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