1 // SPDX-License-Identifier: CDDL-1.0 2 /* 3 * CDDL HEADER START 4 * 5 * The contents of this file are subject to the terms of the 6 * Common Development and Distribution License (the "License"). 7 * You may not use this file except in compliance with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or https://opensource.org/licenses/CDDL-1.0. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2011, 2021 by Delphix. All rights reserved. 25 * Copyright 2016 Gary Mills 26 * Copyright (c) 2017, 2019, Datto Inc. All rights reserved. 27 * Copyright (c) 2015, Nexenta Systems, Inc. All rights reserved. 28 * Copyright 2019 Joyent, Inc. 29 */ 30 31 #include <sys/dsl_scan.h> 32 #include <sys/dsl_pool.h> 33 #include <sys/dsl_dataset.h> 34 #include <sys/dsl_prop.h> 35 #include <sys/dsl_dir.h> 36 #include <sys/dsl_synctask.h> 37 #include <sys/dnode.h> 38 #include <sys/dmu_tx.h> 39 #include <sys/dmu_objset.h> 40 #include <sys/arc.h> 41 #include <sys/arc_impl.h> 42 #include <sys/zap.h> 43 #include <sys/zio.h> 44 #include <sys/zfs_context.h> 45 #include <sys/fs/zfs.h> 46 #include <sys/zfs_znode.h> 47 #include <sys/spa_impl.h> 48 #include <sys/vdev_impl.h> 49 #include <sys/zil_impl.h> 50 #include <sys/zio_checksum.h> 51 #include <sys/brt.h> 52 #include <sys/ddt.h> 53 #include <sys/sa.h> 54 #include <sys/sa_impl.h> 55 #include <sys/zfeature.h> 56 #include <sys/abd.h> 57 #include <sys/range_tree.h> 58 #include <sys/dbuf.h> 59 #ifdef _KERNEL 60 #include <sys/zfs_vfsops.h> 61 #endif 62 63 /* 64 * Grand theory statement on scan queue sorting 65 * 66 * Scanning is implemented by recursively traversing all indirection levels 67 * in an object and reading all blocks referenced from said objects. This 68 * results in us approximately traversing the object from lowest logical 69 * offset to the highest. For best performance, we would want the logical 70 * blocks to be physically contiguous. However, this is frequently not the 71 * case with pools given the allocation patterns of copy-on-write filesystems. 72 * So instead, we put the I/Os into a reordering queue and issue them in a 73 * way that will most benefit physical disks (LBA-order). 74 * 75 * Queue management: 76 * 77 * Ideally, we would want to scan all metadata and queue up all block I/O 78 * prior to starting to issue it, because that allows us to do an optimal 79 * sorting job. This can however consume large amounts of memory. Therefore 80 * we continuously monitor the size of the queues and constrain them to 5% 81 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this 82 * limit, we clear out a few of the largest extents at the head of the queues 83 * to make room for more scanning. Hopefully, these extents will be fairly 84 * large and contiguous, allowing us to approach sequential I/O throughput 85 * even without a fully sorted tree. 86 * 87 * Metadata scanning takes place in dsl_scan_visit(), which is called from 88 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all 89 * metadata on the pool, or we need to make room in memory because our 90 * queues are too large, dsl_scan_visit() is postponed and 91 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies 92 * that metadata scanning and queued I/O issuing are mutually exclusive. This 93 * allows us to provide maximum sequential I/O throughput for the majority of 94 * I/O's issued since sequential I/O performance is significantly negatively 95 * impacted if it is interleaved with random I/O. 96 * 97 * Implementation Notes 98 * 99 * One side effect of the queued scanning algorithm is that the scanning code 100 * needs to be notified whenever a block is freed. This is needed to allow 101 * the scanning code to remove these I/Os from the issuing queue. Additionally, 102 * we do not attempt to queue gang blocks to be issued sequentially since this 103 * is very hard to do and would have an extremely limited performance benefit. 104 * Instead, we simply issue gang I/Os as soon as we find them using the legacy 105 * algorithm. 106 * 107 * Backwards compatibility 108 * 109 * This new algorithm is backwards compatible with the legacy on-disk data 110 * structures (and therefore does not require a new feature flag). 111 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan 112 * will stop scanning metadata (in logical order) and wait for all outstanding 113 * sorted I/O to complete. Once this is done, we write out a checkpoint 114 * bookmark, indicating that we have scanned everything logically before it. 115 * If the pool is imported on a machine without the new sorting algorithm, 116 * the scan simply resumes from the last checkpoint using the legacy algorithm. 117 */ 118 119 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, 120 const zbookmark_phys_t *); 121 122 static scan_cb_t dsl_scan_scrub_cb; 123 124 static int scan_ds_queue_compare(const void *a, const void *b); 125 static int scan_prefetch_queue_compare(const void *a, const void *b); 126 static void scan_ds_queue_clear(dsl_scan_t *scn); 127 static void scan_ds_prefetch_queue_clear(dsl_scan_t *scn); 128 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, 129 uint64_t *txg); 130 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg); 131 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj); 132 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx); 133 static uint64_t dsl_scan_count_data_disks(spa_t *spa); 134 static void read_by_block_level(dsl_scan_t *scn, zbookmark_phys_t zb); 135 136 extern uint_t zfs_vdev_async_write_active_min_dirty_percent; 137 static int zfs_scan_blkstats = 0; 138 139 /* 140 * 'zpool status' uses bytes processed per pass to report throughput and 141 * estimate time remaining. We define a pass to start when the scanning 142 * phase completes for a sequential resilver. Optionally, this value 143 * may be used to reset the pass statistics every N txgs to provide an 144 * estimated completion time based on currently observed performance. 145 */ 146 static uint_t zfs_scan_report_txgs = 0; 147 148 /* 149 * By default zfs will check to ensure it is not over the hard memory 150 * limit before each txg. If finer-grained control of this is needed 151 * this value can be set to 1 to enable checking before scanning each 152 * block. 153 */ 154 static int zfs_scan_strict_mem_lim = B_FALSE; 155 156 /* 157 * Maximum number of parallelly executed bytes per leaf vdev. We attempt 158 * to strike a balance here between keeping the vdev queues full of I/Os 159 * at all times and not overflowing the queues to cause long latency, 160 * which would cause long txg sync times. No matter what, we will not 161 * overload the drives with I/O, since that is protected by 162 * zfs_vdev_scrub_max_active. 163 */ 164 static uint64_t zfs_scan_vdev_limit = 16 << 20; 165 166 static uint_t zfs_scan_issue_strategy = 0; 167 168 /* don't queue & sort zios, go direct */ 169 static int zfs_scan_legacy = B_FALSE; 170 static uint64_t zfs_scan_max_ext_gap = 2 << 20; /* in bytes */ 171 172 /* 173 * fill_weight is non-tunable at runtime, so we copy it at module init from 174 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would 175 * break queue sorting. 176 */ 177 static uint_t zfs_scan_fill_weight = 3; 178 static uint64_t fill_weight; 179 180 /* See dsl_scan_should_clear() for details on the memory limit tunables */ 181 static const uint64_t zfs_scan_mem_lim_min = 16 << 20; /* bytes */ 182 static const uint64_t zfs_scan_mem_lim_soft_max = 128 << 20; /* bytes */ 183 184 185 /* fraction of physmem */ 186 static uint_t zfs_scan_mem_lim_fact = 20; 187 188 /* fraction of mem lim above */ 189 static uint_t zfs_scan_mem_lim_soft_fact = 20; 190 191 /* minimum milliseconds to scrub per txg */ 192 static uint_t zfs_scrub_min_time_ms = 1000; 193 194 /* minimum milliseconds to obsolete per txg */ 195 static uint_t zfs_obsolete_min_time_ms = 500; 196 197 /* minimum milliseconds to free per txg */ 198 static uint_t zfs_free_min_time_ms = 1000; 199 200 /* minimum milliseconds to resilver per txg */ 201 static uint_t zfs_resilver_min_time_ms = 3000; 202 203 static uint_t zfs_scan_checkpoint_intval = 7200; /* in seconds */ 204 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */ 205 static int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */ 206 static int zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */ 207 static const ddt_class_t zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE; 208 /* max number of blocks to free in a single TXG */ 209 static uint64_t zfs_async_block_max_blocks = UINT64_MAX; 210 /* max number of dedup blocks to free in a single TXG */ 211 static uint64_t zfs_max_async_dedup_frees = 100000; 212 213 /* set to disable resilver deferring */ 214 static int zfs_resilver_disable_defer = B_FALSE; 215 216 /* Don't defer a resilver if the one in progress only got this far: */ 217 static uint_t zfs_resilver_defer_percent = 10; 218 219 /* 220 * We wait a few txgs after importing a pool to begin scanning so that 221 * the import / mounting code isn't held up by scrub / resilver IO. 222 * Unfortunately, it is a bit difficult to determine exactly how long 223 * this will take since userspace will trigger fs mounts asynchronously 224 * and the kernel will create zvol minors asynchronously. As a result, 225 * the value provided here is a bit arbitrary, but represents a 226 * reasonable estimate of how many txgs it will take to finish fully 227 * importing a pool 228 */ 229 #define SCAN_IMPORT_WAIT_TXGS 5 230 231 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \ 232 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \ 233 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER) 234 235 #define DSL_SCAN_IS_SCRUB(scn) \ 236 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB) 237 238 #define DSL_SCAN_IS_RESILVER(scn) \ 239 ((scn)->scn_phys.scn_func == POOL_SCAN_RESILVER) 240 241 /* 242 * Enable/disable the processing of the free_bpobj object. 243 */ 244 static int zfs_free_bpobj_enabled = 1; 245 246 /* Error blocks to be scrubbed in one txg. */ 247 static uint_t zfs_scrub_error_blocks_per_txg = 1 << 12; 248 249 /* the order has to match pool_scan_type */ 250 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = { 251 NULL, 252 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */ 253 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */ 254 }; 255 256 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */ 257 typedef struct { 258 uint64_t sds_dsobj; 259 uint64_t sds_txg; 260 avl_node_t sds_node; 261 } scan_ds_t; 262 263 /* 264 * This controls what conditions are placed on dsl_scan_sync_state(): 265 * SYNC_OPTIONAL) write out scn_phys iff scn_queues_pending == 0 266 * SYNC_MANDATORY) write out scn_phys always. scn_queues_pending must be 0. 267 * SYNC_CACHED) if scn_queues_pending == 0, write out scn_phys. Otherwise 268 * write out the scn_phys_cached version. 269 * See dsl_scan_sync_state for details. 270 */ 271 typedef enum { 272 SYNC_OPTIONAL, 273 SYNC_MANDATORY, 274 SYNC_CACHED 275 } state_sync_type_t; 276 277 /* 278 * This struct represents the minimum information needed to reconstruct a 279 * zio for sequential scanning. This is useful because many of these will 280 * accumulate in the sequential IO queues before being issued, so saving 281 * memory matters here. 282 */ 283 typedef struct scan_io { 284 /* fields from blkptr_t */ 285 uint64_t sio_blk_prop; 286 uint64_t sio_phys_birth; 287 uint64_t sio_birth; 288 zio_cksum_t sio_cksum; 289 uint32_t sio_nr_dvas; 290 291 /* fields from zio_t */ 292 uint32_t sio_flags; 293 zbookmark_phys_t sio_zb; 294 295 /* members for queue sorting */ 296 union { 297 avl_node_t sio_addr_node; /* link into issuing queue */ 298 list_node_t sio_list_node; /* link for issuing to disk */ 299 } sio_nodes; 300 301 /* 302 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp, 303 * depending on how many were in the original bp. Only the 304 * first DVA is really used for sorting and issuing purposes. 305 * The other DVAs (if provided) simply exist so that the zio 306 * layer can find additional copies to repair from in the 307 * event of an error. This array must go at the end of the 308 * struct to allow this for the variable number of elements. 309 */ 310 dva_t sio_dva[]; 311 } scan_io_t; 312 313 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x) 314 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x) 315 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0]) 316 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0]) 317 #define SIO_GET_END_OFFSET(sio) \ 318 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio)) 319 #define SIO_GET_MUSED(sio) \ 320 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t))) 321 322 struct dsl_scan_io_queue { 323 dsl_scan_t *q_scn; /* associated dsl_scan_t */ 324 vdev_t *q_vd; /* top-level vdev that this queue represents */ 325 zio_t *q_zio; /* scn_zio_root child for waiting on IO */ 326 327 /* trees used for sorting I/Os and extents of I/Os */ 328 zfs_range_tree_t *q_exts_by_addr; 329 zfs_btree_t q_exts_by_size; 330 avl_tree_t q_sios_by_addr; 331 uint64_t q_sio_memused; 332 uint64_t q_last_ext_addr; 333 334 /* members for zio rate limiting */ 335 uint64_t q_maxinflight_bytes; 336 uint64_t q_inflight_bytes; 337 kcondvar_t q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */ 338 339 /* per txg statistics */ 340 uint64_t q_total_seg_size_this_txg; 341 uint64_t q_segs_this_txg; 342 uint64_t q_total_zio_size_this_txg; 343 uint64_t q_zios_this_txg; 344 }; 345 346 /* private data for dsl_scan_prefetch_cb() */ 347 typedef struct scan_prefetch_ctx { 348 zfs_refcount_t spc_refcnt; /* refcount for memory management */ 349 dsl_scan_t *spc_scn; /* dsl_scan_t for the pool */ 350 boolean_t spc_root; /* is this prefetch for an objset? */ 351 uint8_t spc_indblkshift; /* dn_indblkshift of current dnode */ 352 uint16_t spc_datablkszsec; /* dn_idatablkszsec of current dnode */ 353 } scan_prefetch_ctx_t; 354 355 /* private data for dsl_scan_prefetch() */ 356 typedef struct scan_prefetch_issue_ctx { 357 avl_node_t spic_avl_node; /* link into scn->scn_prefetch_queue */ 358 scan_prefetch_ctx_t *spic_spc; /* spc for the callback */ 359 blkptr_t spic_bp; /* bp to prefetch */ 360 zbookmark_phys_t spic_zb; /* bookmark to prefetch */ 361 } scan_prefetch_issue_ctx_t; 362 363 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags, 364 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue); 365 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, 366 scan_io_t *sio); 367 368 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd); 369 static void scan_io_queues_destroy(dsl_scan_t *scn); 370 371 static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP]; 372 373 /* sio->sio_nr_dvas must be set so we know which cache to free from */ 374 static void 375 sio_free(scan_io_t *sio) 376 { 377 ASSERT3U(sio->sio_nr_dvas, >, 0); 378 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP); 379 380 kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio); 381 } 382 383 /* It is up to the caller to set sio->sio_nr_dvas for freeing */ 384 static scan_io_t * 385 sio_alloc(unsigned short nr_dvas) 386 { 387 ASSERT3U(nr_dvas, >, 0); 388 ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP); 389 390 return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP)); 391 } 392 393 void 394 scan_init(void) 395 { 396 /* 397 * This is used in ext_size_compare() to weight segments 398 * based on how sparse they are. This cannot be changed 399 * mid-scan and the tree comparison functions don't currently 400 * have a mechanism for passing additional context to the 401 * compare functions. Thus we store this value globally and 402 * we only allow it to be set at module initialization time 403 */ 404 fill_weight = zfs_scan_fill_weight; 405 406 for (int i = 0; i < SPA_DVAS_PER_BP; i++) { 407 char name[36]; 408 409 (void) snprintf(name, sizeof (name), "sio_cache_%d", i); 410 sio_cache[i] = kmem_cache_create(name, 411 (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))), 412 0, NULL, NULL, NULL, NULL, NULL, 0); 413 } 414 } 415 416 void 417 scan_fini(void) 418 { 419 for (int i = 0; i < SPA_DVAS_PER_BP; i++) { 420 kmem_cache_destroy(sio_cache[i]); 421 } 422 } 423 424 static inline boolean_t 425 dsl_scan_is_running(const dsl_scan_t *scn) 426 { 427 return (scn->scn_phys.scn_state == DSS_SCANNING); 428 } 429 430 boolean_t 431 dsl_scan_resilvering(dsl_pool_t *dp) 432 { 433 return (dsl_scan_is_running(dp->dp_scan) && 434 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER); 435 } 436 437 static inline void 438 sio2bp(const scan_io_t *sio, blkptr_t *bp) 439 { 440 memset(bp, 0, sizeof (*bp)); 441 bp->blk_prop = sio->sio_blk_prop; 442 BP_SET_PHYSICAL_BIRTH(bp, sio->sio_phys_birth); 443 BP_SET_LOGICAL_BIRTH(bp, sio->sio_birth); 444 bp->blk_fill = 1; /* we always only work with data pointers */ 445 bp->blk_cksum = sio->sio_cksum; 446 447 ASSERT3U(sio->sio_nr_dvas, >, 0); 448 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP); 449 450 memcpy(bp->blk_dva, sio->sio_dva, sio->sio_nr_dvas * sizeof (dva_t)); 451 } 452 453 static inline void 454 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i) 455 { 456 sio->sio_blk_prop = bp->blk_prop; 457 sio->sio_phys_birth = BP_GET_PHYSICAL_BIRTH(bp); 458 sio->sio_birth = BP_GET_LOGICAL_BIRTH(bp); 459 sio->sio_cksum = bp->blk_cksum; 460 sio->sio_nr_dvas = BP_GET_NDVAS(bp); 461 462 /* 463 * Copy the DVAs to the sio. We need all copies of the block so 464 * that the self healing code can use the alternate copies if the 465 * first is corrupted. We want the DVA at index dva_i to be first 466 * in the sio since this is the primary one that we want to issue. 467 */ 468 for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) { 469 sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas]; 470 } 471 } 472 473 int 474 dsl_scan_init(dsl_pool_t *dp, uint64_t txg) 475 { 476 int err; 477 dsl_scan_t *scn; 478 spa_t *spa = dp->dp_spa; 479 uint64_t f; 480 481 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP); 482 scn->scn_dp = dp; 483 484 /* 485 * It's possible that we're resuming a scan after a reboot so 486 * make sure that the scan_async_destroying flag is initialized 487 * appropriately. 488 */ 489 ASSERT(!scn->scn_async_destroying); 490 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa, 491 SPA_FEATURE_ASYNC_DESTROY); 492 493 /* 494 * Calculate the max number of in-flight bytes for pool-wide 495 * scanning operations (minimum 1MB, maximum 1/4 of arc_c_max). 496 * Limits for the issuing phase are done per top-level vdev and 497 * are handled separately. 498 */ 499 scn->scn_maxinflight_bytes = MIN(arc_c_max / 4, MAX(1ULL << 20, 500 zfs_scan_vdev_limit * dsl_scan_count_data_disks(spa))); 501 502 avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t), 503 offsetof(scan_ds_t, sds_node)); 504 mutex_init(&scn->scn_queue_lock, NULL, MUTEX_DEFAULT, NULL); 505 avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare, 506 sizeof (scan_prefetch_issue_ctx_t), 507 offsetof(scan_prefetch_issue_ctx_t, spic_avl_node)); 508 509 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 510 "scrub_func", sizeof (uint64_t), 1, &f); 511 if (err == 0) { 512 /* 513 * There was an old-style scrub in progress. Restart a 514 * new-style scrub from the beginning. 515 */ 516 scn->scn_restart_txg = txg; 517 zfs_dbgmsg("old-style scrub was in progress for %s; " 518 "restarting new-style scrub in txg %llu", 519 spa->spa_name, 520 (longlong_t)scn->scn_restart_txg); 521 522 /* 523 * Load the queue obj from the old location so that it 524 * can be freed by dsl_scan_done(). 525 */ 526 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 527 "scrub_queue", sizeof (uint64_t), 1, 528 &scn->scn_phys.scn_queue_obj); 529 } else { 530 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 531 DMU_POOL_ERRORSCRUB, sizeof (uint64_t), 532 ERRORSCRUB_PHYS_NUMINTS, &scn->errorscrub_phys); 533 534 if (err != 0 && err != ENOENT) 535 return (err); 536 537 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 538 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 539 &scn->scn_phys); 540 541 /* 542 * Detect if the pool contains the signature of #2094. If it 543 * does properly update the scn->scn_phys structure and notify 544 * the administrator by setting an errata for the pool. 545 */ 546 if (err == EOVERFLOW) { 547 uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1]; 548 VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24); 549 VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==, 550 (23 * sizeof (uint64_t))); 551 552 err = zap_lookup(dp->dp_meta_objset, 553 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN, 554 sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp); 555 if (err == 0) { 556 uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS]; 557 558 if (overflow & ~DSL_SCAN_FLAGS_MASK || 559 scn->scn_async_destroying) { 560 spa->spa_errata = 561 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY; 562 return (EOVERFLOW); 563 } 564 565 memcpy(&scn->scn_phys, zaptmp, 566 SCAN_PHYS_NUMINTS * sizeof (uint64_t)); 567 scn->scn_phys.scn_flags = overflow; 568 569 /* Required scrub already in progress. */ 570 if (scn->scn_phys.scn_state == DSS_FINISHED || 571 scn->scn_phys.scn_state == DSS_CANCELED) 572 spa->spa_errata = 573 ZPOOL_ERRATA_ZOL_2094_SCRUB; 574 } 575 } 576 577 if (err == ENOENT) 578 return (0); 579 else if (err) 580 return (err); 581 582 /* 583 * We might be restarting after a reboot, so jump the issued 584 * counter to how far we've scanned. We know we're consistent 585 * up to here. 586 */ 587 scn->scn_issued_before_pass = scn->scn_phys.scn_examined - 588 scn->scn_phys.scn_skipped; 589 590 if (dsl_scan_is_running(scn) && 591 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) { 592 /* 593 * A new-type scrub was in progress on an old 594 * pool, and the pool was accessed by old 595 * software. Restart from the beginning, since 596 * the old software may have changed the pool in 597 * the meantime. 598 */ 599 scn->scn_restart_txg = txg; 600 zfs_dbgmsg("new-style scrub for %s was modified " 601 "by old software; restarting in txg %llu", 602 spa->spa_name, 603 (longlong_t)scn->scn_restart_txg); 604 } else if (dsl_scan_resilvering(dp)) { 605 /* 606 * If a resilver is in progress and there are already 607 * errors, restart it instead of finishing this scan and 608 * then restarting it. If there haven't been any errors 609 * then remember that the incore DTL is valid. 