xref: /titanic_41/usr/src/uts/common/fs/nfs/nfs3_vnops.c (revision 3893cb7fe5bfa1c9a4f7954517a917367f6cf081)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  *	Copyright (c) 1983,1984,1985,1986,1987,1988,1989 AT&T.
28  *	All rights reserved.
29  */
30 
31 #pragma ident	"%Z%%M%	%I%	%E% SMI"
32 
33 #include <sys/param.h>
34 #include <sys/types.h>
35 #include <sys/systm.h>
36 #include <sys/cred.h>
37 #include <sys/time.h>
38 #include <sys/vnode.h>
39 #include <sys/vfs.h>
40 #include <sys/vfs_opreg.h>
41 #include <sys/file.h>
42 #include <sys/filio.h>
43 #include <sys/uio.h>
44 #include <sys/buf.h>
45 #include <sys/mman.h>
46 #include <sys/pathname.h>
47 #include <sys/dirent.h>
48 #include <sys/debug.h>
49 #include <sys/vmsystm.h>
50 #include <sys/fcntl.h>
51 #include <sys/flock.h>
52 #include <sys/swap.h>
53 #include <sys/errno.h>
54 #include <sys/strsubr.h>
55 #include <sys/sysmacros.h>
56 #include <sys/kmem.h>
57 #include <sys/cmn_err.h>
58 #include <sys/pathconf.h>
59 #include <sys/utsname.h>
60 #include <sys/dnlc.h>
61 #include <sys/acl.h>
62 #include <sys/systeminfo.h>
63 #include <sys/atomic.h>
64 #include <sys/policy.h>
65 #include <sys/sdt.h>
66 
67 #include <rpc/types.h>
68 #include <rpc/auth.h>
69 #include <rpc/clnt.h>
70 
71 #include <nfs/nfs.h>
72 #include <nfs/nfs_clnt.h>
73 #include <nfs/rnode.h>
74 #include <nfs/nfs_acl.h>
75 #include <nfs/lm.h>
76 
77 #include <vm/hat.h>
78 #include <vm/as.h>
79 #include <vm/page.h>
80 #include <vm/pvn.h>
81 #include <vm/seg.h>
82 #include <vm/seg_map.h>
83 #include <vm/seg_kpm.h>
84 #include <vm/seg_vn.h>
85 
86 #include <fs/fs_subr.h>
87 
88 #include <sys/ddi.h>
89 
90 static int	nfs3_rdwrlbn(vnode_t *, page_t *, u_offset_t, size_t, int,
91 			cred_t *);
92 static int	nfs3write(vnode_t *, caddr_t, u_offset_t, int, cred_t *,
93 			stable_how *);
94 static int	nfs3read(vnode_t *, caddr_t, offset_t, int, size_t *, cred_t *);
95 static int	nfs3setattr(vnode_t *, struct vattr *, int, cred_t *);
96 static int	nfs3_accessx(void *, int, cred_t *);
97 static int	nfs3lookup_dnlc(vnode_t *, char *, vnode_t **, cred_t *);
98 static int	nfs3lookup_otw(vnode_t *, char *, vnode_t **, cred_t *, int);
99 static int	nfs3create(vnode_t *, char *, struct vattr *, enum vcexcl,
100 			int, vnode_t **, cred_t *, int);
101 static int	nfs3excl_create_settimes(vnode_t *, struct vattr *, cred_t *);
102 static int	nfs3mknod(vnode_t *, char *, struct vattr *, enum vcexcl,
103 			int, vnode_t **, cred_t *);
104 static int	nfs3rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
105 static int	do_nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
106 static void	nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
107 static void	nfs3readdirplus(vnode_t *, rddir_cache *, cred_t *);
108 static int	nfs3_bio(struct buf *, stable_how *, cred_t *);
109 static int	nfs3_getapage(vnode_t *, u_offset_t, size_t, uint_t *,
110 			page_t *[], size_t, struct seg *, caddr_t,
111 			enum seg_rw, cred_t *);
112 static void	nfs3_readahead(vnode_t *, u_offset_t, caddr_t, struct seg *,
113 			cred_t *);
114 static int	nfs3_sync_putapage(vnode_t *, page_t *, u_offset_t, size_t,
115 			int, cred_t *);
116 static int	nfs3_sync_pageio(vnode_t *, page_t *, u_offset_t, size_t,
117 			int, cred_t *);
118 static int	nfs3_commit(vnode_t *, offset3, count3, cred_t *);
119 static void	nfs3_set_mod(vnode_t *);
120 static void	nfs3_get_commit(vnode_t *);
121 static void	nfs3_get_commit_range(vnode_t *, u_offset_t, size_t);
122 #if 0 /* unused */
123 #ifdef DEBUG
124 static int	nfs3_no_uncommitted_pages(vnode_t *);
125 #endif
126 #endif /* unused */
127 static int	nfs3_putpage_commit(vnode_t *, offset_t, size_t, cred_t *);
128 static int	nfs3_commit_vp(vnode_t *, u_offset_t, size_t,  cred_t *);
129 static int	nfs3_sync_commit(vnode_t *, page_t *, offset3, count3,
130 			cred_t *);
131 static void	nfs3_async_commit(vnode_t *, page_t *, offset3, count3,
132 			cred_t *);
133 static void	nfs3_delmap_callback(struct as *, void *, uint_t);
134 
135 /*
136  * Error flags used to pass information about certain special errors
137  * which need to be handled specially.
138  */
139 #define	NFS_EOF			-98
140 #define	NFS_VERF_MISMATCH	-97
141 
142 /* ALIGN64 aligns the given buffer and adjust buffer size to 64 bit */
143 #define	ALIGN64(x, ptr, sz)						\
144 	x = ((uintptr_t)(ptr)) & (sizeof (uint64_t) - 1);		\
145 	if (x) {							\
146 		x = sizeof (uint64_t) - (x);				\
147 		sz -= (x);						\
148 		ptr += (x);						\
149 	}
150 
151 /*
152  * These are the vnode ops routines which implement the vnode interface to
153  * the networked file system.  These routines just take their parameters,
154  * make them look networkish by putting the right info into interface structs,
155  * and then calling the appropriate remote routine(s) to do the work.
156  *
157  * Note on directory name lookup cacheing:  If we detect a stale fhandle,
158  * we purge the directory cache relative to that vnode.  This way, the
159  * user won't get burned by the cache repeatedly.  See <nfs/rnode.h> for
160  * more details on rnode locking.
161  */
162 
163 static int	nfs3_open(vnode_t **, int, cred_t *);
164 static int	nfs3_close(vnode_t *, int, int, offset_t, cred_t *);
165 static int	nfs3_read(vnode_t *, struct uio *, int, cred_t *,
166 			caller_context_t *);
167 static int	nfs3_write(vnode_t *, struct uio *, int, cred_t *,
168 			caller_context_t *);
169 static int	nfs3_ioctl(vnode_t *, int, intptr_t, int, cred_t *, int *);
170 static int	nfs3_getattr(vnode_t *, struct vattr *, int, cred_t *);
171 static int	nfs3_setattr(vnode_t *, struct vattr *, int, cred_t *,
172 			caller_context_t *);
173 static int	nfs3_access(vnode_t *, int, int, cred_t *);
174 static int	nfs3_readlink(vnode_t *, struct uio *, cred_t *);
175 static int	nfs3_fsync(vnode_t *, int, cred_t *);
176 static void	nfs3_inactive(vnode_t *, cred_t *);
177 static int	nfs3_lookup(vnode_t *, char *, vnode_t **,
178 			struct pathname *, int, vnode_t *, cred_t *);
179 static int	nfs3_create(vnode_t *, char *, struct vattr *, enum vcexcl,
180 			int, vnode_t **, cred_t *, int);
181 static int	nfs3_remove(vnode_t *, char *, cred_t *);
182 static int	nfs3_link(vnode_t *, vnode_t *, char *, cred_t *);
183 static int	nfs3_rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
184 static int	nfs3_mkdir(vnode_t *, char *, struct vattr *,
185 			vnode_t **, cred_t *);
186 static int	nfs3_rmdir(vnode_t *, char *, vnode_t *, cred_t *);
187 static int	nfs3_symlink(vnode_t *, char *, struct vattr *, char *,
188 			cred_t *);
189 static int	nfs3_readdir(vnode_t *, struct uio *, cred_t *, int *);
190 static int	nfs3_fid(vnode_t *, fid_t *);
191 static int	nfs3_rwlock(vnode_t *, int, caller_context_t *);
192 static void	nfs3_rwunlock(vnode_t *, int, caller_context_t *);
193 static int	nfs3_seek(vnode_t *, offset_t, offset_t *);
194 static int	nfs3_getpage(vnode_t *, offset_t, size_t, uint_t *,
195 			page_t *[], size_t, struct seg *, caddr_t,
196 			enum seg_rw, cred_t *);
197 static int	nfs3_putpage(vnode_t *, offset_t, size_t, int, cred_t *);
198 static int	nfs3_map(vnode_t *, offset_t, struct as *, caddr_t *,
199 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
200 static int	nfs3_addmap(vnode_t *, offset_t, struct as *, caddr_t,
201 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
202 static int	nfs3_frlock(vnode_t *, int, struct flock64 *, int, offset_t,
203 			struct flk_callback *, cred_t *);
204 static int	nfs3_space(vnode_t *, int, struct flock64 *, int, offset_t,
205 			cred_t *, caller_context_t *);
206 static int	nfs3_realvp(vnode_t *, vnode_t **);
207 static int	nfs3_delmap(vnode_t *, offset_t, struct as *, caddr_t,
208 			size_t, uint_t, uint_t, uint_t, cred_t *);
209 static int	nfs3_pathconf(vnode_t *, int, ulong_t *, cred_t *);
210 static int	nfs3_pageio(vnode_t *, page_t *, u_offset_t, size_t, int,
211 			cred_t *);
212 static void	nfs3_dispose(vnode_t *, page_t *, int, int, cred_t *);
213 static int	nfs3_setsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
214 static int	nfs3_getsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
215 static int	nfs3_shrlock(vnode_t *, int, struct shrlock *, int, cred_t *);
216 
217 struct vnodeops *nfs3_vnodeops;
218 
219 const fs_operation_def_t nfs3_vnodeops_template[] = {
220 	VOPNAME_OPEN,		{ .vop_open = nfs3_open },
221 	VOPNAME_CLOSE,		{ .vop_close = nfs3_close },
222 	VOPNAME_READ,		{ .vop_read = nfs3_read },
223 	VOPNAME_WRITE,		{ .vop_write = nfs3_write },
224 	VOPNAME_IOCTL,		{ .vop_ioctl = nfs3_ioctl },
225 	VOPNAME_GETATTR,	{ .vop_getattr = nfs3_getattr },
226 	VOPNAME_SETATTR,	{ .vop_setattr = nfs3_setattr },
227 	VOPNAME_ACCESS,		{ .vop_access = nfs3_access },
228 	VOPNAME_LOOKUP,		{ .vop_lookup = nfs3_lookup },
229 	VOPNAME_CREATE,		{ .vop_create = nfs3_create },
230 	VOPNAME_REMOVE,		{ .vop_remove = nfs3_remove },
231 	VOPNAME_LINK,		{ .vop_link = nfs3_link },
232 	VOPNAME_RENAME,		{ .vop_rename = nfs3_rename },
233 	VOPNAME_MKDIR,		{ .vop_mkdir = nfs3_mkdir },
234 	VOPNAME_RMDIR,		{ .vop_rmdir = nfs3_rmdir },
235 	VOPNAME_READDIR,	{ .vop_readdir = nfs3_readdir },
236 	VOPNAME_SYMLINK,	{ .vop_symlink = nfs3_symlink },
237 	VOPNAME_READLINK,	{ .vop_readlink = nfs3_readlink },
238 	VOPNAME_FSYNC,		{ .vop_fsync = nfs3_fsync },
239 	VOPNAME_INACTIVE,	{ .vop_inactive = nfs3_inactive },
240 	VOPNAME_FID,		{ .vop_fid = nfs3_fid },
241 	VOPNAME_RWLOCK,		{ .vop_rwlock = nfs3_rwlock },
242 	VOPNAME_RWUNLOCK,	{ .vop_rwunlock = nfs3_rwunlock },
243 	VOPNAME_SEEK,		{ .vop_seek = nfs3_seek },
244 	VOPNAME_FRLOCK,		{ .vop_frlock = nfs3_frlock },
245 	VOPNAME_SPACE,		{ .vop_space = nfs3_space },
246 	VOPNAME_REALVP,		{ .vop_realvp = nfs3_realvp },
247 	VOPNAME_GETPAGE,	{ .vop_getpage = nfs3_getpage },
248 	VOPNAME_PUTPAGE,	{ .vop_putpage = nfs3_putpage },
249 	VOPNAME_MAP,		{ .vop_map = nfs3_map },
250 	VOPNAME_ADDMAP,		{ .vop_addmap = nfs3_addmap },
251 	VOPNAME_DELMAP,		{ .vop_delmap = nfs3_delmap },
252 	/* no separate nfs3_dump */
253 	VOPNAME_DUMP,		{ .vop_dump = nfs_dump },
254 	VOPNAME_PATHCONF,	{ .vop_pathconf = nfs3_pathconf },
255 	VOPNAME_PAGEIO,		{ .vop_pageio = nfs3_pageio },
256 	VOPNAME_DISPOSE,	{ .vop_dispose = nfs3_dispose },
257 	VOPNAME_SETSECATTR,	{ .vop_setsecattr = nfs3_setsecattr },
258 	VOPNAME_GETSECATTR,	{ .vop_getsecattr = nfs3_getsecattr },
259 	VOPNAME_SHRLOCK,	{ .vop_shrlock = nfs3_shrlock },
260 	NULL,			NULL
261 };
262 
263 /*
264  * XXX:  This is referenced in modstubs.s
265  */
266 struct vnodeops *
267 nfs3_getvnodeops(void)
268 {
269 	return (nfs3_vnodeops);
270 }
271 
272 /* ARGSUSED */
273 static int
274 nfs3_open(vnode_t **vpp, int flag, cred_t *cr)
275 {
276 	int error;
277 	struct vattr va;
278 	rnode_t *rp;
279 	vnode_t *vp;
280 
281 	vp = *vpp;
282 	if (nfs_zone() != VTOMI(vp)->mi_zone)
283 		return (EIO);
284 	rp = VTOR(vp);
285 	mutex_enter(&rp->r_statelock);
286 	if (rp->r_cred == NULL) {
287 		crhold(cr);
288 		rp->r_cred = cr;
289 	}
290 	mutex_exit(&rp->r_statelock);
291 
292 	/*
293 	 * If there is no cached data or if close-to-open
294 	 * consistency checking is turned off, we can avoid
295 	 * the over the wire getattr.  Otherwise, if the
296 	 * file system is mounted readonly, then just verify
297 	 * the caches are up to date using the normal mechanism.
298 	 * Else, if the file is not mmap'd, then just mark
299 	 * the attributes as timed out.  They will be refreshed
300 	 * and the caches validated prior to being used.
301 	 * Else, the file system is mounted writeable so
302 	 * force an over the wire GETATTR in order to ensure
303 	 * that all cached data is valid.
304 	 */
305 	if (vp->v_count > 1 ||
306 	    ((vn_has_cached_data(vp) || HAVE_RDDIR_CACHE(rp)) &&
307 	    !(VTOMI(vp)->mi_flags & MI_NOCTO))) {
308 		if (vn_is_readonly(vp))
309 			error = nfs3_validate_caches(vp, cr);
310 		else if (rp->r_mapcnt == 0 && vp->v_count == 1) {
311 			PURGE_ATTRCACHE(vp);
312 			error = 0;
313 		} else {
314 			va.va_mask = AT_ALL;
315 			error = nfs3_getattr_otw(vp, &va, cr);
316 		}
317 	} else
318 		error = 0;
319 
320 	return (error);
321 }
322 
323 static int
324 nfs3_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr)
325 {
326 	rnode_t *rp;
327 	int error;
328 	struct vattr va;
329 
330 	/*
331 	 * zone_enter(2) prevents processes from changing zones with NFS files
332 	 * open; if we happen to get here from the wrong zone we can't do
333 	 * anything over the wire.
334 	 */
335 	if (VTOMI(vp)->mi_zone != nfs_zone()) {
336 		/*
337 		 * We could attempt to clean up locks, except we're sure
338 		 * that the current process didn't acquire any locks on
339 		 * the file: any attempt to lock a file belong to another zone
340 		 * will fail, and one can't lock an NFS file and then change
341 		 * zones, as that fails too.
342 		 *
343 		 * Returning an error here is the sane thing to do.  A
344 		 * subsequent call to VN_RELE() which translates to a
345 		 * nfs3_inactive() will clean up state: if the zone of the
346 		 * vnode's origin is still alive and kicking, an async worker
347 		 * thread will handle the request (from the correct zone), and
348 		 * everything (minus the commit and final nfs3_getattr_otw()
349 		 * call) should be OK. If the zone is going away
350 		 * nfs_async_inactive() will throw away cached pages inline.
351 		 */
352 		return (EIO);
353 	}
354 
355 	/*
356 	 * If we are using local locking for this filesystem, then
357 	 * release all of the SYSV style record locks.  Otherwise,
358 	 * we are doing network locking and we need to release all
359 	 * of the network locks.  All of the locks held by this
360 	 * process on this file are released no matter what the
361 	 * incoming reference count is.
362 	 */
363 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
364 		cleanlocks(vp, ttoproc(curthread)->p_pid, 0);
365 		cleanshares(vp, ttoproc(curthread)->p_pid);
366 	} else
367 		nfs_lockrelease(vp, flag, offset, cr);
368 
369 	if (count > 1)
370 		return (0);
371 
372 	/*
373 	 * If the file has been `unlinked', then purge the
374 	 * DNLC so that this vnode will get reycled quicker
375 	 * and the .nfs* file on the server will get removed.
376 	 */
377 	rp = VTOR(vp);
378 	if (rp->r_unldvp != NULL)
379 		dnlc_purge_vp(vp);
380 
381 	/*
382 	 * If the file was open for write and there are pages,
383 	 * then if the file system was mounted using the "no-close-
384 	 *	to-open" semantics, then start an asynchronous flush
385 	 *	of the all of the pages in the file.
386 	 * else the file system was not mounted using the "no-close-
387 	 *	to-open" semantics, then do a synchronous flush and
388 	 *	commit of all of the dirty and uncommitted pages.
389 	 *
390 	 * The asynchronous flush of the pages in the "nocto" path
391 	 * mostly just associates a cred pointer with the rnode so
392 	 * writes which happen later will have a better chance of
393 	 * working.  It also starts the data being written to the
394 	 * server, but without unnecessarily delaying the application.
395 	 */
396 	if ((flag & FWRITE) && vn_has_cached_data(vp)) {
397 		if (VTOMI(vp)->mi_flags & MI_NOCTO) {
398 			error = nfs3_putpage(vp, (offset_t)0, 0, B_ASYNC, cr);
399 			if (error == EAGAIN)
400 				error = 0;
401 		} else
402 			error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
403 		if (!error) {
404 			mutex_enter(&rp->r_statelock);
405 			error = rp->r_error;
406 			rp->r_error = 0;
407 			mutex_exit(&rp->r_statelock);
408 		}
409 	} else {
410 		mutex_enter(&rp->r_statelock);
411 		error = rp->r_error;
412 		rp->r_error = 0;
413 		mutex_exit(&rp->r_statelock);
414 	}
415 
416 	/*
417 	 * If RWRITEATTR is set, then issue an over the wire GETATTR to
418 	 * refresh the attribute cache with a set of attributes which
419 	 * weren't returned from a WRITE.  This will enable the close-
420 	 * to-open processing to work.
421 	 */
422 	if (rp->r_flags & RWRITEATTR)
423 		(void) nfs3_getattr_otw(vp, &va, cr);
424 
425 	return (error);
426 }
427 
428 /* ARGSUSED */
429 static int
430 nfs3_directio_read(vnode_t *vp, struct uio *uiop, cred_t *cr)
431 {
432 	mntinfo_t *mi;
433 	READ3args args;
434 	READ3uiores res;
435 	int tsize;
436 	offset_t offset;
437 	ssize_t count;
438 	int error;
439 	int douprintf;
440 	failinfo_t fi;
441 	char *sv_hostname;
442 
443 	mi = VTOMI(vp);
444 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
445 	sv_hostname = VTOR(vp)->r_server->sv_hostname;
446 
447 	douprintf = 1;
448 	args.file = *VTOFH3(vp);
449 	fi.vp = vp;
450 	fi.fhp = (caddr_t)&args.file;
451 	fi.copyproc = nfs3copyfh;
452 	fi.lookupproc = nfs3lookup;
453 	fi.xattrdirproc = acl_getxattrdir3;
454 
455 	res.uiop = uiop;
456 
457 	offset = uiop->uio_loffset;
458 	count = uiop->uio_resid;
459 
460 	do {
461 		if (mi->mi_io_kstats) {
462 			mutex_enter(&mi->mi_lock);
463 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
464 			mutex_exit(&mi->mi_lock);
465 		}
466 
467 		do {
468 			tsize = MIN(mi->mi_tsize, count);
469 			args.offset = (offset3)offset;
470 			args.count = (count3)tsize;
471 			res.size = (uint_t)tsize;
472 			error = rfs3call(mi, NFSPROC3_READ,
473 				    xdr_READ3args, (caddr_t)&args,
474 				    xdr_READ3uiores, (caddr_t)&res, cr,
475 				    &douprintf, &res.status, 0, &fi);
476 		} while (error == ENFS_TRYAGAIN);
477 
478 		if (mi->mi_io_kstats) {
479 			mutex_enter(&mi->mi_lock);
480 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
481 			mutex_exit(&mi->mi_lock);
482 		}
483 
484 		if (error)
485 			return (error);
486 
487 		error = geterrno3(res.status);
488 		if (error)
489 			return (error);
490 
491 		if (res.count != res.size) {
492 			zcmn_err(getzoneid(), CE_WARN,
493 "nfs3_directio_read: server %s returned incorrect amount",
494 					sv_hostname);
495 			return (EIO);
496 		}
497 		count -= res.count;
498 		offset += res.count;
499 		if (mi->mi_io_kstats) {
500 			mutex_enter(&mi->mi_lock);
501 			KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
502 			KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
503 			mutex_exit(&mi->mi_lock);
504 		}
505 		lwp_stat_update(LWP_STAT_INBLK, 1);
506 	} while (count && !res.eof);
507 
508 	return (0);
509 }
510 
511 /* ARGSUSED */
512 static int
513 nfs3_read(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
514 	caller_context_t *ct)
515 {
516 	rnode_t *rp;
517 	u_offset_t off;
518 	offset_t diff;
519 	int on;
520 	size_t n;
521 	caddr_t base;
522 	uint_t flags;
523 	int error = 0;
524 	mntinfo_t *mi;
525 
526 	rp = VTOR(vp);
527 	mi = VTOMI(vp);
528 
529 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
530 
531 	if (nfs_zone() != mi->mi_zone)
532 		return (EIO);
533 
534 	if (vp->v_type != VREG)
535 		return (EISDIR);
536 
537 	if (uiop->uio_resid == 0)
538 		return (0);
539 
540 	if (uiop->uio_loffset < 0 || uiop->uio_loffset + uiop->uio_resid < 0)
541 		return (EINVAL);
542 
543 	/*
544 	 * Bypass VM if caching has been disabled (e.g., locking) or if
545 	 * using client-side direct I/O and the file is not mmap'd and
546 	 * there are no cached pages.
547 	 */
548 	if ((vp->v_flag & VNOCACHE) ||
549 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
550 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
551 		return (nfs3_directio_read(vp, uiop, cr));
552 	}
553 
554 	do {
555 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
556 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
557 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
558 
559 		error = nfs3_validate_caches(vp, cr);
560 		if (error)
561 			break;
562 
563 		mutex_enter(&rp->r_statelock);
564 		diff = rp->r_size - uiop->uio_loffset;
565 		mutex_exit(&rp->r_statelock);
566 		if (diff <= 0)
567 			break;
568 		if (diff < n)
569 			n = (size_t)diff;
570 
571 		if (vpm_enable) {
572 			/*
573 			 * Copy data.
574 			 */
575 			error = vpm_data_copy(vp, off + on, n, uiop,
576 						1, NULL, 0, S_READ);
577 		} else {
578 			base = segmap_getmapflt(segkmap, vp, off + on, n, 1,
579 							S_READ);
580 
581 			error = uiomove(base + on, n, UIO_READ, uiop);
582 		}
583 
584 		if (!error) {
585 			/*
586 			 * If read a whole block or read to eof,
587 			 * won't need this buffer again soon.
588 			 */
589 			mutex_enter(&rp->r_statelock);
590 			if (n + on == MAXBSIZE ||
591 			    uiop->uio_loffset == rp->r_size)
592 				flags = SM_DONTNEED;
593 			else
594 				flags = 0;
595 			mutex_exit(&rp->r_statelock);
596 			if (vpm_enable) {
597 				error = vpm_sync_pages(vp, off, n, flags);
598 			} else {
599 				error = segmap_release(segkmap, base, flags);
600 			}
601 		} else {
602 			if (vpm_enable) {
603 				(void) vpm_sync_pages(vp, off, n, 0);
604 			} else {
605 				(void) segmap_release(segkmap, base, 0);
606 			}
607 		}
608 	} while (!error && uiop->uio_resid > 0);
609 
610 	return (error);
611 }
612 
613 /* ARGSUSED */
614 static int
615 nfs3_write(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
616 	caller_context_t *ct)
617 {
618 	rlim64_t limit = uiop->uio_llimit;
619 	rnode_t *rp;
620 	u_offset_t off;
621 	caddr_t base;
622 	uint_t flags;
623 	int remainder;
624 	size_t n;
625 	int on;
626 	int error;
627 	int resid;
628 	offset_t offset;
629 	mntinfo_t *mi;
630 	uint_t bsize;
631 
632 	rp = VTOR(vp);
633 
634 	if (vp->v_type != VREG)
635 		return (EISDIR);
636 
637 	mi = VTOMI(vp);
638 	if (nfs_zone() != mi->mi_zone)
639 		return (EIO);
640 	if (uiop->uio_resid == 0)
641 		return (0);
642 
643 	if (ioflag & FAPPEND) {
644 		struct vattr va;
645 
646 		/*
647 		 * Must serialize if appending.
648 		 */
649 		if (nfs_rw_lock_held(&rp->r_rwlock, RW_READER)) {
650 			nfs_rw_exit(&rp->r_rwlock);
651 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER,
652 			    INTR(vp)))
653 				return (EINTR);
654 		}
655 
656 		va.va_mask = AT_SIZE;
657 		error = nfs3getattr(vp, &va, cr);
658 		if (error)
659 			return (error);
660 		uiop->uio_loffset = va.va_size;
661 	}
662 
663 	offset = uiop->uio_loffset + uiop->uio_resid;
664 
665 	if (uiop->uio_loffset < 0 || offset < 0)
666 		return (EINVAL);
667 
668 	if (limit == RLIM64_INFINITY || limit > MAXOFFSET_T)
669 		limit = MAXOFFSET_T;
670 
671 	/*
672 	 * Check to make sure that the process will not exceed
673 	 * its limit on file size.  It is okay to write up to
674 	 * the limit, but not beyond.  Thus, the write which
675 	 * reaches the limit will be short and the next write
676 	 * will return an error.
677 	 */
678 	remainder = 0;
679 	if (offset > limit) {
680 		remainder = offset - limit;
681 		uiop->uio_resid = limit - uiop->uio_loffset;
682 		if (uiop->uio_resid <= 0) {
683 			proc_t *p = ttoproc(curthread);
684 
685 			uiop->uio_resid += remainder;
686 			mutex_enter(&p->p_lock);
687 			(void) rctl_action(rctlproc_legacy[RLIMIT_FSIZE],
688 			    p->p_rctls, p, RCA_UNSAFE_SIGINFO);
689 			mutex_exit(&p->p_lock);
690 			return (EFBIG);
691 		}
692 	}
693 
694 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
695 		return (EINTR);
696 
697 	/*
698 	 * Bypass VM if caching has been disabled (e.g., locking) or if
699 	 * using client-side direct I/O and the file is not mmap'd and
700 	 * there are no cached pages.
701 	 */
702 	if ((vp->v_flag & VNOCACHE) ||
703 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
704 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
705 		size_t bufsize;
706 		int count;
707 		u_offset_t org_offset;
708 		stable_how stab_comm;
709 
710 nfs3_fwrite:
711 		if (rp->r_flags & RSTALE) {
712 			resid = uiop->uio_resid;
713 			offset = uiop->uio_loffset;
714 			error = rp->r_error;
715 			goto bottom;
716 		}
717 		bufsize = MIN(uiop->uio_resid, mi->mi_stsize);
718 		base = kmem_alloc(bufsize, KM_SLEEP);
719 		do {
720 			if (ioflag & FDSYNC)
721 				stab_comm = DATA_SYNC;
722 			else
723 				stab_comm = FILE_SYNC;
724 			resid = uiop->uio_resid;
725 			offset = uiop->uio_loffset;
726 			count = MIN(uiop->uio_resid, bufsize);
727 			org_offset = uiop->uio_loffset;
728 			error = uiomove(base, count, UIO_WRITE, uiop);
729 			if (!error) {
730 				error = nfs3write(vp, base, org_offset,
731 				    count, cr, &stab_comm);
732 			}
733 		} while (!error && uiop->uio_resid > 0);
734 		kmem_free(base, bufsize);
735 		goto bottom;
736 	}
737 
738 
739 	bsize = vp->v_vfsp->vfs_bsize;
740 
741 	do {
742 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
743 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
744 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
745 
746 		resid = uiop->uio_resid;
747 		offset = uiop->uio_loffset;
748 
749 		if (rp->r_flags & RSTALE) {
750 			error = rp->r_error;
751 			break;
752 		}
753 
754 		/*
755 		 * Don't create dirty pages faster than they
756 		 * can be cleaned so that the system doesn't
757 		 * get imbalanced.  If the async queue is
758 		 * maxed out, then wait for it to drain before
759 		 * creating more dirty pages.  Also, wait for
760 		 * any threads doing pagewalks in the vop_getattr
761 		 * entry points so that they don't block for
762 		 * long periods.
