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 (c) 2017 by Lawrence Livermore National Security, LLC. 23 */ 24 25 #include <sys/abd.h> 26 #include <sys/mmp.h> 27 #include <sys/spa.h> 28 #include <sys/spa_impl.h> 29 #include <sys/time.h> 30 #include <sys/vdev.h> 31 #include <sys/vdev_impl.h> 32 #include <sys/zfs_context.h> 33 #include <sys/callb.h> 34 35 /* 36 * Multi-Modifier Protection (MMP) attempts to prevent a user from importing 37 * or opening a pool on more than one host at a time. In particular, it 38 * prevents "zpool import -f" on a host from succeeding while the pool is 39 * already imported on another host. There are many other ways in which a 40 * device could be used by two hosts for different purposes at the same time 41 * resulting in pool damage. This implementation does not attempt to detect 42 * those cases. 43 * 44 * MMP operates by ensuring there are frequent visible changes on disk (a 45 * "heartbeat") at all times. And by altering the import process to check 46 * for these changes and failing the import when they are detected. This 47 * functionality is enabled by setting the 'multihost' pool property to on. 48 * 49 * Uberblocks written by the txg_sync thread always go into the first 50 * (N-MMP_BLOCKS_PER_LABEL) slots, the remaining slots are reserved for MMP. 51 * They are used to hold uberblocks which are exactly the same as the last 52 * synced uberblock except that the ub_timestamp and mmp_config are frequently 53 * updated. Like all other uberblocks, the slot is written with an embedded 54 * checksum, and slots with invalid checksums are ignored. This provides the 55 * "heartbeat", with no risk of overwriting good uberblocks that must be 56 * preserved, e.g. previous txgs and associated block pointers. 57 * 58 * Three optional fields are added to uberblock structure; ub_mmp_magic, 59 * ub_mmp_config, and ub_mmp_delay. The ub_mmp_magic value allows zfs to tell 60 * whether the other ub_mmp_* fields are valid. The ub_mmp_config field tells 61 * the importing host the settings of zfs_multihost_interval and 62 * zfs_multihost_fail_intervals on the host which last had (or currently has) 63 * the pool imported. These determine how long a host must wait to detect 64 * activity in the pool, before concluding the pool is not in use. The 65 * mmp_delay field is a decaying average of the amount of time between 66 * completion of successive MMP writes, in nanoseconds. It indicates whether 67 * MMP is enabled. 68 * 69 * During import an activity test may now be performed to determine if 70 * the pool is in use. The activity test is typically required if the 71 * ZPOOL_CONFIG_HOSTID does not match the system hostid, the pool state is 72 * POOL_STATE_ACTIVE, and the pool is not a root pool. 73 * 74 * The activity test finds the "best" uberblock (highest txg, timestamp, and, if 75 * ub_mmp_magic is valid, sequence number from ub_mmp_config). It then waits 76 * some time, and finds the "best" uberblock again. If any of the mentioned 77 * fields have different values in the newly read uberblock, the pool is in use 78 * by another host and the import fails. In order to assure the accuracy of the 79 * activity test, the default values result in an activity test duration of 20x 80 * the mmp write interval. 81 * 82 * The duration of the "zpool import" activity test depends on the information 83 * available in the "best" uberblock: 84 * 85 * 1) If uberblock was written by zfs-0.8 or newer and fail_intervals > 0: 86 * ub_mmp_config.fail_intervals * ub_mmp_config.multihost_interval * 2 87 * 88 * In this case, a weak guarantee is provided. Since the host which last had 89 * the pool imported will suspend the pool if no mmp writes land within 90 * fail_intervals * multihost_interval ms, the absence of writes during that 91 * time means either the pool is not imported, or it is imported but the pool 92 * is suspended and no further writes will occur. 93 * 94 * Note that resuming the suspended pool on the remote host would invalidate 95 * this guarantee, and so it is not allowed. 96 * 97 * The factor of 2 provides a conservative safety factor and derives from 98 * MMP_IMPORT_SAFETY_FACTOR; 99 * 100 * 2) If uberblock was written by zfs-0.8 or newer and fail_intervals == 0: 101 * (ub_mmp_config.multihost_interval + ub_mmp_delay) * 102 * zfs_multihost_import_intervals 103 * 104 * In this case no guarantee can provided. However, as long as some devices 105 * are healthy and connected, it is likely that at least one write will land 106 * within (multihost_interval + mmp_delay) because multihost_interval is 107 * enough time for a write to be attempted to each leaf vdev, and mmp_delay 108 * is enough for one to land, based on past delays. Multiplying by 109 * zfs_multihost_import_intervals provides a conservative safety factor. 110 * 111 * 3) If uberblock was written by zfs-0.7: 112 * (zfs_multihost_interval + ub_mmp_delay) * zfs_multihost_import_intervals 113 * 114 * The same logic as case #2 applies, but we do not know remote tunables. 115 * 116 * We use the local value for zfs_multihost_interval because the original MMP 117 * did not record this value in the uberblock. 118 * 119 * ub_mmp_delay >= (zfs_multihost_interval / leaves), so if the other host 120 * has a much larger zfs_multihost_interval set, ub_mmp_delay will reflect 121 * that. We will have waited enough time for zfs_multihost_import_intervals 122 * writes to be issued and all but one to land. 123 * 124 * single device pool example delays 125 * 126 * import_delay = (1 + 1) * 20 = 40s #defaults, no I/O delay 127 * import_delay = (1 + 10) * 20 = 220s #defaults, 10s I/O delay 128 * import_delay = (10 + 10) * 20 = 400s #10s multihost_interval, 129 * no I/O delay 130 * 100 device pool example delays 131 * 132 * import_delay = (1 + .01) * 20 = 20s #defaults, no I/O delay 133 * import_delay = (1 + 10) * 20 = 220s #defaults, 10s I/O delay 134 * import_delay = (10 + .1) * 20 = 202s #10s multihost_interval, 135 * no I/O delay 136 * 137 * 4) Otherwise, this uberblock was written by a pre-MMP zfs: 138 * zfs_multihost_import_intervals * zfs_multihost_interval 139 * 140 * In this case local tunables are used. By default this product = 10s, long 141 * enough for a pool with any activity at all to write at least one 142 * uberblock. No guarantee can be provided. 143 * 144 * Additionally, the duration is then extended by a random 25% to attempt to to 145 * detect simultaneous imports. For example, if both partner hosts are rebooted 146 * at the same time and automatically attempt to import the pool. 147 */ 148 149 /* 150 * Used to control the frequency of mmp writes which are performed when the 151 * 'multihost' pool property is on. This is one factor used to determine the 152 * length of the activity check during import. 153 * 154 * On average an mmp write will be issued for each leaf vdev every 155 * zfs_multihost_interval milliseconds. In practice, the observed period can 156 * vary with the I/O load and this observed value is the ub_mmp_delay which is 157 * stored in the uberblock. The minimum allowed value is 100 ms. 158 */ 159 ulong_t zfs_multihost_interval = MMP_DEFAULT_INTERVAL; 160 161 /* 162 * Used to control the duration of the activity test on import. Smaller values 163 * of zfs_multihost_import_intervals will reduce the import time but increase 164 * the risk of failing to detect an active pool. The total activity check time 165 * is never allowed to drop below one second. A value of 0 is ignored and 166 * treated as if it was set to 1. 167 */ 168 uint_t zfs_multihost_import_intervals = MMP_DEFAULT_IMPORT_INTERVALS; 169 170 /* 171 * Controls the behavior of the pool when mmp write failures or delays are 172 * detected. 173 * 174 * When zfs_multihost_fail_intervals = 0, mmp write failures or delays are 175 * ignored. The failures will still be reported to the ZED which depending on 176 * its configuration may take action such as suspending the pool or taking a 177 * device offline. 178 * 179 * When zfs_multihost_fail_intervals > 0, the pool will be suspended if 180 * zfs_multihost_fail_intervals * zfs_multihost_interval milliseconds pass 181 * without a successful mmp write. This guarantees the activity test will see 182 * mmp writes if the pool is imported. A value of 1 is ignored and treated as 183 * if it was set to 2, because a single leaf vdev pool will issue a write once 184 * per multihost_interval and thus any variation in latency would cause the 185 * pool to be suspended. 186 */ 187 uint_t zfs_multihost_fail_intervals = MMP_DEFAULT_FAIL_INTERVALS; 188 189 char *mmp_tag = "mmp_write_uberblock"; 190 static void mmp_thread(void *arg); 191 192 void 193 mmp_init(spa_t *spa) 194 { 195 mmp_thread_t *mmp = &spa->spa_mmp; 196 197 mutex_init(&mmp->mmp_thread_lock, NULL, MUTEX_DEFAULT, NULL); 198 cv_init(&mmp->mmp_thread_cv, NULL, CV_DEFAULT, NULL); 199 mutex_init(&mmp->mmp_io_lock, NULL, MUTEX_DEFAULT, NULL); 200 mmp->mmp_kstat_id = 1; 201 202 /* 203 * mmp_write_done() calculates mmp_delay based on prior mmp_delay and 204 * the elapsed time since the last write. For the first mmp write, 205 * there is no "last write", so we start with fake non-zero values. 