Print this page
3484 enhance and document tail follow support
Reviewed by: Joshua M. Clulow <jmc@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/nfs/nfs3_vnops.c
+++ new/usr/src/uts/common/fs/nfs/nfs3_vnops.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
|
↓ open down ↓ |
20 lines elided |
↑ open up ↑ |
21 21 /*
22 22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 25
26 26 /*
27 27 * Copyright (c) 1983,1984,1985,1986,1987,1988,1989 AT&T.
28 28 * All rights reserved.
29 29 */
30 30
31 +/*
32 + * Copyright (c) 2013, Joyent, Inc. All rights reserved.
33 + */
34 +
31 35 #include <sys/param.h>
32 36 #include <sys/types.h>
33 37 #include <sys/systm.h>
34 38 #include <sys/cred.h>
35 39 #include <sys/time.h>
36 40 #include <sys/vnode.h>
37 41 #include <sys/vfs.h>
38 42 #include <sys/vfs_opreg.h>
39 43 #include <sys/file.h>
40 44 #include <sys/filio.h>
41 45 #include <sys/uio.h>
42 46 #include <sys/buf.h>
43 47 #include <sys/mman.h>
44 48 #include <sys/pathname.h>
45 49 #include <sys/dirent.h>
46 50 #include <sys/debug.h>
47 51 #include <sys/vmsystm.h>
48 52 #include <sys/fcntl.h>
49 53 #include <sys/flock.h>
50 54 #include <sys/swap.h>
51 55 #include <sys/errno.h>
52 56 #include <sys/strsubr.h>
53 57 #include <sys/sysmacros.h>
54 58 #include <sys/kmem.h>
55 59 #include <sys/cmn_err.h>
56 60 #include <sys/pathconf.h>
57 61 #include <sys/utsname.h>
58 62 #include <sys/dnlc.h>
59 63 #include <sys/acl.h>
60 64 #include <sys/systeminfo.h>
61 65 #include <sys/atomic.h>
62 66 #include <sys/policy.h>
63 67 #include <sys/sdt.h>
64 68 #include <sys/zone.h>
65 69
66 70 #include <rpc/types.h>
67 71 #include <rpc/auth.h>
68 72 #include <rpc/clnt.h>
69 73 #include <rpc/rpc_rdma.h>
70 74
71 75 #include <nfs/nfs.h>
72 76 #include <nfs/nfs_clnt.h>
73 77 #include <nfs/rnode.h>
74 78 #include <nfs/nfs_acl.h>
75 79 #include <nfs/lm.h>
76 80
77 81 #include <vm/hat.h>
78 82 #include <vm/as.h>
79 83 #include <vm/page.h>
80 84 #include <vm/pvn.h>
81 85 #include <vm/seg.h>
82 86 #include <vm/seg_map.h>
83 87 #include <vm/seg_kpm.h>
84 88 #include <vm/seg_vn.h>
85 89
86 90 #include <fs/fs_subr.h>
87 91
88 92 #include <sys/ddi.h>
89 93
90 94 static int nfs3_rdwrlbn(vnode_t *, page_t *, u_offset_t, size_t, int,
91 95 cred_t *);
92 96 static int nfs3write(vnode_t *, caddr_t, u_offset_t, int, cred_t *,
93 97 stable_how *);
94 98 static int nfs3read(vnode_t *, caddr_t, offset_t, int, size_t *, cred_t *);
95 99 static int nfs3setattr(vnode_t *, struct vattr *, int, cred_t *);
96 100 static int nfs3_accessx(void *, int, cred_t *);
97 101 static int nfs3lookup_dnlc(vnode_t *, char *, vnode_t **, cred_t *);
98 102 static int nfs3lookup_otw(vnode_t *, char *, vnode_t **, cred_t *, int);
99 103 static int nfs3create(vnode_t *, char *, struct vattr *, enum vcexcl,
100 104 int, vnode_t **, cred_t *, int);
101 105 static int nfs3excl_create_settimes(vnode_t *, struct vattr *, cred_t *);
102 106 static int nfs3mknod(vnode_t *, char *, struct vattr *, enum vcexcl,
103 107 int, vnode_t **, cred_t *);
104 108 static int nfs3rename(vnode_t *, char *, vnode_t *, char *, cred_t *,
105 109 caller_context_t *);
106 110 static int do_nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
107 111 static void nfs3readdir(vnode_t *, rddir_cache *, cred_t *);
108 112 static void nfs3readdirplus(vnode_t *, rddir_cache *, cred_t *);
109 113 static int nfs3_bio(struct buf *, stable_how *, cred_t *);
110 114 static int nfs3_getapage(vnode_t *, u_offset_t, size_t, uint_t *,
111 115 page_t *[], size_t, struct seg *, caddr_t,
112 116 enum seg_rw, cred_t *);
113 117 static void nfs3_readahead(vnode_t *, u_offset_t, caddr_t, struct seg *,
114 118 cred_t *);
115 119 static int nfs3_sync_putapage(vnode_t *, page_t *, u_offset_t, size_t,
116 120 int, cred_t *);
117 121 static int nfs3_sync_pageio(vnode_t *, page_t *, u_offset_t, size_t,
118 122 int, cred_t *);
119 123 static int nfs3_commit(vnode_t *, offset3, count3, cred_t *);
120 124 static void nfs3_set_mod(vnode_t *);
121 125 static void nfs3_get_commit(vnode_t *);
122 126 static void nfs3_get_commit_range(vnode_t *, u_offset_t, size_t);
123 127 static int nfs3_putpage_commit(vnode_t *, offset_t, size_t, cred_t *);
124 128 static int nfs3_commit_vp(vnode_t *, u_offset_t, size_t, cred_t *);
125 129 static int nfs3_sync_commit(vnode_t *, page_t *, offset3, count3,
126 130 cred_t *);
127 131 static void nfs3_async_commit(vnode_t *, page_t *, offset3, count3,
128 132 cred_t *);
129 133 static void nfs3_delmap_callback(struct as *, void *, uint_t);
130 134
131 135 /*
132 136 * Error flags used to pass information about certain special errors
133 137 * which need to be handled specially.
134 138 */
135 139 #define NFS_EOF -98
136 140 #define NFS_VERF_MISMATCH -97
137 141
138 142 /* ALIGN64 aligns the given buffer and adjust buffer size to 64 bit */
139 143 #define ALIGN64(x, ptr, sz) \
140 144 x = ((uintptr_t)(ptr)) & (sizeof (uint64_t) - 1); \
141 145 if (x) { \
142 146 x = sizeof (uint64_t) - (x); \
143 147 sz -= (x); \
144 148 ptr += (x); \
145 149 }
146 150
147 151 /*
148 152 * These are the vnode ops routines which implement the vnode interface to
149 153 * the networked file system. These routines just take their parameters,
150 154 * make them look networkish by putting the right info into interface structs,
151 155 * and then calling the appropriate remote routine(s) to do the work.
152 156 *
153 157 * Note on directory name lookup cacheing: If we detect a stale fhandle,
154 158 * we purge the directory cache relative to that vnode. This way, the
155 159 * user won't get burned by the cache repeatedly. See <nfs/rnode.h> for
156 160 * more details on rnode locking.
157 161 */
158 162
159 163 static int nfs3_open(vnode_t **, int, cred_t *, caller_context_t *);
160 164 static int nfs3_close(vnode_t *, int, int, offset_t, cred_t *,
161 165 caller_context_t *);
162 166 static int nfs3_read(vnode_t *, struct uio *, int, cred_t *,
163 167 caller_context_t *);
164 168 static int nfs3_write(vnode_t *, struct uio *, int, cred_t *,
165 169 caller_context_t *);
166 170 static int nfs3_ioctl(vnode_t *, int, intptr_t, int, cred_t *, int *,
167 171 caller_context_t *);
168 172 static int nfs3_getattr(vnode_t *, struct vattr *, int, cred_t *,
169 173 caller_context_t *);
170 174 static int nfs3_setattr(vnode_t *, struct vattr *, int, cred_t *,
171 175 caller_context_t *);
172 176 static int nfs3_access(vnode_t *, int, int, cred_t *, caller_context_t *);
173 177 static int nfs3_readlink(vnode_t *, struct uio *, cred_t *,
174 178 caller_context_t *);
175 179 static int nfs3_fsync(vnode_t *, int, cred_t *, caller_context_t *);
176 180 static void nfs3_inactive(vnode_t *, cred_t *, caller_context_t *);
177 181 static int nfs3_lookup(vnode_t *, char *, vnode_t **,
178 182 struct pathname *, int, vnode_t *, cred_t *,
179 183 caller_context_t *, int *, pathname_t *);
180 184 static int nfs3_create(vnode_t *, char *, struct vattr *, enum vcexcl,
181 185 int, vnode_t **, cred_t *, int, caller_context_t *,
182 186 vsecattr_t *);
183 187 static int nfs3_remove(vnode_t *, char *, cred_t *, caller_context_t *,
184 188 int);
185 189 static int nfs3_link(vnode_t *, vnode_t *, char *, cred_t *,
186 190 caller_context_t *, int);
187 191 static int nfs3_rename(vnode_t *, char *, vnode_t *, char *, cred_t *,
188 192 caller_context_t *, int);
189 193 static int nfs3_mkdir(vnode_t *, char *, struct vattr *, vnode_t **,
190 194 cred_t *, caller_context_t *, int, vsecattr_t *);
191 195 static int nfs3_rmdir(vnode_t *, char *, vnode_t *, cred_t *,
192 196 caller_context_t *, int);
193 197 static int nfs3_symlink(vnode_t *, char *, struct vattr *, char *,
194 198 cred_t *, caller_context_t *, int);
195 199 static int nfs3_readdir(vnode_t *, struct uio *, cred_t *, int *,
196 200 caller_context_t *, int);
197 201 static int nfs3_fid(vnode_t *, fid_t *, caller_context_t *);
198 202 static int nfs3_rwlock(vnode_t *, int, caller_context_t *);
199 203 static void nfs3_rwunlock(vnode_t *, int, caller_context_t *);
200 204 static int nfs3_seek(vnode_t *, offset_t, offset_t *, caller_context_t *);
201 205 static int nfs3_getpage(vnode_t *, offset_t, size_t, uint_t *,
202 206 page_t *[], size_t, struct seg *, caddr_t,
203 207 enum seg_rw, cred_t *, caller_context_t *);
204 208 static int nfs3_putpage(vnode_t *, offset_t, size_t, int, cred_t *,
205 209 caller_context_t *);
206 210 static int nfs3_map(vnode_t *, offset_t, struct as *, caddr_t *, size_t,
207 211 uchar_t, uchar_t, uint_t, cred_t *, caller_context_t *);
208 212 static int nfs3_addmap(vnode_t *, offset_t, struct as *, caddr_t, size_t,
209 213 uchar_t, uchar_t, uint_t, cred_t *, caller_context_t *);
210 214 static int nfs3_frlock(vnode_t *, int, struct flock64 *, int, offset_t,
211 215 struct flk_callback *, cred_t *, caller_context_t *);
212 216 static int nfs3_space(vnode_t *, int, struct flock64 *, int, offset_t,
213 217 cred_t *, caller_context_t *);
214 218 static int nfs3_realvp(vnode_t *, vnode_t **, caller_context_t *);
215 219 static int nfs3_delmap(vnode_t *, offset_t, struct as *, caddr_t, size_t,
216 220 uint_t, uint_t, uint_t, cred_t *, caller_context_t *);
217 221 static int nfs3_pathconf(vnode_t *, int, ulong_t *, cred_t *,
218 222 caller_context_t *);
219 223 static int nfs3_pageio(vnode_t *, page_t *, u_offset_t, size_t, int,
220 224 cred_t *, caller_context_t *);
221 225 static void nfs3_dispose(vnode_t *, page_t *, int, int, cred_t *,
222 226 caller_context_t *);
223 227 static int nfs3_setsecattr(vnode_t *, vsecattr_t *, int, cred_t *,
224 228 caller_context_t *);
225 229 static int nfs3_getsecattr(vnode_t *, vsecattr_t *, int, cred_t *,
226 230 caller_context_t *);
227 231 static int nfs3_shrlock(vnode_t *, int, struct shrlock *, int, cred_t *,
228 232 caller_context_t *);
229 233
230 234 struct vnodeops *nfs3_vnodeops;
231 235
232 236 const fs_operation_def_t nfs3_vnodeops_template[] = {
233 237 VOPNAME_OPEN, { .vop_open = nfs3_open },
234 238 VOPNAME_CLOSE, { .vop_close = nfs3_close },
235 239 VOPNAME_READ, { .vop_read = nfs3_read },
236 240 VOPNAME_WRITE, { .vop_write = nfs3_write },
237 241 VOPNAME_IOCTL, { .vop_ioctl = nfs3_ioctl },
238 242 VOPNAME_GETATTR, { .vop_getattr = nfs3_getattr },
239 243 VOPNAME_SETATTR, { .vop_setattr = nfs3_setattr },
240 244 VOPNAME_ACCESS, { .vop_access = nfs3_access },
241 245 VOPNAME_LOOKUP, { .vop_lookup = nfs3_lookup },
242 246 VOPNAME_CREATE, { .vop_create = nfs3_create },
243 247 VOPNAME_REMOVE, { .vop_remove = nfs3_remove },
244 248 VOPNAME_LINK, { .vop_link = nfs3_link },
245 249 VOPNAME_RENAME, { .vop_rename = nfs3_rename },
246 250 VOPNAME_MKDIR, { .vop_mkdir = nfs3_mkdir },
247 251 VOPNAME_RMDIR, { .vop_rmdir = nfs3_rmdir },
248 252 VOPNAME_READDIR, { .vop_readdir = nfs3_readdir },
249 253 VOPNAME_SYMLINK, { .vop_symlink = nfs3_symlink },
250 254 VOPNAME_READLINK, { .vop_readlink = nfs3_readlink },
251 255 VOPNAME_FSYNC, { .vop_fsync = nfs3_fsync },
252 256 VOPNAME_INACTIVE, { .vop_inactive = nfs3_inactive },
253 257 VOPNAME_FID, { .vop_fid = nfs3_fid },
254 258 VOPNAME_RWLOCK, { .vop_rwlock = nfs3_rwlock },
255 259 VOPNAME_RWUNLOCK, { .vop_rwunlock = nfs3_rwunlock },
256 260 VOPNAME_SEEK, { .vop_seek = nfs3_seek },
257 261 VOPNAME_FRLOCK, { .vop_frlock = nfs3_frlock },
258 262 VOPNAME_SPACE, { .vop_space = nfs3_space },
259 263 VOPNAME_REALVP, { .vop_realvp = nfs3_realvp },
260 264 VOPNAME_GETPAGE, { .vop_getpage = nfs3_getpage },
261 265 VOPNAME_PUTPAGE, { .vop_putpage = nfs3_putpage },
262 266 VOPNAME_MAP, { .vop_map = nfs3_map },
263 267 VOPNAME_ADDMAP, { .vop_addmap = nfs3_addmap },
264 268 VOPNAME_DELMAP, { .vop_delmap = nfs3_delmap },
265 269 /* no separate nfs3_dump */
266 270 VOPNAME_DUMP, { .vop_dump = nfs_dump },
267 271 VOPNAME_PATHCONF, { .vop_pathconf = nfs3_pathconf },
268 272 VOPNAME_PAGEIO, { .vop_pageio = nfs3_pageio },
269 273 VOPNAME_DISPOSE, { .vop_dispose = nfs3_dispose },
270 274 VOPNAME_SETSECATTR, { .vop_setsecattr = nfs3_setsecattr },
271 275 VOPNAME_GETSECATTR, { .vop_getsecattr = nfs3_getsecattr },
272 276 VOPNAME_SHRLOCK, { .vop_shrlock = nfs3_shrlock },
273 277 VOPNAME_VNEVENT, { .vop_vnevent = fs_vnevent_support },
274 278 NULL, NULL
275 279 };
276 280
277 281 /*
278 282 * XXX: This is referenced in modstubs.s
279 283 */
280 284 struct vnodeops *
281 285 nfs3_getvnodeops(void)
282 286 {
283 287 return (nfs3_vnodeops);
284 288 }
285 289
286 290 /* ARGSUSED */
287 291 static int
288 292 nfs3_open(vnode_t **vpp, int flag, cred_t *cr, caller_context_t *ct)
289 293 {
290 294 int error;
291 295 struct vattr va;
292 296 rnode_t *rp;
293 297 vnode_t *vp;
294 298
295 299 vp = *vpp;
296 300 if (nfs_zone() != VTOMI(vp)->mi_zone)
297 301 return (EIO);
298 302 rp = VTOR(vp);
299 303 mutex_enter(&rp->r_statelock);
300 304 if (rp->r_cred == NULL) {
301 305 crhold(cr);
302 306 rp->r_cred = cr;
303 307 }
304 308 mutex_exit(&rp->r_statelock);
305 309
306 310 /*
307 311 * If there is no cached data or if close-to-open
308 312 * consistency checking is turned off, we can avoid
309 313 * the over the wire getattr. Otherwise, if the
310 314 * file system is mounted readonly, then just verify
311 315 * the caches are up to date using the normal mechanism.
312 316 * Else, if the file is not mmap'd, then just mark
313 317 * the attributes as timed out. They will be refreshed
314 318 * and the caches validated prior to being used.
315 319 * Else, the file system is mounted writeable so
316 320 * force an over the wire GETATTR in order to ensure
317 321 * that all cached data is valid.
318 322 */
319 323 if (vp->v_count > 1 ||
320 324 ((vn_has_cached_data(vp) || HAVE_RDDIR_CACHE(rp)) &&
321 325 !(VTOMI(vp)->mi_flags & MI_NOCTO))) {
322 326 if (vn_is_readonly(vp))
323 327 error = nfs3_validate_caches(vp, cr);
324 328 else if (rp->r_mapcnt == 0 && vp->v_count == 1) {
325 329 PURGE_ATTRCACHE(vp);
326 330 error = 0;
327 331 } else {
328 332 va.va_mask = AT_ALL;
329 333 error = nfs3_getattr_otw(vp, &va, cr);
330 334 }
331 335 } else
332 336 error = 0;
333 337
334 338 return (error);
335 339 }
336 340
337 341 /* ARGSUSED */
338 342 static int
339 343 nfs3_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr,
340 344 caller_context_t *ct)
341 345 {
342 346 rnode_t *rp;
343 347 int error;
344 348 struct vattr va;
345 349
346 350 /*
347 351 * zone_enter(2) prevents processes from changing zones with NFS files
348 352 * open; if we happen to get here from the wrong zone we can't do
349 353 * anything over the wire.
350 354 */
351 355 if (VTOMI(vp)->mi_zone != nfs_zone()) {
352 356 /*
353 357 * We could attempt to clean up locks, except we're sure
354 358 * that the current process didn't acquire any locks on
355 359 * the file: any attempt to lock a file belong to another zone
356 360 * will fail, and one can't lock an NFS file and then change
357 361 * zones, as that fails too.
358 362 *
359 363 * Returning an error here is the sane thing to do. A
360 364 * subsequent call to VN_RELE() which translates to a
361 365 * nfs3_inactive() will clean up state: if the zone of the
362 366 * vnode's origin is still alive and kicking, an async worker
363 367 * thread will handle the request (from the correct zone), and
364 368 * everything (minus the commit and final nfs3_getattr_otw()
365 369 * call) should be OK. If the zone is going away
366 370 * nfs_async_inactive() will throw away cached pages inline.
367 371 */
368 372 return (EIO);
369 373 }
370 374
371 375 /*
372 376 * If we are using local locking for this filesystem, then
373 377 * release all of the SYSV style record locks. Otherwise,
374 378 * we are doing network locking and we need to release all
375 379 * of the network locks. All of the locks held by this
376 380 * process on this file are released no matter what the
377 381 * incoming reference count is.
378 382 */
379 383 if (VTOMI(vp)->mi_flags & MI_LLOCK) {
380 384 cleanlocks(vp, ttoproc(curthread)->p_pid, 0);
381 385 cleanshares(vp, ttoproc(curthread)->p_pid);
382 386 } else
383 387 nfs_lockrelease(vp, flag, offset, cr);
384 388
385 389 if (count > 1)
386 390 return (0);
387 391
388 392 /*
389 393 * If the file has been `unlinked', then purge the
390 394 * DNLC so that this vnode will get reycled quicker
391 395 * and the .nfs* file on the server will get removed.
392 396 */
393 397 rp = VTOR(vp);
394 398 if (rp->r_unldvp != NULL)
395 399 dnlc_purge_vp(vp);
396 400
397 401 /*
398 402 * If the file was open for write and there are pages,
399 403 * then if the file system was mounted using the "no-close-
400 404 * to-open" semantics, then start an asynchronous flush
401 405 * of the all of the pages in the file.
402 406 * else the file system was not mounted using the "no-close-
403 407 * to-open" semantics, then do a synchronous flush and
404 408 * commit of all of the dirty and uncommitted pages.
405 409 *
406 410 * The asynchronous flush of the pages in the "nocto" path
407 411 * mostly just associates a cred pointer with the rnode so
408 412 * writes which happen later will have a better chance of
409 413 * working. It also starts the data being written to the
410 414 * server, but without unnecessarily delaying the application.
411 415 */
412 416 if ((flag & FWRITE) && vn_has_cached_data(vp)) {
413 417 if (VTOMI(vp)->mi_flags & MI_NOCTO) {
414 418 error = nfs3_putpage(vp, (offset_t)0, 0, B_ASYNC,
415 419 cr, ct);
416 420 if (error == EAGAIN)
417 421 error = 0;
418 422 } else
419 423 error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
420 424 if (!error) {
421 425 mutex_enter(&rp->r_statelock);
422 426 error = rp->r_error;
423 427 rp->r_error = 0;
424 428 mutex_exit(&rp->r_statelock);
425 429 }
426 430 } else {
427 431 mutex_enter(&rp->r_statelock);
428 432 error = rp->r_error;
429 433 rp->r_error = 0;
430 434 mutex_exit(&rp->r_statelock);
431 435 }
432 436
433 437 /*
434 438 * If RWRITEATTR is set, then issue an over the wire GETATTR to
435 439 * refresh the attribute cache with a set of attributes which
436 440 * weren't returned from a WRITE. This will enable the close-
437 441 * to-open processing to work.
438 442 */
439 443 if (rp->r_flags & RWRITEATTR)
440 444 (void) nfs3_getattr_otw(vp, &va, cr);
441 445
442 446 return (error);
443 447 }
444 448
445 449 /* ARGSUSED */
446 450 static int
447 451 nfs3_directio_read(vnode_t *vp, struct uio *uiop, cred_t *cr)
448 452 {
449 453 mntinfo_t *mi;
450 454 READ3args args;
451 455 READ3uiores res;
452 456 int tsize;
453 457 offset_t offset;
454 458 ssize_t count;
455 459 int error;
456 460 int douprintf;
457 461 failinfo_t fi;
458 462 char *sv_hostname;
459 463
460 464 mi = VTOMI(vp);
461 465 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
462 466 sv_hostname = VTOR(vp)->r_server->sv_hostname;
463 467
464 468 douprintf = 1;
465 469 args.file = *VTOFH3(vp);
466 470 fi.vp = vp;
467 471 fi.fhp = (caddr_t)&args.file;
468 472 fi.copyproc = nfs3copyfh;
469 473 fi.lookupproc = nfs3lookup;
470 474 fi.xattrdirproc = acl_getxattrdir3;
471 475
472 476 res.uiop = uiop;
473 477
474 478 res.wlist = NULL;
475 479
476 480 offset = uiop->uio_loffset;
477 481 count = uiop->uio_resid;
478 482
479 483 do {
480 484 if (mi->mi_io_kstats) {
481 485 mutex_enter(&mi->mi_lock);
482 486 kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
483 487 mutex_exit(&mi->mi_lock);
484 488 }
485 489
486 490 do {
487 491 tsize = MIN(mi->mi_tsize, count);
488 492 args.offset = (offset3)offset;
489 493 args.count = (count3)tsize;
490 494 res.size = (uint_t)tsize;
491 495 args.res_uiop = uiop;
492 496 args.res_data_val_alt = NULL;
493 497
494 498 error = rfs3call(mi, NFSPROC3_READ,
495 499 xdr_READ3args, (caddr_t)&args,
496 500 xdr_READ3uiores, (caddr_t)&res, cr,
497 501 &douprintf, &res.status, 0, &fi);
498 502 } while (error == ENFS_TRYAGAIN);
499 503
500 504 if (mi->mi_io_kstats) {
501 505 mutex_enter(&mi->mi_lock);
502 506 kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
503 507 mutex_exit(&mi->mi_lock);
504 508 }
505 509
506 510 if (error)
507 511 return (error);
508 512
509 513 error = geterrno3(res.status);
510 514 if (error)
511 515 return (error);
512 516
513 517 if (res.count != res.size) {
514 518 zcmn_err(getzoneid(), CE_WARN,
515 519 "nfs3_directio_read: server %s returned incorrect amount",
516 520 sv_hostname);
517 521 return (EIO);
518 522 }
519 523 count -= res.count;
520 524 offset += res.count;
521 525 if (mi->mi_io_kstats) {
522 526 mutex_enter(&mi->mi_lock);
523 527 KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
524 528 KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
525 529 mutex_exit(&mi->mi_lock);
526 530 }
527 531 lwp_stat_update(LWP_STAT_INBLK, 1);
528 532 } while (count && !res.eof);
529 533
530 534 return (0);
531 535 }
532 536
533 537 /* ARGSUSED */
534 538 static int
535 539 nfs3_read(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
536 540 caller_context_t *ct)
537 541 {
538 542 rnode_t *rp;
539 543 u_offset_t off;
540 544 offset_t diff;
541 545 int on;
542 546 size_t n;
543 547 caddr_t base;
544 548 uint_t flags;
545 549 int error = 0;
546 550 mntinfo_t *mi;
547 551
548 552 rp = VTOR(vp);
549 553 mi = VTOMI(vp);
550 554
551 555 ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
552 556
553 557 if (nfs_zone() != mi->mi_zone)
554 558 return (EIO);
555 559
556 560 if (vp->v_type != VREG)
557 561 return (EISDIR);
558 562
559 563 if (uiop->uio_resid == 0)
560 564 return (0);
561 565
562 566 if (uiop->uio_loffset < 0 || uiop->uio_loffset + uiop->uio_resid < 0)
563 567 return (EINVAL);
564 568
565 569 /*
566 570 * Bypass VM if caching has been disabled (e.g., locking) or if
567 571 * using client-side direct I/O and the file is not mmap'd and
568 572 * there are no cached pages.
569 573 */
570 574 if ((vp->v_flag & VNOCACHE) ||
571 575 (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
572 576 rp->r_mapcnt == 0 && rp->r_inmap == 0 &&
573 577 !vn_has_cached_data(vp))) {
574 578 return (nfs3_directio_read(vp, uiop, cr));
575 579 }
576 580
577 581 do {
578 582 off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
579 583 on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
580 584 n = MIN(MAXBSIZE - on, uiop->uio_resid);
581 585
582 586 error = nfs3_validate_caches(vp, cr);
583 587 if (error)
584 588 break;
585 589
586 590 mutex_enter(&rp->r_statelock);
587 591 while (rp->r_flags & RINCACHEPURGE) {
588 592 if (!cv_wait_sig(&rp->r_cv, &rp->r_statelock)) {
589 593 mutex_exit(&rp->r_statelock);
590 594 return (EINTR);
591 595 }
592 596 }
593 597 diff = rp->r_size - uiop->uio_loffset;
594 598 mutex_exit(&rp->r_statelock);
595 599 if (diff <= 0)
596 600 break;
597 601 if (diff < n)
598 602 n = (size_t)diff;
599 603
600 604 if (vpm_enable) {
601 605 /*
602 606 * Copy data.
603 607 */
604 608 error = vpm_data_copy(vp, off + on, n, uiop,
605 609 1, NULL, 0, S_READ);
606 610 } else {
607 611 base = segmap_getmapflt(segkmap, vp, off + on, n, 1,
608 612 S_READ);
609 613
610 614 error = uiomove(base + on, n, UIO_READ, uiop);
611 615 }
612 616
613 617 if (!error) {
614 618 /*
615 619 * If read a whole block or read to eof,
616 620 * won't need this buffer again soon.
617 621 */
618 622 mutex_enter(&rp->r_statelock);
619 623 if (n + on == MAXBSIZE ||
620 624 uiop->uio_loffset == rp->r_size)
621 625 flags = SM_DONTNEED;
622 626 else
623 627 flags = 0;
624 628 mutex_exit(&rp->r_statelock);
625 629 if (vpm_enable) {
626 630 error = vpm_sync_pages(vp, off, n, flags);
627 631 } else {
628 632 error = segmap_release(segkmap, base, flags);
629 633 }
630 634 } else {
631 635 if (vpm_enable) {
632 636 (void) vpm_sync_pages(vp, off, n, 0);
633 637 } else {
634 638 (void) segmap_release(segkmap, base, 0);
635 639 }
636 640 }
637 641 } while (!error && uiop->uio_resid > 0);
638 642
639 643 return (error);
640 644 }
641 645
642 646 /* ARGSUSED */
643 647 static int
644 648 nfs3_write(vnode_t *vp, struct uio *uiop, int ioflag, cred_t *cr,
645 649 caller_context_t *ct)
646 650 {
647 651 rlim64_t limit = uiop->uio_llimit;
648 652 rnode_t *rp;
649 653 u_offset_t off;
650 654 caddr_t base;
651 655 uint_t flags;
652 656 int remainder;
653 657 size_t n;
654 658 int on;
655 659 int error;
656 660 int resid;
657 661 offset_t offset;
658 662 mntinfo_t *mi;
659 663 uint_t bsize;
660 664
661 665 rp = VTOR(vp);
662 666
663 667 if (vp->v_type != VREG)
664 668 return (EISDIR);
665 669
666 670 mi = VTOMI(vp);
667 671 if (nfs_zone() != mi->mi_zone)
668 672 return (EIO);
669 673 if (uiop->uio_resid == 0)
670 674 return (0);
671 675
672 676 if (ioflag & FAPPEND) {
673 677 struct vattr va;
674 678
675 679 /*
676 680 * Must serialize if appending.
677 681 */
678 682 if (nfs_rw_lock_held(&rp->r_rwlock, RW_READER)) {
679 683 nfs_rw_exit(&rp->r_rwlock);
680 684 if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER,
681 685 INTR(vp)))
682 686 return (EINTR);
683 687 }
684 688
685 689 va.va_mask = AT_SIZE;
686 690 error = nfs3getattr(vp, &va, cr);
687 691 if (error)
688 692 return (error);
689 693 uiop->uio_loffset = va.va_size;
690 694 }
691 695
692 696 offset = uiop->uio_loffset + uiop->uio_resid;
693 697
694 698 if (uiop->uio_loffset < 0 || offset < 0)
695 699 return (EINVAL);
696 700
697 701 if (limit == RLIM64_INFINITY || limit > MAXOFFSET_T)
698 702 limit = MAXOFFSET_T;
699 703
700 704 /*
701 705 * Check to make sure that the process will not exceed
702 706 * its limit on file size. It is okay to write up to
703 707 * the limit, but not beyond. Thus, the write which
704 708 * reaches the limit will be short and the next write
705 709 * will return an error.
706 710 */
707 711 remainder = 0;
708 712 if (offset > limit) {
709 713 remainder = offset - limit;
710 714 uiop->uio_resid = limit - uiop->uio_loffset;
711 715 if (uiop->uio_resid <= 0) {
712 716 proc_t *p = ttoproc(curthread);
713 717
714 718 uiop->uio_resid += remainder;
715 719 mutex_enter(&p->p_lock);
716 720 (void) rctl_action(rctlproc_legacy[RLIMIT_FSIZE],
717 721 p->p_rctls, p, RCA_UNSAFE_SIGINFO);
718 722 mutex_exit(&p->p_lock);
719 723 return (EFBIG);
720 724 }
721 725 }
722 726
723 727 if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp)))
724 728 return (EINTR);
725 729
726 730 /*
727 731 * Bypass VM if caching has been disabled (e.g., locking) or if
728 732 * using client-side direct I/O and the file is not mmap'd and
729 733 * there are no cached pages.
730 734 */
731 735 if ((vp->v_flag & VNOCACHE) ||
732 736 (((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO)) &&
733 737 rp->r_mapcnt == 0 && rp->r_inmap == 0 &&
734 738 !vn_has_cached_data(vp))) {
735 739 size_t bufsize;
736 740 int count;
737 741 u_offset_t org_offset;
738 742 stable_how stab_comm;
739 743
740 744 nfs3_fwrite:
741 745 if (rp->r_flags & RSTALE) {
742 746 resid = uiop->uio_resid;
743 747 offset = uiop->uio_loffset;
744 748 error = rp->r_error;
745 749 /*
746 750 * A close may have cleared r_error, if so,
747 751 * propagate ESTALE error return properly
748 752 */
749 753 if (error == 0)
750 754 error = ESTALE;
751 755 goto bottom;
752 756 }
753 757 bufsize = MIN(uiop->uio_resid, mi->mi_stsize);
754 758 base = kmem_alloc(bufsize, KM_SLEEP);
755 759 do {
756 760 if (ioflag & FDSYNC)
757 761 stab_comm = DATA_SYNC;
758 762 else
759 763 stab_comm = FILE_SYNC;
760 764 resid = uiop->uio_resid;
761 765 offset = uiop->uio_loffset;
762 766 count = MIN(uiop->uio_resid, bufsize);
763 767 org_offset = uiop->uio_loffset;
764 768 error = uiomove(base, count, UIO_WRITE, uiop);
765 769 if (!error) {
766 770 error = nfs3write(vp, base, org_offset,
767 771 count, cr, &stab_comm);
768 772 }
769 773 } while (!error && uiop->uio_resid > 0);
770 774 kmem_free(base, bufsize);
771 775 goto bottom;
772 776 }
773 777
774 778
775 779 bsize = vp->v_vfsp->vfs_bsize;
776 780
777 781 do {
778 782 off = uiop->uio_loffset & MAXBMASK; /* mapping offset */
779 783 on = uiop->uio_loffset & MAXBOFFSET; /* Relative offset */
780 784 n = MIN(MAXBSIZE - on, uiop->uio_resid);
781 785
782 786 resid = uiop->uio_resid;
783 787 offset = uiop->uio_loffset;
784 788
785 789 if (rp->r_flags & RSTALE) {
786 790 error = rp->r_error;
787 791 /*
788 792 * A close may have cleared r_error, if so,
789 793 * propagate ESTALE error return properly
790 794 */
791 795 if (error == 0)
792 796 error = ESTALE;
793 797 break;
794 798 }
795 799
796 800 /*
797 801 * Don't create dirty pages faster than they
798 802 * can be cleaned so that the system doesn't
799 803 * get imbalanced. If the async queue is
800 804 * maxed out, then wait for it to drain before
801 805 * creating more dirty pages. Also, wait for
802 806 * any threads doing pagewalks in the vop_getattr
803 807 * entry points so that they don't block for
804 808 * long periods.
805 809 */
806 810 mutex_enter(&rp->r_statelock);
807 811 while ((mi->mi_max_threads != 0 &&
808 812 rp->r_awcount > 2 * mi->mi_max_threads) ||
809 813 rp->r_gcount > 0) {
810 814 if (INTR(vp)) {
811 815 klwp_t *lwp = ttolwp(curthread);
812 816
813 817 if (lwp != NULL)
814 818 lwp->lwp_nostop++;
815 819 if (!cv_wait_sig(&rp->r_cv, &rp->r_statelock)) {
816 820 mutex_exit(&rp->r_statelock);
817 821 if (lwp != NULL)
818 822 lwp->lwp_nostop--;
819 823 error = EINTR;
820 824 goto bottom;
821 825 }
822 826 if (lwp != NULL)
823 827 lwp->lwp_nostop--;
824 828 } else
825 829 cv_wait(&rp->r_cv, &rp->r_statelock);
826 830 }
827 831 mutex_exit(&rp->r_statelock);
828 832
829 833 /*
830 834 * Touch the page and fault it in if it is not in core
831 835 * before segmap_getmapflt or vpm_data_copy can lock it.
