xref: /freebsd/sys/fs/nfsclient/nfs_clbio.c (revision 734e82fe33aa764367791a7d603b383996c6b40b)
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 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/bio.h>
41 #include <sys/buf.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/rwlock.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_pager.h>
54 #include <vm/vnode_pager.h>
55 
56 #include <fs/nfs/nfsport.h>
57 #include <fs/nfsclient/nfsmount.h>
58 #include <fs/nfsclient/nfs.h>
59 #include <fs/nfsclient/nfsnode.h>
60 #include <fs/nfsclient/nfs_kdtrace.h>
61 
62 extern int newnfs_directio_allow_mmap;
63 extern struct nfsstatsv1 nfsstatsv1;
64 extern struct mtx ncl_iod_mutex;
65 extern int ncl_numasync;
66 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
67 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
68 extern int newnfs_directio_enable;
69 extern int nfs_keep_dirty_on_error;
70 
71 uma_zone_t ncl_pbuf_zone;
72 
73 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
74     struct thread *td);
75 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
76     struct ucred *cred, int ioflag);
77 
78 /*
79  * Vnode op for VM getpages.
80  */
81 SYSCTL_DECL(_vfs_nfs);
82 static int use_buf_pager = 1;
83 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
84     &use_buf_pager, 0,
85     "Use buffer pager instead of direct readrpc call");
86 
87 static daddr_t
88 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
89 {
90 
91 	return (off / vp->v_bufobj.bo_bsize);
92 }
93 
94 static int
95 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz)
96 {
97 	struct nfsnode *np;
98 	u_quad_t nsize;
99 	int biosize, bcount;
100 
101 	np = VTONFS(vp);
102 	NFSLOCKNODE(np);
103 	nsize = np->n_size;
104 	NFSUNLOCKNODE(np);
105 
106 	biosize = vp->v_bufobj.bo_bsize;
107 	bcount = biosize;
108 	if ((off_t)lbn * biosize >= nsize)
109 		bcount = 0;
110 	else if ((off_t)(lbn + 1) * biosize > nsize)
111 		bcount = nsize - (off_t)lbn * biosize;
112 	*sz = bcount;
113 	return (0);
114 }
115 
116 int
117 ncl_getpages(struct vop_getpages_args *ap)
118 {
119 	int i, error, nextoff, size, toff, count, npages;
120 	struct uio uio;
121 	struct iovec iov;
122 	vm_offset_t kva;
123 	struct buf *bp;
124 	struct vnode *vp;
125 	struct thread *td;
126 	struct ucred *cred;
127 	struct nfsmount *nmp;
128 	vm_object_t object;
129 	vm_page_t *pages;
130 	struct nfsnode *np;
131 
132 	vp = ap->a_vp;
133 	np = VTONFS(vp);
134 	td = curthread;
135 	cred = curthread->td_ucred;
136 	nmp = VFSTONFS(vp->v_mount);
137 	pages = ap->a_m;
138 	npages = ap->a_count;
139 
140 	if ((object = vp->v_object) == NULL) {
141 		printf("ncl_getpages: called with non-merged cache vnode\n");
142 		return (VM_PAGER_ERROR);
143 	}
144 
145 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
146 		NFSLOCKNODE(np);
147 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
148 			NFSUNLOCKNODE(np);
149 			printf("ncl_getpages: called on non-cacheable vnode\n");
150 			return (VM_PAGER_ERROR);
151 		} else
152 			NFSUNLOCKNODE(np);
153 	}
154 
155 	mtx_lock(&nmp->nm_mtx);
156 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
157 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
158 		mtx_unlock(&nmp->nm_mtx);
159 		/* We'll never get here for v4, because we always have fsinfo */
160 		(void)ncl_fsinfo(nmp, vp, cred, td);
161 	} else
162 		mtx_unlock(&nmp->nm_mtx);
163 
164 	if (use_buf_pager)
165 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
166 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
167 
168 	/*
169 	 * If the requested page is partially valid, just return it and
170 	 * allow the pager to zero-out the blanks.  Partially valid pages
171 	 * can only occur at the file EOF.
172 	 *
173 	 * XXXGL: is that true for NFS, where short read can occur???
174 	 */
175 	VM_OBJECT_WLOCK(object);
176 	if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
177 		goto out;
178 	VM_OBJECT_WUNLOCK(object);
179 
180 	/*
181 	 * We use only the kva address for the buffer, but this is extremely
182 	 * convenient and fast.
183 	 */
184 	bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
185 
186 	kva = (vm_offset_t) bp->b_data;
187 	pmap_qenter(kva, pages, npages);
188 	VM_CNT_INC(v_vnodein);
189 	VM_CNT_ADD(v_vnodepgsin, npages);
190 
191 	count = npages << PAGE_SHIFT;
192 	iov.iov_base = (caddr_t) kva;
193 	iov.iov_len = count;
194 	uio.uio_iov = &iov;
195 	uio.uio_iovcnt = 1;
196 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
197 	uio.uio_resid = count;
198 	uio.uio_segflg = UIO_SYSSPACE;
199 	uio.uio_rw = UIO_READ;
200 	uio.uio_td = td;
201 
202 	error = ncl_readrpc(vp, &uio, cred);
203 	pmap_qremove(kva, npages);
204 
205 	uma_zfree(ncl_pbuf_zone, bp);
206 
207 	if (error && (uio.uio_resid == count)) {
208 		printf("ncl_getpages: error %d\n", error);
209 		return (VM_PAGER_ERROR);
210 	}
211 
212 	/*
213 	 * Calculate the number of bytes read and validate only that number
214 	 * of bytes.  Note that due to pending writes, size may be 0.  This
215 	 * does not mean that the remaining data is invalid!
216 	 */
217 
218 	size = count - uio.uio_resid;
219 	VM_OBJECT_WLOCK(object);
220 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
221 		vm_page_t m;
222 		nextoff = toff + PAGE_SIZE;
223 		m = pages[i];
224 
225 		if (nextoff <= size) {
226 			/*
227 			 * Read operation filled an entire page
228 			 */
229 			vm_page_valid(m);
230 			KASSERT(m->dirty == 0,
231 			    ("nfs_getpages: page %p is dirty", m));
232 		} else if (size > toff) {
233 			/*
234 			 * Read operation filled a partial page.
235 			 */
236 			vm_page_invalid(m);
237 			vm_page_set_valid_range(m, 0, size - toff);
238 			KASSERT(m->dirty == 0,
239 			    ("nfs_getpages: page %p is dirty", m));
240 		} else {
241 			/*
242 			 * Read operation was short.  If no error
243 			 * occurred we may have hit a zero-fill
244 			 * section.  We leave valid set to 0, and page
245 			 * is freed by vm_page_readahead_finish() if
246 			 * its index is not equal to requested, or
247 			 * page is zeroed and set valid by
248 			 * vm_pager_get_pages() for requested page.
249 			 */
250 			;
251 		}
252 	}
253 out:
254 	VM_OBJECT_WUNLOCK(object);
255 	if (ap->a_rbehind)
256 		*ap->a_rbehind = 0;
257 	if (ap->a_rahead)
258 		*ap->a_rahead = 0;
259 	return (VM_PAGER_OK);
260 }
261 
262 /*
263  * Vnode op for VM putpages.
264  */
265 int
266 ncl_putpages(struct vop_putpages_args *ap)
267 {
268 	struct uio uio;
269 	struct iovec iov;
270 	int i, error, npages, count;
271 	off_t offset;
272 	int *rtvals;
273 	struct vnode *vp;
274 	struct thread *td;
275 	struct ucred *cred;
276 	struct nfsmount *nmp;
277 	struct nfsnode *np;
278 	vm_page_t *pages;
279 
280 	vp = ap->a_vp;
281 	np = VTONFS(vp);
282 	td = curthread;				/* XXX */
283 	/* Set the cred to n_writecred for the write rpcs. */
284 	if (np->n_writecred != NULL)
285 		cred = crhold(np->n_writecred);
286 	else
287 		cred = crhold(curthread->td_ucred);	/* XXX */
288 	nmp = VFSTONFS(vp->v_mount);
289 	pages = ap->a_m;
290 	count = ap->a_count;
291 	rtvals = ap->a_rtvals;
292 	npages = btoc(count);
293 	offset = IDX_TO_OFF(pages[0]->pindex);
294 
295 	mtx_lock(&nmp->nm_mtx);
296 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
297 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
298 		mtx_unlock(&nmp->nm_mtx);
299 		(void)ncl_fsinfo(nmp, vp, cred, td);
300 	} else
301 		mtx_unlock(&nmp->nm_mtx);
302 
303 	NFSLOCKNODE(np);
304 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
305 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
306 		NFSUNLOCKNODE(np);
307 		printf("ncl_putpages: called on noncache-able vnode\n");
308 		NFSLOCKNODE(np);
309 	}
310 	/*
311 	 * When putting pages, do not extend file past EOF.
