sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_scan.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  24  * Copyright 2016 Gary Mills
  25  */
  26
  27 #include <sys/dsl_scan.h>
  28 #include <sys/dsl_pool.h>
  29 #include <sys/dsl_dataset.h>
  30 #include <sys/dsl_prop.h>
  31 #include <sys/dsl_dir.h>
  32 #include <sys/dsl_synctask.h>
  33 #include <sys/dnode.h>
  34 #include <sys/dmu_tx.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/arc.h>
  37 #include <sys/zap.h>
  38 #include <sys/zio.h>
  39 #include <sys/zfs_context.h>
  40 #include <sys/fs/zfs.h>
  41 #include <sys/zfs_znode.h>
  42 #include <sys/spa_impl.h>
  43 #include <sys/vdev_impl.h>
  44 #include <sys/zil_impl.h>
  45 #include <sys/zio_checksum.h>
  46 #include <sys/ddt.h>
  47 #include <sys/sa.h>
  48 #include <sys/sa_impl.h>
  49 #include <sys/zfeature.h>
  50 #ifdef _KERNEL
  51 #include <sys/zfs_vfsops.h>
  52 #endif
  53
  54 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
  55     const zbookmark_phys_t *);
  56
  57 static scan_cb_t dsl_scan_scrub_cb;
  58 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
  59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *);
  60 static boolean_t dsl_scan_restarting(dsl_scan_t *, dmu_tx_t *);
  61
  62 unsigned int zfs_top_maxinflight = 32;  /* maximum I/Os per top-level */
  63 unsigned int zfs_resilver_delay = 2;    /* number of ticks to delay resilver */
  64 unsigned int zfs_scrub_delay = 4;       /* number of ticks to delay scrub */
  65 unsigned int zfs_scan_idle = 50;        /* idle window in clock ticks */
  66
  67 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  68 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  69 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
  70                                                  per txg */
  71 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  72 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
  73
  74 SYSCTL_DECL(_vfs_zfs);
  75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RWTUN,
  76     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
  77 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RWTUN,
  78     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
  79 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RWTUN,
  80     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
  81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RWTUN,
  82     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
  83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RWTUN,
  84     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
  85 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RWTUN,
  86     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
  87 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RWTUN,
  88     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
  89 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RWTUN,
  90     &zfs_no_scrub_io, 0, "Disable scrub I/O");
  91 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RWTUN,
  92     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
  93
  94 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
  95 /* max number of blocks to free in a single TXG */
  96 uint64_t zfs_free_max_blocks = UINT64_MAX;
  97 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
  98     &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
  99
 100
 101 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 102         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 103         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 104
 105 extern int zfs_txg_timeout;
 106
 107 /*
 108  * Enable/disable the processing of the free_bpobj object.
 109  */
 110 boolean_t zfs_free_bpobj_enabled = B_TRUE;
 111
 112 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
 113     &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
 114
 115 /* the order has to match pool_scan_type */
 116 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 117         NULL,
 118         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 119         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 120 };
 121
 122 int
 123 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 124 {
 125         int err;
 126         dsl_scan_t *scn;
 127         spa_t *spa = dp->dp_spa;
 128         uint64_t f;
 129
 130         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 131         scn->scn_dp = dp;
 132
 133         /*
 134          * It's possible that we're resuming a scan after a reboot so
 135          * make sure that the scan_async_destroying flag is initialized
 136          * appropriately.
 137          */
 138         ASSERT(!scn->scn_async_destroying);
 139         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 140             SPA_FEATURE_ASYNC_DESTROY);
 141
 142         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 143             "scrub_func", sizeof (uint64_t), 1, &f);
 144         if (err == 0) {
 145                 /*
 146                  * There was an old-style scrub in progress.  Restart a
 147                  * new-style scrub from the beginning.
 148                  */
 149                 scn->scn_restart_txg = txg;
 150                 zfs_dbgmsg("old-style scrub was in progress; "
 151                     "restarting new-style scrub in txg %llu",
 152                     scn->scn_restart_txg);
 153
 154                 /*
 155                  * Load the queue obj from the old location so that it
 156                  * can be freed by dsl_scan_done().