832 836 * This is to avoid the deadlock if the buffer is mapped
833 837 * to the same file through mmap which we want to write.
834 838 */
835 839 uio_prefaultpages((long)n, uiop);
836 840
837 841 if (vpm_enable) {
838 842 /*
839 843 * It will use kpm mappings, so no need to
840 844 * pass an address.
841 845 */
842 846 error = writerp(rp, NULL, n, uiop, 0);
843 847 } else {
844 848 if (segmap_kpm) {
845 849 int pon = uiop->uio_loffset & PAGEOFFSET;
846 850 size_t pn = MIN(PAGESIZE - pon,
847 851 uiop->uio_resid);
848 852 int pagecreate;
849 853
850 854 mutex_enter(&rp->r_statelock);
851 855 pagecreate = (pon == 0) && (pn == PAGESIZE ||
852 856 uiop->uio_loffset + pn >= rp->r_size);
853 857 mutex_exit(&rp->r_statelock);
854 858
855 859 base = segmap_getmapflt(segkmap, vp, off + on,
856 860 pn, !pagecreate, S_WRITE);
857 861
858 862 error = writerp(rp, base + pon, n, uiop,
859 863 pagecreate);
860 864
861 865 } else {
862 866 base = segmap_getmapflt(segkmap, vp, off + on,
863 867 n, 0, S_READ);
864 868 error = writerp(rp, base + on, n, uiop, 0);
865 869 }
866 870 }
867 871
868 872 if (!error) {
869 873 if (mi->mi_flags & MI_NOAC)
870 874 flags = SM_WRITE;
871 875 else if ((uiop->uio_loffset % bsize) == 0 ||
872 876 IS_SWAPVP(vp)) {
873 877 /*
874 878 * Have written a whole block.
875 879 * Start an asynchronous write
876 880 * and mark the buffer to
877 881 * indicate that it won't be
878 882 * needed again soon.
879 883 */
880 884 flags = SM_WRITE | SM_ASYNC | SM_DONTNEED;
881 885 } else
882 886 flags = 0;
883 887 if ((ioflag & (FSYNC|FDSYNC)) ||
884 888 (rp->r_flags & ROUTOFSPACE)) {
885 889 flags &= ~SM_ASYNC;
886 890 flags |= SM_WRITE;
887 891 }
888 892 if (vpm_enable) {
889 893 error = vpm_sync_pages(vp, off, n, flags);
890 894 } else {
891 895 error = segmap_release(segkmap, base, flags);
892 896 }
893 897 } else {
894 898 if (vpm_enable) {
895 899 (void) vpm_sync_pages(vp, off, n, 0);
896 900 } else {
897 901 (void) segmap_release(segkmap, base, 0);
898 902 }
899 903 /*
900 904 * In the event that we got an access error while
901 905 * faulting in a page for a write-only file just
902 906 * force a write.
903 907 */
904 908 if (error == EACCES)
905 909 goto nfs3_fwrite;
906 910 }
907 911 } while (!error && uiop->uio_resid > 0);
908 912
909 913 bottom:
910 914 if (error) {
911 915 uiop->uio_resid = resid + remainder;
912 916 uiop->uio_loffset = offset;
913 917 } else
914 918 uiop->uio_resid += remainder;
915 919
916 920 nfs_rw_exit(&rp->r_lkserlock);
917 921
918 922 return (error);
919 923 }
920 924
921 925 /*
922 926 * Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED}
923 927 */
924 928 static int
925 929 nfs3_rdwrlbn(vnode_t *vp, page_t *pp, u_offset_t off, size_t len,
926 930 int flags, cred_t *cr)
927 931 {
928 932 struct buf *bp;
929 933 int error;
930 934 page_t *savepp;
931 935 uchar_t fsdata;
932 936 stable_how stab_comm;
933 937
934 938 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
935 939 bp = pageio_setup(pp, len, vp, flags);
936 940 ASSERT(bp != NULL);
937 941
938 942 /*
939 943 * pageio_setup should have set b_addr to 0. This
940 944 * is correct since we want to do I/O on a page
941 945 * boundary. bp_mapin will use this addr to calculate
942 946 * an offset, and then set b_addr to the kernel virtual
943 947 * address it allocated for us.
944 948 */
945 949 ASSERT(bp->b_un.b_addr == 0);
946 950
947 951 bp->b_edev = 0;
948 952 bp->b_dev = 0;
949 953 bp->b_lblkno = lbtodb(off);
950 954 bp->b_file = vp;
951 955 bp->b_offset = (offset_t)off;
952 956 bp_mapin(bp);
953 957
954 958 /*
955 959 * Calculate the desired level of stability to write data
956 960 * on the server and then mark all of the pages to reflect
957 961 * this.
958 962 */
959 963 if ((flags & (B_WRITE|B_ASYNC)) == (B_WRITE|B_ASYNC) &&
960 964 freemem > desfree) {
961 965 stab_comm = UNSTABLE;
962 966 fsdata = C_DELAYCOMMIT;
963 967 } else {
964 968 stab_comm = FILE_SYNC;
965 969 fsdata = C_NOCOMMIT;
966 970 }
967 971
968 972 savepp = pp;
969 973 do {
970 974 pp->p_fsdata = fsdata;
971 975 } while ((pp = pp->p_next) != savepp);
972 976
973 977 error = nfs3_bio(bp, &stab_comm, cr);
974 978
975 979 bp_mapout(bp);
976 980 pageio_done(bp);
977 981
978 982 /*
979 983 * If the server wrote pages in a more stable fashion than
980 984 * was requested, then clear all of the marks in the pages
981 985 * indicating that COMMIT operations were required.
982 986 */
983 987 if (stab_comm != UNSTABLE && fsdata == C_DELAYCOMMIT) {
984 988 do {
985 989 pp->p_fsdata = C_NOCOMMIT;
986 990 } while ((pp = pp->p_next) != savepp);
987 991 }
988 992
989 993 return (error);
990 994 }
991 995
992 996 /*
993 997 * Write to file. Writes to remote server in largest size
994 998 * chunks that the server can handle. Write is synchronous.
995 999 */
996 1000 static int
997 1001 nfs3write(vnode_t *vp, caddr_t base, u_offset_t offset, int count, cred_t *cr,
998 1002 stable_how *stab_comm)
999 1003 {
1000 1004 mntinfo_t *mi;
1001 1005 WRITE3args args;
1002 1006 WRITE3res res;
1003 1007 int error;
1004 1008 int tsize;
1005 1009 rnode_t *rp;
1006 1010 int douprintf;
1007 1011
1008 1012 rp = VTOR(vp);
1009 1013 mi = VTOMI(vp);
1010 1014
1011 1015 ASSERT(nfs_zone() == mi->mi_zone);
1012 1016
1013 1017 args.file = *VTOFH3(vp);
1014 1018 args.stable = *stab_comm;
1015 1019
1016 1020 *stab_comm = FILE_SYNC;
1017 1021
1018 1022 douprintf = 1;
1019 1023
1020 1024 do {
1021 1025 if ((vp->v_flag & VNOCACHE) ||
1022 1026 (rp->r_flags & RDIRECTIO) ||
1023 1027 (mi->mi_flags & MI_DIRECTIO))
1024 1028 tsize = MIN(mi->mi_stsize, count);
1025 1029 else
1026 1030 tsize = MIN(mi->mi_curwrite, count);
1027 1031 args.offset = (offset3)offset;
1028 1032 args.count = (count3)tsize;
1029 1033 args.data.data_len = (uint_t)tsize;
1030 1034 args.data.data_val = base;
1031 1035
1032 1036 if (mi->mi_io_kstats) {
1033 1037 mutex_enter(&mi->mi_lock);
1034 1038 kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1035 1039 mutex_exit(&mi->mi_lock);
1036 1040 }
1037 1041 args.mblk = NULL;
1038 1042 do {
1039 1043 error = rfs3call(mi, NFSPROC3_WRITE,
1040 1044 xdr_WRITE3args, (caddr_t)&args,
1041 1045 xdr_WRITE3res, (caddr_t)&res, cr,
1042 1046 &douprintf, &res.status, 0, NULL);
1043 1047 } while (error == ENFS_TRYAGAIN);
1044 1048 if (mi->mi_io_kstats) {
1045 1049 mutex_enter(&mi->mi_lock);
1046 1050 kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
1047 1051 mutex_exit(&mi->mi_lock);
1048 1052 }
1049 1053
1050 1054 if (error)
1051 1055 return (error);
1052 1056 error = geterrno3(res.status);
1053 1057 if (!error) {
1054 1058 if (res.resok.count > args.count) {
1055 1059 zcmn_err(getzoneid(), CE_WARN,
1056 1060 "nfs3write: server %s wrote %u, "
1057 1061 "requested was %u",
1058 1062 rp->r_server->sv_hostname,
1059 1063 res.resok.count, args.count);
1060 1064 return (EIO);
1061 1065 }
1062 1066 if (res.resok.committed == UNSTABLE) {
1063 1067 *stab_comm = UNSTABLE;
1064 1068 if (args.stable == DATA_SYNC ||
1065 1069 args.stable == FILE_SYNC) {
1066 1070 zcmn_err(getzoneid(), CE_WARN,
1067 1071 "nfs3write: server %s did not commit to stable storage",
1068 1072 rp->r_server->sv_hostname);
1069 1073 return (EIO);
1070 1074 }
1071 1075 }
1072 1076 tsize = (int)res.resok.count;
1073 1077 count -= tsize;
1074 1078 base += tsize;
1075 1079 offset += tsize;
1076 1080 if (mi->mi_io_kstats) {
1077 1081 mutex_enter(&mi->mi_lock);
1078 1082 KSTAT_IO_PTR(mi->mi_io_kstats)->writes++;
1079 1083 KSTAT_IO_PTR(mi->mi_io_kstats)->nwritten +=
1080 1084 tsize;
1081 1085 mutex_exit(&mi->mi_lock);
1082 1086 }
1083 1087 lwp_stat_update(LWP_STAT_OUBLK, 1);
1084 1088 mutex_enter(&rp->r_statelock);
1085 1089 if (rp->r_flags & RHAVEVERF) {
1086 1090 if (rp->r_verf != res.resok.verf) {
1087 1091 nfs3_set_mod(vp);
1088 1092 rp->r_verf = res.resok.verf;
1089 1093 /*
1090 1094 * If the data was written UNSTABLE,
1091 1095 * then might as well stop because
1092 1096 * the whole block will have to get
1093 1097 * rewritten anyway.
1094 1098 */
1095 1099 if (*stab_comm == UNSTABLE) {
1096 1100 mutex_exit(&rp->r_statelock);
1097 1101 break;
1098 1102 }
1099 1103 }
1100 1104 } else {
1101 1105 rp->r_verf = res.resok.verf;
1102 1106 rp->r_flags |= RHAVEVERF;
1103 1107 }
1104 1108 /*
1105 1109 * Mark the attribute cache as timed out and
1106 1110 * set RWRITEATTR to indicate that the file
1107 1111 * was modified with a WRITE operation and
1108 1112 * that the attributes can not be trusted.
1109 1113 */
1110 1114 PURGE_ATTRCACHE_LOCKED(rp);
1111 1115 rp->r_flags |= RWRITEATTR;
1112 1116 mutex_exit(&rp->r_statelock);
1113 1117 }
1114 1118 } while (!error && count);
1115 1119
1116 1120 return (error);
1117 1121 }
1118 1122
1119 1123 /*
1120 1124 * Read from a file. Reads data in largest chunks our interface can handle.
1121 1125 */
1122 1126 static int
1123 1127 nfs3read(vnode_t *vp, caddr_t base, offset_t offset, int count,
1124 1128 size_t *residp, cred_t *cr)
1125 1129 {
1126 1130 mntinfo_t *mi;
1127 1131 READ3args args;
1128 1132 READ3vres res;
1129 1133 int tsize;
1130 1134 int error;
1131 1135 int douprintf;
1132 1136 failinfo_t fi;
1133 1137 rnode_t *rp;
1134 1138 struct vattr va;
1135 1139 hrtime_t t;
1136 1140
1137 1141 rp = VTOR(vp);
1138 1142 mi = VTOMI(vp);
1139 1143 ASSERT(nfs_zone() == mi->mi_zone);
1140 1144 douprintf = 1;
1141 1145
1142 1146 args.file = *VTOFH3(vp);
1143 1147 fi.vp = vp;
1144 1148 fi.fhp = (caddr_t)&args.file;
1145 1149 fi.copyproc = nfs3copyfh;
1146 1150 fi.lookupproc = nfs3lookup;
1147 1151 fi.xattrdirproc = acl_getxattrdir3;
1148 1152
1149 1153 res.pov.fres.vp = vp;
1150 1154 res.pov.fres.vap = &va;
1151 1155
1152 1156 res.wlist = NULL;
1153 1157 *residp = count;
1154 1158 do {
1155 1159 if (mi->mi_io_kstats) {
1156 1160 mutex_enter(&mi->mi_lock);
1157 1161 kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1158 1162 mutex_exit(&mi->mi_lock);
1159 1163 }
1160 1164
1161 1165 do {
1162 1166 if ((vp->v_flag & VNOCACHE) ||
1163 1167 (rp->r_flags & RDIRECTIO) ||
1164 1168 (mi->mi_flags & MI_DIRECTIO))
1165 1169 tsize = MIN(mi->mi_tsize, count);
1166 1170 else
1167 1171 tsize = MIN(mi->mi_curread, count);
1168 1172 res.data.data_val = base;
1169 1173 res.data.data_len = tsize;
1170 1174 args.offset = (offset3)offset;
1171 1175 args.count = (count3)tsize;
1172 1176 args.res_uiop = NULL;
1173 1177 args.res_data_val_alt = base;
1174 1178
1175 1179 t = gethrtime();
1176 1180 error = rfs3call(mi, NFSPROC3_READ,
1177 1181 xdr_READ3args, (caddr_t)&args,
1178 1182 xdr_READ3vres, (caddr_t)&res, cr,
1179 1183 &douprintf, &res.status, 0, &fi);
1180 1184 } while (error == ENFS_TRYAGAIN);
1181 1185
1182 1186 if (mi->mi_io_kstats) {
1183 1187 mutex_enter(&mi->mi_lock);
1184 1188 kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
1185 1189 mutex_exit(&mi->mi_lock);
1186 1190 }
1187 1191
1188 1192 if (error)
1189 1193 return (error);
1190 1194
1191 1195 error = geterrno3(res.status);
1192 1196 if (error)
1193 1197 return (error);
1194 1198
1195 1199 if (res.count != res.data.data_len) {
1196 1200 zcmn_err(getzoneid(), CE_WARN,
1197 1201 "nfs3read: server %s returned incorrect amount",
1198 1202 rp->r_server->sv_hostname);
1199 1203 return (EIO);
1200 1204 }
1201 1205
1202 1206 count -= res.count;
1203 1207 *residp = count;
1204 1208 base += res.count;
1205 1209 offset += res.count;
1206 1210 if (mi->mi_io_kstats) {
1207 1211 mutex_enter(&mi->mi_lock);
1208 1212 KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
1209 1213 KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.count;
1210 1214 mutex_exit(&mi->mi_lock);
1211 1215 }
1212 1216 lwp_stat_update(LWP_STAT_INBLK, 1);
1213 1217 } while (count && !res.eof);
1214 1218
1215 1219 if (res.pov.attributes) {
1216 1220 mutex_enter(&rp->r_statelock);
1217 1221 if (!CACHE_VALID(rp, va.va_mtime, va.va_size)) {
1218 1222 mutex_exit(&rp->r_statelock);
1219 1223 PURGE_ATTRCACHE(vp);
1220 1224 } else {
1221 1225 if (rp->r_mtime <= t)
1222 1226 nfs_attrcache_va(vp, &va);
1223 1227 mutex_exit(&rp->r_statelock);
1224 1228 }
1225 1229 }
1226 1230
1227 1231 return (0);
1228 1232 }
1229 1233
1230 1234 /* ARGSUSED */
1231 1235 static int
1232 1236 nfs3_ioctl(vnode_t *vp, int cmd, intptr_t arg, int flag, cred_t *cr, int *rvalp,
1233 1237 caller_context_t *ct)
1234 1238 {
1235 1239
1236 1240 if (nfs_zone() != VTOMI(vp)->mi_zone)
1237 1241 return (EIO);
1238 1242 switch (cmd) {
1239 1243 case _FIODIRECTIO:
1240 1244 return (nfs_directio(vp, (int)arg, cr));
1241 1245 default:
1242 1246 return (ENOTTY);
1243 1247 }
1244 1248 }
1245 1249
1246 1250 /* ARGSUSED */
1247 1251 static int
1248 1252 nfs3_getattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr,
1249 1253 caller_context_t *ct)
1250 1254 {
1251 1255 int error;
1252 1256 rnode_t *rp;
1253 1257
1254 1258 if (nfs_zone() != VTOMI(vp)->mi_zone)
1255 1259 return (EIO);
1256 1260 /*
1257 1261 * If it has been specified that the return value will
1258 1262 * just be used as a hint, and we are only being asked
1259 1263 * for size, fsid or rdevid, then return the client's
1260 1264 * notion of these values without checking to make sure
1261 1265 * that the attribute cache is up to date.
1262 1266 * The whole point is to avoid an over the wire GETATTR
1263 1267 * call.
1264 1268 */
1265 1269 rp = VTOR(vp);
1266 1270 if (flags & ATTR_HINT) {
1267 1271 if (vap->va_mask ==
1268 1272 (vap->va_mask & (AT_SIZE | AT_FSID | AT_RDEV))) {
1269 1273 mutex_enter(&rp->r_statelock);
1270 1274 if (vap->va_mask | AT_SIZE)
1271 1275 vap->va_size = rp->r_size;
1272 1276 if (vap->va_mask | AT_FSID)
1273 1277 vap->va_fsid = rp->r_attr.va_fsid;
1274 1278 if (vap->va_mask | AT_RDEV)
1275 1279 vap->va_rdev = rp->r_attr.va_rdev;
1276 1280 mutex_exit(&rp->r_statelock);
1277 1281 return (0);
1278 1282 }
1279 1283 }
1280 1284
1281 1285 /*
1282 1286 * Only need to flush pages if asking for the mtime
1283 1287 * and if there any dirty pages or any outstanding
1284 1288 * asynchronous (write) requests for this file.
1285 1289 */
1286 1290 if (vap->va_mask & AT_MTIME) {
1287 1291 if (vn_has_cached_data(vp) &&
1288 1292 ((rp->r_flags & RDIRTY) || rp->r_awcount > 0)) {
1289 1293 mutex_enter(&rp->r_statelock);
1290 1294 rp->r_gcount++;
1291 1295 mutex_exit(&rp->r_statelock);
1292 1296 error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr, ct);
1293 1297 mutex_enter(&rp->r_statelock);
1294 1298 if (error && (error == ENOSPC || error == EDQUOT)) {
1295 1299 if (!rp->r_error)
1296 1300 rp->r_error = error;
1297 1301 }
1298 1302 if (--rp->r_gcount == 0)
1299 1303 cv_broadcast(&rp->r_cv);
1300 1304 mutex_exit(&rp->r_statelock);
1301 1305 }
1302 1306 }
1303 1307
1304 1308 return (nfs3getattr(vp, vap, cr));
1305 1309 }
1306 1310
1307 1311 /*ARGSUSED4*/
1308 1312 static int
1309 1313 nfs3_setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr,
1310 1314 caller_context_t *ct)
1311 1315 {
1312 1316 int error;
1313 1317 struct vattr va;
1314 1318
1315 1319 if (vap->va_mask & AT_NOSET)
1316 1320 return (EINVAL);
1317 1321 if (nfs_zone() != VTOMI(vp)->mi_zone)
1318 1322 return (EIO);
1319 1323
1320 1324 va.va_mask = AT_UID | AT_MODE;
1321 1325 error = nfs3getattr(vp, &va, cr);
1322 1326 if (error)
1323 1327 return (error);
1324 1328
1325 1329 error = secpolicy_vnode_setattr(cr, vp, vap, &va, flags, nfs3_accessx,
1326 1330 vp);
1327 1331 if (error)
1328 1332 return (error);
1329 1333
1330 1334 return (nfs3setattr(vp, vap, flags, cr));
1331 1335 }
1332 1336
1333 1337 static int
1334 1338 nfs3setattr(vnode_t *vp, struct vattr *vap, int flags, cred_t *cr)
1335 1339 {
1336 1340 int error;
1337 1341 uint_t mask;
1338 1342 SETATTR3args args;
1339 1343 SETATTR3res res;
1340 1344 int douprintf;
1341 1345 rnode_t *rp;
1342 1346 struct vattr va;
1343 1347 mode_t omode;
1344 1348 vsecattr_t *vsp;
1345 1349 hrtime_t t;
1346 1350
1347 1351 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
1348 1352 mask = vap->va_mask;
1349 1353
1350 1354 rp = VTOR(vp);
1351 1355
1352 1356 /*
1353 1357 * Only need to flush pages if there are any pages and
1354 1358 * if the file is marked as dirty in some fashion. The
1355 1359 * file must be flushed so that we can accurately
1356 1360 * determine the size of the file and the cached data
1357 1361 * after the SETATTR returns. A file is considered to
1358 1362 * be dirty if it is either marked with RDIRTY, has
1359 1363 * outstanding i/o's active, or is mmap'd. In this
1360 1364 * last case, we can't tell whether there are dirty
1361 1365 * pages, so we flush just to be sure.
1362 1366 */
1363 1367 if (vn_has_cached_data(vp) &&
1364 1368 ((rp->r_flags & RDIRTY) ||
1365 1369 rp->r_count > 0 ||
1366 1370 rp->r_mapcnt > 0)) {
1367 1371 ASSERT(vp->v_type != VCHR);
1368 1372 error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr, NULL);
1369 1373 if (error && (error == ENOSPC || error == EDQUOT)) {
1370 1374 mutex_enter(&rp->r_statelock);
1371 1375 if (!rp->r_error)
1372 1376 rp->r_error = error;
1373 1377 mutex_exit(&rp->r_statelock);
1374 1378 }
1375 1379 }
1376 1380
1377 1381 args.object = *RTOFH3(rp);
1378 1382 /*
1379 1383 * If the intent is for the server to set the times,
1380 1384 * there is no point in have the mask indicating set mtime or
1381 1385 * atime, because the vap values may be junk, and so result
1382 1386 * in an overflow error. Remove these flags from the vap mask
1383 1387 * before calling in this case, and restore them afterwards.
1384 1388 */
1385 1389 if ((mask & (AT_ATIME | AT_MTIME)) && !(flags & ATTR_UTIME)) {
1386 1390 /* Use server times, so don't set the args time fields */
1387 1391 vap->va_mask &= ~(AT_ATIME | AT_MTIME);
1388 1392 error = vattr_to_sattr3(vap, &args.new_attributes);
1389 1393 vap->va_mask |= (mask & (AT_ATIME | AT_MTIME));
1390 1394 if (mask & AT_ATIME) {
1391 1395 args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
1392 1396 }
1393 1397 if (mask & AT_MTIME) {
1394 1398 args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
1395 1399 }
1396 1400 } else {
1397 1401 /* Either do not set times or use the client specified times */
1398 1402 error = vattr_to_sattr3(vap, &args.new_attributes);
1399 1403 }
1400 1404
1401 1405 if (error) {
1402 1406 /* req time field(s) overflow - return immediately */
1403 1407 return (error);
1404 1408 }
1405 1409
1406 1410 va.va_mask = AT_MODE | AT_CTIME;
1407 1411 error = nfs3getattr(vp, &va, cr);
1408 1412 if (error)
1409 1413 return (error);
1410 1414 omode = va.va_mode;
1411 1415
1412 1416 tryagain:
1413 1417 if (mask & AT_SIZE) {
1414 1418 args.guard.check = TRUE;
1415 1419 args.guard.obj_ctime.seconds = va.va_ctime.tv_sec;
1416 1420 args.guard.obj_ctime.nseconds = va.va_ctime.tv_nsec;
1417 1421 } else
1418 1422 args.guard.check = FALSE;
1419 1423
1420 1424 douprintf = 1;
1421 1425
1422 1426 t = gethrtime();
1423 1427
1424 1428 error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
1425 1429 xdr_SETATTR3args, (caddr_t)&args,
1426 1430 xdr_SETATTR3res, (caddr_t)&res, cr,
1427 1431 &douprintf, &res.status, 0, NULL);
1428 1432
1429 1433 /*
1430 1434 * Purge the access cache and ACL cache if changing either the
1431 1435 * owner of the file, the group owner, or the mode. These may
1432 1436 * change the access permissions of the file, so purge old
1433 1437 * information and start over again.
1434 1438 */
1435 1439 if (mask & (AT_UID | AT_GID | AT_MODE)) {
1436 1440 (void) nfs_access_purge_rp(rp);
1437 1441 if (rp->r_secattr != NULL) {
1438 1442 mutex_enter(&rp->r_statelock);
1439 1443 vsp = rp->r_secattr;
1440 1444 rp->r_secattr = NULL;
1441 1445 mutex_exit(&rp->r_statelock);
1442 1446 if (vsp != NULL)
1443 1447 nfs_acl_free(vsp);
1444 1448 }
1445 1449 }
1446 1450
1447 1451 if (error) {
1448 1452 PURGE_ATTRCACHE(vp);
1449 1453 return (error);
1450 1454 }
1451 1455
1452 1456 error = geterrno3(res.status);
1453 1457 if (!error) {
1454 1458 /*
1455 1459 * If changing the size of the file, invalidate
1456 1460 * any local cached data which is no longer part
1457 1461 * of the file. We also possibly invalidate the
1458 1462 * last page in the file. We could use
1459 1463 * pvn_vpzero(), but this would mark the page as
1460 1464 * modified and require it to be written back to
1461 1465 * the server for no particularly good reason.
1462 1466 * This way, if we access it, then we bring it
1463 1467 * back in. A read should be cheaper than a
1464 1468 * write.
1465 1469 */
1466 1470 if (mask & AT_SIZE) {
1467 1471 nfs_invalidate_pages(vp,
1468 1472 (vap->va_size & PAGEMASK), cr);
1469 1473 }
1470 1474 nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
1471 1475 /*
1472 1476 * Some servers will change the mode to clear the setuid
1473 1477 * and setgid bits when changing the uid or gid. The
1474 1478 * client needs to compensate appropriately.
1475 1479 */
1476 1480 if (mask & (AT_UID | AT_GID)) {
1477 1481 int terror;
1478 1482
1479 1483 va.va_mask = AT_MODE;
1480 1484 terror = nfs3getattr(vp, &va, cr);
1481 1485 if (!terror &&
1482 1486 (((mask & AT_MODE) && va.va_mode != vap->va_mode) ||
1483 1487 (!(mask & AT_MODE) && va.va_mode != omode))) {
1484 1488 va.va_mask = AT_MODE;
1485 1489 if (mask & AT_MODE)
1486 1490 va.va_mode = vap->va_mode;
1487 1491 else
1488 1492 va.va_mode = omode;
1489 1493 (void) nfs3setattr(vp, &va, 0, cr);
1490 1494 }
1491 1495 }
1492 1496 } else {
1493 1497 nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
1494 1498 /*
1495 1499 * If we got back a "not synchronized" error, then
1496 1500 * we need to retry with a new guard value. The
1497 1501 * guard value used is the change time. If the
1498 1502 * server returned post_op_attr, then we can just
1499 1503 * retry because we have the latest attributes.
1500 1504 * Otherwise, we issue a GETATTR to get the latest
1501 1505 * attributes and then retry. If we couldn't get
1502 1506 * the attributes this way either, then we give
1503 1507 * up because we can't complete the operation as
1504 1508 * required.
1505 1509 */
1506 1510 if (res.status == NFS3ERR_NOT_SYNC) {
1507 1511 va.va_mask = AT_CTIME;
1508 1512 if (nfs3getattr(vp, &va, cr) == 0)
1509 1513 goto tryagain;
1510 1514 }
1511 1515 PURGE_STALE_FH(error, vp, cr);
1512 1516 }
1513 1517
1514 1518 return (error);
1515 1519 }
1516 1520
1517 1521 static int
1518 1522 nfs3_accessx(void *vp, int mode, cred_t *cr)
1519 1523 {
1520 1524 ASSERT(nfs_zone() == VTOMI((vnode_t *)vp)->mi_zone);
1521 1525 return (nfs3_access(vp, mode, 0, cr, NULL));
1522 1526 }
1523 1527
1524 1528 /* ARGSUSED */
1525 1529 static int
1526 1530 nfs3_access(vnode_t *vp, int mode, int flags, cred_t *cr, caller_context_t *ct)
1527 1531 {
1528 1532 int error;
1529 1533 ACCESS3args args;
1530 1534 ACCESS3res res;
1531 1535 int douprintf;
1532 1536 uint32 acc;
1533 1537 rnode_t *rp;
1534 1538 cred_t *cred, *ncr, *ncrfree = NULL;
1535 1539 failinfo_t fi;
1536 1540 nfs_access_type_t cacc;
1537 1541 hrtime_t t;
1538 1542
1539 1543 acc = 0;
1540 1544 if (nfs_zone() != VTOMI(vp)->mi_zone)
1541 1545 return (EIO);
1542 1546 if (mode & VREAD)
1543 1547 acc |= ACCESS3_READ;
1544 1548 if (mode & VWRITE) {
1545 1549 if (vn_is_readonly(vp) && !IS_DEVVP(vp))
1546 1550 return (EROFS);
1547 1551 if (vp->v_type == VDIR)
1548 1552 acc |= ACCESS3_DELETE;
1549 1553 acc |= ACCESS3_MODIFY | ACCESS3_EXTEND;
1550 1554 }
1551 1555 if (mode & VEXEC) {
1552 1556 if (vp->v_type == VDIR)
1553 1557 acc |= ACCESS3_LOOKUP;
1554 1558 else
1555 1559 acc |= ACCESS3_EXECUTE;
1556 1560 }
1557 1561
1558 1562 rp = VTOR(vp);
1559 1563 args.object = *VTOFH3(vp);
1560 1564 if (vp->v_type == VDIR) {
1561 1565 args.access = ACCESS3_READ | ACCESS3_DELETE | ACCESS3_MODIFY |
1562 1566 ACCESS3_EXTEND | ACCESS3_LOOKUP;
1563 1567 } else {
1564 1568 args.access = ACCESS3_READ | ACCESS3_MODIFY | ACCESS3_EXTEND |
1565 1569 ACCESS3_EXECUTE;
1566 1570 }
1567 1571 fi.vp = vp;
1568 1572 fi.fhp = (caddr_t)&args.object;
1569 1573 fi.copyproc = nfs3copyfh;
1570 1574 fi.lookupproc = nfs3lookup;
1571 1575 fi.xattrdirproc = acl_getxattrdir3;
1572 1576
1573 1577 cred = cr;
1574 1578 /*
1575 1579 * ncr and ncrfree both initially
1576 1580 * point to the memory area returned
1577 1581 * by crnetadjust();
1578 1582 * ncrfree not NULL when exiting means
1579 1583 * that we need to release it
1580 1584 */
1581 1585 ncr = crnetadjust(cred);
1582 1586 ncrfree = ncr;
1583 1587 tryagain:
1584 1588 if (rp->r_acache != NULL) {
1585 1589 cacc = nfs_access_check(rp, acc, cred);
1586 1590 if (cacc == NFS_ACCESS_ALLOWED) {
1587 1591 if (ncrfree != NULL)
1588 1592 crfree(ncrfree);
1589 1593 return (0);
1590 1594 }
1591 1595 if (cacc == NFS_ACCESS_DENIED) {
1592 1596 /*
1593 1597 * If the cred can be adjusted, try again
1594 1598 * with the new cred.
1595 1599 */
1596 1600 if (ncr != NULL) {
1597 1601 cred = ncr;
1598 1602 ncr = NULL;
1599 1603 goto tryagain;
1600 1604 }
1601 1605 if (ncrfree != NULL)
1602 1606 crfree(ncrfree);
1603 1607 return (EACCES);
1604 1608 }
1605 1609 }
1606 1610
1607 1611 douprintf = 1;
1608 1612
1609 1613 t = gethrtime();
1610 1614
1611 1615 error = rfs3call(VTOMI(vp), NFSPROC3_ACCESS,
1612 1616 xdr_ACCESS3args, (caddr_t)&args,
1613 1617 xdr_ACCESS3res, (caddr_t)&res, cred,
1614 1618 &douprintf, &res.status, 0, &fi);
1615 1619
1616 1620 if (error) {
1617 1621 if (ncrfree != NULL)
1618 1622 crfree(ncrfree);
1619 1623 return (error);
1620 1624 }
1621 1625
1622 1626 error = geterrno3(res.status);
1623 1627 if (!error) {
1624 1628 nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
1625 1629 nfs_access_cache(rp, args.access, res.resok.access, cred);
1626 1630 /*
1627 1631 * we just cached results with cred; if cred is the
1628 1632 * adjusted credentials from crnetadjust, we do not want
1629 1633 * to release them before exiting: hence setting ncrfree
1630 1634 * to NULL
1631 1635 */
1632 1636 if (cred != cr)
1633 1637 ncrfree = NULL;
1634 1638 if ((acc & res.resok.access) != acc) {
1635 1639 /*
1636 1640 * If the cred can be adjusted, try again
1637 1641 * with the new cred.
1638 1642 */
1639 1643 if (ncr != NULL) {
1640 1644 cred = ncr;
1641 1645 ncr = NULL;
1642 1646 goto tryagain;
1643 1647 }
1644 1648 error = EACCES;
1645 1649 }
1646 1650 } else {
1647 1651 nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
1648 1652 PURGE_STALE_FH(error, vp, cr);
1649 1653 }
1650 1654
1651 1655 if (ncrfree != NULL)
1652 1656 crfree(ncrfree);
1653 1657
1654 1658 return (error);
1655 1659 }
1656 1660
1657 1661 static int nfs3_do_symlink_cache = 1;
1658 1662
1659 1663 /* ARGSUSED */
1660 1664 static int
1661 1665 nfs3_readlink(vnode_t *vp, struct uio *uiop, cred_t *cr, caller_context_t *ct)
1662 1666 {
1663 1667 int error;
1664 1668 READLINK3args args;
1665 1669 READLINK3res res;
1666 1670 nfspath3 resdata_backup;
1667 1671 rnode_t *rp;
1668 1672 int douprintf;
1669 1673 int len;
1670 1674 failinfo_t fi;
1671 1675 hrtime_t t;
1672 1676
1673 1677 /*
1674 1678 * Can't readlink anything other than a symbolic link.