312 	 */
313 	if (offset + count > np->n_size) {
314 		count = np->n_size - offset;
315 		if (count < 0)
316 			count = 0;
317 	}
318 	NFSUNLOCKNODE(np);
319 
320 	for (i = 0; i < npages; i++)
321 		rtvals[i] = VM_PAGER_ERROR;
322 
323 	VM_CNT_INC(v_vnodeout);
324 	VM_CNT_ADD(v_vnodepgsout, count);
325 
326 	iov.iov_base = unmapped_buf;
327 	iov.iov_len = count;
328 	uio.uio_iov = &iov;
329 	uio.uio_iovcnt = 1;
330 	uio.uio_offset = offset;
331 	uio.uio_resid = count;
332 	uio.uio_segflg = UIO_NOCOPY;
333 	uio.uio_rw = UIO_WRITE;
334 	uio.uio_td = td;
335 
336 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
337 	    cred);
338 	crfree(cred);
339 
340 	if (error == 0 || !nfs_keep_dirty_on_error) {
341 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
342 		    np->n_size - offset, npages * PAGE_SIZE);
343 	}
344 	return (rtvals[0]);
345 }
346 
347 /*
348  * For nfs, cache consistency can only be maintained approximately.
349  * Although RFC1094 does not specify the criteria, the following is
350  * believed to be compatible with the reference port.
351  * For nfs:
352  * If the file's modify time on the server has changed since the
353  * last read rpc or you have written to the file,
354  * you may have lost data cache consistency with the
355  * server, so flush all of the file's data out of the cache.
356  * Then force a getattr rpc to ensure that you have up to date
357  * attributes.
358  * NB: This implies that cache data can be read when up to
359  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
360  * attributes this could be forced by setting n_attrstamp to 0 before
361  * the VOP_GETATTR() call.
362  */
363 static inline int
364 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
365 {
366 	int error = 0;
367 	struct vattr vattr;
368 	struct nfsnode *np = VTONFS(vp);
369 	bool old_lock;
370 
371 	/*
372 	 * Ensure the exclusove access to the node before checking
373 	 * whether the cache is consistent.
374 	 */
375 	old_lock = ncl_excl_start(vp);
376 	NFSLOCKNODE(np);
377 	if (np->n_flag & NMODIFIED) {
378 		NFSUNLOCKNODE(np);
379 		if (vp->v_type != VREG) {
380 			if (vp->v_type != VDIR)
381 				panic("nfs: bioread, not dir");
382 			ncl_invaldir(vp);
383 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
384 			if (error != 0)
385 				goto out;
386 		}
387 		np->n_attrstamp = 0;
388 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
389 		error = VOP_GETATTR(vp, &vattr, cred);
390 		if (error)
391 			goto out;
392 		NFSLOCKNODE(np);
393 		np->n_mtime = vattr.va_mtime;
394 		NFSUNLOCKNODE(np);
395 	} else {
396 		NFSUNLOCKNODE(np);
397 		error = VOP_GETATTR(vp, &vattr, cred);
398 		if (error)
399 			goto out;
400 		NFSLOCKNODE(np);
401 		if ((np->n_flag & NSIZECHANGED)
402 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
403 			NFSUNLOCKNODE(np);
404 			if (vp->v_type == VDIR)
405 				ncl_invaldir(vp);
406 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
407 			if (error != 0)
408 				goto out;
409 			NFSLOCKNODE(np);
410 			np->n_mtime = vattr.va_mtime;
411 			np->n_flag &= ~NSIZECHANGED;
412 		}
413 		NFSUNLOCKNODE(np);
414 	}
415 out:
416 	ncl_excl_finish(vp, old_lock);
417 	return (error);
418 }
419 
420 /*
421  * Vnode op for read using bio
422  */
423 int
424 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
425 {
426 	struct nfsnode *np = VTONFS(vp);
427 	struct buf *bp, *rabp;
428 	struct thread *td;
429 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
430 	daddr_t lbn, rabn;
431 	int biosize, bcount, error, i, n, nra, on, save2, seqcount;
432 	off_t tmp_off;
433 
434 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
435 	if (uio->uio_resid == 0)
436 		return (0);
437 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
438 		return (EINVAL);
439 	td = uio->uio_td;
440 
441 	mtx_lock(&nmp->nm_mtx);
442 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
443 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
444 		mtx_unlock(&nmp->nm_mtx);
445 		(void)ncl_fsinfo(nmp, vp, cred, td);
446 		mtx_lock(&nmp->nm_mtx);
447 	}
448 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
449 		(void) newnfs_iosize(nmp);
450 
451 	tmp_off = uio->uio_offset + uio->uio_resid;
452 	if (vp->v_type != VDIR &&
453 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
454 		mtx_unlock(&nmp->nm_mtx);
455 		return (EFBIG);
456 	}
457 	mtx_unlock(&nmp->nm_mtx);
458 
459 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
460 		/* No caching/ no readaheads. Just read data into the user buffer */
461 		return ncl_readrpc(vp, uio, cred);
462 
463 	n = 0;
464 	on = 0;
465 	biosize = vp->v_bufobj.bo_bsize;
466 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
467 
468 	error = nfs_bioread_check_cons(vp, td, cred);
469 	if (error)
470 		return error;
471 
472 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
473 	do {
474 	    u_quad_t nsize;
475 
476 	    NFSLOCKNODE(np);
477 	    nsize = np->n_size;
478 	    NFSUNLOCKNODE(np);
479 
480 	    switch (vp->v_type) {
481 	    case VREG:
482 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
483 		lbn = uio->uio_offset / biosize;
484 		on = uio->uio_offset - (lbn * biosize);
485 
486 		/*
487 		 * Start the read ahead(s), as required.
488 		 */
489 		if (nmp->nm_readahead > 0) {
490 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
491 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
492 			rabn = lbn + 1 + nra;
493 			if (incore(&vp->v_bufobj, rabn) == NULL) {
494 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
495 			    if (!rabp) {
496 				error = newnfs_sigintr(nmp, td);
497 				if (error == 0)
498 					error = EINTR;
499 				goto out;
500 			    }
501 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
502 				rabp->b_flags |= B_ASYNC;
503 				rabp->b_iocmd = BIO_READ;
504 				vfs_busy_pages(rabp, 0);
505 				if (ncl_asyncio(nmp, rabp, cred, td)) {
506 				    rabp->b_flags |= B_INVAL;
507 				    rabp->b_ioflags |= BIO_ERROR;
508 				    vfs_unbusy_pages(rabp);
509 				    brelse(rabp);
510 				    break;
511 				}
512 			    } else {
513 				brelse(rabp);
514 			    }
515 			}
516 		    }
517 		}
518 
519 		/* Note that bcount is *not* DEV_BSIZE aligned. */
520 		bcount = biosize;
521 		if ((off_t)lbn * biosize >= nsize) {
522 			bcount = 0;
523 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
524 			bcount = nsize - (off_t)lbn * biosize;
525 		}
526 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
527 
528 		if (!bp) {
529 			error = newnfs_sigintr(nmp, td);
530 			if (error == 0)
531 				error = EINTR;
532 			goto out;
533 		}
534 
535 		/*
536 		 * If B_CACHE is not set, we must issue the read.  If this
537 		 * fails, we return an error.
538 		 */
539 
540 		if ((bp->b_flags & B_CACHE) == 0) {
541 		    bp->b_iocmd = BIO_READ;
542 		    vfs_busy_pages(bp, 0);
543 		    error = ncl_doio(vp, bp, cred, td, 0);
544 		    if (error) {
545 			brelse(bp);
546 			goto out;
547 		    }
548 		}
549 
550 		/*
551 		 * on is the offset into the current bp.  Figure out how many
552 		 * bytes we can copy out of the bp.  Note that bcount is
553 		 * NOT DEV_BSIZE aligned.
