 * section.  We leave valid set to 0, and page
246 			 * is freed by vm_page_readahead_finish() if
247 			 * its index is not equal to requested, or
248 			 * page is zeroed and set valid by
249 			 * vm_pager_get_pages() for requested page.
250 			 */
251 			;
252 		}
253 	}
254 out:
255 	VM_OBJECT_WUNLOCK(object);
256 	if (ap->a_rbehind)
257 		*ap->a_rbehind = 0;
258 	if (ap->a_rahead)
259 		*ap->a_rahead = 0;
260 	return (VM_PAGER_OK);
261 }
262 
263 /*
264  * Vnode op for VM putpages.
265  */
266 int
267 ncl_putpages(struct vop_putpages_args *ap)
268 {
269 	struct uio uio;
270 	struct iovec iov;
271 	int i, error, npages, count;
272 	off_t offset;
273 	int *rtvals;
274 	struct vnode *vp;
275 	struct thread *td;
276 	struct ucred *cred;
277 	struct nfsmount *nmp;
278 	struct nfsnode *np;
279 	vm_page_t *pages;
280 
281 	vp = ap->a_vp;
282 	np = VTONFS(vp);
283 	td = curthread;				/* XXX */
284 	/* Set the cred to n_writecred for the write rpcs. */
285 	if (np->n_writecred != NULL)
286 		cred = crhold(np->n_writecred);
287 	else
288 		cred = crhold(curthread->td_ucred);	/* XXX */
289 	nmp = VFSTONFS(vp->v_mount);
290 	pages = ap->a_m;
291 	count = ap->a_count;
292 	rtvals = ap->a_rtvals;
293 	npages = btoc(count);
294 	offset = IDX_TO_OFF(pages[0]->pindex);
295 
296 	mtx_lock(&nmp->nm_mtx);
297 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
298 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
299 		mtx_unlock(&nmp->nm_mtx);
300 		(void)ncl_fsinfo(nmp, vp, cred, td);
301 	} else
302 		mtx_unlock(&nmp->nm_mtx);
303 
304 	NFSLOCKNODE(np);
305 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
306 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
307 		NFSUNLOCKNODE(np);
308 		printf("ncl_putpages: called on noncache-able vnode\n");
309 		NFSLOCKNODE(np);
310 	}
311 	/*
312 	 * When putting pages, do not extend file past EOF.
313 	 */
314 	if (offset + count > np->n_size) {
315 		count = np->n_size - offset;
316 		if (count < 0)
317 			count = 0;
318 	}
319 	NFSUNLOCKNODE(np);
320 
321 	for (i = 0; i < npages; i++)
322 		rtvals[i] = VM_PAGER_ERROR;
323 
324 	VM_CNT_INC(v_vnodeout);
325 	VM_CNT_ADD(v_vnodepgsout, count);
326 
327 	iov.iov_base = unmapped_buf;
328 	iov.iov_len = count;
329 	uio.uio_iov = &iov;
330 	uio.uio_iovcnt = 1;
331 	uio.uio_offset = offset;
332 	uio.uio_resid = count;
333 	uio.uio_segflg = UIO_NOCOPY;
334 	uio.uio_rw = UIO_WRITE;
335 	uio.uio_td = td;
336 
337 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
338 	    cred);
339 	crfree(cred);
340 
341 	if (error == 0 || !nfs_keep_dirty_on_error) {
342 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
343 		    np->n_size - offset, npages * PAGE_SIZE);
344 	}
345 	return (rtvals[0]);
346 }
347 
348 /*
349  * For nfs, cache consistency can only be maintained approximately.
350  * Although RFC1094 does not specify the criteria, the following is
351  * believed to be compatible with the reference port.
352  * For nfs:
353  * If the file's modify time on the server has changed since the
354  * last read rpc or you have written to the file,
355  * you may have lost data cache consistency with the
356  * server, so flush all of the file's data out of the cache.
357  * Then force a getattr rpc to ensure that you have up to date
358  * attributes.
359  * NB: This implies that cache data can be read when up to
360  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
361  * attributes this could be forced by setting n_attrstamp to 0 before
362  * the VOP_GETATTR() call.
363  */
364 static inline int
365 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
366 {
367 	int error = 0;
368 	struct vattr vattr;
369 	struct nfsnode *np = VTONFS(vp);
370 	bool old_lock;
371 
372 	/*
373 	 * Ensure the exclusove access to the node before checking
374 	 * whether the cache is consistent.
375 	 */
376 	old_lock = ncl_excl_start(vp);
377 	NFSLOCKNODE(np);
378 	if (np->n_flag & NMODIFIED) {
379 		NFSUNLOCKNODE(np);
380 		if (vp->v_type != VREG) {
381 			if (vp->v_type != VDIR)
382 				panic("nfs: bioread, not dir");
383 			ncl_invaldir(vp);
384 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
385 			if (error != 0)
386 				goto out;
387 		}
388 		np->n_attrstamp = 0;
389 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
390 		error = VOP_GETATTR(vp, &vattr, cred);
391 		if (error)
392 			goto out;
393 		NFSLOCKNODE(np);
394 		np->n_mtime = vattr.va_mtime;
395 		NFSUNLOCKNODE(np);
396 	} else {
397 		NFSUNLOCKNODE(np);
398 		error = VOP_GETATTR(vp, &vattr, cred);
399 		if (error)
400 			goto out;
401 		NFSLOCKNODE(np);
402 		if ((np->n_flag & NSIZECHANGED)
403 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
404 			NFSUNLOCKNODE(np);
405 			if (vp->v_type == VDIR)
406 				ncl_invaldir(vp);
407 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
408 			if (error != 0)
409 				goto out;
410 			NFSLOCKNODE(np);
411 			np->n_mtime = vattr.va_mtime;
412 			np->n_flag &= ~NSIZECHANGED;
413 		}
414 		NFSUNLOCKNODE(np);
415 	}
416 out:
417 	ncl_excl_finish(vp, old_lock);
418 	return (error);
419 }
420 
421 /*
422  * Vnode op for read using bio
423  */
424 int
425 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
426 {
427 	struct nfsnode *np = VTONFS(vp);
428 	struct buf *bp, *rabp;
429 	struct thread *td;
430 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
431 	daddr_t lbn, rabn;
432 	int biosize, bcount, error, i, n, nra, on, save2, seqcount;
433 	off_t tmp_off;
434 
435 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
436 	if (uio->uio_resid == 0)
437 		return (0);
438 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
439 		return (EINVAL);
440 	td = uio->uio_td;
441 
442 	mtx_lock(&nmp->nm_mtx);
443 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
444 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
445 		mtx_unlock(&nmp->nm_mtx);
446 		(void)ncl_fsinfo(nmp, vp, cred, td);
447 		mtx_lock(&nmp->nm_mtx);
448 	}
449 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
450 		(void) newnfs_iosize(nmp);
451 
452 	tmp_off = uio->uio_offset + uio->uio_resid;
453 	if (vp->v_type != VDIR &&
454 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
455 		mtx_unlock(&nmp->nm_mtx);
456 		return (EFBIG);
457 	}
458 	mtx_unlock(&nmp->nm_mtx);
459 
460 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
461 		/* No caching/ no readaheads. Just read data into the user buffer */
462 		return ncl_readrpc(vp, uio, cred);
463 
464 	n = 0;
465 	on = 0;
466 	biosize = vp->v_bufobj.bo_bsize;
467 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
468 
469 	error = nfs_bioread_check_cons(vp, td, cred);
470 	if (error)
471 		return error;
472 
473 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
474 	do {
475 	    u_quad_t nsize;
476 
477 	    NFSLOCKNODE(np);
478 	    nsize = np->n_size;
479 	    NFSUNLOCKNODE(np);
480 
481 	    switch (vp->v_type) {
482 	    case VREG:
483 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
484 		lbn = uio->uio_offset / biosize;
485 		on = uio->uio_offset - (lbn * biosize);
486 
487 		/*
488 		 * Start the read ahead(s), as required.