1675 1679 */
1676 1680 if (vp->v_type != VLNK)
1677 1681 return (EINVAL);
1678 1682 if (nfs_zone() != VTOMI(vp)->mi_zone)
1679 1683 return (EIO);
1680 1684
1681 1685 rp = VTOR(vp);
1682 1686 if (nfs3_do_symlink_cache && rp->r_symlink.contents != NULL) {
1683 1687 error = nfs3_validate_caches(vp, cr);
1684 1688 if (error)
1685 1689 return (error);
1686 1690 mutex_enter(&rp->r_statelock);
1687 1691 if (rp->r_symlink.contents != NULL) {
1688 1692 error = uiomove(rp->r_symlink.contents,
1689 1693 rp->r_symlink.len, UIO_READ, uiop);
1690 1694 mutex_exit(&rp->r_statelock);
1691 1695 return (error);
1692 1696 }
1693 1697 mutex_exit(&rp->r_statelock);
1694 1698 }
1695 1699
1696 1700 args.symlink = *VTOFH3(vp);
1697 1701 fi.vp = vp;
1698 1702 fi.fhp = (caddr_t)&args.symlink;
1699 1703 fi.copyproc = nfs3copyfh;
1700 1704 fi.lookupproc = nfs3lookup;
1701 1705 fi.xattrdirproc = acl_getxattrdir3;
1702 1706
1703 1707 res.resok.data = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1704 1708
1705 1709 resdata_backup = res.resok.data;
1706 1710
1707 1711 douprintf = 1;
1708 1712
1709 1713 t = gethrtime();
1710 1714
1711 1715 error = rfs3call(VTOMI(vp), NFSPROC3_READLINK,
1712 1716 xdr_READLINK3args, (caddr_t)&args,
1713 1717 xdr_READLINK3res, (caddr_t)&res, cr,
1714 1718 &douprintf, &res.status, 0, &fi);
1715 1719
1716 1720 if (res.resok.data == nfs3nametoolong)
1717 1721 error = EINVAL;
1718 1722
1719 1723 if (error) {
1720 1724 kmem_free(resdata_backup, MAXPATHLEN);
1721 1725 return (error);
1722 1726 }
1723 1727
1724 1728 error = geterrno3(res.status);
1725 1729 if (!error) {
1726 1730 nfs3_cache_post_op_attr(vp, &res.resok.symlink_attributes, t,
1727 1731 cr);
1728 1732 len = strlen(res.resok.data);
1729 1733 error = uiomove(res.resok.data, len, UIO_READ, uiop);
1730 1734 if (nfs3_do_symlink_cache && rp->r_symlink.contents == NULL) {
1731 1735 mutex_enter(&rp->r_statelock);
1732 1736 if (rp->r_symlink.contents == NULL) {
1733 1737 rp->r_symlink.contents = res.resok.data;
1734 1738 rp->r_symlink.len = len;
1735 1739 rp->r_symlink.size = MAXPATHLEN;
1736 1740 mutex_exit(&rp->r_statelock);
1737 1741 } else {
1738 1742 mutex_exit(&rp->r_statelock);
1739 1743
1740 1744 kmem_free((void *)res.resok.data, MAXPATHLEN);
1741 1745 }
1742 1746 } else {
1743 1747 kmem_free((void *)res.resok.data, MAXPATHLEN);
1744 1748 }
1745 1749 } else {
1746 1750 nfs3_cache_post_op_attr(vp,
1747 1751 &res.resfail.symlink_attributes, t, cr);
1748 1752 PURGE_STALE_FH(error, vp, cr);
1749 1753
1750 1754 kmem_free((void *)res.resok.data, MAXPATHLEN);
1751 1755
1752 1756 }
1753 1757
1754 1758 /*
1755 1759 * The over the wire error for attempting to readlink something
1756 1760 * other than a symbolic link is ENXIO. However, we need to
1757 1761 * return EINVAL instead of ENXIO, so we map it here.
1758 1762 */
1759 1763 return (error == ENXIO ? EINVAL : error);
1760 1764 }
1761 1765
1762 1766 /*
1763 1767 * Flush local dirty pages to stable storage on the server.
1764 1768 *
1765 1769 * If FNODSYNC is specified, then there is nothing to do because
1766 1770 * metadata changes are not cached on the client before being
1767 1771 * sent to the server.
1768 1772 */
1769 1773 /* ARGSUSED */
1770 1774 static int
1771 1775 nfs3_fsync(vnode_t *vp, int syncflag, cred_t *cr, caller_context_t *ct)
1772 1776 {
1773 1777 int error;
1774 1778
1775 1779 if ((syncflag & FNODSYNC) || IS_SWAPVP(vp))
1776 1780 return (0);
1777 1781 if (nfs_zone() != VTOMI(vp)->mi_zone)
1778 1782 return (EIO);
1779 1783
1780 1784 error = nfs3_putpage_commit(vp, (offset_t)0, 0, cr);
1781 1785 if (!error)
1782 1786 error = VTOR(vp)->r_error;
1783 1787 return (error);
1784 1788 }
1785 1789
1786 1790 /*
1787 1791 * Weirdness: if the file was removed or the target of a rename
1788 1792 * operation while it was open, it got renamed instead. Here we
1789 1793 * remove the renamed file.
1790 1794 */
1791 1795 /* ARGSUSED */
1792 1796 static void
1793 1797 nfs3_inactive(vnode_t *vp, cred_t *cr, caller_context_t *ct)
1794 1798 {
1795 1799 rnode_t *rp;
1796 1800
1797 1801 ASSERT(vp != DNLC_NO_VNODE);
1798 1802
1799 1803 /*
1800 1804 * If this is coming from the wrong zone, we let someone in the right
1801 1805 * zone take care of it asynchronously. We can get here due to
1802 1806 * VN_RELE() being called from pageout() or fsflush(). This call may
1803 1807 * potentially turn into an expensive no-op if, for instance, v_count
1804 1808 * gets incremented in the meantime, but it's still correct.
1805 1809 */
1806 1810 if (nfs_zone() != VTOMI(vp)->mi_zone) {
1807 1811 nfs_async_inactive(vp, cr, nfs3_inactive);
1808 1812 return;
1809 1813 }
1810 1814
1811 1815 rp = VTOR(vp);
1812 1816 redo:
1813 1817 if (rp->r_unldvp != NULL) {
1814 1818 /*
1815 1819 * Save the vnode pointer for the directory where the
1816 1820 * unlinked-open file got renamed, then set it to NULL
1817 1821 * to prevent another thread from getting here before
1818 1822 * we're done with the remove. While we have the
1819 1823 * statelock, make local copies of the pertinent rnode
1820 1824 * fields. If we weren't to do this in an atomic way, the
1821 1825 * the unl* fields could become inconsistent with respect
1822 1826 * to each other due to a race condition between this
1823 1827 * code and nfs_remove(). See bug report 1034328.
1824 1828 */
1825 1829 mutex_enter(&rp->r_statelock);
1826 1830 if (rp->r_unldvp != NULL) {
1827 1831 vnode_t *unldvp;
1828 1832 char *unlname;
1829 1833 cred_t *unlcred;
1830 1834 REMOVE3args args;
1831 1835 REMOVE3res res;
1832 1836 int douprintf;
1833 1837 int error;
1834 1838 hrtime_t t;
1835 1839
1836 1840 unldvp = rp->r_unldvp;
1837 1841 rp->r_unldvp = NULL;
1838 1842 unlname = rp->r_unlname;
1839 1843 rp->r_unlname = NULL;
1840 1844 unlcred = rp->r_unlcred;
1841 1845 rp->r_unlcred = NULL;
1842 1846 mutex_exit(&rp->r_statelock);
1843 1847
1844 1848 /*
1845 1849 * If there are any dirty pages left, then flush
1846 1850 * them. This is unfortunate because they just
1847 1851 * may get thrown away during the remove operation,
1848 1852 * but we have to do this for correctness.
1849 1853 */
1850 1854 if (vn_has_cached_data(vp) &&
1851 1855 ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
1852 1856 ASSERT(vp->v_type != VCHR);
1853 1857 error = nfs3_putpage(vp, (offset_t)0, 0, 0,
1854 1858 cr, ct);
1855 1859 if (error) {
1856 1860 mutex_enter(&rp->r_statelock);
1857 1861 if (!rp->r_error)
1858 1862 rp->r_error = error;
1859 1863 mutex_exit(&rp->r_statelock);
1860 1864 }
1861 1865 }
1862 1866
1863 1867 /*
1864 1868 * Do the remove operation on the renamed file
1865 1869 */
1866 1870 setdiropargs3(&args.object, unlname, unldvp);
1867 1871
1868 1872 douprintf = 1;
1869 1873
1870 1874 t = gethrtime();
1871 1875
1872 1876 error = rfs3call(VTOMI(unldvp), NFSPROC3_REMOVE,
1873 1877 xdr_diropargs3, (caddr_t)&args,
1874 1878 xdr_REMOVE3res, (caddr_t)&res, unlcred,
1875 1879 &douprintf, &res.status, 0, NULL);
1876 1880
1877 1881 if (error) {
1878 1882 PURGE_ATTRCACHE(unldvp);
1879 1883 } else {
1880 1884 error = geterrno3(res.status);
1881 1885 if (!error) {
1882 1886 nfs3_cache_wcc_data(unldvp,
1883 1887 &res.resok.dir_wcc, t, cr);
1884 1888 if (HAVE_RDDIR_CACHE(VTOR(unldvp)))
1885 1889 nfs_purge_rddir_cache(unldvp);
1886 1890 } else {
1887 1891 nfs3_cache_wcc_data(unldvp,
1888 1892 &res.resfail.dir_wcc, t, cr);
1889 1893 PURGE_STALE_FH(error, unldvp, cr);
1890 1894 }
1891 1895 }
1892 1896
1893 1897 /*
1894 1898 * Release stuff held for the remove
1895 1899 */
1896 1900 VN_RELE(unldvp);
1897 1901 kmem_free(unlname, MAXNAMELEN);
1898 1902 crfree(unlcred);
1899 1903 goto redo;
1900 1904 }
1901 1905 mutex_exit(&rp->r_statelock);
1902 1906 }
1903 1907
1904 1908 rp_addfree(rp, cr);
1905 1909 }
1906 1910
1907 1911 /*
1908 1912 * Remote file system operations having to do with directory manipulation.
1909 1913 */
1910 1914
1911 1915 /* ARGSUSED */
1912 1916 static int
1913 1917 nfs3_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1914 1918 int flags, vnode_t *rdir, cred_t *cr, caller_context_t *ct,
1915 1919 int *direntflags, pathname_t *realpnp)
1916 1920 {
1917 1921 int error;
1918 1922 vnode_t *vp;
1919 1923 vnode_t *avp = NULL;
1920 1924 rnode_t *drp;
1921 1925
1922 1926 if (nfs_zone() != VTOMI(dvp)->mi_zone)
1923 1927 return (EPERM);
1924 1928
1925 1929 drp = VTOR(dvp);
1926 1930
1927 1931 /*
1928 1932 * Are we looking up extended attributes? If so, "dvp" is
1929 1933 * the file or directory for which we want attributes, and
1930 1934 * we need a lookup of the hidden attribute directory
1931 1935 * before we lookup the rest of the path.
1932 1936 */
1933 1937 if (flags & LOOKUP_XATTR) {
1934 1938 bool_t cflag = ((flags & CREATE_XATTR_DIR) != 0);
1935 1939 mntinfo_t *mi;
1936 1940
1937 1941 mi = VTOMI(dvp);
1938 1942 if (!(mi->mi_flags & MI_EXTATTR))
1939 1943 return (EINVAL);
1940 1944
1941 1945 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp)))
1942 1946 return (EINTR);
1943 1947
1944 1948 (void) nfs3lookup_dnlc(dvp, XATTR_DIR_NAME, &avp, cr);
1945 1949 if (avp == NULL)
1946 1950 error = acl_getxattrdir3(dvp, &avp, cflag, cr, 0);
1947 1951 else
1948 1952 error = 0;
1949 1953
1950 1954 nfs_rw_exit(&drp->r_rwlock);
1951 1955
1952 1956 if (error) {
1953 1957 if (mi->mi_flags & MI_EXTATTR)
1954 1958 return (error);
1955 1959 return (EINVAL);
1956 1960 }
1957 1961 dvp = avp;
1958 1962 drp = VTOR(dvp);
1959 1963 }
1960 1964
1961 1965 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_READER, INTR(dvp))) {
1962 1966 error = EINTR;
1963 1967 goto out;
1964 1968 }
1965 1969
1966 1970 error = nfs3lookup(dvp, nm, vpp, pnp, flags, rdir, cr, 0);
1967 1971
1968 1972 nfs_rw_exit(&drp->r_rwlock);
1969 1973
1970 1974 /*
1971 1975 * If vnode is a device, create special vnode.
1972 1976 */
1973 1977 if (!error && IS_DEVVP(*vpp)) {
1974 1978 vp = *vpp;
1975 1979 *vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
1976 1980 VN_RELE(vp);
1977 1981 }
1978 1982
1979 1983 out:
1980 1984 if (avp != NULL)
1981 1985 VN_RELE(avp);
1982 1986
1983 1987 return (error);
1984 1988 }
1985 1989
1986 1990 static int nfs3_lookup_neg_cache = 1;
1987 1991
1988 1992 #ifdef DEBUG
1989 1993 static int nfs3_lookup_dnlc_hits = 0;
1990 1994 static int nfs3_lookup_dnlc_misses = 0;
1991 1995 static int nfs3_lookup_dnlc_neg_hits = 0;
1992 1996 static int nfs3_lookup_dnlc_disappears = 0;
1993 1997 static int nfs3_lookup_dnlc_lookups = 0;
1994 1998 #endif
1995 1999
1996 2000 /* ARGSUSED */
1997 2001 int
1998 2002 nfs3lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct pathname *pnp,
1999 2003 int flags, vnode_t *rdir, cred_t *cr, int rfscall_flags)
2000 2004 {
2001 2005 int error;
2002 2006 rnode_t *drp;
2003 2007
2004 2008 ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2005 2009 /*
2006 2010 * If lookup is for "", just return dvp. Don't need
2007 2011 * to send it over the wire, look it up in the dnlc,
2008 2012 * or perform any access checks.
2009 2013 */
2010 2014 if (*nm == '\0') {
2011 2015 VN_HOLD(dvp);
2012 2016 *vpp = dvp;
2013 2017 return (0);
2014 2018 }
2015 2019
2016 2020 /*
2017 2021 * Can't do lookups in non-directories.
2018 2022 */
2019 2023 if (dvp->v_type != VDIR)
2020 2024 return (ENOTDIR);
2021 2025
2022 2026 /*
2023 2027 * If we're called with RFSCALL_SOFT, it's important that
2024 2028 * the only rfscall is one we make directly; if we permit
2025 2029 * an access call because we're looking up "." or validating
2026 2030 * a dnlc hit, we'll deadlock because that rfscall will not
2027 2031 * have the RFSCALL_SOFT set.
2028 2032 */
2029 2033 if (rfscall_flags & RFSCALL_SOFT)
2030 2034 goto callit;
2031 2035
2032 2036 /*
2033 2037 * If lookup is for ".", just return dvp. Don't need
2034 2038 * to send it over the wire or look it up in the dnlc,
2035 2039 * just need to check access.
2036 2040 */
2037 2041 if (strcmp(nm, ".") == 0) {
2038 2042 error = nfs3_access(dvp, VEXEC, 0, cr, NULL);
2039 2043 if (error)
2040 2044 return (error);
2041 2045 VN_HOLD(dvp);
2042 2046 *vpp = dvp;
2043 2047 return (0);
2044 2048 }
2045 2049
2046 2050 drp = VTOR(dvp);
2047 2051 if (!(drp->r_flags & RLOOKUP)) {
2048 2052 mutex_enter(&drp->r_statelock);
2049 2053 drp->r_flags |= RLOOKUP;
2050 2054 mutex_exit(&drp->r_statelock);
2051 2055 }
2052 2056
2053 2057 /*
2054 2058 * Lookup this name in the DNLC. If there was a valid entry,
2055 2059 * then return the results of the lookup.
2056 2060 */
2057 2061 error = nfs3lookup_dnlc(dvp, nm, vpp, cr);
2058 2062 if (error || *vpp != NULL)
2059 2063 return (error);
2060 2064
2061 2065 callit:
2062 2066 error = nfs3lookup_otw(dvp, nm, vpp, cr, rfscall_flags);
2063 2067
2064 2068 return (error);
2065 2069 }
2066 2070
2067 2071 static int
2068 2072 nfs3lookup_dnlc(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr)
2069 2073 {
2070 2074 int error;
2071 2075 vnode_t *vp;
2072 2076
2073 2077 ASSERT(*nm != '\0');
2074 2078 ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2075 2079 /*
2076 2080 * Lookup this name in the DNLC. If successful, then validate
2077 2081 * the caches and then recheck the DNLC. The DNLC is rechecked
2078 2082 * just in case this entry got invalidated during the call
2079 2083 * to nfs3_validate_caches.
2080 2084 *
2081 2085 * An assumption is being made that it is safe to say that a
2082 2086 * file exists which may not on the server. Any operations to
2083 2087 * the server will fail with ESTALE.
2084 2088 */
2085 2089 #ifdef DEBUG
2086 2090 nfs3_lookup_dnlc_lookups++;
2087 2091 #endif
2088 2092 vp = dnlc_lookup(dvp, nm);
2089 2093 if (vp != NULL) {
2090 2094 VN_RELE(vp);
2091 2095 if (vp == DNLC_NO_VNODE && !vn_is_readonly(dvp)) {
2092 2096 PURGE_ATTRCACHE(dvp);
2093 2097 }
2094 2098 error = nfs3_validate_caches(dvp, cr);
2095 2099 if (error)
2096 2100 return (error);
2097 2101 vp = dnlc_lookup(dvp, nm);
2098 2102 if (vp != NULL) {
2099 2103 error = nfs3_access(dvp, VEXEC, 0, cr, NULL);
2100 2104 if (error) {
2101 2105 VN_RELE(vp);
2102 2106 return (error);
2103 2107 }
2104 2108 if (vp == DNLC_NO_VNODE) {
2105 2109 VN_RELE(vp);
2106 2110 #ifdef DEBUG
2107 2111 nfs3_lookup_dnlc_neg_hits++;
2108 2112 #endif
2109 2113 return (ENOENT);
2110 2114 }
2111 2115 *vpp = vp;
2112 2116 #ifdef DEBUG
2113 2117 nfs3_lookup_dnlc_hits++;
2114 2118 #endif
2115 2119 return (0);
2116 2120 }
2117 2121 #ifdef DEBUG
2118 2122 nfs3_lookup_dnlc_disappears++;
2119 2123 #endif
2120 2124 }
2121 2125 #ifdef DEBUG
2122 2126 else
2123 2127 nfs3_lookup_dnlc_misses++;
2124 2128 #endif
2125 2129
2126 2130 *vpp = NULL;
2127 2131
2128 2132 return (0);
2129 2133 }
2130 2134
2131 2135 static int
2132 2136 nfs3lookup_otw(vnode_t *dvp, char *nm, vnode_t **vpp, cred_t *cr,
2133 2137 int rfscall_flags)
2134 2138 {
2135 2139 int error;
2136 2140 LOOKUP3args args;
2137 2141 LOOKUP3vres res;
2138 2142 int douprintf;
2139 2143 struct vattr vattr;
2140 2144 struct vattr dvattr;
2141 2145 vnode_t *vp;
2142 2146 failinfo_t fi;
2143 2147 hrtime_t t;
2144 2148
2145 2149 ASSERT(*nm != '\0');
2146 2150 ASSERT(dvp->v_type == VDIR);
2147 2151 ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2148 2152
2149 2153 setdiropargs3(&args.what, nm, dvp);
2150 2154
2151 2155 fi.vp = dvp;
2152 2156 fi.fhp = (caddr_t)&args.what.dir;
2153 2157 fi.copyproc = nfs3copyfh;
2154 2158 fi.lookupproc = nfs3lookup;
2155 2159 fi.xattrdirproc = acl_getxattrdir3;
2156 2160 res.obj_attributes.fres.vp = dvp;
2157 2161 res.obj_attributes.fres.vap = &vattr;
2158 2162 res.dir_attributes.fres.vp = dvp;
2159 2163 res.dir_attributes.fres.vap = &dvattr;
2160 2164
2161 2165 douprintf = 1;
2162 2166
2163 2167 t = gethrtime();
2164 2168
2165 2169 error = rfs3call(VTOMI(dvp), NFSPROC3_LOOKUP,
2166 2170 xdr_diropargs3, (caddr_t)&args,
2167 2171 xdr_LOOKUP3vres, (caddr_t)&res, cr,
2168 2172 &douprintf, &res.status, rfscall_flags, &fi);
2169 2173
2170 2174 if (error)
2171 2175 return (error);
2172 2176
2173 2177 nfs3_cache_post_op_vattr(dvp, &res.dir_attributes, t, cr);
2174 2178
2175 2179 error = geterrno3(res.status);
2176 2180 if (error) {
2177 2181 PURGE_STALE_FH(error, dvp, cr);
2178 2182 if (error == ENOENT && nfs3_lookup_neg_cache)
2179 2183 dnlc_enter(dvp, nm, DNLC_NO_VNODE);
2180 2184 return (error);
2181 2185 }
2182 2186
2183 2187 if (res.obj_attributes.attributes) {
2184 2188 vp = makenfs3node_va(&res.object, res.obj_attributes.fres.vap,
2185 2189 dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2186 2190 } else {
2187 2191 vp = makenfs3node_va(&res.object, NULL,
2188 2192 dvp->v_vfsp, t, cr, VTOR(dvp)->r_path, nm);
2189 2193 if (vp->v_type == VNON) {
2190 2194 vattr.va_mask = AT_TYPE;
2191 2195 error = nfs3getattr(vp, &vattr, cr);
2192 2196 if (error) {
2193 2197 VN_RELE(vp);
2194 2198 return (error);
2195 2199 }
2196 2200 vp->v_type = vattr.va_type;
2197 2201 }
2198 2202 }
2199 2203
2200 2204 if (!(rfscall_flags & RFSCALL_SOFT))
2201 2205 dnlc_update(dvp, nm, vp);
2202 2206
2203 2207 *vpp = vp;
2204 2208
2205 2209 return (error);
2206 2210 }
2207 2211
2208 2212 #ifdef DEBUG
2209 2213 static int nfs3_create_misses = 0;
2210 2214 #endif
2211 2215
2212 2216 /* ARGSUSED */
2213 2217 static int
2214 2218 nfs3_create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2215 2219 int mode, vnode_t **vpp, cred_t *cr, int lfaware, caller_context_t *ct,
2216 2220 vsecattr_t *vsecp)
2217 2221 {
2218 2222 int error;
2219 2223 vnode_t *vp;
2220 2224 rnode_t *rp;
2221 2225 struct vattr vattr;
2222 2226 rnode_t *drp;
2223 2227 vnode_t *tempvp;
2224 2228
2225 2229 drp = VTOR(dvp);
2226 2230 if (nfs_zone() != VTOMI(dvp)->mi_zone)
2227 2231 return (EPERM);
2228 2232 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2229 2233 return (EINTR);
2230 2234
2231 2235 top:
2232 2236 /*
2233 2237 * We make a copy of the attributes because the caller does not
2234 2238 * expect us to change what va points to.
2235 2239 */
2236 2240 vattr = *va;
2237 2241
2238 2242 /*
2239 2243 * If the pathname is "", just use dvp. Don't need
2240 2244 * to send it over the wire, look it up in the dnlc,
2241 2245 * or perform any access checks.
2242 2246 */
2243 2247 if (*nm == '\0') {
2244 2248 error = 0;
2245 2249 VN_HOLD(dvp);
2246 2250 vp = dvp;
2247 2251 /*
2248 2252 * If the pathname is ".", just use dvp. Don't need
2249 2253 * to send it over the wire or look it up in the dnlc,
2250 2254 * just need to check access.
2251 2255 */
2252 2256 } else if (strcmp(nm, ".") == 0) {
2253 2257 error = nfs3_access(dvp, VEXEC, 0, cr, ct);
2254 2258 if (error) {
2255 2259 nfs_rw_exit(&drp->r_rwlock);
2256 2260 return (error);
2257 2261 }
2258 2262 VN_HOLD(dvp);
2259 2263 vp = dvp;
2260 2264 /*
2261 2265 * We need to go over the wire, just to be sure whether the
2262 2266 * file exists or not. Using the DNLC can be dangerous in
2263 2267 * this case when making a decision regarding existence.
2264 2268 */
2265 2269 } else {
2266 2270 error = nfs3lookup_otw(dvp, nm, &vp, cr, 0);
2267 2271 }
2268 2272 if (!error) {
2269 2273 if (exclusive == EXCL)
2270 2274 error = EEXIST;
2271 2275 else if (vp->v_type == VDIR && (mode & VWRITE))
2272 2276 error = EISDIR;
2273 2277 else {
2274 2278 /*
2275 2279 * If vnode is a device, create special vnode.
2276 2280 */
2277 2281 if (IS_DEVVP(vp)) {
2278 2282 tempvp = vp;
2279 2283 vp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2280 2284 VN_RELE(tempvp);
2281 2285 }
2282 2286 if (!(error = VOP_ACCESS(vp, mode, 0, cr, ct))) {
2283 2287 if ((vattr.va_mask & AT_SIZE) &&
2284 2288 vp->v_type == VREG) {
2285 2289 rp = VTOR(vp);
2286 2290 /*
2287 2291 * Check here for large file handled
2288 2292 * by LF-unaware process (as
2289 2293 * ufs_create() does)
2290 2294 */
|
↓ open down ↓ |
2250 lines elided |
↑ open up ↑ |
2291 2295 if (!(lfaware & FOFFMAX)) {
2292 2296 mutex_enter(&rp->r_statelock);
2293 2297 if (rp->r_size > MAXOFF32_T)
2294 2298 error = EOVERFLOW;
2295 2299 mutex_exit(&rp->r_statelock);
2296 2300 }
2297 2301 if (!error) {
2298 2302 vattr.va_mask = AT_SIZE;
2299 2303 error = nfs3setattr(vp,
2300 2304 &vattr, 0, cr);
2305 +
2306 + /*
2307 + * Existing file was truncated;
2308 + * emit a create event.
2309 + */
2310 + vnevent_create(vp, ct);
2301 2311 }
2302 2312 }
2303 2313 }
2304 2314 }
2305 2315 nfs_rw_exit(&drp->r_rwlock);
2306 2316 if (error) {
2307 2317 VN_RELE(vp);
2308 2318 } else {
2309 - /*
2310 - * existing file got truncated, notify.
2311 - */
2312 - vnevent_create(vp, ct);
2313 2319 *vpp = vp;
2314 2320 }
2321 +
2315 2322 return (error);
2316 2323 }
2317 2324
2318 2325 dnlc_remove(dvp, nm);
2319 2326
2320 2327 /*
2321 2328 * Decide what the group-id of the created file should be.
2322 2329 * Set it in attribute list as advisory...
2323 2330 */
2324 2331 error = setdirgid(dvp, &vattr.va_gid, cr);
2325 2332 if (error) {
2326 2333 nfs_rw_exit(&drp->r_rwlock);
2327 2334 return (error);
2328 2335 }
2329 2336 vattr.va_mask |= AT_GID;
2330 2337
2331 2338 ASSERT(vattr.va_mask & AT_TYPE);
2332 2339 if (vattr.va_type == VREG) {
2333 2340 ASSERT(vattr.va_mask & AT_MODE);
2334 2341 if (MANDMODE(vattr.va_mode)) {
2335 2342 nfs_rw_exit(&drp->r_rwlock);
2336 2343 return (EACCES);
2337 2344 }
2338 2345 error = nfs3create(dvp, nm, &vattr, exclusive, mode, vpp, cr,
2339 2346 lfaware);
2340 2347 /*
2341 2348 * If this is not an exclusive create, then the CREATE
2342 2349 * request will be made with the GUARDED mode set. This
2343 2350 * means that the server will return EEXIST if the file
2344 2351 * exists. The file could exist because of a retransmitted
2345 2352 * request. In this case, we recover by starting over and
2346 2353 * checking to see whether the file exists. This second
2347 2354 * time through it should and a CREATE request will not be
2348 2355 * sent.
2349 2356 *
2350 2357 * This handles the problem of a dangling CREATE request
2351 2358 * which contains attributes which indicate that the file
2352 2359 * should be truncated. This retransmitted request could
2353 2360 * possibly truncate valid data in the file if not caught
2354 2361 * by the duplicate request mechanism on the server or if
2355 2362 * not caught by other means. The scenario is:
2356 2363 *
2357 2364 * Client transmits CREATE request with size = 0
2358 2365 * Client times out, retransmits request.
2359 2366 * Response to the first request arrives from the server
2360 2367 * and the client proceeds on.
2361 2368 * Client writes data to the file.
2362 2369 * The server now processes retransmitted CREATE request
2363 2370 * and truncates file.
2364 2371 *
2365 2372 * The use of the GUARDED CREATE request prevents this from
2366 2373 * happening because the retransmitted CREATE would fail
2367 2374 * with EEXIST and would not truncate the file.
2368 2375 */
2369 2376 if (error == EEXIST && exclusive == NONEXCL) {
2370 2377 #ifdef DEBUG
2371 2378 nfs3_create_misses++;
2372 2379 #endif
2373 2380 goto top;
2374 2381 }
2375 2382 nfs_rw_exit(&drp->r_rwlock);
2376 2383 return (error);
2377 2384 }
2378 2385 error = nfs3mknod(dvp, nm, &vattr, exclusive, mode, vpp, cr);
2379 2386 nfs_rw_exit(&drp->r_rwlock);
2380 2387 return (error);
2381 2388 }
2382 2389
2383 2390 /* ARGSUSED */
2384 2391 static int
2385 2392 nfs3create(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2386 2393 int mode, vnode_t **vpp, cred_t *cr, int lfaware)
2387 2394 {
2388 2395 int error;
2389 2396 CREATE3args args;
2390 2397 CREATE3res res;
2391 2398 int douprintf;
2392 2399 vnode_t *vp;
2393 2400 struct vattr vattr;
2394 2401 nfstime3 *verfp;
2395 2402 rnode_t *rp;
2396 2403 timestruc_t now;
2397 2404 hrtime_t t;
2398 2405
2399 2406 ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2400 2407 setdiropargs3(&args.where, nm, dvp);
2401 2408 if (exclusive == EXCL) {
2402 2409 args.how.mode = EXCLUSIVE;
2403 2410 /*
2404 2411 * Construct the create verifier. This verifier needs
2405 2412 * to be unique between different clients. It also needs
2406 2413 * to vary for each exclusive create request generated
2407 2414 * from the client to the server.
2408 2415 *
2409 2416 * The first attempt is made to use the hostid and a
2410 2417 * unique number on the client. If the hostid has not
2411 2418 * been set, the high resolution time that the exclusive
2412 2419 * create request is being made is used. This will work
2413 2420 * unless two different clients, both with the hostid
2414 2421 * not set, attempt an exclusive create request on the
2415 2422 * same file, at exactly the same clock time. The
2416 2423 * chances of this happening seem small enough to be
2417 2424 * reasonable.
2418 2425 */
2419 2426 verfp = (nfstime3 *)&args.how.createhow3_u.verf;
2420 2427 verfp->seconds = zone_get_hostid(NULL);
2421 2428 if (verfp->seconds != 0)
2422 2429 verfp->nseconds = newnum();
2423 2430 else {
2424 2431 gethrestime(&now);
2425 2432 verfp->seconds = now.tv_sec;
2426 2433 verfp->nseconds = now.tv_nsec;
2427 2434 }
2428 2435 /*
2429 2436 * Since the server will use this value for the mtime,
2430 2437 * make sure that it can't overflow. Zero out the MSB.
2431 2438 * The actual value does not matter here, only its uniqeness.
2432 2439 */
2433 2440 verfp->seconds %= INT32_MAX;
2434 2441 } else {
2435 2442 /*
2436 2443 * Issue the non-exclusive create in guarded mode. This
2437 2444 * may result in some false EEXIST responses for
2438 2445 * retransmitted requests, but these will be handled at
2439 2446 * a higher level. By using GUARDED, duplicate requests
2440 2447 * to do file truncation and possible access problems
2441 2448 * can be avoided.
2442 2449 */
2443 2450 args.how.mode = GUARDED;
2444 2451 error = vattr_to_sattr3(va,
2445 2452 &args.how.createhow3_u.obj_attributes);
2446 2453 if (error) {
2447 2454 /* req time field(s) overflow - return immediately */
2448 2455 return (error);
2449 2456 }
2450 2457 }
2451 2458
2452 2459 douprintf = 1;
2453 2460
2454 2461 t = gethrtime();
2455 2462
2456 2463 error = rfs3call(VTOMI(dvp), NFSPROC3_CREATE,
2457 2464 xdr_CREATE3args, (caddr_t)&args,
2458 2465 xdr_CREATE3res, (caddr_t)&res, cr,
2459 2466 &douprintf, &res.status, 0, NULL);
2460 2467
2461 2468 if (error) {
2462 2469 PURGE_ATTRCACHE(dvp);
2463 2470 return (error);
2464 2471 }
2465 2472
2466 2473 error = geterrno3(res.status);
2467 2474 if (!error) {
2468 2475 nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2469 2476 if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2470 2477 nfs_purge_rddir_cache(dvp);
2471 2478
2472 2479 /*
2473 2480 * On exclusive create the times need to be explicitly
2474 2481 * set to clear any potential verifier that may be stored
2475 2482 * in one of these fields (see comment below). This
2476 2483 * is done here to cover the case where no post op attrs
2477 2484 * were returned or a 'invalid' time was returned in
2478 2485 * the attributes.
2479 2486 */
2480 2487 if (exclusive == EXCL)
2481 2488 va->va_mask |= (AT_MTIME | AT_ATIME);
2482 2489
2483 2490 if (!res.resok.obj.handle_follows) {
2484 2491 error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2485 2492 if (error)
2486 2493 return (error);
2487 2494 } else {
2488 2495 if (res.resok.obj_attributes.attributes) {
2489 2496 vp = makenfs3node(&res.resok.obj.handle,
2490 2497 &res.resok.obj_attributes.attr,
2491 2498 dvp->v_vfsp, t, cr, NULL, NULL);
2492 2499 } else {
2493 2500 vp = makenfs3node(&res.resok.obj.handle, NULL,
2494 2501 dvp->v_vfsp, t, cr, NULL, NULL);
2495 2502
2496 2503 /*
2497 2504 * On an exclusive create, it is possible
2498 2505 * that attributes were returned but those
2499 2506 * postop attributes failed to decode
2500 2507 * properly. If this is the case,
2501 2508 * then most likely the atime or mtime
2502 2509 * were invalid for our client; this
2503 2510 * is caused by the server storing the
2504 2511 * create verifier in one of the time
2505 2512 * fields(most likely mtime).
2506 2513 * So... we are going to setattr just the
2507 2514 * atime/mtime to clear things up.
2508 2515 */
2509 2516 if (exclusive == EXCL) {
2510 2517 if (error =
2511 2518 nfs3excl_create_settimes(vp,
2512 2519 va, cr)) {
2513 2520 /*
2514 2521 * Setting the times failed.
2515 2522 * Remove the file and return
2516 2523 * the error.
2517 2524 */
2518 2525 VN_RELE(vp);
2519 2526 (void) nfs3_remove(dvp,
2520 2527 nm, cr, NULL, 0);
2521 2528 return (error);
2522 2529 }
2523 2530 }
2524 2531
2525 2532 /*
2526 2533 * This handles the non-exclusive case
2527 2534 * and the exclusive case where no post op
2528 2535 * attrs were returned.