 157                  */
 158                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 159                     "scrub_queue", sizeof (uint64_t), 1,
 160                     &scn->scn_phys.scn_queue_obj);
 161         } else {
 162                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 163                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 164                     &scn->scn_phys);
 165                 if (err == ENOENT)
 166                         return (0);
 167                 else if (err)
 168                         return (err);
 169
 170                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 171                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 172                         /*
 173                          * A new-type scrub was in progress on an old
 174                          * pool, and the pool was accessed by old
 175                          * software.  Restart from the beginning, since
 176                          * the old software may have changed the pool in
 177                          * the meantime.
 178                          */
 179                         scn->scn_restart_txg = txg;
 180                         zfs_dbgmsg("new-style scrub was modified "
 181                             "by old software; restarting in txg %llu",
 182                             scn->scn_restart_txg);
 183                 }
 184         }
 185
 186         spa_scan_stat_init(spa);
 187         return (0);
 188 }
 189
 190 void
 191 dsl_scan_fini(dsl_pool_t *dp)
 192 {
 193         if (dp->dp_scan) {
 194                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 195                 dp->dp_scan = NULL;
 196         }
 197 }
 198
 199 /* ARGSUSED */
 200 static int
 201 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 202 {
 203         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 204
 205         if (scn->scn_phys.scn_state == DSS_SCANNING)
 206                 return (SET_ERROR(EBUSY));
 207
 208         return (0);
 209 }
 210
 211 static void
 212 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 213 {
 214         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 215         pool_scan_func_t *funcp = arg;
 216         dmu_object_type_t ot = 0;
 217         dsl_pool_t *dp = scn->scn_dp;
 218         spa_t *spa = dp->dp_spa;
 219
 220         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 221         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 222         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 223         scn->scn_phys.scn_func = *funcp;
 224         scn->scn_phys.scn_state = DSS_SCANNING;
 225         scn->scn_phys.scn_min_txg = 0;
 226         scn->scn_phys.scn_max_txg = tx->tx_txg;
 227         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 228         scn->scn_phys.scn_start_time = gethrestime_sec();
 229         scn->scn_phys.scn_errors = 0;
 230         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 231         scn->scn_restart_txg = 0;
 232         scn->scn_done_txg = 0;
 233         spa_scan_stat_init(spa);
 234
 235         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 236                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 237
 238                 /* rewrite all disk labels */
 239                 vdev_config_dirty(spa->spa_root_vdev);
 240
 241                 if (vdev_resilver_needed(spa->spa_root_vdev,
 242                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 243                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 244                 } else {
 245                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 246                 }
 247
 248                 spa->spa_scrub_started = B_TRUE;
 249                 /*
 250                  * If this is an incremental scrub, limit the DDT scrub phase
 251                  * to just the auto-ditto class (for correctness); the rest
 252                  * of the scrub should go faster using top-down pruning.
 253                  */
 254                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 255                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 256
 257         }
 258
 259         /* back to the generic stuff */
 260
 261         if (dp->dp_blkstats == NULL) {
 262                 dp->dp_blkstats =
 263                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 264         }
 265         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 266
 267         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 268                 ot = DMU_OT_ZAP_OTHER;
 269
 270         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 271             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 272
 273         dsl_scan_sync_state(scn, tx);
 274
 275         spa_history_log_internal(spa, "scan setup", tx,
 276             "func=%u mintxg=%llu maxtxg=%llu",
 277             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 278 }
 279
 280 /* ARGSUSED */
 281 static void
 282 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 283 {
 284         static const char *old_names[] = {
 285                 "scrub_bookmark",
 286                 "scrub_ddt_bookmark",
 287                 "scrub_ddt_class_max",
 288                 "scrub_queue",
 289                 "scrub_min_txg",
 290                 "scrub_max_txg",
 291                 "scrub_func",
 292                 "scrub_errors",
 293                 NULL
 294         };
 295
 296         dsl_pool_t *dp = scn->scn_dp;
 297         spa_t *spa = dp->dp_spa;
 298         int i;
 299
 300         /* Remove any remnants of an old-style scrub. */
 301         for (i = 0; old_names[i]; i++) {
 302                 (void) zap_remove(dp->dp_meta_objset,
 303                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 304         }
 305
 306         if (scn->scn_phys.scn_queue_obj != 0) {
 307                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 308                     scn->scn_phys.scn_queue_obj, tx));
 309                 scn->scn_phys.scn_queue_obj = 0;
 310         }
 311
 312         /*
 313          * If we were "restarted" from a stopped state, don't bother
 314          * with anything else.