489 		 */
490 		if (nmp->nm_readahead > 0) {
491 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
492 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
493 			rabn = lbn + 1 + nra;
494 			if (incore(&vp->v_bufobj, rabn) == NULL) {
495 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
496 			    if (!rabp) {
497 				error = newnfs_sigintr(nmp, td);
498 				if (error == 0)
499 					error = EINTR;
500 				goto out;
501 			    }
502 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
503 				rabp->b_flags |= B_ASYNC;
504 				rabp->b_iocmd = BIO_READ;
505 				vfs_busy_pages(rabp, 0);
506 				if (ncl_asyncio(nmp, rabp, cred, td)) {
507 				    rabp->b_flags |= B_INVAL;
508 				    rabp->b_ioflags |= BIO_ERROR;
509 				    vfs_unbusy_pages(rabp);
510 				    brelse(rabp);
511 				    break;
512 				}
513 			    } else {
514 				brelse(rabp);
515 			    }
516 			}
517 		    }
518 		}
519 
520 		/* Note that bcount is *not* DEV_BSIZE aligned. */
521 		bcount = biosize;
522 		if ((off_t)lbn * biosize >= nsize) {
523 			bcount = 0;
524 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
525 			bcount = nsize - (off_t)lbn * biosize;
526 		}
527 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
528 
529 		if (!bp) {
530 			error = newnfs_sigintr(nmp, td);
531 			if (error == 0)
532 				error = EINTR;
533 			goto out;
534 		}
535 
536 		/*
537 		 * If B_CACHE is not set, we must issue the read.  If this
538 		 * fails, we return an error.
539 		 */
540 
541 		if ((bp->b_flags & B_CACHE) == 0) {
542 		    bp->b_iocmd = BIO_READ;
543 		    vfs_busy_pages(bp, 0);
544 		    error = ncl_doio(vp, bp, cred, td, 0);
545 		    if (error) {
546 			brelse(bp);
547 			goto out;
548 		    }
549 		}
550 
551 		/*
552 		 * on is the offset into the current bp.  Figure out how many
553 		 * bytes we can copy out of the bp.  Note that bcount is
554 		 * NOT DEV_BSIZE aligned.
555 		 *
556 		 * Then figure out how many bytes we can copy into the uio.
557 		 */
558 
559 		n = 0;
560 		if (on < bcount)
561 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
562 		break;
563 	    case VLNK:
564 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
565 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
566 		if (!bp) {
567 			error = newnfs_sigintr(nmp, td);
568 			if (error == 0)
569 				error = EINTR;
570 			goto out;
571 		}
572 		if ((bp->b_flags & B_CACHE) == 0) {
573 		    bp->b_iocmd = BIO_READ;
574 		    vfs_busy_pages(bp, 0);
575 		    error = ncl_doio(vp, bp, cred, td, 0);
576 		    if (error) {
577 			bp->b_ioflags |= BIO_ERROR;
578 			brelse(bp);
579 			goto out;
580 		    }
581 		}
582 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
583 		on = 0;
584 		break;
585 	    case VDIR:
586 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
587 		NFSLOCKNODE(np);
588 		if (np->n_direofoffset
589 		    && uio->uio_offset >= np->n_direofoffset) {
590 			NFSUNLOCKNODE(np);
591 			error = 0;
592 			goto out;
593 		}
594 		NFSUNLOCKNODE(np);
595 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
596 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
597 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
598 		if (!bp) {
599 			error = newnfs_sigintr(nmp, td);
600 			if (error == 0)
601 				error = EINTR;
602 			goto out;
603 		}
604 		if ((bp->b_flags & B_CACHE) == 0) {
605 		    bp->b_iocmd = BIO_READ;
606 		    vfs_busy_pages(bp, 0);
607 		    error = ncl_doio(vp, bp, cred, td, 0);
608 		    if (error) {
609 			    brelse(bp);
610 		    }
611 		    while (error == NFSERR_BAD_COOKIE) {
612 			ncl_invaldir(vp);
613 			error = ncl_vinvalbuf(vp, 0, td, 1);
614 
615 			/*
616 			 * Yuck! The directory has been modified on the
617 			 * server. The only way to get the block is by
618 			 * reading from the beginning to get all the
619 			 * offset cookies.
620 			 *
621 			 * Leave the last bp intact unless there is an error.
622 			 * Loop back up to the while if the error is another
623 			 * NFSERR_BAD_COOKIE (double yuch!).
624 			 */
625 			for (i = 0; i <= lbn && !error; i++) {
626 			    NFSLOCKNODE(np);
627 			    if (np->n_direofoffset
628 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
629 				    NFSUNLOCKNODE(np);
630 				    error = 0;
631 				    goto out;
632 			    }
633 			    NFSUNLOCKNODE(np);
634 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
635 			    if (!bp) {
636 				error = newnfs_sigintr(nmp, td);
637 				if (error == 0)
638 					error = EINTR;
639 				goto out;
640 			    }
641 			    if ((bp->b_flags & B_CACHE) == 0) {
642 				    bp->b_iocmd = BIO_READ;
643 				    vfs_busy_pages(bp, 0);
644 				    error = ncl_doio(vp, bp, cred, td, 0);
645 				    /*
646 				     * no error + B_INVAL == directory EOF,
647 				     * use the block.
648 				     */
649 				    if (error == 0 && (bp->b_flags & B_INVAL))
650 					    break;
651 			    }
652 			    /*
653 			     * An error will throw away the block and the
654 			     * for loop will break out.  If no error and this
655 			     * is not the block we want, we throw away the
656 			     * block and go for the next one via the for loop.
657 			     */
658 			    if (error || i < lbn)
659 				    brelse(bp);
660 			}
661 		    }
662 		    /*
663 		     * The above while is repeated if we hit another cookie
664 		     * error.  If we hit an error and it wasn't a cookie error,
665 		     * we give up.
666 		     */
667 		    if (error)
668 			    goto out;
669 		}
670 
671 		/*
672 		 * If not eof and read aheads are enabled, start one.
673 		 * (You need the current block first, so that you have the
674 		 *  directory offset cookie of the next block.)