610 */ 611 if (scn->scn_phys.scn_errors > 0) { 612 scn->scn_restart_txg = txg; 613 zfs_dbgmsg("resilver can't excise DTL_MISSING " 614 "when finished; restarting on %s in txg " 615 "%llu", 616 spa->spa_name, 617 (u_longlong_t)scn->scn_restart_txg); 618 } else { 619 /* it's safe to excise DTL when finished */ 620 spa->spa_scrub_started = B_TRUE; 621 } 622 } 623 } 624 625 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys)); 626 627 /* reload the queue into the in-core state */ 628 if (scn->scn_phys.scn_queue_obj != 0) { 629 zap_cursor_t zc; 630 zap_attribute_t *za = zap_attribute_alloc(); 631 632 for (zap_cursor_init(&zc, dp->dp_meta_objset, 633 scn->scn_phys.scn_queue_obj); 634 zap_cursor_retrieve(&zc, za) == 0; 635 (void) zap_cursor_advance(&zc)) { 636 scan_ds_queue_insert(scn, 637 zfs_strtonum(za->za_name, NULL), 638 za->za_first_integer); 639 } 640 zap_cursor_fini(&zc); 641 zap_attribute_free(za); 642 } 643 644 ddt_walk_init(spa, scn->scn_phys.scn_max_txg); 645 646 spa_scan_stat_init(spa); 647 vdev_scan_stat_init(spa->spa_root_vdev); 648 649 return (0); 650 } 651 652 void 653 dsl_scan_fini(dsl_pool_t *dp) 654 { 655 if (dp->dp_scan != NULL) { 656 dsl_scan_t *scn = dp->dp_scan; 657 658 if (scn->scn_taskq != NULL) 659 taskq_destroy(scn->scn_taskq); 660 661 scan_ds_queue_clear(scn); 662 avl_destroy(&scn->scn_queue); 663 mutex_destroy(&scn->scn_queue_lock); 664 scan_ds_prefetch_queue_clear(scn); 665 avl_destroy(&scn->scn_prefetch_queue); 666 667 kmem_free(dp->dp_scan, sizeof (dsl_scan_t)); 668 dp->dp_scan = NULL; 669 } 670 } 671 672 static boolean_t 673 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx) 674 { 675 return (scn->scn_restart_txg != 0 && 676 scn->scn_restart_txg <= tx->tx_txg); 677 } 678 679 boolean_t 680 dsl_scan_resilver_scheduled(dsl_pool_t *dp) 681 { 682 return ((dp->dp_scan && dp->dp_scan->scn_restart_txg != 0) || 683 (spa_async_tasks(dp->dp_spa) & SPA_ASYNC_RESILVER)); 684 } 685 686 boolean_t 687 dsl_scan_scrubbing(const dsl_pool_t *dp) 688 { 689 dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys; 690 691 return (scn_phys->scn_state == DSS_SCANNING && 692 scn_phys->scn_func == POOL_SCAN_SCRUB); 693 } 694 695 boolean_t 696 dsl_errorscrubbing(const dsl_pool_t *dp) 697 { 698 dsl_errorscrub_phys_t *errorscrub_phys = &dp->dp_scan->errorscrub_phys; 699 700 return (errorscrub_phys->dep_state == DSS_ERRORSCRUBBING && 701 errorscrub_phys->dep_func == POOL_SCAN_ERRORSCRUB); 702 } 703 704 boolean_t 705 dsl_errorscrub_is_paused(const dsl_scan_t *scn) 706 { 707 return (dsl_errorscrubbing(scn->scn_dp) && 708 scn->errorscrub_phys.dep_paused_flags); 709 } 710 711 boolean_t 712 dsl_scan_is_paused_scrub(const dsl_scan_t *scn) 713 { 714 return (dsl_scan_scrubbing(scn->scn_dp) && 715 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED); 716 } 717 718 static void 719 dsl_errorscrub_sync_state(dsl_scan_t *scn, dmu_tx_t *tx) 720 { 721 scn->errorscrub_phys.dep_cursor = 722 zap_cursor_serialize(&scn->errorscrub_cursor); 723 724 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset, 725 DMU_POOL_DIRECTORY_OBJECT, 726 DMU_POOL_ERRORSCRUB, sizeof (uint64_t), ERRORSCRUB_PHYS_NUMINTS, 727 &scn->errorscrub_phys, tx)); 728 } 729 730 static void 731 dsl_errorscrub_setup_sync(void *arg, dmu_tx_t *tx) 732 { 733 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 734 pool_scan_func_t *funcp = arg; 735 dsl_pool_t *dp = scn->scn_dp; 736 spa_t *spa = dp->dp_spa; 737 738 ASSERT(!dsl_scan_is_running(scn)); 739 ASSERT(!dsl_errorscrubbing(scn->scn_dp)); 740 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS); 741 742 memset(&scn->errorscrub_phys, 0, sizeof (scn->errorscrub_phys)); 743 scn->errorscrub_phys.dep_func = *funcp; 744 scn->errorscrub_phys.dep_state = DSS_ERRORSCRUBBING; 745 scn->errorscrub_phys.dep_start_time = gethrestime_sec(); 746 scn->errorscrub_phys.dep_to_examine = spa_get_last_errlog_size(spa); 747 scn->errorscrub_phys.dep_examined = 0; 748 scn->errorscrub_phys.dep_errors = 0; 749 scn->errorscrub_phys.dep_cursor = 0; 750 zap_cursor_init_serialized(&scn->errorscrub_cursor, 751 spa->spa_meta_objset, spa->spa_errlog_last, 752 scn->errorscrub_phys.dep_cursor); 753 754 vdev_config_dirty(spa->spa_root_vdev); 755 spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_START); 756 757 dsl_errorscrub_sync_state(scn, tx); 758 759 spa_history_log_internal(spa, "error scrub setup", tx, 760 "func=%u mintxg=%u maxtxg=%llu", 761 *funcp, 0, (u_longlong_t)tx->tx_txg); 762 } 763 764 static int 765 dsl_errorscrub_setup_check(void *arg, dmu_tx_t *tx) 766 { 767 (void) arg; 768 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 769 770 if (dsl_scan_is_running(scn) || (dsl_errorscrubbing(scn->scn_dp))) { 771 return (SET_ERROR(EBUSY)); 772 } 773 774 if (spa_get_last_errlog_size(scn->scn_dp->dp_spa) == 0) { 775 return (ECANCELED); 776 } 777 return (0); 778 } 779 780 /* 781 * Writes out a persistent dsl_scan_phys_t record to the pool directory. 782 * Because we can be running in the block sorting algorithm, we do not always 783 * want to write out the record, only when it is "safe" to do so. This safety 784 * condition is achieved by making sure that the sorting queues are empty 785 * (scn_queues_pending == 0). When this condition is not true, the sync'd state 786 * is inconsistent with how much actual scanning progress has been made. The 787 * kind of sync to be performed is specified by the sync_type argument. If the 788 * sync is optional, we only sync if the queues are empty. If the sync is 789 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The 790 * third possible state is a "cached" sync. This is done in response to: 791 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been 792 * destroyed, so we wouldn't be able to restart scanning from it. 793 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been 794 * superseded by a newer snapshot. 795 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been 796 * swapped with its clone. 797 * In all cases, a cached sync simply rewrites the last record we've written, 798 * just slightly modified. For the modifications that are performed to the 799 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed, 800 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped. 801 */ 802 static void 803 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type) 804 { 805 int i; 806 spa_t *spa = scn->scn_dp->dp_spa; 807 808 ASSERT(sync_type != SYNC_MANDATORY || scn->scn_queues_pending == 0); 809 if (scn->scn_queues_pending == 0) { 810 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) { 811 vdev_t *vd = spa->spa_root_vdev->vdev_child[i]; 812 dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue; 813 814 if (q == NULL) 815 continue; 816 817 mutex_enter(&vd->vdev_scan_io_queue_lock); 818 ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL); 819 ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==, 820 NULL); 821 ASSERT3P(zfs_range_tree_first(q->q_exts_by_addr), ==, 822 NULL); 823 mutex_exit(&vd->vdev_scan_io_queue_lock); 824 } 825 826 if (scn->scn_phys.scn_queue_obj != 0) 827 scan_ds_queue_sync(scn, tx); 828 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset, 829 DMU_POOL_DIRECTORY_OBJECT, 830 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 831 &scn->scn_phys, tx)); 832 memcpy(&scn->scn_phys_cached, &scn->scn_phys, 833 sizeof (scn->scn_phys)); 834 835 if (scn->scn_checkpointing) 836 zfs_dbgmsg("finish scan checkpoint for %s", 837 spa->spa_name); 838 839 scn->scn_checkpointing = B_FALSE; 840 scn->scn_last_checkpoint = ddi_get_lbolt(); 841 } else if (sync_type == SYNC_CACHED) { 842 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset, 843 DMU_POOL_DIRECTORY_OBJECT, 844 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 845 &scn->scn_phys_cached, tx)); 846 } 847 } 848 849 int 850 dsl_scan_setup_check(void *arg, dmu_tx_t *tx) 851 { 852 (void) arg; 853 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 854 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev; 855 856 if (dsl_scan_is_running(scn) || vdev_rebuild_active(rvd) || 857 dsl_errorscrubbing(scn->scn_dp)) 858 return (SET_ERROR(EBUSY)); 859 860 return (0); 861 } 862 863 void 864 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx) 865 { 866 setup_sync_arg_t *setup_sync_arg = (setup_sync_arg_t *)arg; 867 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 868 dmu_object_type_t ot = 0; 869 dsl_pool_t *dp = scn->scn_dp; 870 spa_t *spa = dp->dp_spa; 871 872 ASSERT(!dsl_scan_is_running(scn)); 873 ASSERT3U(setup_sync_arg->func, >, POOL_SCAN_NONE); 874 ASSERT3U(setup_sync_arg->func, <, POOL_SCAN_FUNCS); 875 memset(&scn->scn_phys, 0, sizeof (scn->scn_phys)); 876 877 /* 878 * If we are starting a fresh scrub, we erase the error scrub 879 * information from disk. 880 */ 881 memset(&scn->errorscrub_phys, 0, sizeof (scn->errorscrub_phys)); 882 dsl_errorscrub_sync_state(scn, tx); 883 884 scn->scn_phys.scn_func = setup_sync_arg->func; 885 scn->scn_phys.scn_state = DSS_SCANNING; 886 scn->scn_phys.scn_min_txg = setup_sync_arg->txgstart; 887 if (setup_sync_arg->txgend == 0) { 888 scn->scn_phys.scn_max_txg = tx->tx_txg; 889 } else { 890 scn->scn_phys.scn_max_txg = setup_sync_arg->txgend; 891 } 892 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */ 893 scn->scn_phys.scn_start_time = gethrestime_sec(); 894 scn->scn_phys.scn_errors = 0; 895 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc; 896 scn->scn_issued_before_pass = 0; 897 scn->scn_restart_txg = 0; 898 scn->scn_done_txg = 0; 899 scn->scn_last_checkpoint = 0; 900 scn->scn_checkpointing = B_FALSE; 901 spa_scan_stat_init(spa); 902 vdev_scan_stat_init(spa->spa_root_vdev); 903 904 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 905 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max; 906 907 /* rewrite all disk labels */ 908 vdev_config_dirty(spa->spa_root_vdev); 909 910 if (vdev_resilver_needed(spa->spa_root_vdev, 911 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) { 912 nvlist_t *aux = fnvlist_alloc(); 913 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE, 914 "healing"); 915 spa_event_notify(spa, NULL, aux, 916 ESC_ZFS_RESILVER_START); 917 nvlist_free(aux); 918 } else { 919 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START); 920 } 921 922 spa->spa_scrub_started = B_TRUE; 923 /* 924 * If this is an incremental scrub, limit the DDT scrub phase 925 * to just the auto-ditto class (for correctness); the rest 926 * of the scrub should go faster using top-down pruning. 927 */ 928 if (scn->scn_phys.scn_min_txg > TXG_INITIAL) 929 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO; 930 931 /* 932 * When starting a resilver clear any existing rebuild state. 933 * This is required to prevent stale rebuild status from 934 * being reported when a rebuild is run, then a resilver and 935 * finally a scrub. In which case only the scrub status 936 * should be reported by 'zpool status'. 937 */ 938 if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) { 939 vdev_t *rvd = spa->spa_root_vdev; 940 for (uint64_t i = 0; i < rvd->vdev_children; i++) { 941 vdev_t *vd = rvd->vdev_child[i]; 942 vdev_rebuild_clear_sync( 943 (void *)(uintptr_t)vd->vdev_id, tx); 944 } 945 } 946 } 947 948 /* back to the generic stuff */ 949 950 if (zfs_scan_blkstats) { 951 if (dp->dp_blkstats == NULL) { 952 dp->dp_blkstats = 953 vmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP); 954 } 955 memset(&dp->dp_blkstats->zab_type, 0, 956 sizeof (dp->dp_blkstats->zab_type)); 957 } else { 958 if (dp->dp_blkstats) { 959 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 960 dp->dp_blkstats = NULL; 961 } 962 } 963 964 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) 965 ot = DMU_OT_ZAP_OTHER; 966 967 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, 968 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx); 969 970 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys)); 971 972 ddt_walk_init(spa, scn->scn_phys.scn_max_txg); 973 974 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY); 975 976 spa_history_log_internal(spa, "scan setup", tx, 977 "func=%u mintxg=%llu maxtxg=%llu", 978 setup_sync_arg->func, (u_longlong_t)scn->scn_phys.scn_min_txg, 979 (u_longlong_t)scn->scn_phys.scn_max_txg); 980 } 981 982 /* 983 * Called by ZFS_IOC_POOL_SCRUB and ZFS_IOC_POOL_SCAN ioctl to start a scrub, 984 * error scrub or resilver. Can also be called to resume a paused scrub or 985 * error scrub. 986 */ 987 int 988 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func, uint64_t txgstart, 989 uint64_t txgend) 990 { 991 spa_t *spa = dp->dp_spa; 992 dsl_scan_t *scn = dp->dp_scan; 993 setup_sync_arg_t setup_sync_arg; 994 995 if (func != POOL_SCAN_SCRUB && (txgstart != 0 || txgend != 0)) { 996 return (EINVAL); 997 } 998 999 /* 1000 * Purge all vdev caches and probe all devices. We do this here 1001 * rather than in sync context because this requires a writer lock 1002 * on the spa_config lock, which we can't do from sync context. The 1003 * spa_scrub_reopen flag indicates that vdev_open() should not 1004 * attempt to start another scrub. 1005 */ 1006 spa_vdev_state_enter(spa, SCL_NONE); 1007 spa->spa_scrub_reopen = B_TRUE; 1008 vdev_reopen(spa->spa_root_vdev); 1009 spa->spa_scrub_reopen = B_FALSE; 1010 (void) spa_vdev_state_exit(spa, NULL, 0); 1011 1012 if (func == POOL_SCAN_RESILVER) { 1013 dsl_scan_restart_resilver(spa->spa_dsl_pool, 0); 1014 return (0); 1015 } 1016 1017 if (func == POOL_SCAN_ERRORSCRUB) { 1018 if (dsl_errorscrub_is_paused(dp->dp_scan)) { 1019 /* 1020 * got error scrub start cmd, resume paused error scrub. 1021 */ 1022 int err = dsl_scrub_set_pause_resume(scn->scn_dp, 1023 POOL_SCRUB_NORMAL); 1024 if (err == 0) { 1025 spa_event_notify(spa, NULL, NULL, 1026 ESC_ZFS_ERRORSCRUB_RESUME); 1027 return (0); 1028 } 1029 return (SET_ERROR(err)); 1030 } 1031 1032 return (dsl_sync_task(spa_name(dp->dp_spa), 1033 dsl_errorscrub_setup_check, dsl_errorscrub_setup_sync, 1034 &func, 0, ZFS_SPACE_CHECK_RESERVED)); 1035 } 1036 1037 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) { 1038 /* got scrub start cmd, resume paused scrub */ 1039 int err = dsl_scrub_set_pause_resume(scn->scn_dp, 1040 POOL_SCRUB_NORMAL); 1041 if (err == 0) { 1042 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME); 1043 return (0); 1044 } 1045 return (SET_ERROR(err)); 1046 } 1047 1048 setup_sync_arg.func = func; 1049 setup_sync_arg.txgstart = txgstart; 1050 setup_sync_arg.txgend = txgend; 1051 1052 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check, 1053 dsl_scan_setup_sync, &setup_sync_arg, 0, 1054 ZFS_SPACE_CHECK_EXTRA_RESERVED)); 1055 } 1056 1057 static void 1058 dsl_errorscrub_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx) 1059 { 1060 dsl_pool_t *dp = scn->scn_dp; 1061 spa_t *spa = dp->dp_spa; 1062 1063 if (complete) { 1064 spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_FINISH); 1065 spa_history_log_internal(spa, "error scrub done", tx, 1066 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa)); 1067 } else { 1068 spa_history_log_internal(spa, "error scrub canceled", tx, 1069 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa)); 1070 } 1071 1072 scn->errorscrub_phys.dep_state = complete ? DSS_FINISHED : DSS_CANCELED; 1073 spa->spa_scrub_active = B_FALSE; 1074 spa_errlog_rotate(spa); 1075 scn->errorscrub_phys.dep_end_time = gethrestime_sec(); 1076 zap_cursor_fini(&scn->errorscrub_cursor); 1077 1078 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB) 1079 spa->spa_errata = 0; 1080 1081 ASSERT(!dsl_errorscrubbing(scn->scn_dp)); 1082 } 1083 1084 static void 1085 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx) 1086 { 1087 static const char *old_names[] = { 1088 "scrub_bookmark", 1089 "scrub_ddt_bookmark", 1090 "scrub_ddt_class_max", 1091 "scrub_queue", 1092 "scrub_min_txg", 1093 "scrub_max_txg", 1094 "scrub_func", 1095 "scrub_errors", 1096 NULL 1097 }; 1098 1099 dsl_pool_t *dp = scn->scn_dp; 1100 spa_t *spa = dp->dp_spa; 1101 int i; 1102 1103 /* Remove any remnants of an old-style scrub. */ 1104 for (i = 0; old_names[i]; i++) { 1105 (void) zap_remove(dp->dp_meta_objset, 1106 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx); 1107 } 1108 1109 if (scn->scn_phys.scn_queue_obj != 0) { 1110 VERIFY0(dmu_object_free(dp->dp_meta_objset, 1111 scn->scn_phys.scn_queue_obj, tx)); 1112 scn->scn_phys.scn_queue_obj = 0; 1113 } 1114 scan_ds_queue_clear(scn); 1115 scan_ds_prefetch_queue_clear(scn); 1116 1117 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED; 1118 1119 /* 1120 * If we were "restarted" from a stopped state, don't bother 1121 * with anything else. 1122 */ 1123 if (!dsl_scan_is_running(scn)) { 1124 ASSERT(!scn->scn_is_sorted); 1125 return; 1126 } 1127 1128 if (scn->scn_is_sorted) { 1129 scan_io_queues_destroy(scn); 1130 scn->scn_is_sorted = B_FALSE; 1131 1132 if (scn->scn_taskq != NULL) { 1133 taskq_destroy(scn->scn_taskq); 1134 scn->scn_taskq = NULL; 1135 } 1136 } 1137 1138 if (dsl_scan_restarting(scn, tx)) { 1139 spa_history_log_internal(spa, "scan aborted, restarting", tx, 1140 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa)); 1141 } else if (!complete) { 1142 spa_history_log_internal(spa, "scan cancelled", tx, 1143 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa)); 1144 } else { 1145 spa_history_log_internal(spa, "scan done", tx, 1146 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa)); 1147 if (DSL_SCAN_IS_SCRUB(scn)) { 1148 VERIFY0(zap_update(dp->dp_meta_objset, 1149 DMU_POOL_DIRECTORY_OBJECT, 1150 DMU_POOL_LAST_SCRUBBED_TXG, 1151 sizeof (uint64_t), 1, 1152 &scn->scn_phys.scn_max_txg, tx)); 1153 spa->spa_scrubbed_last_txg = scn->scn_phys.scn_max_txg; 1154 } 1155 } 1156 1157 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 1158 spa->spa_scrub_active = B_FALSE; 1159 1160 /* 1161 * If the scrub/resilver completed, update all DTLs to 1162 * reflect this. Whether it succeeded or not, vacate 1163 * all temporary scrub DTLs. 1164 * 1165 * As the scrub does not currently support traversing 1166 * data that have been freed but are part of a checkpoint, 1167 * we don't mark the scrub as done in the DTLs as faults 1168 * may still exist in those vdevs. 1169 */ 1170 if (complete && 1171 !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 1172 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg, 1173 scn->scn_phys.scn_max_txg, B_TRUE, B_FALSE); 1174 1175 if (DSL_SCAN_IS_RESILVER(scn)) { 1176 nvlist_t *aux = fnvlist_alloc(); 1177 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE, 1178 "healing"); 1179 spa_event_notify(spa, NULL, aux, 1180 ESC_ZFS_RESILVER_FINISH); 1181 nvlist_free(aux); 1182 } else { 1183 spa_event_notify(spa, NULL, NULL, 1184 ESC_ZFS_SCRUB_FINISH); 1185 } 1186 } else { 1187 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg, 1188 0, B_TRUE, B_FALSE); 1189 } 1190 spa_errlog_rotate(spa); 1191 1192 /* 1193 * Don't clear flag until after vdev_dtl_reassess to ensure that 1194 * DTL_MISSING will get updated when possible. 1195 */ 1196 scn->scn_phys.scn_state = complete ? DSS_FINISHED : 1197 DSS_CANCELED; 1198 scn->scn_phys.scn_end_time = gethrestime_sec(); 1199 spa->spa_scrub_started = B_FALSE; 1200 1201 /* 1202 * We may have finished replacing a device. 1203 * Let the async thread assess this and handle the detach. 1204 */ 1205 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); 1206 1207 /* 1208 * Clear any resilver_deferred flags in the config. 1209 * If there are drives that need resilvering, kick 1210 * off an asynchronous request to start resilver. 1211 * vdev_clear_resilver_deferred() may update the config 1212 * before the resilver can restart. In the event of 1213 * a crash during this period, the spa loading code 1214 * will find the drives that need to be resilvered 1215 * and start the resilver then. 1216 */ 1217 if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER) && 1218 vdev_clear_resilver_deferred(spa->spa_root_vdev, tx)) { 1219 spa_history_log_internal(spa, 1220 "starting deferred resilver", tx, "errors=%llu", 1221 (u_longlong_t)spa_approx_errlog_size(spa)); 1222 spa_async_request(spa, SPA_ASYNC_RESILVER); 1223 } 1224 1225 /* Clear recent error events (i.e. duplicate events tracking) */ 1226 if (complete) 1227 zfs_ereport_clear(spa, NULL); 1228 } else { 1229 scn->scn_phys.scn_state = complete ? DSS_FINISHED : 1230 DSS_CANCELED; 1231 scn->scn_phys.scn_end_time = gethrestime_sec(); 1232 } 1233 1234 spa_notify_waiters(spa); 1235 1236 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB) 1237 spa->spa_errata = 0; 1238 1239 ASSERT(!dsl_scan_is_running(scn)); 1240 } 1241 1242 static int 1243 dsl_errorscrub_pause_resume_check(void *arg, dmu_tx_t *tx) 1244 { 1245 pool_scrub_cmd_t *cmd = arg; 1246 dsl_pool_t *dp = dmu_tx_pool(tx); 1247 dsl_scan_t *scn = dp->dp_scan; 1248 1249 if (*cmd == POOL_SCRUB_PAUSE) { 1250 /* 1251 * can't pause a error scrub when there is no in-progress 1252 * error scrub. 1253 */ 1254 if (!dsl_errorscrubbing(dp)) 1255 return (SET_ERROR(ENOENT)); 1256 1257 /* can't pause a paused error scrub */ 1258 if (dsl_errorscrub_is_paused(scn)) 1259 return (SET_ERROR(EBUSY)); 1260 } else if (*cmd != POOL_SCRUB_NORMAL) { 1261 return (SET_ERROR(ENOTSUP)); 1262 } 1263 1264 return (0); 1265 } 1266 1267 static void 1268 dsl_errorscrub_pause_resume_sync(void *arg, dmu_tx_t *tx) 1269 { 1270 pool_scrub_cmd_t *cmd = arg; 1271 dsl_pool_t *dp = dmu_tx_pool(tx); 1272 spa_t *spa = dp->dp_spa; 1273 dsl_scan_t *scn = dp->dp_scan; 1274 1275 if (*cmd == POOL_SCRUB_PAUSE) { 1276 spa->spa_scan_pass_errorscrub_pause = gethrestime_sec(); 1277 scn->errorscrub_phys.dep_paused_flags = B_TRUE; 1278 dsl_errorscrub_sync_state(scn, tx); 1279 spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_PAUSED); 1280 } else { 1281 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL); 1282 if (dsl_errorscrub_is_paused(scn)) { 1283 /* 1284 * We need to keep track of how much time we spend 1285 * paused per pass so that we can adjust the error scrub 1286 * rate shown in the output of 'zpool status'. 