xref: /titanic_41/usr/src/uts/common/fs/nfs/nfs_vnops.c (revision 3173664e967186de0a5f0e61548c25996fa6d37f)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  *
25  *	Copyright (c) 1983,1984,1985,1986,1987,1988,1989 AT&T.
26  *	All rights reserved.
27  */
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 #include <sys/param.h>
32 #include <sys/types.h>
33 #include <sys/systm.h>
34 #include <sys/cred.h>
35 #include <sys/time.h>
36 #include <sys/vnode.h>
37 #include <sys/vfs.h>
38 #include <sys/file.h>
39 #include <sys/filio.h>
40 #include <sys/uio.h>
41 #include <sys/buf.h>
42 #include <sys/mman.h>
43 #include <sys/pathname.h>
44 #include <sys/dirent.h>
45 #include <sys/debug.h>
46 #include <sys/vmsystm.h>
47 #include <sys/fcntl.h>
48 #include <sys/flock.h>
49 #include <sys/swap.h>
50 #include <sys/errno.h>
51 #include <sys/strsubr.h>
52 #include <sys/sysmacros.h>
53 #include <sys/kmem.h>
54 #include <sys/cmn_err.h>
55 #include <sys/pathconf.h>
56 #include <sys/utsname.h>
57 #include <sys/dnlc.h>
58 #include <sys/acl.h>
59 #include <sys/atomic.h>
60 #include <sys/policy.h>
61 #include <sys/sdt.h>
62 
63 #include <rpc/types.h>
64 #include <rpc/auth.h>
65 #include <rpc/clnt.h>
66 
67 #include <nfs/nfs.h>
68 #include <nfs/nfs_clnt.h>
69 #include <nfs/rnode.h>
70 #include <nfs/nfs_acl.h>
71 #include <nfs/lm.h>
72 
73 #include <vm/hat.h>
74 #include <vm/as.h>
75 #include <vm/page.h>
76 #include <vm/pvn.h>
77 #include <vm/seg.h>
78 #include <vm/seg_map.h>
79 #include <vm/seg_kpm.h>
80 #include <vm/seg_vn.h>
81 
82 #include <fs/fs_subr.h>
83 
84 #include <sys/ddi.h>
85 
86 static int	nfs_rdwrlbn(vnode_t *, page_t *, u_offset_t, size_t, int,
87 			cred_t *);
88 static int	nfswrite(vnode_t *, caddr_t, uint_t, int, cred_t *);
89 static int	nfsread(vnode_t *, caddr_t, uint_t, int, size_t *, cred_t *);
90 static int	nfssetattr(vnode_t *, struct vattr *, int, cred_t *);
91 static int	nfslookup_dnlc(vnode_t *, char *, vnode_t **, cred_t *);
92 static int	nfslookup_otw(vnode_t *, char *, vnode_t **, cred_t *, int);
93 static int	nfsrename(vnode_t *, char *, vnode_t *, char *, cred_t *);
94 static int	nfsreaddir(vnode_t *, rddir_cache *, cred_t *);
95 static int	nfs_bio(struct buf *, cred_t *);
96 static int	nfs_getapage(vnode_t *, u_offset_t, size_t, uint_t *,
97 			page_t *[], size_t, struct seg *, caddr_t,
98 			enum seg_rw, cred_t *);
99 static void	nfs_readahead(vnode_t *, u_offset_t, caddr_t, struct seg *,
100 			cred_t *);
101 static int	nfs_sync_putapage(vnode_t *, page_t *, u_offset_t, size_t,
102 			int, cred_t *);
103 static int	nfs_sync_pageio(vnode_t *, page_t *, u_offset_t, size_t,
104 			int, cred_t *);
105 static void	nfs_delmap_callback(struct as *, void *, uint_t);
106 
107 /*
108  * Error flags used to pass information about certain special errors
109  * which need to be handled specially.
110  */
111 #define	NFS_EOF			-98
112 
113 /*
114  * These are the vnode ops routines which implement the vnode interface to
115  * the networked file system.  These routines just take their parameters,
116  * make them look networkish by putting the right info into interface structs,
117  * and then calling the appropriate remote routine(s) to do the work.
118  *
119  * Note on directory name lookup cacheing:  If we detect a stale fhandle,
120  * we purge the directory cache relative to that vnode.  This way, the
121  * user won't get burned by the cache repeatedly.  See <nfs/rnode.h> for
122  * more details on rnode locking.
123  */
124 
125 static int	nfs_open(vnode_t **, int, cred_t *);
126 static int	nfs_close(vnode_t *, int, int, offset_t, cred_t *);
127 static int	nfs_read(vnode_t *, struct uio *, int, cred_t *,
128 			caller_context_t *);
129 static int	nfs_write(vnode_t *, struct uio *, int, cred_t *,
130 			caller_context_t *);
131 static int	nfs_ioctl(vnode_t *, int, intptr_t, int, cred_t *, int *);
132 static int	nfs_getattr(vnode_t *, struct vattr *, int, cred_t *);
133 static int	nfs_setattr(vnode_t *, struct vattr *, int, cred_t *,
134 			caller_context_t *);
135 static int	nfs_access(vnode_t *, int, int, cred_t *);
136 static int	nfs_accessx(void *, int, cred_t *);
137 static int	nfs_readlink(vnode_t *, struct uio *, cred_t *);
138 static int	nfs_fsync(vnode_t *, int, cred_t *);
139 static void	nfs_inactive(vnode_t *, cred_t *);
140 static int	nfs_lookup(vnode_t *, char *, vnode_t **, struct pathname *,
141 			int, vnode_t *, cred_t *);
142 static int	nfs_create(vnode_t *, char *, struct vattr *, enum vcexcl,
143 			int, vnode_t **, cred_t *, int);
144 static int	nfs_remove(vnode_t *, char *, cred_t *);
145 static int	nfs_link(vnode_t *, vnode_t *, char *, cred_t *);
146 static int	nfs_rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
147 static int	nfs_mkdir(vnode_t *, char *, struct vattr *,
148 			vnode_t **, cred_t *);
149 static int	nfs_rmdir(vnode_t *, char *, vnode_t *, cred_t *);
150 static int	nfs_symlink(vnode_t *, char *, struct vattr *, char *,
151 			cred_t *);
152 static int	nfs_readdir(vnode_t *, struct uio *, cred_t *, int *);
153 static int	nfs_fid(vnode_t *, fid_t *);
154 static int	nfs_rwlock(vnode_t *, int, caller_context_t *);
155 static void	nfs_rwunlock(vnode_t *, int, caller_context_t *);
156 static int	nfs_seek(vnode_t *, offset_t, offset_t *);
157 static int	nfs_getpage(vnode_t *, offset_t, size_t, uint_t *,
158 			page_t *[], size_t, struct seg *, caddr_t,
159 			enum seg_rw, cred_t *);
160 static int	nfs_putpage(vnode_t *, offset_t, size_t, int, cred_t *);
161 static int	nfs_map(vnode_t *, offset_t, struct as *, caddr_t *,
162 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
163 static int	nfs_addmap(vnode_t *, offset_t, struct as *, caddr_t,
164 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
165 static int	nfs_frlock(vnode_t *, int, struct flock64 *, int, offset_t,
166 			struct flk_callback *, cred_t *);
167 static int	nfs_space(vnode_t *, int, struct flock64 *, int, offset_t,
168 			cred_t *, caller_context_t *);
169 static int	nfs_realvp(vnode_t *, vnode_t **);
170 static int	nfs_delmap(vnode_t *, offset_t, struct as *, caddr_t,
171 			size_t, uint_t, uint_t, uint_t, cred_t *);
172 static int	nfs_pathconf(vnode_t *, int, ulong_t *, cred_t *);
173 static int	nfs_pageio(vnode_t *, page_t *, u_offset_t, size_t, int,
174 			cred_t *);
175 static int	nfs_setsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
176 static int	nfs_getsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
177 static int	nfs_shrlock(vnode_t *, int, struct shrlock *, int, cred_t *);
178 
179 struct vnodeops *nfs_vnodeops;
180 
181 const fs_operation_def_t nfs_vnodeops_template[] = {
182 	VOPNAME_OPEN, nfs_open,
183 	VOPNAME_CLOSE, nfs_close,
184 	VOPNAME_READ, nfs_read,
185 	VOPNAME_WRITE, nfs_write,
186 	VOPNAME_IOCTL, nfs_ioctl,
187 	VOPNAME_GETATTR, nfs_getattr,
188 	VOPNAME_SETATTR, nfs_setattr,
189 	VOPNAME_ACCESS, nfs_access,
190 	VOPNAME_LOOKUP, nfs_lookup,
191 	VOPNAME_CREATE, nfs_create,
192 	VOPNAME_REMOVE, nfs_remove,
193 	VOPNAME_LINK, nfs_link,
194 	VOPNAME_RENAME, nfs_rename,
195 	VOPNAME_MKDIR, nfs_mkdir,
196 	VOPNAME_RMDIR, nfs_rmdir,
197 	VOPNAME_READDIR, nfs_readdir,
198 	VOPNAME_SYMLINK, nfs_symlink,
199 	VOPNAME_READLINK, nfs_readlink,
200 	VOPNAME_FSYNC, nfs_fsync,
201 	VOPNAME_INACTIVE, (fs_generic_func_p) nfs_inactive,
202 	VOPNAME_FID, nfs_fid,
203 	VOPNAME_RWLOCK, nfs_rwlock,
204 	VOPNAME_RWUNLOCK, (fs_generic_func_p) nfs_rwunlock,
205 	VOPNAME_SEEK, nfs_seek,
206 	VOPNAME_FRLOCK, nfs_frlock,
207 	VOPNAME_SPACE, nfs_space,
208 	VOPNAME_REALVP, nfs_realvp,
209 	VOPNAME_GETPAGE, nfs_getpage,
210 	VOPNAME_PUTPAGE, nfs_putpage,
211 	VOPNAME_MAP, (fs_generic_func_p) nfs_map,
212 	VOPNAME_ADDMAP, (fs_generic_func_p) nfs_addmap,
213 	VOPNAME_DELMAP, nfs_delmap,
214 	VOPNAME_DUMP, nfs_dump,
215 	VOPNAME_PATHCONF, nfs_pathconf,
216 	VOPNAME_PAGEIO, nfs_pageio,
217 	VOPNAME_SETSECATTR, nfs_setsecattr,
218 	VOPNAME_GETSECATTR, nfs_getsecattr,
219 	VOPNAME_SHRLOCK, nfs_shrlock,
220 	NULL, NULL
221 };
222 
223 /*
224  * XXX:  This is referenced in modstubs.s
225  */
226 struct vnodeops *
227 nfs_getvnodeops(void)
228 {
229 	return (nfs_vnodeops);
230 }
231 
232 /* ARGSUSED */
233 static int
234 nfs_open(vnode_t **vpp, int flag, cred_t *cr)
235 {
236 	int error;
237 	struct vattr va;
238 	rnode_t *rp;
239 	vnode_t *vp;
240 
241 	vp = *vpp;
242 	rp = VTOR(vp);
243 	if (nfs_zone() != VTOMI(vp)->mi_zone)
244 		return (EIO);
245 	mutex_enter(&rp->r_statelock);
246 	if (rp->r_cred == NULL) {
247 		crhold(cr);
248 		rp->r_cred = cr;
249 	}
250 	mutex_exit(&rp->r_statelock);
251 
252 	/*
253 	 * If there is no cached data or if close-to-open
254 	 * consistency checking is turned off, we can avoid
255 	 * the over the wire getattr.  Otherwise, if the
256 	 * file system is mounted readonly, then just verify
257 	 * the caches are up to date using the normal mechanism.
258 	 * Else, if the file is not mmap'd, then just mark
259 	 * the attributes as timed out.  They will be refreshed
260 	 * and the caches validated prior to being used.
261 	 * Else, the file system is mounted writeable so
262 	 * force an over the wire GETATTR in order to ensure
263 	 * that all cached data is valid.
264 	 */
265 	if (vp->v_count > 1 ||
266 	    ((vn_has_cached_data(vp) || HAVE_RDDIR_CACHE(rp)) &&
267 	    !(VTOMI(vp)->mi_flags & MI_NOCTO))) {
268 		if (vn_is_readonly(vp))
269 			error = nfs_validate_caches(vp, cr);
270 		else if (rp->r_mapcnt == 0 && vp->v_count == 1) {
271 			PURGE_ATTRCACHE(vp);
272 			error = 0;
273 		} else {
274 			va.va_mask = AT_ALL;
275 			error = nfs_getattr_otw(vp, &va, cr);
276 		}
277 	} else
278 		error = 0;
279 
280 	return (error);
281 }
282 
283 static int
284 nfs_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr)
285 {
286 	rnode_t *rp;
287 	int error;
288 	struct vattr va;
289 
290 	/*
291 	 * zone_enter(2) prevents processes from changing zones with NFS files
292 	 * open; if we happen to get here from the wrong zone we can't do
293 	 * anything over the wire.
294 	 */
295 	if (VTOMI(vp)->mi_zone != nfs_zone()) {
296 		/*
297 		 * We could attempt to clean up locks, except we're sure
298 		 * that the current process didn't acquire any locks on
299 		 * the file: any attempt to lock a file belong to another zone
300 		 * will fail, and one can't lock an NFS file and then change
301 		 * zones, as that fails too.
302 		 *
303 		 * Returning an error here is the sane thing to do.  A
304 		 * subsequent call to VN_RELE() which translates to a
305 		 * nfs_inactive() will clean up state: if the zone of the
306 		 * vnode's origin is still alive and kicking, an async worker
307 		 * thread will handle the request (from the correct zone), and
308 		 * everything (minus the final nfs_getattr_otw() call) should
309 		 * be OK. If the zone is going away nfs_async_inactive() will
310 		 * throw away cached pages inline.
311 		 */
312 		return (EIO);
313 	}
314 
315 	/*
316 	 * If we are using local locking for this filesystem, then
317 	 * release all of the SYSV style record locks.  Otherwise,
318 	 * we are doing network locking and we need to release all
319 	 * of the network locks.  All of the locks held by this
320 	 * process on this file are released no matter what the
321 	 * incoming reference count is.
322 	 */
323 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
324 		cleanlocks(vp, ttoproc(curthread)->p_pid, 0);
325 		cleanshares(vp, ttoproc(curthread)->p_pid);
326 	} else
327 		nfs_lockrelease(vp, flag, offset, cr);
328 
329 	if (count > 1)
330 		return (0);
331 
332 	/*
333 	 * If the file has been `unlinked', then purge the
334 	 * DNLC so that this vnode will get reycled quicker
335 	 * and the .nfs* file on the server will get removed.
336 	 */
337 	rp = VTOR(vp);
338 	if (rp->r_unldvp != NULL)
339 		dnlc_purge_vp(vp);
340 
341 	/*
342 	 * If the file was open for write and there are pages,
343 	 * then if the file system was mounted using the "no-close-
344 	 *	to-open" semantics, then start an asynchronous flush
345 	 *	of the all of the pages in the file.
346 	 * else the file system was not mounted using the "no-close-
347 	 *	to-open" semantics, then do a synchronous flush and
348 	 *	commit of all of the dirty and uncommitted pages.
349 	 *
350 	 * The asynchronous flush of the pages in the "nocto" path
351 	 * mostly just associates a cred pointer with the rnode so
352 	 * writes which happen later will have a better chance of
353 	 * working.  It also starts the data being written to the
354 	 * server, but without unnecessarily delaying the application.
355 	 */
356 	if ((flag & FWRITE) && vn_has_cached_data(vp)) {
357 		if ((VTOMI(vp)->mi_flags & MI_NOCTO)) {
358 			error = nfs_putpage(vp, (offset_t)0, 0, B_ASYNC, cr);
359 			if (error == EAGAIN)
360 				error = 0;
361 		} else
362 			error = nfs_putpage(vp, (offset_t)0, 0, 0, cr);
363 		if (!error) {
364 			mutex_enter(&rp->r_statelock);
365 			error = rp->r_error;
366 			rp->r_error = 0;
367 			mutex_exit(&rp->r_statelock);
368 		}
369 	} else {
370 		mutex_enter(&rp->r_statelock);
371 		error = rp->r_error;
372 		rp->r_error = 0;
373 		mutex_exit(&rp->r_statelock);
374 	}
375 
376 	/*
377 	 * If RWRITEATTR is set, then issue an over the wire GETATTR to
378 	 * refresh the attribute cache with a set of attributes which
379 	 * weren't returned from a WRITE.  This will enable the close-
380 	 * to-open processing to work.
381 	 */
382 	if (rp->r_flags & RWRITEATTR)
383 		(void) nfs_getattr_otw(vp, &va, cr);
384 
385 	return (error);
386 }
387 
388 /* ARGSUSED */
389 static int
390 nfs_read(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
391 	caller_context_t *ct)
392 {
393 	rnode_t *rp;
394 	u_offset_t off;
395 	offset_t diff;
396 	int on;
397 	size_t n;
398 	caddr_t base;
399 	uint_t flags;
400 	int error;
401 	mntinfo_t *mi;
402 
403 	rp = VTOR(vp);
404 	mi = VTOMI(vp);
405 
406 	if (nfs_zone() != mi->mi_zone)
407 		return (EIO);
408 
409 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
410 
411 	if (vp->v_type != VREG)
412 		return (EISDIR);
413 
414 	if (uiop->uio_resid == 0)
415 		return (0);
416 
417 	if (uiop->uio_loffset > MAXOFF32_T)
418 		return (EFBIG);
419 
420 	if (uiop->uio_loffset < 0 ||
421 	    uiop->uio_loffset + uiop->uio_resid > MAXOFF32_T)
422 		return (EINVAL);
423 
424 	/*
425 	 * Bypass VM if caching has been disabled (e.g., locking) or if
426 	 * using client-side direct I/O and the file is not mmap'd and
427 	 * there are no cached pages.
428 	 */
429 	if ((vp->v_flag & VNOCACHE) ||
430 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
431 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
432 		size_t bufsize;
433 		size_t resid = 0;
434 
435 		/*
436 		 * Let's try to do read in as large a chunk as we can
437 		 * (Filesystem (NFS client) bsize if possible/needed).
438 		 * For V3, this is 32K and for V2, this is 8K.
439 		 */
440 		bufsize = MIN(uiop->uio_resid, VTOMI(vp)->mi_curread);
441 		base = kmem_alloc(bufsize, KM_SLEEP);
442 		do {
443 			n = MIN(uiop->uio_resid, bufsize);
444 			error = nfsread(vp, base, uiop->uio_offset, n,
445 			    &resid, cr);
446 			if (!error) {
447 				n -= resid;
448 				error = uiomove(base, n, UIO_READ, uiop);
449 			}
450 		} while (!error && uiop->uio_resid > 0 && n > 0);
451 		kmem_free(base, bufsize);
452 		return (error);
453 	}
454 
455 	error = 0;
456 
457 	do {
458 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
459 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
460 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
461 
462 		error = nfs_validate_caches(vp, cr);
463 		if (error)
464 			break;
465 
466 		mutex_enter(&rp->r_statelock);
467 		diff = rp->r_size - uiop->uio_loffset;
468 		mutex_exit(&rp->r_statelock);
469 		if (diff <= 0)
470 			break;
471 		if (diff < n)
472 			n = (size_t)diff;
473 
474 		if (vpm_enable) {
475 			/*
476 			 * Copy data.
477 			 */
478 			error = vpm_data_copy(vp, off + on, n, uiop,
479 						1, NULL, 0, S_READ);
480 		} else {
481 			base = segmap_getmapflt(segkmap, vp, off + on, n,
482 								1, S_READ);
483 			error = uiomove(base + on, n, UIO_READ, uiop);
484 		}
485 
486 		if (!error) {
487 			/*
488 			 * If read a whole block or read to eof,
489 			 * won't need this buffer again soon.
490 			 */
491 			mutex_enter(&rp->r_statelock);
492 			if (n + on == MAXBSIZE ||
493 			    uiop->uio_loffset == rp->r_size)
494 				flags = SM_DONTNEED;
495 			else
496 				flags = 0;
497 			mutex_exit(&rp->r_statelock);
498 			if (vpm_enable) {
499 				error = vpm_sync_pages(vp, off, n, flags);
500 			} else {
501 				error = segmap_release(segkmap, base, flags);
502 			}
503 		} else {
504 			if (vpm_enable) {
505 				(void) vpm_sync_pages(vp, off, n, 0);
506 			} else {
507 				(void) segmap_release(segkmap, base, 0);
508 			}
509 		}
510 	} while (!error && uiop->uio_resid > 0);
511 
512 	return (error);
513 }
514 
515 /* ARGSUSED */
516 static int
517 nfs_write(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
518 	caller_context_t *ct)
519 {
520 	rnode_t *rp;
521 	u_offset_t off;
522 	caddr_t base;
523 	uint_t flags;
524 	int remainder;
525 	size_t n;
526 	int on;
527 	int error;
528 	int resid;
529 	offset_t offset;
530 	rlim_t limit;
531 	mntinfo_t *mi;
532 
533 	rp = VTOR(vp);
534 
535 	mi = VTOMI(vp);
536 	if (nfs_zone() != mi->mi_zone)
537 		return (EIO);
538 	if (vp->v_type != VREG)
539 		return (EISDIR);
540 
541 	if (uiop->uio_resid == 0)
542 		return (0);
543 
544 	if (ioflag & FAPPEND) {
545 		struct vattr va;
546 
547 		/*
548 		 * Must serialize if appending.
549 		 */
550 		if (nfs_rw_lock_held(&rp->r_rwlock, RW_READER)) {
551 			nfs_rw_exit(&rp->r_rwlock);
552 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER,
553 			    INTR(vp)))
554 				return (EINTR);
555 		}
556 
557 		va.va_mask = AT_SIZE;
558 		error = nfsgetattr(vp, &va, cr);
559 		if (error)
560 			return (error);
561 		uiop->uio_loffset = va.va_size;
562 	}
563 
564 	if (uiop->uio_loffset > MAXOFF32_T)
565 		return (EFBIG);
566 
567 	offset = uiop->uio_loffset + uiop->uio_resid;
568 
569 	if (uiop->uio_loffset < 0 || offset > MAXOFF32_T)
570 		return (EINVAL);
571 
572 	if (uiop->uio_llimit > (rlim64_t)MAXOFF32_T) {
573 		limit = MAXOFF32_T;
574 	} else {
575 		limit = (rlim_t)uiop->uio_llimit;
576 	}
577 
578 	/*
579 	 * Check to make sure that the process will not exceed
580 	 * its limit on file size.  It is okay to write up to
581 	 * the limit, but not beyond.  Thus, the write which
582 	 * reaches the limit will be short and the next write
583 	 * will return an error.