2529 2536 */
2530 2537 if (vp->v_type == VNON) {
2531 2538 vattr.va_mask = AT_TYPE;
2532 2539 error = nfs3getattr(vp, &vattr, cr);
2533 2540 if (error) {
2534 2541 VN_RELE(vp);
2535 2542 return (error);
2536 2543 }
2537 2544 vp->v_type = vattr.va_type;
2538 2545 }
2539 2546 }
2540 2547 dnlc_update(dvp, nm, vp);
2541 2548 }
2542 2549
2543 2550 rp = VTOR(vp);
2544 2551
2545 2552 /*
2546 2553 * Check here for large file handled by
2547 2554 * LF-unaware process (as ufs_create() does)
2548 2555 */
2549 2556 if ((va->va_mask & AT_SIZE) && vp->v_type == VREG &&
2550 2557 !(lfaware & FOFFMAX)) {
2551 2558 mutex_enter(&rp->r_statelock);
2552 2559 if (rp->r_size > MAXOFF32_T) {
2553 2560 mutex_exit(&rp->r_statelock);
2554 2561 VN_RELE(vp);
2555 2562 return (EOVERFLOW);
2556 2563 }
2557 2564 mutex_exit(&rp->r_statelock);
2558 2565 }
2559 2566
2560 2567 if (exclusive == EXCL &&
2561 2568 (va->va_mask & ~(AT_GID | AT_SIZE))) {
2562 2569 /*
2563 2570 * If doing an exclusive create, then generate
2564 2571 * a SETATTR to set the initial attributes.
2565 2572 * Try to set the mtime and the atime to the
2566 2573 * server's current time. It is somewhat
2567 2574 * expected that these fields will be used to
2568 2575 * store the exclusive create cookie. If not,
2569 2576 * server implementors will need to know that
2570 2577 * a SETATTR will follow an exclusive create
2571 2578 * and the cookie should be destroyed if
2572 2579 * appropriate. This work may have been done
2573 2580 * earlier in this function if post op attrs
2574 2581 * were not available.
2575 2582 *
2576 2583 * The AT_GID and AT_SIZE bits are turned off
2577 2584 * so that the SETATTR request will not attempt
2578 2585 * to process these. The gid will be set
2579 2586 * separately if appropriate. The size is turned
2580 2587 * off because it is assumed that a new file will
2581 2588 * be created empty and if the file wasn't empty,
2582 2589 * then the exclusive create will have failed
2583 2590 * because the file must have existed already.
2584 2591 * Therefore, no truncate operation is needed.
2585 2592 */
2586 2593 va->va_mask &= ~(AT_GID | AT_SIZE);
2587 2594 error = nfs3setattr(vp, va, 0, cr);
2588 2595 if (error) {
2589 2596 /*
2590 2597 * Couldn't correct the attributes of
2591 2598 * the newly created file and the
2592 2599 * attributes are wrong. Remove the
2593 2600 * file and return an error to the
2594 2601 * application.
2595 2602 */
2596 2603 VN_RELE(vp);
2597 2604 (void) nfs3_remove(dvp, nm, cr, NULL, 0);
2598 2605 return (error);
2599 2606 }
2600 2607 }
2601 2608
2602 2609 if (va->va_gid != rp->r_attr.va_gid) {
2603 2610 /*
2604 2611 * If the gid on the file isn't right, then
2605 2612 * generate a SETATTR to attempt to change
2606 2613 * it. This may or may not work, depending
2607 2614 * upon the server's semantics for allowing
2608 2615 * file ownership changes.
2609 2616 */
2610 2617 va->va_mask = AT_GID;
2611 2618 (void) nfs3setattr(vp, va, 0, cr);
2612 2619 }
2613 2620
2614 2621 /*
2615 2622 * If vnode is a device create special vnode
2616 2623 */
2617 2624 if (IS_DEVVP(vp)) {
2618 2625 *vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2619 2626 VN_RELE(vp);
2620 2627 } else
2621 2628 *vpp = vp;
2622 2629 } else {
2623 2630 nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2624 2631 PURGE_STALE_FH(error, dvp, cr);
2625 2632 }
2626 2633
2627 2634 return (error);
2628 2635 }
2629 2636
2630 2637 /*
2631 2638 * Special setattr function to take care of rest of atime/mtime
2632 2639 * after successful exclusive create. This function exists to avoid
2633 2640 * handling attributes from the server; exclusive the atime/mtime fields
2634 2641 * may be 'invalid' in client's view and therefore can not be trusted.
2635 2642 */
2636 2643 static int
2637 2644 nfs3excl_create_settimes(vnode_t *vp, struct vattr *vap, cred_t *cr)
2638 2645 {
2639 2646 int error;
2640 2647 uint_t mask;
2641 2648 SETATTR3args args;
2642 2649 SETATTR3res res;
2643 2650 int douprintf;
2644 2651 rnode_t *rp;
2645 2652 hrtime_t t;
2646 2653
2647 2654 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
2648 2655 /* save the caller's mask so that it can be reset later */
2649 2656 mask = vap->va_mask;
2650 2657
2651 2658 rp = VTOR(vp);
2652 2659
2653 2660 args.object = *RTOFH3(rp);
2654 2661 args.guard.check = FALSE;
2655 2662
2656 2663 /* Use the mask to initialize the arguments */
2657 2664 vap->va_mask = 0;
2658 2665 error = vattr_to_sattr3(vap, &args.new_attributes);
2659 2666
2660 2667 /* We want to set just atime/mtime on this request */
2661 2668 args.new_attributes.atime.set_it = SET_TO_SERVER_TIME;
2662 2669 args.new_attributes.mtime.set_it = SET_TO_SERVER_TIME;
2663 2670
2664 2671 douprintf = 1;
2665 2672
2666 2673 t = gethrtime();
2667 2674
2668 2675 error = rfs3call(VTOMI(vp), NFSPROC3_SETATTR,
2669 2676 xdr_SETATTR3args, (caddr_t)&args,
2670 2677 xdr_SETATTR3res, (caddr_t)&res, cr,
2671 2678 &douprintf, &res.status, 0, NULL);
2672 2679
2673 2680 if (error) {
2674 2681 vap->va_mask = mask;
2675 2682 return (error);
2676 2683 }
2677 2684
2678 2685 error = geterrno3(res.status);
2679 2686 if (!error) {
2680 2687 /*
2681 2688 * It is important to pick up the attributes.
2682 2689 * Since this is the exclusive create path, the
2683 2690 * attributes on the initial create were ignored
2684 2691 * and we need these to have the correct info.
2685 2692 */
2686 2693 nfs3_cache_wcc_data(vp, &res.resok.obj_wcc, t, cr);
2687 2694 /*
2688 2695 * No need to do the atime/mtime work again so clear
2689 2696 * the bits.
2690 2697 */
2691 2698 mask &= ~(AT_ATIME | AT_MTIME);
2692 2699 } else {
2693 2700 nfs3_cache_wcc_data(vp, &res.resfail.obj_wcc, t, cr);
2694 2701 }
2695 2702
2696 2703 vap->va_mask = mask;
2697 2704
2698 2705 return (error);
2699 2706 }
2700 2707
2701 2708 /* ARGSUSED */
2702 2709 static int
2703 2710 nfs3mknod(vnode_t *dvp, char *nm, struct vattr *va, enum vcexcl exclusive,
2704 2711 int mode, vnode_t **vpp, cred_t *cr)
2705 2712 {
2706 2713 int error;
2707 2714 MKNOD3args args;
2708 2715 MKNOD3res res;
2709 2716 int douprintf;
2710 2717 vnode_t *vp;
2711 2718 struct vattr vattr;
2712 2719 hrtime_t t;
2713 2720
2714 2721 ASSERT(nfs_zone() == VTOMI(dvp)->mi_zone);
2715 2722 switch (va->va_type) {
2716 2723 case VCHR:
2717 2724 case VBLK:
2718 2725 setdiropargs3(&args.where, nm, dvp);
2719 2726 args.what.type = (va->va_type == VCHR) ? NF3CHR : NF3BLK;
2720 2727 error = vattr_to_sattr3(va,
2721 2728 &args.what.mknoddata3_u.device.dev_attributes);
2722 2729 if (error) {
2723 2730 /* req time field(s) overflow - return immediately */
2724 2731 return (error);
2725 2732 }
2726 2733 args.what.mknoddata3_u.device.spec.specdata1 =
2727 2734 getmajor(va->va_rdev);
2728 2735 args.what.mknoddata3_u.device.spec.specdata2 =
2729 2736 getminor(va->va_rdev);
2730 2737 break;
2731 2738
2732 2739 case VFIFO:
2733 2740 case VSOCK:
2734 2741 setdiropargs3(&args.where, nm, dvp);
2735 2742 args.what.type = (va->va_type == VFIFO) ? NF3FIFO : NF3SOCK;
2736 2743 error = vattr_to_sattr3(va,
2737 2744 &args.what.mknoddata3_u.pipe_attributes);
2738 2745 if (error) {
2739 2746 /* req time field(s) overflow - return immediately */
2740 2747 return (error);
2741 2748 }
2742 2749 break;
2743 2750
2744 2751 default:
2745 2752 return (EINVAL);
2746 2753 }
2747 2754
2748 2755 douprintf = 1;
2749 2756
2750 2757 t = gethrtime();
2751 2758
2752 2759 error = rfs3call(VTOMI(dvp), NFSPROC3_MKNOD,
2753 2760 xdr_MKNOD3args, (caddr_t)&args,
2754 2761 xdr_MKNOD3res, (caddr_t)&res, cr,
2755 2762 &douprintf, &res.status, 0, NULL);
2756 2763
2757 2764 if (error) {
2758 2765 PURGE_ATTRCACHE(dvp);
2759 2766 return (error);
2760 2767 }
2761 2768
2762 2769 error = geterrno3(res.status);
2763 2770 if (!error) {
2764 2771 nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
2765 2772 if (HAVE_RDDIR_CACHE(VTOR(dvp)))
2766 2773 nfs_purge_rddir_cache(dvp);
2767 2774
2768 2775 if (!res.resok.obj.handle_follows) {
2769 2776 error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2770 2777 if (error)
2771 2778 return (error);
2772 2779 } else {
2773 2780 if (res.resok.obj_attributes.attributes) {
2774 2781 vp = makenfs3node(&res.resok.obj.handle,
2775 2782 &res.resok.obj_attributes.attr,
2776 2783 dvp->v_vfsp, t, cr, NULL, NULL);
2777 2784 } else {
2778 2785 vp = makenfs3node(&res.resok.obj.handle, NULL,
2779 2786 dvp->v_vfsp, t, cr, NULL, NULL);
2780 2787 if (vp->v_type == VNON) {
2781 2788 vattr.va_mask = AT_TYPE;
2782 2789 error = nfs3getattr(vp, &vattr, cr);
2783 2790 if (error) {
2784 2791 VN_RELE(vp);
2785 2792 return (error);
2786 2793 }
2787 2794 vp->v_type = vattr.va_type;
2788 2795 }
2789 2796
2790 2797 }
2791 2798 dnlc_update(dvp, nm, vp);
2792 2799 }
2793 2800
2794 2801 if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
2795 2802 va->va_mask = AT_GID;
2796 2803 (void) nfs3setattr(vp, va, 0, cr);
2797 2804 }
2798 2805
2799 2806 /*
2800 2807 * If vnode is a device create special vnode
2801 2808 */
2802 2809 if (IS_DEVVP(vp)) {
2803 2810 *vpp = specvp(vp, vp->v_rdev, vp->v_type, cr);
2804 2811 VN_RELE(vp);
2805 2812 } else
2806 2813 *vpp = vp;
2807 2814 } else {
2808 2815 nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
2809 2816 PURGE_STALE_FH(error, dvp, cr);
2810 2817 }
2811 2818 return (error);
2812 2819 }
2813 2820
2814 2821 /*
2815 2822 * Weirdness: if the vnode to be removed is open
2816 2823 * we rename it instead of removing it and nfs_inactive
2817 2824 * will remove the new name.
2818 2825 */
2819 2826 /* ARGSUSED */
2820 2827 static int
2821 2828 nfs3_remove(vnode_t *dvp, char *nm, cred_t *cr, caller_context_t *ct, int flags)
2822 2829 {
2823 2830 int error;
2824 2831 REMOVE3args args;
2825 2832 REMOVE3res res;
2826 2833 vnode_t *vp;
2827 2834 char *tmpname;
2828 2835 int douprintf;
2829 2836 rnode_t *rp;
2830 2837 rnode_t *drp;
2831 2838 hrtime_t t;
2832 2839
2833 2840 if (nfs_zone() != VTOMI(dvp)->mi_zone)
2834 2841 return (EPERM);
2835 2842 drp = VTOR(dvp);
2836 2843 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
2837 2844 return (EINTR);
2838 2845
2839 2846 error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
2840 2847 if (error) {
2841 2848 nfs_rw_exit(&drp->r_rwlock);
2842 2849 return (error);
2843 2850 }
2844 2851
2845 2852 if (vp->v_type == VDIR && secpolicy_fs_linkdir(cr, dvp->v_vfsp)) {
2846 2853 VN_RELE(vp);
2847 2854 nfs_rw_exit(&drp->r_rwlock);
2848 2855 return (EPERM);
2849 2856 }
2850 2857
2851 2858 /*
2852 2859 * First just remove the entry from the name cache, as it
2853 2860 * is most likely the only entry for this vp.
2854 2861 */
2855 2862 dnlc_remove(dvp, nm);
2856 2863
2857 2864 /*
2858 2865 * If the file has a v_count > 1 then there may be more than one
2859 2866 * entry in the name cache due multiple links or an open file,
2860 2867 * but we don't have the real reference count so flush all
2861 2868 * possible entries.
2862 2869 */
2863 2870 if (vp->v_count > 1)
2864 2871 dnlc_purge_vp(vp);
2865 2872
2866 2873 /*
2867 2874 * Now we have the real reference count on the vnode
2868 2875 */
2869 2876 rp = VTOR(vp);
2870 2877 mutex_enter(&rp->r_statelock);
2871 2878 if (vp->v_count > 1 &&
2872 2879 (rp->r_unldvp == NULL || strcmp(nm, rp->r_unlname) == 0)) {
2873 2880 mutex_exit(&rp->r_statelock);
2874 2881 tmpname = newname();
2875 2882 error = nfs3rename(dvp, nm, dvp, tmpname, cr, ct);
2876 2883 if (error)
2877 2884 kmem_free(tmpname, MAXNAMELEN);
2878 2885 else {
2879 2886 mutex_enter(&rp->r_statelock);
2880 2887 if (rp->r_unldvp == NULL) {
2881 2888 VN_HOLD(dvp);
2882 2889 rp->r_unldvp = dvp;
2883 2890 if (rp->r_unlcred != NULL)
2884 2891 crfree(rp->r_unlcred);
2885 2892 crhold(cr);
2886 2893 rp->r_unlcred = cr;
2887 2894 rp->r_unlname = tmpname;
2888 2895 } else {
2889 2896 kmem_free(rp->r_unlname, MAXNAMELEN);
2890 2897 rp->r_unlname = tmpname;
2891 2898 }
2892 2899 mutex_exit(&rp->r_statelock);
2893 2900 }
2894 2901 } else {
2895 2902 mutex_exit(&rp->r_statelock);
2896 2903 /*
2897 2904 * We need to flush any dirty pages which happen to
2898 2905 * be hanging around before removing the file. This
2899 2906 * shouldn't happen very often and mostly on file
2900 2907 * systems mounted "nocto".
2901 2908 */
2902 2909 if (vn_has_cached_data(vp) &&
2903 2910 ((rp->r_flags & RDIRTY) || rp->r_count > 0)) {
2904 2911 error = nfs3_putpage(vp, (offset_t)0, 0, 0, cr, ct);
2905 2912 if (error && (error == ENOSPC || error == EDQUOT)) {
2906 2913 mutex_enter(&rp->r_statelock);
2907 2914 if (!rp->r_error)
2908 2915 rp->r_error = error;
2909 2916 mutex_exit(&rp->r_statelock);
2910 2917 }
2911 2918 }
2912 2919
2913 2920 setdiropargs3(&args.object, nm, dvp);
2914 2921
2915 2922 douprintf = 1;
2916 2923
2917 2924 t = gethrtime();
2918 2925
2919 2926 error = rfs3call(VTOMI(dvp), NFSPROC3_REMOVE,
2920 2927 xdr_diropargs3, (caddr_t)&args,
2921 2928 xdr_REMOVE3res, (caddr_t)&res, cr,
2922 2929 &douprintf, &res.status, 0, NULL);
2923 2930
2924 2931 /*
2925 2932 * The xattr dir may be gone after last attr is removed,
2926 2933 * so flush it from dnlc.
2927 2934 */
2928 2935 if (dvp->v_flag & V_XATTRDIR)
2929 2936 dnlc_purge_vp(dvp);
2930 2937
2931 2938 PURGE_ATTRCACHE(vp);
2932 2939
2933 2940 if (error) {
2934 2941 PURGE_ATTRCACHE(dvp);
2935 2942 } else {
2936 2943 error = geterrno3(res.status);
2937 2944 if (!error) {
2938 2945 nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t,
2939 2946 cr);
2940 2947 if (HAVE_RDDIR_CACHE(drp))
2941 2948 nfs_purge_rddir_cache(dvp);
2942 2949 } else {
2943 2950 nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc,
2944 2951 t, cr);
2945 2952 PURGE_STALE_FH(error, dvp, cr);
2946 2953 }
2947 2954 }
2948 2955 }
2949 2956
2950 2957 if (error == 0) {
2951 2958 vnevent_remove(vp, dvp, nm, ct);
2952 2959 }
2953 2960 VN_RELE(vp);
2954 2961
2955 2962 nfs_rw_exit(&drp->r_rwlock);
2956 2963
2957 2964 return (error);
2958 2965 }
2959 2966
2960 2967 /* ARGSUSED */
2961 2968 static int
2962 2969 nfs3_link(vnode_t *tdvp, vnode_t *svp, char *tnm, cred_t *cr,
2963 2970 caller_context_t *ct, int flags)
2964 2971 {
2965 2972 int error;
2966 2973 LINK3args args;
2967 2974 LINK3res res;
2968 2975 vnode_t *realvp;
2969 2976 int douprintf;
2970 2977 mntinfo_t *mi;
2971 2978 rnode_t *tdrp;
2972 2979 hrtime_t t;
2973 2980
2974 2981 if (nfs_zone() != VTOMI(tdvp)->mi_zone)
2975 2982 return (EPERM);
2976 2983 if (VOP_REALVP(svp, &realvp, ct) == 0)
2977 2984 svp = realvp;
2978 2985
2979 2986 mi = VTOMI(svp);
2980 2987
2981 2988 if (!(mi->mi_flags & MI_LINK))
2982 2989 return (EOPNOTSUPP);
2983 2990
2984 2991 args.file = *VTOFH3(svp);
2985 2992 setdiropargs3(&args.link, tnm, tdvp);
2986 2993
2987 2994 tdrp = VTOR(tdvp);
2988 2995 if (nfs_rw_enter_sig(&tdrp->r_rwlock, RW_WRITER, INTR(tdvp)))
2989 2996 return (EINTR);
2990 2997
2991 2998 dnlc_remove(tdvp, tnm);
2992 2999
2993 3000 douprintf = 1;
2994 3001
2995 3002 t = gethrtime();
2996 3003
2997 3004 error = rfs3call(mi, NFSPROC3_LINK,
2998 3005 xdr_LINK3args, (caddr_t)&args,
2999 3006 xdr_LINK3res, (caddr_t)&res, cr,
3000 3007 &douprintf, &res.status, 0, NULL);
3001 3008
3002 3009 if (error) {
3003 3010 PURGE_ATTRCACHE(tdvp);
3004 3011 PURGE_ATTRCACHE(svp);
3005 3012 nfs_rw_exit(&tdrp->r_rwlock);
3006 3013 return (error);
3007 3014 }
3008 3015
3009 3016 error = geterrno3(res.status);
3010 3017
3011 3018 if (!error) {
3012 3019 nfs3_cache_post_op_attr(svp, &res.resok.file_attributes, t, cr);
3013 3020 nfs3_cache_wcc_data(tdvp, &res.resok.linkdir_wcc, t, cr);
3014 3021 if (HAVE_RDDIR_CACHE(tdrp))
3015 3022 nfs_purge_rddir_cache(tdvp);
3016 3023 dnlc_update(tdvp, tnm, svp);
3017 3024 } else {
3018 3025 nfs3_cache_post_op_attr(svp, &res.resfail.file_attributes, t,
3019 3026 cr);
3020 3027 nfs3_cache_wcc_data(tdvp, &res.resfail.linkdir_wcc, t, cr);
3021 3028 if (error == EOPNOTSUPP) {
3022 3029 mutex_enter(&mi->mi_lock);
3023 3030 mi->mi_flags &= ~MI_LINK;
3024 3031 mutex_exit(&mi->mi_lock);
3025 3032 }
3026 3033 }
3027 3034
3028 3035 nfs_rw_exit(&tdrp->r_rwlock);
3029 3036
3030 3037 if (!error) {
3031 3038 /*
3032 3039 * Notify the source file of this link operation.
3033 3040 */
3034 3041 vnevent_link(svp, ct);
3035 3042 }
3036 3043 return (error);
3037 3044 }
3038 3045
3039 3046 /* ARGSUSED */
3040 3047 static int
3041 3048 nfs3_rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr,
3042 3049 caller_context_t *ct, int flags)
3043 3050 {
3044 3051 vnode_t *realvp;
3045 3052
3046 3053 if (nfs_zone() != VTOMI(odvp)->mi_zone)
3047 3054 return (EPERM);
3048 3055 if (VOP_REALVP(ndvp, &realvp, ct) == 0)
3049 3056 ndvp = realvp;
3050 3057
3051 3058 return (nfs3rename(odvp, onm, ndvp, nnm, cr, ct));
3052 3059 }
3053 3060
3054 3061 /*
3055 3062 * nfs3rename does the real work of renaming in NFS Version 3.
3056 3063 */
3057 3064 static int
3058 3065 nfs3rename(vnode_t *odvp, char *onm, vnode_t *ndvp, char *nnm, cred_t *cr,
3059 3066 caller_context_t *ct)
3060 3067 {
3061 3068 int error;
3062 3069 RENAME3args args;
3063 3070 RENAME3res res;
3064 3071 int douprintf;
3065 3072 vnode_t *nvp = NULL;
3066 3073 vnode_t *ovp = NULL;
3067 3074 char *tmpname;
3068 3075 rnode_t *rp;
3069 3076 rnode_t *odrp;
3070 3077 rnode_t *ndrp;
3071 3078 hrtime_t t;
3072 3079
3073 3080 ASSERT(nfs_zone() == VTOMI(odvp)->mi_zone);
3074 3081
3075 3082 if (strcmp(onm, ".") == 0 || strcmp(onm, "..") == 0 ||
3076 3083 strcmp(nnm, ".") == 0 || strcmp(nnm, "..") == 0)
3077 3084 return (EINVAL);
3078 3085
3079 3086 odrp = VTOR(odvp);
3080 3087 ndrp = VTOR(ndvp);
3081 3088 if ((intptr_t)odrp < (intptr_t)ndrp) {
3082 3089 if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp)))
3083 3090 return (EINTR);
3084 3091 if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp))) {
3085 3092 nfs_rw_exit(&odrp->r_rwlock);
3086 3093 return (EINTR);
3087 3094 }
3088 3095 } else {
3089 3096 if (nfs_rw_enter_sig(&ndrp->r_rwlock, RW_WRITER, INTR(ndvp)))
3090 3097 return (EINTR);
3091 3098 if (nfs_rw_enter_sig(&odrp->r_rwlock, RW_WRITER, INTR(odvp))) {
3092 3099 nfs_rw_exit(&ndrp->r_rwlock);
3093 3100 return (EINTR);
3094 3101 }
3095 3102 }
3096 3103
3097 3104 /*
3098 3105 * Lookup the target file. If it exists, it needs to be
3099 3106 * checked to see whether it is a mount point and whether
3100 3107 * it is active (open).
3101 3108 */
3102 3109 error = nfs3lookup(ndvp, nnm, &nvp, NULL, 0, NULL, cr, 0);
3103 3110 if (!error) {
3104 3111 /*
3105 3112 * If this file has been mounted on, then just
3106 3113 * return busy because renaming to it would remove
3107 3114 * the mounted file system from the name space.
3108 3115 */
3109 3116 if (vn_mountedvfs(nvp) != NULL) {
3110 3117 VN_RELE(nvp);
3111 3118 nfs_rw_exit(&odrp->r_rwlock);
3112 3119 nfs_rw_exit(&ndrp->r_rwlock);
3113 3120 return (EBUSY);
3114 3121 }
3115 3122
3116 3123 /*
3117 3124 * Purge the name cache of all references to this vnode
3118 3125 * so that we can check the reference count to infer
3119 3126 * whether it is active or not.
3120 3127 */
3121 3128 /*
3122 3129 * First just remove the entry from the name cache, as it
3123 3130 * is most likely the only entry for this vp.
3124 3131 */
3125 3132 dnlc_remove(ndvp, nnm);
3126 3133 /*
3127 3134 * If the file has a v_count > 1 then there may be more
3128 3135 * than one entry in the name cache due multiple links
3129 3136 * or an open file, but we don't have the real reference
3130 3137 * count so flush all possible entries.
3131 3138 */
3132 3139 if (nvp->v_count > 1)
3133 3140 dnlc_purge_vp(nvp);
3134 3141
3135 3142 /*
3136 3143 * If the vnode is active and is not a directory,
3137 3144 * arrange to rename it to a
3138 3145 * temporary file so that it will continue to be
3139 3146 * accessible. This implements the "unlink-open-file"
3140 3147 * semantics for the target of a rename operation.
3141 3148 * Before doing this though, make sure that the
3142 3149 * source and target files are not already the same.
3143 3150 */
3144 3151 if (nvp->v_count > 1 && nvp->v_type != VDIR) {
3145 3152 /*
3146 3153 * Lookup the source name.
3147 3154 */
3148 3155 error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL,
3149 3156 cr, 0);
3150 3157
3151 3158 /*
3152 3159 * The source name *should* already exist.
3153 3160 */
3154 3161 if (error) {
3155 3162 VN_RELE(nvp);
3156 3163 nfs_rw_exit(&odrp->r_rwlock);
3157 3164 nfs_rw_exit(&ndrp->r_rwlock);
3158 3165 return (error);
3159 3166 }
3160 3167
3161 3168 /*
3162 3169 * Compare the two vnodes. If they are the same,
3163 3170 * just release all held vnodes and return success.
3164 3171 */
3165 3172 if (ovp == nvp) {
3166 3173 VN_RELE(ovp);
3167 3174 VN_RELE(nvp);
3168 3175 nfs_rw_exit(&odrp->r_rwlock);
3169 3176 nfs_rw_exit(&ndrp->r_rwlock);
3170 3177 return (0);
3171 3178 }
3172 3179
3173 3180 /*
3174 3181 * Can't mix and match directories and non-
3175 3182 * directories in rename operations. We already
3176 3183 * know that the target is not a directory. If
3177 3184 * the source is a directory, return an error.
3178 3185 */
3179 3186 if (ovp->v_type == VDIR) {
3180 3187 VN_RELE(ovp);
3181 3188 VN_RELE(nvp);
3182 3189 nfs_rw_exit(&odrp->r_rwlock);
3183 3190 nfs_rw_exit(&ndrp->r_rwlock);
3184 3191 return (ENOTDIR);
3185 3192 }
3186 3193
3187 3194 /*
3188 3195 * The target file exists, is not the same as
3189 3196 * the source file, and is active. Link it
3190 3197 * to a temporary filename to avoid having
3191 3198 * the server removing the file completely.
3192 3199 */
3193 3200 tmpname = newname();
3194 3201 error = nfs3_link(ndvp, nvp, tmpname, cr, NULL, 0);
3195 3202 if (error == EOPNOTSUPP) {
3196 3203 error = nfs3_rename(ndvp, nnm, ndvp, tmpname,
3197 3204 cr, NULL, 0);
3198 3205 }
3199 3206 if (error) {
3200 3207 kmem_free(tmpname, MAXNAMELEN);
3201 3208 VN_RELE(ovp);
3202 3209 VN_RELE(nvp);
3203 3210 nfs_rw_exit(&odrp->r_rwlock);
3204 3211 nfs_rw_exit(&ndrp->r_rwlock);
3205 3212 return (error);
3206 3213 }
3207 3214 rp = VTOR(nvp);
3208 3215 mutex_enter(&rp->r_statelock);
3209 3216 if (rp->r_unldvp == NULL) {
3210 3217 VN_HOLD(ndvp);
3211 3218 rp->r_unldvp = ndvp;
3212 3219 if (rp->r_unlcred != NULL)
3213 3220 crfree(rp->r_unlcred);
3214 3221 crhold(cr);
3215 3222 rp->r_unlcred = cr;
3216 3223 rp->r_unlname = tmpname;
3217 3224 } else {
3218 3225 kmem_free(rp->r_unlname, MAXNAMELEN);
3219 3226 rp->r_unlname = tmpname;
3220 3227 }
3221 3228 mutex_exit(&rp->r_statelock);
3222 3229 }
3223 3230 }
3224 3231
3225 3232 if (ovp == NULL) {
3226 3233 /*
3227 3234 * When renaming directories to be a subdirectory of a
3228 3235 * different parent, the dnlc entry for ".." will no
3229 3236 * longer be valid, so it must be removed.
3230 3237 *
3231 3238 * We do a lookup here to determine whether we are renaming
3232 3239 * a directory and we need to check if we are renaming
3233 3240 * an unlinked file. This might have already been done
3234 3241 * in previous code, so we check ovp == NULL to avoid
3235 3242 * doing it twice.
3236 3243 */
3237 3244
3238 3245 error = nfs3lookup(odvp, onm, &ovp, NULL, 0, NULL, cr, 0);
3239 3246 /*
3240 3247 * The source name *should* already exist.
3241 3248 */
3242 3249 if (error) {
3243 3250 nfs_rw_exit(&odrp->r_rwlock);
3244 3251 nfs_rw_exit(&ndrp->r_rwlock);
3245 3252 if (nvp) {
3246 3253 VN_RELE(nvp);
3247 3254 }
3248 3255 return (error);
3249 3256 }
3250 3257 ASSERT(ovp != NULL);
3251 3258 }
3252 3259
3253 3260 dnlc_remove(odvp, onm);
3254 3261 dnlc_remove(ndvp, nnm);
3255 3262
3256 3263 setdiropargs3(&args.from, onm, odvp);
3257 3264 setdiropargs3(&args.to, nnm, ndvp);
3258 3265
3259 3266 douprintf = 1;
3260 3267
3261 3268 t = gethrtime();
3262 3269
3263 3270 error = rfs3call(VTOMI(odvp), NFSPROC3_RENAME,
3264 3271 xdr_RENAME3args, (caddr_t)&args,
3265 3272 xdr_RENAME3res, (caddr_t)&res, cr,
3266 3273 &douprintf, &res.status, 0, NULL);
3267 3274
3268 3275 if (error) {
3269 3276 PURGE_ATTRCACHE(odvp);
3270 3277 PURGE_ATTRCACHE(ndvp);
3271 3278 VN_RELE(ovp);
3272 3279 nfs_rw_exit(&odrp->r_rwlock);
3273 3280 nfs_rw_exit(&ndrp->r_rwlock);
3274 3281 if (nvp) {
3275 3282 VN_RELE(nvp);
3276 3283 }
3277 3284 return (error);
3278 3285 }
3279 3286
3280 3287 error = geterrno3(res.status);
3281 3288
3282 3289 if (!error) {
3283 3290 nfs3_cache_wcc_data(odvp, &res.resok.fromdir_wcc, t, cr);
3284 3291 if (HAVE_RDDIR_CACHE(odrp))
3285 3292 nfs_purge_rddir_cache(odvp);
3286 3293 if (ndvp != odvp) {
3287 3294 nfs3_cache_wcc_data(ndvp, &res.resok.todir_wcc, t, cr);
3288 3295 if (HAVE_RDDIR_CACHE(ndrp))
3289 3296 nfs_purge_rddir_cache(ndvp);
3290 3297 }
3291 3298 /*
3292 3299 * when renaming directories to be a subdirectory of a
3293 3300 * different parent, the dnlc entry for ".." will no
3294 3301 * longer be valid, so it must be removed
3295 3302 */
3296 3303 rp = VTOR(ovp);
3297 3304 if (ndvp != odvp) {
3298 3305 if (ovp->v_type == VDIR) {
3299 3306 dnlc_remove(ovp, "..");
3300 3307 if (HAVE_RDDIR_CACHE(rp))
3301 3308 nfs_purge_rddir_cache(ovp);
3302 3309 }
3303 3310 }
3304 3311
3305 3312 /*
3306 3313 * If we are renaming the unlinked file, update the
3307 3314 * r_unldvp and r_unlname as needed.
3308 3315 */
3309 3316 mutex_enter(&rp->r_statelock);
3310 3317 if (rp->r_unldvp != NULL) {
3311 3318 if (strcmp(rp->r_unlname, onm) == 0) {
3312 3319 (void) strncpy(rp->r_unlname, nnm, MAXNAMELEN);
3313 3320 rp->r_unlname[MAXNAMELEN - 1] = '\0';
3314 3321
3315 3322 if (ndvp != rp->r_unldvp) {
3316 3323 VN_RELE(rp->r_unldvp);
3317 3324 rp->r_unldvp = ndvp;
3318 3325 VN_HOLD(ndvp);
3319 3326 }
3320 3327 }
3321 3328 }
3322 3329 mutex_exit(&rp->r_statelock);
3323 3330 } else {
3324 3331 nfs3_cache_wcc_data(odvp, &res.resfail.fromdir_wcc, t, cr);
3325 3332 if (ndvp != odvp) {
3326 3333 nfs3_cache_wcc_data(ndvp, &res.resfail.todir_wcc, t,
3327 3334 cr);
3328 3335 }
3329 3336 /*
3330 3337 * System V defines rename to return EEXIST, not
3331 3338 * ENOTEMPTY if the target directory is not empty.
3332 3339 * Over the wire, the error is NFSERR_ENOTEMPTY
3333 3340 * which geterrno maps to ENOTEMPTY.
3334 3341 */
3335 3342 if (error == ENOTEMPTY)
3336 3343 error = EEXIST;
3337 3344 }
3338 3345
3339 3346 if (error == 0) {
3340 3347 if (nvp)
3341 3348 vnevent_rename_dest(nvp, ndvp, nnm, ct);
3342 3349
3343 3350 if (odvp != ndvp)
3344 3351 vnevent_rename_dest_dir(ndvp, ct);
3345 3352 ASSERT(ovp != NULL);
3346 3353 vnevent_rename_src(ovp, odvp, onm, ct);
3347 3354 }
3348 3355
3349 3356 if (nvp) {
3350 3357 VN_RELE(nvp);
3351 3358 }
3352 3359 VN_RELE(ovp);
3353 3360
3354 3361 nfs_rw_exit(&odrp->r_rwlock);
3355 3362 nfs_rw_exit(&ndrp->r_rwlock);
3356 3363
3357 3364 return (error);
3358 3365 }
3359 3366
3360 3367 /* ARGSUSED */
3361 3368 static int
3362 3369 nfs3_mkdir(vnode_t *dvp, char *nm, struct vattr *va, vnode_t **vpp, cred_t *cr,
3363 3370 caller_context_t *ct, int flags, vsecattr_t *vsecp)
3364 3371 {
3365 3372 int error;
3366 3373 MKDIR3args args;
3367 3374 MKDIR3res res;
3368 3375 int douprintf;
3369 3376 struct vattr vattr;
3370 3377 vnode_t *vp;
3371 3378 rnode_t *drp;
3372 3379 hrtime_t t;
3373 3380
3374 3381 if (nfs_zone() != VTOMI(dvp)->mi_zone)
3375 3382 return (EPERM);
3376 3383 setdiropargs3(&args.where, nm, dvp);
3377 3384
3378 3385 /*
3379 3386 * Decide what the group-id and set-gid bit of the created directory
3380 3387 * should be. May have to do a setattr to get the gid right.