1287 */ 1288 spa->spa_scan_pass_errorscrub_spent_paused += 1289 gethrestime_sec() - 1290 spa->spa_scan_pass_errorscrub_pause; 1291 1292 spa->spa_scan_pass_errorscrub_pause = 0; 1293 scn->errorscrub_phys.dep_paused_flags = B_FALSE; 1294 1295 zap_cursor_init_serialized( 1296 &scn->errorscrub_cursor, 1297 spa->spa_meta_objset, spa->spa_errlog_last, 1298 scn->errorscrub_phys.dep_cursor); 1299 1300 dsl_errorscrub_sync_state(scn, tx); 1301 } 1302 } 1303 } 1304 1305 static int 1306 dsl_errorscrub_cancel_check(void *arg, dmu_tx_t *tx) 1307 { 1308 (void) arg; 1309 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 1310 /* can't cancel a error scrub when there is no one in-progress */ 1311 if (!dsl_errorscrubbing(scn->scn_dp)) 1312 return (SET_ERROR(ENOENT)); 1313 return (0); 1314 } 1315 1316 static void 1317 dsl_errorscrub_cancel_sync(void *arg, dmu_tx_t *tx) 1318 { 1319 (void) arg; 1320 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 1321 1322 dsl_errorscrub_done(scn, B_FALSE, tx); 1323 dsl_errorscrub_sync_state(scn, tx); 1324 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, 1325 ESC_ZFS_ERRORSCRUB_ABORT); 1326 } 1327 1328 static int 1329 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx) 1330 { 1331 (void) arg; 1332 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 1333 1334 if (!dsl_scan_is_running(scn)) 1335 return (SET_ERROR(ENOENT)); 1336 return (0); 1337 } 1338 1339 static void 1340 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx) 1341 { 1342 (void) arg; 1343 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 1344 1345 dsl_scan_done(scn, B_FALSE, tx); 1346 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY); 1347 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT); 1348 } 1349 1350 int 1351 dsl_scan_cancel(dsl_pool_t *dp) 1352 { 1353 if (dsl_errorscrubbing(dp)) { 1354 return (dsl_sync_task(spa_name(dp->dp_spa), 1355 dsl_errorscrub_cancel_check, dsl_errorscrub_cancel_sync, 1356 NULL, 3, ZFS_SPACE_CHECK_RESERVED)); 1357 } 1358 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check, 1359 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED)); 1360 } 1361 1362 static int 1363 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx) 1364 { 1365 pool_scrub_cmd_t *cmd = arg; 1366 dsl_pool_t *dp = dmu_tx_pool(tx); 1367 dsl_scan_t *scn = dp->dp_scan; 1368 1369 if (*cmd == POOL_SCRUB_PAUSE) { 1370 /* can't pause a scrub when there is no in-progress scrub */ 1371 if (!dsl_scan_scrubbing(dp)) 1372 return (SET_ERROR(ENOENT)); 1373 1374 /* can't pause a paused scrub */ 1375 if (dsl_scan_is_paused_scrub(scn)) 1376 return (SET_ERROR(EBUSY)); 1377 } else if (*cmd != POOL_SCRUB_NORMAL) { 1378 return (SET_ERROR(ENOTSUP)); 1379 } 1380 1381 return (0); 1382 } 1383 1384 static void 1385 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx) 1386 { 1387 pool_scrub_cmd_t *cmd = arg; 1388 dsl_pool_t *dp = dmu_tx_pool(tx); 1389 spa_t *spa = dp->dp_spa; 1390 dsl_scan_t *scn = dp->dp_scan; 1391 1392 if (*cmd == POOL_SCRUB_PAUSE) { 1393 /* can't pause a scrub when there is no in-progress scrub */ 1394 spa->spa_scan_pass_scrub_pause = gethrestime_sec(); 1395 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED; 1396 scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED; 1397 dsl_scan_sync_state(scn, tx, SYNC_CACHED); 1398 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED); 1399 spa_notify_waiters(spa); 1400 } else { 1401 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL); 1402 if (dsl_scan_is_paused_scrub(scn)) { 1403 /* 1404 * We need to keep track of how much time we spend 1405 * paused per pass so that we can adjust the scrub rate 1406 * shown in the output of 'zpool status' 1407 */ 1408 spa->spa_scan_pass_scrub_spent_paused += 1409 gethrestime_sec() - spa->spa_scan_pass_scrub_pause; 1410 spa->spa_scan_pass_scrub_pause = 0; 1411 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED; 1412 scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED; 1413 dsl_scan_sync_state(scn, tx, SYNC_CACHED); 1414 } 1415 } 1416 } 1417 1418 /* 1419 * Set scrub pause/resume state if it makes sense to do so 1420 */ 1421 int 1422 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd) 1423 { 1424 if (dsl_errorscrubbing(dp)) { 1425 return (dsl_sync_task(spa_name(dp->dp_spa), 1426 dsl_errorscrub_pause_resume_check, 1427 dsl_errorscrub_pause_resume_sync, &cmd, 3, 1428 ZFS_SPACE_CHECK_RESERVED)); 1429 } 1430 return (dsl_sync_task(spa_name(dp->dp_spa), 1431 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3, 1432 ZFS_SPACE_CHECK_RESERVED)); 1433 } 1434 1435 1436 /* start a new scan, or restart an existing one. */ 1437 void 1438 dsl_scan_restart_resilver(dsl_pool_t *dp, uint64_t txg) 1439 { 1440 if (txg == 0) { 1441 dmu_tx_t *tx; 1442 tx = dmu_tx_create_dd(dp->dp_mos_dir); 1443 VERIFY0(dmu_tx_assign(tx, DMU_TX_WAIT | DMU_TX_SUSPEND)); 1444 1445 txg = dmu_tx_get_txg(tx); 1446 dp->dp_scan->scn_restart_txg = txg; 1447 dmu_tx_commit(tx); 1448 } else { 1449 dp->dp_scan->scn_restart_txg = txg; 1450 } 1451 zfs_dbgmsg("restarting resilver for %s at txg=%llu", 1452 dp->dp_spa->spa_name, (longlong_t)txg); 1453 } 1454 1455 void 1456 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp) 1457 { 1458 zio_free(dp->dp_spa, txg, bp); 1459 } 1460 1461 void 1462 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp) 1463 { 1464 ASSERT(dsl_pool_sync_context(dp)); 1465 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags)); 1466 } 1467 1468 static int 1469 scan_ds_queue_compare(const void *a, const void *b) 1470 { 1471 const scan_ds_t *sds_a = a, *sds_b = b; 1472 1473 if (sds_a->sds_dsobj < sds_b->sds_dsobj) 1474 return (-1); 1475 if (sds_a->sds_dsobj == sds_b->sds_dsobj) 1476 return (0); 1477 return (1); 1478 } 1479 1480 static void 1481 scan_ds_queue_clear(dsl_scan_t *scn) 1482 { 1483 void *cookie = NULL; 1484 scan_ds_t *sds; 1485 while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) { 1486 kmem_free(sds, sizeof (*sds)); 1487 } 1488 } 1489 1490 static boolean_t 1491 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg) 1492 { 1493 scan_ds_t srch, *sds; 1494 1495 srch.sds_dsobj = dsobj; 1496 sds = avl_find(&scn->scn_queue, &srch, NULL); 1497 if (sds != NULL && txg != NULL) 1498 *txg = sds->sds_txg; 1499 return (sds != NULL); 1500 } 1501 1502 static void 1503 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg) 1504 { 1505 scan_ds_t *sds; 1506 avl_index_t where; 1507 1508 sds = kmem_zalloc(sizeof (*sds), KM_SLEEP); 1509 sds->sds_dsobj = dsobj; 1510 sds->sds_txg = txg; 1511 1512 VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL); 1513 avl_insert(&scn->scn_queue, sds, where); 1514 } 1515 1516 static void 1517 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj) 1518 { 1519 scan_ds_t srch, *sds; 1520 1521 srch.sds_dsobj = dsobj; 1522 1523 sds = avl_find(&scn->scn_queue, &srch, NULL); 1524 VERIFY(sds != NULL); 1525 avl_remove(&scn->scn_queue, sds); 1526 kmem_free(sds, sizeof (*sds)); 1527 } 1528 1529 static void 1530 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx) 1531 { 1532 dsl_pool_t *dp = scn->scn_dp; 1533 spa_t *spa = dp->dp_spa; 1534 dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ? 1535 DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER; 1536 1537 ASSERT0(scn->scn_queues_pending); 1538 ASSERT(scn->scn_phys.scn_queue_obj != 0); 1539 1540 VERIFY0(dmu_object_free(dp->dp_meta_objset, 1541 scn->scn_phys.scn_queue_obj, tx)); 1542 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot, 1543 DMU_OT_NONE, 0, tx); 1544 for (scan_ds_t *sds = avl_first(&scn->scn_queue); 1545 sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) { 1546 VERIFY0(zap_add_int_key(dp->dp_meta_objset, 1547 scn->scn_phys.scn_queue_obj, sds->sds_dsobj, 1548 sds->sds_txg, tx)); 1549 } 1550 } 1551 1552 /* 1553 * Computes the memory limit state that we're currently in. A sorted scan 1554 * needs quite a bit of memory to hold the sorting queue, so we need to 1555 * reasonably constrain the size so it doesn't impact overall system 1556 * performance. We compute two limits: 1557 * 1) Hard memory limit: if the amount of memory used by the sorting 1558 * queues on a pool gets above this value, we stop the metadata 1559 * scanning portion and start issuing the queued up and sorted 1560 * I/Os to reduce memory usage. 1561 * This limit is calculated as a fraction of physmem (by default 5%). 1562 * We constrain the lower bound of the hard limit to an absolute 1563 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain 1564 * the upper bound to 5% of the total pool size - no chance we'll 1565 * ever need that much memory, but just to keep the value in check. 1566 * 2) Soft memory limit: once we hit the hard memory limit, we start 1567 * issuing I/O to reduce queue memory usage, but we don't want to 1568 * completely empty out the queues, since we might be able to find I/Os 1569 * that will fill in the gaps of our non-sequential IOs at some point 1570 * in the future. So we stop the issuing of I/Os once the amount of 1571 * memory used drops below the soft limit (at which point we stop issuing 1572 * I/O and start scanning metadata again). 1573 * 1574 * This limit is calculated by subtracting a fraction of the hard 1575 * limit from the hard limit. By default this fraction is 5%, so 1576 * the soft limit is 95% of the hard limit. We cap the size of the 1577 * difference between the hard and soft limits at an absolute 1578 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is 1579 * sufficient to not cause too frequent switching between the 1580 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's 1581 * worth of queues is about 1.2 GiB of on-pool data, so scanning 1582 * that should take at least a decent fraction of a second). 1583 */ 1584 static boolean_t 1585 dsl_scan_should_clear(dsl_scan_t *scn) 1586 { 1587 spa_t *spa = scn->scn_dp->dp_spa; 1588 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev; 1589 uint64_t alloc, mlim_hard, mlim_soft, mused; 1590 1591 alloc = metaslab_class_get_alloc(spa_normal_class(spa)); 1592 alloc += metaslab_class_get_alloc(spa_special_class(spa)); 1593 alloc += metaslab_class_get_alloc(spa_dedup_class(spa)); 1594 1595 mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE, 1596 zfs_scan_mem_lim_min); 1597 mlim_hard = MIN(mlim_hard, alloc / 20); 1598 mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact, 1599 zfs_scan_mem_lim_soft_max); 1600 mused = 0; 1601 for (uint64_t i = 0; i < rvd->vdev_children; i++) { 1602 vdev_t *tvd = rvd->vdev_child[i]; 1603 dsl_scan_io_queue_t *queue; 1604 1605 mutex_enter(&tvd->vdev_scan_io_queue_lock); 1606 queue = tvd->vdev_scan_io_queue; 1607 if (queue != NULL) { 1608 /* 1609 * # of extents in exts_by_addr = # in exts_by_size. 1610 * B-tree efficiency is ~75%, but can be as low as 50%. 1611 */ 1612 mused += zfs_btree_numnodes(&queue->q_exts_by_size) * (( 1613 sizeof (zfs_range_seg_gap_t) + sizeof (uint64_t)) * 1614 3 / 2) + queue->q_sio_memused; 1615 } 1616 mutex_exit(&tvd->vdev_scan_io_queue_lock); 1617 } 1618 1619 dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused); 1620 1621 if (mused == 0) 1622 ASSERT0(scn->scn_queues_pending); 1623 1624 /* 1625 * If we are above our hard limit, we need to clear out memory. 1626 * If we are below our soft limit, we need to accumulate sequential IOs. 1627 * Otherwise, we should keep doing whatever we are currently doing. 1628 */ 1629 if (mused >= mlim_hard) 1630 return (B_TRUE); 1631 else if (mused < mlim_soft) 1632 return (B_FALSE); 1633 else 1634 return (scn->scn_clearing); 1635 } 1636 1637 static boolean_t 1638 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb) 1639 { 1640 /* we never skip user/group accounting objects */ 1641 if (zb && (int64_t)zb->zb_object < 0) 1642 return (B_FALSE); 1643 1644 if (scn->scn_suspending) 1645 return (B_TRUE); /* we're already suspending */ 1646 1647 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) 1648 return (B_FALSE); /* we're resuming */ 1649 1650 /* We only know how to resume from level-0 and objset blocks. */ 1651 if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL)) 1652 return (B_FALSE); 1653 1654 /* 1655 * We suspend if: 1656 * - we have scanned for at least the minimum time (default 1 sec 1657 * for scrub, 3 sec for resilver), and either we have sufficient 1658 * dirty data that we are starting to write more quickly 1659 * (default 30%), someone is explicitly waiting for this txg 1660 * to complete, or we have used up all of the time in the txg 1661 * timeout (default 5 sec). 1662 * or 1663 * - the spa is shutting down because this pool is being exported 1664 * or the machine is rebooting. 1665 * or 1666 * - the scan queue has reached its memory use limit 1667 */ 1668 uint64_t curr_time_ns = gethrtime(); 1669 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time; 1670 uint64_t sync_time_ns = curr_time_ns - 1671 scn->scn_dp->dp_spa->spa_sync_starttime; 1672 uint64_t dirty_min_bytes = zfs_dirty_data_max * 1673 zfs_vdev_async_write_active_min_dirty_percent / 100; 1674 uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ? 1675 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms; 1676 1677 if ((NSEC2MSEC(scan_time_ns) > mintime && 1678 (scn->scn_dp->dp_dirty_total >= dirty_min_bytes || 1679 txg_sync_waiting(scn->scn_dp) || 1680 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) || 1681 spa_shutting_down(scn->scn_dp->dp_spa) || 1682 (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn)) || 1683 !ddt_walk_ready(scn->scn_dp->dp_spa)) { 1684 if (zb && zb->zb_level == ZB_ROOT_LEVEL) { 1685 dprintf("suspending at first available bookmark " 1686 "%llx/%llx/%llx/%llx\n", 1687 (longlong_t)zb->zb_objset, 1688 (longlong_t)zb->zb_object, 1689 (longlong_t)zb->zb_level, 1690 (longlong_t)zb->zb_blkid); 1691 SET_BOOKMARK(&scn->scn_phys.scn_bookmark, 1692 zb->zb_objset, 0, 0, 0); 1693 } else if (zb != NULL) { 1694 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n", 1695 (longlong_t)zb->zb_objset, 1696 (longlong_t)zb->zb_object, 1697 (longlong_t)zb->zb_level, 1698 (longlong_t)zb->zb_blkid); 1699 scn->scn_phys.scn_bookmark = *zb; 1700 } else { 1701 #ifdef ZFS_DEBUG 1702 dsl_scan_phys_t *scnp = &scn->scn_phys; 1703 dprintf("suspending at at DDT bookmark " 1704 "%llx/%llx/%llx/%llx\n", 1705 (longlong_t)scnp->scn_ddt_bookmark.ddb_class, 1706 (longlong_t)scnp->scn_ddt_bookmark.ddb_type, 1707 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum, 1708 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor); 1709 #endif 1710 } 1711 scn->scn_suspending = B_TRUE; 1712 return (B_TRUE); 1713 } 1714 return (B_FALSE); 1715 } 1716 1717 static boolean_t 1718 dsl_error_scrub_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb) 1719 { 1720 /* 1721 * We suspend if: 1722 * - we have scrubbed for at least the minimum time (default 1 sec 1723 * for error scrub), someone is explicitly waiting for this txg 1724 * to complete, or we have used up all of the time in the txg 1725 * timeout (default 5 sec). 1726 * or 1727 * - the spa is shutting down because this pool is being exported 1728 * or the machine is rebooting. 1729 */ 1730 uint64_t curr_time_ns = gethrtime(); 1731 uint64_t error_scrub_time_ns = curr_time_ns - scn->scn_sync_start_time; 1732 uint64_t sync_time_ns = curr_time_ns - 1733 scn->scn_dp->dp_spa->spa_sync_starttime; 1734 int mintime = zfs_scrub_min_time_ms; 1735 1736 if ((NSEC2MSEC(error_scrub_time_ns) > mintime && 1737 (txg_sync_waiting(scn->scn_dp) || 1738 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) || 1739 spa_shutting_down(scn->scn_dp->dp_spa)) { 1740 if (zb) { 1741 dprintf("error scrub suspending at bookmark " 1742 "%llx/%llx/%llx/%llx\n", 1743 (longlong_t)zb->zb_objset, 1744 (longlong_t)zb->zb_object, 1745 (longlong_t)zb->zb_level, 1746 (longlong_t)zb->zb_blkid); 1747 } 1748 return (B_TRUE); 1749 } 1750 return (B_FALSE); 1751 } 1752 1753 typedef struct zil_scan_arg { 1754 dsl_pool_t *zsa_dp; 1755 zil_header_t *zsa_zh; 1756 } zil_scan_arg_t; 1757 1758 static int 1759 dsl_scan_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg, 1760 uint64_t claim_txg) 1761 { 1762 (void) zilog; 1763 zil_scan_arg_t *zsa = arg; 1764 dsl_pool_t *dp = zsa->zsa_dp; 1765 dsl_scan_t *scn = dp->dp_scan; 1766 zil_header_t *zh = zsa->zsa_zh; 1767 zbookmark_phys_t zb; 1768 1769 ASSERT(!BP_IS_REDACTED(bp)); 1770 if (BP_IS_HOLE(bp) || 1771 BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg) 1772 return (0); 1773 1774 /* 1775 * One block ("stubby") can be allocated a long time ago; we 1776 * want to visit that one because it has been allocated 1777 * (on-disk) even if it hasn't been claimed (even though for 1778 * scrub there's nothing to do to it). 1779 */ 1780 if (claim_txg == 0 && 1781 BP_GET_LOGICAL_BIRTH(bp) >= spa_min_claim_txg(dp->dp_spa)) 1782 return (0); 1783 1784 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 1785 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 1786 1787 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 1788 return (0); 1789 } 1790 1791 static int 1792 dsl_scan_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg, 1793 uint64_t claim_txg) 1794 { 1795 (void) zilog; 1796 if (lrc->lrc_txtype == TX_WRITE) { 1797 zil_scan_arg_t *zsa = arg; 1798 dsl_pool_t *dp = zsa->zsa_dp; 1799 dsl_scan_t *scn = dp->dp_scan; 1800 zil_header_t *zh = zsa->zsa_zh; 1801 const lr_write_t *lr = (const lr_write_t *)lrc; 1802 const blkptr_t *bp = &lr->lr_blkptr; 1803 zbookmark_phys_t zb; 1804 1805 ASSERT(!BP_IS_REDACTED(bp)); 1806 if (BP_IS_HOLE(bp) || 1807 BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg) 1808 return (0); 1809 1810 /* 1811 * birth can be < claim_txg if this record's txg is 1812 * already txg sync'ed (but this log block contains 1813 * other records that are not synced) 1814 */ 1815 if (claim_txg == 0 || BP_GET_LOGICAL_BIRTH(bp) < claim_txg) 1816 return (0); 1817 1818 ASSERT3U(BP_GET_LSIZE(bp), !=, 0); 1819 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 1820 lr->lr_foid, ZB_ZIL_LEVEL, 1821 lr->lr_offset / BP_GET_LSIZE(bp)); 1822 1823 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 1824 } 1825 return (0); 1826 } 1827 1828 static void 1829 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh) 1830 { 1831 uint64_t claim_txg = zh->zh_claim_txg; 1832 zil_scan_arg_t zsa = { dp, zh }; 1833 zilog_t *zilog; 1834 1835 ASSERT(spa_writeable(dp->dp_spa)); 1836 1837 /* 1838 * We only want to visit blocks that have been claimed but not yet 1839 * replayed (or, in read-only mode, blocks that *would* be claimed). 1840 */ 1841 if (claim_txg == 0) 1842 return; 1843 1844 zilog = zil_alloc(dp->dp_meta_objset, zh); 1845 1846 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa, 1847 claim_txg, B_FALSE); 1848 1849 zil_free(zilog); 1850 } 1851 1852 /* 1853 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea 1854 * here is to sort the AVL tree by the order each block will be needed. 1855 */ 1856 static int 1857 scan_prefetch_queue_compare(const void *a, const void *b) 1858 { 1859 const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b; 1860 const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc; 1861 const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc; 1862 1863 return (zbookmark_compare(spc_a->spc_datablkszsec, 1864 spc_a->spc_indblkshift, spc_b->spc_datablkszsec, 1865 spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb)); 1866 } 1867 1868 static void 1869 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, const void *tag) 1870 { 1871 if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) { 1872 zfs_refcount_destroy(&spc->spc_refcnt); 1873 kmem_free(spc, sizeof (scan_prefetch_ctx_t)); 1874 } 1875 } 1876 1877 static scan_prefetch_ctx_t * 1878 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, const void *tag) 1879 { 1880 scan_prefetch_ctx_t *spc; 1881 1882 spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP); 1883 zfs_refcount_create(&spc->spc_refcnt); 1884 zfs_refcount_add(&spc->spc_refcnt, tag); 1885 spc->spc_scn = scn; 1886 if (dnp != NULL) { 1887 spc->spc_datablkszsec = dnp->dn_datablkszsec; 1888 spc->spc_indblkshift = dnp->dn_indblkshift; 1889 spc->spc_root = B_FALSE; 1890 } else { 1891 spc->spc_datablkszsec = 0; 1892 spc->spc_indblkshift = 0; 1893 spc->spc_root = B_TRUE; 1894 } 1895 1896 return (spc); 1897 } 1898 1899 static void 1900 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, const void *tag) 1901 { 1902 zfs_refcount_add(&spc->spc_refcnt, tag); 1903 } 1904 1905 static void 1906 scan_ds_prefetch_queue_clear(dsl_scan_t *scn) 1907 { 1908 spa_t *spa = scn->scn_dp->dp_spa; 1909 void *cookie = NULL; 1910 scan_prefetch_issue_ctx_t *spic = NULL; 1911 1912 mutex_enter(&spa->spa_scrub_lock); 1913 while ((spic = avl_destroy_nodes(&scn->scn_prefetch_queue, 1914 &cookie)) != NULL) { 1915 scan_prefetch_ctx_rele(spic->spic_spc, scn); 1916 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t)); 1917 } 1918 mutex_exit(&spa->spa_scrub_lock); 1919 } 1920 1921 static boolean_t 1922 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc, 1923 const zbookmark_phys_t *zb) 1924 { 1925 zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark; 1926 dnode_phys_t tmp_dnp; 1927 dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp; 1928 1929 if (zb->zb_objset != last_zb->zb_objset) 1930 return (B_TRUE); 1931 if ((int64_t)zb->zb_object < 0) 1932 return (B_FALSE); 1933 1934 tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec; 1935 tmp_dnp.dn_indblkshift = spc->spc_indblkshift; 1936 1937 if (zbookmark_subtree_completed(dnp, zb, last_zb)) 1938 return (B_TRUE); 1939 1940 return (B_FALSE); 1941 } 1942 1943 static void 1944 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb) 1945 { 1946 avl_index_t idx; 1947 dsl_scan_t *scn = spc->spc_scn; 1948 spa_t *spa = scn->scn_dp->dp_spa; 1949 scan_prefetch_issue_ctx_t *spic; 1950 1951 if (zfs_no_scrub_prefetch || BP_IS_REDACTED(bp)) 1952 return; 1953 1954 if (BP_IS_HOLE(bp) || 1955 BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg || 1956 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE && 1957 BP_GET_TYPE(bp) != DMU_OT_OBJSET)) 1958 return; 1959 1960 if (dsl_scan_check_prefetch_resume(spc, zb)) 1961 return; 1962 1963 scan_prefetch_ctx_add_ref(spc, scn); 1964 spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP); 1965 spic->spic_spc = spc; 1966 spic->spic_bp = *bp; 1967 spic->spic_zb = *zb; 1968 1969 /* 1970 * Add the IO to the queue of blocks to prefetch. This allows us to 1971 * prioritize blocks that we will need first for the main traversal 1972 * thread. 