584 	 */
585 	remainder = 0;
586 	if (offset > limit) {
587 		remainder = offset - limit;
588 		uiop->uio_resid = limit - uiop->uio_offset;
589 		if (uiop->uio_resid <= 0) {
590 			proc_t *p = ttoproc(curthread);
591 
592 			uiop->uio_resid += remainder;
593 			mutex_enter(&p->p_lock);
594 			(void) rctl_action(rctlproc_legacy[RLIMIT_FSIZE],
595 			    p->p_rctls, p, RCA_UNSAFE_SIGINFO);
596 			mutex_exit(&p->p_lock);
597 			return (EFBIG);
598 		}
599 	}
600 
601 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
602 		return (EINTR);
603 
604 	/*
605 	 * Bypass VM if caching has been disabled (e.g., locking) or if
606 	 * using client-side direct I/O and the file is not mmap'd and
607 	 * there are no cached pages.
608 	 */
609 	if ((vp->v_flag & VNOCACHE) ||
610 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
611 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
612 		size_t bufsize;
613 		int count;
614 		uint_t org_offset;
615 
616 nfs_fwrite:
617 		if (rp->r_flags & RSTALE) {
618 			resid = uiop->uio_resid;
619 			offset = uiop->uio_loffset;
620 			error = rp->r_error;
621 			goto bottom;
622 		}
623 		bufsize = MIN(uiop->uio_resid, mi->mi_curwrite);
624 		base = kmem_alloc(bufsize, KM_SLEEP);
625 		do {
626 			resid = uiop->uio_resid;
627 			offset = uiop->uio_loffset;
628 			count = MIN(uiop->uio_resid, bufsize);
629 			org_offset = uiop->uio_offset;
630 			error = uiomove(base, count, UIO_WRITE, uiop);
631 			if (!error) {
632 				error = nfswrite(vp, base, org_offset,
633 				    count, cr);
634 			}
635 		} while (!error && uiop->uio_resid > 0);
636 		kmem_free(base, bufsize);
637 		goto bottom;
638 	}
639 
640 	do {
641 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
642 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
643 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
644 
645 		resid = uiop->uio_resid;
646 		offset = uiop->uio_loffset;
647 
648 		if (rp->r_flags & RSTALE) {
649 			error = rp->r_error;
650 			break;
651 		}
652 
653 		/*
654 		 * Don't create dirty pages faster than they
655 		 * can be cleaned so that the system doesn't
656 		 * get imbalanced.  If the async queue is
657 		 * maxed out, then wait for it to drain before
658 		 * creating more dirty pages.  Also, wait for
659 		 * any threads doing pagewalks in the vop_getattr
660 		 * entry points so that they don't block for
661 		 * long periods.
662 		 */
663 		mutex_enter(&rp->r_statelock);
664 		while ((mi->mi_max_threads != 0 &&
665 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
666 		    rp->r_gcount > 0)
667 			cv_wait(&rp->r_cv, &rp->r_statelock);
668 		mutex_exit(&rp->r_statelock);
669 
670 		if (vpm_enable) {
671 			/*
672 			 * It will use kpm mappings, so no need to
673 			 * pass an address.
674 			 */
675 			error = writerp(rp, NULL, n, uiop, 0);
676 		} else  {
677 			if (segmap_kpm) {
678 				int pon = uiop->uio_loffset & PAGEOFFSET;
679 				size_t pn = MIN(PAGESIZE - pon,
680 							uiop->uio_resid);
681 				int pagecreate;
682 
683 				mutex_enter(&rp->r_statelock);
684 				pagecreate = (pon == 0) && (pn == PAGESIZE ||
685 					uiop->uio_loffset + pn >= rp->r_size);
686 				mutex_exit(&rp->r_statelock);
687 
688 				base = segmap_getmapflt(segkmap, vp, off + on,
689 						pn, !pagecreate, S_WRITE);
690 
691 				error = writerp(rp, base + pon, n, uiop,
692 								pagecreate);
693 
694 			} else {
695 				base = segmap_getmapflt(segkmap, vp, off + on,
696 							n, 0, S_READ);
697 				error = writerp(rp, base + on, n, uiop, 0);
698 			}
699 		}
700 
701 		if (!error) {
702 			if (mi->mi_flags & MI_NOAC)
703 				flags = SM_WRITE;
704 			else if (n + on == MAXBSIZE || IS_SWAPVP(vp)) {
705 				/*
706 				 * Have written a whole block.
707 				 * Start an asynchronous write
708 				 * and mark the buffer to
709 				 * indicate that it won't be
710 				 * needed again soon.
711 				 */
712 				flags = SM_WRITE | SM_ASYNC | SM_DONTNEED;
713 			} else
714 				flags = 0;
715 			if ((ioflag & (FSYNC|FDSYNC)) ||
716 			    (rp->r_flags & ROUTOFSPACE)) {
717 				flags &= ~SM_ASYNC;
718 				flags |= SM_WRITE;
719 			}
720 			if (vpm_enable) {
721 				error = vpm_sync_pages(vp, off, n, flags);
722 			} else {
723 				error = segmap_release(segkmap, base, flags);
724 			}
725 		} else {
726 			if (vpm_enable) {
727 				(void) vpm_sync_pages(vp, off, n, 0);
728 			} else {
729 				(void) segmap_release(segkmap, base, 0);
730 			}
731 			/*
732 			 * In the event that we got an access error while
733 			 * faulting in a page for a write-only file just
734 			 * force a write.
735 			 */
736 			if (error == EACCES)
737 				goto nfs_fwrite;
738 		}
739 	} while (!error && uiop->uio_resid > 0);
740 
741 bottom:
742 	if (error) {
743 		uiop->uio_resid = resid + remainder;
744 		uiop->uio_loffset = offset;
745 	} else
746 		uiop->uio_resid += remainder;
747 
748 	nfs_rw_exit(&rp->r_lkserlock);
749 
750 	return (error);
751 }
752 
753 /*
754  * Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED}
755  */
756 static int
757 nfs_rdwrlbn(vnode_t *vp, page_t *pp, u_offset_t off, size_t len,
758 	int flags, cred_t *cr)
759 {
760 	struct buf *bp;
761 	int error;
762 
763 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
764 	bp = pageio_setup(pp, len, vp, flags);
765 	ASSERT(bp != NULL);
766 
767 	/*
768 	 * pageio_setup should have set b_addr to 0.  This
769 	 * is correct since we want to do I/O on a page
770 	 * boundary.  bp_mapin will use this addr to calculate
771 	 * an offset, and then set b_addr to the kernel virtual
772 	 * address it allocated for us.
773 	 */
774 	ASSERT(bp->b_un.b_addr == 0);
775 
776 	bp->b_edev = 0;
777 	bp->b_dev = 0;
778 	bp->b_lblkno = lbtodb(off);
779 	bp->b_file = vp;
780 	bp->b_offset = (offset_t)off;
781 	bp_mapin(bp);
782 
783 	error = nfs_bio(bp, cr);
784 
785 	bp_mapout(bp);
786 	pageio_done(bp);
787 
788 	return (error);
789 }
790 
791 /*
792  * Write to file.  Writes to remote server in largest size
793  * chunks that the server can handle.  Write is synchronous.
794  */
795 static int
796 nfswrite(vnode_t *vp, caddr_t base, uint_t offset, int count, cred_t *cr)
797 {
798 	rnode_t *rp;
799 	mntinfo_t *mi;
800 	struct nfswriteargs wa;
801 	struct nfsattrstat ns;
802 	int error;
803 	int tsize;
804 	int douprintf;
805 
806 	douprintf = 1;
807 
808 	rp = VTOR(vp);
809 	mi = VTOMI(vp);
810 
811 	ASSERT(nfs_zone() == mi->mi_zone);
812 
813 	wa.wa_args = &wa.wa_args_buf;
814 	wa.wa_fhandle = *VTOFH(vp);
815 
816 	do {
817 		tsize = MIN(mi->mi_curwrite, count);
818 		wa.wa_data = base;
819 		wa.wa_begoff = offset;
820 		wa.wa_totcount = tsize;
821 		wa.wa_count = tsize;
822 		wa.wa_offset = offset;
823 
824 		if (mi->mi_io_kstats) {
825 			mutex_enter(&mi->mi_lock);
826 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
827 			mutex_exit(&mi->mi_lock);
828 		}
829 		wa.wa_mblk = NULL;
830 		do {
831 			error = rfs2call(mi, RFS_WRITE,
832 			    xdr_writeargs, (caddr_t)&wa,
833 			    xdr_attrstat, (caddr_t)&ns, cr,
834 			    &douprintf, &ns.ns_status, 0, NULL);
835 		} while (error == ENFS_TRYAGAIN);
836 		if (mi->mi_io_kstats) {
837 			mutex_enter(&mi->mi_lock);
838 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
839 			mutex_exit(&mi->mi_lock);
840 		}
841 
842 		if (!error) {
843 			error = geterrno(ns.ns_status);
844 			/*
845 			 * Can't check for stale fhandle and purge caches
846 			 * here because pages are held by nfs_getpage.
847 			 * Just mark the attribute cache as timed out
848 			 * and set RWRITEATTR to indicate that the file
849 			 * was modified with a WRITE operation.
850 			 */
851 			if (!error) {
852 				count -= tsize;
853 				base += tsize;
854 				offset += tsize;
855 				if (mi->mi_io_kstats) {
856 					mutex_enter(&mi->mi_lock);
857 				    KSTAT_IO_PTR(mi->mi_io_kstats)->writes++;
858 				    KSTAT_IO_PTR(mi->mi_io_kstats)->nwritten +=
859 					    tsize;
860 					mutex_exit(&mi->mi_lock);
861 				}
862 				lwp_stat_update(LWP_STAT_OUBLK, 1);
863 				mutex_enter(&rp->r_statelock);
864 				PURGE_ATTRCACHE_LOCKED(rp);
865 				rp->r_flags |= RWRITEATTR;
866 				mutex_exit(&rp->r_statelock);
867 			}
868 		}
869 	} while (!error && count);
870 
871 	return (error);
872 }
873 
874 /*
875  * Read from a file.  Reads data in largest chunks our interface can handle.
876  */
877 static int
878 nfsread(vnode_t *vp, caddr_t base, uint_t offset, int count, size_t *residp,
879 	cred_t *cr)
880 {
881 	mntinfo_t *mi;
882 	struct nfsreadargs ra;
883 	struct nfsrdresult rr;
884 	int tsize;
885 	int error;
886 	int douprintf;
887 	failinfo_t fi;
888 	rnode_t *rp;
889 	struct vattr va;
890 	hrtime_t t;
891 
892 	rp = VTOR(vp);
893 	mi = VTOMI(vp);
894 
895 	ASSERT(nfs_zone() == mi->mi_zone);
896 
897 	douprintf = 1;
898 
899 	ra.ra_fhandle = *VTOFH(vp);
900 
901 	fi.vp = vp;
902 	fi.fhp = (caddr_t)&ra.ra_fhandle;
903 	fi.copyproc = nfscopyfh;
904 	fi.lookupproc = nfslookup;
905 	fi.xattrdirproc = acl_getxattrdir2;
906 
907 	do {
908 		if (mi->mi_io_kstats) {
909 			mutex_enter(&mi->mi_lock);
910 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
911 			mutex_exit(&mi->mi_lock);
912 		}
913 
914 		do {
915 			tsize = MIN(mi->mi_curread, count);
916 			rr.rr_data = base;
917 			ra.ra_offset = offset;
918 			ra.ra_totcount = tsize;
919 			ra.ra_count = tsize;
920 			t = gethrtime();
921 			error = rfs2call(mi, RFS_READ,
922 			    xdr_readargs, (caddr_t)&ra,
923 			    xdr_rdresult, (caddr_t)&rr, cr,
924 			    &douprintf, &rr.rr_status, 0, &fi);
925 		} while (error == ENFS_TRYAGAIN);
926 
927 		if (mi->mi_io_kstats) {
928 			mutex_enter(&mi->mi_lock);
929 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
930 			mutex_exit(&mi->mi_lock);
931 		}
932 
933 		if (!error) {
934 			error = geterrno(rr.rr_status);
935 			if (!error) {
936 				count -= rr.rr_count;
937 				base += rr.rr_count;
938 				offset += rr.rr_count;
939 				if (mi->mi_io_kstats) {
940 					mutex_enter(&mi->mi_lock);
941 					KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
942 					KSTAT_IO_PTR(mi->mi_io_kstats)->nread +=
943 					    rr.rr_count;
944 					mutex_exit(&mi->mi_lock);
945 				}
946 				lwp_stat_update(LWP_STAT_INBLK, 1);
947 			}
948 		}
949 	} while (!error && count && rr.rr_count == tsize);
950 
951 	*residp = count;
952 
953 	if (!error) {
954 		/*
955 		 * Since no error occurred, we have the current
956 		 * attributes and we need to do a cache check and then
957 		 * potentially update the cached attributes.  We can't
958 		 * use the normal attribute check and cache mechanisms
959 		 * because they might cause a cache flush which would
960 		 * deadlock.  Instead, we just check the cache to see
961 		 * if the attributes have changed.  If it is, then we
962 		 * just mark the attributes as out of date.  The next
963 		 * time that the attributes are checked, they will be
964 		 * out of date, new attributes will be fetched, and
965 		 * the page cache will be flushed.  If the attributes
966 		 * weren't changed, then we just update the cached
967 		 * attributes with these attributes.
968 		 */
969 		/*
970 		 * If NFS_ACL is supported on the server, then the
971 		 * attributes returned by server may have minimal
972 		 * permissions sometimes denying access to users having
973 		 * proper access.  To get the proper attributes, mark
974 		 * the attributes as expired so that they will be
975 		 * regotten via the NFS_ACL GETATTR2 procedure.
976 		 */
977 		error = nattr_to_vattr(vp, &rr.rr_attr, &va);
978 		mutex_enter(&rp->r_statelock);
979 		if (error || !CACHE_VALID(rp, va.va_mtime, va.va_size) ||
980 		    (mi->mi_flags & MI_ACL)) {
981 			mutex_exit(&rp->r_statelock);
982 			PURGE_ATTRCACHE(vp);
983 		} else {
984 			if (rp->r_mtime <= t) {
985 				nfs_attrcache_va(vp, &va);
986 			}
987 			mutex_exit(&rp->r_statelock);
988 		}
989 	}
990 
991 	return (error);
992 }
993 
994 /* ARGSUSED */
995 static int
996 nfs_ioctl(vnode_t *vp, int cmd, intptr_t arg, int flag, cred_t *cr, int *rvalp)
997 {
998 
999 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1000 		return (EIO);
1001 	switch (cmd) {
1002 		case _FIODIRECTIO:
1003 			return (nfs_directio(vp, (int)arg, cr));
1004 		default:
1005 			return (ENOTTY);
1006 	}
1007 }
1008 
1009 static int
1010 nfs_getattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1011 {
1012 	int error;
1013 	rnode_t *rp;
1014 
1015 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1016 		return (EIO);
1017 	/*
1018 	 * If it has been specified that the return value will
1019 	 * just be used as a hint, and we are only being asked
1020 	 * for size, fsid or rdevid, then return the client's
1021 	 * notion of these values without checking to make sure
1022 	 * that the attribute cache is up to date.
1023 	 * The whole point is to avoid an over the wire GETATTR
1024 	 * call.
1025 	 */
1026 	rp = VTOR(vp);
1027 	if (flags & ATTR_HINT) {
1028 		if (vap->va_mask ==
1029 		    (vap->va_mask & (AT_SIZE | AT_FSID | AT_RDEV))) {
1030 			mutex_enter(&rp->r_statelock);
1031 			if (vap->va_mask | AT_SIZE)
1032 				vap->va_size = rp->r_size;
1033 			if (vap->va_mask | AT_FSID)
1034 				vap->va_fsid = rp->r_attr.va_fsid;
1035 			if (vap->va_mask | AT_RDEV)
1036 				vap->va_rdev = rp->r_attr.va_rdev;
1037 			mutex_exit(&rp->r_statelock);
1038 			return (0);
1039 		}
1040 	}
1041 
1042 	/*
1043 	 * Only need to flush pages if asking for the mtime
1044 	 * and if there any dirty pages or any outstanding
1045 	 * asynchronous (write) requests for this file.
1046 	 */
1047 	if (vap->va_mask & AT_MTIME) {
1048 		if (vn_has_cached_data(vp) &&
1049 		    ((rp->r_flags & RDIRTY) || rp->r_awcount > 0)) {
1050 			mutex_enter(&rp->r_statelock);
1051 			rp->r_gcount++;
1052 			mutex_exit(&rp->r_statelock);
1053 			error = nfs_putpage(vp, (offset_t)0, 0, 0, cr);
1054 			mutex_enter(&rp->r_statelock);
1055 			if (error && (error == ENOSPC || error == EDQUOT)) {
1056 				if (!rp->r_error)
1057 					rp->r_error = error;
1058 			}
1059 			if (--rp->r_gcount == 0)
1060 				cv_broadcast(&rp->r_cv);
1061 			mutex_exit(&rp->r_statelock);
1062 		}
1063 	}
1064 
1065 	return (nfsgetattr(vp, vap, cr));
1066 }
1067 
1068 /*ARGSUSED4*/
1069 static int
1070 nfs_setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr,
1071 		caller_context_t *ct)
1072 {
1073 	int error;
1074 	uint_t mask;
1075 	struct vattr va;
1076 
1077 	mask = vap->va_mask;
1078 
1079 	if (mask & AT_NOSET)
1080 		return (EINVAL);
1081 
1082 	if ((mask & AT_SIZE) &&
1083 	    vap->va_type == VREG &&
1084 	    vap->va_size > MAXOFF32_T)
1085 		return (EFBIG);
1086 
1087 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1088 		return (EIO);
1089 
1090 	va.va_mask = AT_UID | AT_MODE;
1091 
1092 	error = nfsgetattr(vp, &va, cr);
1093 	if (error)
1094 		return (error);
1095 
1096 	error = secpolicy_vnode_setattr(cr, vp, vap, &va, flags, nfs_accessx,
1097 			vp);
1098 
1099 	if (error)
1100 		return (error);
1101 
1102 	return (nfssetattr(vp, vap, flags, cr));
1103 }
1104 
1105 static int
1106 nfssetattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1107 {
1108 	int error;
1109 	uint_t mask;
1110 	struct nfssaargs args;
1111 	struct nfsattrstat ns;
1112 	int douprintf;
1113 	rnode_t *rp;
1114 	struct vattr va;
1115 	mode_t omode;
1116 	mntinfo_t *mi;
1117 	vsecattr_t *vsp;
1118 	hrtime_t t;
1119 
1120 	mask = vap->va_mask;
1121 
1122 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
1123 
1124 	rp = VTOR(vp);
1125 
1126 	/*
1127 	 * Only need to flush pages if there are any pages and
1128 	 * if the file is marked as dirty in some fashion.  The
1129 	 * file must be flushed so that we can accurately
1130 	 * determine the size of the file and the cached data
1131 	 * after the SETATTR returns.  A file is considered to
1132 	 * be dirty if it is either marked with RDIRTY, has
1133 	 * outstanding i/o's active, or is mmap'd.  In this
1134 	 * last case, we can't tell whether there are dirty
1135 	 * pages, so we flush just to be sure.
1136 	 */
1137 	if (vn_has_cached_data(vp) &&
1138 	    ((rp->r_flags & RDIRTY) ||
1139 	    rp->r_count > 0 ||
1140 	    rp->r_mapcnt > 0)) {
1141 		ASSERT(vp->v_type != VCHR);
1142 		error = nfs_putpage(vp, (offset_t)0, 0, 0, cr);
1143 		if (error && (error == ENOSPC || error == EDQUOT)) {
1144 			mutex_enter(&rp->r_statelock);
1145 			if (!rp->r_error)
1146 				rp->r_error = error;
1147 			mutex_exit(&rp->r_statelock);
1148 		}
1149 	}
1150 
1151 	/*
1152 	 * If the system call was utime(2) or utimes(2) and the
1153 	 * application did not specify the times, then set the
1154 	 * mtime nanosecond field to 1 billion.  This will get
1155 	 * translated from 1 billion nanoseconds to 1 million
1156 	 * microseconds in the over the wire request.  The
1157 	 * server will use 1 million in the microsecond field
1158 	 * to tell whether both the mtime and atime should be
1159 	 * set to the server's current time.
1160 	 *
1161 	 * This is an overload of the protocol and should be
1162 	 * documented in the NFS Version 2 protocol specification.
1163 	 */
1164 	if ((mask & AT_MTIME) && !(flags & ATTR_UTIME)) {
1165 		vap->va_mtime.tv_nsec = 1000000000;
1166 		if (NFS_TIME_T_OK(vap->va_mtime.tv_sec) &&
1167 		    NFS_TIME_T_OK(vap->va_atime.tv_sec)) {
1168 			error = vattr_to_sattr(vap, &args.saa_sa);
1169 		} else {
1170 			/*
1171 			 * Use server times. vap time values will not be used.
1172 			 * To ensure no time overflow, make sure vap has
1173 			 * valid values, but retain the original values.
1174 			 */
1175 			timestruc_t	mtime = vap->va_mtime;
1176 			timestruc_t	atime = vap->va_atime;
1177 			time_t		now;
1178 
1179 			now = gethrestime_sec();
1180 			if (NFS_TIME_T_OK(now)) {
1181 				/* Just in case server does not know of this */
1182 				vap->va_mtime.tv_sec = now;
1183 				vap->va_atime.tv_sec = now;
1184 			} else {
1185 				vap->va_mtime.tv_sec = 0;
1186 				vap->va_atime.tv_sec = 0;
1187 			}
1188 			error = vattr_to_sattr(vap, &args.saa_sa);
1189 			/* set vap times back on */
1190 			vap->va_mtime = mtime;
1191 			vap->va_atime = atime;
1192 		}
1193 	} else {
1194 		/* Either do not set times or use the client specified times */
1195 		error = vattr_to_sattr(vap, &args.saa_sa);
1196 	}
1197 	if (error) {
1198 		/* req time field(s) overflow - return immediately */
1199 		return (error);
1200 	}
1201 	args.saa_fh = *VTOFH(vp);
1202 
1203 	va.va_mask = AT_MODE;
1204 	error = nfsgetattr(vp, &va, cr);
1205 	if (error)
1206 		return (error);
1207 	omode = va.va_mode;
1208 
1209 	mi = VTOMI(vp);
1210 
1211 	douprintf = 1;
1212 
1213 	t = gethrtime();
1214 
1215 	error = rfs2call(mi, RFS_SETATTR,
1216 	    xdr_saargs, (caddr_t)&args,
1217 	    xdr_attrstat, (caddr_t)&ns, cr,
1218 	    &douprintf, &ns.ns_status, 0, NULL);
1219 
1220 	/*
1221 	 * Purge the access cache and ACL cache if changing either the
1222 	 * owner of the file, the group owner, or the mode.  These may
1223 	 * change the access permissions of the file, so purge old
1224 	 * information and start over again.