763 		 */
764 		mutex_enter(&rp->r_statelock);
765 		while ((mi->mi_max_threads != 0 &&
766 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
767 		    rp->r_gcount > 0)
768 			cv_wait(&rp->r_cv, &rp->r_statelock);
769 		mutex_exit(&rp->r_statelock);
770 
771 		if (vpm_enable) {
772 			/*
773 			 * It will use kpm mappings, so no need to
774 			 * pass an address.
775 			 */
776 			error = writerp(rp, NULL, n, uiop, 0);
777 		} else  {
778 			if (segmap_kpm) {
779 				int pon = uiop->uio_loffset & PAGEOFFSET;
780 				size_t pn = MIN(PAGESIZE - pon,
781 							uiop->uio_resid);
782 				int pagecreate;
783 
784 				mutex_enter(&rp->r_statelock);
785 				pagecreate = (pon == 0) && (pn == PAGESIZE ||
786 					uiop->uio_loffset + pn >= rp->r_size);
787 				mutex_exit(&rp->r_statelock);
788 
789 				base = segmap_getmapflt(segkmap, vp, off + on,
790 						pn, !pagecreate, S_WRITE);
791 
792 				error = writerp(rp, base + pon, n, uiop,
793 								pagecreate);
794 
795 			} else {
796 				base = segmap_getmapflt(segkmap, vp, off + on,
797 							n, 0, S_READ);
798 				error = writerp(rp, base + on, n, uiop, 0);
799 			}
800 		}
801 
802 		if (!error) {
803 			if (mi->mi_flags & MI_NOAC)
804 				flags = SM_WRITE;
805 			else if ((uiop->uio_loffset % bsize) == 0 ||
806 			    IS_SWAPVP(vp)) {
807 				/*
808 				 * Have written a whole block.
809 				 * Start an asynchronous write
810 				 * and mark the buffer to
811 				 * indicate that it won't be
812 				 * needed again soon.
813 				 */
814 				flags = SM_WRITE | SM_ASYNC | SM_DONTNEED;
815 			} else
816 				flags = 0;
817 			if ((ioflag & (FSYNC|FDSYNC)) ||
818 			    (rp->r_flags & ROUTOFSPACE)) {
819 				flags &= ~SM_ASYNC;
820 				flags |= SM_WRITE;
821 			}
822 			if (vpm_enable) {
823 				error = vpm_sync_pages(vp, off, n, flags);
824 			} else {
825 				error = segmap_release(segkmap, base, flags);
826 			}
827 		} else {
828 			if (vpm_enable) {
829 				(void) vpm_sync_pages(vp, off, n, 0);
830 			} else {
831 				(void) segmap_release(segkmap, base, 0);
832 			}
833 			/*
834 			 * In the event that we got an access error while
835 			 * faulting in a page for a write-only file just
836 			 * force a write.
837 			 */
838 			if (error == EACCES)
839 				goto nfs3_fwrite;
840 		}
841 	} while (!error && uiop->uio_resid > 0);
842 
843 bottom:
844 	if (error) {
845 		uiop->uio_resid = resid + remainder;
846 		uiop->uio_loffset = offset;
847 	} else
848 		uiop->uio_resid += remainder;
849 
850 	nfs_rw_exit(&rp->r_lkserlock);
851 
852 	return (error);
853 }
854 
855 /*
856  * Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED}
857  */
858 static int
859 nfs3_rdwrlbn(vnode_t *vp, page_t *pp, u_offset_t off, size_t len,
860 	int flags, cred_t *cr)
861 {
862 	struct buf *bp;
863 	int error;
864 	page_t *savepp;
865 	uchar_t fsdata;
866 	stable_how stab_comm;
867 
868 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
869 	bp = pageio_setup(pp, len, vp, flags);
870 	ASSERT(bp != NULL);
871 
872 	/*
873 	 * pageio_setup should have set b_addr to 0.  This
874 	 * is correct since we want to do I/O on a page
875 	 * boundary.  bp_mapin will use this addr to calculate
876 	 * an offset, and then set b_addr to the kernel virtual
877 	 * address it allocated for us.
878 	 */
879 	ASSERT(bp->b_un.b_addr == 0);
880 
881 	bp->b_edev = 0;
882 	bp->b_dev = 0;
883 	bp->b_lblkno = lbtodb(off);
884 	bp->b_file = vp;
885 	bp->b_offset = (offset_t)off;
886 	bp_mapin(bp);
887 
888 	/*
889 	 * Calculate the desired level of stability to write data
890 	 * on the server and then mark all of the pages to reflect
891 	 * this.
892 	 */
893 	if ((flags & (B_WRITE|B_ASYNC)) == (B_WRITE|B_ASYNC) &&
894 	    freemem > desfree) {
895 		stab_comm = UNSTABLE;
896 		fsdata = C_DELAYCOMMIT;
897 	} else {
898 		stab_comm = FILE_SYNC;
899 		fsdata = C_NOCOMMIT;
900 	}
901 
902 	savepp = pp;
903 	do {
904 		pp->p_fsdata = fsdata;
905 	} while ((pp = pp->p_next) != savepp);
906 
907 	error = nfs3_bio(bp, &stab_comm, cr);
908 
909 	bp_mapout(bp);
910 	pageio_done(bp);
911 
912 	/*
913 	 * If the server wrote pages in a more stable fashion than
914 	 * was requested, then clear all of the marks in the pages
915 	 * indicating that COMMIT operations were required.
916 	 */
917 	if (stab_comm != UNSTABLE && fsdata == C_DELAYCOMMIT) {
918 		do {
919 			pp->p_fsdata = C_NOCOMMIT;
920 		} while ((pp = pp->p_next) != savepp);
921 	}
922 
923 	return (error);
924 }
925 
926 /*
927  * Write to file.  Writes to remote server in largest size
928  * chunks that the server can handle.  Write is synchronous.
929  */
930 static int
931 nfs3write(vnode_t *vp, caddr_t base, u_offset_t offset, int count, cred_t *cr,
932 	stable_how *stab_comm)
933 {
934 	mntinfo_t *mi;
935 	WRITE3args args;
936 	WRITE3res res;
937 	int error;
938 	int tsize;
939 	rnode_t *rp;
940 	int douprintf;
941 
942 	rp = VTOR(vp);
943 	mi = VTOMI(vp);
944 
945 	ASSERT(nfs_zone() == mi->mi_zone);
946 
947 	args.file = *VTOFH3(vp);
948 	args.stable = *stab_comm;
949 
950 	*stab_comm = FILE_SYNC;
951 
952 	douprintf = 1;
953 
954 	do {
955 		if ((vp->v_flag & VNOCACHE) ||
956 		    (rp->r_flags & RDIRECTIO) ||
957 		    (mi->mi_flags & MI_DIRECTIO))
958 			tsize = MIN(mi->mi_stsize, count);
959 		else
960 			tsize = MIN(mi->mi_curwrite, count);
961 		args.offset = (offset3)offset;
962 		args.count = (count3)tsize;
963 		args.data.data_len = (uint_t)tsize;
964 		args.data.data_val = base;
965 
966 		if (mi->mi_io_kstats) {
967 			mutex_enter(&mi->mi_lock);
968 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
969 			mutex_exit(&mi->mi_lock);
970 		}
971 		args.mblk = NULL;
972 		do {
973 			error = rfs3call(mi, NFSPROC3_WRITE,
974 			    xdr_WRITE3args, (caddr_t)&args,
975 			    xdr_WRITE3res, (caddr_t)&res, cr,
976 			    &douprintf, &res.status, 0, NULL);
977 		} while (error == ENFS_TRYAGAIN);
978 		if (mi->mi_io_kstats) {
979 			mutex_enter(&mi->mi_lock);
980 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
981 			mutex_exit(&mi->mi_lock);
982 		}
983 
984 		if (error)
985 			return (error);
986 		error = geterrno3(res.status);
987 		if (!error) {
988 			if (res.resok.count > args.count) {
989 				zcmn_err(getzoneid(), CE_WARN,
990 				    "nfs3write: server %s wrote %u, "
991 				    "requested was %u",
992 				    rp->r_server->sv_hostname,
993 				    res.resok.count, args.count);
994 				return (EIO);
995 			}
996 			if (res.resok.committed == UNSTABLE) {
997 				*stab_comm = UNSTABLE;
998 				if (args.stable == DATA_SYNC ||
999 				    args.stable == FILE_SYNC) {
1000 					zcmn_err(getzoneid(), CE_WARN,
1001 			"nfs3write: server %s did not commit to stable storage",
1002 					    rp->r_server->sv_hostname);
1003 					return (EIO);
1004 				}
1005 			}
1006 			tsize = (int)res.resok.count;
1007 			count -= tsize;
1008 			base += tsize;
1009 			offset += tsize;
1010 			if (mi->mi_io_kstats) {
1011 				mutex_enter(&mi->mi_lock);
1012 				KSTAT_IO_PTR(mi->mi_io_kstats)->writes++;
1013 				KSTAT_IO_PTR(mi->mi_io_kstats)->nwritten +=
1014 				    tsize;
1015 				mutex_exit(&mi->mi_lock);
1016 			}
1017 			lwp_stat_update(LWP_STAT_OUBLK, 1);
1018 			mutex_enter(&rp->r_statelock);
1019 			if (rp->r_flags & RHAVEVERF) {
1020 				if (rp->r_verf != res.resok.verf) {
1021 					nfs3_set_mod(vp);
1022 					rp->r_verf = res.resok.verf;
1023 					/*
1024 					 * If the data was written UNSTABLE,
1025 					 * then might as well stop because
1026 					 * the whole block will have to get
1027 					 * rewritten anyway.
1028 					 */
1029 					if (*stab_comm == UNSTABLE) {
1030 						mutex_exit(&rp->r_statelock);
1031 						break;
1032 					}
1033 				}
1034 			} else {
1035 				rp->r_verf = res.resok.verf;
1036 				rp->r_flags |= RHAVEVERF;
1037 			}
1038 			/*
1039 			 * Mark the attribute cache as timed out and
1040 			 * set RWRITEATTR to indicate that the file
1041 			 * was modified with a WRITE operation and
1042 			 * that the attributes can not be trusted.
1043 			 */
1044 			PURGE_ATTRCACHE_LOCKED(rp);
1045 			rp->r_flags |= RWRITEATTR;
1046 			mutex_exit(&rp->r_statelock);
1047 		}
1048 	} while (!error && count);
1049 
1050 	return (error);
1051 }
1052 
1053 /*
1054  * Read from a file.  Reads data in largest chunks our interface can handle.
1055  */
1056 static int
1057 nfs3read(vnode_t *vp, caddr_t base, offset_t offset, int count,
1058 	size_t *residp, cred_t *cr)
1059 {
1060 	mntinfo_t *mi;
1061 	READ3args args;
1062 	READ3vres res;
1063 	int tsize;
1064 	int error;
1065 	int douprintf;
1066 	failinfo_t fi;
1067 	rnode_t *rp;
1068 	struct vattr va;
1069 	hrtime_t t;
1070 
1071 	rp = VTOR(vp);
1072 	mi = VTOMI(vp);
1073 	ASSERT(nfs_zone() == mi->mi_zone);
1074 	douprintf = 1;
1075 
1076 	args.file = *VTOFH3(vp);
1077 	fi.vp = vp;
1078 	fi.fhp = (caddr_t)&args.file;
1079 	fi.copyproc = nfs3copyfh;
1080 	fi.lookupproc = nfs3lookup;
1081 	fi.xattrdirproc = acl_getxattrdir3;
1082 
1083 	res.pov.fres.vp = vp;
1084 	res.pov.fres.vap = &va;
1085 
1086 	*residp = count;
1087 	do {
1088 		if (mi->mi_io_kstats) {
1089 			mutex_enter(&mi->mi_lock);
1090 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1091 			mutex_exit(&mi->mi_lock);
1092 		}
1093 
1094 		do {
1095 			if ((vp->v_flag & VNOCACHE) ||
1096 			    (rp->r_flags & RDIRECTIO) ||
1097 			    (mi->mi_flags & MI_DIRECTIO))
1098 				tsize = MIN(mi->mi_tsize, count);
1099 			else
1100 				tsize = MIN(mi->mi_curread, count);
1101 			res.data.data_val = base;
1102 			res.data.data_len = tsize;
1103 			args.offset = (offset3)offset;
1104 			args.count = (count3)tsize;
1105 			t = gethrtime();
1106 			error = rfs3call(mi, NFSPROC3_READ,
1107 			    xdr_READ3args, (caddr_t)&args,
1108 			    xdr_READ3vres, (caddr_t)&res, cr,
1109 			    &douprintf, &res.status, 0, &fi);
1110 		} while (error == ENFS_TRYAGAIN);
1111 
1112 		if (mi->mi_io_kstats) {
1113 			mutex_enter(&mi->mi_lock);
1114 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
1115 			mutex_exit(&mi->mi_lock);
1116 		}
1117 
1118 		if (error)
1119 			return (error);
1120 
1121 		error = geterrno3(res.status);
1122 		if (error)
1123 			return (error);
1124 
1125 		if (res.count != res.data.data_len) {
1126 			zcmn_err(getzoneid(), CE_WARN,
1127 				"nfs3read: server %s returned incorrect amount",
1128 				rp->r_server->sv_hostname);
1129 			return (EIO);
1130 		}
1131 
1132 		count -= res.count;
1133 		*residp = count;
1134 		base += res.count;
1135 		offset += res.count;
1136 		if (mi->mi_io_kstats) {
1137 			mutex_enter(&mi->mi_lock);
1138 			KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
1139 			KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
1140 			mutex_exit(&mi->mi_lock);
1141 		}
1142 		lwp_stat_update(LWP_STAT_INBLK, 1);
1143 	} while (count && !res.eof);
1144 
1145 	if (res.pov.attributes) {
1146 		mutex_enter(&rp->r_statelock);
1147 		if (!CACHE_VALID(rp, va.va_mtime, va.va_size)) {
1148 			mutex_exit(&rp->r_statelock);
1149 			PURGE_ATTRCACHE(vp);
1150 		} else {
1151 			if (rp->r_mtime <= t)
1152 				nfs_attrcache_va(vp, &va);
1153 			mutex_exit(&rp->r_statelock);
1154 		}
1155 	}
1156 
1157 	return (0);
1158 }
1159 
1160 /* ARGSUSED */
1161 static int
1162 nfs3_ioctl(vnode_t *vp, int cmd, intptr_t arg, int flag, cred_t *cr, int *rvalp)
1163 {
1164 
1165 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1166 		return (EIO);
1167 	switch (cmd) {
1168 		case _FIODIRECTIO:
1169 			return (nfs_directio(vp, (int)arg, cr));
1170 		default:
1171 			return (ENOTTY);
1172 	}
1173 }
1174 
1175 static int
1176 nfs3_getattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1177 {
1178 	int error;
1179 	rnode_t *rp;
1180 
1181 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1182 		return (EIO);
1183 	/*
1184 	 * If it has been specified that the return value will
1185 	 * just be used as a hint, and we are only being asked
1186 	 * for size, fsid or rdevid, then return the client's
1187 	 * notion of these values without checking to make sure
1188 	 * that the attribute cache is up to date.
1189 	 * The whole point is to avoid an over the wire GETATTR
1190 	 * call.
1191 	 */
1192 	rp = VTOR(vp);
1193 	if (flags & ATTR_HINT) {
1194 		if (vap->va_mask ==
1195 		    (vap->va_mask & (AT_SIZE | AT_FSID | AT_RDEV))) {
1196 			mutex_enter(&rp->r_statelock);
1197 			if (vap->va_mask | AT_SIZE)
1198 				vap->va_size = rp->r_size;
1199 			if (vap->va_mask | AT_FSID)
1200 				vap->va_fsid = rp->r_attr.va_fsid;
1201 			if (vap->va_mask | AT_RDEV)
1202 				vap->va_rdev = rp->r_attr.va_rdev;
1203 			mutex_exit(&rp->r_statelock);
1204 			return (0);
1205 		}
1206 	}
1207 
1208 	/*
1209 	 * Only need to flush pages if asking for the mtime
1210 	 * and if there any dirty pages or any outstanding
1211 	 * asynchronous (write) requests for this file.
1212 	 */
1213 	if (vap->va_mask & AT_MTIME) {
1214 		if (vn_has_cached_data(vp) &&
1215 		    ((rp->r_flags & RDIRTY) || rp->r_awcount > 0)) {
1216 			mutex_enter(&rp->r_statelock);
1217 			rp->r_gcount++;
1218 			mutex_exit(&rp->r_statelock);
1219 			error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1220 			mutex_enter(&rp->r_statelock);
1221 			if (error && (error == ENOSPC || error == EDQUOT)) {
1222 				if (!rp->r_error)
1223 					rp->r_error = error;
1224 			}
1225 			if (--rp->r_gcount == 0)
1226 				cv_broadcast(&rp->r_cv);
1227 			mutex_exit(&rp->r_statelock);
1228 		}
1229 	}
1230 
1231 	return (nfs3getattr(vp, vap, cr));
1232 }
1233 
1234 /*ARGSUSED4*/
1235 static int
1236 nfs3_setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr,
1237 		caller_context_t *ct)
1238 {
1239 	int error;
1240 	struct vattr va;
1241 
1242 	if (vap->va_mask & AT_NOSET)
1243 		return (EINVAL);
1244 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1245 		return (EIO);
1246 
1247 	va.va_mask = AT_UID | AT_MODE;
1248 	error = nfs3getattr(vp, &va, cr);
1249 	if (error)
1250 		return (error);
1251 
1252 	error = secpolicy_vnode_setattr(cr, vp, vap, &va, flags, nfs3_accessx,
1253 		vp);
1254 	if (error)
1255 		return (error);
1256 
1257 	return (nfs3setattr(vp, vap, flags, cr));
1258 }
1259 
1260 static int
1261 nfs3setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1262 {
1263 	int error;
1264 	uint_t mask;
1265 	SETATTR3args args;
1266 	SETATTR3res res;
1267 	int douprintf;
1268 	rnode_t *rp;
1269 	struct vattr va;
1270 	mode_t omode;
1271 	vsecattr_t *vsp;
1272 	hrtime_t t;
1273 
1274 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
1275 	mask = vap->va_mask;
1276 
1277 	rp = VTOR(vp);
1278 
1279 	/*
1280 	 * Only need to flush pages if there are any pages and
1281 	 * if the file is marked as dirty in some fashion.  The
1282 	 * file must be flushed so that we can accurately
1283 	 * determine the size of the file and the cached data
1284 	 * after the SETATTR returns.  A file is considered to
1285 	 * be dirty if it is either marked with RDIRTY, has
1286 	 * outstanding i/o's active, or is mmap'd.  In this
1287 	 * last case, we can't tell whether there are dirty
1288 	 * pages, so we flush just to be sure.
1289 	 */
1290 	if (vn_has_cached_data(vp) &&
1291 	    ((rp->r_flags & RDIRTY) ||
1292 	    rp->r_count > 0 ||
1293 	    rp->r_mapcnt > 0)) {
1294 		ASSERT(vp->v_type != VCHR);
1295 		error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1296 		if (error && (error == ENOSPC || error == EDQUOT)) {
1297 			mutex_enter(&rp->r_statelock);
1298 			if (!rp->r_error)
1299 				rp->r_error = error;
1300 			mutex_exit(&rp->r_statelock);
1301 		}
1302 	}
1303 
1304 	args.object = *RTOFH3(rp);
1305 	/*
1306 	 * If the intent is for the server to set the times,
1307 	 * there is no point in have the mask indicating set mtime or
1308 	 * atime, because the vap values may be junk, and so result
1309 	 * in an overflow error. Remove these flags from the vap mask
1310 	 * before calling in this case, and restore them afterwards.
1311 	 */
1312 	if ((mask & (AT_ATIME | AT_MTIME)) && !(flags & ATTR_UTIME)) {
1313 		/* Use server times, so don't set the args time fields */
1314 		vap->va_mask &= ~(AT_ATIME | AT_MTIME);
1315 		error = vattr_to_sattr3(vap, &args.new_attributes);
1316 		vap->va_mask |= (mask & (AT_ATIME | AT_MTIME));
1317 		if (mask & AT_ATIME) {
1318 			args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
1319 		}
1320 		if (mask & AT_MTIME) {
1321 			args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
1322 		}
1323 	} else {
1324 		/* Either do not set times or use the client specified times */
1325 		error = vattr_to_sattr3(vap, &args.new_attributes);
1326 	}
1327 
1328 	if (error) {
1329 		/* req time field(s) overflow - return immediately */
1330 		return (error);
1331 	}
1332 
1333 	va.va_mask = AT_MODE | AT_CTIME;
1334 	error = nfs3getattr(vp, &va, cr);
1335 	if (error)
1336 		return (error);
1337 	omode = va.va_mode;
1338 
1339 tryagain:
1340 	if (mask & AT_SIZE) {
1341 		args.guard.check = TRUE;
1342 		args.guard.obj_ctime.seconds = va.va_ctime.tv_sec;
1343 		args.guard.obj_ctime.nseconds = va.va_ctime.tv_nsec;
1344 	} else
1345 		args.guard.check = FALSE;
1346 
1347 	douprintf = 1;
1348 
1349 	t = gethrtime();
1350 
1351 	error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
1352 	    xdr_SETATTR3args, (caddr_t)&args,
1353 	    xdr_SETATTR3res, (caddr_t)&res, cr,
1354 	    &douprintf, &res.status, 0, NULL);
1355 
1356 	/*
1357 	 * Purge the access cache and ACL cache if changing either the
1358 	 * owner of the file, the group owner, or the mode.  These may
1359 	 * change the access permissions of the file, so purge old
1360 	 * information and start over again.
1361 	 */
1362 	if (mask & (AT_UID | AT_GID | AT_MODE)) {
1363 		(void) nfs_access_purge_rp(rp);
1364 		if (rp->r_secattr != NULL) {
1365 			mutex_enter(&rp->r_statelock);
1366 			vsp = rp->r_secattr;
1367 			rp->r_secattr = NULL;
1368 			mutex_exit(&rp->r_statelock);
1369 			if (vsp != NULL)
1370 				nfs_acl_free(vsp);
1371 		}
1372 	}
1373 
1374 	if (error) {
1375 		PURGE_ATTRCACHE(vp);
1376 		return (error);
1377 	}
1378 
1379 	error = geterrno3(res.status);
1380 	if (!error) {
1381 		/*
1382 		 * If changing the size of the file, invalidate
1383 		 * any local cached data which is no longer part
1384 		 * of the file.  We also possibly invalidate the
1385 		 * last page in the file.  We could use
1386 		 * pvn_vpzero(), but this would mark the page as
1387 		 * modified and require it to be written back to
1388 		 * the server for no particularly good reason.
1389 		 * This way, if we access it, then we bring it
1390 		 * back in.  A read should be cheaper than a
1391 		 * write.
1392 		 */
1393 		if (mask & AT_SIZE) {
1394 			nfs_invalidate_pages(vp,
1395 			    (vap->va_size & PAGEMASK), cr);
1396 		}
1397 		nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
1398 		/*
1399 		 * Some servers will change the mode to clear the setuid
1400 		 * and setgid bits when changing the uid or gid.  The
1401 		 * client needs to compensate appropriately.
1402 		 */
1403 		if (mask & (AT_UID | AT_GID)) {
1404 			int terror;
1405 
1406 			va.va_mask = AT_MODE;
1407 			terror = nfs3getattr(vp, &va, cr);
1408 			if (!terror &&
1409 			    (((mask & AT_MODE) && va.va_mode != vap->va_mode) ||
1410 			    (!(mask & AT_MODE) && va.va_mode != omode))) {
1411 				va.va_mask = AT_MODE;
1412 				if (mask & AT_MODE)
1413 					va.va_mode = vap->va_mode;
1414 				else
1415 					va.va_mode = omode;
1416 				(void) nfs3setattr(vp, &va, 0, cr);
1417 			}
1418 		}
1419 	} else {
1420 		nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
1421 		/*
1422 		 * If we got back a "not synchronized" error, then
1423 		 * we need to retry with a new guard value.  The
1424 		 * guard value used is the change time.  If the
1425 		 * server returned post_op_attr, then we can just
1426 		 * retry because we have the latest attributes.
1427 		 * Otherwise, we issue a GETATTR to get the latest
1428 		 * attributes and then retry.  If we couldn't get
1429 		 * the attributes this way either, then we give
1430 		 * up because we can't complete the operation as
1431 		 * required.
1432 		 */
1433 		if (res.status == NFS3ERR_NOT_SYNC) {
1434 			va.va_mask = AT_CTIME;
1435 			if (nfs3getattr(vp, &va, cr) == 0)
1436 				goto tryagain;
1437 		}
1438 		PURGE_STALE_FH(error, vp, cr);
1439 	}
1440 
1441 	return (error);
1442 }
1443 
1444 static int
1445 nfs3_accessx(void *vp, int mode, cred_t *cr)
1446 {
1447 	ASSERT(nfs_zone() == VTOMI((vnode_t *)vp)->mi_zone);
1448 	return (nfs3_access(vp, mode, 0, cr));
1449 }
1450 
1451 /* ARGSUSED */
1452 static int
1453 nfs3_access(vnode_t *vp, int mode, int flags, cred_t *cr)
1454 {
1455 	int error;
1456 	ACCESS3args args;
1457 	ACCESS3res res;
1458 	int douprintf;
1459 	uint32 acc;
1460 	rnode_t *rp;
1461 	cred_t *cred, *ncr, *ncrfree = NULL;
1462 	failinfo_t fi;
1463 	nfs_access_type_t cacc;
1464 	hrtime_t t;
1465 
1466 	acc = 0;
1467 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1468 		return (EIO);
1469 	if (mode & VREAD)
1470 		acc |= ACCESS3_READ;
1471 	if (mode & VWRITE) {
1472 		if (vn_is_readonly(vp) && !IS_DEVVP(vp))
1473 			return (EROFS);
1474 		if (vp->v_type == VDIR)
1475 			acc |= ACCESS3_DELETE;
1476 		acc |= ACCESS3_MODIFY | ACCESS3_EXTEND;
1477 	}
1478 	if (mode & VEXEC) {
1479 		if (vp->v_type == VDIR)
1480 			acc |= ACCESS3_LOOKUP;
1481 		else
1482 			acc |= ACCESS3_EXECUTE;
1483 	}
1484 
1485 	rp = VTOR(vp);
1486 	args.object = *VTOFH3(vp);
1487 	if (vp->v_type == VDIR) {
1488 		args.access = ACCESS3_READ | ACCESS3_DELETE | ACCESS3_MODIFY |
1489 		    ACCESS3_EXTEND | ACCESS3_LOOKUP;
1490 	} else {
1491 		args.access = ACCESS3_READ | ACCESS3_MODIFY | ACCESS3_EXTEND |
1492 		    ACCESS3_EXECUTE;
1493 	}
1494 	fi.vp = vp;
1495 	fi.fhp = (caddr_t)&args.object;
1496 	fi.copyproc = nfs3copyfh;
1497 	fi.lookupproc = nfs3lookup;
1498 	fi.xattrdirproc = acl_getxattrdir3;
1499 
1500 	cred = cr;
1501 	/*
1502 	 * ncr and ncrfree both initially
1503 	 * point to the memory area returned
1504 	 * by crnetadjust();
1505 	 * ncrfree not NULL when exiting means
1506 	 * that we need to release it
1507 	 */
1508 	ncr = crnetadjust(cred);
1509 	ncrfree = ncr;
1510 tryagain:
1511 	if (rp->r_acache != NULL) {
1512 		cacc = nfs_access_check(rp, acc, cred);
1513 		if (cacc == NFS_ACCESS_ALLOWED) {
1514 			if (ncrfree != NULL)
1515 				crfree(ncrfree);
1516 			return (0);
1517 		}
1518 		if (cacc == NFS_ACCESS_DENIED) {
1519 			/*
1520 			 * If the cred can be adjusted, try again
1521 			 * with the new cred.
1522 			 */
1523 			if (ncr != NULL) {
1524 				cred = ncr;
1525 				ncr = NULL;
1526 				goto tryagain;
1527 			}
1528 			if (ncrfree != NULL)
1529 				crfree(ncrfree);
1530 			return (EACCES);
1531 		}
1532 	}
1533 
1534 	douprintf = 1;
1535 
1536 	t = gethrtime();
1537 
1538 	error = rfs3call(VTOMI(vp), NFSPROC3_ACCESS,
1539 	    xdr_ACCESS3args, (caddr_t)&args,
1540 	    xdr_ACCESS3res, (caddr_t)&res, cred,
1541 	    &douprintf, &res.status, 0, &fi);
1542 
1543 	if (error) {
1544 		if (ncrfree != NULL)
1545 			crfree(ncrfree);
1546 		return (error);
1547 	}
1548 
1549 	error = geterrno3(res.status);
1550 	if (!error) {
1551 		nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
1552 		nfs_access_cache(rp, args.access, res.resok.access, cred);
1553 		/*
1554 		 * we just cached results with cred; if cred is the
1555 		 * adjusted credentials from crnetadjust, we do not want
1556 		 * to release them before exiting: hence setting ncrfree
1557 		 * to NULL
1558 		 */
1559 		if (cred != cr)
1560 			ncrfree = NULL;
1561 		if ((acc & res.resok.access) != acc) {
1562 			/*
1563 			 * If the cred can be adjusted, try again
1564 			 * with the new cred.
1565 			 */
1566 			if (ncr != NULL) {
1567 				cred = ncr;
1568 				ncr = NULL;
1569 				goto tryagain;
1570 			}
1571 			error = EACCES;
1572 		}
1573 	} else {
1574 		nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
1575 		PURGE_STALE_FH(error, vp, cr);
1576 	}
1577 
1578 	if (ncrfree != NULL)
1579 		crfree(ncrfree);
1580 
1581 	return (error);
1582 }
1583 
1584 static int nfs3_do_symlink_cache = 1;
1585 
1586 static int
1587 nfs3_readlink(vnode_t *vp, struct uio *uiop, cred_t *cr)
1588 {
1589 	int error;
1590 	READLINK3args args;
1591 	READLINK3res res;
1592 	nfspath3 resdata_backup;
1593 	rnode_t *rp;
1594 	int douprintf;
1595 	int len;
1596 	failinfo_t fi;
1597 	hrtime_t t;
1598 
1599 	/*
1600 	 * Can't readlink anything other than a symbolic link.