3381 3388 */
3382 3389 error = setdirgid(dvp, &va->va_gid, cr);
3383 3390 if (error)
3384 3391 return (error);
3385 3392 error = setdirmode(dvp, &va->va_mode, cr);
3386 3393 if (error)
3387 3394 return (error);
3388 3395 va->va_mask |= AT_MODE|AT_GID;
3389 3396
3390 3397 error = vattr_to_sattr3(va, &args.attributes);
3391 3398 if (error) {
3392 3399 /* req time field(s) overflow - return immediately */
3393 3400 return (error);
3394 3401 }
3395 3402
3396 3403 drp = VTOR(dvp);
3397 3404 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3398 3405 return (EINTR);
3399 3406
3400 3407 dnlc_remove(dvp, nm);
3401 3408
3402 3409 douprintf = 1;
3403 3410
3404 3411 t = gethrtime();
3405 3412
3406 3413 error = rfs3call(VTOMI(dvp), NFSPROC3_MKDIR,
3407 3414 xdr_MKDIR3args, (caddr_t)&args,
3408 3415 xdr_MKDIR3res, (caddr_t)&res, cr,
3409 3416 &douprintf, &res.status, 0, NULL);
3410 3417
3411 3418 if (error) {
3412 3419 PURGE_ATTRCACHE(dvp);
3413 3420 nfs_rw_exit(&drp->r_rwlock);
3414 3421 return (error);
3415 3422 }
3416 3423
3417 3424 error = geterrno3(res.status);
3418 3425 if (!error) {
3419 3426 nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3420 3427 if (HAVE_RDDIR_CACHE(drp))
3421 3428 nfs_purge_rddir_cache(dvp);
3422 3429
3423 3430 if (!res.resok.obj.handle_follows) {
3424 3431 error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3425 3432 if (error) {
3426 3433 nfs_rw_exit(&drp->r_rwlock);
3427 3434 return (error);
3428 3435 }
3429 3436 } else {
3430 3437 if (res.resok.obj_attributes.attributes) {
3431 3438 vp = makenfs3node(&res.resok.obj.handle,
3432 3439 &res.resok.obj_attributes.attr,
3433 3440 dvp->v_vfsp, t, cr, NULL, NULL);
3434 3441 } else {
3435 3442 vp = makenfs3node(&res.resok.obj.handle, NULL,
3436 3443 dvp->v_vfsp, t, cr, NULL, NULL);
3437 3444 if (vp->v_type == VNON) {
3438 3445 vattr.va_mask = AT_TYPE;
3439 3446 error = nfs3getattr(vp, &vattr, cr);
3440 3447 if (error) {
3441 3448 VN_RELE(vp);
3442 3449 nfs_rw_exit(&drp->r_rwlock);
3443 3450 return (error);
3444 3451 }
3445 3452 vp->v_type = vattr.va_type;
3446 3453 }
3447 3454 }
3448 3455 dnlc_update(dvp, nm, vp);
3449 3456 }
3450 3457 if (va->va_gid != VTOR(vp)->r_attr.va_gid) {
3451 3458 va->va_mask = AT_GID;
3452 3459 (void) nfs3setattr(vp, va, 0, cr);
3453 3460 }
3454 3461 *vpp = vp;
3455 3462 } else {
3456 3463 nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3457 3464 PURGE_STALE_FH(error, dvp, cr);
3458 3465 }
3459 3466
3460 3467 nfs_rw_exit(&drp->r_rwlock);
3461 3468
3462 3469 return (error);
3463 3470 }
3464 3471
3465 3472 /* ARGSUSED */
3466 3473 static int
3467 3474 nfs3_rmdir(vnode_t *dvp, char *nm, vnode_t *cdir, cred_t *cr,
3468 3475 caller_context_t *ct, int flags)
3469 3476 {
3470 3477 int error;
3471 3478 RMDIR3args args;
3472 3479 RMDIR3res res;
3473 3480 vnode_t *vp;
3474 3481 int douprintf;
3475 3482 rnode_t *drp;
3476 3483 hrtime_t t;
3477 3484
3478 3485 if (nfs_zone() != VTOMI(dvp)->mi_zone)
3479 3486 return (EPERM);
3480 3487 drp = VTOR(dvp);
3481 3488 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3482 3489 return (EINTR);
3483 3490
3484 3491 /*
3485 3492 * Attempt to prevent a rmdir(".") from succeeding.
3486 3493 */
3487 3494 error = nfs3lookup(dvp, nm, &vp, NULL, 0, NULL, cr, 0);
3488 3495 if (error) {
3489 3496 nfs_rw_exit(&drp->r_rwlock);
3490 3497 return (error);
3491 3498 }
3492 3499
3493 3500 if (vp == cdir) {
3494 3501 VN_RELE(vp);
3495 3502 nfs_rw_exit(&drp->r_rwlock);
3496 3503 return (EINVAL);
3497 3504 }
3498 3505
3499 3506 setdiropargs3(&args.object, nm, dvp);
3500 3507
3501 3508 /*
3502 3509 * First just remove the entry from the name cache, as it
3503 3510 * is most likely an entry for this vp.
3504 3511 */
3505 3512 dnlc_remove(dvp, nm);
3506 3513
3507 3514 /*
3508 3515 * If there vnode reference count is greater than one, then
3509 3516 * there may be additional references in the DNLC which will
3510 3517 * need to be purged. First, trying removing the entry for
3511 3518 * the parent directory and see if that removes the additional
3512 3519 * reference(s). If that doesn't do it, then use dnlc_purge_vp
3513 3520 * to completely remove any references to the directory which
3514 3521 * might still exist in the DNLC.
3515 3522 */
3516 3523 if (vp->v_count > 1) {
3517 3524 dnlc_remove(vp, "..");
3518 3525 if (vp->v_count > 1)
3519 3526 dnlc_purge_vp(vp);
3520 3527 }
3521 3528
3522 3529 douprintf = 1;
3523 3530
3524 3531 t = gethrtime();
3525 3532
3526 3533 error = rfs3call(VTOMI(dvp), NFSPROC3_RMDIR,
3527 3534 xdr_diropargs3, (caddr_t)&args,
3528 3535 xdr_RMDIR3res, (caddr_t)&res, cr,
3529 3536 &douprintf, &res.status, 0, NULL);
3530 3537
3531 3538 PURGE_ATTRCACHE(vp);
3532 3539
3533 3540 if (error) {
3534 3541 PURGE_ATTRCACHE(dvp);
3535 3542 VN_RELE(vp);
3536 3543 nfs_rw_exit(&drp->r_rwlock);
3537 3544 return (error);
3538 3545 }
3539 3546
3540 3547 error = geterrno3(res.status);
3541 3548 if (!error) {
3542 3549 nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3543 3550 if (HAVE_RDDIR_CACHE(drp))
3544 3551 nfs_purge_rddir_cache(dvp);
3545 3552 if (HAVE_RDDIR_CACHE(VTOR(vp)))
3546 3553 nfs_purge_rddir_cache(vp);
3547 3554 } else {
3548 3555 nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3549 3556 PURGE_STALE_FH(error, dvp, cr);
3550 3557 /*
3551 3558 * System V defines rmdir to return EEXIST, not
3552 3559 * ENOTEMPTY if the directory is not empty. Over
3553 3560 * the wire, the error is NFSERR_ENOTEMPTY which
3554 3561 * geterrno maps to ENOTEMPTY.
3555 3562 */
3556 3563 if (error == ENOTEMPTY)
3557 3564 error = EEXIST;
3558 3565 }
3559 3566
3560 3567 if (error == 0) {
3561 3568 vnevent_rmdir(vp, dvp, nm, ct);
3562 3569 }
3563 3570 VN_RELE(vp);
3564 3571
3565 3572 nfs_rw_exit(&drp->r_rwlock);
3566 3573
3567 3574 return (error);
3568 3575 }
3569 3576
3570 3577 /* ARGSUSED */
3571 3578 static int
3572 3579 nfs3_symlink(vnode_t *dvp, char *lnm, struct vattr *tva, char *tnm, cred_t *cr,
3573 3580 caller_context_t *ct, int flags)
3574 3581 {
3575 3582 int error;
3576 3583 SYMLINK3args args;
3577 3584 SYMLINK3res res;
3578 3585 int douprintf;
3579 3586 mntinfo_t *mi;
3580 3587 vnode_t *vp;
3581 3588 rnode_t *rp;
3582 3589 char *contents;
3583 3590 rnode_t *drp;
3584 3591 hrtime_t t;
3585 3592
3586 3593 mi = VTOMI(dvp);
3587 3594
3588 3595 if (nfs_zone() != mi->mi_zone)
3589 3596 return (EPERM);
3590 3597 if (!(mi->mi_flags & MI_SYMLINK))
3591 3598 return (EOPNOTSUPP);
3592 3599
3593 3600 setdiropargs3(&args.where, lnm, dvp);
3594 3601 error = vattr_to_sattr3(tva, &args.symlink.symlink_attributes);
3595 3602 if (error) {
3596 3603 /* req time field(s) overflow - return immediately */
3597 3604 return (error);
3598 3605 }
3599 3606 args.symlink.symlink_data = tnm;
3600 3607
3601 3608 drp = VTOR(dvp);
3602 3609 if (nfs_rw_enter_sig(&drp->r_rwlock, RW_WRITER, INTR(dvp)))
3603 3610 return (EINTR);
3604 3611
3605 3612 dnlc_remove(dvp, lnm);
3606 3613
3607 3614 douprintf = 1;
3608 3615
3609 3616 t = gethrtime();
3610 3617
3611 3618 error = rfs3call(mi, NFSPROC3_SYMLINK,
3612 3619 xdr_SYMLINK3args, (caddr_t)&args,
3613 3620 xdr_SYMLINK3res, (caddr_t)&res, cr,
3614 3621 &douprintf, &res.status, 0, NULL);
3615 3622
3616 3623 if (error) {
3617 3624 PURGE_ATTRCACHE(dvp);
3618 3625 nfs_rw_exit(&drp->r_rwlock);
3619 3626 return (error);
3620 3627 }
3621 3628
3622 3629 error = geterrno3(res.status);
3623 3630 if (!error) {
3624 3631 nfs3_cache_wcc_data(dvp, &res.resok.dir_wcc, t, cr);
3625 3632 if (HAVE_RDDIR_CACHE(drp))
3626 3633 nfs_purge_rddir_cache(dvp);
3627 3634
3628 3635 if (res.resok.obj.handle_follows) {
3629 3636 if (res.resok.obj_attributes.attributes) {
3630 3637 vp = makenfs3node(&res.resok.obj.handle,
3631 3638 &res.resok.obj_attributes.attr,
3632 3639 dvp->v_vfsp, t, cr, NULL, NULL);
3633 3640 } else {
3634 3641 vp = makenfs3node(&res.resok.obj.handle, NULL,
3635 3642 dvp->v_vfsp, t, cr, NULL, NULL);
3636 3643 vp->v_type = VLNK;
3637 3644 vp->v_rdev = 0;
3638 3645 }
3639 3646 dnlc_update(dvp, lnm, vp);
3640 3647 rp = VTOR(vp);
3641 3648 if (nfs3_do_symlink_cache &&
3642 3649 rp->r_symlink.contents == NULL) {
3643 3650
3644 3651 contents = kmem_alloc(MAXPATHLEN,
3645 3652 KM_NOSLEEP);
3646 3653
3647 3654 if (contents != NULL) {
3648 3655 mutex_enter(&rp->r_statelock);
3649 3656 if (rp->r_symlink.contents == NULL) {
3650 3657 rp->r_symlink.len = strlen(tnm);
3651 3658 bcopy(tnm, contents,
3652 3659 rp->r_symlink.len);
3653 3660 rp->r_symlink.contents =
3654 3661 contents;
3655 3662 rp->r_symlink.size = MAXPATHLEN;
3656 3663 mutex_exit(&rp->r_statelock);
3657 3664 } else {
3658 3665 mutex_exit(&rp->r_statelock);
3659 3666 kmem_free((void *)contents,
3660 3667 MAXPATHLEN);
3661 3668 }
3662 3669 }
3663 3670 }
3664 3671 VN_RELE(vp);
3665 3672 }
3666 3673 } else {
3667 3674 nfs3_cache_wcc_data(dvp, &res.resfail.dir_wcc, t, cr);
3668 3675 PURGE_STALE_FH(error, dvp, cr);
3669 3676 if (error == EOPNOTSUPP) {
3670 3677 mutex_enter(&mi->mi_lock);
3671 3678 mi->mi_flags &= ~MI_SYMLINK;
3672 3679 mutex_exit(&mi->mi_lock);
3673 3680 }
3674 3681 }
3675 3682
3676 3683 nfs_rw_exit(&drp->r_rwlock);
3677 3684
3678 3685 return (error);
3679 3686 }
3680 3687
3681 3688 #ifdef DEBUG
3682 3689 static int nfs3_readdir_cache_hits = 0;
3683 3690 static int nfs3_readdir_cache_shorts = 0;
3684 3691 static int nfs3_readdir_cache_waits = 0;
3685 3692 static int nfs3_readdir_cache_misses = 0;
3686 3693 static int nfs3_readdir_readahead = 0;
3687 3694 #endif
3688 3695
3689 3696 static int nfs3_shrinkreaddir = 0;
3690 3697
3691 3698 /*
3692 3699 * Read directory entries.
3693 3700 * There are some weird things to look out for here. The uio_loffset
3694 3701 * field is either 0 or it is the offset returned from a previous
3695 3702 * readdir. It is an opaque value used by the server to find the
3696 3703 * correct directory block to read. The count field is the number
3697 3704 * of blocks to read on the server. This is advisory only, the server
3698 3705 * may return only one block's worth of entries. Entries may be compressed
3699 3706 * on the server.
3700 3707 */
3701 3708 /* ARGSUSED */
3702 3709 static int
3703 3710 nfs3_readdir(vnode_t *vp, struct uio *uiop, cred_t *cr, int *eofp,
3704 3711 caller_context_t *ct, int flags)
3705 3712 {
3706 3713 int error;
3707 3714 size_t count;
3708 3715 rnode_t *rp;
3709 3716 rddir_cache *rdc;
3710 3717 rddir_cache *nrdc;
3711 3718 rddir_cache *rrdc;
3712 3719 #ifdef DEBUG
3713 3720 int missed;
3714 3721 #endif
3715 3722 int doreadahead;
3716 3723 rddir_cache srdc;
3717 3724 avl_index_t where;
3718 3725
3719 3726 if (nfs_zone() != VTOMI(vp)->mi_zone)
3720 3727 return (EIO);
3721 3728 rp = VTOR(vp);
3722 3729
3723 3730 ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_READER));
3724 3731
3725 3732 /*
3726 3733 * Make sure that the directory cache is valid.
3727 3734 */
3728 3735 if (HAVE_RDDIR_CACHE(rp)) {
3729 3736 if (nfs_disable_rddir_cache) {
3730 3737 /*
3731 3738 * Setting nfs_disable_rddir_cache in /etc/system
3732 3739 * allows interoperability with servers that do not
3733 3740 * properly update the attributes of directories.
3734 3741 * Any cached information gets purged before an
3735 3742 * access is made to it.
3736 3743 */
3737 3744 nfs_purge_rddir_cache(vp);
3738 3745 } else {
3739 3746 error = nfs3_validate_caches(vp, cr);
3740 3747 if (error)
3741 3748 return (error);
3742 3749 }
3743 3750 }
3744 3751
3745 3752 /*
3746 3753 * It is possible that some servers may not be able to correctly
3747 3754 * handle a large READDIR or READDIRPLUS request due to bugs in
3748 3755 * their implementation. In order to continue to interoperate
3749 3756 * with them, this workaround is provided to limit the maximum
3750 3757 * size of a READDIRPLUS request to 1024. In any case, the request
3751 3758 * size is limited to MAXBSIZE.
3752 3759 */
3753 3760 count = MIN(uiop->uio_iov->iov_len,
3754 3761 nfs3_shrinkreaddir ? 1024 : MAXBSIZE);
3755 3762
3756 3763 nrdc = NULL;
3757 3764 #ifdef DEBUG
3758 3765 missed = 0;
3759 3766 #endif
3760 3767 top:
3761 3768 /*
3762 3769 * Short circuit last readdir which always returns 0 bytes.
3763 3770 * This can be done after the directory has been read through
3764 3771 * completely at least once. This will set r_direof which
3765 3772 * can be used to find the value of the last cookie.
3766 3773 */
3767 3774 mutex_enter(&rp->r_statelock);
3768 3775 if (rp->r_direof != NULL &&
3769 3776 uiop->uio_loffset == rp->r_direof->nfs3_ncookie) {
3770 3777 mutex_exit(&rp->r_statelock);
3771 3778 #ifdef DEBUG
3772 3779 nfs3_readdir_cache_shorts++;
3773 3780 #endif
3774 3781 if (eofp)
3775 3782 *eofp = 1;
3776 3783 if (nrdc != NULL)
3777 3784 rddir_cache_rele(nrdc);
3778 3785 return (0);
3779 3786 }
3780 3787 /*
3781 3788 * Look for a cache entry. Cache entries are identified
3782 3789 * by the NFS cookie value and the byte count requested.
3783 3790 */
3784 3791 srdc.nfs3_cookie = uiop->uio_loffset;
3785 3792 srdc.buflen = count;
3786 3793 rdc = avl_find(&rp->r_dir, &srdc, &where);
3787 3794 if (rdc != NULL) {
3788 3795 rddir_cache_hold(rdc);
3789 3796 /*
3790 3797 * If the cache entry is in the process of being
3791 3798 * filled in, wait until this completes. The
3792 3799 * RDDIRWAIT bit is set to indicate that someone
3793 3800 * is waiting and then the thread currently
3794 3801 * filling the entry is done, it should do a
3795 3802 * cv_broadcast to wakeup all of the threads
3796 3803 * waiting for it to finish.
3797 3804 */
3798 3805 if (rdc->flags & RDDIR) {
3799 3806 nfs_rw_exit(&rp->r_rwlock);
3800 3807 rdc->flags |= RDDIRWAIT;
3801 3808 #ifdef DEBUG
3802 3809 nfs3_readdir_cache_waits++;
3803 3810 #endif
3804 3811 if (!cv_wait_sig(&rdc->cv, &rp->r_statelock)) {
3805 3812 /*
3806 3813 * We got interrupted, probably
3807 3814 * the user typed ^C or an alarm
3808 3815 * fired. We free the new entry
3809 3816 * if we allocated one.
3810 3817 */
3811 3818 mutex_exit(&rp->r_statelock);
3812 3819 (void) nfs_rw_enter_sig(&rp->r_rwlock,
3813 3820 RW_READER, FALSE);
3814 3821 rddir_cache_rele(rdc);
3815 3822 if (nrdc != NULL)
3816 3823 rddir_cache_rele(nrdc);
3817 3824 return (EINTR);
3818 3825 }
3819 3826 mutex_exit(&rp->r_statelock);
3820 3827 (void) nfs_rw_enter_sig(&rp->r_rwlock,
3821 3828 RW_READER, FALSE);
3822 3829 rddir_cache_rele(rdc);
3823 3830 goto top;
3824 3831 }
3825 3832 /*
3826 3833 * Check to see if a readdir is required to
3827 3834 * fill the entry. If so, mark this entry
3828 3835 * as being filled, remove our reference,
3829 3836 * and branch to the code to fill the entry.
3830 3837 */
3831 3838 if (rdc->flags & RDDIRREQ) {
3832 3839 rdc->flags &= ~RDDIRREQ;
3833 3840 rdc->flags |= RDDIR;
3834 3841 if (nrdc != NULL)
3835 3842 rddir_cache_rele(nrdc);
3836 3843 nrdc = rdc;
3837 3844 mutex_exit(&rp->r_statelock);
3838 3845 goto bottom;
3839 3846 }
3840 3847 #ifdef DEBUG
3841 3848 if (!missed)
3842 3849 nfs3_readdir_cache_hits++;
3843 3850 #endif
3844 3851 /*
3845 3852 * If an error occurred while attempting
3846 3853 * to fill the cache entry, just return it.
3847 3854 */
3848 3855 if (rdc->error) {
3849 3856 error = rdc->error;
3850 3857 mutex_exit(&rp->r_statelock);
3851 3858 rddir_cache_rele(rdc);
3852 3859 if (nrdc != NULL)
3853 3860 rddir_cache_rele(nrdc);
3854 3861 return (error);
3855 3862 }
3856 3863
3857 3864 /*
3858 3865 * The cache entry is complete and good,
3859 3866 * copyout the dirent structs to the calling
3860 3867 * thread.
3861 3868 */
3862 3869 error = uiomove(rdc->entries, rdc->entlen, UIO_READ, uiop);
3863 3870
3864 3871 /*
3865 3872 * If no error occurred during the copyout,
3866 3873 * update the offset in the uio struct to
3867 3874 * contain the value of the next cookie
3868 3875 * and set the eof value appropriately.
3869 3876 */
3870 3877 if (!error) {
3871 3878 uiop->uio_loffset = rdc->nfs3_ncookie;
3872 3879 if (eofp)
3873 3880 *eofp = rdc->eof;
3874 3881 }
3875 3882
3876 3883 /*
3877 3884 * Decide whether to do readahead.
3878 3885 *
3879 3886 * Don't if have already read to the end of
3880 3887 * directory. There is nothing more to read.
3881 3888 *
3882 3889 * Don't if the application is not doing
3883 3890 * lookups in the directory. The readahead
3884 3891 * is only effective if the application can
3885 3892 * be doing work while an async thread is
3886 3893 * handling the over the wire request.
3887 3894 */
3888 3895 if (rdc->eof) {
3889 3896 rp->r_direof = rdc;
3890 3897 doreadahead = FALSE;
3891 3898 } else if (!(rp->r_flags & RLOOKUP))
3892 3899 doreadahead = FALSE;
3893 3900 else
3894 3901 doreadahead = TRUE;
3895 3902
3896 3903 if (!doreadahead) {
3897 3904 mutex_exit(&rp->r_statelock);
3898 3905 rddir_cache_rele(rdc);
3899 3906 if (nrdc != NULL)
3900 3907 rddir_cache_rele(nrdc);
3901 3908 return (error);
3902 3909 }
3903 3910
3904 3911 /*
3905 3912 * Check to see whether we found an entry
3906 3913 * for the readahead. If so, we don't need
3907 3914 * to do anything further, so free the new
3908 3915 * entry if one was allocated. Otherwise,
3909 3916 * allocate a new entry, add it to the cache,
3910 3917 * and then initiate an asynchronous readdir
3911 3918 * operation to fill it.
3912 3919 */
3913 3920 srdc.nfs3_cookie = rdc->nfs3_ncookie;
3914 3921 srdc.buflen = count;
3915 3922 rrdc = avl_find(&rp->r_dir, &srdc, &where);
3916 3923 if (rrdc != NULL) {
3917 3924 if (nrdc != NULL)
3918 3925 rddir_cache_rele(nrdc);
3919 3926 } else {
3920 3927 if (nrdc != NULL)
3921 3928 rrdc = nrdc;
3922 3929 else {
3923 3930 rrdc = rddir_cache_alloc(KM_NOSLEEP);
3924 3931 }
3925 3932 if (rrdc != NULL) {
3926 3933 rrdc->nfs3_cookie = rdc->nfs3_ncookie;
3927 3934 rrdc->buflen = count;
3928 3935 avl_insert(&rp->r_dir, rrdc, where);
3929 3936 rddir_cache_hold(rrdc);
3930 3937 mutex_exit(&rp->r_statelock);
3931 3938 rddir_cache_rele(rdc);
3932 3939 #ifdef DEBUG
3933 3940 nfs3_readdir_readahead++;
3934 3941 #endif
3935 3942 nfs_async_readdir(vp, rrdc, cr, do_nfs3readdir);
3936 3943 return (error);
3937 3944 }
3938 3945 }
3939 3946
3940 3947 mutex_exit(&rp->r_statelock);
3941 3948 rddir_cache_rele(rdc);
3942 3949 return (error);
3943 3950 }
3944 3951
3945 3952 /*
3946 3953 * Didn't find an entry in the cache. Construct a new empty
3947 3954 * entry and link it into the cache. Other processes attempting
3948 3955 * to access this entry will need to wait until it is filled in.
3949 3956 *
3950 3957 * Since kmem_alloc may block, another pass through the cache
3951 3958 * will need to be taken to make sure that another process
3952 3959 * hasn't already added an entry to the cache for this request.
3953 3960 */
3954 3961 if (nrdc == NULL) {
3955 3962 mutex_exit(&rp->r_statelock);
3956 3963 nrdc = rddir_cache_alloc(KM_SLEEP);
3957 3964 nrdc->nfs3_cookie = uiop->uio_loffset;
3958 3965 nrdc->buflen = count;
3959 3966 goto top;
3960 3967 }
3961 3968
3962 3969 /*
3963 3970 * Add this entry to the cache.
3964 3971 */
3965 3972 avl_insert(&rp->r_dir, nrdc, where);
3966 3973 rddir_cache_hold(nrdc);
3967 3974 mutex_exit(&rp->r_statelock);
3968 3975
3969 3976 bottom:
3970 3977 #ifdef DEBUG
3971 3978 missed = 1;
3972 3979 nfs3_readdir_cache_misses++;
3973 3980 #endif
3974 3981 /*
3975 3982 * Do the readdir. This routine decides whether to use
3976 3983 * READDIR or READDIRPLUS.
3977 3984 */
3978 3985 error = do_nfs3readdir(vp, nrdc, cr);
3979 3986
3980 3987 /*
3981 3988 * If this operation failed, just return the error which occurred.
3982 3989 */
3983 3990 if (error != 0)
3984 3991 return (error);
3985 3992
3986 3993 /*
3987 3994 * Since the RPC operation will have taken sometime and blocked
3988 3995 * this process, another pass through the cache will need to be
3989 3996 * taken to find the correct cache entry. It is possible that
3990 3997 * the correct cache entry will not be there (although one was
3991 3998 * added) because the directory changed during the RPC operation
3992 3999 * and the readdir cache was flushed. In this case, just start
3993 4000 * over. It is hoped that this will not happen too often... :-)
3994 4001 */
3995 4002 nrdc = NULL;
3996 4003 goto top;
3997 4004 /* NOTREACHED */
3998 4005 }
3999 4006
4000 4007 static int
4001 4008 do_nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
4002 4009 {
4003 4010 int error;
4004 4011 rnode_t *rp;
4005 4012 mntinfo_t *mi;
4006 4013
4007 4014 rp = VTOR(vp);
4008 4015 mi = VTOMI(vp);
4009 4016 ASSERT(nfs_zone() == mi->mi_zone);
4010 4017 /*
4011 4018 * Issue the proper request.
4012 4019 *
4013 4020 * If the server does not support READDIRPLUS, then use READDIR.
4014 4021 *
4015 4022 * Otherwise --
4016 4023 * Issue a READDIRPLUS if reading to fill an empty cache or if
4017 4024 * an application has performed a lookup in the directory which
4018 4025 * required an over the wire lookup. The use of READDIRPLUS
4019 4026 * will help to (re)populate the DNLC.
4020 4027 */
4021 4028 if (!(mi->mi_flags & MI_READDIRONLY) &&
4022 4029 (rp->r_flags & (RLOOKUP | RREADDIRPLUS))) {
4023 4030 if (rp->r_flags & RREADDIRPLUS) {
4024 4031 mutex_enter(&rp->r_statelock);
4025 4032 rp->r_flags &= ~RREADDIRPLUS;
4026 4033 mutex_exit(&rp->r_statelock);
4027 4034 }
4028 4035 nfs3readdirplus(vp, rdc, cr);
4029 4036 if (rdc->error == EOPNOTSUPP)
4030 4037 nfs3readdir(vp, rdc, cr);
4031 4038 } else
4032 4039 nfs3readdir(vp, rdc, cr);
4033 4040
4034 4041 mutex_enter(&rp->r_statelock);
4035 4042 rdc->flags &= ~RDDIR;
4036 4043 if (rdc->flags & RDDIRWAIT) {
4037 4044 rdc->flags &= ~RDDIRWAIT;
4038 4045 cv_broadcast(&rdc->cv);
4039 4046 }
4040 4047 error = rdc->error;
4041 4048 if (error)
4042 4049 rdc->flags |= RDDIRREQ;
4043 4050 mutex_exit(&rp->r_statelock);
4044 4051
4045 4052 rddir_cache_rele(rdc);
4046 4053
4047 4054 return (error);
4048 4055 }
4049 4056
4050 4057 static void
4051 4058 nfs3readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
4052 4059 {
4053 4060 int error;
4054 4061 READDIR3args args;
4055 4062 READDIR3vres res;
4056 4063 vattr_t dva;
4057 4064 rnode_t *rp;
4058 4065 int douprintf;
4059 4066 failinfo_t fi, *fip = NULL;
4060 4067 mntinfo_t *mi;
4061 4068 hrtime_t t;
4062 4069
4063 4070 rp = VTOR(vp);
4064 4071 mi = VTOMI(vp);
4065 4072 ASSERT(nfs_zone() == mi->mi_zone);
4066 4073
4067 4074 args.dir = *RTOFH3(rp);
4068 4075 args.cookie = (cookie3)rdc->nfs3_cookie;
4069 4076 args.cookieverf = rp->r_cookieverf;
4070 4077 args.count = rdc->buflen;
4071 4078
4072 4079 /*
4073 4080 * NFS client failover support
4074 4081 * suppress failover unless we have a zero cookie
4075 4082 */
4076 4083 if (args.cookie == (cookie3) 0) {
4077 4084 fi.vp = vp;
4078 4085 fi.fhp = (caddr_t)&args.dir;
4079 4086 fi.copyproc = nfs3copyfh;
4080 4087 fi.lookupproc = nfs3lookup;
4081 4088 fi.xattrdirproc = acl_getxattrdir3;
4082 4089 fip = &fi;
4083 4090 }
4084 4091
4085 4092 #ifdef DEBUG
4086 4093 rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
4087 4094 #else
4088 4095 rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
4089 4096 #endif
4090 4097
4091 4098 res.entries = (dirent64_t *)rdc->entries;
4092 4099 res.entries_size = rdc->buflen;
4093 4100 res.dir_attributes.fres.vap = &dva;
4094 4101 res.dir_attributes.fres.vp = vp;
4095 4102 res.loff = rdc->nfs3_cookie;
4096 4103
4097 4104 douprintf = 1;
4098 4105
4099 4106 if (mi->mi_io_kstats) {
4100 4107 mutex_enter(&mi->mi_lock);
4101 4108 kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
4102 4109 mutex_exit(&mi->mi_lock);
4103 4110 }
4104 4111
4105 4112 t = gethrtime();
4106 4113
4107 4114 error = rfs3call(VTOMI(vp), NFSPROC3_READDIR,
4108 4115 xdr_READDIR3args, (caddr_t)&args,
4109 4116 xdr_READDIR3vres, (caddr_t)&res, cr,
4110 4117 &douprintf, &res.status, 0, fip);
4111 4118
4112 4119 if (mi->mi_io_kstats) {
4113 4120 mutex_enter(&mi->mi_lock);
4114 4121 kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
4115 4122 mutex_exit(&mi->mi_lock);
4116 4123 }
4117 4124
4118 4125 if (error)
4119 4126 goto err;
4120 4127
4121 4128 nfs3_cache_post_op_vattr(vp, &res.dir_attributes, t, cr);
4122 4129
4123 4130 error = geterrno3(res.status);
4124 4131 if (error) {
4125 4132 PURGE_STALE_FH(error, vp, cr);
4126 4133 goto err;
4127 4134 }
4128 4135
4129 4136 if (mi->mi_io_kstats) {
4130 4137 mutex_enter(&mi->mi_lock);
4131 4138 KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4132 4139 KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4133 4140 mutex_exit(&mi->mi_lock);
4134 4141 }
4135 4142
4136 4143 rdc->nfs3_ncookie = res.loff;
4137 4144 rp->r_cookieverf = res.cookieverf;
4138 4145 rdc->eof = res.eof ? 1 : 0;
4139 4146 rdc->entlen = res.size;
4140 4147 ASSERT(rdc->entlen <= rdc->buflen);
4141 4148 rdc->error = 0;
4142 4149 return;
4143 4150
4144 4151 err:
4145 4152 kmem_free(rdc->entries, rdc->buflen);
4146 4153 rdc->entries = NULL;
4147 4154 rdc->error = error;
4148 4155 }
4149 4156
4150 4157 /*
4151 4158 * Read directory entries.
4152 4159 * There are some weird things to look out for here. The uio_loffset
4153 4160 * field is either 0 or it is the offset returned from a previous
4154 4161 * readdir. It is an opaque value used by the server to find the
4155 4162 * correct directory block to read. The count field is the number
4156 4163 * of blocks to read on the server. This is advisory only, the server
4157 4164 * may return only one block's worth of entries. Entries may be compressed
4158 4165 * on the server.