1225 	 */
1226 	if ((mask & (AT_UID | AT_GID | AT_MODE)) && (mi->mi_flags & MI_ACL)) {
1227 		(void) nfs_access_purge_rp(rp);
1228 		if (rp->r_secattr != NULL) {
1229 			mutex_enter(&rp->r_statelock);
1230 			vsp = rp->r_secattr;
1231 			rp->r_secattr = NULL;
1232 			mutex_exit(&rp->r_statelock);
1233 			if (vsp != NULL)
1234 				nfs_acl_free(vsp);
1235 		}
1236 	}
1237 
1238 	if (!error) {
1239 		error = geterrno(ns.ns_status);
1240 		if (!error) {
1241 			/*
1242 			 * If changing the size of the file, invalidate
1243 			 * any local cached data which is no longer part
1244 			 * of the file.  We also possibly invalidate the
1245 			 * last page in the file.  We could use
1246 			 * pvn_vpzero(), but this would mark the page as
1247 			 * modified and require it to be written back to
1248 			 * the server for no particularly good reason.
1249 			 * This way, if we access it, then we bring it
1250 			 * back in.  A read should be cheaper than a
1251 			 * write.
1252 			 */
1253 			if (mask & AT_SIZE) {
1254 				nfs_invalidate_pages(vp,
1255 				    (vap->va_size & PAGEMASK), cr);
1256 			}
1257 			(void) nfs_cache_fattr(vp, &ns.ns_attr, &va, t, cr);
1258 			/*
1259 			 * If NFS_ACL is supported on the server, then the
1260 			 * attributes returned by server may have minimal
1261 			 * permissions sometimes denying access to users having
1262 			 * proper access.  To get the proper attributes, mark
1263 			 * the attributes as expired so that they will be
1264 			 * regotten via the NFS_ACL GETATTR2 procedure.
1265 			 */
1266 			if (mi->mi_flags & MI_ACL) {
1267 				PURGE_ATTRCACHE(vp);
1268 			}
1269 			/*
1270 			 * This next check attempts to deal with NFS
1271 			 * servers which can not handle increasing
1272 			 * the size of the file via setattr.  Most
1273 			 * of these servers do not return an error,
1274 			 * but do not change the size of the file.
1275 			 * Hence, this check and then attempt to set
1276 			 * the file size by writing 1 byte at the
1277 			 * offset of the end of the file that we need.
1278 			 */
1279 			if ((mask & AT_SIZE) &&
1280 			    ns.ns_attr.na_size < (uint32_t)vap->va_size) {
1281 				char zb = '\0';
1282 
1283 				error = nfswrite(vp, &zb,
1284 				    vap->va_size - sizeof (zb),
1285 				    sizeof (zb), cr);
1286 			}
1287 			/*
1288 			 * Some servers will change the mode to clear the setuid
1289 			 * and setgid bits when changing the uid or gid.  The
1290 			 * client needs to compensate appropriately.
1291 			 */
1292 			if (mask & (AT_UID | AT_GID)) {
1293 				int terror;
1294 
1295 				va.va_mask = AT_MODE;
1296 				terror = nfsgetattr(vp, &va, cr);
1297 				if (!terror &&
1298 				    (((mask & AT_MODE) &&
1299 				    va.va_mode != vap->va_mode) ||
1300 				    (!(mask & AT_MODE) &&
1301 				    va.va_mode != omode))) {
1302 					va.va_mask = AT_MODE;
1303 					if (mask & AT_MODE)
1304 						va.va_mode = vap->va_mode;
1305 					else
1306 						va.va_mode = omode;
1307 					(void) nfssetattr(vp, &va, 0, cr);
1308 				}
1309 			}
1310 		} else {
1311 			PURGE_ATTRCACHE(vp);
1312 			PURGE_STALE_FH(error, vp, cr);
1313 		}
1314 	} else {
1315 		PURGE_ATTRCACHE(vp);
1316 	}
1317 
1318 	return (error);
1319 }
1320 
1321 static int
1322 nfs_accessx(void *vp, int mode, cred_t *cr)
1323 {
1324 	ASSERT(nfs_zone() == VTOMI((vnode_t *)vp)->mi_zone);
1325 	return (nfs_access(vp, mode, 0, cr));
1326 }
1327 
1328 static int
1329 nfs_access(vnode_t *vp, int mode, int flags, cred_t *cr)
1330 {
1331 	struct vattr va;
1332 	int error;
1333 	mntinfo_t *mi;
1334 	int shift = 0;
1335 
1336 	mi = VTOMI(vp);
1337 
1338 	if (nfs_zone() != mi->mi_zone)
1339 		return (EIO);
1340 	if (mi->mi_flags & MI_ACL) {
1341 		error = acl_access2(vp, mode, flags, cr);
1342 		if (mi->mi_flags & MI_ACL)
1343 			return (error);
1344 	}
1345 
1346 	va.va_mask = AT_MODE | AT_UID | AT_GID;
1347 	error = nfsgetattr(vp, &va, cr);
1348 	if (error)
1349 		return (error);
1350 
1351 	/*
1352 	 * Disallow write attempts on read-only
1353 	 * file systems, unless the file is a
1354 	 * device node.
1355 	 */
1356 	if ((mode & VWRITE) && vn_is_readonly(vp) && !IS_DEVVP(vp))
1357 		return (EROFS);
1358 
1359 	/*
1360 	 * Disallow attempts to access mandatory lock files.
1361 	 */
1362 	if ((mode & (VWRITE | VREAD | VEXEC)) &&
1363 	    MANDLOCK(vp, va.va_mode))
1364 		return (EACCES);
1365 
1366 	/*
1367 	 * Access check is based on only
1368 	 * one of owner, group, public.
1369 	 * If not owner, then check group.
1370 	 * If not a member of the group,
1371 	 * then check public access.
1372 	 */
1373 	if (crgetuid(cr) != va.va_uid) {
1374 		shift += 3;
1375 		if (!groupmember(va.va_gid, cr))
1376 			shift += 3;
1377 	}
1378 found:
1379 	mode &= ~(va.va_mode << shift);
1380 	if (mode == 0)
1381 		return (0);
1382 
1383 	return (secpolicy_vnode_access(cr, vp, va.va_uid, mode));
1384 }
1385 
1386 static int nfs_do_symlink_cache = 1;
1387 
1388 static int
1389 nfs_readlink(vnode_t *vp, struct uio *uiop, cred_t *cr)
1390 {
1391 	int error;
1392 	struct nfsrdlnres rl;
1393 	rnode_t *rp;
1394 	int douprintf;
1395 	failinfo_t fi;
1396 
1397 	/*
1398 	 * We want to be consistent with UFS semantics so we will return
1399 	 * EINVAL instead of ENXIO. This violates the XNFS spec and
1400 	 * the RFC 1094, which are wrong any way. BUGID 1138002.
1401 	 */
1402 	if (vp->v_type != VLNK)
1403 		return (EINVAL);
1404 
1405 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1406 		return (EIO);
1407 
1408 	rp = VTOR(vp);
1409 	if (nfs_do_symlink_cache && rp->r_symlink.contents != NULL) {
1410 		error = nfs_validate_caches(vp, cr);
1411 		if (error)
1412 			return (error);
1413 		mutex_enter(&rp->r_statelock);
1414 		if (rp->r_symlink.contents != NULL) {
1415 			error = uiomove(rp->r_symlink.contents,
1416 			    rp->r_symlink.len, UIO_READ, uiop);
1417 			mutex_exit(&rp->r_statelock);
1418 			return (error);
1419 		}
1420 		mutex_exit(&rp->r_statelock);
1421 	}
1422 
1423 
1424 	rl.rl_data = kmem_alloc(NFS_MAXPATHLEN, KM_SLEEP);
1425 
1426 	fi.vp = vp;
1427 	fi.fhp = NULL;		/* no need to update, filehandle not copied */
1428 	fi.copyproc = nfscopyfh;
1429 	fi.lookupproc = nfslookup;
1430 	fi.xattrdirproc = acl_getxattrdir2;
1431 
1432 	douprintf = 1;
1433 
1434 	error = rfs2call(VTOMI(vp), RFS_READLINK,
1435 	    xdr_fhandle, (caddr_t)VTOFH(vp),
1436 	    xdr_rdlnres, (caddr_t)&rl, cr,
1437 	    &douprintf, &rl.rl_status, 0, &fi);
1438 
1439 	if (error) {
1440 
1441 		kmem_free((void *)rl.rl_data, NFS_MAXPATHLEN);
1442 		return (error);
1443 	}
1444 
1445 	error = geterrno(rl.rl_status);
1446 	if (!error) {
1447 		error = uiomove(rl.rl_data, (int)rl.rl_count, UIO_READ, uiop);
1448 		if (nfs_do_symlink_cache && rp->r_symlink.contents == NULL) {
1449 			mutex_enter(&rp->r_statelock);
1450 			if (rp->r_symlink.contents == NULL) {
1451 				rp->r_symlink.contents = rl.rl_data;
1452 				rp->r_symlink.len = (int)rl.rl_count;
1453 				rp->r_symlink.size = NFS_MAXPATHLEN;
1454 				mutex_exit(&rp->r_statelock);
1455 			} else {
1456 				mutex_exit(&rp->r_statelock);
1457 
1458 				kmem_free((void *)rl.rl_data,
1459 				    NFS_MAXPATHLEN);
1460 			}
1461 		} else {
1462 
1463 			kmem_free((void *)rl.rl_data, NFS_MAXPATHLEN);
1464 		}
1465 	} else {
1466 		PURGE_STALE_FH(error, vp, cr);
1467 
1468 		kmem_free((void *)rl.rl_data, NFS_MAXPATHLEN);
1469 	}
1470 
1471 	/*
1472 	 * Conform to UFS semantics (see comment above)
1473 	 */
1474 	return (error == ENXIO ? EINVAL : error);
1475 }
1476 
1477 /*
1478  * Flush local dirty pages to stable storage on the server.
1479  *
1480  * If FNODSYNC is specified, then there is nothing to do because
1481  * metadata changes are not cached on the client before being
1482  * sent to the server.
1483  */
1484 static int
1485 nfs_fsync(vnode_t *vp, int syncflag, cred_t *cr)
1486 {
1487 	int error;
1488 
1489 	if ((syncflag & FNODSYNC) || IS_SWAPVP(vp))
1490 		return (0);
1491 
1492 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1493 		return (EIO);
1494 
1495 	error = nfs_putpage(vp, (offset_t)0, 0, 0, cr);
1496 	if (!error)
1497 		error = VTOR(vp)->r_error;
1498 	return (error);
1499 }
1500 
1501 
1502 /*
1503  * Weirdness: if the file was removed or the target of a rename
1504  * operation while it was open, it got renamed instead.  Here we
1505  * remove the renamed file.
1506  */
1507 static void
1508 nfs_inactive(vnode_t *vp, cred_t *cr)
1509 {
1510 	rnode_t *rp;
1511 
1512 	ASSERT(vp != DNLC_NO_VNODE);
1513 
1514 	/*
1515 	 * If this is coming from the wrong zone, we let someone in the right
1516 	 * zone take care of it asynchronously.  We can get here due to
1517 	 * VN_RELE() being called from pageout() or fsflush().  This call may
1518 	 * potentially turn into an expensive no-op if, for instance, v_count
1519 	 * gets incremented in the meantime, but it's still correct.
1520 	 */
1521 	if (nfs_zone() != VTOMI(vp)->mi_zone) {
1522 		nfs_async_inactive(vp, cr, nfs_inactive);
1523 		return;
1524 	}
1525 
1526 	rp = VTOR(vp);
1527 redo:
1528 	if (rp->r_unldvp != NULL) {
1529 		/*
1530 		 * Save the vnode pointer for the directory where the
1531 		 * unlinked-open file got renamed, then set it to NULL
1532 		 * to prevent another thread from getting here before
1533 		 * we're done with the remove.  While we have the
1534 		 * statelock, make local copies of the pertinent rnode
1535 		 * fields.  If we weren't to do this in an atomic way, the
1536 		 * the unl* fields could become inconsistent with respect
1537 		 * to each other due to a race condition between this
1538 		 * code and nfs_remove().  See bug report 1034328.
1539 		 */
1540 		mutex_enter(&rp->r_statelock);
1541 		if (rp->r_unldvp != NULL) {
1542 			vnode_t *unldvp;
1543 			char *unlname;
1544 			cred_t *unlcred;
1545 			struct nfsdiropargs da;
1546 			enum nfsstat status;
1547 			int douprintf;
1548 			int error;
1549 
1550 			unldvp = rp->r_unldvp;
1551 			rp->r_unldvp = NULL;
1552 			unlname = rp->r_unlname;
1553 			rp->r_unlname = NULL;
1554 			unlcred = rp->r_unlcred;
1555 			rp->r_unlcred = NULL;
1556 			mutex_exit(&rp->r_statelock);
1557 
1558 			/*
1559 			 * If there are any dirty pages left, then flush
1560 			 * them.  This is unfortunate because they just
1561 			 * may get thrown away during the remove operation,
1562 			 * but we have to do this for correctness.
1563 			 */
1564 			if (vn_has_cached_data(vp) &&
1565 			    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
1566 				ASSERT(vp->v_type != VCHR);
1567 				error = nfs_putpage(vp, (offset_t)0, 0, 0, cr);
1568 				if (error) {
1569 					mutex_enter(&rp->r_statelock);
1570 					if (!rp->r_error)
1571 						rp->r_error = error;
1572 					mutex_exit(&rp->r_statelock);
1573 				}
1574 			}
1575 
1576 			/*
1577 			 * Do the remove operation on the renamed file
1578 			 */
1579 			setdiropargs(&da, unlname, unldvp);
1580 
1581 			douprintf = 1;
1582 
1583 			(void) rfs2call(VTOMI(unldvp), RFS_REMOVE,
1584 			    xdr_diropargs, (caddr_t)&da,
1585 			    xdr_enum, (caddr_t)&status, unlcred,
1586 			    &douprintf, &status, 0, NULL);
1587 
1588 			if (HAVE_RDDIR_CACHE(VTOR(unldvp)))
1589 				nfs_purge_rddir_cache(unldvp);
1590 			PURGE_ATTRCACHE(unldvp);
1591 
1592 			/*
1593 			 * Release stuff held for the remove
1594 			 */
1595 			VN_RELE(unldvp);
1596 			kmem_free(unlname, MAXNAMELEN);
1597 			crfree(unlcred);
1598 			goto redo;
1599 		}
1600 		mutex_exit(&rp->r_statelock);
1601 	}
1602 
1603 	rp_addfree(rp, cr);
1604 }
1605 
1606 /*
1607  * Remote file system operations having to do with directory manipulation.
1608  */
1609 
1610 static int
1611 nfs_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1612 	int flags, vnode_t *rdir, cred_t *cr)
1613 {
1614 	int error;
1615 	vnode_t *vp;
1616 	vnode_t *avp = NULL;
1617 	rnode_t *drp;
1618 
1619 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
1620 		return (EPERM);
1621 
1622 	drp = VTOR(dvp);
1623 
1624 	/*
1625 	 * Are we looking up extended attributes?  If so, "dvp" is
1626 	 * the file or directory for which we want attributes, and
1627 	 * we need a lookup of the hidden attribute directory
1628 	 * before we lookup the rest of the path.
1629 	 */
1630 	if (flags & LOOKUP_XATTR) {
1631 		bool_t cflag = ((flags & CREATE_XATTR_DIR) != 0);
1632 		mntinfo_t *mi;
1633 
1634 		mi = VTOMI(dvp);
1635 		if (!(mi->mi_flags & MI_EXTATTR))
1636 			return (EINVAL);
1637 
1638 		if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp)))
1639 			return (EINTR);
1640 
1641 		(void) nfslookup_dnlc(dvp, XATTR_DIR_NAME, &avp, cr);
1642 		if (avp == NULL)
1643 			error = acl_getxattrdir2(dvp, &avp, cflag, cr, 0);
1644 		else
1645 			error = 0;
1646 
1647 		nfs_rw_exit(&drp->r_rwlock);
1648 
1649 		if (error) {
1650 			if (mi->mi_flags & MI_EXTATTR)
1651 				return (error);
1652 			return (EINVAL);
1653 		}
1654 		dvp = avp;
1655 		drp = VTOR(dvp);
1656 	}
1657 
1658 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp))) {
1659 		error = EINTR;
1660 		goto out;
1661 	}
1662 
1663 	error = nfslookup(dvp, nm, vpp, pnp, flags, rdir, cr, 0);
1664 
1665 	nfs_rw_exit(&drp->r_rwlock);
1666 
1667 	/*
1668 	 * If vnode is a device, create special vnode.
1669 	 */
1670 	if (!error && IS_DEVVP(*vpp)) {
1671 		vp = *vpp;
1672 		*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
1673 		VN_RELE(vp);
1674 	}
1675 
1676 out:
1677 	if (avp != NULL)
1678 		VN_RELE(avp);
1679 
1680 	return (error);
1681 }
1682 
1683 static int nfs_lookup_neg_cache = 1;
1684 
1685 #ifdef DEBUG
1686 static int nfs_lookup_dnlc_hits = 0;
1687 static int nfs_lookup_dnlc_misses = 0;
1688 static int nfs_lookup_dnlc_neg_hits = 0;
1689 static int nfs_lookup_dnlc_disappears = 0;
1690 static int nfs_lookup_dnlc_lookups = 0;
1691 #endif
1692 
1693 /* ARGSUSED */
1694 int
1695 nfslookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1696 	int flags, vnode_t *rdir, cred_t *cr, int rfscall_flags)
1697 {
1698 	int error;
1699 
1700 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1701 
1702 	/*
1703 	 * If lookup is for "", just return dvp.  Don't need
1704 	 * to send it over the wire, look it up in the dnlc,
1705 	 * or perform any access checks.
1706 	 */
1707 	if (*nm == '\0') {
1708 		VN_HOLD(dvp);
1709 		*vpp = dvp;
1710 		return (0);
1711 	}
1712 
1713 	/*
1714 	 * Can't do lookups in non-directories.
1715 	 */
1716 	if (dvp->v_type != VDIR)
1717 		return (ENOTDIR);
1718 
1719 	/*
1720 	 * If we're called with RFSCALL_SOFT, it's important that
1721 	 * the only rfscall is one we make directly; if we permit
1722 	 * an access call because we're looking up "." or validating
1723 	 * a dnlc hit, we'll deadlock because that rfscall will not
1724 	 * have the RFSCALL_SOFT set.
1725 	 */
1726 	if (rfscall_flags & RFSCALL_SOFT)
1727 		goto callit;
1728 
1729 	/*
1730 	 * If lookup is for ".", just return dvp.  Don't need
1731 	 * to send it over the wire or look it up in the dnlc,
1732 	 * just need to check access.
1733 	 */
1734 	if (strcmp(nm, ".") == 0) {
1735 		error = nfs_access(dvp, VEXEC, 0, cr);
1736 		if (error)
1737 			return (error);
1738 		VN_HOLD(dvp);
1739 		*vpp = dvp;
1740 		return (0);
1741 	}
1742 
1743 	/*
1744 	 * Lookup this name in the DNLC.  If there was a valid entry,
1745 	 * then return the results of the lookup.
1746 	 */
1747 	error = nfslookup_dnlc(dvp, nm, vpp, cr);
1748 	if (error || *vpp != NULL)
1749 		return (error);
1750 
1751 callit:
1752 	error = nfslookup_otw(dvp, nm, vpp, cr, rfscall_flags);
1753 
1754 	return (error);
1755 }
1756 
1757 static int
1758 nfslookup_dnlc(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr)
1759 {
1760 	int error;
1761 	vnode_t *vp;
1762 
1763 	ASSERT(*nm != '\0');
1764 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1765 
1766 	/*
1767 	 * Lookup this name in the DNLC.  If successful, then validate
1768 	 * the caches and then recheck the DNLC.  The DNLC is rechecked
1769 	 * just in case this entry got invalidated during the call
1770 	 * to nfs_validate_caches.
1771 	 *
1772 	 * An assumption is being made that it is safe to say that a
1773 	 * file exists which may not on the server.  Any operations to
1774 	 * the server will fail with ESTALE.
1775 	 */
1776 #ifdef DEBUG
1777 	nfs_lookup_dnlc_lookups++;
1778 #endif
1779 	vp = dnlc_lookup(dvp, nm);
1780 	if (vp != NULL) {
1781 		VN_RELE(vp);
1782 		if (vp == DNLC_NO_VNODE && !vn_is_readonly(dvp)) {
1783 			PURGE_ATTRCACHE(dvp);
1784 		}
1785 		error = nfs_validate_caches(dvp, cr);
1786 		if (error)
1787 			return (error);
1788 		vp = dnlc_lookup(dvp, nm);
1789 		if (vp != NULL) {
1790 			error = nfs_access(dvp, VEXEC, 0, cr);
1791 			if (error) {
1792 				VN_RELE(vp);
1793 				return (error);
1794 			}
1795 			if (vp == DNLC_NO_VNODE) {
1796 				VN_RELE(vp);
1797 #ifdef DEBUG
1798 				nfs_lookup_dnlc_neg_hits++;
1799 #endif
1800 				return (ENOENT);
1801 			}
1802 			*vpp = vp;
1803 #ifdef DEBUG
1804 			nfs_lookup_dnlc_hits++;
1805 #endif
1806 			return (0);
1807 		}
1808 #ifdef DEBUG
1809 		nfs_lookup_dnlc_disappears++;
1810 #endif
1811 	}
1812 #ifdef DEBUG
1813 	else
1814 		nfs_lookup_dnlc_misses++;
1815 #endif
1816 
1817 	*vpp = NULL;
1818 
1819 	return (0);
1820 }
1821 
1822 static int
1823 nfslookup_otw(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr,
1824 	int rfscall_flags)
1825 {
1826 	int error;
1827 	struct nfsdiropargs da;
1828 	struct nfsdiropres dr;
1829 	int douprintf;
1830 	failinfo_t fi;
1831 	hrtime_t t;
1832 
1833 	ASSERT(*nm != '\0');
1834 	ASSERT(dvp->v_type == VDIR);
1835 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1836 
1837 	setdiropargs(&da, nm, dvp);
1838 
1839 	fi.vp = dvp;
1840 	fi.fhp = NULL;		/* no need to update, filehandle not copied */
1841 	fi.copyproc = nfscopyfh;
1842 	fi.lookupproc = nfslookup;
1843 	fi.xattrdirproc = acl_getxattrdir2;
1844 
1845 	douprintf = 1;
1846 
1847 	t = gethrtime();
1848 
1849 	error = rfs2call(VTOMI(dvp), RFS_LOOKUP,
1850 	    xdr_diropargs, (caddr_t)&da,
1851 	    xdr_diropres, (caddr_t)&dr, cr,
1852 	    &douprintf, &dr.dr_status, rfscall_flags, &fi);
1853 
1854 	if (!error) {
1855 		error = geterrno(dr.dr_status);
1856 		if (!error) {
1857 			*vpp = makenfsnode(&dr.dr_fhandle, &dr.dr_attr,
1858 			    dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
1859 			/*
1860 			 * If NFS_ACL is supported on the server, then the
1861 			 * attributes returned by server may have minimal
1862 			 * permissions sometimes denying access to users having
1863 			 * proper access.  To get the proper attributes, mark
1864 			 * the attributes as expired so that they will be
1865 			 * regotten via the NFS_ACL GETATTR2 procedure.