675 		 */
676 		NFSLOCKNODE(np);
677 		if (nmp->nm_readahead > 0 &&
678 		    (bp->b_flags & B_INVAL) == 0 &&
679 		    (np->n_direofoffset == 0 ||
680 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
681 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
682 			NFSUNLOCKNODE(np);
683 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
684 			if (rabp) {
685 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
686 				rabp->b_flags |= B_ASYNC;
687 				rabp->b_iocmd = BIO_READ;
688 				vfs_busy_pages(rabp, 0);
689 				if (ncl_asyncio(nmp, rabp, cred, td)) {
690 				    rabp->b_flags |= B_INVAL;
691 				    rabp->b_ioflags |= BIO_ERROR;
692 				    vfs_unbusy_pages(rabp);
693 				    brelse(rabp);
694 				}
695 			    } else {
696 				brelse(rabp);
697 			    }
698 			}
699 			NFSLOCKNODE(np);
700 		}
701 		/*
702 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
703 		 * chopped for the EOF condition, we cannot tell how large
704 		 * NFS directories are going to be until we hit EOF.  So
705 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
706 		 * it just so happens that b_resid will effectively chop it
707 		 * to EOF.  *BUT* this information is lost if the buffer goes
708 		 * away and is reconstituted into a B_CACHE state ( due to
709 		 * being VMIO ) later.  So we keep track of the directory eof
710 		 * in np->n_direofoffset and chop it off as an extra step
711 		 * right here.
712 		 */
713 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
714 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
715 			n = np->n_direofoffset - uio->uio_offset;
716 		NFSUNLOCKNODE(np);
717 		break;
718 	    default:
719 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
720 		bp = NULL;
721 		break;
722 	    }
723 
724 	    if (n > 0) {
725 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
726 	    }
727 	    if (vp->v_type == VLNK)
728 		n = 0;
729 	    if (bp != NULL)
730 		brelse(bp);
731 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
732 out:
733 	curthread_pflags2_restore(save2);
734 	if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
735 		NFSLOCKNODE(np);
736 		ncl_pager_setsize(vp, NULL);
737 	}
738 	return (error);
739 }
740 
741 /*
742  * The NFS write path cannot handle iovecs with len > 1. So we need to
743  * break up iovecs accordingly (restricting them to wsize).
744  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
745  * For the ASYNC case, 2 copies are needed. The first a copy from the
746  * user buffer to a staging buffer and then a second copy from the staging
747  * buffer to mbufs. This can be optimized by copying from the user buffer
748  * directly into mbufs and passing the chain down, but that requires a
749  * fair amount of re-working of the relevant codepaths (and can be done
750  * later).
751  */
752 static int
753 nfs_directio_write(vp, uiop, cred, ioflag)
754 	struct vnode *vp;
755 	struct uio *uiop;
756 	struct ucred *cred;
757 	int ioflag;
758 {
759 	int error;
760 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
761 	struct thread *td = uiop->uio_td;
762 	int size;
763 	int wsize;
764 
765 	mtx_lock(&nmp->nm_mtx);
766 	wsize = nmp->nm_wsize;
767 	mtx_unlock(&nmp->nm_mtx);
768 	if (ioflag & IO_SYNC) {
769 		int iomode, must_commit;
770 		struct uio uio;
771 		struct iovec iov;
772 do_sync:
773 		while (uiop->uio_resid > 0) {
774 			size = MIN(uiop->uio_resid, wsize);
775 			size = MIN(uiop->uio_iov->iov_len, size);
776 			iov.iov_base = uiop->uio_iov->iov_base;
777 			iov.iov_len = size;
778 			uio.uio_iov = &iov;
779 			uio.uio_iovcnt = 1;
780 			uio.uio_offset = uiop->uio_offset;
781 			uio.uio_resid = size;
782 			uio.uio_segflg = UIO_USERSPACE;
783 			uio.uio_rw = UIO_WRITE;
784 			uio.uio_td = td;
785 			iomode = NFSWRITE_FILESYNC;
786 			error = ncl_writerpc(vp, &uio, cred, &iomode,
787 			    &must_commit, 0);
788 			KASSERT((must_commit == 0),
789 				("ncl_directio_write: Did not commit write"));
790 			if (error)
791 				return (error);
792 			uiop->uio_offset += size;
793 			uiop->uio_resid -= size;
794 			if (uiop->uio_iov->iov_len <= size) {
795 				uiop->uio_iovcnt--;
796 				uiop->uio_iov++;
797 			} else {
798 				uiop->uio_iov->iov_base =
799 					(char *)uiop->uio_iov->iov_base + size;
800 				uiop->uio_iov->iov_len -= size;
801 			}
802 		}
803 	} else {
804 		struct uio *t_uio;
805 		struct iovec *t_iov;
806 		struct buf *bp;
807 
808 		/*
809 		 * Break up the write into blocksize chunks and hand these
810 		 * over to nfsiod's for write back.
811 		 * Unfortunately, this incurs a copy of the data. Since
812 		 * the user could modify the buffer before the write is
813 		 * initiated.
814 		 *
815 		 * The obvious optimization here is that one of the 2 copies
816 		 * in the async write path can be eliminated by copying the
817 		 * data here directly into mbufs and passing the mbuf chain
818 		 * down. But that will require a fair amount of re-working
819 		 * of the code and can be done if there's enough interest
820 		 * in NFS directio access.
821 		 */
822 		while (uiop->uio_resid > 0) {
823 			size = MIN(uiop->uio_resid, wsize);
824 			size = MIN(uiop->uio_iov->iov_len, size);
825 			bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
826 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
827 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
828 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
829 			t_iov->iov_len = size;
830 			t_uio->uio_iov = t_iov;
831 			t_uio->uio_iovcnt = 1;
832 			t_uio->uio_offset = uiop->uio_offset;
833 			t_uio->uio_resid = size;
834 			t_uio->uio_segflg = UIO_SYSSPACE;
835 			t_uio->uio_rw = UIO_WRITE;
836 			t_uio->uio_td = td;
837 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
838 			    uiop->uio_segflg == UIO_SYSSPACE,
839 			    ("nfs_directio_write: Bad uio_segflg"));
840 			if (uiop->uio_segflg == UIO_USERSPACE) {
841 				error = copyin(uiop->uio_iov->iov_base,
842 				    t_iov->iov_base, size);
843 				if (error != 0)
844 					goto err_free;
845 			} else
846 				/*
847 				 * UIO_SYSSPACE may never happen, but handle
848 				 * it just in case it does.
