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