1866 			 */
1867 			if (VTOMI(*vpp)->mi_flags & MI_ACL) {
1868 				PURGE_ATTRCACHE(*vpp);
1869 			}
1870 			if (!(rfscall_flags & RFSCALL_SOFT))
1871 				dnlc_update(dvp, nm, *vpp);
1872 		} else {
1873 			PURGE_STALE_FH(error, dvp, cr);
1874 			if (error == ENOENT && nfs_lookup_neg_cache)
1875 				dnlc_enter(dvp, nm, DNLC_NO_VNODE);
1876 		}
1877 	}
1878 
1879 	return (error);
1880 }
1881 
1882 /* ARGSUSED */
1883 static int
1884 nfs_create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
1885 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
1886 {
1887 	int error;
1888 	struct nfscreatargs args;
1889 	struct nfsdiropres dr;
1890 	int douprintf;
1891 	vnode_t *vp;
1892 	rnode_t *rp;
1893 	struct vattr vattr;
1894 	rnode_t *drp;
1895 	vnode_t *tempvp;
1896 	hrtime_t t;
1897 
1898 	drp = VTOR(dvp);
1899 
1900 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
1901 		return (EPERM);
1902 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
1903 		return (EINTR);
1904 
1905 	/*
1906 	 * We make a copy of the attributes because the caller does not
1907 	 * expect us to change what va points to.
1908 	 */
1909 	vattr = *va;
1910 
1911 	/*
1912 	 * If the pathname is "", just use dvp.  Don't need
1913 	 * to send it over the wire, look it up in the dnlc,
1914 	 * or perform any access checks.
1915 	 */
1916 	if (*nm == '\0') {
1917 		error = 0;
1918 		VN_HOLD(dvp);
1919 		vp = dvp;
1920 	/*
1921 	 * If the pathname is ".", just use dvp.  Don't need
1922 	 * to send it over the wire or look it up in the dnlc,
1923 	 * just need to check access.
1924 	 */
1925 	} else if (strcmp(nm, ".") == 0) {
1926 		error = nfs_access(dvp, VEXEC, 0, cr);
1927 		if (error) {
1928 			nfs_rw_exit(&drp->r_rwlock);
1929 			return (error);
1930 		}
1931 		VN_HOLD(dvp);
1932 		vp = dvp;
1933 	/*
1934 	 * We need to go over the wire, just to be sure whether the
1935 	 * file exists or not.  Using the DNLC can be dangerous in
1936 	 * this case when making a decision regarding existence.
1937 	 */
1938 	} else {
1939 		error = nfslookup_otw(dvp, nm, &vp, cr, 0);
1940 	}
1941 	if (!error) {
1942 		if (exclusive == EXCL)
1943 			error = EEXIST;
1944 		else if (vp->v_type == VDIR && (mode & VWRITE))
1945 			error = EISDIR;
1946 		else {
1947 			/*
1948 			 * If vnode is a device, create special vnode.
1949 			 */
1950 			if (IS_DEVVP(vp)) {
1951 				tempvp = vp;
1952 				vp = specvp(vp, vp->v_rdev, vp->v_type, cr);
1953 				VN_RELE(tempvp);
1954 			}
1955 			if (!(error = VOP_ACCESS(vp, mode, 0, cr))) {
1956 				if ((vattr.va_mask & AT_SIZE) &&
1957 				    vp->v_type == VREG) {
1958 					vattr.va_mask = AT_SIZE;
1959 					error = nfssetattr(vp, &vattr, 0, cr);
1960 				}
1961 			}
1962 		}
1963 		nfs_rw_exit(&drp->r_rwlock);
1964 		if (error) {
1965 			VN_RELE(vp);
1966 		} else
1967 			*vpp = vp;
1968 		return (error);
1969 	}
1970 
1971 	ASSERT(vattr.va_mask & AT_TYPE);
1972 	if (vattr.va_type == VREG) {
1973 		ASSERT(vattr.va_mask & AT_MODE);
1974 		if (MANDMODE(vattr.va_mode)) {
1975 			nfs_rw_exit(&drp->r_rwlock);
1976 			return (EACCES);
1977 		}
1978 	}
1979 
1980 	dnlc_remove(dvp, nm);
1981 
1982 	setdiropargs(&args.ca_da, nm, dvp);
1983 
1984 	/*
1985 	 * Decide what the group-id of the created file should be.
1986 	 * Set it in attribute list as advisory...then do a setattr
1987 	 * if the server didn't get it right the first time.
1988 	 */
1989 	error = setdirgid(dvp, &vattr.va_gid, cr);
1990 	if (error) {
1991 		nfs_rw_exit(&drp->r_rwlock);
1992 		return (error);
1993 	}
1994 	vattr.va_mask |= AT_GID;
1995 
1996 	/*
1997 	 * This is a completely gross hack to make mknod
1998 	 * work over the wire until we can wack the protocol
1999 	 */
2000 #define	IFCHR		0020000		/* character special */
2001 #define	IFBLK		0060000		/* block special */
2002 #define	IFSOCK		0140000		/* socket */
2003 
2004 	/*
2005 	 * dev_t is uint_t in 5.x and short in 4.x. Both 4.x
2006 	 * supports 8 bit majors. 5.x supports 14 bit majors. 5.x supports 18
2007 	 * bits in the minor number where 4.x supports 8 bits.  If the 5.x
2008 	 * minor/major numbers <= 8 bits long, compress the device
2009 	 * number before sending it. Otherwise, the 4.x server will not
2010 	 * create the device with the correct device number and nothing can be
2011 	 * done about this.
2012 	 */
2013 	if (vattr.va_type == VCHR || vattr.va_type == VBLK) {
2014 		dev_t d = vattr.va_rdev;
2015 		dev32_t dev32;
2016 
2017 		if (vattr.va_type == VCHR)
2018 			vattr.va_mode |= IFCHR;
2019 		else
2020 			vattr.va_mode |= IFBLK;
2021 
2022 		(void) cmpldev(&dev32, d);
2023 		if (dev32 & ~((SO4_MAXMAJ << L_BITSMINOR32) | SO4_MAXMIN))
2024 			vattr.va_size = (u_offset_t)dev32;
2025 		else
2026 			vattr.va_size = (u_offset_t)nfsv2_cmpdev(d);
2027 
2028 		vattr.va_mask |= AT_MODE|AT_SIZE;
2029 	} else if (vattr.va_type == VFIFO) {
2030 		vattr.va_mode |= IFCHR;		/* xtra kludge for namedpipe */
2031 		vattr.va_size = (u_offset_t)NFS_FIFO_DEV;	/* blech */
2032 		vattr.va_mask |= AT_MODE|AT_SIZE;
2033 	} else if (vattr.va_type == VSOCK) {
2034 		vattr.va_mode |= IFSOCK;
2035 		/*
2036 		 * To avoid triggering bugs in the servers set AT_SIZE
2037 		 * (all other RFS_CREATE calls set this).
2038 		 */
2039 		vattr.va_size = 0;
2040 		vattr.va_mask |= AT_MODE|AT_SIZE;
2041 	}
2042 
2043 	args.ca_sa = &args.ca_sa_buf;
2044 	error = vattr_to_sattr(&vattr, args.ca_sa);
2045 	if (error) {
2046 		/* req time field(s) overflow - return immediately */
2047 		nfs_rw_exit(&drp->r_rwlock);
2048 		return (error);
2049 	}
2050 
2051 	douprintf = 1;
2052 
2053 	t = gethrtime();
2054 
2055 	error = rfs2call(VTOMI(dvp), RFS_CREATE,
2056 	    xdr_creatargs, (caddr_t)&args,
2057 	    xdr_diropres, (caddr_t)&dr, cr,
2058 	    &douprintf, &dr.dr_status, 0, NULL);
2059 
2060 	PURGE_ATTRCACHE(dvp);	/* mod time changed */
2061 
2062 	if (!error) {
2063 		error = geterrno(dr.dr_status);
2064 		if (!error) {
2065 			if (HAVE_RDDIR_CACHE(drp))
2066 				nfs_purge_rddir_cache(dvp);
2067 			vp = makenfsnode(&dr.dr_fhandle, &dr.dr_attr,
2068 			    dvp->v_vfsp, t, cr, NULL, NULL);
2069 			/*
2070 			 * If NFS_ACL is supported on the server, then the
2071 			 * attributes returned by server may have minimal
2072 			 * permissions sometimes denying access to users having
2073 			 * proper access.  To get the proper attributes, mark
2074 			 * the attributes as expired so that they will be
2075 			 * regotten via the NFS_ACL GETATTR2 procedure.
2076 			 */
2077 			if (VTOMI(vp)->mi_flags & MI_ACL) {
2078 				PURGE_ATTRCACHE(vp);
2079 			}
2080 			dnlc_update(dvp, nm, vp);
2081 			rp = VTOR(vp);
2082 			if (vattr.va_size == 0) {
2083 				mutex_enter(&rp->r_statelock);
2084 				rp->r_size = 0;
2085 				mutex_exit(&rp->r_statelock);
2086 				if (vn_has_cached_data(vp)) {
2087 					ASSERT(vp->v_type != VCHR);
2088 					nfs_invalidate_pages(vp,
2089 					    (u_offset_t)0, cr);
2090 				}
2091 			}
2092 
2093 			/*
2094 			 * Make sure the gid was set correctly.
2095 			 * If not, try to set it (but don't lose
2096 			 * any sleep over it).
2097 			 */
2098 			if (vattr.va_gid != rp->r_attr.va_gid) {
2099 				vattr.va_mask = AT_GID;
2100 				(void) nfssetattr(vp, &vattr, 0, cr);
2101 			}
2102 
2103 			/*
2104 			 * If vnode is a device create special vnode
2105 			 */
2106 			if (IS_DEVVP(vp)) {
2107 				*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2108 				VN_RELE(vp);
2109 			} else
2110 				*vpp = vp;
2111 		} else {
2112 			PURGE_STALE_FH(error, dvp, cr);
2113 		}
2114 	}
2115 
2116 	nfs_rw_exit(&drp->r_rwlock);
2117 
2118 	return (error);
2119 }
2120 
2121 /*
2122  * Weirdness: if the vnode to be removed is open
2123  * we rename it instead of removing it and nfs_inactive
2124  * will remove the new name.
2125  */
2126 static int
2127 nfs_remove(vnode_t *dvp, char *nm, cred_t *cr)
2128 {
2129 	int error;
2130 	struct nfsdiropargs da;
2131 	enum nfsstat status;
2132 	vnode_t *vp;
2133 	char *tmpname;
2134 	int douprintf;
2135 	rnode_t *rp;
2136 	rnode_t *drp;
2137 
2138 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2139 		return (EPERM);
2140 	drp = VTOR(dvp);
2141 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2142 		return (EINTR);
2143 
2144 	error = nfslookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2145 	if (error) {
2146 		nfs_rw_exit(&drp->r_rwlock);
2147 		return (error);
2148 	}
2149 
2150 	if (vp->v_type == VDIR && secpolicy_fs_linkdir(cr, dvp->v_vfsp)) {
2151 		VN_RELE(vp);
2152 		nfs_rw_exit(&drp->r_rwlock);
2153 		return (EPERM);
2154 	}
2155 
2156 	/*
2157 	 * First just remove the entry from the name cache, as it
2158 	 * is most likely the only entry for this vp.
2159 	 */
2160 	dnlc_remove(dvp, nm);
2161 
2162 	/*
2163 	 * If the file has a v_count > 1 then there may be more than one
2164 	 * entry in the name cache due multiple links or an open file,
2165 	 * but we don't have the real reference count so flush all
2166 	 * possible entries.
2167 	 */
2168 	if (vp->v_count > 1)
2169 		dnlc_purge_vp(vp);
2170 
2171 	/*
2172 	 * Now we have the real reference count on the vnode
2173 	 */
2174 	rp = VTOR(vp);
2175 	mutex_enter(&rp->r_statelock);
2176 	if (vp->v_count > 1 &&
2177 	    (rp->r_unldvp == NULL || strcmp(nm, rp->r_unlname) == 0)) {
2178 		mutex_exit(&rp->r_statelock);
2179 		tmpname = newname();
2180 		error = nfsrename(dvp, nm, dvp, tmpname, cr);
2181 		if (error)
2182 			kmem_free(tmpname, MAXNAMELEN);
2183 		else {
2184 			mutex_enter(&rp->r_statelock);
2185 			if (rp->r_unldvp == NULL) {
2186 				VN_HOLD(dvp);
2187 				rp->r_unldvp = dvp;
2188 				if (rp->r_unlcred != NULL)
2189 					crfree(rp->r_unlcred);
2190 				crhold(cr);
2191 				rp->r_unlcred = cr;
2192 				rp->r_unlname = tmpname;
2193 			} else {
2194 				kmem_free(rp->r_unlname, MAXNAMELEN);
2195 				rp->r_unlname = tmpname;
2196 			}
2197 			mutex_exit(&rp->r_statelock);
2198 		}
2199 	} else {
2200 		mutex_exit(&rp->r_statelock);
2201 		/*
2202 		 * We need to flush any dirty pages which happen to
2203 		 * be hanging around before removing the file.  This
2204 		 * shouldn't happen very often and mostly on file
2205 		 * systems mounted "nocto".
2206 		 */
2207 		if (vn_has_cached_data(vp) &&
2208 		    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
2209 			error = nfs_putpage(vp, (offset_t)0, 0, 0, cr);
2210 			if (error && (error == ENOSPC || error == EDQUOT)) {
2211 				mutex_enter(&rp->r_statelock);
2212 				if (!rp->r_error)
2213 					rp->r_error = error;
2214 				mutex_exit(&rp->r_statelock);
2215 			}
2216 		}
2217 
2218 		setdiropargs(&da, nm, dvp);
2219 
2220 		douprintf = 1;
2221 
2222 		error = rfs2call(VTOMI(dvp), RFS_REMOVE,
2223 		    xdr_diropargs, (caddr_t)&da,
2224 		    xdr_enum, (caddr_t)&status, cr,
2225 		    &douprintf, &status, 0, NULL);
2226 
2227 		/*
2228 		 * The xattr dir may be gone after last attr is removed,
2229 		 * so flush it from dnlc.
2230 		 */
2231 		if (dvp->v_flag & V_XATTRDIR)
2232 			dnlc_purge_vp(dvp);
2233 
2234 		PURGE_ATTRCACHE(dvp);	/* mod time changed */
2235 		PURGE_ATTRCACHE(vp);	/* link count changed */
2236 
2237 		if (!error) {
2238 			error = geterrno(status);
2239 			if (!error) {
2240 				if (HAVE_RDDIR_CACHE(drp))
2241 					nfs_purge_rddir_cache(dvp);
2242 			} else {
2243 				PURGE_STALE_FH(error, dvp, cr);
2244 			}
2245 		}
2246 	}
2247 
2248 	VN_RELE(vp);
2249 
2250 	nfs_rw_exit(&drp->r_rwlock);
2251 
2252 	return (error);
2253 }
2254 
2255 static int
2256 nfs_link(vnode_t *tdvp, vnode_t *svp, char *tnm, cred_t *cr)
2257 {
2258 	int error;
2259 	struct nfslinkargs args;
2260 	enum nfsstat status;
2261 	vnode_t *realvp;
2262 	int douprintf;
2263 	rnode_t *tdrp;
2264 
2265 	if (nfs_zone() != VTOMI(tdvp)->mi_zone)
2266 		return (EPERM);
2267 	if (VOP_REALVP(svp, &realvp) == 0)
2268 		svp = realvp;
2269 
2270 	args.la_from = VTOFH(svp);
2271 	setdiropargs(&args.la_to, tnm, tdvp);
2272 
2273 	tdrp = VTOR(tdvp);
2274 	if (nfs_rw_enter_sig(&tdrp->r_rwlock, RW_WRITER, INTR(tdvp)))
2275 		return (EINTR);
2276 
2277 	dnlc_remove(tdvp, tnm);
2278 
2279 	douprintf = 1;
2280 
2281 	error = rfs2call(VTOMI(svp), RFS_LINK,
2282 	    xdr_linkargs, (caddr_t)&args,
2283 	    xdr_enum, (caddr_t)&status, cr,
2284 	    &douprintf, &status, 0, NULL);
2285 
2286 	PURGE_ATTRCACHE(tdvp);	/* mod time changed */
2287 	PURGE_ATTRCACHE(svp);	/* link count changed */
2288 
2289 	if (!error) {
2290 		error = geterrno(status);
2291 		if (!error) {
2292 			if (HAVE_RDDIR_CACHE(tdrp))
2293 				nfs_purge_rddir_cache(tdvp);
2294 		}
2295 	}
2296 
2297 	nfs_rw_exit(&tdrp->r_rwlock);
2298 
2299 	return (error);
2300 }
2301 
2302 static int
2303 nfs_rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2304 {
2305 	vnode_t *realvp;
2306 
2307 	if (nfs_zone() != VTOMI(odvp)->mi_zone)
2308 		return (EPERM);
2309 	if (VOP_REALVP(ndvp, &realvp) == 0)
2310 		ndvp = realvp;
2311 
2312 	return (nfsrename(odvp, onm, ndvp, nnm, cr));
2313 }
2314 
2315 /*
2316  * nfsrename does the real work of renaming in NFS Version 2.
2317  */
2318 static int
2319 nfsrename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2320 {
2321 	int error;
2322 	enum nfsstat status;
2323 	struct nfsrnmargs args;
2324 	int douprintf;
2325 	vnode_t *nvp;
2326 	vnode_t *ovp = NULL;
2327 	char *tmpname;
2328 	rnode_t *rp;
2329 	rnode_t *odrp;
2330 	rnode_t *ndrp;
2331 
2332 	ASSERT(nfs_zone() == VTOMI(odvp)->mi_zone);
2333 	if (strcmp(onm, ".") == 0 || strcmp(onm, "..") == 0 ||
2334 	    strcmp(nnm, ".") == 0 || strcmp(nnm, "..") == 0)
2335 		return (EINVAL);
2336 
2337 	odrp = VTOR(odvp);
2338 	ndrp = VTOR(ndvp);
2339 	if ((intptr_t)odrp < (intptr_t)ndrp) {
2340 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp)))
2341 			return (EINTR);
2342 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp))) {
2343 			nfs_rw_exit(&odrp->r_rwlock);
2344 			return (EINTR);
2345 		}
2346 	} else {
2347 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp)))
2348 			return (EINTR);
2349 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp))) {
2350 			nfs_rw_exit(&ndrp->r_rwlock);
2351 			return (EINTR);
2352 		}
2353 	}
2354 
2355 	/*
2356 	 * Lookup the target file.  If it exists, it needs to be
2357 	 * checked to see whether it is a mount point and whether
2358 	 * it is active (open).
2359 	 */
2360 	error = nfslookup(ndvp, nnm, &nvp, NULL, 0, NULL, cr, 0);
2361 	if (!error) {
2362 		/*
2363 		 * If this file has been mounted on, then just
2364 		 * return busy because renaming to it would remove
2365 		 * the mounted file system from the name space.
2366 		 */
2367 		if (vn_mountedvfs(nvp) != NULL) {
2368 			VN_RELE(nvp);
2369 			nfs_rw_exit(&odrp->r_rwlock);
2370 			nfs_rw_exit(&ndrp->r_rwlock);
2371 			return (EBUSY);
2372 		}
2373 
2374 		/*
2375 		 * Purge the name cache of all references to this vnode
2376 		 * so that we can check the reference count to infer
2377 		 * whether it is active or not.
2378 		 */
2379 		/*
2380 		 * First just remove the entry from the name cache, as it
2381 		 * is most likely the only entry for this vp.
2382 		 */
2383 		dnlc_remove(ndvp, nnm);
2384 		/*
2385 		 * If the file has a v_count > 1 then there may be more
2386 		 * than one entry in the name cache due multiple links
2387 		 * or an open file, but we don't have the real reference
2388 		 * count so flush all possible entries.
2389 		 */
2390 		if (nvp->v_count > 1)
2391 			dnlc_purge_vp(nvp);
2392 
2393 		/*
2394 		 * If the vnode is active and is not a directory,
2395 		 * arrange to rename it to a
2396 		 * temporary file so that it will continue to be
2397 		 * accessible.  This implements the "unlink-open-file"
2398 		 * semantics for the target of a rename operation.
2399 		 * Before doing this though, make sure that the
2400 		 * source and target files are not already the same.
2401 		 */
2402 		if (nvp->v_count > 1 && nvp->v_type != VDIR) {
2403 			/*
2404 			 * Lookup the source name.
2405 			 */
2406 			error = nfslookup(odvp, onm, &ovp, NULL, 0, NULL,
2407 			    cr, 0);
2408 
2409 			/*
2410 			 * The source name *should* already exist.
2411 			 */
2412 			if (error) {
2413 				VN_RELE(nvp);
2414 				nfs_rw_exit(&odrp->r_rwlock);
2415 				nfs_rw_exit(&ndrp->r_rwlock);
2416 				return (error);
2417 			}
2418 
2419 			/*
2420 			 * Compare the two vnodes.  If they are the same,
2421 			 * just release all held vnodes and return success.
2422 			 */
2423 			if (ovp == nvp) {
2424 				VN_RELE(ovp);
2425 				VN_RELE(nvp);
2426 				nfs_rw_exit(&odrp->r_rwlock);
2427 				nfs_rw_exit(&ndrp->r_rwlock);
2428 				return (0);
2429 			}
2430 
2431 			/*
2432 			 * Can't mix and match directories and non-
2433 			 * directories in rename operations.  We already
2434 			 * know that the target is not a directory.  If
2435 			 * the source is a directory, return an error.
2436 			 */
2437 			if (ovp->v_type == VDIR) {
2438 				VN_RELE(ovp);
2439 				VN_RELE(nvp);
2440 				nfs_rw_exit(&odrp->r_rwlock);
2441 				nfs_rw_exit(&ndrp->r_rwlock);
2442 				return (ENOTDIR);
2443 			}
2444 
2445 			/*
2446 			 * The target file exists, is not the same as
2447 			 * the source file, and is active.  Link it
2448 			 * to a temporary filename to avoid having
2449 			 * the server removing the file completely.
2450 			 */
2451 			tmpname = newname();
2452 			error = nfs_link(ndvp, nvp, tmpname, cr);
2453 			if (error == EOPNOTSUPP) {
2454 				error = nfs_rename(ndvp, nnm, ndvp, tmpname,
2455 				    cr);
2456 			}
2457 			if (error) {
2458 				kmem_free(tmpname, MAXNAMELEN);
2459 				VN_RELE(ovp);
2460 				VN_RELE(nvp);
2461 				nfs_rw_exit(&odrp->r_rwlock);
2462 				nfs_rw_exit(&ndrp->r_rwlock);
2463 				return (error);
2464 			}
2465 			rp = VTOR(nvp);
2466 			mutex_enter(&rp->r_statelock);
2467 			if (rp->r_unldvp == NULL) {
2468 				VN_HOLD(ndvp);
2469 				rp->r_unldvp = ndvp;
2470 				if (rp->r_unlcred != NULL)
2471 					crfree(rp->r_unlcred);
2472 				crhold(cr);
2473 				rp->r_unlcred = cr;
2474 				rp->r_unlname = tmpname;
2475 			} else {
2476 				kmem_free(rp->r_unlname, MAXNAMELEN);
2477 				rp->r_unlname = tmpname;
2478 			}
2479 			mutex_exit(&rp->r_statelock);
2480 		}
2481 
2482 		VN_RELE(nvp);
2483 	}
2484 
2485 	if (ovp == NULL) {
2486 		/*
2487 		 * When renaming directories to be a subdirectory of a
2488 		 * different parent, the dnlc entry for ".." will no
2489 		 * longer be valid, so it must be removed.
2490 		 *
2491 		 * We do a lookup here to determine whether we are renaming
2492 		 * a directory and we need to check if we are renaming
2493 		 * an unlinked file.  This might have already been done
2494 		 * in previous code, so we check ovp == NULL to avoid
2495 		 * doing it twice.
2496 		 */
2497 
2498 		error = nfslookup(odvp, onm, &ovp, NULL, 0, NULL, cr, 0);
2499 
2500 		/*
2501 		 * The source name *should* already exist.
2502 		 */
2503 		if (error) {
2504 			nfs_rw_exit(&odrp->r_rwlock);
2505 			nfs_rw_exit(&ndrp->r_rwlock);
2506 			return (error);
2507 		}
2508 		ASSERT(ovp != NULL);
2509 	}
2510 
2511 	dnlc_remove(odvp, onm);
2512 	dnlc_remove(ndvp, nnm);
2513 
2514 	setdiropargs(&args.rna_from, onm, odvp);
2515 	setdiropargs(&args.rna_to, nnm, ndvp);
2516 
2517 	douprintf = 1;
2518 
2519 	error = rfs2call(VTOMI(odvp), RFS_RENAME,
2520 	    xdr_rnmargs, (caddr_t)&args,
2521 	    xdr_enum, (caddr_t)&status, cr,
2522 	    &douprintf, &status, 0, NULL);
2523 
2524 	PURGE_ATTRCACHE(odvp);	/* mod time changed */
2525 	PURGE_ATTRCACHE(ndvp);	/* mod time changed */
2526 
2527 	if (!error) {
2528 		error = geterrno(status);
2529 		if (!error) {
2530 			if (HAVE_RDDIR_CACHE(odrp))
2531 				nfs_purge_rddir_cache(odvp);
2532 			if (HAVE_RDDIR_CACHE(ndrp))
2533 				nfs_purge_rddir_cache(ndvp);
2534 			/*
2535 			 * when renaming directories to be a subdirectory of a
2536 			 * different parent, the dnlc entry for ".." will no
2537 			 * longer be valid, so it must be removed
2538 			 */
2539 			rp = VTOR(ovp);
2540 			if (ndvp != odvp) {
2541 				if (ovp->v_type == VDIR) {
2542 					dnlc_remove(ovp, "..");
2543 					if (HAVE_RDDIR_CACHE(rp))
2544 						nfs_purge_rddir_cache(ovp);
2545 				}
2546 			}
2547 
2548 			/*
2549 			 * If we are renaming the unlinked file, update the
2550 			 * r_unldvp and r_unlname as needed.
2551 			 */
2552 			mutex_enter(&rp->r_statelock);
2553 			if (rp->r_unldvp != NULL) {
2554 				if (strcmp(rp->r_unlname, onm) == 0) {
2555 					(void) strncpy(rp->r_unlname,
2556 						    nnm, MAXNAMELEN);
2557 					rp->r_unlname[MAXNAMELEN - 1] = '\0';
2558 
2559 					if (ndvp != rp->r_unldvp) {
2560 						VN_RELE(rp->r_unldvp);
2561 						rp->r_unldvp = ndvp;
2562 						VN_HOLD(ndvp);
2563 					}
2564 				}
2565 			}
2566 			mutex_exit(&rp->r_statelock);
2567 		} else {
2568 			/*
2569 			 * System V defines rename to return EEXIST, not
2570 			 * ENOTEMPTY if the target directory is not empty.
2571 			 * Over the wire, the error is NFSERR_ENOTEMPTY
2572 			 * which geterrno maps to ENOTEMPTY.
2573 			 */
2574 			if (error == ENOTEMPTY)
2575 				error = EEXIST;
2576 		}
2577 	}
2578 
2579 	VN_RELE(ovp);
2580 
2581 	nfs_rw_exit(&odrp->r_rwlock);
2582 	nfs_rw_exit(&ndrp->r_rwlock);
2583 
2584 	return (error);
2585 }
2586 
2587 static int
2588 nfs_mkdir(vnode_t *dvp, char *nm, struct vattr *va, vnode_t **vpp, cred_t *cr)
2589 {
2590 	int error;
2591 	struct nfscreatargs args;
2592 	struct nfsdiropres dr;
2593 	int douprintf;
2594 	rnode_t *drp;
2595 	hrtime_t t;
2596 
2597 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2598 		return (EPERM);
2599 
2600 	setdiropargs(&args.ca_da, nm, dvp);
2601 
2602 	/*
2603 	 * Decide what the group-id and set-gid bit of the created directory
2604 	 * should be.  May have to do a setattr to get the gid right.
2605 	 */
2606 	error = setdirgid(dvp, &va->va_gid, cr);
2607 	if (error)
2608 		return (error);
2609 	error = setdirmode(dvp, &va->va_mode, cr);
2610 	if (error)
2611 		return (error);
2612 	va->va_mask |= AT_MODE|AT_GID;
2613 
2614 	args.ca_sa = &args.ca_sa_buf;
2615 	error = vattr_to_sattr(va, args.ca_sa);
2616 	if (error) {
2617 		/* req time field(s) overflow - return immediately */
2618 		return (error);
2619 	}
2620 
2621 	drp = VTOR(dvp);
2622 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2623 		return (EINTR);
2624 
2625 	dnlc_remove(dvp, nm);
2626 
2627 	douprintf = 1;
2628 
2629 	t = gethrtime();
2630 
2631 	error = rfs2call(VTOMI(dvp), RFS_MKDIR,
2632 	    xdr_creatargs, (caddr_t)&args,
2633 	    xdr_diropres, (caddr_t)&dr, cr,
2634 	    &douprintf, &dr.dr_status, 0, NULL);
2635 
2636 	PURGE_ATTRCACHE(dvp);	/* mod time changed */
2637 
2638 	if (!error) {
2639 		error = geterrno(dr.dr_status);
2640 		if (!error) {
2641 			if (HAVE_RDDIR_CACHE(drp))
2642 				nfs_purge_rddir_cache(dvp);
2643 			/*
2644 			 * The attributes returned by RFS_MKDIR can not
2645 			 * be depended upon, so mark the attribute cache
2646 			 * as purged.  A subsequent GETATTR will get the
2647 			 * correct attributes from the server.
2648 			 */
2649 			*vpp = makenfsnode(&dr.dr_fhandle, &dr.dr_attr,
2650 			    dvp->v_vfsp, t, cr, NULL, NULL);
2651 			PURGE_ATTRCACHE(*vpp);
2652 			dnlc_update(dvp, nm, *vpp);
2653 
2654 			/*
2655 			 * Make sure the gid was set correctly.
2656 			 * If not, try to set it (but don't lose
2657 			 * any sleep over it).
2658 			 */
2659 			if (va->va_gid != VTOR(*vpp)->r_attr.va_gid) {
2660 				va->va_mask = AT_GID;
2661 				(void) nfssetattr(*vpp, va, 0, cr);
2662 			}
2663 		} else {
2664 			PURGE_STALE_FH(error, dvp, cr);
2665 		}
2666 	}
2667 
2668 	nfs_rw_exit(&drp->r_rwlock);
2669 
2670 	return (error);
2671 }
2672 
2673 static int
2674 nfs_rmdir(vnode_t *dvp, char *nm, vnode_t *cdir, cred_t *cr)
2675 {
2676 	int error;
2677 	enum nfsstat status;
2678 	struct nfsdiropargs da;
2679 	vnode_t *vp;
2680 	int douprintf;
2681 	rnode_t *drp;
2682 
2683 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2684 		return (EPERM);
2685 	drp = VTOR(dvp);
2686 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2687 		return (EINTR);
2688 
2689 	/*
2690 	 * Attempt to prevent a rmdir(".") from succeeding.
2691 	 */
2692 	error = nfslookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2693 	if (error) {
2694 		nfs_rw_exit(&drp->r_rwlock);
2695 		return (error);
2696 	}
2697 
2698 	if (vp == cdir) {
2699 		VN_RELE(vp);
2700 		nfs_rw_exit(&drp->r_rwlock);
2701 		return (EINVAL);
2702 	}
2703 
2704 	setdiropargs(&da, nm, dvp);
2705 
2706 	/*
2707 	 * First just remove the entry from the name cache, as it
2708 	 * is most likely an entry for this vp.
2709 	 */
2710 	dnlc_remove(dvp, nm);
2711 
2712 	/*
2713 	 * If there vnode reference count is greater than one, then
2714 	 * there may be additional references in the DNLC which will
2715 	 * need to be purged.  First, trying removing the entry for
2716 	 * the parent directory and see if that removes the additional
2717 	 * reference(s).  If that doesn't do it, then use dnlc_purge_vp
2718 	 * to completely remove any references to the directory which
2719 	 * might still exist in the DNLC.
2720 	 */
2721 	if (vp->v_count > 1) {
2722 		dnlc_remove(vp, "..");
2723 		if (vp->v_count > 1)
2724 			dnlc_purge_vp(vp);
2725 	}
2726 
2727 	douprintf = 1;
2728 
2729 	error = rfs2call(VTOMI(dvp), RFS_RMDIR,
2730 	    xdr_diropargs, (caddr_t)&da,
2731 	    xdr_enum, (caddr_t)&status, cr,
2732 	    &douprintf, &status, 0, NULL);
2733 
2734 	PURGE_ATTRCACHE(dvp);	/* mod time changed */
2735 
2736 	if (error) {
2737 		VN_RELE(vp);
2738 		nfs_rw_exit(&drp->r_rwlock);
2739 		return (error);
2740 	}
2741 
2742 	error = geterrno(status);
2743 	if (!error) {
2744 		if (HAVE_RDDIR_CACHE(drp))
2745 			nfs_purge_rddir_cache(dvp);
2746 		if (HAVE_RDDIR_CACHE(VTOR(vp)))
2747 			nfs_purge_rddir_cache(vp);
2748 	} else {
2749 		PURGE_STALE_FH(error, dvp, cr);
2750 		/*
2751 		 * System V defines rmdir to return EEXIST, not
2752 		 * ENOTEMPTY if the directory is not empty.  Over
2753 		 * the wire, the error is NFSERR_ENOTEMPTY which
2754 		 * geterrno maps to ENOTEMPTY.
2755 		 */
2756 		if (error == ENOTEMPTY)
2757 			error = EEXIST;
2758 	}
2759 
2760 	VN_RELE(vp);
2761 
2762 	nfs_rw_exit(&drp->r_rwlock);
2763 
2764 	return (error);
2765 }
2766 
2767 static int
2768 nfs_symlink(vnode_t *dvp, char *lnm, struct vattr *tva, char *tnm, cred_t *cr)
2769 {
2770 	int error;
2771 	struct nfsslargs args;
2772 	enum nfsstat status;
2773 	int douprintf;
2774 	rnode_t *drp;
2775 
2776 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2777 		return (EPERM);
2778 	setdiropargs(&args.sla_from, lnm, dvp);
2779 	args.sla_sa = &args.sla_sa_buf;
2780 	error = vattr_to_sattr(tva, args.sla_sa);
2781 	if (error) {
2782 		/* req time field(s) overflow - return immediately */
2783 		return (error);
2784 	}
2785 	args.sla_tnm = tnm;
2786 
2787 	drp = VTOR(dvp);
2788 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2789 		return (EINTR);
2790 
2791 	dnlc_remove(dvp, lnm);
2792 
2793 	douprintf = 1;
2794 
2795 	error = rfs2call(VTOMI(dvp), RFS_SYMLINK,
2796 	    xdr_slargs, (caddr_t)&args,
2797 	    xdr_enum, (caddr_t)&status, cr,
2798 	    &douprintf, &status, 0, NULL);
2799 
2800 	PURGE_ATTRCACHE(dvp);	/* mod time changed */
2801 
2802 	if (!error) {
2803 		error = geterrno(status);
2804 		if (!error) {
2805 			if (HAVE_RDDIR_CACHE(drp))
2806 				nfs_purge_rddir_cache(dvp);
2807 		} else {
2808 			PURGE_STALE_FH(error, dvp, cr);
2809 		}
2810 	}
2811 
2812 	nfs_rw_exit(&drp->r_rwlock);
2813 
2814 	return (error);
2815 }
2816 
2817 #ifdef DEBUG
2818 static int nfs_readdir_cache_hits = 0;
2819 static int nfs_readdir_cache_shorts = 0;
2820 static int nfs_readdir_cache_waits = 0;
2821 static int nfs_readdir_cache_misses = 0;
2822 static int nfs_readdir_readahead = 0;
2823 #endif
2824 
2825 static int nfs_shrinkreaddir = 0;
2826 
2827 /*
2828  * Read directory entries.
2829  * There are some weird things to look out for here.  The uio_offset
2830  * field is either 0 or it is the offset returned from a previous
2831  * readdir.  It is an opaque value used by the server to find the
2832  * correct directory block to read. The count field is the number
2833  * of blocks to read on the server.  This is advisory only, the server
2834  * may return only one block's worth of entries.  Entries may be compressed
2835  * on the server.
2836  */
2837 static int
2838 nfs_readdir(vnode_t *vp, struct uio *uiop, cred_t *cr, int *eofp)
2839 {
2840 	int error;
2841 	size_t count;
2842 	rnode_t *rp;
2843 	rddir_cache *rdc;
2844 	rddir_cache *nrdc;
2845 	rddir_cache *rrdc;
2846 #ifdef DEBUG
2847 	int missed;
2848 #endif
2849 	rddir_cache srdc;
2850 	avl_index_t where;
2851 
2852 	rp = VTOR(vp);
2853 
2854 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
2855 	if (nfs_zone() != VTOMI(vp)->mi_zone)
2856 		return (EIO);
2857 	/*
2858 	 * Make sure that the directory cache is valid.
2859 	 */
2860 	if (HAVE_RDDIR_CACHE(rp)) {
2861 		if (nfs_disable_rddir_cache) {
2862 			/*
2863 			 * Setting nfs_disable_rddir_cache in /etc/system
2864 			 * allows interoperability with servers that do not
2865 			 * properly update the attributes of directories.
2866 			 * Any cached information gets purged before an
2867 			 * access is made to it.
2868 			 */
2869 			nfs_purge_rddir_cache(vp);
2870 		} else {
2871 			error = nfs_validate_caches(vp, cr);
2872 			if (error)
2873 				return (error);
2874 		}
2875 	}
2876 
2877 	/*
2878 	 * UGLINESS: SunOS 3.2 servers apparently cannot always handle an
2879 	 * RFS_READDIR request with rda_count set to more than 0x400. So
2880 	 * we reduce the request size here purely for compatibility.
2881 	 *
2882 	 * In general, this is no longer required.  However, if a server
2883 	 * is discovered which can not handle requests larger than 1024,
2884 	 * nfs_shrinkreaddir can be set to 1 to enable this backwards
2885 	 * compatibility.
2886 	 *
2887 	 * In any case, the request size is limited to NFS_MAXDATA bytes.
2888 	 */
2889 	count = MIN(uiop->uio_iov->iov_len,
2890 	    nfs_shrinkreaddir ? 0x400 : NFS_MAXDATA);
2891 
2892 	nrdc = NULL;
2893 #ifdef DEBUG
2894 	missed = 0;
2895 #endif
2896 top:
2897 	/*
2898 	 * Short circuit last readdir which always returns 0 bytes.
2899 	 * This can be done after the directory has been read through
2900 	 * completely at least once.  This will set r_direof which
2901 	 * can be used to find the value of the last cookie.
2902 	 */
2903 	mutex_enter(&rp->r_statelock);
2904 	if (rp->r_direof != NULL &&
2905 	    uiop->uio_offset == rp->r_direof->nfs_ncookie) {
2906 		mutex_exit(&rp->r_statelock);
2907 #ifdef DEBUG
2908 		nfs_readdir_cache_shorts++;
2909 #endif
2910 		if (eofp)
2911 			*eofp = 1;
2912 		if (nrdc != NULL)
2913 			rddir_cache_rele(nrdc);
2914 		return (0);
2915 	}
2916 	/*
2917 	 * Look for a cache entry.  Cache entries are identified
2918 	 * by the NFS cookie value and the byte count requested.
2919 	 */
2920 	srdc.nfs_cookie = uiop->uio_offset;
2921 	srdc.buflen = count;
2922 	rdc = avl_find(&rp->r_dir, &srdc, &where);
2923 	if (rdc != NULL) {
2924 		rddir_cache_hold(rdc);
2925 		/*
2926 		 * If the cache entry is in the process of being
2927 		 * filled in, wait until this completes.  The
2928 		 * RDDIRWAIT bit is set to indicate that someone
2929 		 * is waiting and then the thread currently
2930 		 * filling the entry is done, it should do a
2931 		 * cv_broadcast to wakeup all of the threads
2932 		 * waiting for it to finish.
2933 		 */
2934 		if (rdc->flags & RDDIR) {
2935 			nfs_rw_exit(&rp->r_rwlock);
2936 			rdc->flags |= RDDIRWAIT;
2937 #ifdef DEBUG
2938 			nfs_readdir_cache_waits++;
2939 #endif
2940 			if (!cv_wait_sig(&rdc->cv, &rp->r_statelock)) {
2941 				/*
2942 				 * We got interrupted, probably
2943 				 * the user typed ^C or an alarm
2944 				 * fired.  We free the new entry
2945 				 * if we allocated one.
2946 				 */
2947 				mutex_exit(&rp->r_statelock);
2948 				(void) nfs_rw_enter_sig(&rp->r_rwlock,
2949 					RW_READER, FALSE);
2950 				rddir_cache_rele(rdc);
2951 				if (nrdc != NULL)
2952 					rddir_cache_rele(nrdc);
2953 				return (EINTR);
2954 			}
2955 			mutex_exit(&rp->r_statelock);
2956 			(void) nfs_rw_enter_sig(&rp->r_rwlock,
2957 				RW_READER, FALSE);
2958 			rddir_cache_rele(rdc);
2959 			goto top;
2960 		}
2961 		/*
2962 		 * Check to see if a readdir is required to
2963 		 * fill the entry.  If so, mark this entry
2964 		 * as being filled, remove our reference,
2965 		 * and branch to the code to fill the entry.
2966 		 */
2967 		if (rdc->flags & RDDIRREQ) {
2968 			rdc->flags &= ~RDDIRREQ;
2969 			rdc->flags |= RDDIR;
2970 			if (nrdc != NULL)
2971 				rddir_cache_rele(nrdc);
2972 			nrdc = rdc;
2973 			mutex_exit(&rp->r_statelock);
2974 			goto bottom;
2975 		}
2976 #ifdef DEBUG
2977 		if (!missed)
2978 			nfs_readdir_cache_hits++;
2979 #endif
2980 		/*
2981 		 * If an error occurred while attempting
2982 		 * to fill the cache entry, just return it.
2983 		 */
2984 		if (rdc->error) {
2985 			error = rdc->error;
2986 			mutex_exit(&rp->r_statelock);
2987 			rddir_cache_rele(rdc);
2988 			if (nrdc != NULL)
2989 				rddir_cache_rele(nrdc);
2990 			return (error);
2991 		}
2992 
2993 		/*
2994 		 * The cache entry is complete and good,
2995 		 * copyout the dirent structs to the calling
2996 		 * thread.
2997 		 */
2998 		error = uiomove(rdc->entries, rdc->entlen, UIO_READ, uiop);
2999 
3000 		/*
3001 		 * If no error occurred during the copyout,
3002 		 * update the offset in the uio struct to
3003 		 * contain the value of the next cookie
3004 		 * and set the eof value appropriately.
3005 		 */
3006 		if (!error) {
3007 			uiop->uio_offset = rdc->nfs_ncookie;
3008 			if (eofp)
3009 				*eofp = rdc->eof;
3010 		}
3011 
3012 		/*
3013 		 * Decide whether to do readahead.  Don't if
3014 		 * have already read to the end of directory.
3015 		 */
3016 		if (rdc->eof) {
3017 			rp->r_direof = rdc;
3018 			mutex_exit(&rp->r_statelock);
3019 			rddir_cache_rele(rdc);
3020 			if (nrdc != NULL)
3021 				rddir_cache_rele(nrdc);
3022 			return (error);
3023 		}
3024 
3025 		/*
3026 		 * Check to see whether we found an entry
3027 		 * for the readahead.  If so, we don't need
3028 		 * to do anything further, so free the new
3029 		 * entry if one was allocated.  Otherwise,
3030 		 * allocate a new entry, add it to the cache,
3031 		 * and then initiate an asynchronous readdir
3032 		 * operation to fill it.
3033 		 */
3034 		srdc.nfs_cookie = rdc->nfs_ncookie;
3035 		srdc.buflen = count;
3036 		rrdc = avl_find(&rp->r_dir, &srdc, &where);
3037 		if (rrdc != NULL) {
3038 			if (nrdc != NULL)
3039 				rddir_cache_rele(nrdc);
3040 		} else {
3041 			if (nrdc != NULL)
3042 				rrdc = nrdc;
3043 			else {
3044 				rrdc = rddir_cache_alloc(KM_NOSLEEP);
3045 			}
3046 			if (rrdc != NULL) {
3047 				rrdc->nfs_cookie = rdc->nfs_ncookie;
3048 				rrdc->buflen = count;
3049 				avl_insert(&rp->r_dir, rrdc, where);
3050 				rddir_cache_hold(rrdc);
3051 				mutex_exit(&rp->r_statelock);
3052 				rddir_cache_rele(rdc);
3053 #ifdef DEBUG
3054 				nfs_readdir_readahead++;
3055 #endif
3056 				nfs_async_readdir(vp, rrdc, cr, nfsreaddir);
3057 				return (error);
3058 			}
3059 		}
3060 
3061 		mutex_exit(&rp->r_statelock);
3062 		rddir_cache_rele(rdc);
3063 		return (error);
3064 	}
3065 
3066 	/*
3067 	 * Didn't find an entry in the cache.  Construct a new empty
3068 	 * entry and link it into the cache.  Other processes attempting
3069 	 * to access this entry will need to wait until it is filled in.
3070 	 *
3071 	 * Since kmem_alloc may block, another pass through the cache
3072 	 * will need to be taken to make sure that another process
3073 	 * hasn't already added an entry to the cache for this request.
3074 	 */
3075 	if (nrdc == NULL) {
3076 		mutex_exit(&rp->r_statelock);
3077 		nrdc = rddir_cache_alloc(KM_SLEEP);
3078 		nrdc->nfs_cookie = uiop->uio_offset;
3079 		nrdc->buflen = count;
3080 		goto top;
3081 	}
3082 
3083 	/*
3084 	 * Add this entry to the cache.
3085 	 */
3086 	avl_insert(&rp->r_dir, nrdc, where);
3087 	rddir_cache_hold(nrdc);
3088 	mutex_exit(&rp->r_statelock);
3089 
3090 bottom:
3091 #ifdef DEBUG
3092 	missed = 1;
3093 	nfs_readdir_cache_misses++;
3094 #endif
3095 	/*
3096 	 * Do the readdir.
3097 	 */
3098 	error = nfsreaddir(vp, nrdc, cr);
3099 
3100 	/*
3101 	 * If this operation failed, just return the error which occurred.
3102 	 */
3103 	if (error != 0)
3104 		return (error);
3105 
3106 	/*
3107 	 * Since the RPC operation will have taken sometime and blocked
3108 	 * this process, another pass through the cache will need to be
3109 	 * taken to find the correct cache entry.  It is possible that
3110 	 * the correct cache entry will not be there (although one was
3111 	 * added) because the directory changed during the RPC operation
3112 	 * and the readdir cache was flushed.  In this case, just start
3113 	 * over.  It is hoped that this will not happen too often... :-)
3114 	 */
3115 	nrdc = NULL;
3116 	goto top;
3117 	/* NOTREACHED */
3118 }
3119 
3120 static int
3121 nfsreaddir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3122 {
3123 	int error;
3124 	struct nfsrddirargs rda;
3125 	struct nfsrddirres rd;
3126 	rnode_t *rp;
3127 	mntinfo_t *mi;
3128 	uint_t count;
3129 	int douprintf;
3130 	failinfo_t fi, *fip;
3131 
3132 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
3133 	count = rdc->buflen;
3134 
3135 	rp = VTOR(vp);
3136 	mi = VTOMI(vp);
3137 
3138 	rda.rda_fh = *VTOFH(vp);
3139 	rda.rda_offset = rdc->nfs_cookie;
3140 
3141 	/*
3142 	 * NFS client failover support
3143 	 * suppress failover unless we have a zero cookie
3144 	 */
3145 	if (rdc->nfs_cookie == (off_t)0) {
3146 		fi.vp = vp;
3147 		fi.fhp = (caddr_t)&rda.rda_fh;
3148 		fi.copyproc = nfscopyfh;
3149 		fi.lookupproc = nfslookup;
3150 		fi.xattrdirproc = acl_getxattrdir2;
3151 		fip = &fi;
3152 	} else {
3153 		fip = NULL;
3154 	}
3155 
3156 	rd.rd_entries = kmem_alloc(rdc->buflen, KM_SLEEP);
3157 	rd.rd_size = count;
3158 	rd.rd_offset = rda.rda_offset;
3159 
3160 	douprintf = 1;
3161 
3162 	if (mi->mi_io_kstats) {
3163 		mutex_enter(&mi->mi_lock);
3164 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
3165 		mutex_exit(&mi->mi_lock);
3166 	}
3167 
3168 	do {
3169 		rda.rda_count = MIN(count, mi->mi_curread);
3170 		error = rfs2call(mi, RFS_READDIR,
3171 		    xdr_rddirargs, (caddr_t)&rda,
3172 		    xdr_getrddirres, (caddr_t)&rd, cr,
3173 		    &douprintf, &rd.rd_status, 0, fip);
3174 	} while (error == ENFS_TRYAGAIN);
3175 
3176 	if (mi->mi_io_kstats) {
3177 		mutex_enter(&mi->mi_lock);
3178 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
3179 		mutex_exit(&mi->mi_lock);
3180 	}
3181 
3182 	/*
3183 	 * Since we are actually doing a READDIR RPC, we must have
3184 	 * exclusive access to the cache entry being filled.  Thus,
3185 	 * it is safe to update all fields except for the flags
3186 	 * field.  The r_statelock in the rnode must be held to
3187 	 * prevent two different threads from simultaneously
3188 	 * attempting to update the flags field.  This can happen
3189 	 * if we are turning off RDDIR and the other thread is
3190 	 * trying to set RDDIRWAIT.
3191 	 */
3192 	ASSERT(rdc->flags & RDDIR);
3193 	if (!error) {
3194 		error = geterrno(rd.rd_status);
3195 		if (!error) {
3196 			rdc->nfs_ncookie = rd.rd_offset;
3197 			rdc->eof = rd.rd_eof ? 1 : 0;
3198 			rdc->entlen = rd.rd_size;
3199 			ASSERT(rdc->entlen <= rdc->buflen);
3200 #ifdef DEBUG
3201 			rdc->entries = rddir_cache_buf_alloc(rdc->buflen,
3202 			    KM_SLEEP);
3203 #else
3204 			rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
3205 #endif
3206 			bcopy(rd.rd_entries, rdc->entries, rdc->entlen);
3207 			rdc->error = 0;
3208 			if (mi->mi_io_kstats) {
3209 				mutex_enter(&mi->mi_lock);
3210 				KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
3211 				KSTAT_IO_PTR(mi->mi_io_kstats)->nread +=
3212 				    rd.rd_size;
3213 				mutex_exit(&mi->mi_lock);
3214 			}
3215 		} else {
3216 			PURGE_STALE_FH(error, vp, cr);
3217 		}
3218 	}
3219 	if (error) {
3220 		rdc->entries = NULL;
3221 		rdc->error = error;
3222 	}
3223 	kmem_free(rd.rd_entries, rdc->buflen);
3224 
3225 	mutex_enter(&rp->r_statelock);
3226 	rdc->flags &= ~RDDIR;
3227 	if (rdc->flags & RDDIRWAIT) {
3228 		rdc->flags &= ~RDDIRWAIT;
3229 		cv_broadcast(&rdc->cv);
3230 	}
3231 	if (error)
3232 		rdc->flags |= RDDIRREQ;
3233 	mutex_exit(&rp->r_statelock);
3234 
3235 	rddir_cache_rele(rdc);
3236 
3237 	return (error);
3238 }
3239 
3240 #ifdef DEBUG
3241 static int nfs_bio_do_stop = 0;
3242 #endif
3243 
3244 static int
3245 nfs_bio(struct buf *bp, cred_t *cr)
3246 {
3247 	rnode_t *rp = VTOR(bp->b_vp);
3248 	int count;
3249 	int error;
3250 	cred_t *cred;
3251 	uint_t offset;
3252 
3253 	DTRACE_IO1(start, struct buf *, bp);
3254 
3255 	ASSERT(nfs_zone() == VTOMI(bp->b_vp)->mi_zone);
3256 	offset = dbtob(bp->b_blkno);
3257 
3258 	if (bp->b_flags & B_READ) {
3259 		mutex_enter(&rp->r_statelock);
3260 		if (rp->r_cred != NULL) {
3261 			cred = rp->r_cred;
3262 			crhold(cred);
3263 		} else {
3264 			rp->r_cred = cr;
3265 			crhold(cr);
3266 			cred = cr;
3267 			crhold(cred);
3268 		}
3269 		mutex_exit(&rp->r_statelock);
3270 	read_again:
3271 		error = bp->b_error = nfsread(bp->b_vp, bp->b_un.b_addr,
3272 		    offset, bp->b_bcount, &bp->b_resid, cred);
3273 		crfree(cred);
3274 		if (!error) {
3275 			if (bp->b_resid) {
3276 				/*
3277 				 * Didn't get it all because we hit EOF,
3278 				 * zero all the memory beyond the EOF.
3279 				 */
3280 				/* bzero(rdaddr + */
3281 				bzero(bp->b_un.b_addr +
3282 				    bp->b_bcount - bp->b_resid, bp->b_resid);
3283 			}
3284 			mutex_enter(&rp->r_statelock);
3285 			if (bp->b_resid == bp->b_bcount &&
3286 			    offset >= rp->r_size) {
3287 				/*
3288 				 * We didn't read anything at all as we are
3289 				 * past EOF.  Return an error indicator back
3290 				 * but don't destroy the pages (yet).
3291 				 */
3292 				error = NFS_EOF;
3293 			}
3294 			mutex_exit(&rp->r_statelock);
3295 		} else if (error == EACCES) {
3296 			mutex_enter(&rp->r_statelock);
3297 			if (cred != cr) {
3298 				if (rp->r_cred != NULL)
3299 					crfree(rp->r_cred);
3300 				rp->r_cred = cr;
3301 				crhold(cr);
3302 				cred = cr;
3303 				crhold(cred);
3304 				mutex_exit(&rp->r_statelock);
3305 				goto read_again;
3306 			}
3307 			mutex_exit(&rp->r_statelock);
3308 		}
3309 	} else {
3310 		if (!(rp->r_flags & RSTALE)) {
3311 			mutex_enter(&rp->r_statelock);
3312 			if (rp->r_cred != NULL) {
3313 				cred = rp->r_cred;
3314 				crhold(cred);
3315 			} else {
3316 				rp->r_cred = cr;
3317 				crhold(cr);
3318 				cred = cr;
3319 				crhold(cred);
3320 			}
3321 			mutex_exit(&rp->r_statelock);
3322 		write_again:
3323 			mutex_enter(&rp->r_statelock);
3324 			count = MIN(bp->b_bcount, rp->r_size - offset);
3325 			mutex_exit(&rp->r_statelock);
3326 			if (count < 0)
3327 				cmn_err(CE_PANIC, "nfs_bio: write count < 0");
3328 #ifdef DEBUG
3329 			if (count == 0) {
3330 				zcmn_err(getzoneid(), CE_WARN,
3331 				    "nfs_bio: zero length write at %d",
3332 				    offset);
3333 				nfs_printfhandle(&rp->r_fh);
3334 				if (nfs_bio_do_stop)
3335 					debug_enter("nfs_bio");
3336 			}
3337 #endif
3338 			error = nfswrite(bp->b_vp, bp->b_un.b_addr, offset,
3339 			    count, cred);
3340 			if (error == EACCES) {
3341 				mutex_enter(&rp->r_statelock);
3342 				if (cred != cr) {
3343 					if (rp->r_cred != NULL)
3344 						crfree(rp->r_cred);
3345 					rp->r_cred = cr;
3346 					crhold(cr);
3347 					crfree(cred);
3348 					cred = cr;
3349 					crhold(cred);
3350 					mutex_exit(&rp->r_statelock);
3351 					goto write_again;
3352 				}
3353 				mutex_exit(&rp->r_statelock);
3354 			}
3355 			bp->b_error = error;
3356 			if (error && error != EINTR) {
3357 				/*
3358 				 * Don't print EDQUOT errors on the console.
3359 				 * Don't print asynchronous EACCES errors.
3360 				 * Don't print EFBIG errors.
3361 				 * Print all other write errors.
3362 				 */
3363 				if (error != EDQUOT && error != EFBIG &&
3364 				    (error != EACCES ||
3365 				    !(bp->b_flags & B_ASYNC)))
3366 					nfs_write_error(bp->b_vp, error, cred);
3367 				/*
3368 				 * Update r_error and r_flags as appropriate.
3369 				 * If the error was ESTALE, then mark the
3370 				 * rnode as not being writeable and save
3371 				 * the error status.  Otherwise, save any
3372 				 * errors which occur from asynchronous
3373 				 * page invalidations.  Any errors occurring
3374 				 * from other operations should be saved
3375 				 * by the caller.
3376 				 */
3377 				mutex_enter(&rp->r_statelock);
3378 				if (error == ESTALE) {
3379 					rp->r_flags |= RSTALE;
3380 					if (!rp->r_error)
3381 						rp->r_error = error;
3382 				} else if (!rp->r_error &&
3383 				    (bp->b_flags &
3384 				    (B_INVAL|B_FORCE|B_ASYNC)) ==
3385 				    (B_INVAL|B_FORCE|B_ASYNC)) {
3386 					rp->r_error = error;
3387 				}
3388 				mutex_exit(&rp->r_statelock);
3389 			}
3390 			crfree(cred);
3391 		} else
3392 			error = rp->r_error;
3393 	}
3394 
3395 	if (error != 0 && error != NFS_EOF)
3396 		bp->b_flags |= B_ERROR;
3397 
3398 	DTRACE_IO1(done, struct buf *, bp);
3399 
3400 	return (error);
3401 }
3402 
3403 static int
3404 nfs_fid(vnode_t *vp, fid_t *fidp)
3405 {
3406 	struct nfs_fid *fp;
3407 	rnode_t *rp;
3408 
3409 	rp = VTOR(vp);
3410 
3411 	if (fidp->fid_len < (sizeof (struct nfs_fid) - sizeof (short))) {
3412 		fidp->fid_len = sizeof (struct nfs_fid) - sizeof (short);
3413 		return (ENOSPC);
3414 	}
3415 	fp = (struct nfs_fid *)fidp;
3416 	fp->nf_pad = 0;
3417 	fp->nf_len = sizeof (struct nfs_fid) - sizeof (short);
3418 	bcopy(rp->r_fh.fh_buf, fp->nf_data, NFS_FHSIZE);
3419 	return (0);
3420 }
3421 
3422 /* ARGSUSED2 */
3423 static int
3424 nfs_rwlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
3425 {
3426 	rnode_t *rp = VTOR(vp);
3427 
3428 	if (!write_lock) {
3429 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
3430 		return (V_WRITELOCK_FALSE);
3431 	}
3432 
3433 	if ((rp->r_flags & RDIRECTIO) || (VTOMI(vp)->mi_flags & MI_DIRECTIO)) {
3434 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
3435 		if (rp->r_mapcnt == 0 && !vn_has_cached_data(vp))
3436 			return (V_WRITELOCK_FALSE);
3437 		nfs_rw_exit(&rp->r_rwlock);
3438 	}
3439 
3440 	(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, FALSE);
3441 	return (V_WRITELOCK_TRUE);
3442 }
3443 
3444 /* ARGSUSED */
3445 static void
3446 nfs_rwunlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
3447 {
3448 	rnode_t *rp = VTOR(vp);
3449 
3450 	nfs_rw_exit(&rp->r_rwlock);
3451 }
3452 
3453 /* ARGSUSED */
3454 static int
3455 nfs_seek(vnode_t *vp, offset_t ooff, offset_t *noffp)
3456 {
3457 
3458 	/*
3459 	 * Because we stuff the readdir cookie into the offset field
3460 	 * someone may attempt to do an lseek with the cookie which
3461 	 * we want to succeed.
3462 	 */
3463 	if (vp->v_type == VDIR)
3464 		return (0);
3465 	if (*noffp < 0 || *noffp > MAXOFF32_T)
3466 		return (EINVAL);
3467 	return (0);
3468 }
3469 
3470 /*
3471  * number of NFS_MAXDATA blocks to read ahead
3472  * optimized for 100 base-T.
3473  */
3474 static int nfs_nra = 4;
3475 
3476 #ifdef DEBUG
3477 static int nfs_lostpage = 0;	/* number of times we lost original page */
3478 #endif
3479 
3480 /*
3481  * Return all the pages from [off..off+len) in file
3482  */
3483 static int
3484 nfs_getpage(vnode_t *vp, offset_t off, size_t len, uint_t *protp,
3485 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
3486 	enum seg_rw rw, cred_t *cr)
3487 {
3488 	rnode_t *rp;
3489 	int error;
3490 	mntinfo_t *mi;
3491 
3492 	if (vp->v_flag & VNOMAP)
3493 		return (ENOSYS);
3494 
3495 	ASSERT(off <= MAXOFF32_T);
3496 	if (nfs_zone() != VTOMI(vp)->mi_zone)
3497 		return (EIO);
3498 	if (protp != NULL)
3499 		*protp = PROT_ALL;
3500 
3501 	/*
3502 	 * Now valididate that the caches are up to date.
3503 	 */
3504 	error = nfs_validate_caches(vp, cr);
3505 	if (error)
3506 		return (error);
3507 
3508 	rp = VTOR(vp);
3509 	mi = VTOMI(vp);
3510 retry:
3511 	mutex_enter(&rp->r_statelock);
3512 
3513 	/*
3514 	 * Don't create dirty pages faster than they
3515 	 * can be cleaned so that the system doesn't
3516 	 * get imbalanced.  If the async queue is
3517 	 * maxed out, then wait for it to drain before
3518 	 * creating more dirty pages.  Also, wait for
3519 	 * any threads doing pagewalks in the vop_getattr
3520 	 * entry points so that they don't block for
3521 	 * long periods.
3522 	 */
3523 	if (rw == S_CREATE) {
3524 		while ((mi->mi_max_threads != 0 &&
3525 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
3526 		    rp->r_gcount > 0)
3527 			cv_wait(&rp->r_cv, &rp->r_statelock);
3528 	}
3529 
3530 	/*
3531 	 * If we are getting called as a side effect of an nfs_write()
3532 	 * operation the local file size might not be extended yet.
3533 	 * In this case we want to be able to return pages of zeroes.
3534 	 */
3535 	if (off + len > rp->r_size + PAGEOFFSET && seg != segkmap) {
3536 		mutex_exit(&rp->r_statelock);
3537 		return (EFAULT);		/* beyond EOF */
3538 	}
3539 
3540 	mutex_exit(&rp->r_statelock);
3541 
3542 	if (len <= PAGESIZE) {
3543 		error = nfs_getapage(vp, off, len, protp, pl, plsz,
3544 		    seg, addr, rw, cr);
3545 	} else {
3546 		error = pvn_getpages(nfs_getapage, vp, off, len, protp,
3547 		    pl, plsz, seg, addr, rw, cr);
3548 	}
3549 
3550 	switch (error) {
3551 	case NFS_EOF:
3552 		nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
3553 		goto retry;
3554 	case ESTALE:
3555 		PURGE_STALE_FH(error, vp, cr);
3556 	}
3557 
3558 	return (error);
3559 }
3560 
3561 /*
3562  * Called from pvn_getpages or nfs_getpage to get a particular page.
3563  */
3564 /* ARGSUSED */
3565 static int
3566 nfs_getapage(vnode_t *vp, u_offset_t off, size_t len, uint_t *protp,
3567 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
3568 	enum seg_rw rw, cred_t *cr)
3569 {
3570 	rnode_t *rp;
3571 	uint_t bsize;
3572 	struct buf *bp;
3573 	page_t *pp;
3574 	u_offset_t lbn;
3575 	u_offset_t io_off;
3576 	u_offset_t blkoff;
3577 	u_offset_t rablkoff;
3578 	size_t io_len;
3579 	uint_t blksize;
3580 	int error;
3581 	int readahead;
3582 	int readahead_issued = 0;
3583 	int ra_window; /* readahead window */
3584 	page_t *pagefound;
3585 
3586 	if (nfs_zone() != VTOMI(vp)->mi_zone)
3587 		return (EIO);
3588 	rp = VTOR(vp);
3589 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
3590 
3591 reread:
3592 	bp = NULL;
3593 	pp = NULL;
3594 	pagefound = NULL;
3595 
3596 	if (pl != NULL)
3597 		pl[0] = NULL;
3598 
3599 	error = 0;
3600 	lbn = off / bsize;
3601 	blkoff = lbn * bsize;
3602 
3603 	/*
3604 	 * Queueing up the readahead before doing the synchronous read
3605 	 * results in a significant increase in read throughput because
3606 	 * of the increased parallelism between the async threads and
3607 	 * the process context.
3608 	 */
3609 	if ((off & ((vp->v_vfsp->vfs_bsize) - 1)) == 0 &&
3610 	    rw != S_CREATE &&
3611 	    !(vp->v_flag & VNOCACHE)) {
3612 		mutex_enter(&rp->r_statelock);
3613 
3614 		/*
3615 		 * Calculate the number of readaheads to do.
3616 		 * a) No readaheads at offset = 0.
3617 		 * b) Do maximum(nfs_nra) readaheads when the readahead
3618 		 *    window is closed.
3619 		 * c) Do readaheads between 1 to (nfs_nra - 1) depending
3620 		 *    upon how far the readahead window is open or close.
3621 		 * d) No readaheads if rp->r_nextr is not within the scope
3622 		 *    of the readahead window (random i/o).
3623 		 */
3624 
3625 		if (off == 0)
3626 			readahead = 0;
3627 		else if (blkoff == rp->r_nextr)
3628 			readahead = nfs_nra;
3629 		else if (rp->r_nextr > blkoff &&
3630 				((ra_window = (rp->r_nextr - blkoff) / bsize)
3631 					<= (nfs_nra - 1)))
3632 			readahead = nfs_nra - ra_window;
3633 		else
3634 			readahead = 0;
3635 
3636 		rablkoff = rp->r_nextr;
3637 		while (readahead > 0 && rablkoff + bsize < rp->r_size) {
3638 			mutex_exit(&rp->r_statelock);
3639 			if (nfs_async_readahead(vp, rablkoff + bsize,
3640 			    addr + (rablkoff + bsize - off), seg, cr,
3641 			    nfs_readahead) < 0) {
3642 				mutex_enter(&rp->r_statelock);
3643 				break;
3644 			}
3645 			readahead--;
3646 			rablkoff += bsize;
3647 			/*
3648 			 * Indicate that we did a readahead so
3649 			 * readahead offset is not updated
3650 			 * by the synchronous read below.
3651 			 */
3652 			readahead_issued = 1;
3653 			mutex_enter(&rp->r_statelock);
3654 			/*
3655 			 * set readahead offset to
3656 			 * offset of last async readahead
3657 			 * request.
3658 			 */
3659 			rp->r_nextr = rablkoff;
3660 		}
3661 		mutex_exit(&rp->r_statelock);
3662 	}
3663 
3664 again:
3665 	if ((pagefound = page_exists(vp, off)) == NULL) {
3666 		if (pl == NULL) {
3667 			(void) nfs_async_readahead(vp, blkoff, addr, seg, cr,
3668 			    nfs_readahead);
3669 		} else if (rw == S_CREATE) {
3670 			/*
3671 			 * Block for this page is not allocated, or the offset
3672 			 * is beyond the current allocation size, or we're
3673 			 * allocating a swap slot and the page was not found,
3674 			 * so allocate it and return a zero page.
3675 			 */
3676 			if ((pp = page_create_va(vp, off,
3677 			    PAGESIZE, PG_WAIT, seg, addr)) == NULL)
3678 				cmn_err(CE_PANIC, "nfs_getapage: page_create");
3679 			io_len = PAGESIZE;
3680 			mutex_enter(&rp->r_statelock);
3681 			rp->r_nextr = off + PAGESIZE;
3682 			mutex_exit(&rp->r_statelock);
3683 		} else {
3684 			/*
3685 			 * Need to go to server to get a BLOCK, exception to
3686 			 * that being while reading at offset = 0 or doing
3687 			 * random i/o, in that case read only a PAGE.
3688 			 */
3689 			mutex_enter(&rp->r_statelock);
3690 			if (blkoff < rp->r_size &&
3691 			    blkoff + bsize >= rp->r_size) {
3692 				/*
3693 				 * If only a block or less is left in
3694 				 * the file, read all that is remaining.
3695 				 */
3696 				if (rp->r_size <= off) {
3697 					/*
3698 					 * Trying to access beyond EOF,
3699 					 * set up to get at least one page.
