xref: /titanic_41/usr/src/uts/common/fs/nfs/nfs3_vnops.c (revision 4496171313bed39e96f21bc2f9faf2868e267ae3)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
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/file.h>
41 #include <sys/filio.h>
42 #include <sys/uio.h>
43 #include <sys/buf.h>
44 #include <sys/mman.h>
45 #include <sys/pathname.h>
46 #include <sys/dirent.h>
47 #include <sys/debug.h>
48 #include <sys/vmsystm.h>
49 #include <sys/fcntl.h>
50 #include <sys/flock.h>
51 #include <sys/swap.h>
52 #include <sys/errno.h>
53 #include <sys/strsubr.h>
54 #include <sys/sysmacros.h>
55 #include <sys/kmem.h>
56 #include <sys/cmn_err.h>
57 #include <sys/pathconf.h>
58 #include <sys/utsname.h>
59 #include <sys/dnlc.h>
60 #include <sys/acl.h>
61 #include <sys/systeminfo.h>
62 #include <sys/atomic.h>
63 #include <sys/policy.h>
64 #include <sys/sdt.h>
65 
66 #include <rpc/types.h>
67 #include <rpc/auth.h>
68 #include <rpc/clnt.h>
69 
70 #include <nfs/nfs.h>
71 #include <nfs/nfs_clnt.h>
72 #include <nfs/rnode.h>
73 #include <nfs/nfs_acl.h>
74 #include <nfs/lm.h>
75 
76 #include <vm/hat.h>
77 #include <vm/as.h>
78 #include <vm/page.h>
79 #include <vm/pvn.h>
80 #include <vm/seg.h>
81 #include <vm/seg_map.h>
82 #include <vm/seg_kpm.h>
83 #include <vm/seg_vn.h>
84 
85 #include <fs/fs_subr.h>
86 
87 #include <sys/ddi.h>
88 
89 static int	nfs3_rdwrlbn(vnode_t *, page_t *, u_offset_t, size_t, int,
90 			cred_t *);
91 static int	nfs3write(vnode_t *, caddr_t, u_offset_t, int, cred_t *,
92 			stable_how *);
93 static int	nfs3read(vnode_t *, caddr_t, offset_t, int, size_t *, cred_t *);
94 static int	nfs3setattr(vnode_t *, struct vattr *, int, cred_t *);
95 static int	nfs3_accessx(void *, int, cred_t *);
96 static int	nfs3lookup_dnlc(vnode_t *, char *, vnode_t **, cred_t *);
97 static int	nfs3lookup_otw(vnode_t *, char *, vnode_t **, cred_t *, int);
98 static int	nfs3create(vnode_t *, char *, struct vattr *, enum vcexcl,
99 			int, vnode_t **, cred_t *, int);
100 static int	nfs3excl_create_settimes(vnode_t *, struct vattr *, cred_t *);
101 static int	nfs3mknod(vnode_t *, char *, struct vattr *, enum vcexcl,
102 			int, vnode_t **, cred_t *);
103 static int	nfs3rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
104 static int	do_nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
105 static void	nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
106 static void	nfs3readdirplus(vnode_t *, rddir_cache *, cred_t *);
107 static int	nfs3_bio(struct buf *, stable_how *, cred_t *);
108 static int	nfs3_getapage(vnode_t *, u_offset_t, size_t, uint_t *,
109 			page_t *[], size_t, struct seg *, caddr_t,
110 			enum seg_rw, cred_t *);
111 static void	nfs3_readahead(vnode_t *, u_offset_t, caddr_t, struct seg *,
112 			cred_t *);
113 static int	nfs3_sync_putapage(vnode_t *, page_t *, u_offset_t, size_t,
114 			int, cred_t *);
115 static int	nfs3_sync_pageio(vnode_t *, page_t *, u_offset_t, size_t,
116 			int, cred_t *);
117 static int	nfs3_commit(vnode_t *, offset3, count3, cred_t *);
118 static void	nfs3_set_mod(vnode_t *);
119 static void	nfs3_get_commit(vnode_t *);
120 static void	nfs3_get_commit_range(vnode_t *, u_offset_t, size_t);
121 #if 0 /* unused */
122 #ifdef DEBUG
123 static int	nfs3_no_uncommitted_pages(vnode_t *);
124 #endif
125 #endif /* unused */
126 static int	nfs3_putpage_commit(vnode_t *, offset_t, size_t, cred_t *);
127 static int	nfs3_commit_vp(vnode_t *, u_offset_t, size_t,  cred_t *);
128 static int	nfs3_sync_commit(vnode_t *, page_t *, offset3, count3,
129 			cred_t *);
130 static void	nfs3_async_commit(vnode_t *, page_t *, offset3, count3,
131 			cred_t *);
132 static void	nfs3_delmap_callback(struct as *, void *, uint_t);
133 
134 /*
135  * Error flags used to pass information about certain special errors
136  * which need to be handled specially.
137  */
138 #define	NFS_EOF			-98
139 #define	NFS_VERF_MISMATCH	-97
140 
141 /* ALIGN64 aligns the given buffer and adjust buffer size to 64 bit */
142 #define	ALIGN64(x, ptr, sz)						\
143 	x = ((uintptr_t)(ptr)) & (sizeof (uint64_t) - 1);		\
144 	if (x) {							\
145 		x = sizeof (uint64_t) - (x);				\
146 		sz -= (x);						\
147 		ptr += (x);						\
148 	}
149 
150 /*
151  * These are the vnode ops routines which implement the vnode interface to
152  * the networked file system.  These routines just take their parameters,
153  * make them look networkish by putting the right info into interface structs,
154  * and then calling the appropriate remote routine(s) to do the work.
155  *
156  * Note on directory name lookup cacheing:  If we detect a stale fhandle,
157  * we purge the directory cache relative to that vnode.  This way, the
158  * user won't get burned by the cache repeatedly.  See <nfs/rnode.h> for
159  * more details on rnode locking.
160  */
161 
162 static int	nfs3_open(vnode_t **, int, cred_t *);
163 static int	nfs3_close(vnode_t *, int, int, offset_t, cred_t *);
164 static int	nfs3_read(vnode_t *, struct uio *, int, cred_t *,
165 			caller_context_t *);
166 static int	nfs3_write(vnode_t *, struct uio *, int, cred_t *,
167 			caller_context_t *);
168 static int	nfs3_ioctl(vnode_t *, int, intptr_t, int, cred_t *, int *);
169 static int	nfs3_getattr(vnode_t *, struct vattr *, int, cred_t *);
170 static int	nfs3_setattr(vnode_t *, struct vattr *, int, cred_t *,
171 			caller_context_t *);
172 static int	nfs3_access(vnode_t *, int, int, cred_t *);
173 static int	nfs3_readlink(vnode_t *, struct uio *, cred_t *);
174 static int	nfs3_fsync(vnode_t *, int, cred_t *);
175 static void	nfs3_inactive(vnode_t *, cred_t *);
176 static int	nfs3_lookup(vnode_t *, char *, vnode_t **,
177 			struct pathname *, int, vnode_t *, cred_t *);
178 static int	nfs3_create(vnode_t *, char *, struct vattr *, enum vcexcl,
179 			int, vnode_t **, cred_t *, int);
180 static int	nfs3_remove(vnode_t *, char *, cred_t *);
181 static int	nfs3_link(vnode_t *, vnode_t *, char *, cred_t *);
182 static int	nfs3_rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
183 static int	nfs3_mkdir(vnode_t *, char *, struct vattr *,
184 			vnode_t **, cred_t *);
185 static int	nfs3_rmdir(vnode_t *, char *, vnode_t *, cred_t *);
186 static int	nfs3_symlink(vnode_t *, char *, struct vattr *, char *,
187 			cred_t *);
188 static int	nfs3_readdir(vnode_t *, struct uio *, cred_t *, int *);
189 static int	nfs3_fid(vnode_t *, fid_t *);
190 static int	nfs3_rwlock(vnode_t *, int, caller_context_t *);
191 static void	nfs3_rwunlock(vnode_t *, int, caller_context_t *);
192 static int	nfs3_seek(vnode_t *, offset_t, offset_t *);
193 static int	nfs3_getpage(vnode_t *, offset_t, size_t, uint_t *,
194 			page_t *[], size_t, struct seg *, caddr_t,
195 			enum seg_rw, cred_t *);
196 static int	nfs3_putpage(vnode_t *, offset_t, size_t, int, cred_t *);
197 static int	nfs3_map(vnode_t *, offset_t, struct as *, caddr_t *,
198 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
199 static int	nfs3_addmap(vnode_t *, offset_t, struct as *, caddr_t,
200 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
201 static int	nfs3_frlock(vnode_t *, int, struct flock64 *, int, offset_t,
202 			struct flk_callback *, cred_t *);
203 static int	nfs3_space(vnode_t *, int, struct flock64 *, int, offset_t,
204 			cred_t *, caller_context_t *);
205 static int	nfs3_realvp(vnode_t *, vnode_t **);
206 static int	nfs3_delmap(vnode_t *, offset_t, struct as *, caddr_t,
207 			size_t, uint_t, uint_t, uint_t, cred_t *);
208 static int	nfs3_pathconf(vnode_t *, int, ulong_t *, cred_t *);
209 static int	nfs3_pageio(vnode_t *, page_t *, u_offset_t, size_t, int,
210 			cred_t *);
211 static void	nfs3_dispose(vnode_t *, page_t *, int, int, cred_t *);
212 static int	nfs3_setsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
213 static int	nfs3_getsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
214 static int	nfs3_shrlock(vnode_t *, int, struct shrlock *, int, cred_t *);
215 
216 struct vnodeops *nfs3_vnodeops;
217 
218 const fs_operation_def_t nfs3_vnodeops_template[] = {
219 	VOPNAME_OPEN, nfs3_open,
220 	VOPNAME_CLOSE, nfs3_close,
221 	VOPNAME_READ, nfs3_read,
222 	VOPNAME_WRITE, nfs3_write,
223 	VOPNAME_IOCTL, nfs3_ioctl,
224 	VOPNAME_GETATTR, nfs3_getattr,
225 	VOPNAME_SETATTR, nfs3_setattr,
226 	VOPNAME_ACCESS, nfs3_access,
227 	VOPNAME_LOOKUP, nfs3_lookup,
228 	VOPNAME_CREATE, nfs3_create,
229 	VOPNAME_REMOVE, nfs3_remove,
230 	VOPNAME_LINK, nfs3_link,
231 	VOPNAME_RENAME, nfs3_rename,
232 	VOPNAME_MKDIR, nfs3_mkdir,
233 	VOPNAME_RMDIR, nfs3_rmdir,
234 	VOPNAME_READDIR, nfs3_readdir,
235 	VOPNAME_SYMLINK, nfs3_symlink,
236 	VOPNAME_READLINK, nfs3_readlink,
237 	VOPNAME_FSYNC, nfs3_fsync,
238 	VOPNAME_INACTIVE, (fs_generic_func_p) nfs3_inactive,
239 	VOPNAME_FID, nfs3_fid,
240 	VOPNAME_RWLOCK, nfs3_rwlock,
241 	VOPNAME_RWUNLOCK, (fs_generic_func_p) nfs3_rwunlock,
242 	VOPNAME_SEEK, nfs3_seek,
243 	VOPNAME_FRLOCK, nfs3_frlock,
244 	VOPNAME_SPACE, nfs3_space,
245 	VOPNAME_REALVP, nfs3_realvp,
246 	VOPNAME_GETPAGE, nfs3_getpage,
247 	VOPNAME_PUTPAGE, nfs3_putpage,
248 	VOPNAME_MAP, (fs_generic_func_p) nfs3_map,
249 	VOPNAME_ADDMAP, (fs_generic_func_p) nfs3_addmap,
250 	VOPNAME_DELMAP, nfs3_delmap,
251 	VOPNAME_DUMP, nfs_dump,		/* there is no separate nfs3_dump */
252 	VOPNAME_PATHCONF, nfs3_pathconf,
253 	VOPNAME_PAGEIO, nfs3_pageio,
254 	VOPNAME_DISPOSE, (fs_generic_func_p) nfs3_dispose,
255 	VOPNAME_SETSECATTR, nfs3_setsecattr,
256 	VOPNAME_GETSECATTR, nfs3_getsecattr,
257 	VOPNAME_SHRLOCK, nfs3_shrlock,
258 	NULL, NULL
259 };
260 
261 /*
262  * XXX:  This is referenced in modstubs.s
263  */
264 struct vnodeops *
265 nfs3_getvnodeops(void)
266 {
267 	return (nfs3_vnodeops);
268 }
269 
270 /* ARGSUSED */
271 static int
272 nfs3_open(vnode_t **vpp, int flag, cred_t *cr)
273 {
274 	int error;
275 	struct vattr va;
276 	rnode_t *rp;
277 	vnode_t *vp;
278 
279 	vp = *vpp;
280 	if (nfs_zone() != VTOMI(vp)->mi_zone)
281 		return (EIO);
282 	rp = VTOR(vp);
283 	mutex_enter(&rp->r_statelock);
284 	if (rp->r_cred == NULL) {
285 		crhold(cr);
286 		rp->r_cred = cr;
287 	}
288 	mutex_exit(&rp->r_statelock);
289 
290 	/*
291 	 * If there is no cached data or if close-to-open
292 	 * consistency checking is turned off, we can avoid
293 	 * the over the wire getattr.  Otherwise, if the
294 	 * file system is mounted readonly, then just verify
295 	 * the caches are up to date using the normal mechanism.
296 	 * Else, if the file is not mmap'd, then just mark
297 	 * the attributes as timed out.  They will be refreshed
298 	 * and the caches validated prior to being used.
299 	 * Else, the file system is mounted writeable so
300 	 * force an over the wire GETATTR in order to ensure
301 	 * that all cached data is valid.
302 	 */
303 	if (vp->v_count > 1 ||
304 	    ((vn_has_cached_data(vp) || HAVE_RDDIR_CACHE(rp)) &&
305 	    !(VTOMI(vp)->mi_flags & MI_NOCTO))) {
306 		if (vn_is_readonly(vp))
307 			error = nfs3_validate_caches(vp, cr);
308 		else if (rp->r_mapcnt == 0 && vp->v_count == 1) {
309 			PURGE_ATTRCACHE(vp);
310 			error = 0;
311 		} else {
312 			va.va_mask = AT_ALL;
313 			error = nfs3_getattr_otw(vp, &va, cr);
314 		}
315 	} else
316 		error = 0;
317 
318 	return (error);
319 }
320 
321 static int
322 nfs3_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr)
323 {
324 	rnode_t *rp;
325 	int error;
326 	struct vattr va;
327 
328 	/*
329 	 * zone_enter(2) prevents processes from changing zones with NFS files
330 	 * open; if we happen to get here from the wrong zone we can't do
331 	 * anything over the wire.
332 	 */
333 	if (VTOMI(vp)->mi_zone != nfs_zone()) {
334 		/*
335 		 * We could attempt to clean up locks, except we're sure
336 		 * that the current process didn't acquire any locks on
337 		 * the file: any attempt to lock a file belong to another zone
338 		 * will fail, and one can't lock an NFS file and then change
339 		 * zones, as that fails too.
340 		 *
341 		 * Returning an error here is the sane thing to do.  A
342 		 * subsequent call to VN_RELE() which translates to a
343 		 * nfs3_inactive() will clean up state: if the zone of the
344 		 * vnode's origin is still alive and kicking, an async worker
345 		 * thread will handle the request (from the correct zone), and
346 		 * everything (minus the commit and final nfs3_getattr_otw()
347 		 * call) should be OK. If the zone is going away
348 		 * nfs_async_inactive() will throw away cached pages inline.
349 		 */
350 		return (EIO);
351 	}
352 
353 	/*
354 	 * If we are using local locking for this filesystem, then
355 	 * release all of the SYSV style record locks.  Otherwise,
356 	 * we are doing network locking and we need to release all
357 	 * of the network locks.  All of the locks held by this
358 	 * process on this file are released no matter what the
359 	 * incoming reference count is.
360 	 */
361 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
362 		cleanlocks(vp, ttoproc(curthread)->p_pid, 0);
363 		cleanshares(vp, ttoproc(curthread)->p_pid);
364 	} else
365 		nfs_lockrelease(vp, flag, offset, cr);
366 
367 	if (count > 1)
368 		return (0);
369 
370 	/*
371 	 * If the file has been `unlinked', then purge the
372 	 * DNLC so that this vnode will get reycled quicker
373 	 * and the .nfs* file on the server will get removed.
374 	 */
375 	rp = VTOR(vp);
376 	if (rp->r_unldvp != NULL)
377 		dnlc_purge_vp(vp);
378 
379 	/*
380 	 * If the file was open for write and there are pages,
381 	 * then if the file system was mounted using the "no-close-
382 	 *	to-open" semantics, then start an asynchronous flush
383 	 *	of the all of the pages in the file.
384 	 * else the file system was not mounted using the "no-close-
385 	 *	to-open" semantics, then do a synchronous flush and
386 	 *	commit of all of the dirty and uncommitted pages.
387 	 *
388 	 * The asynchronous flush of the pages in the "nocto" path
389 	 * mostly just associates a cred pointer with the rnode so
390 	 * writes which happen later will have a better chance of
391 	 * working.  It also starts the data being written to the
392 	 * server, but without unnecessarily delaying the application.
393 	 */
394 	if ((flag & FWRITE) && vn_has_cached_data(vp)) {
395 		if (VTOMI(vp)->mi_flags & MI_NOCTO) {
396 			error = nfs3_putpage(vp, (offset_t)0, 0, B_ASYNC, cr);
397 			if (error == EAGAIN)
398 				error = 0;
399 		} else
400 			error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
401 		if (!error) {
402 			mutex_enter(&rp->r_statelock);
403 			error = rp->r_error;
404 			rp->r_error = 0;
405 			mutex_exit(&rp->r_statelock);
406 		}
407 	} else {
408 		mutex_enter(&rp->r_statelock);
409 		error = rp->r_error;
410 		rp->r_error = 0;
411 		mutex_exit(&rp->r_statelock);
412 	}
413 
414 	/*
415 	 * If RWRITEATTR is set, then issue an over the wire GETATTR to
416 	 * refresh the attribute cache with a set of attributes which
417 	 * weren't returned from a WRITE.  This will enable the close-
418 	 * to-open processing to work.
419 	 */
420 	if (rp->r_flags & RWRITEATTR)
421 		(void) nfs3_getattr_otw(vp, &va, cr);
422 
423 	return (error);
424 }
425 
426 /* ARGSUSED */
427 static int
428 nfs3_directio_read(vnode_t *vp, struct uio *uiop, cred_t *cr)
429 {
430 	mntinfo_t *mi;
431 	READ3args args;
432 	READ3uiores res;
433 	int tsize;
434 	offset_t offset;
435 	ssize_t count;
436 	int error;
437 	int douprintf;
438 	failinfo_t fi;
439 	char *sv_hostname;
440 
441 	mi = VTOMI(vp);
442 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
443 	sv_hostname = VTOR(vp)->r_server->sv_hostname;
444 
445 	douprintf = 1;
446 	args.file = *VTOFH3(vp);
447 	fi.vp = vp;
448 	fi.fhp = (caddr_t)&args.file;
449 	fi.copyproc = nfs3copyfh;
450 	fi.lookupproc = nfs3lookup;
451 	fi.xattrdirproc = acl_getxattrdir3;
452 
453 	res.uiop = uiop;
454 
455 	offset = uiop->uio_loffset;
456 	count = uiop->uio_resid;
457 
458 	do {
459 		if (mi->mi_io_kstats) {
460 			mutex_enter(&mi->mi_lock);
461 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
462 			mutex_exit(&mi->mi_lock);
463 		}
464 
465 		do {
466 			tsize = MIN(mi->mi_tsize, count);
467 			args.offset = (offset3)offset;
468 			args.count = (count3)tsize;
469 			res.size = (uint_t)tsize;
470 			error = rfs3call(mi, NFSPROC3_READ,
471 				    xdr_READ3args, (caddr_t)&args,
472 				    xdr_READ3uiores, (caddr_t)&res, cr,
473 				    &douprintf, &res.status, 0, &fi);
474 		} while (error == ENFS_TRYAGAIN);
475 
476 		if (mi->mi_io_kstats) {
477 			mutex_enter(&mi->mi_lock);
478 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
479 			mutex_exit(&mi->mi_lock);
480 		}
481 
482 		if (error)
483 			return (error);
484 
485 		error = geterrno3(res.status);
486 		if (error)
487 			return (error);
488 
489 		if (res.count != res.size) {
490 			zcmn_err(getzoneid(), CE_WARN,
491 "nfs3_directio_read: server %s returned incorrect amount",
492 					sv_hostname);
493 			return (EIO);
494 		}
495 		count -= res.count;
496 		offset += res.count;
497 		if (mi->mi_io_kstats) {
498 			mutex_enter(&mi->mi_lock);
499 			KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
500 			KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
501 			mutex_exit(&mi->mi_lock);
502 		}
503 		lwp_stat_update(LWP_STAT_INBLK, 1);
504 	} while (count && !res.eof);
505 
506 	return (0);
507 }
508 
509 /* ARGSUSED */
510 static int
511 nfs3_read(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
512 	caller_context_t *ct)
513 {
514 	rnode_t *rp;
515 	u_offset_t off;
516 	offset_t diff;
517 	int on;
518 	size_t n;
519 	caddr_t base;
520 	uint_t flags;
521 	int error = 0;
522 	mntinfo_t *mi;
523 
524 	rp = VTOR(vp);
525 	mi = VTOMI(vp);
526 
527 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
528 
529 	if (nfs_zone() != mi->mi_zone)
530 		return (EIO);
531 
532 	if (vp->v_type != VREG)
533 		return (EISDIR);
534 
535 	if (uiop->uio_resid == 0)
536 		return (0);
537 
538 	if (uiop->uio_loffset < 0 || uiop->uio_loffset + uiop->uio_resid < 0)
539 		return (EINVAL);
540 
541 	/*
542 	 * Bypass VM if caching has been disabled (e.g., locking) or if
543 	 * using client-side direct I/O and the file is not mmap'd and
544 	 * there are no cached pages.
545 	 */
546 	if ((vp->v_flag & VNOCACHE) ||
547 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
548 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
549 		return (nfs3_directio_read(vp, uiop, cr));
550 	}
551 
552 	do {
553 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
554 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
555 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
556 
557 		error = nfs3_validate_caches(vp, cr);
558 		if (error)
559 			break;
560 
561 		mutex_enter(&rp->r_statelock);
562 		diff = rp->r_size - uiop->uio_loffset;
563 		mutex_exit(&rp->r_statelock);
564 		if (diff <= 0)
565 			break;
566 		if (diff < n)
567 			n = (size_t)diff;
568 
569 		if (vpm_enable) {
570 			/*
571 			 * Copy data.
572 			 */
573 			error = vpm_data_copy(vp, off + on, n, uiop,
574 						1, NULL, 0, S_READ);
575 		} else {
576 			base = segmap_getmapflt(segkmap, vp, off + on, n, 1,
577 							S_READ);
578 
579 			error = uiomove(base + on, n, UIO_READ, uiop);
580 		}
581 
582 		if (!error) {
583 			/*
584 			 * If read a whole block or read to eof,
585 			 * won't need this buffer again soon.
586 			 */
587 			mutex_enter(&rp->r_statelock);
588 			if (n + on == MAXBSIZE ||
589 			    uiop->uio_loffset == rp->r_size)
590 				flags = SM_DONTNEED;
591 			else
592 				flags = 0;
593 			mutex_exit(&rp->r_statelock);
594 			if (vpm_enable) {
595 				error = vpm_sync_pages(vp, off, n, flags);
596 			} else {
597 				error = segmap_release(segkmap, base, flags);
598 			}
599 		} else {
600 			if (vpm_enable) {
601 				(void) vpm_sync_pages(vp, off, n, 0);
602 			} else {
603 				(void) segmap_release(segkmap, base, 0);
604 			}
605 		}
606 	} while (!error && uiop->uio_resid > 0);
607 
608 	return (error);
609 }
610 
611 /* ARGSUSED */
612 static int
613 nfs3_write(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
614 	caller_context_t *ct)
615 {
616 	rlim64_t limit = uiop->uio_llimit;
617 	rnode_t *rp;
618 	u_offset_t off;
619 	caddr_t base;
620 	uint_t flags;
621 	int remainder;
622 	size_t n;
623 	int on;
624 	int error;
625 	int resid;
626 	offset_t offset;
627 	mntinfo_t *mi;
628 	uint_t bsize;
629 
630 	rp = VTOR(vp);
631 
632 	if (vp->v_type != VREG)
633 		return (EISDIR);
634 
635 	mi = VTOMI(vp);
636 	if (nfs_zone() != mi->mi_zone)
637 		return (EIO);
638 	if (uiop->uio_resid == 0)
639 		return (0);
640 
641 	if (ioflag & FAPPEND) {
642 		struct vattr va;
643 
644 		/*
645 		 * Must serialize if appending.
646 		 */
647 		if (nfs_rw_lock_held(&rp->r_rwlock, RW_READER)) {
648 			nfs_rw_exit(&rp->r_rwlock);
649 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER,
650 			    INTR(vp)))
651 				return (EINTR);
652 		}
653 
654 		va.va_mask = AT_SIZE;
655 		error = nfs3getattr(vp, &va, cr);
656 		if (error)
657 			return (error);
658 		uiop->uio_loffset = va.va_size;
659 	}
660 
661 	offset = uiop->uio_loffset + uiop->uio_resid;
662 
663 	if (uiop->uio_loffset < 0 || offset < 0)
664 		return (EINVAL);
665 
666 	if (limit == RLIM64_INFINITY || limit > MAXOFFSET_T)
667 		limit = MAXOFFSET_T;
668 
669 	/*
670 	 * Check to make sure that the process will not exceed
671 	 * its limit on file size.  It is okay to write up to
672 	 * the limit, but not beyond.  Thus, the write which
673 	 * reaches the limit will be short and the next write
674 	 * will return an error.
675 	 */
676 	remainder = 0;
677 	if (offset > limit) {
678 		remainder = offset - limit;
679 		uiop->uio_resid = limit - uiop->uio_loffset;
680 		if (uiop->uio_resid <= 0) {
681 			proc_t *p = ttoproc(curthread);
682 
683 			uiop->uio_resid += remainder;
684 			mutex_enter(&p->p_lock);
685 			(void) rctl_action(rctlproc_legacy[RLIMIT_FSIZE],
686 			    p->p_rctls, p, RCA_UNSAFE_SIGINFO);
687 			mutex_exit(&p->p_lock);
688 			return (EFBIG);
689 		}
690 	}
691 
692 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
693 		return (EINTR);
694 
695 	/*
696 	 * Bypass VM if caching has been disabled (e.g., locking) or if
697 	 * using client-side direct I/O and the file is not mmap'd and
698 	 * there are no cached pages.
699 	 */
700 	if ((vp->v_flag & VNOCACHE) ||
701 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
702 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
703 		size_t bufsize;
704 		int count;
705 		u_offset_t org_offset;
706 		stable_how stab_comm;
707 
708 nfs3_fwrite:
709 		if (rp->r_flags & RSTALE) {
710 			resid = uiop->uio_resid;
711 			offset = uiop->uio_loffset;
712 			error = rp->r_error;
713 			goto bottom;
714 		}
715 		bufsize = MIN(uiop->uio_resid, mi->mi_stsize);
716 		base = kmem_alloc(bufsize, KM_SLEEP);
717 		do {
718 			if (ioflag & FDSYNC)
719 				stab_comm = DATA_SYNC;
720 			else
721 				stab_comm = FILE_SYNC;
722 			resid = uiop->uio_resid;
723 			offset = uiop->uio_loffset;
724 			count = MIN(uiop->uio_resid, bufsize);
725 			org_offset = uiop->uio_loffset;
726 			error = uiomove(base, count, UIO_WRITE, uiop);
727 			if (!error) {
728 				error = nfs3write(vp, base, org_offset,
729 				    count, cr, &stab_comm);
730 			}
731 		} while (!error && uiop->uio_resid > 0);
732 		kmem_free(base, bufsize);
733 		goto bottom;
734 	}
735 
736 
737 	bsize = vp->v_vfsp->vfs_bsize;
738 
739 	do {
740 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
741 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
742 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
743 
744 		resid = uiop->uio_resid;
745 		offset = uiop->uio_loffset;
746 
747 		if (rp->r_flags & RSTALE) {
748 			error = rp->r_error;
749 			break;
750 		}
751 
752 		/*
753 		 * Don't create dirty pages faster than they
754 		 * can be cleaned so that the system doesn't
755 		 * get imbalanced.  If the async queue is
756 		 * maxed out, then wait for it to drain before
757 		 * creating more dirty pages.  Also, wait for
758 		 * any threads doing pagewalks in the vop_getattr
759 		 * entry points so that they don't block for
760 		 * long periods.
761 		 */
762 		mutex_enter(&rp->r_statelock);
763 		while ((mi->mi_max_threads != 0 &&
764 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
765 		    rp->r_gcount > 0)
766 			cv_wait(&rp->r_cv, &rp->r_statelock);
767 		mutex_exit(&rp->r_statelock);
768 
769 		if (vpm_enable) {
770 			/*
771 			 * It will use kpm mappings, so no need to
772 			 * pass an address.
773 			 */
774 			error = writerp(rp, NULL, n, uiop, 0);
775 		} else  {
776 			if (segmap_kpm) {
777 				int pon = uiop->uio_loffset & PAGEOFFSET;
778 				size_t pn = MIN(PAGESIZE - pon,
779 							uiop->uio_resid);
780 				int pagecreate;
781 
782 				mutex_enter(&rp->r_statelock);
783 				pagecreate = (pon == 0) && (pn == PAGESIZE ||
784 					uiop->uio_loffset + pn >= rp->r_size);
785 				mutex_exit(&rp->r_statelock);
786 
787 				base = segmap_getmapflt(segkmap, vp, off + on,
788 						pn, !pagecreate, S_WRITE);
789 
790 				error = writerp(rp, base + pon, n, uiop,
791 								pagecreate);
792 
793 			} else {
794 				base = segmap_getmapflt(segkmap, vp, off + on,
795 							n, 0, S_READ);
796 				error = writerp(rp, base + on, n, uiop, 0);
797 			}
798 		}
799 
800 		if (!error) {
801 			if (mi->mi_flags & MI_NOAC)
802 				flags = SM_WRITE;
803 			else if ((uiop->uio_loffset % bsize) == 0 ||
804 			    IS_SWAPVP(vp)) {
805 				/*
806 				 * Have written a whole block.