4159 4166 */
4160 4167 static void
4161 4168 nfs3readdirplus(vnode_t *vp, rddir_cache *rdc, cred_t *cr)
4162 4169 {
4163 4170 int error;
4164 4171 READDIRPLUS3args args;
4165 4172 READDIRPLUS3vres res;
4166 4173 vattr_t dva;
4167 4174 rnode_t *rp;
4168 4175 mntinfo_t *mi;
4169 4176 int douprintf;
4170 4177 failinfo_t fi, *fip = NULL;
4171 4178
4172 4179 rp = VTOR(vp);
4173 4180 mi = VTOMI(vp);
4174 4181 ASSERT(nfs_zone() == mi->mi_zone);
4175 4182
4176 4183 args.dir = *RTOFH3(rp);
4177 4184 args.cookie = (cookie3)rdc->nfs3_cookie;
4178 4185 args.cookieverf = rp->r_cookieverf;
4179 4186 args.dircount = rdc->buflen;
4180 4187 args.maxcount = mi->mi_tsize;
4181 4188
4182 4189 /*
4183 4190 * NFS client failover support
4184 4191 * suppress failover unless we have a zero cookie
4185 4192 */
4186 4193 if (args.cookie == (cookie3)0) {
4187 4194 fi.vp = vp;
4188 4195 fi.fhp = (caddr_t)&args.dir;
4189 4196 fi.copyproc = nfs3copyfh;
4190 4197 fi.lookupproc = nfs3lookup;
4191 4198 fi.xattrdirproc = acl_getxattrdir3;
4192 4199 fip = &fi;
4193 4200 }
4194 4201
4195 4202 #ifdef DEBUG
4196 4203 rdc->entries = rddir_cache_buf_alloc(rdc->buflen, KM_SLEEP);
4197 4204 #else
4198 4205 rdc->entries = kmem_alloc(rdc->buflen, KM_SLEEP);
4199 4206 #endif
4200 4207
4201 4208 res.entries = (dirent64_t *)rdc->entries;
4202 4209 res.entries_size = rdc->buflen;
4203 4210 res.dir_attributes.fres.vap = &dva;
4204 4211 res.dir_attributes.fres.vp = vp;
4205 4212 res.loff = rdc->nfs3_cookie;
4206 4213 res.credentials = cr;
4207 4214
4208 4215 douprintf = 1;
4209 4216
4210 4217 if (mi->mi_io_kstats) {
4211 4218 mutex_enter(&mi->mi_lock);
4212 4219 kstat_runq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
4213 4220 mutex_exit(&mi->mi_lock);
4214 4221 }
4215 4222
4216 4223 res.time = gethrtime();
4217 4224
4218 4225 error = rfs3call(mi, NFSPROC3_READDIRPLUS,
4219 4226 xdr_READDIRPLUS3args, (caddr_t)&args,
4220 4227 xdr_READDIRPLUS3vres, (caddr_t)&res, cr,
4221 4228 &douprintf, &res.status, 0, fip);
4222 4229
4223 4230 if (mi->mi_io_kstats) {
4224 4231 mutex_enter(&mi->mi_lock);
4225 4232 kstat_runq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
4226 4233 mutex_exit(&mi->mi_lock);
4227 4234 }
4228 4235
4229 4236 if (error) {
4230 4237 goto err;
4231 4238 }
4232 4239
4233 4240 nfs3_cache_post_op_vattr(vp, &res.dir_attributes, res.time, cr);
4234 4241
4235 4242 error = geterrno3(res.status);
4236 4243 if (error) {
4237 4244 PURGE_STALE_FH(error, vp, cr);
4238 4245 if (error == EOPNOTSUPP) {
4239 4246 mutex_enter(&mi->mi_lock);
4240 4247 mi->mi_flags |= MI_READDIRONLY;
4241 4248 mutex_exit(&mi->mi_lock);
4242 4249 }
4243 4250 goto err;
4244 4251 }
4245 4252
4246 4253 if (mi->mi_io_kstats) {
4247 4254 mutex_enter(&mi->mi_lock);
4248 4255 KSTAT_IO_PTR(mi->mi_io_kstats)->reads++;
4249 4256 KSTAT_IO_PTR(mi->mi_io_kstats)->nread += res.size;
4250 4257 mutex_exit(&mi->mi_lock);
4251 4258 }
4252 4259
4253 4260 rdc->nfs3_ncookie = res.loff;
4254 4261 rp->r_cookieverf = res.cookieverf;
4255 4262 rdc->eof = res.eof ? 1 : 0;
4256 4263 rdc->entlen = res.size;
4257 4264 ASSERT(rdc->entlen <= rdc->buflen);
4258 4265 rdc->error = 0;
4259 4266
4260 4267 return;
4261 4268
4262 4269 err:
4263 4270 kmem_free(rdc->entries, rdc->buflen);
4264 4271 rdc->entries = NULL;
4265 4272 rdc->error = error;
4266 4273 }
4267 4274
4268 4275 #ifdef DEBUG
4269 4276 static int nfs3_bio_do_stop = 0;
4270 4277 #endif
4271 4278
4272 4279 static int
4273 4280 nfs3_bio(struct buf *bp, stable_how *stab_comm, cred_t *cr)
4274 4281 {
4275 4282 rnode_t *rp = VTOR(bp->b_vp);
4276 4283 int count;
4277 4284 int error;
4278 4285 cred_t *cred;
4279 4286 offset_t offset;
4280 4287
4281 4288 ASSERT(nfs_zone() == VTOMI(bp->b_vp)->mi_zone);
4282 4289 offset = ldbtob(bp->b_lblkno);
4283 4290
4284 4291 DTRACE_IO1(start, struct buf *, bp);
4285 4292
4286 4293 if (bp->b_flags & B_READ) {
4287 4294 mutex_enter(&rp->r_statelock);
4288 4295 if (rp->r_cred != NULL) {
4289 4296 cred = rp->r_cred;
4290 4297 crhold(cred);
4291 4298 } else {
4292 4299 rp->r_cred = cr;
4293 4300 crhold(cr);
4294 4301 cred = cr;
4295 4302 crhold(cred);
4296 4303 }
4297 4304 mutex_exit(&rp->r_statelock);
4298 4305 read_again:
4299 4306 error = bp->b_error = nfs3read(bp->b_vp, bp->b_un.b_addr,
4300 4307 offset, bp->b_bcount, &bp->b_resid, cred);
4301 4308 crfree(cred);
4302 4309 if (!error) {
4303 4310 if (bp->b_resid) {
4304 4311 /*
4305 4312 * Didn't get it all because we hit EOF,
4306 4313 * zero all the memory beyond the EOF.
4307 4314 */
4308 4315 /* bzero(rdaddr + */
4309 4316 bzero(bp->b_un.b_addr +
4310 4317 bp->b_bcount - bp->b_resid, bp->b_resid);
4311 4318 }
4312 4319 mutex_enter(&rp->r_statelock);
4313 4320 if (bp->b_resid == bp->b_bcount &&
4314 4321 offset >= rp->r_size) {
4315 4322 /*
4316 4323 * We didn't read anything at all as we are
4317 4324 * past EOF. Return an error indicator back
4318 4325 * but don't destroy the pages (yet).
4319 4326 */
4320 4327 error = NFS_EOF;
4321 4328 }
4322 4329 mutex_exit(&rp->r_statelock);
4323 4330 } else if (error == EACCES) {
4324 4331 mutex_enter(&rp->r_statelock);
4325 4332 if (cred != cr) {
4326 4333 if (rp->r_cred != NULL)
4327 4334 crfree(rp->r_cred);
4328 4335 rp->r_cred = cr;
4329 4336 crhold(cr);
4330 4337 cred = cr;
4331 4338 crhold(cred);
4332 4339 mutex_exit(&rp->r_statelock);
4333 4340 goto read_again;
4334 4341 }
4335 4342 mutex_exit(&rp->r_statelock);
4336 4343 }
4337 4344 } else {
4338 4345 if (!(rp->r_flags & RSTALE)) {
4339 4346 mutex_enter(&rp->r_statelock);
4340 4347 if (rp->r_cred != NULL) {
4341 4348 cred = rp->r_cred;
4342 4349 crhold(cred);
4343 4350 } else {
4344 4351 rp->r_cred = cr;
4345 4352 crhold(cr);
4346 4353 cred = cr;
4347 4354 crhold(cred);
4348 4355 }
4349 4356 mutex_exit(&rp->r_statelock);
4350 4357 write_again:
4351 4358 mutex_enter(&rp->r_statelock);
4352 4359 count = MIN(bp->b_bcount, rp->r_size - offset);
4353 4360 mutex_exit(&rp->r_statelock);
4354 4361 if (count < 0)
4355 4362 cmn_err(CE_PANIC, "nfs3_bio: write count < 0");
4356 4363 #ifdef DEBUG
4357 4364 if (count == 0) {
4358 4365 zcmn_err(getzoneid(), CE_WARN,
4359 4366 "nfs3_bio: zero length write at %lld",
4360 4367 offset);
4361 4368 nfs_printfhandle(&rp->r_fh);
4362 4369 if (nfs3_bio_do_stop)
4363 4370 debug_enter("nfs3_bio");
4364 4371 }
4365 4372 #endif
4366 4373 error = nfs3write(bp->b_vp, bp->b_un.b_addr, offset,
4367 4374 count, cred, stab_comm);
4368 4375 if (error == EACCES) {
4369 4376 mutex_enter(&rp->r_statelock);
4370 4377 if (cred != cr) {
4371 4378 if (rp->r_cred != NULL)
4372 4379 crfree(rp->r_cred);
4373 4380 rp->r_cred = cr;
4374 4381 crhold(cr);
4375 4382 crfree(cred);
4376 4383 cred = cr;
4377 4384 crhold(cred);
4378 4385 mutex_exit(&rp->r_statelock);
4379 4386 goto write_again;
4380 4387 }
4381 4388 mutex_exit(&rp->r_statelock);
4382 4389 }
4383 4390 bp->b_error = error;
4384 4391 if (error && error != EINTR) {
4385 4392 /*
4386 4393 * Don't print EDQUOT errors on the console.
4387 4394 * Don't print asynchronous EACCES errors.
4388 4395 * Don't print EFBIG errors.
4389 4396 * Print all other write errors.
4390 4397 */
4391 4398 if (error != EDQUOT && error != EFBIG &&
4392 4399 (error != EACCES ||
4393 4400 !(bp->b_flags & B_ASYNC)))
4394 4401 nfs_write_error(bp->b_vp, error, cred);
4395 4402 /*
4396 4403 * Update r_error and r_flags as appropriate.
4397 4404 * If the error was ESTALE, then mark the
4398 4405 * rnode as not being writeable and save
4399 4406 * the error status. Otherwise, save any
4400 4407 * errors which occur from asynchronous
4401 4408 * page invalidations. Any errors occurring
4402 4409 * from other operations should be saved
4403 4410 * by the caller.
4404 4411 */
4405 4412 mutex_enter(&rp->r_statelock);
4406 4413 if (error == ESTALE) {
4407 4414 rp->r_flags |= RSTALE;
4408 4415 if (!rp->r_error)
4409 4416 rp->r_error = error;
4410 4417 } else if (!rp->r_error &&
4411 4418 (bp->b_flags &
4412 4419 (B_INVAL|B_FORCE|B_ASYNC)) ==
4413 4420 (B_INVAL|B_FORCE|B_ASYNC)) {
4414 4421 rp->r_error = error;
4415 4422 }
4416 4423 mutex_exit(&rp->r_statelock);
4417 4424 }
4418 4425 crfree(cred);
4419 4426 } else {
4420 4427 error = rp->r_error;
4421 4428 /*
4422 4429 * A close may have cleared r_error, if so,
4423 4430 * propagate ESTALE error return properly
4424 4431 */
4425 4432 if (error == 0)
4426 4433 error = ESTALE;
4427 4434 }
4428 4435 }
4429 4436
4430 4437 if (error != 0 && error != NFS_EOF)
4431 4438 bp->b_flags |= B_ERROR;
4432 4439
4433 4440 DTRACE_IO1(done, struct buf *, bp);
4434 4441
4435 4442 return (error);
4436 4443 }
4437 4444
4438 4445 /* ARGSUSED */
4439 4446 static int
4440 4447 nfs3_fid(vnode_t *vp, fid_t *fidp, caller_context_t *ct)
4441 4448 {
4442 4449 rnode_t *rp;
4443 4450
4444 4451 if (nfs_zone() != VTOMI(vp)->mi_zone)
4445 4452 return (EIO);
4446 4453 rp = VTOR(vp);
4447 4454
4448 4455 if (fidp->fid_len < (ushort_t)rp->r_fh.fh_len) {
4449 4456 fidp->fid_len = rp->r_fh.fh_len;
4450 4457 return (ENOSPC);
4451 4458 }
4452 4459 fidp->fid_len = rp->r_fh.fh_len;
4453 4460 bcopy(rp->r_fh.fh_buf, fidp->fid_data, fidp->fid_len);
4454 4461 return (0);
4455 4462 }
4456 4463
4457 4464 /* ARGSUSED2 */
4458 4465 static int
4459 4466 nfs3_rwlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4460 4467 {
4461 4468 rnode_t *rp = VTOR(vp);
4462 4469
4463 4470 if (!write_lock) {
4464 4471 (void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4465 4472 return (V_WRITELOCK_FALSE);
4466 4473 }
4467 4474
4468 4475 if ((rp->r_flags & RDIRECTIO) || (VTOMI(vp)->mi_flags & MI_DIRECTIO)) {
4469 4476 (void) nfs_rw_enter_sig(&rp->r_rwlock, RW_READER, FALSE);
4470 4477 if (rp->r_mapcnt == 0 && !vn_has_cached_data(vp))
4471 4478 return (V_WRITELOCK_FALSE);
4472 4479 nfs_rw_exit(&rp->r_rwlock);
4473 4480 }
4474 4481
4475 4482 (void) nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, FALSE);
4476 4483 return (V_WRITELOCK_TRUE);
4477 4484 }
4478 4485
4479 4486 /* ARGSUSED */
4480 4487 static void
4481 4488 nfs3_rwunlock(vnode_t *vp, int write_lock, caller_context_t *ctp)
4482 4489 {
4483 4490 rnode_t *rp = VTOR(vp);
4484 4491
4485 4492 nfs_rw_exit(&rp->r_rwlock);
4486 4493 }
4487 4494
4488 4495 /* ARGSUSED */
4489 4496 static int
4490 4497 nfs3_seek(vnode_t *vp, offset_t ooff, offset_t *noffp, caller_context_t *ct)
4491 4498 {
4492 4499
4493 4500 /*
4494 4501 * Because we stuff the readdir cookie into the offset field
4495 4502 * someone may attempt to do an lseek with the cookie which
4496 4503 * we want to succeed.
4497 4504 */
4498 4505 if (vp->v_type == VDIR)
4499 4506 return (0);
4500 4507 if (*noffp < 0)
4501 4508 return (EINVAL);
4502 4509 return (0);
4503 4510 }
4504 4511
4505 4512 /*
4506 4513 * number of nfs3_bsize blocks to read ahead.
4507 4514 */
4508 4515 static int nfs3_nra = 4;
4509 4516
4510 4517 #ifdef DEBUG
4511 4518 static int nfs3_lostpage = 0; /* number of times we lost original page */
4512 4519 #endif
4513 4520
4514 4521 /*
4515 4522 * Return all the pages from [off..off+len) in file
4516 4523 */
4517 4524 /* ARGSUSED */
4518 4525 static int
4519 4526 nfs3_getpage(vnode_t *vp, offset_t off, size_t len, uint_t *protp,
4520 4527 page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4521 4528 enum seg_rw rw, cred_t *cr, caller_context_t *ct)
4522 4529 {
4523 4530 rnode_t *rp;
4524 4531 int error;
4525 4532 mntinfo_t *mi;
4526 4533
4527 4534 if (vp->v_flag & VNOMAP)
4528 4535 return (ENOSYS);
4529 4536
4530 4537 if (nfs_zone() != VTOMI(vp)->mi_zone)
4531 4538 return (EIO);
4532 4539 if (protp != NULL)
4533 4540 *protp = PROT_ALL;
4534 4541
4535 4542 /*
4536 4543 * Now valididate that the caches are up to date.
4537 4544 */
4538 4545 error = nfs3_validate_caches(vp, cr);
4539 4546 if (error)
4540 4547 return (error);
4541 4548
4542 4549 rp = VTOR(vp);
4543 4550 mi = VTOMI(vp);
4544 4551 retry:
4545 4552 mutex_enter(&rp->r_statelock);
4546 4553
4547 4554 /*
4548 4555 * Don't create dirty pages faster than they
4549 4556 * can be cleaned so that the system doesn't
4550 4557 * get imbalanced. If the async queue is
4551 4558 * maxed out, then wait for it to drain before
4552 4559 * creating more dirty pages. Also, wait for
4553 4560 * any threads doing pagewalks in the vop_getattr
4554 4561 * entry points so that they don't block for
4555 4562 * long periods.
4556 4563 */
4557 4564 if (rw == S_CREATE) {
4558 4565 while ((mi->mi_max_threads != 0 &&
4559 4566 rp->r_awcount > 2 * mi->mi_max_threads) ||
4560 4567 rp->r_gcount > 0)
4561 4568 cv_wait(&rp->r_cv, &rp->r_statelock);
4562 4569 }
4563 4570
4564 4571 /*
4565 4572 * If we are getting called as a side effect of an nfs_write()
4566 4573 * operation the local file size might not be extended yet.
4567 4574 * In this case we want to be able to return pages of zeroes.
4568 4575 */
4569 4576 if (off + len > rp->r_size + PAGEOFFSET && seg != segkmap) {
4570 4577 mutex_exit(&rp->r_statelock);
4571 4578 return (EFAULT); /* beyond EOF */
4572 4579 }
4573 4580
4574 4581 mutex_exit(&rp->r_statelock);
4575 4582
4576 4583 if (len <= PAGESIZE) {
4577 4584 error = nfs3_getapage(vp, off, len, protp, pl, plsz,
4578 4585 seg, addr, rw, cr);
4579 4586 } else {
4580 4587 error = pvn_getpages(nfs3_getapage, vp, off, len, protp,
4581 4588 pl, plsz, seg, addr, rw, cr);
4582 4589 }
4583 4590
4584 4591 switch (error) {
4585 4592 case NFS_EOF:
4586 4593 nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
4587 4594 goto retry;
4588 4595 case ESTALE:
4589 4596 PURGE_STALE_FH(error, vp, cr);
4590 4597 }
4591 4598
4592 4599 return (error);
4593 4600 }
4594 4601
4595 4602 /*
4596 4603 * Called from pvn_getpages or nfs3_getpage to get a particular page.
4597 4604 */
4598 4605 /* ARGSUSED */
4599 4606 static int
4600 4607 nfs3_getapage(vnode_t *vp, u_offset_t off, size_t len, uint_t *protp,
4601 4608 page_t *pl[], size_t plsz, struct seg *seg, caddr_t addr,
4602 4609 enum seg_rw rw, cred_t *cr)
4603 4610 {
4604 4611 rnode_t *rp;
4605 4612 uint_t bsize;
4606 4613 struct buf *bp;
4607 4614 page_t *pp;
4608 4615 u_offset_t lbn;
4609 4616 u_offset_t io_off;
4610 4617 u_offset_t blkoff;
4611 4618 u_offset_t rablkoff;
4612 4619 size_t io_len;
4613 4620 uint_t blksize;
4614 4621 int error;
4615 4622 int readahead;
4616 4623 int readahead_issued = 0;
4617 4624 int ra_window; /* readahead window */
4618 4625 page_t *pagefound;
4619 4626 page_t *savepp;
4620 4627
4621 4628 if (nfs_zone() != VTOMI(vp)->mi_zone)
4622 4629 return (EIO);
4623 4630 rp = VTOR(vp);
4624 4631 bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4625 4632
4626 4633 reread:
4627 4634 bp = NULL;
4628 4635 pp = NULL;
4629 4636 pagefound = NULL;
4630 4637
4631 4638 if (pl != NULL)
4632 4639 pl[0] = NULL;
4633 4640
4634 4641 error = 0;
4635 4642 lbn = off / bsize;
4636 4643 blkoff = lbn * bsize;
4637 4644
4638 4645 /*
4639 4646 * Queueing up the readahead before doing the synchronous read
4640 4647 * results in a significant increase in read throughput because
4641 4648 * of the increased parallelism between the async threads and
4642 4649 * the process context.
4643 4650 */
4644 4651 if ((off & ((vp->v_vfsp->vfs_bsize) - 1)) == 0 &&
4645 4652 rw != S_CREATE &&
4646 4653 !(vp->v_flag & VNOCACHE)) {
4647 4654 mutex_enter(&rp->r_statelock);
4648 4655
4649 4656 /*
4650 4657 * Calculate the number of readaheads to do.
4651 4658 * a) No readaheads at offset = 0.
4652 4659 * b) Do maximum(nfs3_nra) readaheads when the readahead
4653 4660 * window is closed.
4654 4661 * c) Do readaheads between 1 to (nfs3_nra - 1) depending
4655 4662 * upon how far the readahead window is open or close.
4656 4663 * d) No readaheads if rp->r_nextr is not within the scope
4657 4664 * of the readahead window (random i/o).
4658 4665 */
4659 4666
4660 4667 if (off == 0)
4661 4668 readahead = 0;
4662 4669 else if (blkoff == rp->r_nextr)
4663 4670 readahead = nfs3_nra;
4664 4671 else if (rp->r_nextr > blkoff &&
4665 4672 ((ra_window = (rp->r_nextr - blkoff) / bsize)
4666 4673 <= (nfs3_nra - 1)))
4667 4674 readahead = nfs3_nra - ra_window;
4668 4675 else
4669 4676 readahead = 0;
4670 4677
4671 4678 rablkoff = rp->r_nextr;
4672 4679 while (readahead > 0 && rablkoff + bsize < rp->r_size) {
4673 4680 mutex_exit(&rp->r_statelock);
4674 4681 if (nfs_async_readahead(vp, rablkoff + bsize,
4675 4682 addr + (rablkoff + bsize - off), seg, cr,
4676 4683 nfs3_readahead) < 0) {
4677 4684 mutex_enter(&rp->r_statelock);
4678 4685 break;
4679 4686 }
4680 4687 readahead--;
4681 4688 rablkoff += bsize;
4682 4689 /*
4683 4690 * Indicate that we did a readahead so
4684 4691 * readahead offset is not updated
4685 4692 * by the synchronous read below.
4686 4693 */
4687 4694 readahead_issued = 1;
4688 4695 mutex_enter(&rp->r_statelock);
4689 4696 /*
4690 4697 * set readahead offset to
4691 4698 * offset of last async readahead
4692 4699 * request.
4693 4700 */
4694 4701 rp->r_nextr = rablkoff;
4695 4702 }
4696 4703 mutex_exit(&rp->r_statelock);
4697 4704 }
4698 4705
4699 4706 again:
4700 4707 if ((pagefound = page_exists(vp, off)) == NULL) {
4701 4708 if (pl == NULL) {
4702 4709 (void) nfs_async_readahead(vp, blkoff, addr, seg, cr,
4703 4710 nfs3_readahead);
4704 4711 } else if (rw == S_CREATE) {
4705 4712 /*
4706 4713 * Block for this page is not allocated, or the offset
4707 4714 * is beyond the current allocation size, or we're
4708 4715 * allocating a swap slot and the page was not found,
4709 4716 * so allocate it and return a zero page.
4710 4717 */
4711 4718 if ((pp = page_create_va(vp, off,
4712 4719 PAGESIZE, PG_WAIT, seg, addr)) == NULL)
4713 4720 cmn_err(CE_PANIC, "nfs3_getapage: page_create");
4714 4721 io_len = PAGESIZE;
4715 4722 mutex_enter(&rp->r_statelock);
4716 4723 rp->r_nextr = off + PAGESIZE;
4717 4724 mutex_exit(&rp->r_statelock);
4718 4725 } else {
4719 4726 /*
4720 4727 * Need to go to server to get a BLOCK, exception to
4721 4728 * that being while reading at offset = 0 or doing
4722 4729 * random i/o, in that case read only a PAGE.
4723 4730 */
4724 4731 mutex_enter(&rp->r_statelock);
4725 4732 if (blkoff < rp->r_size &&
4726 4733 blkoff + bsize >= rp->r_size) {
4727 4734 /*
4728 4735 * If only a block or less is left in
4729 4736 * the file, read all that is remaining.
4730 4737 */
4731 4738 if (rp->r_size <= off) {
4732 4739 /*
4733 4740 * Trying to access beyond EOF,
4734 4741 * set up to get at least one page.
4735 4742 */
4736 4743 blksize = off + PAGESIZE - blkoff;
4737 4744 } else
4738 4745 blksize = rp->r_size - blkoff;
4739 4746 } else if ((off == 0) ||
4740 4747 (off != rp->r_nextr && !readahead_issued)) {
4741 4748 blksize = PAGESIZE;
4742 4749 blkoff = off; /* block = page here */
4743 4750 } else
4744 4751 blksize = bsize;
4745 4752 mutex_exit(&rp->r_statelock);
4746 4753
4747 4754 pp = pvn_read_kluster(vp, off, seg, addr, &io_off,
4748 4755 &io_len, blkoff, blksize, 0);
4749 4756
4750 4757 /*
4751 4758 * Some other thread has entered the page,
4752 4759 * so just use it.
4753 4760 */
4754 4761 if (pp == NULL)
4755 4762 goto again;
4756 4763
4757 4764 /*
4758 4765 * Now round the request size up to page boundaries.
4759 4766 * This ensures that the entire page will be
4760 4767 * initialized to zeroes if EOF is encountered.
4761 4768 */
4762 4769 io_len = ptob(btopr(io_len));
4763 4770
4764 4771 bp = pageio_setup(pp, io_len, vp, B_READ);
4765 4772 ASSERT(bp != NULL);
4766 4773
4767 4774 /*
4768 4775 * pageio_setup should have set b_addr to 0. This
4769 4776 * is correct since we want to do I/O on a page
4770 4777 * boundary. bp_mapin will use this addr to calculate
4771 4778 * an offset, and then set b_addr to the kernel virtual
4772 4779 * address it allocated for us.
4773 4780 */
4774 4781 ASSERT(bp->b_un.b_addr == 0);
4775 4782
4776 4783 bp->b_edev = 0;
4777 4784 bp->b_dev = 0;
4778 4785 bp->b_lblkno = lbtodb(io_off);
4779 4786 bp->b_file = vp;
4780 4787 bp->b_offset = (offset_t)off;
4781 4788 bp_mapin(bp);
4782 4789
4783 4790 /*
4784 4791 * If doing a write beyond what we believe is EOF,
4785 4792 * don't bother trying to read the pages from the
4786 4793 * server, we'll just zero the pages here. We
4787 4794 * don't check that the rw flag is S_WRITE here
4788 4795 * because some implementations may attempt a
4789 4796 * read access to the buffer before copying data.
4790 4797 */
4791 4798 mutex_enter(&rp->r_statelock);
4792 4799 if (io_off >= rp->r_size && seg == segkmap) {
4793 4800 mutex_exit(&rp->r_statelock);
4794 4801 bzero(bp->b_un.b_addr, io_len);
4795 4802 } else {
4796 4803 mutex_exit(&rp->r_statelock);
4797 4804 error = nfs3_bio(bp, NULL, cr);
4798 4805 }
4799 4806
4800 4807 /*
4801 4808 * Unmap the buffer before freeing it.
4802 4809 */
4803 4810 bp_mapout(bp);
4804 4811 pageio_done(bp);
4805 4812
4806 4813 savepp = pp;
4807 4814 do {
4808 4815 pp->p_fsdata = C_NOCOMMIT;
4809 4816 } while ((pp = pp->p_next) != savepp);
4810 4817
4811 4818 if (error == NFS_EOF) {
4812 4819 /*
4813 4820 * If doing a write system call just return
4814 4821 * zeroed pages, else user tried to get pages
4815 4822 * beyond EOF, return error. We don't check
4816 4823 * that the rw flag is S_WRITE here because
4817 4824 * some implementations may attempt a read
4818 4825 * access to the buffer before copying data.
4819 4826 */
4820 4827 if (seg == segkmap)
4821 4828 error = 0;
4822 4829 else
4823 4830 error = EFAULT;
4824 4831 }
4825 4832
4826 4833 if (!readahead_issued && !error) {
4827 4834 mutex_enter(&rp->r_statelock);
4828 4835 rp->r_nextr = io_off + io_len;
4829 4836 mutex_exit(&rp->r_statelock);
4830 4837 }
4831 4838 }
4832 4839 }
4833 4840
4834 4841 out:
4835 4842 if (pl == NULL)
4836 4843 return (error);
4837 4844
4838 4845 if (error) {
4839 4846 if (pp != NULL)
4840 4847 pvn_read_done(pp, B_ERROR);
4841 4848 return (error);
4842 4849 }
4843 4850
4844 4851 if (pagefound) {
4845 4852 se_t se = (rw == S_CREATE ? SE_EXCL : SE_SHARED);
4846 4853
4847 4854 /*
4848 4855 * Page exists in the cache, acquire the appropriate lock.
4849 4856 * If this fails, start all over again.
4850 4857 */
4851 4858 if ((pp = page_lookup(vp, off, se)) == NULL) {
4852 4859 #ifdef DEBUG
4853 4860 nfs3_lostpage++;
4854 4861 #endif
4855 4862 goto reread;
4856 4863 }
4857 4864 pl[0] = pp;
4858 4865 pl[1] = NULL;
4859 4866 return (0);
4860 4867 }
4861 4868
4862 4869 if (pp != NULL)
4863 4870 pvn_plist_init(pp, pl, plsz, off, io_len, rw);
4864 4871
4865 4872 return (error);
4866 4873 }
4867 4874
4868 4875 static void
4869 4876 nfs3_readahead(vnode_t *vp, u_offset_t blkoff, caddr_t addr, struct seg *seg,
4870 4877 cred_t *cr)
4871 4878 {
4872 4879 int error;
4873 4880 page_t *pp;
4874 4881 u_offset_t io_off;
4875 4882 size_t io_len;
4876 4883 struct buf *bp;
4877 4884 uint_t bsize, blksize;
4878 4885 rnode_t *rp = VTOR(vp);
4879 4886 page_t *savepp;
4880 4887
4881 4888 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
4882 4889 bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
4883 4890
4884 4891 mutex_enter(&rp->r_statelock);
4885 4892 if (blkoff < rp->r_size && blkoff + bsize > rp->r_size) {
4886 4893 /*
4887 4894 * If less than a block left in file read less
4888 4895 * than a block.
4889 4896 */
4890 4897 blksize = rp->r_size - blkoff;
4891 4898 } else
4892 4899 blksize = bsize;
4893 4900 mutex_exit(&rp->r_statelock);
4894 4901
4895 4902 pp = pvn_read_kluster(vp, blkoff, segkmap, addr,
4896 4903 &io_off, &io_len, blkoff, blksize, 1);
4897 4904 /*
4898 4905 * The isra flag passed to the kluster function is 1, we may have
4899 4906 * gotten a return value of NULL for a variety of reasons (# of free
4900 4907 * pages < minfree, someone entered the page on the vnode etc). In all
4901 4908 * cases, we want to punt on the readahead.
4902 4909 */
4903 4910 if (pp == NULL)
4904 4911 return;
4905 4912
4906 4913 /*
4907 4914 * Now round the request size up to page boundaries.
4908 4915 * This ensures that the entire page will be
4909 4916 * initialized to zeroes if EOF is encountered.
4910 4917 */
4911 4918 io_len = ptob(btopr(io_len));
4912 4919
4913 4920 bp = pageio_setup(pp, io_len, vp, B_READ);
4914 4921 ASSERT(bp != NULL);
4915 4922
4916 4923 /*
4917 4924 * pageio_setup should have set b_addr to 0. This is correct since
4918 4925 * we want to do I/O on a page boundary. bp_mapin() will use this addr
4919 4926 * to calculate an offset, and then set b_addr to the kernel virtual
4920 4927 * address it allocated for us.
4921 4928 */
4922 4929 ASSERT(bp->b_un.b_addr == 0);
4923 4930
4924 4931 bp->b_edev = 0;
4925 4932 bp->b_dev = 0;
4926 4933 bp->b_lblkno = lbtodb(io_off);
4927 4934 bp->b_file = vp;
4928 4935 bp->b_offset = (offset_t)blkoff;
4929 4936 bp_mapin(bp);
4930 4937
4931 4938 /*
4932 4939 * If doing a write beyond what we believe is EOF, don't bother trying
4933 4940 * to read the pages from the server, we'll just zero the pages here.
4934 4941 * We don't check that the rw flag is S_WRITE here because some
4935 4942 * implementations may attempt a read access to the buffer before
4936 4943 * copying data.
4937 4944 */
4938 4945 mutex_enter(&rp->r_statelock);
4939 4946 if (io_off >= rp->r_size && seg == segkmap) {
4940 4947 mutex_exit(&rp->r_statelock);
4941 4948 bzero(bp->b_un.b_addr, io_len);
4942 4949 error = 0;
4943 4950 } else {
4944 4951 mutex_exit(&rp->r_statelock);
4945 4952 error = nfs3_bio(bp, NULL, cr);
4946 4953 if (error == NFS_EOF)
4947 4954 error = 0;
4948 4955 }
4949 4956
4950 4957 /*
4951 4958 * Unmap the buffer before freeing it.
4952 4959 */
4953 4960 bp_mapout(bp);
4954 4961 pageio_done(bp);
4955 4962
4956 4963 savepp = pp;
4957 4964 do {
4958 4965 pp->p_fsdata = C_NOCOMMIT;
4959 4966 } while ((pp = pp->p_next) != savepp);
4960 4967
4961 4968 pvn_read_done(pp, error ? B_READ | B_ERROR : B_READ);
4962 4969
4963 4970 /*
4964 4971 * In case of error set readahead offset
4965 4972 * to the lowest offset.
4966 4973 * pvn_read_done() calls VN_DISPOSE to destroy the pages
4967 4974 */
4968 4975 if (error && rp->r_nextr > io_off) {
4969 4976 mutex_enter(&rp->r_statelock);
4970 4977 if (rp->r_nextr > io_off)
4971 4978 rp->r_nextr = io_off;
4972 4979 mutex_exit(&rp->r_statelock);
4973 4980 }
4974 4981 }
4975 4982
4976 4983 /*
4977 4984 * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
4978 4985 * If len == 0, do from off to EOF.
4979 4986 *
4980 4987 * The normal cases should be len == 0 && off == 0 (entire vp list),
4981 4988 * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
4982 4989 * (from pageout).
4983 4990 */
4984 4991 /* ARGSUSED */
4985 4992 static int
4986 4993 nfs3_putpage(vnode_t *vp, offset_t off, size_t len, int flags, cred_t *cr,
4987 4994 caller_context_t *ct)
4988 4995 {
4989 4996 int error;
4990 4997 rnode_t *rp;
4991 4998
4992 4999 ASSERT(cr != NULL);
4993 5000
4994 5001 /*
4995 5002 * XXX - Why should this check be made here?
4996 5003 */
4997 5004 if (vp->v_flag & VNOMAP)
4998 5005 return (ENOSYS);
4999 5006 if (len == 0 && !(flags & B_INVAL) && vn_is_readonly(vp))
5000 5007 return (0);
5001 5008 if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
5002 5009 return (EIO);
5003 5010
5004 5011 rp = VTOR(vp);
5005 5012 mutex_enter(&rp->r_statelock);
5006 5013 rp->r_count++;
5007 5014 mutex_exit(&rp->r_statelock);
5008 5015 error = nfs_putpages(vp, off, len, flags, cr);
5009 5016 mutex_enter(&rp->r_statelock);
5010 5017 rp->r_count--;
5011 5018 cv_broadcast(&rp->r_cv);
5012 5019 mutex_exit(&rp->r_statelock);
5013 5020
5014 5021 return (error);
5015 5022 }
5016 5023
5017 5024 /*
5018 5025 * Write out a single page, possibly klustering adjacent dirty pages.
5019 5026 */
5020 5027 int
5021 5028 nfs3_putapage(vnode_t *vp, page_t *pp, u_offset_t *offp, size_t *lenp,
5022 5029 int flags, cred_t *cr)
5023 5030 {
5024 5031 u_offset_t io_off;
5025 5032 u_offset_t lbn_off;
5026 5033 u_offset_t lbn;
5027 5034 size_t io_len;
5028 5035 uint_t bsize;
5029 5036 int error;
5030 5037 rnode_t *rp;
5031 5038
5032 5039 ASSERT(!vn_is_readonly(vp));
5033 5040 ASSERT(pp != NULL);
5034 5041 ASSERT(cr != NULL);
5035 5042 ASSERT((flags & B_ASYNC) || nfs_zone() == VTOMI(vp)->mi_zone);
5036 5043
5037 5044 rp = VTOR(vp);
5038 5045 ASSERT(rp->r_count > 0);
5039 5046
5040 5047 bsize = MAX(vp->v_vfsp->vfs_bsize, PAGESIZE);
5041 5048 lbn = pp->p_offset / bsize;
5042 5049 lbn_off = lbn * bsize;
5043 5050
5044 5051 /*
5045 5052 * Find a kluster that fits in one block, or in
5046 5053 * one page if pages are bigger than blocks. If
5047 5054 * there is less file space allocated than a whole
5048 5055 * page, we'll shorten the i/o request below.
5049 5056 */
5050 5057 pp = pvn_write_kluster(vp, pp, &io_off, &io_len, lbn_off,
5051 5058 roundup(bsize, PAGESIZE), flags);
5052 5059
5053 5060 /*
5054 5061 * pvn_write_kluster shouldn't have returned a page with offset
5055 5062 * behind the original page we were given. Verify that.
5056 5063 */
5057 5064 ASSERT((pp->p_offset / bsize) >= lbn);
5058 5065
5059 5066 /*
5060 5067 * Now pp will have the list of kept dirty pages marked for
5061 5068 * write back. It will also handle invalidation and freeing
5062 5069 * of pages that are not dirty. Check for page length rounding
5063 5070 * problems.