3700 					 */
3701 					blksize = off + PAGESIZE - blkoff;
3702 				} else
3703 					blksize = rp->r_size - blkoff;
3704 			} else if ((off == 0) ||
3705 				(off != rp->r_nextr && !readahead_issued)) {
3706 				blksize = PAGESIZE;
3707 				blkoff = off; /* block = page here */
3708 			} else
3709 				blksize = bsize;
3710 			mutex_exit(&rp->r_statelock);
3711 
3712 			pp = pvn_read_kluster(vp, off, seg, addr, &io_off,
3713 			    &io_len, blkoff, blksize, 0);
3714 
3715 			/*
3716 			 * Some other thread has entered the page,
3717 			 * so just use it.
3718 			 */
3719 			if (pp == NULL)
3720 				goto again;
3721 
3722 			/*
3723 			 * Now round the request size up to page boundaries.
3724 			 * This ensures that the entire page will be
3725 			 * initialized to zeroes if EOF is encountered.
3726 			 */
3727 			io_len = ptob(btopr(io_len));
3728 
3729 			bp = pageio_setup(pp, io_len, vp, B_READ);
3730 			ASSERT(bp != NULL);
3731 
3732 			/*
3733 			 * pageio_setup should have set b_addr to 0.  This
3734 			 * is correct since we want to do I/O on a page
3735 			 * boundary.  bp_mapin will use this addr to calculate
3736 			 * an offset, and then set b_addr to the kernel virtual
3737 			 * address it allocated for us.
3738 			 */
3739 			ASSERT(bp->b_un.b_addr == 0);
3740 
3741 			bp->b_edev = 0;
3742 			bp->b_dev = 0;
3743 			bp->b_lblkno = lbtodb(io_off);
3744 			bp->b_file = vp;
3745 			bp->b_offset = (offset_t)off;
3746 			bp_mapin(bp);
3747 
3748 			/*
3749 			 * If doing a write beyond what we believe is EOF,
3750 			 * don't bother trying to read the pages from the
3751 			 * server, we'll just zero the pages here.  We
3752 			 * don't check that the rw flag is S_WRITE here
3753 			 * because some implementations may attempt a
3754 			 * read access to the buffer before copying data.
3755 			 */
3756 			mutex_enter(&rp->r_statelock);
3757 			if (io_off >= rp->r_size && seg == segkmap) {
3758 				mutex_exit(&rp->r_statelock);
3759 				bzero(bp->b_un.b_addr, io_len);
3760 			} else {
3761 				mutex_exit(&rp->r_statelock);
3762 				error = nfs_bio(bp, cr);
3763 			}
3764 
3765 			/*
3766 			 * Unmap the buffer before freeing it.
3767 			 */
3768 			bp_mapout(bp);
3769 			pageio_done(bp);
3770 
3771 			if (error == NFS_EOF) {
3772 				/*
3773 				 * If doing a write system call just return
3774 				 * zeroed pages, else user tried to get pages
3775 				 * beyond EOF, return error.  We don't check
3776 				 * that the rw flag is S_WRITE here because
3777 				 * some implementations may attempt a read
3778 				 * access to the buffer before copying data.
3779 				 */
3780 				if (seg == segkmap)
3781 					error = 0;
3782 				else
3783 					error = EFAULT;
3784 			}
3785 
3786 			if (!readahead_issued && !error) {
3787 			    mutex_enter(&rp->r_statelock);
3788 			    rp->r_nextr = io_off + io_len;
3789 			    mutex_exit(&rp->r_statelock);
3790 			}
3791 		}
3792 	}
3793 
3794 out:
3795 	if (pl == NULL)
3796 		return (error);
3797 
3798 	if (error) {
3799 		if (pp != NULL)
3800 			pvn_read_done(pp, B_ERROR);
3801 		return (error);
3802 	}
3803 
3804 	if (pagefound) {
3805 		se_t se = (rw == S_CREATE ? SE_EXCL : SE_SHARED);
3806 
3807 		/*
3808 		 * Page exists in the cache, acquire the appropriate lock.
3809 		 * If this fails, start all over again.
3810 		 */
3811 		if ((pp = page_lookup(vp, off, se)) == NULL) {
3812 #ifdef DEBUG
3813 			nfs_lostpage++;
3814 #endif
3815 			goto reread;
3816 		}
3817 		pl[0] = pp;
3818 		pl[1] = NULL;
3819 		return (0);
3820 	}
3821 
3822 	if (pp != NULL)
3823 		pvn_plist_init(pp, pl, plsz, off, io_len, rw);
3824 
3825 	return (error);
3826 }
3827 
3828 static void
3829 nfs_readahead(vnode_t *vp, u_offset_t blkoff, caddr_t addr, struct seg *seg,
3830 	cred_t *cr)
3831 {
3832 	int error;
3833 	page_t *pp;
3834 	u_offset_t io_off;
3835 	size_t io_len;
3836 	struct buf *bp;
3837 	uint_t bsize, blksize;
3838 	rnode_t *rp = VTOR(vp);
3839 
3840 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
3841 
3842 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
3843 
3844 	mutex_enter(&rp->r_statelock);
3845 	if (blkoff < rp->r_size && blkoff + bsize > rp->r_size) {
3846 		/*
3847 		 * If less than a block left in file read less
3848 		 * than a block.
3849 		 */
3850 		blksize = rp->r_size - blkoff;
3851 	} else
3852 		blksize = bsize;
3853 	mutex_exit(&rp->r_statelock);
3854 
3855 	pp = pvn_read_kluster(vp, blkoff, segkmap, addr,
3856 	    &io_off, &io_len, blkoff, blksize, 1);
3857 	/*
3858 	 * The isra flag passed to the kluster function is 1, we may have
3859 	 * gotten a return value of NULL for a variety of reasons (# of free
3860 	 * pages < minfree, someone entered the page on the vnode etc). In all
3861 	 * cases, we want to punt on the readahead.
3862 	 */
3863 	if (pp == NULL)
3864 		return;
3865 
3866 	/*
3867 	 * Now round the request size up to page boundaries.
3868 	 * This ensures that the entire page will be
3869 	 * initialized to zeroes if EOF is encountered.
3870 	 */
3871 	io_len = ptob(btopr(io_len));
3872 
3873 	bp = pageio_setup(pp, io_len, vp, B_READ);
3874 	ASSERT(bp != NULL);
3875 
3876 	/*
3877 	 * pageio_setup should have set b_addr to 0.  This is correct since
3878 	 * we want to do I/O on a page boundary. bp_mapin() will use this addr
3879 	 * to calculate an offset, and then set b_addr to the kernel virtual
3880 	 * address it allocated for us.
3881 	 */
3882 	ASSERT(bp->b_un.b_addr == 0);
3883 
3884 	bp->b_edev = 0;
3885 	bp->b_dev = 0;
3886 	bp->b_lblkno = lbtodb(io_off);
3887 	bp->b_file = vp;
3888 	bp->b_offset = (offset_t)blkoff;
3889 	bp_mapin(bp);
3890 
3891 	/*
3892 	 * If doing a write beyond what we believe is EOF, don't bother trying
3893 	 * to read the pages from the server, we'll just zero the pages here.
3894 	 * We don't check that the rw flag is S_WRITE here because some
3895 	 * implementations may attempt a read access to the buffer before
3896 	 * copying data.
3897 	 */
3898 	mutex_enter(&rp->r_statelock);
3899 	if (io_off >= rp->r_size && seg == segkmap) {
3900 		mutex_exit(&rp->r_statelock);
3901 		bzero(bp->b_un.b_addr, io_len);
3902 		error = 0;
3903 	} else {
3904 		mutex_exit(&rp->r_statelock);
3905 		error = nfs_bio(bp, cr);
3906 		if (error == NFS_EOF)
3907 			error = 0;
3908 	}
3909 
3910 	/*
3911 	 * Unmap the buffer before freeing it.
3912 	 */
3913 	bp_mapout(bp);
3914 	pageio_done(bp);
3915 
3916 	pvn_read_done(pp, error ? B_READ | B_ERROR : B_READ);
3917 
3918 	/*
3919 	 * In case of error set readahead offset
3920 	 * to the lowest offset.
3921 	 * pvn_read_done() calls VN_DISPOSE to destroy the pages
3922 	 */
3923 	if (error && rp->r_nextr > io_off) {
3924 		mutex_enter(&rp->r_statelock);
3925 		if (rp->r_nextr > io_off)
3926 			rp->r_nextr = io_off;
3927 		mutex_exit(&rp->r_statelock);
3928 	}
3929 }
3930 
3931 /*
3932  * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
3933  * If len == 0, do from off to EOF.
3934  *
3935  * The normal cases should be len == 0 && off == 0 (entire vp list),
3936  * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
3937  * (from pageout).
3938  */
3939 static int
3940 nfs_putpage(vnode_t *vp, offset_t off, size_t len, int flags, cred_t *cr)
3941 {
3942 	int error;
3943 	rnode_t *rp;
3944 
3945 	ASSERT(cr != NULL);
3946 
3947 	/*
3948 	 * XXX - Why should this check be made here?
3949 	 */
3950 	if (vp->v_flag & VNOMAP)
3951 		return (ENOSYS);
3952 
3953 	if (len == 0 && !(flags & B_INVAL) && vn_is_readonly(vp))
3954 		return (0);
3955 
3956 	if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
3957 		return (EIO);
3958 	ASSERT(off <= MAXOFF32_T);
3959 
3960 	rp = VTOR(vp);
3961 	mutex_enter(&rp->r_statelock);
3962 	rp->r_count++;
3963 	mutex_exit(&rp->r_statelock);
3964 	error = nfs_putpages(vp, off, len, flags, cr);
3965 	mutex_enter(&rp->r_statelock);
3966 	rp->r_count--;
3967 	cv_broadcast(&rp->r_cv);
3968 	mutex_exit(&rp->r_statelock);
3969 
3970 	return (error);
3971 }
3972 
3973 /*
3974  * Write out a single page, possibly klustering adjacent dirty pages.
3975  */
3976 int
3977 nfs_putapage(vnode_t *vp, page_t *pp, u_offset_t *offp, size_t *lenp,
3978 	int flags, cred_t *cr)
3979 {
3980 	u_offset_t io_off;
3981 	u_offset_t lbn_off;
3982 	u_offset_t lbn;
3983 	size_t io_len;
3984 	uint_t bsize;
3985 	int error;
3986 	rnode_t *rp;
3987 
3988 	ASSERT(!vn_is_readonly(vp));
3989 	ASSERT(pp != NULL);
3990 	ASSERT(cr != NULL);
3991 	ASSERT((flags & B_ASYNC) || nfs_zone() == VTOMI(vp)->mi_zone);
3992 
3993 	rp = VTOR(vp);
3994 	ASSERT(rp->r_count > 0);
3995 
3996 	ASSERT(pp->p_offset <= MAXOFF32_T);
3997 
3998 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
3999 	lbn = pp->p_offset / bsize;
4000 	lbn_off = lbn * bsize;
4001 
4002 	/*
4003 	 * Find a kluster that fits in one block, or in
4004 	 * one page if pages are bigger than blocks.  If
4005 	 * there is less file space allocated than a whole
4006 	 * page, we'll shorten the i/o request below.
4007 	 */
4008 	pp = pvn_write_kluster(vp, pp, &io_off, &io_len, lbn_off,
4009 	    roundup(bsize, PAGESIZE), flags);
4010 
4011 	/*
4012 	 * pvn_write_kluster shouldn't have returned a page with offset
4013 	 * behind the original page we were given.  Verify that.
4014 	 */
4015 	ASSERT((pp->p_offset / bsize) >= lbn);
4016 
4017 	/*
4018 	 * Now pp will have the list of kept dirty pages marked for
4019 	 * write back.  It will also handle invalidation and freeing
4020 	 * of pages that are not dirty.  Check for page length rounding
4021 	 * problems.
4022 	 */
4023 	if (io_off + io_len > lbn_off + bsize) {
4024 		ASSERT((io_off + io_len) - (lbn_off + bsize) < PAGESIZE);
4025 		io_len = lbn_off + bsize - io_off;
4026 	}
4027 	/*
4028 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4029 	 * consistent value of r_size. RMODINPROGRESS is set in writerp().
4030 	 * When RMODINPROGRESS is set it indicates that a uiomove() is in
4031 	 * progress and the r_size has not been made consistent with the
4032 	 * new size of the file. When the uiomove() completes the r_size is
4033 	 * updated and the RMODINPROGRESS flag is cleared.
4034 	 *
4035 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4036 	 * consistent value of r_size. Without this handshaking, it is
4037 	 * possible that nfs(3)_bio() picks  up the old value of r_size
4038 	 * before the uiomove() in writerp() completes. This will result
4039 	 * in the write through nfs(3)_bio() being dropped.
4040 	 *
4041 	 * More precisely, there is a window between the time the uiomove()
4042 	 * completes and the time the r_size is updated. If a VOP_PUTPAGE()
4043 	 * operation intervenes in this window, the page will be picked up,
4044 	 * because it is dirty (it will be unlocked, unless it was
4045 	 * pagecreate'd). When the page is picked up as dirty, the dirty
4046 	 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
4047 	 * checked. This will still be the old size. Therefore the page will
4048 	 * not be written out. When segmap_release() calls VOP_PUTPAGE(),
4049 	 * the page will be found to be clean and the write will be dropped.
4050 	 */
4051 	if (rp->r_flags & RMODINPROGRESS) {
4052 		mutex_enter(&rp->r_statelock);
4053 		if ((rp->r_flags & RMODINPROGRESS) &&
4054 		    rp->r_modaddr + MAXBSIZE > io_off &&
4055 		    rp->r_modaddr < io_off + io_len) {
4056 			page_t *plist;
4057 			/*
4058 			 * A write is in progress for this region of the file.
4059 			 * If we did not detect RMODINPROGRESS here then this
4060 			 * path through nfs_putapage() would eventually go to
4061 			 * nfs(3)_bio() and may not write out all of the data
4062 			 * in the pages. We end up losing data. So we decide
4063 			 * to set the modified bit on each page in the page
4064 			 * list and mark the rnode with RDIRTY. This write
4065 			 * will be restarted at some later time.
4066 			 */
4067 			plist = pp;
4068 			while (plist != NULL) {
4069 				pp = plist;
4070 				page_sub(&plist, pp);
4071 				hat_setmod(pp);
4072 				page_io_unlock(pp);
4073 				page_unlock(pp);
4074 			}
4075 			rp->r_flags |= RDIRTY;
4076 			mutex_exit(&rp->r_statelock);
4077 			if (offp)
4078 				*offp = io_off;
4079 			if (lenp)
4080 				*lenp = io_len;
4081 			return (0);
4082 		}
4083 		mutex_exit(&rp->r_statelock);
4084 	}
4085 
4086 	if (flags & B_ASYNC) {
4087 		error = nfs_async_putapage(vp, pp, io_off, io_len, flags, cr,
4088 		    nfs_sync_putapage);
4089 	} else
4090 		error = nfs_sync_putapage(vp, pp, io_off, io_len, flags, cr);
4091 
4092 	if (offp)
4093 		*offp = io_off;
4094 	if (lenp)
4095 		*lenp = io_len;
4096 	return (error);
4097 }
4098 
4099 static int
4100 nfs_sync_putapage(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
4101 	int flags, cred_t *cr)
4102 {
4103 	int error;
4104 	rnode_t *rp;
4105 
4106 	flags |= B_WRITE;
4107 
4108 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
4109 	error = nfs_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
4110 
4111 	rp = VTOR(vp);
4112 
4113 	if ((error == ENOSPC || error == EDQUOT || error == EACCES) &&
4114 	    (flags & (B_INVAL|B_FORCE)) != (B_INVAL|B_FORCE)) {
4115 		if (!(rp->r_flags & ROUTOFSPACE)) {
4116 			mutex_enter(&rp->r_statelock);
4117 			rp->r_flags |= ROUTOFSPACE;
4118 			mutex_exit(&rp->r_statelock);
4119 		}
4120 		flags |= B_ERROR;
4121 		pvn_write_done(pp, flags);
4122 		/*
4123 		 * If this was not an async thread, then try again to
4124 		 * write out the pages, but this time, also destroy
4125 		 * them whether or not the write is successful.  This
4126 		 * will prevent memory from filling up with these
4127 		 * pages and destroying them is the only alternative
4128 		 * if they can't be written out.
4129 		 *
4130 		 * Don't do this if this is an async thread because
4131 		 * when the pages are unlocked in pvn_write_done,
4132 		 * some other thread could have come along, locked
4133 		 * them, and queued for an async thread.  It would be
4134 		 * possible for all of the async threads to be tied
4135 		 * up waiting to lock the pages again and they would
4136 		 * all already be locked and waiting for an async
4137 		 * thread to handle them.  Deadlock.
4138 		 */
4139 		if (!(flags & B_ASYNC)) {
4140 			error = nfs_putpage(vp, io_off, io_len,
4141 			    B_INVAL | B_FORCE, cr);
4142 		}
4143 	} else {
4144 		if (error)
4145 			flags |= B_ERROR;
4146 		else if (rp->r_flags & ROUTOFSPACE) {
4147 			mutex_enter(&rp->r_statelock);
4148 			rp->r_flags &= ~ROUTOFSPACE;
4149 			mutex_exit(&rp->r_statelock);
4150 		}
4151 		pvn_write_done(pp, flags);
4152 	}
4153 
4154 	return (error);
4155 }
4156 
4157 static int
4158 nfs_map(vnode_t *vp, offset_t off, struct as *as, caddr_t *addrp,
4159 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
4160 {
4161 	struct segvn_crargs vn_a;
4162 	int error;
4163 	rnode_t *rp;
4164 	struct vattr va;
4165 
4166 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4167 		return (EIO);
4168 
4169 	if (vp->v_flag & VNOMAP)
4170 		return (ENOSYS);
4171 
4172 	if (off > MAXOFF32_T)
4173 		return (EFBIG);
4174 
4175 	if (off < 0 || off + len < 0)
4176 		return (ENXIO);
4177 
4178 	if (vp->v_type != VREG)
4179 		return (ENODEV);
4180 
4181 	/*
4182 	 * If there is cached data and if close-to-open consistency
4183 	 * checking is not turned off and if the file system is not
4184 	 * mounted readonly, then force an over the wire getattr.
4185 	 * Otherwise, just invoke nfsgetattr to get a copy of the
4186 	 * attributes.  The attribute cache will be used unless it
4187 	 * is timed out and if it is, then an over the wire getattr
4188 	 * will be issued.
4189 	 */
4190 	va.va_mask = AT_ALL;
4191 	if (vn_has_cached_data(vp) &&
4192 	    !(VTOMI(vp)->mi_flags & MI_NOCTO) && !vn_is_readonly(vp))
4193 		error = nfs_getattr_otw(vp, &va, cr);
4194 	else
4195 		error = nfsgetattr(vp, &va, cr);
4196 	if (error)
4197 		return (error);
4198 
4199 	/*
4200 	 * Check to see if the vnode is currently marked as not cachable.
4201 	 * This means portions of the file are locked (through VOP_FRLOCK).
4202 	 * In this case the map request must be refused.  We use
4203 	 * rp->r_lkserlock to avoid a race with concurrent lock requests.
4204 	 */
4205 	rp = VTOR(vp);
4206 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
4207 		return (EINTR);
4208 
4209 	if (vp->v_flag & VNOCACHE) {
4210 		error = EAGAIN;
4211 		goto done;
4212 	}
4213 
4214 	/*
4215 	 * Don't allow concurrent locks and mapping if mandatory locking is
4216 	 * enabled.
4217 	 */
4218 	if ((flk_has_remote_locks(vp) || lm_has_sleep(vp)) &&
4219 	    MANDLOCK(vp, va.va_mode)) {
4220 		error = EAGAIN;
4221 		goto done;
4222 	}
4223 
4224 	as_rangelock(as);
4225 	if (!(flags & MAP_FIXED)) {
4226 		map_addr(addrp, len, off, 1, flags);
4227 		if (*addrp == NULL) {
4228 			as_rangeunlock(as);
4229 			error = ENOMEM;
4230 			goto done;
4231 		}
4232 	} else {
4233 		/*
4234 		 * User specified address - blow away any previous mappings
4235 		 */
4236 		(void) as_unmap(as, *addrp, len);
4237 	}
4238 
4239 	vn_a.vp = vp;
4240 	vn_a.offset = off;
4241 	vn_a.type = (flags & MAP_TYPE);
4242 	vn_a.prot = (uchar_t)prot;
4243 	vn_a.maxprot = (uchar_t)maxprot;
4244 	vn_a.flags = (flags & ~MAP_TYPE);
4245 	vn_a.cred = cr;
4246 	vn_a.amp = NULL;
4247 	vn_a.szc = 0;
4248 	vn_a.lgrp_mem_policy_flags = 0;
4249 
4250 	error = as_map(as, *addrp, len, segvn_create, &vn_a);
4251 	as_rangeunlock(as);
4252 
4253 done:
4254 	nfs_rw_exit(&rp->r_lkserlock);
4255 	return (error);
4256 }
4257 
4258 /* ARGSUSED */
4259 static int
4260 nfs_addmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
4261 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
4262 {
4263 	rnode_t *rp;
4264 
4265 	if (vp->v_flag & VNOMAP)
4266 		return (ENOSYS);
4267 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4268 		return (EIO);
4269 
4270 	/*
4271 	 * Need to hold rwlock while incrementing the mapcnt so that
4272 	 * mmap'ing can be serialized with writes so that the caching
4273 	 * can be handled correctly.
4274 	 */
4275 	rp = VTOR(vp);
4276 	if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, INTR(vp)))
4277 		return (EINTR);
4278 	atomic_add_long((ulong_t *)&rp->r_mapcnt, btopr(len));
4279 	nfs_rw_exit(&rp->r_rwlock);
4280 
4281 	return (0);
4282 }
4283 
4284 static int
4285 nfs_frlock(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
4286 	offset_t offset, struct flk_callback *flk_cbp, cred_t *cr)
4287 {
4288 	netobj lm_fh;
4289 	int rc;
4290 	u_offset_t start, end;
4291 	rnode_t *rp;
4292 	int error = 0, intr = INTR(vp);
4293 
4294 	/* check for valid cmd parameter */
4295 	if (cmd != F_GETLK && cmd != F_SETLK && cmd != F_SETLKW)
4296 		return (EINVAL);
4297 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4298 		return (EIO);
4299 
4300 	/* Verify l_type. */
4301 	switch (bfp->l_type) {
4302 	case F_RDLCK:
4303 		if (cmd != F_GETLK && !(flag & FREAD))
4304 			return (EBADF);
4305 		break;
4306 	case F_WRLCK:
4307 		if (cmd != F_GETLK && !(flag & FWRITE))
4308 			return (EBADF);
4309 		break;
4310 	case F_UNLCK:
4311 		intr = 0;
4312 		break;
4313 
4314 	default:
4315 		return (EINVAL);
4316 	}
4317 
4318 	/* check the validity of the lock range */
4319 	if (rc = flk_convert_lock_data(vp, bfp, &start, &end, offset))
4320 		return (rc);
4321 	if (rc = flk_check_lock_data(start, end, MAXOFF32_T))
4322 		return (rc);
4323 
4324 	/*
4325 	 * If the filesystem is mounted using local locking, pass the
4326 	 * request off to the local locking code.