1973 */ 1974 mutex_enter(&spa->spa_scrub_lock); 1975 if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) { 1976 /* this block is already queued for prefetch */ 1977 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t)); 1978 scan_prefetch_ctx_rele(spc, scn); 1979 mutex_exit(&spa->spa_scrub_lock); 1980 return; 1981 } 1982 1983 avl_insert(&scn->scn_prefetch_queue, spic, idx); 1984 cv_broadcast(&spa->spa_scrub_io_cv); 1985 mutex_exit(&spa->spa_scrub_lock); 1986 } 1987 1988 static void 1989 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp, 1990 uint64_t objset, uint64_t object) 1991 { 1992 int i; 1993 zbookmark_phys_t zb; 1994 scan_prefetch_ctx_t *spc; 1995 1996 if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1997 return; 1998 1999 SET_BOOKMARK(&zb, objset, object, 0, 0); 2000 2001 spc = scan_prefetch_ctx_create(scn, dnp, FTAG); 2002 2003 for (i = 0; i < dnp->dn_nblkptr; i++) { 2004 zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]); 2005 zb.zb_blkid = i; 2006 dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb); 2007 } 2008 2009 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 2010 zb.zb_level = 0; 2011 zb.zb_blkid = DMU_SPILL_BLKID; 2012 dsl_scan_prefetch(spc, DN_SPILL_BLKPTR(dnp), &zb); 2013 } 2014 2015 scan_prefetch_ctx_rele(spc, FTAG); 2016 } 2017 2018 static void 2019 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, 2020 arc_buf_t *buf, void *private) 2021 { 2022 (void) zio; 2023 scan_prefetch_ctx_t *spc = private; 2024 dsl_scan_t *scn = spc->spc_scn; 2025 spa_t *spa = scn->scn_dp->dp_spa; 2026 2027 /* broadcast that the IO has completed for rate limiting purposes */ 2028 mutex_enter(&spa->spa_scrub_lock); 2029 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp)); 2030 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp); 2031 cv_broadcast(&spa->spa_scrub_io_cv); 2032 mutex_exit(&spa->spa_scrub_lock); 2033 2034 /* if there was an error or we are done prefetching, just cleanup */ 2035 if (buf == NULL || scn->scn_prefetch_stop) 2036 goto out; 2037 2038 if (BP_GET_LEVEL(bp) > 0) { 2039 int i; 2040 blkptr_t *cbp; 2041 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; 2042 zbookmark_phys_t czb; 2043 2044 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) { 2045 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, 2046 zb->zb_level - 1, zb->zb_blkid * epb + i); 2047 dsl_scan_prefetch(spc, cbp, &czb); 2048 } 2049 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { 2050 dnode_phys_t *cdnp; 2051 int i; 2052 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; 2053 2054 for (i = 0, cdnp = buf->b_data; i < epb; 2055 i += cdnp->dn_extra_slots + 1, 2056 cdnp += cdnp->dn_extra_slots + 1) { 2057 dsl_scan_prefetch_dnode(scn, cdnp, 2058 zb->zb_objset, zb->zb_blkid * epb + i); 2059 } 2060 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { 2061 objset_phys_t *osp = buf->b_data; 2062 2063 dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode, 2064 zb->zb_objset, DMU_META_DNODE_OBJECT); 2065 2066 if (OBJSET_BUF_HAS_USERUSED(buf)) { 2067 if (OBJSET_BUF_HAS_PROJECTUSED(buf)) { 2068 dsl_scan_prefetch_dnode(scn, 2069 &osp->os_projectused_dnode, zb->zb_objset, 2070 DMU_PROJECTUSED_OBJECT); 2071 } 2072 dsl_scan_prefetch_dnode(scn, 2073 &osp->os_groupused_dnode, zb->zb_objset, 2074 DMU_GROUPUSED_OBJECT); 2075 dsl_scan_prefetch_dnode(scn, 2076 &osp->os_userused_dnode, zb->zb_objset, 2077 DMU_USERUSED_OBJECT); 2078 } 2079 } 2080 2081 out: 2082 if (buf != NULL) 2083 arc_buf_destroy(buf, private); 2084 scan_prefetch_ctx_rele(spc, scn); 2085 } 2086 2087 static void 2088 dsl_scan_prefetch_thread(void *arg) 2089 { 2090 dsl_scan_t *scn = arg; 2091 spa_t *spa = scn->scn_dp->dp_spa; 2092 scan_prefetch_issue_ctx_t *spic; 2093 2094 /* loop until we are told to stop */ 2095 while (!scn->scn_prefetch_stop) { 2096 arc_flags_t flags = ARC_FLAG_NOWAIT | 2097 ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH; 2098 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD; 2099 2100 mutex_enter(&spa->spa_scrub_lock); 2101 2102 /* 2103 * Wait until we have an IO to issue and are not above our 2104 * maximum in flight limit. 2105 */ 2106 while (!scn->scn_prefetch_stop && 2107 (avl_numnodes(&scn->scn_prefetch_queue) == 0 || 2108 spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) { 2109 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 2110 } 2111 2112 /* recheck if we should stop since we waited for the cv */ 2113 if (scn->scn_prefetch_stop) { 2114 mutex_exit(&spa->spa_scrub_lock); 2115 break; 2116 } 2117 2118 /* remove the prefetch IO from the tree */ 2119 spic = avl_first(&scn->scn_prefetch_queue); 2120 spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp); 2121 avl_remove(&scn->scn_prefetch_queue, spic); 2122 2123 mutex_exit(&spa->spa_scrub_lock); 2124 2125 if (BP_IS_PROTECTED(&spic->spic_bp)) { 2126 ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE || 2127 BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET); 2128 ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0); 2129 zio_flags |= ZIO_FLAG_RAW; 2130 } 2131 2132 /* We don't need data L1 buffer since we do not prefetch L0. */ 2133 blkptr_t *bp = &spic->spic_bp; 2134 if (BP_GET_LEVEL(bp) == 1 && BP_GET_TYPE(bp) != DMU_OT_DNODE && 2135 BP_GET_TYPE(bp) != DMU_OT_OBJSET) 2136 flags |= ARC_FLAG_NO_BUF; 2137 2138 /* issue the prefetch asynchronously */ 2139 (void) arc_read(scn->scn_zio_root, spa, bp, 2140 dsl_scan_prefetch_cb, spic->spic_spc, ZIO_PRIORITY_SCRUB, 2141 zio_flags, &flags, &spic->spic_zb); 2142 2143 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t)); 2144 } 2145 2146 ASSERT(scn->scn_prefetch_stop); 2147 2148 /* free any prefetches we didn't get to complete */ 2149 mutex_enter(&spa->spa_scrub_lock); 2150 while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) { 2151 avl_remove(&scn->scn_prefetch_queue, spic); 2152 scan_prefetch_ctx_rele(spic->spic_spc, scn); 2153 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t)); 2154 } 2155 ASSERT0(avl_numnodes(&scn->scn_prefetch_queue)); 2156 mutex_exit(&spa->spa_scrub_lock); 2157 } 2158 2159 static boolean_t 2160 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp, 2161 const zbookmark_phys_t *zb) 2162 { 2163 /* 2164 * We never skip over user/group accounting objects (obj<0) 2165 */ 2166 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) && 2167 (int64_t)zb->zb_object >= 0) { 2168 /* 2169 * If we already visited this bp & everything below (in 2170 * a prior txg sync), don't bother doing it again. 2171 */ 2172 if (zbookmark_subtree_completed(dnp, zb, 2173 &scn->scn_phys.scn_bookmark)) 2174 return (B_TRUE); 2175 2176 /* 2177 * If we found the block we're trying to resume from, or 2178 * we went past it, zero it out to indicate that it's OK 2179 * to start checking for suspending again. 2180 */ 2181 if (zbookmark_subtree_tbd(dnp, zb, 2182 &scn->scn_phys.scn_bookmark)) { 2183 dprintf("resuming at %llx/%llx/%llx/%llx\n", 2184 (longlong_t)zb->zb_objset, 2185 (longlong_t)zb->zb_object, 2186 (longlong_t)zb->zb_level, 2187 (longlong_t)zb->zb_blkid); 2188 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (*zb)); 2189 } 2190 } 2191 return (B_FALSE); 2192 } 2193 2194 static void dsl_scan_visitbp(const blkptr_t *bp, const zbookmark_phys_t *zb, 2195 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn, 2196 dmu_objset_type_t ostype, dmu_tx_t *tx); 2197 inline __attribute__((always_inline)) static void dsl_scan_visitdnode( 2198 dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype, 2199 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx); 2200 2201 /* 2202 * Return nonzero on i/o error. 2203 * Return new buf to write out in *bufp. 2204 */ 2205 inline __attribute__((always_inline)) static int 2206 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype, 2207 dnode_phys_t *dnp, const blkptr_t *bp, 2208 const zbookmark_phys_t *zb, dmu_tx_t *tx) 2209 { 2210 dsl_pool_t *dp = scn->scn_dp; 2211 spa_t *spa = dp->dp_spa; 2212 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD; 2213 int err; 2214 2215 ASSERT(!BP_IS_REDACTED(bp)); 2216 2217 /* 2218 * There is an unlikely case of encountering dnodes with contradicting 2219 * dn_bonuslen and DNODE_FLAG_SPILL_BLKPTR flag before in files created 2220 * or modified before commit 4254acb was merged. As it is not possible 2221 * to know which of the two is correct, report an error. 2222 */ 2223 if (dnp != NULL && 2224 dnp->dn_bonuslen > DN_MAX_BONUS_LEN(dnp)) { 2225 scn->scn_phys.scn_errors++; 2226 spa_log_error(spa, zb, BP_GET_LOGICAL_BIRTH(bp)); 2227 return (SET_ERROR(EINVAL)); 2228 } 2229 2230 if (BP_GET_LEVEL(bp) > 0) { 2231 arc_flags_t flags = ARC_FLAG_WAIT; 2232 int i; 2233 blkptr_t *cbp; 2234 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; 2235 arc_buf_t *buf; 2236 2237 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf, 2238 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb); 2239 if (err) { 2240 scn->scn_phys.scn_errors++; 2241 return (err); 2242 } 2243 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) { 2244 zbookmark_phys_t czb; 2245 2246 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, 2247 zb->zb_level - 1, 2248 zb->zb_blkid * epb + i); 2249 dsl_scan_visitbp(cbp, &czb, dnp, 2250 ds, scn, ostype, tx); 2251 } 2252 arc_buf_destroy(buf, &buf); 2253 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { 2254 arc_flags_t flags = ARC_FLAG_WAIT; 2255 dnode_phys_t *cdnp; 2256 int i; 2257 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; 2258 arc_buf_t *buf; 2259 2260 if (BP_IS_PROTECTED(bp)) { 2261 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 2262 zio_flags |= ZIO_FLAG_RAW; 2263 } 2264 2265 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf, 2266 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb); 2267 if (err) { 2268 scn->scn_phys.scn_errors++; 2269 return (err); 2270 } 2271 for (i = 0, cdnp = buf->b_data; i < epb; 2272 i += cdnp->dn_extra_slots + 1, 2273 cdnp += cdnp->dn_extra_slots + 1) { 2274 dsl_scan_visitdnode(scn, ds, ostype, 2275 cdnp, zb->zb_blkid * epb + i, tx); 2276 } 2277 2278 arc_buf_destroy(buf, &buf); 2279 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { 2280 arc_flags_t flags = ARC_FLAG_WAIT; 2281 objset_phys_t *osp; 2282 arc_buf_t *buf; 2283 2284 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf, 2285 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb); 2286 if (err) { 2287 scn->scn_phys.scn_errors++; 2288 return (err); 2289 } 2290 2291 osp = buf->b_data; 2292 2293 dsl_scan_visitdnode(scn, ds, osp->os_type, 2294 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx); 2295 2296 if (OBJSET_BUF_HAS_USERUSED(buf)) { 2297 /* 2298 * We also always visit user/group/project accounting 2299 * objects, and never skip them, even if we are 2300 * suspending. This is necessary so that the 2301 * space deltas from this txg get integrated. 2302 */ 2303 if (OBJSET_BUF_HAS_PROJECTUSED(buf)) 2304 dsl_scan_visitdnode(scn, ds, osp->os_type, 2305 &osp->os_projectused_dnode, 2306 DMU_PROJECTUSED_OBJECT, tx); 2307 dsl_scan_visitdnode(scn, ds, osp->os_type, 2308 &osp->os_groupused_dnode, 2309 DMU_GROUPUSED_OBJECT, tx); 2310 dsl_scan_visitdnode(scn, ds, osp->os_type, 2311 &osp->os_userused_dnode, 2312 DMU_USERUSED_OBJECT, tx); 2313 } 2314 arc_buf_destroy(buf, &buf); 2315 } else if (zfs_blkptr_verify(spa, bp, 2316 BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) { 2317 /* 2318 * Sanity check the block pointer contents, this is handled 2319 * by arc_read() for the cases above. 2320 */ 2321 scn->scn_phys.scn_errors++; 2322 spa_log_error(spa, zb, BP_GET_LOGICAL_BIRTH(bp)); 2323 return (SET_ERROR(EINVAL)); 2324 } 2325 2326 return (0); 2327 } 2328 2329 inline __attribute__((always_inline)) static void 2330 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds, 2331 dmu_objset_type_t ostype, dnode_phys_t *dnp, 2332 uint64_t object, dmu_tx_t *tx) 2333 { 2334 int j; 2335 2336 for (j = 0; j < dnp->dn_nblkptr; j++) { 2337 zbookmark_phys_t czb; 2338 2339 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 2340 dnp->dn_nlevels - 1, j); 2341 dsl_scan_visitbp(&dnp->dn_blkptr[j], 2342 &czb, dnp, ds, scn, ostype, tx); 2343 } 2344 2345 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 2346 zbookmark_phys_t czb; 2347 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 2348 0, DMU_SPILL_BLKID); 2349 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp), 2350 &czb, dnp, ds, scn, ostype, tx); 2351 } 2352 } 2353 2354 /* 2355 * The arguments are in this order because mdb can only print the 2356 * first 5; we want them to be useful. 2357 */ 2358 static void 2359 dsl_scan_visitbp(const blkptr_t *bp, const zbookmark_phys_t *zb, 2360 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn, 2361 dmu_objset_type_t ostype, dmu_tx_t *tx) 2362 { 2363 dsl_pool_t *dp = scn->scn_dp; 2364 2365 if (dsl_scan_check_suspend(scn, zb)) 2366 return; 2367 2368 if (dsl_scan_check_resume(scn, dnp, zb)) 2369 return; 2370 2371 scn->scn_visited_this_txg++; 2372 2373 if (BP_IS_HOLE(bp)) { 2374 scn->scn_holes_this_txg++; 2375 return; 2376 } 2377 2378 if (BP_IS_REDACTED(bp)) { 2379 ASSERT(dsl_dataset_feature_is_active(ds, 2380 SPA_FEATURE_REDACTED_DATASETS)); 2381 return; 2382 } 2383 2384 /* 2385 * Check if this block contradicts any filesystem flags. 2386 */ 2387 spa_feature_t f = SPA_FEATURE_LARGE_BLOCKS; 2388 if (BP_GET_LSIZE(bp) > SPA_OLD_MAXBLOCKSIZE) 2389 ASSERT(dsl_dataset_feature_is_active(ds, f)); 2390 2391 f = zio_checksum_to_feature(BP_GET_CHECKSUM(bp)); 2392 if (f != SPA_FEATURE_NONE) 2393 ASSERT(dsl_dataset_feature_is_active(ds, f)); 2394 2395 f = zio_compress_to_feature(BP_GET_COMPRESS(bp)); 2396 if (f != SPA_FEATURE_NONE) 2397 ASSERT(dsl_dataset_feature_is_active(ds, f)); 2398 2399 if (BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg) { 2400 scn->scn_lt_min_this_txg++; 2401 return; 2402 } 2403 2404 if (dsl_scan_recurse(scn, ds, ostype, dnp, bp, zb, tx) != 0) 2405 return; 2406 2407 /* 2408 * If dsl_scan_ddt() has already visited this block, it will have 2409 * already done any translations or scrubbing, so don't call the 2410 * callback again. 2411 */ 2412 if (ddt_class_contains(dp->dp_spa, 2413 scn->scn_phys.scn_ddt_class_max, bp)) { 2414 scn->scn_ddt_contained_this_txg++; 2415 return; 2416 } 2417 2418 /* 2419 * If this block is from the future (after cur_max_txg), then we 2420 * are doing this on behalf of a deleted snapshot, and we will 2421 * revisit the future block on the next pass of this dataset. 2422 * Don't scan it now unless we need to because something 2423 * under it was modified. 2424 */ 2425 if (BP_GET_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) { 2426 scn->scn_gt_max_this_txg++; 2427 return; 2428 } 2429 2430 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb); 2431 } 2432 2433 static void 2434 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp, 2435 dmu_tx_t *tx) 2436 { 2437 zbookmark_phys_t zb; 2438 scan_prefetch_ctx_t *spc; 2439 2440 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 2441 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); 2442 2443 if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) { 2444 SET_BOOKMARK(&scn->scn_prefetch_bookmark, 2445 zb.zb_objset, 0, 0, 0); 2446 } else { 2447 scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark; 2448 } 2449 2450 scn->scn_objsets_visited_this_txg++; 2451 2452 spc = scan_prefetch_ctx_create(scn, NULL, FTAG); 2453 dsl_scan_prefetch(spc, bp, &zb); 2454 scan_prefetch_ctx_rele(spc, FTAG); 2455 2456 dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx); 2457 2458 dprintf_ds(ds, "finished scan%s", ""); 2459 } 2460 2461 static void 2462 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys) 2463 { 2464 if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) { 2465 if (ds->ds_is_snapshot) { 2466 /* 2467 * Note: 2468 * - scn_cur_{min,max}_txg stays the same. 2469 * - Setting the flag is not really necessary if 2470 * scn_cur_max_txg == scn_max_txg, because there 2471 * is nothing after this snapshot that we care 2472 * about. However, we set it anyway and then 2473 * ignore it when we retraverse it in 2474 * dsl_scan_visitds(). 2475 */ 2476 scn_phys->scn_bookmark.zb_objset = 2477 dsl_dataset_phys(ds)->ds_next_snap_obj; 2478 zfs_dbgmsg("destroying ds %llu on %s; currently " 2479 "traversing; reset zb_objset to %llu", 2480 (u_longlong_t)ds->ds_object, 2481 ds->ds_dir->dd_pool->dp_spa->spa_name, 2482 (u_longlong_t)dsl_dataset_phys(ds)-> 2483 ds_next_snap_obj); 2484 scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN; 2485 } else { 2486 SET_BOOKMARK(&scn_phys->scn_bookmark, 2487 ZB_DESTROYED_OBJSET, 0, 0, 0); 2488 zfs_dbgmsg("destroying ds %llu on %s; currently " 2489 "traversing; reset bookmark to -1,0,0,0", 2490 (u_longlong_t)ds->ds_object, 2491 ds->ds_dir->dd_pool->dp_spa->spa_name); 2492 } 2493 } 2494 } 2495 2496 /* 2497 * Invoked when a dataset is destroyed. We need to make sure that: 2498 * 2499 * 1) If it is the dataset that was currently being scanned, we write 2500 * a new dsl_scan_phys_t and marking the objset reference in it 2501 * as destroyed. 2502 * 2) Remove it from the work queue, if it was present. 2503 * 2504 * If the dataset was actually a snapshot, instead of marking the dataset 2505 * as destroyed, we instead substitute the next snapshot in line. 2506 */ 2507 void 2508 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx) 2509 { 2510 dsl_pool_t *dp = ds->ds_dir->dd_pool; 2511 dsl_scan_t *scn = dp->dp_scan; 2512 uint64_t mintxg; 2513 2514 if (!dsl_scan_is_running(scn)) 2515 return; 2516 2517 ds_destroyed_scn_phys(ds, &scn->scn_phys); 2518 ds_destroyed_scn_phys(ds, &scn->scn_phys_cached); 2519 2520 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) { 2521 scan_ds_queue_remove(scn, ds->ds_object); 2522 if (ds->ds_is_snapshot) 2523 scan_ds_queue_insert(scn, 2524 dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg); 2525 } 2526 2527 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 2528 ds->ds_object, &mintxg) == 0) { 2529 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1); 2530 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 2531 scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 2532 if (ds->ds_is_snapshot) { 2533 /* 2534 * We keep the same mintxg; it could be > 2535 * ds_creation_txg if the previous snapshot was 2536 * deleted too. 2537 */ 2538 VERIFY(zap_add_int_key(dp->dp_meta_objset, 2539 scn->scn_phys.scn_queue_obj, 2540 dsl_dataset_phys(ds)->ds_next_snap_obj, 2541 mintxg, tx) == 0); 2542 zfs_dbgmsg("destroying ds %llu on %s; in queue; " 2543 "replacing with %llu", 2544 (u_longlong_t)ds->ds_object, 2545 dp->dp_spa->spa_name, 2546 (u_longlong_t)dsl_dataset_phys(ds)-> 2547 ds_next_snap_obj); 2548 } else { 2549 zfs_dbgmsg("destroying ds %llu on %s; in queue; " 2550 "removing", 2551 (u_longlong_t)ds->ds_object, 2552 dp->dp_spa->spa_name); 2553 } 2554 } 2555 2556 /* 2557 * dsl_scan_sync() should be called after this, and should sync 2558 * out our changed state, but just to be safe, do it here. 2559 */ 2560 dsl_scan_sync_state(scn, tx, SYNC_CACHED); 2561 } 2562 2563 static void 2564 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark) 2565 { 2566 if (scn_bookmark->zb_objset == ds->ds_object) { 2567 scn_bookmark->zb_objset = 2568 dsl_dataset_phys(ds)->ds_prev_snap_obj; 2569 zfs_dbgmsg("snapshotting ds %llu on %s; currently traversing; " 2570 "reset zb_objset to %llu", 2571 (u_longlong_t)ds->ds_object, 2572 ds->ds_dir->dd_pool->dp_spa->spa_name, 2573 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj); 2574 } 2575 } 2576 2577 /* 2578 * Called when a dataset is snapshotted. If we were currently traversing 2579 * this snapshot, we reset our bookmark to point at the newly created 2580 * snapshot. We also modify our work queue to remove the old snapshot and 2581 * replace with the new one. 2582 */ 2583 void 2584 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx) 2585 { 2586 dsl_pool_t *dp = ds->ds_dir->dd_pool; 2587 dsl_scan_t *scn = dp->dp_scan; 2588 uint64_t mintxg; 2589 2590 if (!dsl_scan_is_running(scn)) 2591 return; 2592 2593 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0); 2594 2595 ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark); 2596 ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark); 2597 2598 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) { 2599 scan_ds_queue_remove(scn, ds->ds_object); 2600 scan_ds_queue_insert(scn, 2601 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg); 2602 } 2603 2604 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 2605 ds->ds_object, &mintxg) == 0) { 2606 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 2607 scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 2608 VERIFY(zap_add_int_key(dp->dp_meta_objset, 2609 scn->scn_phys.scn_queue_obj, 2610 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0); 2611 zfs_dbgmsg("snapshotting ds %llu on %s; in queue; " 2612 "replacing with %llu", 2613 (u_longlong_t)ds->ds_object, 2614 dp->dp_spa->spa_name, 2615 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj); 2616 } 2617 2618 dsl_scan_sync_state(scn, tx, SYNC_CACHED); 2619 } 2620 2621 static void 2622 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2, 2623 zbookmark_phys_t *scn_bookmark) 2624 { 2625 if (scn_bookmark->zb_objset == ds1->ds_object) { 2626 scn_bookmark->zb_objset = ds2->ds_object; 2627 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; " 2628 "reset zb_objset to %llu", 2629 (u_longlong_t)ds1->ds_object, 2630 ds1->ds_dir->dd_pool->dp_spa->spa_name, 2631 (u_longlong_t)ds2->ds_object); 2632 } else if (scn_bookmark->zb_objset == ds2->ds_object) { 2633 scn_bookmark->zb_objset = ds1->ds_object; 2634 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; " 2635 "reset zb_objset to %llu", 2636 (u_longlong_t)ds2->ds_object, 2637 ds2->ds_dir->dd_pool->dp_spa->spa_name, 2638 (u_longlong_t)ds1->ds_object); 2639 } 2640 } 2641 2642 /* 2643 * Called when an origin dataset and its clone are swapped. If we were 2644 * currently traversing the dataset, we need to switch to traversing the 2645 * newly promoted clone. 2646 */ 2647 void 2648 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx) 2649 { 2650 dsl_pool_t *dp = ds1->ds_dir->dd_pool; 2651 dsl_scan_t *scn = dp->dp_scan; 2652 uint64_t mintxg1, mintxg2; 2653 boolean_t ds1_queued, ds2_queued; 2654 2655 if (!dsl_scan_is_running(scn)) 2656 return; 2657 2658 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark); 2659 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark); 2660 2661 /* 2662 * Handle the in-memory scan queue. 2663 */ 2664 ds1_queued = scan_ds_queue_contains(scn, ds1->ds_object, &mintxg1); 2665 ds2_queued = scan_ds_queue_contains(scn, ds2->ds_object, &mintxg2); 2666 2667 /* Sanity checking. */ 2668 if (ds1_queued) { 2669 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg); 2670 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg); 2671 } 2672 if (ds2_queued) { 2673 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg); 2674 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg); 2675 } 2676 2677 if (ds1_queued && ds2_queued) { 2678 /* 2679 * If both are queued, we don't need to do anything. 