5064 5071 */
5065 5072 if (io_off + io_len > lbn_off + bsize) {
5066 5073 ASSERT((io_off + io_len) - (lbn_off + bsize) < PAGESIZE);
5067 5074 io_len = lbn_off + bsize - io_off;
5068 5075 }
5069 5076 /*
5070 5077 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
5071 5078 * consistent value of r_size. RMODINPROGRESS is set in writerp().
5072 5079 * When RMODINPROGRESS is set it indicates that a uiomove() is in
5073 5080 * progress and the r_size has not been made consistent with the
5074 5081 * new size of the file. When the uiomove() completes the r_size is
5075 5082 * updated and the RMODINPROGRESS flag is cleared.
5076 5083 *
5077 5084 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
5078 5085 * consistent value of r_size. Without this handshaking, it is
5079 5086 * possible that nfs(3)_bio() picks up the old value of r_size
5080 5087 * before the uiomove() in writerp() completes. This will result
5081 5088 * in the write through nfs(3)_bio() being dropped.
5082 5089 *
5083 5090 * More precisely, there is a window between the time the uiomove()
5084 5091 * completes and the time the r_size is updated. If a VOP_PUTPAGE()
5085 5092 * operation intervenes in this window, the page will be picked up,
5086 5093 * because it is dirty (it will be unlocked, unless it was
5087 5094 * pagecreate'd). When the page is picked up as dirty, the dirty
5088 5095 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
5089 5096 * checked. This will still be the old size. Therefore the page will
5090 5097 * not be written out. When segmap_release() calls VOP_PUTPAGE(),
5091 5098 * the page will be found to be clean and the write will be dropped.
5092 5099 */
5093 5100 if (rp->r_flags & RMODINPROGRESS) {
5094 5101 mutex_enter(&rp->r_statelock);
5095 5102 if ((rp->r_flags & RMODINPROGRESS) &&
5096 5103 rp->r_modaddr + MAXBSIZE > io_off &&
5097 5104 rp->r_modaddr < io_off + io_len) {
5098 5105 page_t *plist;
5099 5106 /*
5100 5107 * A write is in progress for this region of the file.
5101 5108 * If we did not detect RMODINPROGRESS here then this
5102 5109 * path through nfs_putapage() would eventually go to
5103 5110 * nfs(3)_bio() and may not write out all of the data
5104 5111 * in the pages. We end up losing data. So we decide
5105 5112 * to set the modified bit on each page in the page
5106 5113 * list and mark the rnode with RDIRTY. This write
5107 5114 * will be restarted at some later time.
5108 5115 */
5109 5116 plist = pp;
5110 5117 while (plist != NULL) {
5111 5118 pp = plist;
5112 5119 page_sub(&plist, pp);
5113 5120 hat_setmod(pp);
5114 5121 page_io_unlock(pp);
5115 5122 page_unlock(pp);
5116 5123 }
5117 5124 rp->r_flags |= RDIRTY;
5118 5125 mutex_exit(&rp->r_statelock);
5119 5126 if (offp)
5120 5127 *offp = io_off;
5121 5128 if (lenp)
5122 5129 *lenp = io_len;
5123 5130 return (0);
5124 5131 }
5125 5132 mutex_exit(&rp->r_statelock);
5126 5133 }
5127 5134
5128 5135 if (flags & B_ASYNC) {
5129 5136 error = nfs_async_putapage(vp, pp, io_off, io_len, flags, cr,
5130 5137 nfs3_sync_putapage);
5131 5138 } else
5132 5139 error = nfs3_sync_putapage(vp, pp, io_off, io_len, flags, cr);
5133 5140
5134 5141 if (offp)
5135 5142 *offp = io_off;
5136 5143 if (lenp)
5137 5144 *lenp = io_len;
5138 5145 return (error);
5139 5146 }
5140 5147
5141 5148 static int
5142 5149 nfs3_sync_putapage(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5143 5150 int flags, cred_t *cr)
5144 5151 {
5145 5152 int error;
5146 5153 rnode_t *rp;
5147 5154
5148 5155 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5149 5156
5150 5157 flags |= B_WRITE;
5151 5158
5152 5159 error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5153 5160
5154 5161 rp = VTOR(vp);
5155 5162
5156 5163 if ((error == ENOSPC || error == EDQUOT || error == EFBIG ||
5157 5164 error == EACCES) &&
5158 5165 (flags & (B_INVAL|B_FORCE)) != (B_INVAL|B_FORCE)) {
5159 5166 if (!(rp->r_flags & ROUTOFSPACE)) {
5160 5167 mutex_enter(&rp->r_statelock);
5161 5168 rp->r_flags |= ROUTOFSPACE;
5162 5169 mutex_exit(&rp->r_statelock);
5163 5170 }
5164 5171 flags |= B_ERROR;
5165 5172 pvn_write_done(pp, flags);
5166 5173 /*
5167 5174 * If this was not an async thread, then try again to
5168 5175 * write out the pages, but this time, also destroy
5169 5176 * them whether or not the write is successful. This
5170 5177 * will prevent memory from filling up with these
5171 5178 * pages and destroying them is the only alternative
5172 5179 * if they can't be written out.
5173 5180 *
5174 5181 * Don't do this if this is an async thread because
5175 5182 * when the pages are unlocked in pvn_write_done,
5176 5183 * some other thread could have come along, locked
5177 5184 * them, and queued for an async thread. It would be
5178 5185 * possible for all of the async threads to be tied
5179 5186 * up waiting to lock the pages again and they would
5180 5187 * all already be locked and waiting for an async
5181 5188 * thread to handle them. Deadlock.
5182 5189 */
5183 5190 if (!(flags & B_ASYNC)) {
5184 5191 error = nfs3_putpage(vp, io_off, io_len,
5185 5192 B_INVAL | B_FORCE, cr, NULL);
5186 5193 }
5187 5194 } else {
5188 5195 if (error)
5189 5196 flags |= B_ERROR;
5190 5197 else if (rp->r_flags & ROUTOFSPACE) {
5191 5198 mutex_enter(&rp->r_statelock);
5192 5199 rp->r_flags &= ~ROUTOFSPACE;
5193 5200 mutex_exit(&rp->r_statelock);
5194 5201 }
5195 5202 pvn_write_done(pp, flags);
5196 5203 if (freemem < desfree)
5197 5204 (void) nfs3_commit_vp(vp, (u_offset_t)0, 0, cr);
5198 5205 }
5199 5206
5200 5207 return (error);
5201 5208 }
5202 5209
5203 5210 /* ARGSUSED */
5204 5211 static int
5205 5212 nfs3_map(vnode_t *vp, offset_t off, struct as *as, caddr_t *addrp,
5206 5213 size_t len, uchar_t prot, uchar_t maxprot, uint_t flags,
5207 5214 cred_t *cr, caller_context_t *ct)
5208 5215 {
5209 5216 struct segvn_crargs vn_a;
5210 5217 int error;
5211 5218 rnode_t *rp;
5212 5219 struct vattr va;
5213 5220
5214 5221 if (nfs_zone() != VTOMI(vp)->mi_zone)
5215 5222 return (EIO);
5216 5223
5217 5224 if (vp->v_flag & VNOMAP)
5218 5225 return (ENOSYS);
5219 5226
5220 5227 if (off < 0 || off + len < 0)
5221 5228 return (ENXIO);
5222 5229
5223 5230 if (vp->v_type != VREG)
5224 5231 return (ENODEV);
5225 5232
5226 5233 /*
5227 5234 * If there is cached data and if close-to-open consistency
5228 5235 * checking is not turned off and if the file system is not
5229 5236 * mounted readonly, then force an over the wire getattr.
5230 5237 * Otherwise, just invoke nfs3getattr to get a copy of the
5231 5238 * attributes. The attribute cache will be used unless it
5232 5239 * is timed out and if it is, then an over the wire getattr
5233 5240 * will be issued.
5234 5241 */
5235 5242 va.va_mask = AT_ALL;
5236 5243 if (vn_has_cached_data(vp) &&
5237 5244 !(VTOMI(vp)->mi_flags & MI_NOCTO) && !vn_is_readonly(vp))
5238 5245 error = nfs3_getattr_otw(vp, &va, cr);
5239 5246 else
5240 5247 error = nfs3getattr(vp, &va, cr);
5241 5248 if (error)
5242 5249 return (error);
5243 5250
5244 5251 /*
5245 5252 * Check to see if the vnode is currently marked as not cachable.
5246 5253 * This means portions of the file are locked (through VOP_FRLOCK).
5247 5254 * In this case the map request must be refused. We use
5248 5255 * rp->r_lkserlock to avoid a race with concurrent lock requests.
5249 5256 */
5250 5257 rp = VTOR(vp);
5251 5258
5252 5259 /*
5253 5260 * Atomically increment r_inmap after acquiring r_rwlock. The
5254 5261 * idea here is to acquire r_rwlock to block read/write and
5255 5262 * not to protect r_inmap. r_inmap will inform nfs3_read/write()
5256 5263 * that we are in nfs3_map(). Now, r_rwlock is acquired in order
5257 5264 * and we can prevent the deadlock that would have occurred
5258 5265 * when nfs3_addmap() would have acquired it out of order.
5259 5266 *
5260 5267 * Since we are not protecting r_inmap by any lock, we do not
5261 5268 * hold any lock when we decrement it. We atomically decrement
5262 5269 * r_inmap after we release r_lkserlock.
5263 5270 */
5264 5271
5265 5272 if (nfs_rw_enter_sig(&rp->r_rwlock, RW_WRITER, INTR(vp)))
5266 5273 return (EINTR);
5267 5274 atomic_add_int(&rp->r_inmap, 1);
5268 5275 nfs_rw_exit(&rp->r_rwlock);
5269 5276
5270 5277 if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_READER, INTR(vp))) {
5271 5278 atomic_add_int(&rp->r_inmap, -1);
5272 5279 return (EINTR);
5273 5280 }
5274 5281
5275 5282 if (vp->v_flag & VNOCACHE) {
5276 5283 error = EAGAIN;
5277 5284 goto done;
5278 5285 }
5279 5286
5280 5287 /*
5281 5288 * Don't allow concurrent locks and mapping if mandatory locking is
5282 5289 * enabled.
5283 5290 */
5284 5291 if ((flk_has_remote_locks(vp) || lm_has_sleep(vp)) &&
5285 5292 MANDLOCK(vp, va.va_mode)) {
5286 5293 error = EAGAIN;
5287 5294 goto done;
5288 5295 }
5289 5296
5290 5297 as_rangelock(as);
5291 5298 error = choose_addr(as, addrp, len, off, ADDR_VACALIGN, flags);
5292 5299 if (error != 0) {
5293 5300 as_rangeunlock(as);
5294 5301 goto done;
5295 5302 }
5296 5303
5297 5304 vn_a.vp = vp;
5298 5305 vn_a.offset = off;
5299 5306 vn_a.type = (flags & MAP_TYPE);
5300 5307 vn_a.prot = (uchar_t)prot;
5301 5308 vn_a.maxprot = (uchar_t)maxprot;
5302 5309 vn_a.flags = (flags & ~MAP_TYPE);
5303 5310 vn_a.cred = cr;
5304 5311 vn_a.amp = NULL;
5305 5312 vn_a.szc = 0;
5306 5313 vn_a.lgrp_mem_policy_flags = 0;
5307 5314
5308 5315 error = as_map(as, *addrp, len, segvn_create, &vn_a);
5309 5316 as_rangeunlock(as);
5310 5317
5311 5318 done:
5312 5319 nfs_rw_exit(&rp->r_lkserlock);
5313 5320 atomic_add_int(&rp->r_inmap, -1);
5314 5321 return (error);
5315 5322 }
5316 5323
5317 5324 /* ARGSUSED */
5318 5325 static int
5319 5326 nfs3_addmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5320 5327 size_t len, uchar_t prot, uchar_t maxprot, uint_t flags,
5321 5328 cred_t *cr, caller_context_t *ct)
5322 5329 {
5323 5330 rnode_t *rp;
5324 5331
5325 5332 if (vp->v_flag & VNOMAP)
5326 5333 return (ENOSYS);
5327 5334 if (nfs_zone() != VTOMI(vp)->mi_zone)
5328 5335 return (EIO);
5329 5336
5330 5337 rp = VTOR(vp);
5331 5338 atomic_add_long((ulong_t *)&rp->r_mapcnt, btopr(len));
5332 5339
5333 5340 return (0);
5334 5341 }
5335 5342
5336 5343 /* ARGSUSED */
5337 5344 static int
5338 5345 nfs3_frlock(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5339 5346 offset_t offset, struct flk_callback *flk_cbp, cred_t *cr,
5340 5347 caller_context_t *ct)
5341 5348 {
5342 5349 netobj lm_fh3;
5343 5350 int rc;
5344 5351 u_offset_t start, end;
5345 5352 rnode_t *rp;
5346 5353 int error = 0, intr = INTR(vp);
5347 5354
5348 5355 if (nfs_zone() != VTOMI(vp)->mi_zone)
5349 5356 return (EIO);
5350 5357 /* check for valid cmd parameter */
5351 5358 if (cmd != F_GETLK && cmd != F_SETLK && cmd != F_SETLKW)
5352 5359 return (EINVAL);
5353 5360
5354 5361 /* Verify l_type. */
5355 5362 switch (bfp->l_type) {
5356 5363 case F_RDLCK:
5357 5364 if (cmd != F_GETLK && !(flag & FREAD))
5358 5365 return (EBADF);
5359 5366 break;
5360 5367 case F_WRLCK:
5361 5368 if (cmd != F_GETLK && !(flag & FWRITE))
5362 5369 return (EBADF);
5363 5370 break;
5364 5371 case F_UNLCK:
5365 5372 intr = 0;
5366 5373 break;
5367 5374
5368 5375 default:
5369 5376 return (EINVAL);
5370 5377 }
5371 5378
5372 5379 /* check the validity of the lock range */
5373 5380 if (rc = flk_convert_lock_data(vp, bfp, &start, &end, offset))
5374 5381 return (rc);
5375 5382 if (rc = flk_check_lock_data(start, end, MAXEND))
5376 5383 return (rc);
5377 5384
5378 5385 /*
5379 5386 * If the filesystem is mounted using local locking, pass the
5380 5387 * request off to the local locking code.
5381 5388 */
5382 5389 if (VTOMI(vp)->mi_flags & MI_LLOCK) {
5383 5390 if (cmd == F_SETLK || cmd == F_SETLKW) {
5384 5391 /*
5385 5392 * For complete safety, we should be holding
5386 5393 * r_lkserlock. However, we can't call
5387 5394 * lm_safelock and then fs_frlock while
5388 5395 * holding r_lkserlock, so just invoke
5389 5396 * lm_safelock and expect that this will
5390 5397 * catch enough of the cases.
5391 5398 */
5392 5399 if (!lm_safelock(vp, bfp, cr))
5393 5400 return (EAGAIN);
5394 5401 }
5395 5402 return (fs_frlock(vp, cmd, bfp, flag, offset, flk_cbp, cr, ct));
5396 5403 }
5397 5404
5398 5405 rp = VTOR(vp);
5399 5406
5400 5407 /*
5401 5408 * Check whether the given lock request can proceed, given the
5402 5409 * current file mappings.
5403 5410 */
5404 5411 if (nfs_rw_enter_sig(&rp->r_lkserlock, RW_WRITER, intr))
5405 5412 return (EINTR);
5406 5413 if (cmd == F_SETLK || cmd == F_SETLKW) {
5407 5414 if (!lm_safelock(vp, bfp, cr)) {
5408 5415 rc = EAGAIN;
5409 5416 goto done;
5410 5417 }
5411 5418 }
5412 5419
5413 5420 /*
5414 5421 * Flush the cache after waiting for async I/O to finish. For new
5415 5422 * locks, this is so that the process gets the latest bits from the
5416 5423 * server. For unlocks, this is so that other clients see the
5417 5424 * latest bits once the file has been unlocked. If currently dirty
5418 5425 * pages can't be flushed, then don't allow a lock to be set. But
5419 5426 * allow unlocks to succeed, to avoid having orphan locks on the
5420 5427 * server.
5421 5428 */
5422 5429 if (cmd != F_GETLK) {
5423 5430 mutex_enter(&rp->r_statelock);
5424 5431 while (rp->r_count > 0) {
5425 5432 if (intr) {
5426 5433 klwp_t *lwp = ttolwp(curthread);
5427 5434
5428 5435 if (lwp != NULL)
5429 5436 lwp->lwp_nostop++;
5430 5437 if (cv_wait_sig(&rp->r_cv,
5431 5438 &rp->r_statelock) == 0) {
5432 5439 if (lwp != NULL)
5433 5440 lwp->lwp_nostop--;
5434 5441 rc = EINTR;
5435 5442 break;
5436 5443 }
5437 5444 if (lwp != NULL)
5438 5445 lwp->lwp_nostop--;
5439 5446 } else
5440 5447 cv_wait(&rp->r_cv, &rp->r_statelock);
5441 5448 }
5442 5449 mutex_exit(&rp->r_statelock);
5443 5450 if (rc != 0)
5444 5451 goto done;
5445 5452 error = nfs3_putpage(vp, (offset_t)0, 0, B_INVAL, cr, ct);
5446 5453 if (error) {
5447 5454 if (error == ENOSPC || error == EDQUOT) {
5448 5455 mutex_enter(&rp->r_statelock);
5449 5456 if (!rp->r_error)
5450 5457 rp->r_error = error;
5451 5458 mutex_exit(&rp->r_statelock);
5452 5459 }
5453 5460 if (bfp->l_type != F_UNLCK) {
5454 5461 rc = ENOLCK;
5455 5462 goto done;
5456 5463 }
5457 5464 }
5458 5465 }
5459 5466
5460 5467 lm_fh3.n_len = VTOFH3(vp)->fh3_length;
5461 5468 lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
5462 5469
5463 5470 /*
5464 5471 * Call the lock manager to do the real work of contacting
5465 5472 * the server and obtaining the lock.
5466 5473 */
5467 5474 rc = lm4_frlock(vp, cmd, bfp, flag, offset, cr, &lm_fh3, flk_cbp);
5468 5475
5469 5476 if (rc == 0)
5470 5477 nfs_lockcompletion(vp, cmd);
5471 5478
5472 5479 done:
5473 5480 nfs_rw_exit(&rp->r_lkserlock);
5474 5481 return (rc);
5475 5482 }
5476 5483
5477 5484 /*
5478 5485 * Free storage space associated with the specified vnode. The portion
5479 5486 * to be freed is specified by bfp->l_start and bfp->l_len (already
5480 5487 * normalized to a "whence" of 0).
5481 5488 *
5482 5489 * This is an experimental facility whose continued existence is not
5483 5490 * guaranteed. Currently, we only support the special case
5484 5491 * of l_len == 0, meaning free to end of file.
5485 5492 */
5486 5493 /* ARGSUSED */
5487 5494 static int
5488 5495 nfs3_space(vnode_t *vp, int cmd, struct flock64 *bfp, int flag,
5489 5496 offset_t offset, cred_t *cr, caller_context_t *ct)
5490 5497 {
5491 5498 int error;
5492 5499
5493 5500 ASSERT(vp->v_type == VREG);
5494 5501 if (cmd != F_FREESP)
5495 5502 return (EINVAL);
5496 5503 if (nfs_zone() != VTOMI(vp)->mi_zone)
5497 5504 return (EIO);
5498 5505
5499 5506 error = convoff(vp, bfp, 0, offset);
5500 5507 if (!error) {
5501 5508 ASSERT(bfp->l_start >= 0);
5502 5509 if (bfp->l_len == 0) {
5503 5510 struct vattr va;
5504 5511
5505 5512 /*
5506 5513 * ftruncate should not change the ctime and
5507 5514 * mtime if we truncate the file to its
5508 5515 * previous size.
5509 5516 */
5510 5517 va.va_mask = AT_SIZE;
5511 5518 error = nfs3getattr(vp, &va, cr);
5512 5519 if (error || va.va_size == bfp->l_start)
5513 5520 return (error);
5514 5521 va.va_mask = AT_SIZE;
5515 5522 va.va_size = bfp->l_start;
5516 5523 error = nfs3setattr(vp, &va, 0, cr);
5517 5524 } else
5518 5525 error = EINVAL;
5519 5526 }
5520 5527
5521 5528 return (error);
5522 5529 }
5523 5530
5524 5531 /* ARGSUSED */
5525 5532 static int
5526 5533 nfs3_realvp(vnode_t *vp, vnode_t **vpp, caller_context_t *ct)
5527 5534 {
5528 5535
5529 5536 return (EINVAL);
5530 5537 }
5531 5538
5532 5539 /*
5533 5540 * Setup and add an address space callback to do the work of the delmap call.
5534 5541 * The callback will (and must be) deleted in the actual callback function.
5535 5542 *
5536 5543 * This is done in order to take care of the problem that we have with holding
5537 5544 * the address space's a_lock for a long period of time (e.g. if the NFS server
5538 5545 * is down). Callbacks will be executed in the address space code while the
5539 5546 * a_lock is not held. Holding the address space's a_lock causes things such
5540 5547 * as ps and fork to hang because they are trying to acquire this lock as well.
5541 5548 */
5542 5549 /* ARGSUSED */
5543 5550 static int
5544 5551 nfs3_delmap(vnode_t *vp, offset_t off, struct as *as, caddr_t addr,
5545 5552 size_t len, uint_t prot, uint_t maxprot, uint_t flags,
5546 5553 cred_t *cr, caller_context_t *ct)
5547 5554 {
5548 5555 int caller_found;
5549 5556 int error;
5550 5557 rnode_t *rp;
5551 5558 nfs_delmap_args_t *dmapp;
5552 5559 nfs_delmapcall_t *delmap_call;
5553 5560
5554 5561 if (vp->v_flag & VNOMAP)
5555 5562 return (ENOSYS);
5556 5563 /*
5557 5564 * A process may not change zones if it has NFS pages mmap'ed
5558 5565 * in, so we can't legitimately get here from the wrong zone.
5559 5566 */
5560 5567 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5561 5568
5562 5569 rp = VTOR(vp);
5563 5570
5564 5571 /*
5565 5572 * The way that the address space of this process deletes its mapping
5566 5573 * of this file is via the following call chains:
5567 5574 * - as_free()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5568 5575 * - as_unmap()->SEGOP_UNMAP()/segvn_unmap()->VOP_DELMAP()/nfs3_delmap()
5569 5576 *
5570 5577 * With the use of address space callbacks we are allowed to drop the
5571 5578 * address space lock, a_lock, while executing the NFS operations that
5572 5579 * need to go over the wire. Returning EAGAIN to the caller of this
5573 5580 * function is what drives the execution of the callback that we add
5574 5581 * below. The callback will be executed by the address space code
5575 5582 * after dropping the a_lock. When the callback is finished, since
5576 5583 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
5577 5584 * is called again on the same segment to finish the rest of the work
5578 5585 * that needs to happen during unmapping.
5579 5586 *
5580 5587 * This action of calling back into the segment driver causes
5581 5588 * nfs3_delmap() to get called again, but since the callback was
5582 5589 * already executed at this point, it already did the work and there
5583 5590 * is nothing left for us to do.
5584 5591 *
5585 5592 * To Summarize:
5586 5593 * - The first time nfs3_delmap is called by the current thread is when
5587 5594 * we add the caller associated with this delmap to the delmap caller
5588 5595 * list, add the callback, and return EAGAIN.
5589 5596 * - The second time in this call chain when nfs3_delmap is called we
5590 5597 * will find this caller in the delmap caller list and realize there
5591 5598 * is no more work to do thus removing this caller from the list and
5592 5599 * returning the error that was set in the callback execution.
5593 5600 */
5594 5601 caller_found = nfs_find_and_delete_delmapcall(rp, &error);
5595 5602 if (caller_found) {
5596 5603 /*
5597 5604 * 'error' is from the actual delmap operations. To avoid
5598 5605 * hangs, we need to handle the return of EAGAIN differently
5599 5606 * since this is what drives the callback execution.
5600 5607 * In this case, we don't want to return EAGAIN and do the
5601 5608 * callback execution because there are none to execute.
5602 5609 */
5603 5610 if (error == EAGAIN)
5604 5611 return (0);
5605 5612 else
5606 5613 return (error);
5607 5614 }
5608 5615
5609 5616 /* current caller was not in the list */
5610 5617 delmap_call = nfs_init_delmapcall();
5611 5618
5612 5619 mutex_enter(&rp->r_statelock);
5613 5620 list_insert_tail(&rp->r_indelmap, delmap_call);
5614 5621 mutex_exit(&rp->r_statelock);
5615 5622
5616 5623 dmapp = kmem_alloc(sizeof (nfs_delmap_args_t), KM_SLEEP);
5617 5624
5618 5625 dmapp->vp = vp;
5619 5626 dmapp->off = off;
5620 5627 dmapp->addr = addr;
5621 5628 dmapp->len = len;
5622 5629 dmapp->prot = prot;
5623 5630 dmapp->maxprot = maxprot;
5624 5631 dmapp->flags = flags;
5625 5632 dmapp->cr = cr;
5626 5633 dmapp->caller = delmap_call;
5627 5634
5628 5635 error = as_add_callback(as, nfs3_delmap_callback, dmapp,
5629 5636 AS_UNMAP_EVENT, addr, len, KM_SLEEP);
5630 5637
5631 5638 return (error ? error : EAGAIN);
5632 5639 }
5633 5640
5634 5641 /*
5635 5642 * Remove some pages from an mmap'd vnode. Just update the
5636 5643 * count of pages. If doing close-to-open, then flush and
5637 5644 * commit all of the pages associated with this file.
5638 5645 * Otherwise, start an asynchronous page flush to write out
5639 5646 * any dirty pages. This will also associate a credential
5640 5647 * with the rnode which can be used to write the pages.
5641 5648 */
5642 5649 /* ARGSUSED */
5643 5650 static void
5644 5651 nfs3_delmap_callback(struct as *as, void *arg, uint_t event)
5645 5652 {
5646 5653 int error;
5647 5654 rnode_t *rp;
5648 5655 mntinfo_t *mi;
5649 5656 nfs_delmap_args_t *dmapp = (nfs_delmap_args_t *)arg;
5650 5657
5651 5658 rp = VTOR(dmapp->vp);
5652 5659 mi = VTOMI(dmapp->vp);
5653 5660
5654 5661 atomic_add_long((ulong_t *)&rp->r_mapcnt, -btopr(dmapp->len));
5655 5662 ASSERT(rp->r_mapcnt >= 0);
5656 5663
5657 5664 /*
5658 5665 * Initiate a page flush and potential commit if there are
5659 5666 * pages, the file system was not mounted readonly, the segment
5660 5667 * was mapped shared, and the pages themselves were writeable.
5661 5668 */
5662 5669 if (vn_has_cached_data(dmapp->vp) && !vn_is_readonly(dmapp->vp) &&
5663 5670 dmapp->flags == MAP_SHARED && (dmapp->maxprot & PROT_WRITE)) {
5664 5671 mutex_enter(&rp->r_statelock);
5665 5672 rp->r_flags |= RDIRTY;
5666 5673 mutex_exit(&rp->r_statelock);
5667 5674 /*
5668 5675 * If this is a cross-zone access a sync putpage won't work, so
5669 5676 * the best we can do is try an async putpage. That seems
5670 5677 * better than something more draconian such as discarding the
5671 5678 * dirty pages.
5672 5679 */
5673 5680 if ((mi->mi_flags & MI_NOCTO) ||
5674 5681 nfs_zone() != mi->mi_zone)
5675 5682 error = nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5676 5683 B_ASYNC, dmapp->cr, NULL);
5677 5684 else
5678 5685 error = nfs3_putpage_commit(dmapp->vp, dmapp->off,
5679 5686 dmapp->len, dmapp->cr);
5680 5687 if (!error) {
5681 5688 mutex_enter(&rp->r_statelock);
5682 5689 error = rp->r_error;
5683 5690 rp->r_error = 0;
5684 5691 mutex_exit(&rp->r_statelock);
5685 5692 }
5686 5693 } else
5687 5694 error = 0;
5688 5695
5689 5696 if ((rp->r_flags & RDIRECTIO) || (mi->mi_flags & MI_DIRECTIO))
5690 5697 (void) nfs3_putpage(dmapp->vp, dmapp->off, dmapp->len,
5691 5698 B_INVAL, dmapp->cr, NULL);
5692 5699
5693 5700 dmapp->caller->error = error;
5694 5701 (void) as_delete_callback(as, arg);
5695 5702 kmem_free(dmapp, sizeof (nfs_delmap_args_t));
5696 5703 }
5697 5704
5698 5705 static int nfs3_pathconf_disable_cache = 0;
5699 5706
5700 5707 #ifdef DEBUG
5701 5708 static int nfs3_pathconf_cache_hits = 0;
5702 5709 static int nfs3_pathconf_cache_misses = 0;
5703 5710 #endif
5704 5711
5705 5712 /* ARGSUSED */
5706 5713 static int
5707 5714 nfs3_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr,
5708 5715 caller_context_t *ct)
5709 5716 {
5710 5717 int error;
5711 5718 PATHCONF3args args;
5712 5719 PATHCONF3res res;
5713 5720 int douprintf;
5714 5721 failinfo_t fi;
5715 5722 rnode_t *rp;
5716 5723 hrtime_t t;
5717 5724
5718 5725 if (nfs_zone() != VTOMI(vp)->mi_zone)
5719 5726 return (EIO);
5720 5727 /*
5721 5728 * Large file spec - need to base answer on info stored
5722 5729 * on original FSINFO response.
5723 5730 */
5724 5731 if (cmd == _PC_FILESIZEBITS) {
5725 5732 unsigned long long ll;
5726 5733 long l = 1;
5727 5734
5728 5735 ll = VTOMI(vp)->mi_maxfilesize;
5729 5736
5730 5737 if (ll == 0) {
5731 5738 *valp = 0;
5732 5739 return (0);
5733 5740 }
5734 5741
5735 5742 if (ll & 0xffffffff00000000) {
5736 5743 l += 32; ll >>= 32;
5737 5744 }
5738 5745 if (ll & 0xffff0000) {
5739 5746 l += 16; ll >>= 16;
5740 5747 }
5741 5748 if (ll & 0xff00) {
5742 5749 l += 8; ll >>= 8;
5743 5750 }
5744 5751 if (ll & 0xf0) {
5745 5752 l += 4; ll >>= 4;
5746 5753 }
5747 5754 if (ll & 0xc) {
5748 5755 l += 2; ll >>= 2;
5749 5756 }
5750 5757 if (ll & 0x2)
5751 5758 l += 2;
5752 5759 else if (ll & 0x1)
5753 5760 l += 1;
5754 5761 *valp = l;
5755 5762 return (0);
5756 5763 }
5757 5764
5758 5765 if (cmd == _PC_ACL_ENABLED) {
5759 5766 *valp = _ACL_ACLENT_ENABLED;
5760 5767 return (0);
5761 5768 }
5762 5769
5763 5770 if (cmd == _PC_XATTR_EXISTS) {
5764 5771 error = 0;
5765 5772 *valp = 0;
5766 5773 if (vp->v_vfsp->vfs_flag & VFS_XATTR) {
5767 5774 vnode_t *avp;
5768 5775 rnode_t *rp;
5769 5776 int error = 0;
5770 5777 mntinfo_t *mi = VTOMI(vp);
5771 5778
5772 5779 if (!(mi->mi_flags & MI_EXTATTR))
5773 5780 return (0);
5774 5781
5775 5782 rp = VTOR(vp);
5776 5783 if (nfs_rw_enter_sig(&rp->r_rwlock, RW_READER,
5777 5784 INTR(vp)))
5778 5785 return (EINTR);
5779 5786
5780 5787 error = nfs3lookup_dnlc(vp, XATTR_DIR_NAME, &avp, cr);
5781 5788 if (error || avp == NULL)
5782 5789 error = acl_getxattrdir3(vp, &avp, 0, cr, 0);
5783 5790
5784 5791 nfs_rw_exit(&rp->r_rwlock);
5785 5792
5786 5793 if (error == 0 && avp != NULL) {
5787 5794 error = do_xattr_exists_check(avp, valp, cr);
5788 5795 VN_RELE(avp);
5789 5796 } else if (error == ENOENT) {
5790 5797 error = 0;
5791 5798 *valp = 0;
5792 5799 }
5793 5800 }
5794 5801 return (error);
5795 5802 }
5796 5803
5797 5804 rp = VTOR(vp);
5798 5805 if (rp->r_pathconf != NULL) {
5799 5806 mutex_enter(&rp->r_statelock);
5800 5807 if (rp->r_pathconf != NULL && nfs3_pathconf_disable_cache) {
5801 5808 kmem_free(rp->r_pathconf, sizeof (*rp->r_pathconf));
5802 5809 rp->r_pathconf = NULL;
5803 5810 }
5804 5811 if (rp->r_pathconf != NULL) {
5805 5812 error = 0;
5806 5813 switch (cmd) {
5807 5814 case _PC_LINK_MAX:
5808 5815 *valp = rp->r_pathconf->link_max;
5809 5816 break;
5810 5817 case _PC_NAME_MAX:
5811 5818 *valp = rp->r_pathconf->name_max;
5812 5819 break;
5813 5820 case _PC_PATH_MAX:
5814 5821 case _PC_SYMLINK_MAX:
5815 5822 *valp = MAXPATHLEN;
5816 5823 break;
5817 5824 case _PC_CHOWN_RESTRICTED:
5818 5825 *valp = rp->r_pathconf->chown_restricted;
5819 5826 break;
5820 5827 case _PC_NO_TRUNC:
5821 5828 *valp = rp->r_pathconf->no_trunc;
5822 5829 break;
5823 5830 default:
5824 5831 error = EINVAL;
5825 5832 break;
5826 5833 }
5827 5834 mutex_exit(&rp->r_statelock);
5828 5835 #ifdef DEBUG
5829 5836 nfs3_pathconf_cache_hits++;
5830 5837 #endif
5831 5838 return (error);
5832 5839 }
5833 5840 mutex_exit(&rp->r_statelock);
5834 5841 }
5835 5842 #ifdef DEBUG
5836 5843 nfs3_pathconf_cache_misses++;
5837 5844 #endif
5838 5845
5839 5846 args.object = *VTOFH3(vp);
5840 5847 fi.vp = vp;
5841 5848 fi.fhp = (caddr_t)&args.object;
5842 5849 fi.copyproc = nfs3copyfh;
5843 5850 fi.lookupproc = nfs3lookup;
5844 5851 fi.xattrdirproc = acl_getxattrdir3;
5845 5852
5846 5853 douprintf = 1;
5847 5854
5848 5855 t = gethrtime();
5849 5856
5850 5857 error = rfs3call(VTOMI(vp), NFSPROC3_PATHCONF,
5851 5858 xdr_nfs_fh3, (caddr_t)&args,
5852 5859 xdr_PATHCONF3res, (caddr_t)&res, cr,
5853 5860 &douprintf, &res.status, 0, &fi);
5854 5861
5855 5862 if (error)
5856 5863 return (error);
5857 5864
5858 5865 error = geterrno3(res.status);
5859 5866
5860 5867 if (!error) {
5861 5868 nfs3_cache_post_op_attr(vp, &res.resok.obj_attributes, t, cr);
5862 5869 if (!nfs3_pathconf_disable_cache) {
5863 5870 mutex_enter(&rp->r_statelock);
5864 5871 if (rp->r_pathconf == NULL) {
5865 5872 rp->r_pathconf = kmem_alloc(
5866 5873 sizeof (*rp->r_pathconf), KM_NOSLEEP);
5867 5874 if (rp->r_pathconf != NULL)
5868 5875 *rp->r_pathconf = res.resok.info;
5869 5876 }
5870 5877 mutex_exit(&rp->r_statelock);
5871 5878 }
5872 5879 switch (cmd) {
5873 5880 case _PC_LINK_MAX:
5874 5881 *valp = res.resok.info.link_max;
5875 5882 break;
5876 5883 case _PC_NAME_MAX:
5877 5884 *valp = res.resok.info.name_max;
5878 5885 break;
5879 5886 case _PC_PATH_MAX:
5880 5887 case _PC_SYMLINK_MAX:
5881 5888 *valp = MAXPATHLEN;
5882 5889 break;
5883 5890 case _PC_CHOWN_RESTRICTED:
5884 5891 *valp = res.resok.info.chown_restricted;
5885 5892 break;
5886 5893 case _PC_NO_TRUNC:
5887 5894 *valp = res.resok.info.no_trunc;
5888 5895 break;
5889 5896 default:
5890 5897 return (EINVAL);
5891 5898 }
5892 5899 } else {
5893 5900 nfs3_cache_post_op_attr(vp, &res.resfail.obj_attributes, t, cr);
5894 5901 PURGE_STALE_FH(error, vp, cr);
5895 5902 }
5896 5903
5897 5904 return (error);
5898 5905 }
5899 5906
5900 5907 /*
5901 5908 * Called by async thread to do synchronous pageio. Do the i/o, wait
5902 5909 * for it to complete, and cleanup the page list when done.