4327 	 */
4328 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
4329 		if (offset > MAXOFF32_T)
4330 			return (EFBIG);
4331 		if (cmd == F_SETLK || cmd == F_SETLKW) {
4332 			/*
4333 			 * For complete safety, we should be holding
4334 			 * r_lkserlock.  However, we can't call
4335 			 * lm_safelock and then fs_frlock while
4336 			 * holding r_lkserlock, so just invoke
4337 			 * lm_safelock and expect that this will
4338 			 * catch enough of the cases.
4339 			 */
4340 			if (!lm_safelock(vp, bfp, cr))
4341 				return (EAGAIN);
4342 		}
4343 		return (fs_frlock(vp, cmd, bfp, flag, offset, flk_cbp, cr));
4344 	}
4345 
4346 	rp = VTOR(vp);
4347 
4348 	/*
4349 	 * Check whether the given lock request can proceed, given the
4350 	 * current file mappings.
4351 	 */
4352 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_WRITER, intr))
4353 		return (EINTR);
4354 	if (cmd == F_SETLK || cmd == F_SETLKW) {
4355 		if (!lm_safelock(vp, bfp, cr)) {
4356 			rc = EAGAIN;
4357 			goto done;
4358 		}
4359 	}
4360 
4361 	/*
4362 	 * Flush the cache after waiting for async I/O to finish.  For new
4363 	 * locks, this is so that the process gets the latest bits from the
4364 	 * server.  For unlocks, this is so that other clients see the
4365 	 * latest bits once the file has been unlocked.  If currently dirty
4366 	 * pages can't be flushed, then don't allow a lock to be set.  But
4367 	 * allow unlocks to succeed, to avoid having orphan locks on the
4368 	 * server.
4369 	 */
4370 	if (cmd != F_GETLK) {
4371 		mutex_enter(&rp->r_statelock);
4372 		while (rp->r_count > 0) {
4373 		    if (intr) {
4374 			klwp_t *lwp = ttolwp(curthread);
4375 
4376 			if (lwp != NULL)
4377 				lwp->lwp_nostop++;
4378 			if (cv_wait_sig(&rp->r_cv, &rp->r_statelock) == 0) {
4379 				if (lwp != NULL)
4380 					lwp->lwp_nostop--;
4381 				rc = EINTR;
4382 				break;
4383 			}
4384 			if (lwp != NULL)
4385 				lwp->lwp_nostop--;
4386 		    } else
4387 			cv_wait(&rp->r_cv, &rp->r_statelock);
4388 		}
4389 		mutex_exit(&rp->r_statelock);
4390 		if (rc != 0)
4391 			goto done;
4392 		error = nfs_putpage(vp, (offset_t)0, 0, B_INVAL, cr);
4393 		if (error) {
4394 			if (error == ENOSPC || error == EDQUOT) {
4395 				mutex_enter(&rp->r_statelock);
4396 				if (!rp->r_error)
4397 					rp->r_error = error;
4398 				mutex_exit(&rp->r_statelock);
4399 			}
4400 			if (bfp->l_type != F_UNLCK) {
4401 				rc = ENOLCK;
4402 				goto done;
4403 			}
4404 		}
4405 	}
4406 
4407 	lm_fh.n_len = sizeof (fhandle_t);
4408 	lm_fh.n_bytes = (char *)VTOFH(vp);
4409 
4410 	/*
4411 	 * Call the lock manager to do the real work of contacting
4412 	 * the server and obtaining the lock.
4413 	 */
4414 	rc = lm_frlock(vp, cmd, bfp, flag, offset, cr, &lm_fh, flk_cbp);
4415 
4416 	if (rc == 0)
4417 		nfs_lockcompletion(vp, cmd);
4418 
4419 done:
4420 	nfs_rw_exit(&rp->r_lkserlock);
4421 	return (rc);
4422 }
4423 
4424 /*
4425  * Free storage space associated with the specified vnode.  The portion
4426  * to be freed is specified by bfp->l_start and bfp->l_len (already
4427  * normalized to a "whence" of 0).
4428  *
4429  * This is an experimental facility whose continued existence is not
4430  * guaranteed.  Currently, we only support the special case
4431  * of l_len == 0, meaning free to end of file.
4432  */
4433 /* ARGSUSED */
4434 static int
4435 nfs_space(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
4436 	offset_t offset, cred_t *cr, caller_context_t *ct)
4437 {
4438 	int error;
4439 
4440 	ASSERT(vp->v_type == VREG);
4441 	if (cmd != F_FREESP)
4442 		return (EINVAL);
4443 
4444 	if (offset > MAXOFF32_T)
4445 		return (EFBIG);
4446 
4447 	if ((bfp->l_start > MAXOFF32_T) || (bfp->l_end > MAXOFF32_T) ||
4448 	    (bfp->l_len > MAXOFF32_T))
4449 		return (EFBIG);
4450 
4451 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4452 		return (EIO);
4453 
4454 	error = convoff(vp, bfp, 0, offset);
4455 	if (!error) {
4456 		ASSERT(bfp->l_start >= 0);
4457 		if (bfp->l_len == 0) {
4458 			struct vattr va;
4459 
4460 			/*
4461 			 * ftruncate should not change the ctime and
4462 			 * mtime if we truncate the file to its
4463 			 * previous size.
4464 			 */
4465 			va.va_mask = AT_SIZE;
4466 			error = nfsgetattr(vp, &va, cr);
4467 			if (error || va.va_size == bfp->l_start)
4468 				return (error);
4469 			va.va_mask = AT_SIZE;
4470 			va.va_size = bfp->l_start;
4471 			error = nfssetattr(vp, &va, 0, cr);
4472 		} else
4473 			error = EINVAL;
4474 	}
4475 
4476 	return (error);
4477 }
4478 
4479 /* ARGSUSED */
4480 static int
4481 nfs_realvp(vnode_t *vp, vnode_t **vpp)
4482 {
4483 
4484 	return (EINVAL);
4485 }
4486 
4487 /*
4488  * Setup and add an address space callback to do the work of the delmap call.
4489  * The callback will (and must be) deleted in the actual callback function.
4490  *
4491  * This is done in order to take care of the problem that we have with holding
4492  * the address space's a_lock for a long period of time (e.g. if the NFS server
4493  * is down).  Callbacks will be executed in the address space code while the
4494  * a_lock is not held.	Holding the address space's a_lock causes things such
4495  * as ps and fork to hang because they are trying to acquire this lock as well.
4496  */
4497 /* ARGSUSED */
4498 static int
4499 nfs_delmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
4500 	size_t len, uint_t prot, uint_t maxprot, uint_t flags, cred_t *cr)
4501 {
4502 	int			caller_found;
4503 	int			error;
4504 	rnode_t			*rp;
4505 	nfs_delmap_args_t	*dmapp;
4506 	nfs_delmapcall_t	*delmap_call;
4507 
4508 	if (vp->v_flag & VNOMAP)
4509 		return (ENOSYS);
4510 	/*
4511 	 * A process may not change zones if it has NFS pages mmap'ed
4512 	 * in, so we can't legitimately get here from the wrong zone.
4513 	 */
4514 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
4515 
4516 	rp = VTOR(vp);
4517 
4518 	/*
4519 	 * The way that the address space of this process deletes its mapping
4520 	 * of this file is via the following call chains:
4521 	 * - as_free()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs_delmap()
4522 	 * - as_unmap()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs_delmap()
4523 	 *
4524 	 * With the use of address space callbacks we are allowed to drop the
4525 	 * address space lock, a_lock, while executing the NFS operations that
4526 	 * need to go over the wire.  Returning EAGAIN to the caller of this
4527 	 * function is what drives the execution of the callback that we add
4528 	 * below.  The callback will be executed by the address space code
4529 	 * after dropping the a_lock.  When the callback is finished, since
4530 	 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
4531 	 * is called again on the same segment to finish the rest of the work
4532 	 * that needs to happen during unmapping.
4533 	 *
4534 	 * This action of calling back into the segment driver causes
4535 	 * nfs_delmap() to get called again, but since the callback was
4536 	 * already executed at this point, it already did the work and there
4537 	 * is nothing left for us to do.
4538 	 *
4539 	 * To Summarize:
4540 	 * - The first time nfs_delmap is called by the current thread is when
4541 	 * we add the caller associated with this delmap to the delmap caller
4542 	 * list, add the callback, and return EAGAIN.
4543 	 * - The second time in this call chain when nfs_delmap is called we
4544 	 * will find this caller in the delmap caller list and realize there
4545 	 * is no more work to do thus removing this caller from the list and
4546 	 * returning the error that was set in the callback execution.
4547 	 */
4548 	caller_found = nfs_find_and_delete_delmapcall(rp, &error);
4549 	if (caller_found) {
4550 		/*
4551 		 * 'error' is from the actual delmap operations.  To avoid
4552 		 * hangs, we need to handle the return of EAGAIN differently
4553 		 * since this is what drives the callback execution.
4554 		 * In this case, we don't want to return EAGAIN and do the
4555 		 * callback execution because there are none to execute.
4556 		 */
4557 		if (error == EAGAIN)
4558 			return (0);
4559 		else
4560 			return (error);
4561 	}
4562 
4563 	/* current caller was not in the list */
4564 	delmap_call = nfs_init_delmapcall();
4565 
4566 	mutex_enter(&rp->r_statelock);
4567 	list_insert_tail(&rp->r_indelmap, delmap_call);
4568 	mutex_exit(&rp->r_statelock);
4569 
4570 	dmapp = kmem_alloc(sizeof (nfs_delmap_args_t), KM_SLEEP);
4571 
4572 	dmapp->vp = vp;
4573 	dmapp->off = off;
4574 	dmapp->addr = addr;
4575 	dmapp->len = len;
4576 	dmapp->prot = prot;
4577 	dmapp->maxprot = maxprot;
4578 	dmapp->flags = flags;
4579 	dmapp->cr = cr;
4580 	dmapp->caller = delmap_call;
4581 
4582 	error = as_add_callback(as, nfs_delmap_callback, dmapp,
4583 	    AS_UNMAP_EVENT, addr, len, KM_SLEEP);
4584 
4585 	return (error ? error : EAGAIN);
4586 }
4587 
4588 /*
4589  * Remove some pages from an mmap'd vnode.  Just update the
4590  * count of pages.  If doing close-to-open, then flush all
4591  * of the pages associated with this file.  Otherwise, start
4592  * an asynchronous page flush to write out any dirty pages.
4593  * This will also associate a credential with the rnode which
4594  * can be used to write the pages.
4595  */
4596 /* ARGSUSED */
4597 static void
4598 nfs_delmap_callback(struct as *as, void *arg, uint_t event)
4599 {
4600 	int			error;
4601 	rnode_t			*rp;
4602 	mntinfo_t		*mi;
4603 	nfs_delmap_args_t	*dmapp = (nfs_delmap_args_t *)arg;
4604 
4605 	rp = VTOR(dmapp->vp);
4606 	mi = VTOMI(dmapp->vp);
4607 
4608 	atomic_add_long((ulong_t *)&rp->r_mapcnt, -btopr(dmapp->len));
4609 	ASSERT(rp->r_mapcnt >= 0);
4610 
4611 	/*
4612 	 * Initiate a page flush if there are pages, the file system
4613 	 * was not mounted readonly, the segment was mapped shared, and
4614 	 * the pages themselves were writeable.
4615 	 */
4616 	if (vn_has_cached_data(dmapp->vp) && !vn_is_readonly(dmapp->vp) &&
4617 	    dmapp->flags == MAP_SHARED && (dmapp->maxprot & PROT_WRITE)) {
4618 		mutex_enter(&rp->r_statelock);
4619 		rp->r_flags |= RDIRTY;
4620 		mutex_exit(&rp->r_statelock);
4621 		/*
4622 		 * If this is a cross-zone access a sync putpage won't work, so
4623 		 * the best we can do is try an async putpage.  That seems
4624 		 * better than something more draconian such as discarding the
4625 		 * dirty pages.
4626 		 */
4627 		if ((mi->mi_flags & MI_NOCTO) ||
4628 		    nfs_zone() != mi->mi_zone)
4629 			error = nfs_putpage(dmapp->vp, dmapp->off, dmapp->len,
4630 			    B_ASYNC, dmapp->cr);
4631 		else
4632 			error = nfs_putpage(dmapp->vp, dmapp->off, dmapp->len,
4633 			    0, dmapp->cr);
4634 		if (!error) {
4635 			mutex_enter(&rp->r_statelock);
4636 			error = rp->r_error;
4637 			rp->r_error = 0;
4638 			mutex_exit(&rp->r_statelock);
4639 		}
4640 	} else
4641 		error = 0;
4642 
4643 	if ((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO))
4644 		(void) nfs_putpage(dmapp->vp, dmapp->off, dmapp->len,
4645 		    B_INVAL, dmapp->cr);
4646 
4647 	dmapp->caller->error = error;
4648 	(void) as_delete_callback(as, arg);
4649 	kmem_free(dmapp, sizeof (nfs_delmap_args_t));
4650 }
4651 
4652 /* ARGSUSED */
4653 static int
4654 nfs_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr)
4655 {
4656 	int error = 0;
4657 
4658 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4659 		return (EIO);
4660 	/*
4661 	 * This looks a little weird because it's written in a general
4662 	 * manner but we make little use of cases.  If cntl() ever gets
4663 	 * widely used, the outer switch will make more sense.
4664 	 */
4665 
4666 	switch (cmd) {
4667 
4668 	/*
4669 	 * Large file spec - need to base answer new query with
4670 	 * hardcoded constant based on the protocol.
4671 	 */
4672 	case _PC_FILESIZEBITS:
4673 		*valp = 32;
4674 		return (0);
4675 
4676 	case _PC_LINK_MAX:
4677 	case _PC_NAME_MAX:
4678 	case _PC_PATH_MAX:
4679 	case _PC_SYMLINK_MAX:
4680 	case _PC_CHOWN_RESTRICTED:
4681 	case _PC_NO_TRUNC: {
4682 		mntinfo_t *mi;
4683 		struct pathcnf *pc;
4684 
4685 		if ((mi = VTOMI(vp)) == NULL || (pc = mi->mi_pathconf) == NULL)
4686 			return (EINVAL);
4687 		error = _PC_ISSET(cmd, pc->pc_mask);    /* error or bool */
4688 		switch (cmd) {
4689 		case _PC_LINK_MAX:
4690 			*valp = pc->pc_link_max;
4691 			break;
4692 		case _PC_NAME_MAX:
4693 			*valp = pc->pc_name_max;
4694 			break;
4695 		case _PC_PATH_MAX:
4696 		case _PC_SYMLINK_MAX:
4697 			*valp = pc->pc_path_max;
4698 			break;
4699 		case _PC_CHOWN_RESTRICTED:
4700 			/*
4701 			 * if we got here, error is really a boolean which
4702 			 * indicates whether cmd is set or not.
4703 			 */
4704 			*valp = error ? 1 : 0;	/* see above */
4705 			error = 0;
4706 			break;
4707 		case _PC_NO_TRUNC:
4708 			/*
4709 			 * if we got here, error is really a boolean which
4710 			 * indicates whether cmd is set or not.
4711 			 */
4712 			*valp = error ? 1 : 0;	/* see above */
4713 			error = 0;
4714 			break;
4715 		}
4716 		return (error ? EINVAL : 0);
4717 	    }
4718 
4719 	case _PC_XATTR_EXISTS:
4720 		*valp = 0;
4721 		if (vp->v_vfsp->vfs_flag & VFS_XATTR) {
4722 			vnode_t *avp;
4723 			rnode_t *rp;
4724 			mntinfo_t *mi = VTOMI(vp);
4725 
4726 			if (!(mi->mi_flags & MI_EXTATTR))
4727 				return (0);
4728 
4729 			rp = VTOR(vp);
4730 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_READER,
4731 			    INTR(vp)))
4732 				return (EINTR);
4733 
4734 			error = nfslookup_dnlc(vp, XATTR_DIR_NAME, &avp, cr);
4735 			if (error || avp == NULL)
4736 				error = acl_getxattrdir2(vp, &avp, 0, cr, 0);
4737 
4738 			nfs_rw_exit(&rp->r_rwlock);
4739 
4740 			if (error == 0 && avp != NULL) {
4741 				VN_RELE(avp);
4742 				*valp = 1;
4743 			}
4744 		}
4745 		return (error ? EINVAL : 0);
4746 
4747 	case _PC_ACL_ENABLED:
4748 		*valp = _ACL_ACLENT_ENABLED;
4749 		return (0);
4750 
4751 	default:
4752 		return (EINVAL);
4753 	}
4754 }
4755 
4756 /*
4757  * Called by async thread to do synchronous pageio. Do the i/o, wait
4758  * for it to complete, and cleanup the page list when done.
4759  */
4760 static int
4761 nfs_sync_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
4762 	int flags, cred_t *cr)
4763 {
4764 	int error;
4765 
4766 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
4767 	error = nfs_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
4768 	if (flags & B_READ)
4769 		pvn_read_done(pp, (error ? B_ERROR : 0) | flags);
4770 	else
4771 		pvn_write_done(pp, (error ? B_ERROR : 0) | flags);
4772 	return (error);
4773 }
4774 
4775 static int
4776 nfs_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
4777 	int flags, cred_t *cr)
4778 {
4779 	int error;
4780 	rnode_t *rp;
4781 
4782 	if (pp == NULL)
4783 		return (EINVAL);
4784 
4785 	if (io_off > MAXOFF32_T)
4786 		return (EFBIG);
4787 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4788 		return (EIO);
4789 	rp = VTOR(vp);
4790 	mutex_enter(&rp->r_statelock);
4791 	rp->r_count++;
4792 	mutex_exit(&rp->r_statelock);
4793 
4794 	if (flags & B_ASYNC) {
4795 		error = nfs_async_pageio(vp, pp, io_off, io_len, flags, cr,
4796 		    nfs_sync_pageio);
4797 	} else
4798 		error = nfs_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
4799 	mutex_enter(&rp->r_statelock);
4800 	rp->r_count--;
4801 	cv_broadcast(&rp->r_cv);
4802 	mutex_exit(&rp->r_statelock);
4803 	return (error);
4804 }
4805 
4806 static int
4807 nfs_setsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
4808 {
4809 	int error;
4810 	mntinfo_t *mi;
4811 
4812 	mi = VTOMI(vp);
4813 
4814 	if (nfs_zone() != mi->mi_zone)
4815 		return (EIO);
4816 	if (mi->mi_flags & MI_ACL) {
4817 		error = acl_setacl2(vp, vsecattr, flag, cr);
4818 		if (mi->mi_flags & MI_ACL)
4819 			return (error);
4820 	}
4821 
4822 	return (ENOSYS);
4823 }
4824 
4825 static int
4826 nfs_getsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
4827 {
4828 	int error;
4829 	mntinfo_t *mi;
4830 
4831 	mi = VTOMI(vp);
4832 
4833 	if (nfs_zone() != mi->mi_zone)
4834 		return (EIO);
4835 	if (mi->mi_flags & MI_ACL) {
4836 		error = acl_getacl2(vp, vsecattr, flag, cr);
4837 		if (mi->mi_flags & MI_ACL)
4838 			return (error);
4839 	}
4840 
4841 	return (fs_fab_acl(vp, vsecattr, flag, cr));
4842 }
4843 
4844 static int
4845 nfs_shrlock(vnode_t *vp, int cmd, struct shrlock *shr, int flag, cred_t *cr)
4846 {
4847 	int error;
4848 	struct shrlock nshr;
4849 	struct nfs_owner nfs_owner;
4850 	netobj lm_fh;
4851 
4852 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4853 		return (EIO);
4854 
4855 	/*
4856 	 * check for valid cmd parameter
4857 	 */
4858 	if (cmd != F_SHARE && cmd != F_UNSHARE && cmd != F_HASREMOTELOCKS)
4859 		return (EINVAL);
4860 
4861 	/*
4862 	 * Check access permissions
4863 	 */
4864 	if (cmd == F_SHARE &&
4865 	    (((shr->s_access & F_RDACC) && !(flag & FREAD)) ||
4866 	    ((shr->s_access & F_WRACC) && !(flag & FWRITE))))
4867 		return (EBADF);
4868 
4869 	/*
4870 	 * If the filesystem is mounted using local locking, pass the
4871 	 * request off to the local share code.
4872 	 */
4873 	if (VTOMI(vp)->mi_flags & MI_LLOCK)
4874 		return (fs_shrlock(vp, cmd, shr, flag, cr));
4875 
4876 	switch (cmd) {
4877 	case F_SHARE:
4878 	case F_UNSHARE:
4879 		lm_fh.n_len = sizeof (fhandle_t);
4880 		lm_fh.n_bytes = (char *)VTOFH(vp);
4881 
4882 		/*
4883 		 * If passed an owner that is too large to fit in an
4884 		 * nfs_owner it is likely a recursive call from the
4885 		 * lock manager client and pass it straight through.  If
4886 		 * it is not a nfs_owner then simply return an error.
4887 		 */
4888 		if (shr->s_own_len > sizeof (nfs_owner.lowner)) {
4889 			if (((struct nfs_owner *)shr->s_owner)->magic !=
4890 			    NFS_OWNER_MAGIC)
4891 				return (EINVAL);
4892 
4893 			if (error = lm_shrlock(vp, cmd, shr, flag, &lm_fh)) {
4894 				error = set_errno(error);
4895 			}
4896 			return (error);
4897 		}
4898 		/*
4899 		 * Remote share reservations owner is a combination of
4900 		 * a magic number, hostname, and the local owner
4901 		 */
4902 		bzero(&nfs_owner, sizeof (nfs_owner));
4903 		nfs_owner.magic = NFS_OWNER_MAGIC;
4904 		(void) strncpy(nfs_owner.hname, uts_nodename(),
4905 		    sizeof (nfs_owner.hname));
4906 		bcopy(shr->s_owner, nfs_owner.lowner, shr->s_own_len);
4907 		nshr.s_access = shr->s_access;
4908 		nshr.s_deny = shr->s_deny;
4909 		nshr.s_sysid = 0;
4910 		nshr.s_pid = ttoproc(curthread)->p_pid;
4911 		nshr.s_own_len = sizeof (nfs_owner);
4912 		nshr.s_owner = (caddr_t)&nfs_owner;
4913 
4914 		if (error = lm_shrlock(vp, cmd, &nshr, flag, &lm_fh)) {
4915 			error = set_errno(error);
4916 		}
4917 
4918 		break;
4919 
4920 	case F_HASREMOTELOCKS:
4921 		/*
4922 		 * NFS client can't store remote locks itself
4923 		 */
4924 		shr->s_access = 0;
4925 		error = 0;
4926 		break;
4927 
4928 	default:
4929 		error = EINVAL;
4930 		break;
4931 	}
4932 
4933 	return (error);
4934 }
4935