807 				 * Start an asynchronous write
808 				 * and mark the buffer to
809 				 * indicate that it won't be
810 				 * needed again soon.
811 				 */
812 				flags = SM_WRITE | SM_ASYNC | SM_DONTNEED;
813 			} else
814 				flags = 0;
815 			if ((ioflag & (FSYNC|FDSYNC)) ||
816 			    (rp->r_flags & ROUTOFSPACE)) {
817 				flags &= ~SM_ASYNC;
818 				flags |= SM_WRITE;
819 			}
820 			if (vpm_enable) {
821 				error = vpm_sync_pages(vp, off, n, flags);
822 			} else {
823 				error = segmap_release(segkmap, base, flags);
824 			}
825 		} else {
826 			if (vpm_enable) {
827 				(void) vpm_sync_pages(vp, off, n, 0);
828 			} else {
829 				(void) segmap_release(segkmap, base, 0);
830 			}
831 			/*
832 			 * In the event that we got an access error while
833 			 * faulting in a page for a write-only file just
834 			 * force a write.
835 			 */
836 			if (error == EACCES)
837 				goto nfs3_fwrite;
838 		}
839 	} while (!error && uiop->uio_resid > 0);
840 
841 bottom:
842 	if (error) {
843 		uiop->uio_resid = resid + remainder;
844 		uiop->uio_loffset = offset;
845 	} else
846 		uiop->uio_resid += remainder;
847 
848 	nfs_rw_exit(&rp->r_lkserlock);
849 
850 	return (error);
851 }
852 
853 /*
854  * Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED}
855  */
856 static int
857 nfs3_rdwrlbn(vnode_t *vp, page_t *pp, u_offset_t off, size_t len,
858 	int flags, cred_t *cr)
859 {
860 	struct buf *bp;
861 	int error;
862 	page_t *savepp;
863 	uchar_t fsdata;
864 	stable_how stab_comm;
865 
866 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
867 	bp = pageio_setup(pp, len, vp, flags);
868 	ASSERT(bp != NULL);
869 
870 	/*
871 	 * pageio_setup should have set b_addr to 0.  This
872 	 * is correct since we want to do I/O on a page
873 	 * boundary.  bp_mapin will use this addr to calculate
874 	 * an offset, and then set b_addr to the kernel virtual
875 	 * address it allocated for us.
876 	 */
877 	ASSERT(bp->b_un.b_addr == 0);
878 
879 	bp->b_edev = 0;
880 	bp->b_dev = 0;
881 	bp->b_lblkno = lbtodb(off);
882 	bp->b_file = vp;
883 	bp->b_offset = (offset_t)off;
884 	bp_mapin(bp);
885 
886 	/*
887 	 * Calculate the desired level of stability to write data
888 	 * on the server and then mark all of the pages to reflect
889 	 * this.
890 	 */
891 	if ((flags & (B_WRITE|B_ASYNC)) == (B_WRITE|B_ASYNC) &&
892 	    freemem > desfree) {
893 		stab_comm = UNSTABLE;
894 		fsdata = C_DELAYCOMMIT;
895 	} else {
896 		stab_comm = FILE_SYNC;
897 		fsdata = C_NOCOMMIT;
898 	}
899 
900 	savepp = pp;
901 	do {
902 		pp->p_fsdata = fsdata;
903 	} while ((pp = pp->p_next) != savepp);
904 
905 	error = nfs3_bio(bp, &stab_comm, cr);
906 
907 	bp_mapout(bp);
908 	pageio_done(bp);
909 
910 	/*
911 	 * If the server wrote pages in a more stable fashion than
912 	 * was requested, then clear all of the marks in the pages
913 	 * indicating that COMMIT operations were required.
914 	 */
915 	if (stab_comm != UNSTABLE && fsdata == C_DELAYCOMMIT) {
916 		do {
917 			pp->p_fsdata = C_NOCOMMIT;
918 		} while ((pp = pp->p_next) != savepp);
919 	}
920 
921 	return (error);
922 }
923 
924 /*
925  * Write to file.  Writes to remote server in largest size
926  * chunks that the server can handle.  Write is synchronous.
927  */
928 static int
929 nfs3write(vnode_t *vp, caddr_t base, u_offset_t offset, int count, cred_t *cr,
930 	stable_how *stab_comm)
931 {
932 	mntinfo_t *mi;
933 	WRITE3args args;
934 	WRITE3res res;
935 	int error;
936 	int tsize;
937 	rnode_t *rp;
938 	int douprintf;
939 
940 	rp = VTOR(vp);
941 	mi = VTOMI(vp);
942 
943 	ASSERT(nfs_zone() == mi->mi_zone);
944 
945 	args.file = *VTOFH3(vp);
946 	args.stable = *stab_comm;
947 
948 	*stab_comm = FILE_SYNC;
949 
950 	douprintf = 1;
951 
952 	do {
953 		if ((vp->v_flag & VNOCACHE) ||
954 		    (rp->r_flags & RDIRECTIO) ||
955 		    (mi->mi_flags & MI_DIRECTIO))
956 			tsize = MIN(mi->mi_stsize, count);
957 		else
958 			tsize = MIN(mi->mi_curwrite, count);
959 		args.offset = (offset3)offset;
960 		args.count = (count3)tsize;
961 		args.data.data_len = (uint_t)tsize;
962 		args.data.data_val = base;
963 
964 		if (mi->mi_io_kstats) {
965 			mutex_enter(&mi->mi_lock);
966 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
967 			mutex_exit(&mi->mi_lock);
968 		}
969 		args.mblk = NULL;
970 		do {
971 			error = rfs3call(mi, NFSPROC3_WRITE,
972 			    xdr_WRITE3args, (caddr_t)&args,
973 			    xdr_WRITE3res, (caddr_t)&res, cr,
974 			    &douprintf, &res.status, 0, NULL);
975 		} while (error == ENFS_TRYAGAIN);
976 		if (mi->mi_io_kstats) {
977 			mutex_enter(&mi->mi_lock);
978 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
979 			mutex_exit(&mi->mi_lock);
980 		}
981 
982 		if (error)
983 			return (error);
984 		error = geterrno3(res.status);
985 		if (!error) {
986 			if (res.resok.count > args.count) {
987 				zcmn_err(getzoneid(), CE_WARN,
988 				    "nfs3write: server %s wrote %u, "
989 				    "requested was %u",
990 				    rp->r_server->sv_hostname,
991 				    res.resok.count, args.count);
992 				return (EIO);
993 			}
994 			if (res.resok.committed == UNSTABLE) {
995 				*stab_comm = UNSTABLE;
996 				if (args.stable == DATA_SYNC ||
997 				    args.stable == FILE_SYNC) {
998 					zcmn_err(getzoneid(), CE_WARN,
999 			"nfs3write: server %s did not commit to stable storage",
1000 					    rp->r_server->sv_hostname);
1001 					return (EIO);
1002 				}
1003 			}
1004 			tsize = (int)res.resok.count;
1005 			count -= tsize;
1006 			base += tsize;
1007 			offset += tsize;
1008 			if (mi->mi_io_kstats) {
1009 				mutex_enter(&mi->mi_lock);
1010 				KSTAT_IO_PTR(mi->mi_io_kstats)->writes++;
1011 				KSTAT_IO_PTR(mi->mi_io_kstats)->nwritten +=
1012 				    tsize;
1013 				mutex_exit(&mi->mi_lock);
1014 			}
1015 			lwp_stat_update(LWP_STAT_OUBLK, 1);
1016 			mutex_enter(&rp->r_statelock);
1017 			if (rp->r_flags & RHAVEVERF) {
1018 				if (rp->r_verf != res.resok.verf) {
1019 					nfs3_set_mod(vp);
1020 					rp->r_verf = res.resok.verf;
1021 					/*
1022 					 * If the data was written UNSTABLE,
1023 					 * then might as well stop because
1024 					 * the whole block will have to get
1025 					 * rewritten anyway.
1026 					 */
1027 					if (*stab_comm == UNSTABLE) {
1028 						mutex_exit(&rp->r_statelock);
1029 						break;
1030 					}
1031 				}
1032 			} else {
1033 				rp->r_verf = res.resok.verf;
1034 				rp->r_flags |= RHAVEVERF;
1035 			}
1036 			/*
1037 			 * Mark the attribute cache as timed out and
1038 			 * set RWRITEATTR to indicate that the file
1039 			 * was modified with a WRITE operation and
1040 			 * that the attributes can not be trusted.
1041 			 */
1042 			PURGE_ATTRCACHE_LOCKED(rp);
1043 			rp->r_flags |= RWRITEATTR;
1044 			mutex_exit(&rp->r_statelock);
1045 		}
1046 	} while (!error && count);
1047 
1048 	return (error);
1049 }
1050 
1051 /*
1052  * Read from a file.  Reads data in largest chunks our interface can handle.
1053  */
1054 static int
1055 nfs3read(vnode_t *vp, caddr_t base, offset_t offset, int count,
1056 	size_t *residp, cred_t *cr)
1057 {
1058 	mntinfo_t *mi;
1059 	READ3args args;
1060 	READ3vres res;
1061 	int tsize;
1062 	int error;
1063 	int douprintf;
1064 	failinfo_t fi;
1065 	rnode_t *rp;
1066 	struct vattr va;
1067 	hrtime_t t;
1068 
1069 	rp = VTOR(vp);
1070 	mi = VTOMI(vp);
1071 	ASSERT(nfs_zone() == mi->mi_zone);
1072 	douprintf = 1;
1073 
1074 	args.file = *VTOFH3(vp);
1075 	fi.vp = vp;
1076 	fi.fhp = (caddr_t)&args.file;
1077 	fi.copyproc = nfs3copyfh;
1078 	fi.lookupproc = nfs3lookup;
1079 	fi.xattrdirproc = acl_getxattrdir3;
1080 
1081 	res.pov.fres.vp = vp;
1082 	res.pov.fres.vap = &va;
1083 
1084 	*residp = count;
1085 	do {
1086 		if (mi->mi_io_kstats) {
1087 			mutex_enter(&mi->mi_lock);
1088 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1089 			mutex_exit(&mi->mi_lock);
1090 		}
1091 
1092 		do {
1093 			if ((vp->v_flag & VNOCACHE) ||
1094 			    (rp->r_flags & RDIRECTIO) ||
1095 			    (mi->mi_flags & MI_DIRECTIO))
1096 				tsize = MIN(mi->mi_tsize, count);
1097 			else
1098 				tsize = MIN(mi->mi_curread, count);
1099 			res.data.data_val = base;
1100 			res.data.data_len = tsize;
1101 			args.offset = (offset3)offset;
1102 			args.count = (count3)tsize;
1103 			t = gethrtime();
1104 			error = rfs3call(mi, NFSPROC3_READ,
1105 			    xdr_READ3args, (caddr_t)&args,
1106 			    xdr_READ3vres, (caddr_t)&res, cr,
1107 			    &douprintf, &res.status, 0, &fi);
1108 		} while (error == ENFS_TRYAGAIN);
1109 
1110 		if (mi->mi_io_kstats) {
1111 			mutex_enter(&mi->mi_lock);
1112 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
1113 			mutex_exit(&mi->mi_lock);
1114 		}
1115 
1116 		if (error)
1117 			return (error);
1118 
1119 		error = geterrno3(res.status);
1120 		if (error)
1121 			return (error);
1122 
1123 		if (res.count != res.data.data_len) {
1124 			zcmn_err(getzoneid(), CE_WARN,
1125 				"nfs3read: server %s returned incorrect amount",
1126 				rp->r_server->sv_hostname);
1127 			return (EIO);
1128 		}
1129 
1130 		count -= res.count;
1131 		*residp = count;
1132 		base += res.count;
1133 		offset += res.count;
1134 		if (mi->mi_io_kstats) {
1135 			mutex_enter(&mi->mi_lock);
1136 			KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
1137 			KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
1138 			mutex_exit(&mi->mi_lock);
1139 		}
1140 		lwp_stat_update(LWP_STAT_INBLK, 1);
1141 	} while (count && !res.eof);
1142 
1143 	if (res.pov.attributes) {
1144 		mutex_enter(&rp->r_statelock);
1145 		if (!CACHE_VALID(rp, va.va_mtime, va.va_size)) {
1146 			mutex_exit(&rp->r_statelock);
1147 			PURGE_ATTRCACHE(vp);
1148 		} else {
1149 			if (rp->r_mtime <= t)
1150 				nfs_attrcache_va(vp, &va);
1151 			mutex_exit(&rp->r_statelock);
1152 		}
1153 	}
1154 
1155 	return (0);
1156 }
1157 
1158 /* ARGSUSED */
1159 static int
1160 nfs3_ioctl(vnode_t *vp, int cmd, intptr_t arg, int flag, cred_t *cr, int *rvalp)
1161 {
1162 
1163 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1164 		return (EIO);
1165 	switch (cmd) {
1166 		case _FIODIRECTIO:
1167 			return (nfs_directio(vp, (int)arg, cr));
1168 		default:
1169 			return (ENOTTY);
1170 	}
1171 }
1172 
1173 static int
1174 nfs3_getattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1175 {
1176 	int error;
1177 	rnode_t *rp;
1178 
1179 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1180 		return (EIO);
1181 	/*
1182 	 * If it has been specified that the return value will
1183 	 * just be used as a hint, and we are only being asked
1184 	 * for size, fsid or rdevid, then return the client's
1185 	 * notion of these values without checking to make sure
1186 	 * that the attribute cache is up to date.
1187 	 * The whole point is to avoid an over the wire GETATTR
1188 	 * call.
1189 	 */
1190 	rp = VTOR(vp);
1191 	if (flags & ATTR_HINT) {
1192 		if (vap->va_mask ==
1193 		    (vap->va_mask & (AT_SIZE | AT_FSID | AT_RDEV))) {
1194 			mutex_enter(&rp->r_statelock);
1195 			if (vap->va_mask | AT_SIZE)
1196 				vap->va_size = rp->r_size;
1197 			if (vap->va_mask | AT_FSID)
1198 				vap->va_fsid = rp->r_attr.va_fsid;
1199 			if (vap->va_mask | AT_RDEV)
1200 				vap->va_rdev = rp->r_attr.va_rdev;
1201 			mutex_exit(&rp->r_statelock);
1202 			return (0);
1203 		}
1204 	}
1205 
1206 	/*
1207 	 * Only need to flush pages if asking for the mtime
1208 	 * and if there any dirty pages or any outstanding
1209 	 * asynchronous (write) requests for this file.
1210 	 */
1211 	if (vap->va_mask & AT_MTIME) {
1212 		if (vn_has_cached_data(vp) &&
1213 		    ((rp->r_flags & RDIRTY) || rp->r_awcount > 0)) {
1214 			mutex_enter(&rp->r_statelock);
1215 			rp->r_gcount++;
1216 			mutex_exit(&rp->r_statelock);
1217 			error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1218 			mutex_enter(&rp->r_statelock);
1219 			if (error && (error == ENOSPC || error == EDQUOT)) {
1220 				if (!rp->r_error)
1221 					rp->r_error = error;
1222 			}
1223 			if (--rp->r_gcount == 0)
1224 				cv_broadcast(&rp->r_cv);
1225 			mutex_exit(&rp->r_statelock);
1226 		}
1227 	}
1228 
1229 	return (nfs3getattr(vp, vap, cr));
1230 }
1231 
1232 /*ARGSUSED4*/
1233 static int
1234 nfs3_setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr,
1235 		caller_context_t *ct)
1236 {
1237 	int error;
1238 	struct vattr va;
1239 
1240 	if (vap->va_mask & AT_NOSET)
1241 		return (EINVAL);
1242 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1243 		return (EIO);
1244 
1245 	va.va_mask = AT_UID | AT_MODE;
1246 	error = nfs3getattr(vp, &va, cr);
1247 	if (error)
1248 		return (error);
1249 
1250 	error = secpolicy_vnode_setattr(cr, vp, vap, &va, flags, nfs3_accessx,
1251 		vp);
1252 	if (error)
1253 		return (error);
1254 
1255 	return (nfs3setattr(vp, vap, flags, cr));
1256 }
1257 
1258 static int
1259 nfs3setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1260 {
1261 	int error;
1262 	uint_t mask;
1263 	SETATTR3args args;
1264 	SETATTR3res res;
1265 	int douprintf;
1266 	rnode_t *rp;
1267 	struct vattr va;
1268 	mode_t omode;
1269 	vsecattr_t *vsp;
1270 	hrtime_t t;
1271 
1272 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
1273 	mask = vap->va_mask;
1274 
1275 	rp = VTOR(vp);
1276 
1277 	/*
1278 	 * Only need to flush pages if there are any pages and
1279 	 * if the file is marked as dirty in some fashion.  The
1280 	 * file must be flushed so that we can accurately
1281 	 * determine the size of the file and the cached data
1282 	 * after the SETATTR returns.  A file is considered to
1283 	 * be dirty if it is either marked with RDIRTY, has
1284 	 * outstanding i/o's active, or is mmap'd.  In this
1285 	 * last case, we can't tell whether there are dirty
1286 	 * pages, so we flush just to be sure.
1287 	 */
1288 	if (vn_has_cached_data(vp) &&
1289 	    ((rp->r_flags & RDIRTY) ||
1290 	    rp->r_count > 0 ||
1291 	    rp->r_mapcnt > 0)) {
1292 		ASSERT(vp->v_type != VCHR);
1293 		error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1294 		if (error && (error == ENOSPC || error == EDQUOT)) {
1295 			mutex_enter(&rp->r_statelock);
1296 			if (!rp->r_error)
1297 				rp->r_error = error;
1298 			mutex_exit(&rp->r_statelock);
1299 		}
1300 	}
1301 
1302 	args.object = *RTOFH3(rp);
1303 	/*
1304 	 * If the intent is for the server to set the times,
1305 	 * there is no point in have the mask indicating set mtime or
1306 	 * atime, because the vap values may be junk, and so result
1307 	 * in an overflow error. Remove these flags from the vap mask
1308 	 * before calling in this case, and restore them afterwards.
1309 	 */
1310 	if ((mask & (AT_ATIME | AT_MTIME)) && !(flags & ATTR_UTIME)) {
1311 		/* Use server times, so don't set the args time fields */
1312 		vap->va_mask &= ~(AT_ATIME | AT_MTIME);
1313 		error = vattr_to_sattr3(vap, &args.new_attributes);
1314 		vap->va_mask |= (mask & (AT_ATIME | AT_MTIME));
1315 		if (mask & AT_ATIME) {
1316 			args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
1317 		}
1318 		if (mask & AT_MTIME) {
1319 			args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
1320 		}
1321 	} else {
1322 		/* Either do not set times or use the client specified times */
1323 		error = vattr_to_sattr3(vap, &args.new_attributes);
1324 	}
1325 
1326 	if (error) {
1327 		/* req time field(s) overflow - return immediately */
1328 		return (error);
1329 	}
1330 
1331 	va.va_mask = AT_MODE | AT_CTIME;
1332 	error = nfs3getattr(vp, &va, cr);
1333 	if (error)
1334 		return (error);
1335 	omode = va.va_mode;
1336 
1337 tryagain:
1338 	if (mask & AT_SIZE) {
1339 		args.guard.check = TRUE;
1340 		args.guard.obj_ctime.seconds = va.va_ctime.tv_sec;
1341 		args.guard.obj_ctime.nseconds = va.va_ctime.tv_nsec;
1342 	} else
1343 		args.guard.check = FALSE;
1344 
1345 	douprintf = 1;
1346 
1347 	t = gethrtime();
1348 
1349 	error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
1350 	    xdr_SETATTR3args, (caddr_t)&args,
1351 	    xdr_SETATTR3res, (caddr_t)&res, cr,
1352 	    &douprintf, &res.status, 0, NULL);
1353 
1354 	/*
1355 	 * Purge the access cache and ACL cache if changing either the
1356 	 * owner of the file, the group owner, or the mode.  These may
1357 	 * change the access permissions of the file, so purge old
1358 	 * information and start over again.
1359 	 */
1360 	if (mask & (AT_UID | AT_GID | AT_MODE)) {
1361 		(void) nfs_access_purge_rp(rp);
1362 		if (rp->r_secattr != NULL) {
1363 			mutex_enter(&rp->r_statelock);
1364 			vsp = rp->r_secattr;
1365 			rp->r_secattr = NULL;
1366 			mutex_exit(&rp->r_statelock);
1367 			if (vsp != NULL)
1368 				nfs_acl_free(vsp);
1369 		}
1370 	}
1371 
1372 	if (error) {
1373 		PURGE_ATTRCACHE(vp);
1374 		return (error);
1375 	}
1376 
1377 	error = geterrno3(res.status);
1378 	if (!error) {
1379 		/*
1380 		 * If changing the size of the file, invalidate
1381 		 * any local cached data which is no longer part
1382 		 * of the file.  We also possibly invalidate the
1383 		 * last page in the file.  We could use
1384 		 * pvn_vpzero(), but this would mark the page as
1385 		 * modified and require it to be written back to
1386 		 * the server for no particularly good reason.
1387 		 * This way, if we access it, then we bring it
1388 		 * back in.  A read should be cheaper than a
1389 		 * write.
1390 		 */
1391 		if (mask & AT_SIZE) {
1392 			nfs_invalidate_pages(vp,
1393 			    (vap->va_size & PAGEMASK), cr);
1394 		}
1395 		nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
1396 		/*
1397 		 * Some servers will change the mode to clear the setuid
1398 		 * and setgid bits when changing the uid or gid.  The
1399 		 * client needs to compensate appropriately.
1400 		 */
1401 		if (mask & (AT_UID | AT_GID)) {
1402 			int terror;
1403 
1404 			va.va_mask = AT_MODE;
1405 			terror = nfs3getattr(vp, &va, cr);
1406 			if (!terror &&
1407 			    (((mask & AT_MODE) && va.va_mode != vap->va_mode) ||
1408 			    (!(mask & AT_MODE) && va.va_mode != omode))) {
1409 				va.va_mask = AT_MODE;
1410 				if (mask & AT_MODE)
1411 					va.va_mode = vap->va_mode;
1412 				else
1413 					va.va_mode = omode;
1414 				(void) nfs3setattr(vp, &va, 0, cr);
1415 			}
1416 		}
1417 	} else {
1418 		nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
1419 		/*
1420 		 * If we got back a "not synchronized" error, then
1421 		 * we need to retry with a new guard value.  The
1422 		 * guard value used is the change time.  If the
1423 		 * server returned post_op_attr, then we can just
1424 		 * retry because we have the latest attributes.
1425 		 * Otherwise, we issue a GETATTR to get the latest
1426 		 * attributes and then retry.  If we couldn't get
1427 		 * the attributes this way either, then we give
1428 		 * up because we can't complete the operation as
1429 		 * required.
1430 		 */
1431 		if (res.status == NFS3ERR_NOT_SYNC) {
1432 			va.va_mask = AT_CTIME;
1433 			if (nfs3getattr(vp, &va, cr) == 0)
1434 				goto tryagain;
1435 		}
1436 		PURGE_STALE_FH(error, vp, cr);
1437 	}
1438 
1439 	return (error);
1440 }
1441 
1442 static int
1443 nfs3_accessx(void *vp, int mode, cred_t *cr)
1444 {
1445 	ASSERT(nfs_zone() == VTOMI((vnode_t *)vp)->mi_zone);
1446 	return (nfs3_access(vp, mode, 0, cr));
1447 }
1448 
1449 /* ARGSUSED */
1450 static int
1451 nfs3_access(vnode_t *vp, int mode, int flags, cred_t *cr)
1452 {
1453 	int error;
1454 	ACCESS3args args;
1455 	ACCESS3res res;
1456 	int douprintf;
1457 	uint32 acc;
1458 	rnode_t *rp;
1459 	cred_t *cred, *ncr, *ncrfree = NULL;
1460 	failinfo_t fi;
1461 	nfs_access_type_t cacc;
1462 	hrtime_t t;
1463 
1464 	acc = 0;
1465 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1466 		return (EIO);
1467 	if (mode & VREAD)
1468 		acc |= ACCESS3_READ;
1469 	if (mode & VWRITE) {
1470 		if (vn_is_readonly(vp) && !IS_DEVVP(vp))
1471 			return (EROFS);
1472 		if (vp->v_type == VDIR)
1473 			acc |= ACCESS3_DELETE;
1474 		acc |= ACCESS3_MODIFY | ACCESS3_EXTEND;
1475 	}
1476 	if (mode & VEXEC) {
1477 		if (vp->v_type == VDIR)
1478 			acc |= ACCESS3_LOOKUP;
1479 		else
1480 			acc |= ACCESS3_EXECUTE;
1481 	}
1482 
1483 	rp = VTOR(vp);
1484 	args.object = *VTOFH3(vp);
1485 	if (vp->v_type == VDIR) {
1486 		args.access = ACCESS3_READ | ACCESS3_DELETE | ACCESS3_MODIFY |
1487 		    ACCESS3_EXTEND | ACCESS3_LOOKUP;
1488 	} else {
1489 		args.access = ACCESS3_READ | ACCESS3_MODIFY | ACCESS3_EXTEND |
1490 		    ACCESS3_EXECUTE;
1491 	}
1492 	fi.vp = vp;
1493 	fi.fhp = (caddr_t)&args.object;
1494 	fi.copyproc = nfs3copyfh;
1495 	fi.lookupproc = nfs3lookup;
1496 	fi.xattrdirproc = acl_getxattrdir3;
1497 
1498 	cred = cr;
1499 	/*
1500 	 * ncr and ncrfree both initially
1501 	 * point to the memory area returned
1502 	 * by crnetadjust();
1503 	 * ncrfree not NULL when exiting means
1504 	 * that we need to release it
1505 	 */
1506 	ncr = crnetadjust(cred);
1507 	ncrfree = ncr;
1508 tryagain:
1509 	if (rp->r_acache != NULL) {
1510 		cacc = nfs_access_check(rp, acc, cred);
1511 		if (cacc == NFS_ACCESS_ALLOWED) {
1512 			if (ncrfree != NULL)
1513 				crfree(ncrfree);
1514 			return (0);
1515 		}
1516 		if (cacc == NFS_ACCESS_DENIED) {
1517 			/*
1518 			 * If the cred can be adjusted, try again
1519 			 * with the new cred.
1520 			 */
1521 			if (ncr != NULL) {
1522 				cred = ncr;
1523 				ncr = NULL;
1524 				goto tryagain;
1525 			}
1526 			if (ncrfree != NULL)
1527 				crfree(ncrfree);
1528 			return (EACCES);
1529 		}
1530 	}
1531 
1532 	douprintf = 1;
1533 
1534 	t = gethrtime();
1535 
1536 	error = rfs3call(VTOMI(vp), NFSPROC3_ACCESS,
1537 	    xdr_ACCESS3args, (caddr_t)&args,
1538 	    xdr_ACCESS3res, (caddr_t)&res, cred,
1539 	    &douprintf, &res.status, 0, &fi);
1540 
1541 	if (error) {
1542 		if (ncrfree != NULL)
1543 			crfree(ncrfree);
1544 		return (error);
1545 	}
1546 
1547 	error = geterrno3(res.status);
1548 	if (!error) {
1549 		nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
1550 		nfs_access_cache(rp, args.access, res.resok.access, cred);
1551 		/*
1552 		 * we just cached results with cred; if cred is the
1553 		 * adjusted credentials from crnetadjust, we do not want
1554 		 * to release them before exiting: hence setting ncrfree
1555 		 * to NULL
1556 		 */
1557 		if (cred != cr)
1558 			ncrfree = NULL;
1559 		if ((acc & res.resok.access) != acc) {
1560 			/*
1561 			 * If the cred can be adjusted, try again
1562 			 * with the new cred.
1563 			 */
1564 			if (ncr != NULL) {
1565 				cred = ncr;
1566 				ncr = NULL;
1567 				goto tryagain;
1568 			}
1569 			error = EACCES;
1570 		}
1571 	} else {
1572 		nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
1573 		PURGE_STALE_FH(error, vp, cr);
1574 	}
1575 
1576 	if (ncrfree != NULL)
1577 		crfree(ncrfree);
1578 
1579 	return (error);
1580 }
1581 
1582 static int nfs3_do_symlink_cache = 1;
1583 
1584 static int
1585 nfs3_readlink(vnode_t *vp, struct uio *uiop, cred_t *cr)
1586 {
1587 	int error;
1588 	READLINK3args args;
1589 	READLINK3res res;
1590 	nfspath3 resdata_backup;
1591 	rnode_t *rp;
1592 	int douprintf;
1593 	int len;
1594 	failinfo_t fi;
1595 	hrtime_t t;
1596 
1597 	/*
1598 	 * Can't readlink anything other than a symbolic link.