5903 5910 */
5904 5911 static int
5905 5912 nfs3_sync_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5906 5913 int flags, cred_t *cr)
5907 5914 {
5908 5915 int error;
5909 5916
5910 5917 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
5911 5918 error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5912 5919 if (flags & B_READ)
5913 5920 pvn_read_done(pp, (error ? B_ERROR : 0) | flags);
5914 5921 else
5915 5922 pvn_write_done(pp, (error ? B_ERROR : 0) | flags);
5916 5923 return (error);
5917 5924 }
5918 5925
5919 5926 /* ARGSUSED */
5920 5927 static int
5921 5928 nfs3_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
5922 5929 int flags, cred_t *cr, caller_context_t *ct)
5923 5930 {
5924 5931 int error;
5925 5932 rnode_t *rp;
5926 5933
5927 5934 if (pp == NULL)
5928 5935 return (EINVAL);
5929 5936 if (!(flags & B_ASYNC) && nfs_zone() != VTOMI(vp)->mi_zone)
5930 5937 return (EIO);
5931 5938
5932 5939 rp = VTOR(vp);
5933 5940 mutex_enter(&rp->r_statelock);
5934 5941 rp->r_count++;
5935 5942 mutex_exit(&rp->r_statelock);
5936 5943
5937 5944 if (flags & B_ASYNC) {
5938 5945 error = nfs_async_pageio(vp, pp, io_off, io_len, flags, cr,
5939 5946 nfs3_sync_pageio);
5940 5947 } else
5941 5948 error = nfs3_rdwrlbn(vp, pp, io_off, io_len, flags, cr);
5942 5949 mutex_enter(&rp->r_statelock);
5943 5950 rp->r_count--;
5944 5951 cv_broadcast(&rp->r_cv);
5945 5952 mutex_exit(&rp->r_statelock);
5946 5953 return (error);
5947 5954 }
5948 5955
5949 5956 /* ARGSUSED */
5950 5957 static void
5951 5958 nfs3_dispose(vnode_t *vp, page_t *pp, int fl, int dn, cred_t *cr,
5952 5959 caller_context_t *ct)
5953 5960 {
5954 5961 int error;
5955 5962 rnode_t *rp;
5956 5963 page_t *plist;
5957 5964 page_t *pptr;
5958 5965 offset3 offset;
5959 5966 count3 len;
5960 5967 k_sigset_t smask;
5961 5968
5962 5969 /*
5963 5970 * We should get called with fl equal to either B_FREE or
5964 5971 * B_INVAL. Any other value is illegal.
5965 5972 *
5966 5973 * The page that we are either supposed to free or destroy
5967 5974 * should be exclusive locked and its io lock should not
5968 5975 * be held.
5969 5976 */
5970 5977 ASSERT(fl == B_FREE || fl == B_INVAL);
5971 5978 ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr);
5972 5979 rp = VTOR(vp);
5973 5980
5974 5981 /*
5975 5982 * If the page doesn't need to be committed or we shouldn't
5976 5983 * even bother attempting to commit it, then just make sure
5977 5984 * that the p_fsdata byte is clear and then either free or
5978 5985 * destroy the page as appropriate.
5979 5986 */
5980 5987 if (pp->p_fsdata == C_NOCOMMIT || (rp->r_flags & RSTALE)) {
5981 5988 pp->p_fsdata = C_NOCOMMIT;
5982 5989 if (fl == B_FREE)
5983 5990 page_free(pp, dn);
5984 5991 else
5985 5992 page_destroy(pp, dn);
5986 5993 return;
5987 5994 }
5988 5995
5989 5996 /*
5990 5997 * If there is a page invalidation operation going on, then
5991 5998 * if this is one of the pages being destroyed, then just
5992 5999 * clear the p_fsdata byte and then either free or destroy
5993 6000 * the page as appropriate.
5994 6001 */
5995 6002 mutex_enter(&rp->r_statelock);
5996 6003 if ((rp->r_flags & RTRUNCATE) && pp->p_offset >= rp->r_truncaddr) {
5997 6004 mutex_exit(&rp->r_statelock);
5998 6005 pp->p_fsdata = C_NOCOMMIT;
5999 6006 if (fl == B_FREE)
6000 6007 page_free(pp, dn);
6001 6008 else
6002 6009 page_destroy(pp, dn);
6003 6010 return;
6004 6011 }
6005 6012
6006 6013 /*
6007 6014 * If we are freeing this page and someone else is already
6008 6015 * waiting to do a commit, then just unlock the page and
6009 6016 * return. That other thread will take care of commiting
6010 6017 * this page. The page can be freed sometime after the
6011 6018 * commit has finished. Otherwise, if the page is marked
6012 6019 * as delay commit, then we may be getting called from
6013 6020 * pvn_write_done, one page at a time. This could result
6014 6021 * in one commit per page, so we end up doing lots of small
6015 6022 * commits instead of fewer larger commits. This is bad,
6016 6023 * we want do as few commits as possible.
6017 6024 */
6018 6025 if (fl == B_FREE) {
6019 6026 if (rp->r_flags & RCOMMITWAIT) {
6020 6027 page_unlock(pp);
6021 6028 mutex_exit(&rp->r_statelock);
6022 6029 return;
6023 6030 }
6024 6031 if (pp->p_fsdata == C_DELAYCOMMIT) {
6025 6032 pp->p_fsdata = C_COMMIT;
6026 6033 page_unlock(pp);
6027 6034 mutex_exit(&rp->r_statelock);
6028 6035 return;
6029 6036 }
6030 6037 }
6031 6038
6032 6039 /*
6033 6040 * Check to see if there is a signal which would prevent an
6034 6041 * attempt to commit the pages from being successful. If so,
6035 6042 * then don't bother with all of the work to gather pages and
6036 6043 * generate the unsuccessful RPC. Just return from here and
6037 6044 * let the page be committed at some later time.
6038 6045 */
6039 6046 sigintr(&smask, VTOMI(vp)->mi_flags & MI_INT);
6040 6047 if (ttolwp(curthread) != NULL && ISSIG(curthread, JUSTLOOKING)) {
6041 6048 sigunintr(&smask);
6042 6049 page_unlock(pp);
6043 6050 mutex_exit(&rp->r_statelock);
6044 6051 return;
6045 6052 }
6046 6053 sigunintr(&smask);
6047 6054
6048 6055 /*
6049 6056 * We are starting to need to commit pages, so let's try
6050 6057 * to commit as many as possible at once to reduce the
6051 6058 * overhead.
6052 6059 *
6053 6060 * Set the `commit inprogress' state bit. We must
6054 6061 * first wait until any current one finishes. Then
6055 6062 * we initialize the c_pages list with this page.
6056 6063 */
6057 6064 while (rp->r_flags & RCOMMIT) {
6058 6065 rp->r_flags |= RCOMMITWAIT;
6059 6066 cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
6060 6067 rp->r_flags &= ~RCOMMITWAIT;
6061 6068 }
6062 6069 rp->r_flags |= RCOMMIT;
6063 6070 mutex_exit(&rp->r_statelock);
6064 6071 ASSERT(rp->r_commit.c_pages == NULL);
6065 6072 rp->r_commit.c_pages = pp;
6066 6073 rp->r_commit.c_commbase = (offset3)pp->p_offset;
6067 6074 rp->r_commit.c_commlen = PAGESIZE;
6068 6075
6069 6076 /*
6070 6077 * Gather together all other pages which can be committed.
6071 6078 * They will all be chained off r_commit.c_pages.
6072 6079 */
6073 6080 nfs3_get_commit(vp);
6074 6081
6075 6082 /*
6076 6083 * Clear the `commit inprogress' status and disconnect
6077 6084 * the list of pages to be committed from the rnode.
6078 6085 * At this same time, we also save the starting offset
6079 6086 * and length of data to be committed on the server.
6080 6087 */
6081 6088 plist = rp->r_commit.c_pages;
6082 6089 rp->r_commit.c_pages = NULL;
6083 6090 offset = rp->r_commit.c_commbase;
6084 6091 len = rp->r_commit.c_commlen;
6085 6092 mutex_enter(&rp->r_statelock);
6086 6093 rp->r_flags &= ~RCOMMIT;
6087 6094 cv_broadcast(&rp->r_commit.c_cv);
6088 6095 mutex_exit(&rp->r_statelock);
6089 6096
6090 6097 if (curproc == proc_pageout || curproc == proc_fsflush ||
6091 6098 nfs_zone() != VTOMI(vp)->mi_zone) {
6092 6099 nfs_async_commit(vp, plist, offset, len, cr, nfs3_async_commit);
6093 6100 return;
6094 6101 }
6095 6102
6096 6103 /*
6097 6104 * Actually generate the COMMIT3 over the wire operation.
6098 6105 */
6099 6106 error = nfs3_commit(vp, offset, len, cr);
6100 6107
6101 6108 /*
6102 6109 * If we got an error during the commit, just unlock all
6103 6110 * of the pages. The pages will get retransmitted to the
6104 6111 * server during a putpage operation.
6105 6112 */
6106 6113 if (error) {
6107 6114 while (plist != NULL) {
6108 6115 pptr = plist;
6109 6116 page_sub(&plist, pptr);
6110 6117 page_unlock(pptr);
6111 6118 }
6112 6119 return;
6113 6120 }
6114 6121
6115 6122 /*
6116 6123 * We've tried as hard as we can to commit the data to stable
6117 6124 * storage on the server. We release the rest of the pages
6118 6125 * and clear the commit required state. They will be put
6119 6126 * onto the tail of the cachelist if they are nolonger
6120 6127 * mapped.
6121 6128 */
6122 6129 while (plist != pp) {
6123 6130 pptr = plist;
6124 6131 page_sub(&plist, pptr);
6125 6132 pptr->p_fsdata = C_NOCOMMIT;
6126 6133 (void) page_release(pptr, 1);
6127 6134 }
6128 6135
6129 6136 /*
6130 6137 * It is possible that nfs3_commit didn't return error but
6131 6138 * some other thread has modified the page we are going
6132 6139 * to free/destroy.
6133 6140 * In this case we need to rewrite the page. Do an explicit check
6134 6141 * before attempting to free/destroy the page. If modified, needs to
6135 6142 * be rewritten so unlock the page and return.
6136 6143 */
6137 6144 if (hat_ismod(pp)) {
6138 6145 pp->p_fsdata = C_NOCOMMIT;
6139 6146 page_unlock(pp);
6140 6147 return;
6141 6148 }
6142 6149
6143 6150 /*
6144 6151 * Now, as appropriate, either free or destroy the page
6145 6152 * that we were called with.
6146 6153 */
6147 6154 pp->p_fsdata = C_NOCOMMIT;
6148 6155 if (fl == B_FREE)
6149 6156 page_free(pp, dn);
6150 6157 else
6151 6158 page_destroy(pp, dn);
6152 6159 }
6153 6160
6154 6161 static int
6155 6162 nfs3_commit(vnode_t *vp, offset3 offset, count3 count, cred_t *cr)
6156 6163 {
6157 6164 int error;
6158 6165 rnode_t *rp;
6159 6166 COMMIT3args args;
6160 6167 COMMIT3res res;
6161 6168 int douprintf;
6162 6169 cred_t *cred;
6163 6170
6164 6171 rp = VTOR(vp);
6165 6172 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6166 6173
6167 6174 mutex_enter(&rp->r_statelock);
6168 6175 if (rp->r_cred != NULL) {
6169 6176 cred = rp->r_cred;
6170 6177 crhold(cred);
6171 6178 } else {
6172 6179 rp->r_cred = cr;
6173 6180 crhold(cr);
6174 6181 cred = cr;
6175 6182 crhold(cred);
6176 6183 }
6177 6184 mutex_exit(&rp->r_statelock);
6178 6185
6179 6186 args.file = *VTOFH3(vp);
6180 6187 args.offset = offset;
6181 6188 args.count = count;
6182 6189
6183 6190 doitagain:
6184 6191 douprintf = 1;
6185 6192 error = rfs3call(VTOMI(vp), NFSPROC3_COMMIT,
6186 6193 xdr_COMMIT3args, (caddr_t)&args,
6187 6194 xdr_COMMIT3res, (caddr_t)&res, cred,
6188 6195 &douprintf, &res.status, 0, NULL);
6189 6196
6190 6197 crfree(cred);
6191 6198
6192 6199 if (error)
6193 6200 return (error);
6194 6201
6195 6202 error = geterrno3(res.status);
6196 6203 if (!error) {
6197 6204 ASSERT(rp->r_flags & RHAVEVERF);
6198 6205 mutex_enter(&rp->r_statelock);
6199 6206 if (rp->r_verf == res.resok.verf) {
6200 6207 mutex_exit(&rp->r_statelock);
6201 6208 return (0);
6202 6209 }
6203 6210 nfs3_set_mod(vp);
6204 6211 rp->r_verf = res.resok.verf;
6205 6212 mutex_exit(&rp->r_statelock);
6206 6213 error = NFS_VERF_MISMATCH;
6207 6214 } else {
6208 6215 if (error == EACCES) {
6209 6216 mutex_enter(&rp->r_statelock);
6210 6217 if (cred != cr) {
6211 6218 if (rp->r_cred != NULL)
6212 6219 crfree(rp->r_cred);
6213 6220 rp->r_cred = cr;
6214 6221 crhold(cr);
6215 6222 cred = cr;
6216 6223 crhold(cred);
6217 6224 mutex_exit(&rp->r_statelock);
6218 6225 goto doitagain;
6219 6226 }
6220 6227 mutex_exit(&rp->r_statelock);
6221 6228 }
6222 6229 /*
6223 6230 * Can't do a PURGE_STALE_FH here because this
6224 6231 * can cause a deadlock. nfs3_commit can
6225 6232 * be called from nfs3_dispose which can be called
6226 6233 * indirectly via pvn_vplist_dirty. PURGE_STALE_FH
6227 6234 * can call back to pvn_vplist_dirty.
6228 6235 */
6229 6236 if (error == ESTALE) {
6230 6237 mutex_enter(&rp->r_statelock);
6231 6238 rp->r_flags |= RSTALE;
6232 6239 if (!rp->r_error)
6233 6240 rp->r_error = error;
6234 6241 mutex_exit(&rp->r_statelock);
6235 6242 PURGE_ATTRCACHE(vp);
6236 6243 } else {
6237 6244 mutex_enter(&rp->r_statelock);
6238 6245 if (!rp->r_error)
6239 6246 rp->r_error = error;
6240 6247 mutex_exit(&rp->r_statelock);
6241 6248 }
6242 6249 }
6243 6250
6244 6251 return (error);
6245 6252 }
6246 6253
6247 6254 static void
6248 6255 nfs3_set_mod(vnode_t *vp)
6249 6256 {
6250 6257 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6251 6258
6252 6259 pvn_vplist_setdirty(vp, nfs_setmod_check);
6253 6260 }
6254 6261
6255 6262 /*
6256 6263 * This routine is used to gather together a page list of the pages
6257 6264 * which are to be committed on the server. This routine must not
6258 6265 * be called if the calling thread holds any locked pages.
6259 6266 *
6260 6267 * The calling thread must have set RCOMMIT. This bit is used to
6261 6268 * serialize access to the commit structure in the rnode. As long
6262 6269 * as the thread has set RCOMMIT, then it can manipulate the commit
6263 6270 * structure without requiring any other locks.
6264 6271 */
6265 6272 static void
6266 6273 nfs3_get_commit(vnode_t *vp)
6267 6274 {
6268 6275 rnode_t *rp;
6269 6276 page_t *pp;
6270 6277 kmutex_t *vphm;
6271 6278
6272 6279 rp = VTOR(vp);
6273 6280
6274 6281 ASSERT(rp->r_flags & RCOMMIT);
6275 6282
6276 6283 vphm = page_vnode_mutex(vp);
6277 6284 mutex_enter(vphm);
6278 6285
6279 6286 /*
6280 6287 * If there are no pages associated with this vnode, then
6281 6288 * just return.
6282 6289 */
6283 6290 if ((pp = vp->v_pages) == NULL) {
6284 6291 mutex_exit(vphm);
6285 6292 return;
6286 6293 }
6287 6294
6288 6295 /*
6289 6296 * Step through all of the pages associated with this vnode
6290 6297 * looking for pages which need to be committed.
6291 6298 */
6292 6299 do {
6293 6300 /* Skip marker pages. */
6294 6301 if (pp->p_hash == PVN_VPLIST_HASH_TAG)
6295 6302 continue;
6296 6303
6297 6304 /*
6298 6305 * If this page does not need to be committed or is
6299 6306 * modified, then just skip it.
6300 6307 */
6301 6308 if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp))
6302 6309 continue;
6303 6310
6304 6311 /*
6305 6312 * Attempt to lock the page. If we can't, then
6306 6313 * someone else is messing with it and we will
6307 6314 * just skip it.
6308 6315 */
6309 6316 if (!page_trylock(pp, SE_EXCL))
6310 6317 continue;
6311 6318
6312 6319 /*
6313 6320 * If this page does not need to be committed or is
6314 6321 * modified, then just skip it. Recheck now that
6315 6322 * the page is locked.
6316 6323 */
6317 6324 if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6318 6325 page_unlock(pp);
6319 6326 continue;
6320 6327 }
6321 6328
6322 6329 if (PP_ISFREE(pp)) {
6323 6330 cmn_err(CE_PANIC, "nfs3_get_commit: %p is free",
6324 6331 (void *)pp);
6325 6332 }
6326 6333
6327 6334 /*
6328 6335 * The page needs to be committed and we locked it.
6329 6336 * Update the base and length parameters and add it
6330 6337 * to r_pages.
6331 6338 */
6332 6339 if (rp->r_commit.c_pages == NULL) {
6333 6340 rp->r_commit.c_commbase = (offset3)pp->p_offset;
6334 6341 rp->r_commit.c_commlen = PAGESIZE;
6335 6342 } else if (pp->p_offset < rp->r_commit.c_commbase) {
6336 6343 rp->r_commit.c_commlen = rp->r_commit.c_commbase -
6337 6344 (offset3)pp->p_offset + rp->r_commit.c_commlen;
6338 6345 rp->r_commit.c_commbase = (offset3)pp->p_offset;
6339 6346 } else if ((rp->r_commit.c_commbase + rp->r_commit.c_commlen)
6340 6347 <= pp->p_offset) {
6341 6348 rp->r_commit.c_commlen = (offset3)pp->p_offset -
6342 6349 rp->r_commit.c_commbase + PAGESIZE;
6343 6350 }
6344 6351 page_add(&rp->r_commit.c_pages, pp);
6345 6352 } while ((pp = pp->p_vpnext) != vp->v_pages);
6346 6353
6347 6354 mutex_exit(vphm);
6348 6355 }
6349 6356
6350 6357 /*
6351 6358 * This routine is used to gather together a page list of the pages
6352 6359 * which are to be committed on the server. This routine must not
6353 6360 * be called if the calling thread holds any locked pages.
6354 6361 *
6355 6362 * The calling thread must have set RCOMMIT. This bit is used to
6356 6363 * serialize access to the commit structure in the rnode. As long
6357 6364 * as the thread has set RCOMMIT, then it can manipulate the commit
6358 6365 * structure without requiring any other locks.
6359 6366 */
6360 6367 static void
6361 6368 nfs3_get_commit_range(vnode_t *vp, u_offset_t soff, size_t len)
6362 6369 {
6363 6370
6364 6371 rnode_t *rp;
6365 6372 page_t *pp;
6366 6373 u_offset_t end;
6367 6374 u_offset_t off;
6368 6375
6369 6376 ASSERT(len != 0);
6370 6377
6371 6378 rp = VTOR(vp);
6372 6379
6373 6380 ASSERT(rp->r_flags & RCOMMIT);
6374 6381 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6375 6382
6376 6383 /*
6377 6384 * If there are no pages associated with this vnode, then
6378 6385 * just return.
6379 6386 */
6380 6387 if ((pp = vp->v_pages) == NULL)
6381 6388 return;
6382 6389
6383 6390 /*
6384 6391 * Calculate the ending offset.
6385 6392 */
6386 6393 end = soff + len;
6387 6394
6388 6395 for (off = soff; off < end; off += PAGESIZE) {
6389 6396 /*
6390 6397 * Lookup each page by vp, offset.
6391 6398 */
6392 6399 if ((pp = page_lookup_nowait(vp, off, SE_EXCL)) == NULL)
6393 6400 continue;
6394 6401
6395 6402 /*
6396 6403 * If this page does not need to be committed or is
6397 6404 * modified, then just skip it.
6398 6405 */
6399 6406 if (pp->p_fsdata == C_NOCOMMIT || hat_ismod(pp)) {
6400 6407 page_unlock(pp);
6401 6408 continue;
6402 6409 }
6403 6410
6404 6411 ASSERT(PP_ISFREE(pp) == 0);
6405 6412
6406 6413 /*
6407 6414 * The page needs to be committed and we locked it.
6408 6415 * Update the base and length parameters and add it
6409 6416 * to r_pages.
6410 6417 */
6411 6418 if (rp->r_commit.c_pages == NULL) {
6412 6419 rp->r_commit.c_commbase = (offset3)pp->p_offset;
6413 6420 rp->r_commit.c_commlen = PAGESIZE;
6414 6421 } else {
6415 6422 rp->r_commit.c_commlen = (offset3)pp->p_offset -
6416 6423 rp->r_commit.c_commbase + PAGESIZE;
6417 6424 }
6418 6425 page_add(&rp->r_commit.c_pages, pp);
6419 6426 }
6420 6427 }
6421 6428
6422 6429 static int
6423 6430 nfs3_putpage_commit(vnode_t *vp, offset_t poff, size_t plen, cred_t *cr)
6424 6431 {
6425 6432 int error;
6426 6433 writeverf3 write_verf;
6427 6434 rnode_t *rp = VTOR(vp);
6428 6435
6429 6436 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6430 6437 /*
6431 6438 * Flush the data portion of the file and then commit any
6432 6439 * portions which need to be committed. This may need to
6433 6440 * be done twice if the server has changed state since
6434 6441 * data was last written. The data will need to be
6435 6442 * rewritten to the server and then a new commit done.
6436 6443 *
6437 6444 * In fact, this may need to be done several times if the
6438 6445 * server is having problems and crashing while we are
6439 6446 * attempting to do this.
6440 6447 */
6441 6448
6442 6449 top:
6443 6450 /*
6444 6451 * Do a flush based on the poff and plen arguments. This
6445 6452 * will asynchronously write out any modified pages in the
6446 6453 * range specified by (poff, plen). This starts all of the
6447 6454 * i/o operations which will be waited for in the next
6448 6455 * call to nfs3_putpage
6449 6456 */
6450 6457
6451 6458 mutex_enter(&rp->r_statelock);
6452 6459 write_verf = rp->r_verf;
6453 6460 mutex_exit(&rp->r_statelock);
6454 6461
6455 6462 error = nfs3_putpage(vp, poff, plen, B_ASYNC, cr, NULL);
6456 6463 if (error == EAGAIN)
6457 6464 error = 0;
6458 6465
6459 6466 /*
6460 6467 * Do a flush based on the poff and plen arguments. This
6461 6468 * will synchronously write out any modified pages in the
6462 6469 * range specified by (poff, plen) and wait until all of
6463 6470 * the asynchronous i/o's in that range are done as well.
6464 6471 */
6465 6472 if (!error)
6466 6473 error = nfs3_putpage(vp, poff, plen, 0, cr, NULL);
6467 6474
6468 6475 if (error)
6469 6476 return (error);
6470 6477
6471 6478 mutex_enter(&rp->r_statelock);
6472 6479 if (rp->r_verf != write_verf) {
6473 6480 mutex_exit(&rp->r_statelock);
6474 6481 goto top;
6475 6482 }
6476 6483 mutex_exit(&rp->r_statelock);
6477 6484
6478 6485 /*
6479 6486 * Now commit any pages which might need to be committed.
6480 6487 * If the error, NFS_VERF_MISMATCH, is returned, then
6481 6488 * start over with the flush operation.
6482 6489 */
6483 6490
6484 6491 error = nfs3_commit_vp(vp, poff, plen, cr);
6485 6492
6486 6493 if (error == NFS_VERF_MISMATCH)
6487 6494 goto top;
6488 6495
6489 6496 return (error);
6490 6497 }
6491 6498
6492 6499 static int
6493 6500 nfs3_commit_vp(vnode_t *vp, u_offset_t poff, size_t plen, cred_t *cr)
6494 6501 {
6495 6502 rnode_t *rp;
6496 6503 page_t *plist;
6497 6504 offset3 offset;
6498 6505 count3 len;
6499 6506
6500 6507
6501 6508 rp = VTOR(vp);
6502 6509
6503 6510 if (nfs_zone() != VTOMI(vp)->mi_zone)
6504 6511 return (EIO);
6505 6512 /*
6506 6513 * Set the `commit inprogress' state bit. We must
6507 6514 * first wait until any current one finishes.
6508 6515 */
6509 6516 mutex_enter(&rp->r_statelock);
6510 6517 while (rp->r_flags & RCOMMIT) {
6511 6518 rp->r_flags |= RCOMMITWAIT;
6512 6519 cv_wait(&rp->r_commit.c_cv, &rp->r_statelock);
6513 6520 rp->r_flags &= ~RCOMMITWAIT;
6514 6521 }
6515 6522 rp->r_flags |= RCOMMIT;
6516 6523 mutex_exit(&rp->r_statelock);
6517 6524
6518 6525 /*
6519 6526 * Gather together all of the pages which need to be
6520 6527 * committed.
6521 6528 */
6522 6529 if (plen == 0)
6523 6530 nfs3_get_commit(vp);
6524 6531 else
6525 6532 nfs3_get_commit_range(vp, poff, plen);
6526 6533
6527 6534 /*
6528 6535 * Clear the `commit inprogress' bit and disconnect the
6529 6536 * page list which was gathered together in nfs3_get_commit.
6530 6537 */
6531 6538 plist = rp->r_commit.c_pages;
6532 6539 rp->r_commit.c_pages = NULL;
6533 6540 offset = rp->r_commit.c_commbase;
6534 6541 len = rp->r_commit.c_commlen;
6535 6542 mutex_enter(&rp->r_statelock);
6536 6543 rp->r_flags &= ~RCOMMIT;
6537 6544 cv_broadcast(&rp->r_commit.c_cv);
6538 6545 mutex_exit(&rp->r_statelock);
6539 6546
6540 6547 /*
6541 6548 * If any pages need to be committed, commit them and
6542 6549 * then unlock them so that they can be freed some
6543 6550 * time later.
6544 6551 */
6545 6552 if (plist != NULL) {
6546 6553 /*
6547 6554 * No error occurred during the flush portion
6548 6555 * of this operation, so now attempt to commit
6549 6556 * the data to stable storage on the server.
6550 6557 *
6551 6558 * This will unlock all of the pages on the list.
6552 6559 */
6553 6560 return (nfs3_sync_commit(vp, plist, offset, len, cr));
6554 6561 }
6555 6562 return (0);
6556 6563 }
6557 6564
6558 6565 static int
6559 6566 nfs3_sync_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6560 6567 cred_t *cr)
6561 6568 {
6562 6569 int error;
6563 6570 page_t *pp;
6564 6571
6565 6572 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6566 6573 error = nfs3_commit(vp, offset, count, cr);
6567 6574
6568 6575 /*
6569 6576 * If we got an error, then just unlock all of the pages
6570 6577 * on the list.
6571 6578 */
6572 6579 if (error) {
6573 6580 while (plist != NULL) {
6574 6581 pp = plist;
6575 6582 page_sub(&plist, pp);
6576 6583 page_unlock(pp);
6577 6584 }
6578 6585 return (error);
6579 6586 }
6580 6587 /*
6581 6588 * We've tried as hard as we can to commit the data to stable
6582 6589 * storage on the server. We just unlock the pages and clear
6583 6590 * the commit required state. They will get freed later.
6584 6591 */
6585 6592 while (plist != NULL) {
6586 6593 pp = plist;
6587 6594 page_sub(&plist, pp);
6588 6595 pp->p_fsdata = C_NOCOMMIT;
6589 6596 page_unlock(pp);
6590 6597 }
6591 6598
6592 6599 return (error);
6593 6600 }
6594 6601
6595 6602 static void
6596 6603 nfs3_async_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
6597 6604 cred_t *cr)
6598 6605 {
6599 6606 ASSERT(nfs_zone() == VTOMI(vp)->mi_zone);
6600 6607 (void) nfs3_sync_commit(vp, plist, offset, count, cr);
6601 6608 }
6602 6609
6603 6610 /* ARGSUSED */
6604 6611 static int
6605 6612 nfs3_setsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr,
6606 6613 caller_context_t *ct)
6607 6614 {
6608 6615 int error;
6609 6616 mntinfo_t *mi;
6610 6617
6611 6618 mi = VTOMI(vp);
6612 6619
6613 6620 if (nfs_zone() != mi->mi_zone)
6614 6621 return (EIO);
6615 6622
6616 6623 if (mi->mi_flags & MI_ACL) {
6617 6624 error = acl_setacl3(vp, vsecattr, flag, cr);
6618 6625 if (mi->mi_flags & MI_ACL)
6619 6626 return (error);
6620 6627 }
6621 6628
6622 6629 return (ENOSYS);
6623 6630 }
6624 6631
6625 6632 /* ARGSUSED */
6626 6633 static int
6627 6634 nfs3_getsecattr(vnode_t *vp, vsecattr_t *vsecattr, int flag, cred_t *cr,
6628 6635 caller_context_t *ct)
6629 6636 {
6630 6637 int error;
6631 6638 mntinfo_t *mi;
6632 6639
6633 6640 mi = VTOMI(vp);
6634 6641
6635 6642 if (nfs_zone() != mi->mi_zone)
6636 6643 return (EIO);
6637 6644
6638 6645 if (mi->mi_flags & MI_ACL) {
6639 6646 error = acl_getacl3(vp, vsecattr, flag, cr);
6640 6647 if (mi->mi_flags & MI_ACL)
6641 6648 return (error);
6642 6649 }
6643 6650
6644 6651 return (fs_fab_acl(vp, vsecattr, flag, cr, ct));
6645 6652 }
6646 6653
6647 6654 /* ARGSUSED */
6648 6655 static int
6649 6656 nfs3_shrlock(vnode_t *vp, int cmd, struct shrlock *shr, int flag, cred_t *cr,
6650 6657 caller_context_t *ct)
6651 6658 {
6652 6659 int error;
6653 6660 struct shrlock nshr;
6654 6661 struct nfs_owner nfs_owner;
6655 6662 netobj lm_fh3;
6656 6663
6657 6664 if (nfs_zone() != VTOMI(vp)->mi_zone)
6658 6665 return (EIO);
6659 6666
6660 6667 /*
6661 6668 * check for valid cmd parameter
6662 6669 */
6663 6670 if (cmd != F_SHARE && cmd != F_UNSHARE && cmd != F_HASREMOTELOCKS)
6664 6671 return (EINVAL);
6665 6672
6666 6673 /*
6667 6674 * Check access permissions
6668 6675 */
6669 6676 if (cmd == F_SHARE &&
6670 6677 (((shr->s_access & F_RDACC) && !(flag & FREAD)) ||
6671 6678 ((shr->s_access & F_WRACC) && !(flag & FWRITE))))
6672 6679 return (EBADF);
6673 6680
6674 6681 /*
6675 6682 * If the filesystem is mounted using local locking, pass the
6676 6683 * request off to the local share code.
6677 6684 */
6678 6685 if (VTOMI(vp)->mi_flags & MI_LLOCK)
6679 6686 return (fs_shrlock(vp, cmd, shr, flag, cr, ct));
6680 6687
6681 6688 switch (cmd) {
6682 6689 case F_SHARE:
6683 6690 case F_UNSHARE:
6684 6691 lm_fh3.n_len = VTOFH3(vp)->fh3_length;
6685 6692 lm_fh3.n_bytes = (char *)&(VTOFH3(vp)->fh3_u.data);
6686 6693
6687 6694 /*
6688 6695 * If passed an owner that is too large to fit in an
6689 6696 * nfs_owner it is likely a recursive call from the
6690 6697 * lock manager client and pass it straight through. If
6691 6698 * it is not a nfs_owner then simply return an error.
6692 6699 */
6693 6700 if (shr->s_own_len > sizeof (nfs_owner.lowner)) {
6694 6701 if (((struct nfs_owner *)shr->s_owner)->magic !=
6695 6702 NFS_OWNER_MAGIC)
6696 6703 return (EINVAL);
6697 6704
6698 6705 if (error = lm4_shrlock(vp, cmd, shr, flag, &lm_fh3)) {
6699 6706 error = set_errno(error);
6700 6707 }
6701 6708 return (error);
6702 6709 }
6703 6710 /*
6704 6711 * Remote share reservations owner is a combination of
6705 6712 * a magic number, hostname, and the local owner
6706 6713 */
6707 6714 bzero(&nfs_owner, sizeof (nfs_owner));
6708 6715 nfs_owner.magic = NFS_OWNER_MAGIC;
6709 6716 (void) strncpy(nfs_owner.hname, uts_nodename(),
6710 6717 sizeof (nfs_owner.hname));
6711 6718 bcopy(shr->s_owner, nfs_owner.lowner, shr->s_own_len);
6712 6719 nshr.s_access = shr->s_access;
6713 6720 nshr.s_deny = shr->s_deny;
6714 6721 nshr.s_sysid = 0;
6715 6722 nshr.s_pid = ttoproc(curthread)->p_pid;
6716 6723 nshr.s_own_len = sizeof (nfs_owner);
6717 6724 nshr.s_owner = (caddr_t)&nfs_owner;
6718 6725
6719 6726 if (error = lm4_shrlock(vp, cmd, &nshr, flag, &lm_fh3)) {
6720 6727 error = set_errno(error);
6721 6728 }
6722 6729
6723 6730 break;
6724 6731
6725 6732 case F_HASREMOTELOCKS:
6726 6733 /*
6727 6734 * NFS client can't store remote locks itself
6728 6735 */
6729 6736 shr->s_access = 0;
6730 6737 error = 0;
6731 6738 break;
6732 6739
6733 6740 default:
6734 6741 error = EINVAL;
6735 6742 break;
6736 6743 }
6737 6744
6738 6745 return (error);
6739 6746 }
|
↓ open down ↓ |
4415 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX