1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* 27 * ZFS fault injection 28 * 29 * To handle fault injection, we keep track of a series of zinject_record_t 30 * structures which describe which logical block(s) should be injected with a 31 * fault. These are kept in a global list. Each record corresponds to a given 32 * spa_t and maintains a special hold on the spa_t so that it cannot be deleted 33 * or exported while the injection record exists. 34 * 35 * Device level injection is done using the 'zi_guid' field. If this is set, it 36 * means that the error is destined for a particular device, not a piece of 37 * data. 38 * 39 * This is a rather poor data structure and algorithm, but we don't expect more 40 * than a few faults at any one time, so it should be sufficient for our needs. 41 */ 42 43 #include <sys/arc.h> 44 #include <sys/zio_impl.h> 45 #include <sys/zfs_ioctl.h> 46 #include <sys/vdev_impl.h> 47 #include <sys/dmu_objset.h> 48 #include <sys/fs/zfs.h> 49 50 uint32_t zio_injection_enabled; 51 52 typedef struct inject_handler { 53 int zi_id; 54 spa_t *zi_spa; 55 zinject_record_t zi_record; 56 list_node_t zi_link; 57 } inject_handler_t; 58 59 static list_t inject_handlers; 60 static krwlock_t inject_lock; 61 static int inject_next_id = 1; 62 63 /* 64 * Returns true if the given record matches the I/O in progress. 65 */ 66 static boolean_t 67 zio_match_handler(zbookmark_t *zb, uint64_t type, 68 zinject_record_t *record, int error) 69 { 70 /* 71 * Check for a match against the MOS, which is based on type 72 */ 73 if (zb->zb_objset == DMU_META_OBJSET && 74 record->zi_objset == DMU_META_OBJSET && 75 record->zi_object == DMU_META_DNODE_OBJECT) { 76 if (record->zi_type == DMU_OT_NONE || 77 type == record->zi_type) 78 return (record->zi_freq == 0 || 79 spa_get_random(100) < record->zi_freq); 80 else 81 return (B_FALSE); 82 } 83 84 /* 85 * Check for an exact match. 86 */ 87 if (zb->zb_objset == record->zi_objset && 88 zb->zb_object == record->zi_object && 89 zb->zb_level == record->zi_level && 90 zb->zb_blkid >= record->zi_start && 91 zb->zb_blkid <= record->zi_end && 92 error == record->zi_error) 93 return (record->zi_freq == 0 || 94 spa_get_random(100) < record->zi_freq); 95 96 return (B_FALSE); 97 } 98 99 /* 100 * Panic the system when a config change happens in the function 101 * specified by tag. 102 */ 103 void 104 zio_handle_panic_injection(spa_t *spa, char *tag) 105 { 106 inject_handler_t *handler; 107 108 rw_enter(&inject_lock, RW_READER); 109 110 for (handler = list_head(&inject_handlers); handler != NULL; 111 handler = list_next(&inject_handlers, handler)) { 112 113 if (spa != handler->zi_spa) 114 continue; 115 116 if (strcmp(tag, handler->zi_record.zi_func) == 0) 117 panic("Panic requested in function %s\n", tag); 118 } 119 120 rw_exit(&inject_lock); 121 } 122 123 /* 124 * Determine if the I/O in question should return failure. Returns the errno 125 * to be returned to the caller. 126 */ 127 int 128 zio_handle_fault_injection(zio_t *zio, int error) 129 { 130 int ret = 0; 131 inject_handler_t *handler; 132 133 /* 134 * Ignore I/O not associated with any logical data. 135 */ 136 if (zio->io_logical == NULL) 137 return (0); 138 139 /* 140 * Currently, we only support fault injection on reads. 141 */ 142 if (zio->io_type != ZIO_TYPE_READ) 143 return (0); 144 145 rw_enter(&inject_lock, RW_READER); 146 147 for (handler = list_head(&inject_handlers); handler != NULL; 148 handler = list_next(&inject_handlers, handler)) { 149 150 /* Ignore errors not destined for this pool */ 151 if (zio->io_spa != handler->zi_spa) 152 continue; 153 154 /* Ignore device errors and panic injection */ 155 if (handler->zi_record.zi_guid != 0 || 156 handler->zi_record.zi_func[0] != '\0' || 157 handler->zi_record.zi_duration != 0) 158 continue; 159 160 /* If this handler matches, return EIO */ 161 if (zio_match_handler(&zio->io_logical->io_bookmark, 162 zio->io_bp ? BP_GET_TYPE(zio->io_bp) : DMU_OT_NONE, 163 &handler->zi_record, error)) { 164 ret = error; 165 break; 166 } 167 } 168 169 rw_exit(&inject_lock); 170 171 return (ret); 172 } 173 174 /* 175 * Determine if the zio is part of a label update and has an injection 176 * handler associated with that portion of the label. Currently, we 177 * allow error injection in either the nvlist or the uberblock region of 178 * of the vdev label. 179 */ 180 int 181 zio_handle_label_injection(zio_t *zio, int error) 182 { 183 inject_handler_t *handler; 184 vdev_t *vd = zio->io_vd; 185 uint64_t offset = zio->io_offset; 186 int label; 187 int ret = 0; 188 189 if (offset >= VDEV_LABEL_START_SIZE && 190 offset < vd->vdev_psize - VDEV_LABEL_END_SIZE) 191 return (0); 192 193 rw_enter(&inject_lock, RW_READER); 194 195 for (handler = list_head(&inject_handlers); handler != NULL; 196 handler = list_next(&inject_handlers, handler)) { 197 uint64_t start = handler->zi_record.zi_start; 198 uint64_t end = handler->zi_record.zi_end; 199 200 /* Ignore device only faults or panic injection */ 201 if (handler->zi_record.zi_start == 0 || 202 handler->zi_record.zi_func[0] != '\0' || 203 handler->zi_record.zi_duration != 0) 204 continue; 205 206 /* 207 * The injection region is the relative offsets within a 208 * vdev label. We must determine the label which is being 209 * updated and adjust our region accordingly. 210 */ 211 label = vdev_label_number(vd->vdev_psize, offset); 212 start = vdev_label_offset(vd->vdev_psize, label, start); 213 end = vdev_label_offset(vd->vdev_psize, label, end); 214 215 if (zio->io_vd->vdev_guid == handler->zi_record.zi_guid && 216 (offset >= start && offset <= end)) { 217 ret = error; 218 break; 219 } 220 } 221 rw_exit(&inject_lock); 222 return (ret); 223 } 224 225 226 int 227 zio_handle_device_injection(vdev_t *vd, zio_t *zio, int error) 228 { 229 inject_handler_t *handler; 230 int ret = 0; 231 232 /* 233 * We skip over faults in the labels unless it's during 234 * device open (i.e. zio == NULL). 235 */ 236 if (zio != NULL) { 237 uint64_t offset = zio->io_offset; 238 239 if (offset < VDEV_LABEL_START_SIZE || 240 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE) 241 return (0); 242 } 243 244 rw_enter(&inject_lock, RW_READER); 245 246 for (handler = list_head(&inject_handlers); handler != NULL; 247 handler = list_next(&inject_handlers, handler)) { 248 249 /* 250 * Ignore label specific faults, panic injection 251 * or fake writes 252 */ 253 if (handler->zi_record.zi_start != 0 || 254 handler->zi_record.zi_func[0] != '\0' || 255 handler->zi_record.zi_duration != 0) 256 continue; 257 258 if (vd->vdev_guid == handler->zi_record.zi_guid) { 259 if (handler->zi_record.zi_failfast && 260 (zio == NULL || (zio->io_flags & 261 (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))) { 262 continue; 263 } 264 265 /* Handle type specific I/O failures */ 266 if (zio != NULL && 267 handler->zi_record.zi_iotype != ZIO_TYPES && 268 handler->zi_record.zi_iotype != zio->io_type) 269 continue; 270 271 if (handler->zi_record.zi_error == error) { 272 /* 273 * For a failed open, pretend like the device 274 * has gone away. 275 */ 276 if (error == ENXIO) 277 vd->vdev_stat.vs_aux = 278 VDEV_AUX_OPEN_FAILED; 279 ret = error; 280 break; 281 } 282 if (handler->zi_record.zi_error == ENXIO) { 283 ret = EIO; 284 break; 285 } 286 } 287 } 288 289 rw_exit(&inject_lock); 290 291 return (ret); 292 } 293 294 /* 295 * Simulate hardware that ignores cache flushes. For requested number 296 * of seconds nix the actual writing to disk. 297 */ 298 void 299 zio_handle_ignored_writes(zio_t *zio) 300 { 301 inject_handler_t *handler; 302 303 rw_enter(&inject_lock, RW_READER); 304 305 for (handler = list_head(&inject_handlers); handler != NULL; 306 handler = list_next(&inject_handlers, handler)) { 307 308 /* Ignore errors not destined for this pool */ 309 if (zio->io_spa != handler->zi_spa) 310 continue; 311 312 if (handler->zi_record.zi_duration == 0) 313 continue; 314 315 /* 316 * Positive duration implies # of seconds, negative 317 * a number of txgs 318 */ 319 if (handler->zi_record.zi_timer == 0) { 320 if (handler->zi_record.zi_duration > 0) 321 handler->zi_record.zi_timer = lbolt64; 322 else 323 handler->zi_record.zi_timer = zio->io_txg; 324 } 325 326 /* Have a "problem" writing 60% of the time */ 327 if (spa_get_random(100) < 60) 328 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 329 break; 330 } 331 332 rw_exit(&inject_lock); 333 } 334 335 void 336 spa_handle_ignored_writes(spa_t *spa) 337 { 338 inject_handler_t *handler; 339 340 if (zio_injection_enabled == 0) 341 return; 342 343 rw_enter(&inject_lock, RW_READER); 344 345 for (handler = list_head(&inject_handlers); handler != NULL; 346 handler = list_next(&inject_handlers, handler)) { 347 348 /* Ignore errors not destined for this pool */ 349 if (spa != handler->zi_spa) 350 continue; 351 352 if (handler->zi_record.zi_duration == 0) 353 continue; 354 355 if (handler->zi_record.zi_duration > 0) { 356 VERIFY(handler->zi_record.zi_timer == 0 || 357 handler->zi_record.zi_timer + 358 handler->zi_record.zi_duration * hz > lbolt64); 359 } else { 360 /* duration is negative so the subtraction here adds */ 361 VERIFY(handler->zi_record.zi_timer == 0 || 362 handler->zi_record.zi_timer - 363 handler->zi_record.zi_duration >= 364 spa_syncing_txg(spa)); 365 } 366 } 367 368 rw_exit(&inject_lock); 369 } 370 371 /* 372 * Create a new handler for the given record. We add it to the list, adding 373 * a reference to the spa_t in the process. We increment zio_injection_enabled, 374 * which is the switch to trigger all fault injection. 375 */ 376 int 377 zio_inject_fault(char *name, int flags, int *id, zinject_record_t *record) 378 { 379 inject_handler_t *handler; 380 int error; 381 spa_t *spa; 382 383 /* 384 * If this is pool-wide metadata, make sure we unload the corresponding 385 * spa_t, so that the next attempt to load it will trigger the fault. 386 * We call spa_reset() to unload the pool appropriately. 387 */ 388 if (flags & ZINJECT_UNLOAD_SPA) 389 if ((error = spa_reset(name)) != 0) 390 return (error); 391 392 if (!(flags & ZINJECT_NULL)) { 393 /* 394 * spa_inject_ref() will add an injection reference, which will 395 * prevent the pool from being removed from the namespace while 396 * still allowing it to be unloaded. 397 */ 398 if ((spa = spa_inject_addref(name)) == NULL) 399 return (ENOENT); 400 401 handler = kmem_alloc(sizeof (inject_handler_t), KM_SLEEP); 402 403 rw_enter(&inject_lock, RW_WRITER); 404 405 *id = handler->zi_id = inject_next_id++; 406 handler->zi_spa = spa; 407 handler->zi_record = *record; 408 list_insert_tail(&inject_handlers, handler); 409 atomic_add_32(&zio_injection_enabled, 1); 410 411 rw_exit(&inject_lock); 412 } 413 414 /* 415 * Flush the ARC, so that any attempts to read this data will end up 416 * going to the ZIO layer. Note that this is a little overkill, but 417 * we don't have the necessary ARC interfaces to do anything else, and 418 * fault injection isn't a performance critical path. 419 */ 420 if (flags & ZINJECT_FLUSH_ARC) 421 arc_flush(NULL); 422 423 return (0); 424 } 425 426 /* 427 * Returns the next record with an ID greater than that supplied to the 428 * function. Used to iterate over all handlers in the system. 429 */ 430 int 431 zio_inject_list_next(int *id, char *name, size_t buflen, 432 zinject_record_t *record) 433 { 434 inject_handler_t *handler; 435 int ret; 436 437 mutex_enter(&spa_namespace_lock); 438 rw_enter(&inject_lock, RW_READER); 439 440 for (handler = list_head(&inject_handlers); handler != NULL; 441 handler = list_next(&inject_handlers, handler)) 442 if (handler->zi_id > *id) 443 break; 444 445 if (handler) { 446 *record = handler->zi_record; 447 *id = handler->zi_id; 448 (void) strncpy(name, spa_name(handler->zi_spa), buflen); 449 ret = 0; 450 } else { 451 ret = ENOENT; 452 } 453 454 rw_exit(&inject_lock); 455 mutex_exit(&spa_namespace_lock); 456 457 return (ret); 458 } 459 460 /* 461 * Clear the fault handler with the given identifier, or return ENOENT if none 462 * exists. 463 */ 464 int 465 zio_clear_fault(int id) 466 { 467 inject_handler_t *handler; 468 int ret; 469 470 rw_enter(&inject_lock, RW_WRITER); 471 472 for (handler = list_head(&inject_handlers); handler != NULL; 473 handler = list_next(&inject_handlers, handler)) 474 if (handler->zi_id == id) 475 break; 476 477 if (handler == NULL) { 478 ret = ENOENT; 479 } else { 480 list_remove(&inject_handlers, handler); 481 spa_inject_delref(handler->zi_spa); 482 kmem_free(handler, sizeof (inject_handler_t)); 483 atomic_add_32(&zio_injection_enabled, -1); 484 ret = 0; 485 } 486 487 rw_exit(&inject_lock); 488 489 return (ret); 490 } 491 492 void 493 zio_inject_init(void) 494 { 495 rw_init(&inject_lock, NULL, RW_DEFAULT, NULL); 496 list_create(&inject_handlers, sizeof (inject_handler_t), 497 offsetof(inject_handler_t, zi_link)); 498 } 499 500 void 501 zio_inject_fini(void) 502 { 503 list_destroy(&inject_handlers); 504 rw_destroy(&inject_lock); 505 } 506