1599 	 */
1600 	if (vp->v_type != VLNK)
1601 		return (EINVAL);
1602 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1603 		return (EIO);
1604 
1605 	rp = VTOR(vp);
1606 	if (nfs3_do_symlink_cache && rp->r_symlink.contents != NULL) {
1607 		error = nfs3_validate_caches(vp, cr);
1608 		if (error)
1609 			return (error);
1610 		mutex_enter(&rp->r_statelock);
1611 		if (rp->r_symlink.contents != NULL) {
1612 			error = uiomove(rp->r_symlink.contents,
1613 			    rp->r_symlink.len, UIO_READ, uiop);
1614 			mutex_exit(&rp->r_statelock);
1615 			return (error);
1616 		}
1617 		mutex_exit(&rp->r_statelock);
1618 	}
1619 
1620 	args.symlink = *VTOFH3(vp);
1621 	fi.vp = vp;
1622 	fi.fhp = (caddr_t)&args.symlink;
1623 	fi.copyproc = nfs3copyfh;
1624 	fi.lookupproc = nfs3lookup;
1625 	fi.xattrdirproc = acl_getxattrdir3;
1626 
1627 	res.resok.data = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1628 
1629 	resdata_backup = res.resok.data;
1630 
1631 	douprintf = 1;
1632 
1633 	t = gethrtime();
1634 
1635 	error = rfs3call(VTOMI(vp), NFSPROC3_READLINK,
1636 	    xdr_nfs_fh3, (caddr_t)&args,
1637 	    xdr_READLINK3res, (caddr_t)&res, cr,
1638 	    &douprintf, &res.status, 0, &fi);
1639 
1640 	if (res.resok.data == nfs3nametoolong)
1641 		error = EINVAL;
1642 
1643 	if (error) {
1644 		kmem_free(resdata_backup, MAXPATHLEN);
1645 		return (error);
1646 	}
1647 
1648 	error = geterrno3(res.status);
1649 	if (!error) {
1650 		nfs3_cache_post_op_attr(vp, &res.resok.symlink_attributes, t,
1651 		    cr);
1652 		len = strlen(res.resok.data);
1653 		error = uiomove(res.resok.data, len, UIO_READ, uiop);
1654 		if (nfs3_do_symlink_cache && rp->r_symlink.contents == NULL) {
1655 			mutex_enter(&rp->r_statelock);
1656 				if (rp->r_symlink.contents == NULL) {
1657 				rp->r_symlink.contents = res.resok.data;
1658 				rp->r_symlink.len = len;
1659 				rp->r_symlink.size = MAXPATHLEN;
1660 				mutex_exit(&rp->r_statelock);
1661 			} else {
1662 				mutex_exit(&rp->r_statelock);
1663 
1664 				kmem_free((void *)res.resok.data, MAXPATHLEN);
1665 			}
1666 		} else {
1667 			kmem_free((void *)res.resok.data, MAXPATHLEN);
1668 		}
1669 	} else {
1670 		nfs3_cache_post_op_attr(vp,
1671 		    &res.resfail.symlink_attributes, t, cr);
1672 		PURGE_STALE_FH(error, vp, cr);
1673 
1674 		kmem_free((void *)res.resok.data, MAXPATHLEN);
1675 
1676 	}
1677 
1678 	/*
1679 	 * The over the wire error for attempting to readlink something
1680 	 * other than a symbolic link is ENXIO.  However, we need to
1681 	 * return EINVAL instead of ENXIO, so we map it here.
1682 	 */
1683 	return (error == ENXIO ? EINVAL : error);
1684 }
1685 
1686 /*
1687  * Flush local dirty pages to stable storage on the server.
1688  *
1689  * If FNODSYNC is specified, then there is nothing to do because
1690  * metadata changes are not cached on the client before being
1691  * sent to the server.
1692  */
1693 static int
1694 nfs3_fsync(vnode_t *vp, int syncflag, cred_t *cr)
1695 {
1696 	int error;
1697 
1698 	if ((syncflag & FNODSYNC) || IS_SWAPVP(vp))
1699 		return (0);
1700 	if (nfs_zone() != VTOMI(vp)->mi_zone)
1701 		return (EIO);
1702 
1703 	error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
1704 	if (!error)
1705 		error = VTOR(vp)->r_error;
1706 	return (error);
1707 }
1708 
1709 /*
1710  * Weirdness: if the file was removed or the target of a rename
1711  * operation while it was open, it got renamed instead.  Here we
1712  * remove the renamed file.
1713  */
1714 static void
1715 nfs3_inactive(vnode_t *vp, cred_t *cr)
1716 {
1717 	rnode_t *rp;
1718 
1719 	ASSERT(vp != DNLC_NO_VNODE);
1720 
1721 	/*
1722 	 * If this is coming from the wrong zone, we let someone in the right
1723 	 * zone take care of it asynchronously.  We can get here due to
1724 	 * VN_RELE() being called from pageout() or fsflush().  This call may
1725 	 * potentially turn into an expensive no-op if, for instance, v_count
1726 	 * gets incremented in the meantime, but it's still correct.
1727 	 */
1728 	if (nfs_zone() != VTOMI(vp)->mi_zone) {
1729 		nfs_async_inactive(vp, cr, nfs3_inactive);
1730 		return;
1731 	}
1732 
1733 	rp = VTOR(vp);
1734 redo:
1735 	if (rp->r_unldvp != NULL) {
1736 		/*
1737 		 * Save the vnode pointer for the directory where the
1738 		 * unlinked-open file got renamed, then set it to NULL
1739 		 * to prevent another thread from getting here before
1740 		 * we're done with the remove.  While we have the
1741 		 * statelock, make local copies of the pertinent rnode
1742 		 * fields.  If we weren't to do this in an atomic way, the
1743 		 * the unl* fields could become inconsistent with respect
1744 		 * to each other due to a race condition between this
1745 		 * code and nfs_remove().  See bug report 1034328.
1746 		 */
1747 		mutex_enter(&rp->r_statelock);
1748 		if (rp->r_unldvp != NULL) {
1749 			vnode_t *unldvp;
1750 			char *unlname;
1751 			cred_t *unlcred;
1752 			REMOVE3args args;
1753 			REMOVE3res res;
1754 			int douprintf;
1755 			int error;
1756 			hrtime_t t;
1757 
1758 			unldvp = rp->r_unldvp;
1759 			rp->r_unldvp = NULL;
1760 			unlname = rp->r_unlname;
1761 			rp->r_unlname = NULL;
1762 			unlcred = rp->r_unlcred;
1763 			rp->r_unlcred = NULL;
1764 			mutex_exit(&rp->r_statelock);
1765 
1766 			/*
1767 			 * If there are any dirty pages left, then flush
1768 			 * them.  This is unfortunate because they just
1769 			 * may get thrown away during the remove operation,
1770 			 * but we have to do this for correctness.
1771 			 */
1772 			if (vn_has_cached_data(vp) &&
1773 			    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
1774 				ASSERT(vp->v_type != VCHR);
1775 				error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1776 				if (error) {
1777 					mutex_enter(&rp->r_statelock);
1778 					if (!rp->r_error)
1779 						rp->r_error = error;
1780 					mutex_exit(&rp->r_statelock);
1781 				}
1782 			}
1783 
1784 			/*
1785 			 * Do the remove operation on the renamed file
1786 			 */
1787 			setdiropargs3(&args.object, unlname, unldvp);
1788 
1789 			douprintf = 1;
1790 
1791 			t = gethrtime();
1792 
1793 			error = rfs3call(VTOMI(unldvp), NFSPROC3_REMOVE,
1794 			    xdr_diropargs3, (caddr_t)&args,
1795 			    xdr_REMOVE3res, (caddr_t)&res, unlcred,
1796 			    &douprintf, &res.status, 0, NULL);
1797 
1798 			if (error) {
1799 				PURGE_ATTRCACHE(unldvp);
1800 			} else {
1801 				error = geterrno3(res.status);
1802 				if (!error) {
1803 					nfs3_cache_wcc_data(unldvp,
1804 					    &res.resok.dir_wcc, t, cr);
1805 					if (HAVE_RDDIR_CACHE(VTOR(unldvp)))
1806 						nfs_purge_rddir_cache(unldvp);
1807 				} else {
1808 					nfs3_cache_wcc_data(unldvp,
1809 					    &res.resfail.dir_wcc, t, cr);
1810 					PURGE_STALE_FH(error, unldvp, cr);
1811 				}
1812 			}
1813 
1814 			/*
1815 			 * Release stuff held for the remove
1816 			 */
1817 			VN_RELE(unldvp);
1818 			kmem_free(unlname, MAXNAMELEN);
1819 			crfree(unlcred);
1820 			goto redo;
1821 		}
1822 		mutex_exit(&rp->r_statelock);
1823 	}
1824 
1825 	rp_addfree(rp, cr);
1826 }
1827 
1828 /*
1829  * Remote file system operations having to do with directory manipulation.
1830  */
1831 
1832 static int
1833 nfs3_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1834 	int flags, vnode_t *rdir, cred_t *cr)
1835 {
1836 	int error;
1837 	vnode_t *vp;
1838 	vnode_t *avp = NULL;
1839 	rnode_t *drp;
1840 
1841 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
1842 		return (EPERM);
1843 
1844 	drp = VTOR(dvp);
1845 
1846 	/*
1847 	 * Are we looking up extended attributes?  If so, "dvp" is
1848 	 * the file or directory for which we want attributes, and
1849 	 * we need a lookup of the hidden attribute directory
1850 	 * before we lookup the rest of the path.
1851 	 */
1852 	if (flags & LOOKUP_XATTR) {
1853 		bool_t cflag = ((flags & CREATE_XATTR_DIR) != 0);
1854 		mntinfo_t *mi;
1855 
1856 		mi = VTOMI(dvp);
1857 		if (!(mi->mi_flags & MI_EXTATTR))
1858 			return (EINVAL);
1859 
1860 		if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp)))
1861 			return (EINTR);
1862 
1863 		(void) nfs3lookup_dnlc(dvp, XATTR_DIR_NAME, &avp, cr);
1864 		if (avp == NULL)
1865 			error = acl_getxattrdir3(dvp, &avp, cflag, cr, 0);
1866 		else
1867 			error = 0;
1868 
1869 		nfs_rw_exit(&drp->r_rwlock);
1870 
1871 		if (error) {
1872 			if (mi->mi_flags & MI_EXTATTR)
1873 				return (error);
1874 			return (EINVAL);
1875 		}
1876 		dvp = avp;
1877 		drp = VTOR(dvp);
1878 	}
1879 
1880 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp))) {
1881 		error = EINTR;
1882 		goto out;
1883 	}
1884 
1885 	error = nfs3lookup(dvp, nm, vpp, pnp, flags, rdir, cr, 0);
1886 
1887 	nfs_rw_exit(&drp->r_rwlock);
1888 
1889 	/*
1890 	 * If vnode is a device, create special vnode.
1891 	 */
1892 	if (!error && IS_DEVVP(*vpp)) {
1893 		vp = *vpp;
1894 		*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
1895 		VN_RELE(vp);
1896 	}
1897 
1898 out:
1899 	if (avp != NULL)
1900 		VN_RELE(avp);
1901 
1902 	return (error);
1903 }
1904 
1905 static int nfs3_lookup_neg_cache = 1;
1906 
1907 #ifdef DEBUG
1908 static int nfs3_lookup_dnlc_hits = 0;
1909 static int nfs3_lookup_dnlc_misses = 0;
1910 static int nfs3_lookup_dnlc_neg_hits = 0;
1911 static int nfs3_lookup_dnlc_disappears = 0;
1912 static int nfs3_lookup_dnlc_lookups = 0;
1913 #endif
1914 
1915 /* ARGSUSED */
1916 int
1917 nfs3lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1918 	int flags, vnode_t *rdir, cred_t *cr, int rfscall_flags)
1919 {
1920 	int error;
1921 	rnode_t *drp;
1922 
1923 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1924 	/*
1925 	 * If lookup is for "", just return dvp.  Don't need
1926 	 * to send it over the wire, look it up in the dnlc,
1927 	 * or perform any access checks.
1928 	 */
1929 	if (*nm == '\0') {
1930 		VN_HOLD(dvp);
1931 		*vpp = dvp;
1932 		return (0);
1933 	}
1934 
1935 	/*
1936 	 * Can't do lookups in non-directories.
1937 	 */
1938 	if (dvp->v_type != VDIR)
1939 		return (ENOTDIR);
1940 
1941 	/*
1942 	 * If we're called with RFSCALL_SOFT, it's important that
1943 	 * the only rfscall is one we make directly; if we permit
1944 	 * an access call because we're looking up "." or validating
1945 	 * a dnlc hit, we'll deadlock because that rfscall will not
1946 	 * have the RFSCALL_SOFT set.
1947 	 */
1948 	if (rfscall_flags & RFSCALL_SOFT)
1949 		goto callit;
1950 
1951 	/*
1952 	 * If lookup is for ".", just return dvp.  Don't need
1953 	 * to send it over the wire or look it up in the dnlc,
1954 	 * just need to check access.
1955 	 */
1956 	if (strcmp(nm, ".") == 0) {
1957 		error = nfs3_access(dvp, VEXEC, 0, cr);
1958 		if (error)
1959 			return (error);
1960 		VN_HOLD(dvp);
1961 		*vpp = dvp;
1962 		return (0);
1963 	}
1964 
1965 	drp = VTOR(dvp);
1966 	if (!(drp->r_flags & RLOOKUP)) {
1967 		mutex_enter(&drp->r_statelock);
1968 		drp->r_flags |= RLOOKUP;
1969 		mutex_exit(&drp->r_statelock);
1970 	}
1971 
1972 	/*
1973 	 * Lookup this name in the DNLC.  If there was a valid entry,
1974 	 * then return the results of the lookup.
1975 	 */
1976 	error = nfs3lookup_dnlc(dvp, nm, vpp, cr);
1977 	if (error || *vpp != NULL)
1978 		return (error);
1979 
1980 callit:
1981 	error = nfs3lookup_otw(dvp, nm, vpp, cr, rfscall_flags);
1982 
1983 	return (error);
1984 }
1985 
1986 static int
1987 nfs3lookup_dnlc(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr)
1988 {
1989 	int error;
1990 	vnode_t *vp;
1991 
1992 	ASSERT(*nm != '\0');
1993 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
1994 	/*
1995 	 * Lookup this name in the DNLC.  If successful, then validate
1996 	 * the caches and then recheck the DNLC.  The DNLC is rechecked
1997 	 * just in case this entry got invalidated during the call
1998 	 * to nfs3_validate_caches.
1999 	 *
2000 	 * An assumption is being made that it is safe to say that a
2001 	 * file exists which may not on the server.  Any operations to
2002 	 * the server will fail with ESTALE.
2003 	 */
2004 #ifdef DEBUG
2005 	nfs3_lookup_dnlc_lookups++;
2006 #endif
2007 	vp = dnlc_lookup(dvp, nm);
2008 	if (vp != NULL) {
2009 		VN_RELE(vp);
2010 		if (vp == DNLC_NO_VNODE && !vn_is_readonly(dvp)) {
2011 			PURGE_ATTRCACHE(dvp);
2012 		}
2013 		error = nfs3_validate_caches(dvp, cr);
2014 		if (error)
2015 			return (error);
2016 		vp = dnlc_lookup(dvp, nm);
2017 		if (vp != NULL) {
2018 			error = nfs3_access(dvp, VEXEC, 0, cr);
2019 			if (error) {
2020 				VN_RELE(vp);
2021 				return (error);
2022 			}
2023 			if (vp == DNLC_NO_VNODE) {
2024 				VN_RELE(vp);
2025 #ifdef DEBUG
2026 				nfs3_lookup_dnlc_neg_hits++;
2027 #endif
2028 				return (ENOENT);
2029 			}
2030 			*vpp = vp;
2031 #ifdef DEBUG
2032 			nfs3_lookup_dnlc_hits++;
2033 #endif
2034 			return (0);
2035 		}
2036 #ifdef DEBUG
2037 		nfs3_lookup_dnlc_disappears++;
2038 #endif
2039 	}
2040 #ifdef DEBUG
2041 	else
2042 		nfs3_lookup_dnlc_misses++;
2043 #endif
2044 
2045 	*vpp = NULL;
2046 
2047 	return (0);
2048 }
2049 
2050 static int
2051 nfs3lookup_otw(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr,
2052 	int rfscall_flags)
2053 {
2054 	int error;
2055 	LOOKUP3args args;
2056 	LOOKUP3vres res;
2057 	int douprintf;
2058 	struct vattr vattr;
2059 	struct vattr dvattr;
2060 	vnode_t *vp;
2061 	failinfo_t fi;
2062 	hrtime_t t;
2063 
2064 	ASSERT(*nm != '\0');
2065 	ASSERT(dvp->v_type == VDIR);
2066 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2067 
2068 	setdiropargs3(&args.what, nm, dvp);
2069 
2070 	fi.vp = dvp;
2071 	fi.fhp = (caddr_t)&args.what.dir;
2072 	fi.copyproc = nfs3copyfh;
2073 	fi.lookupproc = nfs3lookup;
2074 	fi.xattrdirproc = acl_getxattrdir3;
2075 	res.obj_attributes.fres.vp = dvp;
2076 	res.obj_attributes.fres.vap = &vattr;
2077 	res.dir_attributes.fres.vp = dvp;
2078 	res.dir_attributes.fres.vap = &dvattr;
2079 
2080 	douprintf = 1;
2081 
2082 	t = gethrtime();
2083 
2084 	error = rfs3call(VTOMI(dvp), NFSPROC3_LOOKUP,
2085 	    xdr_diropargs3, (caddr_t)&args,
2086 	    xdr_LOOKUP3vres, (caddr_t)&res, cr,
2087 	    &douprintf, &res.status, rfscall_flags, &fi);
2088 
2089 	if (error)
2090 		return (error);
2091 
2092 	nfs3_cache_post_op_vattr(dvp, &res.dir_attributes, t, cr);
2093 
2094 	error = geterrno3(res.status);
2095 	if (error) {
2096 		PURGE_STALE_FH(error, dvp, cr);
2097 		if (error == ENOENT && nfs3_lookup_neg_cache)
2098 			dnlc_enter(dvp, nm, DNLC_NO_VNODE);
2099 		return (error);
2100 	}
2101 
2102 	if (res.obj_attributes.attributes) {
2103 		vp = makenfs3node_va(&res.object, res.obj_attributes.fres.vap,
2104 				dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2105 	} else {
2106 		vp = makenfs3node_va(&res.object, NULL,
2107 				dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2108 		if (vp->v_type == VNON) {
2109 			vattr.va_mask = AT_TYPE;
2110 			error = nfs3getattr(vp, &vattr, cr);
2111 			if (error) {
2112 				VN_RELE(vp);
2113 				return (error);
2114 			}
2115 			vp->v_type = vattr.va_type;
2116 		}
2117 	}
2118 
2119 	if (!(rfscall_flags & RFSCALL_SOFT))
2120 		dnlc_update(dvp, nm, vp);
2121 
2122 	*vpp = vp;
2123 
2124 	return (error);
2125 }
2126 
2127 #ifdef DEBUG
2128 static int nfs3_create_misses = 0;
2129 #endif
2130 
2131 /* ARGSUSED */
2132 static int
2133 nfs3_create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2134 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2135 {
2136 	int error;
2137 	vnode_t *vp;
2138 	rnode_t *rp;
2139 	struct vattr vattr;
2140 	rnode_t *drp;
2141 	vnode_t *tempvp;
2142 
2143 	drp = VTOR(dvp);
2144 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2145 		return (EPERM);
2146 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2147 		return (EINTR);
2148 
2149 top:
2150 	/*
2151 	 * We make a copy of the attributes because the caller does not
2152 	 * expect us to change what va points to.
2153 	 */
2154 	vattr = *va;
2155 
2156 	/*
2157 	 * If the pathname is "", just use dvp.  Don't need
2158 	 * to send it over the wire, look it up in the dnlc,
2159 	 * or perform any access checks.
2160 	 */
2161 	if (*nm == '\0') {
2162 		error = 0;
2163 		VN_HOLD(dvp);
2164 		vp = dvp;
2165 	/*
2166 	 * If the pathname is ".", just use dvp.  Don't need
2167 	 * to send it over the wire or look it up in the dnlc,
2168 	 * just need to check access.
2169 	 */
2170 	} else if (strcmp(nm, ".") == 0) {
2171 		error = nfs3_access(dvp, VEXEC, 0, cr);
2172 		if (error) {
2173 			nfs_rw_exit(&drp->r_rwlock);
2174 			return (error);
2175 		}
2176 		VN_HOLD(dvp);
2177 		vp = dvp;
2178 	/*
2179 	 * We need to go over the wire, just to be sure whether the
2180 	 * file exists or not.  Using the DNLC can be dangerous in
2181 	 * this case when making a decision regarding existence.
2182 	 */
2183 	} else {
2184 		error = nfs3lookup_otw(dvp, nm, &vp, cr, 0);
2185 	}
2186 	if (!error) {
2187 		if (exclusive == EXCL)
2188 			error = EEXIST;
2189 		else if (vp->v_type == VDIR && (mode & VWRITE))
2190 			error = EISDIR;
2191 		else {
2192 			/*
2193 			 * If vnode is a device, create special vnode.
2194 			 */
2195 			if (IS_DEVVP(vp)) {
2196 				tempvp = vp;
2197 				vp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2198 				VN_RELE(tempvp);
2199 			}
2200 			if (!(error = VOP_ACCESS(vp, mode, 0, cr))) {
2201 				if ((vattr.va_mask & AT_SIZE) &&
2202 				    vp->v_type == VREG) {
2203 					rp = VTOR(vp);
2204 					/*
2205 					 * Check here for large file handled
2206 					 * by LF-unaware process (as
2207 					 * ufs_create() does)
2208 					 */
2209 					if (!(lfaware & FOFFMAX)) {
2210 						mutex_enter(&rp->r_statelock);
2211 						if (rp->r_size > MAXOFF32_T)
2212 							error = EOVERFLOW;
2213 						mutex_exit(&rp->r_statelock);
2214 					}
2215 					if (!error) {
2216 						vattr.va_mask = AT_SIZE;
2217 						error = nfs3setattr(vp,
2218 						    &vattr, 0, cr);
2219 					}
2220 				}
2221 			}
2222 		}
2223 		nfs_rw_exit(&drp->r_rwlock);
2224 		if (error) {
2225 			VN_RELE(vp);
2226 		} else
2227 			*vpp = vp;
2228 		return (error);
2229 	}
2230 
2231 	dnlc_remove(dvp, nm);
2232 
2233 	/*
2234 	 * Decide what the group-id of the created file should be.
2235 	 * Set it in attribute list as advisory...
2236 	 */
2237 	error = setdirgid(dvp, &vattr.va_gid, cr);
2238 	if (error) {
2239 		nfs_rw_exit(&drp->r_rwlock);
2240 		return (error);
2241 	}
2242 	vattr.va_mask |= AT_GID;
2243 
2244 	ASSERT(vattr.va_mask & AT_TYPE);
2245 	if (vattr.va_type == VREG) {
2246 		ASSERT(vattr.va_mask & AT_MODE);
2247 		if (MANDMODE(vattr.va_mode)) {
2248 			nfs_rw_exit(&drp->r_rwlock);
2249 			return (EACCES);
2250 		}
2251 		error = nfs3create(dvp, nm, &vattr, exclusive, mode, vpp, cr,
2252 		    lfaware);
2253 		/*
2254 		 * If this is not an exclusive create, then the CREATE
2255 		 * request will be made with the GUARDED mode set.  This
2256 		 * means that the server will return EEXIST if the file
2257 		 * exists.  The file could exist because of a retransmitted
2258 		 * request.  In this case, we recover by starting over and
2259 		 * checking to see whether the file exists.  This second
2260 		 * time through it should and a CREATE request will not be
2261 		 * sent.
2262 		 *
2263 		 * This handles the problem of a dangling CREATE request
2264 		 * which contains attributes which indicate that the file
2265 		 * should be truncated.  This retransmitted request could
2266 		 * possibly truncate valid data in the file if not caught
2267 		 * by the duplicate request mechanism on the server or if
2268 		 * not caught by other means.  The scenario is:
2269 		 *
2270 		 * Client transmits CREATE request with size = 0
2271 		 * Client times out, retransmits request.
2272 		 * Response to the first request arrives from the server
2273 		 *  and the client proceeds on.
2274 		 * Client writes data to the file.
2275 		 * The server now processes retransmitted CREATE request
2276 		 *  and truncates file.
2277 		 *
2278 		 * The use of the GUARDED CREATE request prevents this from
2279 		 * happening because the retransmitted CREATE would fail
2280 		 * with EEXIST and would not truncate the file.
2281 		 */
2282 		if (error == EEXIST && exclusive == NONEXCL) {
2283 #ifdef DEBUG
2284 			nfs3_create_misses++;
2285 #endif
2286 			goto top;
2287 		}
2288 		nfs_rw_exit(&drp->r_rwlock);
2289 		return (error);
2290 	}
2291 	error = nfs3mknod(dvp, nm, &vattr, exclusive, mode, vpp, cr);
2292 	nfs_rw_exit(&drp->r_rwlock);
2293 	return (error);
2294 }
2295 
2296 /* ARGSUSED */
2297 static int
2298 nfs3create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2299 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2300 {
2301 	int error;
2302 	CREATE3args args;
2303 	CREATE3res res;
2304 	int douprintf;
2305 	vnode_t *vp;
2306 	struct vattr vattr;
2307 	nfstime3 *verfp;
2308 	rnode_t *rp;
2309 	timestruc_t now;
2310 	hrtime_t t;
2311 
2312 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2313 	setdiropargs3(&args.where, nm, dvp);
2314 	if (exclusive == EXCL) {
2315 		args.how.mode = EXCLUSIVE;
2316 		/*
2317 		 * Construct the create verifier.  This verifier needs
2318 		 * to be unique between different clients.  It also needs
2319 		 * to vary for each exclusive create request generated
2320 		 * from the client to the server.
2321 		 *
2322 		 * The first attempt is made to use the hostid and a
2323 		 * unique number on the client.  If the hostid has not
2324 		 * been set, the high resolution time that the exclusive
2325 		 * create request is being made is used.  This will work
2326 		 * unless two different clients, both with the hostid
2327 		 * not set, attempt an exclusive create request on the
2328 		 * same file, at exactly the same clock time.  The
2329 		 * chances of this happening seem small enough to be
2330 		 * reasonable.
2331 		 */
2332 		verfp = (nfstime3 *)&args.how.createhow3_u.verf;
2333 		verfp->seconds = nfs_atoi(hw_serial);
2334 		if (verfp->seconds != 0)
2335 			verfp->nseconds = newnum();
2336 		else {
2337 			gethrestime(&now);
2338 			verfp->seconds = now.tv_sec;
2339 			verfp->nseconds = now.tv_nsec;
2340 		}
2341 		/*
2342 		 * Since the server will use this value for the mtime,
2343 		 * make sure that it can't overflow. Zero out the MSB.
2344 		 * The actual value does not matter here, only its uniqeness.
2345 		 */
2346 		verfp->seconds %= INT32_MAX;
2347 	} else {
2348 		/*
2349 		 * Issue the non-exclusive create in guarded mode.  This
2350 		 * may result in some false EEXIST responses for
2351 		 * retransmitted requests, but these will be handled at
2352 		 * a higher level.  By using GUARDED, duplicate requests
2353 		 * to do file truncation and possible access problems
2354 		 * can be avoided.
2355 		 */
2356 		args.how.mode = GUARDED;
2357 		error = vattr_to_sattr3(va,
2358 				&args.how.createhow3_u.obj_attributes);
2359 		if (error) {
2360 			/* req time field(s) overflow - return immediately */
2361 			return (error);
2362 		}
2363 	}
2364 
2365 	douprintf = 1;
2366 
2367 	t = gethrtime();
2368 
2369 	error = rfs3call(VTOMI(dvp), NFSPROC3_CREATE,
2370 	    xdr_CREATE3args, (caddr_t)&args,
2371 	    xdr_CREATE3res, (caddr_t)&res, cr,
2372 	    &douprintf, &res.status, 0, NULL);
2373 
2374 	if (error) {
2375 		PURGE_ATTRCACHE(dvp);
2376 		return (error);
2377 	}
2378 
2379 	error = geterrno3(res.status);
2380 	if (!error) {
2381 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2382 		if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2383 			nfs_purge_rddir_cache(dvp);
2384 
2385 		/*
2386 		 * On exclusive create the times need to be explicitly
2387 		 * set to clear any potential verifier that may be stored
2388 		 * in one of these fields (see comment below).  This
2389 		 * is done here to cover the case where no post op attrs
2390 		 * were returned or a 'invalid' time was returned in
2391 		 * the attributes.
2392 		 */
2393 		if (exclusive == EXCL)
2394 			va->va_mask |= (AT_MTIME | AT_ATIME);
2395 
2396 		if (!res.resok.obj.handle_follows) {
2397 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2398 			if (error)
2399 				return (error);
2400 		} else {
2401 			if (res.resok.obj_attributes.attributes) {
2402 				vp = makenfs3node(&res.resok.obj.handle,
2403 				    &res.resok.obj_attributes.attr,
2404 				    dvp->v_vfsp, t, cr, NULL, NULL);
2405 			} else {
2406 				vp = makenfs3node(&res.resok.obj.handle, NULL,
2407 				    dvp->v_vfsp, t, cr, NULL, NULL);
2408 
2409 				/*
2410 				 * On an exclusive create, it is possible
2411 				 * that attributes were returned but those
2412 				 * postop attributes failed to decode
2413 				 * properly.  If this is the case,
2414 				 * then most likely the atime or mtime
2415 				 * were invalid for our client; this
2416 				 * is caused by the server storing the
2417 				 * create verifier in one of the time
2418 				 * fields(most likely mtime).
2419 				 * So... we are going to setattr just the
2420 				 * atime/mtime to clear things up.
2421 				 */
2422 				if (exclusive == EXCL) {
2423 					if (error =
2424 						nfs3excl_create_settimes(vp,
2425 							va, cr)) {
2426 						/*
2427 						 * Setting the times failed.
2428 						 * Remove the file and return
2429 						 * the error.
2430 						 */
2431 						VN_RELE(vp);
2432 						(void) nfs3_remove(dvp,
2433 							nm, cr);
2434 						return (error);
2435 					}
2436 				}
2437 
2438 				/*
2439 				 * This handles the non-exclusive case
2440 				 * and the exclusive case where no post op
2441 				 * attrs were returned.
2442 				 */
2443 				if (vp->v_type == VNON) {
2444 					vattr.va_mask = AT_TYPE;
2445 					error = nfs3getattr(vp, &vattr, cr);
2446 					if (error) {
2447 						VN_RELE(vp);
2448 						return (error);
2449 					}
2450 					vp->v_type = vattr.va_type;
2451 				}
2452 			}
2453 			dnlc_update(dvp, nm, vp);
2454 		}
2455 
2456 		rp = VTOR(vp);
2457 
2458 		/*
2459 		 * Check here for large file handled by
2460 		 * LF-unaware process (as ufs_create() does)
2461 		 */
2462 		if ((va->va_mask & AT_SIZE) && vp->v_type == VREG &&
2463 		    !(lfaware & FOFFMAX)) {
2464 			mutex_enter(&rp->r_statelock);
2465 			if (rp->r_size > MAXOFF32_T) {
2466 				mutex_exit(&rp->r_statelock);
2467 				VN_RELE(vp);
2468 				return (EOVERFLOW);
2469 			}
2470 			mutex_exit(&rp->r_statelock);
2471 		}
2472 
2473 		if (exclusive == EXCL &&
2474 			(va->va_mask & ~(AT_GID | AT_SIZE))) {
2475 			/*
2476 			 * If doing an exclusive create, then generate
2477 			 * a SETATTR to set the initial attributes.
2478 			 * Try to set the mtime and the atime to the
2479 			 * server's current time.  It is somewhat
2480 			 * expected that these fields will be used to
2481 			 * store the exclusive create cookie.  If not,
2482 			 * server implementors will need to know that
2483 			 * a SETATTR will follow an exclusive create
2484 			 * and the cookie should be destroyed if
2485 			 * appropriate. This work may have been done
2486 			 * earlier in this function if post op attrs
2487 			 * were not available.
2488 			 *
2489 			 * The AT_GID and AT_SIZE bits are turned off
2490 			 * so that the SETATTR request will not attempt
2491 			 * to process these.  The gid will be set
2492 			 * separately if appropriate.  The size is turned
2493 			 * off because it is assumed that a new file will
2494 			 * be created empty and if the file wasn't empty,
2495 			 * then the exclusive create will have failed
2496 			 * because the file must have existed already.
2497 			 * Therefore, no truncate operation is needed.
2498 			 */
2499 			va->va_mask &= ~(AT_GID | AT_SIZE);
2500 			error = nfs3setattr(vp, va, 0, cr);
2501 			if (error) {
2502 				/*
2503 				 * Couldn't correct the attributes of
2504 				 * the newly created file and the
2505 				 * attributes are wrong.  Remove the
2506 				 * file and return an error to the
2507 				 * application.
2508 				 */
2509 				VN_RELE(vp);
2510 				(void) nfs3_remove(dvp, nm, cr);
2511 				return (error);
2512 			}
2513 		}
2514 
2515 		if (va->va_gid != rp->r_attr.va_gid) {
2516 			/*
2517 			 * If the gid on the file isn't right, then
2518 			 * generate a SETATTR to attempt to change
2519 			 * it.  This may or may not work, depending
2520 			 * upon the server's semantics for allowing
2521 			 * file ownership changes.
2522 			 */
2523 			va->va_mask = AT_GID;
2524 			(void) nfs3setattr(vp, va, 0, cr);
2525 		}
2526 
2527 		/*
2528 		 * If vnode is a device create special vnode
2529 		 */
2530 		if (IS_DEVVP(vp)) {
2531 			*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2532 			VN_RELE(vp);
2533 		} else
2534 			*vpp = vp;
2535 	} else {
2536 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2537 		PURGE_STALE_FH(error, dvp, cr);
2538 	}
2539 
2540 	return (error);
2541 }
2542 
2543 /*
2544  * Special setattr function to take care of rest of atime/mtime
2545  * after successful exclusive create.  This function exists to avoid
2546  * handling attributes from the server; exclusive the atime/mtime fields
2547  * may be 'invalid' in client's view and therefore can not be trusted.
2548  */
2549 static int
2550 nfs3excl_create_settimes(vnode_t *vp, struct vattr *vap, cred_t *cr)
2551 {
2552 	int error;
2553 	uint_t mask;
2554 	SETATTR3args args;
2555 	SETATTR3res res;
2556 	int douprintf;
2557 	rnode_t *rp;
2558 	hrtime_t t;
2559 
2560 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
2561 	/* save the caller's mask so that it can be reset later */
2562 	mask = vap->va_mask;
2563 
2564 	rp = VTOR(vp);
2565 
2566 	args.object = *RTOFH3(rp);
2567 	args.guard.check = FALSE;
2568 
2569 	/* Use the mask to initialize the arguments */
2570 	vap->va_mask = 0;
2571 	error = vattr_to_sattr3(vap, &args.new_attributes);
2572 
2573 	/* We want to set just atime/mtime on this request */
2574 	args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
2575 	args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
2576 
2577 	douprintf = 1;
2578 
2579 	t = gethrtime();
2580 
2581 	error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
2582 	    xdr_SETATTR3args, (caddr_t)&args,
2583 	    xdr_SETATTR3res, (caddr_t)&res, cr,
2584 	    &douprintf, &res.status, 0, NULL);
2585 
2586 	if (error) {
2587 		vap->va_mask = mask;
2588 		return (error);
2589 	}
2590 
2591 	error = geterrno3(res.status);
2592 	if (!error) {
2593 		/*
2594 		 * It is important to pick up the attributes.
2595 		 * Since this is the exclusive create path, the
2596 		 * attributes on the initial create were ignored
2597 		 * and we need these to have the correct info.
2598 		 */
2599 		nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
2600 		/*
2601 		 * No need to do the atime/mtime work again so clear
2602 		 * the bits.
2603 		 */
2604 		mask &= ~(AT_ATIME | AT_MTIME);
2605 	} else {
2606 		nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
2607 	}
2608 
2609 	vap->va_mask = mask;
2610 
2611 	return (error);
2612 }
2613 
2614 /* ARGSUSED */
2615 static int
2616 nfs3mknod(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2617 	int mode, vnode_t **vpp, cred_t *cr)
2618 {
2619 	int error;
2620 	MKNOD3args args;
2621 	MKNOD3res res;
2622 	int douprintf;
2623 	vnode_t *vp;
2624 	struct vattr vattr;
2625 	hrtime_t t;
2626 
2627 	ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2628 	switch (va->va_type) {
2629 	case VCHR:
2630 	case VBLK:
2631 		setdiropargs3(&args.where, nm, dvp);
2632 		args.what.type = (va->va_type == VCHR) ? NF3CHR : NF3BLK;
2633 		error = vattr_to_sattr3(va,
2634 		    &args.what.mknoddata3_u.device.dev_attributes);
2635 		if (error) {
2636 			/* req time field(s) overflow - return immediately */
2637 			return (error);
2638 		}
2639 		args.what.mknoddata3_u.device.spec.specdata1 =
2640 		    getmajor(va->va_rdev);
2641 		args.what.mknoddata3_u.device.spec.specdata2 =
2642 		    getminor(va->va_rdev);
2643 		break;
2644 
2645 	case VFIFO:
2646 	case VSOCK:
2647 		setdiropargs3(&args.where, nm, dvp);
2648 		args.what.type = (va->va_type == VFIFO) ? NF3FIFO : NF3SOCK;
2649 		error = vattr_to_sattr3(va,
2650 				&args.what.mknoddata3_u.pipe_attributes);
2651 		if (error) {
2652 			/* req time field(s) overflow - return immediately */
2653 			return (error);
2654 		}
2655 		break;
2656 
2657 	default:
2658 		return (EINVAL);
2659 	}
2660 
2661 	douprintf = 1;
2662 
2663 	t = gethrtime();
2664 
2665 	error = rfs3call(VTOMI(dvp), NFSPROC3_MKNOD,
2666 	    xdr_MKNOD3args, (caddr_t)&args,
2667 	    xdr_MKNOD3res, (caddr_t)&res, cr,
2668 	    &douprintf, &res.status, 0, NULL);
2669 
2670 	if (error) {
2671 		PURGE_ATTRCACHE(dvp);
2672 		return (error);
2673 	}
2674 
2675 	error = geterrno3(res.status);
2676 	if (!error) {
2677 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2678 		if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2679 			nfs_purge_rddir_cache(dvp);
2680 
2681 		if (!res.resok.obj.handle_follows) {
2682 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2683 			if (error)
2684 				return (error);
2685 		} else {
2686 			if (res.resok.obj_attributes.attributes) {
2687 				vp = makenfs3node(&res.resok.obj.handle,
2688 				    &res.resok.obj_attributes.attr,
2689 				    dvp->v_vfsp, t, cr, NULL, NULL);
2690 			} else {
2691 				vp = makenfs3node(&res.resok.obj.handle, NULL,
2692 				    dvp->v_vfsp, t, cr, NULL, NULL);
2693 				if (vp->v_type == VNON) {
2694 					vattr.va_mask = AT_TYPE;
2695 					error = nfs3getattr(vp, &vattr, cr);
2696 					if (error) {
2697 						VN_RELE(vp);
2698 						return (error);
2699 					}
2700 					vp->v_type = vattr.va_type;
2701 				}
2702 
2703 			}
2704 			dnlc_update(dvp, nm, vp);
2705 		}
2706 
2707 		if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
2708 			va->va_mask = AT_GID;
2709 			(void) nfs3setattr(vp, va, 0, cr);
2710 		}
2711 
2712 		/*
2713 		 * If vnode is a device create special vnode
2714 		 */
2715 		if (IS_DEVVP(vp)) {
2716 			*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2717 			VN_RELE(vp);
2718 		} else
2719 			*vpp = vp;
2720 	} else {
2721 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2722 		PURGE_STALE_FH(error, dvp, cr);
2723 	}
2724 	return (error);
2725 }
2726 
2727 /*
2728  * Weirdness: if the vnode to be removed is open
2729  * we rename it instead of removing it and nfs_inactive
2730  * will remove the new name.
2731  */
2732 static int
2733 nfs3_remove(vnode_t *dvp, char *nm, cred_t *cr)
2734 {
2735 	int error;
2736 	REMOVE3args args;
2737 	REMOVE3res res;
2738 	vnode_t *vp;
2739 	char *tmpname;
2740 	int douprintf;
2741 	rnode_t *rp;
2742 	rnode_t *drp;
2743 	hrtime_t t;
2744 
2745 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
2746 		return (EPERM);
2747 	drp = VTOR(dvp);
2748 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2749 		return (EINTR);
2750 
2751 	error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2752 	if (error) {
2753 		nfs_rw_exit(&drp->r_rwlock);
2754 		return (error);
2755 	}
2756 
2757 	if (vp->v_type == VDIR && secpolicy_fs_linkdir(cr, dvp->v_vfsp)) {
2758 		VN_RELE(vp);
2759 		nfs_rw_exit(&drp->r_rwlock);
2760 		return (EPERM);
2761 	}
2762 
2763 	/*
2764 	 * First just remove the entry from the name cache, as it
2765 	 * is most likely the only entry for this vp.
2766 	 */
2767 	dnlc_remove(dvp, nm);
2768 
2769 	/*
2770 	 * If the file has a v_count > 1 then there may be more than one
2771 	 * entry in the name cache due multiple links or an open file,
2772 	 * but we don't have the real reference count so flush all
2773 	 * possible entries.
2774 	 */
2775 	if (vp->v_count > 1)
2776 		dnlc_purge_vp(vp);
2777 
2778 	/*
2779 	 * Now we have the real reference count on the vnode
2780 	 */
2781 	rp = VTOR(vp);
2782 	mutex_enter(&rp->r_statelock);
2783 	if (vp->v_count > 1 &&
2784 	    (rp->r_unldvp == NULL || strcmp(nm, rp->r_unlname) == 0)) {
2785 		mutex_exit(&rp->r_statelock);
2786 		tmpname = newname();
2787 		error = nfs3rename(dvp, nm, dvp, tmpname, cr);
2788 		if (error)
2789 			kmem_free(tmpname, MAXNAMELEN);
2790 		else {
2791 			mutex_enter(&rp->r_statelock);
2792 			if (rp->r_unldvp == NULL) {
2793 				VN_HOLD(dvp);
2794 				rp->r_unldvp = dvp;
2795 				if (rp->r_unlcred != NULL)
2796 					crfree(rp->r_unlcred);
2797 				crhold(cr);
2798 				rp->r_unlcred = cr;
2799 				rp->r_unlname = tmpname;
2800 			} else {
2801 				kmem_free(rp->r_unlname, MAXNAMELEN);
2802 				rp->r_unlname = tmpname;
2803 			}
2804 			mutex_exit(&rp->r_statelock);
2805 		}
2806 	} else {
2807 		mutex_exit(&rp->r_statelock);
2808 		/*
2809 		 * We need to flush any dirty pages which happen to
2810 		 * be hanging around before removing the file.  This
2811 		 * shouldn't happen very often and mostly on file
2812 		 * systems mounted "nocto".
2813 		 */
2814 		if (vn_has_cached_data(vp) &&
2815 		    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
2816 			error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
2817 			if (error && (error == ENOSPC || error == EDQUOT)) {
2818 				mutex_enter(&rp->r_statelock);
2819 				if (!rp->r_error)
2820 					rp->r_error = error;
2821 				mutex_exit(&rp->r_statelock);
2822 			}
2823 		}
2824 
2825 		setdiropargs3(&args.object, nm, dvp);
2826 
2827 		douprintf = 1;
2828 
2829 		t = gethrtime();
2830 
2831 		error = rfs3call(VTOMI(dvp), NFSPROC3_REMOVE,
2832 		    xdr_diropargs3, (caddr_t)&args,
2833 		    xdr_REMOVE3res, (caddr_t)&res, cr,
2834 		    &douprintf, &res.status, 0, NULL);
2835 
2836 		/*
2837 		 * The xattr dir may be gone after last attr is removed,
2838 		 * so flush it from dnlc.
2839 		 */
2840 		if (dvp->v_flag & V_XATTRDIR)
2841 			dnlc_purge_vp(dvp);
2842 
2843 		PURGE_ATTRCACHE(vp);
2844 
2845 		if (error) {
2846 			PURGE_ATTRCACHE(dvp);
2847 		} else {
2848 			error = geterrno3(res.status);
2849 			if (!error) {
2850 				nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t,
2851 				    cr);
2852 				if (HAVE_RDDIR_CACHE(drp))
2853 					nfs_purge_rddir_cache(dvp);
2854 			} else {
2855 				nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc,
2856 				    t, cr);
2857 				PURGE_STALE_FH(error, dvp, cr);
2858 			}
2859 		}
2860 	}
2861 
2862 	VN_RELE(vp);
2863 
2864 	nfs_rw_exit(&drp->r_rwlock);
2865 
2866 	return (error);
2867 }
2868 
2869 static int
2870 nfs3_link(vnode_t *tdvp, vnode_t *svp, char *tnm, cred_t *cr)
2871 {
2872 	int error;
2873 	LINK3args args;
2874 	LINK3res res;
2875 	vnode_t *realvp;
2876 	int douprintf;
2877 	mntinfo_t *mi;
2878 	rnode_t *tdrp;
2879 	hrtime_t t;
2880 
2881 	if (nfs_zone() != VTOMI(tdvp)->mi_zone)
2882 		return (EPERM);
2883 	if (VOP_REALVP(svp, &realvp) == 0)
2884 		svp = realvp;
2885 
2886 	mi = VTOMI(svp);
2887 
2888 	if (!(mi->mi_flags & MI_LINK))
2889 		return (EOPNOTSUPP);
2890 
2891 	args.file = *VTOFH3(svp);
2892 	setdiropargs3(&args.link, tnm, tdvp);
2893 
2894 	tdrp = VTOR(tdvp);
2895 	if (nfs_rw_enter_sig(&tdrp->r_rwlock, RW_WRITER, INTR(tdvp)))
2896 		return (EINTR);
2897 
2898 	dnlc_remove(tdvp, tnm);
2899 
2900 	douprintf = 1;
2901 
2902 	t = gethrtime();
2903 
2904 	error = rfs3call(mi, NFSPROC3_LINK,
2905 	    xdr_LINK3args, (caddr_t)&args,
2906 	    xdr_LINK3res, (caddr_t)&res, cr,
2907 	    &douprintf, &res.status, 0, NULL);
2908 
2909 	if (error) {
2910 		PURGE_ATTRCACHE(tdvp);
2911 		PURGE_ATTRCACHE(svp);
2912 		nfs_rw_exit(&tdrp->r_rwlock);
2913 		return (error);
2914 	}
2915 
2916 	error = geterrno3(res.status);
2917 
2918 	if (!error) {
2919 		nfs3_cache_post_op_attr(svp, &res.resok.file_attributes, t, cr);
2920 		nfs3_cache_wcc_data(tdvp, &res.resok.linkdir_wcc, t, cr);
2921 		if (HAVE_RDDIR_CACHE(tdrp))
2922 			nfs_purge_rddir_cache(tdvp);
2923 		dnlc_update(tdvp, tnm, svp);
2924 	} else {
2925 		nfs3_cache_post_op_attr(svp, &res.resfail.file_attributes, t,
2926 		    cr);
2927 		nfs3_cache_wcc_data(tdvp, &res.resfail.linkdir_wcc, t, cr);
2928 		if (error == EOPNOTSUPP) {
2929 			mutex_enter(&mi->mi_lock);
2930 			mi->mi_flags &= ~MI_LINK;
2931 			mutex_exit(&mi->mi_lock);
2932 		}
2933 	}
2934 
2935 	nfs_rw_exit(&tdrp->r_rwlock);
2936 
2937 	return (error);
2938 }
2939 
2940 static int
2941 nfs3_rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2942 {
2943 	vnode_t *realvp;
2944 
2945 	if (nfs_zone() != VTOMI(odvp)->mi_zone)
2946 		return (EPERM);
2947 	if (VOP_REALVP(ndvp, &realvp) == 0)
2948 		ndvp = realvp;
2949 
2950 	return (nfs3rename(odvp, onm, ndvp, nnm, cr));
2951 }
2952 
2953 /*
2954  * nfs3rename does the real work of renaming in NFS Version 3.
2955  */
2956 static int
2957 nfs3rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2958 {
2959 	int error;
2960 	RENAME3args args;
2961 	RENAME3res res;
2962 	int douprintf;
2963 	vnode_t *nvp;
2964 	vnode_t *ovp = NULL;
2965 	char *tmpname;
2966 	rnode_t *rp;
2967 	rnode_t *odrp;
2968 	rnode_t *ndrp;
2969 	hrtime_t t;
2970 
2971 	ASSERT(nfs_zone() == VTOMI(odvp)->mi_zone);
2972 
2973 	if (strcmp(onm, ".") == 0 || strcmp(onm, "..") == 0 ||
2974 	    strcmp(nnm, ".") == 0 || strcmp(nnm, "..") == 0)
2975 		return (EINVAL);
2976 
2977 	odrp = VTOR(odvp);
2978 	ndrp = VTOR(ndvp);
2979 	if ((intptr_t)odrp < (intptr_t)ndrp) {
2980 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp)))
2981 			return (EINTR);
2982 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp))) {
2983 			nfs_rw_exit(&odrp->r_rwlock);
2984 			return (EINTR);
2985 		}
2986 	} else {
2987 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp)))
2988 			return (EINTR);
2989 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp))) {
2990 			nfs_rw_exit(&ndrp->r_rwlock);
2991 			return (EINTR);
2992 		}
2993 	}
2994 
2995 	/*
2996 	 * Lookup the target file.  If it exists, it needs to be
2997 	 * checked to see whether it is a mount point and whether
2998 	 * it is active (open).
2999 	 */
3000 	error = nfs3lookup(ndvp, nnm, &nvp, NULL, 0, NULL, cr, 0);
3001 	if (!error) {
3002 		/*
3003 		 * If this file has been mounted on, then just
3004 		 * return busy because renaming to it would remove
3005 		 * the mounted file system from the name space.
3006 		 */
3007 		if (vn_mountedvfs(nvp) != NULL) {
3008 			VN_RELE(nvp);
3009 			nfs_rw_exit(&odrp->r_rwlock);
3010 			nfs_rw_exit(&ndrp->r_rwlock);
3011 			return (EBUSY);
3012 		}
3013 
3014 		/*
3015 		 * Purge the name cache of all references to this vnode
3016 		 * so that we can check the reference count to infer
3017 		 * whether it is active or not.
3018 		 */
3019 		/*
3020 		 * First just remove the entry from the name cache, as it
3021 		 * is most likely the only entry for this vp.
3022 		 */
3023 		dnlc_remove(ndvp, nnm);
3024 		/*
3025 		 * If the file has a v_count > 1 then there may be more
3026 		 * than one entry in the name cache due multiple links
3027 		 * or an open file, but we don't have the real reference
3028 		 * count so flush all possible entries.
3029 		 */
3030 		if (nvp->v_count > 1)
3031 			dnlc_purge_vp(nvp);
3032 
3033 		/*
3034 		 * If the vnode is active and is not a directory,
3035 		 * arrange to rename it to a
3036 		 * temporary file so that it will continue to be
3037 		 * accessible.  This implements the "unlink-open-file"
3038 		 * semantics for the target of a rename operation.
3039 		 * Before doing this though, make sure that the
3040 		 * source and target files are not already the same.
3041 		 */
3042 		if (nvp->v_count > 1 && nvp->v_type != VDIR) {
3043 			/*
3044 			 * Lookup the source name.
3045 			 */
3046 			error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL,
3047 			    cr, 0);
3048 
3049 			/*
3050 			 * The source name *should* already exist.
3051 			 */
3052 			if (error) {
3053 				VN_RELE(nvp);
3054 				nfs_rw_exit(&odrp->r_rwlock);
3055 				nfs_rw_exit(&ndrp->r_rwlock);
3056 				return (error);
3057 			}
3058 
3059 			/*
3060 			 * Compare the two vnodes.  If they are the same,
3061 			 * just release all held vnodes and return success.
3062 			 */
3063 			if (ovp == nvp) {
3064 				VN_RELE(ovp);
3065 				VN_RELE(nvp);
3066 				nfs_rw_exit(&odrp->r_rwlock);
3067 				nfs_rw_exit(&ndrp->r_rwlock);
3068 				return (0);
3069 			}
3070 
3071 			/*
3072 			 * Can't mix and match directories and non-
3073 			 * directories in rename operations.  We already
3074 			 * know that the target is not a directory.  If
3075 			 * the source is a directory, return an error.
3076 			 */
3077 			if (ovp->v_type == VDIR) {
3078 				VN_RELE(ovp);
3079 				VN_RELE(nvp);
3080 				nfs_rw_exit(&odrp->r_rwlock);
3081 				nfs_rw_exit(&ndrp->r_rwlock);
3082 				return (ENOTDIR);
3083 			}
3084 
3085 			/*
3086 			 * The target file exists, is not the same as
3087 			 * the source file, and is active.  Link it
3088 			 * to a temporary filename to avoid having
3089 			 * the server removing the file completely.
3090 			 */
3091 			tmpname = newname();
3092 			error = nfs3_link(ndvp, nvp, tmpname, cr);
3093 			if (error == EOPNOTSUPP) {
3094 				error = nfs3_rename(ndvp, nnm, ndvp, tmpname,
3095 				    cr);
3096 			}
3097 			if (error) {
3098 				kmem_free(tmpname, MAXNAMELEN);
3099 				VN_RELE(ovp);
3100 				VN_RELE(nvp);
3101 				nfs_rw_exit(&odrp->r_rwlock);
3102 				nfs_rw_exit(&ndrp->r_rwlock);
3103 				return (error);
3104 			}
3105 			rp = VTOR(nvp);
3106 			mutex_enter(&rp->r_statelock);
3107 			if (rp->r_unldvp == NULL) {
3108 				VN_HOLD(ndvp);
3109 				rp->r_unldvp = ndvp;
3110 				if (rp->r_unlcred != NULL)
3111 					crfree(rp->r_unlcred);
3112 				crhold(cr);
3113 				rp->r_unlcred = cr;
3114 				rp->r_unlname = tmpname;
3115 			} else {
3116 				kmem_free(rp->r_unlname, MAXNAMELEN);
3117 				rp->r_unlname = tmpname;
3118 			}
3119 			mutex_exit(&rp->r_statelock);
3120 		}
3121 
3122 		VN_RELE(nvp);
3123 	}
3124 
3125 	if (ovp == NULL) {
3126 		/*
3127 		 * When renaming directories to be a subdirectory of a
3128 		 * different parent, the dnlc entry for ".." will no
3129 		 * longer be valid, so it must be removed.
3130 		 *
3131 		 * We do a lookup here to determine whether we are renaming
3132 		 * a directory and we need to check if we are renaming
3133 		 * an unlinked file.  This might have already been done
3134 		 * in previous code, so we check ovp == NULL to avoid
3135 		 * doing it twice.
3136 		 */
3137 
3138 		error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL, cr, 0);
3139 		/*
3140 		 * The source name *should* already exist.
3141 		 */
3142 		if (error) {
3143 			nfs_rw_exit(&odrp->r_rwlock);
3144 			nfs_rw_exit(&ndrp->r_rwlock);
3145 			return (error);
3146 		}
3147 		ASSERT(ovp != NULL);
3148 	}
3149 
3150 	dnlc_remove(odvp, onm);
3151 	dnlc_remove(ndvp, nnm);
3152 
3153 	setdiropargs3(&args.from, onm, odvp);
3154 	setdiropargs3(&args.to, nnm, ndvp);
3155 
3156 	douprintf = 1;
3157 
3158 	t = gethrtime();
3159 
3160 	error = rfs3call(VTOMI(odvp), NFSPROC3_RENAME,
3161 	    xdr_RENAME3args, (caddr_t)&args,
3162 	    xdr_RENAME3res, (caddr_t)&res, cr,
3163 	    &douprintf, &res.status, 0, NULL);
3164 
3165 	if (error) {
3166 		PURGE_ATTRCACHE(odvp);
3167 		PURGE_ATTRCACHE(ndvp);
3168 		VN_RELE(ovp);
3169 		nfs_rw_exit(&odrp->r_rwlock);
3170 		nfs_rw_exit(&ndrp->r_rwlock);
3171 		return (error);
3172 	}
3173 
3174 	error = geterrno3(res.status);
3175 
3176 	if (!error) {
3177 		nfs3_cache_wcc_data(odvp, &res.resok.fromdir_wcc, t, cr);
3178 		if (HAVE_RDDIR_CACHE(odrp))
3179 			nfs_purge_rddir_cache(odvp);
3180 		if (ndvp != odvp) {
3181 			nfs3_cache_wcc_data(ndvp, &res.resok.todir_wcc, t, cr);
3182 			if (HAVE_RDDIR_CACHE(ndrp))
3183 				nfs_purge_rddir_cache(ndvp);
3184 		}
3185 		/*
3186 		 * when renaming directories to be a subdirectory of a
3187 		 * different parent, the dnlc entry for ".." will no
3188 		 * longer be valid, so it must be removed
3189 		 */
3190 		rp = VTOR(ovp);
3191 		if (ndvp != odvp) {
3192 			if (ovp->v_type == VDIR) {
3193 				dnlc_remove(ovp, "..");
3194 				if (HAVE_RDDIR_CACHE(rp))
3195 					nfs_purge_rddir_cache(ovp);
3196 			}
3197 		}
3198 
3199 		/*
3200 		 * If we are renaming the unlinked file, update the
3201 		 * r_unldvp and r_unlname as needed.
3202 		 */
3203 		mutex_enter(&rp->r_statelock);
3204 		if (rp->r_unldvp != NULL) {
3205 			if (strcmp(rp->r_unlname, onm) == 0) {
3206 				(void) strncpy(rp->r_unlname, nnm, MAXNAMELEN);
3207 				rp->r_unlname[MAXNAMELEN - 1] = '\0';
3208 
3209 				if (ndvp != rp->r_unldvp) {
3210 					VN_RELE(rp->r_unldvp);
3211 					rp->r_unldvp = ndvp;
3212 					VN_HOLD(ndvp);
3213 				}
3214 			}
3215 		}
3216 		mutex_exit(&rp->r_statelock);
3217 	} else {
3218 		nfs3_cache_wcc_data(odvp, &res.resfail.fromdir_wcc, t, cr);
3219 		if (ndvp != odvp) {
3220 			nfs3_cache_wcc_data(ndvp, &res.resfail.todir_wcc, t,
3221 			    cr);
3222 		}
3223 		/*
3224 		 * System V defines rename to return EEXIST, not
3225 		 * ENOTEMPTY if the target directory is not empty.
3226 		 * Over the wire, the error is NFSERR_ENOTEMPTY
3227 		 * which geterrno maps to ENOTEMPTY.
3228 		 */
3229 		if (error == ENOTEMPTY)
3230 			error = EEXIST;
3231 	}
3232 
3233 	VN_RELE(ovp);
3234 
3235 	nfs_rw_exit(&odrp->r_rwlock);
3236 	nfs_rw_exit(&ndrp->r_rwlock);
3237 
3238 	return (error);
3239 }
3240 
3241 static int
3242 nfs3_mkdir(vnode_t *dvp, char *nm, struct vattr *va, vnode_t **vpp, cred_t *cr)
3243 {
3244 	int error;
3245 	MKDIR3args args;
3246 	MKDIR3res res;
3247 	int douprintf;
3248 	struct vattr vattr;
3249 	vnode_t *vp;
3250 	rnode_t *drp;
3251 	hrtime_t t;
3252 
3253 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
3254 		return (EPERM);
3255 	setdiropargs3(&args.where, nm, dvp);
3256 
3257 	/*
3258 	 * Decide what the group-id and set-gid bit of the created directory
3259 	 * should be.  May have to do a setattr to get the gid right.
3260 	 */
3261 	error = setdirgid(dvp, &va->va_gid, cr);
3262 	if (error)
3263 		return (error);
3264 	error = setdirmode(dvp, &va->va_mode, cr);
3265 	if (error)
3266 		return (error);
3267 	va->va_mask |= AT_MODE|AT_GID;
3268 
3269 	error = vattr_to_sattr3(va, &args.attributes);
3270 	if (error) {
3271 		/* req time field(s) overflow - return immediately */
3272 		return (error);
3273 	}
3274 
3275 	drp = VTOR(dvp);
3276 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3277 		return (EINTR);
3278 
3279 	dnlc_remove(dvp, nm);
3280 
3281 	douprintf = 1;
3282 
3283 	t = gethrtime();
3284 
3285 	error = rfs3call(VTOMI(dvp), NFSPROC3_MKDIR,
3286 	    xdr_MKDIR3args, (caddr_t)&args,
3287 	    xdr_MKDIR3res, (caddr_t)&res, cr,
3288 	    &douprintf, &res.status, 0, NULL);
3289 
3290 	if (error) {
3291 		PURGE_ATTRCACHE(dvp);
3292 		nfs_rw_exit(&drp->r_rwlock);
3293 		return (error);
3294 	}
3295 
3296 	error = geterrno3(res.status);
3297 	if (!error) {
3298 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3299 		if (HAVE_RDDIR_CACHE(drp))
3300 			nfs_purge_rddir_cache(dvp);
3301 
3302 		if (!res.resok.obj.handle_follows) {
3303 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3304 			if (error) {
3305 				nfs_rw_exit(&drp->r_rwlock);
3306 				return (error);
3307 			}
3308 		} else {
3309 			if (res.resok.obj_attributes.attributes) {
3310 				vp = makenfs3node(&res.resok.obj.handle,
3311 				    &res.resok.obj_attributes.attr,
3312 				    dvp->v_vfsp, t, cr, NULL, NULL);
3313 			} else {
3314 				vp = makenfs3node(&res.resok.obj.handle, NULL,
3315 				    dvp->v_vfsp, t, cr, NULL, NULL);
3316 				if (vp->v_type == VNON) {
3317 					vattr.va_mask = AT_TYPE;
3318 					error = nfs3getattr(vp, &vattr, cr);
3319 					if (error) {
3320 						VN_RELE(vp);
3321 						nfs_rw_exit(&drp->r_rwlock);
3322 						return (error);
3323 					}
3324 					vp->v_type = vattr.va_type;
3325 				}
3326 			}
3327 			dnlc_update(dvp, nm, vp);
3328 		}
3329 		if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
3330 			va->va_mask = AT_GID;
3331 			(void) nfs3setattr(vp, va, 0, cr);
3332 		}
3333 		*vpp = vp;
3334 	} else {
3335 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3336 		PURGE_STALE_FH(error, dvp, cr);
3337 	}
3338 
3339 	nfs_rw_exit(&drp->r_rwlock);
3340 
3341 	return (error);
3342 }
3343 
3344 static int
3345 nfs3_rmdir(vnode_t *dvp, char *nm, vnode_t *cdir, cred_t *cr)
3346 {
3347 	int error;
3348 	RMDIR3args args;
3349 	RMDIR3res res;
3350 	vnode_t *vp;
3351 	int douprintf;
3352 	rnode_t *drp;
3353 	hrtime_t t;
3354 
3355 	if (nfs_zone() != VTOMI(dvp)->mi_zone)
3356 		return (EPERM);
3357 	drp = VTOR(dvp);
3358 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3359 		return (EINTR);
3360 
3361 	/*
3362 	 * Attempt to prevent a rmdir(".") from succeeding.
3363 	 */
3364 	error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3365 	if (error) {
3366 		nfs_rw_exit(&drp->r_rwlock);
3367 		return (error);
3368 	}
3369 
3370 	if (vp == cdir) {
3371 		VN_RELE(vp);
3372 		nfs_rw_exit(&drp->r_rwlock);
3373 		return (EINVAL);
3374 	}
3375 
3376 	setdiropargs3(&args.object, nm, dvp);
3377 
3378 	/*
3379 	 * First just remove the entry from the name cache, as it
3380 	 * is most likely an entry for this vp.
3381 	 */
3382 	dnlc_remove(dvp, nm);
3383 
3384 	/*
3385 	 * If there vnode reference count is greater than one, then
3386 	 * there may be additional references in the DNLC which will
3387 	 * need to be purged.  First, trying removing the entry for
3388 	 * the parent directory and see if that removes the additional
3389 	 * reference(s).  If that doesn't do it, then use dnlc_purge_vp
3390 	 * to completely remove any references to the directory which
3391 	 * might still exist in the DNLC.
3392 	 */
3393 	if (vp->v_count > 1) {
3394 		dnlc_remove(vp, "..");
3395 		if (vp->v_count > 1)
3396 			dnlc_purge_vp(vp);
3397 	}
3398 
3399 	douprintf = 1;
3400 
3401 	t = gethrtime();
3402 
3403 	error = rfs3call(VTOMI(dvp), NFSPROC3_RMDIR,
3404 	    xdr_diropargs3, (caddr_t)&args,
3405 	    xdr_RMDIR3res, (caddr_t)&res, cr,
3406 	    &douprintf, &res.status, 0, NULL);
3407 
3408 	PURGE_ATTRCACHE(vp);
3409 
3410 	if (error) {
3411 		PURGE_ATTRCACHE(dvp);
3412 		VN_RELE(vp);
3413 		nfs_rw_exit(&drp->r_rwlock);
3414 		return (error);
3415 	}
3416 
3417 	error = geterrno3(res.status);
3418 	if (!error) {
3419 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3420 		if (HAVE_RDDIR_CACHE(drp))
3421 			nfs_purge_rddir_cache(dvp);
3422 		if (HAVE_RDDIR_CACHE(VTOR(vp)))
3423 			nfs_purge_rddir_cache(vp);
3424 	} else {
3425 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3426 		PURGE_STALE_FH(error, dvp, cr);
3427 		/*
3428 		 * System V defines rmdir to return EEXIST, not
3429 		 * ENOTEMPTY if the directory is not empty.  Over
3430 		 * the wire, the error is NFSERR_ENOTEMPTY which
3431 		 * geterrno maps to ENOTEMPTY.
3432 		 */
3433 		if (error == ENOTEMPTY)
3434 			error = EEXIST;
3435 	}
3436 
3437 	VN_RELE(vp);
3438 
3439 	nfs_rw_exit(&drp->r_rwlock);
3440 
3441 	return (error);
3442 }
3443 
3444 static int
3445 nfs3_symlink(vnode_t *dvp, char *lnm, struct vattr *tva, char *tnm, cred_t *cr)
3446 {
3447 	int error;
3448 	SYMLINK3args args;
3449 	SYMLINK3res res;
3450 	int douprintf;
3451 	mntinfo_t *mi;
3452 	vnode_t *vp;
3453 	rnode_t *rp;
3454 	char *contents;
3455 	rnode_t *drp;
3456 	hrtime_t t;
3457 
3458 	mi = VTOMI(dvp);
3459 
3460 	if (nfs_zone() != mi->mi_zone)
3461 		return (EPERM);
3462 	if (!(mi->mi_flags & MI_SYMLINK))
3463 		return (EOPNOTSUPP);
3464 
3465 	setdiropargs3(&args.where, lnm, dvp);
3466 	error = vattr_to_sattr3(tva, &args.symlink.symlink_attributes);
3467 	if (error) {
3468 		/* req time field(s) overflow - return immediately */
3469 		return (error);
3470 	}
3471 	args.symlink.symlink_data = tnm;
3472 
3473 	drp = VTOR(dvp);
3474 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3475 		return (EINTR);
3476 
3477 	dnlc_remove(dvp, lnm);
3478 
3479 	douprintf = 1;
3480 
3481 	t = gethrtime();
3482 
3483 	error = rfs3call(mi, NFSPROC3_SYMLINK,
3484 	    xdr_SYMLINK3args, (caddr_t)&args,
3485 	    xdr_SYMLINK3res, (caddr_t)&res, cr,
3486 	    &douprintf, &res.status, 0, NULL);
3487 
3488 	if (error) {
3489 		PURGE_ATTRCACHE(dvp);
3490 		nfs_rw_exit(&drp->r_rwlock);
3491 		return (error);
3492 	}
3493 
3494 	error = geterrno3(res.status);
3495 	if (!error) {
3496 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3497 		if (HAVE_RDDIR_CACHE(drp))
3498 			nfs_purge_rddir_cache(dvp);
3499 
3500 		if (res.resok.obj.handle_follows) {
3501 			if (res.resok.obj_attributes.attributes) {
3502 				vp = makenfs3node(&res.resok.obj.handle,
3503 				    &res.resok.obj_attributes.attr,
3504 				    dvp->v_vfsp, t, cr, NULL, NULL);
3505 			} else {
3506 				vp = makenfs3node(&res.resok.obj.handle, NULL,
3507 				    dvp->v_vfsp, t, cr, NULL, NULL);
3508 				vp->v_type = VLNK;
3509 				vp->v_rdev = 0;
3510 			}
3511 			dnlc_update(dvp, lnm, vp);
3512 			rp = VTOR(vp);
3513 			if (nfs3_do_symlink_cache &&
3514 			    rp->r_symlink.contents == NULL) {
3515 
3516 				contents = kmem_alloc(MAXPATHLEN,
3517 				    KM_NOSLEEP);
3518 
3519 				if (contents != NULL) {
3520 					mutex_enter(&rp->r_statelock);
3521 					if (rp->r_symlink.contents == NULL) {
3522 						rp->r_symlink.len = strlen(tnm);
3523 						bcopy(tnm, contents,
3524 						    rp->r_symlink.len);
3525 						rp->r_symlink.contents =
3526 						    contents;
3527 						rp->r_symlink.size = MAXPATHLEN;
3528 						mutex_exit(&rp->r_statelock);
3529 					} else {
3530 						mutex_exit(&rp->r_statelock);
3531 						kmem_free((void *)contents,
3532 							    MAXPATHLEN);
3533 					}
3534 				}
3535 			}
3536 			VN_RELE(vp);
3537 		}
3538 	} else {
3539 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3540 		PURGE_STALE_FH(error, dvp, cr);
3541 		if (error == EOPNOTSUPP) {
3542 			mutex_enter(&mi->mi_lock);
3543 			mi->mi_flags &= ~MI_SYMLINK;
3544 			mutex_exit(&mi->mi_lock);
3545 		}
3546 	}
3547 
3548 	nfs_rw_exit(&drp->r_rwlock);
3549 
3550 	return (error);
3551 }
3552 
3553 #ifdef DEBUG
3554 static int nfs3_readdir_cache_hits = 0;
3555 static int nfs3_readdir_cache_shorts = 0;
3556 static int nfs3_readdir_cache_waits = 0;
3557 static int nfs3_readdir_cache_misses = 0;
3558 static int nfs3_readdir_readahead = 0;
3559 #endif
3560 
3561 static int nfs3_shrinkreaddir = 0;
3562 
3563 /*
3564  * Read directory entries.
3565  * There are some weird things to look out for here.  The uio_loffset
3566  * field is either 0 or it is the offset returned from a previous
3567  * readdir.  It is an opaque value used by the server to find the
3568  * correct directory block to read. The count field is the number
3569  * of blocks to read on the server.  This is advisory only, the server
3570  * may return only one block's worth of entries.  Entries may be compressed
3571  * on the server.
3572  */
3573 static int
3574 nfs3_readdir(vnode_t *vp, struct uio *uiop, cred_t *cr, int *eofp)
3575 {
3576 	int error;
3577 	size_t count;
3578 	rnode_t *rp;
3579 	rddir_cache *rdc;
3580 	rddir_cache *nrdc;
3581 	rddir_cache *rrdc;
3582 #ifdef DEBUG
3583 	int missed;
3584 #endif
3585 	int doreadahead;
3586 	rddir_cache srdc;
3587 	avl_index_t where;
3588 
3589 	if (nfs_zone() != VTOMI(vp)->mi_zone)
3590 		return (EIO);
3591 	rp = VTOR(vp);
3592 
3593 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
3594 
3595 	/*
3596 	 * Make sure that the directory cache is valid.
3597 	 */
3598 	if (HAVE_RDDIR_CACHE(rp)) {
3599 		if (nfs_disable_rddir_cache) {
3600 			/*
3601 			 * Setting nfs_disable_rddir_cache in /etc/system
3602 			 * allows interoperability with servers that do not
3603 			 * properly update the attributes of directories.
3604 			 * Any cached information gets purged before an
3605 			 * access is made to it.
3606 			 */
3607 			nfs_purge_rddir_cache(vp);
3608 		} else {
3609 			error = nfs3_validate_caches(vp, cr);
3610 			if (error)
3611 				return (error);
3612 		}
3613 	}
3614 
3615 	/*
3616 	 * It is possible that some servers may not be able to correctly
3617 	 * handle a large READDIR or READDIRPLUS request due to bugs in
3618 	 * their implementation.  In order to continue to interoperate
3619 	 * with them, this workaround is provided to limit the maximum
3620 	 * size of a READDIRPLUS request to 1024.  In any case, the request
3621 	 * size is limited to MAXBSIZE.
3622 	 */
3623 	count = MIN(uiop->uio_iov->iov_len,
3624 	    nfs3_shrinkreaddir ? 1024 : MAXBSIZE);
3625 
3626 	nrdc = NULL;
3627 #ifdef DEBUG
3628 	missed = 0;
3629 #endif
3630 top:
3631 	/*
3632 	 * Short circuit last readdir which always returns 0 bytes.
3633 	 * This can be done after the directory has been read through
3634 	 * completely at least once.  This will set r_direof which
3635 	 * can be used to find the value of the last cookie.
3636 	 */
3637 	mutex_enter(&rp->r_statelock);
3638 	if (rp->r_direof != NULL &&
3639 	    uiop->uio_loffset == rp->r_direof->nfs3_ncookie) {
3640 		mutex_exit(&rp->r_statelock);
3641 #ifdef DEBUG
3642 		nfs3_readdir_cache_shorts++;
3643 #endif
3644 		if (eofp)
3645 			*eofp = 1;
3646 		if (nrdc != NULL)
3647 			rddir_cache_rele(nrdc);
3648 		return (0);
3649 	}
3650 	/*
3651 	 * Look for a cache entry.  Cache entries are identified
3652 	 * by the NFS cookie value and the byte count requested.
3653 	 */
3654 	srdc.nfs3_cookie = uiop->uio_loffset;
3655 	srdc.buflen = count;
3656 	rdc = avl_find(&rp->r_dir, &srdc, &where);
3657 	if (rdc != NULL) {
3658 		rddir_cache_hold(rdc);
3659 		/*
3660 		 * If the cache entry is in the process of being
3661 		 * filled in, wait until this completes.  The
3662 		 * RDDIRWAIT bit is set to indicate that someone
3663 		 * is waiting and then the thread currently
3664 		 * filling the entry is done, it should do a
3665 		 * cv_broadcast to wakeup all of the threads
3666 		 * waiting for it to finish.
3667 		 */
3668 		if (rdc->flags & RDDIR) {
3669 			nfs_rw_exit(&rp->r_rwlock);
3670 			rdc->flags |= RDDIRWAIT;
3671 #ifdef DEBUG
3672 			nfs3_readdir_cache_waits++;
3673 #endif
3674 			if (!cv_wait_sig(&rdc->cv, &rp->r_statelock)) {
3675 				/*
3676 				 * We got interrupted, probably
3677 				 * the user typed ^C or an alarm
3678 				 * fired.  We free the new entry
3679 				 * if we allocated one.
3680 				 */
3681 				mutex_exit(&rp->r_statelock);
3682 				(void) nfs_rw_enter_sig(&rp->r_rwlock,
3683 					RW_READER, FALSE);
3684 				rddir_cache_rele(rdc);
3685 				if (nrdc != NULL)
3686 					rddir_cache_rele(nrdc);
3687 				return (EINTR);
3688 			}
3689 			mutex_exit(&rp->r_statelock);
3690 			(void) nfs_rw_enter_sig(&rp->r_rwlock,
3691 				RW_READER, FALSE);
3692 			rddir_cache_rele(rdc);
3693 			goto top;
3694 		}
3695 		/*
3696 		 * Check to see if a readdir is required to
3697 		 * fill the entry.  If so, mark this entry
3698 		 * as being filled, remove our reference,
3699 		 * and branch to the code to fill the entry.
3700 		 */
3701 		if (rdc->flags & RDDIRREQ) {
3702 			rdc->flags &= ~RDDIRREQ;
3703 			rdc->flags |= RDDIR;
3704 			if (nrdc != NULL)
3705 				rddir_cache_rele(nrdc);
3706 			nrdc = rdc;
3707 			mutex_exit(&rp->r_statelock);
3708 			goto bottom;
3709 		}
3710 #ifdef DEBUG
3711 		if (!missed)
3712 			nfs3_readdir_cache_hits++;
3713 #endif
3714 		/*
3715 		 * If an error occurred while attempting
3716 		 * to fill the cache entry, just return it.
3717 		 */
3718 		if (rdc->error) {
3719 			error = rdc->error;
3720 			mutex_exit(&rp->r_statelock);
3721 			rddir_cache_rele(rdc);
3722 			if (nrdc != NULL)
3723 				rddir_cache_rele(nrdc);
3724 			return (error);
3725 		}
3726 
3727 		/*
3728 		 * The cache entry is complete and good,
3729 		 * copyout the dirent structs to the calling
3730 		 * thread.
3731 		 */
3732 		error = uiomove(rdc->entries, rdc->entlen, UIO_READ, uiop);
3733 
3734 		/*
3735 		 * If no error occurred during the copyout,
3736 		 * update the offset in the uio struct to
3737 		 * contain the value of the next cookie
3738 		 * and set the eof value appropriately.
3739 		 */
3740 		if (!error) {
3741 			uiop->uio_loffset = rdc->nfs3_ncookie;
3742 			if (eofp)
3743 				*eofp = rdc->eof;
3744 		}
3745 
3746 		/*
3747 		 * Decide whether to do readahead.
3748 		 *
3749 		 * Don't if have already read to the end of
3750 		 * directory.  There is nothing more to read.
3751 		 *
3752 		 * Don't if the application is not doing
3753 		 * lookups in the directory.  The readahead
3754 		 * is only effective if the application can
3755 		 * be doing work while an async thread is
3756 		 * handling the over the wire request.
3757 		 */
3758 		if (rdc->eof) {
3759 			rp->r_direof = rdc;
3760 			doreadahead = FALSE;
3761 		} else if (!(rp->r_flags & RLOOKUP))
3762 			doreadahead = FALSE;
3763 		else
3764 			doreadahead = TRUE;
3765 
3766 		if (!doreadahead) {
3767 			mutex_exit(&rp->r_statelock);
3768 			rddir_cache_rele(rdc);
3769 			if (nrdc != NULL)
3770 				rddir_cache_rele(nrdc);
3771 			return (error);
3772 		}
3773 
3774 		/*
3775 		 * Check to see whether we found an entry
3776 		 * for the readahead.  If so, we don't need
3777 		 * to do anything further, so free the new
3778 		 * entry if one was allocated.  Otherwise,
3779 		 * allocate a new entry, add it to the cache,
3780 		 * and then initiate an asynchronous readdir
3781 		 * operation to fill it.
3782 		 */
3783 		srdc.nfs3_cookie = rdc->nfs3_ncookie;
3784 		srdc.buflen = count;
3785 		rrdc = avl_find(&rp->r_dir, &srdc, &where);
3786 		if (rrdc != NULL) {
3787 			if (nrdc != NULL)
3788 				rddir_cache_rele(nrdc);
3789 		} else {
3790 			if (nrdc != NULL)
3791 				rrdc = nrdc;
3792 			else {
3793 				rrdc = rddir_cache_alloc(KM_NOSLEEP);
3794 			}
3795 			if (rrdc != NULL) {
3796 				rrdc->nfs3_cookie = rdc->nfs3_ncookie;
3797 				rrdc->buflen = count;
3798 				avl_insert(&rp->r_dir, rrdc, where);
3799 				rddir_cache_hold(rrdc);
3800 				mutex_exit(&rp->r_statelock);
3801 				rddir_cache_rele(rdc);
3802 #ifdef DEBUG
3803 				nfs3_readdir_readahead++;
3804 #endif
3805 				nfs_async_readdir(vp, rrdc, cr, do_nfs3readdir);
3806 				return (error);
3807 			}
3808 		}
3809 
3810 		mutex_exit(&rp->r_statelock);
3811 		rddir_cache_rele(rdc);
3812 		return (error);
3813 	}
3814 
3815 	/*
3816 	 * Didn't find an entry in the cache.  Construct a new empty
3817 	 * entry and link it into the cache.  Other processes attempting
3818 	 * to access this entry will need to wait until it is filled in.
3819 	 *
3820 	 * Since kmem_alloc may block, another pass through the cache
3821 	 * will need to be taken to make sure that another process
3822 	 * hasn't already added an entry to the cache for this request.
3823 	 */
3824 	if (nrdc == NULL) {
3825 		mutex_exit(&rp->r_statelock);
3826 		nrdc = rddir_cache_alloc(KM_SLEEP);
3827 		nrdc->nfs3_cookie = uiop->uio_loffset;
3828 		nrdc->buflen = count;
3829 		goto top;
3830 	}
3831 
3832 	/*
3833 	 * Add this entry to the cache.
3834 	 */
3835 	avl_insert(&rp->r_dir, nrdc, where);
3836 	rddir_cache_hold(nrdc);
3837 	mutex_exit(&rp->r_statelock);
3838 
3839 bottom:
3840 #ifdef DEBUG
3841 	missed = 1;
3842 	nfs3_readdir_cache_misses++;
3843 #endif
3844 	/*
3845 	 * Do the readdir.  This routine decides whether to use
3846 	 * READDIR or READDIRPLUS.
3847 	 */
3848 	error = do_nfs3readdir(vp, nrdc, cr);
3849 
3850 	/*
3851 	 * If this operation failed, just return the error which occurred.
3852 	 */
3853 	if (error != 0)
3854 		return (error);
3855 
3856 	/*
3857 	 * Since the RPC operation will have taken sometime and blocked
3858 	 * this process, another pass through the cache will need to be
3859 	 * taken to find the correct cache entry.  It is possible that
3860 	 * the correct cache entry will not be there (although one was
3861 	 * added) because the directory changed during the RPC operation
3862 	 * and the readdir cache was flushed.  In this case, just start
3863 	 * over.  It is hoped that this will not happen too often... :-)
3864 	 */
3865 	nrdc = NULL;
3866 	goto top;
3867 	/* NOTREACHED */
3868 }
3869 
3870 static int
3871 do_nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3872 {
3873 	int error;
3874 	rnode_t *rp;
3875 	mntinfo_t *mi;
3876 
3877 	rp = VTOR(vp);
3878 	mi = VTOMI(vp);
3879 	ASSERT(nfs_zone() == mi->mi_zone);
3880 	/*
3881 	 * Issue the proper request.
3882 	 *
3883 	 * If the server does not support READDIRPLUS, then use READDIR.
3884 	 *
3885 	 * Otherwise --
3886 	 * Issue a READDIRPLUS if reading to fill an empty cache or if
3887 	 * an application has performed a lookup in the directory which
3888 	 * required an over the wire lookup.  The use of READDIRPLUS
3889 	 * will help to (re)populate the DNLC.
3890 	 */
3891 	if (!(mi->mi_flags & MI_READDIRONLY) &&
3892 	    (rp->r_flags & (RLOOKUP | RREADDIRPLUS))) {
3893 		if (rp->r_flags & RREADDIRPLUS) {
3894 			mutex_enter(&rp->r_statelock);
3895 			rp->r_flags &= ~RREADDIRPLUS;
3896 			mutex_exit(&rp->r_statelock);
3897 		}
3898 		nfs3readdirplus(vp, rdc, cr);
3899 		if (rdc->error == EOPNOTSUPP)
3900 			nfs3readdir(vp, rdc, cr);
3901 	} else
3902 		nfs3readdir(vp, rdc, cr);
3903 
3904 	mutex_enter(&rp->r_statelock);
3905 	rdc->flags &= ~RDDIR;
3906 	if (rdc->flags & RDDIRWAIT) {
3907 		rdc->flags &= ~RDDIRWAIT;
3908 		cv_broadcast(&rdc->cv);
3909 	}
3910 	error = rdc->error;
3911 	if (error)
3912 		rdc->flags |= RDDIRREQ;
3913 	mutex_exit(&rp->r_statelock);
3914 
3915 	rddir_cache_rele(rdc);
3916 
3917 	return (error);
3918 }
3919 
3920 static void
3921 nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3922 {
3923 	int error;
3924 	READDIR3args args;
3925 	READDIR3vres res;
3926 	vattr_t dva;
3927 	rnode_t *rp;
3928 	int douprintf;
3929 	failinfo_t fi, *fip = NULL;
3930 	mntinfo_t *mi;
3931 	hrtime_t t;
3932 
3933 	rp = VTOR(vp);
3934 	mi = VTOMI(vp);
3935 	ASSERT(nfs_zone() == mi->mi_zone);
3936 
3937 	args.dir = *RTOFH3(rp);
3938 	args.cookie = (cookie3)rdc->nfs3_cookie;
3939 	args.cookieverf = rp->r_cookieverf;
3940 	args.count = rdc->buflen;
3941 
3942 	/*
3943 	 * NFS client failover support
3944 	 * suppress failover unless we have a zero cookie
3945 	 */
3946 	if (args.cookie == (cookie3) 0) {
3947 		fi.vp = vp;
3948 		fi.fhp = (caddr_t)&args.dir;
3949 		fi.copyproc = nfs3copyfh;
3950 		fi.lookupproc = nfs3lookup;
3951 		fi.xattrdirproc = acl_getxattrdir3;
3952 		fip = &fi;
3953 	}
3954 
3955 #ifdef DEBUG
3956 	rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
3957 #else
3958 	rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
3959 #endif
3960 
3961 	res.entries = (dirent64_t *)rdc->entries;
3962 	res.entries_size = rdc->buflen;
3963 	res.dir_attributes.fres.vap = &dva;
3964 	res.dir_attributes.fres.vp = vp;
3965 	res.loff = rdc->nfs3_cookie;
3966 
3967 	douprintf = 1;
3968 
3969 	if (mi->mi_io_kstats) {
3970 		mutex_enter(&mi->mi_lock);
3971 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
3972 		mutex_exit(&mi->mi_lock);
3973 	}
3974 
3975 	t = gethrtime();
3976 
3977 	error = rfs3call(VTOMI(vp), NFSPROC3_READDIR,
3978 	    xdr_READDIR3args, (caddr_t)&args,
3979 	    xdr_READDIR3vres, (caddr_t)&res, cr,
3980 	    &douprintf, &res.status, 0, fip);
3981 
3982 	if (mi->mi_io_kstats) {
3983 		mutex_enter(&mi->mi_lock);
3984 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
3985 		mutex_exit(&mi->mi_lock);
3986 	}
3987 
3988 	if (error)
3989 		goto err;
3990 
3991 	nfs3_cache_post_op_vattr(vp, &res.dir_attributes, t, cr);
3992 
3993 	error = geterrno3(res.status);
3994 	if (error) {
3995 		PURGE_STALE_FH(error, vp, cr);
3996 		goto err;
3997 	}
3998 
3999 	if (mi->mi_io_kstats) {
4000 		mutex_enter(&mi->mi_lock);
4001 		KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4002 		KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4003 		mutex_exit(&mi->mi_lock);
4004 	}
4005 
4006 	rdc->nfs3_ncookie = res.loff;
4007 	rp->r_cookieverf = res.cookieverf;
4008 	rdc->eof = res.eof ? 1 : 0;
4009 	rdc->entlen = res.size;
4010 	ASSERT(rdc->entlen <= rdc->buflen);
4011 	rdc->error = 0;
4012 	return;
4013 
4014 err:
4015 	kmem_free(rdc->entries, rdc->buflen);
4016 	rdc->entries = NULL;
4017 	rdc->error = error;
4018 }
4019 
4020 /*
4021  * Read directory entries.
4022  * There are some weird things to look out for here.  The uio_loffset
4023  * field is either 0 or it is the offset returned from a previous
4024  * readdir.  It is an opaque value used by the server to find the
4025  * correct directory block to read. The count field is the number
4026  * of blocks to read on the server.  This is advisory only, the server
4027  * may return only one block's worth of entries.  Entries may be compressed
4028  * on the server.
4029  */
4030 static void
4031 nfs3readdirplus(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
4032 {
4033 	int error;
4034 	READDIRPLUS3args args;
4035 	READDIRPLUS3vres res;
4036 	vattr_t dva;
4037 	rnode_t *rp;
4038 	mntinfo_t *mi;
4039 	int douprintf;
4040 	failinfo_t fi, *fip = NULL;
4041 
4042 	rp = VTOR(vp);
4043 	mi = VTOMI(vp);
4044 	ASSERT(nfs_zone() == mi->mi_zone);
4045 
4046 	args.dir = *RTOFH3(rp);
4047 	args.cookie = (cookie3)rdc->nfs3_cookie;
4048 	args.cookieverf = rp->r_cookieverf;
4049 	args.dircount = rdc->buflen;
4050 	args.maxcount = mi->mi_tsize;
4051 
4052 	/*
4053 	 * NFS client failover support
4054 	 * suppress failover unless we have a zero cookie
4055 	 */
4056 	if (args.cookie == (cookie3)0) {
4057 		fi.vp = vp;
4058 		fi.fhp = (caddr_t)&args.dir;
4059 		fi.copyproc = nfs3copyfh;
4060 		fi.lookupproc = nfs3lookup;
4061 		fi.xattrdirproc = acl_getxattrdir3;
4062 		fip = &fi;
4063 	}
4064 
4065 #ifdef DEBUG
4066 	rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
4067 #else
4068 	rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
4069 #endif
4070 
4071 	res.entries = (dirent64_t *)rdc->entries;
4072 	res.entries_size = rdc->buflen;
4073 	res.dir_attributes.fres.vap = &dva;
4074 	res.dir_attributes.fres.vp = vp;
4075 	res.loff = rdc->nfs3_cookie;
4076 	res.credentials = cr;
4077 
4078 	douprintf = 1;
4079 
4080 	if (mi->mi_io_kstats) {
4081 		mutex_enter(&mi->mi_lock);
4082 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
4083 		mutex_exit(&mi->mi_lock);
4084 	}
4085 
4086 	res.time = gethrtime();
4087 
4088 	error = rfs3call(mi, NFSPROC3_READDIRPLUS,
4089 	    xdr_READDIRPLUS3args, (caddr_t)&args,
4090 	    xdr_READDIRPLUS3vres, (caddr_t)&res, cr,
4091 	    &douprintf, &res.status, 0, fip);
4092 
4093 	if (mi->mi_io_kstats) {
4094 		mutex_enter(&mi->mi_lock);
4095 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
4096 		mutex_exit(&mi->mi_lock);
4097 	}
4098 
4099 	if (error) {
4100 		goto err;
4101 	}
4102 
4103 	nfs3_cache_post_op_vattr(vp, &res.dir_attributes, res.time, cr);
4104 
4105 	error = geterrno3(res.status);
4106 	if (error) {
4107 		PURGE_STALE_FH(error, vp, cr);
4108 		if (error == EOPNOTSUPP) {
4109 			mutex_enter(&mi->mi_lock);
4110 			mi->mi_flags |= MI_READDIRONLY;
4111 			mutex_exit(&mi->mi_lock);
4112 		}
4113 		goto err;
4114 	}
4115 
4116 	if (mi->mi_io_kstats) {
4117 		mutex_enter(&mi->mi_lock);
4118 		KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4119 		KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4120 		mutex_exit(&mi->mi_lock);
4121 	}
4122 
4123 	rdc->nfs3_ncookie = res.loff;
4124 	rp->r_cookieverf = res.cookieverf;
4125 	rdc->eof = res.eof ? 1 : 0;
4126 	rdc->entlen = res.size;
4127 	ASSERT(rdc->entlen <= rdc->buflen);
4128 	rdc->error = 0;
4129 
4130 	return;
4131 
4132 err:
4133 	kmem_free(rdc->entries, rdc->buflen);
4134 	rdc->entries = NULL;
4135 	rdc->error = error;
4136 }
4137 
4138 #ifdef DEBUG
4139 static int nfs3_bio_do_stop = 0;
4140 #endif
4141 
4142 static int
4143 nfs3_bio(struct buf *bp, stable_how *stab_comm, cred_t *cr)
4144 {
4145 	rnode_t *rp = VTOR(bp->b_vp);
4146 	int count;
4147 	int error;
4148 	cred_t *cred;
4149 	offset_t offset;
4150 
4151 	ASSERT(nfs_zone() == VTOMI(bp->b_vp)->mi_zone);
4152 	offset = ldbtob(bp->b_lblkno);
4153 
4154 	DTRACE_IO1(start, struct buf *, bp);
4155 
4156 	if (bp->b_flags & B_READ) {
4157 		mutex_enter(&rp->r_statelock);
4158 		if (rp->r_cred != NULL) {
4159 			cred = rp->r_cred;
4160 			crhold(cred);
4161 		} else {
4162 			rp->r_cred = cr;
4163 			crhold(cr);
4164 			cred = cr;
4165 			crhold(cred);
4166 		}
4167 		mutex_exit(&rp->r_statelock);
4168 	read_again:
4169 		error = bp->b_error = nfs3read(bp->b_vp, bp->b_un.b_addr,
4170 		    offset, bp->b_bcount, &bp->b_resid, cred);
4171 		crfree(cred);
4172 		if (!error) {
4173 			if (bp->b_resid) {
4174 				/*
4175 				 * Didn't get it all because we hit EOF,
4176 				 * zero all the memory beyond the EOF.
4177 				 */
4178 				/* bzero(rdaddr + */
4179 				bzero(bp->b_un.b_addr +
4180 				    bp->b_bcount - bp->b_resid, bp->b_resid);
4181 			}
4182 			mutex_enter(&rp->r_statelock);
4183 			if (bp->b_resid == bp->b_bcount &&
4184 			    offset >= rp->r_size) {
4185 				/*
4186 				 * We didn't read anything at all as we are
4187 				 * past EOF.  Return an error indicator back
4188 				 * but don't destroy the pages (yet).
4189 				 */
4190 				error = NFS_EOF;
4191 			}
4192 			mutex_exit(&rp->r_statelock);
4193 		} else if (error == EACCES) {
4194 			mutex_enter(&rp->r_statelock);
4195 			if (cred != cr) {
4196 				if (rp->r_cred != NULL)
4197 					crfree(rp->r_cred);
4198 				rp->r_cred = cr;
4199 				crhold(cr);
4200 				cred = cr;
4201 				crhold(cred);
4202 				mutex_exit(&rp->r_statelock);
4203 				goto read_again;
4204 			}
4205 			mutex_exit(&rp->r_statelock);
4206 		}
4207 	} else {
4208 		if (!(rp->r_flags & RSTALE)) {
4209 			mutex_enter(&rp->r_statelock);
4210 			if (rp->r_cred != NULL) {
4211 				cred = rp->r_cred;
4212 				crhold(cred);
4213 			} else {
4214 				rp->r_cred = cr;
4215 				crhold(cr);
4216 				cred = cr;
4217 				crhold(cred);
4218 			}
4219 			mutex_exit(&rp->r_statelock);
4220 		write_again:
4221 			mutex_enter(&rp->r_statelock);
4222 			count = MIN(bp->b_bcount, rp->r_size - offset);
4223 			mutex_exit(&rp->r_statelock);
4224 			if (count < 0)
4225 				cmn_err(CE_PANIC, "nfs3_bio: write count < 0");
4226 #ifdef DEBUG
4227 			if (count == 0) {
4228 				zcmn_err(getzoneid(), CE_WARN,
4229 				    "nfs3_bio: zero length write at %lld",
4230 				    offset);
4231 				nfs_printfhandle(&rp->r_fh);
4232 				if (nfs3_bio_do_stop)
4233 					debug_enter("nfs3_bio");
4234 			}
4235 #endif
4236 			error = nfs3write(bp->b_vp, bp->b_un.b_addr, offset,
4237 			    count, cred, stab_comm);
4238 			if (error == EACCES) {
4239 				mutex_enter(&rp->r_statelock);
4240 				if (cred != cr) {
4241 					if (rp->r_cred != NULL)
4242 						crfree(rp->r_cred);
4243 					rp->r_cred = cr;
4244 					crhold(cr);
4245 					crfree(cred);
4246 					cred = cr;
4247 					crhold(cred);
4248 					mutex_exit(&rp->r_statelock);
4249 					goto write_again;
4250 				}
4251 				mutex_exit(&rp->r_statelock);
4252 			}
4253 			bp->b_error = error;
4254 			if (error && error != EINTR) {
4255 				/*
4256 				 * Don't print EDQUOT errors on the console.
4257 				 * Don't print asynchronous EACCES errors.
4258 				 * Don't print EFBIG errors.
4259 				 * Print all other write errors.
4260 				 */
4261 				if (error != EDQUOT && error != EFBIG &&
4262 				    (error != EACCES ||
4263 				    !(bp->b_flags & B_ASYNC)))
4264 					nfs_write_error(bp->b_vp, error, cred);
4265 				/*
4266 				 * Update r_error and r_flags as appropriate.
4267 				 * If the error was ESTALE, then mark the
4268 				 * rnode as not being writeable and save
4269 				 * the error status.  Otherwise, save any
4270 				 * errors which occur from asynchronous
4271 				 * page invalidations.  Any errors occurring
4272 				 * from other operations should be saved
4273 				 * by the caller.
4274 				 */
4275 				mutex_enter(&rp->r_statelock);
4276 				if (error == ESTALE) {
4277 					rp->r_flags |= RSTALE;
4278 					if (!rp->r_error)
4279 						rp->r_error = error;
4280 				} else if (!rp->r_error &&
4281 				    (bp->b_flags &
4282 				    (B_INVAL|B_FORCE|B_ASYNC)) ==
4283 				    (B_INVAL|B_FORCE|B_ASYNC)) {
4284 					rp->r_error = error;
4285 				}
4286 				mutex_exit(&rp->r_statelock);
4287 			}
4288 			crfree(cred);
4289 		} else
4290 			error = rp->r_error;
4291 	}
4292 
4293 	if (error != 0 && error != NFS_EOF)
4294 		bp->b_flags |= B_ERROR;
4295 
4296 	DTRACE_IO1(done, struct buf *, bp);
4297 
4298 	return (error);
4299 }
4300 
4301 static int
4302 nfs3_fid(vnode_t *vp, fid_t *fidp)
4303 {
4304 	rnode_t *rp;
4305 
4306 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4307 		return (EIO);
4308 	rp = VTOR(vp);
4309 
4310 	if (fidp->fid_len < (ushort_t)rp->r_fh.fh_len) {
4311 		fidp->fid_len = rp->r_fh.fh_len;
4312 		return (ENOSPC);
4313 	}
4314 	fidp->fid_len = rp->r_fh.fh_len;
4315 	bcopy(rp->r_fh.fh_buf, fidp->fid_data, fidp->fid_len);
4316 	return (0);
4317 }
4318 
4319 /* ARGSUSED2 */
4320 static int
4321 nfs3_rwlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4322 {
4323 	rnode_t *rp = VTOR(vp);
4324 
4325 	if (!write_lock) {
4326 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4327 		return (V_WRITELOCK_FALSE);
4328 	}
4329 
4330 	if ((rp->r_flags & RDIRECTIO) || (VTOMI(vp)->mi_flags & MI_DIRECTIO)) {
4331 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4332 		if (rp->r_mapcnt == 0 && !vn_has_cached_data(vp))
4333 			return (V_WRITELOCK_FALSE);
4334 		nfs_rw_exit(&rp->r_rwlock);
4335 	}
4336 
4337 	(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, FALSE);
4338 	return (V_WRITELOCK_TRUE);
4339 }
4340 
4341 /* ARGSUSED */
4342 static void
4343 nfs3_rwunlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4344 {
4345 	rnode_t *rp = VTOR(vp);
4346 
4347 	nfs_rw_exit(&rp->r_rwlock);
4348 }
4349 
4350 /* ARGSUSED */
4351 static int
4352 nfs3_seek(vnode_t *vp, offset_t ooff, offset_t *noffp)
4353 {
4354 
4355 	/*
4356 	 * Because we stuff the readdir cookie into the offset field
4357 	 * someone may attempt to do an lseek with the cookie which
4358 	 * we want to succeed.
4359 	 */
4360 	if (vp->v_type == VDIR)
4361 		return (0);
4362 	if (*noffp < 0)
4363 		return (EINVAL);
4364 	return (0);
4365 }
4366 
4367 /*
4368  * number of nfs3_bsize blocks to read ahead.
4369  */
4370 static int nfs3_nra = 4;
4371 
4372 #ifdef DEBUG
4373 static int nfs3_lostpage = 0;	/* number of times we lost original page */
4374 #endif
4375 
4376 /*
4377  * Return all the pages from [off..off+len) in file
4378  */
4379 static int
4380 nfs3_getpage(vnode_t *vp, offset_t off, size_t len, uint_t *protp,
4381 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4382 	enum seg_rw rw, cred_t *cr)
4383 {
4384 	rnode_t *rp;
4385 	int error;
4386 	mntinfo_t *mi;
4387 
4388 	if (vp->v_flag & VNOMAP)
4389 		return (ENOSYS);
4390 
4391 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4392 		return (EIO);
4393 	if (protp != NULL)
4394 		*protp = PROT_ALL;
4395 
4396 	/*
4397 	 * Now valididate that the caches are up to date.
4398 	 */
4399 	error = nfs3_validate_caches(vp, cr);
4400 	if (error)
4401 		return (error);
4402 
4403 	rp = VTOR(vp);
4404 	mi = VTOMI(vp);
4405 retry:
4406 	mutex_enter(&rp->r_statelock);
4407 
4408 	/*
4409 	 * Don't create dirty pages faster than they
4410 	 * can be cleaned so that the system doesn't
4411 	 * get imbalanced.  If the async queue is
4412 	 * maxed out, then wait for it to drain before
4413 	 * creating more dirty pages.  Also, wait for
4414 	 * any threads doing pagewalks in the vop_getattr
4415 	 * entry points so that they don't block for
4416 	 * long periods.
4417 	 */
4418 	if (rw == S_CREATE) {
4419 		while ((mi->mi_max_threads != 0 &&
4420 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
4421 		    rp->r_gcount > 0)
4422 			cv_wait(&rp->r_cv, &rp->r_statelock);
4423 	}
4424 
4425 	/*
4426 	 * If we are getting called as a side effect of an nfs_write()
4427 	 * operation the local file size might not be extended yet.
4428 	 * In this case we want to be able to return pages of zeroes.
4429 	 */
4430 	if (off + len > rp->r_size + PAGEOFFSET && seg != segkmap) {
4431 		mutex_exit(&rp->r_statelock);
4432 		return (EFAULT);		/* beyond EOF */
4433 	}
4434 
4435 	mutex_exit(&rp->r_statelock);
4436 
4437 	if (len <= PAGESIZE) {
4438 		error = nfs3_getapage(vp, off, len, protp, pl, plsz,
4439 		    seg, addr, rw, cr);
4440 	} else {
4441 		error = pvn_getpages(nfs3_getapage, vp, off, len, protp,
4442 		    pl, plsz, seg, addr, rw, cr);
4443 	}
4444 
4445 	switch (error) {
4446 	case NFS_EOF:
4447 		nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
4448 		goto retry;
4449 	case ESTALE:
4450 		PURGE_STALE_FH(error, vp, cr);
4451 	}
4452 
4453 	return (error);
4454 }
4455 
4456 /*
4457  * Called from pvn_getpages or nfs3_getpage to get a particular page.
4458  */
4459 /* ARGSUSED */
4460 static int
4461 nfs3_getapage(vnode_t *vp, u_offset_t off, size_t len, uint_t *protp,
4462 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4463 	enum seg_rw rw, cred_t *cr)
4464 {
4465 	rnode_t *rp;
4466 	uint_t bsize;
4467 	struct buf *bp;
4468 	page_t *pp;
4469 	u_offset_t lbn;
4470 	u_offset_t io_off;
4471 	u_offset_t blkoff;
4472 	u_offset_t rablkoff;
4473 	size_t io_len;
4474 	uint_t blksize;
4475 	int error;
4476 	int readahead;
4477 	int readahead_issued = 0;
4478 	int ra_window; /* readahead window */
4479 	page_t *pagefound;
4480 	page_t *savepp;
4481 
4482 	if (nfs_zone() != VTOMI(vp)->mi_zone)
4483 		return (EIO);
4484 	rp = VTOR(vp);
4485 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4486 
4487 reread:
4488 	bp = NULL;
4489 	pp = NULL;
4490 	pagefound = NULL;
4491 
4492 	if (pl != NULL)
4493 		pl[0] = NULL;
4494 
4495 	error = 0;
4496 	lbn = off / bsize;
4497 	blkoff = lbn * bsize;
4498 
4499 	/*
4500 	 * Queueing up the readahead before doing the synchronous read
4501 	 * results in a significant increase in read throughput because
4502 	 * of the increased parallelism between the async threads and
4503 	 * the process context.
4504 	 */
4505 	if ((off & ((vp->v_vfsp->vfs_bsize) - 1)) == 0 &&
4506 	    rw != S_CREATE &&
4507 	    !(vp->v_flag & VNOCACHE)) {
4508 		mutex_enter(&rp->r_statelock);
4509 
4510 		/*
4511 		 * Calculate the number of readaheads to do.
4512 		 * a) No readaheads at offset = 0.
4513 		 * b) Do maximum(nfs3_nra) readaheads when the readahead
4514 		 *    window is closed.
4515 		 * c) Do readaheads between 1 to (nfs3_nra - 1) depending
4516 		 *    upon how far the readahead window is open or close.
4517 		 * d) No readaheads if rp->r_nextr is not within the scope
4518 		 *    of the readahead window (random i/o).
4519 		 */
4520 
4521 		if (off == 0)
4522 			readahead = 0;
4523 		else if (blkoff == rp->r_nextr)
4524 			readahead = nfs3_nra;
4525 		else if (rp->r_nextr > blkoff &&
4526 				((ra_window = (rp->r_nextr - blkoff) / bsize)
4527 					<= (nfs3_nra - 1)))
4528 			readahead = nfs3_nra - ra_window;
4529 		else
4530 			readahead = 0;
4531 
4532 		rablkoff = rp->r_nextr;
4533 		while (readahead > 0 && rablkoff + bsize < rp->r_size) {
4534 			mutex_exit(&rp->r_statelock);
4535 			if (nfs_async_readahead(vp, rablkoff + bsize,
4536 			    addr + (rablkoff + bsize - off), seg, cr,
4537 			    nfs3_readahead) < 0) {
4538 				mutex_enter(&rp->r_statelock);
4539 				break;
4540 			}
4541 			readahead--;
4542 			rablkoff += bsize;
4543 			/*
4544 			 * Indicate that we did a readahead so
4545 			 * readahead offset is not updated
4546 			 * by the synchronous read below.
4547 			 */
4548 			readahead_issued = 1;
4549 			mutex_enter(&rp->r_statelock);
4550 			/*
4551 			 * set readahead offset to
4552 			 * offset of last async readahead
4553 			 * request.
4554 			 */
4555 			rp->r_nextr = rablkoff;
4556 		}
4557 		mutex_exit(&rp->r_statelock);
4558 	}
4559 
4560 again:
4561 	if ((pagefound = page_exists(vp, off)) == NULL) {
4562 		if (pl == NULL) {
4563 			(void) nfs_async_readahead(vp, blkoff, addr, seg, cr,
4564 			    nfs3_readahead);
4565 		} else if (rw == S_CREATE) {
4566 			/*
4567 			 * Block for this page is not allocated, or the offset
4568 			 * is beyond the current allocation size, or we're
4569 			 * allocating a swap slot and the page was not found,
4570 			 * so allocate it and return a zero page.
4571 			 */
4572 			if ((pp = page_create_va(vp, off,
4573 			    PAGESIZE, PG_WAIT, seg, addr)) == NULL)
4574 				cmn_err(CE_PANIC, "nfs3_getapage: page_create");
4575 			io_len = PAGESIZE;
4576 			mutex_enter(&rp->r_statelock);
4577 			rp->r_nextr = off + PAGESIZE;
4578 			mutex_exit(&rp->r_statelock);
4579 		} else {
4580 			/*
4581 			 * Need to go to server to get a BLOCK, exception to
4582 			 * that being while reading at offset = 0 or doing
4583 			 * random i/o, in that case read only a PAGE.
4584 			 */
4585 			mutex_enter(&rp->r_statelock);
4586 			if (blkoff < rp->r_size &&
4587 			    blkoff + bsize >= rp->r_size) {
4588 				/*
4589 				 * If only a block or less is left in
4590 				 * the file, read all that is remaining.
4591 				 */
4592 				if (rp->r_size <= off) {
4593 					/*
4594 					 * Trying to access beyond EOF,
4595 					 * set up to get at least one page.
4596 					 */
4597 					blksize = off + PAGESIZE - blkoff;
4598 				} else
4599 					blksize = rp->r_size - blkoff;
4600 			} else if ((off == 0) ||
4601 				(off != rp->r_nextr && !readahead_issued)) {
4602 				blksize = PAGESIZE;
4603 				blkoff = off; /* block = page here */
4604 			} else
4605 				blksize = bsize;
4606 			mutex_exit(&rp->r_statelock);
4607 
4608 			pp = pvn_read_kluster(vp, off, seg, addr, &io_off,
4609 			    &io_len, blkoff, blksize, 0);
4610 
4611 			/*
4612 			 * Some other thread has entered the page,
4613 			 * so just use it.
4614 			 */
4615 			if (pp == NULL)
4616 				goto again;
4617 
4618 			/*
4619 			 * Now round the request size up to page boundaries.
4620 			 * This ensures that the entire page will be
4621 			 * initialized to zeroes if EOF is encountered.
4622 			 */
4623 			io_len = ptob(btopr(io_len));
4624 
4625 			bp = pageio_setup(pp, io_len, vp, B_READ);
4626 			ASSERT(bp != NULL);
4627 
4628 			/*
4629 			 * pageio_setup should have set b_addr to 0.  This
4630 			 * is correct since we want to do I/O on a page
4631 			 * boundary.  bp_mapin will use this addr to calculate
4632 			 * an offset, and then set b_addr to the kernel virtual
4633 			 * address it allocated for us.
4634 			 */
4635 			ASSERT(bp->b_un.b_addr == 0);
4636 
4637 			bp->b_edev = 0;
4638 			bp->b_dev = 0;
4639 			bp->b_lblkno = lbtodb(io_off);
4640 			bp->b_file = vp;
4641 			bp->b_offset = (offset_t)off;
4642 			bp_mapin(bp);
4643 
4644 			/*
4645 			 * If doing a write beyond what we believe is EOF,
4646 			 * don't bother trying to read the pages from the
4647 			 * server, we'll just zero the pages here.  We
4648 			 * don't check that the rw flag is S_WRITE here
4649 			 * because some implementations may attempt a
4650 			 * read access to the buffer before copying data.
4651 			 */
4652 			mutex_enter(&rp->r_statelock);
4653 			if (io_off >= rp->r_size && seg == segkmap) {
4654 				mutex_exit(&rp->r_statelock);
4655 				bzero(bp->b_un.b_addr, io_len);
4656 			} else {
4657 				mutex_exit(&rp->r_statelock);
4658 				error = nfs3_bio(bp, NULL, cr);
4659 			}
4660 
4661 			/*
4662 			 * Unmap the buffer before freeing it.
4663 			 */
4664 			bp_mapout(bp);
4665 			pageio_done(bp);
4666 
4667 			savepp = pp;
4668 			do {
4669 				pp->p_fsdata = C_NOCOMMIT;
4670 			} while ((pp = pp->p_next) != savepp);
4671 
4672 			if (error == NFS_EOF) {
4673 				/*
4674 				 * If doing a write system call just return
4675 				 * zeroed pages, else user tried to get pages
4676 				 * beyond EOF, return error.  We don't check
4677 				 * that the rw flag is S_WRITE here because
4678 				 * some implementations may attempt a read
4679 				 * access to the buffer before copying data.
4680 				 */
4681 				if (seg == segkmap)
4682 					error = 0;
4683 				else
4684 					error = EFAULT;
4685 			}
4686 
4687 			if (!readahead_issued && !error) {
4688 			    mutex_enter(&rp->r_statelock);
4689 			    rp->r_nextr = io_off + io_len;
4690 			    mutex_exit(&rp->r_statelock);
4691 			}
4692 		}
4693 	}
4694 
4695 out:
4696 	if (pl == NULL)
4697 		return (error);
4698 
4699 	if (error) {
4700 		if (pp != NULL)
4701 			pvn_read_done(pp, B_ERROR);
4702 		return (error);
4703 	}
4704 
4705 	if (pagefound) {
4706 		se_t se = (rw == S_CREATE ? SE_EXCL : SE_SHARED);
4707 
4708 		/*
4709 		 * Page exists in the cache, acquire the appropriate lock.
4710 		 * If this fails, start all over again.
4711 		 */
4712 		if ((pp = page_lookup(vp, off, se)) == NULL) {
4713 #ifdef DEBUG
4714 			nfs3_lostpage++;
4715 #endif
4716 			goto reread;
4717 		}
4718 		pl[0] = pp;
4719 		pl[1] = NULL;
4720 		return (0);
4721 	}
4722 
4723 	if (pp != NULL)
4724 		pvn_plist_init(pp, pl, plsz, off, io_len, rw);
4725 
4726 	return (error);
4727 }
4728 
4729 static void
4730 nfs3_readahead(vnode_t *vp, u_offset_t blkoff, caddr_t addr, struct seg *seg,
4731 	cred_t *cr)
4732 {
4733 	int error;
4734 	page_t *pp;
4735 	u_offset_t io_off;
4736 	size_t io_len;
4737 	struct buf *bp;
4738 	uint_t bsize, blksize;
4739 	rnode_t *rp = VTOR(vp);
4740 	page_t *savepp;
4741 
4742 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
4743 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4744 
4745 	mutex_enter(&rp->r_statelock);
4746 	if (blkoff < rp->r_size && blkoff + bsize > rp->r_size) {
4747 		/*
4748 		 * If less than a block left in file read less
4749 		 * than a block.
4750 		 */
4751 		blksize = rp->r_size - blkoff;
4752 	} else
4753 		blksize = bsize;
4754 	mutex_exit(&rp->r_statelock);
4755 
4756 	pp = pvn_read_kluster(vp, blkoff, segkmap, addr,
4757 	    &io_off, &io_len, blkoff, blksize, 1);
4758 	/*
4759 	 * The isra flag passed to the kluster function is 1, we may have
4760 	 * gotten a return value of NULL for a variety of reasons (# of free
4761 	 * pages < minfree, someone entered the page on the vnode etc). In all
4762 	 * cases, we want to punt on the readahead.
4763 	 */
4764 	if (pp == NULL)
4765 		return;
4766 
4767 	/*
4768 	 * Now round the request size up to page boundaries.
4769 	 * This ensures that the entire page will be
4770 	 * initialized to zeroes if EOF is encountered.
4771 	 */
4772 	io_len = ptob(btopr(io_len));
4773 
4774 	bp = pageio_setup(pp, io_len, vp, B_READ);
4775 	ASSERT(bp != NULL);
4776 
4777 	/*
4778 	 * pageio_setup should have set b_addr to 0.  This is correct since
4779 	 * we want to do I/O on a page boundary. bp_mapin() will use this addr
4780 	 * to calculate an offset, and then set b_addr to the kernel virtual
4781 	 * address it allocated for us.
4782 	 */
4783 	ASSERT(bp->b_un.b_addr == 0);
4784 
4785 	bp->b_edev = 0;
4786 	bp->b_dev = 0;
4787 	bp->b_lblkno = lbtodb(io_off);
4788 	bp->b_file = vp;
4789 	bp->b_offset = (offset_t)blkoff;
4790 	bp_mapin(bp);
4791 
4792 	/*
4793 	 * If doing a write beyond what we believe is EOF, don't bother trying
4794 	 * to read the pages from the server, we'll just zero the pages here.
4795 	 * We don't check that the rw flag is S_WRITE here because some
4796 	 * implementations may attempt a read access to the buffer before
4797 	 * copying data.
4798 	 */
4799 	mutex_enter(&rp->r_statelock);
4800 	if (io_off >= rp->r_size && seg == segkmap) {
4801 		mutex_exit(&rp->r_statelock);
4802 		bzero(bp->b_un.b_addr, io_len);
4803 		error = 0;
4804 	} else {
4805 		mutex_exit(&rp->r_statelock);
4806 		error = nfs3_bio(bp, NULL, cr);
4807 		if (error == NFS_EOF)
4808 			error = 0;
4809 	}
4810 
4811 	/*
4812 	 * Unmap the buffer before freeing it.
4813 	 */
4814 	bp_mapout(bp);
4815 	pageio_done(bp);
4816 
4817 	savepp = pp;
4818 	do {
4819 		pp->p_fsdata = C_NOCOMMIT;
4820 	} while ((pp = pp->p_next) != savepp);
4821 
4822 	pvn_read_done(pp, error ? B_READ | B_ERROR : B_READ);
4823 
4824 	/*
4825 	 * In case of error set readahead offset
4826 	 * to the lowest offset.
4827 	 * pvn_read_done() calls VN_DISPOSE to destroy the pages
4828 	 */
4829 	if (error && rp->r_nextr > io_off) {
4830 		mutex_enter(&rp->r_statelock);
4831 		if (rp->r_nextr > io_off)
4832 			rp->r_nextr = io_off;
4833 		mutex_exit(&rp->r_statelock);
4834 	}
4835 }
4836 
4837 /*
4838  * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
4839  * If len == 0, do from off to EOF.
4840  *
4841  * The normal cases should be len == 0 && off == 0 (entire vp list),
4842  * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
4843  * (from pageout).
4844  */
4845 static int
4846 nfs3_putpage(vnode_t *vp, offset_t off, size_t len, int flags, cred_t *cr)
4847 {
4848 	int error;
4849 	rnode_t *rp;
4850 
4851 	ASSERT(cr != NULL);
4852 
4853 	/*
4854 	 * XXX - Why should this check be made here?
4855 	 */
4856 	if (vp->v_flag & VNOMAP)
4857 		return (ENOSYS);
4858 	if (len == 0 && !(flags & B_INVAL) && vn_is_readonly(vp))
4859 		return (0);
4860 	if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
4861 		return (EIO);
4862 
4863 	rp = VTOR(vp);
4864 	mutex_enter(&rp->r_statelock);
4865 	rp->r_count++;
4866 	mutex_exit(&rp->r_statelock);
4867 	error = nfs_putpages(vp, off, len, flags, cr);
4868 	mutex_enter(&rp->r_statelock);
4869 	rp->r_count--;
4870 	cv_broadcast(&rp->r_cv);
4871 	mutex_exit(&rp->r_statelock);
4872 
4873 	return (error);
4874 }
4875 
4876 /*
4877  * Write out a single page, possibly klustering adjacent dirty pages.
4878  */
4879 int
4880 nfs3_putapage(vnode_t *vp, page_t *pp, u_offset_t *offp, size_t *lenp,
4881 	int flags, cred_t *cr)
4882 {
4883 	u_offset_t io_off;
4884 	u_offset_t lbn_off;
4885 	u_offset_t lbn;
4886 	size_t io_len;
4887 	uint_t bsize;
4888 	int error;
4889 	rnode_t *rp;
4890 
4891 	ASSERT(!vn_is_readonly(vp));
4892 	ASSERT(pp != NULL);
4893 	ASSERT(cr != NULL);
4894 	ASSERT((flags & B_ASYNC) || nfs_zone() == VTOMI(vp)->mi_zone);
4895 
4896 	rp = VTOR(vp);
4897 	ASSERT(rp->r_count > 0);
4898 
4899 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4900 	lbn = pp->p_offset / bsize;
4901 	lbn_off = lbn * bsize;
4902 
4903 	/*
4904 	 * Find a kluster that fits in one block, or in
4905 	 * one page if pages are bigger than blocks.  If
4906 	 * there is less file space allocated than a whole
4907 	 * page, we'll shorten the i/o request below.
4908 	 */
4909 	pp = pvn_write_kluster(vp, pp, &io_off, &io_len, lbn_off,
4910 	    roundup(bsize, PAGESIZE), flags);
4911 
4912 	/*
4913 	 * pvn_write_kluster shouldn't have returned a page with offset
4914 	 * behind the original page we were given.  Verify that.
4915 	 */
4916 	ASSERT((pp->p_offset / bsize) >= lbn);
4917 
4918 	/*
4919 	 * Now pp will have the list of kept dirty pages marked for
4920 	 * write back.  It will also handle invalidation and freeing
4921 	 * of pages that are not dirty.  Check for page length rounding
4922 	 * problems.
4923 	 */
4924 	if (io_off + io_len > lbn_off + bsize) {
4925 		ASSERT((io_off + io_len) - (lbn_off + bsize) < PAGESIZE);
4926 		io_len = lbn_off + bsize - io_off;
4927 	}
4928 	/*
4929 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4930 	 * consistent value of r_size. RMODINPROGRESS is set in writerp().
4931 	 * When RMODINPROGRESS is set it indicates that a uiomove() is in
4932 	 * progress and the r_size has not been made consistent with the
4933 	 * new size of the file. When the uiomove() completes the r_size is
4934 	 * updated and the RMODINPROGRESS flag is cleared.
4935 	 *
4936 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4937 	 * consistent value of r_size. Without this handshaking, it is
4938 	 * possible that nfs(3)_bio() picks  up the old value of r_size
4939 	 * before the uiomove() in writerp() completes. This will result
4940 	 * in the write through nfs(3)_bio() being dropped.
4941 	 *
4942 	 * More precisely, there is a window between the time the uiomove()
4943 	 * completes and the time the r_size is updated. If a VOP_PUTPAGE()
4944 	 * operation intervenes in this window, the page will be picked up,
4945 	 * because it is dirty (it will be unlocked, unless it was
4946 	 * pagecreate'd). When the page is picked up as dirty, the dirty
4947 	 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
4948 	 * checked. This will still be the old size. Therefore the page will
4949 	 * not be written out. When segmap_release() calls VOP_PUTPAGE(),
4950 	 * the page will be found to be clean and the write will be dropped.
4951 	 */
4952 	if (rp->r_flags & RMODINPROGRESS) {
4953 		mutex_enter(&rp->r_statelock);
4954 		if ((rp->r_flags & RMODINPROGRESS) &&
4955 		    rp->r_modaddr + MAXBSIZE > io_off &&
4956 		    rp->r_modaddr < io_off + io_len) {
4957 			page_t *plist;
4958 			/*
4959 			 * A write is in progress for this region of the file.
4960 			 * If we did not detect RMODINPROGRESS here then this
4961 			 * path through nfs_putapage() would eventually go to
4962 			 * nfs(3)_bio() and may not write out all of the data
4963 			 * in the pages. We end up losing data. So we decide
4964 			 * to set the modified bit on each page in the page
4965 			 * list and mark the rnode with RDIRTY. This write
4966 			 * will be restarted at some later time.
4967 			 */
4968 			plist = pp;
4969 			while (plist != NULL) {
4970 				pp = plist;
4971 				page_sub(&plist, pp);
4972 				hat_setmod(pp);
4973 				page_io_unlock(pp);
4974 				page_unlock(pp);
4975 			}
4976 			rp->r_flags |= RDIRTY;
4977 			mutex_exit(&rp->r_statelock);
4978 			if (offp)
4979 				*offp = io_off;
4980 			if (lenp)
4981 				*lenp = io_len;
4982 			return (0);
4983 		}
4984 		mutex_exit(&rp->r_statelock);
4985 	}
4986 
4987 	if (flags & B_ASYNC) {
4988 		error = nfs_async_putapage(vp, pp, io_off, io_len, flags, cr,
4989 		    nfs3_sync_putapage);
4990 	} else
4991 		error = nfs3_sync_putapage(vp, pp, io_off, io_len, flags, cr);
4992 
4993 	if (offp)
4994 		*offp = io_off;
4995 	if (lenp)
4996 		*lenp = io_len;
4997 	return (error);
4998 }
4999 
5000 static int
5001 nfs3_sync_putapage(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5002 	int flags, cred_t *cr)
5003 {
5004 	int error;
5005 	rnode_t *rp;
5006 
5007 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5008 
5009 	flags |= B_WRITE;
5010 
5011 	error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5012 
5013 	rp = VTOR(vp);
5014 
5015 	if ((error == ENOSPC || error == EDQUOT || error == EFBIG ||
5016 	    error == EACCES) &&
5017 	    (flags & (B_INVAL|B_FORCE)) != (B_INVAL|B_FORCE)) {
5018 		if (!(rp->r_flags & ROUTOFSPACE)) {
5019 			mutex_enter(&rp->r_statelock);
5020 			rp->r_flags |= ROUTOFSPACE;
5021 			mutex_exit(&rp->r_statelock);
5022 		}
5023 		flags |= B_ERROR;
5024 		pvn_write_done(pp, flags);
5025 		/*
5026 		 * If this was not an async thread, then try again to
5027 		 * write out the pages, but this time, also destroy
5028 		 * them whether or not the write is successful.  This
5029 		 * will prevent memory from filling up with these
5030 		 * pages and destroying them is the only alternative
5031 		 * if they can't be written out.
5032 		 *
5033 		 * Don't do this if this is an async thread because
5034 		 * when the pages are unlocked in pvn_write_done,
5035 		 * some other thread could have come along, locked
5036 		 * them, and queued for an async thread.  It would be
5037 		 * possible for all of the async threads to be tied
5038 		 * up waiting to lock the pages again and they would
5039 		 * all already be locked and waiting for an async
5040 		 * thread to handle them.  Deadlock.
5041 		 */
5042 		if (!(flags & B_ASYNC)) {
5043 			error = nfs3_putpage(vp, io_off, io_len,
5044 			    B_INVAL | B_FORCE, cr);
5045 		}
5046 	} else {
5047 		if (error)
5048 			flags |= B_ERROR;
5049 		else if (rp->r_flags & ROUTOFSPACE) {
5050 			mutex_enter(&rp->r_statelock);
5051 			rp->r_flags &= ~ROUTOFSPACE;
5052 			mutex_exit(&rp->r_statelock);
5053 		}
5054 		pvn_write_done(pp, flags);
5055 		if (freemem < desfree)
5056 			(void) nfs3_commit_vp(vp, (u_offset_t)0, 0, cr);
5057 	}
5058 
5059 	return (error);
5060 }
5061 
5062 static int
5063 nfs3_map(vnode_t *vp, offset_t off, struct as *as, caddr_t *addrp,
5064 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
5065 {
5066 	struct segvn_crargs vn_a;
5067 	int error;
5068 	rnode_t *rp;
5069 	struct vattr va;
5070 
5071 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5072 		return (EIO);
5073 
5074 	if (vp->v_flag & VNOMAP)
5075 		return (ENOSYS);
5076 
5077 	if (off < 0 || off + len < 0)
5078 		return (ENXIO);
5079 
5080 	if (vp->v_type != VREG)
5081 		return (ENODEV);
5082 
5083 	/*
5084 	 * If there is cached data and if close-to-open consistency
5085 	 * checking is not turned off and if the file system is not
5086 	 * mounted readonly, then force an over the wire getattr.
5087 	 * Otherwise, just invoke nfs3getattr to get a copy of the
5088 	 * attributes.  The attribute cache will be used unless it
5089 	 * is timed out and if it is, then an over the wire getattr
5090 	 * will be issued.
5091 	 */
5092 	va.va_mask = AT_ALL;
5093 	if (vn_has_cached_data(vp) &&
5094 	    !(VTOMI(vp)->mi_flags & MI_NOCTO) && !vn_is_readonly(vp))
5095 		error = nfs3_getattr_otw(vp, &va, cr);
5096 	else
5097 		error = nfs3getattr(vp, &va, cr);
5098 	if (error)
5099 		return (error);
5100 
5101 	/*
5102 	 * Check to see if the vnode is currently marked as not cachable.
5103 	 * This means portions of the file are locked (through VOP_FRLOCK).
5104 	 * In this case the map request must be refused.  We use
5105 	 * rp->r_lkserlock to avoid a race with concurrent lock requests.
5106 	 */
5107 	rp = VTOR(vp);
5108 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
5109 		return (EINTR);
5110 
5111 	if (vp->v_flag & VNOCACHE) {
5112 		error = EAGAIN;
5113 		goto done;
5114 	}
5115 
5116 	/*
5117 	 * Don't allow concurrent locks and mapping if mandatory locking is
5118 	 * enabled.
5119 	 */
5120 	if ((flk_has_remote_locks(vp) || lm_has_sleep(vp)) &&
5121 	    MANDLOCK(vp, va.va_mode)) {
5122 		error = EAGAIN;
5123 		goto done;
5124 	}
5125 
5126 	as_rangelock(as);
5127 	if (!(flags & MAP_FIXED)) {
5128 		map_addr(addrp, len, off, 1, flags);
5129 		if (*addrp == NULL) {
5130 			as_rangeunlock(as);
5131 			error = ENOMEM;
5132 			goto done;
5133 		}
5134 	} else {
5135 		/*
5136 		 * User specified address - blow away any previous mappings
5137 		 */
5138 		(void) as_unmap(as, *addrp, len);
5139 	}
5140 
5141 	vn_a.vp = vp;
5142 	vn_a.offset = off;
5143 	vn_a.type = (flags & MAP_TYPE);
5144 	vn_a.prot = (uchar_t)prot;
5145 	vn_a.maxprot = (uchar_t)maxprot;
5146 	vn_a.flags = (flags & ~MAP_TYPE);
5147 	vn_a.cred = cr;
5148 	vn_a.amp = NULL;
5149 	vn_a.szc = 0;
5150 	vn_a.lgrp_mem_policy_flags = 0;
5151 
5152 	error = as_map(as, *addrp, len, segvn_create, &vn_a);
5153 	as_rangeunlock(as);
5154 
5155 done:
5156 	nfs_rw_exit(&rp->r_lkserlock);
5157 	return (error);
5158 }
5159 
5160 /* ARGSUSED */
5161 static int
5162 nfs3_addmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5163 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
5164 {
5165 	rnode_t *rp;
5166 
5167 	if (vp->v_flag & VNOMAP)
5168 		return (ENOSYS);
5169 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5170 		return (EIO);
5171 
5172 	/*
5173 	 * Need to hold rwlock while incrementing the mapcnt so that
5174 	 * mmap'ing can be serialized with writes so that the caching
5175 	 * can be handled correctly.
5176 	 */
5177 	rp = VTOR(vp);
5178 	if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, INTR(vp)))
5179 		return (EINTR);
5180 	atomic_add_long((ulong_t *)&rp->r_mapcnt, btopr(len));
5181 	nfs_rw_exit(&rp->r_rwlock);
5182 
5183 	return (0);
5184 }
5185 
5186 static int
5187 nfs3_frlock(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5188 	offset_t offset, struct flk_callback *flk_cbp, cred_t *cr)
5189 {
5190 	netobj lm_fh3;
5191 	int rc;
5192 	u_offset_t start, end;
5193 	rnode_t *rp;
5194 	int error = 0, intr = INTR(vp);
5195 
5196 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5197 		return (EIO);
5198 	/* check for valid cmd parameter */
5199 	if (cmd != F_GETLK && cmd != F_SETLK && cmd != F_SETLKW)
5200 		return (EINVAL);
5201 
5202 	/* Verify l_type. */
5203 	switch (bfp->l_type) {
5204 	case F_RDLCK:
5205 		if (cmd != F_GETLK && !(flag & FREAD))
5206 			return (EBADF);
5207 		break;
5208 	case F_WRLCK:
5209 		if (cmd != F_GETLK && !(flag & FWRITE))
5210 			return (EBADF);
5211 		break;
5212 	case F_UNLCK:
5213 		intr = 0;
5214 		break;
5215 
5216 	default:
5217 		return (EINVAL);
5218 	}
5219 
5220 	/* check the validity of the lock range */
5221 	if (rc = flk_convert_lock_data(vp, bfp, &start, &end, offset))
5222 		return (rc);
5223 	if (rc = flk_check_lock_data(start, end, MAXEND))
5224 		return (rc);
5225 
5226 	/*
5227 	 * If the filesystem is mounted using local locking, pass the
5228 	 * request off to the local locking code.
5229 	 */
5230 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
5231 		if (cmd == F_SETLK || cmd == F_SETLKW) {
5232 			/*
5233 			 * For complete safety, we should be holding
5234 			 * r_lkserlock.  However, we can't call
5235 			 * lm_safelock and then fs_frlock while
5236 			 * holding r_lkserlock, so just invoke
5237 			 * lm_safelock and expect that this will
5238 			 * catch enough of the cases.
5239 			 */
5240 			if (!lm_safelock(vp, bfp, cr))
5241 				return (EAGAIN);
5242 		}
5243 		return (fs_frlock(vp, cmd, bfp, flag, offset, flk_cbp, cr));
5244 	}
5245 
5246 	rp = VTOR(vp);
5247 
5248 	/*
5249 	 * Check whether the given lock request can proceed, given the
5250 	 * current file mappings.
5251 	 */
5252 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_WRITER, intr))
5253 		return (EINTR);
5254 	if (cmd == F_SETLK || cmd == F_SETLKW) {
5255 		if (!lm_safelock(vp, bfp, cr)) {
5256 			rc = EAGAIN;
5257 			goto done;
5258 		}
5259 	}
5260 
5261 	/*
5262 	 * Flush the cache after waiting for async I/O to finish.  For new
5263 	 * locks, this is so that the process gets the latest bits from the
5264 	 * server.  For unlocks, this is so that other clients see the
5265 	 * latest bits once the file has been unlocked.  If currently dirty
5266 	 * pages can't be flushed, then don't allow a lock to be set.  But
5267 	 * allow unlocks to succeed, to avoid having orphan locks on the
5268 	 * server.
5269 	 */
5270 	if (cmd != F_GETLK) {
5271 		mutex_enter(&rp->r_statelock);
5272 		while (rp->r_count > 0) {
5273 		    if (intr) {
5274 			klwp_t *lwp = ttolwp(curthread);
5275 
5276 			if (lwp != NULL)
5277 				lwp->lwp_nostop++;
5278 			if (cv_wait_sig(&rp->r_cv, &rp->r_statelock) == 0) {
5279 				if (lwp != NULL)
5280 					lwp->lwp_nostop--;
5281 				rc = EINTR;
5282 				break;
5283 			}
5284 			if (lwp != NULL)
5285 				lwp->lwp_nostop--;
5286 		    } else
5287 			cv_wait(&rp->r_cv, &rp->r_statelock);
5288 		}
5289 		mutex_exit(&rp->r_statelock);
5290 		if (rc != 0)
5291 			goto done;
5292 		error = nfs3_putpage(vp, (offset_t)0, 0, B_INVAL, cr);
5293 		if (error) {
5294 			if (error == ENOSPC || error == EDQUOT) {
5295 				mutex_enter(&rp->r_statelock);
5296 				if (!rp->r_error)
5297 					rp->r_error = error;
5298 				mutex_exit(&rp->r_statelock);
5299 			}
5300 			if (bfp->l_type != F_UNLCK) {
5301 				rc = ENOLCK;
5302 				goto done;
5303 			}
5304 		}
5305 	}
5306 
5307 	lm_fh3.n_len = VTOFH3(vp)->fh3_length;
5308 	lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
5309 
5310 	/*
5311 	 * Call the lock manager to do the real work of contacting
5312 	 * the server and obtaining the lock.
5313 	 */
5314 	rc = lm4_frlock(vp, cmd, bfp, flag, offset, cr, &lm_fh3, flk_cbp);
5315 
5316 	if (rc == 0)
5317 		nfs_lockcompletion(vp, cmd);
5318 
5319 done:
5320 	nfs_rw_exit(&rp->r_lkserlock);
5321 	return (rc);
5322 }
5323 
5324 /*
5325  * Free storage space associated with the specified vnode.  The portion
5326  * to be freed is specified by bfp->l_start and bfp->l_len (already
5327  * normalized to a "whence" of 0).
5328  *
5329  * This is an experimental facility whose continued existence is not
5330  * guaranteed.  Currently, we only support the special case
5331  * of l_len == 0, meaning free to end of file.
5332  */
5333 /* ARGSUSED */
5334 static int
5335 nfs3_space(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5336 	offset_t offset, cred_t *cr, caller_context_t *ct)
5337 {
5338 	int error;
5339 
5340 	ASSERT(vp->v_type == VREG);
5341 	if (cmd != F_FREESP)
5342 		return (EINVAL);
5343 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5344 		return (EIO);
5345 
5346 	error = convoff(vp, bfp, 0, offset);
5347 	if (!error) {
5348 		ASSERT(bfp->l_start >= 0);
5349 		if (bfp->l_len == 0) {
5350 			struct vattr va;
5351 
5352 			/*
5353 			 * ftruncate should not change the ctime and
5354 			 * mtime if we truncate the file to its
5355 			 * previous size.
5356 			 */
5357 			va.va_mask = AT_SIZE;
5358 			error = nfs3getattr(vp, &va, cr);
5359 			if (error || va.va_size == bfp->l_start)
5360 				return (error);
5361 			va.va_mask = AT_SIZE;
5362 			va.va_size = bfp->l_start;
5363 			error = nfs3setattr(vp, &va, 0, cr);
5364 		} else
5365 			error = EINVAL;
5366 	}
5367 
5368 	return (error);
5369 }
5370 
5371 /* ARGSUSED */
5372 static int
5373 nfs3_realvp(vnode_t *vp, vnode_t **vpp)
5374 {
5375 
5376 	return (EINVAL);
5377 }
5378 
5379 /*
5380  * Setup and add an address space callback to do the work of the delmap call.
5381  * The callback will (and must be) deleted in the actual callback function.
5382  *
5383  * This is done in order to take care of the problem that we have with holding
5384  * the address space's a_lock for a long period of time (e.g. if the NFS server
5385  * is down).  Callbacks will be executed in the address space code while the
5386  * a_lock is not held.	Holding the address space's a_lock causes things such
5387  * as ps and fork to hang because they are trying to acquire this lock as well.
5388  */
5389 /* ARGSUSED */
5390 static int
5391 nfs3_delmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5392 	size_t len, uint_t prot, uint_t maxprot, uint_t flags, cred_t *cr)
5393 {
5394 	int			caller_found;
5395 	int			error;
5396 	rnode_t			*rp;
5397 	nfs_delmap_args_t	*dmapp;
5398 	nfs_delmapcall_t	*delmap_call;
5399 
5400 	if (vp->v_flag & VNOMAP)
5401 		return (ENOSYS);
5402 	/*
5403 	 * A process may not change zones if it has NFS pages mmap'ed
5404 	 * in, so we can't legitimately get here from the wrong zone.
5405 	 */
5406 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5407 
5408 	rp = VTOR(vp);
5409 
5410 	/*
5411 	 * The way that the address space of this process deletes its mapping
5412 	 * of this file is via the following call chains:
5413 	 * - as_free()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5414 	 * - as_unmap()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5415 	 *
5416 	 * With the use of address space callbacks we are allowed to drop the
5417 	 * address space lock, a_lock, while executing the NFS operations that
5418 	 * need to go over the wire.  Returning EAGAIN to the caller of this
5419 	 * function is what drives the execution of the callback that we add
5420 	 * below.  The callback will be executed by the address space code
5421 	 * after dropping the a_lock.  When the callback is finished, since
5422 	 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
5423 	 * is called again on the same segment to finish the rest of the work
5424 	 * that needs to happen during unmapping.
5425 	 *
5426 	 * This action of calling back into the segment driver causes
5427 	 * nfs3_delmap() to get called again, but since the callback was
5428 	 * already executed at this point, it already did the work and there
5429 	 * is nothing left for us to do.
5430 	 *
5431 	 * To Summarize:
5432 	 * - The first time nfs3_delmap is called by the current thread is when
5433 	 * we add the caller associated with this delmap to the delmap caller
5434 	 * list, add the callback, and return EAGAIN.
5435 	 * - The second time in this call chain when nfs3_delmap is called we
5436 	 * will find this caller in the delmap caller list and realize there
5437 	 * is no more work to do thus removing this caller from the list and
5438 	 * returning the error that was set in the callback execution.
5439 	 */
5440 	caller_found = nfs_find_and_delete_delmapcall(rp, &error);
5441 	if (caller_found) {
5442 		/*
5443 		 * 'error' is from the actual delmap operations.  To avoid
5444 		 * hangs, we need to handle the return of EAGAIN differently
5445 		 * since this is what drives the callback execution.
5446 		 * In this case, we don't want to return EAGAIN and do the
5447 		 * callback execution because there are none to execute.
5448 		 */
5449 		if (error == EAGAIN)
5450 			return (0);
5451 		else
5452 			return (error);
5453 	}
5454 
5455 	/* current caller was not in the list */
5456 	delmap_call = nfs_init_delmapcall();
5457 
5458 	mutex_enter(&rp->r_statelock);
5459 	list_insert_tail(&rp->r_indelmap, delmap_call);
5460 	mutex_exit(&rp->r_statelock);
5461 
5462 	dmapp = kmem_alloc(sizeof (nfs_delmap_args_t), KM_SLEEP);
5463 
5464 	dmapp->vp = vp;
5465 	dmapp->off = off;
5466 	dmapp->addr = addr;
5467 	dmapp->len = len;
5468 	dmapp->prot = prot;
5469 	dmapp->maxprot = maxprot;
5470 	dmapp->flags = flags;
5471 	dmapp->cr = cr;
5472 	dmapp->caller = delmap_call;
5473 
5474 	error = as_add_callback(as, nfs3_delmap_callback, dmapp,
5475 	AS_UNMAP_EVENT, addr, len, KM_SLEEP);
5476 
5477 	return (error ? error : EAGAIN);
5478 }
5479 
5480 /*
5481  * Remove some pages from an mmap'd vnode.  Just update the
5482  * count of pages.  If doing close-to-open, then flush and
5483  * commit all of the pages associated with this file.
5484  * Otherwise, start an asynchronous page flush to write out
5485  * any dirty pages.  This will also associate a credential
5486  * with the rnode which can be used to write the pages.
5487  */
5488 /* ARGSUSED */
5489 static void
5490 nfs3_delmap_callback(struct as *as, void *arg, uint_t event)
5491 {
5492 	int			error;
5493 	rnode_t			*rp;
5494 	mntinfo_t		*mi;
5495 	nfs_delmap_args_t	*dmapp = (nfs_delmap_args_t *)arg;
5496 
5497 	rp = VTOR(dmapp->vp);
5498 	mi = VTOMI(dmapp->vp);
5499 
5500 	atomic_add_long((ulong_t *)&rp->r_mapcnt, -btopr(dmapp->len));
5501 	ASSERT(rp->r_mapcnt >= 0);
5502 
5503 	/*
5504 	 * Initiate a page flush and potential commit if there are
5505 	 * pages, the file system was not mounted readonly, the segment
5506 	 * was mapped shared, and the pages themselves were writeable.
5507 	 */
5508 	if (vn_has_cached_data(dmapp->vp) && !vn_is_readonly(dmapp->vp) &&
5509 	    dmapp->flags == MAP_SHARED && (dmapp->maxprot & PROT_WRITE)) {
5510 		mutex_enter(&rp->r_statelock);
5511 		rp->r_flags |= RDIRTY;
5512 		mutex_exit(&rp->r_statelock);
5513 		/*
5514 		 * If this is a cross-zone access a sync putpage won't work, so
5515 		 * the best we can do is try an async putpage.  That seems
5516 		 * better than something more draconian such as discarding the
5517 		 * dirty pages.
5518 		 */
5519 		if ((mi->mi_flags & MI_NOCTO) ||
5520 		    nfs_zone() != mi->mi_zone)
5521 			error = nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5522 			    B_ASYNC, dmapp->cr);
5523 		else
5524 			error = nfs3_putpage_commit(dmapp->vp, dmapp->off,
5525 			    dmapp->len, dmapp->cr);
5526 		if (!error) {
5527 			mutex_enter(&rp->r_statelock);
5528 			error = rp->r_error;
5529 			rp->r_error = 0;
5530 			mutex_exit(&rp->r_statelock);
5531 		}
5532 	} else
5533 		error = 0;
5534 
5535 	if ((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO))
5536 		(void) nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5537 		    B_INVAL, dmapp->cr);
5538 
5539 	dmapp->caller->error = error;
5540 	(void) as_delete_callback(as, arg);
5541 	kmem_free(dmapp, sizeof (nfs_delmap_args_t));
5542 }
5543 
5544 static int nfs3_pathconf_disable_cache = 0;
5545 
5546 #ifdef DEBUG
5547 static int nfs3_pathconf_cache_hits = 0;
5548 static int nfs3_pathconf_cache_misses = 0;
5549 #endif
5550 
5551 static int
5552 nfs3_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr)
5553 {
5554 	int error;
5555 	PATHCONF3args args;
5556 	PATHCONF3res res;
5557 	int douprintf;
5558 	failinfo_t fi;
5559 	rnode_t *rp;
5560 	hrtime_t t;
5561 
5562 	if (nfs_zone() != VTOMI(vp)->mi_zone)
5563 		return (EIO);
5564 	/*
5565 	 * Large file spec - need to base answer on info stored
5566 	 * on original FSINFO response.
5567 	 */
5568 	if (cmd == _PC_FILESIZEBITS) {
5569 		unsigned long long ll;
5570 		long l = 1;
5571 
5572 		ll = VTOMI(vp)->mi_maxfilesize;
5573 
5574 		if (ll == 0) {
5575 			*valp = 0;
5576 			return (0);
5577 		}
5578 
5579 		if (ll & 0xffffffff00000000) {
5580 			l += 32; ll >>= 32;
5581 		}
5582 		if (ll & 0xffff0000) {
5583 			l += 16; ll >>= 16;
5584 		}
5585 		if (ll & 0xff00) {
5586 			l += 8; ll >>= 8;
5587 		}
5588 		if (ll & 0xf0) {
5589 			l += 4; ll >>= 4;
5590 		}
5591 		if (ll & 0xc) {
5592 			l += 2; ll >>= 2;
5593 		}
5594 		if (ll & 0x2)
5595 			l += 2;
5596 		else if (ll & 0x1)
5597 			l += 1;
5598 		*valp = l;
5599 		return (0);
5600 	}
5601 
5602 	if (cmd == _PC_ACL_ENABLED) {
5603 		*valp = _ACL_ACLENT_ENABLED;
5604 		return (0);
5605 	}
5606 
5607 	if (cmd == _PC_XATTR_EXISTS) {
5608 		error = 0;
5609 		*valp = 0;
5610 		if (vp->v_vfsp->vfs_flag & VFS_XATTR) {
5611 			vnode_t *avp;
5612 			rnode_t *rp;
5613 			int error = 0;
5614 			mntinfo_t *mi = VTOMI(vp);
5615 
5616 			if (!(mi->mi_flags & MI_EXTATTR))
5617 				return (0);
5618 
5619 			rp = VTOR(vp);
5620 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_READER,
5621 			    INTR(vp)))
5622 				return (EINTR);
5623 
5624 			error = nfs3lookup_dnlc(vp, XATTR_DIR_NAME, &avp, cr);
5625 			if (error || avp == NULL)
5626 				error = acl_getxattrdir3(vp, &avp, 0, cr, 0);
5627 
5628 			nfs_rw_exit(&rp->r_rwlock);
5629 
5630 			if (error == 0 && avp != NULL) {
5631 				VN_RELE(avp);
5632 				*valp = 1;
5633 			} else if (error == ENOENT)
5634 				error = 0;
5635 		}
5636 		return (error);
5637 	}
5638 
5639 	rp = VTOR(vp);
5640 	if (rp->r_pathconf != NULL) {
5641 		mutex_enter(&rp->r_statelock);
5642 		if (rp->r_pathconf != NULL && nfs3_pathconf_disable_cache) {
5643 			kmem_free(rp->r_pathconf, sizeof (*rp->r_pathconf));
5644 			rp->r_pathconf = NULL;
5645 		}
5646 		if (rp->r_pathconf != NULL) {
5647 			error = 0;
5648 			switch (cmd) {
5649 			case _PC_LINK_MAX:
5650 				*valp = rp->r_pathconf->link_max;
5651 				break;
5652 			case _PC_NAME_MAX:
5653 				*valp = rp->r_pathconf->name_max;
5654 				break;
5655 			case _PC_PATH_MAX:
5656 			case _PC_SYMLINK_MAX:
5657 				*valp = MAXPATHLEN;
5658 				break;
5659 			case _PC_CHOWN_RESTRICTED:
5660 				*valp = rp->r_pathconf->chown_restricted;
5661 				break;
5662 			case _PC_NO_TRUNC:
5663 				*valp = rp->r_pathconf->no_trunc;
5664 				break;
5665 			default:
5666 				error = EINVAL;
5667 				break;
5668 			}
5669 			mutex_exit(&rp->r_statelock);
5670 #ifdef DEBUG
5671 			nfs3_pathconf_cache_hits++;
5672 #endif
5673 			return (error);
5674 		}
5675 		mutex_exit(&rp->r_statelock);
5676 	}
5677 #ifdef DEBUG
5678 	nfs3_pathconf_cache_misses++;
5679 #endif
5680 
5681 	args.object = *VTOFH3(vp);
5682 	fi.vp = vp;
5683 	fi.fhp = (caddr_t)&args.object;
5684 	fi.copyproc = nfs3copyfh;
5685 	fi.lookupproc = nfs3lookup;
5686 	fi.xattrdirproc = acl_getxattrdir3;
5687 
5688 	douprintf = 1;
5689 
5690 	t = gethrtime();
5691 
5692 	error = rfs3call(VTOMI(vp), NFSPROC3_PATHCONF,
5693 	    xdr_nfs_fh3, (caddr_t)&args,
5694 	    xdr_PATHCONF3res, (caddr_t)&res, cr,
5695 	    &douprintf, &res.status, 0, &fi);
5696 
5697 	if (error)
5698 		return (error);
5699 
5700 	error = geterrno3(res.status);
5701 
5702 	if (!error) {
5703 		nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
5704 		if (!nfs3_pathconf_disable_cache) {
5705 			mutex_enter(&rp->r_statelock);
5706 			if (rp->r_pathconf == NULL) {
5707 				rp->r_pathconf = kmem_alloc(
5708 				    sizeof (*rp->r_pathconf), KM_NOSLEEP);
5709 				if (rp->r_pathconf != NULL)
5710 					*rp->r_pathconf = res.resok.info;
5711 			}
5712 			mutex_exit(&rp->r_statelock);
5713 		}
5714 		switch (cmd) {
5715 		case _PC_LINK_MAX:
5716 			*valp = res.resok.info.link_max;
5717 			break;
5718 		case _PC_NAME_MAX:
5719 			*valp = res.resok.info.name_max;
5720 			break;
5721 		case _PC_PATH_MAX:
5722 		case _PC_SYMLINK_MAX:
5723 			*valp = MAXPATHLEN;
5724 			break;
5725 		case _PC_CHOWN_RESTRICTED:
5726 			*valp = res.resok.info.chown_restricted;
5727 			break;
5728 		case _PC_NO_TRUNC:
5729 			*valp = res.resok.info.no_trunc;
5730 			break;
5731 		default:
5732 			return (EINVAL);
5733 		}
5734 	} else {
5735 		nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
5736 		PURGE_STALE_FH(error, vp, cr);
5737 	}
5738 
5739 	return (error);
5740 }
5741 
5742 /*
5743  * Called by async thread to do synchronous pageio. Do the i/o, wait
5744  * for it to complete, and cleanup the page list when done.
5745  */
5746 static int
5747 nfs3_sync_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5748 	int flags, cred_t *cr)
5749 {
5750 	int error;
5751 
5752 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5753 	error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5754 	if (flags & B_READ)
5755 		pvn_read_done(pp, (error ? B_ERROR : 0) | flags);
5756 	else
5757 		pvn_write_done(pp, (error ? B_ERROR : 0) | flags);
5758 	return (error);
5759 }
5760 
5761 static int
5762 nfs3_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5763 	int flags, cred_t *cr)
5764 {
5765 	int error;
5766 	rnode_t *rp;
5767 
5768 	if (pp == NULL)
5769 		return (EINVAL);
5770 	if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
5771 		return (EIO);
5772 
5773 	rp = VTOR(vp);
5774 	mutex_enter(&rp->r_statelock);
5775 	rp->r_count++;
5776 	mutex_exit(&rp->r_statelock);
5777 
5778 	if (flags & B_ASYNC) {
5779 		error = nfs_async_pageio(vp, pp, io_off, io_len, flags, cr,
5780 		    nfs3_sync_pageio);
5781 	} else
5782 		error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5783 	mutex_enter(&rp->r_statelock);
5784 	rp->r_count--;
5785 	cv_broadcast(&rp->r_cv);
5786 	mutex_exit(&rp->r_statelock);
5787 	return (error);
5788 }
5789 
5790 static void
5791 nfs3_dispose(vnode_t *vp, page_t *pp, int fl, int dn, cred_t *cr)
5792 {
5793 	int error;
5794 	rnode_t *rp;
5795 	page_t *plist;
5796 	page_t *pptr;
5797 	offset3 offset;
5798 	count3 len;
5799 	k_sigset_t smask;
5800 
5801 	/*
5802 	 * We should get called with fl equal to either B_FREE or
5803 	 * B_INVAL.  Any other value is illegal.
5804 	 *
5805 	 * The page that we are either supposed to free or destroy
5806 	 * should be exclusive locked and its io lock should not
5807 	 * be held.
5808 	 */
5809 	ASSERT(fl == B_FREE || fl == B_INVAL);
5810 	ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr);
5811 	rp = VTOR(vp);
5812 
5813 	/*
5814 	 * If the page doesn't need to be committed or we shouldn't
5815 	 * even bother attempting to commit it, then just make sure
5816 	 * that the p_fsdata byte is clear and then either free or
5817 	 * destroy the page as appropriate.
5818 	 */
5819 	if (pp->p_fsdata == C_NOCOMMIT || (rp->r_flags & RSTALE)) {
5820 		pp->p_fsdata = C_NOCOMMIT;
5821 		if (fl == B_FREE)
5822 			page_free(pp, dn);
5823 		else
5824 			page_destroy(pp, dn);
5825 		return;
5826 	}
5827 
5828 	/*
5829 	 * If there is a page invalidation operation going on, then
5830 	 * if this is one of the pages being destroyed, then just
5831 	 * clear the p_fsdata byte and then either free or destroy
5832 	 * the page as appropriate.
5833 	 */
5834 	mutex_enter(&rp->r_statelock);
5835 	if ((rp->r_flags & RTRUNCATE) && pp->p_offset >= rp->r_truncaddr) {
5836 		mutex_exit(&rp->r_statelock);
5837 		pp->p_fsdata = C_NOCOMMIT;
5838 		if (fl == B_FREE)
5839 			page_free(pp, dn);
5840 		else
5841 			page_destroy(pp, dn);
5842 		return;
5843 	}
5844 
5845 	/*
5846 	 * If we are freeing this page and someone else is already
5847 	 * waiting to do a commit, then just unlock the page and
5848 	 * return.  That other thread will take care of commiting
5849 	 * this page.  The page can be freed sometime after the
5850 	 * commit has finished.  Otherwise, if the page is marked
5851 	 * as delay commit, then we may be getting called from
5852 	 * pvn_write_done, one page at a time.   This could result
5853 	 * in one commit per page, so we end up doing lots of small
5854 	 * commits instead of fewer larger commits.  This is bad,
5855 	 * we want do as few commits as possible.
5856 	 */
5857 	if (fl == B_FREE) {
5858 		if (rp->r_flags & RCOMMITWAIT) {
5859 			page_unlock(pp);
5860 			mutex_exit(&rp->r_statelock);
5861 			return;
5862 		}
5863 		if (pp->p_fsdata == C_DELAYCOMMIT) {
5864 			pp->p_fsdata = C_COMMIT;
5865 			page_unlock(pp);
5866 			mutex_exit(&rp->r_statelock);
5867 			return;
5868 		}
5869 	}
5870 
5871 	/*
5872 	 * Check to see if there is a signal which would prevent an
5873 	 * attempt to commit the pages from being successful.  If so,
5874 	 * then don't bother with all of the work to gather pages and
5875 	 * generate the unsuccessful RPC.  Just return from here and
5876 	 * let the page be committed at some later time.
5877 	 */
5878 	sigintr(&smask, VTOMI(vp)->mi_flags & MI_INT);
5879 	if (ttolwp(curthread) != NULL && ISSIG(curthread, JUSTLOOKING)) {
5880 		sigunintr(&smask);
5881 		page_unlock(pp);
5882 		mutex_exit(&rp->r_statelock);
5883 		return;
5884 	}
5885 	sigunintr(&smask);
5886 
5887 	/*
5888 	 * We are starting to need to commit pages, so let's try
5889 	 * to commit as many as possible at once to reduce the
5890 	 * overhead.
5891 	 *
5892 	 * Set the `commit inprogress' state bit.  We must
5893 	 * first wait until any current one finishes.  Then
5894 	 * we initialize the c_pages list with this page.
5895 	 */
5896 	while (rp->r_flags & RCOMMIT) {
5897 		rp->r_flags |= RCOMMITWAIT;
5898 		cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
5899 		rp->r_flags &= ~RCOMMITWAIT;
5900 	}
5901 	rp->r_flags |= RCOMMIT;
5902 	mutex_exit(&rp->r_statelock);
5903 	ASSERT(rp->r_commit.c_pages == NULL);
5904 	rp->r_commit.c_pages = pp;
5905 	rp->r_commit.c_commbase = (offset3)pp->p_offset;
5906 	rp->r_commit.c_commlen = PAGESIZE;
5907 
5908 	/*
5909 	 * Gather together all other pages which can be committed.
5910 	 * They will all be chained off r_commit.c_pages.
5911 	 */
5912 	nfs3_get_commit(vp);
5913 
5914 	/*
5915 	 * Clear the `commit inprogress' status and disconnect
5916 	 * the list of pages to be committed from the rnode.
5917 	 * At this same time, we also save the starting offset
5918 	 * and length of data to be committed on the server.
5919 	 */
5920 	plist = rp->r_commit.c_pages;
5921 	rp->r_commit.c_pages = NULL;
5922 	offset = rp->r_commit.c_commbase;
5923 	len = rp->r_commit.c_commlen;
5924 	mutex_enter(&rp->r_statelock);
5925 	rp->r_flags &= ~RCOMMIT;
5926 	cv_broadcast(&rp->r_commit.c_cv);
5927 	mutex_exit(&rp->r_statelock);
5928 
5929 	if (curproc == proc_pageout || curproc == proc_fsflush ||
5930 	    nfs_zone() != VTOMI(vp)->mi_zone) {
5931 		nfs_async_commit(vp, plist, offset, len, cr, nfs3_async_commit);
5932 		return;
5933 	}
5934 
5935 	/*
5936 	 * Actually generate the COMMIT3 over the wire operation.
5937 	 */
5938 	error = nfs3_commit(vp, offset, len, cr);
5939 
5940 	/*
5941 	 * If we got an error during the commit, just unlock all
5942 	 * of the pages.  The pages will get retransmitted to the
5943 	 * server during a putpage operation.
5944 	 */
5945 	if (error) {
5946 		while (plist != NULL) {
5947 			pptr = plist;
5948 			page_sub(&plist, pptr);
5949 			page_unlock(pptr);
5950 		}
5951 		return;
5952 	}
5953 
5954 	/*
5955 	 * We've tried as hard as we can to commit the data to stable
5956 	 * storage on the server.  We release the rest of the pages
5957 	 * and clear the commit required state.  They will be put
5958 	 * onto the tail of the cachelist if they are nolonger
5959 	 * mapped.
5960 	 */
5961 	while (plist != pp) {
5962 		pptr = plist;
5963 		page_sub(&plist, pptr);
5964 		pptr->p_fsdata = C_NOCOMMIT;
5965 		(void) page_release(pptr, 1);
5966 	}
5967 
5968 	/*
5969 	 * It is possible that nfs3_commit didn't return error but
5970 	 * some other thread has modified the page we are going
5971 	 * to free/destroy.
5972 	 *    In this case we need to rewrite the page. Do an explicit check
5973 	 * before attempting to free/destroy the page. If modified, needs to
5974 	 * be rewritten so unlock the page and return.
5975 	 */
5976 	if (hat_ismod(pp)) {
5977 		pp->p_fsdata = C_NOCOMMIT;
5978 		page_unlock(pp);
5979 		return;
5980 	}
5981 
5982 	/*
5983 	 * Now, as appropriate, either free or destroy the page
5984 	 * that we were called with.
5985 	 */
5986 	pp->p_fsdata = C_NOCOMMIT;
5987 	if (fl == B_FREE)
5988 		page_free(pp, dn);
5989 	else
5990 		page_destroy(pp, dn);
5991 }
5992 
5993 static int
5994 nfs3_commit(vnode_t *vp, offset3 offset, count3 count, cred_t *cr)
5995 {
5996 	int error;
5997 	rnode_t *rp;
5998 	COMMIT3args args;
5999 	COMMIT3res res;
6000 	int douprintf;
6001 	cred_t *cred;
6002 
6003 	rp = VTOR(vp);
6004 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6005 
6006 	mutex_enter(&rp->r_statelock);
6007 	if (rp->r_cred != NULL) {
6008 		cred = rp->r_cred;
6009 		crhold(cred);
6010 	} else {
6011 		rp->r_cred = cr;
6012 		crhold(cr);
6013 		cred = cr;
6014 		crhold(cred);
6015 	}
6016 	mutex_exit(&rp->r_statelock);
6017 
6018 	args.file = *VTOFH3(vp);
6019 	args.offset = offset;
6020 	args.count = count;
6021 
6022 doitagain:
6023 	douprintf = 1;
6024 	error = rfs3call(VTOMI(vp), NFSPROC3_COMMIT,
6025 	    xdr_COMMIT3args, (caddr_t)&args,
6026 	    xdr_COMMIT3res, (caddr_t)&res, cred,
6027 	    &douprintf, &res.status, 0, NULL);
6028 
6029 	crfree(cred);
6030 
6031 	if (error)
6032 		return (error);
6033 
6034 	error = geterrno3(res.status);
6035 	if (!error) {
6036 		ASSERT(rp->r_flags & RHAVEVERF);
6037 		mutex_enter(&rp->r_statelock);
6038 		if (rp->r_verf == res.resok.verf) {
6039 			mutex_exit(&rp->r_statelock);
6040 			return (0);
6041 		}
6042 		nfs3_set_mod(vp);
6043 		rp->r_verf = res.resok.verf;
6044 		mutex_exit(&rp->r_statelock);
6045 		error = NFS_VERF_MISMATCH;
6046 	} else {
6047 		if (error == EACCES) {
6048 			mutex_enter(&rp->r_statelock);
6049 			if (cred != cr) {
6050 				if (rp->r_cred != NULL)
6051 					crfree(rp->r_cred);
6052 				rp->r_cred = cr;
6053 				crhold(cr);
6054 				cred = cr;
6055 				crhold(cred);
6056 				mutex_exit(&rp->r_statelock);
6057 				goto doitagain;
6058 			}
6059 			mutex_exit(&rp->r_statelock);
6060 		}
6061 		/*
6062 		 * Can't do a PURGE_STALE_FH here because this
6063 		 * can cause a deadlock.  nfs3_commit can
6064 		 * be called from nfs3_dispose which can be called
6065 		 * indirectly via pvn_vplist_dirty.  PURGE_STALE_FH
6066 		 * can call back to pvn_vplist_dirty.
6067 		 */
6068 		if (error == ESTALE) {
6069 			mutex_enter(&rp->r_statelock);
6070 			rp->r_flags |= RSTALE;
6071 			if (!rp->r_error)
6072 				rp->r_error = error;
6073 			mutex_exit(&rp->r_statelock);
6074 			PURGE_ATTRCACHE(vp);
6075 		} else {
6076 			mutex_enter(&rp->r_statelock);
6077 			if (!rp->r_error)
6078 				rp->r_error = error;
6079 			mutex_exit(&rp->r_statelock);
6080 		}
6081 	}
6082 
6083 	return (error);
6084 }
6085 
6086 static void
6087 nfs3_set_mod(vnode_t *vp)
6088 {
6089 	page_t *pp;
6090 	kmutex_t *vphm;
6091 
6092 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6093 	vphm = page_vnode_mutex(vp);
6094 	mutex_enter(vphm);
6095 	if ((pp = vp->v_pages) != NULL) {
6096 		do {
6097 			if (pp->p_fsdata != C_NOCOMMIT) {
6098 				hat_setmod(pp);
6099 				pp->p_fsdata = C_NOCOMMIT;
6100 			}
6101 		} while ((pp = pp->p_vpnext) != vp->v_pages);
6102 	}
6103 	mutex_exit(vphm);
6104 }
6105 
6106 
6107 /*
6108  * This routine is used to gather together a page list of the pages
6109  * which are to be committed on the server.  This routine must not
6110  * be called if the calling thread holds any locked pages.
6111  *
6112  * The calling thread must have set RCOMMIT.  This bit is used to
6113  * serialize access to the commit structure in the rnode.  As long
6114  * as the thread has set RCOMMIT, then it can manipulate the commit
6115  * structure without requiring any other locks.
6116  */
6117 static void
6118 nfs3_get_commit(vnode_t *vp)
6119 {
6120 	rnode_t *rp;
6121 	page_t *pp;
6122 	kmutex_t *vphm;
6123 
6124 	rp = VTOR(vp);
6125 
6126 	ASSERT(rp->r_flags & RCOMMIT);
6127 
6128 	vphm = page_vnode_mutex(vp);
6129 	mutex_enter(vphm);
6130 
6131 	/*
6132 	 * If there are no pages associated with this vnode, then
6133 	 * just return.
6134 	 */
6135 	if ((pp = vp->v_pages) == NULL) {
6136 		mutex_exit(vphm);
6137 		return;
6138 	}
6139 
6140 	/*
6141 	 * Step through all of the pages associated with this vnode
6142 	 * looking for pages which need to be committed.
6143 	 */
6144 	do {
6145 		/*
6146 		 * If this page does not need to be committed or is
6147 		 * modified, then just skip it.
6148 		 */
6149 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp))
6150 			continue;
6151 
6152 		/*
6153 		 * Attempt to lock the page.  If we can't, then
6154 		 * someone else is messing with it and we will
6155 		 * just skip it.
6156 		 */
6157 		if (!page_trylock(pp, SE_EXCL))
6158 			continue;
6159 
6160 		/*
6161 		 * If this page does not need to be committed or is
6162 		 * modified, then just skip it.  Recheck now that
6163 		 * the page is locked.
6164 		 */
6165 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6166 			page_unlock(pp);
6167 			continue;
6168 		}
6169 
6170 		if (PP_ISFREE(pp)) {
6171 			cmn_err(CE_PANIC, "nfs3_get_commit: %p is free",
6172 			    (void *)pp);
6173 		}
6174 
6175 		/*
6176 		 * The page needs to be committed and we locked it.
6177 		 * Update the base and length parameters and add it
6178 		 * to r_pages.
6179 		 */
6180 		if (rp->r_commit.c_pages == NULL) {
6181 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6182 			rp->r_commit.c_commlen = PAGESIZE;
6183 		} else if (pp->p_offset < rp->r_commit.c_commbase) {
6184 			rp->r_commit.c_commlen = rp->r_commit.c_commbase -
6185 			    (offset3)pp->p_offset + rp->r_commit.c_commlen;
6186 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6187 		} else if ((rp->r_commit.c_commbase + rp->r_commit.c_commlen)
6188 			    <= pp->p_offset) {
6189 			rp->r_commit.c_commlen = (offset3)pp->p_offset -
6190 			    rp->r_commit.c_commbase + PAGESIZE;
6191 		}
6192 		page_add(&rp->r_commit.c_pages, pp);
6193 	} while ((pp = pp->p_vpnext) != vp->v_pages);
6194 
6195 	mutex_exit(vphm);
6196 }
6197 
6198 /*
6199  * This routine is used to gather together a page list of the pages
6200  * which are to be committed on the server.  This routine must not
6201  * be called if the calling thread holds any locked pages.
6202  *
6203  * The calling thread must have set RCOMMIT.  This bit is used to
6204  * serialize access to the commit structure in the rnode.  As long
6205  * as the thread has set RCOMMIT, then it can manipulate the commit
6206  * structure without requiring any other locks.
6207  */
6208 static void
6209 nfs3_get_commit_range(vnode_t *vp, u_offset_t soff, size_t len)
6210 {
6211 
6212 	rnode_t *rp;
6213 	page_t *pp;
6214 	u_offset_t end;
6215 	u_offset_t off;
6216 
6217 	ASSERT(len != 0);
6218 
6219 	rp = VTOR(vp);
6220 
6221 	ASSERT(rp->r_flags & RCOMMIT);
6222 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6223 
6224 	/*
6225 	 * If there are no pages associated with this vnode, then
6226 	 * just return.
6227 	 */
6228 	if ((pp = vp->v_pages) == NULL)
6229 		return;
6230 
6231 	/*
6232 	 * Calculate the ending offset.
6233 	 */
6234 	end = soff + len;
6235 
6236 	for (off = soff; off < end; off += PAGESIZE) {
6237 		/*
6238 		 * Lookup each page by vp, offset.
6239 		 */
6240 		if ((pp = page_lookup_nowait(vp, off, SE_EXCL)) == NULL)
6241 			continue;
6242 
6243 		/*
6244 		 * If this page does not need to be committed or is
6245 		 * modified, then just skip it.
6246 		 */
6247 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6248 			page_unlock(pp);
6249 			continue;
6250 		}
6251 
6252 		ASSERT(PP_ISFREE(pp) == 0);
6253 
6254 		/*
6255 		 * The page needs to be committed and we locked it.
6256 		 * Update the base and length parameters and add it
6257 		 * to r_pages.
6258 		 */
6259 		if (rp->r_commit.c_pages == NULL) {
6260 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6261 			rp->r_commit.c_commlen = PAGESIZE;
6262 		} else {
6263 			rp->r_commit.c_commlen = (offset3)pp->p_offset -
6264 					rp->r_commit.c_commbase + PAGESIZE;
6265 		}
6266 		page_add(&rp->r_commit.c_pages, pp);
6267 	}
6268 }
6269 
6270 #if 0	/* unused */
6271 #ifdef DEBUG
6272 static int
6273 nfs3_no_uncommitted_pages(vnode_t *vp)
6274 {
6275 	page_t *pp;
6276 	kmutex_t *vphm;
6277 
6278 	vphm = page_vnode_mutex(vp);
6279 	mutex_enter(vphm);
6280 	if ((pp = vp->v_pages) != NULL) {
6281 		do {
6282 			if (pp->p_fsdata != C_NOCOMMIT) {
6283 				mutex_exit(vphm);
6284 				return (0);
6285 			}
6286 		} while ((pp = pp->p_vpnext) != vp->v_pages);
6287 	}
6288 	mutex_exit(vphm);
6289 
6290 	return (1);
6291 }
6292 #endif
6293 #endif
6294 
6295 static int
6296 nfs3_putpage_commit(vnode_t *vp, offset_t poff, size_t plen, cred_t *cr)
6297 {
6298 	int error;
6299 	writeverf3 write_verf;
6300 	rnode_t *rp = VTOR(vp);
6301 
6302 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6303 	/*
6304 	 * Flush the data portion of the file and then commit any
6305 	 * portions which need to be committed.  This may need to
6306 	 * be done twice if the server has changed state since
6307 	 * data was last written.  The data will need to be
6308 	 * rewritten to the server and then a new commit done.
6309 	 *
6310 	 * In fact, this may need to be done several times if the
6311 	 * server is having problems and crashing while we are
6312 	 * attempting to do this.
6313 	 */
6314 
6315 top:
6316 	/*
6317 	 * Do a flush based on the poff and plen arguments.  This
6318 	 * will asynchronously write out any modified pages in the
6319 	 * range specified by (poff, plen).  This starts all of the
6320 	 * i/o operations which will be waited for in the next
6321 	 * call to nfs3_putpage
6322 	 */
6323 
6324 	mutex_enter(&rp->r_statelock);
6325 	write_verf = rp->r_verf;
6326 	mutex_exit(&rp->r_statelock);
6327 
6328 	error = nfs3_putpage(vp, poff, plen, B_ASYNC, cr);
6329 	if (error == EAGAIN)
6330 		error = 0;
6331 
6332 	/*
6333 	 * Do a flush based on the poff and plen arguments.  This
6334 	 * will synchronously write out any modified pages in the
6335 	 * range specified by (poff, plen) and wait until all of
6336 	 * the asynchronous i/o's in that range are done as well.
6337 	 */
6338 	if (!error)
6339 		error = nfs3_putpage(vp, poff, plen, 0, cr);
6340 
6341 	if (error)
6342 		return (error);
6343 
6344 	mutex_enter(&rp->r_statelock);
6345 	if (rp->r_verf != write_verf) {
6346 		mutex_exit(&rp->r_statelock);
6347 		goto top;
6348 	}
6349 	mutex_exit(&rp->r_statelock);
6350 
6351 	/*
6352 	 * Now commit any pages which might need to be committed.
6353 	 * If the error, NFS_VERF_MISMATCH, is returned, then
6354 	 * start over with the flush operation.
6355 	 */
6356 
6357 	error = nfs3_commit_vp(vp, poff, plen, cr);
6358 
6359 	if (error == NFS_VERF_MISMATCH)
6360 		goto top;
6361 
6362 	return (error);
6363 }
6364 
6365 static int
6366 nfs3_commit_vp(vnode_t *vp, u_offset_t poff, size_t plen, cred_t *cr)
6367 {
6368 	rnode_t *rp;
6369 	page_t *plist;
6370 	offset3 offset;
6371 	count3 len;
6372 
6373 
6374 	rp = VTOR(vp);
6375 
6376 	if (nfs_zone() != VTOMI(vp)->mi_zone)
6377 		return (EIO);
6378 	/*
6379 	 * Set the `commit inprogress' state bit.  We must
6380 	 * first wait until any current one finishes.
6381 	 */
6382 	mutex_enter(&rp->r_statelock);
6383 	while (rp->r_flags & RCOMMIT) {
6384 		rp->r_flags |= RCOMMITWAIT;
6385 		cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
6386 		rp->r_flags &= ~RCOMMITWAIT;
6387 	}
6388 	rp->r_flags |= RCOMMIT;
6389 	mutex_exit(&rp->r_statelock);
6390 
6391 	/*
6392 	 * Gather together all of the pages which need to be
6393 	 * committed.
6394 	 */
6395 	if (plen == 0)
6396 		nfs3_get_commit(vp);
6397 	else
6398 		nfs3_get_commit_range(vp, poff, plen);
6399 
6400 	/*
6401 	 * Clear the `commit inprogress' bit and disconnect the
6402 	 * page list which was gathered together in nfs3_get_commit.
6403 	 */
6404 	plist = rp->r_commit.c_pages;
6405 	rp->r_commit.c_pages = NULL;
6406 	offset = rp->r_commit.c_commbase;
6407 	len = rp->r_commit.c_commlen;
6408 	mutex_enter(&rp->r_statelock);
6409 	rp->r_flags &= ~RCOMMIT;
6410 	cv_broadcast(&rp->r_commit.c_cv);
6411 	mutex_exit(&rp->r_statelock);
6412 
6413 	/*
6414 	 * If any pages need to be committed, commit them and
6415 	 * then unlock them so that they can be freed some
6416 	 * time later.
6417 	 */
6418 	if (plist != NULL) {
6419 		/*
6420 		 * No error occurred during the flush portion
6421 		 * of this operation, so now attempt to commit
6422 		 * the data to stable storage on the server.
6423 		 *
6424 		 * This will unlock all of the pages on the list.
6425 		 */
6426 		return (nfs3_sync_commit(vp, plist, offset, len, cr));
6427 	}
6428 	return (0);
6429 }
6430 
6431 static int
6432 nfs3_sync_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6433 	cred_t *cr)
6434 {
6435 	int error;
6436 	page_t *pp;
6437 
6438 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6439 	error = nfs3_commit(vp, offset, count, cr);
6440 
6441 	/*
6442 	 * If we got an error, then just unlock all of the pages
6443 	 * on the list.
6444 	 */
6445 	if (error) {
6446 		while (plist != NULL) {
6447 			pp = plist;
6448 			page_sub(&plist, pp);
6449 			page_unlock(pp);
6450 		}
6451 		return (error);
6452 	}
6453 	/*
6454 	 * We've tried as hard as we can to commit the data to stable
6455 	 * storage on the server.  We just unlock the pages and clear
6456 	 * the commit required state.  They will get freed later.
6457 	 */
6458 	while (plist != NULL) {
6459 		pp = plist;
6460 		page_sub(&plist, pp);
6461 		pp->p_fsdata = C_NOCOMMIT;
6462 		page_unlock(pp);
6463 	}
6464 
6465 	return (error);
6466 }
6467 
6468 static void
6469 nfs3_async_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6470 	cred_t *cr)
6471 {
6472 	ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6473 	(void) nfs3_sync_commit(vp, plist, offset, count, cr);
6474 }
6475 
6476 static int
6477 nfs3_setsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
6478 {
6479 	int error;
6480 	mntinfo_t *mi;
6481 
6482 	mi = VTOMI(vp);
6483 
6484 	if (nfs_zone() != mi->mi_zone)
6485 		return (EIO);
6486 
6487 	if (mi->mi_flags & MI_ACL) {
6488 		error = acl_setacl3(vp, vsecattr, flag, cr);
6489 		if (mi->mi_flags & MI_ACL)
6490 			return (error);
6491 	}
6492 
6493 	return (ENOSYS);
6494 }
6495 
6496 static int
6497 nfs3_getsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
6498 {
6499 	int error;
6500 	mntinfo_t *mi;
6501 
6502 	mi = VTOMI(vp);
6503 
6504 	if (nfs_zone() != mi->mi_zone)
6505 		return (EIO);
6506 
6507 	if (mi->mi_flags & MI_ACL) {
6508 		error = acl_getacl3(vp, vsecattr, flag, cr);
6509 		if (mi->mi_flags & MI_ACL)
6510 			return (error);
6511 	}
6512 
6513 	return (fs_fab_acl(vp, vsecattr, flag, cr));
6514 }
6515 
6516 static int
6517 nfs3_shrlock(vnode_t *vp, int cmd, struct shrlock *shr, int flag, cred_t *cr)
6518 {
6519 	int error;
6520 	struct shrlock nshr;
6521 	struct nfs_owner nfs_owner;
6522 	netobj lm_fh3;
6523 
6524 	if (nfs_zone() != VTOMI(vp)->mi_zone)
6525 		return (EIO);
6526 
6527 	/*
6528 	 * check for valid cmd parameter
6529 	 */
6530 	if (cmd != F_SHARE && cmd != F_UNSHARE && cmd != F_HASREMOTELOCKS)
6531 		return (EINVAL);
6532 
6533 	/*
6534 	 * Check access permissions
6535 	 */
6536 	if (cmd == F_SHARE &&
6537 	    (((shr->s_access & F_RDACC) && !(flag & FREAD)) ||
6538 	    ((shr->s_access & F_WRACC) && !(flag & FWRITE))))
6539 		return (EBADF);
6540 
6541 	/*
6542 	 * If the filesystem is mounted using local locking, pass the
6543 	 * request off to the local share code.
6544 	 */
6545 	if (VTOMI(vp)->mi_flags & MI_LLOCK)
6546 		return (fs_shrlock(vp, cmd, shr, flag, cr));
6547 
6548 	switch (cmd) {
6549 	case F_SHARE:
6550 	case F_UNSHARE:
6551 		lm_fh3.n_len = VTOFH3(vp)->fh3_length;
6552 		lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
6553 
6554 		/*
6555 		 * If passed an owner that is too large to fit in an
6556 		 * nfs_owner it is likely a recursive call from the
6557 		 * lock manager client and pass it straight through.  If
6558 		 * it is not a nfs_owner then simply return an error.
6559 		 */
6560 		if (shr->s_own_len > sizeof (nfs_owner.lowner)) {
6561 			if (((struct nfs_owner *)shr->s_owner)->magic !=
6562 			    NFS_OWNER_MAGIC)
6563 				return (EINVAL);
6564 
6565 			if (error = lm4_shrlock(vp, cmd, shr, flag, &lm_fh3)) {
6566 				error = set_errno(error);
6567 			}
6568 			return (error);
6569 		}
6570 		/*
6571 		 * Remote share reservations owner is a combination of
6572 		 * a magic number, hostname, and the local owner
6573 		 */
6574 		bzero(&nfs_owner, sizeof (nfs_owner));
6575 		nfs_owner.magic = NFS_OWNER_MAGIC;
6576 		(void) strncpy(nfs_owner.hname, uts_nodename(),
6577 		    sizeof (nfs_owner.hname));
6578 		bcopy(shr->s_owner, nfs_owner.lowner, shr->s_own_len);
6579 		nshr.s_access = shr->s_access;
6580 		nshr.s_deny = shr->s_deny;
6581 		nshr.s_sysid = 0;
6582 		nshr.s_pid = ttoproc(curthread)->p_pid;
6583 		nshr.s_own_len = sizeof (nfs_owner);
6584 		nshr.s_owner = (caddr_t)&nfs_owner;
6585 
6586 		if (error = lm4_shrlock(vp, cmd, &nshr, flag, &lm_fh3)) {
6587 			error = set_errno(error);
6588 		}
6589 
6590 		break;
6591 
6592 	case F_HASREMOTELOCKS:
6593 		/*
6594 		 * NFS client can't store remote locks itself
6595 		 */
6596 		shr->s_access = 0;
6597 		error = 0;
6598 		break;
6599 
6600 	default:
6601 		error = EINVAL;
6602 		break;
6603 	}
6604 
6605 	return (error);
6606 }
6607