 315          */
 316         if (scn->scn_phys.scn_state != DSS_SCANNING)
 317                 return;
 318
 319         if (complete)
 320                 scn->scn_phys.scn_state = DSS_FINISHED;
 321         else
 322                 scn->scn_phys.scn_state = DSS_CANCELED;
 323
 324         if (dsl_scan_restarting(scn, tx))
 325                 spa_history_log_internal(spa, "scan aborted, restarting", tx,
 326                     "errors=%llu", spa_get_errlog_size(spa));
 327         else if (!complete)
 328                 spa_history_log_internal(spa, "scan cancelled", tx,
 329                     "errors=%llu", spa_get_errlog_size(spa));
 330         else
 331                 spa_history_log_internal(spa, "scan done", tx,
 332                     "errors=%llu", spa_get_errlog_size(spa));
 333
 334         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 335                 mutex_enter(&spa->spa_scrub_lock);
 336                 while (spa->spa_scrub_inflight > 0) {
 337                         cv_wait(&spa->spa_scrub_io_cv,
 338                             &spa->spa_scrub_lock);
 339                 }
 340                 mutex_exit(&spa->spa_scrub_lock);
 341                 spa->spa_scrub_started = B_FALSE;
 342                 spa->spa_scrub_active = B_FALSE;
 343
 344                 /*
 345                  * If the scrub/resilver completed, update all DTLs to
 346                  * reflect this.  Whether it succeeded or not, vacate
 347                  * all temporary scrub DTLs.
 348                  */
 349                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 350                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 351                 if (complete) {
 352                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 353                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 354                 }
 355                 spa_errlog_rotate(spa);
 356
 357                 /*
 358                  * We may have finished replacing a device.
 359                  * Let the async thread assess this and handle the detach.
 360                  */
 361                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 362         }
 363
 364         scn->scn_phys.scn_end_time = gethrestime_sec();
 365 }
 366
 367 /* ARGSUSED */
 368 static int
 369 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 370 {
 371         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 372
 373         if (scn->scn_phys.scn_state != DSS_SCANNING)
 374                 return (SET_ERROR(ENOENT));
 375         return (0);
 376 }
 377
 378 /* ARGSUSED */
 379 static void
 380 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 381 {
 382         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 383
 384         dsl_scan_done(scn, B_FALSE, tx);
 385         dsl_scan_sync_state(scn, tx);
 386 }
 387
 388 int
 389 dsl_scan_cancel(dsl_pool_t *dp)
 390 {
 391         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 392             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 393 }
 394
 395 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 396     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 397     dmu_objset_type_t ostype, dmu_tx_t *tx);
 398 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 399     dmu_objset_type_t ostype,
 400     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
 401
 402 void
 403 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 404 {
 405         zio_free(dp->dp_spa, txg, bp);
 406 }
 407
 408 void
 409 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 410 {
 411         ASSERT(dsl_pool_sync_context(dp));
 412         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 413             pio->io_flags));
 414 }
 415
 416 static uint64_t
 417 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 418 {
 419         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 420         if (ds->ds_is_snapshot)
 421                 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
 422         return (smt);
 423 }
 424
 425 static void
 426 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 427 {
 428         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 429             DMU_POOL_DIRECTORY_OBJECT,
 430             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 431             &scn->scn_phys, tx));
 432 }
 433
 434 extern int zfs_vdev_async_write_active_min_dirty_percent;
 435
 436 static boolean_t
 437 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 438 {
 439         /* we never skip user/group accounting objects */
 440         if (zb && (int64_t)zb->zb_object < 0)
 441                 return (B_FALSE);
 442
 443         if (scn->scn_pausing)
 444                 return (B_TRUE); /* we're already pausing */
 445
 446         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 447                 return (B_FALSE); /* we're resuming */
 448
 449         /* We only know how to resume from level-0 blocks. */
 450         if (zb && zb->zb_level != 0)
 451                 return (B_FALSE);
 452
 453         /*
 454          * We pause if:
 455          *  - we have scanned for the maximum time: an entire txg
 456          *    timeout (default 5 sec)
 457          *  or
 458          *  - we have scanned for at least the minimum time (default 1 sec
 459          *    for scrub, 3 sec for resilver), and either we have sufficient
 460          *    dirty data that we are starting to write more quickly
 461          *    (default 30%), or someone is explicitly waiting for this txg
 462          *    to complete.