206 */ 207 mmp->mmp_last_write = gethrtime(); 208 mmp->mmp_delay = MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval)); 209 } 210 211 void 212 mmp_fini(spa_t *spa) 213 { 214 mmp_thread_t *mmp = &spa->spa_mmp; 215 216 mutex_destroy(&mmp->mmp_thread_lock); 217 cv_destroy(&mmp->mmp_thread_cv); 218 mutex_destroy(&mmp->mmp_io_lock); 219 } 220 221 static void 222 mmp_thread_enter(mmp_thread_t *mmp, callb_cpr_t *cpr) 223 { 224 CALLB_CPR_INIT(cpr, &mmp->mmp_thread_lock, callb_generic_cpr, FTAG); 225 mutex_enter(&mmp->mmp_thread_lock); 226 } 227 228 static void 229 mmp_thread_exit(mmp_thread_t *mmp, kthread_t **mpp, callb_cpr_t *cpr) 230 { 231 ASSERT(*mpp != NULL); 232 *mpp = NULL; 233 cv_broadcast(&mmp->mmp_thread_cv); 234 CALLB_CPR_EXIT(cpr); /* drops &mmp->mmp_thread_lock */ 235 thread_exit(); 236 } 237 238 void 239 mmp_thread_start(spa_t *spa) 240 { 241 mmp_thread_t *mmp = &spa->spa_mmp; 242 243 if (spa_writeable(spa)) { 244 mutex_enter(&mmp->mmp_thread_lock); 245 if (!mmp->mmp_thread) { 246 mmp->mmp_thread = thread_create(NULL, 0, mmp_thread, 247 spa, 0, &p0, TS_RUN, defclsyspri); 248 zfs_dbgmsg("MMP thread started pool '%s' " 249 "gethrtime %llu", spa_name(spa), gethrtime()); 250 } 251 mutex_exit(&mmp->mmp_thread_lock); 252 } 253 } 254 255 void 256 mmp_thread_stop(spa_t *spa) 257 { 258 mmp_thread_t *mmp = &spa->spa_mmp; 259 260 mutex_enter(&mmp->mmp_thread_lock); 261 mmp->mmp_thread_exiting = 1; 262 cv_broadcast(&mmp->mmp_thread_cv); 263 264 while (mmp->mmp_thread) { 265 cv_wait(&mmp->mmp_thread_cv, &mmp->mmp_thread_lock); 266 } 267 mutex_exit(&mmp->mmp_thread_lock); 268 zfs_dbgmsg("MMP thread stopped pool '%s' gethrtime %llu", 269 spa_name(spa), gethrtime()); 270 271 ASSERT(mmp->mmp_thread == NULL); 272 mmp->mmp_thread_exiting = 0; 273 } 274 275 typedef enum mmp_vdev_state_flag { 276 MMP_FAIL_NOT_WRITABLE = (1 << 0), 277 MMP_FAIL_WRITE_PENDING = (1 << 1), 278 } mmp_vdev_state_flag_t; 279 280 /* 281 * Find a leaf vdev to write an MMP block to. It must not have an outstanding 282 * mmp write (if so a new write will also likely block). If there is no usable 283 * leaf, a nonzero error value is returned. The error value returned is a bit 284 * field. 285 * 286 * MMP_FAIL_WRITE_PENDING One or more leaf vdevs are writeable, but have an 287 * outstanding MMP write. 288 * MMP_FAIL_NOT_WRITABLE One or more leaf vdevs are not writeable. 289 */ 290 291 static int 292 mmp_next_leaf(spa_t *spa) 293 { 294 vdev_t *leaf; 295 vdev_t *starting_leaf; 296 int fail_mask = 0; 297 298 ASSERT(MUTEX_HELD(&spa->spa_mmp.mmp_io_lock)); 299 ASSERT(spa_config_held(spa, SCL_STATE, RW_READER)); 300 ASSERT(list_link_active(&spa->spa_leaf_list.list_head) == B_TRUE); 301 ASSERT(!list_is_empty(&spa->spa_leaf_list)); 302 303 if (spa->spa_mmp.mmp_leaf_last_gen != spa->spa_leaf_list_gen) { 304 spa->spa_mmp.mmp_last_leaf = list_head(&spa->spa_leaf_list); 305 spa->spa_mmp.mmp_leaf_last_gen = spa->spa_leaf_list_gen; 306 } 307 308 leaf = spa->spa_mmp.mmp_last_leaf; 309 if (leaf == NULL) 310 leaf = list_head(&spa->spa_leaf_list); 311 starting_leaf = leaf; 312 313 do { 314 leaf = list_next(&spa->spa_leaf_list, leaf); 315 if (leaf == NULL) 316 leaf = list_head(&spa->spa_leaf_list); 317 318 if (!vdev_writeable(leaf)) { 319 fail_mask |= MMP_FAIL_NOT_WRITABLE; 320 } else if (leaf->vdev_mmp_pending != 0) { 321 fail_mask |= MMP_FAIL_WRITE_PENDING; 322 } else { 323 spa->spa_mmp.mmp_last_leaf = leaf; 324 return (0); 325 } 326 } while (leaf != starting_leaf); 327 328 ASSERT(fail_mask); 329 330 return (fail_mask); 331 } 332 333 /* 334 * MMP writes are issued on a fixed schedule, but may complete at variable, 