xref: /titanic_52/usr/src/uts/common/fs/nfs/nfs3_vnops.c (revision bdfc6d18da790deeec2e0eb09c625902defe2498)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  *	Copyright (c) 1983,1984,1985,1986,1987,1988,1989 AT&T.
29  *	All rights reserved.
30  */
31 
32 #pragma ident	"%Z%%M%	%I%	%E% SMI"
33 
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/systm.h>
37 #include <sys/cred.h>
38 #include <sys/time.h>
39 #include <sys/vnode.h>
40 #include <sys/vfs.h>
41 #include <sys/file.h>
42 #include <sys/filio.h>
43 #include <sys/uio.h>
44 #include <sys/buf.h>
45 #include <sys/mman.h>
46 #include <sys/pathname.h>
47 #include <sys/dirent.h>
48 #include <sys/debug.h>
49 #include <sys/vmsystm.h>
50 #include <sys/fcntl.h>
51 #include <sys/flock.h>
52 #include <sys/swap.h>
53 #include <sys/errno.h>
54 #include <sys/strsubr.h>
55 #include <sys/sysmacros.h>
56 #include <sys/kmem.h>
57 #include <sys/cmn_err.h>
58 #include <sys/pathconf.h>
59 #include <sys/utsname.h>
60 #include <sys/dnlc.h>
61 #include <sys/acl.h>
62 #include <sys/systeminfo.h>
63 #include <sys/atomic.h>
64 #include <sys/policy.h>
65 #include <sys/sdt.h>
66 
67 #include <rpc/types.h>
68 #include <rpc/auth.h>
69 #include <rpc/clnt.h>
70 
71 #include <nfs/nfs.h>
72 #include <nfs/nfs_clnt.h>
73 #include <nfs/rnode.h>
74 #include <nfs/nfs_acl.h>
75 #include <nfs/lm.h>
76 
77 #include <vm/hat.h>
78 #include <vm/as.h>
79 #include <vm/page.h>
80 #include <vm/pvn.h>
81 #include <vm/seg.h>
82 #include <vm/seg_map.h>
83 #include <vm/seg_kpm.h>
84 #include <vm/seg_vn.h>
85 
86 #include <fs/fs_subr.h>
87 
88 #include <sys/ddi.h>
89 
90 static int	nfs3_rdwrlbn(vnode_t *, page_t *, u_offset_t, size_t, int,
91 			cred_t *);
92 static int	nfs3write(vnode_t *, caddr_t, u_offset_t, int, cred_t *,
93 			stable_how *);
94 static int	nfs3read(vnode_t *, caddr_t, offset_t, int, size_t *, cred_t *);
95 static int	nfs3setattr(vnode_t *, struct vattr *, int, cred_t *);
96 static int	nfs3_accessx(void *, int, cred_t *);
97 static int	nfs3lookup_dnlc(vnode_t *, char *, vnode_t **, cred_t *);
98 static int	nfs3lookup_otw(vnode_t *, char *, vnode_t **, cred_t *, int);
99 static int	nfs3create(vnode_t *, char *, struct vattr *, enum vcexcl,
100 			int, vnode_t **, cred_t *, int);
101 static int	nfs3excl_create_settimes(vnode_t *, struct vattr *, cred_t *);
102 static int	nfs3mknod(vnode_t *, char *, struct vattr *, enum vcexcl,
103 			int, vnode_t **, cred_t *);
104 static int	nfs3rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
105 static int	do_nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
106 static void	nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
107 static void	nfs3readdirplus(vnode_t *, rddir_cache *, cred_t *);
108 static int	nfs3_bio(struct buf *, stable_how *, cred_t *);
109 static int	nfs3_getapage(vnode_t *, u_offset_t, size_t, uint_t *,
110 			page_t *[], size_t, struct seg *, caddr_t,
111 			enum seg_rw, cred_t *);
112 static void	nfs3_readahead(vnode_t *, u_offset_t, caddr_t, struct seg *,
113 			cred_t *);
114 static int	nfs3_sync_putapage(vnode_t *, page_t *, u_offset_t, size_t,
115 			int, cred_t *);
116 static int	nfs3_sync_pageio(vnode_t *, page_t *, u_offset_t, size_t,
117 			int, cred_t *);
118 static int	nfs3_commit(vnode_t *, offset3, count3, cred_t *);
119 static void	nfs3_set_mod(vnode_t *);
120 static void	nfs3_get_commit(vnode_t *);
121 static void	nfs3_get_commit_range(vnode_t *, u_offset_t, size_t);
122 #if 0 /* unused */
123 #ifdef DEBUG
124 static int	nfs3_no_uncommitted_pages(vnode_t *);
125 #endif
126 #endif /* unused */
127 static int	nfs3_putpage_commit(vnode_t *, offset_t, size_t, cred_t *);
128 static int	nfs3_commit_vp(vnode_t *, u_offset_t, size_t,  cred_t *);
129 static int	nfs3_sync_commit(vnode_t *, page_t *, offset3, count3,
130 			cred_t *);
131 static void	nfs3_async_commit(vnode_t *, page_t *, offset3, count3,
132 			cred_t *);
133 static void	nfs3_delmap_callback(struct as *, void *, uint_t);
134 
135 /*
136  * Error flags used to pass information about certain special errors
137  * which need to be handled specially.
138  */
139 #define	NFS_EOF			-98
140 #define	NFS_VERF_MISMATCH	-97
141 
142 /* ALIGN64 aligns the given buffer and adjust buffer size to 64 bit */
143 #define	ALIGN64(x, ptr, sz)						\
144 	x = ((uintptr_t)(ptr)) & (sizeof (uint64_t) - 1);		\
145 	if (x) {							\
146 		x = sizeof (uint64_t) - (x);				\
147 		sz -= (x);						\
148 		ptr += (x);						\
149 	}
150 
151 /*
152  * These are the vnode ops routines which implement the vnode interface to
153  * the networked file system.  These routines just take their parameters,
154  * make them look networkish by putting the right info into interface structs,
155  * and then calling the appropriate remote routine(s) to do the work.
156  *
157  * Note on directory name lookup cacheing:  If we detect a stale fhandle,
158  * we purge the directory cache relative to that vnode.  This way, the
159  * user won't get burned by the cache repeatedly.  See <nfs/rnode.h> for
160  * more details on rnode locking.
161  */
162 
163 static int	nfs3_open(vnode_t **, int, cred_t *);
164 static int	nfs3_close(vnode_t *, int, int, offset_t, cred_t *);
165 static int	nfs3_read(vnode_t *, struct uio *, int, cred_t *,
166 			caller_context_t *);
167 static int	nfs3_write(vnode_t *, struct uio *, int, cred_t *,
168 			caller_context_t *);
169 static int	nfs3_ioctl(vnode_t *, int, intptr_t, int, cred_t *, int *);
170 static int	nfs3_getattr(vnode_t *, struct vattr *, int, cred_t *);
171 static int	nfs3_setattr(vnode_t *, struct vattr *, int, cred_t *,
172 			caller_context_t *);
173 static int	nfs3_access(vnode_t *, int, int, cred_t *);
174 static int	nfs3_readlink(vnode_t *, struct uio *, cred_t *);
175 static int	nfs3_fsync(vnode_t *, int, cred_t *);
176 static void	nfs3_inactive(vnode_t *, cred_t *);
177 static int	nfs3_lookup(vnode_t *, char *, vnode_t **,
178 			struct pathname *, int, vnode_t *, cred_t *);
179 static int	nfs3_create(vnode_t *, char *, struct vattr *, enum vcexcl,
180 			int, vnode_t **, cred_t *, int);
181 static int	nfs3_remove(vnode_t *, char *, cred_t *);
182 static int	nfs3_link(vnode_t *, vnode_t *, char *, cred_t *);
183 static int	nfs3_rename(vnode_t *, char *, vnode_t *, char *, cred_t *);
184 static int	nfs3_mkdir(vnode_t *, char *, struct vattr *,
185 			vnode_t **, cred_t *);
186 static int	nfs3_rmdir(vnode_t *, char *, vnode_t *, cred_t *);
187 static int	nfs3_symlink(vnode_t *, char *, struct vattr *, char *,
188 			cred_t *);
189 static int	nfs3_readdir(vnode_t *, struct uio *, cred_t *, int *);
190 static int	nfs3_fid(vnode_t *, fid_t *);
191 static int	nfs3_rwlock(vnode_t *, int, caller_context_t *);
192 static void	nfs3_rwunlock(vnode_t *, int, caller_context_t *);
193 static int	nfs3_seek(vnode_t *, offset_t, offset_t *);
194 static int	nfs3_getpage(vnode_t *, offset_t, size_t, uint_t *,
195 			page_t *[], size_t, struct seg *, caddr_t,
196 			enum seg_rw, cred_t *);
197 static int	nfs3_putpage(vnode_t *, offset_t, size_t, int, cred_t *);
198 static int	nfs3_map(vnode_t *, offset_t, struct as *, caddr_t *,
199 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
200 static int	nfs3_addmap(vnode_t *, offset_t, struct as *, caddr_t,
201 			size_t, uchar_t, uchar_t, uint_t, cred_t *);
202 static int	nfs3_frlock(vnode_t *, int, struct flock64 *, int, offset_t,
203 			struct flk_callback *, cred_t *);
204 static int	nfs3_space(vnode_t *, int, struct flock64 *, int, offset_t,
205 			cred_t *, caller_context_t *);
206 static int	nfs3_realvp(vnode_t *, vnode_t **);
207 static int	nfs3_delmap(vnode_t *, offset_t, struct as *, caddr_t,
208 			size_t, uint_t, uint_t, uint_t, cred_t *);
209 static int	nfs3_pathconf(vnode_t *, int, ulong_t *, cred_t *);
210 static int	nfs3_pageio(vnode_t *, page_t *, u_offset_t, size_t, int,
211 			cred_t *);
212 static void	nfs3_dispose(vnode_t *, page_t *, int, int, cred_t *);
213 static int	nfs3_setsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
214 static int	nfs3_getsecattr(vnode_t *, vsecattr_t *, int, cred_t *);
215 static int	nfs3_shrlock(vnode_t *, int, struct shrlock *, int, cred_t *);
216 
217 struct vnodeops *nfs3_vnodeops;
218 
219 const fs_operation_def_t nfs3_vnodeops_template[] = {
220 	VOPNAME_OPEN, nfs3_open,
221 	VOPNAME_CLOSE, nfs3_close,
222 	VOPNAME_READ, nfs3_read,
223 	VOPNAME_WRITE, nfs3_write,
224 	VOPNAME_IOCTL, nfs3_ioctl,
225 	VOPNAME_GETATTR, nfs3_getattr,
226 	VOPNAME_SETATTR, nfs3_setattr,
227 	VOPNAME_ACCESS, nfs3_access,
228 	VOPNAME_LOOKUP, nfs3_lookup,
229 	VOPNAME_CREATE, nfs3_create,
230 	VOPNAME_REMOVE, nfs3_remove,
231 	VOPNAME_LINK, nfs3_link,
232 	VOPNAME_RENAME, nfs3_rename,
233 	VOPNAME_MKDIR, nfs3_mkdir,
234 	VOPNAME_RMDIR, nfs3_rmdir,
235 	VOPNAME_READDIR, nfs3_readdir,
236 	VOPNAME_SYMLINK, nfs3_symlink,
237 	VOPNAME_READLINK, nfs3_readlink,
238 	VOPNAME_FSYNC, nfs3_fsync,
239 	VOPNAME_INACTIVE, (fs_generic_func_p) nfs3_inactive,
240 	VOPNAME_FID, nfs3_fid,
241 	VOPNAME_RWLOCK, nfs3_rwlock,
242 	VOPNAME_RWUNLOCK, (fs_generic_func_p) nfs3_rwunlock,
243 	VOPNAME_SEEK, nfs3_seek,
244 	VOPNAME_FRLOCK, nfs3_frlock,
245 	VOPNAME_SPACE, nfs3_space,
246 	VOPNAME_REALVP, nfs3_realvp,
247 	VOPNAME_GETPAGE, nfs3_getpage,
248 	VOPNAME_PUTPAGE, nfs3_putpage,
249 	VOPNAME_MAP, (fs_generic_func_p) nfs3_map,
250 	VOPNAME_ADDMAP, (fs_generic_func_p) nfs3_addmap,
251 	VOPNAME_DELMAP, nfs3_delmap,
252 	VOPNAME_DUMP, nfs_dump,		/* there is no separate nfs3_dump */
253 	VOPNAME_PATHCONF, nfs3_pathconf,
254 	VOPNAME_PAGEIO, nfs3_pageio,
255 	VOPNAME_DISPOSE, (fs_generic_func_p) nfs3_dispose,
256 	VOPNAME_SETSECATTR, nfs3_setsecattr,
257 	VOPNAME_GETSECATTR, nfs3_getsecattr,
258 	VOPNAME_SHRLOCK, nfs3_shrlock,
259 	NULL, NULL
260 };
261 
262 /*
263  * XXX:  This is referenced in modstubs.s
264  */
265 struct vnodeops *
266 nfs3_getvnodeops(void)
267 {
268 	return (nfs3_vnodeops);
269 }
270 
271 /* ARGSUSED */
272 static int
273 nfs3_open(vnode_t **vpp, int flag, cred_t *cr)
274 {
275 	int error;
276 	struct vattr va;
277 	rnode_t *rp;
278 	vnode_t *vp;
279 
280 	vp = *vpp;
281 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
282 		return (EIO);
283 	rp = VTOR(vp);
284 	mutex_enter(&rp->r_statelock);
285 	if (rp->r_cred == NULL) {
286 		crhold(cr);
287 		rp->r_cred = cr;
288 	}
289 	mutex_exit(&rp->r_statelock);
290 
291 	/*
292 	 * If there is no cached data or if close-to-open
293 	 * consistency checking is turned off, we can avoid
294 	 * the over the wire getattr.  Otherwise, if the
295 	 * file system is mounted readonly, then just verify
296 	 * the caches are up to date using the normal mechanism.
297 	 * Else, if the file is not mmap'd, then just mark
298 	 * the attributes as timed out.  They will be refreshed
299 	 * and the caches validated prior to being used.
300 	 * Else, the file system is mounted writeable so
301 	 * force an over the wire GETATTR in order to ensure
302 	 * that all cached data is valid.
303 	 */
304 	if (vp->v_count > 1 ||
305 	    ((vn_has_cached_data(vp) || HAVE_RDDIR_CACHE(rp)) &&
306 	    !(VTOMI(vp)->mi_flags & MI_NOCTO))) {
307 		if (vn_is_readonly(vp))
308 			error = nfs3_validate_caches(vp, cr);
309 		else if (rp->r_mapcnt == 0 && vp->v_count == 1) {
310 			PURGE_ATTRCACHE(vp);
311 			error = 0;
312 		} else {
313 			va.va_mask = AT_ALL;
314 			error = nfs3_getattr_otw(vp, &va, cr);
315 		}
316 	} else
317 		error = 0;
318 
319 	return (error);
320 }
321 
322 static int
323 nfs3_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr)
324 {
325 	rnode_t *rp;
326 	int error;
327 	struct vattr va;
328 
329 	/*
330 	 * zone_enter(2) prevents processes from changing zones with NFS files
331 	 * open; if we happen to get here from the wrong zone we can't do
332 	 * anything over the wire.
333 	 */
334 	if (VTOMI(vp)->mi_zone != curproc->p_zone) {
335 		/*
336 		 * We could attempt to clean up locks, except we're sure
337 		 * that the current process didn't acquire any locks on
338 		 * the file: any attempt to lock a file belong to another zone
339 		 * will fail, and one can't lock an NFS file and then change
340 		 * zones, as that fails too.
341 		 *
342 		 * Returning an error here is the sane thing to do.  A
343 		 * subsequent call to VN_RELE() which translates to a
344 		 * nfs3_inactive() will clean up state: if the zone of the
345 		 * vnode's origin is still alive and kicking, an async worker
346 		 * thread will handle the request (from the correct zone), and
347 		 * everything (minus the commit and final nfs3_getattr_otw()
348 		 * call) should be OK. If the zone is going away
349 		 * nfs_async_inactive() will throw away cached pages inline.
350 		 */
351 		return (EIO);
352 	}
353 
354 	/*
355 	 * If we are using local locking for this filesystem, then
356 	 * release all of the SYSV style record locks.  Otherwise,
357 	 * we are doing network locking and we need to release all
358 	 * of the network locks.  All of the locks held by this
359 	 * process on this file are released no matter what the
360 	 * incoming reference count is.
361 	 */
362 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
363 		cleanlocks(vp, ttoproc(curthread)->p_pid, 0);
364 		cleanshares(vp, ttoproc(curthread)->p_pid);
365 	} else
366 		nfs_lockrelease(vp, flag, offset, cr);
367 
368 	if (count > 1)
369 		return (0);
370 
371 	/*
372 	 * If the file has been `unlinked', then purge the
373 	 * DNLC so that this vnode will get reycled quicker
374 	 * and the .nfs* file on the server will get removed.
375 	 */
376 	rp = VTOR(vp);
377 	if (rp->r_unldvp != NULL)
378 		dnlc_purge_vp(vp);
379 
380 	/*
381 	 * If the file was open for write and there are pages,
382 	 * then if the file system was mounted using the "no-close-
383 	 *	to-open" semantics, then start an asynchronous flush
384 	 *	of the all of the pages in the file.
385 	 * else the file system was not mounted using the "no-close-
386 	 *	to-open" semantics, then do a synchronous flush and
387 	 *	commit of all of the dirty and uncommitted pages.
388 	 *
389 	 * The asynchronous flush of the pages in the "nocto" path
390 	 * mostly just associates a cred pointer with the rnode so
391 	 * writes which happen later will have a better chance of
392 	 * working.  It also starts the data being written to the
393 	 * server, but without unnecessarily delaying the application.
394 	 */
395 	if ((flag & FWRITE) && vn_has_cached_data(vp)) {
396 		if (VTOMI(vp)->mi_flags & MI_NOCTO) {
397 			error = nfs3_putpage(vp, (offset_t)0, 0, B_ASYNC, cr);
398 			if (error == EAGAIN)
399 				error = 0;
400 		} else
401 			error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
402 		if (!error) {
403 			mutex_enter(&rp->r_statelock);
404 			error = rp->r_error;
405 			rp->r_error = 0;
406 			mutex_exit(&rp->r_statelock);
407 		}
408 	} else {
409 		mutex_enter(&rp->r_statelock);
410 		error = rp->r_error;
411 		rp->r_error = 0;
412 		mutex_exit(&rp->r_statelock);
413 	}
414 
415 	/*
416 	 * If RWRITEATTR is set, then issue an over the wire GETATTR to
417 	 * refresh the attribute cache with a set of attributes which
418 	 * weren't returned from a WRITE.  This will enable the close-
419 	 * to-open processing to work.
420 	 */
421 	if (rp->r_flags & RWRITEATTR)
422 		(void) nfs3_getattr_otw(vp, &va, cr);
423 
424 	return (error);
425 }
426 
427 /* ARGSUSED */
428 static int
429 nfs3_directio_read(vnode_t *vp, struct uio *uiop, cred_t *cr)
430 {
431 	mntinfo_t *mi;
432 	READ3args args;
433 	READ3uiores res;
434 	int tsize;
435 	offset_t offset;
436 	ssize_t count;
437 	int error;
438 	int douprintf;
439 	failinfo_t fi;
440 	char *sv_hostname;
441 
442 	mi = VTOMI(vp);
443 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
444 	sv_hostname = VTOR(vp)->r_server->sv_hostname;
445 
446 	douprintf = 1;
447 	args.file = *VTOFH3(vp);
448 	fi.vp = vp;
449 	fi.fhp = (caddr_t)&args.file;
450 	fi.copyproc = nfs3copyfh;
451 	fi.lookupproc = nfs3lookup;
452 	fi.xattrdirproc = acl_getxattrdir3;
453 
454 	res.uiop = uiop;
455 
456 	offset = uiop->uio_loffset;
457 	count = uiop->uio_resid;
458 
459 	do {
460 		if (mi->mi_io_kstats) {
461 			mutex_enter(&mi->mi_lock);
462 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
463 			mutex_exit(&mi->mi_lock);
464 		}
465 
466 		do {
467 			tsize = MIN(mi->mi_tsize, count);
468 			args.offset = (offset3)offset;
469 			args.count = (count3)tsize;
470 			res.size = (uint_t)tsize;
471 			error = rfs3call(mi, NFSPROC3_READ,
472 				    xdr_READ3args, (caddr_t)&args,
473 				    xdr_READ3uiores, (caddr_t)&res, cr,
474 				    &douprintf, &res.status, 0, &fi);
475 		} while (error == ENFS_TRYAGAIN);
476 
477 		if (mi->mi_io_kstats) {
478 			mutex_enter(&mi->mi_lock);
479 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
480 			mutex_exit(&mi->mi_lock);
481 		}
482 
483 		if (error)
484 			return (error);
485 
486 		error = geterrno3(res.status);
487 		if (error)
488 			return (error);
489 
490 		if (res.count != res.size) {
491 			zcmn_err(getzoneid(), CE_WARN,
492 "nfs3_directio_read: server %s returned incorrect amount",
493 					sv_hostname);
494 			return (EIO);
495 		}
496 		count -= res.count;
497 		offset += res.count;
498 		if (mi->mi_io_kstats) {
499 			mutex_enter(&mi->mi_lock);
500 			KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
501 			KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
502 			mutex_exit(&mi->mi_lock);
503 		}
504 		lwp_stat_update(LWP_STAT_INBLK, 1);
505 	} while (count && !res.eof);
506 
507 	return (0);
508 }
509 
510 /* ARGSUSED */
511 static int
512 nfs3_read(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
513 	caller_context_t *ct)
514 {
515 	rnode_t *rp;
516 	u_offset_t off;
517 	offset_t diff;
518 	int on;
519 	size_t n;
520 	caddr_t base;
521 	uint_t flags;
522 	int error = 0;
523 	mntinfo_t *mi;
524 
525 	rp = VTOR(vp);
526 	mi = VTOMI(vp);
527 
528 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
529 
530 	if (curproc->p_zone != mi->mi_zone)
531 		return (EIO);
532 
533 	if (vp->v_type != VREG)
534 		return (EISDIR);
535 
536 	if (uiop->uio_resid == 0)
537 		return (0);
538 
539 	if (uiop->uio_loffset < 0 || uiop->uio_loffset + uiop->uio_resid < 0)
540 		return (EINVAL);
541 
542 	/*
543 	 * Bypass VM if caching has been disabled (e.g., locking) or if
544 	 * using client-side direct I/O and the file is not mmap'd and
545 	 * there are no cached pages.
546 	 */
547 	if ((vp->v_flag & VNOCACHE) ||
548 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
549 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
550 		return (nfs3_directio_read(vp, uiop, cr));
551 	}
552 
553 	do {
554 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
555 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
556 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
557 
558 		error = nfs3_validate_caches(vp, cr);
559 		if (error)
560 			break;
561 
562 		mutex_enter(&rp->r_statelock);
563 		diff = rp->r_size - uiop->uio_loffset;
564 		mutex_exit(&rp->r_statelock);
565 		if (diff <= 0)
566 			break;
567 		if (diff < n)
568 			n = (size_t)diff;
569 
570 		base = segmap_getmapflt(segkmap, vp, off + on, n, 1, S_READ);
571 
572 		error = uiomove(base + on, n, UIO_READ, uiop);
573 
574 		if (!error) {
575 			/*
576 			 * If read a whole block or read to eof,
577 			 * won't need this buffer again soon.
578 			 */
579 			mutex_enter(&rp->r_statelock);
580 			if (n + on == MAXBSIZE ||
581 			    uiop->uio_loffset == rp->r_size)
582 				flags = SM_DONTNEED;
583 			else
584 				flags = 0;
585 			mutex_exit(&rp->r_statelock);
586 			error = segmap_release(segkmap, base, flags);
587 		} else
588 			(void) segmap_release(segkmap, base, 0);
589 	} while (!error && uiop->uio_resid > 0);
590 
591 	return (error);
592 }
593 
594 /* ARGSUSED */
595 static int
596 nfs3_write(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
597 	caller_context_t *ct)
598 {
599 	rlim64_t limit = uiop->uio_llimit;
600 	rnode_t *rp;
601 	u_offset_t off;
602 	caddr_t base;
603 	uint_t flags;
604 	int remainder;
605 	size_t n;
606 	int on;
607 	int error;
608 	int resid;
609 	offset_t offset;
610 	mntinfo_t *mi;
611 	uint_t bsize;
612 
613 	rp = VTOR(vp);
614 
615 	if (vp->v_type != VREG)
616 		return (EISDIR);
617 
618 	mi = VTOMI(vp);
619 	if (curproc->p_zone != mi->mi_zone)
620 		return (EIO);
621 	if (uiop->uio_resid == 0)
622 		return (0);
623 
624 	if (ioflag & FAPPEND) {
625 		struct vattr va;
626 
627 		/*
628 		 * Must serialize if appending.
629 		 */
630 		if (nfs_rw_lock_held(&rp->r_rwlock, RW_READER)) {
631 			nfs_rw_exit(&rp->r_rwlock);
632 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER,
633 			    INTR(vp)))
634 				return (EINTR);
635 		}
636 
637 		va.va_mask = AT_SIZE;
638 		error = nfs3getattr(vp, &va, cr);
639 		if (error)
640 			return (error);
641 		uiop->uio_loffset = va.va_size;
642 	}
643 
644 	offset = uiop->uio_loffset + uiop->uio_resid;
645 
646 	if (uiop->uio_loffset < 0 || offset < 0)
647 		return (EINVAL);
648 
649 	if (limit == RLIM64_INFINITY || limit > MAXOFFSET_T)
650 		limit = MAXOFFSET_T;
651 
652 	/*
653 	 * Check to make sure that the process will not exceed
654 	 * its limit on file size.  It is okay to write up to
655 	 * the limit, but not beyond.  Thus, the write which
656 	 * reaches the limit will be short and the next write
657 	 * will return an error.
658 	 */
659 	remainder = 0;
660 	if (offset > limit) {
661 		remainder = offset - limit;
662 		uiop->uio_resid = limit - uiop->uio_loffset;
663 		if (uiop->uio_resid <= 0) {
664 			proc_t *p = ttoproc(curthread);
665 
666 			uiop->uio_resid += remainder;
667 			mutex_enter(&p->p_lock);
668 			(void) rctl_action(rctlproc_legacy[RLIMIT_FSIZE],
669 			    p->p_rctls, p, RCA_UNSAFE_SIGINFO);
670 			mutex_exit(&p->p_lock);
671 			return (EFBIG);
672 		}
673 	}
674 
675 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
676 		return (EINTR);
677 
678 	/*
679 	 * Bypass VM if caching has been disabled (e.g., locking) or if
680 	 * using client-side direct I/O and the file is not mmap'd and
681 	 * there are no cached pages.
682 	 */
683 	if ((vp->v_flag & VNOCACHE) ||
684 	    (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
685 	    rp->r_mapcnt == 0 && !vn_has_cached_data(vp))) {
686 		size_t bufsize;
687 		int count;
688 		u_offset_t org_offset;
689 		stable_how stab_comm;
690 
691 nfs3_fwrite:
692 		if (rp->r_flags & RSTALE) {
693 			resid = uiop->uio_resid;
694 			offset = uiop->uio_loffset;
695 			error = rp->r_error;
696 			goto bottom;
697 		}
698 		bufsize = MIN(uiop->uio_resid, mi->mi_stsize);
699 		base = kmem_alloc(bufsize, KM_SLEEP);
700 		do {
701 			if (ioflag & FDSYNC)
702 				stab_comm = DATA_SYNC;
703 			else
704 				stab_comm = FILE_SYNC;
705 			resid = uiop->uio_resid;
706 			offset = uiop->uio_loffset;
707 			count = MIN(uiop->uio_resid, bufsize);
708 			org_offset = uiop->uio_loffset;
709 			error = uiomove(base, count, UIO_WRITE, uiop);
710 			if (!error) {
711 				error = nfs3write(vp, base, org_offset,
712 				    count, cr, &stab_comm);
713 			}
714 		} while (!error && uiop->uio_resid > 0);
715 		kmem_free(base, bufsize);
716 		goto bottom;
717 	}
718 
719 
720 	bsize = vp->v_vfsp->vfs_bsize;
721 
722 	do {
723 		off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
724 		on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
725 		n = MIN(MAXBSIZE - on, uiop->uio_resid);
726 
727 		resid = uiop->uio_resid;
728 		offset = uiop->uio_loffset;
729 
730 		if (rp->r_flags & RSTALE) {
731 			error = rp->r_error;
732 			break;
733 		}
734 
735 		/*
736 		 * Don't create dirty pages faster than they
737 		 * can be cleaned so that the system doesn't
738 		 * get imbalanced.  If the async queue is
739 		 * maxed out, then wait for it to drain before
740 		 * creating more dirty pages.  Also, wait for
741 		 * any threads doing pagewalks in the vop_getattr
742 		 * entry points so that they don't block for
743 		 * long periods.
744 		 */
745 		mutex_enter(&rp->r_statelock);
746 		while ((mi->mi_max_threads != 0 &&
747 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
748 		    rp->r_gcount > 0)
749 			cv_wait(&rp->r_cv, &rp->r_statelock);
750 		mutex_exit(&rp->r_statelock);
751 
752 		if (segmap_kpm) {
753 			int pon = uiop->uio_loffset & PAGEOFFSET;
754 			size_t pn = MIN(PAGESIZE - pon, uiop->uio_resid);
755 			int pagecreate;
756 
757 			mutex_enter(&rp->r_statelock);
758 			pagecreate = (pon == 0) && (pn == PAGESIZE ||
759 				uiop->uio_loffset + pn >= rp->r_size);
760 			mutex_exit(&rp->r_statelock);
761 
762 			base = segmap_getmapflt(segkmap, vp, off + on,
763 						pn, !pagecreate, S_WRITE);
764 
765 			error = writerp(rp, base + pon, n, uiop, pagecreate);
766 
767 		} else {
768 			base = segmap_getmapflt(segkmap, vp, off + on,
769 						n, 0, S_READ);
770 			error = writerp(rp, base + on, n, uiop, 0);
771 		}
772 
773 		if (!error) {
774 			if (mi->mi_flags & MI_NOAC)
775 				flags = SM_WRITE;
776 			else if ((uiop->uio_loffset % bsize) == 0 ||
777 			    IS_SWAPVP(vp)) {
778 				/*
779 				 * Have written a whole block.