 463          *  or
 464          *  - the spa is shutting down because this pool is being exported
 465          *    or the machine is rebooting.
 466          */
 467         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 468             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 469         uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 470         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
 471         if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
 472             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 473             (txg_sync_waiting(scn->scn_dp) ||
 474             dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
 475             spa_shutting_down(scn->scn_dp->dp_spa)) {
 476                 if (zb) {
 477                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 478                             (longlong_t)zb->zb_objset,
 479                             (longlong_t)zb->zb_object,
 480                             (longlong_t)zb->zb_level,
 481                             (longlong_t)zb->zb_blkid);
 482                         scn->scn_phys.scn_bookmark = *zb;
 483                 }
 484                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 485                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 486                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 487                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 488                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 489                 scn->scn_pausing = B_TRUE;
 490                 return (B_TRUE);
 491         }
 492         return (B_FALSE);
 493 }
 494
 495 typedef struct zil_scan_arg {
 496         dsl_pool_t      *zsa_dp;
 497         zil_header_t    *zsa_zh;
 498 } zil_scan_arg_t;
 499
 500 /* ARGSUSED */
 501 static int
 502 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 503 {
 504         zil_scan_arg_t *zsa = arg;
 505         dsl_pool_t *dp = zsa->zsa_dp;
 506         dsl_scan_t *scn = dp->dp_scan;
 507         zil_header_t *zh = zsa->zsa_zh;
 508         zbookmark_phys_t zb;
 509
 510         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 511                 return (0);
 512
 513         /*
 514          * One block ("stubby") can be allocated a long time ago; we
 515          * want to visit that one because it has been allocated
 516          * (on-disk) even if it hasn't been claimed (even though for
 517          * scrub there's nothing to do to it).
 518          */
 519         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 520                 return (0);
 521
 522         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 523             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 524
 525         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 526         return (0);
 527 }
 528
 529 /* ARGSUSED */
 530 static int
 531 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 532 {
 533         if (lrc->lrc_txtype == TX_WRITE) {
 534                 zil_scan_arg_t *zsa = arg;
 535                 dsl_pool_t *dp = zsa->zsa_dp;
 536                 dsl_scan_t *scn = dp->dp_scan;
 537                 zil_header_t *zh = zsa->zsa_zh;
 538                 lr_write_t *lr = (lr_write_t *)lrc;
 539                 blkptr_t *bp = &lr->lr_blkptr;
 540                 zbookmark_phys_t zb;
 541
 542                 if (BP_IS_HOLE(bp) ||
 543                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 544                         return (0);
 545
 546                 /*
 547                  * birth can be < claim_txg if this record's txg is
 548                  * already txg sync'ed (but this log block contains
 549                  * other records that are not synced)
 550                  */
 551                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 552                         return (0);
 553
 554                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 555                     lr->lr_foid, ZB_ZIL_LEVEL,
 556                     lr->lr_offset / BP_GET_LSIZE(bp));
 557
 558                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 559         }
 560         return (0);
 561 }
 562
 563 static void
 564 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 565 {
 566         uint64_t claim_txg = zh->zh_claim_txg;
 567         zil_scan_arg_t zsa = { dp, zh };
 568         zilog_t *zilog;
 569
 570         /*
 571          * We only want to visit blocks that have been claimed but not yet
 572          * replayed (or, in read-only mode, blocks that *would* be claimed).