2680 * The swapping code below would not handle this case correctly, 2681 * since we can't insert ds2 if it is already there. That's 2682 * because scan_ds_queue_insert() prohibits a duplicate insert 2683 * and panics. 2684 */ 2685 } else if (ds1_queued) { 2686 scan_ds_queue_remove(scn, ds1->ds_object); 2687 scan_ds_queue_insert(scn, ds2->ds_object, mintxg1); 2688 } else if (ds2_queued) { 2689 scan_ds_queue_remove(scn, ds2->ds_object); 2690 scan_ds_queue_insert(scn, ds1->ds_object, mintxg2); 2691 } 2692 2693 /* 2694 * Handle the on-disk scan queue. 2695 * The on-disk state is an out-of-date version of the in-memory state, 2696 * so the in-memory and on-disk values for ds1_queued and ds2_queued may 2697 * be different. Therefore we need to apply the swap logic to the 2698 * on-disk state independently of the in-memory state. 2699 */ 2700 ds1_queued = zap_lookup_int_key(dp->dp_meta_objset, 2701 scn->scn_phys.scn_queue_obj, ds1->ds_object, &mintxg1) == 0; 2702 ds2_queued = zap_lookup_int_key(dp->dp_meta_objset, 2703 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg2) == 0; 2704 2705 /* Sanity checking. */ 2706 if (ds1_queued) { 2707 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg); 2708 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg); 2709 } 2710 if (ds2_queued) { 2711 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg); 2712 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg); 2713 } 2714 2715 if (ds1_queued && ds2_queued) { 2716 /* 2717 * If both are queued, we don't need to do anything. 2718 * Alternatively, we could check for EEXIST from 2719 * zap_add_int_key() and back out to the original state, but 2720 * that would be more work than checking for this case upfront. 2721 */ 2722 } else if (ds1_queued) { 2723 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset, 2724 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx)); 2725 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset, 2726 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg1, tx)); 2727 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; " 2728 "replacing with %llu", 2729 (u_longlong_t)ds1->ds_object, 2730 dp->dp_spa->spa_name, 2731 (u_longlong_t)ds2->ds_object); 2732 } else if (ds2_queued) { 2733 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset, 2734 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx)); 2735 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset, 2736 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg2, tx)); 2737 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; " 2738 "replacing with %llu", 2739 (u_longlong_t)ds2->ds_object, 2740 dp->dp_spa->spa_name, 2741 (u_longlong_t)ds1->ds_object); 2742 } 2743 2744 dsl_scan_sync_state(scn, tx, SYNC_CACHED); 2745 } 2746 2747 static int 2748 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 2749 { 2750 uint64_t originobj = *(uint64_t *)arg; 2751 dsl_dataset_t *ds; 2752 int err; 2753 dsl_scan_t *scn = dp->dp_scan; 2754 2755 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj) 2756 return (0); 2757 2758 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 2759 if (err) 2760 return (err); 2761 2762 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) { 2763 dsl_dataset_t *prev; 2764 err = dsl_dataset_hold_obj(dp, 2765 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev); 2766 2767 dsl_dataset_rele(ds, FTAG); 2768 if (err) 2769 return (err); 2770 ds = prev; 2771 } 2772 mutex_enter(&scn->scn_queue_lock); 2773 scan_ds_queue_insert(scn, ds->ds_object, 2774 dsl_dataset_phys(ds)->ds_prev_snap_txg); 2775 mutex_exit(&scn->scn_queue_lock); 2776 dsl_dataset_rele(ds, FTAG); 2777 return (0); 2778 } 2779 2780 static void 2781 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx) 2782 { 2783 dsl_pool_t *dp = scn->scn_dp; 2784 dsl_dataset_t *ds; 2785 2786 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 2787 2788 if (scn->scn_phys.scn_cur_min_txg >= 2789 scn->scn_phys.scn_max_txg) { 2790 /* 2791 * This can happen if this snapshot was created after the 2792 * scan started, and we already completed a previous snapshot 2793 * that was created after the scan started. This snapshot 2794 * only references blocks with: 2795 * 2796 * birth < our ds_creation_txg 2797 * cur_min_txg is no less than ds_creation_txg. 2798 * We have already visited these blocks. 2799 * or 2800 * birth > scn_max_txg 2801 * The scan requested not to visit these blocks. 2802 * 2803 * Subsequent snapshots (and clones) can reference our 2804 * blocks, or blocks with even higher birth times. 2805 * Therefore we do not need to visit them either, 2806 * so we do not add them to the work queue. 2807 * 2808 * Note that checking for cur_min_txg >= cur_max_txg 2809 * is not sufficient, because in that case we may need to 2810 * visit subsequent snapshots. This happens when min_txg > 0, 2811 * which raises cur_min_txg. In this case we will visit 2812 * this dataset but skip all of its blocks, because the 2813 * rootbp's birth time is < cur_min_txg. Then we will 2814 * add the next snapshots/clones to the work queue. 2815 */ 2816 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP); 2817 dsl_dataset_name(ds, dsname); 2818 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because " 2819 "cur_min_txg (%llu) >= max_txg (%llu)", 2820 (longlong_t)dsobj, dsname, 2821 (longlong_t)scn->scn_phys.scn_cur_min_txg, 2822 (longlong_t)scn->scn_phys.scn_max_txg); 2823 kmem_free(dsname, MAXNAMELEN); 2824 2825 goto out; 2826 } 2827 2828 /* 2829 * Only the ZIL in the head (non-snapshot) is valid. Even though 2830 * snapshots can have ZIL block pointers (which may be the same 2831 * BP as in the head), they must be ignored. In addition, $ORIGIN 2832 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't 2833 * need to look for a ZIL in it either. So we traverse the ZIL here, 2834 * rather than in scan_recurse(), because the regular snapshot 2835 * block-sharing rules don't apply to it. 2836 */ 2837 if (!dsl_dataset_is_snapshot(ds) && 2838 (dp->dp_origin_snap == NULL || 2839 ds->ds_dir != dp->dp_origin_snap->ds_dir)) { 2840 objset_t *os; 2841 if (dmu_objset_from_ds(ds, &os) != 0) { 2842 goto out; 2843 } 2844 dsl_scan_zil(dp, &os->os_zil_header); 2845 } 2846 2847 /* 2848 * Iterate over the bps in this ds. 2849 */ 2850 dmu_buf_will_dirty(ds->ds_dbuf, tx); 2851 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 2852 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx); 2853 rrw_exit(&ds->ds_bp_rwlock, FTAG); 2854 2855 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP); 2856 dsl_dataset_name(ds, dsname); 2857 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; " 2858 "suspending=%u", 2859 (longlong_t)dsobj, dsname, 2860 (longlong_t)scn->scn_phys.scn_cur_min_txg, 2861 (longlong_t)scn->scn_phys.scn_cur_max_txg, 2862 (int)scn->scn_suspending); 2863 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN); 2864 2865 if (scn->scn_suspending) 2866 goto out; 2867 2868 /* 2869 * We've finished this pass over this dataset. 2870 */ 2871 2872 /* 2873 * If we did not completely visit this dataset, do another pass. 2874 */ 2875 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) { 2876 zfs_dbgmsg("incomplete pass on %s; visiting again", 2877 dp->dp_spa->spa_name); 2878 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN; 2879 scan_ds_queue_insert(scn, ds->ds_object, 2880 scn->scn_phys.scn_cur_max_txg); 2881 goto out; 2882 } 2883 2884 /* 2885 * Add descendant datasets to work queue. 2886 */ 2887 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) { 2888 scan_ds_queue_insert(scn, 2889 dsl_dataset_phys(ds)->ds_next_snap_obj, 2890 dsl_dataset_phys(ds)->ds_creation_txg); 2891 } 2892 if (dsl_dataset_phys(ds)->ds_num_children > 1) { 2893 boolean_t usenext = B_FALSE; 2894 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) { 2895 uint64_t count; 2896 /* 2897 * A bug in a previous version of the code could 2898 * cause upgrade_clones_cb() to not set 2899 * ds_next_snap_obj when it should, leading to a 2900 * missing entry. Therefore we can only use the 2901 * next_clones_obj when its count is correct. 2902 */ 2903 int err = zap_count(dp->dp_meta_objset, 2904 dsl_dataset_phys(ds)->ds_next_clones_obj, &count); 2905 if (err == 0 && 2906 count == dsl_dataset_phys(ds)->ds_num_children - 1) 2907 usenext = B_TRUE; 2908 } 2909 2910 if (usenext) { 2911 zap_cursor_t zc; 2912 zap_attribute_t *za = zap_attribute_alloc(); 2913 for (zap_cursor_init(&zc, dp->dp_meta_objset, 2914 dsl_dataset_phys(ds)->ds_next_clones_obj); 2915 zap_cursor_retrieve(&zc, za) == 0; 2916 (void) zap_cursor_advance(&zc)) { 2917 scan_ds_queue_insert(scn, 2918 zfs_strtonum(za->za_name, NULL), 2919 dsl_dataset_phys(ds)->ds_creation_txg); 2920 } 2921 zap_cursor_fini(&zc); 2922 zap_attribute_free(za); 2923 } else { 2924 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 2925 enqueue_clones_cb, &ds->ds_object, 2926 DS_FIND_CHILDREN)); 2927 } 2928 } 2929 2930 out: 2931 dsl_dataset_rele(ds, FTAG); 2932 } 2933 2934 static int 2935 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 2936 { 2937 (void) arg; 2938 dsl_dataset_t *ds; 2939 int err; 2940 dsl_scan_t *scn = dp->dp_scan; 2941 2942 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 2943 if (err) 2944 return (err); 2945 2946 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) { 2947 dsl_dataset_t *prev; 2948 err = dsl_dataset_hold_obj(dp, 2949 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev); 2950 if (err) { 2951 dsl_dataset_rele(ds, FTAG); 2952 return (err); 2953 } 2954 2955 /* 2956 * If this is a clone, we don't need to worry about it for now. 2957 */ 2958 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) { 2959 dsl_dataset_rele(ds, FTAG); 2960 dsl_dataset_rele(prev, FTAG); 2961 return (0); 2962 } 2963 dsl_dataset_rele(ds, FTAG); 2964 ds = prev; 2965 } 2966 2967 mutex_enter(&scn->scn_queue_lock); 2968 scan_ds_queue_insert(scn, ds->ds_object, 2969 dsl_dataset_phys(ds)->ds_prev_snap_txg); 2970 mutex_exit(&scn->scn_queue_lock); 2971 dsl_dataset_rele(ds, FTAG); 2972 return (0); 2973 } 2974 2975 void 2976 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum, 2977 ddt_t *ddt, ddt_lightweight_entry_t *ddlwe, dmu_tx_t *tx) 2978 { 2979 (void) tx; 2980 const ddt_key_t *ddk = &ddlwe->ddlwe_key; 2981 blkptr_t bp; 2982 zbookmark_phys_t zb = { 0 }; 2983 2984 if (!dsl_scan_is_running(scn)) 2985 return; 2986 2987 /* 2988 * This function is special because it is the only thing 2989 * that can add scan_io_t's to the vdev scan queues from 2990 * outside dsl_scan_sync(). For the most part this is ok 2991 * as long as it is called from within syncing context. 2992 * However, dsl_scan_sync() expects that no new sio's will 2993 * be added between when all the work for a scan is done 2994 * and the next txg when the scan is actually marked as 2995 * completed. This check ensures we do not issue new sio's 2996 * during this period. 2997 */ 2998 if (scn->scn_done_txg != 0) 2999 return; 3000 3001 for (int p = 0; p < DDT_NPHYS(ddt); p++) { 3002 ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p); 3003 uint64_t phys_birth = ddt_phys_birth(&ddlwe->ddlwe_phys, v); 3004 3005 if (phys_birth == 0 || phys_birth > scn->scn_phys.scn_max_txg) 3006 continue; 3007 ddt_bp_create(checksum, ddk, &ddlwe->ddlwe_phys, v, &bp); 3008 3009 scn->scn_visited_this_txg++; 3010 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb); 3011 } 3012 } 3013 3014 /* 3015 * Scrub/dedup interaction. 3016 * 3017 * If there are N references to a deduped block, we don't want to scrub it 3018 * N times -- ideally, we should scrub it exactly once. 3019 * 3020 * We leverage the fact that the dde's replication class (ddt_class_t) 3021 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest 3022 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order. 3023 * 3024 * To prevent excess scrubbing, the scrub begins by walking the DDT 3025 * to find all blocks with refcnt > 1, and scrubs each of these once. 3026 * Since there are two replication classes which contain blocks with 3027 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first. 3028 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1. 3029 * 3030 * There would be nothing more to say if a block's refcnt couldn't change 3031 * during a scrub, but of course it can so we must account for changes 3032 * in a block's replication class. 3033 * 3034 * Here's an example of what can occur: 3035 * 3036 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1 3037 * when visited during the top-down scrub phase, it will be scrubbed twice. 3038 * This negates our scrub optimization, but is otherwise harmless. 3039 * 3040 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1 3041 * on each visit during the top-down scrub phase, it will never be scrubbed. 3042 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's 3043 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to 3044 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1 3045 * while a scrub is in progress, it scrubs the block right then. 3046 */ 3047 static void 3048 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx) 3049 { 3050 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark; 3051 ddt_lightweight_entry_t ddlwe = {0}; 3052 int error; 3053 uint64_t n = 0; 3054 3055 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &ddlwe)) == 0) { 3056 ddt_t *ddt; 3057 3058 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max) 3059 break; 3060 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n", 3061 (longlong_t)ddb->ddb_class, 3062 (longlong_t)ddb->ddb_type, 3063 (longlong_t)ddb->ddb_checksum, 3064 (longlong_t)ddb->ddb_cursor); 3065 3066 /* There should be no pending changes to the dedup table */ 3067 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum]; 3068 ASSERT(avl_first(&ddt->ddt_tree) == NULL); 3069 3070 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, ddt, &ddlwe, tx); 3071 n++; 3072 3073 if (dsl_scan_check_suspend(scn, NULL)) 3074 break; 3075 } 3076 3077 if (error == EAGAIN) { 3078 dsl_scan_check_suspend(scn, NULL); 3079 error = 0; 3080 3081 zfs_dbgmsg("waiting for ddt to become ready for scan " 3082 "on %s with class_max = %u; suspending=%u", 3083 scn->scn_dp->dp_spa->spa_name, 3084 (int)scn->scn_phys.scn_ddt_class_max, 3085 (int)scn->scn_suspending); 3086 } else 3087 zfs_dbgmsg("scanned %llu ddt entries on %s with " 3088 "class_max = %u; suspending=%u", (longlong_t)n, 3089 scn->scn_dp->dp_spa->spa_name, 3090 (int)scn->scn_phys.scn_ddt_class_max, 3091 (int)scn->scn_suspending); 3092 3093 ASSERT(error == 0 || error == ENOENT); 3094 ASSERT(error != ENOENT || 3095 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max); 3096 } 3097 3098 static uint64_t 3099 dsl_scan_ds_maxtxg(dsl_dataset_t *ds) 3100 { 3101 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg; 3102 if (ds->ds_is_snapshot) 3103 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg)); 3104 return (smt); 3105 } 3106 3107 static void 3108 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx) 3109 { 3110 scan_ds_t *sds; 3111 dsl_pool_t *dp = scn->scn_dp; 3112 3113 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 3114 scn->scn_phys.scn_ddt_class_max) { 3115 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 3116 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 3117 dsl_scan_ddt(scn, tx); 3118 if (scn->scn_suspending) 3119 return; 3120 } 3121 3122 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) { 3123 /* First do the MOS & ORIGIN */ 3124 3125 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 3126 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 3127 dsl_scan_visit_rootbp(scn, NULL, 3128 &dp->dp_meta_rootbp, tx); 3129 if (scn->scn_suspending) 3130 return; 3131 3132 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) { 3133 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 3134 enqueue_cb, NULL, DS_FIND_CHILDREN)); 3135 } else { 3136 dsl_scan_visitds(scn, 3137 dp->dp_origin_snap->ds_object, tx); 3138 } 3139 ASSERT(!scn->scn_suspending); 3140 } else if (scn->scn_phys.scn_bookmark.zb_objset != 3141 ZB_DESTROYED_OBJSET) { 3142 uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset; 3143 /* 3144 * If we were suspended, continue from here. Note if the 3145 * ds we were suspended on was deleted, the zb_objset may 3146 * be -1, so we will skip this and find a new objset 3147 * below. 3148 */ 3149 dsl_scan_visitds(scn, dsobj, tx); 3150 if (scn->scn_suspending) 3151 return; 3152 } 3153 3154 /* 3155 * In case we suspended right at the end of the ds, zero the 3156 * bookmark so we don't think that we're still trying to resume. 3157 */ 3158 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (zbookmark_phys_t)); 3159 3160 /* 3161 * Keep pulling things out of the dataset avl queue. Updates to the 3162 * persistent zap-object-as-queue happen only at checkpoints. 3163 */ 3164 while ((sds = avl_first(&scn->scn_queue)) != NULL) { 3165 dsl_dataset_t *ds; 3166 uint64_t dsobj = sds->sds_dsobj; 3167 uint64_t txg = sds->sds_txg; 3168 3169 /* dequeue and free the ds from the queue */ 3170 scan_ds_queue_remove(scn, dsobj); 3171 sds = NULL; 3172 3173 /* set up min / max txg */ 3174 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 3175 if (txg != 0) { 3176 scn->scn_phys.scn_cur_min_txg = 3177 MAX(scn->scn_phys.scn_min_txg, txg); 3178 } else { 3179 scn->scn_phys.scn_cur_min_txg = 3180 MAX(scn->scn_phys.scn_min_txg, 3181 dsl_dataset_phys(ds)->ds_prev_snap_txg); 3182 } 3183 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds); 3184 dsl_dataset_rele(ds, FTAG); 3185 3186 dsl_scan_visitds(scn, dsobj, tx); 3187 if (scn->scn_suspending) 3188 return; 3189 } 3190 3191 /* No more objsets to fetch, we're done */ 3192 scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET; 3193 ASSERT0(scn->scn_suspending); 3194 } 3195 3196 static uint64_t 3197 dsl_scan_count_data_disks(spa_t *spa) 3198 { 3199 vdev_t *rvd = spa->spa_root_vdev; 3200 uint64_t i, leaves = 0; 3201 3202 for (i = 0; i < rvd->vdev_children; i++) { 3203 vdev_t *vd = rvd->vdev_child[i]; 3204 if (vd->vdev_islog || vd->vdev_isspare || vd->vdev_isl2cache) 3205 continue; 3206 leaves += vdev_get_ndisks(vd) - vdev_get_nparity(vd); 3207 } 3208 return (leaves); 3209 } 3210 3211 static void 3212 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp) 3213 { 3214 int i; 3215 uint64_t cur_size = 0; 3216 3217 for (i = 0; i < BP_GET_NDVAS(bp); i++) { 3218 cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]); 3219 } 3220 3221 q->q_total_zio_size_this_txg += cur_size; 3222 q->q_zios_this_txg++; 3223 } 3224 3225 static void 3226 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start, 3227 uint64_t end) 3228 { 3229 q->q_total_seg_size_this_txg += end - start; 3230 q->q_segs_this_txg++; 3231 } 3232 3233 static boolean_t 3234 scan_io_queue_check_suspend(dsl_scan_t *scn) 3235 { 3236 /* See comment in dsl_scan_check_suspend() */ 3237 uint64_t curr_time_ns = gethrtime(); 3238 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time; 3239 uint64_t sync_time_ns = curr_time_ns - 3240 scn->scn_dp->dp_spa->spa_sync_starttime; 3241 uint64_t dirty_min_bytes = zfs_dirty_data_max * 3242 zfs_vdev_async_write_active_min_dirty_percent / 100; 3243 uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ? 3244 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms; 3245 3246 return ((NSEC2MSEC(scan_time_ns) > mintime && 3247 (scn->scn_dp->dp_dirty_total >= dirty_min_bytes || 3248 txg_sync_waiting(scn->scn_dp) || 3249 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) || 3250 spa_shutting_down(scn->scn_dp->dp_spa)); 3251 } 3252 3253 /* 3254 * Given a list of scan_io_t's in io_list, this issues the I/Os out to 3255 * disk. This consumes the io_list and frees the scan_io_t's. This is 3256 * called when emptying queues, either when we're up against the memory 3257 * limit or when we have finished scanning. Returns B_TRUE if we stopped 3258 * processing the list before we finished. Any sios that were not issued 3259 * will remain in the io_list. 3260 */ 3261 static boolean_t 3262 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list) 3263 { 3264 dsl_scan_t *scn = queue->q_scn; 3265 scan_io_t *sio; 3266 boolean_t suspended = B_FALSE; 3267 3268 while ((sio = list_head(io_list)) != NULL) { 3269 blkptr_t bp; 3270 3271 if (scan_io_queue_check_suspend(scn)) { 3272 suspended = B_TRUE; 3273 break; 3274 } 3275 3276 sio2bp(sio, &bp); 3277 scan_exec_io(scn->scn_dp, &bp, sio->sio_flags, 3278 &sio->sio_zb, queue); 3279 (void) list_remove_head(io_list); 3280 scan_io_queues_update_zio_stats(queue, &bp); 3281 sio_free(sio); 3282 } 3283 return (suspended); 3284 } 3285 3286 /* 3287 * This function removes sios from an IO queue which reside within a given 3288 * zfs_range_seg_t and inserts them (in offset order) into a list. Note that 3289 * we only ever return a maximum of 32 sios at once. If there are more sios 3290 * to process within this segment that did not make it onto the list we 3291 * return B_TRUE and otherwise B_FALSE. 3292 */ 3293 static boolean_t 3294 scan_io_queue_gather(dsl_scan_io_queue_t *queue, zfs_range_seg_t *rs, 3295 list_t *list) 3296 { 3297 scan_io_t *srch_sio, *sio, *next_sio; 3298 avl_index_t idx; 3299 uint_t num_sios = 0; 3300 int64_t bytes_issued = 0; 3301 3302 ASSERT(rs != NULL); 3303 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock)); 3304 3305 srch_sio = sio_alloc(1); 3306 srch_sio->sio_nr_dvas = 1; 3307 SIO_SET_OFFSET(srch_sio, zfs_rs_get_start(rs, queue->q_exts_by_addr)); 3308 3309 /* 3310 * The exact start of the extent might not contain any matching zios, 3311 * so if that's the case, examine the next one in the tree. 3312 */ 3313 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx); 3314 sio_free(srch_sio); 3315 3316 if (sio == NULL) 3317 sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER); 3318 3319 while (sio != NULL && SIO_GET_OFFSET(sio) < zfs_rs_get_end(rs, 3320 queue->q_exts_by_addr) && num_sios <= 32) { 3321 ASSERT3U(SIO_GET_OFFSET(sio), >=, zfs_rs_get_start(rs, 3322 queue->q_exts_by_addr)); 3323 ASSERT3U(SIO_GET_END_OFFSET(sio), <=, zfs_rs_get_end(rs, 3324 queue->q_exts_by_addr)); 3325 3326 next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio); 3327 avl_remove(&queue->q_sios_by_addr, sio); 3328 if (avl_is_empty(&queue->q_sios_by_addr)) 3329 atomic_add_64(&queue->q_scn->scn_queues_pending, -1); 3330 queue->q_sio_memused -= SIO_GET_MUSED(sio); 3331 3332 bytes_issued += SIO_GET_ASIZE(sio); 3333 num_sios++; 3334 list_insert_tail(list, sio); 3335 sio = next_sio; 3336 } 3337 3338 /* 3339 * We limit the number of sios we process at once to 32 to avoid 3340 * biting off more than we can chew. If we didn't take everything 3341 * in the segment we update it to reflect the work we were able to 3342 * complete. Otherwise, we remove it from the range tree entirely. 