780 				 * Start an asynchronous write
781 				 * and mark the buffer to
782 				 * indicate that it won't be
783 				 * needed again soon.
784 				 */
785 				flags = SM_WRITE | SM_ASYNC | SM_DONTNEED;
786 			} else
787 				flags = 0;
788 			if ((ioflag & (FSYNC|FDSYNC)) ||
789 			    (rp->r_flags & ROUTOFSPACE)) {
790 				flags &= ~SM_ASYNC;
791 				flags |= SM_WRITE;
792 			}
793 			error = segmap_release(segkmap, base, flags);
794 		} else {
795 			(void) segmap_release(segkmap, base, 0);
796 			/*
797 			 * In the event that we got an access error while
798 			 * faulting in a page for a write-only file just
799 			 * force a write.
800 			 */
801 			if (error == EACCES)
802 				goto nfs3_fwrite;
803 		}
804 	} while (!error && uiop->uio_resid > 0);
805 
806 bottom:
807 	if (error) {
808 		uiop->uio_resid = resid + remainder;
809 		uiop->uio_loffset = offset;
810 	} else
811 		uiop->uio_resid += remainder;
812 
813 	nfs_rw_exit(&rp->r_lkserlock);
814 
815 	return (error);
816 }
817 
818 /*
819  * Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED}
820  */
821 static int
822 nfs3_rdwrlbn(vnode_t *vp, page_t *pp, u_offset_t off, size_t len,
823 	int flags, cred_t *cr)
824 {
825 	struct buf *bp;
826 	int error;
827 	page_t *savepp;
828 	uchar_t fsdata;
829 	stable_how stab_comm;
830 
831 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
832 	bp = pageio_setup(pp, len, vp, flags);
833 	ASSERT(bp != NULL);
834 
835 	/*
836 	 * pageio_setup should have set b_addr to 0.  This
837 	 * is correct since we want to do I/O on a page
838 	 * boundary.  bp_mapin will use this addr to calculate
839 	 * an offset, and then set b_addr to the kernel virtual
840 	 * address it allocated for us.
841 	 */
842 	ASSERT(bp->b_un.b_addr == 0);
843 
844 	bp->b_edev = 0;
845 	bp->b_dev = 0;
846 	bp->b_lblkno = lbtodb(off);
847 	bp->b_file = vp;
848 	bp->b_offset = (offset_t)off;
849 	bp_mapin(bp);
850 
851 	/*
852 	 * Calculate the desired level of stability to write data
853 	 * on the server and then mark all of the pages to reflect
854 	 * this.
855 	 */
856 	if ((flags & (B_WRITE|B_ASYNC)) == (B_WRITE|B_ASYNC) &&
857 	    freemem > desfree) {
858 		stab_comm = UNSTABLE;
859 		fsdata = C_DELAYCOMMIT;
860 	} else {
861 		stab_comm = FILE_SYNC;
862 		fsdata = C_NOCOMMIT;
863 	}
864 
865 	savepp = pp;
866 	do {
867 		pp->p_fsdata = fsdata;
868 	} while ((pp = pp->p_next) != savepp);
869 
870 	error = nfs3_bio(bp, &stab_comm, cr);
871 
872 	bp_mapout(bp);
873 	pageio_done(bp);
874 
875 	/*
876 	 * If the server wrote pages in a more stable fashion than
877 	 * was requested, then clear all of the marks in the pages
878 	 * indicating that COMMIT operations were required.
879 	 */
880 	if (stab_comm != UNSTABLE && fsdata == C_DELAYCOMMIT) {
881 		do {
882 			pp->p_fsdata = C_NOCOMMIT;
883 		} while ((pp = pp->p_next) != savepp);
884 	}
885 
886 	return (error);
887 }
888 
889 /*
890  * Write to file.  Writes to remote server in largest size
891  * chunks that the server can handle.  Write is synchronous.
892  */
893 static int
894 nfs3write(vnode_t *vp, caddr_t base, u_offset_t offset, int count, cred_t *cr,
895 	stable_how *stab_comm)
896 {
897 	mntinfo_t *mi;
898 	WRITE3args args;
899 	WRITE3res res;
900 	int error;
901 	int tsize;
902 	rnode_t *rp;
903 	int douprintf;
904 
905 	rp = VTOR(vp);
906 	mi = VTOMI(vp);
907 
908 	ASSERT(curproc->p_zone == mi->mi_zone);
909 
910 	args.file = *VTOFH3(vp);
911 	args.stable = *stab_comm;
912 
913 	*stab_comm = FILE_SYNC;
914 
915 	douprintf = 1;
916 
917 	do {
918 		if ((vp->v_flag & VNOCACHE) ||
919 		    (rp->r_flags & RDIRECTIO) ||
920 		    (mi->mi_flags & MI_DIRECTIO))
921 			tsize = MIN(mi->mi_stsize, count);
922 		else
923 			tsize = MIN(mi->mi_curwrite, count);
924 		args.offset = (offset3)offset;
925 		args.count = (count3)tsize;
926 		args.data.data_len = (uint_t)tsize;
927 		args.data.data_val = base;
928 
929 		if (mi->mi_io_kstats) {
930 			mutex_enter(&mi->mi_lock);
931 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
932 			mutex_exit(&mi->mi_lock);
933 		}
934 		args.mblk = NULL;
935 		do {
936 			error = rfs3call(mi, NFSPROC3_WRITE,
937 			    xdr_WRITE3args, (caddr_t)&args,
938 			    xdr_WRITE3res, (caddr_t)&res, cr,
939 			    &douprintf, &res.status, 0, NULL);
940 		} while (error == ENFS_TRYAGAIN);
941 		if (mi->mi_io_kstats) {
942 			mutex_enter(&mi->mi_lock);
943 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
944 			mutex_exit(&mi->mi_lock);
945 		}
946 
947 		if (error)
948 			return (error);
949 		error = geterrno3(res.status);
950 		if (!error) {
951 			if (res.resok.count > args.count) {
952 				zcmn_err(getzoneid(), CE_WARN,
953 				    "nfs3write: server %s wrote %u, "
954 				    "requested was %u",
955 				    rp->r_server->sv_hostname,
956 				    res.resok.count, args.count);
957 				return (EIO);
958 			}
959 			if (res.resok.committed == UNSTABLE) {
960 				*stab_comm = UNSTABLE;
961 				if (args.stable == DATA_SYNC ||
962 				    args.stable == FILE_SYNC) {
963 					zcmn_err(getzoneid(), CE_WARN,
964 			"nfs3write: server %s did not commit to stable storage",
965 					    rp->r_server->sv_hostname);
966 					return (EIO);
967 				}
968 			}
969 			tsize = (int)res.resok.count;
970 			count -= tsize;
971 			base += tsize;
972 			offset += tsize;
973 			if (mi->mi_io_kstats) {
974 				mutex_enter(&mi->mi_lock);
975 				KSTAT_IO_PTR(mi->mi_io_kstats)->writes++;
976 				KSTAT_IO_PTR(mi->mi_io_kstats)->nwritten +=
977 				    tsize;
978 				mutex_exit(&mi->mi_lock);
979 			}
980 			lwp_stat_update(LWP_STAT_OUBLK, 1);
981 			mutex_enter(&rp->r_statelock);
982 			if (rp->r_flags & RHAVEVERF) {
983 				if (rp->r_verf != res.resok.verf) {
984 					nfs3_set_mod(vp);
985 					rp->r_verf = res.resok.verf;
986 					/*
987 					 * If the data was written UNSTABLE,
988 					 * then might as well stop because
989 					 * the whole block will have to get
990 					 * rewritten anyway.
991 					 */
992 					if (*stab_comm == UNSTABLE) {
993 						mutex_exit(&rp->r_statelock);
994 						break;
995 					}
996 				}
997 			} else {
998 				rp->r_verf = res.resok.verf;
999 				rp->r_flags |= RHAVEVERF;
1000 			}
1001 			/*
1002 			 * Mark the attribute cache as timed out and
1003 			 * set RWRITEATTR to indicate that the file
1004 			 * was modified with a WRITE operation and
1005 			 * that the attributes can not be trusted.
1006 			 */
1007 			PURGE_ATTRCACHE_LOCKED(rp);
1008 			rp->r_flags |= RWRITEATTR;
1009 			mutex_exit(&rp->r_statelock);
1010 		}
1011 	} while (!error && count);
1012 
1013 	return (error);
1014 }
1015 
1016 /*
1017  * Read from a file.  Reads data in largest chunks our interface can handle.
1018  */
1019 static int
1020 nfs3read(vnode_t *vp, caddr_t base, offset_t offset, int count,
1021 	size_t *residp, cred_t *cr)
1022 {
1023 	mntinfo_t *mi;
1024 	READ3args args;
1025 	READ3vres res;
1026 	int tsize;
1027 	int error;
1028 	int douprintf;
1029 	failinfo_t fi;
1030 	rnode_t *rp;
1031 	struct vattr va;
1032 	hrtime_t t;
1033 
1034 	rp = VTOR(vp);
1035 	mi = VTOMI(vp);
1036 	ASSERT(curproc->p_zone == mi->mi_zone);
1037 	douprintf = 1;
1038 
1039 	args.file = *VTOFH3(vp);
1040 	fi.vp = vp;
1041 	fi.fhp = (caddr_t)&args.file;
1042 	fi.copyproc = nfs3copyfh;
1043 	fi.lookupproc = nfs3lookup;
1044 	fi.xattrdirproc = acl_getxattrdir3;
1045 
1046 	res.pov.fres.vp = vp;
1047 	res.pov.fres.vap = &va;
1048 
1049 	*residp = count;
1050 	do {
1051 		if (mi->mi_io_kstats) {
1052 			mutex_enter(&mi->mi_lock);
1053 			kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1054 			mutex_exit(&mi->mi_lock);
1055 		}
1056 
1057 		do {
1058 			if ((vp->v_flag & VNOCACHE) ||
1059 			    (rp->r_flags & RDIRECTIO) ||
1060 			    (mi->mi_flags & MI_DIRECTIO))
1061 				tsize = MIN(mi->mi_tsize, count);
1062 			else
1063 				tsize = MIN(mi->mi_curread, count);
1064 			res.data.data_val = base;
1065 			res.data.data_len = tsize;
1066 			args.offset = (offset3)offset;
1067 			args.count = (count3)tsize;
1068 			t = gethrtime();
1069 			error = rfs3call(mi, NFSPROC3_READ,
1070 			    xdr_READ3args, (caddr_t)&args,
1071 			    xdr_READ3vres, (caddr_t)&res, cr,
1072 			    &douprintf, &res.status, 0, &fi);
1073 		} while (error == ENFS_TRYAGAIN);
1074 
1075 		if (mi->mi_io_kstats) {
1076 			mutex_enter(&mi->mi_lock);
1077 			kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
1078 			mutex_exit(&mi->mi_lock);
1079 		}
1080 
1081 		if (error)
1082 			return (error);
1083 
1084 		error = geterrno3(res.status);
1085 		if (error)
1086 			return (error);
1087 
1088 		if (res.count != res.data.data_len) {
1089 			zcmn_err(getzoneid(), CE_WARN,
1090 				"nfs3read: server %s returned incorrect amount",
1091 				rp->r_server->sv_hostname);
1092 			return (EIO);
1093 		}
1094 
1095 		count -= res.count;
1096 		*residp = count;
1097 		base += res.count;
1098 		offset += res.count;
1099 		if (mi->mi_io_kstats) {
1100 			mutex_enter(&mi->mi_lock);
1101 			KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
1102 			KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
1103 			mutex_exit(&mi->mi_lock);
1104 		}
1105 		lwp_stat_update(LWP_STAT_INBLK, 1);
1106 	} while (count && !res.eof);
1107 
1108 	if (res.pov.attributes) {
1109 		mutex_enter(&rp->r_statelock);
1110 		if (!CACHE_VALID(rp, va.va_mtime, va.va_size)) {
1111 			mutex_exit(&rp->r_statelock);
1112 			PURGE_ATTRCACHE(vp);
1113 		} else {
1114 			if (rp->r_mtime <= t)
1115 				nfs_attrcache_va(vp, &va);
1116 			mutex_exit(&rp->r_statelock);
1117 		}
1118 	}
1119 
1120 	return (0);
1121 }
1122 
1123 /* ARGSUSED */
1124 static int
1125 nfs3_ioctl(vnode_t *vp, int cmd, intptr_t arg, int flag, cred_t *cr, int *rvalp)
1126 {
1127 
1128 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
1129 		return (EIO);
1130 	switch (cmd) {
1131 		case _FIODIRECTIO:
1132 			return (nfs_directio(vp, (int)arg, cr));
1133 		default:
1134 			return (ENOTTY);
1135 	}
1136 }
1137 
1138 static int
1139 nfs3_getattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1140 {
1141 	int error;
1142 	rnode_t *rp;
1143 
1144 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
1145 		return (EIO);
1146 	/*
1147 	 * If it has been specified that the return value will
1148 	 * just be used as a hint, and we are only being asked
1149 	 * for size, fsid or rdevid, then return the client's
1150 	 * notion of these values without checking to make sure
1151 	 * that the attribute cache is up to date.
1152 	 * The whole point is to avoid an over the wire GETATTR
1153 	 * call.
1154 	 */
1155 	rp = VTOR(vp);
1156 	if (flags & ATTR_HINT) {
1157 		if (vap->va_mask ==
1158 		    (vap->va_mask & (AT_SIZE | AT_FSID | AT_RDEV))) {
1159 			mutex_enter(&rp->r_statelock);
1160 			if (vap->va_mask | AT_SIZE)
1161 				vap->va_size = rp->r_size;
1162 			if (vap->va_mask | AT_FSID)
1163 				vap->va_fsid = rp->r_attr.va_fsid;
1164 			if (vap->va_mask | AT_RDEV)
1165 				vap->va_rdev = rp->r_attr.va_rdev;
1166 			mutex_exit(&rp->r_statelock);
1167 			return (0);
1168 		}
1169 	}
1170 
1171 	/*
1172 	 * Only need to flush pages if asking for the mtime
1173 	 * and if there any dirty pages or any outstanding
1174 	 * asynchronous (write) requests for this file.
1175 	 */
1176 	if (vap->va_mask & AT_MTIME) {
1177 		if (vn_has_cached_data(vp) &&
1178 		    ((rp->r_flags & RDIRTY) || rp->r_awcount > 0)) {
1179 			mutex_enter(&rp->r_statelock);
1180 			rp->r_gcount++;
1181 			mutex_exit(&rp->r_statelock);
1182 			error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1183 			mutex_enter(&rp->r_statelock);
1184 			if (error && (error == ENOSPC || error == EDQUOT)) {
1185 				if (!rp->r_error)
1186 					rp->r_error = error;
1187 			}
1188 			if (--rp->r_gcount == 0)
1189 				cv_broadcast(&rp->r_cv);
1190 			mutex_exit(&rp->r_statelock);
1191 		}
1192 	}
1193 
1194 	return (nfs3getattr(vp, vap, cr));
1195 }
1196 
1197 /*ARGSUSED4*/
1198 static int
1199 nfs3_setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr,
1200 		caller_context_t *ct)
1201 {
1202 	int error;
1203 	struct vattr va;
1204 
1205 	if (vap->va_mask & AT_NOSET)
1206 		return (EINVAL);
1207 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
1208 		return (EIO);
1209 
1210 	va.va_mask = AT_UID | AT_MODE;
1211 	error = nfs3getattr(vp, &va, cr);
1212 	if (error)
1213 		return (error);
1214 
1215 	error = secpolicy_vnode_setattr(cr, vp, vap, &va, flags, nfs3_accessx,
1216 		vp);
1217 	if (error)
1218 		return (error);
1219 
1220 	return (nfs3setattr(vp, vap, flags, cr));
1221 }
1222 
1223 static int
1224 nfs3setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1225 {
1226 	int error;
1227 	uint_t mask;
1228 	SETATTR3args args;
1229 	SETATTR3res res;
1230 	int douprintf;
1231 	rnode_t *rp;
1232 	struct vattr va;
1233 	mode_t omode;
1234 	vsecattr_t *vsp;
1235 	hrtime_t t;
1236 
1237 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
1238 	mask = vap->va_mask;
1239 
1240 	rp = VTOR(vp);
1241 
1242 	/*
1243 	 * Only need to flush pages if there are any pages and
1244 	 * if the file is marked as dirty in some fashion.  The
1245 	 * file must be flushed so that we can accurately
1246 	 * determine the size of the file and the cached data
1247 	 * after the SETATTR returns.  A file is considered to
1248 	 * be dirty if it is either marked with RDIRTY, has
1249 	 * outstanding i/o's active, or is mmap'd.  In this
1250 	 * last case, we can't tell whether there are dirty
1251 	 * pages, so we flush just to be sure.
1252 	 */
1253 	if (vn_has_cached_data(vp) &&
1254 	    ((rp->r_flags & RDIRTY) ||
1255 	    rp->r_count > 0 ||
1256 	    rp->r_mapcnt > 0)) {
1257 		ASSERT(vp->v_type != VCHR);
1258 		error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1259 		if (error && (error == ENOSPC || error == EDQUOT)) {
1260 			mutex_enter(&rp->r_statelock);
1261 			if (!rp->r_error)
1262 				rp->r_error = error;
1263 			mutex_exit(&rp->r_statelock);
1264 		}
1265 	}
1266 
1267 	args.object = *RTOFH3(rp);
1268 	/*
1269 	 * If the intent is for the server to set the times,
1270 	 * there is no point in have the mask indicating set mtime or
1271 	 * atime, because the vap values may be junk, and so result
1272 	 * in an overflow error. Remove these flags from the vap mask
1273 	 * before calling in this case, and restore them afterwards.
1274 	 */
1275 	if ((mask & (AT_ATIME | AT_MTIME)) && !(flags & ATTR_UTIME)) {
1276 		/* Use server times, so don't set the args time fields */
1277 		vap->va_mask &= ~(AT_ATIME | AT_MTIME);
1278 		error = vattr_to_sattr3(vap, &args.new_attributes);
1279 		vap->va_mask |= (mask & (AT_ATIME | AT_MTIME));
1280 		if (mask & AT_ATIME) {
1281 			args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
1282 		}
1283 		if (mask & AT_MTIME) {
1284 			args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
1285 		}
1286 	} else {
1287 		/* Either do not set times or use the client specified times */
1288 		error = vattr_to_sattr3(vap, &args.new_attributes);
1289 	}
1290 
1291 	if (error) {
1292 		/* req time field(s) overflow - return immediately */
1293 		return (error);
1294 	}
1295 
1296 	va.va_mask = AT_MODE | AT_CTIME;
1297 	error = nfs3getattr(vp, &va, cr);
1298 	if (error)
1299 		return (error);
1300 	omode = va.va_mode;
1301 
1302 tryagain:
1303 	if (mask & AT_SIZE) {
1304 		args.guard.check = TRUE;
1305 		args.guard.obj_ctime.seconds = va.va_ctime.tv_sec;
1306 		args.guard.obj_ctime.nseconds = va.va_ctime.tv_nsec;
1307 	} else
1308 		args.guard.check = FALSE;
1309 
1310 	douprintf = 1;
1311 
1312 	t = gethrtime();
1313 
1314 	error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
1315 	    xdr_SETATTR3args, (caddr_t)&args,
1316 	    xdr_SETATTR3res, (caddr_t)&res, cr,
1317 	    &douprintf, &res.status, 0, NULL);
1318 
1319 	/*
1320 	 * Purge the access cache and ACL cache if changing either the
1321 	 * owner of the file, the group owner, or the mode.  These may
1322 	 * change the access permissions of the file, so purge old
1323 	 * information and start over again.
1324 	 */
1325 	if (mask & (AT_UID | AT_GID | AT_MODE)) {
1326 		(void) nfs_access_purge_rp(rp);
1327 		if (rp->r_secattr != NULL) {
1328 			mutex_enter(&rp->r_statelock);
1329 			vsp = rp->r_secattr;
1330 			rp->r_secattr = NULL;
1331 			mutex_exit(&rp->r_statelock);
1332 			if (vsp != NULL)
1333 				nfs_acl_free(vsp);
1334 		}
1335 	}
1336 
1337 	if (error) {
1338 		PURGE_ATTRCACHE(vp);
1339 		return (error);
1340 	}
1341 
1342 	error = geterrno3(res.status);
1343 	if (!error) {
1344 		/*
1345 		 * If changing the size of the file, invalidate
1346 		 * any local cached data which is no longer part
1347 		 * of the file.  We also possibly invalidate the
1348 		 * last page in the file.  We could use
1349 		 * pvn_vpzero(), but this would mark the page as
1350 		 * modified and require it to be written back to
1351 		 * the server for no particularly good reason.
1352 		 * This way, if we access it, then we bring it
1353 		 * back in.  A read should be cheaper than a
1354 		 * write.
1355 		 */
1356 		if (mask & AT_SIZE) {
1357 			nfs_invalidate_pages(vp,
1358 			    (vap->va_size & PAGEMASK), cr);
1359 		}
1360 		nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
1361 		/*
1362 		 * Some servers will change the mode to clear the setuid
1363 		 * and setgid bits when changing the uid or gid.  The
1364 		 * client needs to compensate appropriately.
1365 		 */
1366 		if (mask & (AT_UID | AT_GID)) {
1367 			int terror;
1368 
1369 			va.va_mask = AT_MODE;
1370 			terror = nfs3getattr(vp, &va, cr);
1371 			if (!terror &&
1372 			    (((mask & AT_MODE) && va.va_mode != vap->va_mode) ||
1373 			    (!(mask & AT_MODE) && va.va_mode != omode))) {
1374 				va.va_mask = AT_MODE;
1375 				if (mask & AT_MODE)
1376 					va.va_mode = vap->va_mode;
1377 				else
1378 					va.va_mode = omode;
1379 				(void) nfs3setattr(vp, &va, 0, cr);
1380 			}
1381 		}
1382 	} else {
1383 		nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
1384 		/*
1385 		 * If we got back a "not synchronized" error, then
1386 		 * we need to retry with a new guard value.  The
1387 		 * guard value used is the change time.  If the
1388 		 * server returned post_op_attr, then we can just
1389 		 * retry because we have the latest attributes.
1390 		 * Otherwise, we issue a GETATTR to get the latest
1391 		 * attributes and then retry.  If we couldn't get
1392 		 * the attributes this way either, then we give
1393 		 * up because we can't complete the operation as
1394 		 * required.
1395 		 */
1396 		if (res.status == NFS3ERR_NOT_SYNC) {
1397 			va.va_mask = AT_CTIME;
1398 			if (nfs3getattr(vp, &va, cr) == 0)
1399 				goto tryagain;
1400 		}
1401 		PURGE_STALE_FH(error, vp, cr);
1402 	}
1403 
1404 	return (error);
1405 }
1406 
1407 static int
1408 nfs3_accessx(void *vp, int mode, cred_t *cr)
1409 {
1410 	ASSERT(curproc->p_zone == VTOMI((vnode_t *)vp)->mi_zone);
1411 	return (nfs3_access(vp, mode, 0, cr));
1412 }
1413 
1414 /* ARGSUSED */
1415 static int
1416 nfs3_access(vnode_t *vp, int mode, int flags, cred_t *cr)
1417 {
1418 	int error;
1419 	ACCESS3args args;
1420 	ACCESS3res res;
1421 	int douprintf;
1422 	uint32 acc;
1423 	rnode_t *rp;
1424 	cred_t *cred;
1425 	failinfo_t fi;
1426 	nfs_access_type_t cacc;
1427 	hrtime_t t;
1428 
1429 	acc = 0;
1430 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
1431 		return (EIO);
1432 	if (mode & VREAD)
1433 		acc |= ACCESS3_READ;
1434 	if (mode & VWRITE) {
1435 		if (vn_is_readonly(vp) && !IS_DEVVP(vp))
1436 			return (EROFS);
1437 		if (vp->v_type == VDIR)
1438 			acc |= ACCESS3_DELETE;
1439 		acc |= ACCESS3_MODIFY | ACCESS3_EXTEND;
1440 	}
1441 	if (mode & VEXEC) {
1442 		if (vp->v_type == VDIR)
1443 			acc |= ACCESS3_LOOKUP;
1444 		else
1445 			acc |= ACCESS3_EXECUTE;
1446 	}
1447 
1448 	rp = VTOR(vp);
1449 	if (rp->r_acache != NULL) {
1450 		cacc = nfs_access_check(rp, acc, cr);
1451 		if (cacc == NFS_ACCESS_ALLOWED)
1452 			return (0);
1453 		if (cacc == NFS_ACCESS_DENIED)
1454 			return (EACCES);
1455 	}
1456 
1457 	args.object = *VTOFH3(vp);
1458 	if (vp->v_type == VDIR) {
1459 		args.access = ACCESS3_READ | ACCESS3_DELETE | ACCESS3_MODIFY |
1460 		    ACCESS3_EXTEND | ACCESS3_LOOKUP;
1461 	} else {
1462 		args.access = ACCESS3_READ | ACCESS3_MODIFY | ACCESS3_EXTEND |
1463 		    ACCESS3_EXECUTE;
1464 	}
1465 	fi.vp = vp;
1466 	fi.fhp = (caddr_t)&args.object;
1467 	fi.copyproc = nfs3copyfh;
1468 	fi.lookupproc = nfs3lookup;
1469 	fi.xattrdirproc = acl_getxattrdir3;
1470 
1471 	cred = cr;
1472 tryagain:
1473 	douprintf = 1;
1474 
1475 	t = gethrtime();
1476 
1477 	error = rfs3call(VTOMI(vp), NFSPROC3_ACCESS,
1478 	    xdr_ACCESS3args, (caddr_t)&args,
1479 	    xdr_ACCESS3res, (caddr_t)&res, cred,
1480 	    &douprintf, &res.status, 0, &fi);
1481 
1482 	if (error) {
1483 		if (cred != cr)
1484 			crfree(cred);
1485 		return (error);
1486 	}
1487 
1488 	error = geterrno3(res.status);
1489 	if (!error) {
1490 		nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
1491 		if ((acc & res.resok.access) != acc) {
1492 			cred_t *ncr = crnetadjust(cred);
1493 
1494 			if (ncr != NULL) {
1495 				cred = ncr;
1496 				goto tryagain;
1497 			}
1498 			error = EACCES;
1499 		}
1500 		nfs_access_cache(rp, args.access, res.resok.access, cr);
1501 	} else {
1502 		nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
1503 		PURGE_STALE_FH(error, vp, cr);
1504 	}
1505 
1506 	if (cred != cr)
1507 		crfree(cred);
1508 
1509 	return (error);
1510 }
1511 
1512 static int nfs3_do_symlink_cache = 1;
1513 
1514 static int
1515 nfs3_readlink(vnode_t *vp, struct uio *uiop, cred_t *cr)
1516 {
1517 	int error;
1518 	READLINK3args args;
1519 	READLINK3res res;
1520 	nfspath3 resdata_backup;
1521 	rnode_t *rp;
1522 	int douprintf;
1523 	int len;
1524 	failinfo_t fi;
1525 	hrtime_t t;
1526 
1527 	/*
1528 	 * Can't readlink anything other than a symbolic link.
1529 	 */
1530 	if (vp->v_type != VLNK)
1531 		return (EINVAL);
1532 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
1533 		return (EIO);
1534 
1535 	rp = VTOR(vp);
1536 	if (nfs3_do_symlink_cache && rp->r_symlink.contents != NULL) {
1537 		error = nfs3_validate_caches(vp, cr);
1538 		if (error)
1539 			return (error);
1540 		mutex_enter(&rp->r_statelock);
1541 		if (rp->r_symlink.contents != NULL) {
1542 			error = uiomove(rp->r_symlink.contents,
1543 			    rp->r_symlink.len, UIO_READ, uiop);
1544 			mutex_exit(&rp->r_statelock);
1545 			return (error);
1546 		}
1547 		mutex_exit(&rp->r_statelock);
1548 	}
1549 
1550 	args.symlink = *VTOFH3(vp);
1551 	fi.vp = vp;
1552 	fi.fhp = (caddr_t)&args.symlink;
1553 	fi.copyproc = nfs3copyfh;
1554 	fi.lookupproc = nfs3lookup;
1555 	fi.xattrdirproc = acl_getxattrdir3;
1556 
1557 	res.resok.data = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1558 
1559 	resdata_backup = res.resok.data;
1560 
1561 	douprintf = 1;
1562 
1563 	t = gethrtime();
1564 
1565 	error = rfs3call(VTOMI(vp), NFSPROC3_READLINK,
1566 	    xdr_nfs_fh3, (caddr_t)&args,
1567 	    xdr_READLINK3res, (caddr_t)&res, cr,
1568 	    &douprintf, &res.status, 0, &fi);
1569 
1570 	if (res.resok.data == nfs3nametoolong)
1571 		error = EINVAL;
1572 
1573 	if (error) {
1574 		kmem_free(resdata_backup, MAXPATHLEN);
1575 		return (error);
1576 	}
1577 
1578 	error = geterrno3(res.status);
1579 	if (!error) {
1580 		nfs3_cache_post_op_attr(vp, &res.resok.symlink_attributes, t,
1581 		    cr);
1582 		len = strlen(res.resok.data);
1583 		error = uiomove(res.resok.data, len, UIO_READ, uiop);
1584 		if (nfs3_do_symlink_cache && rp->r_symlink.contents == NULL) {
1585 			mutex_enter(&rp->r_statelock);
1586 				if (rp->r_symlink.contents == NULL) {
1587 				rp->r_symlink.contents = res.resok.data;
1588 				rp->r_symlink.len = len;
1589 				rp->r_symlink.size = MAXPATHLEN;
1590 				mutex_exit(&rp->r_statelock);
1591 			} else {
1592 				mutex_exit(&rp->r_statelock);
1593 
1594 				kmem_free((void *)res.resok.data, MAXPATHLEN);
1595 			}
1596 		} else {
1597 			kmem_free((void *)res.resok.data, MAXPATHLEN);
1598 		}
1599 	} else {
1600 		nfs3_cache_post_op_attr(vp,
1601 		    &res.resfail.symlink_attributes, t, cr);
1602 		PURGE_STALE_FH(error, vp, cr);
1603 
1604 		kmem_free((void *)res.resok.data, MAXPATHLEN);
1605 
1606 	}
1607 
1608 	/*
1609 	 * The over the wire error for attempting to readlink something
1610 	 * other than a symbolic link is ENXIO.  However, we need to
1611 	 * return EINVAL instead of ENXIO, so we map it here.
1612 	 */
1613 	return (error == ENXIO ? EINVAL : error);
1614 }
1615 
1616 /*
1617  * Flush local dirty pages to stable storage on the server.
1618  *
1619  * If FNODSYNC is specified, then there is nothing to do because
1620  * metadata changes are not cached on the client before being
1621  * sent to the server.
1622  */
1623 static int
1624 nfs3_fsync(vnode_t *vp, int syncflag, cred_t *cr)
1625 {
1626 	int error;
1627 
1628 	if ((syncflag & FNODSYNC) || IS_SWAPVP(vp))
1629 		return (0);
1630 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
1631 		return (EIO);
1632 
1633 	error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
1634 	if (!error)
1635 		error = VTOR(vp)->r_error;
1636 	return (error);
1637 }
1638 
1639 /*
1640  * Weirdness: if the file was removed or the target of a rename
1641  * operation while it was open, it got renamed instead.  Here we
1642  * remove the renamed file.
1643  */
1644 static void
1645 nfs3_inactive(vnode_t *vp, cred_t *cr)
1646 {
1647 	rnode_t *rp;
1648 
1649 	ASSERT(vp != DNLC_NO_VNODE);
1650 
1651 	/*
1652 	 * If this is coming from the wrong zone, we let someone in the right
1653 	 * zone take care of it asynchronously.  We can get here due to
1654 	 * VN_RELE() being called from pageout() or fsflush().  This call may
1655 	 * potentially turn into an expensive no-op if, for instance, v_count
1656 	 * gets incremented in the meantime, but it's still correct.
1657 	 */
1658 	if (curproc->p_zone != VTOMI(vp)->mi_zone) {
1659 		nfs_async_inactive(vp, cr, nfs3_inactive);
1660 		return;
1661 	}
1662 
1663 	rp = VTOR(vp);
1664 redo:
1665 	if (rp->r_unldvp != NULL) {
1666 		/*
1667 		 * Save the vnode pointer for the directory where the
1668 		 * unlinked-open file got renamed, then set it to NULL
1669 		 * to prevent another thread from getting here before
1670 		 * we're done with the remove.  While we have the
1671 		 * statelock, make local copies of the pertinent rnode
1672 		 * fields.  If we weren't to do this in an atomic way, the
1673 		 * the unl* fields could become inconsistent with respect
1674 		 * to each other due to a race condition between this
1675 		 * code and nfs_remove().  See bug report 1034328.
1676 		 */
1677 		mutex_enter(&rp->r_statelock);
1678 		if (rp->r_unldvp != NULL) {
1679 			vnode_t *unldvp;
1680 			char *unlname;
1681 			cred_t *unlcred;
1682 			REMOVE3args args;
1683 			REMOVE3res res;
1684 			int douprintf;
1685 			int error;
1686 			hrtime_t t;
1687 
1688 			unldvp = rp->r_unldvp;
1689 			rp->r_unldvp = NULL;
1690 			unlname = rp->r_unlname;
1691 			rp->r_unlname = NULL;
1692 			unlcred = rp->r_unlcred;
1693 			rp->r_unlcred = NULL;
1694 			mutex_exit(&rp->r_statelock);
1695 
1696 			/*
1697 			 * If there are any dirty pages left, then flush
1698 			 * them.  This is unfortunate because they just
1699 			 * may get thrown away during the remove operation,
1700 			 * but we have to do this for correctness.
1701 			 */
1702 			if (vn_has_cached_data(vp) &&
1703 			    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
1704 				ASSERT(vp->v_type != VCHR);
1705 				error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
1706 				if (error) {
1707 					mutex_enter(&rp->r_statelock);
1708 					if (!rp->r_error)
1709 						rp->r_error = error;
1710 					mutex_exit(&rp->r_statelock);
1711 				}
1712 			}
1713 
1714 			/*
1715 			 * Do the remove operation on the renamed file
1716 			 */
1717 			setdiropargs3(&args.object, unlname, unldvp);
1718 
1719 			douprintf = 1;
1720 
1721 			t = gethrtime();
1722 
1723 			error = rfs3call(VTOMI(unldvp), NFSPROC3_REMOVE,
1724 			    xdr_diropargs3, (caddr_t)&args,
1725 			    xdr_REMOVE3res, (caddr_t)&res, unlcred,
1726 			    &douprintf, &res.status, 0, NULL);
1727 
1728 			if (error) {
1729 				PURGE_ATTRCACHE(unldvp);
1730 			} else {
1731 				error = geterrno3(res.status);
1732 				if (!error) {
1733 					nfs3_cache_wcc_data(unldvp,
1734 					    &res.resok.dir_wcc, t, cr);
1735 					if (HAVE_RDDIR_CACHE(VTOR(unldvp)))
1736 						nfs_purge_rddir_cache(unldvp);
1737 				} else {
1738 					nfs3_cache_wcc_data(unldvp,
1739 					    &res.resfail.dir_wcc, t, cr);
1740 					PURGE_STALE_FH(error, unldvp, cr);
1741 				}
1742 			}
1743 
1744 			/*
1745 			 * Release stuff held for the remove
1746 			 */
1747 			VN_RELE(unldvp);
1748 			kmem_free(unlname, MAXNAMELEN);
1749 			crfree(unlcred);
1750 			goto redo;
1751 		}
1752 		mutex_exit(&rp->r_statelock);
1753 	}
1754 
1755 	rp_addfree(rp, cr);
1756 }
1757 
1758 /*
1759  * Remote file system operations having to do with directory manipulation.
1760  */
1761 
1762 static int
1763 nfs3_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1764 	int flags, vnode_t *rdir, cred_t *cr)
1765 {
1766 	int error;
1767 	vnode_t *vp;
1768 	vnode_t *avp = NULL;
1769 	rnode_t *drp;
1770 
1771 	if (curproc->p_zone != VTOMI(dvp)->mi_zone)
1772 		return (EPERM);
1773 
1774 	drp = VTOR(dvp);
1775 
1776 	/*
1777 	 * Are we looking up extended attributes?  If so, "dvp" is
1778 	 * the file or directory for which we want attributes, and
1779 	 * we need a lookup of the hidden attribute directory
1780 	 * before we lookup the rest of the path.
1781 	 */
1782 	if (flags & LOOKUP_XATTR) {
1783 		bool_t cflag = ((flags & CREATE_XATTR_DIR) != 0);
1784 		mntinfo_t *mi;
1785 
1786 		mi = VTOMI(dvp);
1787 		if (!(mi->mi_flags & MI_EXTATTR))
1788 			return (EINVAL);
1789 
1790 		if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp)))
1791 			return (EINTR);
1792 
1793 		(void) nfs3lookup_dnlc(dvp, XATTR_DIR_NAME, &avp, cr);
1794 		if (avp == NULL)
1795 			error = acl_getxattrdir3(dvp, &avp, cflag, cr, 0);
1796 		else
1797 			error = 0;
1798 
1799 		nfs_rw_exit(&drp->r_rwlock);
1800 
1801 		if (error) {
1802 			if (mi->mi_flags & MI_EXTATTR)
1803 				return (error);
1804 			return (EINVAL);
1805 		}
1806 		dvp = avp;
1807 		drp = VTOR(dvp);
1808 	}
1809 
1810 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp))) {
1811 		error = EINTR;
1812 		goto out;
1813 	}
1814 
1815 	error = nfs3lookup(dvp, nm, vpp, pnp, flags, rdir, cr, 0);
1816 
1817 	nfs_rw_exit(&drp->r_rwlock);
1818 
1819 	/*
1820 	 * If vnode is a device, create special vnode.
1821 	 */
1822 	if (!error && IS_DEVVP(*vpp)) {
1823 		vp = *vpp;
1824 		*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
1825 		VN_RELE(vp);
1826 	}
1827 
1828 out:
1829 	if (avp != NULL)
1830 		VN_RELE(avp);
1831 
1832 	return (error);
1833 }
1834 
1835 static int nfs3_lookup_neg_cache = 1;
1836 
1837 #ifdef DEBUG
1838 static int nfs3_lookup_dnlc_hits = 0;
1839 static int nfs3_lookup_dnlc_misses = 0;
1840 static int nfs3_lookup_dnlc_neg_hits = 0;
1841 static int nfs3_lookup_dnlc_disappears = 0;
1842 static int nfs3_lookup_dnlc_lookups = 0;
1843 #endif
1844 
1845 /* ARGSUSED */
1846 int
1847 nfs3lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1848 	int flags, vnode_t *rdir, cred_t *cr, int rfscall_flags)
1849 {
1850 	int error;
1851 	rnode_t *drp;
1852 
1853 	ASSERT(curproc->p_zone == VTOMI(dvp)->mi_zone);
1854 	/*
1855 	 * If lookup is for "", just return dvp.  Don't need
1856 	 * to send it over the wire, look it up in the dnlc,
1857 	 * or perform any access checks.
1858 	 */
1859 	if (*nm == '\0') {
1860 		VN_HOLD(dvp);
1861 		*vpp = dvp;
1862 		return (0);
1863 	}
1864 
1865 	/*
1866 	 * Can't do lookups in non-directories.
1867 	 */
1868 	if (dvp->v_type != VDIR)
1869 		return (ENOTDIR);
1870 
1871 	/*
1872 	 * If we're called with RFSCALL_SOFT, it's important that
1873 	 * the only rfscall is one we make directly; if we permit
1874 	 * an access call because we're looking up "." or validating
1875 	 * a dnlc hit, we'll deadlock because that rfscall will not
1876 	 * have the RFSCALL_SOFT set.
1877 	 */
1878 	if (rfscall_flags & RFSCALL_SOFT)
1879 		goto callit;
1880 
1881 	/*
1882 	 * If lookup is for ".", just return dvp.  Don't need
1883 	 * to send it over the wire or look it up in the dnlc,
1884 	 * just need to check access.
1885 	 */
1886 	if (strcmp(nm, ".") == 0) {
1887 		error = nfs3_access(dvp, VEXEC, 0, cr);
1888 		if (error)
1889 			return (error);
1890 		VN_HOLD(dvp);
1891 		*vpp = dvp;
1892 		return (0);
1893 	}
1894 
1895 	drp = VTOR(dvp);
1896 	if (!(drp->r_flags & RLOOKUP)) {
1897 		mutex_enter(&drp->r_statelock);
1898 		drp->r_flags |= RLOOKUP;
1899 		mutex_exit(&drp->r_statelock);
1900 	}
1901 
1902 	/*
1903 	 * Lookup this name in the DNLC.  If there was a valid entry,
1904 	 * then return the results of the lookup.
1905 	 */
1906 	error = nfs3lookup_dnlc(dvp, nm, vpp, cr);
1907 	if (error || *vpp != NULL)
1908 		return (error);
1909 
1910 callit:
1911 	error = nfs3lookup_otw(dvp, nm, vpp, cr, rfscall_flags);
1912 
1913 	return (error);
1914 }
1915 
1916 static int
1917 nfs3lookup_dnlc(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr)
1918 {
1919 	int error;
1920 	vnode_t *vp;
1921 
1922 	ASSERT(*nm != '\0');
1923 	ASSERT(curproc->p_zone == VTOMI(dvp)->mi_zone);
1924 	/*
1925 	 * Lookup this name in the DNLC.  If successful, then validate
1926 	 * the caches and then recheck the DNLC.  The DNLC is rechecked
1927 	 * just in case this entry got invalidated during the call
1928 	 * to nfs3_validate_caches.
1929 	 *
1930 	 * An assumption is being made that it is safe to say that a
1931 	 * file exists which may not on the server.  Any operations to
1932 	 * the server will fail with ESTALE.
1933 	 */
1934 #ifdef DEBUG
1935 	nfs3_lookup_dnlc_lookups++;
1936 #endif
1937 	vp = dnlc_lookup(dvp, nm);
1938 	if (vp != NULL) {
1939 		VN_RELE(vp);
1940 		if (vp == DNLC_NO_VNODE && !vn_is_readonly(dvp)) {
1941 			PURGE_ATTRCACHE(dvp);
1942 		}
1943 		error = nfs3_validate_caches(dvp, cr);
1944 		if (error)
1945 			return (error);
1946 		vp = dnlc_lookup(dvp, nm);
1947 		if (vp != NULL) {
1948 			error = nfs3_access(dvp, VEXEC, 0, cr);
1949 			if (error) {
1950 				VN_RELE(vp);
1951 				return (error);
1952 			}
1953 			if (vp == DNLC_NO_VNODE) {
1954 				VN_RELE(vp);
1955 #ifdef DEBUG
1956 				nfs3_lookup_dnlc_neg_hits++;
1957 #endif
1958 				return (ENOENT);
1959 			}
1960 			*vpp = vp;
1961 #ifdef DEBUG
1962 			nfs3_lookup_dnlc_hits++;
1963 #endif
1964 			return (0);
1965 		}
1966 #ifdef DEBUG
1967 		nfs3_lookup_dnlc_disappears++;
1968 #endif
1969 	}
1970 #ifdef DEBUG
1971 	else
1972 		nfs3_lookup_dnlc_misses++;
1973 #endif
1974 
1975 	*vpp = NULL;
1976 
1977 	return (0);
1978 }
1979 
1980 static int
1981 nfs3lookup_otw(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr,
1982 	int rfscall_flags)
1983 {
1984 	int error;
1985 	LOOKUP3args args;
1986 	LOOKUP3vres res;
1987 	int douprintf;
1988 	struct vattr vattr;
1989 	struct vattr dvattr;
1990 	vnode_t *vp;
1991 	failinfo_t fi;
1992 	hrtime_t t;
1993 
1994 	ASSERT(*nm != '\0');
1995 	ASSERT(dvp->v_type == VDIR);
1996 	ASSERT(curproc->p_zone == VTOMI(dvp)->mi_zone);
1997 
1998 	setdiropargs3(&args.what, nm, dvp);
1999 
2000 	fi.vp = dvp;
2001 	fi.fhp = (caddr_t)&args.what.dir;
2002 	fi.copyproc = nfs3copyfh;
2003 	fi.lookupproc = nfs3lookup;
2004 	fi.xattrdirproc = acl_getxattrdir3;
2005 	res.obj_attributes.fres.vp = dvp;
2006 	res.obj_attributes.fres.vap = &vattr;
2007 	res.dir_attributes.fres.vp = dvp;
2008 	res.dir_attributes.fres.vap = &dvattr;
2009 
2010 	douprintf = 1;
2011 
2012 	t = gethrtime();
2013 
2014 	error = rfs3call(VTOMI(dvp), NFSPROC3_LOOKUP,
2015 	    xdr_diropargs3, (caddr_t)&args,
2016 	    xdr_LOOKUP3vres, (caddr_t)&res, cr,
2017 	    &douprintf, &res.status, rfscall_flags, &fi);
2018 
2019 	if (error)
2020 		return (error);
2021 
2022 	nfs3_cache_post_op_vattr(dvp, &res.dir_attributes, t, cr);
2023 
2024 	error = geterrno3(res.status);
2025 	if (error) {
2026 		PURGE_STALE_FH(error, dvp, cr);
2027 		if (error == ENOENT && nfs3_lookup_neg_cache)
2028 			dnlc_enter(dvp, nm, DNLC_NO_VNODE);
2029 		return (error);
2030 	}
2031 
2032 	if (res.obj_attributes.attributes) {
2033 		vp = makenfs3node_va(&res.object, res.obj_attributes.fres.vap,
2034 				dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2035 	} else {
2036 		vp = makenfs3node_va(&res.object, NULL,
2037 				dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2038 		if (vp->v_type == VNON) {
2039 			vattr.va_mask = AT_TYPE;
2040 			error = nfs3getattr(vp, &vattr, cr);
2041 			if (error) {
2042 				VN_RELE(vp);
2043 				return (error);
2044 			}
2045 			vp->v_type = vattr.va_type;
2046 		}
2047 	}
2048 
2049 	if (!(rfscall_flags & RFSCALL_SOFT))
2050 		dnlc_update(dvp, nm, vp);
2051 
2052 	*vpp = vp;
2053 
2054 	return (error);
2055 }
2056 
2057 #ifdef DEBUG
2058 static int nfs3_create_misses = 0;
2059 #endif
2060 
2061 /* ARGSUSED */
2062 static int
2063 nfs3_create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2064 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2065 {
2066 	int error;
2067 	vnode_t *vp;
2068 	rnode_t *rp;
2069 	struct vattr vattr;
2070 	rnode_t *drp;
2071 	vnode_t *tempvp;
2072 
2073 	drp = VTOR(dvp);
2074 	if (curproc->p_zone != VTOMI(dvp)->mi_zone)
2075 		return (EPERM);
2076 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2077 		return (EINTR);
2078 
2079 top:
2080 	/*
2081 	 * We make a copy of the attributes because the caller does not
2082 	 * expect us to change what va points to.
2083 	 */
2084 	vattr = *va;
2085 
2086 	/*
2087 	 * If the pathname is "", just use dvp.  Don't need
2088 	 * to send it over the wire, look it up in the dnlc,
2089 	 * or perform any access checks.
2090 	 */
2091 	if (*nm == '\0') {
2092 		error = 0;
2093 		VN_HOLD(dvp);
2094 		vp = dvp;
2095 	/*
2096 	 * If the pathname is ".", just use dvp.  Don't need
2097 	 * to send it over the wire or look it up in the dnlc,
2098 	 * just need to check access.
2099 	 */
2100 	} else if (strcmp(nm, ".") == 0) {
2101 		error = nfs3_access(dvp, VEXEC, 0, cr);
2102 		if (error) {
2103 			nfs_rw_exit(&drp->r_rwlock);
2104 			return (error);
2105 		}
2106 		VN_HOLD(dvp);
2107 		vp = dvp;
2108 	/*
2109 	 * We need to go over the wire, just to be sure whether the
2110 	 * file exists or not.  Using the DNLC can be dangerous in
2111 	 * this case when making a decision regarding existence.
2112 	 */
2113 	} else {
2114 		error = nfs3lookup_otw(dvp, nm, &vp, cr, 0);
2115 	}
2116 	if (!error) {
2117 		if (exclusive == EXCL)
2118 			error = EEXIST;
2119 		else if (vp->v_type == VDIR && (mode & VWRITE))
2120 			error = EISDIR;
2121 		else {
2122 			/*
2123 			 * If vnode is a device, create special vnode.
2124 			 */
2125 			if (IS_DEVVP(vp)) {
2126 				tempvp = vp;
2127 				vp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2128 				VN_RELE(tempvp);
2129 			}
2130 			if (!(error = VOP_ACCESS(vp, mode, 0, cr))) {
2131 				if ((vattr.va_mask & AT_SIZE) &&
2132 				    vp->v_type == VREG) {
2133 					rp = VTOR(vp);
2134 					/*
2135 					 * Check here for large file handled
2136 					 * by LF-unaware process (as
2137 					 * ufs_create() does)
2138 					 */
2139 					if (!(lfaware & FOFFMAX)) {
2140 						mutex_enter(&rp->r_statelock);
2141 						if (rp->r_size > MAXOFF32_T)
2142 							error = EOVERFLOW;
2143 						mutex_exit(&rp->r_statelock);
2144 					}
2145 					if (!error) {
2146 						vattr.va_mask = AT_SIZE;
2147 						error = nfs3setattr(vp,
2148 						    &vattr, 0, cr);
2149 					}
2150 				}
2151 			}
2152 		}
2153 		nfs_rw_exit(&drp->r_rwlock);
2154 		if (error) {
2155 			VN_RELE(vp);
2156 		} else
2157 			*vpp = vp;
2158 		return (error);
2159 	}
2160 
2161 	dnlc_remove(dvp, nm);
2162 
2163 	/*
2164 	 * Decide what the group-id of the created file should be.
2165 	 * Set it in attribute list as advisory...
2166 	 */
2167 	error = setdirgid(dvp, &vattr.va_gid, cr);
2168 	if (error) {
2169 		nfs_rw_exit(&drp->r_rwlock);
2170 		return (error);
2171 	}
2172 	vattr.va_mask |= AT_GID;
2173 
2174 	ASSERT(vattr.va_mask & AT_TYPE);
2175 	if (vattr.va_type == VREG) {
2176 		ASSERT(vattr.va_mask & AT_MODE);
2177 		if (MANDMODE(vattr.va_mode)) {
2178 			nfs_rw_exit(&drp->r_rwlock);
2179 			return (EACCES);
2180 		}
2181 		error = nfs3create(dvp, nm, &vattr, exclusive, mode, vpp, cr,
2182 		    lfaware);
2183 		/*
2184 		 * If this is not an exclusive create, then the CREATE
2185 		 * request will be made with the GUARDED mode set.  This
2186 		 * means that the server will return EEXIST if the file
2187 		 * exists.  The file could exist because of a retransmitted
2188 		 * request.  In this case, we recover by starting over and
2189 		 * checking to see whether the file exists.  This second
2190 		 * time through it should and a CREATE request will not be
2191 		 * sent.
2192 		 *
2193 		 * This handles the problem of a dangling CREATE request
2194 		 * which contains attributes which indicate that the file
2195 		 * should be truncated.  This retransmitted request could
2196 		 * possibly truncate valid data in the file if not caught
2197 		 * by the duplicate request mechanism on the server or if
2198 		 * not caught by other means.  The scenario is:
2199 		 *
2200 		 * Client transmits CREATE request with size = 0
2201 		 * Client times out, retransmits request.
2202 		 * Response to the first request arrives from the server
2203 		 *  and the client proceeds on.
2204 		 * Client writes data to the file.
2205 		 * The server now processes retransmitted CREATE request
2206 		 *  and truncates file.
2207 		 *
2208 		 * The use of the GUARDED CREATE request prevents this from
2209 		 * happening because the retransmitted CREATE would fail
2210 		 * with EEXIST and would not truncate the file.
2211 		 */
2212 		if (error == EEXIST && exclusive == NONEXCL) {
2213 #ifdef DEBUG
2214 			nfs3_create_misses++;
2215 #endif
2216 			goto top;
2217 		}
2218 		nfs_rw_exit(&drp->r_rwlock);
2219 		return (error);
2220 	}
2221 	error = nfs3mknod(dvp, nm, &vattr, exclusive, mode, vpp, cr);
2222 	nfs_rw_exit(&drp->r_rwlock);
2223 	return (error);
2224 }
2225 
2226 /* ARGSUSED */
2227 static int
2228 nfs3create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2229 	int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2230 {
2231 	int error;
2232 	CREATE3args args;
2233 	CREATE3res res;
2234 	int douprintf;
2235 	vnode_t *vp;
2236 	struct vattr vattr;
2237 	nfstime3 *verfp;
2238 	rnode_t *rp;
2239 	timestruc_t now;
2240 	hrtime_t t;
2241 
2242 	ASSERT(curproc->p_zone == VTOMI(dvp)->mi_zone);
2243 	setdiropargs3(&args.where, nm, dvp);
2244 	if (exclusive == EXCL) {
2245 		args.how.mode = EXCLUSIVE;
2246 		/*
2247 		 * Construct the create verifier.  This verifier needs
2248 		 * to be unique between different clients.  It also needs
2249 		 * to vary for each exclusive create request generated
2250 		 * from the client to the server.
2251 		 *
2252 		 * The first attempt is made to use the hostid and a
2253 		 * unique number on the client.  If the hostid has not
2254 		 * been set, the high resolution time that the exclusive
2255 		 * create request is being made is used.  This will work
2256 		 * unless two different clients, both with the hostid
2257 		 * not set, attempt an exclusive create request on the
2258 		 * same file, at exactly the same clock time.  The
2259 		 * chances of this happening seem small enough to be
2260 		 * reasonable.
2261 		 */
2262 		verfp = (nfstime3 *)&args.how.createhow3_u.verf;
2263 		verfp->seconds = nfs_atoi(hw_serial);
2264 		if (verfp->seconds != 0)
2265 			verfp->nseconds = newnum();
2266 		else {
2267 			gethrestime(&now);
2268 			verfp->seconds = now.tv_sec;
2269 			verfp->nseconds = now.tv_nsec;
2270 		}
2271 		/*
2272 		 * Since the server will use this value for the mtime,
2273 		 * make sure that it can't overflow. Zero out the MSB.
2274 		 * The actual value does not matter here, only its uniqeness.
2275 		 */
2276 		verfp->seconds %= INT32_MAX;
2277 	} else {
2278 		/*
2279 		 * Issue the non-exclusive create in guarded mode.  This
2280 		 * may result in some false EEXIST responses for
2281 		 * retransmitted requests, but these will be handled at
2282 		 * a higher level.  By using GUARDED, duplicate requests
2283 		 * to do file truncation and possible access problems
2284 		 * can be avoided.
2285 		 */
2286 		args.how.mode = GUARDED;
2287 		error = vattr_to_sattr3(va,
2288 				&args.how.createhow3_u.obj_attributes);
2289 		if (error) {
2290 			/* req time field(s) overflow - return immediately */
2291 			return (error);
2292 		}
2293 	}
2294 
2295 	douprintf = 1;
2296 
2297 	t = gethrtime();
2298 
2299 	error = rfs3call(VTOMI(dvp), NFSPROC3_CREATE,
2300 	    xdr_CREATE3args, (caddr_t)&args,
2301 	    xdr_CREATE3res, (caddr_t)&res, cr,
2302 	    &douprintf, &res.status, 0, NULL);
2303 
2304 	if (error) {
2305 		PURGE_ATTRCACHE(dvp);
2306 		return (error);
2307 	}
2308 
2309 	error = geterrno3(res.status);
2310 	if (!error) {
2311 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2312 		if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2313 			nfs_purge_rddir_cache(dvp);
2314 
2315 		/*
2316 		 * On exclusive create the times need to be explicitly
2317 		 * set to clear any potential verifier that may be stored
2318 		 * in one of these fields (see comment below).  This
2319 		 * is done here to cover the case where no post op attrs
2320 		 * were returned or a 'invalid' time was returned in
2321 		 * the attributes.
2322 		 */
2323 		if (exclusive == EXCL)
2324 			va->va_mask |= (AT_MTIME | AT_ATIME);
2325 
2326 		if (!res.resok.obj.handle_follows) {
2327 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2328 			if (error)
2329 				return (error);
2330 		} else {
2331 			if (res.resok.obj_attributes.attributes) {
2332 				vp = makenfs3node(&res.resok.obj.handle,
2333 				    &res.resok.obj_attributes.attr,
2334 				    dvp->v_vfsp, t, cr, NULL, NULL);
2335 			} else {
2336 				vp = makenfs3node(&res.resok.obj.handle, NULL,
2337 				    dvp->v_vfsp, t, cr, NULL, NULL);
2338 
2339 				/*
2340 				 * On an exclusive create, it is possible
2341 				 * that attributes were returned but those
2342 				 * postop attributes failed to decode
2343 				 * properly.  If this is the case,
2344 				 * then most likely the atime or mtime
2345 				 * were invalid for our client; this
2346 				 * is caused by the server storing the
2347 				 * create verifier in one of the time
2348 				 * fields(most likely mtime).
2349 				 * So... we are going to setattr just the
2350 				 * atime/mtime to clear things up.
2351 				 */
2352 				if (exclusive == EXCL) {
2353 					if (error =
2354 						nfs3excl_create_settimes(vp,
2355 							va, cr)) {
2356 						/*
2357 						 * Setting the times failed.
2358 						 * Remove the file and return
2359 						 * the error.
2360 						 */
2361 						VN_RELE(vp);
2362 						(void) nfs3_remove(dvp,
2363 							nm, cr);
2364 						return (error);
2365 					}
2366 				}
2367 
2368 				/*
2369 				 * This handles the non-exclusive case
2370 				 * and the exclusive case where no post op
2371 				 * attrs were returned.
2372 				 */
2373 				if (vp->v_type == VNON) {
2374 					vattr.va_mask = AT_TYPE;
2375 					error = nfs3getattr(vp, &vattr, cr);
2376 					if (error) {
2377 						VN_RELE(vp);
2378 						return (error);
2379 					}
2380 					vp->v_type = vattr.va_type;
2381 				}
2382 			}
2383 			dnlc_update(dvp, nm, vp);
2384 		}
2385 
2386 		rp = VTOR(vp);
2387 
2388 		/*
2389 		 * Check here for large file handled by
2390 		 * LF-unaware process (as ufs_create() does)
2391 		 */
2392 		if ((va->va_mask & AT_SIZE) && vp->v_type == VREG &&
2393 		    !(lfaware & FOFFMAX)) {
2394 			mutex_enter(&rp->r_statelock);
2395 			if (rp->r_size > MAXOFF32_T) {
2396 				mutex_exit(&rp->r_statelock);
2397 				VN_RELE(vp);
2398 				return (EOVERFLOW);
2399 			}
2400 			mutex_exit(&rp->r_statelock);
2401 		}
2402 
2403 		if (exclusive == EXCL &&
2404 			(va->va_mask & ~(AT_GID | AT_SIZE))) {
2405 			/*
2406 			 * If doing an exclusive create, then generate
2407 			 * a SETATTR to set the initial attributes.
2408 			 * Try to set the mtime and the atime to the
2409 			 * server's current time.  It is somewhat
2410 			 * expected that these fields will be used to
2411 			 * store the exclusive create cookie.  If not,
2412 			 * server implementors will need to know that
2413 			 * a SETATTR will follow an exclusive create
2414 			 * and the cookie should be destroyed if
2415 			 * appropriate. This work may have been done
2416 			 * earlier in this function if post op attrs
2417 			 * were not available.
2418 			 *
2419 			 * The AT_GID and AT_SIZE bits are turned off
2420 			 * so that the SETATTR request will not attempt
2421 			 * to process these.  The gid will be set
2422 			 * separately if appropriate.  The size is turned
2423 			 * off because it is assumed that a new file will
2424 			 * be created empty and if the file wasn't empty,
2425 			 * then the exclusive create will have failed
2426 			 * because the file must have existed already.