 573          */
 574         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 575                 return;
 576
 577         zilog = zil_alloc(dp->dp_meta_objset, zh);
 578
 579         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 580             claim_txg);
 581
 582         zil_free(zilog);
 583 }
 584
 585 /* ARGSUSED */
 586 static void
 587 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 588     uint64_t objset, uint64_t object, uint64_t blkid)
 589 {
 590         zbookmark_phys_t czb;
 591         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 592
 593         if (zfs_no_scrub_prefetch)
 594                 return;
 595
 596         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 597             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 598                 return;
 599
 600         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 601
 602         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 603             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 604             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 605 }
 606
 607 static boolean_t
 608 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 609     const zbookmark_phys_t *zb)
 610 {
 611         /*
 612          * We never skip over user/group accounting objects (obj<0)
 613          */
 614         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 615             (int64_t)zb->zb_object >= 0) {
 616                 /*
 617                  * If we already visited this bp & everything below (in
 618                  * a prior txg sync), don't bother doing it again.
 619                  */
 620                 if (zbookmark_subtree_completed(dnp, zb,
 621                     &scn->scn_phys.scn_bookmark))
 622                         return (B_TRUE);
 623
 624                 /*
 625                  * If we found the block we're trying to resume from, or
 626                  * we went past it to a different object, zero it out to
 627                  * indicate that it's OK to start checking for pausing
 628                  * again.
 629                  */
 630                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 631                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 632                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 633                             (longlong_t)zb->zb_objset,
 634                             (longlong_t)zb->zb_object,
 635                             (longlong_t)zb->zb_level,
 636                             (longlong_t)zb->zb_blkid);
 637                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 638                 }
 639         }
 640         return (B_FALSE);
 641 }
 642
 643 /*
 644  * Return nonzero on i/o error.
 645  * Return new buf to write out in *bufp.
 646  */
 647 static int
 648 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 649     dnode_phys_t *dnp, const blkptr_t *bp,
 650     const zbookmark_phys_t *zb, dmu_tx_t *tx)
 651 {
 652         dsl_pool_t *dp = scn->scn_dp;
 653         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 654         int err;
 655
 656         if (BP_GET_LEVEL(bp) > 0) {
 657                 arc_flags_t flags = ARC_FLAG_WAIT;
 658                 int i;
 659                 blkptr_t *cbp;
 660                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 661                 arc_buf_t *buf;
 662
 663                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 664                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 665                 if (err) {
 666                         scn->scn_phys.scn_errors++;
 667                         return (err);
 668                 }
 669                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 670                         dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
 671                             zb->zb_object, zb->zb_blkid * epb + i);
 672                 }
 673                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 674                         zbookmark_phys_t czb;
 675
 676                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 677                             zb->zb_level - 1,
 678                             zb->zb_blkid * epb + i);
 679                         dsl_scan_visitbp(cbp, &czb, dnp,
 680                             ds, scn, ostype, tx);
 681                 }
 682                 arc_buf_destroy(buf, &buf);
 683         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 684                 arc_flags_t flags = ARC_FLAG_WAIT;
 685                 dnode_phys_t *cdnp;
 686                 int i, j;
 687                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 688                 arc_buf_t *buf;
 689
 690                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 691                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 692                 if (err) {
 693                         scn->scn_phys.scn_errors++;
 694                         return (err);
 695                 }
 696                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 697                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 698                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 699                                 dsl_scan_prefetch(scn, buf, cbp,
 700                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 701                         }
 702                 }
 703                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 704                         dsl_scan_visitdnode(scn, ds, ostype,
 705                             cdnp, zb->zb_blkid * epb + i, tx);
 706                 }
 707
 708                 arc_buf_destroy(buf, &buf);
 709         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 710                 arc_flags_t flags = ARC_FLAG_WAIT;
 711                 objset_phys_t *osp;
 712                 arc_buf_t *buf;
 713
 714                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 715                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 716                 if (err) {
 717                         scn->scn_phys.scn_errors++;
 718                         return (err);
 719                 }
 720
 721                 osp = buf->b_data;
 722
 723                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 724                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
 725
 726                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 727                         /*
 728                          * We also always visit user/group accounting
 729                          * objects, and never skip them, even if we are
 730                          * pausing.  This is necessary so that the space
 731                          * deltas from this txg get integrated.