1601 	 */
1602 	if (vp->v_type != VLNK)
1603 		return (EINVAL);
1604 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1605 		return (EIO);
1606 
1607 	rp = VTOR(vp);
1608 	if (nfs3_do_symlink_cache && rp->r_symlink.contents != NULL) {
1609 		error = nfs3_validate_caches(vp, cr);
1610 		if (error)
1611 			return (error);
1612 		mutex_enter(&rp->r_statelock);
1613 		if (rp->r_symlink.contents != NULL) {
1614 			error = uiomove(rp->r_symlink.contents,
1615 			    rp->r_symlink.len, UIO_READ, uiop);
1616 			mutex_exit(&rp->r_statelock);
1617 			return (error);
1618 		}
1619 		mutex_exit(&rp->r_statelock);
1620 	}
1621 
1622 	args.symlink = *VTOFH3(vp);
1623 	fi.vp = vp;
1624 	fi.fhp = (caddr_t)&args.symlink;
1625 	fi.copyproc = nfs3copyfh;
1626 	fi.lookupproc = nfs3lookup;
1627 	fi.xattrdirproc = acl_getxattrdir3;
1628 
1629 	res.resok.data = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1630 
1631 	resdata_backup = res.resok.data;
1632 
1633 	douprintf = 1;
1634 
1635 	t = gethrtime();
1636 
1637 	error = rfs3call(VTOMI(vp), NFSPROC3_READLINK,
1638 	    xdr_nfs_fh3, (caddr_t)&args,
1639 	    xdr_READLINK3res, (caddr_t)&res, cr,
1640 	    &douprintf, &res.status, 0, &fi);
1641 
1642 	if (res.resok.data == nfs3nametoolong)
1643 		error = EINVAL;
1644 
1645 	if (error) {
1646 		kmem_free(resdata_backup, MAXPATHLEN);
1647 		return (error);
1648 	}
1649 
1650 	error = geterrno3(res.status);
1651 	if (!error) {
1652 		nfs3_cache_post_op_attr(vp, &res.resok.symlink_attributes, t,
1653 		    cr);
1654 		len = strlen(res.resok.data);
1655 		error = uiomove(res.resok.data, len, UIO_READ, uiop);
1656 		if (nfs3_do_symlink_cache && rp->r_symlink.contents == NULL) {
1657 			mutex_enter(&rp->r_statelock);
1658 				if (rp->r_symlink.contents == NULL) {
1659 				rp->r_symlink.contents = res.resok.data;
1660 				rp->r_symlink.len = len;
1661 				rp->r_symlink.size = MAXPATHLEN;
1662 				mutex_exit(&rp->r_statelock);
1663 			} else {
1664 				mutex_exit(&rp->r_statelock);
1665 
1666 				kmem_free((void *)res.resok.data, MAXPATHLEN);
1667 			}
1668 		} else {
1669 			kmem_free((void *)res.resok.data, MAXPATHLEN);
1670 		}
1671 	} else {
1672 		nfs3_cache_post_op_attr(vp,
1673 		    &res.resfail.symlink_attributes, t, cr);
1674 		PURGE_STALE_FH(error, vp, cr);
1675 
1676 		kmem_free((void *)res.resok.data, MAXPATHLEN);
1677 
1678 	}
1679 
1680 	/*
1681 	 * The over the wire error for attempting to readlink something
1682 	 * other than a symbolic link is ENXIO.  However, we need to
1683 	 * return EINVAL instead of ENXIO, so we map it here.
1684 	 */
1685 	return (error == ENXIO ? EINVAL : error);
1686 }
1687 
1688 /*
1689  * Flush local dirty pages to stable storage on the server.
1690  *
1691  * If FNODSYNC is specified, then there is nothing to do because
1692  * metadata changes are not cached on the client before being
1693  * sent to the server.
1694  */
1695 static int
1696 nfs3_fsync(vnode_t *vp, int syncflag, cred_t *cr)
1697 {
1698 	int error;
1699 
1700 	if ((syncflag & FNODSYNC) || IS_SWAPVP(vp))
1701 		return (0);
1702 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1703 		return (EIO);
1704 
1705 	error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
1706 	if (!error)
1707 		error = VTOR(vp)->r_error;
1708 	return (error);
1709 }
1710 
1711 /*
1712  * Weirdness: if the file was removed or the target of a rename
1713  * operation while it was open, it got renamed instead.  Here we
1714  * remove the renamed file.
1715  */
1716 static void
1717 nfs3_inactive(vnode_t *vp, cred_t *cr)
1718 {
1719 	rnode_t *rp;
1720 
1721 	ASSERT(vp != DNLC_NO_VNODE);
1722 
1723 	/*
1724 	 * If this is coming from the wrong zone, we let someone in the right
1725 	 * zone take care of it asynchronously.  We can get here due to
1726 	 * VN_RELE() being called from pageout() or fsflush().  This call may
1727 	 * potentially turn into an expensive no-op if, for instance, v_count
1728 	 * gets incremented in the meantime, but it's still correct.
1729 	 */
1730 	if (nfs_zone() != VTOMI(vp)->mi_zone) {
1731 		nfs_async_inactive(vp, cr, nfs3_inactive);
1732 		return;
1733 	}
1734 
1735 	rp = VTOR(vp);
1736 redo:
1737 	if (rp->r_unldvp != NULL) {
1738 		/*
1739 		 * Save the vnode pointer for the directory where the
1740 		 * unlinked-open file got renamed, then set it to NULL
1741 		 * to prevent another thread from getting here before
1742 		 * we're done with the remove.  While we have the
1743 		 * statelock, make local copies of the pertinent rnode
1744 		 * fields.  If we weren't to do this in an atomic way, the
1745 		 * the unl* fields could become inconsistent with respect
1746 		 * to each other due to a race condition between this
1747 		 * code and nfs_remove().  See bug report 1034328.
1748 		 */
1749 		mutex_enter(&rp->r_statelock);
1750 		if (rp->r_unldvp != NULL) {
1751 			vnode_t *unldvp;
1752 			char *unlname;
1753 			cred_t *unlcred;
1754 			REMOVE3args args;
1755 			REMOVE3res res;
1756 			int douprintf;
1757 			int error;
1758 			hrtime_t t;
1759 
1760 			unldvp = rp->r_unldvp;
1761 			rp->r_unldvp = NULL;
1762 			unlname = rp->r_unlname;
1763 			rp->r_unlname = NULL;
1764 			unlcred = rp->r_unlcred;
1765 			rp->r_unlcred = NULL;
1766 			mutex_exit(&rp->r_statelock);
1767 
1768 			/*
1769 			 * If there are any dirty pages left, then flush
1770 			 * them.  This is unfortunate because they just
1771 			 * may get thrown away during the remove operation,
1772 			 * but we have to do this for correctness.
1773 			 */
1774 			if (vn_has_cached_data(vp) &&
1775 			    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
1776 				ASSERT(vp->v_type != VCHR);
1777 				error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1778 				if (error) {
1779 					mutex_enter(&rp->r_statelock);
1780 					if (!rp->r_error)
1781 						rp->r_error = error;
1782 					mutex_exit(&rp->r_statelock);
1783 				}
1784 			}
1785 
1786 			/*
1787 			 * Do the remove operation on the renamed file
1788 			 */
1789 			setdiropargs3(&args.object, unlname, unldvp);
1790 
1791 			douprintf = 1;
1792 
1793 			t = gethrtime();
1794 
1795 			error = rfs3call(VTOMI(unldvp), NFSPROC3_REMOVE,
1796 			    xdr_diropargs3, (caddr_t)&args,
1797 			    xdr_REMOVE3res, (caddr_t)&res, unlcred,
1798 			    &douprintf, &res.status, 0, NULL);
1799 
1800 			if (error) {
1801 				PURGE_ATTRCACHE(unldvp);
1802 			} else {
1803 				error = geterrno3(res.status);
1804 				if (!error) {
1805 					nfs3_cache_wcc_data(unldvp,
1806 					    &res.resok.dir_wcc, t, cr);
1807 					if (HAVE_RDDIR_CACHE(VTOR(unldvp)))
1808 						nfs_purge_rddir_cache(unldvp);
1809 				} else {
1810 					nfs3_cache_wcc_data(unldvp,
1811 					    &res.resfail.dir_wcc, t, cr);
1812 					PURGE_STALE_FH(error, unldvp, cr);
1813 				}
1814 			}
1815 
1816 			/*
1817 			 * Release stuff held for the remove
1818 			 */
1819 			VN_RELE(unldvp);
1820 			kmem_free(unlname, MAXNAMELEN);
1821 			crfree(unlcred);
1822 			goto redo;
1823 		}
1824 		mutex_exit(&rp->r_statelock);
1825 	}
1826 
1827 	rp_addfree(rp, cr);
1828 }
1829 
1830 /*
1831  * Remote file system operations having to do with directory manipulation.
1832  */
1833 
1834 static int
1835 nfs3_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1836 	int flags, vnode_t *rdir, cred_t *cr)
1837 {
1838 	int error;
1839 	vnode_t *vp;
1840 	vnode_t *avp = NULL;
1841 	rnode_t *drp;
1842 
1843 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
1844 		return (EPERM);
1845 
1846 	drp = VTOR(dvp);
1847 
1848 	/*
1849 	 * Are we looking up extended attributes?  If so, "dvp" is
1850 	 * the file or directory for which we want attributes, and
1851 	 * we need a lookup of the hidden attribute directory
1852 	 * before we lookup the rest of the path.
1853 	 */
1854 	if (flags & LOOKUP_XATTR) {
1855 		bool_t cflag = ((flags & CREATE_XATTR_DIR) != 0);
1856 		mntinfo_t *mi;
1857 
1858 		mi = VTOMI(dvp);
1859 		if (!(mi->mi_flags & MI_EXTATTR))
1860 			return (EINVAL);
1861 
1862 		if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp)))
1863 			return (EINTR);
1864 
1865 		(void) nfs3lookup_dnlc(dvp, XATTR_DIR_NAME, &avp, cr);
1866 		if (avp == NULL)
1867 			error = acl_getxattrdir3(dvp, &avp, cflag, cr, 0);
1868 		else
1869 			error = 0;
1870 
1871 		nfs_rw_exit(&drp->r_rwlock);
1872 
1873 		if (error) {
1874 			if (mi->mi_flags & MI_EXTATTR)
1875 				return (error);
1876 			return (EINVAL);
1877 		}
1878 		dvp = avp;
1879 		drp = VTOR(dvp);
1880 	}
1881 
1882 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp))) {
1883 		error = EINTR;
1884 		goto out;
1885 	}
1886 
1887 	error = nfs3lookup(dvp, nm, vpp, pnp, flags, rdir, cr, 0);
1888 
1889 	nfs_rw_exit(&drp->r_rwlock);
1890 
1891 	/*
1892 	 * If vnode is a device, create special vnode.
1893 	 */
1894 	if (!error && IS_DEVVP(*vpp)) {
1895 		vp = *vpp;
1896 		*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
1897 		VN_RELE(vp);
1898 	}
1899 
1900 out:
1901 	if (avp != NULL)
1902 		VN_RELE(avp);
1903 
1904 	return (error);
1905 }
1906 
1907 static int nfs3_lookup_neg_cache = 1;
1908 
1909 #ifdef DEBUG
1910 static int nfs3_lookup_dnlc_hits = 0;
1911 static int nfs3_lookup_dnlc_misses = 0;
1912 static int nfs3_lookup_dnlc_neg_hits = 0;
1913 static int nfs3_lookup_dnlc_disappears = 0;
1914 static int nfs3_lookup_dnlc_lookups = 0;
1915 #endif
1916 
1917 /* ARGSUSED */
1918 int
1919 nfs3lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1920 	int flags, vnode_t *rdir, cred_t *cr, int rfscall_flags)
1921 {
1922 	int error;
1923 	rnode_t *drp;
1924 
1925 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1926 	/*
1927 	 * If lookup is for "", just return dvp.  Don't need
1928 	 * to send it over the wire, look it up in the dnlc,
1929 	 * or perform any access checks.
1930 	 */
1931 	if (*nm == '\0') {
1932 		VN_HOLD(dvp);
1933 		*vpp = dvp;
1934 		return (0);
1935 	}
1936 
1937 	/*
1938 	 * Can't do lookups in non-directories.
1939 	 */
1940 	if (dvp->v_type != VDIR)
1941 		return (ENOTDIR);
1942 
1943 	/*
1944 	 * If we're called with RFSCALL_SOFT, it's important that
1945 	 * the only rfscall is one we make directly; if we permit
1946 	 * an access call because we're looking up "." or validating
1947 	 * a dnlc hit, we'll deadlock because that rfscall will not
1948 	 * have the RFSCALL_SOFT set.
1949 	 */
1950 	if (rfscall_flags & RFSCALL_SOFT)
1951 		goto callit;
1952 
1953 	/*
1954 	 * If lookup is for ".", just return dvp.  Don't need
1955 	 * to send it over the wire or look it up in the dnlc,
1956 	 * just need to check access.
1957 	 */
1958 	if (strcmp(nm, ".") == 0) {
1959 		error = nfs3_access(dvp, VEXEC, 0, cr);
1960 		if (error)
1961 			return (error);
1962 		VN_HOLD(dvp);
1963 		*vpp = dvp;
1964 		return (0);
1965 	}
1966 
1967 	drp = VTOR(dvp);
1968 	if (!(drp->r_flags & RLOOKUP)) {
1969 		mutex_enter(&drp->r_statelock);
1970 		drp->r_flags |= RLOOKUP;
1971 		mutex_exit(&drp->r_statelock);
1972 	}
1973 
1974 	/*
1975 	 * Lookup this name in the DNLC.  If there was a valid entry,
1976 	 * then return the results of the lookup.
1977 	 */
1978 	error = nfs3lookup_dnlc(dvp, nm, vpp, cr);
1979 	if (error || *vpp != NULL)
1980 		return (error);
1981 
1982 callit:
1983 	error = nfs3lookup_otw(dvp, nm, vpp, cr, rfscall_flags);
1984 
1985 	return (error);
1986 }
1987 
1988 static int
1989 nfs3lookup_dnlc(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr)
1990 {
1991 	int error;
1992 	vnode_t *vp;
1993 
1994 	ASSERT(*nm != '\0');
1995 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1996 	/*
1997 	 * Lookup this name in the DNLC.  If successful, then validate
1998 	 * the caches and then recheck the DNLC.  The DNLC is rechecked
1999 	 * just in case this entry got invalidated during the call
2000 	 * to nfs3_validate_caches.
2001 	 *
2002 	 * An assumption is being made that it is safe to say that a
2003 	 * file exists which may not on the server.  Any operations to
2004 	 * the server will fail with ESTALE.
2005 	 */
2006 #ifdef DEBUG
2007 	nfs3_lookup_dnlc_lookups++;
2008 #endif
2009 	vp = dnlc_lookup(dvp, nm);
2010 	if (vp != NULL) {
2011 		VN_RELE(vp);
2012 		if (vp == DNLC_NO_VNODE && !vn_is_readonly(dvp)) {
2013 			PURGE_ATTRCACHE(dvp);
2014 		}
2015 		error = nfs3_validate_caches(dvp, cr);
2016 		if (error)
2017 			return (error);
2018 		vp = dnlc_lookup(dvp, nm);
2019 		if (vp != NULL) {
2020 			error = nfs3_access(dvp, VEXEC, 0, cr);
2021 			if (error) {
2022 				VN_RELE(vp);
2023 				return (error);
2024 			}
2025 			if (vp == DNLC_NO_VNODE) {
2026 				VN_RELE(vp);
2027 #ifdef DEBUG
2028 				nfs3_lookup_dnlc_neg_hits++;
2029 #endif
2030 				return (ENOENT);
2031 			}
2032 			*vpp = vp;
2033 #ifdef DEBUG
2034 			nfs3_lookup_dnlc_hits++;
2035 #endif
2036 			return (0);
2037 		}
2038 #ifdef DEBUG
2039 		nfs3_lookup_dnlc_disappears++;
2040 #endif
2041 	}
2042 #ifdef DEBUG
2043 	else
2044 		nfs3_lookup_dnlc_misses++;
2045 #endif
2046 
2047 	*vpp = NULL;
2048 
2049 	return (0);
2050 }
2051 
2052 static int
2053 nfs3lookup_otw(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr,
2054 	int rfscall_flags)
2055 {
2056 	int error;
2057 	LOOKUP3args args;
2058 	LOOKUP3vres res;
2059 	int douprintf;
2060 	struct vattr vattr;
2061 	struct vattr dvattr;
2062 	vnode_t *vp;
2063 	failinfo_t fi;
2064 	hrtime_t t;
2065 
2066 	ASSERT(*nm != '\0');
2067 	ASSERT(dvp->v_type == VDIR);
2068 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2069 
2070 	setdiropargs3(&args.what, nm, dvp);
2071 
2072 	fi.vp = dvp;
2073 	fi.fhp = (caddr_t)&args.what.dir;
2074 	fi.copyproc = nfs3copyfh;
2075 	fi.lookupproc = nfs3lookup;
2076 	fi.xattrdirproc = acl_getxattrdir3;
2077 	res.obj_attributes.fres.vp = dvp;
2078 	res.obj_attributes.fres.vap = &vattr;
2079 	res.dir_attributes.fres.vp = dvp;
2080 	res.dir_attributes.fres.vap = &dvattr;
2081 
2082 	douprintf = 1;
2083 
2084 	t = gethrtime();
2085 
2086 	error = rfs3call(VTOMI(dvp), NFSPROC3_LOOKUP,
2087 	    xdr_diropargs3, (caddr_t)&args,
2088 	    xdr_LOOKUP3vres, (caddr_t)&res, cr,
2089 	    &douprintf, &res.status, rfscall_flags, &fi);
2090 
2091 	if (error)
2092 		return (error);
2093 
2094 	nfs3_cache_post_op_vattr(dvp, &res.dir_attributes, t, cr);
2095 
2096 	error = geterrno3(res.status);
2097 	if (error) {
2098 		PURGE_STALE_FH(error, dvp, cr);
2099 		if (error == ENOENT && nfs3_lookup_neg_cache)
2100 			dnlc_enter(dvp, nm, DNLC_NO_VNODE);
2101 		return (error);
2102 	}
2103 
2104 	if (res.obj_attributes.attributes) {
2105 		vp = makenfs3node_va(&res.object, res.obj_attributes.fres.vap,
2106 				dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2107 	} else {
2108 		vp = makenfs3node_va(&res.object, NULL,
2109 				dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2110 		if (vp->v_type == VNON) {
2111 			vattr.va_mask = AT_TYPE;
2112 			error = nfs3getattr(vp, &vattr, cr);
2113 			if (error) {
2114 				VN_RELE(vp);
2115 				return (error);
2116 			}
2117 			vp->v_type = vattr.va_type;
2118 		}
2119 	}
2120 
2121 	if (!(rfscall_flags & RFSCALL_SOFT))
2122 		dnlc_update(dvp, nm, vp);
2123 
2124 	*vpp = vp;
2125 
2126 	return (error);
2127 }
2128 
2129 #ifdef DEBUG
2130 static int nfs3_create_misses = 0;
2131 #endif
2132 
2133 /* ARGSUSED */
2134 static int
2135 nfs3_create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2136 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2137 {
2138 	int error;
2139 	vnode_t *vp;
2140 	rnode_t *rp;
2141 	struct vattr vattr;
2142 	rnode_t *drp;
2143 	vnode_t *tempvp;
2144 
2145 	drp = VTOR(dvp);
2146 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2147 		return (EPERM);
2148 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2149 		return (EINTR);
2150 
2151 top:
2152 	/*
2153 	 * We make a copy of the attributes because the caller does not
2154 	 * expect us to change what va points to.
2155 	 */
2156 	vattr = *va;
2157 
2158 	/*
2159 	 * If the pathname is "", just use dvp.  Don't need
2160 	 * to send it over the wire, look it up in the dnlc,
2161 	 * or perform any access checks.
2162 	 */
2163 	if (*nm == '\0') {
2164 		error = 0;
2165 		VN_HOLD(dvp);
2166 		vp = dvp;
2167 	/*
2168 	 * If the pathname is ".", just use dvp.  Don't need
2169 	 * to send it over the wire or look it up in the dnlc,
2170 	 * just need to check access.
2171 	 */
2172 	} else if (strcmp(nm, ".") == 0) {
2173 		error = nfs3_access(dvp, VEXEC, 0, cr);
2174 		if (error) {
2175 			nfs_rw_exit(&drp->r_rwlock);
2176 			return (error);
2177 		}
2178 		VN_HOLD(dvp);
2179 		vp = dvp;
2180 	/*
2181 	 * We need to go over the wire, just to be sure whether the
2182 	 * file exists or not.  Using the DNLC can be dangerous in
2183 	 * this case when making a decision regarding existence.
2184 	 */
2185 	} else {
2186 		error = nfs3lookup_otw(dvp, nm, &vp, cr, 0);
2187 	}
2188 	if (!error) {
2189 		if (exclusive == EXCL)
2190 			error = EEXIST;
2191 		else if (vp->v_type == VDIR && (mode & VWRITE))
2192 			error = EISDIR;
2193 		else {
2194 			/*
2195 			 * If vnode is a device, create special vnode.
2196 			 */
2197 			if (IS_DEVVP(vp)) {
2198 				tempvp = vp;
2199 				vp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2200 				VN_RELE(tempvp);
2201 			}
2202 			if (!(error = VOP_ACCESS(vp, mode, 0, cr))) {
2203 				if ((vattr.va_mask & AT_SIZE) &&
2204 				    vp->v_type == VREG) {
2205 					rp = VTOR(vp);
2206 					/*
2207 					 * Check here for large file handled
2208 					 * by LF-unaware process (as
2209 					 * ufs_create() does)
2210 					 */
2211 					if (!(lfaware & FOFFMAX)) {
2212 						mutex_enter(&rp->r_statelock);
2213 						if (rp->r_size > MAXOFF32_T)
2214 							error = EOVERFLOW;
2215 						mutex_exit(&rp->r_statelock);
2216 					}
2217 					if (!error) {
2218 						vattr.va_mask = AT_SIZE;
2219 						error = nfs3setattr(vp,
2220 						    &vattr, 0, cr);
2221 					}
2222 				}
2223 			}
2224 		}
2225 		nfs_rw_exit(&drp->r_rwlock);
2226 		if (error) {
2227 			VN_RELE(vp);
2228 		} else
2229 			*vpp = vp;
2230 		return (error);
2231 	}
2232 
2233 	dnlc_remove(dvp, nm);
2234 
2235 	/*
2236 	 * Decide what the group-id of the created file should be.
2237 	 * Set it in attribute list as advisory...
2238 	 */
2239 	error = setdirgid(dvp, &vattr.va_gid, cr);
2240 	if (error) {
2241 		nfs_rw_exit(&drp->r_rwlock);
2242 		return (error);
2243 	}
2244 	vattr.va_mask |= AT_GID;
2245 
2246 	ASSERT(vattr.va_mask & AT_TYPE);
2247 	if (vattr.va_type == VREG) {
2248 		ASSERT(vattr.va_mask & AT_MODE);
2249 		if (MANDMODE(vattr.va_mode)) {
2250 			nfs_rw_exit(&drp->r_rwlock);
2251 			return (EACCES);
2252 		}
2253 		error = nfs3create(dvp, nm, &vattr, exclusive, mode, vpp, cr,
2254 		    lfaware);
2255 		/*
2256 		 * If this is not an exclusive create, then the CREATE
2257 		 * request will be made with the GUARDED mode set.  This
2258 		 * means that the server will return EEXIST if the file
2259 		 * exists.  The file could exist because of a retransmitted
2260 		 * request.  In this case, we recover by starting over and
2261 		 * checking to see whether the file exists.  This second
2262 		 * time through it should and a CREATE request will not be
2263 		 * sent.
2264 		 *
2265 		 * This handles the problem of a dangling CREATE request
2266 		 * which contains attributes which indicate that the file
2267 		 * should be truncated.  This retransmitted request could
2268 		 * possibly truncate valid data in the file if not caught
2269 		 * by the duplicate request mechanism on the server or if
2270 		 * not caught by other means.  The scenario is:
2271 		 *
2272 		 * Client transmits CREATE request with size = 0
2273 		 * Client times out, retransmits request.
2274 		 * Response to the first request arrives from the server
2275 		 *  and the client proceeds on.
2276 		 * Client writes data to the file.
2277 		 * The server now processes retransmitted CREATE request
2278 		 *  and truncates file.
2279 		 *
2280 		 * The use of the GUARDED CREATE request prevents this from
2281 		 * happening because the retransmitted CREATE would fail
2282 		 * with EEXIST and would not truncate the file.
2283 		 */
2284 		if (error == EEXIST && exclusive == NONEXCL) {
2285 #ifdef DEBUG
2286 			nfs3_create_misses++;
2287 #endif
2288 			goto top;
2289 		}
2290 		nfs_rw_exit(&drp->r_rwlock);
2291 		return (error);
2292 	}
2293 	error = nfs3mknod(dvp, nm, &vattr, exclusive, mode, vpp, cr);
2294 	nfs_rw_exit(&drp->r_rwlock);
2295 	return (error);
2296 }
2297 
2298 /* ARGSUSED */
2299 static int
2300 nfs3create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2301 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2302 {
2303 	int error;
2304 	CREATE3args args;
2305 	CREATE3res res;
2306 	int douprintf;
2307 	vnode_t *vp;
2308 	struct vattr vattr;
2309 	nfstime3 *verfp;
2310 	rnode_t *rp;
2311 	timestruc_t now;
2312 	hrtime_t t;
2313 
2314 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2315 	setdiropargs3(&args.where, nm, dvp);
2316 	if (exclusive == EXCL) {
2317 		args.how.mode = EXCLUSIVE;
2318 		/*
2319 		 * Construct the create verifier.  This verifier needs
2320 		 * to be unique between different clients.  It also needs
2321 		 * to vary for each exclusive create request generated
2322 		 * from the client to the server.
2323 		 *
2324 		 * The first attempt is made to use the hostid and a
2325 		 * unique number on the client.  If the hostid has not
2326 		 * been set, the high resolution time that the exclusive
2327 		 * create request is being made is used.  This will work
2328 		 * unless two different clients, both with the hostid
2329 		 * not set, attempt an exclusive create request on the
2330 		 * same file, at exactly the same clock time.  The
2331 		 * chances of this happening seem small enough to be
2332 		 * reasonable.
2333 		 */
2334 		verfp = (nfstime3 *)&args.how.createhow3_u.verf;
2335 		verfp->seconds = nfs_atoi(hw_serial);
2336 		if (verfp->seconds != 0)
2337 			verfp->nseconds = newnum();
2338 		else {
2339 			gethrestime(&now);
2340 			verfp->seconds = now.tv_sec;
2341 			verfp->nseconds = now.tv_nsec;
2342 		}
2343 		/*
2344 		 * Since the server will use this value for the mtime,
2345 		 * make sure that it can't overflow. Zero out the MSB.
2346 		 * The actual value does not matter here, only its uniqeness.
2347 		 */
2348 		verfp->seconds %= INT32_MAX;
2349 	} else {
2350 		/*
2351 		 * Issue the non-exclusive create in guarded mode.  This
2352 		 * may result in some false EEXIST responses for
2353 		 * retransmitted requests, but these will be handled at
2354 		 * a higher level.  By using GUARDED, duplicate requests
2355 		 * to do file truncation and possible access problems
2356 		 * can be avoided.
2357 		 */
2358 		args.how.mode = GUARDED;
2359 		error = vattr_to_sattr3(va,
2360 				&args.how.createhow3_u.obj_attributes);
2361 		if (error) {
2362 			/* req time field(s) overflow - return immediately */
2363 			return (error);
2364 		}
2365 	}
2366 
2367 	douprintf = 1;
2368 
2369 	t = gethrtime();
2370 
2371 	error = rfs3call(VTOMI(dvp), NFSPROC3_CREATE,
2372 	    xdr_CREATE3args, (caddr_t)&args,
2373 	    xdr_CREATE3res, (caddr_t)&res, cr,
2374 	    &douprintf, &res.status, 0, NULL);
2375 
2376 	if (error) {
2377 		PURGE_ATTRCACHE(dvp);
2378 		return (error);
2379 	}
2380 
2381 	error = geterrno3(res.status);
2382 	if (!error) {
2383 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2384 		if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2385 			nfs_purge_rddir_cache(dvp);
2386 
2387 		/*
2388 		 * On exclusive create the times need to be explicitly
2389 		 * set to clear any potential verifier that may be stored
2390 		 * in one of these fields (see comment below).  This
2391 		 * is done here to cover the case where no post op attrs
2392 		 * were returned or a 'invalid' time was returned in
2393 		 * the attributes.
2394 		 */
2395 		if (exclusive == EXCL)
2396 			va->va_mask |= (AT_MTIME | AT_ATIME);
2397 
2398 		if (!res.resok.obj.handle_follows) {
2399 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2400 			if (error)
2401 				return (error);
2402 		} else {
2403 			if (res.resok.obj_attributes.attributes) {
2404 				vp = makenfs3node(&res.resok.obj.handle,
2405 				    &res.resok.obj_attributes.attr,
2406 				    dvp->v_vfsp, t, cr, NULL, NULL);
2407 			} else {
2408 				vp = makenfs3node(&res.resok.obj.handle, NULL,
2409 				    dvp->v_vfsp, t, cr, NULL, NULL);
2410 
2411 				/*
2412 				 * On an exclusive create, it is possible
2413 				 * that attributes were returned but those
2414 				 * postop attributes failed to decode
2415 				 * properly.  If this is the case,
2416 				 * then most likely the atime or mtime
2417 				 * were invalid for our client; this
2418 				 * is caused by the server storing the
2419 				 * create verifier in one of the time
2420 				 * fields(most likely mtime).
2421 				 * So... we are going to setattr just the
2422 				 * atime/mtime to clear things up.
2423 				 */
2424 				if (exclusive == EXCL) {
2425 					if (error =
2426 						nfs3excl_create_settimes(vp,
2427 							va, cr)) {
2428 						/*
2429 						 * Setting the times failed.
2430 						 * Remove the file and return
2431 						 * the error.
2432 						 */
2433 						VN_RELE(vp);
2434 						(void) nfs3_remove(dvp,
2435 							nm, cr);
2436 						return (error);
2437 					}
2438 				}
2439 
2440 				/*
2441 				 * This handles the non-exclusive case
2442 				 * and the exclusive case where no post op
2443 				 * attrs were returned.