335 * much longer, intervals. The mmp_delay captures long periods between 336 * successful writes for any reason, including disk latency, scheduling delays, 337 * etc. 338 * 339 * The mmp_delay is usually calculated as a decaying average, but if the latest 340 * delay is higher we do not average it, so that we do not hide sudden spikes 341 * which the importing host must wait for. 342 * 343 * If writes are occurring frequently, such as due to a high rate of txg syncs, 344 * the mmp_delay could become very small. Since those short delays depend on 345 * activity we cannot count on, we never allow mmp_delay to get lower than rate 346 * expected if only mmp_thread writes occur. 347 * 348 * If an mmp write was skipped or fails, and we have already waited longer than 349 * mmp_delay, we need to update it so the next write reflects the longer delay. 350 * 351 * Do not set mmp_delay if the multihost property is not on, so as not to 352 * trigger an activity check on import. 353 */ 354 static void 355 mmp_delay_update(spa_t *spa, boolean_t write_completed) 356 { 357 mmp_thread_t *mts = &spa->spa_mmp; 358 hrtime_t delay = gethrtime() - mts->mmp_last_write; 359 360 ASSERT(MUTEX_HELD(&mts->mmp_io_lock)); 361 362 if (spa_multihost(spa) == B_FALSE) { 363 mts->mmp_delay = 0; 364 return; 365 } 366 367 if (delay > mts->mmp_delay) 368 mts->mmp_delay = delay; 369 370 if (write_completed == B_FALSE) 371 return; 372 373 mts->mmp_last_write = gethrtime(); 374 375 /* 376 * strictly less than, in case delay was changed above. 377 */ 378 if (delay < mts->mmp_delay) { 379 hrtime_t min_delay = 380 MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval)) / 381 MAX(1, vdev_count_leaves(spa)); 382 mts->mmp_delay = MAX(((delay + mts->mmp_delay * 127) / 128), 383 min_delay); 384 } 385 } 386 387 static void 388 mmp_write_done(zio_t *zio) 389 { 390 spa_t *spa = zio->io_spa; 391 vdev_t *vd = zio->io_vd; 392 mmp_thread_t *mts = zio->io_private; 393 394 mutex_enter(&mts->mmp_io_lock); 395 uint64_t mmp_kstat_id = vd->vdev_mmp_kstat_id; 396 hrtime_t mmp_write_duration = gethrtime() - vd->vdev_mmp_pending; 397 398 mmp_delay_update(spa, (zio->io_error == 0)); 399 400 vd->vdev_mmp_pending = 0; 401 vd->vdev_mmp_kstat_id = 0; 402 403 mutex_exit(&mts->mmp_io_lock); 404 spa_config_exit(spa, SCL_STATE, mmp_tag); 405 406 spa_mmp_history_set(spa, mmp_kstat_id, zio->io_error, 407 mmp_write_duration); 408 409 abd_free(zio->io_abd); 410 } 411 412 /* 413 * When the uberblock on-disk is updated by a spa_sync, 414 * creating a new "best" uberblock, update the one stored 415 * in the mmp thread state, used for mmp writes. 416 */ 417 void 418 mmp_update_uberblock(spa_t *spa, uberblock_t *ub) 419 { 420 mmp_thread_t *mmp = &spa->spa_mmp; 421 422 mutex_enter(&mmp->mmp_io_lock); 423 mmp->mmp_ub = *ub; 424 mmp->mmp_seq = 1; 425 mmp->mmp_ub.ub_timestamp = gethrestime_sec(); 426 mmp_delay_update(spa, B_TRUE); 427 mutex_exit(&mmp->mmp_io_lock); 428 } 429 430 /* 431 * Choose a random vdev, label, and MMP block, and write over it 432 * with a copy of the last-synced uberblock, whose timestamp 433 * has been updated to reflect that the pool is in use. 434 */ 435 static void 436 mmp_write_uberblock(spa_t *spa) 437 { 438 int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL; 439 mmp_thread_t *mmp = &spa->spa_mmp; 440 uberblock_t *ub; 441 vdev_t *vd = NULL; 442 int label, error; 443 uint64_t offset; 444 445 hrtime_t lock_acquire_time = gethrtime(); 446 spa_config_enter(spa, SCL_STATE, mmp_tag, RW_READER); 447 lock_acquire_time = gethrtime() - lock_acquire_time; 448 if (lock_acquire_time > (MSEC2NSEC(MMP_MIN_INTERVAL) / 10)) 449 zfs_dbgmsg("MMP SCL_STATE acquisition pool '%s' took %llu ns " 450 "gethrtime %llu", spa_name(spa), lock_acquire_time, 451 gethrtime()); 452 453 mutex_enter(&mmp->mmp_io_lock); 454 455 error = mmp_next_leaf(spa); 456 457 /* 458 * spa_mmp_history has two types of entries: 459 * Issued MMP write: records time issued, error status, etc. 460 * Skipped MMP write: an MMP write could not be issued because no 461 * suitable leaf vdev was available. See comment above struct 462 * spa_mmp_history for details. 463 */ 464 465 if (error) { 466 mmp_delay_update(spa, B_FALSE); 467 if (mmp->mmp_skip_error == error) { 468 spa_mmp_history_set_skip(spa, mmp->mmp_kstat_id - 1); 469 } else { 470 mmp->mmp_skip_error = error; 471 spa_mmp_history_add(spa, mmp->mmp_ub.ub_txg, 472 gethrestime_sec(), mmp->mmp_delay, NULL, 0, 473 mmp->mmp_kstat_id++, error); 474 zfs_dbgmsg("MMP error choosing leaf pool '%s' " 475 "gethrtime %llu fail_mask %#x", spa_name(spa), 476 gethrtime(), error); 477 } 478 mutex_exit(&mmp->mmp_io_lock); 479 spa_config_exit(spa, SCL_STATE, mmp_tag); 480 return; 481 } 482 483 vd = spa->spa_mmp.mmp_last_leaf; 484 if (mmp->mmp_skip_error != 0) { 485 mmp->mmp_skip_error = 0; 486 zfs_dbgmsg("MMP write after skipping due to unavailable " 487 "leaves, pool '%s' gethrtime %llu leaf %#llu", 488 spa_name(spa), gethrtime(), vd->vdev_guid); 489 } 490 491 if (mmp->mmp_zio_root == NULL) 492 mmp->mmp_zio_root = zio_root(spa, NULL, NULL, 493 flags | ZIO_FLAG_GODFATHER); 494 495 if (mmp->mmp_ub.ub_timestamp != gethrestime_sec()) { 496 /* 497 * Want to reset mmp_seq when timestamp advances because after 498 * an mmp_seq wrap new values will not be chosen by 499 * uberblock_compare() as the "best". 500 */ 501 mmp->mmp_ub.ub_timestamp = gethrestime_sec(); 502 mmp->mmp_seq = 1; 503 } 504 505 ub = &mmp->mmp_ub; 506 ub->ub_mmp_magic = MMP_MAGIC; 507 ub->ub_mmp_delay = mmp->mmp_delay; 508 ub->ub_mmp_config = MMP_SEQ_SET(mmp->mmp_seq) | 509 MMP_INTERVAL_SET(MMP_INTERVAL_OK(zfs_multihost_interval)) | 510 MMP_FAIL_INT_SET(MMP_FAIL_INTVS_OK( 511 zfs_multihost_fail_intervals)); 512 vd->vdev_mmp_pending = gethrtime(); 513 vd->vdev_mmp_kstat_id = mmp->mmp_kstat_id; 514 515 zio_t *zio = zio_null(mmp->mmp_zio_root, spa, NULL, NULL, NULL, flags); 516 abd_t *ub_abd = abd_alloc_for_io(VDEV_UBERBLOCK_SIZE(vd), B_TRUE); 517 abd_zero(ub_abd, VDEV_UBERBLOCK_SIZE(vd)); 518 abd_copy_from_buf(ub_abd, ub, sizeof (uberblock_t)); 519 520 mmp->mmp_seq++; 521 mmp->mmp_kstat_id++; 522 mutex_exit(&mmp->mmp_io_lock); 523 524 offset = VDEV_UBERBLOCK_OFFSET(vd, VDEV_UBERBLOCK_COUNT(vd) - 525 MMP_BLOCKS_PER_LABEL + spa_get_random(MMP_BLOCKS_PER_LABEL)); 526 527 label = spa_get_random(VDEV_LABELS); 528 vdev_label_write(zio, vd, label, ub_abd, offset, 529 VDEV_UBERBLOCK_SIZE(vd), mmp_write_done, mmp, 530 flags | ZIO_FLAG_DONT_PROPAGATE); 531 532 (void) spa_mmp_history_add(spa, ub->ub_txg, ub->ub_timestamp, 533 ub->ub_mmp_delay, vd, label, vd->vdev_mmp_kstat_id, 0); 534 535 zio_nowait(zio); 536 } 537 538 static void 539 mmp_thread(void *arg) 540 { 541 spa_t *spa = (spa_t *)arg; 542 mmp_thread_t *mmp = &spa->spa_mmp; 543 boolean_t suspended = spa_suspended(spa); 544 boolean_t multihost = spa_multihost(spa); 545 uint64_t mmp_interval = MSEC2NSEC(MMP_INTERVAL_OK( 546 zfs_multihost_interval)); 547 uint32_t mmp_fail_intervals = MMP_FAIL_INTVS_OK( 548 zfs_multihost_fail_intervals); 549 hrtime_t mmp_fail_ns = mmp_fail_intervals * mmp_interval; 550 boolean_t last_spa_suspended = suspended; 551 boolean_t last_spa_multihost = multihost; 552 uint64_t last_mmp_interval = mmp_interval; 553 uint32_t last_mmp_fail_intervals = mmp_fail_intervals; 554 hrtime_t last_mmp_fail_ns = mmp_fail_ns; 555 callb_cpr_t cpr; 556 int skip_wait = 0; 557 558 mmp_thread_enter(mmp, &cpr); 559 560 while (!mmp->mmp_thread_exiting) { 561 hrtime_t next_time = gethrtime() + 562 MSEC2NSEC(MMP_DEFAULT_INTERVAL); 563 int leaves = MAX(vdev_count_leaves(spa), 1); 564 565 /* Detect changes in tunables or state */ 566 567 last_spa_suspended = suspended; 568 last_spa_multihost = multihost; 569 suspended = spa_suspended(spa); 570 multihost = spa_multihost(spa); 571 572 last_mmp_interval = mmp_interval; 573 last_mmp_fail_intervals = mmp_fail_intervals; 574 last_mmp_fail_ns = mmp_fail_ns; 575 mmp_interval = MSEC2NSEC(MMP_INTERVAL_OK( 576 zfs_multihost_interval)); 577 mmp_fail_intervals = MMP_FAIL_INTVS_OK( 578 zfs_multihost_fail_intervals); 579 580 /* Smooth so pool is not suspended when reducing tunables */ 581 if (mmp_fail_intervals * mmp_interval < mmp_fail_ns) { 582 mmp_fail_ns = (mmp_fail_ns * 31 + 583 mmp_fail_intervals * mmp_interval) / 32; 584 } else { 585 mmp_fail_ns = mmp_fail_intervals * 586 mmp_interval; 587 } 588 589 if (mmp_interval != last_mmp_interval || 590 mmp_fail_intervals != last_mmp_fail_intervals) { 591 /* 592 * We want other hosts to see new tunables as quickly as 593 * possible. Write out at higher frequency than usual. 594 */ 595 skip_wait += leaves; 596 } 597 598 if (multihost) 599 next_time = gethrtime() + mmp_interval / leaves; 600 601 if (mmp_fail_ns != last_mmp_fail_ns) { 602 zfs_dbgmsg("MMP interval change pool '%s' " 603 "gethrtime %llu last_mmp_interval %llu " 604 "mmp_interval %llu last_mmp_fail_intervals %u " 605 "mmp_fail_intervals %u mmp_fail_ns %llu " 606 "skip_wait %d leaves %d next_time %llu", 607 spa_name(spa), gethrtime(), last_mmp_interval, 608 mmp_interval, last_mmp_fail_intervals, 609 mmp_fail_intervals, mmp_fail_ns, skip_wait, leaves, 610 next_time); 611 } 612 613 /* 614 * MMP off => on, or suspended => !suspended: 615 * No writes occurred recently. Update mmp_last_write to give 616 * us some time to try. 617 */ 618 if ((!last_spa_multihost && multihost) || 619 (last_spa_suspended && !suspended)) { 620 zfs_dbgmsg("MMP state change pool '%s': gethrtime %llu " 621 "last_spa_multihost %u multihost %u " 622 "last_spa_suspended %u suspended %u", 623 spa_name(spa), last_spa_multihost, multihost, 624 last_spa_suspended, suspended); 625 mutex_enter(&mmp->mmp_io_lock); 626 mmp->mmp_last_write = gethrtime(); 627 mmp->mmp_delay = mmp_interval; 628 mutex_exit(&mmp->mmp_io_lock); 629 } 630 631 /* 632 * MMP on => off: 633 * mmp_delay == 0 tells importing node to skip activity check. 