2427 			 * Therefore, no truncate operation is needed.
2428 			 */
2429 			va->va_mask &= ~(AT_GID | AT_SIZE);
2430 			error = nfs3setattr(vp, va, 0, cr);
2431 			if (error) {
2432 				/*
2433 				 * Couldn't correct the attributes of
2434 				 * the newly created file and the
2435 				 * attributes are wrong.  Remove the
2436 				 * file and return an error to the
2437 				 * application.
2438 				 */
2439 				VN_RELE(vp);
2440 				(void) nfs3_remove(dvp, nm, cr);
2441 				return (error);
2442 			}
2443 		}
2444 
2445 		if (va->va_gid != rp->r_attr.va_gid) {
2446 			/*
2447 			 * If the gid on the file isn't right, then
2448 			 * generate a SETATTR to attempt to change
2449 			 * it.  This may or may not work, depending
2450 			 * upon the server's semantics for allowing
2451 			 * file ownership changes.
2452 			 */
2453 			va->va_mask = AT_GID;
2454 			(void) nfs3setattr(vp, va, 0, cr);
2455 		}
2456 
2457 		/*
2458 		 * If vnode is a device create special vnode
2459 		 */
2460 		if (IS_DEVVP(vp)) {
2461 			*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2462 			VN_RELE(vp);
2463 		} else
2464 			*vpp = vp;
2465 	} else {
2466 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2467 		PURGE_STALE_FH(error, dvp, cr);
2468 	}
2469 
2470 	return (error);
2471 }
2472 
2473 /*
2474  * Special setattr function to take care of rest of atime/mtime
2475  * after successful exclusive create.  This function exists to avoid
2476  * handling attributes from the server; exclusive the atime/mtime fields
2477  * may be 'invalid' in client's view and therefore can not be trusted.
2478  */
2479 static int
2480 nfs3excl_create_settimes(vnode_t *vp, struct vattr *vap, cred_t *cr)
2481 {
2482 	int error;
2483 	uint_t mask;
2484 	SETATTR3args args;
2485 	SETATTR3res res;
2486 	int douprintf;
2487 	rnode_t *rp;
2488 	hrtime_t t;
2489 
2490 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
2491 	/* save the caller's mask so that it can be reset later */
2492 	mask = vap->va_mask;
2493 
2494 	rp = VTOR(vp);
2495 
2496 	args.object = *RTOFH3(rp);
2497 	args.guard.check = FALSE;
2498 
2499 	/* Use the mask to initialize the arguments */
2500 	vap->va_mask = 0;
2501 	error = vattr_to_sattr3(vap, &args.new_attributes);
2502 
2503 	/* We want to set just atime/mtime on this request */
2504 	args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
2505 	args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
2506 
2507 	douprintf = 1;
2508 
2509 	t = gethrtime();
2510 
2511 	error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
2512 	    xdr_SETATTR3args, (caddr_t)&args,
2513 	    xdr_SETATTR3res, (caddr_t)&res, cr,
2514 	    &douprintf, &res.status, 0, NULL);
2515 
2516 	if (error) {
2517 		vap->va_mask = mask;
2518 		return (error);
2519 	}
2520 
2521 	error = geterrno3(res.status);
2522 	if (!error) {
2523 		/*
2524 		 * It is important to pick up the attributes.
2525 		 * Since this is the exclusive create path, the
2526 		 * attributes on the initial create were ignored
2527 		 * and we need these to have the correct info.
2528 		 */
2529 		nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
2530 		/*
2531 		 * No need to do the atime/mtime work again so clear
2532 		 * the bits.
2533 		 */
2534 		mask &= ~(AT_ATIME | AT_MTIME);
2535 	} else {
2536 		nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
2537 	}
2538 
2539 	vap->va_mask = mask;
2540 
2541 	return (error);
2542 }
2543 
2544 /* ARGSUSED */
2545 static int
2546 nfs3mknod(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2547 	int mode, vnode_t **vpp, cred_t *cr)
2548 {
2549 	int error;
2550 	MKNOD3args args;
2551 	MKNOD3res res;
2552 	int douprintf;
2553 	vnode_t *vp;
2554 	struct vattr vattr;
2555 	hrtime_t t;
2556 
2557 	ASSERT(curproc->p_zone == VTOMI(dvp)->mi_zone);
2558 	switch (va->va_type) {
2559 	case VCHR:
2560 	case VBLK:
2561 		setdiropargs3(&args.where, nm, dvp);
2562 		args.what.type = (va->va_type == VCHR) ? NF3CHR : NF3BLK;
2563 		error = vattr_to_sattr3(va,
2564 		    &args.what.mknoddata3_u.device.dev_attributes);
2565 		if (error) {
2566 			/* req time field(s) overflow - return immediately */
2567 			return (error);
2568 		}
2569 		args.what.mknoddata3_u.device.spec.specdata1 =
2570 		    getmajor(va->va_rdev);
2571 		args.what.mknoddata3_u.device.spec.specdata2 =
2572 		    getminor(va->va_rdev);
2573 		break;
2574 
2575 	case VFIFO:
2576 	case VSOCK:
2577 		setdiropargs3(&args.where, nm, dvp);
2578 		args.what.type = (va->va_type == VFIFO) ? NF3FIFO : NF3SOCK;
2579 		error = vattr_to_sattr3(va,
2580 				&args.what.mknoddata3_u.pipe_attributes);
2581 		if (error) {
2582 			/* req time field(s) overflow - return immediately */
2583 			return (error);
2584 		}
2585 		break;
2586 
2587 	default:
2588 		return (EINVAL);
2589 	}
2590 
2591 	douprintf = 1;
2592 
2593 	t = gethrtime();
2594 
2595 	error = rfs3call(VTOMI(dvp), NFSPROC3_MKNOD,
2596 	    xdr_MKNOD3args, (caddr_t)&args,
2597 	    xdr_MKNOD3res, (caddr_t)&res, cr,
2598 	    &douprintf, &res.status, 0, NULL);
2599 
2600 	if (error) {
2601 		PURGE_ATTRCACHE(dvp);
2602 		return (error);
2603 	}
2604 
2605 	error = geterrno3(res.status);
2606 	if (!error) {
2607 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2608 		if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2609 			nfs_purge_rddir_cache(dvp);
2610 
2611 		if (!res.resok.obj.handle_follows) {
2612 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2613 			if (error)
2614 				return (error);
2615 		} else {
2616 			if (res.resok.obj_attributes.attributes) {
2617 				vp = makenfs3node(&res.resok.obj.handle,
2618 				    &res.resok.obj_attributes.attr,
2619 				    dvp->v_vfsp, t, cr, NULL, NULL);
2620 			} else {
2621 				vp = makenfs3node(&res.resok.obj.handle, NULL,
2622 				    dvp->v_vfsp, t, cr, NULL, NULL);
2623 				if (vp->v_type == VNON) {
2624 					vattr.va_mask = AT_TYPE;
2625 					error = nfs3getattr(vp, &vattr, cr);
2626 					if (error) {
2627 						VN_RELE(vp);
2628 						return (error);
2629 					}
2630 					vp->v_type = vattr.va_type;
2631 				}
2632 
2633 			}
2634 			dnlc_update(dvp, nm, vp);
2635 		}
2636 
2637 		if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
2638 			va->va_mask = AT_GID;
2639 			(void) nfs3setattr(vp, va, 0, cr);
2640 		}
2641 
2642 		/*
2643 		 * If vnode is a device create special vnode
2644 		 */
2645 		if (IS_DEVVP(vp)) {
2646 			*vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2647 			VN_RELE(vp);
2648 		} else
2649 			*vpp = vp;
2650 	} else {
2651 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2652 		PURGE_STALE_FH(error, dvp, cr);
2653 	}
2654 	return (error);
2655 }
2656 
2657 /*
2658  * Weirdness: if the vnode to be removed is open
2659  * we rename it instead of removing it and nfs_inactive
2660  * will remove the new name.
2661  */
2662 static int
2663 nfs3_remove(vnode_t *dvp, char *nm, cred_t *cr)
2664 {
2665 	int error;
2666 	REMOVE3args args;
2667 	REMOVE3res res;
2668 	vnode_t *vp;
2669 	char *tmpname;
2670 	int douprintf;
2671 	rnode_t *rp;
2672 	rnode_t *drp;
2673 	hrtime_t t;
2674 
2675 	if (curproc->p_zone != VTOMI(dvp)->mi_zone)
2676 		return (EPERM);
2677 	drp = VTOR(dvp);
2678 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2679 		return (EINTR);
2680 
2681 	error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2682 	if (error) {
2683 		nfs_rw_exit(&drp->r_rwlock);
2684 		return (error);
2685 	}
2686 
2687 	if (vp->v_type == VDIR && secpolicy_fs_linkdir(cr, dvp->v_vfsp)) {
2688 		VN_RELE(vp);
2689 		nfs_rw_exit(&drp->r_rwlock);
2690 		return (EPERM);
2691 	}
2692 
2693 	/*
2694 	 * First just remove the entry from the name cache, as it
2695 	 * is most likely the only entry for this vp.
2696 	 */
2697 	dnlc_remove(dvp, nm);
2698 
2699 	/*
2700 	 * If the file has a v_count > 1 then there may be more than one
2701 	 * entry in the name cache due multiple links or an open file,
2702 	 * but we don't have the real reference count so flush all
2703 	 * possible entries.
2704 	 */
2705 	if (vp->v_count > 1)
2706 		dnlc_purge_vp(vp);
2707 
2708 	/*
2709 	 * Now we have the real reference count on the vnode
2710 	 */
2711 	rp = VTOR(vp);
2712 	mutex_enter(&rp->r_statelock);
2713 	if (vp->v_count > 1 &&
2714 	    (rp->r_unldvp == NULL || strcmp(nm, rp->r_unlname) == 0)) {
2715 		mutex_exit(&rp->r_statelock);
2716 		tmpname = newname();
2717 		error = nfs3rename(dvp, nm, dvp, tmpname, cr);
2718 		if (error)
2719 			kmem_free(tmpname, MAXNAMELEN);
2720 		else {
2721 			mutex_enter(&rp->r_statelock);
2722 			if (rp->r_unldvp == NULL) {
2723 				VN_HOLD(dvp);
2724 				rp->r_unldvp = dvp;
2725 				if (rp->r_unlcred != NULL)
2726 					crfree(rp->r_unlcred);
2727 				crhold(cr);
2728 				rp->r_unlcred = cr;
2729 				rp->r_unlname = tmpname;
2730 			} else {
2731 				kmem_free(rp->r_unlname, MAXNAMELEN);
2732 				rp->r_unlname = tmpname;
2733 			}
2734 			mutex_exit(&rp->r_statelock);
2735 		}
2736 	} else {
2737 		mutex_exit(&rp->r_statelock);
2738 		/*
2739 		 * We need to flush any dirty pages which happen to
2740 		 * be hanging around before removing the file.  This
2741 		 * shouldn't happen very often and mostly on file
2742 		 * systems mounted "nocto".
2743 		 */
2744 		if (vn_has_cached_data(vp) &&
2745 		    ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
2746 			error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr);
2747 			if (error && (error == ENOSPC || error == EDQUOT)) {
2748 				mutex_enter(&rp->r_statelock);
2749 				if (!rp->r_error)
2750 					rp->r_error = error;
2751 				mutex_exit(&rp->r_statelock);
2752 			}
2753 		}
2754 
2755 		setdiropargs3(&args.object, nm, dvp);
2756 
2757 		douprintf = 1;
2758 
2759 		t = gethrtime();
2760 
2761 		error = rfs3call(VTOMI(dvp), NFSPROC3_REMOVE,
2762 		    xdr_diropargs3, (caddr_t)&args,
2763 		    xdr_REMOVE3res, (caddr_t)&res, cr,
2764 		    &douprintf, &res.status, 0, NULL);
2765 
2766 		/*
2767 		 * The xattr dir may be gone after last attr is removed,
2768 		 * so flush it from dnlc.
2769 		 */
2770 		if (dvp->v_flag & V_XATTRDIR)
2771 			dnlc_purge_vp(dvp);
2772 
2773 		PURGE_ATTRCACHE(vp);
2774 
2775 		if (error) {
2776 			PURGE_ATTRCACHE(dvp);
2777 		} else {
2778 			error = geterrno3(res.status);
2779 			if (!error) {
2780 				nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t,
2781 				    cr);
2782 				if (HAVE_RDDIR_CACHE(drp))
2783 					nfs_purge_rddir_cache(dvp);
2784 			} else {
2785 				nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc,
2786 				    t, cr);
2787 				PURGE_STALE_FH(error, dvp, cr);
2788 			}
2789 		}
2790 	}
2791 
2792 	VN_RELE(vp);
2793 
2794 	nfs_rw_exit(&drp->r_rwlock);
2795 
2796 	return (error);
2797 }
2798 
2799 static int
2800 nfs3_link(vnode_t *tdvp, vnode_t *svp, char *tnm, cred_t *cr)
2801 {
2802 	int error;
2803 	LINK3args args;
2804 	LINK3res res;
2805 	vnode_t *realvp;
2806 	int douprintf;
2807 	mntinfo_t *mi;
2808 	rnode_t *tdrp;
2809 	hrtime_t t;
2810 
2811 	if (curproc->p_zone != VTOMI(tdvp)->mi_zone)
2812 		return (EPERM);
2813 	if (VOP_REALVP(svp, &realvp) == 0)
2814 		svp = realvp;
2815 
2816 	mi = VTOMI(svp);
2817 
2818 	if (!(mi->mi_flags & MI_LINK))
2819 		return (EOPNOTSUPP);
2820 
2821 	args.file = *VTOFH3(svp);
2822 	setdiropargs3(&args.link, tnm, tdvp);
2823 
2824 	tdrp = VTOR(tdvp);
2825 	if (nfs_rw_enter_sig(&tdrp->r_rwlock, RW_WRITER, INTR(tdvp)))
2826 		return (EINTR);
2827 
2828 	dnlc_remove(tdvp, tnm);
2829 
2830 	douprintf = 1;
2831 
2832 	t = gethrtime();
2833 
2834 	error = rfs3call(mi, NFSPROC3_LINK,
2835 	    xdr_LINK3args, (caddr_t)&args,
2836 	    xdr_LINK3res, (caddr_t)&res, cr,
2837 	    &douprintf, &res.status, 0, NULL);
2838 
2839 	if (error) {
2840 		PURGE_ATTRCACHE(tdvp);
2841 		PURGE_ATTRCACHE(svp);
2842 		nfs_rw_exit(&tdrp->r_rwlock);
2843 		return (error);
2844 	}
2845 
2846 	error = geterrno3(res.status);
2847 
2848 	if (!error) {
2849 		nfs3_cache_post_op_attr(svp, &res.resok.file_attributes, t, cr);
2850 		nfs3_cache_wcc_data(tdvp, &res.resok.linkdir_wcc, t, cr);
2851 		if (HAVE_RDDIR_CACHE(tdrp))
2852 			nfs_purge_rddir_cache(tdvp);
2853 		dnlc_update(tdvp, tnm, svp);
2854 	} else {
2855 		nfs3_cache_post_op_attr(svp, &res.resfail.file_attributes, t,
2856 		    cr);
2857 		nfs3_cache_wcc_data(tdvp, &res.resfail.linkdir_wcc, t, cr);
2858 		if (error == EOPNOTSUPP) {
2859 			mutex_enter(&mi->mi_lock);
2860 			mi->mi_flags &= ~MI_LINK;
2861 			mutex_exit(&mi->mi_lock);
2862 		}
2863 	}
2864 
2865 	nfs_rw_exit(&tdrp->r_rwlock);
2866 
2867 	return (error);
2868 }
2869 
2870 static int
2871 nfs3_rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2872 {
2873 	vnode_t *realvp;
2874 
2875 	if (curproc->p_zone != VTOMI(odvp)->mi_zone)
2876 		return (EPERM);
2877 	if (VOP_REALVP(ndvp, &realvp) == 0)
2878 		ndvp = realvp;
2879 
2880 	return (nfs3rename(odvp, onm, ndvp, nnm, cr));
2881 }
2882 
2883 /*
2884  * nfs3rename does the real work of renaming in NFS Version 3.
2885  */
2886 static int
2887 nfs3rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr)
2888 {
2889 	int error;
2890 	RENAME3args args;
2891 	RENAME3res res;
2892 	int douprintf;
2893 	vnode_t *nvp;
2894 	vnode_t *ovp = NULL;
2895 	char *tmpname;
2896 	rnode_t *rp;
2897 	rnode_t *odrp;
2898 	rnode_t *ndrp;
2899 	hrtime_t t;
2900 
2901 	ASSERT(curproc->p_zone == VTOMI(odvp)->mi_zone);
2902 
2903 	if (strcmp(onm, ".") == 0 || strcmp(onm, "..") == 0 ||
2904 	    strcmp(nnm, ".") == 0 || strcmp(nnm, "..") == 0)
2905 		return (EINVAL);
2906 
2907 	odrp = VTOR(odvp);
2908 	ndrp = VTOR(ndvp);
2909 	if ((intptr_t)odrp < (intptr_t)ndrp) {
2910 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp)))
2911 			return (EINTR);
2912 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp))) {
2913 			nfs_rw_exit(&odrp->r_rwlock);
2914 			return (EINTR);
2915 		}
2916 	} else {
2917 		if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp)))
2918 			return (EINTR);
2919 		if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp))) {
2920 			nfs_rw_exit(&ndrp->r_rwlock);
2921 			return (EINTR);
2922 		}
2923 	}
2924 
2925 	/*
2926 	 * Lookup the target file.  If it exists, it needs to be
2927 	 * checked to see whether it is a mount point and whether
2928 	 * it is active (open).
2929 	 */
2930 	error = nfs3lookup(ndvp, nnm, &nvp, NULL, 0, NULL, cr, 0);
2931 	if (!error) {
2932 		/*
2933 		 * If this file has been mounted on, then just
2934 		 * return busy because renaming to it would remove
2935 		 * the mounted file system from the name space.
2936 		 */
2937 		if (vn_mountedvfs(nvp) != NULL) {
2938 			VN_RELE(nvp);
2939 			nfs_rw_exit(&odrp->r_rwlock);
2940 			nfs_rw_exit(&ndrp->r_rwlock);
2941 			return (EBUSY);
2942 		}
2943 
2944 		/*
2945 		 * Purge the name cache of all references to this vnode
2946 		 * so that we can check the reference count to infer
2947 		 * whether it is active or not.
2948 		 */
2949 		/*
2950 		 * First just remove the entry from the name cache, as it
2951 		 * is most likely the only entry for this vp.
2952 		 */
2953 		dnlc_remove(ndvp, nnm);
2954 		/*
2955 		 * If the file has a v_count > 1 then there may be more
2956 		 * than one entry in the name cache due multiple links
2957 		 * or an open file, but we don't have the real reference
2958 		 * count so flush all possible entries.
2959 		 */
2960 		if (nvp->v_count > 1)
2961 			dnlc_purge_vp(nvp);
2962 
2963 		/*
2964 		 * If the vnode is active and is not a directory,
2965 		 * arrange to rename it to a
2966 		 * temporary file so that it will continue to be
2967 		 * accessible.  This implements the "unlink-open-file"
2968 		 * semantics for the target of a rename operation.
2969 		 * Before doing this though, make sure that the
2970 		 * source and target files are not already the same.
2971 		 */
2972 		if (nvp->v_count > 1 && nvp->v_type != VDIR) {
2973 			/*
2974 			 * Lookup the source name.
2975 			 */
2976 			error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL,
2977 			    cr, 0);
2978 
2979 			/*
2980 			 * The source name *should* already exist.
2981 			 */
2982 			if (error) {
2983 				VN_RELE(nvp);
2984 				nfs_rw_exit(&odrp->r_rwlock);
2985 				nfs_rw_exit(&ndrp->r_rwlock);
2986 				return (error);
2987 			}
2988 
2989 			/*
2990 			 * Compare the two vnodes.  If they are the same,
2991 			 * just release all held vnodes and return success.
2992 			 */
2993 			if (ovp == nvp) {
2994 				VN_RELE(ovp);
2995 				VN_RELE(nvp);
2996 				nfs_rw_exit(&odrp->r_rwlock);
2997 				nfs_rw_exit(&ndrp->r_rwlock);
2998 				return (0);
2999 			}
3000 
3001 			/*
3002 			 * Can't mix and match directories and non-
3003 			 * directories in rename operations.  We already
3004 			 * know that the target is not a directory.  If
3005 			 * the source is a directory, return an error.
3006 			 */
3007 			if (ovp->v_type == VDIR) {
3008 				VN_RELE(ovp);
3009 				VN_RELE(nvp);
3010 				nfs_rw_exit(&odrp->r_rwlock);
3011 				nfs_rw_exit(&ndrp->r_rwlock);
3012 				return (ENOTDIR);
3013 			}
3014 
3015 			/*
3016 			 * The target file exists, is not the same as
3017 			 * the source file, and is active.  Link it
3018 			 * to a temporary filename to avoid having
3019 			 * the server removing the file completely.
3020 			 */
3021 			tmpname = newname();
3022 			error = nfs3_link(ndvp, nvp, tmpname, cr);
3023 			if (error == EOPNOTSUPP) {
3024 				error = nfs3_rename(ndvp, nnm, ndvp, tmpname,
3025 				    cr);
3026 			}
3027 			if (error) {
3028 				kmem_free(tmpname, MAXNAMELEN);
3029 				VN_RELE(ovp);
3030 				VN_RELE(nvp);
3031 				nfs_rw_exit(&odrp->r_rwlock);
3032 				nfs_rw_exit(&ndrp->r_rwlock);
3033 				return (error);
3034 			}
3035 			rp = VTOR(nvp);
3036 			mutex_enter(&rp->r_statelock);
3037 			if (rp->r_unldvp == NULL) {
3038 				VN_HOLD(ndvp);
3039 				rp->r_unldvp = ndvp;
3040 				if (rp->r_unlcred != NULL)
3041 					crfree(rp->r_unlcred);
3042 				crhold(cr);
3043 				rp->r_unlcred = cr;
3044 				rp->r_unlname = tmpname;
3045 			} else {
3046 				kmem_free(rp->r_unlname, MAXNAMELEN);
3047 				rp->r_unlname = tmpname;
3048 			}
3049 			mutex_exit(&rp->r_statelock);
3050 		}
3051 
3052 		VN_RELE(nvp);
3053 	}
3054 
3055 	if (ovp == NULL) {
3056 		/*
3057 		 * When renaming directories to be a subdirectory of a
3058 		 * different parent, the dnlc entry for ".." will no
3059 		 * longer be valid, so it must be removed.
3060 		 *
3061 		 * We do a lookup here to determine whether we are renaming
3062 		 * a directory and we need to check if we are renaming
3063 		 * an unlinked file.  This might have already been done
3064 		 * in previous code, so we check ovp == NULL to avoid
3065 		 * doing it twice.
3066 		 */
3067 
3068 		error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL, cr, 0);
3069 		/*
3070 		 * The source name *should* already exist.
3071 		 */
3072 		if (error) {
3073 			nfs_rw_exit(&odrp->r_rwlock);
3074 			nfs_rw_exit(&ndrp->r_rwlock);
3075 			return (error);
3076 		}
3077 		ASSERT(ovp != NULL);
3078 	}
3079 
3080 	dnlc_remove(odvp, onm);
3081 	dnlc_remove(ndvp, nnm);
3082 
3083 	setdiropargs3(&args.from, onm, odvp);
3084 	setdiropargs3(&args.to, nnm, ndvp);
3085 
3086 	douprintf = 1;
3087 
3088 	t = gethrtime();
3089 
3090 	error = rfs3call(VTOMI(odvp), NFSPROC3_RENAME,
3091 	    xdr_RENAME3args, (caddr_t)&args,
3092 	    xdr_RENAME3res, (caddr_t)&res, cr,
3093 	    &douprintf, &res.status, 0, NULL);
3094 
3095 	if (error) {
3096 		PURGE_ATTRCACHE(odvp);
3097 		PURGE_ATTRCACHE(ndvp);
3098 		VN_RELE(ovp);
3099 		nfs_rw_exit(&odrp->r_rwlock);
3100 		nfs_rw_exit(&ndrp->r_rwlock);
3101 		return (error);
3102 	}
3103 
3104 	error = geterrno3(res.status);
3105 
3106 	if (!error) {
3107 		nfs3_cache_wcc_data(odvp, &res.resok.fromdir_wcc, t, cr);
3108 		if (HAVE_RDDIR_CACHE(odrp))
3109 			nfs_purge_rddir_cache(odvp);
3110 		if (ndvp != odvp) {
3111 			nfs3_cache_wcc_data(ndvp, &res.resok.todir_wcc, t, cr);
3112 			if (HAVE_RDDIR_CACHE(ndrp))
3113 				nfs_purge_rddir_cache(ndvp);
3114 		}
3115 		/*
3116 		 * when renaming directories to be a subdirectory of a
3117 		 * different parent, the dnlc entry for ".." will no
3118 		 * longer be valid, so it must be removed
3119 		 */
3120 		rp = VTOR(ovp);
3121 		if (ndvp != odvp) {
3122 			if (ovp->v_type == VDIR) {
3123 				dnlc_remove(ovp, "..");
3124 				if (HAVE_RDDIR_CACHE(rp))
3125 					nfs_purge_rddir_cache(ovp);
3126 			}
3127 		}
3128 
3129 		/*
3130 		 * If we are renaming the unlinked file, update the
3131 		 * r_unldvp and r_unlname as needed.
3132 		 */
3133 		mutex_enter(&rp->r_statelock);
3134 		if (rp->r_unldvp != NULL) {
3135 			if (strcmp(rp->r_unlname, onm) == 0) {
3136 				(void) strncpy(rp->r_unlname, nnm, MAXNAMELEN);
3137 				rp->r_unlname[MAXNAMELEN - 1] = '\0';
3138 
3139 				if (ndvp != rp->r_unldvp) {
3140 					VN_RELE(rp->r_unldvp);
3141 					rp->r_unldvp = ndvp;
3142 					VN_HOLD(ndvp);
3143 				}
3144 			}
3145 		}
3146 		mutex_exit(&rp->r_statelock);
3147 	} else {
3148 		nfs3_cache_wcc_data(odvp, &res.resfail.fromdir_wcc, t, cr);
3149 		if (ndvp != odvp) {
3150 			nfs3_cache_wcc_data(ndvp, &res.resfail.todir_wcc, t,
3151 			    cr);
3152 		}
3153 		/*
3154 		 * System V defines rename to return EEXIST, not
3155 		 * ENOTEMPTY if the target directory is not empty.
3156 		 * Over the wire, the error is NFSERR_ENOTEMPTY
3157 		 * which geterrno maps to ENOTEMPTY.
3158 		 */
3159 		if (error == ENOTEMPTY)
3160 			error = EEXIST;
3161 	}
3162 
3163 	VN_RELE(ovp);
3164 
3165 	nfs_rw_exit(&odrp->r_rwlock);
3166 	nfs_rw_exit(&ndrp->r_rwlock);
3167 
3168 	return (error);
3169 }
3170 
3171 static int
3172 nfs3_mkdir(vnode_t *dvp, char *nm, struct vattr *va, vnode_t **vpp, cred_t *cr)
3173 {
3174 	int error;
3175 	MKDIR3args args;
3176 	MKDIR3res res;
3177 	int douprintf;
3178 	struct vattr vattr;
3179 	vnode_t *vp;
3180 	rnode_t *drp;
3181 	hrtime_t t;
3182 
3183 	if (curproc->p_zone != VTOMI(dvp)->mi_zone)
3184 		return (EPERM);
3185 	setdiropargs3(&args.where, nm, dvp);
3186 
3187 	/*
3188 	 * Decide what the group-id and set-gid bit of the created directory
3189 	 * should be.  May have to do a setattr to get the gid right.
3190 	 */
3191 	error = setdirgid(dvp, &va->va_gid, cr);
3192 	if (error)
3193 		return (error);
3194 	error = setdirmode(dvp, &va->va_mode, cr);
3195 	if (error)
3196 		return (error);
3197 	va->va_mask |= AT_MODE|AT_GID;
3198 
3199 	error = vattr_to_sattr3(va, &args.attributes);
3200 	if (error) {
3201 		/* req time field(s) overflow - return immediately */
3202 		return (error);
3203 	}
3204 
3205 	drp = VTOR(dvp);
3206 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3207 		return (EINTR);
3208 
3209 	dnlc_remove(dvp, nm);
3210 
3211 	douprintf = 1;
3212 
3213 	t = gethrtime();
3214 
3215 	error = rfs3call(VTOMI(dvp), NFSPROC3_MKDIR,
3216 	    xdr_MKDIR3args, (caddr_t)&args,
3217 	    xdr_MKDIR3res, (caddr_t)&res, cr,
3218 	    &douprintf, &res.status, 0, NULL);
3219 
3220 	if (error) {
3221 		PURGE_ATTRCACHE(dvp);
3222 		nfs_rw_exit(&drp->r_rwlock);
3223 		return (error);
3224 	}
3225 
3226 	error = geterrno3(res.status);
3227 	if (!error) {
3228 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3229 		if (HAVE_RDDIR_CACHE(drp))
3230 			nfs_purge_rddir_cache(dvp);
3231 
3232 		if (!res.resok.obj.handle_follows) {
3233 			error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3234 			if (error) {
3235 				nfs_rw_exit(&drp->r_rwlock);
3236 				return (error);
3237 			}
3238 		} else {
3239 			if (res.resok.obj_attributes.attributes) {
3240 				vp = makenfs3node(&res.resok.obj.handle,
3241 				    &res.resok.obj_attributes.attr,
3242 				    dvp->v_vfsp, t, cr, NULL, NULL);
3243 			} else {
3244 				vp = makenfs3node(&res.resok.obj.handle, NULL,
3245 				    dvp->v_vfsp, t, cr, NULL, NULL);
3246 				if (vp->v_type == VNON) {
3247 					vattr.va_mask = AT_TYPE;
3248 					error = nfs3getattr(vp, &vattr, cr);
3249 					if (error) {
3250 						VN_RELE(vp);
3251 						nfs_rw_exit(&drp->r_rwlock);
3252 						return (error);
3253 					}
3254 					vp->v_type = vattr.va_type;
3255 				}
3256 			}
3257 			dnlc_update(dvp, nm, vp);
3258 		}
3259 		if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
3260 			va->va_mask = AT_GID;
3261 			(void) nfs3setattr(vp, va, 0, cr);
3262 		}
3263 		*vpp = vp;
3264 	} else {
3265 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3266 		PURGE_STALE_FH(error, dvp, cr);
3267 	}
3268 
3269 	nfs_rw_exit(&drp->r_rwlock);
3270 
3271 	return (error);
3272 }
3273 
3274 static int
3275 nfs3_rmdir(vnode_t *dvp, char *nm, vnode_t *cdir, cred_t *cr)
3276 {
3277 	int error;
3278 	RMDIR3args args;
3279 	RMDIR3res res;
3280 	vnode_t *vp;
3281 	int douprintf;
3282 	rnode_t *drp;
3283 	hrtime_t t;
3284 
3285 	if (curproc->p_zone != VTOMI(dvp)->mi_zone)
3286 		return (EPERM);
3287 	drp = VTOR(dvp);
3288 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3289 		return (EINTR);
3290 
3291 	/*
3292 	 * Attempt to prevent a rmdir(".") from succeeding.
3293 	 */
3294 	error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3295 	if (error) {
3296 		nfs_rw_exit(&drp->r_rwlock);
3297 		return (error);
3298 	}
3299 
3300 	if (vp == cdir) {
3301 		VN_RELE(vp);
3302 		nfs_rw_exit(&drp->r_rwlock);
3303 		return (EINVAL);
3304 	}
3305 
3306 	setdiropargs3(&args.object, nm, dvp);
3307 
3308 	/*
3309 	 * First just remove the entry from the name cache, as it
3310 	 * is most likely an entry for this vp.
3311 	 */
3312 	dnlc_remove(dvp, nm);
3313 
3314 	/*
3315 	 * If there vnode reference count is greater than one, then
3316 	 * there may be additional references in the DNLC which will
3317 	 * need to be purged.  First, trying removing the entry for
3318 	 * the parent directory and see if that removes the additional
3319 	 * reference(s).  If that doesn't do it, then use dnlc_purge_vp
3320 	 * to completely remove any references to the directory which
3321 	 * might still exist in the DNLC.
3322 	 */
3323 	if (vp->v_count > 1) {
3324 		dnlc_remove(vp, "..");
3325 		if (vp->v_count > 1)
3326 			dnlc_purge_vp(vp);
3327 	}
3328 
3329 	douprintf = 1;
3330 
3331 	t = gethrtime();
3332 
3333 	error = rfs3call(VTOMI(dvp), NFSPROC3_RMDIR,
3334 	    xdr_diropargs3, (caddr_t)&args,
3335 	    xdr_RMDIR3res, (caddr_t)&res, cr,
3336 	    &douprintf, &res.status, 0, NULL);
3337 
3338 	PURGE_ATTRCACHE(vp);
3339 
3340 	if (error) {
3341 		PURGE_ATTRCACHE(dvp);
3342 		VN_RELE(vp);
3343 		nfs_rw_exit(&drp->r_rwlock);
3344 		return (error);
3345 	}
3346 
3347 	error = geterrno3(res.status);
3348 	if (!error) {
3349 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3350 		if (HAVE_RDDIR_CACHE(drp))
3351 			nfs_purge_rddir_cache(dvp);
3352 		if (HAVE_RDDIR_CACHE(VTOR(vp)))
3353 			nfs_purge_rddir_cache(vp);
3354 	} else {
3355 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3356 		PURGE_STALE_FH(error, dvp, cr);
3357 		/*
3358 		 * System V defines rmdir to return EEXIST, not
3359 		 * ENOTEMPTY if the directory is not empty.  Over
3360 		 * the wire, the error is NFSERR_ENOTEMPTY which
3361 		 * geterrno maps to ENOTEMPTY.
3362 		 */
3363 		if (error == ENOTEMPTY)
3364 			error = EEXIST;
3365 	}
3366 
3367 	VN_RELE(vp);
3368 
3369 	nfs_rw_exit(&drp->r_rwlock);
3370 
3371 	return (error);
3372 }
3373 
3374 static int
3375 nfs3_symlink(vnode_t *dvp, char *lnm, struct vattr *tva, char *tnm, cred_t *cr)
3376 {
3377 	int error;
3378 	SYMLINK3args args;
3379 	SYMLINK3res res;
3380 	int douprintf;
3381 	mntinfo_t *mi;
3382 	vnode_t *vp;
3383 	rnode_t *rp;
3384 	char *contents;
3385 	rnode_t *drp;
3386 	hrtime_t t;
3387 
3388 	mi = VTOMI(dvp);
3389 
3390 	if (curproc->p_zone != mi->mi_zone)
3391 		return (EPERM);
3392 	if (!(mi->mi_flags & MI_SYMLINK))
3393 		return (EOPNOTSUPP);
3394 
3395 	setdiropargs3(&args.where, lnm, dvp);
3396 	error = vattr_to_sattr3(tva, &args.symlink.symlink_attributes);
3397 	if (error) {
3398 		/* req time field(s) overflow - return immediately */
3399 		return (error);
3400 	}
3401 	args.symlink.symlink_data = tnm;
3402 
3403 	drp = VTOR(dvp);
3404 	if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3405 		return (EINTR);
3406 
3407 	dnlc_remove(dvp, lnm);
3408 
3409 	douprintf = 1;
3410 
3411 	t = gethrtime();
3412 
3413 	error = rfs3call(mi, NFSPROC3_SYMLINK,
3414 	    xdr_SYMLINK3args, (caddr_t)&args,
3415 	    xdr_SYMLINK3res, (caddr_t)&res, cr,
3416 	    &douprintf, &res.status, 0, NULL);
3417 
3418 	if (error) {
3419 		PURGE_ATTRCACHE(dvp);
3420 		nfs_rw_exit(&drp->r_rwlock);
3421 		return (error);
3422 	}
3423 
3424 	error = geterrno3(res.status);
3425 	if (!error) {
3426 		nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3427 		if (HAVE_RDDIR_CACHE(drp))
3428 			nfs_purge_rddir_cache(dvp);
3429 
3430 		if (res.resok.obj.handle_follows) {
3431 			if (res.resok.obj_attributes.attributes) {
3432 				vp = makenfs3node(&res.resok.obj.handle,
3433 				    &res.resok.obj_attributes.attr,
3434 				    dvp->v_vfsp, t, cr, NULL, NULL);
3435 			} else {
3436 				vp = makenfs3node(&res.resok.obj.handle, NULL,
3437 				    dvp->v_vfsp, t, cr, NULL, NULL);
3438 				vp->v_type = VLNK;
3439 				vp->v_rdev = 0;
3440 			}
3441 			dnlc_update(dvp, lnm, vp);
3442 			rp = VTOR(vp);
3443 			if (nfs3_do_symlink_cache &&
3444 			    rp->r_symlink.contents == NULL) {
3445 
3446 				contents = kmem_alloc(MAXPATHLEN,
3447 				    KM_NOSLEEP);
3448 
3449 				if (contents != NULL) {
3450 					mutex_enter(&rp->r_statelock);
3451 					if (rp->r_symlink.contents == NULL) {
3452 						rp->r_symlink.len = strlen(tnm);
3453 						bcopy(tnm, contents,
3454 						    rp->r_symlink.len);
3455 						rp->r_symlink.contents =
3456 						    contents;
3457 						rp->r_symlink.size = MAXPATHLEN;
3458 						mutex_exit(&rp->r_statelock);
3459 					} else {
3460 						mutex_exit(&rp->r_statelock);
3461 						kmem_free((void *)contents,
3462 							    MAXPATHLEN);
3463 					}
3464 				}
3465 			}
3466 			VN_RELE(vp);
3467 		}
3468 	} else {
3469 		nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3470 		PURGE_STALE_FH(error, dvp, cr);
3471 		if (error == EOPNOTSUPP) {
3472 			mutex_enter(&mi->mi_lock);
3473 			mi->mi_flags &= ~MI_SYMLINK;
3474 			mutex_exit(&mi->mi_lock);
3475 		}
3476 	}
3477 
3478 	nfs_rw_exit(&drp->r_rwlock);
3479 
3480 	return (error);
3481 }
3482 
3483 #ifdef DEBUG
3484 static int nfs3_readdir_cache_hits = 0;
3485 static int nfs3_readdir_cache_shorts = 0;
3486 static int nfs3_readdir_cache_waits = 0;
3487 static int nfs3_readdir_cache_misses = 0;
3488 static int nfs3_readdir_readahead = 0;
3489 #endif
3490 
3491 static int nfs3_shrinkreaddir = 0;
3492 
3493 /*
3494  * Read directory entries.
3495  * There are some weird things to look out for here.  The uio_loffset
3496  * field is either 0 or it is the offset returned from a previous
3497  * readdir.  It is an opaque value used by the server to find the
3498  * correct directory block to read. The count field is the number
3499  * of blocks to read on the server.  This is advisory only, the server
3500  * may return only one block's worth of entries.  Entries may be compressed
3501  * on the server.
3502  */
3503 static int
3504 nfs3_readdir(vnode_t *vp, struct uio *uiop, cred_t *cr, int *eofp)
3505 {
3506 	int error;
3507 	size_t count;
3508 	rnode_t *rp;
3509 	rddir_cache *rdc;
3510 	rddir_cache *nrdc;
3511 	rddir_cache *rrdc;
3512 #ifdef DEBUG
3513 	int missed;
3514 #endif
3515 	int doreadahead;
3516 	rddir_cache srdc;
3517 	avl_index_t where;
3518 
3519 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
3520 		return (EIO);
3521 	rp = VTOR(vp);
3522 
3523 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
3524 
3525 	/*
3526 	 * Make sure that the directory cache is valid.
3527 	 */
3528 	if (HAVE_RDDIR_CACHE(rp)) {
3529 		if (nfs_disable_rddir_cache) {
3530 			/*
3531 			 * Setting nfs_disable_rddir_cache in /etc/system
3532 			 * allows interoperability with servers that do not
3533 			 * properly update the attributes of directories.
3534 			 * Any cached information gets purged before an
3535 			 * access is made to it.
3536 			 */
3537 			nfs_purge_rddir_cache(vp);
3538 		} else {
3539 			error = nfs3_validate_caches(vp, cr);
3540 			if (error)
3541 				return (error);
3542 		}
3543 	}
3544 
3545 	/*
3546 	 * It is possible that some servers may not be able to correctly
3547 	 * handle a large READDIR or READDIRPLUS request due to bugs in
3548 	 * their implementation.  In order to continue to interoperate
3549 	 * with them, this workaround is provided to limit the maximum
3550 	 * size of a READDIRPLUS request to 1024.  In any case, the request
3551 	 * size is limited to MAXBSIZE.
3552 	 */
3553 	count = MIN(uiop->uio_iov->iov_len,
3554 	    nfs3_shrinkreaddir ? 1024 : MAXBSIZE);
3555 
3556 	nrdc = NULL;
3557 #ifdef DEBUG
3558 	missed = 0;
3559 #endif
3560 top:
3561 	/*
3562 	 * Short circuit last readdir which always returns 0 bytes.
3563 	 * This can be done after the directory has been read through
3564 	 * completely at least once.  This will set r_direof which
3565 	 * can be used to find the value of the last cookie.
3566 	 */
3567 	mutex_enter(&rp->r_statelock);
3568 	if (rp->r_direof != NULL &&
3569 	    uiop->uio_loffset == rp->r_direof->nfs3_ncookie) {
3570 		mutex_exit(&rp->r_statelock);
3571 #ifdef DEBUG
3572 		nfs3_readdir_cache_shorts++;
3573 #endif
3574 		if (eofp)
3575 			*eofp = 1;
3576 		if (nrdc != NULL)
3577 			rddir_cache_rele(nrdc);
3578 		return (0);
3579 	}
3580 	/*
3581 	 * Look for a cache entry.  Cache entries are identified
3582 	 * by the NFS cookie value and the byte count requested.
3583 	 */
3584 	srdc.nfs3_cookie = uiop->uio_loffset;
3585 	srdc.buflen = count;
3586 	rdc = avl_find(&rp->r_dir, &srdc, &where);
3587 	if (rdc != NULL) {
3588 		rddir_cache_hold(rdc);
3589 		/*
3590 		 * If the cache entry is in the process of being
3591 		 * filled in, wait until this completes.  The
3592 		 * RDDIRWAIT bit is set to indicate that someone
3593 		 * is waiting and then the thread currently
3594 		 * filling the entry is done, it should do a
3595 		 * cv_broadcast to wakeup all of the threads
3596 		 * waiting for it to finish.
3597 		 */
3598 		if (rdc->flags & RDDIR) {
3599 			nfs_rw_exit(&rp->r_rwlock);
3600 			rdc->flags |= RDDIRWAIT;
3601 #ifdef DEBUG
3602 			nfs3_readdir_cache_waits++;
3603 #endif
3604 			if (!cv_wait_sig(&rdc->cv, &rp->r_statelock)) {
3605 				/*
3606 				 * We got interrupted, probably
3607 				 * the user typed ^C or an alarm
3608 				 * fired.  We free the new entry
3609 				 * if we allocated one.
3610 				 */
3611 				mutex_exit(&rp->r_statelock);
3612 				(void) nfs_rw_enter_sig(&rp->r_rwlock,
3613 					RW_READER, FALSE);
3614 				rddir_cache_rele(rdc);
3615 				if (nrdc != NULL)
3616 					rddir_cache_rele(nrdc);
3617 				return (EINTR);
3618 			}
3619 			mutex_exit(&rp->r_statelock);
3620 			(void) nfs_rw_enter_sig(&rp->r_rwlock,
3621 				RW_READER, FALSE);
3622 			rddir_cache_rele(rdc);
3623 			goto top;
3624 		}
3625 		/*
3626 		 * Check to see if a readdir is required to
3627 		 * fill the entry.  If so, mark this entry
3628 		 * as being filled, remove our reference,
3629 		 * and branch to the code to fill the entry.
3630 		 */
3631 		if (rdc->flags & RDDIRREQ) {
3632 			rdc->flags &= ~RDDIRREQ;
3633 			rdc->flags |= RDDIR;
3634 			if (nrdc != NULL)
3635 				rddir_cache_rele(nrdc);
3636 			nrdc = rdc;
3637 			mutex_exit(&rp->r_statelock);
3638 			goto bottom;
3639 		}
3640 #ifdef DEBUG
3641 		if (!missed)
3642 			nfs3_readdir_cache_hits++;
3643 #endif
3644 		/*
3645 		 * If an error occurred while attempting
3646 		 * to fill the cache entry, just return it.
3647 		 */
3648 		if (rdc->error) {
3649 			error = rdc->error;
3650 			mutex_exit(&rp->r_statelock);
3651 			rddir_cache_rele(rdc);
3652 			if (nrdc != NULL)
3653 				rddir_cache_rele(nrdc);
3654 			return (error);
3655 		}
3656 
3657 		/*
3658 		 * The cache entry is complete and good,
3659 		 * copyout the dirent structs to the calling
3660 		 * thread.
3661 		 */
3662 		error = uiomove(rdc->entries, rdc->entlen, UIO_READ, uiop);
3663 
3664 		/*
3665 		 * If no error occurred during the copyout,
3666 		 * update the offset in the uio struct to
3667 		 * contain the value of the next cookie
3668 		 * and set the eof value appropriately.
3669 		 */
3670 		if (!error) {
3671 			uiop->uio_loffset = rdc->nfs3_ncookie;
3672 			if (eofp)
3673 				*eofp = rdc->eof;
3674 		}
3675 
3676 		/*
3677 		 * Decide whether to do readahead.
3678 		 *
3679 		 * Don't if have already read to the end of
3680 		 * directory.  There is nothing more to read.
3681 		 *
3682 		 * Don't if the application is not doing
3683 		 * lookups in the directory.  The readahead
3684 		 * is only effective if the application can
3685 		 * be doing work while an async thread is
3686 		 * handling the over the wire request.
3687 		 */
3688 		if (rdc->eof) {
3689 			rp->r_direof = rdc;
3690 			doreadahead = FALSE;
3691 		} else if (!(rp->r_flags & RLOOKUP))
3692 			doreadahead = FALSE;
3693 		else
3694 			doreadahead = TRUE;
3695 
3696 		if (!doreadahead) {
3697 			mutex_exit(&rp->r_statelock);
3698 			rddir_cache_rele(rdc);
3699 			if (nrdc != NULL)
3700 				rddir_cache_rele(nrdc);
3701 			return (error);
3702 		}
3703 
3704 		/*
3705 		 * Check to see whether we found an entry
3706 		 * for the readahead.  If so, we don't need
3707 		 * to do anything further, so free the new
3708 		 * entry if one was allocated.  Otherwise,
3709 		 * allocate a new entry, add it to the cache,
3710 		 * and then initiate an asynchronous readdir
3711 		 * operation to fill it.
3712 		 */
3713 		srdc.nfs3_cookie = rdc->nfs3_ncookie;
3714 		srdc.buflen = count;
3715 		rrdc = avl_find(&rp->r_dir, &srdc, &where);
3716 		if (rrdc != NULL) {
3717 			if (nrdc != NULL)
3718 				rddir_cache_rele(nrdc);
3719 		} else {
3720 			if (nrdc != NULL)
3721 				rrdc = nrdc;
3722 			else {
3723 				rrdc = rddir_cache_alloc(KM_NOSLEEP);
3724 			}
3725 			if (rrdc != NULL) {
3726 				rrdc->nfs3_cookie = rdc->nfs3_ncookie;
3727 				rrdc->buflen = count;
3728 				avl_insert(&rp->r_dir, rrdc, where);
3729 				rddir_cache_hold(rrdc);
3730 				mutex_exit(&rp->r_statelock);
3731 				rddir_cache_rele(rdc);
3732 #ifdef DEBUG
3733 				nfs3_readdir_readahead++;
3734 #endif
3735 				nfs_async_readdir(vp, rrdc, cr, do_nfs3readdir);
3736 				return (error);
3737 			}
3738 		}
3739 
3740 		mutex_exit(&rp->r_statelock);
3741 		rddir_cache_rele(rdc);
3742 		return (error);
3743 	}
3744 
3745 	/*
3746 	 * Didn't find an entry in the cache.  Construct a new empty
3747 	 * entry and link it into the cache.  Other processes attempting
3748 	 * to access this entry will need to wait until it is filled in.
3749 	 *
3750 	 * Since kmem_alloc may block, another pass through the cache
3751 	 * will need to be taken to make sure that another process
3752 	 * hasn't already added an entry to the cache for this request.
3753 	 */
3754 	if (nrdc == NULL) {
3755 		mutex_exit(&rp->r_statelock);
3756 		nrdc = rddir_cache_alloc(KM_SLEEP);
3757 		nrdc->nfs3_cookie = uiop->uio_loffset;
3758 		nrdc->buflen = count;
3759 		goto top;
3760 	}
3761 
3762 	/*
3763 	 * Add this entry to the cache.
3764 	 */
3765 	avl_insert(&rp->r_dir, nrdc, where);
3766 	rddir_cache_hold(nrdc);
3767 	mutex_exit(&rp->r_statelock);
3768 
3769 bottom:
3770 #ifdef DEBUG
3771 	missed = 1;
3772 	nfs3_readdir_cache_misses++;
3773 #endif
3774 	/*
3775 	 * Do the readdir.  This routine decides whether to use
3776 	 * READDIR or READDIRPLUS.
3777 	 */
3778 	error = do_nfs3readdir(vp, nrdc, cr);
3779 
3780 	/*
3781 	 * If this operation failed, just return the error which occurred.
3782 	 */
3783 	if (error != 0)
3784 		return (error);
3785 
3786 	/*
3787 	 * Since the RPC operation will have taken sometime and blocked
3788 	 * this process, another pass through the cache will need to be
3789 	 * taken to find the correct cache entry.  It is possible that
3790 	 * the correct cache entry will not be there (although one was
3791 	 * added) because the directory changed during the RPC operation
3792 	 * and the readdir cache was flushed.  In this case, just start
3793 	 * over.  It is hoped that this will not happen too often... :-)
3794 	 */
3795 	nrdc = NULL;
3796 	goto top;
3797 	/* NOTREACHED */
3798 }
3799 
3800 static int
3801 do_nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3802 {
3803 	int error;
3804 	rnode_t *rp;
3805 	mntinfo_t *mi;
3806 
3807 	rp = VTOR(vp);
3808 	mi = VTOMI(vp);
3809 	ASSERT(curproc->p_zone == mi->mi_zone);
3810 	/*
3811 	 * Issue the proper request.
3812 	 *
3813 	 * If the server does not support READDIRPLUS, then use READDIR.
3814 	 *
3815 	 * Otherwise --
3816 	 * Issue a READDIRPLUS if reading to fill an empty cache or if
3817 	 * an application has performed a lookup in the directory which
3818 	 * required an over the wire lookup.  The use of READDIRPLUS
3819 	 * will help to (re)populate the DNLC.
3820 	 */
3821 	if (!(mi->mi_flags & MI_READDIRONLY) &&
3822 	    (rp->r_flags & (RLOOKUP | RREADDIRPLUS))) {
3823 		if (rp->r_flags & RREADDIRPLUS) {
3824 			mutex_enter(&rp->r_statelock);
3825 			rp->r_flags &= ~RREADDIRPLUS;
3826 			mutex_exit(&rp->r_statelock);
3827 		}
3828 		nfs3readdirplus(vp, rdc, cr);
3829 		if (rdc->error == EOPNOTSUPP)
3830 			nfs3readdir(vp, rdc, cr);
3831 	} else
3832 		nfs3readdir(vp, rdc, cr);
3833 
3834 	mutex_enter(&rp->r_statelock);
3835 	rdc->flags &= ~RDDIR;
3836 	if (rdc->flags & RDDIRWAIT) {
3837 		rdc->flags &= ~RDDIRWAIT;
3838 		cv_broadcast(&rdc->cv);
3839 	}
3840 	error = rdc->error;
3841 	if (error)
3842 		rdc->flags |= RDDIRREQ;
3843 	mutex_exit(&rp->r_statelock);
3844 
3845 	rddir_cache_rele(rdc);
3846 
3847 	return (error);
3848 }
3849 
3850 static void
3851 nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3852 {
3853 	int error;
3854 	READDIR3args args;
3855 	READDIR3vres res;
3856 	vattr_t dva;
3857 	rnode_t *rp;
3858 	int douprintf;
3859 	failinfo_t fi, *fip = NULL;
3860 	mntinfo_t *mi;
3861 	hrtime_t t;
3862 
3863 	rp = VTOR(vp);
3864 	mi = VTOMI(vp);
3865 	ASSERT(curproc->p_zone == mi->mi_zone);
3866 
3867 	args.dir = *RTOFH3(rp);
3868 	args.cookie = (cookie3)rdc->nfs3_cookie;
3869 	args.cookieverf = rp->r_cookieverf;
3870 	args.count = rdc->buflen;
3871 
3872 	/*
3873 	 * NFS client failover support
3874 	 * suppress failover unless we have a zero cookie
3875 	 */
3876 	if (args.cookie == (cookie3) 0) {
3877 		fi.vp = vp;
3878 		fi.fhp = (caddr_t)&args.dir;
3879 		fi.copyproc = nfs3copyfh;
3880 		fi.lookupproc = nfs3lookup;
3881 		fi.xattrdirproc = acl_getxattrdir3;
3882 		fip = &fi;
3883 	}
3884 
3885 #ifdef DEBUG
3886 	rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
3887 #else
3888 	rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
3889 #endif
3890 
3891 	res.entries = (dirent64_t *)rdc->entries;
3892 	res.entries_size = rdc->buflen;
3893 	res.dir_attributes.fres.vap = &dva;
3894 	res.dir_attributes.fres.vp = vp;
3895 	res.loff = rdc->nfs3_cookie;
3896 
3897 	douprintf = 1;
3898 
3899 	if (mi->mi_io_kstats) {
3900 		mutex_enter(&mi->mi_lock);
3901 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
3902 		mutex_exit(&mi->mi_lock);
3903 	}
3904 
3905 	t = gethrtime();
3906 
3907 	error = rfs3call(VTOMI(vp), NFSPROC3_READDIR,
3908 	    xdr_READDIR3args, (caddr_t)&args,
3909 	    xdr_READDIR3vres, (caddr_t)&res, cr,
3910 	    &douprintf, &res.status, 0, fip);
3911 
3912 	if (mi->mi_io_kstats) {
3913 		mutex_enter(&mi->mi_lock);
3914 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
3915 		mutex_exit(&mi->mi_lock);
3916 	}
3917 
3918 	if (error)
3919 		goto err;
3920 
3921 	nfs3_cache_post_op_vattr(vp, &res.dir_attributes, t, cr);
3922 
3923 	error = geterrno3(res.status);
3924 	if (error) {
3925 		PURGE_STALE_FH(error, vp, cr);
3926 		goto err;
3927 	}
3928 
3929 	if (mi->mi_io_kstats) {
3930 		mutex_enter(&mi->mi_lock);
3931 		KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
3932 		KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
3933 		mutex_exit(&mi->mi_lock);
3934 	}
3935 
3936 	rdc->nfs3_ncookie = res.loff;
3937 	rp->r_cookieverf = res.cookieverf;
3938 	rdc->eof = res.eof ? 1 : 0;
3939 	rdc->entlen = res.size;
3940 	ASSERT(rdc->entlen <= rdc->buflen);
3941 	rdc->error = 0;
3942 	return;
3943 
3944 err:
3945 	kmem_free(rdc->entries, rdc->buflen);
3946 	rdc->entries = NULL;
3947 	rdc->error = error;
3948 }
3949 
3950 /*
3951  * Read directory entries.
3952  * There are some weird things to look out for here.  The uio_loffset
3953  * field is either 0 or it is the offset returned from a previous
3954  * readdir.  It is an opaque value used by the server to find the
3955  * correct directory block to read. The count field is the number
3956  * of blocks to read on the server.  This is advisory only, the server
3957  * may return only one block's worth of entries.  Entries may be compressed
3958  * on the server.
3959  */
3960 static void
3961 nfs3readdirplus(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
3962 {
3963 	int error;
3964 	READDIRPLUS3args args;
3965 	READDIRPLUS3vres res;
3966 	vattr_t dva;
3967 	rnode_t *rp;
3968 	mntinfo_t *mi;
3969 	int douprintf;
3970 	failinfo_t fi, *fip = NULL;
3971 
3972 	rp = VTOR(vp);
3973 	mi = VTOMI(vp);
3974 	ASSERT(curproc->p_zone == mi->mi_zone);
3975 
3976 	args.dir = *RTOFH3(rp);
3977 	args.cookie = (cookie3)rdc->nfs3_cookie;
3978 	args.cookieverf = rp->r_cookieverf;
3979 	args.dircount = rdc->buflen;
3980 	args.maxcount = mi->mi_tsize;
3981 
3982 	/*
3983 	 * NFS client failover support
3984 	 * suppress failover unless we have a zero cookie
3985 	 */
3986 	if (args.cookie == (cookie3)0) {
3987 		fi.vp = vp;
3988 		fi.fhp = (caddr_t)&args.dir;
3989 		fi.copyproc = nfs3copyfh;
3990 		fi.lookupproc = nfs3lookup;
3991 		fi.xattrdirproc = acl_getxattrdir3;
3992 		fip = &fi;
3993 	}
3994 
3995 #ifdef DEBUG
3996 	rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
3997 #else
3998 	rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
3999 #endif
4000 
4001 	res.entries = (dirent64_t *)rdc->entries;
4002 	res.entries_size = rdc->buflen;
4003 	res.dir_attributes.fres.vap = &dva;
4004 	res.dir_attributes.fres.vp = vp;
4005 	res.loff = rdc->nfs3_cookie;
4006 	res.credentials = cr;
4007 
4008 	douprintf = 1;
4009 
4010 	if (mi->mi_io_kstats) {
4011 		mutex_enter(&mi->mi_lock);
4012 		kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
4013 		mutex_exit(&mi->mi_lock);
4014 	}
4015 
4016 	res.time = gethrtime();
4017 
4018 	error = rfs3call(mi, NFSPROC3_READDIRPLUS,
4019 	    xdr_READDIRPLUS3args, (caddr_t)&args,
4020 	    xdr_READDIRPLUS3vres, (caddr_t)&res, cr,
4021 	    &douprintf, &res.status, 0, fip);
4022 
4023 	if (mi->mi_io_kstats) {
4024 		mutex_enter(&mi->mi_lock);
4025 		kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
4026 		mutex_exit(&mi->mi_lock);
4027 	}
4028 
4029 	if (error) {
4030 		goto err;
4031 	}
4032 
4033 	nfs3_cache_post_op_vattr(vp, &res.dir_attributes, res.time, cr);
4034 
4035 	error = geterrno3(res.status);
4036 	if (error) {
4037 		PURGE_STALE_FH(error, vp, cr);
4038 		if (error == EOPNOTSUPP) {
4039 			mutex_enter(&mi->mi_lock);
4040 			mi->mi_flags |= MI_READDIRONLY;
4041 			mutex_exit(&mi->mi_lock);
4042 		}
4043 		goto err;
4044 	}
4045 
4046 	if (mi->mi_io_kstats) {
4047 		mutex_enter(&mi->mi_lock);
4048 		KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4049 		KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4050 		mutex_exit(&mi->mi_lock);
4051 	}
4052 
4053 	rdc->nfs3_ncookie = res.loff;
4054 	rp->r_cookieverf = res.cookieverf;
4055 	rdc->eof = res.eof ? 1 : 0;
4056 	rdc->entlen = res.size;
4057 	ASSERT(rdc->entlen <= rdc->buflen);
4058 	rdc->error = 0;
4059 
4060 	return;
4061 
4062 err:
4063 	kmem_free(rdc->entries, rdc->buflen);
4064 	rdc->entries = NULL;
4065 	rdc->error = error;
4066 }
4067 
4068 #ifdef DEBUG
4069 static int nfs3_bio_do_stop = 0;
4070 #endif
4071 
4072 static int
4073 nfs3_bio(struct buf *bp, stable_how *stab_comm, cred_t *cr)
4074 {
4075 	rnode_t *rp = VTOR(bp->b_vp);
4076 	int count;
4077 	int error;
4078 	cred_t *cred;
4079 	offset_t offset;
4080 
4081 	ASSERT(curproc->p_zone == VTOMI(bp->b_vp)->mi_zone);
4082 	offset = ldbtob(bp->b_lblkno);
4083 
4084 	DTRACE_IO1(start, struct buf *, bp);
4085 
4086 	if (bp->b_flags & B_READ) {
4087 		mutex_enter(&rp->r_statelock);
4088 		if (rp->r_cred != NULL) {
4089 			cred = rp->r_cred;
4090 			crhold(cred);
4091 		} else {
4092 			rp->r_cred = cr;
4093 			crhold(cr);
4094 			cred = cr;
4095 			crhold(cred);
4096 		}
4097 		mutex_exit(&rp->r_statelock);
4098 	read_again:
4099 		error = bp->b_error = nfs3read(bp->b_vp, bp->b_un.b_addr,
4100 		    offset, bp->b_bcount, &bp->b_resid, cred);
4101 		crfree(cred);
4102 		if (!error) {
4103 			if (bp->b_resid) {
4104 				/*
4105 				 * Didn't get it all because we hit EOF,
4106 				 * zero all the memory beyond the EOF.
4107 				 */
4108 				/* bzero(rdaddr + */
4109 				bzero(bp->b_un.b_addr +
4110 				    bp->b_bcount - bp->b_resid, bp->b_resid);
4111 			}
4112 			mutex_enter(&rp->r_statelock);
4113 			if (bp->b_resid == bp->b_bcount &&
4114 			    offset >= rp->r_size) {
4115 				/*
4116 				 * We didn't read anything at all as we are
4117 				 * past EOF.  Return an error indicator back
4118 				 * but don't destroy the pages (yet).
4119 				 */
4120 				error = NFS_EOF;
4121 			}
4122 			mutex_exit(&rp->r_statelock);
4123 		} else if (error == EACCES) {
4124 			mutex_enter(&rp->r_statelock);
4125 			if (cred != cr) {
4126 				if (rp->r_cred != NULL)
4127 					crfree(rp->r_cred);
4128 				rp->r_cred = cr;
4129 				crhold(cr);
4130 				cred = cr;
4131 				crhold(cred);
4132 				mutex_exit(&rp->r_statelock);
4133 				goto read_again;
4134 			}
4135 			mutex_exit(&rp->r_statelock);
4136 		}
4137 	} else {
4138 		if (!(rp->r_flags & RSTALE)) {
4139 			mutex_enter(&rp->r_statelock);
4140 			if (rp->r_cred != NULL) {
4141 				cred = rp->r_cred;
4142 				crhold(cred);
4143 			} else {
4144 				rp->r_cred = cr;
4145 				crhold(cr);
4146 				cred = cr;
4147 				crhold(cred);
4148 			}
4149 			mutex_exit(&rp->r_statelock);
4150 		write_again:
4151 			mutex_enter(&rp->r_statelock);
4152 			count = MIN(bp->b_bcount, rp->r_size - offset);
4153 			mutex_exit(&rp->r_statelock);
4154 			if (count < 0)
4155 				cmn_err(CE_PANIC, "nfs3_bio: write count < 0");
4156 #ifdef DEBUG
4157 			if (count == 0) {
4158 				zcmn_err(getzoneid(), CE_WARN,
4159 				    "nfs3_bio: zero length write at %lld",
4160 				    offset);
4161 				nfs_printfhandle(&rp->r_fh);
4162 				if (nfs3_bio_do_stop)
4163 					debug_enter("nfs3_bio");
4164 			}
4165 #endif
4166 			error = nfs3write(bp->b_vp, bp->b_un.b_addr, offset,
4167 			    count, cred, stab_comm);
4168 			if (error == EACCES) {
4169 				mutex_enter(&rp->r_statelock);
4170 				if (cred != cr) {
4171 					if (rp->r_cred != NULL)
4172 						crfree(rp->r_cred);
4173 					rp->r_cred = cr;
4174 					crhold(cr);
4175 					crfree(cred);
4176 					cred = cr;
4177 					crhold(cred);
4178 					mutex_exit(&rp->r_statelock);
4179 					goto write_again;
4180 				}
4181 				mutex_exit(&rp->r_statelock);
4182 			}
4183 			bp->b_error = error;
4184 			if (error && error != EINTR) {
4185 				/*
4186 				 * Don't print EDQUOT errors on the console.
4187 				 * Don't print asynchronous EACCES errors.
4188 				 * Don't print EFBIG errors.
4189 				 * Print all other write errors.
4190 				 */
4191 				if (error != EDQUOT && error != EFBIG &&
4192 				    (error != EACCES ||
4193 				    !(bp->b_flags & B_ASYNC)))
4194 					nfs_write_error(bp->b_vp, error, cred);
4195 				/*
4196 				 * Update r_error and r_flags as appropriate.
4197 				 * If the error was ESTALE, then mark the
4198 				 * rnode as not being writeable and save
4199 				 * the error status.  Otherwise, save any
4200 				 * errors which occur from asynchronous
4201 				 * page invalidations.  Any errors occurring
4202 				 * from other operations should be saved
4203 				 * by the caller.
4204 				 */
4205 				mutex_enter(&rp->r_statelock);
4206 				if (error == ESTALE) {
4207 					rp->r_flags |= RSTALE;
4208 					if (!rp->r_error)
4209 						rp->r_error = error;
4210 				} else if (!rp->r_error &&
4211 				    (bp->b_flags &
4212 				    (B_INVAL|B_FORCE|B_ASYNC)) ==
4213 				    (B_INVAL|B_FORCE|B_ASYNC)) {
4214 					rp->r_error = error;
4215 				}
4216 				mutex_exit(&rp->r_statelock);
4217 			}
4218 			crfree(cred);
4219 		} else
4220 			error = rp->r_error;
4221 	}
4222 
4223 	if (error != 0 && error != NFS_EOF)
4224 		bp->b_flags |= B_ERROR;
4225 
4226 	DTRACE_IO1(done, struct buf *, bp);
4227 
4228 	return (error);
4229 }
4230 
4231 static int
4232 nfs3_fid(vnode_t *vp, fid_t *fidp)
4233 {
4234 	rnode_t *rp;
4235 
4236 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
4237 		return (EIO);
4238 	rp = VTOR(vp);
4239 
4240 	if (fidp->fid_len < (ushort_t)rp->r_fh.fh_len) {
4241 		fidp->fid_len = rp->r_fh.fh_len;
4242 		return (ENOSPC);
4243 	}
4244 	fidp->fid_len = rp->r_fh.fh_len;
4245 	bcopy(rp->r_fh.fh_buf, fidp->fid_data, fidp->fid_len);
4246 	return (0);
4247 }
4248 
4249 /* ARGSUSED2 */
4250 static int
4251 nfs3_rwlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4252 {
4253 	rnode_t *rp = VTOR(vp);
4254 
4255 	if (!write_lock) {
4256 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4257 		return (V_WRITELOCK_FALSE);
4258 	}
4259 
4260 	if ((rp->r_flags & RDIRECTIO) || (VTOMI(vp)->mi_flags & MI_DIRECTIO)) {
4261 		(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4262 		if (rp->r_mapcnt == 0 && !vn_has_cached_data(vp))
4263 			return (V_WRITELOCK_FALSE);
4264 		nfs_rw_exit(&rp->r_rwlock);
4265 	}
4266 
4267 	(void) nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, FALSE);
4268 	return (V_WRITELOCK_TRUE);
4269 }
4270 
4271 /* ARGSUSED */
4272 static void
4273 nfs3_rwunlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4274 {
4275 	rnode_t *rp = VTOR(vp);
4276 
4277 	nfs_rw_exit(&rp->r_rwlock);
4278 }
4279 
4280 /* ARGSUSED */
4281 static int
4282 nfs3_seek(vnode_t *vp, offset_t ooff, offset_t *noffp)
4283 {
4284 
4285 	/*
4286 	 * Because we stuff the readdir cookie into the offset field
4287 	 * someone may attempt to do an lseek with the cookie which
4288 	 * we want to succeed.
4289 	 */
4290 	if (vp->v_type == VDIR)
4291 		return (0);
4292 	if (*noffp < 0)
4293 		return (EINVAL);
4294 	return (0);
4295 }
4296 
4297 /*
4298  * number of nfs3_bsize blocks to read ahead.
4299  */
4300 static int nfs3_nra = 4;
4301 
4302 #ifdef DEBUG
4303 static int nfs3_lostpage = 0;	/* number of times we lost original page */
4304 #endif
4305 
4306 /*
4307  * Return all the pages from [off..off+len) in file
4308  */
4309 static int
4310 nfs3_getpage(vnode_t *vp, offset_t off, size_t len, uint_t *protp,
4311 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4312 	enum seg_rw rw, cred_t *cr)
4313 {
4314 	rnode_t *rp;
4315 	int error;
4316 	mntinfo_t *mi;
4317 
4318 	if (vp->v_flag & VNOMAP)
4319 		return (ENOSYS);
4320 
4321 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
4322 		return (EIO);
4323 	if (protp != NULL)
4324 		*protp = PROT_ALL;
4325 
4326 	/*
4327 	 * Now valididate that the caches are up to date.
4328 	 */
4329 	error = nfs3_validate_caches(vp, cr);
4330 	if (error)
4331 		return (error);
4332 
4333 	rp = VTOR(vp);
4334 	mi = VTOMI(vp);
4335 retry:
4336 	mutex_enter(&rp->r_statelock);
4337 
4338 	/*
4339 	 * Don't create dirty pages faster than they
4340 	 * can be cleaned so that the system doesn't
4341 	 * get imbalanced.  If the async queue is
4342 	 * maxed out, then wait for it to drain before
4343 	 * creating more dirty pages.  Also, wait for
4344 	 * any threads doing pagewalks in the vop_getattr
4345 	 * entry points so that they don't block for
4346 	 * long periods.
4347 	 */
4348 	if (rw == S_CREATE) {
4349 		while ((mi->mi_max_threads != 0 &&
4350 		    rp->r_awcount > 2 * mi->mi_max_threads) ||
4351 		    rp->r_gcount > 0)
4352 			cv_wait(&rp->r_cv, &rp->r_statelock);
4353 	}
4354 
4355 	/*
4356 	 * If we are getting called as a side effect of an nfs_write()
4357 	 * operation the local file size might not be extended yet.
4358 	 * In this case we want to be able to return pages of zeroes.
4359 	 */
4360 	if (off + len > rp->r_size + PAGEOFFSET && seg != segkmap) {
4361 		mutex_exit(&rp->r_statelock);
4362 		return (EFAULT);		/* beyond EOF */
4363 	}
4364 
4365 	mutex_exit(&rp->r_statelock);
4366 
4367 	if (len <= PAGESIZE) {
4368 		error = nfs3_getapage(vp, off, len, protp, pl, plsz,
4369 		    seg, addr, rw, cr);
4370 	} else {
4371 		error = pvn_getpages(nfs3_getapage, vp, off, len, protp,
4372 		    pl, plsz, seg, addr, rw, cr);
4373 	}
4374 
4375 	switch (error) {
4376 	case NFS_EOF:
4377 		nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
4378 		goto retry;
4379 	case ESTALE:
4380 		PURGE_STALE_FH(error, vp, cr);
4381 	}
4382 
4383 	return (error);
4384 }
4385 
4386 /*
4387  * Called from pvn_getpages or nfs3_getpage to get a particular page.
4388  */
4389 /* ARGSUSED */
4390 static int
4391 nfs3_getapage(vnode_t *vp, u_offset_t off, size_t len, uint_t *protp,
4392 	page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4393 	enum seg_rw rw, cred_t *cr)
4394 {
4395 	rnode_t *rp;
4396 	uint_t bsize;
4397 	struct buf *bp;
4398 	page_t *pp;
4399 	u_offset_t lbn;
4400 	u_offset_t io_off;
4401 	u_offset_t blkoff;
4402 	u_offset_t rablkoff;
4403 	size_t io_len;
4404 	uint_t blksize;
4405 	int error;
4406 	int readahead;
4407 	int readahead_issued = 0;
4408 	int ra_window; /* readahead window */
4409 	page_t *pagefound;
4410 	page_t *savepp;
4411 
4412 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
4413 		return (EIO);
4414 	rp = VTOR(vp);
4415 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4416 
4417 reread:
4418 	bp = NULL;
4419 	pp = NULL;
4420 	pagefound = NULL;
4421 
4422 	if (pl != NULL)
4423 		pl[0] = NULL;
4424 
4425 	error = 0;
4426 	lbn = off / bsize;
4427 	blkoff = lbn * bsize;
4428 
4429 	/*
4430 	 * Queueing up the readahead before doing the synchronous read
4431 	 * results in a significant increase in read throughput because
4432 	 * of the increased parallelism between the async threads and
4433 	 * the process context.
4434 	 */
4435 	if ((off & ((vp->v_vfsp->vfs_bsize) - 1)) == 0 &&
4436 	    rw != S_CREATE &&
4437 	    !(vp->v_flag & VNOCACHE)) {
4438 		mutex_enter(&rp->r_statelock);
4439 
4440 		/*
4441 		 * Calculate the number of readaheads to do.
4442 		 * a) No readaheads at offset = 0.
4443 		 * b) Do maximum(nfs3_nra) readaheads when the readahead
4444 		 *    window is closed.
4445 		 * c) Do readaheads between 1 to (nfs3_nra - 1) depending
4446 		 *    upon how far the readahead window is open or close.
4447 		 * d) No readaheads if rp->r_nextr is not within the scope
4448 		 *    of the readahead window (random i/o).
4449 		 */
4450 
4451 		if (off == 0)
4452 			readahead = 0;
4453 		else if (blkoff == rp->r_nextr)
4454 			readahead = nfs3_nra;
4455 		else if (rp->r_nextr > blkoff &&
4456 				((ra_window = (rp->r_nextr - blkoff) / bsize)
4457 					<= (nfs3_nra - 1)))
4458 			readahead = nfs3_nra - ra_window;
4459 		else
4460 			readahead = 0;
4461 
4462 		rablkoff = rp->r_nextr;
4463 		while (readahead > 0 && rablkoff + bsize < rp->r_size) {
4464 			mutex_exit(&rp->r_statelock);
4465 			if (nfs_async_readahead(vp, rablkoff + bsize,
4466 			    addr + (rablkoff + bsize - off), seg, cr,
4467 			    nfs3_readahead) < 0) {
4468 				mutex_enter(&rp->r_statelock);
4469 				break;
4470 			}
4471 			readahead--;
4472 			rablkoff += bsize;
4473 			/*
4474 			 * Indicate that we did a readahead so
4475 			 * readahead offset is not updated
4476 			 * by the synchronous read below.
4477 			 */
4478 			readahead_issued = 1;
4479 			mutex_enter(&rp->r_statelock);
4480 			/*
4481 			 * set readahead offset to
4482 			 * offset of last async readahead
4483 			 * request.
4484 			 */
4485 			rp->r_nextr = rablkoff;
4486 		}
4487 		mutex_exit(&rp->r_statelock);
4488 	}
4489 
4490 again:
4491 	if ((pagefound = page_exists(vp, off)) == NULL) {
4492 		if (pl == NULL) {
4493 			(void) nfs_async_readahead(vp, blkoff, addr, seg, cr,
4494 			    nfs3_readahead);
4495 		} else if (rw == S_CREATE) {
4496 			/*
4497 			 * Block for this page is not allocated, or the offset
4498 			 * is beyond the current allocation size, or we're
4499 			 * allocating a swap slot and the page was not found,
4500 			 * so allocate it and return a zero page.
4501 			 */
4502 			if ((pp = page_create_va(vp, off,
4503 			    PAGESIZE, PG_WAIT, seg, addr)) == NULL)
4504 				cmn_err(CE_PANIC, "nfs3_getapage: page_create");
4505 			io_len = PAGESIZE;
4506 			mutex_enter(&rp->r_statelock);
4507 			rp->r_nextr = off + PAGESIZE;
4508 			mutex_exit(&rp->r_statelock);
4509 		} else {
4510 			/*
4511 			 * Need to go to server to get a BLOCK, exception to
4512 			 * that being while reading at offset = 0 or doing
4513 			 * random i/o, in that case read only a PAGE.
4514 			 */
4515 			mutex_enter(&rp->r_statelock);
4516 			if (blkoff < rp->r_size &&
4517 			    blkoff + bsize >= rp->r_size) {
4518 				/*
4519 				 * If only a block or less is left in
4520 				 * the file, read all that is remaining.
4521 				 */
4522 				if (rp->r_size <= off) {
4523 					/*
4524 					 * Trying to access beyond EOF,
4525 					 * set up to get at least one page.
4526 					 */
4527 					blksize = off + PAGESIZE - blkoff;
4528 				} else
4529 					blksize = rp->r_size - blkoff;
4530 			} else if ((off == 0) ||
4531 				(off != rp->r_nextr && !readahead_issued)) {
4532 				blksize = PAGESIZE;
4533 				blkoff = off; /* block = page here */
4534 			} else
4535 				blksize = bsize;
4536 			mutex_exit(&rp->r_statelock);
4537 
4538 			pp = pvn_read_kluster(vp, off, seg, addr, &io_off,
4539 			    &io_len, blkoff, blksize, 0);
4540 
4541 			/*
4542 			 * Some other thread has entered the page,
4543 			 * so just use it.
4544 			 */
4545 			if (pp == NULL)
4546 				goto again;
4547 
4548 			/*
4549 			 * Now round the request size up to page boundaries.
4550 			 * This ensures that the entire page will be
4551 			 * initialized to zeroes if EOF is encountered.
4552 			 */
4553 			io_len = ptob(btopr(io_len));
4554 
4555 			bp = pageio_setup(pp, io_len, vp, B_READ);
4556 			ASSERT(bp != NULL);
4557 
4558 			/*
4559 			 * pageio_setup should have set b_addr to 0.  This
4560 			 * is correct since we want to do I/O on a page
4561 			 * boundary.  bp_mapin will use this addr to calculate
4562 			 * an offset, and then set b_addr to the kernel virtual
4563 			 * address it allocated for us.
4564 			 */
4565 			ASSERT(bp->b_un.b_addr == 0);
4566 
4567 			bp->b_edev = 0;
4568 			bp->b_dev = 0;
4569 			bp->b_lblkno = lbtodb(io_off);
4570 			bp->b_file = vp;
4571 			bp->b_offset = (offset_t)off;
4572 			bp_mapin(bp);
4573 
4574 			/*
4575 			 * If doing a write beyond what we believe is EOF,
4576 			 * don't bother trying to read the pages from the
4577 			 * server, we'll just zero the pages here.  We
4578 			 * don't check that the rw flag is S_WRITE here
4579 			 * because some implementations may attempt a
4580 			 * read access to the buffer before copying data.
4581 			 */
4582 			mutex_enter(&rp->r_statelock);
4583 			if (io_off >= rp->r_size && seg == segkmap) {
4584 				mutex_exit(&rp->r_statelock);
4585 				bzero(bp->b_un.b_addr, io_len);
4586 			} else {
4587 				mutex_exit(&rp->r_statelock);
4588 				error = nfs3_bio(bp, NULL, cr);
4589 			}
4590 
4591 			/*
4592 			 * Unmap the buffer before freeing it.
4593 			 */
4594 			bp_mapout(bp);
4595 			pageio_done(bp);
4596 
4597 			savepp = pp;
4598 			do {
4599 				pp->p_fsdata = C_NOCOMMIT;
4600 			} while ((pp = pp->p_next) != savepp);
4601 
4602 			if (error == NFS_EOF) {
4603 				/*
4604 				 * If doing a write system call just return
4605 				 * zeroed pages, else user tried to get pages
4606 				 * beyond EOF, return error.  We don't check
4607 				 * that the rw flag is S_WRITE here because
4608 				 * some implementations may attempt a read
4609 				 * access to the buffer before copying data.
4610 				 */
4611 				if (seg == segkmap)
4612 					error = 0;
4613 				else
4614 					error = EFAULT;
4615 			}
4616 
4617 			if (!readahead_issued && !error) {
4618 			    mutex_enter(&rp->r_statelock);
4619 			    rp->r_nextr = io_off + io_len;
4620 			    mutex_exit(&rp->r_statelock);
4621 			}
4622 		}
4623 	}
4624 
4625 out:
4626 	if (pl == NULL)
4627 		return (error);
4628 
4629 	if (error) {
4630 		if (pp != NULL)
4631 			pvn_read_done(pp, B_ERROR);
4632 		return (error);
4633 	}
4634 
4635 	if (pagefound) {
4636 		se_t se = (rw == S_CREATE ? SE_EXCL : SE_SHARED);
4637 
4638 		/*
4639 		 * Page exists in the cache, acquire the appropriate lock.
4640 		 * If this fails, start all over again.
4641 		 */
4642 		if ((pp = page_lookup(vp, off, se)) == NULL) {
4643 #ifdef DEBUG
4644 			nfs3_lostpage++;
4645 #endif
4646 			goto reread;
4647 		}
4648 		pl[0] = pp;
4649 		pl[1] = NULL;
4650 		return (0);
4651 	}
4652 
4653 	if (pp != NULL)
4654 		pvn_plist_init(pp, pl, plsz, off, io_len, rw);
4655 
4656 	return (error);
4657 }
4658 
4659 static void
4660 nfs3_readahead(vnode_t *vp, u_offset_t blkoff, caddr_t addr, struct seg *seg,
4661 	cred_t *cr)
4662 {
4663 	int error;
4664 	page_t *pp;
4665 	u_offset_t io_off;
4666 	size_t io_len;
4667 	struct buf *bp;
4668 	uint_t bsize, blksize;
4669 	rnode_t *rp = VTOR(vp);
4670 	page_t *savepp;
4671 
4672 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
4673 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4674 
4675 	mutex_enter(&rp->r_statelock);
4676 	if (blkoff < rp->r_size && blkoff + bsize > rp->r_size) {
4677 		/*
4678 		 * If less than a block left in file read less
4679 		 * than a block.
4680 		 */
4681 		blksize = rp->r_size - blkoff;
4682 	} else
4683 		blksize = bsize;
4684 	mutex_exit(&rp->r_statelock);
4685 
4686 	pp = pvn_read_kluster(vp, blkoff, segkmap, addr,
4687 	    &io_off, &io_len, blkoff, blksize, 1);
4688 	/*
4689 	 * The isra flag passed to the kluster function is 1, we may have
4690 	 * gotten a return value of NULL for a variety of reasons (# of free
4691 	 * pages < minfree, someone entered the page on the vnode etc). In all
4692 	 * cases, we want to punt on the readahead.
4693 	 */
4694 	if (pp == NULL)
4695 		return;
4696 
4697 	/*
4698 	 * Now round the request size up to page boundaries.
4699 	 * This ensures that the entire page will be
4700 	 * initialized to zeroes if EOF is encountered.
4701 	 */
4702 	io_len = ptob(btopr(io_len));
4703 
4704 	bp = pageio_setup(pp, io_len, vp, B_READ);
4705 	ASSERT(bp != NULL);
4706 
4707 	/*
4708 	 * pageio_setup should have set b_addr to 0.  This is correct since
4709 	 * we want to do I/O on a page boundary. bp_mapin() will use this addr
4710 	 * to calculate an offset, and then set b_addr to the kernel virtual
4711 	 * address it allocated for us.
4712 	 */
4713 	ASSERT(bp->b_un.b_addr == 0);
4714 
4715 	bp->b_edev = 0;
4716 	bp->b_dev = 0;
4717 	bp->b_lblkno = lbtodb(io_off);
4718 	bp->b_file = vp;
4719 	bp->b_offset = (offset_t)blkoff;
4720 	bp_mapin(bp);
4721 
4722 	/*
4723 	 * If doing a write beyond what we believe is EOF, don't bother trying
4724 	 * to read the pages from the server, we'll just zero the pages here.
4725 	 * We don't check that the rw flag is S_WRITE here because some
4726 	 * implementations may attempt a read access to the buffer before
4727 	 * copying data.
4728 	 */
4729 	mutex_enter(&rp->r_statelock);
4730 	if (io_off >= rp->r_size && seg == segkmap) {
4731 		mutex_exit(&rp->r_statelock);
4732 		bzero(bp->b_un.b_addr, io_len);
4733 		error = 0;
4734 	} else {
4735 		mutex_exit(&rp->r_statelock);
4736 		error = nfs3_bio(bp, NULL, cr);
4737 		if (error == NFS_EOF)
4738 			error = 0;
4739 	}
4740 
4741 	/*
4742 	 * Unmap the buffer before freeing it.
4743 	 */
4744 	bp_mapout(bp);
4745 	pageio_done(bp);
4746 
4747 	savepp = pp;
4748 	do {
4749 		pp->p_fsdata = C_NOCOMMIT;
4750 	} while ((pp = pp->p_next) != savepp);
4751 
4752 	pvn_read_done(pp, error ? B_READ | B_ERROR : B_READ);
4753 
4754 	/*
4755 	 * In case of error set readahead offset
4756 	 * to the lowest offset.
4757 	 * pvn_read_done() calls VN_DISPOSE to destroy the pages
4758 	 */
4759 	if (error && rp->r_nextr > io_off) {
4760 		mutex_enter(&rp->r_statelock);
4761 		if (rp->r_nextr > io_off)
4762 			rp->r_nextr = io_off;
4763 		mutex_exit(&rp->r_statelock);
4764 	}
4765 }
4766 
4767 /*
4768  * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
4769  * If len == 0, do from off to EOF.
4770  *
4771  * The normal cases should be len == 0 && off == 0 (entire vp list),
4772  * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
4773  * (from pageout).
4774  */
4775 static int
4776 nfs3_putpage(vnode_t *vp, offset_t off, size_t len, int flags, cred_t *cr)
4777 {
4778 	int error;
4779 	rnode_t *rp;
4780 
4781 	ASSERT(cr != NULL);
4782 
4783 	/*
4784 	 * XXX - Why should this check be made here?
4785 	 */
4786 	if (vp->v_flag & VNOMAP)
4787 		return (ENOSYS);
4788 	if (len == 0 && !(flags & B_INVAL) && vn_is_readonly(vp))
4789 		return (0);
4790 	if (!(flags & B_ASYNC) && curproc->p_zone != VTOMI(vp)->mi_zone)
4791 		return (EIO);
4792 
4793 	rp = VTOR(vp);
4794 	mutex_enter(&rp->r_statelock);
4795 	rp->r_count++;
4796 	mutex_exit(&rp->r_statelock);
4797 	error = nfs_putpages(vp, off, len, flags, cr);
4798 	mutex_enter(&rp->r_statelock);
4799 	rp->r_count--;
4800 	cv_broadcast(&rp->r_cv);
4801 	mutex_exit(&rp->r_statelock);
4802 
4803 	return (error);
4804 }
4805 
4806 /*
4807  * Write out a single page, possibly klustering adjacent dirty pages.
4808  */
4809 int
4810 nfs3_putapage(vnode_t *vp, page_t *pp, u_offset_t *offp, size_t *lenp,
4811 	int flags, cred_t *cr)
4812 {
4813 	u_offset_t io_off;
4814 	u_offset_t lbn_off;
4815 	u_offset_t lbn;
4816 	size_t io_len;
4817 	uint_t bsize;
4818 	int error;
4819 	rnode_t *rp;
4820 
4821 	ASSERT(!vn_is_readonly(vp));
4822 	ASSERT(pp != NULL);
4823 	ASSERT(cr != NULL);
4824 	ASSERT((flags & B_ASYNC) || curproc->p_zone == VTOMI(vp)->mi_zone);
4825 
4826 	rp = VTOR(vp);
4827 	ASSERT(rp->r_count > 0);
4828 
4829 	bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4830 	lbn = pp->p_offset / bsize;
4831 	lbn_off = lbn * bsize;
4832 
4833 	/*
4834 	 * Find a kluster that fits in one block, or in
4835 	 * one page if pages are bigger than blocks.  If
4836 	 * there is less file space allocated than a whole
4837 	 * page, we'll shorten the i/o request below.
4838 	 */
4839 	pp = pvn_write_kluster(vp, pp, &io_off, &io_len, lbn_off,
4840 	    roundup(bsize, PAGESIZE), flags);
4841 
4842 	/*
4843 	 * pvn_write_kluster shouldn't have returned a page with offset
4844 	 * behind the original page we were given.  Verify that.
4845 	 */
4846 	ASSERT((pp->p_offset / bsize) >= lbn);
4847 
4848 	/*
4849 	 * Now pp will have the list of kept dirty pages marked for
4850 	 * write back.  It will also handle invalidation and freeing
4851 	 * of pages that are not dirty.  Check for page length rounding
4852 	 * problems.
4853 	 */
4854 	if (io_off + io_len > lbn_off + bsize) {
4855 		ASSERT((io_off + io_len) - (lbn_off + bsize) < PAGESIZE);
4856 		io_len = lbn_off + bsize - io_off;
4857 	}
4858 	/*
4859 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4860 	 * consistent value of r_size. RMODINPROGRESS is set in writerp().
4861 	 * When RMODINPROGRESS is set it indicates that a uiomove() is in
4862 	 * progress and the r_size has not been made consistent with the
4863 	 * new size of the file. When the uiomove() completes the r_size is
4864 	 * updated and the RMODINPROGRESS flag is cleared.
4865 	 *
4866 	 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
4867 	 * consistent value of r_size. Without this handshaking, it is
4868 	 * possible that nfs(3)_bio() picks  up the old value of r_size
4869 	 * before the uiomove() in writerp() completes. This will result
4870 	 * in the write through nfs(3)_bio() being dropped.
4871 	 *
4872 	 * More precisely, there is a window between the time the uiomove()
4873 	 * completes and the time the r_size is updated. If a VOP_PUTPAGE()
4874 	 * operation intervenes in this window, the page will be picked up,
4875 	 * because it is dirty (it will be unlocked, unless it was
4876 	 * pagecreate'd). When the page is picked up as dirty, the dirty
4877 	 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
4878 	 * checked. This will still be the old size. Therefore the page will
4879 	 * not be written out. When segmap_release() calls VOP_PUTPAGE(),
4880 	 * the page will be found to be clean and the write will be dropped.