2444 				 */
2445 				if (vp->v_type == VNON) {
2446 					vattr.va_mask = AT_TYPE;
2447 					error = nfs3getattr(vp, &vattr, cr);
2448 					if (error) {
2449 						VN_RELE(vp);
2450 						return (error);
2451 					}
2452 					vp->v_type = vattr.va_type;
2453 				}
2454 			}
2455 			dnlc_update(dvp, nm, vp);
2456 		}
2457 
2458 		rp = VTOR(vp);
2459 
2460 		/*
2461 		 * Check here for large file handled by
2462 		 * LF-unaware process (as ufs_create() does)
2463 		 */
2464 		if ((va->va_mask & AT_SIZE) && vp->v_type == VREG &&
2465 		    !(lfaware & FOFFMAX)) {
2466 			mutex_enter(&rp->r_statelock);
2467 			if (rp->r_size > MAXOFF32_T) {
2468 				mutex_exit(&rp->r_statelock);
2469 				VN_RELE(vp);
2470 				return (EOVERFLOW);
2471 			}
2472 			mutex_exit(&rp->r_statelock);
2473 		}
2474 
2475 		if (exclusive == EXCL &&
2476 			(va->va_mask & ~(AT_GID | AT_SIZE))) {
2477 			/*
2478 			 * If doing an exclusive create, then generate
2479 			 * a SETATTR to set the initial attributes.
2480 			 * Try to set the mtime and the atime to the
2481 			 * server's current time.  It is somewhat
2482 			 * expected that these fields will be used to
2483 			 * store the exclusive create cookie.  If not,
2484 			 * server implementors will need to know that
2485 			 * a SETATTR will follow an exclusive create
2486 			 * and the cookie should be destroyed if
2487 			 * appropriate. This work may have been done
2488 			 * earlier in this function if post op attrs
2489 			 * were not available.
2490 			 *
2491 			 * The AT_GID and AT_SIZE bits are turned off
2492 			 * so that the SETATTR request will not attempt
2493 			 * to process these.  The gid will be set
2494 			 * separately if appropriate.  The size is turned
2495 			 * off because it is assumed that a new file will
2496 			 * be created empty and if the file wasn't empty,
2497 			 * then the exclusive create will have failed
2498 			 * because the file must have existed already.
2499 			 * Therefore, no truncate operation is needed.
2500 			 */
2501 			va->va_mask &= ~(AT_GID | AT_SIZE);
2502 			error = nfs3setattr(vp, va, 0, cr);
2503 			if (error) {
2504 				/*
2505 				 * Couldn't correct the attributes of
2506 				 * the newly created file and the
2507 				 * attributes are wrong.  Remove the
2508 				 * file and return an error to the
2509 				 * application.
2510 				 */
2511 				VN_RELE(vp);
2512 				(void) nfs3_remove(dvp, nm, cr);
2513 				return (error);
2514 			}
2515 		}
2516 
2517 		if (va->va_gid != rp->r_attr.va_gid) {
2518 			/*
2519 			 * If the gid on the file isn't right, then
2520 			 * generate a SETATTR to attempt to change
2521 			 * it.  This may or may not work, depending
2522 			 * upon the server's semantics for allowing
2523 			 * file ownership changes.
2524 			 */
2525 			va->va_mask = AT_GID;
2526 			(void) nfs3setattr(vp, va, 0, cr);
2527 		}
2528 
2529 		/*
2530 		 * If vnode is a device create special vnode
2531 		 */
2532 		if (IS_DEVVP(vp)) {
2533 			*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2534 			VN_RELE(vp);
2535 		} else
2536 			*vpp = vp;
2537 	} else {
2538 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2539 		PURGE_STALE_FH(error, dvp, cr);
2540 	}
2541 
2542 	return (error);
2543 }
2544 
2545 /*
2546  * Special setattr function to take care of rest of atime/mtime
2547  * after successful exclusive create.  This function exists to avoid
2548  * handling attributes from the server; exclusive the atime/mtime fields
2549  * may be 'invalid' in client's view and therefore can not be trusted.
2550  */
2551 static int
2552 nfs3excl_create_settimes(vnode_t *vp, struct vattr *vap, cred_t *cr)
2553 {
2554 	int error;
2555 	uint_t mask;
2556 	SETATTR3args args;
2557 	SETATTR3res res;
2558 	int douprintf;
2559 	rnode_t *rp;
2560 	hrtime_t t;
2561 
2562 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
2563 	/* save the caller's mask so that it can be reset later */
2564 	mask = vap->va_mask;
2565 
2566 	rp = VTOR(vp);
2567 
2568 	args.object = *RTOFH3(rp);
2569 	args.guard.check = FALSE;
2570 
2571 	/* Use the mask to initialize the arguments */
2572 	vap->va_mask = 0;
2573 	error = vattr_to_sattr3(vap, &args.new_attributes);
2574 
2575 	/* We want to set just atime/mtime on this request */
2576 	args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
2577 	args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
2578 
2579 	douprintf = 1;
2580 
2581 	t = gethrtime();
2582 
2583 	error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
2584 	    xdr_SETATTR3args, (caddr_t)&args,
2585 	    xdr_SETATTR3res, (caddr_t)&res, cr,
2586 	    &douprintf, &res.status, 0, NULL);
2587 
2588 	if (error) {
2589 		vap->va_mask = mask;
2590 		return (error);
2591 	}
2592 
2593 	error = geterrno3(res.status);
2594 	if (!error) {
2595 		/*
2596 		 * It is important to pick up the attributes.
2597 		 * Since this is the exclusive create path, the
2598 		 * attributes on the initial create were ignored
2599 		 * and we need these to have the correct info.
2600 		 */
2601 		nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
2602 		/*
2603 		 * No need to do the atime/mtime work again so clear
2604 		 * the bits.
2605 		 */
2606 		mask &= ~(AT_ATIME | AT_MTIME);
2607 	} else {
2608 		nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
2609 	}
2610 
2611 	vap->va_mask = mask;
2612 
2613 	return (error);
2614 }
2615 
2616 /* ARGSUSED */
2617 static int
2618 nfs3mknod(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2619 	int mode, vnode_t **vpp, cred_t *cr)
2620 {
2621 	int error;
2622 	MKNOD3args args;
2623 	MKNOD3res res;
2624 	int douprintf;
2625 	vnode_t *vp;
2626 	struct vattr vattr;
2627 	hrtime_t t;
2628 
2629 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2630 	switch (va->va_type) {
2631 	case VCHR:
2632 	case VBLK:
2633 		setdiropargs3(&args.where, nm, dvp);
2634 		args.what.type = (va->va_type == VCHR) ? NF3CHR : NF3BLK;
2635 		error = vattr_to_sattr3(va,
2636 		    &args.what.mknoddata3_u.device.dev_attributes);
2637 		if (error) {
2638 			/* req time field(s) overflow - return immediately */
2639 			return (error);
2640 		}
2641 		args.what.mknoddata3_u.device.spec.specdata1 =
2642 		    getmajor(va->va_rdev);
2643 		args.what.mknoddata3_u.device.spec.specdata2 =
2644 		    getminor(va->va_rdev);
2645 		break;
2646 
2647 	case VFIFO:
2648 	case VSOCK:
2649 		setdiropargs3(&args.where, nm, dvp);
2650 		args.what.type = (va->va_type == VFIFO) ? NF3FIFO : NF3SOCK;
2651 		error = vattr_to_sattr3(va,
2652 				&args.what.mknoddata3_u.pipe_attributes);
2653 		if (error) {
2654 			/* req time field(s) overflow - return immediately */
2655 			return (error);
2656 		}
2657 		break;
2658 
2659 	default:
2660 		return (EINVAL);
2661 	}
2662 
2663 	douprintf = 1;
2664 
2665 	t = gethrtime();
2666 
2667 	error = rfs3call(VTOMI(dvp), NFSPROC3_MKNOD,
2668 	    xdr_MKNOD3args, (caddr_t)&args,
2669 	    xdr_MKNOD3res, (caddr_t)&res, cr,
2670 	    &douprintf, &res.status, 0, NULL);
2671 
2672 	if (error) {
2673 		PURGE_ATTRCACHE(dvp);
2674 		return (error);
2675 	}
2676 
2677 	error = geterrno3(res.status);
2678 	if (!error) {
2679 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2680 		if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2681 			nfs_purge_rddir_cache(dvp);
2682 
2683 		if (!res.resok.obj.handle_follows) {
2684 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2685 			if (error)
2686 				return (error);
2687 		} else {
2688 			if (res.resok.obj_attributes.attributes) {
2689 				vp = makenfs3node(&res.resok.obj.handle,
2690 				    &res.resok.obj_attributes.attr,
2691 				    dvp->v_vfsp, t, cr, NULL, NULL);
2692 			} else {
2693 				vp = makenfs3node(&res.resok.obj.handle, NULL,
2694 				    dvp->v_vfsp, t, cr, NULL, NULL);
2695 				if (vp->v_type == VNON) {
2696 					vattr.va_mask = AT_TYPE;
2697 					error = nfs3getattr(vp, &vattr, cr);
2698 					if (error) {
2699 						VN_RELE(vp);
2700 						return (error);
2701 					}
2702 					vp->v_type = vattr.va_type;
2703 				}
2704 
2705 			}
2706 			dnlc_update(dvp, nm, vp);
2707 		}
2708 
2709 		if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
2710 			va->va_mask = AT_GID;
2711 			(void) nfs3setattr(vp, va, 0, cr);
2712 		}
2713 
2714 		/*
2715 		 * If vnode is a device create special vnode
2716 		 */
2717 		if (IS_DEVVP(vp)) {
2718 			*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2719 			VN_RELE(vp);
2720 		} else
2721 			*vpp = vp;
2722 	} else {
2723 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2724 		PURGE_STALE_FH(error, dvp, cr);
2725 	}
2726 	return (error);
2727 }
2728 
2729 /*
2730  * Weirdness: if the vnode to be removed is open
2731  * we rename it instead of removing it and nfs_inactive
2732  * will remove the new name.
2733  */
2734 static int
2735 nfs3_remove(vnode_t *dvp, char *nm, cred_t *cr)
2736 {
2737 	int error;
2738 	REMOVE3args args;
2739 	REMOVE3res res;
2740 	vnode_t *vp;
2741 	char *tmpname;
2742 	int douprintf;
2743 	rnode_t *rp;
2744 	rnode_t *drp;
2745 	hrtime_t t;
2746 
2747 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2748 		return (EPERM);
2749 	drp = VTOR(dvp);
2750 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2751 		return (EINTR);
2752 
2753 	error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2754 	if (error) {
2755 		nfs_rw_exit(&drp->r_rwlock);
2756 		return (error);
2757 	}
2758 
2759 	if (vp->v_type == VDIR && secpolicy_fs_linkdir(cr, dvp->v_vfsp)) {
2760 		VN_RELE(vp);
2761 		nfs_rw_exit(&drp->r_rwlock);
2762 		return (EPERM);
2763 	}
2764 
2765 	/*
2766 	 * First just remove the entry from the name cache, as it
2767 	 * is most likely the only entry for this vp.
2768 	 */
2769 	dnlc_remove(dvp, nm);
2770 
2771 	/*
2772 	 * If the file has a v_count > 1 then there may be more than one
2773 	 * entry in the name cache due multiple links or an open file,
2774 	 * but we don't have the real reference count so flush all
2775 	 * possible entries.
2776 	 */
2777 	if (vp->v_count > 1)
2778 		dnlc_purge_vp(vp);
2779 
2780 	/*
2781 	 * Now we have the real reference count on the vnode
2782 	 */
2783 	rp = VTOR(vp);
2784 	mutex_enter(&rp->r_statelock);
2785 	if (vp->v_count > 1 &&
2786 	    (rp->r_unldvp == NULL || strcmp(nm, rp->r_unlname) == 0)) {
2787 		mutex_exit(&rp->r_statelock);
2788 		tmpname = newname();
2789 		error = nfs3rename(dvp, nm, dvp, tmpname, cr);
2790 		if (error)
2791 			kmem_free(tmpname, MAXNAMELEN);
2792 		else {
2793 			mutex_enter(&rp->r_statelock);
2794 			if (rp->r_unldvp == NULL) {
2795 				VN_HOLD(dvp);
2796 				rp->r_unldvp = dvp;
2797 				if (rp->r_unlcred != NULL)
2798 					crfree(rp->r_unlcred);
2799 				crhold(cr);
2800 				rp->r_unlcred = cr;
2801 				rp->r_unlname = tmpname;
2802 			} else {
2803 				kmem_free(rp->r_unlname, MAXNAMELEN);
2804 				rp->r_unlname = tmpname;
2805 			}
2806 			mutex_exit(&rp->r_statelock);
2807 		}
2808 	} else {
2809 		mutex_exit(&rp->r_statelock);
2810 		/*
2811 		 * We need to flush any dirty pages which happen to
2812 		 * be hanging around before removing the file.  This
2813 		 * shouldn't happen very often and mostly on file
2814 		 * systems mounted "nocto".
2815 		 */
2816 		if (vn_has_cached_data(vp) &&
2817 		    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
2818 			error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
2819 			if (error && (error == ENOSPC || error == EDQUOT)) {
2820 				mutex_enter(&rp->r_statelock);
2821 				if (!rp->r_error)
2822 					rp->r_error = error;
2823 				mutex_exit(&rp->r_statelock);
2824 			}
2825 		}
2826 
2827 		setdiropargs3(&args.object, nm, dvp);
2828 
2829 		douprintf = 1;
2830 
2831 		t = gethrtime();
2832 
2833 		error = rfs3call(VTOMI(dvp), NFSPROC3_REMOVE,
2834 		    xdr_diropargs3, (caddr_t)&args,
2835 		    xdr_REMOVE3res, (caddr_t)&res, cr,
2836 		    &douprintf, &res.status, 0, NULL);
2837 
2838 		/*
2839 		 * The xattr dir may be gone after last attr is removed,
2840 		 * so flush it from dnlc.
2841 		 */
2842 		if (dvp->v_flag & V_XATTRDIR)
2843 			dnlc_purge_vp(dvp);
2844 
2845 		PURGE_ATTRCACHE(vp);
2846 
2847 		if (error) {
2848 			PURGE_ATTRCACHE(dvp);
2849 		} else {
2850 			error = geterrno3(res.status);
2851 			if (!error) {
2852 				nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t,
2853 				    cr);
2854 				if (HAVE_RDDIR_CACHE(drp))
2855 					nfs_purge_rddir_cache(dvp);
2856 			} else {
2857 				nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc,
2858 				    t, cr);
2859 				PURGE_STALE_FH(error, dvp, cr);
2860 			}
2861 		}
2862 	}
2863 
2864 	VN_RELE(vp);
2865 
2866 	nfs_rw_exit(&drp->r_rwlock);
2867 
2868 	return (error);
2869 }
2870 
2871 static int
2872 nfs3_link(vnode_t *tdvp, vnode_t *svp, char *tnm, cred_t *cr)
2873 {
2874 	int error;
2875 	LINK3args args;
2876 	LINK3res res;
2877 	vnode_t *realvp;
2878 	int douprintf;
2879 	mntinfo_t *mi;
2880 	rnode_t *tdrp;
2881 	hrtime_t t;
2882 
2883 	if (nfs_zone() != VTOMI(tdvp)->mi_zone)
2884 		return (EPERM);
2885 	if (VOP_REALVP(svp, &realvp) == 0)
2886 		svp = realvp;
2887 
2888 	mi = VTOMI(svp);
2889 
2890 	if (!(mi->mi_flags & MI_LINK))
2891 		return (EOPNOTSUPP);
2892 
2893 	args.file = *VTOFH3(svp);
2894 	setdiropargs3(&args.link, tnm, tdvp);
2895 
2896 	tdrp = VTOR(tdvp);
2897 	if (nfs_rw_enter_sig(&tdrp->r_rwlock, RW_WRITER, INTR(tdvp)))
2898 		return (EINTR);
2899 
2900 	dnlc_remove(tdvp, tnm);
2901 
2902 	douprintf = 1;
2903 
2904 	t = gethrtime();
2905 
2906 	error = rfs3call(mi, NFSPROC3_LINK,
2907 	    xdr_LINK3args, (caddr_t)&args,
2908 	    xdr_LINK3res, (caddr_t)&res, cr,
2909 	    &douprintf, &res.status, 0, NULL);
2910 
2911 	if (error) {
2912 		PURGE_ATTRCACHE(tdvp);
2913 		PURGE_ATTRCACHE(svp);
2914 		nfs_rw_exit(&tdrp->r_rwlock);
2915 		return (error);
2916 	}
2917 
2918 	error = geterrno3(res.status);
2919 
2920 	if (!error) {
2921 		nfs3_cache_post_op_attr(svp, &res.resok.file_attributes, t, cr);
2922 		nfs3_cache_wcc_data(tdvp, &res.resok.linkdir_wcc, t, cr);
2923 		if (HAVE_RDDIR_CACHE(tdrp))
2924 			nfs_purge_rddir_cache(tdvp);
2925 		dnlc_update(tdvp, tnm, svp);
2926 	} else {
2927 		nfs3_cache_post_op_attr(svp, &res.resfail.file_attributes, t,
2928 		    cr);
2929 		nfs3_cache_wcc_data(tdvp, &res.resfail.linkdir_wcc, t, cr);
2930 		if (error == EOPNOTSUPP) {
2931 			mutex_enter(&mi->mi_lock);
2932 			mi->mi_flags &= ~MI_LINK;
2933 			mutex_exit(&mi->mi_lock);
2934 		}
2935 	}
2936 
2937 	nfs_rw_exit(&tdrp->r_rwlock);
2938 
2939 	return (error);
2940 }
2941 
2942 static int
2943 nfs3_rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2944 {
2945 	vnode_t *realvp;
2946 
2947 	if (nfs_zone() != VTOMI(odvp)->mi_zone)
2948 		return (EPERM);
2949 	if (VOP_REALVP(ndvp, &realvp) == 0)
2950 		ndvp = realvp;
2951 
2952 	return (nfs3rename(odvp, onm, ndvp, nnm, cr));
2953 }
2954 
2955 /*
2956  * nfs3rename does the real work of renaming in NFS Version 3.
2957  */
2958 static int
2959 nfs3rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2960 {
2961 	int error;
2962 	RENAME3args args;
2963 	RENAME3res res;
2964 	int douprintf;
2965 	vnode_t *nvp;
2966 	vnode_t *ovp = NULL;
2967 	char *tmpname;
2968 	rnode_t *rp;
2969 	rnode_t *odrp;
2970 	rnode_t *ndrp;
2971 	hrtime_t t;
2972 
2973 	ASSERT(nfs_zone() == VTOMI(odvp)->mi_zone);
2974 
2975 	if (strcmp(onm, ".") == 0 || strcmp(onm, "..") == 0 ||
2976 	    strcmp(nnm, ".") == 0 || strcmp(nnm, "..") == 0)
2977 		return (EINVAL);
2978 
2979 	odrp = VTOR(odvp);
2980 	ndrp = VTOR(ndvp);
2981 	if ((intptr_t)odrp < (intptr_t)ndrp) {
2982 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp)))
2983 			return (EINTR);
2984 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp))) {
2985 			nfs_rw_exit(&odrp->r_rwlock);
2986 			return (EINTR);
2987 		}
2988 	} else {
2989 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp)))
2990 			return (EINTR);
2991 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp))) {
2992 			nfs_rw_exit(&ndrp->r_rwlock);
2993 			return (EINTR);
2994 		}
2995 	}
2996 
2997 	/*
2998 	 * Lookup the target file.  If it exists, it needs to be
2999 	 * checked to see whether it is a mount point and whether
3000 	 * it is active (open).
3001 	 */
3002 	error = nfs3lookup(ndvp, nnm, &nvp, NULL, 0, NULL, cr, 0);
3003 	if (!error) {
3004 		/*
3005 		 * If this file has been mounted on, then just
3006 		 * return busy because renaming to it would remove
3007 		 * the mounted file system from the name space.
3008 		 */
3009 		if (vn_mountedvfs(nvp) != NULL) {
3010 			VN_RELE(nvp);
3011 			nfs_rw_exit(&odrp->r_rwlock);
3012 			nfs_rw_exit(&ndrp->r_rwlock);
3013 			return (EBUSY);
3014 		}
3015 
3016 		/*
3017 		 * Purge the name cache of all references to this vnode
3018 		 * so that we can check the reference count to infer
3019 		 * whether it is active or not.
3020 		 */
3021 		/*
3022 		 * First just remove the entry from the name cache, as it
3023 		 * is most likely the only entry for this vp.
3024 		 */
3025 		dnlc_remove(ndvp, nnm);
3026 		/*
3027 		 * If the file has a v_count > 1 then there may be more
3028 		 * than one entry in the name cache due multiple links
3029 		 * or an open file, but we don't have the real reference
3030 		 * count so flush all possible entries.
3031 		 */
3032 		if (nvp->v_count > 1)
3033 			dnlc_purge_vp(nvp);
3034 
3035 		/*
3036 		 * If the vnode is active and is not a directory,
3037 		 * arrange to rename it to a
3038 		 * temporary file so that it will continue to be
3039 		 * accessible.  This implements the "unlink-open-file"
3040 		 * semantics for the target of a rename operation.
3041 		 * Before doing this though, make sure that the
3042 		 * source and target files are not already the same.
3043 		 */
3044 		if (nvp->v_count > 1 && nvp->v_type != VDIR) {
3045 			/*
3046 			 * Lookup the source name.
3047 			 */
3048 			error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL,
3049 			    cr, 0);
3050 
3051 			/*
3052 			 * The source name *should* already exist.
3053 			 */
3054 			if (error) {
3055 				VN_RELE(nvp);
3056 				nfs_rw_exit(&odrp->r_rwlock);
3057 				nfs_rw_exit(&ndrp->r_rwlock);
3058 				return (error);
3059 			}
3060 
3061 			/*
3062 			 * Compare the two vnodes.  If they are the same,
3063 			 * just release all held vnodes and return success.
3064 			 */
3065 			if (ovp == nvp) {
3066 				VN_RELE(ovp);
3067 				VN_RELE(nvp);
3068 				nfs_rw_exit(&odrp->r_rwlock);
3069 				nfs_rw_exit(&ndrp->r_rwlock);
3070 				return (0);
3071 			}
3072 
3073 			/*
3074 			 * Can't mix and match directories and non-
3075 			 * directories in rename operations.  We already
3076 			 * know that the target is not a directory.  If
3077 			 * the source is a directory, return an error.
3078 			 */
3079 			if (ovp->v_type == VDIR) {
3080 				VN_RELE(ovp);
3081 				VN_RELE(nvp);
3082 				nfs_rw_exit(&odrp->r_rwlock);
3083 				nfs_rw_exit(&ndrp->r_rwlock);
3084 				return (ENOTDIR);
3085 			}
3086 
3087 			/*
3088 			 * The target file exists, is not the same as
3089 			 * the source file, and is active.  Link it
3090 			 * to a temporary filename to avoid having
3091 			 * the server removing the file completely.
3092 			 */
3093 			tmpname = newname();
3094 			error = nfs3_link(ndvp, nvp, tmpname, cr);
3095 			if (error == EOPNOTSUPP) {
3096 				error = nfs3_rename(ndvp, nnm, ndvp, tmpname,
3097 				    cr);
3098 			}
3099 			if (error) {
3100 				kmem_free(tmpname, MAXNAMELEN);
3101 				VN_RELE(ovp);
3102 				VN_RELE(nvp);
3103 				nfs_rw_exit(&odrp->r_rwlock);
3104 				nfs_rw_exit(&ndrp->r_rwlock);
3105 				return (error);
3106 			}
3107 			rp = VTOR(nvp);
3108 			mutex_enter(&rp->r_statelock);
3109 			if (rp->r_unldvp == NULL) {
3110 				VN_HOLD(ndvp);
3111 				rp->r_unldvp = ndvp;
3112 				if (rp->r_unlcred != NULL)
3113 					crfree(rp->r_unlcred);
3114 				crhold(cr);
3115 				rp->r_unlcred = cr;
3116 				rp->r_unlname = tmpname;
3117 			} else {
3118 				kmem_free(rp->r_unlname, MAXNAMELEN);
3119 				rp->r_unlname = tmpname;
3120 			}
3121 			mutex_exit(&rp->r_statelock);
3122 		}
3123 
3124 		VN_RELE(nvp);
3125 	}
3126 
3127 	if (ovp == NULL) {
3128 		/*
3129 		 * When renaming directories to be a subdirectory of a
3130 		 * different parent, the dnlc entry for ".." will no
3131 		 * longer be valid, so it must be removed.
3132 		 *
3133 		 * We do a lookup here to determine whether we are renaming
3134 		 * a directory and we need to check if we are renaming
3135 		 * an unlinked file.  This might have already been done
3136 		 * in previous code, so we check ovp == NULL to avoid
3137 		 * doing it twice.
3138 		 */
3139 
3140 		error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL, cr, 0);
3141 		/*
3142 		 * The source name *should* already exist.
3143 		 */
3144 		if (error) {
3145 			nfs_rw_exit(&odrp->r_rwlock);
3146 			nfs_rw_exit(&ndrp->r_rwlock);
3147 			return (error);
3148 		}
3149 		ASSERT(ovp != NULL);
3150 	}
3151 
3152 	dnlc_remove(odvp, onm);
3153 	dnlc_remove(ndvp, nnm);
3154 
3155 	setdiropargs3(&args.from, onm, odvp);
3156 	setdiropargs3(&args.to, nnm, ndvp);
3157 
3158 	douprintf = 1;
3159 
3160 	t = gethrtime();
3161 
3162 	error = rfs3call(VTOMI(odvp), NFSPROC3_RENAME,
3163 	    xdr_RENAME3args, (caddr_t)&args,
3164 	    xdr_RENAME3res, (caddr_t)&res, cr,
3165 	    &douprintf, &res.status, 0, NULL);
3166 
3167 	if (error) {
3168 		PURGE_ATTRCACHE(odvp);
3169 		PURGE_ATTRCACHE(ndvp);
3170 		VN_RELE(ovp);
3171 		nfs_rw_exit(&odrp->r_rwlock);
3172 		nfs_rw_exit(&ndrp->r_rwlock);
3173 		return (error);
3174 	}
3175 
3176 	error = geterrno3(res.status);
3177 
3178 	if (!error) {
3179 		nfs3_cache_wcc_data(odvp, &res.resok.fromdir_wcc, t, cr);
3180 		if (HAVE_RDDIR_CACHE(odrp))
3181 			nfs_purge_rddir_cache(odvp);
3182 		if (ndvp != odvp) {
3183 			nfs3_cache_wcc_data(ndvp, &res.resok.todir_wcc, t, cr);
3184 			if (HAVE_RDDIR_CACHE(ndrp))
3185 				nfs_purge_rddir_cache(ndvp);
3186 		}
3187 		/*
3188 		 * when renaming directories to be a subdirectory of a
3189 		 * different parent, the dnlc entry for ".." will no
3190 		 * longer be valid, so it must be removed
3191 		 */
3192 		rp = VTOR(ovp);
3193 		if (ndvp != odvp) {
3194 			if (ovp->v_type == VDIR) {
3195 				dnlc_remove(ovp, "..");
3196 				if (HAVE_RDDIR_CACHE(rp))
3197 					nfs_purge_rddir_cache(ovp);
3198 			}
3199 		}
3200 
3201 		/*
3202 		 * If we are renaming the unlinked file, update the
3203 		 * r_unldvp and r_unlname as needed.
3204 		 */
3205 		mutex_enter(&rp->r_statelock);
3206 		if (rp->r_unldvp != NULL) {
3207 			if (strcmp(rp->r_unlname, onm) == 0) {
3208 				(void) strncpy(rp->r_unlname, nnm, MAXNAMELEN);
3209 				rp->r_unlname[MAXNAMELEN - 1] = '\0';
3210 
3211 				if (ndvp != rp->r_unldvp) {
3212 					VN_RELE(rp->r_unldvp);
3213 					rp->r_unldvp = ndvp;
3214 					VN_HOLD(ndvp);
3215 				}
3216 			}
3217 		}
3218 		mutex_exit(&rp->r_statelock);
3219 	} else {
3220 		nfs3_cache_wcc_data(odvp, &res.resfail.fromdir_wcc, t, cr);
3221 		if (ndvp != odvp) {
3222 			nfs3_cache_wcc_data(ndvp, &res.resfail.todir_wcc, t,
3223 			    cr);
3224 		}
3225 		/*
3226 		 * System V defines rename to return EEXIST, not
3227 		 * ENOTEMPTY if the target directory is not empty.
3228 		 * Over the wire, the error is NFSERR_ENOTEMPTY
3229 		 * which geterrno maps to ENOTEMPTY.
3230 		 */
3231 		if (error == ENOTEMPTY)
3232 			error = EEXIST;
3233 	}
3234 
3235 	VN_RELE(ovp);
3236 
3237 	nfs_rw_exit(&odrp->r_rwlock);
3238 	nfs_rw_exit(&ndrp->r_rwlock);
3239 
3240 	return (error);
3241 }
3242 
3243 static int
3244 nfs3_mkdir(vnode_t *dvp, char *nm, struct vattr *va, vnode_t **vpp, cred_t *cr)
3245 {
3246 	int error;
3247 	MKDIR3args args;
3248 	MKDIR3res res;
3249 	int douprintf;
3250 	struct vattr vattr;
3251 	vnode_t *vp;
3252 	rnode_t *drp;
3253 	hrtime_t t;
3254 
3255 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
3256 		return (EPERM);
3257 	setdiropargs3(&args.where, nm, dvp);
3258 
3259 	/*
3260 	 * Decide what the group-id and set-gid bit of the created directory
3261 	 * should be.  May have to do a setattr to get the gid right.
3262 	 */
3263 	error = setdirgid(dvp, &va->va_gid, cr);
3264 	if (error)
3265 		return (error);
3266 	error = setdirmode(dvp, &va->va_mode, cr);
3267 	if (error)
3268 		return (error);
3269 	va->va_mask |= AT_MODE|AT_GID;
3270 
3271 	error = vattr_to_sattr3(va, &args.attributes);
3272 	if (error) {
3273 		/* req time field(s) overflow - return immediately */
3274 		return (error);
3275 	}
3276 
3277 	drp = VTOR(dvp);
3278 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3279 		return (EINTR);
3280 
3281 	dnlc_remove(dvp, nm);
3282 
3283 	douprintf = 1;
3284 
3285 	t = gethrtime();
3286 
3287 	error = rfs3call(VTOMI(dvp), NFSPROC3_MKDIR,
3288 	    xdr_MKDIR3args, (caddr_t)&args,
3289 	    xdr_MKDIR3res, (caddr_t)&res, cr,
3290 	    &douprintf, &res.status, 0, NULL);
3291 
3292 	if (error) {
3293 		PURGE_ATTRCACHE(dvp);
3294 		nfs_rw_exit(&drp->r_rwlock);
3295 		return (error);
3296 	}
3297 
3298 	error = geterrno3(res.status);
3299 	if (!error) {
3300 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3301 		if (HAVE_RDDIR_CACHE(drp))
3302 			nfs_purge_rddir_cache(dvp);
3303 
3304 		if (!res.resok.obj.handle_follows) {
3305 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3306 			if (error) {
3307 				nfs_rw_exit(&drp->r_rwlock);
3308 				return (error);
3309 			}
3310 		} else {
3311 			if (res.resok.obj_attributes.attributes) {
3312 				vp = makenfs3node(&res.resok.obj.handle,
3313 				    &res.resok.obj_attributes.attr,
3314 				    dvp->v_vfsp, t, cr, NULL, NULL);
3315 			} else {
3316 				vp = makenfs3node(&res.resok.obj.handle, NULL,
3317 				    dvp->v_vfsp, t, cr, NULL, NULL);
3318 				if (vp->v_type == VNON) {
3319 					vattr.va_mask = AT_TYPE;
3320 					error = nfs3getattr(vp, &vattr, cr);
3321 					if (error) {
3322 						VN_RELE(vp);
3323 						nfs_rw_exit(&drp->r_rwlock);
3324 						return (error);
3325 					}
3326 					vp->v_type = vattr.va_type;
3327 				}
3328 			}
3329 			dnlc_update(dvp, nm, vp);
3330 		}
3331 		if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
3332 			va->va_mask = AT_GID;
3333 			(void) nfs3setattr(vp, va, 0, cr);
3334 		}
3335 		*vpp = vp;
3336 	} else {
3337 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3338 		PURGE_STALE_FH(error, dvp, cr);
3339 	}
3340 
3341 	nfs_rw_exit(&drp->r_rwlock);
3342 
3343 	return (error);
3344 }
3345 
3346 static int
3347 nfs3_rmdir(vnode_t *dvp, char *nm, vnode_t *cdir, cred_t *cr)
3348 {
3349 	int error;
3350 	RMDIR3args args;
3351 	RMDIR3res res;
3352 	vnode_t *vp;
3353 	int douprintf;
3354 	rnode_t *drp;
3355 	hrtime_t t;
3356 
3357 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
3358 		return (EPERM);
3359 	drp = VTOR(dvp);
3360 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3361 		return (EINTR);
3362 
3363 	/*
3364 	 * Attempt to prevent a rmdir(".") from succeeding.
3365 	 */
3366 	error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3367 	if (error) {
3368 		nfs_rw_exit(&drp->r_rwlock);
3369 		return (error);
3370 	}
3371 
3372 	if (vp == cdir) {
3373 		VN_RELE(vp);
3374 		nfs_rw_exit(&drp->r_rwlock);
3375 		return (EINVAL);
3376 	}
3377 
3378 	setdiropargs3(&args.object, nm, dvp);
3379 
3380 	/*
3381 	 * First just remove the entry from the name cache, as it
3382 	 * is most likely an entry for this vp.
3383 	 */
3384 	dnlc_remove(dvp, nm);
3385 
3386 	/*
3387 	 * If there vnode reference count is greater than one, then
3388 	 * there may be additional references in the DNLC which will
3389 	 * need to be purged.  First, trying removing the entry for
3390 	 * the parent directory and see if that removes the additional
3391 	 * reference(s).  If that doesn't do it, then use dnlc_purge_vp
3392 	 * to completely remove any references to the directory which
3393 	 * might still exist in the DNLC.
3394 	 */
3395 	if (vp->v_count > 1) {
3396 		dnlc_remove(vp, "..");
3397 		if (vp->v_count > 1)
3398 			dnlc_purge_vp(vp);
3399 	}
3400 
3401 	douprintf = 1;
3402 
3403 	t = gethrtime();
3404 
3405 	error = rfs3call(VTOMI(dvp), NFSPROC3_RMDIR,
3406 	    xdr_diropargs3, (caddr_t)&args,
3407 	    xdr_RMDIR3res, (caddr_t)&res, cr,
3408 	    &douprintf, &res.status, 0, NULL);
3409 
3410 	PURGE_ATTRCACHE(vp);
3411 
3412 	if (error) {
3413 		PURGE_ATTRCACHE(dvp);
3414 		VN_RELE(vp);
3415 		nfs_rw_exit(&drp->r_rwlock);
3416 		return (error);
3417 	}
3418 
3419 	error = geterrno3(res.status);
3420 	if (!error) {
3421 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3422 		if (HAVE_RDDIR_CACHE(drp))
3423 			nfs_purge_rddir_cache(dvp);
3424 		if (HAVE_RDDIR_CACHE(VTOR(vp)))
3425 			nfs_purge_rddir_cache(vp);
3426 	} else {
3427 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3428 		PURGE_STALE_FH(error, dvp, cr);
3429 		/*
3430 		 * System V defines rmdir to return EEXIST, not
3431 		 * ENOTEMPTY if the directory is not empty.  Over
3432 		 * the wire, the error is NFSERR_ENOTEMPTY which
3433 		 * geterrno maps to ENOTEMPTY.
3434 		 */
3435 		if (error == ENOTEMPTY)
3436 			error = EEXIST;
3437 	}
3438 
3439 	VN_RELE(vp);
3440 
3441 	nfs_rw_exit(&drp->r_rwlock);
3442 
3443 	return (error);
3444 }
3445 
3446 static int
3447 nfs3_symlink(vnode_t *dvp, char *lnm, struct vattr *tva, char *tnm, cred_t *cr)
3448 {
3449 	int error;
3450 	SYMLINK3args args;
3451 	SYMLINK3res res;
3452 	int douprintf;
3453 	mntinfo_t *mi;
3454 	vnode_t *vp;
3455 	rnode_t *rp;
3456 	char *contents;
3457 	rnode_t *drp;
3458 	hrtime_t t;
3459 
3460 	mi = VTOMI(dvp);
3461 
3462 	if (nfs_zone() != mi->mi_zone)
3463 		return (EPERM);
3464 	if (!(mi->mi_flags & MI_SYMLINK))
3465 		return (EOPNOTSUPP);
3466 
3467 	setdiropargs3(&args.where, lnm, dvp);
3468 	error = vattr_to_sattr3(tva, &args.symlink.symlink_attributes);
3469 	if (error) {
3470 		/* req time field(s) overflow - return immediately */
3471 		return (error);
3472 	}
3473 	args.symlink.symlink_data = tnm;
3474 
3475 	drp = VTOR(dvp);
3476 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3477 		return (EINTR);
3478 
3479 	dnlc_remove(dvp, lnm);
3480 
3481 	douprintf = 1;
3482 
3483 	t = gethrtime();
3484 
3485 	error = rfs3call(mi, NFSPROC3_SYMLINK,
3486 	    xdr_SYMLINK3args, (caddr_t)&args,
3487 	    xdr_SYMLINK3res, (caddr_t)&res, cr,
3488 	    &douprintf, &res.status, 0, NULL);
3489 
3490 	if (error) {
3491 		PURGE_ATTRCACHE(dvp);
3492 		nfs_rw_exit(&drp->r_rwlock);
3493 		return (error);
3494 	}
3495 
3496 	error = geterrno3(res.status);
3497 	if (!error) {
3498 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3499 		if (HAVE_RDDIR_CACHE(drp))
3500 			nfs_purge_rddir_cache(dvp);
3501 
3502 		if (res.resok.obj.handle_follows) {
3503 			if (res.resok.obj_attributes.attributes) {
3504 				vp = makenfs3node(&res.resok.obj.handle,
3505 				    &res.resok.obj_attributes.attr,
3506 				    dvp->v_vfsp, t, cr, NULL, NULL);
3507 			} else {
3508 				vp = makenfs3node(&res.resok.obj.handle, NULL,
3509 				    dvp->v_vfsp, t, cr, NULL, NULL);
3510 				vp->v_type = VLNK;
3511 				vp->v_rdev = 0;
3512 			}
3513 			dnlc_update(dvp, lnm, vp);
3514 			rp = VTOR(vp);
3515 			if (nfs3_do_symlink_cache &&
3516 			    rp->r_symlink.contents == NULL) {
3517 
3518 				contents = kmem_alloc(MAXPATHLEN,
3519 				    KM_NOSLEEP);
3520 
3521 				if (contents != NULL) {
3522 					mutex_enter(&rp->r_statelock);
3523 					if (rp->r_symlink.contents == NULL) {
3524 						rp->r_symlink.len = strlen(tnm);
3525 						bcopy(tnm, contents,
3526 						    rp->r_symlink.len);
3527 						rp->r_symlink.contents =
3528 						    contents;
3529 						rp->r_symlink.size = MAXPATHLEN;
3530 						mutex_exit(&rp->r_statelock);
3531 					} else {
3532 						mutex_exit(&rp->r_statelock);
3533 						kmem_free((void *)contents,
3534 							    MAXPATHLEN);
3535 					}
3536 				}
3537 			}
3538 			VN_RELE(vp);
3539 		}
3540 	} else {
3541 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3542 		PURGE_STALE_FH(error, dvp, cr);
3543 		if (error == EOPNOTSUPP) {
3544 			mutex_enter(&mi->mi_lock);
3545 			mi->mi_flags &= ~MI_SYMLINK;
3546 			mutex_exit(&mi->mi_lock);
3547 		}
3548 	}
3549 
3550 	nfs_rw_exit(&drp->r_rwlock);
3551 
3552 	return (error);
3553 }
3554 
3555 #ifdef DEBUG
3556 static int nfs3_readdir_cache_hits = 0;
3557 static int nfs3_readdir_cache_shorts = 0;
3558 static int nfs3_readdir_cache_waits = 0;
3559 static int nfs3_readdir_cache_misses = 0;
3560 static int nfs3_readdir_readahead = 0;
3561 #endif
3562 
3563 static int nfs3_shrinkreaddir = 0;
3564 
3565 /*
3566  * Read directory entries.
3567  * There are some weird things to look out for here.  The uio_loffset
3568  * field is either 0 or it is the offset returned from a previous
3569  * readdir.  It is an opaque value used by the server to find the
3570  * correct directory block to read. The count field is the number
3571  * of blocks to read on the server.  This is advisory only, the server
3572  * may return only one block's worth of entries.  Entries may be compressed
3573  * on the server.
3574  */
3575 static int
3576 nfs3_readdir(vnode_t *vp, struct uio *uiop, cred_t *cr, int *eofp)
3577 {
3578 	int error;
3579 	size_t count;
3580 	rnode_t *rp;
3581 	rddir_cache *rdc;
3582 	rddir_cache *nrdc;
3583 	rddir_cache *rrdc;
3584 #ifdef DEBUG
3585 	int missed;
3586 #endif
3587 	int doreadahead;
3588 	rddir_cache srdc;
3589 	avl_index_t where;
3590 
3591 	if (nfs_zone() != VTOMI(vp)->mi_zone)
3592 		return (EIO);
3593 	rp = VTOR(vp);
3594 
3595 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
3596 
3597 	/*
3598 	 * Make sure that the directory cache is valid.
3599 	 */
3600 	if (HAVE_RDDIR_CACHE(rp)) {
3601 		if (nfs_disable_rddir_cache) {
3602 			/*
3603 			 * Setting nfs_disable_rddir_cache in /etc/system
3604 			 * allows interoperability with servers that do not
3605 			 * properly update the attributes of directories.
3606 			 * Any cached information gets purged before an
3607 			 * access is made to it.
3608 			 */
3609 			nfs_purge_rddir_cache(vp);
3610 		} else {
3611 			error = nfs3_validate_caches(vp, cr);
3612 			if (error)
3613 				return (error);
3614 		}
3615 	}
3616 
3617 	/*
3618 	 * It is possible that some servers may not be able to correctly
3619 	 * handle a large READDIR or READDIRPLUS request due to bugs in
3620 	 * their implementation.  In order to continue to interoperate
3621 	 * with them, this workaround is provided to limit the maximum
3622 	 * size of a READDIRPLUS request to 1024.  In any case, the request
3623 	 * size is limited to MAXBSIZE.
3624 	 */
3625 	count = MIN(uiop->uio_iov->iov_len,
3626 	    nfs3_shrinkreaddir ? 1024 : MAXBSIZE);
3627 
3628 	nrdc = NULL;
3629 #ifdef DEBUG
3630 	missed = 0;
3631 #endif
3632 top:
3633 	/*
3634 	 * Short circuit last readdir which always returns 0 bytes.
3635 	 * This can be done after the directory has been read through
3636 	 * completely at least once.  This will set r_direof which
3637 	 * can be used to find the value of the last cookie.
3638 	 */
3639 	mutex_enter(&rp->r_statelock);
3640 	if (rp->r_direof != NULL &&
3641 	    uiop->uio_loffset == rp->r_direof->nfs3_ncookie) {
3642 		mutex_exit(&rp->r_statelock);
3643 #ifdef DEBUG
3644 		nfs3_readdir_cache_shorts++;
3645 #endif
3646 		if (eofp)
3647 			*eofp = 1;
3648 		if (nrdc != NULL)
3649 			rddir_cache_rele(nrdc);
3650 		return (0);
3651 	}
3652 	/*
3653 	 * Look for a cache entry.  Cache entries are identified
3654 	 * by the NFS cookie value and the byte count requested.
3655 	 */
3656 	srdc.nfs3_cookie = uiop->uio_loffset;
3657 	srdc.buflen = count;
3658 	rdc = avl_find(&rp->r_dir, &srdc, &where);
3659 	if (rdc != NULL) {
3660 		rddir_cache_hold(rdc);
3661 		/*
3662 		 * If the cache entry is in the process of being
3663 		 * filled in, wait until this completes.  The
3664 		 * RDDIRWAIT bit is set to indicate that someone
3665 		 * is waiting and then the thread currently
3666 		 * filling the entry is done, it should do a
3667 		 * cv_broadcast to wakeup all of the threads
3668 		 * waiting for it to finish.
3669 		 */
3670 		if (rdc->flags & RDDIR) {
3671 			nfs_rw_exit(&rp->r_rwlock);
3672 			rdc->flags |= RDDIRWAIT;
3673 #ifdef DEBUG
3674 			nfs3_readdir_cache_waits++;
3675 #endif
3676 			if (!cv_wait_sig(&rdc->cv, &rp->r_statelock)) {
3677 				/*
3678 				 * We got interrupted, probably
3679 				 * the user typed ^C or an alarm
3680 				 * fired.  We free the new entry
3681 				 * if we allocated one.
3682 				 */
3683 				mutex_exit(&rp->r_statelock);
3684 				(void) nfs_rw_enter_sig(&rp->r_rwlock,
3685 					RW_READER, FALSE);
3686 				rddir_cache_rele(rdc);
3687 				if (nrdc != NULL)
3688 					rddir_cache_rele(nrdc);
3689 				return (EINTR);
3690 			}
3691 			mutex_exit(&rp->r_statelock);
3692 			(void) nfs_rw_enter_sig(&rp->r_rwlock,
3693 				RW_READER, FALSE);
3694 			rddir_cache_rele(rdc);
3695 			goto top;
3696 		}
3697 		/*
3698 		 * Check to see if a readdir is required to
3699 		 * fill the entry.  If so, mark this entry
3700 		 * as being filled, remove our reference,
3701 		 * and branch to the code to fill the entry.
3702 		 */
3703 		if (rdc->flags & RDDIRREQ) {
3704 			rdc->flags &= ~RDDIRREQ;
3705 			rdc->flags |= RDDIR;
3706 			if (nrdc != NULL)
3707 				rddir_cache_rele(nrdc);
3708 			nrdc = rdc;
3709 			mutex_exit(&rp->r_statelock);
3710 			goto bottom;
3711 		}
3712 #ifdef DEBUG
3713 		if (!missed)
3714 			nfs3_readdir_cache_hits++;
3715 #endif
3716 		/*
3717 		 * If an error occurred while attempting
3718 		 * to fill the cache entry, just return it.
3719 		 */
3720 		if (rdc->error) {
3721 			error = rdc->error;
3722 			mutex_exit(&rp->r_statelock);
3723 			rddir_cache_rele(rdc);
3724 			if (nrdc != NULL)
3725 				rddir_cache_rele(nrdc);
3726 			return (error);
3727 		}
3728 
3729 		/*
3730 		 * The cache entry is complete and good,
3731 		 * copyout the dirent structs to the calling
3732 		 * thread.
3733 		 */
3734 		error = uiomove(rdc->entries, rdc->entlen, UIO_READ, uiop);
3735 
3736 		/*
3737 		 * If no error occurred during the copyout,
3738 		 * update the offset in the uio struct to
3739 		 * contain the value of the next cookie
3740 		 * and set the eof value appropriately.
3741 		 */
3742 		if (!error) {
3743 			uiop->uio_loffset = rdc->nfs3_ncookie;
3744 			if (eofp)
3745 				*eofp = rdc->eof;
3746 		}
3747 
3748 		/*
3749 		 * Decide whether to do readahead.
3750 		 *
3751 		 * Don't if have already read to the end of
3752 		 * directory.  There is nothing more to read.
3753 		 *
3754 		 * Don't if the application is not doing
3755 		 * lookups in the directory.  The readahead
3756 		 * is only effective if the application can
3757 		 * be doing work while an async thread is
3758 		 * handling the over the wire request.
3759 		 */
3760 		if (rdc->eof) {
3761 			rp->r_direof = rdc;
3762 			doreadahead = FALSE;
3763 		} else if (!(rp->r_flags & RLOOKUP))
3764 			doreadahead = FALSE;
3765 		else
3766 			doreadahead = TRUE;
3767 
3768 		if (!doreadahead) {
3769 			mutex_exit(&rp->r_statelock);
3770 			rddir_cache_rele(rdc);
3771 			if (nrdc != NULL)
3772 				rddir_cache_rele(nrdc);
3773 			return (error);
3774 		}
3775 
3776 		/*
3777 		 * Check to see whether we found an entry
3778 		 * for the readahead.  If so, we don't need
3779 		 * to do anything further, so free the new
3780 		 * entry if one was allocated.  Otherwise,
3781 		 * allocate a new entry, add it to the cache,
3782 		 * and then initiate an asynchronous readdir
3783 		 * operation to fill it.
3784 		 */
3785 		srdc.nfs3_cookie = rdc->nfs3_ncookie;
3786 		srdc.buflen = count;
3787 		rrdc = avl_find(&rp->r_dir, &srdc, &where);
3788 		if (rrdc != NULL) {
3789 			if (nrdc != NULL)
3790 				rddir_cache_rele(nrdc);
3791 		} else {
3792 			if (nrdc != NULL)
3793 				rrdc = nrdc;
3794 			else {
3795 				rrdc = rddir_cache_alloc(KM_NOSLEEP);
3796 			}
3797 			if (rrdc != NULL) {
3798 				rrdc->nfs3_cookie = rdc->nfs3_ncookie;
3799 				rrdc->buflen = count;
3800 				avl_insert(&rp->r_dir, rrdc, where);
3801 				rddir_cache_hold(rrdc);
3802 				mutex_exit(&rp->r_statelock);
3803 				rddir_cache_rele(rdc);
3804 #ifdef DEBUG
3805 				nfs3_readdir_readahead++;
3806 #endif
3807 				nfs_async_readdir(vp, rrdc, cr, do_nfs3readdir);
3808 				return (error);
3809 			}
3810 		}
3811 
3812 		mutex_exit(&rp->r_statelock);
3813 		rddir_cache_rele(rdc);
3814 		return (error);
3815 	}
3816 
3817 	/*
3818 	 * Didn't find an entry in the cache.  Construct a new empty
3819 	 * entry and link it into the cache.  Other processes attempting
3820 	 * to access this entry will need to wait until it is filled in.
3821 	 *
3822 	 * Since kmem_alloc may block, another pass through the cache
3823 	 * will need to be taken to make sure that another process
3824 	 * hasn't already added an entry to the cache for this request.
3825 	 */
3826 	if (nrdc == NULL) {
3827 		mutex_exit(&rp->r_statelock);
3828 		nrdc = rddir_cache_alloc(KM_SLEEP);
3829 		nrdc->nfs3_cookie = uiop->uio_loffset;
3830 		nrdc->buflen = count;
3831 		goto top;
3832 	}
3833 
3834 	/*
3835 	 * Add this entry to the cache.
3836 	 */
3837 	avl_insert(&rp->r_dir, nrdc, where);
3838 	rddir_cache_hold(nrdc);
3839 	mutex_exit(&rp->r_statelock);
3840 
3841 bottom:
3842 #ifdef DEBUG
3843 	missed = 1;
3844 	nfs3_readdir_cache_misses++;
3845 #endif
3846 	/*
3847 	 * Do the readdir.  This routine decides whether to use
3848 	 * READDIR or READDIRPLUS.
3849 	 */
3850 	error = do_nfs3readdir(vp, nrdc, cr);
3851 
3852 	/*
3853 	 * If this operation failed, just return the error which occurred.
3854 	 */
3855 	if (error != 0)
3856 		return (error);
3857 
3858 	/*
3859 	 * Since the RPC operation will have taken sometime and blocked
3860 	 * this process, another pass through the cache will need to be
3861 	 * taken to find the correct cache entry.  It is possible that
3862 	 * the correct cache entry will not be there (although one was
3863 	 * added) because the directory changed during the RPC operation
3864 	 * and the readdir cache was flushed.  In this case, just start
3865 	 * over.  It is hoped that this will not happen too often... :-)
3866 	 */
3867 	nrdc = NULL;
3868 	goto top;
3869 	/* NOTREACHED */
3870 }
3871 
3872 static int
3873 do_nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3874 {
3875 	int error;
3876 	rnode_t *rp;
3877 	mntinfo_t *mi;
3878 
3879 	rp = VTOR(vp);
3880 	mi = VTOMI(vp);
3881 	ASSERT(nfs_zone() == mi->mi_zone);
3882 	/*
3883 	 * Issue the proper request.
3884 	 *
3885 	 * If the server does not support READDIRPLUS, then use READDIR.
3886 	 *
3887 	 * Otherwise --
3888 	 * Issue a READDIRPLUS if reading to fill an empty cache or if
3889 	 * an application has performed a lookup in the directory which
3890 	 * required an over the wire lookup.  The use of READDIRPLUS
3891 	 * will help to (re)populate the DNLC.
3892 	 */
3893 	if (!(mi->mi_flags & MI_READDIRONLY) &&
3894 	    (rp->r_flags & (RLOOKUP | RREADDIRPLUS))) {
3895 		if (rp->r_flags & RREADDIRPLUS) {
3896 			mutex_enter(&rp->r_statelock);
3897 			rp->r_flags &= ~RREADDIRPLUS;
3898 			mutex_exit(&rp->r_statelock);
3899 		}
3900 		nfs3readdirplus(vp, rdc, cr);
3901 		if (rdc->error == EOPNOTSUPP)
3902 			nfs3readdir(vp, rdc, cr);
3903 	} else
3904 		nfs3readdir(vp, rdc, cr);
3905 
3906 	mutex_enter(&rp->r_statelock);
3907 	rdc->flags &= ~RDDIR;
3908 	if (rdc->flags & RDDIRWAIT) {
3909 		rdc->flags &= ~RDDIRWAIT;
3910 		cv_broadcast(&rdc->cv);
3911 	}
3912 	error = rdc->error;
3913 	if (error)
3914 		rdc->flags |= RDDIRREQ;
3915 	mutex_exit(&rp->r_statelock);
3916 
3917 	rddir_cache_rele(rdc);
3918 
3919 	return (error);
3920 }
3921 
3922 static void
3923 nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3924 {
3925 	int error;
3926 	READDIR3args args;
3927 	READDIR3vres res;
3928 	vattr_t dva;
3929 	rnode_t *rp;
3930 	int douprintf;
3931 	failinfo_t fi, *fip = NULL;
3932 	mntinfo_t *mi;
3933 	hrtime_t t;
3934 
3935 	rp = VTOR(vp);
3936 	mi = VTOMI(vp);
3937 	ASSERT(nfs_zone() == mi->mi_zone);
3938 
3939 	args.dir = *RTOFH3(rp);
3940 	args.cookie = (cookie3)rdc->nfs3_cookie;
3941 	args.cookieverf = rp->r_cookieverf;
3942 	args.count = rdc->buflen;
3943 
3944 	/*
3945 	 * NFS client failover support
3946 	 * suppress failover unless we have a zero cookie
3947 	 */
3948 	if (args.cookie == (cookie3) 0) {
3949 		fi.vp = vp;
3950 		fi.fhp = (caddr_t)&args.dir;
3951 		fi.copyproc = nfs3copyfh;
3952 		fi.lookupproc = nfs3lookup;
3953 		fi.xattrdirproc = acl_getxattrdir3;
3954 		fip = &fi;
3955 	}
3956 
3957 #ifdef DEBUG
3958 	rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
3959 #else
3960 	rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
3961 #endif
3962 
3963 	res.entries = (dirent64_t *)rdc->entries;
3964 	res.entries_size = rdc->buflen;
3965 	res.dir_attributes.fres.vap = &dva;
3966 	res.dir_attributes.fres.vp = vp;
3967 	res.loff = rdc->nfs3_cookie;
3968 
3969 	douprintf = 1;
3970 
3971 	if (mi->mi_io_kstats) {
3972 		mutex_enter(&mi->mi_lock);
3973 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
3974 		mutex_exit(&mi->mi_lock);
3975 	}
3976 
3977 	t = gethrtime();
3978 
3979 	error = rfs3call(VTOMI(vp), NFSPROC3_READDIR,
3980 	    xdr_READDIR3args, (caddr_t)&args,
3981 	    xdr_READDIR3vres, (caddr_t)&res, cr,
3982 	    &douprintf, &res.status, 0, fip);
3983 
3984 	if (mi->mi_io_kstats) {
3985 		mutex_enter(&mi->mi_lock);
3986 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
3987 		mutex_exit(&mi->mi_lock);
3988 	}
3989 
3990 	if (error)
3991 		goto err;
3992 
3993 	nfs3_cache_post_op_vattr(vp, &res.dir_attributes, t, cr);
3994 
3995 	error = geterrno3(res.status);
3996 	if (error) {
3997 		PURGE_STALE_FH(error, vp, cr);
3998 		goto err;
3999 	}
4000 
4001 	if (mi->mi_io_kstats) {
4002 		mutex_enter(&mi->mi_lock);
4003 		KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4004 		KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4005 		mutex_exit(&mi->mi_lock);
4006 	}
4007 
4008 	rdc->nfs3_ncookie = res.loff;
4009 	rp->r_cookieverf = res.cookieverf;
4010 	rdc->eof = res.eof ? 1 : 0;
4011 	rdc->entlen = res.size;
4012 	ASSERT(rdc->entlen <= rdc->buflen);
4013 	rdc->error = 0;
4014 	return;
4015 
4016 err:
4017 	kmem_free(rdc->entries, rdc->buflen);
4018 	rdc->entries = NULL;
4019 	rdc->error = error;
4020 }
4021 
4022 /*
4023  * Read directory entries.
4024  * There are some weird things to look out for here.  The uio_loffset
4025  * field is either 0 or it is the offset returned from a previous
4026  * readdir.  It is an opaque value used by the server to find the
4027  * correct directory block to read. The count field is the number
4028  * of blocks to read on the server.  This is advisory only, the server
4029  * may return only one block's worth of entries.  Entries may be compressed
4030  * on the server.
4031  */
4032 static void
4033 nfs3readdirplus(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
4034 {
4035 	int error;
4036 	READDIRPLUS3args args;
4037 	READDIRPLUS3vres res;
4038 	vattr_t dva;
4039 	rnode_t *rp;
4040 	mntinfo_t *mi;
4041 	int douprintf;
4042 	failinfo_t fi, *fip = NULL;
4043 
4044 	rp = VTOR(vp);
4045 	mi = VTOMI(vp);
4046 	ASSERT(nfs_zone() == mi->mi_zone);
4047 
4048 	args.dir = *RTOFH3(rp);
4049 	args.cookie = (cookie3)rdc->nfs3_cookie;
4050 	args.cookieverf = rp->r_cookieverf;
4051 	args.dircount = rdc->buflen;
4052 	args.maxcount = mi->mi_tsize;
4053 
4054 	/*
4055 	 * NFS client failover support
4056 	 * suppress failover unless we have a zero cookie
4057 	 */
4058 	if (args.cookie == (cookie3)0) {
4059 		fi.vp = vp;
4060 		fi.fhp = (caddr_t)&args.dir;
4061 		fi.copyproc = nfs3copyfh;
4062 		fi.lookupproc = nfs3lookup;
4063 		fi.xattrdirproc = acl_getxattrdir3;
4064 		fip = &fi;
4065 	}
4066 
4067 #ifdef DEBUG
4068 	rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
4069 #else
4070 	rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
4071 #endif
4072 
4073 	res.entries = (dirent64_t *)rdc->entries;
4074 	res.entries_size = rdc->buflen;
4075 	res.dir_attributes.fres.vap = &dva;
4076 	res.dir_attributes.fres.vp = vp;
4077 	res.loff = rdc->nfs3_cookie;
4078 	res.credentials = cr;
4079 
4080 	douprintf = 1;
4081 
4082 	if (mi->mi_io_kstats) {
4083 		mutex_enter(&mi->mi_lock);
4084 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
4085 		mutex_exit(&mi->mi_lock);
4086 	}
4087 
4088 	res.time = gethrtime();
4089 
4090 	error = rfs3call(mi, NFSPROC3_READDIRPLUS,
4091 	    xdr_READDIRPLUS3args, (caddr_t)&args,
4092 	    xdr_READDIRPLUS3vres, (caddr_t)&res, cr,
4093 	    &douprintf, &res.status, 0, fip);
4094 
4095 	if (mi->mi_io_kstats) {
4096 		mutex_enter(&mi->mi_lock);
4097 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
4098 		mutex_exit(&mi->mi_lock);
4099 	}
4100 
4101 	if (error) {
4102 		goto err;
4103 	}
4104 
4105 	nfs3_cache_post_op_vattr(vp, &res.dir_attributes, res.time, cr);
4106 
4107 	error = geterrno3(res.status);
4108 	if (error) {
4109 		PURGE_STALE_FH(error, vp, cr);
4110 		if (error == EOPNOTSUPP) {
4111 			mutex_enter(&mi->mi_lock);
4112 			mi->mi_flags |= MI_READDIRONLY;
4113 			mutex_exit(&mi->mi_lock);
4114 		}
4115 		goto err;
4116 	}
4117 
4118 	if (mi->mi_io_kstats) {
4119 		mutex_enter(&mi->mi_lock);
4120 		KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4121 		KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4122 		mutex_exit(&mi->mi_lock);
4123 	}
4124 
4125 	rdc->nfs3_ncookie = res.loff;
4126 	rp->r_cookieverf = res.cookieverf;
4127 	rdc->eof = res.eof ? 1 : 0;
4128 	rdc->entlen = res.size;
4129 	ASSERT(rdc->entlen <= rdc->buflen);
4130 	rdc->error = 0;
4131 
4132 	return;
4133 
4134 err:
4135 	kmem_free(rdc->entries, rdc->buflen);
4136 	rdc->entries = NULL;
4137 	rdc->error = error;
4138 }
4139 
4140 #ifdef DEBUG
4141 static int nfs3_bio_do_stop = 0;
4142 #endif
4143 
4144 static int
4145 nfs3_bio(struct buf *bp, stable_how *stab_comm, cred_t *cr)
4146 {
4147 	rnode_t *rp = VTOR(bp->b_vp);
4148 	int count;
4149 	int error;
4150 	cred_t *cred;
4151 	offset_t offset;
4152 
4153 	ASSERT(nfs_zone() == VTOMI(bp->b_vp)->mi_zone);
4154 	offset = ldbtob(bp->b_lblkno);
4155 
4156 	DTRACE_IO1(start, struct buf *, bp);
4157 
4158 	if (bp->b_flags & B_READ) {
4159 		mutex_enter(&rp->r_statelock);
4160 		if (rp->r_cred != NULL) {
4161 			cred = rp->r_cred;
4162 			crhold(cred);
4163 		} else {
4164 			rp->r_cred = cr;
4165 			crhold(cr);
4166 			cred = cr;
4167 			crhold(cred);
4168 		}
4169 		mutex_exit(&rp->r_statelock);
4170 	read_again:
4171 		error = bp->b_error = nfs3read(bp->b_vp, bp->b_un.b_addr,
4172 		    offset, bp->b_bcount, &bp->b_resid, cred);
4173 		crfree(cred);
4174 		if (!error) {
4175 			if (bp->b_resid) {
4176 				/*
4177 				 * Didn't get it all because we hit EOF,
4178 				 * zero all the memory beyond the EOF.
4179 				 */
4180 				/* bzero(rdaddr + */
4181 				bzero(bp->b_un.b_addr +
4182 				    bp->b_bcount - bp->b_resid, bp->b_resid);
4183 			}
4184 			mutex_enter(&rp->r_statelock);
4185 			if (bp->b_resid == bp->b_bcount &&
4186 			    offset >= rp->r_size) {
4187 				/*
4188 				 * We didn't read anything at all as we are
4189 				 * past EOF.  Return an error indicator back
4190 				 * but don't destroy the pages (yet).
4191 				 */
4192 				error = NFS_EOF;
4193 			}
4194 			mutex_exit(&rp->r_statelock);
4195 		} else if (error == EACCES) {
4196 			mutex_enter(&rp->r_statelock);
4197 			if (cred != cr) {
4198 				if (rp->r_cred != NULL)
4199 					crfree(rp->r_cred);
4200 				rp->r_cred = cr;
4201 				crhold(cr);
4202 				cred = cr;
4203 				crhold(cred);
4204 				mutex_exit(&rp->r_statelock);
4205 				goto read_again;
4206 			}
4207 			mutex_exit(&rp->r_statelock);
4208 		}
4209 	} else {
4210 		if (!(rp->r_flags & RSTALE)) {
4211 			mutex_enter(&rp->r_statelock);
4212 			if (rp->r_cred != NULL) {
4213 				cred = rp->r_cred;
4214 				crhold(cred);
4215 			} else {
4216 				rp->r_cred = cr;
4217 				crhold(cr);
4218 				cred = cr;
4219 				crhold(cred);
4220 			}
4221 			mutex_exit(&rp->r_statelock);
4222 		write_again:
4223 			mutex_enter(&rp->r_statelock);
4224 			count = MIN(bp->b_bcount, rp->r_size - offset);
4225 			mutex_exit(&rp->r_statelock);
4226 			if (count < 0)
4227 				cmn_err(CE_PANIC, "nfs3_bio: write count < 0");
4228 #ifdef DEBUG
4229 			if (count == 0) {
4230 				zcmn_err(getzoneid(), CE_WARN,
4231 				    "nfs3_bio: zero length write at %lld",
4232 				    offset);
4233 				nfs_printfhandle(&rp->r_fh);
4234 				if (nfs3_bio_do_stop)
4235 					debug_enter("nfs3_bio");
4236 			}
4237 #endif
4238 			error = nfs3write(bp->b_vp, bp->b_un.b_addr, offset,
4239 			    count, cred, stab_comm);
4240 			if (error == EACCES) {
4241 				mutex_enter(&rp->r_statelock);
4242 				if (cred != cr) {
4243 					if (rp->r_cred != NULL)
4244 						crfree(rp->r_cred);
4245 					rp->r_cred = cr;
4246 					crhold(cr);
4247 					crfree(cred);
4248 					cred = cr;
4249 					crhold(cred);
4250 					mutex_exit(&rp->r_statelock);
4251 					goto write_again;
4252 				}
4253 				mutex_exit(&rp->r_statelock);
4254 			}
4255 			bp->b_error = error;
4256 			if (error && error != EINTR) {
4257 				/*
4258 				 * Don't print EDQUOT errors on the console.
4259 				 * Don't print asynchronous EACCES errors.
4260 				 * Don't print EFBIG errors.
4261 				 * Print all other write errors.
4262 				 */
4263 				if (error != EDQUOT && error != EFBIG &&
4264 				    (error != EACCES ||
4265 				    !(bp->b_flags & B_ASYNC)))
4266 					nfs_write_error(bp->b_vp, error, cred);
4267 				/*
4268 				 * Update r_error and r_flags as appropriate.
4269 				 * If the error was ESTALE, then mark the
4270 				 * rnode as not being writeable and save
4271 				 * the error status.  Otherwise, save any
4272 				 * errors which occur from asynchronous
4273 				 * page invalidations.  Any errors occurring
4274 				 * from other operations should be saved
4275 				 * by the caller.
4276 				 */
4277 				mutex_enter(&rp->r_statelock);
4278 				if (error == ESTALE) {
4279 					rp->r_flags |= RSTALE;
4280 					if (!rp->r_error)
4281 						rp->r_error = error;
4282 				} else if (!rp->r_error &&
4283 				    (bp->b_flags &
4284 				    (B_INVAL|B_FORCE|B_ASYNC)) ==
4285 				    (B_INVAL|B_FORCE|B_ASYNC)) {
4286 					rp->r_error = error;
4287 				}
4288 				mutex_exit(&rp->r_statelock);
4289 			}
4290 			crfree(cred);
4291 		} else
4292 			error = rp->r_error;
4293 	}
4294 
4295 	if (error != 0 && error != NFS_EOF)
4296 		bp->b_flags |= B_ERROR;
4297 
4298 	DTRACE_IO1(done, struct buf *, bp);
4299 
4300 	return (error);
4301 }
4302 
4303 static int
4304 nfs3_fid(vnode_t *vp, fid_t *fidp)
4305 {
4306 	rnode_t *rp;
4307 
4308 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4309 		return (EIO);
4310 	rp = VTOR(vp);
4311 
4312 	if (fidp->fid_len < (ushort_t)rp->r_fh.fh_len) {
4313 		fidp->fid_len = rp->r_fh.fh_len;
4314 		return (ENOSPC);
4315 	}
4316 	fidp->fid_len = rp->r_fh.fh_len;
4317 	bcopy(rp->r_fh.fh_buf, fidp->fid_data, fidp->fid_len);
4318 	return (0);
4319 }
4320 
4321 /* ARGSUSED2 */
4322 static int
4323 nfs3_rwlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4324 {
4325 	rnode_t *rp = VTOR(vp);
4326 
4327 	if (!write_lock) {
4328 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4329 		return (V_WRITELOCK_FALSE);
4330 	}
4331 
4332 	if ((rp->r_flags & RDIRECTIO) || (VTOMI(vp)->mi_flags & MI_DIRECTIO)) {
4333 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4334 		if (rp->r_mapcnt == 0 && !vn_has_cached_data(vp))
4335 			return (V_WRITELOCK_FALSE);
4336 		nfs_rw_exit(&rp->r_rwlock);
4337 	}
4338 
4339 	(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, FALSE);
4340 	return (V_WRITELOCK_TRUE);
4341 }
4342 
4343 /* ARGSUSED */
4344 static void
4345 nfs3_rwunlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4346 {
4347 	rnode_t *rp = VTOR(vp);
4348 
4349 	nfs_rw_exit(&rp->r_rwlock);
4350 }
4351 
4352 /* ARGSUSED */
4353 static int
4354 nfs3_seek(vnode_t *vp, offset_t ooff, offset_t *noffp)
4355 {
4356 
4357 	/*
4358 	 * Because we stuff the readdir cookie into the offset field
4359 	 * someone may attempt to do an lseek with the cookie which
4360 	 * we want to succeed.
4361 	 */
4362 	if (vp->v_type == VDIR)
4363 		return (0);
4364 	if (*noffp < 0)
4365 		return (EINVAL);
4366 	return (0);
4367 }
4368 
4369 /*
4370  * number of nfs3_bsize blocks to read ahead.
4371  */
4372 static int nfs3_nra = 4;
4373 
4374 #ifdef DEBUG
4375 static int nfs3_lostpage = 0;	/* number of times we lost original page */
4376 #endif
4377 
4378 /*
4379  * Return all the pages from [off..off+len) in file
4380  */
4381 static int
4382 nfs3_getpage(vnode_t *vp, offset_t off, size_t len, uint_t *protp,
4383 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4384 	enum seg_rw rw, cred_t *cr)
4385 {
4386 	rnode_t *rp;
4387 	int error;
4388 	mntinfo_t *mi;
4389 
4390 	if (vp->v_flag & VNOMAP)
4391 		return (ENOSYS);
4392 
4393 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4394 		return (EIO);
4395 	if (protp != NULL)
4396 		*protp = PROT_ALL;
4397 
4398 	/*
4399 	 * Now valididate that the caches are up to date.
4400 	 */
4401 	error = nfs3_validate_caches(vp, cr);
4402 	if (error)
4403 		return (error);
4404 
4405 	rp = VTOR(vp);
4406 	mi = VTOMI(vp);
4407 retry:
4408 	mutex_enter(&rp->r_statelock);
4409 
4410 	/*
4411 	 * Don't create dirty pages faster than they
4412 	 * can be cleaned so that the system doesn't
4413 	 * get imbalanced.  If the async queue is
4414 	 * maxed out, then wait for it to drain before
4415 	 * creating more dirty pages.  Also, wait for
4416 	 * any threads doing pagewalks in the vop_getattr
4417 	 * entry points so that they don't block for
4418 	 * long periods.
4419 	 */
4420 	if (rw == S_CREATE) {
4421 		while ((mi->mi_max_threads != 0 &&
4422 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
4423 		    rp->r_gcount > 0)
4424 			cv_wait(&rp->r_cv, &rp->r_statelock);
4425 	}
4426 
4427 	/*
4428 	 * If we are getting called as a side effect of an nfs_write()
4429 	 * operation the local file size might not be extended yet.
4430 	 * In this case we want to be able to return pages of zeroes.
4431 	 */
4432 	if (off + len > rp->r_size + PAGEOFFSET && seg != segkmap) {
4433 		mutex_exit(&rp->r_statelock);
4434 		return (EFAULT);		/* beyond EOF */
4435 	}
4436 
4437 	mutex_exit(&rp->r_statelock);
4438 
4439 	if (len <= PAGESIZE) {
4440 		error = nfs3_getapage(vp, off, len, protp, pl, plsz,
4441 		    seg, addr, rw, cr);
4442 	} else {
4443 		error = pvn_getpages(nfs3_getapage, vp, off, len, protp,
4444 		    pl, plsz, seg, addr, rw, cr);
4445 	}
4446 
4447 	switch (error) {
4448 	case NFS_EOF:
4449 		nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
4450 		goto retry;
4451 	case ESTALE:
4452 		PURGE_STALE_FH(error, vp, cr);
4453 	}
4454 
4455 	return (error);
4456 }
4457 
4458 /*
4459  * Called from pvn_getpages or nfs3_getpage to get a particular page.
4460  */
4461 /* ARGSUSED */
4462 static int
4463 nfs3_getapage(vnode_t *vp, u_offset_t off, size_t len, uint_t *protp,
4464 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4465 	enum seg_rw rw, cred_t *cr)
4466 {
4467 	rnode_t *rp;
4468 	uint_t bsize;
4469 	struct buf *bp;
4470 	page_t *pp;
4471 	u_offset_t lbn;
4472 	u_offset_t io_off;
4473 	u_offset_t blkoff;
4474 	u_offset_t rablkoff;
4475 	size_t io_len;
4476 	uint_t blksize;
4477 	int error;
4478 	int readahead;
4479 	int readahead_issued = 0;
4480 	int ra_window; /* readahead window */
4481 	page_t *pagefound;
4482 	page_t *savepp;
4483 
4484 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4485 		return (EIO);
4486 	rp = VTOR(vp);
4487 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4488 
4489 reread:
4490 	bp = NULL;
4491 	pp = NULL;
4492 	pagefound = NULL;
4493 
4494 	if (pl != NULL)
4495 		pl[0] = NULL;
4496 
4497 	error = 0;
4498 	lbn = off / bsize;
4499 	blkoff = lbn * bsize;
4500 
4501 	/*
4502 	 * Queueing up the readahead before doing the synchronous read
4503 	 * results in a significant increase in read throughput because
4504 	 * of the increased parallelism between the async threads and
4505 	 * the process context.
4506 	 */
4507 	if ((off & ((vp->v_vfsp->vfs_bsize) - 1)) == 0 &&
4508 	    rw != S_CREATE &&
4509 	    !(vp->v_flag & VNOCACHE)) {
4510 		mutex_enter(&rp->r_statelock);
4511 
4512 		/*
4513 		 * Calculate the number of readaheads to do.
4514 		 * a) No readaheads at offset = 0.
4515 		 * b) Do maximum(nfs3_nra) readaheads when the readahead
4516 		 *    window is closed.
4517 		 * c) Do readaheads between 1 to (nfs3_nra - 1) depending
4518 		 *    upon how far the readahead window is open or close.
4519 		 * d) No readaheads if rp->r_nextr is not within the scope
4520 		 *    of the readahead window (random i/o).
4521 		 */
4522 
4523 		if (off == 0)
4524 			readahead = 0;
4525 		else if (blkoff == rp->r_nextr)
4526 			readahead = nfs3_nra;
4527 		else if (rp->r_nextr > blkoff &&
4528 				((ra_window = (rp->r_nextr - blkoff) / bsize)
4529 					<= (nfs3_nra - 1)))
4530 			readahead = nfs3_nra - ra_window;
4531 		else
4532 			readahead = 0;
4533 
4534 		rablkoff = rp->r_nextr;
4535 		while (readahead > 0 && rablkoff + bsize < rp->r_size) {
4536 			mutex_exit(&rp->r_statelock);
4537 			if (nfs_async_readahead(vp, rablkoff + bsize,
4538 			    addr + (rablkoff + bsize - off), seg, cr,
4539 			    nfs3_readahead) < 0) {
4540 				mutex_enter(&rp->r_statelock);
4541 				break;
4542 			}
4543 			readahead--;
4544 			rablkoff += bsize;
4545 			/*
4546 			 * Indicate that we did a readahead so
4547 			 * readahead offset is not updated
4548 			 * by the synchronous read below.
4549 			 */
4550 			readahead_issued = 1;
4551 			mutex_enter(&rp->r_statelock);
4552 			/*
4553 			 * set readahead offset to
4554 			 * offset of last async readahead
4555 			 * request.
4556 			 */
4557 			rp->r_nextr = rablkoff;
4558 		}
4559 		mutex_exit(&rp->r_statelock);
4560 	}
4561 
4562 again:
4563 	if ((pagefound = page_exists(vp, off)) == NULL) {
4564 		if (pl == NULL) {
4565 			(void) nfs_async_readahead(vp, blkoff, addr, seg, cr,
4566 			    nfs3_readahead);
4567 		} else if (rw == S_CREATE) {
4568 			/*
4569 			 * Block for this page is not allocated, or the offset
4570 			 * is beyond the current allocation size, or we're
4571 			 * allocating a swap slot and the page was not found,
4572 			 * so allocate it and return a zero page.
4573 			 */
4574 			if ((pp = page_create_va(vp, off,
4575 			    PAGESIZE, PG_WAIT, seg, addr)) == NULL)
4576 				cmn_err(CE_PANIC, "nfs3_getapage: page_create");
4577 			io_len = PAGESIZE;
4578 			mutex_enter(&rp->r_statelock);
4579 			rp->r_nextr = off + PAGESIZE;
4580 			mutex_exit(&rp->r_statelock);
4581 		} else {
4582 			/*
4583 			 * Need to go to server to get a BLOCK, exception to
4584 			 * that being while reading at offset = 0 or doing
4585 			 * random i/o, in that case read only a PAGE.
4586 			 */
4587 			mutex_enter(&rp->r_statelock);
4588 			if (blkoff < rp->r_size &&
4589 			    blkoff + bsize >= rp->r_size) {
4590 				/*
4591 				 * If only a block or less is left in
4592 				 * the file, read all that is remaining.
4593 				 */
4594 				if (rp->r_size <= off) {
4595 					/*
4596 					 * Trying to access beyond EOF,
4597 					 * set up to get at least one page.
4598 					 */
4599 					blksize = off + PAGESIZE - blkoff;
4600 				} else
4601 					blksize = rp->r_size - blkoff;
4602 			} else if ((off == 0) ||
4603 				(off != rp->r_nextr && !readahead_issued)) {
4604 				blksize = PAGESIZE;
4605 				blkoff = off; /* block = page here */
4606 			} else
4607 				blksize = bsize;
4608 			mutex_exit(&rp->r_statelock);
4609 
4610 			pp = pvn_read_kluster(vp, off, seg, addr, &io_off,
4611 			    &io_len, blkoff, blksize, 0);
4612 
4613 			/*
4614 			 * Some other thread has entered the page,
4615 			 * so just use it.
4616 			 */
4617 			if (pp == NULL)
4618 				goto again;
4619 
4620 			/*
4621 			 * Now round the request size up to page boundaries.
4622 			 * This ensures that the entire page will be
4623 			 * initialized to zeroes if EOF is encountered.
4624 			 */
4625 			io_len = ptob(btopr(io_len));
4626 
4627 			bp = pageio_setup(pp, io_len, vp, B_READ);
4628 			ASSERT(bp != NULL);
4629 
4630 			/*
4631 			 * pageio_setup should have set b_addr to 0.  This
4632 			 * is correct since we want to do I/O on a page
4633 			 * boundary.  bp_mapin will use this addr to calculate
4634 			 * an offset, and then set b_addr to the kernel virtual
4635 			 * address it allocated for us.
4636 			 */
4637 			ASSERT(bp->b_un.b_addr == 0);
4638 
4639 			bp->b_edev = 0;
4640 			bp->b_dev = 0;
4641 			bp->b_lblkno = lbtodb(io_off);
4642 			bp->b_file = vp;
4643 			bp->b_offset = (offset_t)off;
4644 			bp_mapin(bp);
4645 
4646 			/*
4647 			 * If doing a write beyond what we believe is EOF,
4648 			 * don't bother trying to read the pages from the
4649 			 * server, we'll just zero the pages here.  We
4650 			 * don't check that the rw flag is S_WRITE here
4651 			 * because some implementations may attempt a
4652 			 * read access to the buffer before copying data.
4653 			 */
4654 			mutex_enter(&rp->r_statelock);
4655 			if (io_off >= rp->r_size && seg == segkmap) {
4656 				mutex_exit(&rp->r_statelock);
4657 				bzero(bp->b_un.b_addr, io_len);
4658 			} else {
4659 				mutex_exit(&rp->r_statelock);
4660 				error = nfs3_bio(bp, NULL, cr);
4661 			}
4662 
4663 			/*
4664 			 * Unmap the buffer before freeing it.
4665 			 */
4666 			bp_mapout(bp);
4667 			pageio_done(bp);
4668 
4669 			savepp = pp;
4670 			do {
4671 				pp->p_fsdata = C_NOCOMMIT;
4672 			} while ((pp = pp->p_next) != savepp);
4673 
4674 			if (error == NFS_EOF) {
4675 				/*
4676 				 * If doing a write system call just return
4677 				 * zeroed pages, else user tried to get pages
4678 				 * beyond EOF, return error.  We don't check
4679 				 * that the rw flag is S_WRITE here because
4680 				 * some implementations may attempt a read
4681 				 * access to the buffer before copying data.
4682 				 */
4683 				if (seg == segkmap)
4684 					error = 0;
4685 				else
4686 					error = EFAULT;
4687 			}
4688 
4689 			if (!readahead_issued && !error) {
4690 			    mutex_enter(&rp->r_statelock);
4691 			    rp->r_nextr = io_off + io_len;
4692 			    mutex_exit(&rp->r_statelock);
4693 			}
4694 		}
4695 	}
4696 
4697 out:
4698 	if (pl == NULL)
4699 		return (error);
4700 
4701 	if (error) {
4702 		if (pp != NULL)
4703 			pvn_read_done(pp, B_ERROR);
4704 		return (error);
4705 	}
4706 
4707 	if (pagefound) {
4708 		se_t se = (rw == S_CREATE ? SE_EXCL : SE_SHARED);
4709 
4710 		/*
4711 		 * Page exists in the cache, acquire the appropriate lock.
4712 		 * If this fails, start all over again.
4713 		 */
4714 		if ((pp = page_lookup(vp, off, se)) == NULL) {
4715 #ifdef DEBUG
4716 			nfs3_lostpage++;
4717 #endif
4718 			goto reread;
4719 		}
4720 		pl[0] = pp;
4721 		pl[1] = NULL;
4722 		return (0);
4723 	}
4724 
4725 	if (pp != NULL)
4726 		pvn_plist_init(pp, pl, plsz, off, io_len, rw);
4727 
4728 	return (error);
4729 }
4730 
4731 static void
4732 nfs3_readahead(vnode_t *vp, u_offset_t blkoff, caddr_t addr, struct seg *seg,
4733 	cred_t *cr)
4734 {
4735 	int error;
4736 	page_t *pp;
4737 	u_offset_t io_off;
4738 	size_t io_len;
4739 	struct buf *bp;
4740 	uint_t bsize, blksize;
4741 	rnode_t *rp = VTOR(vp);
4742 	page_t *savepp;
4743 
4744 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
4745 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4746 
4747 	mutex_enter(&rp->r_statelock);
4748 	if (blkoff < rp->r_size && blkoff + bsize > rp->r_size) {
4749 		/*
4750 		 * If less than a block left in file read less
4751 		 * than a block.
4752 		 */
4753 		blksize = rp->r_size - blkoff;
4754 	} else
4755 		blksize = bsize;
4756 	mutex_exit(&rp->r_statelock);
4757 
4758 	pp = pvn_read_kluster(vp, blkoff, segkmap, addr,
4759 	    &io_off, &io_len, blkoff, blksize, 1);
4760 	/*
4761 	 * The isra flag passed to the kluster function is 1, we may have
4762 	 * gotten a return value of NULL for a variety of reasons (# of free
4763 	 * pages < minfree, someone entered the page on the vnode etc). In all
4764 	 * cases, we want to punt on the readahead.
4765 	 */
4766 	if (pp == NULL)
4767 		return;
4768 
4769 	/*
4770 	 * Now round the request size up to page boundaries.
4771 	 * This ensures that the entire page will be
4772 	 * initialized to zeroes if EOF is encountered.
4773 	 */
4774 	io_len = ptob(btopr(io_len));
4775 
4776 	bp = pageio_setup(pp, io_len, vp, B_READ);
4777 	ASSERT(bp != NULL);
4778 
4779 	/*
4780 	 * pageio_setup should have set b_addr to 0.  This is correct since
4781 	 * we want to do I/O on a page boundary. bp_mapin() will use this addr
4782 	 * to calculate an offset, and then set b_addr to the kernel virtual
4783 	 * address it allocated for us.
4784 	 */
4785 	ASSERT(bp->b_un.b_addr == 0);
4786 
4787 	bp->b_edev = 0;
4788 	bp->b_dev = 0;
4789 	bp->b_lblkno = lbtodb(io_off);
4790 	bp->b_file = vp;
4791 	bp->b_offset = (offset_t)blkoff;
4792 	bp_mapin(bp);
4793 
4794 	/*
4795 	 * If doing a write beyond what we believe is EOF, don't bother trying
4796 	 * to read the pages from the server, we'll just zero the pages here.
4797 	 * We don't check that the rw flag is S_WRITE here because some
4798 	 * implementations may attempt a read access to the buffer before
4799 	 * copying data.
4800 	 */
4801 	mutex_enter(&rp->r_statelock);
4802 	if (io_off >= rp->r_size && seg == segkmap) {
4803 		mutex_exit(&rp->r_statelock);
4804 		bzero(bp->b_un.b_addr, io_len);
4805 		error = 0;
4806 	} else {
4807 		mutex_exit(&rp->r_statelock);
4808 		error = nfs3_bio(bp, NULL, cr);
4809 		if (error == NFS_EOF)
4810 			error = 0;
4811 	}
4812 
4813 	/*
4814 	 * Unmap the buffer before freeing it.
4815 	 */
4816 	bp_mapout(bp);
4817 	pageio_done(bp);
4818 
4819 	savepp = pp;
4820 	do {
4821 		pp->p_fsdata = C_NOCOMMIT;
4822 	} while ((pp = pp->p_next) != savepp);
4823 
4824 	pvn_read_done(pp, error ? B_READ | B_ERROR : B_READ);
4825 
4826 	/*
4827 	 * In case of error set readahead offset
4828 	 * to the lowest offset.
4829 	 * pvn_read_done() calls VN_DISPOSE to destroy the pages
4830 	 */
4831 	if (error && rp->r_nextr > io_off) {
4832 		mutex_enter(&rp->r_statelock);
4833 		if (rp->r_nextr > io_off)
4834 			rp->r_nextr = io_off;
4835 		mutex_exit(&rp->r_statelock);
4836 	}
4837 }
4838 
4839 /*
4840  * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
4841  * If len == 0, do from off to EOF.
4842  *
4843  * The normal cases should be len == 0 && off == 0 (entire vp list),
4844  * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
4845  * (from pageout).
4846  */
4847 static int
4848 nfs3_putpage(vnode_t *vp, offset_t off, size_t len, int flags, cred_t *cr)
4849 {
4850 	int error;
4851 	rnode_t *rp;
4852 
4853 	ASSERT(cr != NULL);
4854 
4855 	/*
4856 	 * XXX - Why should this check be made here?
4857 	 */
4858 	if (vp->v_flag & VNOMAP)
4859 		return (ENOSYS);
4860 	if (len == 0 && !(flags & B_INVAL) && vn_is_readonly(vp))
4861 		return (0);
4862 	if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
4863 		return (EIO);
4864 
4865 	rp = VTOR(vp);
4866 	mutex_enter(&rp->r_statelock);
4867 	rp->r_count++;
4868 	mutex_exit(&rp->r_statelock);
4869 	error = nfs_putpages(vp, off, len, flags, cr);
4870 	mutex_enter(&rp->r_statelock);
4871 	rp->r_count--;
4872 	cv_broadcast(&rp->r_cv);
4873 	mutex_exit(&rp->r_statelock);
4874 
4875 	return (error);
4876 }
4877 
4878 /*
4879  * Write out a single page, possibly klustering adjacent dirty pages.
4880  */
4881 int
4882 nfs3_putapage(vnode_t *vp, page_t *pp, u_offset_t *offp, size_t *lenp,
4883 	int flags, cred_t *cr)
4884 {
4885 	u_offset_t io_off;
4886 	u_offset_t lbn_off;
4887 	u_offset_t lbn;
4888 	size_t io_len;
4889 	uint_t bsize;
4890 	int error;
4891 	rnode_t *rp;
4892 
4893 	ASSERT(!vn_is_readonly(vp));
4894 	ASSERT(pp != NULL);
4895 	ASSERT(cr != NULL);
4896 	ASSERT((flags & B_ASYNC) || nfs_zone() == VTOMI(vp)->mi_zone);
4897 
4898 	rp = VTOR(vp);
4899 	ASSERT(rp->r_count > 0);
4900 
4901 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4902 	lbn = pp->p_offset / bsize;
4903 	lbn_off = lbn * bsize;
4904 
4905 	/*
4906 	 * Find a kluster that fits in one block, or in
4907 	 * one page if pages are bigger than blocks.  If
4908 	 * there is less file space allocated than a whole
4909 	 * page, we'll shorten the i/o request below.
4910 	 */
4911 	pp = pvn_write_kluster(vp, pp, &io_off, &io_len, lbn_off,
4912 	    roundup(bsize, PAGESIZE), flags);
4913 
4914 	/*
4915 	 * pvn_write_kluster shouldn't have returned a page with offset
4916 	 * behind the original page we were given.  Verify that.
4917 	 */
4918 	ASSERT((pp->p_offset / bsize) >= lbn);
4919 
4920 	/*
4921 	 * Now pp will have the list of kept dirty pages marked for
4922 	 * write back.  It will also handle invalidation and freeing
4923 	 * of pages that are not dirty.  Check for page length rounding
4924 	 * problems.
4925 	 */
4926 	if (io_off + io_len > lbn_off + bsize) {
4927 		ASSERT((io_off + io_len) - (lbn_off + bsize) < PAGESIZE);
4928 		io_len = lbn_off + bsize - io_off;
4929 	}
4930 	/*
4931 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4932 	 * consistent value of r_size. RMODINPROGRESS is set in writerp().
4933 	 * When RMODINPROGRESS is set it indicates that a uiomove() is in
4934 	 * progress and the r_size has not been made consistent with the
4935 	 * new size of the file. When the uiomove() completes the r_size is
4936 	 * updated and the RMODINPROGRESS flag is cleared.