634 */ 635 if (last_spa_multihost && !multihost) { 636 mutex_enter(&mmp->mmp_io_lock); 637 mmp->mmp_delay = 0; 638 mutex_exit(&mmp->mmp_io_lock); 639 } 640 641 /* 642 * Suspend the pool if no MMP write has succeeded in over 643 * mmp_interval * mmp_fail_intervals nanoseconds. 644 */ 645 if (multihost && !suspended && mmp_fail_intervals && 646 (gethrtime() - mmp->mmp_last_write) > mmp_fail_ns) { 647 zfs_dbgmsg("MMP suspending pool '%s': gethrtime %llu " 648 "mmp_last_write %llu mmp_interval %llu " 649 "mmp_fail_intervals %llu mmp_fail_ns %llu", 650 spa_name(spa), (u_longlong_t)gethrtime(), 651 (u_longlong_t)mmp->mmp_last_write, 652 (u_longlong_t)mmp_interval, 653 (u_longlong_t)mmp_fail_intervals, 654 (u_longlong_t)mmp_fail_ns); 655 cmn_err(CE_WARN, "MMP writes to pool '%s' have not " 656 "succeeded in over %llu ms; suspending pool. " 657 "Hrtime %llu", 658 spa_name(spa), 659 NSEC2MSEC(gethrtime() - mmp->mmp_last_write), 660 gethrtime()); 661 zio_suspend(spa, NULL, ZIO_SUSPEND_MMP); 662 } 663 664 if (multihost && !suspended) 665 mmp_write_uberblock(spa); 666 667 if (skip_wait > 0) { 668 next_time = gethrtime() + MSEC2NSEC(MMP_MIN_INTERVAL) / 669 leaves; 670 skip_wait--; 671 } 672 673 CALLB_CPR_SAFE_BEGIN(&cpr); 674 (void) cv_timedwait_sig_hires(&mmp->mmp_thread_cv, 675 &mmp->mmp_thread_lock, next_time, USEC2NSEC(100), 676 CALLOUT_FLAG_ABSOLUTE); 677 CALLB_CPR_SAFE_END(&cpr, &mmp->mmp_thread_lock); 678 } 679 680 /* Outstanding writes are allowed to complete. */ 681 zio_wait(mmp->mmp_zio_root); 682 683 mmp->mmp_zio_root = NULL; 684 mmp_thread_exit(mmp, &mmp->mmp_thread, &cpr); 685 } 686 687 /* 688 * Signal the MMP thread to wake it, when it is sleeping on 689 * its cv. Used when some module parameter has changed and 690 * we want the thread to know about it. 691 * Only signal if the pool is active and mmp thread is 692 * running, otherwise there is no thread to wake. 693 */ 694 static void 695 mmp_signal_thread(spa_t *spa) 696 { 697 mmp_thread_t *mmp = &spa->spa_mmp; 698 699 mutex_enter(&mmp->mmp_thread_lock); 700 if (mmp->mmp_thread) 701 cv_broadcast(&mmp->mmp_thread_cv); 702 mutex_exit(&mmp->mmp_thread_lock); 703 } 704 705 void 706 mmp_signal_all_threads(void) 707 { 708 spa_t *spa = NULL; 709 710 mutex_enter(&spa_namespace_lock); 711 while ((spa = spa_next(spa))) { 712 if (spa->spa_state == POOL_STATE_ACTIVE) 713 mmp_signal_thread(spa); 714 } 715 mutex_exit(&spa_namespace_lock); 716 } 717 718 /* BEGIN CSTYLED */ 719 ZFS_MODULE_PARAM_CALL(zfs_multihost, zfs_multihost_, interval, 720 param_set_multihost_interval, param_get_ulong, ZMOD_RW, 721 "Milliseconds between mmp writes to each leaf"); 722 /* END CSTYLED */ 723 724 ZFS_MODULE_PARAM(zfs_multihost, zfs_multihost_, fail_intervals, UINT, ZMOD_RW, 725 "Max allowed period without a successful mmp write"); 726 727 ZFS_MODULE_PARAM(zfs_multihost, zfs_multihost_, import_intervals, UINT, ZMOD_RW, 728 "Number of zfs_multihost_interval periods to wait for activity"); 729