554 		 *
555 		 * Then figure out how many bytes we can copy into the uio.
556 		 */
557 
558 		n = 0;
559 		if (on < bcount)
560 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
561 		break;
562 	    case VLNK:
563 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
564 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
565 		if (!bp) {
566 			error = newnfs_sigintr(nmp, td);
567 			if (error == 0)
568 				error = EINTR;
569 			goto out;
570 		}
571 		if ((bp->b_flags & B_CACHE) == 0) {
572 		    bp->b_iocmd = BIO_READ;
573 		    vfs_busy_pages(bp, 0);
574 		    error = ncl_doio(vp, bp, cred, td, 0);
575 		    if (error) {
576 			bp->b_ioflags |= BIO_ERROR;
577 			brelse(bp);
578 			goto out;
579 		    }
580 		}
581 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
582 		on = 0;
583 		break;
584 	    case VDIR:
585 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
586 		NFSLOCKNODE(np);
587 		if (np->n_direofoffset
588 		    && uio->uio_offset >= np->n_direofoffset) {
589 			NFSUNLOCKNODE(np);
590 			error = 0;
591 			goto out;
592 		}
593 		NFSUNLOCKNODE(np);
594 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
595 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
596 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
597 		if (!bp) {
598 			error = newnfs_sigintr(nmp, td);
599 			if (error == 0)
600 				error = EINTR;
601 			goto out;
602 		}
603 		if ((bp->b_flags & B_CACHE) == 0) {
604 		    bp->b_iocmd = BIO_READ;
605 		    vfs_busy_pages(bp, 0);
606 		    error = ncl_doio(vp, bp, cred, td, 0);
607 		    if (error) {
608 			    brelse(bp);
609 		    }
610 		    while (error == NFSERR_BAD_COOKIE) {
611 			ncl_invaldir(vp);
612 			error = ncl_vinvalbuf(vp, 0, td, 1);
613 
614 			/*
615 			 * Yuck! The directory has been modified on the
616 			 * server. The only way to get the block is by
617 			 * reading from the beginning to get all the
618 			 * offset cookies.
619 			 *
620 			 * Leave the last bp intact unless there is an error.
621 			 * Loop back up to the while if the error is another
622 			 * NFSERR_BAD_COOKIE (double yuch!).
623 			 */
624 			for (i = 0; i <= lbn && !error; i++) {
625 			    NFSLOCKNODE(np);
626 			    if (np->n_direofoffset
627 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
628 				    NFSUNLOCKNODE(np);
629 				    error = 0;
630 				    goto out;
631 			    }
632 			    NFSUNLOCKNODE(np);
633 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
634 			    if (!bp) {
635 				error = newnfs_sigintr(nmp, td);
636 				if (error == 0)
637 					error = EINTR;
638 				goto out;
639 			    }
640 			    if ((bp->b_flags & B_CACHE) == 0) {
641 				    bp->b_iocmd = BIO_READ;
642 				    vfs_busy_pages(bp, 0);
643 				    error = ncl_doio(vp, bp, cred, td, 0);
644 				    /*
645 				     * no error + B_INVAL == directory EOF,
646 				     * use the block.
647 				     */
648 				    if (error == 0 && (bp->b_flags & B_INVAL))
649 					    break;
650 			    }
651 			    /*
652 			     * An error will throw away the block and the
653 			     * for loop will break out.  If no error and this
654 			     * is not the block we want, we throw away the
655 			     * block and go for the next one via the for loop.
656 			     */
657 			    if (error || i < lbn)
658 				    brelse(bp);
659 			}
660 		    }
661 		    /*
662 		     * The above while is repeated if we hit another cookie
663 		     * error.  If we hit an error and it wasn't a cookie error,
664 		     * we give up.
665 		     */
666 		    if (error)
667 			    goto out;
668 		}
669 
670 		/*
671 		 * If not eof and read aheads are enabled, start one.
672 		 * (You need the current block first, so that you have the
673 		 *  directory offset cookie of the next block.)
674 		 */
675 		NFSLOCKNODE(np);
676 		if (nmp->nm_readahead > 0 &&
677 		    (bp->b_flags & B_INVAL) == 0 &&
678 		    (np->n_direofoffset == 0 ||
679 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
680 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
681 			NFSUNLOCKNODE(np);
682 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
683 			if (rabp) {
684 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
685 				rabp->b_flags |= B_ASYNC;
686 				rabp->b_iocmd = BIO_READ;
687 				vfs_busy_pages(rabp, 0);
688 				if (ncl_asyncio(nmp, rabp, cred, td)) {
689 				    rabp->b_flags |= B_INVAL;
690 				    rabp->b_ioflags |= BIO_ERROR;
691 				    vfs_unbusy_pages(rabp);
692 				    brelse(rabp);
693 				}
694 			    } else {
695 				brelse(rabp);
696 			    }
697 			}
698 			NFSLOCKNODE(np);
699 		}
700 		/*
701 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
702 		 * chopped for the EOF condition, we cannot tell how large
703 		 * NFS directories are going to be until we hit EOF.  So
704 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
705 		 * it just so happens that b_resid will effectively chop it
706 		 * to EOF.  *BUT* this information is lost if the buffer goes
707 		 * away and is reconstituted into a B_CACHE state ( due to
708 		 * being VMIO ) later.  So we keep track of the directory eof
709 		 * in np->n_direofoffset and chop it off as an extra step
710 		 * right here.
711 		 */
712 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
713 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
714 			n = np->n_direofoffset - uio->uio_offset;
715 		NFSUNLOCKNODE(np);
716 		break;
717 	    default:
718 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
719 		bp = NULL;
720 		break;
721 	    }
722 
723 	    if (n > 0) {
724 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
725 	    }
726 	    if (vp->v_type == VLNK)
727 		n = 0;
728 	    if (bp != NULL)
729 		brelse(bp);
730 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
731 out:
732 	curthread_pflags2_restore(save2);
733 	if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
734 		NFSLOCKNODE(np);
735 		ncl_pager_setsize(vp, NULL);
736 	}
737 	return (error);
738 }
739 
740 /*
741  * The NFS write path cannot handle iovecs with len > 1. So we need to
742  * break up iovecs accordingly (restricting them to wsize).
743  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
744  * For the ASYNC case, 2 copies are needed. The first a copy from the
745  * user buffer to a staging buffer and then a second copy from the staging
746  * buffer to mbufs. This can be optimized by copying from the user buffer
747  * directly into mbufs and passing the chain down, but that requires a
748  * fair amount of re-working of the relevant codepaths (and can be done
749  * later).
750  */
751 static int
752 nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred,
753     int ioflag)
754 {
755 	int error;
756 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
757 	struct thread *td = uiop->uio_td;
758 	int size;
759 	int wsize;
760 
761 	mtx_lock(&nmp->nm_mtx);
762 	wsize = nmp->nm_wsize;
763 	mtx_unlock(&nmp->nm_mtx);
764 	if (ioflag & IO_SYNC) {
765 		int iomode, must_commit;
766 		struct uio uio;
767 		struct iovec iov;
768 do_sync:
769 		while (uiop->uio_resid > 0) {
770 			size = MIN(uiop->uio_resid, wsize);
771 			size = MIN(uiop->uio_iov->iov_len, size);
772 			iov.iov_base = uiop->uio_iov->iov_base;
773 			iov.iov_len = size;
774 			uio.uio_iov = &iov;
775 			uio.uio_iovcnt = 1;
776 			uio.uio_offset = uiop->uio_offset;
777 			uio.uio_resid = size;
778 			uio.uio_segflg = uiop->uio_segflg;
779 			uio.uio_rw = UIO_WRITE;
780 			uio.uio_td = td;
781 			iomode = NFSWRITE_FILESYNC;
782 			/*
783 			 * When doing direct I/O we do not care if the
784 			 * server's write verifier has changed, but we
785 			 * do not want to update the verifier if it has
786 			 * changed, since that hides the change from
787 			 * writes being done through the buffer cache.
788 			 * By passing must_commit in set to two, the code
789 			 * in nfsrpc_writerpc() will not update the
790 			 * verifier on the mount point.
791 			 */
792 			must_commit = 2;
793 			error = ncl_writerpc(vp, &uio, cred, &iomode,
794 			    &must_commit, 0, ioflag);
795 			KASSERT((must_commit == 2),
796 			    ("ncl_directio_write: Updated write verifier"));
797 			if (error)
798 				return (error);
799 			if (iomode != NFSWRITE_FILESYNC)
800 				printf("nfs_directio_write: Broken server "
801 				    "did not reply FILE_SYNC\n");
802 			uiop->uio_offset += size;
803 			uiop->uio_resid -= size;
804 			if (uiop->uio_iov->iov_len <= size) {
805 				uiop->uio_iovcnt--;
806 				uiop->uio_iov++;
807 			} else {
808 				uiop->uio_iov->iov_base =
809 					(char *)uiop->uio_iov->iov_base + size;
810 				uiop->uio_iov->iov_len -= size;
811 			}
812 		}
813 	} else {
814 		struct uio *t_uio;
815 		struct iovec *t_iov;
816 		struct buf *bp;
817 
818 		/*
819 		 * Break up the write into blocksize chunks and hand these
820 		 * over to nfsiod's for write back.
821 		 * Unfortunately, this incurs a copy of the data. Since
822 		 * the user could modify the buffer before the write is
823 		 * initiated.
824 		 *
825 		 * The obvious optimization here is that one of the 2 copies
826 		 * in the async write path can be eliminated by copying the
827 		 * data here directly into mbufs and passing the mbuf chain
828 		 * down. But that will require a fair amount of re-working
829 		 * of the code and can be done if there's enough interest
830 		 * in NFS directio access.
831 		 */
832 		while (uiop->uio_resid > 0) {
833 			size = MIN(uiop->uio_resid, wsize);
834 			size = MIN(uiop->uio_iov->iov_len, size);
835 			bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
836 			t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
837 			t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
838 			t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
839 			t_iov->iov_len = size;
840 			t_uio->uio_iov = t_iov;
841 			t_uio->uio_iovcnt = 1;
842 			t_uio->uio_offset = uiop->uio_offset;
843 			t_uio->uio_resid = size;
844 			t_uio->uio_segflg = UIO_SYSSPACE;
845 			t_uio->uio_rw = UIO_WRITE;
846 			t_uio->uio_td = td;
847 			KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
848 			    uiop->uio_segflg == UIO_SYSSPACE,
849 			    ("nfs_directio_write: Bad uio_segflg"));
850 			if (uiop->uio_segflg == UIO_USERSPACE) {
851 				error = copyin(uiop->uio_iov->iov_base,
852 				    t_iov->iov_base, size);
853 				if (error != 0)
854 					goto err_free;
855 			} else
856 				/*
857 				 * UIO_SYSSPACE may never happen, but handle
858 				 * it just in case it does.
859 				 */
860 				bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
861 				    size);
862 			bp->b_flags |= B_DIRECT;
863 			bp->b_iocmd = BIO_WRITE;
864 			if (cred != NOCRED) {
865 				crhold(cred);
866 				bp->b_wcred = cred;
867 			} else
868 				bp->b_wcred = NOCRED;
869 			bp->b_caller1 = (void *)t_uio;
870 			bp->b_vp = vp;
871 			error = ncl_asyncio(nmp, bp, NOCRED, td);
872 err_free:
873 			if (error) {
874 				free(t_iov->iov_base, M_NFSDIRECTIO);
875 				free(t_iov, M_NFSDIRECTIO);
876 				free(t_uio, M_NFSDIRECTIO);
877 				bp->b_vp = NULL;
878 				uma_zfree(ncl_pbuf_zone, bp);
879 				if (error == EINTR)
880 					return (error);
881 				goto do_sync;
882 			}
883 			uiop->uio_offset += size;
884 			uiop->uio_resid -= size;
885 			if (uiop->uio_iov->iov_len <= size) {
886 				uiop->uio_iovcnt--;
887 				uiop->uio_iov++;
888 			} else {
889 				uiop->uio_iov->iov_base =
890 					(char *)uiop->uio_iov->iov_base + size;
891 				uiop->uio_iov->iov_len -= size;
892 			}
893 		}
894 	}
895 	return (0);
896 }
897 
898 /*
899  * Vnode op for write using bio
900  */
901 int
902 ncl_write(struct vop_write_args *ap)
903 {
904 	int biosize;
905 	struct uio *uio = ap->a_uio;
906 	struct thread *td = uio->uio_td;
907 	struct vnode *vp = ap->a_vp;
908 	struct nfsnode *np = VTONFS(vp);
909 	struct ucred *cred = ap->a_cred;
910 	int ioflag = ap->a_ioflag;
911 	struct buf *bp;
912 	struct vattr vattr;
913 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
914 	daddr_t lbn;
915 	int bcount, noncontig_write, obcount;
916 	int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
917 	size_t orig_resid, local_resid;
918 	off_t orig_size, tmp_off;
919 	struct timespec ts;
920 
921 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
922 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
923 	    ("ncl_write proc"));
924 	if (vp->v_type != VREG)
925 		return (EIO);
926 	NFSLOCKNODE(np);
927 	if (np->n_flag & NWRITEERR) {
928 		np->n_flag &= ~NWRITEERR;
929 		NFSUNLOCKNODE(np);
930 		return (np->n_error);
931 	} else
932 		NFSUNLOCKNODE(np);
933 	mtx_lock(&nmp->nm_mtx);
934 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
935 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
936 		mtx_unlock(&nmp->nm_mtx);
937 		(void)ncl_fsinfo(nmp, vp, cred, td);
938 		mtx_lock(&nmp->nm_mtx);
939 	}
940 	if (nmp->nm_wsize == 0)
941 		(void) newnfs_iosize(nmp);
942 	mtx_unlock(&nmp->nm_mtx);
943 
944 	/*
945 	 * Synchronously flush pending buffers if we are in synchronous
946 	 * mode or if we are appending.
947 	 */
948 	if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
949 	    NMODIFIED))) {
950 		/*
951 		 * For the case where IO_APPEND is being done using a
952 		 * direct output (to the NFS server) RPC and
953 		 * newnfs_directio_enable is 0, all buffer cache buffers,
954 		 * including ones not modified, must be invalidated.
955 		 * This ensures that stale data is not read out of the
956 		 * buffer cache.  The call also invalidates all mapped
957 		 * pages and, since the exclusive lock is held on the vnode,
958 		 * new pages cannot be faulted in.
959 		 *
960 		 * For the case where newnfs_directio_enable is set
961 		 * (which is not the default), it is not obvious that
962 		 * stale data should be left in the buffer cache, but
963 		 * the code has been this way for over a decade without
964 		 * complaints.  Note that, unlike doing IO_APPEND via
965 		 * a direct write RPC when newnfs_directio_enable is not set,
966 		 * when newnfs_directio_enable is set, reading is done via
967 		 * direct to NFS server RPCs as well.
968 		 */
969 		np->n_attrstamp = 0;
970 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
971 		error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
972 		    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
973 		if (error != 0)
974 			return (error);
975 	}
976 
977 	orig_resid = uio->uio_resid;
978 	NFSLOCKNODE(np);
979 	orig_size = np->n_size;
980 	NFSUNLOCKNODE(np);
981 
982 	/*
983 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
984 	 * get the append lock.
985 	 */
986 	if (ioflag & IO_APPEND) {
987 		/*
988 		 * For NFSv4, the AppendWrite will Verify the size against
989 		 * the file's size on the server.  If not the same, the
990 		 * write will then be retried, using the file size returned
991 		 * by the AppendWrite.  However, for NFSv2 and NFSv3, the
992 		 * size must be acquired here via a Getattr RPC.
993 		 * The AppendWrite is not done for a pNFS mount.
994 		 */
995 		if (!NFSHASNFSV4(nmp) || NFSHASPNFS(nmp)) {
996 			np->n_attrstamp = 0;
997 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
998 			error = VOP_GETATTR(vp, &vattr, cred);
999 			if (error)
1000 				return (error);
1001 		}
1002 		NFSLOCKNODE(np);
1003 		uio->uio_offset = np->n_size;
1004 		NFSUNLOCKNODE(np);
1005 	}
1006 
1007 	if (uio->uio_offset < 0)
1008 		return (EINVAL);
1009 	tmp_off = uio->uio_offset + uio->uio_resid;
1010 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
1011 		return (EFBIG);
1012 	if (uio->uio_resid == 0)
1013 		return (0);
1014 
1015 	/*
1016 	 * Do IO_APPEND writing via a synchronous direct write.