4881 	 */
4882 	if (rp->r_flags & RMODINPROGRESS) {
4883 		mutex_enter(&rp->r_statelock);
4884 		if ((rp->r_flags & RMODINPROGRESS) &&
4885 		    rp->r_modaddr + MAXBSIZE > io_off &&
4886 		    rp->r_modaddr < io_off + io_len) {
4887 			page_t *plist;
4888 			/*
4889 			 * A write is in progress for this region of the file.
4890 			 * If we did not detect RMODINPROGRESS here then this
4891 			 * path through nfs_putapage() would eventually go to
4892 			 * nfs(3)_bio() and may not write out all of the data
4893 			 * in the pages. We end up losing data. So we decide
4894 			 * to set the modified bit on each page in the page
4895 			 * list and mark the rnode with RDIRTY. This write
4896 			 * will be restarted at some later time.
4897 			 */
4898 			plist = pp;
4899 			while (plist != NULL) {
4900 				pp = plist;
4901 				page_sub(&plist, pp);
4902 				hat_setmod(pp);
4903 				page_io_unlock(pp);
4904 				page_unlock(pp);
4905 			}
4906 			rp->r_flags |= RDIRTY;
4907 			mutex_exit(&rp->r_statelock);
4908 			if (offp)
4909 				*offp = io_off;
4910 			if (lenp)
4911 				*lenp = io_len;
4912 			return (0);
4913 		}
4914 		mutex_exit(&rp->r_statelock);
4915 	}
4916 
4917 	if (flags & B_ASYNC) {
4918 		error = nfs_async_putapage(vp, pp, io_off, io_len, flags, cr,
4919 		    nfs3_sync_putapage);
4920 	} else
4921 		error = nfs3_sync_putapage(vp, pp, io_off, io_len, flags, cr);
4922 
4923 	if (offp)
4924 		*offp = io_off;
4925 	if (lenp)
4926 		*lenp = io_len;
4927 	return (error);
4928 }
4929 
4930 static int
4931 nfs3_sync_putapage(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
4932 	int flags, cred_t *cr)
4933 {
4934 	int error;
4935 	rnode_t *rp;
4936 
4937 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
4938 
4939 	flags |= B_WRITE;
4940 
4941 	error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
4942 
4943 	rp = VTOR(vp);
4944 
4945 	if ((error == ENOSPC || error == EDQUOT || error == EFBIG ||
4946 	    error == EACCES) &&
4947 	    (flags & (B_INVAL|B_FORCE)) != (B_INVAL|B_FORCE)) {
4948 		if (!(rp->r_flags & ROUTOFSPACE)) {
4949 			mutex_enter(&rp->r_statelock);
4950 			rp->r_flags |= ROUTOFSPACE;
4951 			mutex_exit(&rp->r_statelock);
4952 		}
4953 		flags |= B_ERROR;
4954 		pvn_write_done(pp, flags);
4955 		/*
4956 		 * If this was not an async thread, then try again to
4957 		 * write out the pages, but this time, also destroy
4958 		 * them whether or not the write is successful.  This
4959 		 * will prevent memory from filling up with these
4960 		 * pages and destroying them is the only alternative
4961 		 * if they can't be written out.
4962 		 *
4963 		 * Don't do this if this is an async thread because
4964 		 * when the pages are unlocked in pvn_write_done,
4965 		 * some other thread could have come along, locked
4966 		 * them, and queued for an async thread.  It would be
4967 		 * possible for all of the async threads to be tied
4968 		 * up waiting to lock the pages again and they would
4969 		 * all already be locked and waiting for an async
4970 		 * thread to handle them.  Deadlock.
4971 		 */
4972 		if (!(flags & B_ASYNC)) {
4973 			error = nfs3_putpage(vp, io_off, io_len,
4974 			    B_INVAL | B_FORCE, cr);
4975 		}
4976 	} else {
4977 		if (error)
4978 			flags |= B_ERROR;
4979 		else if (rp->r_flags & ROUTOFSPACE) {
4980 			mutex_enter(&rp->r_statelock);
4981 			rp->r_flags &= ~ROUTOFSPACE;
4982 			mutex_exit(&rp->r_statelock);
4983 		}
4984 		pvn_write_done(pp, flags);
4985 		if (freemem < desfree)
4986 			(void) nfs3_commit_vp(vp, (u_offset_t)0, 0, cr);
4987 	}
4988 
4989 	return (error);
4990 }
4991 
4992 static int
4993 nfs3_map(vnode_t *vp, offset_t off, struct as *as, caddr_t *addrp,
4994 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
4995 {
4996 	struct segvn_crargs vn_a;
4997 	int error;
4998 	rnode_t *rp;
4999 	struct vattr va;
5000 
5001 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
5002 		return (EIO);
5003 
5004 	if (vp->v_flag & VNOMAP)
5005 		return (ENOSYS);
5006 
5007 	if (off < 0 || off + len < 0)
5008 		return (ENXIO);
5009 
5010 	if (vp->v_type != VREG)
5011 		return (ENODEV);
5012 
5013 	/*
5014 	 * If there is cached data and if close-to-open consistency
5015 	 * checking is not turned off and if the file system is not
5016 	 * mounted readonly, then force an over the wire getattr.
5017 	 * Otherwise, just invoke nfs3getattr to get a copy of the
5018 	 * attributes.  The attribute cache will be used unless it
5019 	 * is timed out and if it is, then an over the wire getattr
5020 	 * will be issued.
5021 	 */
5022 	va.va_mask = AT_ALL;
5023 	if (vn_has_cached_data(vp) &&
5024 	    !(VTOMI(vp)->mi_flags & MI_NOCTO) && !vn_is_readonly(vp))
5025 		error = nfs3_getattr_otw(vp, &va, cr);
5026 	else
5027 		error = nfs3getattr(vp, &va, cr);
5028 	if (error)
5029 		return (error);
5030 
5031 	/*
5032 	 * Check to see if the vnode is currently marked as not cachable.
5033 	 * This means portions of the file are locked (through VOP_FRLOCK).
5034 	 * In this case the map request must be refused.  We use
5035 	 * rp->r_lkserlock to avoid a race with concurrent lock requests.
5036 	 */
5037 	rp = VTOR(vp);
5038 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
5039 		return (EINTR);
5040 
5041 	if (vp->v_flag & VNOCACHE) {
5042 		error = EAGAIN;
5043 		goto done;
5044 	}
5045 
5046 	/*
5047 	 * Don't allow concurrent locks and mapping if mandatory locking is
5048 	 * enabled.
5049 	 */
5050 	if ((flk_has_remote_locks(vp) || lm_has_sleep(vp)) &&
5051 	    MANDLOCK(vp, va.va_mode)) {
5052 		error = EAGAIN;
5053 		goto done;
5054 	}
5055 
5056 	as_rangelock(as);
5057 	if (!(flags & MAP_FIXED)) {
5058 		map_addr(addrp, len, off, 1, flags);
5059 		if (*addrp == NULL) {
5060 			as_rangeunlock(as);
5061 			error = ENOMEM;
5062 			goto done;
5063 		}
5064 	} else {
5065 		/*
5066 		 * User specified address - blow away any previous mappings
5067 		 */
5068 		(void) as_unmap(as, *addrp, len);
5069 	}
5070 
5071 	vn_a.vp = vp;
5072 	vn_a.offset = off;
5073 	vn_a.type = (flags & MAP_TYPE);
5074 	vn_a.prot = (uchar_t)prot;
5075 	vn_a.maxprot = (uchar_t)maxprot;
5076 	vn_a.flags = (flags & ~MAP_TYPE);
5077 	vn_a.cred = cr;
5078 	vn_a.amp = NULL;
5079 	vn_a.szc = 0;
5080 	vn_a.lgrp_mem_policy_flags = 0;
5081 
5082 	error = as_map(as, *addrp, len, segvn_create, &vn_a);
5083 	as_rangeunlock(as);
5084 
5085 done:
5086 	nfs_rw_exit(&rp->r_lkserlock);
5087 	return (error);
5088 }
5089 
5090 /* ARGSUSED */
5091 static int
5092 nfs3_addmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5093 	size_t len, uchar_t prot, uchar_t maxprot, uint_t flags, cred_t *cr)
5094 {
5095 	rnode_t *rp;
5096 
5097 	if (vp->v_flag & VNOMAP)
5098 		return (ENOSYS);
5099 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
5100 		return (EIO);
5101 
5102 	/*
5103 	 * Need to hold rwlock while incrementing the mapcnt so that
5104 	 * mmap'ing can be serialized with writes so that the caching
5105 	 * can be handled correctly.
5106 	 */
5107 	rp = VTOR(vp);
5108 	if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, INTR(vp)))
5109 		return (EINTR);
5110 	atomic_add_long((ulong_t *)&rp->r_mapcnt, btopr(len));
5111 	nfs_rw_exit(&rp->r_rwlock);
5112 
5113 	return (0);
5114 }
5115 
5116 static int
5117 nfs3_frlock(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5118 	offset_t offset, struct flk_callback *flk_cbp, cred_t *cr)
5119 {
5120 	netobj lm_fh3;
5121 	int rc;
5122 	u_offset_t start, end;
5123 	rnode_t *rp;
5124 	int error = 0, intr = INTR(vp);
5125 
5126 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
5127 		return (EIO);
5128 	/* check for valid cmd parameter */
5129 	if (cmd != F_GETLK && cmd != F_SETLK && cmd != F_SETLKW)
5130 		return (EINVAL);
5131 
5132 	/* Verify l_type. */
5133 	switch (bfp->l_type) {
5134 	case F_RDLCK:
5135 		if (cmd != F_GETLK && !(flag & FREAD))
5136 			return (EBADF);
5137 		break;
5138 	case F_WRLCK:
5139 		if (cmd != F_GETLK && !(flag & FWRITE))
5140 			return (EBADF);
5141 		break;
5142 	case F_UNLCK:
5143 		intr = 0;
5144 		break;
5145 
5146 	default:
5147 		return (EINVAL);
5148 	}
5149 
5150 	/* check the validity of the lock range */
5151 	if (rc = flk_convert_lock_data(vp, bfp, &start, &end, offset))
5152 		return (rc);
5153 	if (rc = flk_check_lock_data(start, end, MAXEND))
5154 		return (rc);
5155 
5156 	/*
5157 	 * If the filesystem is mounted using local locking, pass the
5158 	 * request off to the local locking code.
5159 	 */
5160 	if (VTOMI(vp)->mi_flags & MI_LLOCK) {
5161 		if (cmd == F_SETLK || cmd == F_SETLKW) {
5162 			/*
5163 			 * For complete safety, we should be holding
5164 			 * r_lkserlock.  However, we can't call
5165 			 * lm_safelock and then fs_frlock while
5166 			 * holding r_lkserlock, so just invoke
5167 			 * lm_safelock and expect that this will
5168 			 * catch enough of the cases.
5169 			 */
5170 			if (!lm_safelock(vp, bfp, cr))
5171 				return (EAGAIN);
5172 		}
5173 		return (fs_frlock(vp, cmd, bfp, flag, offset, flk_cbp, cr));
5174 	}
5175 
5176 	rp = VTOR(vp);
5177 
5178 	/*
5179 	 * Check whether the given lock request can proceed, given the
5180 	 * current file mappings.
5181 	 */
5182 	if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_WRITER, intr))
5183 		return (EINTR);
5184 	if (cmd == F_SETLK || cmd == F_SETLKW) {
5185 		if (!lm_safelock(vp, bfp, cr)) {
5186 			rc = EAGAIN;
5187 			goto done;
5188 		}
5189 	}
5190 
5191 	/*
5192 	 * Flush the cache after waiting for async I/O to finish.  For new
5193 	 * locks, this is so that the process gets the latest bits from the
5194 	 * server.  For unlocks, this is so that other clients see the
5195 	 * latest bits once the file has been unlocked.  If currently dirty
5196 	 * pages can't be flushed, then don't allow a lock to be set.  But
5197 	 * allow unlocks to succeed, to avoid having orphan locks on the
5198 	 * server.
5199 	 */
5200 	if (cmd != F_GETLK) {
5201 		mutex_enter(&rp->r_statelock);
5202 		while (rp->r_count > 0) {
5203 		    if (intr) {
5204 			klwp_t *lwp = ttolwp(curthread);
5205 
5206 			if (lwp != NULL)
5207 				lwp->lwp_nostop++;
5208 			if (cv_wait_sig(&rp->r_cv, &rp->r_statelock) == 0) {
5209 				if (lwp != NULL)
5210 					lwp->lwp_nostop--;
5211 				rc = EINTR;
5212 				break;
5213 			}
5214 			if (lwp != NULL)
5215 				lwp->lwp_nostop--;
5216 		    } else
5217 			cv_wait(&rp->r_cv, &rp->r_statelock);
5218 		}
5219 		mutex_exit(&rp->r_statelock);
5220 		if (rc != 0)
5221 			goto done;
5222 		error = nfs3_putpage(vp, (offset_t)0, 0, B_INVAL, cr);
5223 		if (error) {
5224 			if (error == ENOSPC || error == EDQUOT) {
5225 				mutex_enter(&rp->r_statelock);
5226 				if (!rp->r_error)
5227 					rp->r_error = error;
5228 				mutex_exit(&rp->r_statelock);
5229 			}
5230 			if (bfp->l_type != F_UNLCK) {
5231 				rc = ENOLCK;
5232 				goto done;
5233 			}
5234 		}
5235 	}
5236 
5237 	lm_fh3.n_len = VTOFH3(vp)->fh3_length;
5238 	lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
5239 
5240 	/*
5241 	 * Call the lock manager to do the real work of contacting
5242 	 * the server and obtaining the lock.
5243 	 */
5244 	rc = lm4_frlock(vp, cmd, bfp, flag, offset, cr, &lm_fh3, flk_cbp);
5245 
5246 	if (rc == 0)
5247 		nfs_lockcompletion(vp, cmd);
5248 
5249 done:
5250 	nfs_rw_exit(&rp->r_lkserlock);
5251 	return (rc);
5252 }
5253 
5254 /*
5255  * Free storage space associated with the specified vnode.  The portion
5256  * to be freed is specified by bfp->l_start and bfp->l_len (already
5257  * normalized to a "whence" of 0).
5258  *
5259  * This is an experimental facility whose continued existence is not
5260  * guaranteed.  Currently, we only support the special case
5261  * of l_len == 0, meaning free to end of file.
5262  */
5263 /* ARGSUSED */
5264 static int
5265 nfs3_space(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5266 	offset_t offset, cred_t *cr, caller_context_t *ct)
5267 {
5268 	int error;
5269 
5270 	ASSERT(vp->v_type == VREG);
5271 	if (cmd != F_FREESP)
5272 		return (EINVAL);
5273 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
5274 		return (EIO);
5275 
5276 	error = convoff(vp, bfp, 0, offset);
5277 	if (!error) {
5278 		ASSERT(bfp->l_start >= 0);
5279 		if (bfp->l_len == 0) {
5280 			struct vattr va;
5281 
5282 			/*
5283 			 * ftruncate should not change the ctime and
5284 			 * mtime if we truncate the file to its
5285 			 * previous size.
5286 			 */
5287 			va.va_mask = AT_SIZE;
5288 			error = nfs3getattr(vp, &va, cr);
5289 			if (error || va.va_size == bfp->l_start)
5290 				return (error);
5291 			va.va_mask = AT_SIZE;
5292 			va.va_size = bfp->l_start;
5293 			error = nfs3setattr(vp, &va, 0, cr);
5294 		} else
5295 			error = EINVAL;
5296 	}
5297 
5298 	return (error);
5299 }
5300 
5301 /* ARGSUSED */
5302 static int
5303 nfs3_realvp(vnode_t *vp, vnode_t **vpp)
5304 {
5305 
5306 	return (EINVAL);
5307 }
5308 
5309 /*
5310  * Setup and add an address space callback to do the work of the delmap call.
5311  * The callback will (and must be) deleted in the actual callback function.
5312  *
5313  * This is done in order to take care of the problem that we have with holding
5314  * the address space's a_lock for a long period of time (e.g. if the NFS server
5315  * is down).  Callbacks will be executed in the address space code while the
5316  * a_lock is not held.	Holding the address space's a_lock causes things such
5317  * as ps and fork to hang because they are trying to acquire this lock as well.
5318  */
5319 /* ARGSUSED */
5320 static int
5321 nfs3_delmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5322 	size_t len, uint_t prot, uint_t maxprot, uint_t flags, cred_t *cr)
5323 {
5324 	int			caller_found;
5325 	int			error;
5326 	rnode_t			*rp;
5327 	nfs_delmap_args_t	*dmapp;
5328 	nfs_delmapcall_t	*delmap_call;
5329 
5330 	if (vp->v_flag & VNOMAP)
5331 		return (ENOSYS);
5332 	/*
5333 	 * A process may not change zones if it has NFS pages mmap'ed
5334 	 * in, so we can't legitimately get here from the wrong zone.
5335 	 */
5336 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
5337 
5338 	rp = VTOR(vp);
5339 
5340 	/*
5341 	 * The way that the address space of this process deletes its mapping
5342 	 * of this file is via the following call chains:
5343 	 * - as_free()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5344 	 * - as_unmap()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5345 	 *
5346 	 * With the use of address space callbacks we are allowed to drop the
5347 	 * address space lock, a_lock, while executing the NFS operations that
5348 	 * need to go over the wire.  Returning EAGAIN to the caller of this
5349 	 * function is what drives the execution of the callback that we add
5350 	 * below.  The callback will be executed by the address space code
5351 	 * after dropping the a_lock.  When the callback is finished, since
5352 	 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
5353 	 * is called again on the same segment to finish the rest of the work
5354 	 * that needs to happen during unmapping.
5355 	 *
5356 	 * This action of calling back into the segment driver causes
5357 	 * nfs3_delmap() to get called again, but since the callback was
5358 	 * already executed at this point, it already did the work and there
5359 	 * is nothing left for us to do.
5360 	 *
5361 	 * To Summarize:
5362 	 * - The first time nfs3_delmap is called by the current thread is when
5363 	 * we add the caller associated with this delmap to the delmap caller
5364 	 * list, add the callback, and return EAGAIN.
5365 	 * - The second time in this call chain when nfs3_delmap is called we
5366 	 * will find this caller in the delmap caller list and realize there
5367 	 * is no more work to do thus removing this caller from the list and
5368 	 * returning the error that was set in the callback execution.
5369 	 */
5370 	caller_found = nfs_find_and_delete_delmapcall(rp, &error);
5371 	if (caller_found) {
5372 		/*
5373 		 * 'error' is from the actual delmap operations.  To avoid
5374 		 * hangs, we need to handle the return of EAGAIN differently
5375 		 * since this is what drives the callback execution.
5376 		 * In this case, we don't want to return EAGAIN and do the
5377 		 * callback execution because there are none to execute.
5378 		 */
5379 		if (error == EAGAIN)
5380 			return (0);
5381 		else
5382 			return (error);
5383 	}
5384 
5385 	/* current caller was not in the list */
5386 	delmap_call = nfs_init_delmapcall();
5387 
5388 	mutex_enter(&rp->r_statelock);
5389 	list_insert_tail(&rp->r_indelmap, delmap_call);
5390 	mutex_exit(&rp->r_statelock);
5391 
5392 	dmapp = kmem_alloc(sizeof (nfs_delmap_args_t), KM_SLEEP);
5393 
5394 	dmapp->vp = vp;
5395 	dmapp->off = off;
5396 	dmapp->addr = addr;
5397 	dmapp->len = len;
5398 	dmapp->prot = prot;
5399 	dmapp->maxprot = maxprot;
5400 	dmapp->flags = flags;
5401 	dmapp->cr = cr;
5402 	dmapp->caller = delmap_call;
5403 
5404 	error = as_add_callback(as, nfs3_delmap_callback, dmapp,
5405 	AS_UNMAP_EVENT, addr, len, KM_SLEEP);
5406 
5407 	return (error ? error : EAGAIN);
5408 }
5409 
5410 /*
5411  * Remove some pages from an mmap'd vnode.  Just update the
5412  * count of pages.  If doing close-to-open, then flush and
5413  * commit all of the pages associated with this file.
5414  * Otherwise, start an asynchronous page flush to write out
5415  * any dirty pages.  This will also associate a credential
5416  * with the rnode which can be used to write the pages.
5417  */
5418 /* ARGSUSED */
5419 static void
5420 nfs3_delmap_callback(struct as *as, void *arg, uint_t event)
5421 {
5422 	int			error;
5423 	rnode_t			*rp;
5424 	mntinfo_t		*mi;
5425 	nfs_delmap_args_t	*dmapp = (nfs_delmap_args_t *)arg;
5426 
5427 	rp = VTOR(dmapp->vp);
5428 	mi = VTOMI(dmapp->vp);
5429 
5430 	atomic_add_long((ulong_t *)&rp->r_mapcnt, -btopr(dmapp->len));
5431 	ASSERT(rp->r_mapcnt >= 0);
5432 
5433 	/*
5434 	 * Initiate a page flush and potential commit if there are
5435 	 * pages, the file system was not mounted readonly, the segment
5436 	 * was mapped shared, and the pages themselves were writeable.
5437 	 */
5438 	if (vn_has_cached_data(dmapp->vp) && !vn_is_readonly(dmapp->vp) &&
5439 	    dmapp->flags == MAP_SHARED && (dmapp->maxprot & PROT_WRITE)) {
5440 		mutex_enter(&rp->r_statelock);
5441 		rp->r_flags |= RDIRTY;
5442 		mutex_exit(&rp->r_statelock);
5443 		/*
5444 		 * If this is a cross-zone access a sync putpage won't work, so
5445 		 * the best we can do is try an async putpage.  That seems
5446 		 * better than something more draconian such as discarding the
5447 		 * dirty pages.
5448 		 */
5449 		if ((mi->mi_flags & MI_NOCTO) ||
5450 		    curproc->p_zone != mi->mi_zone)
5451 			error = nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5452 			    B_ASYNC, dmapp->cr);
5453 		else
5454 			error = nfs3_putpage_commit(dmapp->vp, dmapp->off,
5455 			    dmapp->len, dmapp->cr);
5456 		if (!error) {
5457 			mutex_enter(&rp->r_statelock);
5458 			error = rp->r_error;
5459 			rp->r_error = 0;
5460 			mutex_exit(&rp->r_statelock);
5461 		}
5462 	} else
5463 		error = 0;
5464 
5465 	if ((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO))
5466 		(void) nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5467 		    B_INVAL, dmapp->cr);
5468 
5469 	dmapp->caller->error = error;
5470 	(void) as_delete_callback(as, arg);
5471 	kmem_free(dmapp, sizeof (nfs_delmap_args_t));
5472 }
5473 
5474 static int nfs3_pathconf_disable_cache = 0;
5475 
5476 #ifdef DEBUG
5477 static int nfs3_pathconf_cache_hits = 0;
5478 static int nfs3_pathconf_cache_misses = 0;
5479 #endif
5480 
5481 static int
5482 nfs3_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr)
5483 {
5484 	int error;
5485 	PATHCONF3args args;
5486 	PATHCONF3res res;
5487 	int douprintf;
5488 	failinfo_t fi;
5489 	rnode_t *rp;
5490 	hrtime_t t;
5491 
5492 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
5493 		return (EIO);
5494 	/*
5495 	 * Large file spec - need to base answer on info stored
5496 	 * on original FSINFO response.
5497 	 */
5498 	if (cmd == _PC_FILESIZEBITS) {
5499 		unsigned long long ll;
5500 		long l = 1;
5501 
5502 		ll = VTOMI(vp)->mi_maxfilesize;
5503 
5504 		if (ll == 0) {
5505 			*valp = 0;
5506 			return (0);
5507 		}
5508 
5509 		if (ll & 0xffffffff00000000) {
5510 			l += 32; ll >>= 32;
5511 		}
5512 		if (ll & 0xffff0000) {
5513 			l += 16; ll >>= 16;
5514 		}
5515 		if (ll & 0xff00) {
5516 			l += 8; ll >>= 8;
5517 		}
5518 		if (ll & 0xf0) {
5519 			l += 4; ll >>= 4;
5520 		}
5521 		if (ll & 0xc) {
5522 			l += 2; ll >>= 2;
5523 		}
5524 		if (ll & 0x2)
5525 			l += 2;
5526 		else if (ll & 0x1)
5527 			l += 1;
5528 		*valp = l;
5529 		return (0);
5530 	}
5531 
5532 	if (cmd == _PC_ACL_ENABLED) {
5533 		*valp = _ACL_ACLENT_ENABLED;
5534 		return (0);
5535 	}
5536 
5537 	if (cmd == _PC_XATTR_EXISTS) {
5538 		error = 0;
5539 		*valp = 0;
5540 		if (vp->v_vfsp->vfs_flag & VFS_XATTR) {
5541 			vnode_t *avp;
5542 			rnode_t *rp;
5543 			int error = 0;
5544 			mntinfo_t *mi = VTOMI(vp);
5545 
5546 			if (!(mi->mi_flags & MI_EXTATTR))
5547 				return (0);
5548 
5549 			rp = VTOR(vp);
5550 			if (nfs_rw_enter_sig(&rp->r_rwlock, RW_READER,
5551 			    INTR(vp)))
5552 				return (EINTR);
5553 
5554 			error = nfs3lookup_dnlc(vp, XATTR_DIR_NAME, &avp, cr);
5555 			if (error || avp == NULL)
5556 				error = acl_getxattrdir3(vp, &avp, 0, cr, 0);
5557 
5558 			nfs_rw_exit(&rp->r_rwlock);
5559 
5560 			if (error == 0 && avp != NULL) {
5561 				VN_RELE(avp);
5562 				*valp = 1;
5563 			} else if (error == ENOENT)
5564 				error = 0;
5565 		}
5566 		return (error);
5567 	}
5568 
5569 	rp = VTOR(vp);
5570 	if (rp->r_pathconf != NULL) {
5571 		mutex_enter(&rp->r_statelock);
5572 		if (rp->r_pathconf != NULL && nfs3_pathconf_disable_cache) {
5573 			kmem_free(rp->r_pathconf, sizeof (*rp->r_pathconf));
5574 			rp->r_pathconf = NULL;
5575 		}
5576 		if (rp->r_pathconf != NULL) {
5577 			error = 0;
5578 			switch (cmd) {
5579 			case _PC_LINK_MAX:
5580 				*valp = rp->r_pathconf->link_max;
5581 				break;
5582 			case _PC_NAME_MAX:
5583 				*valp = rp->r_pathconf->name_max;
5584 				break;
5585 			case _PC_PATH_MAX:
5586 			case _PC_SYMLINK_MAX:
5587 				*valp = MAXPATHLEN;
5588 				break;
5589 			case _PC_CHOWN_RESTRICTED:
5590 				*valp = rp->r_pathconf->chown_restricted;
5591 				break;
5592 			case _PC_NO_TRUNC:
5593 				*valp = rp->r_pathconf->no_trunc;
5594 				break;
5595 			default:
5596 				error = EINVAL;
5597 				break;
5598 			}
5599 			mutex_exit(&rp->r_statelock);
5600 #ifdef DEBUG
5601 			nfs3_pathconf_cache_hits++;
5602 #endif
5603 			return (error);
5604 		}
5605 		mutex_exit(&rp->r_statelock);
5606 	}
5607 #ifdef DEBUG
5608 	nfs3_pathconf_cache_misses++;
5609 #endif
5610 
5611 	args.object = *VTOFH3(vp);
5612 	fi.vp = vp;
5613 	fi.fhp = (caddr_t)&args.object;
5614 	fi.copyproc = nfs3copyfh;
5615 	fi.lookupproc = nfs3lookup;
5616 	fi.xattrdirproc = acl_getxattrdir3;
5617 
5618 	douprintf = 1;
5619 
5620 	t = gethrtime();
5621 
5622 	error = rfs3call(VTOMI(vp), NFSPROC3_PATHCONF,
5623 	    xdr_nfs_fh3, (caddr_t)&args,
5624 	    xdr_PATHCONF3res, (caddr_t)&res, cr,
5625 	    &douprintf, &res.status, 0, &fi);
5626 
5627 	if (error)
5628 		return (error);
5629 
5630 	error = geterrno3(res.status);
5631 
5632 	if (!error) {
5633 		nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
5634 		if (!nfs3_pathconf_disable_cache) {
5635 			mutex_enter(&rp->r_statelock);
5636 			if (rp->r_pathconf == NULL) {
5637 				rp->r_pathconf = kmem_alloc(
5638 				    sizeof (*rp->r_pathconf), KM_NOSLEEP);
5639 				if (rp->r_pathconf != NULL)
5640 					*rp->r_pathconf = res.resok.info;
5641 			}
5642 			mutex_exit(&rp->r_statelock);
5643 		}
5644 		switch (cmd) {
5645 		case _PC_LINK_MAX:
5646 			*valp = res.resok.info.link_max;
5647 			break;
5648 		case _PC_NAME_MAX:
5649 			*valp = res.resok.info.name_max;
5650 			break;
5651 		case _PC_PATH_MAX:
5652 		case _PC_SYMLINK_MAX:
5653 			*valp = MAXPATHLEN;
5654 			break;
5655 		case _PC_CHOWN_RESTRICTED:
5656 			*valp = res.resok.info.chown_restricted;
5657 			break;
5658 		case _PC_NO_TRUNC:
5659 			*valp = res.resok.info.no_trunc;
5660 			break;
5661 		default:
5662 			return (EINVAL);
5663 		}
5664 	} else {
5665 		nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
5666 		PURGE_STALE_FH(error, vp, cr);
5667 	}
5668 
5669 	return (error);
5670 }
5671 
5672 /*
5673  * Called by async thread to do synchronous pageio. Do the i/o, wait
5674  * for it to complete, and cleanup the page list when done.
5675  */
5676 static int
5677 nfs3_sync_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5678 	int flags, cred_t *cr)
5679 {
5680 	int error;
5681 
5682 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
5683 	error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5684 	if (flags & B_READ)
5685 		pvn_read_done(pp, (error ? B_ERROR : 0) | flags);
5686 	else
5687 		pvn_write_done(pp, (error ? B_ERROR : 0) | flags);
5688 	return (error);
5689 }
5690 
5691 static int
5692 nfs3_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5693 	int flags, cred_t *cr)
5694 {
5695 	int error;
5696 	rnode_t *rp;
5697 
5698 	if (pp == NULL)
5699 		return (EINVAL);
5700 	if (!(flags & B_ASYNC) && curproc->p_zone != VTOMI(vp)->mi_zone)
5701 		return (EIO);
5702 
5703 	rp = VTOR(vp);
5704 	mutex_enter(&rp->r_statelock);
5705 	rp->r_count++;
5706 	mutex_exit(&rp->r_statelock);
5707 
5708 	if (flags & B_ASYNC) {
5709 		error = nfs_async_pageio(vp, pp, io_off, io_len, flags, cr,
5710 		    nfs3_sync_pageio);
5711 	} else
5712 		error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5713 	mutex_enter(&rp->r_statelock);
5714 	rp->r_count--;
5715 	cv_broadcast(&rp->r_cv);
5716 	mutex_exit(&rp->r_statelock);
5717 	return (error);
5718 }
5719 
5720 static void
5721 nfs3_dispose(vnode_t *vp, page_t *pp, int fl, int dn, cred_t *cr)
5722 {
5723 	int error;
5724 	rnode_t *rp;
5725 	page_t *plist;
5726 	page_t *pptr;
5727 	offset3 offset;
5728 	count3 len;
5729 	k_sigset_t smask;
5730 
5731 	/*
5732 	 * We should get called with fl equal to either B_FREE or
5733 	 * B_INVAL.  Any other value is illegal.
5734 	 *
5735 	 * The page that we are either supposed to free or destroy
5736 	 * should be exclusive locked and its io lock should not
5737 	 * be held.
5738 	 */
5739 	ASSERT(fl == B_FREE || fl == B_INVAL);
5740 	ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr);
5741 	rp = VTOR(vp);
5742 
5743 	/*
5744 	 * If the page doesn't need to be committed or we shouldn't
5745 	 * even bother attempting to commit it, then just make sure
5746 	 * that the p_fsdata byte is clear and then either free or
5747 	 * destroy the page as appropriate.
5748 	 */
5749 	if (pp->p_fsdata == C_NOCOMMIT || (rp->r_flags & RSTALE)) {
5750 		pp->p_fsdata = C_NOCOMMIT;
5751 		if (fl == B_FREE)
5752 			page_free(pp, dn);
5753 		else
5754 			page_destroy(pp, dn);
5755 		return;
5756 	}
5757 
5758 	/*
5759 	 * If there is a page invalidation operation going on, then
5760 	 * if this is one of the pages being destroyed, then just
5761 	 * clear the p_fsdata byte and then either free or destroy
5762 	 * the page as appropriate.
5763 	 */
5764 	mutex_enter(&rp->r_statelock);
5765 	if ((rp->r_flags & RTRUNCATE) && pp->p_offset >= rp->r_truncaddr) {
5766 		mutex_exit(&rp->r_statelock);
5767 		pp->p_fsdata = C_NOCOMMIT;
5768 		if (fl == B_FREE)
5769 			page_free(pp, dn);
5770 		else
5771 			page_destroy(pp, dn);
5772 		return;
5773 	}
5774 
5775 	/*
5776 	 * If we are freeing this page and someone else is already
5777 	 * waiting to do a commit, then just unlock the page and
5778 	 * return.  That other thread will take care of commiting
5779 	 * this page.  The page can be freed sometime after the
5780 	 * commit has finished.  Otherwise, if the page is marked
5781 	 * as delay commit, then we may be getting called from
5782 	 * pvn_write_done, one page at a time.   This could result
5783 	 * in one commit per page, so we end up doing lots of small
5784 	 * commits instead of fewer larger commits.  This is bad,
5785 	 * we want do as few commits as possible.
5786 	 */
5787 	if (fl == B_FREE) {
5788 		if (rp->r_flags & RCOMMITWAIT) {
5789 			page_unlock(pp);
5790 			mutex_exit(&rp->r_statelock);
5791 			return;
5792 		}
5793 		if (pp->p_fsdata == C_DELAYCOMMIT) {
5794 			pp->p_fsdata = C_COMMIT;
5795 			page_unlock(pp);
5796 			mutex_exit(&rp->r_statelock);
5797 			return;
5798 		}
5799 	}
5800 
5801 	/*
5802 	 * Check to see if there is a signal which would prevent an
5803 	 * attempt to commit the pages from being successful.  If so,
5804 	 * then don't bother with all of the work to gather pages and
5805 	 * generate the unsuccessful RPC.  Just return from here and
5806 	 * let the page be committed at some later time.
5807 	 */
5808 	sigintr(&smask, VTOMI(vp)->mi_flags & MI_INT);
5809 	if (ttolwp(curthread) != NULL && ISSIG(curthread, JUSTLOOKING)) {
5810 		sigunintr(&smask);
5811 		page_unlock(pp);
5812 		mutex_exit(&rp->r_statelock);
5813 		return;
5814 	}
5815 	sigunintr(&smask);
5816 
5817 	/*
5818 	 * We are starting to need to commit pages, so let's try
5819 	 * to commit as many as possible at once to reduce the
5820 	 * overhead.
5821 	 *
5822 	 * Set the `commit inprogress' state bit.  We must
5823 	 * first wait until any current one finishes.  Then
5824 	 * we initialize the c_pages list with this page.
5825 	 */
5826 	while (rp->r_flags & RCOMMIT) {
5827 		rp->r_flags |= RCOMMITWAIT;
5828 		cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
5829 		rp->r_flags &= ~RCOMMITWAIT;
5830 	}
5831 	rp->r_flags |= RCOMMIT;
5832 	mutex_exit(&rp->r_statelock);
5833 	ASSERT(rp->r_commit.c_pages == NULL);
5834 	rp->r_commit.c_pages = pp;
5835 	rp->r_commit.c_commbase = (offset3)pp->p_offset;
5836 	rp->r_commit.c_commlen = PAGESIZE;
5837 
5838 	/*
5839 	 * Gather together all other pages which can be committed.
5840 	 * They will all be chained off r_commit.c_pages.
5841 	 */
5842 	nfs3_get_commit(vp);
5843 
5844 	/*
5845 	 * Clear the `commit inprogress' status and disconnect
5846 	 * the list of pages to be committed from the rnode.
5847 	 * At this same time, we also save the starting offset
5848 	 * and length of data to be committed on the server.
5849 	 */
5850 	plist = rp->r_commit.c_pages;
5851 	rp->r_commit.c_pages = NULL;
5852 	offset = rp->r_commit.c_commbase;
5853 	len = rp->r_commit.c_commlen;
5854 	mutex_enter(&rp->r_statelock);
5855 	rp->r_flags &= ~RCOMMIT;
5856 	cv_broadcast(&rp->r_commit.c_cv);
5857 	mutex_exit(&rp->r_statelock);
5858 
5859 	if (curproc == proc_pageout || curproc == proc_fsflush ||
5860 	    curproc->p_zone != VTOMI(vp)->mi_zone) {
5861 		nfs_async_commit(vp, plist, offset, len, cr, nfs3_async_commit);
5862 		return;
5863 	}
5864 
5865 	/*
5866 	 * Actually generate the COMMIT3 over the wire operation.
5867 	 */
5868 	error = nfs3_commit(vp, offset, len, cr);
5869 
5870 	/*
5871 	 * If we got an error during the commit, just unlock all
5872 	 * of the pages.  The pages will get retransmitted to the
5873 	 * server during a putpage operation.
5874 	 */
5875 	if (error) {
5876 		while (plist != NULL) {
5877 			pptr = plist;
5878 			page_sub(&plist, pptr);
5879 			page_unlock(pptr);
5880 		}
5881 		return;
5882 	}
5883 
5884 	/*
5885 	 * We've tried as hard as we can to commit the data to stable
5886 	 * storage on the server.  We release the rest of the pages
5887 	 * and clear the commit required state.  They will be put
5888 	 * onto the tail of the cachelist if they are nolonger
5889 	 * mapped.
5890 	 */
5891 	while (plist != pp) {
5892 		pptr = plist;
5893 		page_sub(&plist, pptr);
5894 		pptr->p_fsdata = C_NOCOMMIT;
5895 		(void) page_release(pptr, 1);
5896 	}
5897 
5898 	/*
5899 	 * It is possible that nfs3_commit didn't return error but
5900 	 * some other thread has modified the page we are going
5901 	 * to free/destroy.
5902 	 *    In this case we need to rewrite the page. Do an explicit check
5903 	 * before attempting to free/destroy the page. If modified, needs to
5904 	 * be rewritten so unlock the page and return.
5905 	 */
5906 	if (hat_ismod(pp)) {
5907 		pp->p_fsdata = C_NOCOMMIT;
5908 		page_unlock(pp);
5909 		return;
5910 	}
5911 
5912 	/*
5913 	 * Now, as appropriate, either free or destroy the page
5914 	 * that we were called with.
5915 	 */
5916 	pp->p_fsdata = C_NOCOMMIT;
5917 	if (fl == B_FREE)
5918 		page_free(pp, dn);
5919 	else
5920 		page_destroy(pp, dn);
5921 }
5922 
5923 static int
5924 nfs3_commit(vnode_t *vp, offset3 offset, count3 count, cred_t *cr)
5925 {
5926 	int error;
5927 	rnode_t *rp;
5928 	COMMIT3args args;
5929 	COMMIT3res res;
5930 	int douprintf;
5931 	cred_t *cred;
5932 
5933 	rp = VTOR(vp);
5934 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
5935 
5936 	mutex_enter(&rp->r_statelock);
5937 	if (rp->r_cred != NULL) {
5938 		cred = rp->r_cred;
5939 		crhold(cred);
5940 	} else {
5941 		rp->r_cred = cr;
5942 		crhold(cr);
5943 		cred = cr;
5944 		crhold(cred);
5945 	}
5946 	mutex_exit(&rp->r_statelock);
5947 
5948 	args.file = *VTOFH3(vp);
5949 	args.offset = offset;
5950 	args.count = count;
5951 
5952 doitagain:
5953 	douprintf = 1;
5954 	error = rfs3call(VTOMI(vp), NFSPROC3_COMMIT,
5955 	    xdr_COMMIT3args, (caddr_t)&args,
5956 	    xdr_COMMIT3res, (caddr_t)&res, cred,
5957 	    &douprintf, &res.status, 0, NULL);
5958 
5959 	crfree(cred);
5960 
5961 	if (error)
5962 		return (error);
5963 
5964 	error = geterrno3(res.status);
5965 	if (!error) {
5966 		ASSERT(rp->r_flags & RHAVEVERF);
5967 		mutex_enter(&rp->r_statelock);
5968 		if (rp->r_verf == res.resok.verf) {
5969 			mutex_exit(&rp->r_statelock);
5970 			return (0);
5971 		}
5972 		nfs3_set_mod(vp);
5973 		rp->r_verf = res.resok.verf;
5974 		mutex_exit(&rp->r_statelock);
5975 		error = NFS_VERF_MISMATCH;
5976 	} else {
5977 		if (error == EACCES) {
5978 			mutex_enter(&rp->r_statelock);
5979 			if (cred != cr) {
5980 				if (rp->r_cred != NULL)
5981 					crfree(rp->r_cred);
5982 				rp->r_cred = cr;
5983 				crhold(cr);
5984 				cred = cr;
5985 				crhold(cred);
5986 				mutex_exit(&rp->r_statelock);
5987 				goto doitagain;
5988 			}
5989 			mutex_exit(&rp->r_statelock);
5990 		}
5991 		/*
5992 		 * Can't do a PURGE_STALE_FH here because this
5993 		 * can cause a deadlock.  nfs3_commit can
5994 		 * be called from nfs3_dispose which can be called
5995 		 * indirectly via pvn_vplist_dirty.  PURGE_STALE_FH
5996 		 * can call back to pvn_vplist_dirty.
5997 		 */
5998 		if (error == ESTALE) {
5999 			mutex_enter(&rp->r_statelock);
6000 			rp->r_flags |= RSTALE;
6001 			if (!rp->r_error)
6002 				rp->r_error = error;
6003 			mutex_exit(&rp->r_statelock);
6004 			PURGE_ATTRCACHE(vp);
6005 		} else {
6006 			mutex_enter(&rp->r_statelock);
6007 			if (!rp->r_error)
6008 				rp->r_error = error;
6009 			mutex_exit(&rp->r_statelock);
6010 		}
6011 	}
6012 
6013 	return (error);
6014 }
6015 
6016 static void
6017 nfs3_set_mod(vnode_t *vp)
6018 {
6019 	page_t *pp;
6020 	kmutex_t *vphm;
6021 
6022 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
6023 	vphm = page_vnode_mutex(vp);
6024 	mutex_enter(vphm);
6025 	if ((pp = vp->v_pages) != NULL) {
6026 		do {
6027 			if (pp->p_fsdata != C_NOCOMMIT) {
6028 				hat_setmod(pp);
6029 				pp->p_fsdata = C_NOCOMMIT;
6030 			}
6031 		} while ((pp = pp->p_vpnext) != vp->v_pages);
6032 	}
6033 	mutex_exit(vphm);
6034 }
6035 
6036 
6037 /*
6038  * This routine is used to gather together a page list of the pages
6039  * which are to be committed on the server.  This routine must not
6040  * be called if the calling thread holds any locked pages.
6041  *
6042  * The calling thread must have set RCOMMIT.  This bit is used to
6043  * serialize access to the commit structure in the rnode.  As long
6044  * as the thread has set RCOMMIT, then it can manipulate the commit
6045  * structure without requiring any other locks.
6046  */
6047 static void
6048 nfs3_get_commit(vnode_t *vp)
6049 {
6050 	rnode_t *rp;
6051 	page_t *pp;
6052 	kmutex_t *vphm;
6053 
6054 	rp = VTOR(vp);
6055 
6056 	ASSERT(rp->r_flags & RCOMMIT);
6057 
6058 	vphm = page_vnode_mutex(vp);
6059 	mutex_enter(vphm);
6060 
6061 	/*
6062 	 * If there are no pages associated with this vnode, then
6063 	 * just return.
6064 	 */
6065 	if ((pp = vp->v_pages) == NULL) {
6066 		mutex_exit(vphm);
6067 		return;
6068 	}
6069 
6070 	/*
6071 	 * Step through all of the pages associated with this vnode
6072 	 * looking for pages which need to be committed.
6073 	 */
6074 	do {
6075 		/*
6076 		 * If this page does not need to be committed or is
6077 		 * modified, then just skip it.
6078 		 */
6079 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp))
6080 			continue;
6081 
6082 		/*
6083 		 * Attempt to lock the page.  If we can't, then
6084 		 * someone else is messing with it and we will
6085 		 * just skip it.
6086 		 */
6087 		if (!page_trylock(pp, SE_EXCL))
6088 			continue;
6089 
6090 		/*
6091 		 * If this page does not need to be committed or is
6092 		 * modified, then just skip it.  Recheck now that
6093 		 * the page is locked.
6094 		 */
6095 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6096 			page_unlock(pp);
6097 			continue;
6098 		}
6099 
6100 		if (PP_ISFREE(pp)) {
6101 			cmn_err(CE_PANIC, "nfs3_get_commit: %p is free",
6102 			    (void *)pp);
6103 		}
6104 
6105 		/*
6106 		 * The page needs to be committed and we locked it.
6107 		 * Update the base and length parameters and add it
6108 		 * to r_pages.
6109 		 */
6110 		if (rp->r_commit.c_pages == NULL) {
6111 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6112 			rp->r_commit.c_commlen = PAGESIZE;
6113 		} else if (pp->p_offset < rp->r_commit.c_commbase) {
6114 			rp->r_commit.c_commlen = rp->r_commit.c_commbase -
6115 			    (offset3)pp->p_offset + rp->r_commit.c_commlen;
6116 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6117 		} else if ((rp->r_commit.c_commbase + rp->r_commit.c_commlen)
6118 			    <= pp->p_offset) {
6119 			rp->r_commit.c_commlen = (offset3)pp->p_offset -
6120 			    rp->r_commit.c_commbase + PAGESIZE;
6121 		}
6122 		page_add(&rp->r_commit.c_pages, pp);
6123 	} while ((pp = pp->p_vpnext) != vp->v_pages);
6124 
6125 	mutex_exit(vphm);
6126 }
6127 
6128 /*
6129  * This routine is used to gather together a page list of the pages
6130  * which are to be committed on the server.  This routine must not
6131  * be called if the calling thread holds any locked pages.
6132  *
6133  * The calling thread must have set RCOMMIT.  This bit is used to
6134  * serialize access to the commit structure in the rnode.  As long
6135  * as the thread has set RCOMMIT, then it can manipulate the commit
6136  * structure without requiring any other locks.
6137  */
6138 static void
6139 nfs3_get_commit_range(vnode_t *vp, u_offset_t soff, size_t len)
6140 {
6141 
6142 	rnode_t *rp;
6143 	page_t *pp;
6144 	u_offset_t end;
6145 	u_offset_t off;
6146 
6147 	ASSERT(len != 0);
6148 
6149 	rp = VTOR(vp);
6150 
6151 	ASSERT(rp->r_flags & RCOMMIT);
6152 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
6153 
6154 	/*
6155 	 * If there are no pages associated with this vnode, then
6156 	 * just return.
6157 	 */
6158 	if ((pp = vp->v_pages) == NULL)
6159 		return;
6160 
6161 	/*
6162 	 * Calculate the ending offset.
6163 	 */
6164 	end = soff + len;
6165 
6166 	for (off = soff; off < end; off += PAGESIZE) {
6167 		/*
6168 		 * Lookup each page by vp, offset.
6169 		 */
6170 		if ((pp = page_lookup_nowait(vp, off, SE_EXCL)) == NULL)
6171 			continue;
6172 
6173 		/*
6174 		 * If this page does not need to be committed or is
6175 		 * modified, then just skip it.
6176 		 */
6177 		if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6178 			page_unlock(pp);
6179 			continue;
6180 		}
6181 
6182 		ASSERT(PP_ISFREE(pp) == 0);
6183 
6184 		/*
6185 		 * The page needs to be committed and we locked it.
6186 		 * Update the base and length parameters and add it
6187 		 * to r_pages.
6188 		 */
6189 		if (rp->r_commit.c_pages == NULL) {
6190 			rp->r_commit.c_commbase = (offset3)pp->p_offset;
6191 			rp->r_commit.c_commlen = PAGESIZE;
6192 		} else {
6193 			rp->r_commit.c_commlen = (offset3)pp->p_offset -
6194 					rp->r_commit.c_commbase + PAGESIZE;
6195 		}
6196 		page_add(&rp->r_commit.c_pages, pp);
6197 	}
6198 }
6199 
6200 #if 0	/* unused */
6201 #ifdef DEBUG
6202 static int
6203 nfs3_no_uncommitted_pages(vnode_t *vp)
6204 {
6205 	page_t *pp;
6206 	kmutex_t *vphm;
6207 
6208 	vphm = page_vnode_mutex(vp);
6209 	mutex_enter(vphm);
6210 	if ((pp = vp->v_pages) != NULL) {
6211 		do {
6212 			if (pp->p_fsdata != C_NOCOMMIT) {
6213 				mutex_exit(vphm);
6214 				return (0);
6215 			}
6216 		} while ((pp = pp->p_vpnext) != vp->v_pages);
6217 	}
6218 	mutex_exit(vphm);
6219 
6220 	return (1);
6221 }
6222 #endif
6223 #endif
6224 
6225 static int
6226 nfs3_putpage_commit(vnode_t *vp, offset_t poff, size_t plen, cred_t *cr)
6227 {
6228 	int error;
6229 	writeverf3 write_verf;
6230 	rnode_t *rp = VTOR(vp);
6231 
6232 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
6233 	/*
6234 	 * Flush the data portion of the file and then commit any
6235 	 * portions which need to be committed.  This may need to
6236 	 * be done twice if the server has changed state since
6237 	 * data was last written.  The data will need to be
6238 	 * rewritten to the server and then a new commit done.
6239 	 *
6240 	 * In fact, this may need to be done several times if the
6241 	 * server is having problems and crashing while we are
6242 	 * attempting to do this.
6243 	 */
6244 
6245 top:
6246 	/*
6247 	 * Do a flush based on the poff and plen arguments.  This
6248 	 * will asynchronously write out any modified pages in the
6249 	 * range specified by (poff, plen).  This starts all of the
6250 	 * i/o operations which will be waited for in the next
6251 	 * call to nfs3_putpage
6252 	 */
6253 
6254 	mutex_enter(&rp->r_statelock);
6255 	write_verf = rp->r_verf;
6256 	mutex_exit(&rp->r_statelock);
6257 
6258 	error = nfs3_putpage(vp, poff, plen, B_ASYNC, cr);
6259 	if (error == EAGAIN)
6260 		error = 0;
6261 
6262 	/*
6263 	 * Do a flush based on the poff and plen arguments.  This
6264 	 * will synchronously write out any modified pages in the
6265 	 * range specified by (poff, plen) and wait until all of
6266 	 * the asynchronous i/o's in that range are done as well.
6267 	 */
6268 	if (!error)
6269 		error = nfs3_putpage(vp, poff, plen, 0, cr);
6270 
6271 	if (error)
6272 		return (error);
6273 
6274 	mutex_enter(&rp->r_statelock);
6275 	if (rp->r_verf != write_verf) {
6276 		mutex_exit(&rp->r_statelock);
6277 		goto top;
6278 	}
6279 	mutex_exit(&rp->r_statelock);
6280 
6281 	/*
6282 	 * Now commit any pages which might need to be committed.
6283 	 * If the error, NFS_VERF_MISMATCH, is returned, then
6284 	 * start over with the flush operation.
6285 	 */
6286 
6287 	error = nfs3_commit_vp(vp, poff, plen, cr);
6288 
6289 	if (error == NFS_VERF_MISMATCH)
6290 		goto top;
6291 
6292 	return (error);
6293 }
6294 
6295 static int
6296 nfs3_commit_vp(vnode_t *vp, u_offset_t poff, size_t plen, cred_t *cr)
6297 {
6298 	rnode_t *rp;
6299 	page_t *plist;
6300 	offset3 offset;
6301 	count3 len;
6302 
6303 
6304 	rp = VTOR(vp);
6305 
6306 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
6307 		return (EIO);
6308 	/*
6309 	 * Set the `commit inprogress' state bit.  We must
6310 	 * first wait until any current one finishes.
6311 	 */
6312 	mutex_enter(&rp->r_statelock);
6313 	while (rp->r_flags & RCOMMIT) {
6314 		rp->r_flags |= RCOMMITWAIT;
6315 		cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
6316 		rp->r_flags &= ~RCOMMITWAIT;
6317 	}
6318 	rp->r_flags |= RCOMMIT;
6319 	mutex_exit(&rp->r_statelock);
6320 
6321 	/*
6322 	 * Gather together all of the pages which need to be
6323 	 * committed.
6324 	 */
6325 	if (plen == 0)
6326 		nfs3_get_commit(vp);
6327 	else
6328 		nfs3_get_commit_range(vp, poff, plen);
6329 
6330 	/*
6331 	 * Clear the `commit inprogress' bit and disconnect the
6332 	 * page list which was gathered together in nfs3_get_commit.
6333 	 */
6334 	plist = rp->r_commit.c_pages;
6335 	rp->r_commit.c_pages = NULL;
6336 	offset = rp->r_commit.c_commbase;
6337 	len = rp->r_commit.c_commlen;
6338 	mutex_enter(&rp->r_statelock);
6339 	rp->r_flags &= ~RCOMMIT;
6340 	cv_broadcast(&rp->r_commit.c_cv);
6341 	mutex_exit(&rp->r_statelock);
6342 
6343 	/*
6344 	 * If any pages need to be committed, commit them and
6345 	 * then unlock them so that they can be freed some
6346 	 * time later.
6347 	 */
6348 	if (plist != NULL) {
6349 		/*
6350 		 * No error occurred during the flush portion
6351 		 * of this operation, so now attempt to commit
6352 		 * the data to stable storage on the server.
6353 		 *
6354 		 * This will unlock all of the pages on the list.
6355 		 */
6356 		return (nfs3_sync_commit(vp, plist, offset, len, cr));
6357 	}
6358 	return (0);
6359 }
6360 
6361 static int
6362 nfs3_sync_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6363 	cred_t *cr)
6364 {
6365 	int error;
6366 	page_t *pp;
6367 
6368 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
6369 	error = nfs3_commit(vp, offset, count, cr);
6370 
6371 	/*
6372 	 * If we got an error, then just unlock all of the pages
6373 	 * on the list.
6374 	 */
6375 	if (error) {
6376 		while (plist != NULL) {
6377 			pp = plist;
6378 			page_sub(&plist, pp);
6379 			page_unlock(pp);
6380 		}
6381 		return (error);
6382 	}
6383 	/*
6384 	 * We've tried as hard as we can to commit the data to stable
6385 	 * storage on the server.  We just unlock the pages and clear
6386 	 * the commit required state.  They will get freed later.
6387 	 */
6388 	while (plist != NULL) {
6389 		pp = plist;
6390 		page_sub(&plist, pp);
6391 		pp->p_fsdata = C_NOCOMMIT;
6392 		page_unlock(pp);
6393 	}
6394 
6395 	return (error);
6396 }
6397 
6398 static void
6399 nfs3_async_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6400 	cred_t *cr)
6401 {
6402 	ASSERT(curproc->p_zone == VTOMI(vp)->mi_zone);
6403 	(void) nfs3_sync_commit(vp, plist, offset, count, cr);
6404 }
6405 
6406 static int
6407 nfs3_setsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
6408 {
6409 	int error;
6410 	mntinfo_t *mi;
6411 
6412 	mi = VTOMI(vp);
6413 
6414 	if (curproc->p_zone != mi->mi_zone)
6415 		return (EIO);
6416 
6417 	if (mi->mi_flags & MI_ACL) {
6418 		error = acl_setacl3(vp, vsecattr, flag, cr);
6419 		if (mi->mi_flags & MI_ACL)
6420 			return (error);
6421 	}
6422 
6423 	return (ENOSYS);
6424 }
6425 
6426 static int
6427 nfs3_getsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr)
6428 {
6429 	int error;
6430 	mntinfo_t *mi;
6431 
6432 	mi = VTOMI(vp);
6433 
6434 	if (curproc->p_zone != mi->mi_zone)
6435 		return (EIO);
6436 
6437 	if (mi->mi_flags & MI_ACL) {
6438 		error = acl_getacl3(vp, vsecattr, flag, cr);
6439 		if (mi->mi_flags & MI_ACL)
6440 			return (error);
6441 	}
6442 
6443 	return (fs_fab_acl(vp, vsecattr, flag, cr));
6444 }
6445 
6446 static int
6447 nfs3_shrlock(vnode_t *vp, int cmd, struct shrlock *shr, int flag, cred_t *cr)
6448 {
6449 	int error;
6450 	struct shrlock nshr;
6451 	struct nfs_owner nfs_owner;
6452 	netobj lm_fh3;
6453 
6454 	if (curproc->p_zone != VTOMI(vp)->mi_zone)
6455 		return (EIO);
6456 
6457 	/*
6458 	 * check for valid cmd parameter
6459 	 */
6460 	if (cmd != F_SHARE && cmd != F_UNSHARE && cmd != F_HASREMOTELOCKS)
6461 		return (EINVAL);
6462 
6463 	/*
6464 	 * Check access permissions
6465 	 */
6466 	if (cmd == F_SHARE &&
6467 	    (((shr->s_access & F_RDACC) && !(flag & FREAD)) ||
6468 	    ((shr->s_access & F_WRACC) && !(flag & FWRITE))))
6469 		return (EBADF);
6470 
6471 	/*
6472 	 * If the filesystem is mounted using local locking, pass the
6473 	 * request off to the local share code.
6474 	 */
6475 	if (VTOMI(vp)->mi_flags & MI_LLOCK)
6476 		return (fs_shrlock(vp, cmd, shr, flag, cr));
6477 
6478 	switch (cmd) {
6479 	case F_SHARE:
6480 	case F_UNSHARE:
6481 		lm_fh3.n_len = VTOFH3(vp)->fh3_length;
6482 		lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
6483 
6484 		/*
6485 		 * If passed an owner that is too large to fit in an
6486 		 * nfs_owner it is likely a recursive call from the
6487 		 * lock manager client and pass it straight through.  If
6488 		 * it is not a nfs_owner then simply return an error.
6489 		 */
6490 		if (shr->s_own_len > sizeof (nfs_owner.lowner)) {
6491 			if (((struct nfs_owner *)shr->s_owner)->magic !=
6492 			    NFS_OWNER_MAGIC)
6493 				return (EINVAL);
6494 
6495 			if (error = lm4_shrlock(vp, cmd, shr, flag, &lm_fh3)) {
6496 				error = set_errno(error);
6497 			}
6498 			return (error);
6499 		}
6500 		/*
6501 		 * Remote share reservations owner is a combination of
6502 		 * a magic number, hostname, and the local owner
6503 		 */
6504 		bzero(&nfs_owner, sizeof (nfs_owner));
6505 		nfs_owner.magic = NFS_OWNER_MAGIC;
6506 		(void) strncpy(nfs_owner.hname, uts_nodename(),
6507 		    sizeof (nfs_owner.hname));
6508 		bcopy(shr->s_owner, nfs_owner.lowner, shr->s_own_len);
6509 		nshr.s_access = shr->s_access;
6510 		nshr.s_deny = shr->s_deny;
6511 		nshr.s_sysid = 0;
6512 		nshr.s_pid = ttoproc(curthread)->p_pid;
6513 		nshr.s_own_len = sizeof (nfs_owner);
6514 		nshr.s_owner = (caddr_t)&nfs_owner;
6515 
6516 		if (error = lm4_shrlock(vp, cmd, &nshr, flag, &lm_fh3)) {
6517 			error = set_errno(error);
6518 		}
6519 
6520 		break;
6521 
6522 	case F_HASREMOTELOCKS:
6523 		/*
6524 		 * NFS client can't store remote locks itself
6525 		 */
6526 		shr->s_access = 0;
6527 		error = 0;
6528 		break;
6529 
6530 	default:
6531 		error = EINVAL;
6532 		break;
6533 	}
6534 
6535 	return (error);
6536 }
6537