849 				 */
850 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
851 				    size);
852 			bp->b_flags |= B_DIRECT;
853 			bp->b_iocmd = BIO_WRITE;
854 			if (cred != NOCRED) {
855 				crhold(cred);
856 				bp->b_wcred = cred;
857 			} else
858 				bp->b_wcred = NOCRED;
859 			bp->b_caller1 = (void *)t_uio;
860 			bp->b_vp = vp;
861 			error = ncl_asyncio(nmp, bp, NOCRED, td);
862 err_free:
863 			if (error) {
864 				free(t_iov->iov_base, M_NFSDIRECTIO);
865 				free(t_iov, M_NFSDIRECTIO);
866 				free(t_uio, M_NFSDIRECTIO);
867 				bp->b_vp = NULL;
868 				uma_zfree(ncl_pbuf_zone, bp);
869 				if (error == EINTR)
870 					return (error);
871 				goto do_sync;
872 			}
873 			uiop->uio_offset += size;
874 			uiop->uio_resid -= size;
875 			if (uiop->uio_iov->iov_len <= size) {
876 				uiop->uio_iovcnt--;
877 				uiop->uio_iov++;
878 			} else {
879 				uiop->uio_iov->iov_base =
880 					(char *)uiop->uio_iov->iov_base + size;
881 				uiop->uio_iov->iov_len -= size;
882 			}
883 		}
884 	}
885 	return (0);
886 }
887 
888 /*
889  * Vnode op for write using bio
890  */
891 int
892 ncl_write(struct vop_write_args *ap)
893 {
894 	int biosize;
895 	struct uio *uio = ap->a_uio;
896 	struct thread *td = uio->uio_td;
897 	struct vnode *vp = ap->a_vp;
898 	struct nfsnode *np = VTONFS(vp);
899 	struct ucred *cred = ap->a_cred;
900 	int ioflag = ap->a_ioflag;
901 	struct buf *bp;
902 	struct vattr vattr;
903 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
904 	daddr_t lbn;
905 	int bcount, noncontig_write, obcount;
906 	int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
907 	size_t orig_resid, local_resid;
908 	off_t orig_size, tmp_off;
909 
910 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
911 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
912 	    ("ncl_write proc"));
913 	if (vp->v_type != VREG)
914 		return (EIO);
915 	NFSLOCKNODE(np);
916 	if (np->n_flag & NWRITEERR) {
917 		np->n_flag &= ~NWRITEERR;
918 		NFSUNLOCKNODE(np);
919 		return (np->n_error);
920 	} else
921 		NFSUNLOCKNODE(np);
922 	mtx_lock(&nmp->nm_mtx);
923 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
924 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
925 		mtx_unlock(&nmp->nm_mtx);
926 		(void)ncl_fsinfo(nmp, vp, cred, td);
927 		mtx_lock(&nmp->nm_mtx);
928 	}
929 	if (nmp->nm_wsize == 0)
930 		(void) newnfs_iosize(nmp);
931 	mtx_unlock(&nmp->nm_mtx);
932 
933 	/*
934 	 * Synchronously flush pending buffers if we are in synchronous
935 	 * mode or if we are appending.
936 	 */
937 	if (ioflag & (IO_APPEND | IO_SYNC)) {
938 		NFSLOCKNODE(np);
939 		if (np->n_flag & NMODIFIED) {
940 			NFSUNLOCKNODE(np);
941 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
942 			/*
943 			 * Require non-blocking, synchronous writes to
944 			 * dirty files to inform the program it needs
945 			 * to fsync(2) explicitly.
946 			 */
947 			if (ioflag & IO_NDELAY)
948 				return (EAGAIN);
949 #endif
950 			np->n_attrstamp = 0;
951 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
952 			error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
953 			    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
954 			if (error != 0)
955 				return (error);
956 		} else
957 			NFSUNLOCKNODE(np);
958 	}
959 
960 	orig_resid = uio->uio_resid;
961 	NFSLOCKNODE(np);
962 	orig_size = np->n_size;
963 	NFSUNLOCKNODE(np);
964 
965 	/*
966 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
967 	 * get the append lock.
968 	 */
969 	if (ioflag & IO_APPEND) {
970 		np->n_attrstamp = 0;
971 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
972 		error = VOP_GETATTR(vp, &vattr, cred);
973 		if (error)
974 			return (error);
975 		NFSLOCKNODE(np);
976 		uio->uio_offset = np->n_size;
977 		NFSUNLOCKNODE(np);
978 	}
979 
980 	if (uio->uio_offset < 0)
981 		return (EINVAL);
982 	tmp_off = uio->uio_offset + uio->uio_resid;
983 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
984 		return (EFBIG);
985 	if (uio->uio_resid == 0)
986 		return (0);
987 
988 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
989 		return nfs_directio_write(vp, uio, cred, ioflag);
990 
991 	/*
992 	 * Maybe this should be above the vnode op call, but so long as
993 	 * file servers have no limits, i don't think it matters
994 	 */
995 	if (vn_rlimit_fsize(vp, uio, td))
996 		return (EFBIG);
997 
998 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
999 	biosize = vp->v_bufobj.bo_bsize;
1000 	/*
1001 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
1002 	 * would exceed the local maximum per-file write commit size when
1003 	 * combined with those, we must decide whether to flush,
1004 	 * go synchronous, or return error.  We don't bother checking
1005 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
1006 	 * no point optimizing for something that really won't ever happen.
1007 	 */
1008 	wouldcommit = 0;
1009 	if (!(ioflag & IO_SYNC)) {
1010 		int nflag;
1011 
1012 		NFSLOCKNODE(np);
1013 		nflag = np->n_flag;
1014 		NFSUNLOCKNODE(np);
1015 		if (nflag & NMODIFIED) {
1016 			BO_LOCK(&vp->v_bufobj);
1017 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1018 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1019 				    b_bobufs) {
1020 					if (bp->b_flags & B_NEEDCOMMIT)
1021 						wouldcommit += bp->b_bcount;
1022 				}
1023 			}
1024 			BO_UNLOCK(&vp->v_bufobj);
1025 		}
1026 	}
1027 
1028 	do {
1029 		if (!(ioflag & IO_SYNC)) {
1030 			wouldcommit += biosize;
1031 			if (wouldcommit > nmp->nm_wcommitsize) {
1032 				np->n_attrstamp = 0;
1033 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1034 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1035 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1036 				if (error != 0)
1037 					goto out;
1038 				wouldcommit = biosize;
1039 			}
1040 		}
1041 
1042 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1043 		lbn = uio->uio_offset / biosize;
1044 		on = uio->uio_offset - (lbn * biosize);
1045 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1046 again:
1047 		/*
1048 		 * Handle direct append and file extension cases, calculate
1049 		 * unaligned buffer size.
1050 		 */
1051 		NFSLOCKNODE(np);
1052 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1053 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1054 			noncontig_write = 1;
1055 		else
1056 			noncontig_write = 0;
1057 		if ((uio->uio_offset == np->n_size ||
1058 		    (noncontig_write != 0 &&
1059 		    lbn == (np->n_size / biosize) &&
1060 		    uio->uio_offset + n > np->n_size)) && n) {
1061 			NFSUNLOCKNODE(np);
1062 			/*
1063 			 * Get the buffer (in its pre-append state to maintain
1064 			 * B_CACHE if it was previously set).  Resize the
1065 			 * nfsnode after we have locked the buffer to prevent
1066 			 * readers from reading garbage.