3343 */ 3344 if (sio != NULL && SIO_GET_OFFSET(sio) < zfs_rs_get_end(rs, 3345 queue->q_exts_by_addr)) { 3346 zfs_range_tree_adjust_fill(queue->q_exts_by_addr, rs, 3347 -bytes_issued); 3348 zfs_range_tree_resize_segment(queue->q_exts_by_addr, rs, 3349 SIO_GET_OFFSET(sio), zfs_rs_get_end(rs, 3350 queue->q_exts_by_addr) - SIO_GET_OFFSET(sio)); 3351 queue->q_last_ext_addr = SIO_GET_OFFSET(sio); 3352 return (B_TRUE); 3353 } else { 3354 uint64_t rstart = zfs_rs_get_start(rs, queue->q_exts_by_addr); 3355 uint64_t rend = zfs_rs_get_end(rs, queue->q_exts_by_addr); 3356 zfs_range_tree_remove(queue->q_exts_by_addr, rstart, rend - 3357 rstart); 3358 queue->q_last_ext_addr = -1; 3359 return (B_FALSE); 3360 } 3361 } 3362 3363 /* 3364 * This is called from the queue emptying thread and selects the next 3365 * extent from which we are to issue I/Os. The behavior of this function 3366 * depends on the state of the scan, the current memory consumption and 3367 * whether or not we are performing a scan shutdown. 3368 * 1) We select extents in an elevator algorithm (LBA-order) if the scan 3369 * needs to perform a checkpoint 3370 * 2) We select the largest available extent if we are up against the 3371 * memory limit. 3372 * 3) Otherwise we don't select any extents. 3373 */ 3374 static zfs_range_seg_t * 3375 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue) 3376 { 3377 dsl_scan_t *scn = queue->q_scn; 3378 zfs_range_tree_t *rt = queue->q_exts_by_addr; 3379 3380 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock)); 3381 ASSERT(scn->scn_is_sorted); 3382 3383 if (!scn->scn_checkpointing && !scn->scn_clearing) 3384 return (NULL); 3385 3386 /* 3387 * During normal clearing, we want to issue our largest segments 3388 * first, keeping IO as sequential as possible, and leaving the 3389 * smaller extents for later with the hope that they might eventually 3390 * grow to larger sequential segments. However, when the scan is 3391 * checkpointing, no new extents will be added to the sorting queue, 3392 * so the way we are sorted now is as good as it will ever get. 3393 * In this case, we instead switch to issuing extents in LBA order. 3394 */ 3395 if ((zfs_scan_issue_strategy < 1 && scn->scn_checkpointing) || 3396 zfs_scan_issue_strategy == 1) 3397 return (zfs_range_tree_first(rt)); 3398 3399 /* 3400 * Try to continue previous extent if it is not completed yet. After 3401 * shrink in scan_io_queue_gather() it may no longer be the best, but 3402 * otherwise we leave shorter remnant every txg. 3403 */ 3404 uint64_t start; 3405 uint64_t size = 1ULL << rt->rt_shift; 3406 zfs_range_seg_t *addr_rs; 3407 if (queue->q_last_ext_addr != -1) { 3408 start = queue->q_last_ext_addr; 3409 addr_rs = zfs_range_tree_find(rt, start, size); 3410 if (addr_rs != NULL) 3411 return (addr_rs); 3412 } 3413 3414 /* 3415 * Nothing to continue, so find new best extent. 3416 */ 3417 uint64_t *v = zfs_btree_first(&queue->q_exts_by_size, NULL); 3418 if (v == NULL) 3419 return (NULL); 3420 queue->q_last_ext_addr = start = *v << rt->rt_shift; 3421 3422 /* 3423 * We need to get the original entry in the by_addr tree so we can 3424 * modify it. 3425 */ 3426 addr_rs = zfs_range_tree_find(rt, start, size); 3427 ASSERT3P(addr_rs, !=, NULL); 3428 ASSERT3U(zfs_rs_get_start(addr_rs, rt), ==, start); 3429 ASSERT3U(zfs_rs_get_end(addr_rs, rt), >, start); 3430 return (addr_rs); 3431 } 3432 3433 static void 3434 scan_io_queues_run_one(void *arg) 3435 { 3436 dsl_scan_io_queue_t *queue = arg; 3437 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock; 3438 boolean_t suspended = B_FALSE; 3439 zfs_range_seg_t *rs; 3440 scan_io_t *sio; 3441 zio_t *zio; 3442 list_t sio_list; 3443 3444 ASSERT(queue->q_scn->scn_is_sorted); 3445 3446 list_create(&sio_list, sizeof (scan_io_t), 3447 offsetof(scan_io_t, sio_nodes.sio_list_node)); 3448 zio = zio_null(queue->q_scn->scn_zio_root, queue->q_scn->scn_dp->dp_spa, 3449 NULL, NULL, NULL, ZIO_FLAG_CANFAIL); 3450 mutex_enter(q_lock); 3451 queue->q_zio = zio; 3452 3453 /* Calculate maximum in-flight bytes for this vdev. */ 3454 queue->q_maxinflight_bytes = MAX(1, zfs_scan_vdev_limit * 3455 (vdev_get_ndisks(queue->q_vd) - vdev_get_nparity(queue->q_vd))); 3456 3457 /* reset per-queue scan statistics for this txg */ 3458 queue->q_total_seg_size_this_txg = 0; 3459 queue->q_segs_this_txg = 0; 3460 queue->q_total_zio_size_this_txg = 0; 3461 queue->q_zios_this_txg = 0; 3462 3463 /* loop until we run out of time or sios */ 3464 while ((rs = scan_io_queue_fetch_ext(queue)) != NULL) { 3465 uint64_t seg_start = 0, seg_end = 0; 3466 boolean_t more_left; 3467 3468 ASSERT(list_is_empty(&sio_list)); 3469 3470 /* loop while we still have sios left to process in this rs */ 3471 do { 3472 scan_io_t *first_sio, *last_sio; 3473 3474 /* 3475 * We have selected which extent needs to be 3476 * processed next. Gather up the corresponding sios. 3477 */ 3478 more_left = scan_io_queue_gather(queue, rs, &sio_list); 3479 ASSERT(!list_is_empty(&sio_list)); 3480 first_sio = list_head(&sio_list); 3481 last_sio = list_tail(&sio_list); 3482 3483 seg_end = SIO_GET_END_OFFSET(last_sio); 3484 if (seg_start == 0) 3485 seg_start = SIO_GET_OFFSET(first_sio); 3486 3487 /* 3488 * Issuing sios can take a long time so drop the 3489 * queue lock. The sio queue won't be updated by 3490 * other threads since we're in syncing context so 3491 * we can be sure that our trees will remain exactly 3492 * as we left them. 3493 */ 3494 mutex_exit(q_lock); 3495 suspended = scan_io_queue_issue(queue, &sio_list); 3496 mutex_enter(q_lock); 3497 3498 if (suspended) 3499 break; 3500 } while (more_left); 3501 3502 /* update statistics for debugging purposes */ 3503 scan_io_queues_update_seg_stats(queue, seg_start, seg_end); 3504 3505 if (suspended) 3506 break; 3507 } 3508 3509 /* 3510 * If we were suspended in the middle of processing, 3511 * requeue any unfinished sios and exit. 3512 */ 3513 while ((sio = list_remove_head(&sio_list)) != NULL) 3514 scan_io_queue_insert_impl(queue, sio); 3515 3516 queue->q_zio = NULL; 3517 mutex_exit(q_lock); 3518 zio_nowait(zio); 3519 list_destroy(&sio_list); 3520 } 3521 3522 /* 3523 * Performs an emptying run on all scan queues in the pool. This just 3524 * punches out one thread per top-level vdev, each of which processes 3525 * only that vdev's scan queue. We can parallelize the I/O here because 3526 * we know that each queue's I/Os only affect its own top-level vdev. 3527 * 3528 * This function waits for the queue runs to complete, and must be 3529 * called from dsl_scan_sync (or in general, syncing context). 3530 */ 3531 static void 3532 scan_io_queues_run(dsl_scan_t *scn) 3533 { 3534 spa_t *spa = scn->scn_dp->dp_spa; 3535 3536 ASSERT(scn->scn_is_sorted); 3537 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 3538 3539 if (scn->scn_queues_pending == 0) 3540 return; 3541 3542 if (scn->scn_taskq == NULL) { 3543 int nthreads = spa->spa_root_vdev->vdev_children; 3544 3545 /* 3546 * We need to make this taskq *always* execute as many 3547 * threads in parallel as we have top-level vdevs and no 3548 * less, otherwise strange serialization of the calls to 3549 * scan_io_queues_run_one can occur during spa_sync runs 3550 * and that significantly impacts performance. 3551 */ 3552 scn->scn_taskq = taskq_create("dsl_scan_iss", nthreads, 3553 minclsyspri, nthreads, nthreads, TASKQ_PREPOPULATE); 3554 } 3555 3556 for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) { 3557 vdev_t *vd = spa->spa_root_vdev->vdev_child[i]; 3558 3559 mutex_enter(&vd->vdev_scan_io_queue_lock); 3560 if (vd->vdev_scan_io_queue != NULL) { 3561 VERIFY(taskq_dispatch(scn->scn_taskq, 3562 scan_io_queues_run_one, vd->vdev_scan_io_queue, 3563 TQ_SLEEP) != TASKQID_INVALID); 3564 } 3565 mutex_exit(&vd->vdev_scan_io_queue_lock); 3566 } 3567 3568 /* 3569 * Wait for the queues to finish issuing their IOs for this run 3570 * before we return. There may still be IOs in flight at this 3571 * point. 3572 */ 3573 taskq_wait(scn->scn_taskq); 3574 } 3575 3576 static boolean_t 3577 dsl_scan_async_block_should_pause(dsl_scan_t *scn) 3578 { 3579 uint64_t elapsed_nanosecs; 3580 3581 if (zfs_recover) 3582 return (B_FALSE); 3583 3584 if (zfs_async_block_max_blocks != 0 && 3585 scn->scn_visited_this_txg >= zfs_async_block_max_blocks) { 3586 return (B_TRUE); 3587 } 3588 3589 if (zfs_max_async_dedup_frees != 0 && 3590 scn->scn_dedup_frees_this_txg >= zfs_max_async_dedup_frees) { 3591 return (B_TRUE); 3592 } 3593 3594 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 3595 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 3596 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms && 3597 txg_sync_waiting(scn->scn_dp)) || 3598 spa_shutting_down(scn->scn_dp->dp_spa)); 3599 } 3600 3601 static int 3602 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 3603 { 3604 dsl_scan_t *scn = arg; 3605 3606 if (!scn->scn_is_bptree || 3607 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) { 3608 if (dsl_scan_async_block_should_pause(scn)) 3609 return (SET_ERROR(ERESTART)); 3610 } 3611 3612 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa, 3613 dmu_tx_get_txg(tx), bp, 0)); 3614 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, 3615 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp), 3616 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); 3617 scn->scn_visited_this_txg++; 3618 if (BP_GET_DEDUP(bp)) 3619 scn->scn_dedup_frees_this_txg++; 3620 return (0); 3621 } 3622 3623 static void 3624 dsl_scan_update_stats(dsl_scan_t *scn) 3625 { 3626 spa_t *spa = scn->scn_dp->dp_spa; 3627 uint64_t i; 3628 uint64_t seg_size_total = 0, zio_size_total = 0; 3629 uint64_t seg_count_total = 0, zio_count_total = 0; 3630 3631 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) { 3632 vdev_t *vd = spa->spa_root_vdev->vdev_child[i]; 3633 dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue; 3634 3635 if (queue == NULL) 3636 continue; 3637 3638 seg_size_total += queue->q_total_seg_size_this_txg; 3639 zio_size_total += queue->q_total_zio_size_this_txg; 3640 seg_count_total += queue->q_segs_this_txg; 3641 zio_count_total += queue->q_zios_this_txg; 3642 } 3643 3644 if (seg_count_total == 0 || zio_count_total == 0) { 3645 scn->scn_avg_seg_size_this_txg = 0; 3646 scn->scn_avg_zio_size_this_txg = 0; 3647 scn->scn_segs_this_txg = 0; 3648 scn->scn_zios_this_txg = 0; 3649 return; 3650 } 3651 3652 scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total; 3653 scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total; 3654 scn->scn_segs_this_txg = seg_count_total; 3655 scn->scn_zios_this_txg = zio_count_total; 3656 } 3657 3658 static int 3659 bpobj_dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 3660 dmu_tx_t *tx) 3661 { 3662 ASSERT(!bp_freed); 3663 return (dsl_scan_free_block_cb(arg, bp, tx)); 3664 } 3665 3666 static int 3667 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 3668 dmu_tx_t *tx) 3669 { 3670 ASSERT(!bp_freed); 3671 dsl_scan_t *scn = arg; 3672 const dva_t *dva = &bp->blk_dva[0]; 3673 3674 if (dsl_scan_async_block_should_pause(scn)) 3675 return (SET_ERROR(ERESTART)); 3676 3677 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa, 3678 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva), 3679 DVA_GET_ASIZE(dva), tx); 3680 scn->scn_visited_this_txg++; 3681 return (0); 3682 } 3683 3684 boolean_t 3685 dsl_scan_active(dsl_scan_t *scn) 3686 { 3687 spa_t *spa = scn->scn_dp->dp_spa; 3688 uint64_t used = 0, comp, uncomp; 3689 boolean_t clones_left; 3690 3691 if (spa->spa_load_state != SPA_LOAD_NONE) 3692 return (B_FALSE); 3693 if (spa_shutting_down(spa)) 3694 return (B_FALSE); 3695 if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) || 3696 (scn->scn_async_destroying && !scn->scn_async_stalled)) 3697 return (B_TRUE); 3698 3699 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 3700 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj, 3701 &used, &comp, &uncomp); 3702 } 3703 clones_left = spa_livelist_delete_check(spa); 3704 return ((used != 0) || (clones_left)); 3705 } 3706 3707 boolean_t 3708 dsl_errorscrub_active(dsl_scan_t *scn) 3709 { 3710 spa_t *spa = scn->scn_dp->dp_spa; 3711 if (spa->spa_load_state != SPA_LOAD_NONE) 3712 return (B_FALSE); 3713 if (spa_shutting_down(spa)) 3714 return (B_FALSE); 3715 if (dsl_errorscrubbing(scn->scn_dp)) 3716 return (B_TRUE); 3717 return (B_FALSE); 3718 } 3719 3720 static boolean_t 3721 dsl_scan_check_deferred(vdev_t *vd) 3722 { 3723 boolean_t need_resilver = B_FALSE; 3724 3725 for (int c = 0; c < vd->vdev_children; c++) { 3726 need_resilver |= 3727 dsl_scan_check_deferred(vd->vdev_child[c]); 3728 } 3729 3730 if (!vdev_is_concrete(vd) || vd->vdev_aux || 3731 !vd->vdev_ops->vdev_op_leaf) 3732 return (need_resilver); 3733 3734 if (!vd->vdev_resilver_deferred) 3735 need_resilver = B_TRUE; 3736 3737 return (need_resilver); 3738 } 3739 3740 static boolean_t 3741 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize, 3742 uint64_t phys_birth) 3743 { 3744 vdev_t *vd; 3745 3746 vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva)); 3747 3748 if (vd->vdev_ops == &vdev_indirect_ops) { 3749 /* 3750 * The indirect vdev can point to multiple 3751 * vdevs. For simplicity, always create 3752 * the resilver zio_t. zio_vdev_io_start() 3753 * will bypass the child resilver i/o's if 3754 * they are on vdevs that don't have DTL's. 3755 */ 3756 return (B_TRUE); 3757 } 3758 3759 if (DVA_GET_GANG(dva)) { 3760 /* 3761 * Gang members may be spread across multiple 3762 * vdevs, so the best estimate we have is the 3763 * scrub range, which has already been checked. 3764 * XXX -- it would be better to change our 3765 * allocation policy to ensure that all 3766 * gang members reside on the same vdev. 3767 */ 3768 return (B_TRUE); 3769 } 3770 3771 /* 3772 * Check if the top-level vdev must resilver this offset. 3773 * When the offset does not intersect with a dirty leaf DTL 3774 * then it may be possible to skip the resilver IO. The psize 3775 * is provided instead of asize to simplify the check for RAIDZ. 3776 */ 3777 if (!vdev_dtl_need_resilver(vd, dva, psize, phys_birth)) 3778 return (B_FALSE); 3779 3780 /* 3781 * Check that this top-level vdev has a device under it which 3782 * is resilvering and is not deferred. 3783 */ 3784 if (!dsl_scan_check_deferred(vd)) 3785 return (B_FALSE); 3786 3787 return (B_TRUE); 3788 } 3789 3790 static int 3791 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx) 3792 { 3793 dsl_scan_t *scn = dp->dp_scan; 3794 spa_t *spa = dp->dp_spa; 3795 int err = 0; 3796 3797 if (spa_suspend_async_destroy(spa)) 3798 return (0); 3799 3800 if (zfs_free_bpobj_enabled && 3801 spa_version(spa) >= SPA_VERSION_DEADLISTS) { 3802 scn->scn_is_bptree = B_FALSE; 3803 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms; 3804 scn->scn_zio_root = zio_root(spa, NULL, 3805 NULL, ZIO_FLAG_MUSTSUCCEED); 3806 err = bpobj_iterate(&dp->dp_free_bpobj, 3807 bpobj_dsl_scan_free_block_cb, scn, tx); 3808 VERIFY0(zio_wait(scn->scn_zio_root)); 3809 scn->scn_zio_root = NULL; 3810 3811 if (err != 0 && err != ERESTART) 3812 zfs_panic_recover("error %u from bpobj_iterate()", err); 3813 } 3814 3815 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) { 3816 ASSERT(scn->scn_async_destroying); 3817 scn->scn_is_bptree = B_TRUE; 3818 scn->scn_zio_root = zio_root(spa, NULL, 3819 NULL, ZIO_FLAG_MUSTSUCCEED); 3820 err = bptree_iterate(dp->dp_meta_objset, 3821 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx); 3822 VERIFY0(zio_wait(scn->scn_zio_root)); 3823 scn->scn_zio_root = NULL; 3824 3825 if (err == EIO || err == ECKSUM) { 3826 err = 0; 3827 } else if (err != 0 && err != ERESTART) { 3828 zfs_panic_recover("error %u from " 3829 "traverse_dataset_destroyed()", err); 3830 } 3831 3832 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) { 3833 /* finished; deactivate async destroy feature */ 3834 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx); 3835 ASSERT(!spa_feature_is_active(spa, 3836 SPA_FEATURE_ASYNC_DESTROY)); 3837 VERIFY0(zap_remove(dp->dp_meta_objset, 3838 DMU_POOL_DIRECTORY_OBJECT, 3839 DMU_POOL_BPTREE_OBJ, tx)); 3840 VERIFY0(bptree_free(dp->dp_meta_objset, 3841 dp->dp_bptree_obj, tx)); 3842 dp->dp_bptree_obj = 0; 3843 scn->scn_async_destroying = B_FALSE; 3844 scn->scn_async_stalled = B_FALSE; 3845 } else { 3846 /* 3847 * If we didn't make progress, mark the async 3848 * destroy as stalled, so that we will not initiate 3849 * a spa_sync() on its behalf. Note that we only 3850 * check this if we are not finished, because if the 3851 * bptree had no blocks for us to visit, we can 3852 * finish without "making progress". 3853 */ 3854 scn->scn_async_stalled = 3855 (scn->scn_visited_this_txg == 0); 3856 } 3857 } 3858 if (scn->scn_visited_this_txg) { 3859 zfs_dbgmsg("freed %llu blocks in %llums from " 3860 "free_bpobj/bptree on %s in txg %llu; err=%u", 3861 (longlong_t)scn->scn_visited_this_txg, 3862 (longlong_t) 3863 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time), 3864 spa->spa_name, (longlong_t)tx->tx_txg, err); 3865 scn->scn_visited_this_txg = 0; 3866 scn->scn_dedup_frees_this_txg = 0; 3867 3868 /* 3869 * Write out changes to the DDT and the BRT that may be required 3870 * as a result of the blocks freed. This ensures that the DDT 3871 * and the BRT are clean when a scrub/resilver runs. 3872 */ 3873 ddt_sync(spa, tx->tx_txg); 3874 brt_sync(spa, tx->tx_txg); 3875 } 3876 if (err != 0) 3877 return (err); 3878 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying && 3879 zfs_free_leak_on_eio && 3880 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 || 3881 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 || 3882 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) { 3883 /* 3884 * We have finished background destroying, but there is still 3885 * some space left in the dp_free_dir. Transfer this leaked 3886 * space to the dp_leak_dir. 3887 */ 3888 if (dp->dp_leak_dir == NULL) { 3889 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 3890 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 3891 LEAK_DIR_NAME, tx); 3892 VERIFY0(dsl_pool_open_special_dir(dp, 3893 LEAK_DIR_NAME, &dp->dp_leak_dir)); 3894 rrw_exit(&dp->dp_config_rwlock, FTAG); 3895 } 3896 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD, 3897 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes, 3898 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes, 3899 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx); 3900 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD, 3901 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes, 3902 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes, 3903 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx); 3904 } 3905 3906 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying && 3907 !spa_livelist_delete_check(spa)) { 3908 /* finished; verify that space accounting went to zero */ 3909 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes); 3910 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes); 3911 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes); 3912 } 3913 3914 spa_notify_waiters(spa); 3915 3916 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj), 3917 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 3918 DMU_POOL_OBSOLETE_BPOBJ)); 3919 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) { 3920 ASSERT(spa_feature_is_active(dp->dp_spa, 3921 SPA_FEATURE_OBSOLETE_COUNTS)); 3922 3923 scn->scn_is_bptree = B_FALSE; 3924 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms; 3925 err = bpobj_iterate(&dp->dp_obsolete_bpobj, 3926 dsl_scan_obsolete_block_cb, scn, tx); 3927 if (err != 0 && err != ERESTART) 3928 zfs_panic_recover("error %u from bpobj_iterate()", err); 3929 3930 if (bpobj_is_empty(&dp->dp_obsolete_bpobj)) 3931 dsl_pool_destroy_obsolete_bpobj(dp, tx); 3932 } 3933 return (0); 3934 } 3935 3936 static void 3937 name_to_bookmark(char *buf, zbookmark_phys_t *zb) 3938 { 3939 zb->zb_objset = zfs_strtonum(buf, &buf); 3940 ASSERT(*buf == ':'); 3941 zb->zb_object = zfs_strtonum(buf + 1, &buf); 3942 ASSERT(*buf == ':'); 3943 zb->zb_level = (int)zfs_strtonum(buf + 1, &buf); 3944 ASSERT(*buf == ':'); 3945 zb->zb_blkid = zfs_strtonum(buf + 1, &buf); 3946 ASSERT(*buf == '\0'); 3947 } 3948 3949 static void 3950 name_to_object(char *buf, uint64_t *obj) 3951 { 3952 *obj = zfs_strtonum(buf, &buf); 3953 ASSERT(*buf == '\0'); 3954 } 3955 3956 static void 3957 read_by_block_level(dsl_scan_t *scn, zbookmark_phys_t zb) 3958 { 3959 dsl_pool_t *dp = scn->scn_dp; 3960 dsl_dataset_t *ds; 3961 objset_t *os; 3962 if (dsl_dataset_hold_obj(dp, zb.zb_objset, FTAG, &ds) != 0) 3963 return; 3964 3965 if (dmu_objset_from_ds(ds, &os) != 0) { 3966 dsl_dataset_rele(ds, FTAG); 3967 return; 3968 } 3969 3970 /* 3971 * If the key is not loaded dbuf_dnode_findbp() will error out with 3972 * EACCES. However in that case dnode_hold() will eventually call 3973 * dbuf_read()->zio_wait() which may call spa_log_error(). This will 3974 * lead to a deadlock due to us holding the mutex spa_errlist_lock. 3975 * Avoid this by checking here if the keys are loaded, if not return. 3976 * If the keys are not loaded the head_errlog feature is meaningless 3977 * as we cannot figure out the birth txg of the block pointer. 3978 */ 3979 if (dsl_dataset_get_keystatus(ds->ds_dir) == 3980 ZFS_KEYSTATUS_UNAVAILABLE) { 3981 dsl_dataset_rele(ds, FTAG); 3982 return; 3983 } 3984 3985 dnode_t *dn; 3986 blkptr_t bp; 3987 3988 if (dnode_hold(os, zb.zb_object, FTAG, &dn) != 0) { 3989 dsl_dataset_rele(ds, FTAG); 3990 return; 3991 } 3992 3993 rw_enter(&dn->dn_struct_rwlock, RW_READER); 3994 int error = dbuf_dnode_findbp(dn, zb.zb_level, zb.zb_blkid, &bp, NULL, 3995 NULL); 3996 3997 if (error) { 3998 rw_exit(&dn->dn_struct_rwlock); 3999 dnode_rele(dn, FTAG); 4000 dsl_dataset_rele(ds, FTAG); 4001 return; 4002 } 4003 4004 if (!error && BP_IS_HOLE(&bp)) { 4005 rw_exit(&dn->dn_struct_rwlock); 4006 dnode_rele(dn, FTAG); 4007 dsl_dataset_rele(ds, FTAG); 4008 return; 4009 } 4010 4011 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | 4012 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCRUB; 4013 4014 /* If it's an intent log block, failure is expected. */ 4015 if (zb.zb_level == ZB_ZIL_LEVEL) 4016 zio_flags |= ZIO_FLAG_SPECULATIVE; 4017 4018 ASSERT(!BP_IS_EMBEDDED(&bp)); 4019 scan_exec_io(dp, &bp, zio_flags, &zb, NULL); 4020 rw_exit(&dn->dn_struct_rwlock); 4021 dnode_rele(dn, FTAG); 4022 dsl_dataset_rele(ds, FTAG); 4023 } 4024 4025 /* 4026 * We keep track of the scrubbed error blocks in "count". This will be used 4027 * when deciding whether we exceeded zfs_scrub_error_blocks_per_txg. This 4028 * function is modelled after check_filesystem(). 4029 */ 4030 static int 4031 scrub_filesystem(spa_t *spa, uint64_t fs, zbookmark_err_phys_t *zep, 4032 int *count) 4033 { 4034 dsl_dataset_t *ds; 4035 dsl_pool_t *dp = spa->spa_dsl_pool; 4036 dsl_scan_t *scn = dp->dp_scan; 4037 4038 int error = dsl_dataset_hold_obj(dp, fs, FTAG, &ds); 4039 if (error != 0) 4040 return (error); 4041 4042 uint64_t latest_txg; 4043 uint64_t txg_to_consider = spa->spa_syncing_txg; 4044 boolean_t check_snapshot = B_TRUE; 4045 4046 error = find_birth_txg(ds, zep, &latest_txg); 4047 4048 /* 4049 * If find_birth_txg() errors out, then err on the side of caution and 4050 * proceed. In worst case scenario scrub all objects. If zep->zb_birth 4051 * is 0 (e.g. in case of encryption with unloaded keys) also proceed to 4052 * scrub all objects. 4053 */ 4054 if (error == 0 && zep->zb_birth == latest_txg) { 4055 /* Block neither free nor re written. */ 4056 zbookmark_phys_t zb; 4057 zep_to_zb(fs, zep, &zb); 4058 scn->scn_zio_root = zio_root(spa, NULL, NULL, 4059 ZIO_FLAG_CANFAIL); 4060 /* We have already acquired the config lock for spa */ 4061 read_by_block_level(scn, zb); 4062 4063 (void) zio_wait(scn->scn_zio_root); 4064 scn->scn_zio_root = NULL; 4065 4066 scn->errorscrub_phys.dep_examined++; 4067 scn->errorscrub_phys.dep_to_examine--; 4068 (*count)++; 4069 if ((*count) == zfs_scrub_error_blocks_per_txg || 4070 dsl_error_scrub_check_suspend(scn, &zb)) { 4071 dsl_dataset_rele(ds, FTAG); 4072 return (SET_ERROR(EFAULT)); 4073 } 4074 4075 check_snapshot = B_FALSE; 4076 } else if (error == 0) { 4077 txg_to_consider = latest_txg; 4078 } 4079 4080 /* 4081 * Retrieve the number of snapshots if the dataset is not a snapshot. 4082 */ 4083 uint64_t snap_count = 0; 4084 if (dsl_dataset_phys(ds)->ds_snapnames_zapobj != 0) { 4085 4086 error = zap_count(spa->spa_meta_objset, 4087 dsl_dataset_phys(ds)->ds_snapnames_zapobj, &snap_count); 4088 4089 if (error != 0) { 4090 dsl_dataset_rele(ds, FTAG); 4091 return (error); 4092 } 4093 } 4094 4095 if (snap_count == 0) { 4096 /* Filesystem without snapshots. */ 4097 dsl_dataset_rele(ds, FTAG); 4098 return (0); 4099 } 4100 4101 uint64_t snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj; 4102 uint64_t snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg; 4103 4104 dsl_dataset_rele(ds, FTAG); 4105 4106 /* Check only snapshots created from this file system. */ 4107 while (snap_obj != 0 && zep->zb_birth < snap_obj_txg && 4108 snap_obj_txg <= txg_to_consider) { 4109 4110 error = dsl_dataset_hold_obj(dp, snap_obj, FTAG, &ds); 4111 if (error != 0) 4112 return (error); 4113 4114 if (dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj != fs) { 4115 snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj; 4116 snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg; 4117 dsl_dataset_rele(ds, FTAG); 4118 continue; 4119 } 4120 4121 boolean_t affected = B_TRUE; 4122 if (check_snapshot) { 4123 uint64_t blk_txg; 4124 error = find_birth_txg(ds, zep, &blk_txg); 4125 4126 /* 4127 * Scrub the snapshot also when zb_birth == 0 or when 4128 * find_birth_txg() returns an error. 4129 */ 4130 affected = (error == 0 && zep->zb_birth == blk_txg) || 4131 (error != 0) || (zep->zb_birth == 0); 4132 } 4133 4134 /* Scrub snapshots. */ 4135 if (affected) { 4136 zbookmark_phys_t zb; 4137 zep_to_zb(snap_obj, zep, &zb); 4138 scn->scn_zio_root = zio_root(spa, NULL, NULL, 4139 ZIO_FLAG_CANFAIL); 4140 /* We have already acquired the config lock for spa */ 4141 read_by_block_level(scn, zb); 4142 4143 (void) zio_wait(scn->scn_zio_root); 4144 scn->scn_zio_root = NULL; 4145 4146 scn->errorscrub_phys.dep_examined++; 4147 scn->errorscrub_phys.dep_to_examine--; 4148 (*count)++; 4149 if ((*count) == zfs_scrub_error_blocks_per_txg || 4150 dsl_error_scrub_check_suspend(scn, &zb)) { 4151 dsl_dataset_rele(ds, FTAG); 4152 return (EFAULT); 4153 } 4154 } 4155 snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg; 4156 snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj; 4157 dsl_dataset_rele(ds, FTAG); 4158 } 4159 return (0); 4160 } 4161 4162 void 4163 dsl_errorscrub_sync(dsl_pool_t *dp, dmu_tx_t *tx) 4164 { 4165 spa_t *spa = dp->dp_spa; 4166 dsl_scan_t *scn = dp->dp_scan; 4167 4168 /* 4169 * Only process scans in sync pass 1. 4170 */ 4171 4172 if (spa_sync_pass(spa) > 1) 4173 return; 4174 4175 /* 4176 * If the spa is shutting down, then stop scanning. This will 4177 * ensure that the scan does not dirty any new data during the 4178 * shutdown phase. 4179 */ 4180 if (spa_shutting_down(spa)) 4181 return; 4182 4183 if (!dsl_errorscrub_active(scn) || dsl_errorscrub_is_paused(scn)) { 4184 return; 4185 } 4186 4187 if (dsl_scan_resilvering(scn->scn_dp)) { 4188 /* cancel the error scrub if resilver started */ 4189 dsl_scan_cancel(scn->scn_dp); 4190 return; 4191 } 4192 4193 spa->spa_scrub_active = B_TRUE; 4194 scn->scn_sync_start_time = gethrtime(); 4195 4196 /* 4197 * zfs_scan_suspend_progress can be set to disable scrub progress. 4198 * See more detailed comment in dsl_scan_sync(). 4199 */ 4200 if (zfs_scan_suspend_progress) { 4201 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time; 4202 int mintime = zfs_scrub_min_time_ms; 4203 4204 while (zfs_scan_suspend_progress && 4205 !txg_sync_waiting(scn->scn_dp) && 4206 !spa_shutting_down(scn->scn_dp->dp_spa) && 4207 NSEC2MSEC(scan_time_ns) < mintime) { 4208 delay(hz); 4209 scan_time_ns = gethrtime() - scn->scn_sync_start_time; 4210 } 4211 return; 4212 } 4213 4214 int i = 0; 4215 zap_attribute_t *za; 4216 zbookmark_phys_t *zb; 4217 boolean_t limit_exceeded = B_FALSE; 4218 4219 za = zap_attribute_alloc(); 4220 zb = kmem_zalloc(sizeof (zbookmark_phys_t), KM_SLEEP); 4221 4222 if (!spa_feature_is_enabled(spa, SPA_FEATURE_HEAD_ERRLOG)) { 4223 for (; zap_cursor_retrieve(&scn->errorscrub_cursor, za) == 0; 4224 zap_cursor_advance(&scn->errorscrub_cursor)) { 4225 name_to_bookmark(za->za_name, zb); 4226 4227 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 4228 NULL, ZIO_FLAG_CANFAIL); 4229 dsl_pool_config_enter(dp, FTAG); 4230 read_by_block_level(scn, *zb); 4231 dsl_pool_config_exit(dp, FTAG); 4232 4233 (void) zio_wait(scn->scn_zio_root); 4234 scn->scn_zio_root = NULL; 4235 4236 scn->errorscrub_phys.dep_examined += 1; 4237 scn->errorscrub_phys.dep_to_examine -= 1; 4238 i++; 4239 if (i == zfs_scrub_error_blocks_per_txg || 4240 dsl_error_scrub_check_suspend(scn, zb)) { 4241 limit_exceeded = B_TRUE; 4242 break; 4243 } 4244 } 4245 4246 if (!limit_exceeded) 4247 dsl_errorscrub_done(scn, B_TRUE, tx); 4248 4249 dsl_errorscrub_sync_state(scn, tx); 4250 zap_attribute_free(za); 4251 kmem_free(zb, sizeof (*zb)); 4252 return; 4253 } 4254 4255 int error = 0; 4256 for (; zap_cursor_retrieve(&scn->errorscrub_cursor, za) == 0; 4257 zap_cursor_advance(&scn->errorscrub_cursor)) { 4258 4259 zap_cursor_t *head_ds_cursor; 4260 zap_attribute_t *head_ds_attr; 4261 zbookmark_err_phys_t head_ds_block; 4262 4263 head_ds_cursor = kmem_zalloc(sizeof (zap_cursor_t), KM_SLEEP); 4264 head_ds_attr = zap_attribute_alloc(); 4265 4266 uint64_t head_ds_err_obj = za->za_first_integer; 4267 uint64_t head_ds; 4268 name_to_object(za->za_name, &head_ds); 4269 boolean_t config_held = B_FALSE; 4270 uint64_t top_affected_fs; 4271 4272 for (zap_cursor_init(head_ds_cursor, spa->spa_meta_objset, 4273 head_ds_err_obj); zap_cursor_retrieve(head_ds_cursor, 4274 head_ds_attr) == 0; zap_cursor_advance(head_ds_cursor)) { 4275 4276 name_to_errphys(head_ds_attr->za_name, &head_ds_block); 4277 4278 /* 4279 * In case we are called from spa_sync the pool 4280 * config is already held. 4281 */ 4282 if (!dsl_pool_config_held(dp)) { 4283 dsl_pool_config_enter(dp, FTAG); 4284 config_held = B_TRUE; 4285 } 4286 4287 error = find_top_affected_fs(spa, 4288 head_ds, &head_ds_block, &top_affected_fs); 4289 if (error) 4290 break; 4291 4292 error = scrub_filesystem(spa, top_affected_fs, 4293 &head_ds_block, &i); 4294 4295 if (error == SET_ERROR(EFAULT)) { 4296 limit_exceeded = B_TRUE; 4297 break; 4298 } 4299 } 4300 4301 zap_cursor_fini(head_ds_cursor); 4302 kmem_free(head_ds_cursor, sizeof (*head_ds_cursor)); 4303 zap_attribute_free(head_ds_attr); 4304 4305 if (config_held) 4306 dsl_pool_config_exit(dp, FTAG); 4307 } 4308 4309 zap_attribute_free(za); 4310 kmem_free(zb, sizeof (*zb)); 4311 if (!limit_exceeded) 4312 dsl_errorscrub_done(scn, B_TRUE, tx); 4313 4314 dsl_errorscrub_sync_state(scn, tx); 4315 } 4316 4317 /* 4318 * This is the primary entry point for scans that is called from syncing 4319 * context. Scans must happen entirely during syncing context so that we 4320 * can guarantee that blocks we are currently scanning will not change out 4321 * from under us. While a scan is active, this function controls how quickly 4322 * transaction groups proceed, instead of the normal handling provided by 4323 * txg_sync_thread(). 4324 */ 4325 void 4326 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx) 4327 { 4328 int err = 0; 4329 dsl_scan_t *scn = dp->dp_scan; 4330 spa_t *spa = dp->dp_spa; 4331 state_sync_type_t sync_type = SYNC_OPTIONAL; 4332 int restart_early = 0; 4333 4334 if (spa->spa_resilver_deferred) { 4335 uint64_t to_issue, issued; 4336 4337 if (!spa_feature_is_active(dp->dp_spa, 4338 SPA_FEATURE_RESILVER_DEFER)) 4339 spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx); 4340 4341 /* 4342 * See print_scan_scrub_resilver_status() issued/total_i 4343 * @ cmd/zpool/zpool_main.c 4344 */ 4345 to_issue = 4346 scn->scn_phys.scn_to_examine - scn->scn_phys.scn_skipped; 4347 issued = 4348 scn->scn_issued_before_pass + spa->spa_scan_pass_issued; 4349 restart_early = 4350 zfs_resilver_disable_defer || 4351 (issued < (to_issue * zfs_resilver_defer_percent / 100)); 4352 } 4353 4354 /* 4355 * Only process scans in sync pass 1. 4356 */ 4357 if (spa_sync_pass(spa) > 1) 4358 return; 4359 4360 4361 /* 4362 * Check for scn_restart_txg before checking spa_load_state, so 4363 * that we can restart an old-style scan while the pool is being 4364 * imported (see dsl_scan_init). We also restart scans if there 4365 * is a deferred resilver and the user has manually disabled 4366 * deferred resilvers via zfs_resilver_disable_defer, or if the 4367 * current scan progress is below zfs_resilver_defer_percent. 4368 */ 4369 if (dsl_scan_restarting(scn, tx) || restart_early) { 4370 setup_sync_arg_t setup_sync_arg = { 4371 .func = POOL_SCAN_SCRUB, 4372 .txgstart = 0, 4373 .txgend = 0, 4374 }; 4375 dsl_scan_done(scn, B_FALSE, tx); 4376 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) 4377 setup_sync_arg.func = POOL_SCAN_RESILVER; 4378 zfs_dbgmsg("restarting scan func=%u on %s txg=%llu early=%d", 4379 setup_sync_arg.func, dp->dp_spa->spa_name, 4380 (longlong_t)tx->tx_txg, restart_early); 4381 dsl_scan_setup_sync(&setup_sync_arg, tx); 4382 } 4383 4384 /* 4385 * If the spa is shutting down, then stop scanning. This will 4386 * ensure that the scan does not dirty any new data during the 4387 * shutdown phase. 4388 */ 4389 if (spa_shutting_down(spa)) 4390 return; 4391 4392 /* 4393 * If the scan is inactive due to a stalled async destroy, try again. 4394 */ 4395 if (!scn->scn_async_stalled && !dsl_scan_active(scn)) 4396 return; 4397 4398 /* reset scan statistics */ 4399 scn->scn_visited_this_txg = 0; 4400 scn->scn_dedup_frees_this_txg = 0; 4401 scn->scn_holes_this_txg = 0; 4402 scn->scn_lt_min_this_txg = 0; 4403 scn->scn_gt_max_this_txg = 0; 4404 scn->scn_ddt_contained_this_txg = 0; 4405 scn->scn_objsets_visited_this_txg = 0; 4406 scn->scn_avg_seg_size_this_txg = 0; 4407 scn->scn_segs_this_txg = 0; 4408 scn->scn_avg_zio_size_this_txg = 0; 4409 scn->scn_zios_this_txg = 0; 4410 scn->scn_suspending = B_FALSE; 4411 scn->scn_sync_start_time = gethrtime(); 4412 spa->spa_scrub_active = B_TRUE; 4413 4414 /* 4415 * First process the async destroys. If we suspend, don't do 4416 * any scrubbing or resilvering. This ensures that there are no 4417 * async destroys while we are scanning, so the scan code doesn't 4418 * have to worry about traversing it. It is also faster to free the 4419 * blocks than to scrub them. 4420 */ 4421 err = dsl_process_async_destroys(dp, tx); 4422 if (err != 0) 4423 return; 4424 4425 if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn)) 4426 return; 4427 4428 /* 4429 * Wait a few txgs after importing to begin scanning so that 4430 * we can get the pool imported quickly. 4431 */ 4432 if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS) 4433 return; 4434 4435 /* 4436 * zfs_scan_suspend_progress can be set to disable scan progress. 4437 * We don't want to spin the txg_sync thread, so we add a delay 4438 * here to simulate the time spent doing a scan. This is mostly 4439 * useful for testing and debugging. 4440 */ 4441 if (zfs_scan_suspend_progress) { 4442 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time; 4443 uint_t mintime = (scn->scn_phys.scn_func == 4444 POOL_SCAN_RESILVER) ? zfs_resilver_min_time_ms : 4445 zfs_scrub_min_time_ms; 4446 4447 while (zfs_scan_suspend_progress && 4448 !txg_sync_waiting(scn->scn_dp) && 4449 !spa_shutting_down(scn->scn_dp->dp_spa) && 4450 NSEC2MSEC(scan_time_ns) < mintime) { 4451 delay(hz); 4452 scan_time_ns = gethrtime() - scn->scn_sync_start_time; 4453 } 4454 return; 4455 } 4456 4457 /* 4458 * Disabled by default, set zfs_scan_report_txgs to report 4459 * average performance over the last zfs_scan_report_txgs TXGs. 4460 */ 4461 if (zfs_scan_report_txgs != 0 && 4462 tx->tx_txg % zfs_scan_report_txgs == 0) { 4463 scn->scn_issued_before_pass += spa->spa_scan_pass_issued; 4464 spa_scan_stat_init(spa); 4465 } 4466 4467 /* 4468 * It is possible to switch from unsorted to sorted at any time, 4469 * but afterwards the scan will remain sorted unless reloaded from 4470 * a checkpoint after a reboot. 4471 */ 4472 if (!zfs_scan_legacy) { 4473 scn->scn_is_sorted = B_TRUE; 4474 if (scn->scn_last_checkpoint == 0) 4475 scn->scn_last_checkpoint = ddi_get_lbolt(); 4476 } 4477 4478 /* 4479 * For sorted scans, determine what kind of work we will be doing 4480 * this txg based on our memory limitations and whether or not we 4481 * need to perform a checkpoint. 4482 */ 4483 if (scn->scn_is_sorted) { 4484 /* 4485 * If we are over our checkpoint interval, set scn_clearing 4486 * so that we can begin checkpointing immediately. The 4487 * checkpoint allows us to save a consistent bookmark 4488 * representing how much data we have scrubbed so far. 4489 * Otherwise, use the memory limit to determine if we should 4490 * scan for metadata or start issue scrub IOs. We accumulate 4491 * metadata until we hit our hard memory limit at which point 4492 * we issue scrub IOs until we are at our soft memory limit. 4493 */ 4494 if (scn->scn_checkpointing || 4495 ddi_get_lbolt() - scn->scn_last_checkpoint > 4496 SEC_TO_TICK(zfs_scan_checkpoint_intval)) { 4497 if (!scn->scn_checkpointing) 4498 zfs_dbgmsg("begin scan checkpoint for %s", 4499 spa->spa_name); 4500 4501 scn->scn_checkpointing = B_TRUE; 4502 scn->scn_clearing = B_TRUE; 4503 } else { 4504 boolean_t should_clear = dsl_scan_should_clear(scn); 4505 if (should_clear && !scn->scn_clearing) { 4506 zfs_dbgmsg("begin scan clearing for %s", 4507 spa->spa_name); 4508 scn->scn_clearing = B_TRUE; 4509 } else if (!should_clear && scn->scn_clearing) { 4510 zfs_dbgmsg("finish scan clearing for %s", 4511 spa->spa_name); 4512 scn->scn_clearing = B_FALSE; 4513 } 4514 } 4515 } else { 4516 ASSERT0(scn->scn_checkpointing); 4517 ASSERT0(scn->scn_clearing); 4518 } 4519 4520 if (!scn->scn_clearing && scn->scn_done_txg == 0) { 4521 /* Need to scan metadata for more blocks to scrub */ 4522 dsl_scan_phys_t *scnp = &scn->scn_phys; 4523 taskqid_t prefetch_tqid; 4524 4525 /* 4526 * Calculate the max number of in-flight bytes for pool-wide 4527 * scanning operations (minimum 1MB, maximum 1/4 of arc_c_max). 4528 * Limits for the issuing phase are done per top-level vdev and 4529 * are handled separately. 4530 */ 4531 scn->scn_maxinflight_bytes = MIN(arc_c_max / 4, MAX(1ULL << 20, 4532 zfs_scan_vdev_limit * dsl_scan_count_data_disks(spa))); 4533 4534 if (scnp->scn_ddt_bookmark.ddb_class <= 4535 scnp->scn_ddt_class_max) { 4536 ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark)); 4537 zfs_dbgmsg("doing scan sync for %s txg %llu; " 4538 "ddt bm=%llu/%llu/%llu/%llx", 4539 spa->spa_name, 4540 (longlong_t)tx->tx_txg, 4541 (longlong_t)scnp->scn_ddt_bookmark.ddb_class, 4542 (longlong_t)scnp->scn_ddt_bookmark.ddb_type, 4543 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum, 4544 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor); 4545 } else { 4546 zfs_dbgmsg("doing scan sync for %s txg %llu; " 4547 "bm=%llu/%llu/%llu/%llu", 4548 spa->spa_name, 4549 (longlong_t)tx->tx_txg, 4550 (longlong_t)scnp->scn_bookmark.zb_objset, 4551 (longlong_t)scnp->scn_bookmark.zb_object, 4552 (longlong_t)scnp->scn_bookmark.zb_level, 4553 (longlong_t)scnp->scn_bookmark.zb_blkid); 4554 } 4555 4556 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 4557 NULL, ZIO_FLAG_CANFAIL); 4558 4559 scn->scn_prefetch_stop = B_FALSE; 4560 prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq, 4561 dsl_scan_prefetch_thread, scn, TQ_SLEEP); 4562 ASSERT(prefetch_tqid != TASKQID_INVALID); 4563 4564 dsl_pool_config_enter(dp, FTAG); 4565 dsl_scan_visit(scn, tx); 4566 dsl_pool_config_exit(dp, FTAG); 4567 4568 mutex_enter(&dp->dp_spa->spa_scrub_lock); 4569 scn->scn_prefetch_stop = B_TRUE; 4570 cv_broadcast(&spa->spa_scrub_io_cv); 4571 mutex_exit(&dp->dp_spa->spa_scrub_lock); 4572 4573 taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid); 4574 (void) zio_wait(scn->scn_zio_root); 4575 scn->scn_zio_root = NULL; 4576 4577 zfs_dbgmsg("scan visited %llu blocks of %s in %llums " 4578 "(%llu os's, %llu holes, %llu < mintxg, " 4579 "%llu in ddt, %llu > maxtxg)", 4580 (longlong_t)scn->scn_visited_this_txg, 4581 spa->spa_name, 4582 (longlong_t)NSEC2MSEC(gethrtime() - 4583 scn->scn_sync_start_time), 4584 (longlong_t)scn->scn_objsets_visited_this_txg, 4585 (longlong_t)scn->scn_holes_this_txg, 4586 (longlong_t)scn->scn_lt_min_this_txg, 4587 (longlong_t)scn->scn_ddt_contained_this_txg, 4588 (longlong_t)scn->scn_gt_max_this_txg); 4589 4590 if (!scn->scn_suspending) { 4591 ASSERT0(avl_numnodes(&scn->scn_queue)); 4592 scn->scn_done_txg = tx->tx_txg + 1; 4593 if (scn->scn_is_sorted) { 4594 scn->scn_checkpointing = B_TRUE; 4595 scn->scn_clearing = B_TRUE; 4596 scn->scn_issued_before_pass += 4597 spa->spa_scan_pass_issued; 4598 spa_scan_stat_init(spa); 4599 } 4600 zfs_dbgmsg("scan complete for %s txg %llu", 4601 spa->spa_name, 4602 (longlong_t)tx->tx_txg); 4603 } 4604 } else if (scn->scn_is_sorted && scn->scn_queues_pending != 0) { 4605 ASSERT(scn->scn_clearing); 4606 4607 /* need to issue scrubbing IOs from per-vdev queues */ 4608 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 4609 NULL, ZIO_FLAG_CANFAIL); 4610 scan_io_queues_run(scn); 4611 (void) zio_wait(scn->scn_zio_root); 4612 scn->scn_zio_root = NULL; 4613 4614 /* calculate and dprintf the current memory usage */ 4615 (void) dsl_scan_should_clear(scn); 4616 dsl_scan_update_stats(scn); 4617 4618 zfs_dbgmsg("scan issued %llu blocks for %s (%llu segs) " 4619 "in %llums (avg_block_size = %llu, avg_seg_size = %llu)", 4620 (longlong_t)scn->scn_zios_this_txg, 4621 spa->spa_name, 4622 (longlong_t)scn->scn_segs_this_txg, 4623 (longlong_t)NSEC2MSEC(gethrtime() - 4624 scn->scn_sync_start_time), 4625 (longlong_t)scn->scn_avg_zio_size_this_txg, 4626 (longlong_t)scn->scn_avg_seg_size_this_txg); 4627 } else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) { 4628 /* Finished with everything. Mark the scrub as complete */ 4629 zfs_dbgmsg("scan issuing complete txg %llu for %s", 4630 (longlong_t)tx->tx_txg, 4631 spa->spa_name); 4632 ASSERT3U(scn->scn_done_txg, !=, 0); 4633 ASSERT0(spa->spa_scrub_inflight); 4634 ASSERT0(scn->scn_queues_pending); 4635 dsl_scan_done(scn, B_TRUE, tx); 4636 sync_type = SYNC_MANDATORY; 4637 } 4638 4639 dsl_scan_sync_state(scn, tx, sync_type); 4640 } 4641 4642 static void 4643 count_block_issued(spa_t *spa, const blkptr_t *bp, boolean_t all) 4644 { 4645 /* 4646 * Don't count embedded bp's, since we already did the work of 4647 * scanning these when we scanned the containing block. 4648 */ 4649 if (BP_IS_EMBEDDED(bp)) 4650 return; 4651 4652 /* 4653 * Update the spa's stats on how many bytes we have issued. 4654 * Sequential scrubs create a zio for each DVA of the bp. Each 4655 * of these will include all DVAs for repair purposes, but the 4656 * zio code will only try the first one unless there is an issue. 4657 * Therefore, we should only count the first DVA for these IOs. 4658 */ 4659 atomic_add_64(&spa->spa_scan_pass_issued, 4660 all ? BP_GET_ASIZE(bp) : DVA_GET_ASIZE(&bp->blk_dva[0])); 4661 } 4662 4663 static void 4664 count_block_skipped(dsl_scan_t *scn, const blkptr_t *bp, boolean_t all) 4665 { 4666 if (BP_IS_EMBEDDED(bp)) 4667 return; 4668 atomic_add_64(&scn->scn_phys.scn_skipped, 4669 all ? BP_GET_ASIZE(bp) : DVA_GET_ASIZE(&bp->blk_dva[0])); 4670 } 4671 4672 static void 4673 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp) 4674 { 4675 /* 4676 * If we resume after a reboot, zab will be NULL; don't record 4677 * incomplete stats in that case. 4678 */ 4679 if (zab == NULL) 4680 return; 4681 4682 for (int i = 0; i < 4; i++) { 4683 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS; 4684 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL; 4685 4686 if (t & DMU_OT_NEWTYPE) 4687 t = DMU_OT_OTHER; 4688 zfs_blkstat_t *zb = &zab->zab_type[l][t]; 4689 int equal; 4690 4691 zb->zb_count++; 4692 zb->zb_asize += BP_GET_ASIZE(bp); 4693 zb->zb_lsize += BP_GET_LSIZE(bp); 4694 zb->zb_psize += BP_GET_PSIZE(bp); 4695 zb->zb_gangs += BP_COUNT_GANG(bp); 4696 4697 switch (BP_GET_NDVAS(bp)) { 4698 case 2: 4699 if (DVA_GET_VDEV(&bp->blk_dva[0]) == 4700 DVA_GET_VDEV(&bp->blk_dva[1])) 4701 zb->zb_ditto_2_of_2_samevdev++; 4702 break; 4703 case 3: 4704 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) == 4705 DVA_GET_VDEV(&bp->blk_dva[1])) + 4706 (DVA_GET_VDEV(&bp->blk_dva[0]) == 4707 DVA_GET_VDEV(&bp->blk_dva[2])) + 4708 (DVA_GET_VDEV(&bp->blk_dva[1]) == 4709 DVA_GET_VDEV(&bp->blk_dva[2])); 4710 if (equal == 1) 4711 zb->zb_ditto_2_of_3_samevdev++; 4712 else if (equal == 3) 4713 zb->zb_ditto_3_of_3_samevdev++; 4714 break; 4715 } 4716 } 4717 } 4718 4719 static void 4720 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio) 4721 { 4722 avl_index_t idx; 4723 dsl_scan_t *scn = queue->q_scn; 4724 4725 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock)); 4726 4727 if (unlikely(avl_is_empty(&queue->q_sios_by_addr))) 4728 atomic_add_64(&scn->scn_queues_pending, 1); 4729 if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) { 4730 /* block is already scheduled for reading */ 4731 sio_free(sio); 4732 return; 4733 } 4734 avl_insert(&queue->q_sios_by_addr, sio, idx); 4735 queue->q_sio_memused += SIO_GET_MUSED(sio); 4736 zfs_range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio), 4737 SIO_GET_ASIZE(sio)); 4738 } 4739 4740 /* 4741 * Given all the info we got from our metadata scanning process, we 4742 * construct a scan_io_t and insert it into the scan sorting queue. The 4743 * I/O must already be suitable for us to process. This is controlled 4744 * by dsl_scan_enqueue(). 4745 */ 4746 static void 4747 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i, 4748 int zio_flags, const zbookmark_phys_t *zb) 4749 { 4750 scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp)); 4751 4752 ASSERT0(BP_IS_GANG(bp)); 4753 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock)); 4754 4755 bp2sio(bp, sio, dva_i); 4756 sio->sio_flags = zio_flags; 4757 sio->sio_zb = *zb; 4758 4759 queue->q_last_ext_addr = -1; 4760 scan_io_queue_insert_impl(queue, sio); 4761 } 4762 4763 /* 4764 * Given a set of I/O parameters as discovered by the metadata traversal 4765 * process, attempts to place the I/O into the sorted queues (if allowed), 4766 * or immediately executes the I/O. 4767 */ 4768 static void 4769 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags, 4770 const zbookmark_phys_t *zb) 4771 { 4772 spa_t *spa = dp->dp_spa; 4773 4774 ASSERT(!BP_IS_EMBEDDED(bp)); 4775 4776 /* 4777 * Gang blocks are hard to issue sequentially, so we just issue them 4778 * here immediately instead of queuing them. 4779 */ 4780 if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) { 4781 scan_exec_io(dp, bp, zio_flags, zb, NULL); 4782 return; 4783 } 4784 4785 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 4786 dva_t dva; 4787 vdev_t *vdev; 4788 4789 dva = bp->blk_dva[i]; 4790 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva)); 4791 ASSERT(vdev != NULL); 4792 4793 mutex_enter(&vdev->vdev_scan_io_queue_lock); 4794 if (vdev->vdev_scan_io_queue == NULL) 4795 vdev->vdev_scan_io_queue = scan_io_queue_create(vdev); 4796 ASSERT(dp->dp_scan != NULL); 4797 scan_io_queue_insert(vdev->vdev_scan_io_queue, bp, 4798 i, zio_flags, zb); 4799 mutex_exit(&vdev->vdev_scan_io_queue_lock); 4800 } 4801 } 4802 4803 static int 4804 dsl_scan_scrub_cb(dsl_pool_t *dp, 4805 const blkptr_t *bp, const zbookmark_phys_t *zb) 4806 { 4807 dsl_scan_t *scn = dp->dp_scan; 4808 spa_t *spa = dp->dp_spa; 4809 uint64_t phys_birth = BP_GET_BIRTH(bp); 4810 size_t psize = BP_GET_PSIZE(bp); 4811 boolean_t needs_io = B_FALSE; 4812 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL; 4813 4814 count_block(dp->dp_blkstats, bp); 4815 if (phys_birth <= scn->scn_phys.scn_min_txg || 4816 phys_birth >= scn->scn_phys.scn_max_txg) { 4817 count_block_skipped(scn, bp, B_TRUE); 4818 return (0); 4819 } 4820 4821 /* Embedded BP's have phys_birth==0, so we reject them above. */ 4822 ASSERT(!BP_IS_EMBEDDED(bp)); 4823 4824 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn)); 4825 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) { 4826 zio_flags |= ZIO_FLAG_SCRUB; 4827 needs_io = B_TRUE; 4828 } else { 4829 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER); 4830 zio_flags |= ZIO_FLAG_RESILVER; 4831 needs_io = B_FALSE; 4832 } 4833 4834 /* If it's an intent log block, failure is expected. */ 4835 if (zb->zb_level == ZB_ZIL_LEVEL) 4836 zio_flags |= ZIO_FLAG_SPECULATIVE; 4837 4838 for (int d = 0; d < BP_GET_NDVAS(bp); d++) { 4839 const dva_t *dva = &bp->blk_dva[d]; 4840 4841 /* 4842 * Keep track of how much data we've examined so that 4843 * zpool(8) status can make useful progress reports. 4844 */ 4845 uint64_t asize = DVA_GET_ASIZE(dva); 4846 scn->scn_phys.scn_examined += asize; 4847 spa->spa_scan_pass_exam += asize; 4848 4849 /* if it's a resilver, this may not be in the target range */ 4850 if (!needs_io) 4851 needs_io = dsl_scan_need_resilver(spa, dva, psize, 4852 phys_birth); 4853 } 4854 4855 if (needs_io && !zfs_no_scrub_io) { 4856 dsl_scan_enqueue(dp, bp, zio_flags, zb); 4857 } else { 4858 count_block_skipped(scn, bp, B_TRUE); 4859 } 4860 4861 /* do not relocate this block */ 4862 return (0); 4863 } 4864 4865 static void 4866 dsl_scan_scrub_done(zio_t *zio) 4867 { 4868 spa_t *spa = zio->io_spa; 4869 blkptr_t *bp = zio->io_bp; 4870 dsl_scan_io_queue_t *queue = zio->io_private; 4871 4872 abd_free(zio->io_abd); 4873 4874 if (queue == NULL) { 4875 mutex_enter(&spa->spa_scrub_lock); 4876 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp)); 4877 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp); 4878 cv_broadcast(&spa->spa_scrub_io_cv); 4879 mutex_exit(&spa->spa_scrub_lock); 4880 } else { 4881 mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock); 4882 ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp)); 4883 queue->q_inflight_bytes -= BP_GET_PSIZE(bp); 4884 cv_broadcast(&queue->q_zio_cv); 4885 mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock); 4886 } 4887 4888 if (zio->io_error && (zio->io_error != ECKSUM || 4889 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) { 4890 if (dsl_errorscrubbing(spa->spa_dsl_pool) && 4891 !dsl_errorscrub_is_paused(spa->spa_dsl_pool->dp_scan)) { 4892 atomic_inc_64(&spa->spa_dsl_pool->dp_scan 4893 ->errorscrub_phys.dep_errors); 4894 } else { 4895 atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys 4896 .scn_errors); 4897 } 4898 } 4899 } 4900 4901 /* 4902 * Given a scanning zio's information, executes the zio. The zio need 4903 * not necessarily be only sortable, this function simply executes the 4904 * zio, no matter what it is. The optional queue argument allows the 4905 * caller to specify that they want per top level vdev IO rate limiting 4906 * instead of the legacy global limiting. 4907 */ 4908 static void 4909 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags, 4910 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue) 4911 { 4912 spa_t *spa = dp->dp_spa; 4913 dsl_scan_t *scn = dp->dp_scan; 4914 size_t size = BP_GET_PSIZE(bp); 4915 abd_t *data = abd_alloc_for_io(size, B_FALSE); 4916 zio_t *pio; 4917 4918 if (queue == NULL) { 4919 ASSERT3U(scn->scn_maxinflight_bytes, >, 0); 4920 mutex_enter(&spa->spa_scrub_lock); 4921 while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes) 4922 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 4923 spa->spa_scrub_inflight += BP_GET_PSIZE(bp); 4924 mutex_exit(&spa->spa_scrub_lock); 4925 pio = scn->scn_zio_root; 4926 } else { 4927 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock; 4928 4929 ASSERT3U(queue->q_maxinflight_bytes, >, 0); 4930 mutex_enter(q_lock); 4931 while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes) 4932 cv_wait(&queue->q_zio_cv, q_lock); 4933 queue->q_inflight_bytes += BP_GET_PSIZE(bp); 4934 pio = queue->q_zio; 4935 mutex_exit(q_lock); 4936 } 4937 4938 ASSERT(pio != NULL); 4939 count_block_issued(spa, bp, queue == NULL); 4940 zio_nowait(zio_read(pio, spa, bp, data, size, dsl_scan_scrub_done, 4941 queue, ZIO_PRIORITY_SCRUB, zio_flags, zb)); 4942 } 4943 4944 /* 4945 * This is the primary extent sorting algorithm. We balance two parameters: 4946 * 1) how many bytes of I/O are in an extent 4947 * 2) how well the extent is filled with I/O (as a fraction of its total size) 4948 * Since we allow extents to have gaps between their constituent I/Os, it's 4949 * possible to have a fairly large extent that contains the same amount of 4950 * I/O bytes than a much smaller extent, which just packs the I/O more tightly. 4951 * The algorithm sorts based on a score calculated from the extent's size, 4952 * the relative fill volume (in %) and a "fill weight" parameter that controls 4953 * the split between whether we prefer larger extents or more well populated 4954 * extents: 4955 * 4956 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT) 4957 * 4958 * Example: 4959 * 1) assume extsz = 64 MiB 4960 * 2) assume fill = 32 MiB (extent is half full) 4961 * 3) assume fill_weight = 3 4962 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100 4963 * SCORE = 32M + (50 * 3 * 32M) / 100 4964 * SCORE = 32M + (4800M / 100) 4965 * SCORE = 32M + 48M 4966 * ^ ^ 4967 * | +--- final total relative fill-based score 4968 * +--------- final total fill-based score 4969 * SCORE = 80M 4970 * 4971 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards 4972 * extents that are more completely filled (in a 3:2 ratio) vs just larger. 4973 * Note that as an optimization, we replace multiplication and division by 4974 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128). 4975 * 4976 * Since we do not care if one extent is only few percent better than another, 4977 * compress the score into 6 bits via binary logarithm AKA highbit64() and 4978 * put into otherwise unused due to ashift high bits of offset. This allows 4979 * to reduce q_exts_by_size B-tree elements to only 64 bits and compare them 4980 * with single operation. Plus it makes scrubs more sequential and reduces 4981 * chances that minor extent change move it within the B-tree. 4982 */ 4983 __attribute__((always_inline)) inline 4984 static int 4985 ext_size_compare(const void *x, const void *y) 4986 { 4987 const uint64_t *a = x, *b = y; 4988 4989 return (TREE_CMP(*a, *b)); 4990 } 4991 4992 ZFS_BTREE_FIND_IN_BUF_FUNC(ext_size_find_in_buf, uint64_t, 4993 ext_size_compare) 4994 4995 static void 4996 ext_size_create(zfs_range_tree_t *rt, void *arg) 4997 { 4998 (void) rt; 4999 zfs_btree_t *size_tree = arg; 5000 5001 zfs_btree_create(size_tree, ext_size_compare, ext_size_find_in_buf, 5002 sizeof (uint64_t)); 5003 } 5004 5005 static void 5006 ext_size_destroy(zfs_range_tree_t *rt, void *arg) 5007 { 5008 (void) rt; 5009 zfs_btree_t *size_tree = arg; 5010 ASSERT0(zfs_btree_numnodes(size_tree)); 5011 5012 zfs_btree_destroy(size_tree); 5013 } 5014 5015 static uint64_t 5016 ext_size_value(zfs_range_tree_t *rt, zfs_range_seg_gap_t *rsg) 5017 { 5018 (void) rt; 5019 uint64_t size = rsg->rs_end - rsg->rs_start; 5020 uint64_t score = rsg->rs_fill + ((((rsg->rs_fill << 7) / size) * 5021 fill_weight * rsg->rs_fill) >> 7); 5022 ASSERT3U(rt->rt_shift, >=, 8); 5023 return (((uint64_t)(64 - highbit64(score)) << 56) | rsg->rs_start); 5024 } 5025 5026 static void 5027 ext_size_add(zfs_range_tree_t *rt, zfs_range_seg_t *rs, void *arg) 5028 { 5029 zfs_btree_t *size_tree = arg; 5030 ASSERT3U(rt->rt_type, ==, ZFS_RANGE_SEG_GAP); 5031 uint64_t v = ext_size_value(rt, (zfs_range_seg_gap_t *)rs); 5032 zfs_btree_add(size_tree, &v); 5033 } 5034 5035 static void 5036 ext_size_remove(zfs_range_tree_t *rt, zfs_range_seg_t *rs, void *arg) 5037 { 5038 zfs_btree_t *size_tree = arg; 5039 ASSERT3U(rt->rt_type, ==, ZFS_RANGE_SEG_GAP); 5040 uint64_t v = ext_size_value(rt, (zfs_range_seg_gap_t *)rs); 5041 zfs_btree_remove(size_tree, &v); 5042 } 5043 5044 static void 5045 ext_size_vacate(zfs_range_tree_t *rt, void *arg) 5046 { 5047 zfs_btree_t *size_tree = arg; 5048 zfs_btree_clear(size_tree); 5049 zfs_btree_destroy(size_tree); 5050 5051 ext_size_create(rt, arg); 5052 } 5053 5054 static const zfs_range_tree_ops_t ext_size_ops = { 5055 .rtop_create = ext_size_create, 5056 .rtop_destroy = ext_size_destroy, 5057 .rtop_add = ext_size_add, 5058 .rtop_remove = ext_size_remove, 5059 .rtop_vacate = ext_size_vacate 5060 }; 5061 5062 /* 5063 * Comparator for the q_sios_by_addr tree. Sorting is simply performed 5064 * based on LBA-order (from lowest to highest). 5065 */ 5066 static int 5067 sio_addr_compare(const void *x, const void *y) 5068 { 5069 const scan_io_t *a = x, *b = y; 5070 5071 return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b))); 5072 } 5073 5074 /* IO queues are created on demand when they are needed. */ 5075 static dsl_scan_io_queue_t * 5076 scan_io_queue_create(vdev_t *vd) 5077 { 5078 dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan; 5079 dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP); 5080 5081 q->q_scn = scn; 5082 q->q_vd = vd; 5083 q->q_sio_memused = 0; 5084 q->q_last_ext_addr = -1; 5085 cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL); 5086 q->q_exts_by_addr = zfs_range_tree_create_gap(&ext_size_ops, 5087 ZFS_RANGE_SEG_GAP, &q->q_exts_by_size, 0, vd->vdev_ashift, 5088 zfs_scan_max_ext_gap); 5089 avl_create(&q->q_sios_by_addr, sio_addr_compare, 5090 sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node)); 5091 5092 return (q); 5093 } 5094 5095 /* 5096 * Destroys a scan queue and all segments and scan_io_t's contained in it. 5097 * No further execution of I/O occurs, anything pending in the queue is 5098 * simply freed without being executed. 5099 */ 5100 void 5101 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue) 5102 { 5103 dsl_scan_t *scn = queue->q_scn; 5104 scan_io_t *sio; 5105 void *cookie = NULL; 5106 5107 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock)); 5108 5109 if (!avl_is_empty(&queue->q_sios_by_addr)) 5110 atomic_add_64(&scn->scn_queues_pending, -1); 5111 while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) != 5112 NULL) { 5113 ASSERT(zfs_range_tree_contains(queue->q_exts_by_addr, 5114 SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio))); 5115 queue->q_sio_memused -= SIO_GET_MUSED(sio); 5116 sio_free(sio); 5117 } 5118 5119 ASSERT0(queue->q_sio_memused); 5120 zfs_range_tree_vacate(queue->q_exts_by_addr, NULL, queue); 5121 zfs_range_tree_destroy(queue->q_exts_by_addr); 5122 avl_destroy(&queue->q_sios_by_addr); 5123 cv_destroy(&queue->q_zio_cv); 5124 5125 kmem_free(queue, sizeof (*queue)); 5126 } 5127 5128 /* 5129 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is 5130 * called on behalf of vdev_top_transfer when creating or destroying 5131 * a mirror vdev due to zpool attach/detach. 5132 */ 5133 void 5134 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd) 5135 { 5136 mutex_enter(&svd->vdev_scan_io_queue_lock); 5137 mutex_enter(&tvd->vdev_scan_io_queue_lock); 5138 5139 VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL); 5140 tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue; 5141 svd->vdev_scan_io_queue = NULL; 5142 if (tvd->vdev_scan_io_queue != NULL) 5143 tvd->vdev_scan_io_queue->q_vd = tvd; 5144 5145 mutex_exit(&tvd->vdev_scan_io_queue_lock); 5146 mutex_exit(&svd->vdev_scan_io_queue_lock); 5147 } 5148 5149 static void 5150 scan_io_queues_destroy(dsl_scan_t *scn) 5151 { 5152 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev; 5153 5154 for (uint64_t i = 0; i < rvd->vdev_children; i++) { 5155 vdev_t *tvd = rvd->vdev_child[i]; 5156 5157 mutex_enter(&tvd->vdev_scan_io_queue_lock); 5158 if (tvd->vdev_scan_io_queue != NULL) 5159 dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue); 5160 tvd->vdev_scan_io_queue = NULL; 5161 mutex_exit(&tvd->vdev_scan_io_queue_lock); 5162 } 5163 } 5164 5165 static void 5166 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i) 5167 { 5168 dsl_pool_t *dp = spa->spa_dsl_pool; 5169 dsl_scan_t *scn = dp->dp_scan; 5170 vdev_t *vdev; 5171 kmutex_t *q_lock; 5172 dsl_scan_io_queue_t *queue; 5173 scan_io_t *srch_sio, *sio; 5174 avl_index_t idx; 5175 uint64_t start, size; 5176 5177 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i])); 5178 ASSERT(vdev != NULL); 5179 q_lock = &vdev->vdev_scan_io_queue_lock; 5180 queue = vdev->vdev_scan_io_queue; 5181 5182 mutex_enter(q_lock); 5183 if (queue == NULL) { 5184 mutex_exit(q_lock); 5185 return; 5186 } 5187 5188 srch_sio = sio_alloc(BP_GET_NDVAS(bp)); 5189 bp2sio(bp, srch_sio, dva_i); 5190 start = SIO_GET_OFFSET(srch_sio); 5191 size = SIO_GET_ASIZE(srch_sio); 5192 5193 /* 5194 * We can find the zio in two states: 5195 * 1) Cold, just sitting in the queue of zio's to be issued at 5196 * some point in the future. In this case, all we do is 5197 * remove the zio from the q_sios_by_addr tree, decrement 5198 * its data volume from the containing zfs_range_seg_t and 5199 * resort the q_exts_by_size tree to reflect that the 5200 * zfs_range_seg_t has lost some of its 'fill'. We don't shorten 5201 * the zfs_range_seg_t - this is usually rare enough not to be 5202 * worth the extra hassle of trying keep track of precise 5203 * extent boundaries. 5204 * 2) Hot, where the zio is currently in-flight in 5205 * dsl_scan_issue_ios. In this case, we can't simply 5206 * reach in and stop the in-flight zio's, so we instead 5207 * block the caller. Eventually, dsl_scan_issue_ios will 5208 * be done with issuing the zio's it gathered and will 5209 * signal us. 5210 */ 5211 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx); 5212 sio_free(srch_sio); 5213 5214 if (sio != NULL) { 5215 blkptr_t tmpbp; 5216 5217 /* Got it while it was cold in the queue */ 5218 ASSERT3U(start, ==, SIO_GET_OFFSET(sio)); 5219 ASSERT3U(size, ==, SIO_GET_ASIZE(sio)); 5220 avl_remove(&queue->q_sios_by_addr, sio); 5221 if (avl_is_empty(&queue->q_sios_by_addr)) 5222 atomic_add_64(&scn->scn_queues_pending, -1); 5223 queue->q_sio_memused -= SIO_GET_MUSED(sio); 5224 5225 ASSERT(zfs_range_tree_contains(queue->q_exts_by_addr, start, 5226 size)); 5227 zfs_range_tree_remove_fill(queue->q_exts_by_addr, start, size); 5228 5229 /* count the block as though we skipped it */ 5230 sio2bp(sio, &tmpbp); 5231 count_block_skipped(scn, &tmpbp, B_FALSE); 5232 5233 sio_free(sio); 5234 } 5235 mutex_exit(q_lock); 5236 } 5237 5238 /* 5239 * Callback invoked when a zio_free() zio is executing. This needs to be 5240 * intercepted to prevent the zio from deallocating a particular portion 5241 * of disk space and it then getting reallocated and written to, while we 5242 * still have it queued up for processing. 5243 */ 5244 void 5245 dsl_scan_freed(spa_t *spa, const blkptr_t *bp) 5246 { 5247 dsl_pool_t *dp = spa->spa_dsl_pool; 5248 dsl_scan_t *scn = dp->dp_scan; 5249 5250 ASSERT(!BP_IS_EMBEDDED(bp)); 5251 ASSERT(scn != NULL); 5252 if (!dsl_scan_is_running(scn)) 5253 return; 5254 5255 for (int i = 0; i < BP_GET_NDVAS(bp); i++) 5256 dsl_scan_freed_dva(spa, bp, i); 5257 } 5258 5259 /* 5260 * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has 5261 * not started, start it. Otherwise, only restart if max txg in DTL range is 5262 * greater than the max txg in the current scan. If the DTL max is less than 5263 * the scan max, then the vdev has not missed any new data since the resilver 5264 * started, so a restart is not needed. 5265 */ 5266 void 5267 dsl_scan_assess_vdev(dsl_pool_t *dp, vdev_t *vd) 5268 { 5269 uint64_t min, max; 5270 5271 if (!vdev_resilver_needed(vd, &min, &max)) 5272 return; 5273 5274 if (!dsl_scan_resilvering(dp)) { 5275 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER); 5276 return; 5277 } 5278 5279 if (max <= dp->dp_scan->scn_phys.scn_max_txg) 5280 return; 5281 5282 /* restart is needed, check if it can be deferred */ 5283 if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER)) 5284 vdev_defer_resilver(vd); 5285 else 5286 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER); 5287 } 5288 5289 ZFS_MODULE_PARAM(zfs, zfs_, scan_vdev_limit, U64, ZMOD_RW, 5290 "Max bytes in flight per leaf vdev for scrubs and resilvers"); 5291 5292 ZFS_MODULE_PARAM(zfs, zfs_, scrub_min_time_ms, UINT, ZMOD_RW, 5293 "Min millisecs to scrub per txg"); 5294 5295 ZFS_MODULE_PARAM(zfs, zfs_, obsolete_min_time_ms, UINT, ZMOD_RW, 5296 "Min millisecs to obsolete per txg"); 5297 5298 ZFS_MODULE_PARAM(zfs, zfs_, free_min_time_ms, UINT, ZMOD_RW, 5299 "Min millisecs to free per txg"); 5300 5301 ZFS_MODULE_PARAM(zfs, zfs_, resilver_min_time_ms, UINT, ZMOD_RW, 5302 "Min millisecs to resilver per txg"); 5303 5304 ZFS_MODULE_PARAM(zfs, zfs_, scan_suspend_progress, INT, ZMOD_RW, 5305 "Set to prevent scans from progressing"); 5306 5307 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_io, INT, ZMOD_RW, 5308 "Set to disable scrub I/O"); 5309 5310 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_prefetch, INT, ZMOD_RW, 5311 "Set to disable scrub prefetching"); 5312 5313 ZFS_MODULE_PARAM(zfs, zfs_, async_block_max_blocks, U64, ZMOD_RW, 5314 "Max number of blocks freed in one txg"); 5315 5316 ZFS_MODULE_PARAM(zfs, zfs_, max_async_dedup_frees, U64, ZMOD_RW, 5317 "Max number of dedup blocks freed in one txg"); 5318 5319 ZFS_MODULE_PARAM(zfs, zfs_, free_bpobj_enabled, INT, ZMOD_RW, 5320 "Enable processing of the free_bpobj"); 5321 5322 ZFS_MODULE_PARAM(zfs, zfs_, scan_blkstats, INT, ZMOD_RW, 5323 "Enable block statistics calculation during scrub"); 5324 5325 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_fact, UINT, ZMOD_RW, 5326 "Fraction of RAM for scan hard limit"); 5327 5328 ZFS_MODULE_PARAM(zfs, zfs_, scan_issue_strategy, UINT, ZMOD_RW, 5329 "IO issuing strategy during scrubbing. 0 = default, 1 = LBA, 2 = size"); 5330 5331 ZFS_MODULE_PARAM(zfs, zfs_, scan_legacy, INT, ZMOD_RW, 5332 "Scrub using legacy non-sequential method"); 5333 5334 ZFS_MODULE_PARAM(zfs, zfs_, scan_checkpoint_intval, UINT, ZMOD_RW, 5335 "Scan progress on-disk checkpointing interval"); 5336 5337 ZFS_MODULE_PARAM(zfs, zfs_, scan_max_ext_gap, U64, ZMOD_RW, 5338 "Max gap in bytes between sequential scrub / resilver I/Os"); 5339 5340 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_soft_fact, UINT, ZMOD_RW, 5341 "Fraction of hard limit used as soft limit"); 5342 5343 ZFS_MODULE_PARAM(zfs, zfs_, scan_strict_mem_lim, INT, ZMOD_RW, 5344 "Tunable to attempt to reduce lock contention"); 5345 5346 ZFS_MODULE_PARAM(zfs, zfs_, scan_fill_weight, UINT, ZMOD_RW, 5347 "Tunable to adjust bias towards more filled segments during scans"); 5348 5349 ZFS_MODULE_PARAM(zfs, zfs_, scan_report_txgs, UINT, ZMOD_RW, 5350 "Tunable to report resilver performance over the last N txgs"); 5351 5352 ZFS_MODULE_PARAM(zfs, zfs_, resilver_disable_defer, INT, ZMOD_RW, 5353 "Process all resilvers immediately"); 5354 5355 ZFS_MODULE_PARAM(zfs, zfs_, resilver_defer_percent, UINT, ZMOD_RW, 5356 "Issued IO percent complete after which resilvers are deferred"); 5357 5358 ZFS_MODULE_PARAM(zfs, zfs_, scrub_error_blocks_per_txg, UINT, ZMOD_RW, 5359 "Error blocks to be scrubbed in one txg"); 5360