1017 	 * This can result in a significant performance improvement.
1018 	 */
1019 	if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
1020 	    (ioflag & IO_APPEND)) {
1021 		/*
1022 		 * Direct writes to the server must be done NFSWRITE_FILESYNC,
1023 		 * because the write data is not cached and, therefore, the
1024 		 * write cannot be redone after a server reboot.
1025 		 * Set IO_SYNC to make this happen.
1026 		 */
1027 		ioflag |= IO_SYNC;
1028 		return (nfs_directio_write(vp, uio, cred, ioflag));
1029 	}
1030 
1031 	/*
1032 	 * Maybe this should be above the vnode op call, but so long as
1033 	 * file servers have no limits, i don't think it matters
1034 	 */
1035 	error = vn_rlimit_fsize(vp, uio, td);
1036 	if (error != 0)
1037 		return (error);
1038 
1039 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
1040 	biosize = vp->v_bufobj.bo_bsize;
1041 	/*
1042 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
1043 	 * would exceed the local maximum per-file write commit size when
1044 	 * combined with those, we must decide whether to flush,
1045 	 * go synchronous, or return error.  We don't bother checking
1046 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
1047 	 * no point optimizing for something that really won't ever happen.
1048 	 */
1049 	wouldcommit = 0;
1050 	if (!(ioflag & IO_SYNC)) {
1051 		int nflag;
1052 
1053 		NFSLOCKNODE(np);
1054 		nflag = np->n_flag;
1055 		NFSUNLOCKNODE(np);
1056 		if (nflag & NMODIFIED) {
1057 			BO_LOCK(&vp->v_bufobj);
1058 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1059 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1060 				    b_bobufs) {
1061 					if (bp->b_flags & B_NEEDCOMMIT)
1062 						wouldcommit += bp->b_bcount;
1063 				}
1064 			}
1065 			BO_UNLOCK(&vp->v_bufobj);
1066 		}
1067 	}
1068 
1069 	do {
1070 		if (!(ioflag & IO_SYNC)) {
1071 			wouldcommit += biosize;
1072 			if (wouldcommit > nmp->nm_wcommitsize) {
1073 				np->n_attrstamp = 0;
1074 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1075 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1076 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1077 				if (error != 0)
1078 					goto out;
1079 				wouldcommit = biosize;
1080 			}
1081 		}
1082 
1083 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1084 		lbn = uio->uio_offset / biosize;
1085 		on = uio->uio_offset - (lbn * biosize);
1086 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1087 again:
1088 		/*
1089 		 * Handle direct append and file extension cases, calculate
1090 		 * unaligned buffer size.
1091 		 */
1092 		NFSLOCKNODE(np);
1093 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1094 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1095 			noncontig_write = 1;
1096 		else
1097 			noncontig_write = 0;
1098 		if ((uio->uio_offset == np->n_size ||
1099 		    (noncontig_write != 0 &&
1100 		    lbn == (np->n_size / biosize) &&
1101 		    uio->uio_offset + n > np->n_size)) && n) {
1102 			NFSUNLOCKNODE(np);
1103 			/*
1104 			 * Get the buffer (in its pre-append state to maintain
1105 			 * B_CACHE if it was previously set).  Resize the
1106 			 * nfsnode after we have locked the buffer to prevent
1107 			 * readers from reading garbage.
1108 			 */
1109 			obcount = np->n_size - (lbn * biosize);
1110 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1111 
1112 			if (bp != NULL) {
1113 				long save;
1114 
1115 				NFSLOCKNODE(np);
1116 				np->n_size = uio->uio_offset + n;
1117 				np->n_flag |= NMODIFIED;
1118 				np->n_flag &= ~NVNSETSZSKIP;
1119 				vnode_pager_setsize(vp, np->n_size);
1120 				NFSUNLOCKNODE(np);
1121 
1122 				save = bp->b_flags & B_CACHE;
1123 				bcount = on + n;
1124 				allocbuf(bp, bcount);
1125 				bp->b_flags |= save;
1126 				if (noncontig_write != 0 && on > obcount)
1127 					vfs_bio_bzero_buf(bp, obcount, on -
1128 					    obcount);
1129 			}
1130 		} else {
1131 			/*
1132 			 * Obtain the locked cache block first, and then
1133 			 * adjust the file's size as appropriate.
1134 			 */
1135 			bcount = on + n;
1136 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1137 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1138 					bcount = biosize;
1139 				else
1140 					bcount = np->n_size - (off_t)lbn * biosize;
1141 			}
1142 			NFSUNLOCKNODE(np);
1143 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1144 			NFSLOCKNODE(np);
1145 			if (uio->uio_offset + n > np->n_size) {
1146 				np->n_size = uio->uio_offset + n;
1147 				np->n_flag |= NMODIFIED;
1148 				np->n_flag &= ~NVNSETSZSKIP;
1149 				vnode_pager_setsize(vp, np->n_size);
1150 			}
1151 			NFSUNLOCKNODE(np);
1152 		}
1153 
1154 		if (!bp) {
1155 			error = newnfs_sigintr(nmp, td);
1156 			if (!error)
1157 				error = EINTR;
1158 			break;
1159 		}
1160 
1161 		/*
1162 		 * Issue a READ if B_CACHE is not set.  In special-append
1163 		 * mode, B_CACHE is based on the buffer prior to the write
1164 		 * op and is typically set, avoiding the read.  If a read
1165 		 * is required in special append mode, the server will
1166 		 * probably send us a short-read since we extended the file
1167 		 * on our end, resulting in b_resid == 0 and, thusly,
1168 		 * B_CACHE getting set.
1169 		 *
1170 		 * We can also avoid issuing the read if the write covers
1171 		 * the entire buffer.  We have to make sure the buffer state
1172 		 * is reasonable in this case since we will not be initiating
1173 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1174 		 * more information.
1175 		 *
1176 		 * B_CACHE may also be set due to the buffer being cached
1177 		 * normally.
1178 		 */
1179 
1180 		bp_cached = 1;
1181 		if (on == 0 && n == bcount) {
1182 			if ((bp->b_flags & B_CACHE) == 0)
1183 				bp_cached = 0;
1184 			bp->b_flags |= B_CACHE;
1185 			bp->b_flags &= ~B_INVAL;
1186 			bp->b_ioflags &= ~BIO_ERROR;
1187 		}
1188 
1189 		if ((bp->b_flags & B_CACHE) == 0) {
1190 			bp->b_iocmd = BIO_READ;
1191 			vfs_busy_pages(bp, 0);
1192 			error = ncl_doio(vp, bp, cred, td, 0);
1193 			if (error) {
1194 				brelse(bp);
1195 				break;
1196 			}
1197 		}
1198 		if (bp->b_wcred == NOCRED)
1199 			bp->b_wcred = crhold(cred);
1200 		NFSLOCKNODE(np);
1201 		np->n_flag |= NMODIFIED;
1202 		NFSUNLOCKNODE(np);
1203 
1204 		/*
1205 		 * If dirtyend exceeds file size, chop it down.  This should
1206 		 * not normally occur but there is an append race where it
1207 		 * might occur XXX, so we log it.
1208 		 *
1209 		 * If the chopping creates a reverse-indexed or degenerate
1210 		 * situation with dirtyoff/end, we 0 both of them.
1211 		 */
1212 
1213 		if (bp->b_dirtyend > bcount) {
1214 			printf("NFS append race @%lx:%d\n",
1215 			    (long)bp->b_blkno * DEV_BSIZE,
1216 			    bp->b_dirtyend - bcount);
1217 			bp->b_dirtyend = bcount;
1218 		}
1219 
1220 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1221 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1222 
1223 		/*
1224 		 * If the new write will leave a contiguous dirty
1225 		 * area, just update the b_dirtyoff and b_dirtyend,
1226 		 * otherwise force a write rpc of the old dirty area.
1227 		 *
1228 		 * If there has been a file lock applied to this file
1229 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1230 		 * While it is possible to merge discontiguous writes due to
1231 		 * our having a B_CACHE buffer ( and thus valid read data
1232 		 * for the hole), we don't because it could lead to
1233 		 * significant cache coherency problems with multiple clients,
1234 		 * especially if locking is implemented later on.
1235 		 *
1236 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1237 		 * not been file locking done on this file:
1238 		 * Relax coherency a bit for the sake of performance and
1239 		 * expand the current dirty region to contain the new
1240 		 * write even if it means we mark some non-dirty data as
1241 		 * dirty.
1242 		 */
1243 
1244 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1245 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1246 			if (bwrite(bp) == EINTR) {
1247 				error = EINTR;
1248 				break;
1249 			}
1250 			goto again;
1251 		}
1252 
1253 		local_resid = uio->uio_resid;
1254 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1255 
1256 		if (error != 0 && !bp_cached) {
1257 			/*
1258 			 * This block has no other content then what
1259 			 * possibly was written by the faulty uiomove.
1260 			 * Release it, forgetting the data pages, to
1261 			 * prevent the leak of uninitialized data to
1262 			 * usermode.
1263 			 */
1264 			bp->b_ioflags |= BIO_ERROR;
1265 			brelse(bp);
1266 			uio->uio_offset -= local_resid - uio->uio_resid;
1267 			uio->uio_resid = local_resid;
1268 			break;
1269 		}
1270 
1271 		/*
1272 		 * Since this block is being modified, it must be written
1273 		 * again and not just committed.  Since write clustering does
1274 		 * not work for the stage 1 data write, only the stage 2
1275 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1276 		 */
1277 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1278 
1279 		/*
1280 		 * Get the partial update on the progress made from
1281 		 * uiomove, if an error occurred.
1282 		 */
1283 		if (error != 0)
1284 			n = local_resid - uio->uio_resid;
1285 
1286 		/*
1287 		 * Only update dirtyoff/dirtyend if not a degenerate
1288 		 * condition.
1289 		 */
1290 		if (n > 0) {
1291 			if (bp->b_dirtyend > 0) {
1292 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1293 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1294 			} else {
1295 				bp->b_dirtyoff = on;
1296 				bp->b_dirtyend = on + n;
1297 			}
1298 			vfs_bio_set_valid(bp, on, n);
1299 		}
1300 
1301 		/*
1302 		 * If IO_SYNC do bwrite().
1303 		 *
1304 		 * IO_INVAL appears to be unused.  The idea appears to be
1305 		 * to turn off caching in this case.  Very odd.  XXX
1306 		 */
1307 		if ((ioflag & IO_SYNC)) {
1308 			if (ioflag & IO_INVAL)
1309 				bp->b_flags |= B_NOCACHE;
1310 			error1 = bwrite(bp);
1311 			if (error1 != 0) {
1312 				if (error == 0)
1313 					error = error1;
1314 				break;
1315 			}
1316 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1317 			bp->b_flags |= B_ASYNC;
1318 			(void) ncl_writebp(bp, 0, NULL);
1319 		} else {
1320 			bdwrite(bp);
1321 		}
1322 
1323 		if (error != 0)
1324 			break;
1325 	} while (uio->uio_resid > 0 && n > 0);
1326 
1327 	if (error == 0) {
1328 		nanouptime(&ts);
1329 		NFSLOCKNODE(np);
1330 		np->n_localmodtime = ts;
1331 		NFSUNLOCKNODE(np);
1332 	} else {
1333 		if (ioflag & IO_UNIT) {
1334 			VATTR_NULL(&vattr);
1335 			vattr.va_size = orig_size;
1336 			/* IO_SYNC is handled implicitely */
1337 			(void)VOP_SETATTR(vp, &vattr, cred);
1338 			uio->uio_offset -= orig_resid - uio->uio_resid;
1339 			uio->uio_resid = orig_resid;
1340 		}
1341 	}
1342 
1343 out:
1344 	curthread_pflags2_restore(save2);
1345 	return (error);
1346 }
1347 
1348 /*
1349  * Get an nfs cache block.
1350  *
1351  * Allocate a new one if the block isn't currently in the cache
1352  * and return the block marked busy. If the calling process is
1353  * interrupted by a signal for an interruptible mount point, return
1354  * NULL.
1355  *
1356  * The caller must carefully deal with the possible B_INVAL state of
1357  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1358  * indirectly), so synchronous reads can be issued without worrying about
1359  * the B_INVAL state.  We have to be a little more careful when dealing
1360  * with writes (see comments in nfs_write()) when extending a file past
1361  * its EOF.
1362  */
1363 static struct buf *
1364 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1365 {
1366 	struct buf *bp;
1367 	struct mount *mp;
1368 	struct nfsmount *nmp;
1369 
1370 	mp = vp->v_mount;
1371 	nmp = VFSTONFS(mp);
1372 
1373 	if (nmp->nm_flag & NFSMNT_INT) {
1374 		sigset_t oldset;
1375 
1376 		newnfs_set_sigmask(td, &oldset);
1377 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1378 		newnfs_restore_sigmask(td, &oldset);
1379 		while (bp == NULL) {
1380 			if (newnfs_sigintr(nmp, td))
1381 				return (NULL);
1382 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1383 		}
1384 	} else {
1385 		bp = getblk(vp, bn, size, 0, 0, 0);
1386 	}
1387 
1388 	if (vp->v_type == VREG)
1389 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1390 	return (bp);
1391 }
1392 
1393 /*
1394  * Flush and invalidate all dirty buffers. If another process is already
1395  * doing the flush, just wait for completion.
1396  */
1397 int
1398 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1399 {
1400 	struct nfsnode *np = VTONFS(vp);
1401 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1402 	int error = 0, slpflag, slptimeo;
1403 	bool old_lock;
1404 	struct timespec ts;
1405 
1406 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1407 
1408 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1409 		intrflg = 0;
1410 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1411 		intrflg = 1;
1412 	if (intrflg) {
1413 		slpflag = PCATCH;
1414 		slptimeo = 2 * hz;
1415 	} else {
1416 		slpflag = 0;
1417 		slptimeo = 0;
1418 	}
1419 
1420 	old_lock = ncl_excl_start(vp);
1421 	if (old_lock)
1422 		flags |= V_ALLOWCLEAN;
1423 
1424 	/*
1425 	 * Now, flush as required.
1426 	 */
1427 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1428 	     vp->v_bufobj.bo_object != NULL) {
1429 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1430 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1431 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1432 		/*
1433 		 * If the page clean was interrupted, fail the invalidation.
1434 		 * Not doing so, we run the risk of losing dirty pages in the
1435 		 * vinvalbuf() call below.
1436 		 */
1437 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1438 			goto out;
1439 	}
1440 
1441 	error = vinvalbuf(vp, flags, slpflag, 0);
1442 	while (error) {
1443 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1444 			goto out;
1445 		error = vinvalbuf(vp, flags, 0, slptimeo);
1446 	}
1447 	if (NFSHASPNFS(nmp)) {
1448 		nfscl_layoutcommit(vp, td);
1449 		nanouptime(&ts);
1450 		/*
1451 		 * Invalidate the attribute cache, since writes to a DS
1452 		 * won't update the size attribute.
1453 		 */
1454 		NFSLOCKNODE(np);
1455 		np->n_attrstamp = 0;
1456 	} else {
1457 		nanouptime(&ts);
1458 		NFSLOCKNODE(np);
1459 	}
1460 	if (np->n_directio_asyncwr == 0 && (np->n_flag & NMODIFIED) != 0) {
1461 		np->n_localmodtime = ts;
1462 		np->n_flag &= ~NMODIFIED;
1463 	}
1464 	NFSUNLOCKNODE(np);
1465 out:
1466 	ncl_excl_finish(vp, old_lock);
1467 	return error;
1468 }
1469 
1470 /*
1471  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1472  * This is mainly to avoid queueing async I/O requests when the nfsiods
1473  * are all hung on a dead server.
1474  *
1475  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1476  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1477  */
1478 int
1479 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1480 {
1481 	int iod;
1482 	int gotiod;
1483 	int slpflag = 0;
1484 	int slptimeo = 0;
1485 	int error, error2;
1486 
1487 	/*
1488 	 * Commits are usually short and sweet so lets save some cpu and
1489 	 * leave the async daemons for more important rpc's (such as reads
1490 	 * and writes).
1491 	 *
1492 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1493 	 * in the directory in order to update attributes. This can deadlock
1494 	 * with another thread that is waiting for async I/O to be done by
1495 	 * an nfsiod thread while holding a lock on one of these vnodes.
1496 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1497 	 * perform Readdirplus RPCs.
1498 	 */
1499 	NFSLOCKIOD();
1500 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1501 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1502 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1503 		NFSUNLOCKIOD();
1504 		return(EIO);
1505 	}
1506 again:
1507 	if (nmp->nm_flag & NFSMNT_INT)
1508 		slpflag = PCATCH;
1509 	gotiod = FALSE;
1510 
1511 	/*
1512 	 * Find a free iod to process this request.
1513 	 */
1514 	for (iod = 0; iod < ncl_numasync; iod++)
1515 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1516 			gotiod = TRUE;
1517 			break;
1518 		}
1519 
1520 	/*
1521 	 * Try to create one if none are free.
1522 	 */
1523 	if (!gotiod)
1524 		ncl_nfsiodnew();
1525 	else {
1526 		/*
1527 		 * Found one, so wake it up and tell it which
1528 		 * mount to process.
1529 		 */
1530 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1531 		    iod, nmp));
1532 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1533 		ncl_iodmount[iod] = nmp;
1534 		nmp->nm_bufqiods++;
1535 		wakeup(&ncl_iodwant[iod]);
1536 	}
1537 
1538 	/*
1539 	 * If none are free, we may already have an iod working on this mount
1540 	 * point.  If so, it will process our request.
1541 	 */
1542 	if (!gotiod) {
1543 		if (nmp->nm_bufqiods > 0) {
1544 			NFS_DPF(ASYNCIO,
1545 				("ncl_asyncio: %d iods are already processing mount %p\n",
1546 				 nmp->nm_bufqiods, nmp));
1547 			gotiod = TRUE;
1548 		}
1549 	}
1550 
1551 	/*
1552 	 * If we have an iod which can process the request, then queue
1553 	 * the buffer.
1554 	 */
1555 	if (gotiod) {
1556 		/*
1557 		 * Ensure that the queue never grows too large.  We still want
1558 		 * to asynchronize so we block rather then return EIO.
1559 		 */
1560 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1561 			NFS_DPF(ASYNCIO,
1562 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1563 			nmp->nm_bufqwant = TRUE;
1564 			error = newnfs_msleep(td, &nmp->nm_bufq,
1565 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1566 			   slptimeo);
1567 			if (error) {
1568 				error2 = newnfs_sigintr(nmp, td);
1569 				if (error2) {
1570 					NFSUNLOCKIOD();
1571 					return (error2);
1572 				}
1573 				if (slpflag == PCATCH) {
1574 					slpflag = 0;
1575 					slptimeo = 2 * hz;
1576 				}
1577 			}
1578 			/*
1579 			 * We might have lost our iod while sleeping,
1580 			 * so check and loop if necessary.
1581 			 */
1582 			goto again;
1583 		}
1584 
1585 		/* We might have lost our nfsiod */
1586 		if (nmp->nm_bufqiods == 0) {
1587 			NFS_DPF(ASYNCIO,
1588 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1589 			goto again;
1590 		}
1591 
1592 		if (bp->b_iocmd == BIO_READ) {
1593 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1594 				bp->b_rcred = crhold(cred);
1595 		} else {
1596 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1597 				bp->b_wcred = crhold(cred);
1598 		}
1599 
1600 		if (bp->b_flags & B_REMFREE)
1601 			bremfreef(bp);
1602 		BUF_KERNPROC(bp);
1603 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1604 		nmp->nm_bufqlen++;
1605 		if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1606 			NFSLOCKNODE(VTONFS(bp->b_vp));
1607 			VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1608 			VTONFS(bp->b_vp)->n_directio_asyncwr++;
1609 			NFSUNLOCKNODE(VTONFS(bp->b_vp));
1610 		}
1611 		NFSUNLOCKIOD();
1612 		return (0);
1613 	}
1614 
1615 	NFSUNLOCKIOD();
1616 
1617 	/*
1618 	 * All the iods are busy on other mounts, so return EIO to
1619 	 * force the caller to process the i/o synchronously.
1620 	 */
1621 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1622 	return (EIO);
1623 }
1624 
1625 void
1626 ncl_doio_directwrite(struct buf *bp)
1627 {
1628 	int iomode, must_commit;
1629 	struct uio *uiop = (struct uio *)bp->b_caller1;
1630 	char *iov_base = uiop->uio_iov->iov_base;
1631 
1632 	iomode = NFSWRITE_FILESYNC;
1633 	uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1634 	/*
1635 	 * When doing direct I/O we do not care if the
1636 	 * server's write verifier has changed, but we
1637 	 * do not want to update the verifier if it has
1638 	 * changed, since that hides the change from
1639 	 * writes being done through the buffer cache.
1640 	 * By passing must_commit in set to two, the code
1641 	 * in nfsrpc_writerpc() will not update the
1642 	 * verifier on the mount point.
1643 	 */
1644 	must_commit = 2;
1645 	ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0, 0);
1646 	KASSERT((must_commit == 2), ("ncl_doio_directwrite: Updated write"
1647 	    " verifier"));
1648 	if (iomode != NFSWRITE_FILESYNC)
1649 		printf("ncl_doio_directwrite: Broken server "
1650 		    "did not reply FILE_SYNC\n");
1651 	free(iov_base, M_NFSDIRECTIO);
1652 	free(uiop->uio_iov, M_NFSDIRECTIO);
1653 	free(uiop, M_NFSDIRECTIO);
1654 	if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1655 		struct nfsnode *np = VTONFS(bp->b_vp);
1656 		NFSLOCKNODE(np);
1657 		if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1658 			/*
1659 			 * Invalidate the attribute cache, since writes to a DS
1660 			 * won't update the size attribute.
1661 			 */
1662 			np->n_attrstamp = 0;
1663 		}
1664 		np->n_directio_asyncwr--;
1665 		if (np->n_directio_asyncwr == 0) {
1666 			np->n_flag &= ~NMODIFIED;
1667 			if ((np->n_flag & NFSYNCWAIT)) {
1668 				np->n_flag &= ~NFSYNCWAIT;
1669 				wakeup((caddr_t)&np->n_directio_asyncwr);
1670 			}
1671 		}
1672 		NFSUNLOCKNODE(np);
1673 	}
1674 	bp->b_vp = NULL;
1675 	uma_zfree(ncl_pbuf_zone, bp);
1676 }
1677 
1678 /*
1679  * Do an I/O operation to/from a cache block. This may be called
1680  * synchronously or from an nfsiod.
1681  */
1682 int
1683 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1684     int called_from_strategy)
1685 {
1686 	struct uio *uiop;
1687 	struct nfsnode *np;
1688 	struct nfsmount *nmp;
1689 	int error = 0, iomode, must_commit = 0;
1690 	struct uio uio;
1691 	struct iovec io;
1692 	struct proc *p = td ? td->td_proc : NULL;
1693 	uint8_t	iocmd;
1694 
1695 	np = VTONFS(vp);
1696 	nmp = VFSTONFS(vp->v_mount);
1697 	uiop = &uio;
1698 	uiop->uio_iov = &io;
1699 	uiop->uio_iovcnt = 1;
1700 	uiop->uio_segflg = UIO_SYSSPACE;
1701 	uiop->uio_td = td;
1702 
1703 	/*
1704 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1705 	 * do this here so we do not have to do it in all the code that
1706 	 * calls us.
1707 	 */
1708 	bp->b_flags &= ~B_INVAL;
1709 	bp->b_ioflags &= ~BIO_ERROR;
1710 
1711 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1712 	iocmd = bp->b_iocmd;
1713 	if (iocmd == BIO_READ) {
1714 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1715 	    io.iov_base = bp->b_data;
1716 	    uiop->uio_rw = UIO_READ;
1717 
1718 	    switch (vp->v_type) {
1719 	    case VREG:
1720 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1721 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1722 		error = ncl_readrpc(vp, uiop, cr);
1723 
1724 		if (!error) {
1725 		    if (uiop->uio_resid) {
1726 			/*
1727 			 * If we had a short read with no error, we must have
1728 			 * hit a file hole.  We should zero-fill the remainder.
1729 			 * This can also occur if the server hits the file EOF.
1730 			 *
1731 			 * Holes used to be able to occur due to pending
1732 			 * writes, but that is not possible any longer.
1733 			 */
1734 			int nread = bp->b_bcount - uiop->uio_resid;
1735 			ssize_t left = uiop->uio_resid;
1736 
1737 			if (left > 0)
1738 				bzero((char *)bp->b_data + nread, left);
1739 			uiop->uio_resid = 0;
1740 		    }
1741 		}
1742 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1743 		if (p && vp->v_writecount <= -1) {
1744 			NFSLOCKNODE(np);
1745 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1746 				NFSUNLOCKNODE(np);
1747 				PROC_LOCK(p);
1748 				killproc(p, "text file modification");
1749 				PROC_UNLOCK(p);
1750 			} else
1751 				NFSUNLOCKNODE(np);
1752 		}
1753 		break;
1754 	    case VLNK:
1755 		uiop->uio_offset = (off_t)0;
1756 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1757 		error = ncl_readlinkrpc(vp, uiop, cr);
1758 		break;
1759 	    case VDIR:
1760 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1761 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1762 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1763 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1764 			if (error == NFSERR_NOTSUPP)
1765 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1766 		}
1767 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1768 			error = ncl_readdirrpc(vp, uiop, cr, td);
1769 		/*
1770 		 * end-of-directory sets B_INVAL but does not generate an
1771 		 * error.
1772 		 */
1773 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1774 			bp->b_flags |= B_INVAL;
1775 		break;
1776 	    default:
1777 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1778 		break;
1779 	    }
1780 	    if (error) {
1781 		bp->b_ioflags |= BIO_ERROR;
1782 		bp->b_error = error;
1783 	    }
1784 	} else {
1785 	    /*
1786 	     * If we only need to commit, try to commit
1787 	     */
1788 	    if (bp->b_flags & B_NEEDCOMMIT) {
1789 		    int retv;
1790 		    off_t off;
1791 
1792 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1793 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1794 			bp->b_wcred, td);
1795 		    if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1796 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1797 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1798 			    bp->b_resid = 0;
1799 			    bufdone(bp);
1800 			    return (0);
1801 		    }
1802 		    if (retv == NFSERR_STALEWRITEVERF) {
1803 			    ncl_clearcommit(vp->v_mount);
1804 		    }
1805 	    }
1806 
1807 	    /*
1808 	     * Setup for actual write
1809 	     */
1810 	    NFSLOCKNODE(np);
1811 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1812 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1813 	    NFSUNLOCKNODE(np);
1814 
1815 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1816 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1817 		    - bp->b_dirtyoff;
1818 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1819 		    + bp->b_dirtyoff;
1820 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1821 		uiop->uio_rw = UIO_WRITE;
1822 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1823 
1824 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1825 		    iomode = NFSWRITE_UNSTABLE;
1826 		else
1827 		    iomode = NFSWRITE_FILESYNC;
1828 
1829 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1830 		    called_from_strategy, 0);
1831 
1832 		/*
1833 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1834 		 * to cluster the buffers needing commit.  This will allow
1835 		 * the system to submit a single commit rpc for the whole
1836 		 * cluster.  We can do this even if the buffer is not 100%
1837 		 * dirty (relative to the NFS blocksize), so we optimize the
1838 		 * append-to-file-case.
1839 		 *
1840 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1841 		 * cleared because write clustering only works for commit
1842 		 * rpc's, not for the data portion of the write).
1843 		 */
1844 
1845 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1846 		    bp->b_flags |= B_NEEDCOMMIT;
1847 		    if (bp->b_dirtyoff == 0
1848 			&& bp->b_dirtyend == bp->b_bcount)
1849 			bp->b_flags |= B_CLUSTEROK;
1850 		} else {
1851 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1852 		}
1853 
1854 		/*
1855 		 * For an interrupted write, the buffer is still valid
1856 		 * and the write hasn't been pushed to the server yet,
1857 		 * so we can't set BIO_ERROR and report the interruption
1858 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1859 		 * is not relevant, so the rpc attempt is essentially
1860 		 * a noop.  For the case of a V3 write rpc not being
1861 		 * committed to stable storage, the block is still
1862 		 * dirty and requires either a commit rpc or another
1863 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1864 		 * the block is reused. This is indicated by setting
1865 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1866 		 *
1867 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1868 		 * write error and is handled as above, except that
1869 		 * B_EINTR isn't set. One cause of this is a stale stateid
1870 		 * error for the RPC that indicates recovery is required,
1871 		 * when called with called_from_strategy != 0.
1872 		 *
1873 		 * If the buffer is marked B_PAGING, it does not reside on
1874 		 * the vp's paging queues so we cannot call bdirty().  The
1875 		 * bp in this case is not an NFS cache block so we should
1876 		 * be safe. XXX
1877 		 *
1878 		 * The logic below breaks up errors into recoverable and
1879 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1880 		 * and keep the buffer around for potential write retries.
1881 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1882 		 * and save the error in the nfsnode. This is less than ideal
1883 		 * but necessary. Keeping such buffers around could potentially
1884 		 * cause buffer exhaustion eventually (they can never be written
1885 		 * out, so will get constantly be re-dirtied). It also causes
1886 		 * all sorts of vfs panics. For non-recoverable write errors,
1887 		 * also invalidate the attrcache, so we'll be forced to go over
1888 		 * the wire for this object, returning an error to user on next
1889 		 * call (most of the time).
1890 		 */
1891 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1892 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1893 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1894 			if ((bp->b_flags & B_PAGING) == 0) {
1895 			    bdirty(bp);
1896 			    bp->b_flags &= ~B_DONE;
1897 			}
1898 			if ((error == EINTR || error == ETIMEDOUT) &&
1899 			    (bp->b_flags & B_ASYNC) == 0)
1900 			    bp->b_flags |= B_EINTR;
1901 		} else {
1902 		    if (error) {
1903 			bp->b_ioflags |= BIO_ERROR;
1904 			bp->b_flags |= B_INVAL;
1905 			bp->b_error = np->n_error = error;
1906 			NFSLOCKNODE(np);
1907 			np->n_flag |= NWRITEERR;
1908 			np->n_attrstamp = 0;
1909 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1910 			NFSUNLOCKNODE(np);
1911 		    }
1912 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1913 		}
1914 	    } else {
1915 		bp->b_resid = 0;
1916 		bufdone(bp);
1917 		return (0);
1918 	    }
1919 	}
1920 	bp->b_resid = uiop->uio_resid;
1921 	if (must_commit == 1)
1922 	    ncl_clearcommit(vp->v_mount);
1923 	bufdone(bp);
1924 	return (error);
1925 }
1926 
1927 /*
1928  * Used to aid in handling ftruncate() operations on the NFS client side.
1929  * Truncation creates a number of special problems for NFS.  We have to
1930  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1931  * we have to properly handle VM pages or (potentially dirty) buffers
1932  * that straddle the truncation point.
1933  */
1934 
1935 int
1936 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1937 {
1938 	struct nfsnode *np = VTONFS(vp);
1939 	u_quad_t tsize;
1940 	int biosize = vp->v_bufobj.bo_bsize;
1941 	int error = 0;
1942 
1943 	NFSLOCKNODE(np);
1944 	tsize = np->n_size;
1945 	np->n_size = nsize;
1946 	NFSUNLOCKNODE(np);
1947 
1948 	if (nsize < tsize) {
1949 		struct buf *bp;
1950 		daddr_t lbn;
1951 		int bufsize;
1952 
1953 		/*
1954 		 * vtruncbuf() doesn't get the buffer overlapping the
1955 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1956 		 * buffer that now needs to be truncated.
1957 		 */
1958 		error = vtruncbuf(vp, nsize, biosize);
1959 		lbn = nsize / biosize;
1960 		bufsize = nsize - (lbn * biosize);
1961 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1962 		if (!bp)
1963 			return EINTR;
1964 		if (bp->b_dirtyoff > bp->b_bcount)
1965 			bp->b_dirtyoff = bp->b_bcount;
1966 		if (bp->b_dirtyend > bp->b_bcount)
1967 			bp->b_dirtyend = bp->b_bcount;
1968 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1969 		brelse(bp);
1970 	} else {
1971 		vnode_pager_setsize(vp, nsize);
1972 	}
1973 	return(error);
1974 }
1975