1067 			 */
1068 			obcount = np->n_size - (lbn * biosize);
1069 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1070 
1071 			if (bp != NULL) {
1072 				long save;
1073 
1074 				NFSLOCKNODE(np);
1075 				np->n_size = uio->uio_offset + n;
1076 				np->n_flag |= NMODIFIED;
1077 				np->n_flag &= ~NVNSETSZSKIP;
1078 				vnode_pager_setsize(vp, np->n_size);
1079 				NFSUNLOCKNODE(np);
1080 
1081 				save = bp->b_flags & B_CACHE;
1082 				bcount = on + n;
1083 				allocbuf(bp, bcount);
1084 				bp->b_flags |= save;
1085 				if (noncontig_write != 0 && on > obcount)
1086 					vfs_bio_bzero_buf(bp, obcount, on -
1087 					    obcount);
1088 			}
1089 		} else {
1090 			/*
1091 			 * Obtain the locked cache block first, and then
1092 			 * adjust the file's size as appropriate.
1093 			 */
1094 			bcount = on + n;
1095 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1096 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1097 					bcount = biosize;
1098 				else
1099 					bcount = np->n_size - (off_t)lbn * biosize;
1100 			}
1101 			NFSUNLOCKNODE(np);
1102 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1103 			NFSLOCKNODE(np);
1104 			if (uio->uio_offset + n > np->n_size) {
1105 				np->n_size = uio->uio_offset + n;
1106 				np->n_flag |= NMODIFIED;
1107 				np->n_flag &= ~NVNSETSZSKIP;
1108 				vnode_pager_setsize(vp, np->n_size);
1109 			}
1110 			NFSUNLOCKNODE(np);
1111 		}
1112 
1113 		if (!bp) {
1114 			error = newnfs_sigintr(nmp, td);
1115 			if (!error)
1116 				error = EINTR;
1117 			break;
1118 		}
1119 
1120 		/*
1121 		 * Issue a READ if B_CACHE is not set.  In special-append
1122 		 * mode, B_CACHE is based on the buffer prior to the write
1123 		 * op and is typically set, avoiding the read.  If a read
1124 		 * is required in special append mode, the server will
1125 		 * probably send us a short-read since we extended the file
1126 		 * on our end, resulting in b_resid == 0 and, thusly,
1127 		 * B_CACHE getting set.
1128 		 *
1129 		 * We can also avoid issuing the read if the write covers
1130 		 * the entire buffer.  We have to make sure the buffer state
1131 		 * is reasonable in this case since we will not be initiating
1132 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1133 		 * more information.
1134 		 *
1135 		 * B_CACHE may also be set due to the buffer being cached
1136 		 * normally.
1137 		 */
1138 
1139 		bp_cached = 1;
1140 		if (on == 0 && n == bcount) {
1141 			if ((bp->b_flags & B_CACHE) == 0)
1142 				bp_cached = 0;
1143 			bp->b_flags |= B_CACHE;
1144 			bp->b_flags &= ~B_INVAL;
1145 			bp->b_ioflags &= ~BIO_ERROR;
1146 		}
1147 
1148 		if ((bp->b_flags & B_CACHE) == 0) {
1149 			bp->b_iocmd = BIO_READ;
1150 			vfs_busy_pages(bp, 0);
1151 			error = ncl_doio(vp, bp, cred, td, 0);
1152 			if (error) {
1153 				brelse(bp);
1154 				break;
1155 			}
1156 		}
1157 		if (bp->b_wcred == NOCRED)
1158 			bp->b_wcred = crhold(cred);
1159 		NFSLOCKNODE(np);
1160 		np->n_flag |= NMODIFIED;
1161 		NFSUNLOCKNODE(np);
1162 
1163 		/*
1164 		 * If dirtyend exceeds file size, chop it down.  This should
1165 		 * not normally occur but there is an append race where it
1166 		 * might occur XXX, so we log it.
1167 		 *
1168 		 * If the chopping creates a reverse-indexed or degenerate
1169 		 * situation with dirtyoff/end, we 0 both of them.
1170 		 */
1171 
1172 		if (bp->b_dirtyend > bcount) {
1173 			printf("NFS append race @%lx:%d\n",
1174 			    (long)bp->b_blkno * DEV_BSIZE,
1175 			    bp->b_dirtyend - bcount);
1176 			bp->b_dirtyend = bcount;
1177 		}
1178 
1179 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1180 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1181 
1182 		/*
1183 		 * If the new write will leave a contiguous dirty
1184 		 * area, just update the b_dirtyoff and b_dirtyend,
1185 		 * otherwise force a write rpc of the old dirty area.
1186 		 *
1187 		 * If there has been a file lock applied to this file
1188 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1189 		 * While it is possible to merge discontiguous writes due to
1190 		 * our having a B_CACHE buffer ( and thus valid read data
1191 		 * for the hole), we don't because it could lead to
1192 		 * significant cache coherency problems with multiple clients,
1193 		 * especially if locking is implemented later on.
1194 		 *
1195 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1196 		 * not been file locking done on this file:
1197 		 * Relax coherency a bit for the sake of performance and
1198 		 * expand the current dirty region to contain the new
1199 		 * write even if it means we mark some non-dirty data as
1200 		 * dirty.
1201 		 */
1202 
1203 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1204 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1205 			if (bwrite(bp) == EINTR) {
1206 				error = EINTR;
1207 				break;
1208 			}
1209 			goto again;
1210 		}
1211 
1212 		local_resid = uio->uio_resid;
1213 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1214 
1215 		if (error != 0 && !bp_cached) {
1216 			/*
1217 			 * This block has no other content then what
1218 			 * possibly was written by the faulty uiomove.
1219 			 * Release it, forgetting the data pages, to
1220 			 * prevent the leak of uninitialized data to
1221 			 * usermode.
1222 			 */
1223 			bp->b_ioflags |= BIO_ERROR;
1224 			brelse(bp);
1225 			uio->uio_offset -= local_resid - uio->uio_resid;
1226 			uio->uio_resid = local_resid;
1227 			break;
1228 		}
1229 
1230 		/*
1231 		 * Since this block is being modified, it must be written
1232 		 * again and not just committed.  Since write clustering does
1233 		 * not work for the stage 1 data write, only the stage 2
1234 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1235 		 */
1236 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1237 
1238 		/*
1239 		 * Get the partial update on the progress made from
1240 		 * uiomove, if an error occurred.
1241 		 */
1242 		if (error != 0)
1243 			n = local_resid - uio->uio_resid;
1244 
1245 		/*
1246 		 * Only update dirtyoff/dirtyend if not a degenerate
1247 		 * condition.
1248 		 */
1249 		if (n > 0) {
1250 			if (bp->b_dirtyend > 0) {
1251 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1252 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1253 			} else {
1254 				bp->b_dirtyoff = on;
1255 				bp->b_dirtyend = on + n;
1256 			}
1257 			vfs_bio_set_valid(bp, on, n);
1258 		}
1259 
1260 		/*
1261 		 * If IO_SYNC do bwrite().
1262 		 *
1263 		 * IO_INVAL appears to be unused.  The idea appears to be
1264 		 * to turn off caching in this case.  Very odd.  XXX
1265 		 */
1266 		if ((ioflag & IO_SYNC)) {
1267 			if (ioflag & IO_INVAL)
1268 				bp->b_flags |= B_NOCACHE;
1269 			error1 = bwrite(bp);
1270 			if (error1 != 0) {
1271 				if (error == 0)
1272 					error = error1;
1273 				break;
1274 			}
1275 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1276 			bp->b_flags |= B_ASYNC;
1277 			(void) ncl_writebp(bp, 0, NULL);
1278 		} else {
1279 			bdwrite(bp);
1280 		}
1281 
1282 		if (error != 0)
1283 			break;
1284 	} while (uio->uio_resid > 0 && n > 0);
1285 
1286 	if (error != 0) {
1287 		if (ioflag & IO_UNIT) {
1288 			VATTR_NULL(&vattr);
1289 			vattr.va_size = orig_size;
1290 			/* IO_SYNC is handled implicitely */
1291 			(void)VOP_SETATTR(vp, &vattr, cred);
1292 			uio->uio_offset -= orig_resid - uio->uio_resid;
1293 			uio->uio_resid = orig_resid;
1294 		}
1295 	}
1296 
1297 out:
1298 	curthread_pflags2_restore(save2);
1299 	return (error);
1300 }
1301 
1302 /*
1303  * Get an nfs cache block.
1304  *
1305  * Allocate a new one if the block isn't currently in the cache
1306  * and return the block marked busy. If the calling process is
1307  * interrupted by a signal for an interruptible mount point, return
1308  * NULL.
1309  *
1310  * The caller must carefully deal with the possible B_INVAL state of
1311  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1312  * indirectly), so synchronous reads can be issued without worrying about
1313  * the B_INVAL state.  We have to be a little more careful when dealing
1314  * with writes (see comments in nfs_write()) when extending a file past
1315  * its EOF.
1316  */
1317 static struct buf *
1318 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1319 {
1320 	struct buf *bp;
1321 	struct mount *mp;
1322 	struct nfsmount *nmp;
1323 
1324 	mp = vp->v_mount;
1325 	nmp = VFSTONFS(mp);
1326 
1327 	if (nmp->nm_flag & NFSMNT_INT) {
1328 		sigset_t oldset;
1329 
1330 		newnfs_set_sigmask(td, &oldset);
1331 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1332 		newnfs_restore_sigmask(td, &oldset);
1333 		while (bp == NULL) {
1334 			if (newnfs_sigintr(nmp, td))
1335 				return (NULL);
1336 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1337 		}
1338 	} else {
1339 		bp = getblk(vp, bn, size, 0, 0, 0);
1340 	}
1341 
1342 	if (vp->v_type == VREG)
1343 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1344 	return (bp);
1345 }
1346 
1347 /*
1348  * Flush and invalidate all dirty buffers. If another process is already
1349  * doing the flush, just wait for completion.
1350  */
1351 int
1352 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1353 {
1354 	struct nfsnode *np = VTONFS(vp);
1355 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1356 	int error = 0, slpflag, slptimeo;
1357 	bool old_lock;
1358 
1359 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1360 
1361 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1362 		intrflg = 0;
1363 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1364 		intrflg = 1;
1365 	if (intrflg) {
1366 		slpflag = PCATCH;
1367 		slptimeo = 2 * hz;
1368 	} else {
1369 		slpflag = 0;
1370 		slptimeo = 0;
1371 	}
1372 
1373 	old_lock = ncl_excl_start(vp);
1374 	if (old_lock)
1375 		flags |= V_ALLOWCLEAN;
1376 
1377 	/*
1378 	 * Now, flush as required.
1379 	 */
1380 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1381 	     vp->v_bufobj.bo_object != NULL) {
1382 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1383 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1384 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1385 		/*
1386 		 * If the page clean was interrupted, fail the invalidation.
1387 		 * Not doing so, we run the risk of losing dirty pages in the
1388 		 * vinvalbuf() call below.
1389 		 */
1390 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1391 			goto out;
1392 	}
1393 
1394 	error = vinvalbuf(vp, flags, slpflag, 0);
1395 	while (error) {
1396 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1397 			goto out;
1398 		error = vinvalbuf(vp, flags, 0, slptimeo);
1399 	}
1400 	if (NFSHASPNFS(nmp)) {
1401 		nfscl_layoutcommit(vp, td);
1402 		/*
1403 		 * Invalidate the attribute cache, since writes to a DS
1404 		 * won't update the size attribute.
1405 		 */
1406 		NFSLOCKNODE(np);
1407 		np->n_attrstamp = 0;
1408 	} else
1409 		NFSLOCKNODE(np);
1410 	if (np->n_directio_asyncwr == 0)
1411 		np->n_flag &= ~NMODIFIED;
1412 	NFSUNLOCKNODE(np);
1413 out:
1414 	ncl_excl_finish(vp, old_lock);
1415 	return error;
1416 }
1417 
1418 /*
1419  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1420  * This is mainly to avoid queueing async I/O requests when the nfsiods
1421  * are all hung on a dead server.
1422  *
1423  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1424  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1425  */
1426 int
1427 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1428 {
1429 	int iod;
1430 	int gotiod;
1431 	int slpflag = 0;
1432 	int slptimeo = 0;
1433 	int error, error2;
1434 
1435 	/*
1436 	 * Commits are usually short and sweet so lets save some cpu and
1437 	 * leave the async daemons for more important rpc's (such as reads
1438 	 * and writes).
1439 	 *
1440 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1441 	 * in the directory in order to update attributes. This can deadlock
1442 	 * with another thread that is waiting for async I/O to be done by
1443 	 * an nfsiod thread while holding a lock on one of these vnodes.
1444 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1445 	 * perform Readdirplus RPCs.
1446 	 */
1447 	NFSLOCKIOD();
1448 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1449 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1450 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1451 		NFSUNLOCKIOD();
1452 		return(EIO);
1453 	}
1454 again:
1455 	if (nmp->nm_flag & NFSMNT_INT)
1456 		slpflag = PCATCH;
1457 	gotiod = FALSE;
1458 
1459 	/*
1460 	 * Find a free iod to process this request.
1461 	 */
1462 	for (iod = 0; iod < ncl_numasync; iod++)
1463 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1464 			gotiod = TRUE;
1465 			break;
1466 		}
1467 
1468 	/*
1469 	 * Try to create one if none are free.
1470 	 */
1471 	if (!gotiod)
1472 		ncl_nfsiodnew();
1473 	else {
1474 		/*
1475 		 * Found one, so wake it up and tell it which
1476 		 * mount to process.
1477 		 */
1478 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1479 		    iod, nmp));
1480 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1481 		ncl_iodmount[iod] = nmp;
1482 		nmp->nm_bufqiods++;
1483 		wakeup(&ncl_iodwant[iod]);
1484 	}
1485 
1486 	/*
1487 	 * If none are free, we may already have an iod working on this mount
1488 	 * point.  If so, it will process our request.
1489 	 */
1490 	if (!gotiod) {
1491 		if (nmp->nm_bufqiods > 0) {
1492 			NFS_DPF(ASYNCIO,
1493 				("ncl_asyncio: %d iods are already processing mount %p\n",
1494 				 nmp->nm_bufqiods, nmp));
1495 			gotiod = TRUE;
1496 		}
1497 	}
1498 
1499 	/*
1500 	 * If we have an iod which can process the request, then queue
1501 	 * the buffer.
1502 	 */
1503 	if (gotiod) {
1504 		/*
1505 		 * Ensure that the queue never grows too large.  We still want
1506 		 * to asynchronize so we block rather then return EIO.
1507 		 */
1508 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1509 			NFS_DPF(ASYNCIO,
1510 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1511 			nmp->nm_bufqwant = TRUE;
1512 			error = newnfs_msleep(td, &nmp->nm_bufq,
1513 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1514 			   slptimeo);
1515 			if (error) {
1516 				error2 = newnfs_sigintr(nmp, td);
1517 				if (error2) {
1518 					NFSUNLOCKIOD();
1519 					return (error2);
1520 				}
1521 				if (slpflag == PCATCH) {
1522 					slpflag = 0;
1523 					slptimeo = 2 * hz;
1524 				}
1525 			}
1526 			/*
1527 			 * We might have lost our iod while sleeping,
1528 			 * so check and loop if necessary.
1529 			 */
1530 			goto again;
1531 		}
1532 
1533 		/* We might have lost our nfsiod */
1534 		if (nmp->nm_bufqiods == 0) {
1535 			NFS_DPF(ASYNCIO,
1536 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1537 			goto again;
1538 		}
1539 
1540 		if (bp->b_iocmd == BIO_READ) {
1541 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1542 				bp->b_rcred = crhold(cred);
1543 		} else {
1544 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1545 				bp->b_wcred = crhold(cred);
1546 		}
1547 
1548 		if (bp->b_flags & B_REMFREE)
1549 			bremfreef(bp);
1550 		BUF_KERNPROC(bp);
1551 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1552 		nmp->nm_bufqlen++;
1553 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1554 			NFSLOCKNODE(VTONFS(bp->b_vp));
1555 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1556 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1557 			NFSUNLOCKNODE(VTONFS(bp->b_vp));
1558 		}
1559 		NFSUNLOCKIOD();
1560 		return (0);
1561 	}
1562 
1563 	NFSUNLOCKIOD();
1564 
1565 	/*
1566 	 * All the iods are busy on other mounts, so return EIO to
1567 	 * force the caller to process the i/o synchronously.
1568 	 */
1569 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1570 	return (EIO);
1571 }
1572 
1573 void
1574 ncl_doio_directwrite(struct buf *bp)
1575 {
1576 	int iomode, must_commit;
1577 	struct uio *uiop = (struct uio *)bp->b_caller1;
1578 	char *iov_base = uiop->uio_iov->iov_base;
1579 
1580 	iomode = NFSWRITE_FILESYNC;
1581 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1582 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1583 	KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1584 	free(iov_base, M_NFSDIRECTIO);
1585 	free(uiop->uio_iov, M_NFSDIRECTIO);
1586 	free(uiop, M_NFSDIRECTIO);
1587 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1588 		struct nfsnode *np = VTONFS(bp->b_vp);
1589 		NFSLOCKNODE(np);
1590 		if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1591 			/*
1592 			 * Invalidate the attribute cache, since writes to a DS
1593 			 * won't update the size attribute.
1594 			 */
1595 			np->n_attrstamp = 0;
1596 		}
1597 		np->n_directio_asyncwr--;
1598 		if (np->n_directio_asyncwr == 0) {
1599 			np->n_flag &= ~NMODIFIED;
1600 			if ((np->n_flag & NFSYNCWAIT)) {
1601 				np->n_flag &= ~NFSYNCWAIT;
1602 				wakeup((caddr_t)&np->n_directio_asyncwr);
1603 			}
1604 		}
1605 		NFSUNLOCKNODE(np);
1606 	}
1607 	bp->b_vp = NULL;
1608 	uma_zfree(ncl_pbuf_zone, bp);
1609 }
1610 
1611 /*
1612  * Do an I/O operation to/from a cache block. This may be called
1613  * synchronously or from an nfsiod.
1614  */
1615 int
1616 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1617     int called_from_strategy)
1618 {
1619 	struct uio *uiop;
1620 	struct nfsnode *np;
1621 	struct nfsmount *nmp;
1622 	int error = 0, iomode, must_commit = 0;
1623 	struct uio uio;
1624 	struct iovec io;
1625 	struct proc *p = td ? td->td_proc : NULL;
1626 	uint8_t	iocmd;
1627 
1628 	np = VTONFS(vp);
1629 	nmp = VFSTONFS(vp->v_mount);
1630 	uiop = &uio;
1631 	uiop->uio_iov = &io;
1632 	uiop->uio_iovcnt = 1;
1633 	uiop->uio_segflg = UIO_SYSSPACE;
1634 	uiop->uio_td = td;
1635 
1636 	/*
1637 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1638 	 * do this here so we do not have to do it in all the code that
1639 	 * calls us.
1640 	 */
1641 	bp->b_flags &= ~B_INVAL;
1642 	bp->b_ioflags &= ~BIO_ERROR;
1643 
1644 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1645 	iocmd = bp->b_iocmd;
1646 	if (iocmd == BIO_READ) {
1647 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1648 	    io.iov_base = bp->b_data;
1649 	    uiop->uio_rw = UIO_READ;
1650 
1651 	    switch (vp->v_type) {
1652 	    case VREG:
1653 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1654 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1655 		error = ncl_readrpc(vp, uiop, cr);
1656 
1657 		if (!error) {
1658 		    if (uiop->uio_resid) {
1659 			/*
1660 			 * If we had a short read with no error, we must have
1661 			 * hit a file hole.  We should zero-fill the remainder.
1662 			 * This can also occur if the server hits the file EOF.
1663 			 *
1664 			 * Holes used to be able to occur due to pending
1665 			 * writes, but that is not possible any longer.
1666 			 */
1667 			int nread = bp->b_bcount - uiop->uio_resid;
1668 			ssize_t left = uiop->uio_resid;
1669 
1670 			if (left > 0)
1671 				bzero((char *)bp->b_data + nread, left);
1672 			uiop->uio_resid = 0;
1673 		    }
1674 		}
1675 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1676 		if (p && vp->v_writecount <= -1) {
1677 			NFSLOCKNODE(np);
1678 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1679 				NFSUNLOCKNODE(np);
1680 				PROC_LOCK(p);
1681 				killproc(p, "text file modification");
1682 				PROC_UNLOCK(p);
1683 			} else
1684 				NFSUNLOCKNODE(np);
1685 		}
1686 		break;
1687 	    case VLNK:
1688 		uiop->uio_offset = (off_t)0;
1689 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1690 		error = ncl_readlinkrpc(vp, uiop, cr);
1691 		break;
1692 	    case VDIR:
1693 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1694 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1695 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1696 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1697 			if (error == NFSERR_NOTSUPP)
1698 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1699 		}
1700 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1701 			error = ncl_readdirrpc(vp, uiop, cr, td);
1702 		/*
1703 		 * end-of-directory sets B_INVAL but does not generate an
1704 		 * error.
1705 		 */
1706 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1707 			bp->b_flags |= B_INVAL;
1708 		break;
1709 	    default:
1710 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1711 		break;
1712 	    }
1713 	    if (error) {
1714 		bp->b_ioflags |= BIO_ERROR;
1715 		bp->b_error = error;
1716 	    }
1717 	} else {
1718 	    /*
1719 	     * If we only need to commit, try to commit
1720 	     */
1721 	    if (bp->b_flags & B_NEEDCOMMIT) {
1722 		    int retv;
1723 		    off_t off;
1724 
1725 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1726 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1727 			bp->b_wcred, td);
1728 		    if (retv == 0) {
1729 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1730 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1731 			    bp->b_resid = 0;
1732 			    bufdone(bp);
1733 			    return (0);
1734 		    }
1735 		    if (retv == NFSERR_STALEWRITEVERF) {
1736 			    ncl_clearcommit(vp->v_mount);
1737 		    }
1738 	    }
1739 
1740 	    /*
1741 	     * Setup for actual write
1742 	     */
1743 	    NFSLOCKNODE(np);
1744 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1745 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1746 	    NFSUNLOCKNODE(np);
1747 
1748 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1749 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1750 		    - bp->b_dirtyoff;
1751 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1752 		    + bp->b_dirtyoff;
1753 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1754 		uiop->uio_rw = UIO_WRITE;
1755 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1756 
1757 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1758 		    iomode = NFSWRITE_UNSTABLE;
1759 		else
1760 		    iomode = NFSWRITE_FILESYNC;
1761 
1762 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1763 		    called_from_strategy);
1764 
1765 		/*
1766 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1767 		 * to cluster the buffers needing commit.  This will allow
1768 		 * the system to submit a single commit rpc for the whole
1769 		 * cluster.  We can do this even if the buffer is not 100%
1770 		 * dirty (relative to the NFS blocksize), so we optimize the
1771 		 * append-to-file-case.
1772 		 *
1773 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1774 		 * cleared because write clustering only works for commit
1775 		 * rpc's, not for the data portion of the write).
1776 		 */
1777 
1778 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1779 		    bp->b_flags |= B_NEEDCOMMIT;
1780 		    if (bp->b_dirtyoff == 0
1781 			&& bp->b_dirtyend == bp->b_bcount)
1782 			bp->b_flags |= B_CLUSTEROK;
1783 		} else {
1784 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1785 		}
1786 
1787 		/*
1788 		 * For an interrupted write, the buffer is still valid
1789 		 * and the write hasn't been pushed to the server yet,
1790 		 * so we can't set BIO_ERROR and report the interruption
1791 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1792 		 * is not relevant, so the rpc attempt is essentially
1793 		 * a noop.  For the case of a V3 write rpc not being
1794 		 * committed to stable storage, the block is still
1795 		 * dirty and requires either a commit rpc or another
1796 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1797 		 * the block is reused. This is indicated by setting
1798 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1799 		 *
1800 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1801 		 * write error and is handled as above, except that
1802 		 * B_EINTR isn't set. One cause of this is a stale stateid
1803 		 * error for the RPC that indicates recovery is required,
1804 		 * when called with called_from_strategy != 0.
1805 		 *
1806 		 * If the buffer is marked B_PAGING, it does not reside on
1807 		 * the vp's paging queues so we cannot call bdirty().  The
1808 		 * bp in this case is not an NFS cache block so we should
1809 		 * be safe. XXX
1810 		 *
1811 		 * The logic below breaks up errors into recoverable and
1812 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1813 		 * and keep the buffer around for potential write retries.
1814 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1815 		 * and save the error in the nfsnode. This is less than ideal
1816 		 * but necessary. Keeping such buffers around could potentially
1817 		 * cause buffer exhaustion eventually (they can never be written
1818 		 * out, so will get constantly be re-dirtied). It also causes
1819 		 * all sorts of vfs panics. For non-recoverable write errors,
1820 		 * also invalidate the attrcache, so we'll be forced to go over
1821 		 * the wire for this object, returning an error to user on next
1822 		 * call (most of the time).
1823 		 */
1824 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1825 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1826 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1827 			if ((bp->b_flags & B_PAGING) == 0) {
1828 			    bdirty(bp);
1829 			    bp->b_flags &= ~B_DONE;
1830 			}
1831 			if ((error == EINTR || error == ETIMEDOUT) &&
1832 			    (bp->b_flags & B_ASYNC) == 0)
1833 			    bp->b_flags |= B_EINTR;
1834 		} else {
1835 		    if (error) {
1836 			bp->b_ioflags |= BIO_ERROR;
1837 			bp->b_flags |= B_INVAL;
1838 			bp->b_error = np->n_error = error;
1839 			NFSLOCKNODE(np);
1840 			np->n_flag |= NWRITEERR;
1841 			np->n_attrstamp = 0;
1842 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1843 			NFSUNLOCKNODE(np);
1844 		    }
1845 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1846 		}
1847 	    } else {
1848 		bp->b_resid = 0;
1849 		bufdone(bp);
1850 		return (0);
1851 	    }
1852 	}
1853 	bp->b_resid = uiop->uio_resid;
1854 	if (must_commit)
1855 	    ncl_clearcommit(vp->v_mount);
1856 	bufdone(bp);
1857 	return (error);
1858 }
1859 
1860 /*
1861  * Used to aid in handling ftruncate() operations on the NFS client side.
1862  * Truncation creates a number of special problems for NFS.  We have to
1863  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1864  * we have to properly handle VM pages or (potentially dirty) buffers
1865  * that straddle the truncation point.
1866  */
1867 
1868 int
1869 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1870 {
1871 	struct nfsnode *np = VTONFS(vp);
1872 	u_quad_t tsize;
1873 	int biosize = vp->v_bufobj.bo_bsize;
1874 	int error = 0;
1875 
1876 	NFSLOCKNODE(np);
1877 	tsize = np->n_size;
1878 	np->n_size = nsize;
1879 	NFSUNLOCKNODE(np);
1880 
1881 	if (nsize < tsize) {
1882 		struct buf *bp;
1883 		daddr_t lbn;
1884 		int bufsize;
1885 
1886 		/*
1887 		 * vtruncbuf() doesn't get the buffer overlapping the
1888 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1889 		 * buffer that now needs to be truncated.
1890 		 */
1891 		error = vtruncbuf(vp, nsize, biosize);
1892 		lbn = nsize / biosize;
1893 		bufsize = nsize - (lbn * biosize);
1894 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1895 		if (!bp)
1896 			return EINTR;
1897 		if (bp->b_dirtyoff > bp->b_bcount)
1898 			bp->b_dirtyoff = bp->b_bcount;
1899 		if (bp->b_dirtyend > bp->b_bcount)
1900 			bp->b_dirtyend = bp->b_bcount;
1901 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1902 		brelse(bp);
1903 	} else {
1904 		vnode_pager_setsize(vp, nsize);
1905 	}
1906 	return(error);
1907 }
1908