4937 	 *
4938 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4939 	 * consistent value of r_size. Without this handshaking, it is
4940 	 * possible that nfs(3)_bio() picks  up the old value of r_size
4941 	 * before the uiomove() in writerp() completes. This will result
4942 	 * in the write through nfs(3)_bio() being dropped.
4943 	 *
4944 	 * More precisely, there is a window between the time the uiomove()
4945 	 * completes and the time the r_size is updated. If a VOP_PUTPAGE()
4946 	 * operation intervenes in this window, the page will be picked up,
4947 	 * because it is dirty (it will be unlocked, unless it was
4948 	 * pagecreate'd). When the page is picked up as dirty, the dirty
4949 	 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
4950 	 * checked. This will still be the old size. Therefore the page will
4951 	 * not be written out. When segmap_release() calls VOP_PUTPAGE(),
4952 	 * the page will be found to be clean and the write will be dropped.
4953 	 */
4954 	if (rp->r_flags & RMODINPROGRESS) {
4955 		mutex_enter(&rp->r_statelock);
4956 		if ((rp->r_flags & RMODINPROGRESS) &&
4957 		    rp->r_modaddr + MAXBSIZE > io_off &&
4958 		    rp->r_modaddr < io_off + io_len) {
4959 			page_t *plist;
4960 			/*
4961 			 * A write is in progress for this region of the file.
4962 			 * If we did not detect RMODINPROGRESS here then this
4963 			 * path through nfs_putapage() would eventually go to
4964 			 * nfs(3)_bio() and may not write out all of the data
4965 			 * in the pages. We end up losing data. So we decide
4966 			 * to set the modified bit on each page in the page
4967 			 * list and mark the rnode with RDIRTY. This write
4968 			 * will be restarted at some later time.
4969 			 */
4970 			plist = pp;
4971 			while (plist != NULL) {
4972 				pp = plist;
4973 				page_sub(&plist, pp);
4974 				hat_setmod(pp);
4975 				page_io_unlock(pp);
4976 				page_unlock(pp);
4977 			}
4978 			rp->r_flags |= RDIRTY;
4979 			mutex_exit(&rp->r_statelock);
4980 			if (offp)
4981 				*offp = io_off;
4982 			if (lenp)
4983 				*lenp = io_len;
4984 			return (0);
4985 		}
4986 		mutex_exit(&rp->r_statelock);
4987 	}
4988 
4989 	if (flags & B_ASYNC) {
4990 		error = nfs_async_putapage(vp, pp, io_off, io_len, flags, cr,
4991 		    nfs3_sync_putapage);
4992 	} else
4993 		error = nfs3_sync_putapage(vp, pp, io_off, io_len, flags, cr);
4994 
4995 	if (offp)
4996 		*offp = io_off;
4997 	if (lenp)
4998 		*lenp = io_len;
4999 	return (error);
5000 }
5001 
5002 static int
5003 nfs3_sync_putapage(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5004 	int flags, cred_t *cr)
5005 {
5006 	int error;
5007 	rnode_t *rp;
5008 
5009 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5010 
5011 	flags |= B_WRITE;
5012 
5013 	error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5014 
5015 	rp = VTOR(vp);
5016 
5017 	if ((error == ENOSPC || error == EDQUOT || error == EFBIG ||
5018 	    error == EACCES) &&
5019 	    (flags & (B_INVAL|B_FORCE)) != (B_INVAL|B_FORCE)) {
5020 		if (!(rp->r_flags & ROUTOFSPACE)) {
5021 			mutex_enter(&rp->r_statelock);
5022 			rp->r_flags |= ROUTOFSPACE;
5023 			mutex_exit(&rp->r_statelock);
5024 		}
5025 		flags |= B_ERROR;
5026 		pvn_write_done(pp, flags);
5027 		/*
5028 		 * If this was not an async thread, then try again to
5029 		 * write out the pages, but this time, also destroy
5030 		 * them whether or not the write is successful.  This
5031 		 * will prevent memory from filling up with these
5032 		 * pages and destroying them is the only alternative
5033 		 * if they can't be written out.
5034 		 *
5035 		 * Don't do this if this is an async thread because
5036 		 * when the pages are unlocked in pvn_write_done,
5037 		 * some other thread could have come along, locked
5038 		 * them, and queued for an async thread.  It would be
5039 		 * possible for all of the async threads to be tied
5040 		 * up waiting to lock the pages again and they would
5041 		 * all already be locked and waiting for an async
5042 		 * thread to handle them.  Deadlock.
5043 		 */
5044 		if (!(flags & B_ASYNC)) {
5045 			error = nfs3_putpage(vp, io_off, io_len,
5046 			    B_INVAL | B_FORCE, cr);
5047 		}
5048 	} else {
5049 		if (error)
5050 			flags |= B_ERROR;
5051 		else if (rp->r_flags & ROUTOFSPACE) {
5052 			mutex_enter(&rp->r_statelock);
5053 			rp->r_flags &= ~ROUTOFSPACE;
5054 			mutex_exit(&rp->r_statelock);
5055 		}
5056 		pvn_write_done(pp, flags);
5057 		if (freemem < desfree)
5058 			(void) nfs3_commit_vp(vp, (u_offset_t)0, 0, cr);
5059 	}
5060 
5061 	return (error);
5062 }
5063 
5064 static int
5065 nfs3_map(vnode_t *vp, offset_t off, struct as *as, caddr_t *addrp,
5066 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
5067 {
5068 	struct segvn_crargs vn_a;
5069 	int error;
5070 	rnode_t *rp;
5071 	struct vattr va;
5072 
5073 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5074 		return (EIO);
5075 
5076 	if (vp->v_flag & VNOMAP)
5077 		return (ENOSYS);
5078 
5079 	if (off < 0 || off + len < 0)
5080 		return (ENXIO);
5081 
5082 	if (vp->v_type != VREG)
5083 		return (ENODEV);
5084 
5085 	/*
5086 	 * If there is cached data and if close-to-open consistency
5087 	 * checking is not turned off and if the file system is not
5088 	 * mounted readonly, then force an over the wire getattr.
5089 	 * Otherwise, just invoke nfs3getattr to get a copy of the
5090 	 * attributes.  The attribute cache will be used unless it
5091 	 * is timed out and if it is, then an over the wire getattr
5092 	 * will be issued.
5093 	 */
5094 	va.va_mask = AT_ALL;
5095 	if (vn_has_cached_data(vp) &&
5096 	    !(VTOMI(vp)->mi_flags & MI_NOCTO) && !vn_is_readonly(vp))
5097 		error = nfs3_getattr_otw(vp, &va, cr);
5098 	else
5099 		error = nfs3getattr(vp, &va, cr);
5100 	if (error)
5101 		return (error);
5102 
5103 	/*
5104 	 * Check to see if the vnode is currently marked as not cachable.
5105 	 * This means portions of the file are locked (through VOP_FRLOCK).
5106 	 * In this case the map request must be refused.  We use
5107 	 * rp->r_lkserlock to avoid a race with concurrent lock requests.
5108 	 */
5109 	rp = VTOR(vp);
5110 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
5111 		return (EINTR);
5112 
5113 	if (vp->v_flag & VNOCACHE) {
5114 		error = EAGAIN;
5115 		goto done;
5116 	}
5117 
5118 	/*
5119 	 * Don't allow concurrent locks and mapping if mandatory locking is
5120 	 * enabled.
5121 	 */
5122 	if ((flk_has_remote_locks(vp) || lm_has_sleep(vp)) &&
5123 	    MANDLOCK(vp, va.va_mode)) {
5124 		error = EAGAIN;
5125 		goto done;
5126 	}
5127 
5128 	as_rangelock(as);
5129 	if (!(flags & MAP_FIXED)) {
5130 		map_addr(addrp, len, off, 1, flags);
5131 		if (*addrp == NULL) {
5132 			as_rangeunlock(as);
5133 			error = ENOMEM;
5134 			goto done;
5135 		}
5136 	} else {
5137 		/*
5138 		 * User specified address - blow away any previous mappings
5139 		 */
5140 		(void) as_unmap(as, *addrp, len);
5141 	}
5142 
5143 	vn_a.vp = vp;
5144 	vn_a.offset = off;
5145 	vn_a.type = (flags & MAP_TYPE);
5146 	vn_a.prot = (uchar_t)prot;
5147 	vn_a.maxprot = (uchar_t)maxprot;
5148 	vn_a.flags = (flags & ~MAP_TYPE);
5149 	vn_a.cred = cr;
5150 	vn_a.amp = NULL;
5151 	vn_a.szc = 0;
5152 	vn_a.lgrp_mem_policy_flags = 0;
5153 
5154 	error = as_map(as, *addrp, len, segvn_create, &vn_a);
5155 	as_rangeunlock(as);
5156 
5157 done:
5158 	nfs_rw_exit(&rp->r_lkserlock);
5159 	return (error);
5160 }
5161 
5162 /* ARGSUSED */
5163 static int
5164 nfs3_addmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5165 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
5166 {
5167 	rnode_t *rp;
5168 
5169 	if (vp->v_flag & VNOMAP)
5170 		return (ENOSYS);
5171 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5172 		return (EIO);
5173 
5174 	/*
5175 	 * Need to hold rwlock while incrementing the mapcnt so that
5176 	 * mmap'ing can be serialized with writes so that the caching
5177 	 * can be handled correctly.
5178 	 */
5179 	rp = VTOR(vp);
5180 	if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, INTR(vp)))
5181 		return (EINTR);
5182 	atomic_add_long((ulong_t *)&rp->r_mapcnt, btopr(len));
5183 	nfs_rw_exit(&rp->r_rwlock);
5184 
5185 	return (0);
5186 }
5187 
5188 static int
5189 nfs3_frlock(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5190 	offset_t offset, struct flk_callback *flk_cbp, cred_t *cr)
5191 {
5192 	netobj lm_fh3;
5193 	int rc;
5194 	u_offset_t start, end;
5195 	rnode_t *rp;
5196 	int error = 0, intr = INTR(vp);
5197 
5198 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5199 		return (EIO);
5200 	/* check for valid cmd parameter */
5201 	if (cmd != F_GETLK && cmd != F_SETLK && cmd != F_SETLKW)
5202 		return (EINVAL);
5203 
5204 	/* Verify l_type. */
5205 	switch (bfp->l_type) {
5206 	case F_RDLCK:
5207 		if (cmd != F_GETLK && !(flag & FREAD))
5208 			return (EBADF);
5209 		break;
5210 	case F_WRLCK:
5211 		if (cmd != F_GETLK && !(flag & FWRITE))
5212 			return (EBADF);
5213 		break;
5214 	case F_UNLCK:
5215 		intr = 0;
5216 		break;
5217 
5218 	default:
5219 		return (EINVAL);
5220 	}
5221 
5222 	/* check the validity of the lock range */
5223 	if (rc = flk_convert_lock_data(vp, bfp, &start, &end, offset))
5224 		return (rc);
5225 	if (rc = flk_check_lock_data(start, end, MAXEND))
5226 		return (rc);
5227 
5228 	/*
5229 	 * If the filesystem is mounted using local locking, pass the
5230 	 * request off to the local locking code.
5231 	 */
5232 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
5233 		if (cmd == F_SETLK || cmd == F_SETLKW) {
5234 			/*
5235 			 * For complete safety, we should be holding
5236 			 * r_lkserlock.  However, we can't call
5237 			 * lm_safelock and then fs_frlock while
5238 			 * holding r_lkserlock, so just invoke
5239 			 * lm_safelock and expect that this will
5240 			 * catch enough of the cases.
5241 			 */
5242 			if (!lm_safelock(vp, bfp, cr))
5243 				return (EAGAIN);
5244 		}
5245 		return (fs_frlock(vp, cmd, bfp, flag, offset, flk_cbp, cr));
5246 	}
5247 
5248 	rp = VTOR(vp);
5249 
5250 	/*
5251 	 * Check whether the given lock request can proceed, given the
5252 	 * current file mappings.
5253 	 */
5254 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_WRITER, intr))
5255 		return (EINTR);
5256 	if (cmd == F_SETLK || cmd == F_SETLKW) {
5257 		if (!lm_safelock(vp, bfp, cr)) {
5258 			rc = EAGAIN;
5259 			goto done;
5260 		}
5261 	}
5262 
5263 	/*
5264 	 * Flush the cache after waiting for async I/O to finish.  For new
5265 	 * locks, this is so that the process gets the latest bits from the
5266 	 * server.  For unlocks, this is so that other clients see the
5267 	 * latest bits once the file has been unlocked.  If currently dirty
5268 	 * pages can't be flushed, then don't allow a lock to be set.  But
5269 	 * allow unlocks to succeed, to avoid having orphan locks on the
5270 	 * server.
5271 	 */
5272 	if (cmd != F_GETLK) {
5273 		mutex_enter(&rp->r_statelock);
5274 		while (rp->r_count > 0) {
5275 		    if (intr) {
5276 			klwp_t *lwp = ttolwp(curthread);
5277 
5278 			if (lwp != NULL)
5279 				lwp->lwp_nostop++;
5280 			if (cv_wait_sig(&rp->r_cv, &rp->r_statelock) == 0) {
5281 				if (lwp != NULL)
5282 					lwp->lwp_nostop--;
5283 				rc = EINTR;
5284 				break;
5285 			}
5286 			if (lwp != NULL)
5287 				lwp->lwp_nostop--;
5288 		    } else
5289 			cv_wait(&rp->r_cv, &rp->r_statelock);
5290 		}
5291 		mutex_exit(&rp->r_statelock);
5292 		if (rc != 0)
5293 			goto done;
5294 		error = nfs3_putpage(vp, (offset_t)0, 0, B_INVAL, cr);
5295 		if (error) {
5296 			if (error == ENOSPC || error == EDQUOT) {
5297 				mutex_enter(&rp->r_statelock);
5298 				if (!rp->r_error)
5299 					rp->r_error = error;
5300 				mutex_exit(&rp->r_statelock);
5301 			}
5302 			if (bfp->l_type != F_UNLCK) {
5303 				rc = ENOLCK;
5304 				goto done;
5305 			}
5306 		}
5307 	}
5308 
5309 	lm_fh3.n_len = VTOFH3(vp)->fh3_length;
5310 	lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
5311 
5312 	/*
5313 	 * Call the lock manager to do the real work of contacting
5314 	 * the server and obtaining the lock.
5315 	 */
5316 	rc = lm4_frlock(vp, cmd, bfp, flag, offset, cr, &lm_fh3, flk_cbp);
5317 
5318 	if (rc == 0)
5319 		nfs_lockcompletion(vp, cmd);
5320 
5321 done:
5322 	nfs_rw_exit(&rp->r_lkserlock);
5323 	return (rc);
5324 }
5325 
5326 /*
5327  * Free storage space associated with the specified vnode.  The portion
5328  * to be freed is specified by bfp->l_start and bfp->l_len (already
5329  * normalized to a "whence" of 0).
5330  *
5331  * This is an experimental facility whose continued existence is not
5332  * guaranteed.  Currently, we only support the special case
5333  * of l_len == 0, meaning free to end of file.
5334  */
5335 /* ARGSUSED */
5336 static int
5337 nfs3_space(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5338 	offset_t offset, cred_t *cr, caller_context_t *ct)
5339 {
5340 	int error;
5341 
5342 	ASSERT(vp->v_type == VREG);
5343 	if (cmd != F_FREESP)
5344 		return (EINVAL);
5345 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5346 		return (EIO);
5347 
5348 	error = convoff(vp, bfp, 0, offset);
5349 	if (!error) {
5350 		ASSERT(bfp->l_start >= 0);
5351 		if (bfp->l_len == 0) {
5352 			struct vattr va;
5353 
5354 			/*
5355 			 * ftruncate should not change the ctime and
5356 			 * mtime if we truncate the file to its
5357 			 * previous size.
5358 			 */
5359 			va.va_mask = AT_SIZE;
5360 			error = nfs3getattr(vp, &va, cr);
5361 			if (error || va.va_size == bfp->l_start)
5362 				return (error);
5363 			va.va_mask = AT_SIZE;
5364 			va.va_size = bfp->l_start;
5365 			error = nfs3setattr(vp, &va, 0, cr);
5366 		} else
5367 			error = EINVAL;
5368 	}
5369 
5370 	return (error);
5371 }
5372 
5373 /* ARGSUSED */
5374 static int
5375 nfs3_realvp(vnode_t *vp, vnode_t **vpp)
5376 {
5377 
5378 	return (EINVAL);
5379 }
5380 
5381 /*
5382  * Setup and add an address space callback to do the work of the delmap call.
5383  * The callback will (and must be) deleted in the actual callback function.
5384  *
5385  * This is done in order to take care of the problem that we have with holding
5386  * the address space's a_lock for a long period of time (e.g. if the NFS server
5387  * is down).  Callbacks will be executed in the address space code while the
5388  * a_lock is not held.	Holding the address space's a_lock causes things such
5389  * as ps and fork to hang because they are trying to acquire this lock as well.
5390  */
5391 /* ARGSUSED */
5392 static int
5393 nfs3_delmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5394 	size_t len, uint_t prot, uint_t maxprot, uint_t flags, cred_t *cr)
5395 {
5396 	int			caller_found;
5397 	int			error;
5398 	rnode_t			*rp;
5399 	nfs_delmap_args_t	*dmapp;
5400 	nfs_delmapcall_t	*delmap_call;
5401 
5402 	if (vp->v_flag & VNOMAP)
5403 		return (ENOSYS);
5404 	/*
5405 	 * A process may not change zones if it has NFS pages mmap'ed
5406 	 * in, so we can't legitimately get here from the wrong zone.
5407 	 */
5408 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5409 
5410 	rp = VTOR(vp);
5411 
5412 	/*
5413 	 * The way that the address space of this process deletes its mapping
5414 	 * of this file is via the following call chains:
5415 	 * - as_free()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5416 	 * - as_unmap()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5417 	 *
5418 	 * With the use of address space callbacks we are allowed to drop the
5419 	 * address space lock, a_lock, while executing the NFS operations that
5420 	 * need to go over the wire.  Returning EAGAIN to the caller of this
5421 	 * function is what drives the execution of the callback that we add
5422 	 * below.  The callback will be executed by the address space code
5423 	 * after dropping the a_lock.  When the callback is finished, since
5424 	 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
5425 	 * is called again on the same segment to finish the rest of the work
5426 	 * that needs to happen during unmapping.
5427 	 *
5428 	 * This action of calling back into the segment driver causes
5429 	 * nfs3_delmap() to get called again, but since the callback was
5430 	 * already executed at this point, it already did the work and there
5431 	 * is nothing left for us to do.
5432 	 *
5433 	 * To Summarize:
5434 	 * - The first time nfs3_delmap is called by the current thread is when
5435 	 * we add the caller associated with this delmap to the delmap caller
5436 	 * list, add the callback, and return EAGAIN.
5437 	 * - The second time in this call chain when nfs3_delmap is called we
5438 	 * will find this caller in the delmap caller list and realize there
5439 	 * is no more work to do thus removing this caller from the list and
5440 	 * returning the error that was set in the callback execution.
5441 	 */
5442 	caller_found = nfs_find_and_delete_delmapcall(rp, &error);
5443 	if (caller_found) {
5444 		/*
5445 		 * 'error' is from the actual delmap operations.  To avoid
5446 		 * hangs, we need to handle the return of EAGAIN differently
5447 		 * since this is what drives the callback execution.
5448 		 * In this case, we don't want to return EAGAIN and do the
5449 		 * callback execution because there are none to execute.
5450 		 */
5451 		if (error == EAGAIN)
5452 			return (0);
5453 		else
5454 			return (error);
5455 	}
5456 
5457 	/* current caller was not in the list */
5458 	delmap_call = nfs_init_delmapcall();
5459 
5460 	mutex_enter(&rp->r_statelock);
5461 	list_insert_tail(&rp->r_indelmap, delmap_call);
5462 	mutex_exit(&rp->r_statelock);
5463 
5464 	dmapp = kmem_alloc(sizeof (nfs_delmap_args_t), KM_SLEEP);
5465 
5466 	dmapp->vp = vp;
5467 	dmapp->off = off;
5468 	dmapp->addr = addr;
5469 	dmapp->len = len;
5470 	dmapp->prot = prot;
5471 	dmapp->maxprot = maxprot;
5472 	dmapp->flags = flags;
5473 	dmapp->cr = cr;
5474 	dmapp->caller = delmap_call;
5475 
5476 	error = as_add_callback(as, nfs3_delmap_callback, dmapp,
5477 	AS_UNMAP_EVENT, addr, len, KM_SLEEP);
5478 
5479 	return (error ? error : EAGAIN);
5480 }
5481 
5482 /*
5483  * Remove some pages from an mmap'd vnode.  Just update the
5484  * count of pages.  If doing close-to-open, then flush and
5485  * commit all of the pages associated with this file.
5486  * Otherwise, start an asynchronous page flush to write out
5487  * any dirty pages.  This will also associate a credential
5488  * with the rnode which can be used to write the pages.
5489  */
5490 /* ARGSUSED */
5491 static void
5492 nfs3_delmap_callback(struct as *as, void *arg, uint_t event)
5493 {
5494 	int			error;
5495 	rnode_t			*rp;
5496 	mntinfo_t		*mi;
5497 	nfs_delmap_args_t	*dmapp = (nfs_delmap_args_t *)arg;
5498 
5499 	rp = VTOR(dmapp->vp);
5500 	mi = VTOMI(dmapp->vp);
5501 
5502 	atomic_add_long((ulong_t *)&rp->r_mapcnt, -btopr(dmapp->len));
5503 	ASSERT(rp->r_mapcnt >= 0);
5504 
5505 	/*
5506 	 * Initiate a page flush and potential commit if there are
5507 	 * pages, the file system was not mounted readonly, the segment
5508 	 * was mapped shared, and the pages themselves were writeable.
5509 	 */
5510 	if (vn_has_cached_data(dmapp->vp) && !vn_is_readonly(dmapp->vp) &&
5511 	    dmapp->flags == MAP_SHARED && (dmapp->maxprot & PROT_WRITE)) {
5512 		mutex_enter(&rp->r_statelock);
5513 		rp->r_flags |= RDIRTY;
5514 		mutex_exit(&rp->r_statelock);
5515 		/*
5516 		 * If this is a cross-zone access a sync putpage won't work, so
5517 		 * the best we can do is try an async putpage.  That seems
5518 		 * better than something more draconian such as discarding the
5519 		 * dirty pages.
5520 		 */
5521 		if ((mi->mi_flags & MI_NOCTO) ||
5522 		    nfs_zone() != mi->mi_zone)
5523 			error = nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5524 			    B_ASYNC, dmapp->cr);
5525 		else
5526 			error = nfs3_putpage_commit(dmapp->vp, dmapp->off,
5527 			    dmapp->len, dmapp->cr);
5528 		if (!error) {
5529 			mutex_enter(&rp->r_statelock);
5530 			error = rp->r_error;
5531 			rp->r_error = 0;
5532 			mutex_exit(&rp->r_statelock);
5533 		}
5534 	} else
5535 		error = 0;
5536 
5537 	if ((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO))
5538 		(void) nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5539 		    B_INVAL, dmapp->cr);
5540 
5541 	dmapp->caller->error = error;
5542 	(void) as_delete_callback(as, arg);
5543 	kmem_free(dmapp, sizeof (nfs_delmap_args_t));
5544 }
5545 
5546 static int nfs3_pathconf_disable_cache = 0;
5547 
5548 #ifdef DEBUG
5549 static int nfs3_pathconf_cache_hits = 0;
5550 static int nfs3_pathconf_cache_misses = 0;
5551 #endif
5552 
5553 static int
5554 nfs3_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr)
5555 {
5556 	int error;
5557 	PATHCONF3args args;
5558 	PATHCONF3res res;
5559 	int douprintf;
5560 	failinfo_t fi;
5561 	rnode_t *rp;
5562 	hrtime_t t;
5563 
5564 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5565 		return (EIO);
5566 	/*
5567 	 * Large file spec - need to base answer on info stored
5568 	 * on original FSINFO response.
5569 	 */
5570 	if (cmd == _PC_FILESIZEBITS) {
5571 		unsigned long long ll;
5572 		long l = 1;
5573 
5574 		ll = VTOMI(vp)->mi_maxfilesize;
5575 
5576 		if (ll == 0) {
5577 			*valp = 0;
5578 			return (0);
5579 		}
5580 
5581 		if (ll & 0xffffffff00000000) {
5582 			l += 32; ll >>= 32;
5583 		}
5584 		if (ll & 0xffff0000) {
5585 			l += 16; ll >>= 16;
5586 		}
5587 		if (ll & 0xff00) {
5588 			l += 8; ll >>= 8;
5589 		}
5590 		if (ll & 0xf0) {
5591 			l += 4; ll >>= 4;
5592 		}
5593 		if (ll & 0xc) {
5594 			l += 2; ll >>= 2;
5595 		}
5596 		if (ll & 0x2)
5597 			l += 2;
5598 		else if (ll & 0x1)
5599 			l += 1;
5600 		*valp = l;
5601 		return (0);
5602 	}
5603 
5604 	if (cmd == _PC_ACL_ENABLED) {
5605 		*valp = _ACL_ACLENT_ENABLED;
5606 		return (0);
5607 	}
5608 
5609 	if (cmd == _PC_XATTR_EXISTS) {
5610 		error = 0;
5611 		*valp = 0;
5612 		if (vp->v_vfsp->vfs_flag & VFS_XATTR) {
5613 			vnode_t *avp;
5614 			rnode_t *rp;
5615 			int error = 0;
5616 			mntinfo_t *mi = VTOMI(vp);
5617 
5618 			if (!(mi->mi_flags & MI_EXTATTR))
5619 				return (0);
5620 
5621 			rp = VTOR(vp);
5622 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_READER,
5623 			    INTR(vp)))
5624 				return (EINTR);
5625 
5626 			error = nfs3lookup_dnlc(vp, XATTR_DIR_NAME, &avp, cr);
5627 			if (error || avp == NULL)
5628 				error = acl_getxattrdir3(vp, &avp, 0, cr, 0);
5629 
5630 			nfs_rw_exit(&rp->r_rwlock);
5631 
5632 			if (error == 0 && avp != NULL) {
5633 				VN_RELE(avp);
5634 				*valp = 1;
5635 			} else if (error == ENOENT)
5636 				error = 0;
5637 		}
5638 		return (error);
5639 	}
5640 
5641 	rp = VTOR(vp);
5642 	if (rp->r_pathconf != NULL) {
5643 		mutex_enter(&rp->r_statelock);
5644 		if (rp->r_pathconf != NULL && nfs3_pathconf_disable_cache) {
5645 			kmem_free(rp->r_pathconf, sizeof (*rp->r_pathconf));
5646 			rp->r_pathconf = NULL;
5647 		}
5648 		if (rp->r_pathconf != NULL) {
5649 			error = 0;
5650 			switch (cmd) {
5651 			case _PC_LINK_MAX:
5652 				*valp = rp->r_pathconf->link_max;
5653 				break;
5654 			case _PC_NAME_MAX:
5655 				*valp = rp->r_pathconf->name_max;
5656 				break;
5657 			case _PC_PATH_MAX:
5658 			case _PC_SYMLINK_MAX:
5659 				*valp = MAXPATHLEN;
5660 				break;
5661 			case _PC_CHOWN_RESTRICTED:
5662 				*valp = rp->r_pathconf->chown_restricted;
5663 				break;
5664 			case _PC_NO_TRUNC:
5665 				*valp = rp->r_pathconf->no_trunc;
5666 				break;
5667 			default:
5668 				error = EINVAL;
5669 				break;
5670 			}
5671 			mutex_exit(&rp->r_statelock);
5672 #ifdef DEBUG
5673 			nfs3_pathconf_cache_hits++;
5674 #endif
5675 			return (error);
5676 		}
5677 		mutex_exit(&rp->r_statelock);
5678 	}
5679 #ifdef DEBUG
5680 	nfs3_pathconf_cache_misses++;
5681 #endif
5682 
5683 	args.object = *VTOFH3(vp);
5684 	fi.vp = vp;
5685 	fi.fhp = (caddr_t)&args.object;
5686 	fi.copyproc = nfs3copyfh;
5687 	fi.lookupproc = nfs3lookup;
5688 	fi.xattrdirproc = acl_getxattrdir3;
5689 
5690 	douprintf = 1;
5691 
5692 	t = gethrtime();
5693 
5694 	error = rfs3call(VTOMI(vp), NFSPROC3_PATHCONF,
5695 	    xdr_nfs_fh3, (caddr_t)&args,
5696 	    xdr_PATHCONF3res, (caddr_t)&res, cr,
5697 	    &douprintf, &res.status, 0, &fi);
5698 
5699 	if (error)
5700 		return (error);
5701 
5702 	error = geterrno3(res.status);
5703 
5704 	if (!error) {
5705 		nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
5706 		if (!nfs3_pathconf_disable_cache) {
5707 			mutex_enter(&rp->r_statelock);
5708 			if (rp->r_pathconf == NULL) {
5709 				rp->r_pathconf = kmem_alloc(
5710 				    sizeof (*rp->r_pathconf), KM_NOSLEEP);
5711 				if (rp->r_pathconf != NULL)
5712 					*rp->r_pathconf = res.resok.info;
5713 			}
5714 			mutex_exit(&rp->r_statelock);
5715 		}
5716 		switch (cmd) {
5717 		case _PC_LINK_MAX:
5718 			*valp = res.resok.info.link_max;
5719 			break;
5720 		case _PC_NAME_MAX:
5721 			*valp = res.resok.info.name_max;
5722 			break;
5723 		case _PC_PATH_MAX:
5724 		case _PC_SYMLINK_MAX:
5725 			*valp = MAXPATHLEN;
5726 			break;
5727 		case _PC_CHOWN_RESTRICTED:
5728 			*valp = res.resok.info.chown_restricted;
5729 			break;
5730 		case _PC_NO_TRUNC:
5731 			*valp = res.resok.info.no_trunc;
5732 			break;
5733 		default:
5734 			return (EINVAL);
5735 		}
5736 	} else {
5737 		nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
5738 		PURGE_STALE_FH(error, vp, cr);
5739 	}
5740 
5741 	return (error);
5742 }
5743 
5744 /*
5745  * Called by async thread to do synchronous pageio. Do the i/o, wait
5746  * for it to complete, and cleanup the page list when done.
5747  */
5748 static int
5749 nfs3_sync_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5750 	int flags, cred_t *cr)
5751 {
5752 	int error;
5753 
5754 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5755 	error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5756 	if (flags & B_READ)
5757 		pvn_read_done(pp, (error ? B_ERROR : 0) | flags);
5758 	else
5759 		pvn_write_done(pp, (error ? B_ERROR : 0) | flags);
5760 	return (error);
5761 }
5762 
5763 static int
5764 nfs3_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5765 	int flags, cred_t *cr)
5766 {
5767 	int error;
5768 	rnode_t *rp;
5769 
5770 	if (pp == NULL)
5771 		return (EINVAL);
5772 	if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
5773 		return (EIO);
5774 
5775 	rp = VTOR(vp);
5776 	mutex_enter(&rp->r_statelock);
5777 	rp->r_count++;
5778 	mutex_exit(&rp->r_statelock);
5779 
5780 	if (flags & B_ASYNC) {
5781 		error = nfs_async_pageio(vp, pp, io_off, io_len, flags, cr,
5782 		    nfs3_sync_pageio);
5783 	} else
5784 		error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5785 	mutex_enter(&rp->r_statelock);
5786 	rp->r_count--;
5787 	cv_broadcast(&rp->r_cv);
5788 	mutex_exit(&rp->r_statelock);
5789 	return (error);
5790 }
5791 
5792 static void
5793 nfs3_dispose(vnode_t *vp, page_t *pp, int fl, int dn, cred_t *cr)
5794 {
5795 	int error;
5796 	rnode_t *rp;
5797 	page_t *plist;
5798 	page_t *pptr;
5799 	offset3 offset;
5800 	count3 len;
5801 	k_sigset_t smask;
5802 
5803 	/*
5804 	 * We should get called with fl equal to either B_FREE or
5805 	 * B_INVAL.  Any other value is illegal.
5806 	 *
5807 	 * The page that we are either supposed to free or destroy
5808 	 * should be exclusive locked and its io lock should not
5809 	 * be held.
5810 	 */
5811 	ASSERT(fl == B_FREE || fl == B_INVAL);
5812 	ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr);
5813 	rp = VTOR(vp);
5814 
5815 	/*
5816 	 * If the page doesn't need to be committed or we shouldn't
5817 	 * even bother attempting to commit it, then just make sure
5818 	 * that the p_fsdata byte is clear and then either free or
5819 	 * destroy the page as appropriate.
5820 	 */
5821 	if (pp->p_fsdata == C_NOCOMMIT || (rp->r_flags & RSTALE)) {
5822 		pp->p_fsdata = C_NOCOMMIT;
5823 		if (fl == B_FREE)
5824 			page_free(pp, dn);
5825 		else
5826 			page_destroy(pp, dn);
5827 		return;
5828 	}
5829 
5830 	/*
5831 	 * If there is a page invalidation operation going on, then
5832 	 * if this is one of the pages being destroyed, then just
5833 	 * clear the p_fsdata byte and then either free or destroy
5834 	 * the page as appropriate.
5835 	 */
5836 	mutex_enter(&rp->r_statelock);
5837 	if ((rp->r_flags & RTRUNCATE) && pp->p_offset >= rp->r_truncaddr) {
5838 		mutex_exit(&rp->r_statelock);
5839 		pp->p_fsdata = C_NOCOMMIT;
5840 		if (fl == B_FREE)
5841 			page_free(pp, dn);
5842 		else
5843 			page_destroy(pp, dn);
5844 		return;
5845 	}
5846 
5847 	/*
5848 	 * If we are freeing this page and someone else is already
5849 	 * waiting to do a commit, then just unlock the page and
5850 	 * return.  That other thread will take care of commiting
5851 	 * this page.  The page can be freed sometime after the
5852 	 * commit has finished.  Otherwise, if the page is marked
5853 	 * as delay commit, then we may be getting called from
5854 	 * pvn_write_done, one page at a time.   This could result
5855 	 * in one commit per page, so we end up doing lots of small
5856 	 * commits instead of fewer larger commits.  This is bad,
5857 	 * we want do as few commits as possible.
5858 	 */
5859 	if (fl == B_FREE) {
5860 		if (rp->r_flags & RCOMMITWAIT) {
5861 			page_unlock(pp);
5862 			mutex_exit(&rp->r_statelock);
5863 			return;
5864 		}
5865 		if (pp->p_fsdata == C_DELAYCOMMIT) {
5866 			pp->p_fsdata = C_COMMIT;
5867 			page_unlock(pp);
5868 			mutex_exit(&rp->r_statelock);
5869 			return;
5870 		}
5871 	}
5872 
5873 	/*
5874 	 * Check to see if there is a signal which would prevent an
5875 	 * attempt to commit the pages from being successful.  If so,
5876 	 * then don't bother with all of the work to gather pages and
5877 	 * generate the unsuccessful RPC.  Just return from here and
5878 	 * let the page be committed at some later time.
5879 	 */
5880 	sigintr(&smask, VTOMI(vp)->mi_flags & MI_INT);
5881 	if (ttolwp(curthread) != NULL && ISSIG(curthread, JUSTLOOKING)) {
5882 		sigunintr(&smask);
5883 		page_unlock(pp);
5884 		mutex_exit(&rp->r_statelock);
5885 		return;
5886 	}
5887 	sigunintr(&smask);
5888 
5889 	/*
5890 	 * We are starting to need to commit pages, so let's try
5891 	 * to commit as many as possible at once to reduce the
5892 	 * overhead.
5893 	 *
5894 	 * Set the `commit inprogress' state bit.  We must
5895 	 * first wait until any current one finishes.  Then
5896 	 * we initialize the c_pages list with this page.
5897 	 */
5898 	while (rp->r_flags & RCOMMIT) {
5899 		rp->r_flags |= RCOMMITWAIT;
5900 		cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
5901 		rp->r_flags &= ~RCOMMITWAIT;
5902 	}
5903 	rp->r_flags |= RCOMMIT;
5904 	mutex_exit(&rp->r_statelock);
5905 	ASSERT(rp->r_commit.c_pages == NULL);
5906 	rp->r_commit.c_pages = pp;
5907 	rp->r_commit.c_commbase = (offset3)pp->p_offset;
5908 	rp->r_commit.c_commlen = PAGESIZE;
5909 
5910 	/*
5911 	 * Gather together all other pages which can be committed.
5912 	 * They will all be chained off r_commit.c_pages.
5913 	 */
5914 	nfs3_get_commit(vp);
5915 
5916 	/*
5917 	 * Clear the `commit inprogress' status and disconnect
5918 	 * the list of pages to be committed from the rnode.
5919 	 * At this same time, we also save the starting offset
5920 	 * and length of data to be committed on the server.
5921 	 */
5922 	plist = rp->r_commit.c_pages;
5923 	rp->r_commit.c_pages = NULL;
5924 	offset = rp->r_commit.c_commbase;
5925 	len = rp->r_commit.c_commlen;
5926 	mutex_enter(&rp->r_statelock);
5927 	rp->r_flags &= ~RCOMMIT;
5928 	cv_broadcast(&rp->r_commit.c_cv);
5929 	mutex_exit(&rp->r_statelock);
5930 
5931 	if (curproc == proc_pageout || curproc == proc_fsflush ||
5932 	    nfs_zone() != VTOMI(vp)->mi_zone) {
5933 		nfs_async_commit(vp, plist, offset, len, cr, nfs3_async_commit);
5934 		return;
5935 	}
5936 
5937 	/*
5938 	 * Actually generate the COMMIT3 over the wire operation.
5939 	 */
5940 	error = nfs3_commit(vp, offset, len, cr);
5941 
5942 	/*
5943 	 * If we got an error during the commit, just unlock all
5944 	 * of the pages.  The pages will get retransmitted to the
5945 	 * server during a putpage operation.
5946 	 */
5947 	if (error) {
5948 		while (plist != NULL) {
5949 			pptr = plist;
5950 			page_sub(&plist, pptr);
5951 			page_unlock(pptr);
5952 		}
5953 		return;
5954 	}
5955 
5956 	/*
5957 	 * We've tried as hard as we can to commit the data to stable
5958 	 * storage on the server.  We release the rest of the pages
5959 	 * and clear the commit required state.  They will be put
5960 	 * onto the tail of the cachelist if they are nolonger
5961 	 * mapped.
5962 	 */
5963 	while (plist != pp) {
5964 		pptr = plist;
5965 		page_sub(&plist, pptr);
5966 		pptr->p_fsdata = C_NOCOMMIT;
5967 		(void) page_release(pptr, 1);
5968 	}
5969 
5970 	/*
5971 	 * It is possible that nfs3_commit didn't return error but
5972 	 * some other thread has modified the page we are going
5973 	 * to free/destroy.
5974 	 *    In this case we need to rewrite the page. Do an explicit check
5975 	 * before attempting to free/destroy the page. If modified, needs to
5976 	 * be rewritten so unlock the page and return.
5977 	 */
5978 	if (hat_ismod(pp)) {
5979 		pp->p_fsdata = C_NOCOMMIT;
5980 		page_unlock(pp);
5981 		return;
5982 	}
5983 
5984 	/*
5985 	 * Now, as appropriate, either free or destroy the page
5986 	 * that we were called with.
5987 	 */
5988 	pp->p_fsdata = C_NOCOMMIT;
5989 	if (fl == B_FREE)
5990 		page_free(pp, dn);
5991 	else
5992 		page_destroy(pp, dn);
5993 }
5994 
5995 static int
5996 nfs3_commit(vnode_t *vp, offset3 offset, count3 count, cred_t *cr)
5997 {
5998 	int error;
5999 	rnode_t *rp;
6000 	COMMIT3args args;
6001 	COMMIT3res res;
6002 	int douprintf;
6003 	cred_t *cred;
6004 
6005 	rp = VTOR(vp);
6006 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6007 
6008 	mutex_enter(&rp->r_statelock);
6009 	if (rp->r_cred != NULL) {
6010 		cred = rp->r_cred;
6011 		crhold(cred);
6012 	} else {
6013 		rp->r_cred = cr;
6014 		crhold(cr);
6015 		cred = cr;
6016 		crhold(cred);
6017 	}
6018 	mutex_exit(&rp->r_statelock);
6019 
6020 	args.file = *VTOFH3(vp);
6021 	args.offset = offset;
6022 	args.count = count;
6023 
6024 doitagain:
6025 	douprintf = 1;
6026 	error = rfs3call(VTOMI(vp), NFSPROC3_COMMIT,
6027 	    xdr_COMMIT3args, (caddr_t)&args,
6028 	    xdr_COMMIT3res, (caddr_t)&res, cred,
6029 	    &douprintf, &res.status, 0, NULL);
6030 
6031 	crfree(cred);
6032 
6033 	if (error)
6034 		return (error);
6035 
6036 	error = geterrno3(res.status);
6037 	if (!error) {
6038 		ASSERT(rp->r_flags & RHAVEVERF);
6039 		mutex_enter(&rp->r_statelock);
6040 		if (rp->r_verf == res.resok.verf) {
6041 			mutex_exit(&rp->r_statelock);
6042 			return (0);
6043 		}
6044 		nfs3_set_mod(vp);
6045 		rp->r_verf = res.resok.verf;
6046 		mutex_exit(&rp->r_statelock);
6047 		error = NFS_VERF_MISMATCH;
6048 	} else {
6049 		if (error == EACCES) {
6050 			mutex_enter(&rp->r_statelock);
6051 			if (cred != cr) {
6052 				if (rp->r_cred != NULL)
6053 					crfree(rp->r_cred);
6054 				rp->r_cred = cr;
6055 				crhold(cr);
6056 				cred = cr;
6057 				crhold(cred);
6058 				mutex_exit(&rp->r_statelock);
6059 				goto doitagain;
6060 			}
6061 			mutex_exit(&rp->r_statelock);
6062 		}
6063 		/*
6064 		 * Can't do a PURGE_STALE_FH here because this
6065 		 * can cause a deadlock.  nfs3_commit can
6066 		 * be called from nfs3_dispose which can be called
6067 		 * indirectly via pvn_vplist_dirty.  PURGE_STALE_FH
6068 		 * can call back to pvn_vplist_dirty.
6069 		 */
6070 		if (error == ESTALE) {
6071 			mutex_enter(&rp->r_statelock);
6072 			rp->r_flags |= RSTALE;
6073 			if (!rp->r_error)
6074 				rp->r_error = error;
6075 			mutex_exit(&rp->r_statelock);
6076 			PURGE_ATTRCACHE(vp);
6077 		} else {
6078 			mutex_enter(&rp->r_statelock);
6079 			if (!rp->r_error)
6080 				rp->r_error = error;
6081 			mutex_exit(&rp->r_statelock);
6082 		}
6083 	}
6084 
6085 	return (error);
6086 }
6087 
6088 static void
6089 nfs3_set_mod(vnode_t *vp)
6090 {
6091 	page_t *pp;
6092 	kmutex_t *vphm;
6093 
6094 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6095 	vphm = page_vnode_mutex(vp);
6096 	mutex_enter(vphm);
6097 	if ((pp = vp->v_pages) != NULL) {
6098 		do {
6099 			if (pp->p_fsdata != C_NOCOMMIT) {
6100 				hat_setmod(pp);
6101 				pp->p_fsdata = C_NOCOMMIT;
6102 			}
6103 		} while ((pp = pp->p_vpnext) != vp->v_pages);
6104 	}
6105 	mutex_exit(vphm);
6106 }
6107 
6108 
6109 /*
6110  * This routine is used to gather together a page list of the pages
6111  * which are to be committed on the server.  This routine must not
6112  * be called if the calling thread holds any locked pages.
6113  *
6114  * The calling thread must have set RCOMMIT.  This bit is used to
6115  * serialize access to the commit structure in the rnode.  As long
6116  * as the thread has set RCOMMIT, then it can manipulate the commit
6117  * structure without requiring any other locks.
6118  */
6119 static void
6120 nfs3_get_commit(vnode_t *vp)
6121 {
6122 	rnode_t *rp;
6123 	page_t *pp;
6124 	kmutex_t *vphm;
6125 
6126 	rp = VTOR(vp);
6127 
6128 	ASSERT(rp->r_flags & RCOMMIT);
6129 
6130 	vphm = page_vnode_mutex(vp);
6131 	mutex_enter(vphm);
6132 
6133 	/*
6134 	 * If there are no pages associated with this vnode, then
6135 	 * just return.
6136 	 */
6137 	if ((pp = vp->v_pages) == NULL) {
6138 		mutex_exit(vphm);
6139 		return;
6140 	}
6141 
6142 	/*
6143 	 * Step through all of the pages associated with this vnode
6144 	 * looking for pages which need to be committed.
6145 	 */
6146 	do {
6147 		/*
6148 		 * If this page does not need to be committed or is
6149 		 * modified, then just skip it.
6150 		 */
6151 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp))
6152 			continue;
6153 
6154 		/*
6155 		 * Attempt to lock the page.  If we can't, then
6156 		 * someone else is messing with it and we will
6157 		 * just skip it.
6158 		 */
6159 		if (!page_trylock(pp, SE_EXCL))
6160 			continue;
6161 
6162 		/*
6163 		 * If this page does not need to be committed or is
6164 		 * modified, then just skip it.  Recheck now that
6165 		 * the page is locked.
6166 		 */
6167 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6168 			page_unlock(pp);
6169 			continue;
6170 		}
6171 
6172 		if (PP_ISFREE(pp)) {
6173 			cmn_err(CE_PANIC, "nfs3_get_commit: %p is free",
6174 			    (void *)pp);
6175 		}
6176 
6177 		/*
6178 		 * The page needs to be committed and we locked it.
6179 		 * Update the base and length parameters and add it
6180 		 * to r_pages.
6181 		 */
6182 		if (rp->r_commit.c_pages == NULL) {
6183 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6184 			rp->r_commit.c_commlen = PAGESIZE;
6185 		} else if (pp->p_offset < rp->r_commit.c_commbase) {
6186 			rp->r_commit.c_commlen = rp->r_commit.c_commbase -
6187 			    (offset3)pp->p_offset + rp->r_commit.c_commlen;
6188 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6189 		} else if ((rp->r_commit.c_commbase + rp->r_commit.c_commlen)
6190 			    <= pp->p_offset) {
6191 			rp->r_commit.c_commlen = (offset3)pp->p_offset -
6192 			    rp->r_commit.c_commbase + PAGESIZE;
6193 		}
6194 		page_add(&rp->r_commit.c_pages, pp);
6195 	} while ((pp = pp->p_vpnext) != vp->v_pages);
6196 
6197 	mutex_exit(vphm);
6198 }
6199 
6200 /*
6201  * This routine is used to gather together a page list of the pages
6202  * which are to be committed on the server.  This routine must not
6203  * be called if the calling thread holds any locked pages.
6204  *
6205  * The calling thread must have set RCOMMIT.  This bit is used to
6206  * serialize access to the commit structure in the rnode.  As long
6207  * as the thread has set RCOMMIT, then it can manipulate the commit
6208  * structure without requiring any other locks.
6209  */
6210 static void
6211 nfs3_get_commit_range(vnode_t *vp, u_offset_t soff, size_t len)
6212 {
6213 
6214 	rnode_t *rp;
6215 	page_t *pp;
6216 	u_offset_t end;
6217 	u_offset_t off;
6218 
6219 	ASSERT(len != 0);
6220 
6221 	rp = VTOR(vp);
6222 
6223 	ASSERT(rp->r_flags & RCOMMIT);
6224 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6225 
6226 	/*
6227 	 * If there are no pages associated with this vnode, then
6228 	 * just return.
6229 	 */
6230 	if ((pp = vp->v_pages) == NULL)
6231 		return;
6232 
6233 	/*
6234 	 * Calculate the ending offset.
6235 	 */
6236 	end = soff + len;
6237 
6238 	for (off = soff; off < end; off += PAGESIZE) {
6239 		/*
6240 		 * Lookup each page by vp, offset.
6241 		 */
6242 		if ((pp = page_lookup_nowait(vp, off, SE_EXCL)) == NULL)
6243 			continue;
6244 
6245 		/*
6246 		 * If this page does not need to be committed or is
6247 		 * modified, then just skip it.
6248 		 */
6249 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6250 			page_unlock(pp);
6251 			continue;
6252 		}
6253 
6254 		ASSERT(PP_ISFREE(pp) == 0);
6255 
6256 		/*
6257 		 * The page needs to be committed and we locked it.
6258 		 * Update the base and length parameters and add it
6259 		 * to r_pages.
6260 		 */
6261 		if (rp->r_commit.c_pages == NULL) {
6262 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6263 			rp->r_commit.c_commlen = PAGESIZE;
6264 		} else {
6265 			rp->r_commit.c_commlen = (offset3)pp->p_offset -
6266 					rp->r_commit.c_commbase + PAGESIZE;
6267 		}
6268 		page_add(&rp->r_commit.c_pages, pp);
6269 	}
6270 }
6271 
6272 #if 0	/* unused */
6273 #ifdef DEBUG
6274 static int
6275 nfs3_no_uncommitted_pages(vnode_t *vp)
6276 {
6277 	page_t *pp;
6278 	kmutex_t *vphm;
6279 
6280 	vphm = page_vnode_mutex(vp);
6281 	mutex_enter(vphm);
6282 	if ((pp = vp->v_pages) != NULL) {
6283 		do {
6284 			if (pp->p_fsdata != C_NOCOMMIT) {
6285 				mutex_exit(vphm);
6286 				return (0);
6287 			}
6288 		} while ((pp = pp->p_vpnext) != vp->v_pages);
6289 	}
6290 	mutex_exit(vphm);
6291 
6292 	return (1);
6293 }
6294 #endif
6295 #endif
6296 
6297 static int
6298 nfs3_putpage_commit(vnode_t *vp, offset_t poff, size_t plen, cred_t *cr)
6299 {
6300 	int error;
6301 	writeverf3 write_verf;
6302 	rnode_t *rp = VTOR(vp);
6303 
6304 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6305 	/*
6306 	 * Flush the data portion of the file and then commit any
6307 	 * portions which need to be committed.  This may need to
6308 	 * be done twice if the server has changed state since
6309 	 * data was last written.  The data will need to be
6310 	 * rewritten to the server and then a new commit done.
6311 	 *
6312 	 * In fact, this may need to be done several times if the
6313 	 * server is having problems and crashing while we are
6314 	 * attempting to do this.
6315 	 */
6316 
6317 top:
6318 	/*
6319 	 * Do a flush based on the poff and plen arguments.  This
6320 	 * will asynchronously write out any modified pages in the
6321 	 * range specified by (poff, plen).  This starts all of the
6322 	 * i/o operations which will be waited for in the next
6323 	 * call to nfs3_putpage
6324 	 */
6325 
6326 	mutex_enter(&rp->r_statelock);
6327 	write_verf = rp->r_verf;
6328 	mutex_exit(&rp->r_statelock);
6329 
6330 	error = nfs3_putpage(vp, poff, plen, B_ASYNC, cr);
6331 	if (error == EAGAIN)
6332 		error = 0;
6333 
6334 	/*
6335 	 * Do a flush based on the poff and plen arguments.  This
6336 	 * will synchronously write out any modified pages in the
6337 	 * range specified by (poff, plen) and wait until all of
6338 	 * the asynchronous i/o's in that range are done as well.
6339 	 */
6340 	if (!error)
6341 		error = nfs3_putpage(vp, poff, plen, 0, cr);
6342 
6343 	if (error)
6344 		return (error);
6345 
6346 	mutex_enter(&rp->r_statelock);
6347 	if (rp->r_verf != write_verf) {
6348 		mutex_exit(&rp->r_statelock);
6349 		goto top;
6350 	}
6351 	mutex_exit(&rp->r_statelock);
6352 
6353 	/*
6354 	 * Now commit any pages which might need to be committed.
6355 	 * If the error, NFS_VERF_MISMATCH, is returned, then
6356 	 * start over with the flush operation.
6357 	 */
6358 
6359 	error = nfs3_commit_vp(vp, poff, plen, cr);
6360 
6361 	if (error == NFS_VERF_MISMATCH)
6362 		goto top;
6363 
6364 	return (error);
6365 }
6366 
6367 static int
6368 nfs3_commit_vp(vnode_t *vp, u_offset_t poff, size_t plen, cred_t *cr)
6369 {
6370 	rnode_t *rp;
6371 	page_t *plist;
6372 	offset3 offset;
6373 	count3 len;
6374 
6375 
6376 	rp = VTOR(vp);
6377 
6378 	if (nfs_zone() != VTOMI(vp)->mi_zone)
6379 		return (EIO);
6380 	/*
6381 	 * Set the `commit inprogress' state bit.  We must
6382 	 * first wait until any current one finishes.
6383 	 */
6384 	mutex_enter(&rp->r_statelock);
6385 	while (rp->r_flags & RCOMMIT) {
6386 		rp->r_flags |= RCOMMITWAIT;
6387 		cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
6388 		rp->r_flags &= ~RCOMMITWAIT;
6389 	}
6390 	rp->r_flags |= RCOMMIT;
6391 	mutex_exit(&rp->r_statelock);
6392 
6393 	/*
6394 	 * Gather together all of the pages which need to be
6395 	 * committed.
6396 	 */
6397 	if (plen == 0)
6398 		nfs3_get_commit(vp);
6399 	else
6400 		nfs3_get_commit_range(vp, poff, plen);
6401 
6402 	/*
6403 	 * Clear the `commit inprogress' bit and disconnect the
6404 	 * page list which was gathered together in nfs3_get_commit.
6405 	 */
6406 	plist = rp->r_commit.c_pages;
6407 	rp->r_commit.c_pages = NULL;
6408 	offset = rp->r_commit.c_commbase;
6409 	len = rp->r_commit.c_commlen;
6410 	mutex_enter(&rp->r_statelock);
6411 	rp->r_flags &= ~RCOMMIT;
6412 	cv_broadcast(&rp->r_commit.c_cv);
6413 	mutex_exit(&rp->r_statelock);
6414 
6415 	/*
6416 	 * If any pages need to be committed, commit them and
6417 	 * then unlock them so that they can be freed some
6418 	 * time later.
6419 	 */
6420 	if (plist != NULL) {
6421 		/*
6422 		 * No error occurred during the flush portion
6423 		 * of this operation, so now attempt to commit
6424 		 * the data to stable storage on the server.
6425 		 *
6426 		 * This will unlock all of the pages on the list.
6427 		 */
6428 		return (nfs3_sync_commit(vp, plist, offset, len, cr));
6429 	}
6430 	return (0);
6431 }
6432 
6433 static int
6434 nfs3_sync_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6435 	cred_t *cr)
6436 {
6437 	int error;
6438 	page_t *pp;
6439 
6440 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6441 	error = nfs3_commit(vp, offset, count, cr);
6442 
6443 	/*
6444 	 * If we got an error, then just unlock all of the pages
6445 	 * on the list.
6446 	 */
6447 	if (error) {
6448 		while (plist != NULL) {
6449 			pp = plist;
6450 			page_sub(&plist, pp);
6451 			page_unlock(pp);
6452 		}
6453 		return (error);
6454 	}
6455 	/*
6456 	 * We've tried as hard as we can to commit the data to stable
6457 	 * storage on the server.  We just unlock the pages and clear
6458 	 * the commit required state.  They will get freed later.
6459 	 */
6460 	while (plist != NULL) {
6461 		pp = plist;
6462 		page_sub(&plist, pp);
6463 		pp->p_fsdata = C_NOCOMMIT;
6464 		page_unlock(pp);
6465 	}
6466 
6467 	return (error);
6468 }
6469 
6470 static void
6471 nfs3_async_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6472 	cred_t *cr)
6473 {
6474 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6475 	(void) nfs3_sync_commit(vp, plist, offset, count, cr);
6476 }
6477 
6478 static int
6479 nfs3_setsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
6480 {
6481 	int error;
6482 	mntinfo_t *mi;
6483 
6484 	mi = VTOMI(vp);
6485 
6486 	if (nfs_zone() != mi->mi_zone)
6487 		return (EIO);
6488 
6489 	if (mi->mi_flags & MI_ACL) {
6490 		error = acl_setacl3(vp, vsecattr, flag, cr);
6491 		if (mi->mi_flags & MI_ACL)
6492 			return (error);
6493 	}
6494 
6495 	return (ENOSYS);
6496 }
6497 
6498 static int
6499 nfs3_getsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
6500 {
6501 	int error;
6502 	mntinfo_t *mi;
6503 
6504 	mi = VTOMI(vp);
6505 
6506 	if (nfs_zone() != mi->mi_zone)
6507 		return (EIO);
6508 
6509 	if (mi->mi_flags & MI_ACL) {
6510 		error = acl_getacl3(vp, vsecattr, flag, cr);
6511 		if (mi->mi_flags & MI_ACL)
6512 			return (error);
6513 	}
6514 
6515 	return (fs_fab_acl(vp, vsecattr, flag, cr));
6516 }
6517 
6518 static int
6519 nfs3_shrlock(vnode_t *vp, int cmd, struct shrlock *shr, int flag, cred_t *cr)
6520 {
6521 	int error;
6522 	struct shrlock nshr;
6523 	struct nfs_owner nfs_owner;
6524 	netobj lm_fh3;
6525 
6526 	if (nfs_zone() != VTOMI(vp)->mi_zone)
6527 		return (EIO);
6528 
6529 	/*
6530 	 * check for valid cmd parameter
6531 	 */
6532 	if (cmd != F_SHARE && cmd != F_UNSHARE && cmd != F_HASREMOTELOCKS)
6533 		return (EINVAL);
6534 
6535 	/*
6536 	 * Check access permissions
6537 	 */
6538 	if (cmd == F_SHARE &&
6539 	    (((shr->s_access & F_RDACC) && !(flag & FREAD)) ||
6540 	    ((shr->s_access & F_WRACC) && !(flag & FWRITE))))
6541 		return (EBADF);
6542 
6543 	/*
6544 	 * If the filesystem is mounted using local locking, pass the
6545 	 * request off to the local share code.
6546 	 */
6547 	if (VTOMI(vp)->mi_flags & MI_LLOCK)
6548 		return (fs_shrlock(vp, cmd, shr, flag, cr));
6549 
6550 	switch (cmd) {
6551 	case F_SHARE:
6552 	case F_UNSHARE:
6553 		lm_fh3.n_len = VTOFH3(vp)->fh3_length;
6554 		lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
6555 
6556 		/*
6557 		 * If passed an owner that is too large to fit in an
6558 		 * nfs_owner it is likely a recursive call from the
6559 		 * lock manager client and pass it straight through.  If
6560 		 * it is not a nfs_owner then simply return an error.
6561 		 */
6562 		if (shr->s_own_len > sizeof (nfs_owner.lowner)) {
6563 			if (((struct nfs_owner *)shr->s_owner)->magic !=
6564 			    NFS_OWNER_MAGIC)
6565 				return (EINVAL);
6566 
6567 			if (error = lm4_shrlock(vp, cmd, shr, flag, &lm_fh3)) {
6568 				error = set_errno(error);
6569 			}
6570 			return (error);
6571 		}
6572 		/*
6573 		 * Remote share reservations owner is a combination of
6574 		 * a magic number, hostname, and the local owner
6575 		 */
6576 		bzero(&nfs_owner, sizeof (nfs_owner));
6577 		nfs_owner.magic = NFS_OWNER_MAGIC;
6578 		(void) strncpy(nfs_owner.hname, uts_nodename(),
6579 		    sizeof (nfs_owner.hname));
6580 		bcopy(shr->s_owner, nfs_owner.lowner, shr->s_own_len);
6581 		nshr.s_access = shr->s_access;
6582 		nshr.s_deny = shr->s_deny;
6583 		nshr.s_sysid = 0;
6584 		nshr.s_pid = ttoproc(curthread)->p_pid;
6585 		nshr.s_own_len = sizeof (nfs_owner);
6586 		nshr.s_owner = (caddr_t)&nfs_owner;
6587 
6588 		if (error = lm4_shrlock(vp, cmd, &nshr, flag, &lm_fh3)) {
6589 			error = set_errno(error);
6590 		}
6591 
6592 		break;
6593 
6594 	case F_HASREMOTELOCKS:
6595 		/*
6596 		 * NFS client can't store remote locks itself
6597 		 */
6598 		shr->s_access = 0;
6599 		error = 0;
6600 		break;
6601 
6602 	default:
6603 		error = EINVAL;
6604 		break;
6605 	}
6606 
6607 	return (error);
6608 }
6609