 732                          */
 733                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 734                             &osp->os_groupused_dnode,
 735                             DMU_GROUPUSED_OBJECT, tx);
 736                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 737                             &osp->os_userused_dnode,
 738                             DMU_USERUSED_OBJECT, tx);
 739                 }
 740                 arc_buf_destroy(buf, &buf);
 741         }
 742
 743         return (0);
 744 }
 745
 746 static void
 747 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 748     dmu_objset_type_t ostype, dnode_phys_t *dnp,
 749     uint64_t object, dmu_tx_t *tx)
 750 {
 751         int j;
 752
 753         for (j = 0; j < dnp->dn_nblkptr; j++) {
 754                 zbookmark_phys_t czb;
 755
 756                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 757                     dnp->dn_nlevels - 1, j);
 758                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 759                     &czb, dnp, ds, scn, ostype, tx);
 760         }
 761
 762         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 763                 zbookmark_phys_t czb;
 764                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 765                     0, DMU_SPILL_BLKID);
 766                 dsl_scan_visitbp(&dnp->dn_spill,
 767                     &czb, dnp, ds, scn, ostype, tx);
 768         }
 769 }
 770
 771 /*
 772  * The arguments are in this order because mdb can only print the
 773  * first 5; we want them to be useful.
 774  */
 775 static void
 776 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 777     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 778     dmu_objset_type_t ostype, dmu_tx_t *tx)
 779 {
 780         dsl_pool_t *dp = scn->scn_dp;
 781         arc_buf_t *buf = NULL;
 782         blkptr_t bp_toread = *bp;
 783
 784         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 785
 786         if (dsl_scan_check_pause(scn, zb))
 787                 return;
 788
 789         if (dsl_scan_check_resume(scn, dnp, zb))
 790                 return;
 791
 792         if (BP_IS_HOLE(bp))
 793                 return;
 794
 795         scn->scn_visited_this_txg++;
 796
 797         dprintf_bp(bp,
 798             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
 799             ds, ds ? ds->ds_object : 0,
 800             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 801             bp);
 802
 803         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 804                 return;
 805
 806         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
 807                 return;
 808
 809         /*
 810          * If dsl_scan_ddt() has aready visited this block, it will have
 811          * already done any translations or scrubbing, so don't call the
 812          * callback again.
 813          */
 814         if (ddt_class_contains(dp->dp_spa,
 815             scn->scn_phys.scn_ddt_class_max, bp)) {
 816                 ASSERT(buf == NULL);
 817                 return;
 818         }
 819
 820         /*
 821          * If this block is from the future (after cur_max_txg), then we
 822          * are doing this on behalf of a deleted snapshot, and we will
 823          * revisit the future block on the next pass of this dataset.
 824          * Don't scan it now unless we need to because something
 825          * under it was modified.
 826          */
 827         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 828                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 829         }
 830 }
 831
 832 static void
 833 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 834     dmu_tx_t *tx)
 835 {
 836         zbookmark_phys_t zb;
 837
 838         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 839             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 840         dsl_scan_visitbp(bp, &zb, NULL,
 841             ds, scn, DMU_OST_NONE, tx);
 842
 843         dprintf_ds(ds, "finished scan%s", "");
 844 }
 845
 846 void
 847 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 848 {
 849         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 850         dsl_scan_t *scn = dp->dp_scan;
 851         uint64_t mintxg;
 852
 853         if (scn->scn_phys.scn_state != DSS_SCANNING)
 854                 return;
 855
 856         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 857                 if (ds->ds_is_snapshot) {
 858                         /*
 859                          * Note:
 860                          *  - scn_cur_{min,max}_txg stays the same.
 861                          *  - Setting the flag is not really necessary if
 862                          *    scn_cur_max_txg == scn_max_txg, because there
 863                          *    is nothing after this snapshot that we care
 864                          *    about.  However, we set it anyway and then
 865                          *    ignore it when we retraverse it in
 866                          *    dsl_scan_visitds().
 867                          */
 868                         scn->scn_phys.scn_bookmark.zb_objset =
 869                             dsl_dataset_phys(ds)->ds_next_snap_obj;
 870                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 871                             "reset zb_objset to %llu",
 872                             (u_longlong_t)ds->ds_object,
 873                             (u_longlong_t)dsl_dataset_phys(ds)->
 874                             ds_next_snap_obj);
 875                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 876                 } else {
 877                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 878                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 879                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 880                             "reset bookmark to -1,0,0,0",
 881                             (u_longlong_t)ds->ds_object);
 882                 }
 883         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 884             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 885                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
 886                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 887                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 888                 if (ds->ds_is_snapshot) {
 889                         /*
 890                          * We keep the same mintxg; it could be >
 891                          * ds_creation_txg if the previous snapshot was
 892                          * deleted too.
 893                          */
 894                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 895                             scn->scn_phys.scn_queue_obj,
 896                             dsl_dataset_phys(ds)->ds_next_snap_obj,
 897                             mintxg, tx) == 0);
 898                         zfs_dbgmsg("destroying ds %llu; in queue; "
 899                             "replacing with %llu",
 900                             (u_longlong_t)ds->ds_object,
 901                             (u_longlong_t)dsl_dataset_phys(ds)->
 902                             ds_next_snap_obj);
 903                 } else {
 904                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 905                             (u_longlong_t)ds->ds_object);
 906                 }
 907         }
 908
 909         /*
 910          * dsl_scan_sync() should be called after this, and should sync
 911          * out our changed state, but just to be safe, do it here.
 912          */
 913         dsl_scan_sync_state(scn, tx);
 914 }
 915
 916 void
 917 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 918 {
 919         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 920         dsl_scan_t *scn = dp->dp_scan;
 921         uint64_t mintxg;
 922
 923         if (scn->scn_phys.scn_state != DSS_SCANNING)
 924                 return;
 925
 926         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 927
 928         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 929                 scn->scn_phys.scn_bookmark.zb_objset =
 930                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
 931                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 932                     "reset zb_objset to %llu",
 933                     (u_longlong_t)ds->ds_object,
 934                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 935         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 936             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 937                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 938                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 939                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 940                     scn->scn_phys.scn_queue_obj,
 941                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
 942                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 943                     "replacing with %llu",
 944                     (u_longlong_t)ds->ds_object,
 945                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 946         }
 947         dsl_scan_sync_state(scn, tx);
 948 }
 949
 950 void
 951 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 952 {
 953         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 954         dsl_scan_t *scn = dp->dp_scan;
 955         uint64_t mintxg;
 956
 957         if (scn->scn_phys.scn_state != DSS_SCANNING)
 958                 return;
 959
 960         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 961                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 962                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 963                     "reset zb_objset to %llu",
 964                     (u_longlong_t)ds1->ds_object,
 965                     (u_longlong_t)ds2->ds_object);
 966         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 967                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 968                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 969                     "reset zb_objset to %llu",
 970                     (u_longlong_t)ds2->ds_object,
 971                     (u_longlong_t)ds1->ds_object);
 972         }
 973
 974         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 975             ds1->ds_object, &mintxg) == 0) {
 976                 int err;
 977
 978                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 979                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 980                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 981                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 982                 err = zap_add_int_key(dp->dp_meta_objset,
 983                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 984                 VERIFY(err == 0 || err == EEXIST);
 985                 if (err == EEXIST) {
 986                         /* Both were there to begin with */
 987                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 988                             scn->scn_phys.scn_queue_obj,
 989                             ds1->ds_object, mintxg, tx));
 990                 }
 991                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 992                     "replacing with %llu",
 993                     (u_longlong_t)ds1->ds_object,
 994                     (u_longlong_t)ds2->ds_object);
 995         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 996             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 997                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 998                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 999                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1000                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1001                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1002                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1003                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1004                     "replacing with %llu",
1005                     (u_longlong_t)ds2->ds_object,
1006                     (u_longlong_t)ds1->ds_object);
1007         }
1008
1009         dsl_scan_sync_state(scn, tx);
1010 }
1011
1012 struct enqueue_clones_arg {
1013         dmu_tx_t *tx;
1014         uint64_t originobj;
1015 };
1016
1017 /* ARGSUSED */
1018 static int
1019 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1020 {
1021         struct enqueue_clones_arg *eca = arg;
1022         dsl_dataset_t *ds;
1023         int err;
1024         dsl_scan_t *scn = dp->dp_scan;
1025
1026         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1027                 return (0);
1028
1029         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1030         if (err)
1031                 return (err);
1032
1033         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1034                 dsl_dataset_t *prev;
1035                 err = dsl_dataset_hold_obj(dp,
1036                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1037
1038                 dsl_dataset_rele(ds, FTAG);
1039                 if (err)
1040                         return (err);
1041                 ds = prev;
1042         }
1043         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1044             scn->scn_phys.scn_queue_obj, ds->ds_object,
1045             dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1046         dsl_dataset_rele(ds, FTAG);
1047         return (0);
1048 }
1049
1050 static void
1051 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1052 {
1053         dsl_pool_t *dp = scn->scn_dp;
1054         dsl_dataset_t *ds;
1055         objset_t *os;
1056
1057         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1058
1059         if (scn->scn_phys.scn_cur_min_txg >=
1060             scn->scn_phys.scn_max_txg) {
1061                 /*
1062                  * This can happen if this snapshot was created after the
1063                  * scan started, and we already completed a previous snapshot
1064                  * that was created after the scan started.  This snapshot
1065                  * only references blocks with:
1066                  *
1067                  *      birth < our ds_creation_txg
1068                  *      cur_min_txg is no less than ds_creation_txg.
1069                  *      We have already visited these blocks.
1070                  * or
1071                  *      birth > scn_max_txg
1072                  *      The scan requested not to visit these blocks.
1073                  *
1074                  * Subsequent snapshots (and clones) can reference our
1075                  * blocks, or blocks with even higher birth times.
1076                  * Therefore we do not need to visit them either,
1077                  * so we do not add them to the work queue.
1078                  *
1079                  * Note that checking for cur_min_txg >= cur_max_txg
1080                  * is not sufficient, because in that case we may need to
1081                  * visit subsequent snapshots.  This happens when min_txg > 0,
1082                  * which raises cur_min_txg.  In this case we will visit
1083                  * this dataset but skip all of its blocks, because the
1084                  * rootbp's birth time is < cur_min_txg.  Then we will
1085                  * add the next snapshots/clones to the work queue.
1086                  */
1087                 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1088                 dsl_dataset_name(ds, dsname);
1089                 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1090                     "cur_min_txg (%llu) >= max_txg (%llu)",
1091                     dsobj, dsname,
1092                     scn->scn_phys.scn_cur_min_txg,
1093                     scn->scn_phys.scn_max_txg);
1094                 kmem_free(dsname, MAXNAMELEN);
1095
1096                 goto out;
1097         }
1098
1099         if (dmu_objset_from_ds(ds, &os))
1100                 goto out;
1101
1102         /*
1103          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1104          * snapshots can have ZIL block pointers (which may be the same
1105          * BP as in the head), they must be ignored.  So we traverse the
1106          * ZIL here, rather than in scan_recurse(), because the regular
1107          * snapshot block-sharing rules don't apply to it.
1108          */
1109         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1110                 dsl_scan_zil(dp, &os->os_zil_header);
1111
1112         /*
1113          * Iterate over the bps in this ds.
1114          */
1115         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1116         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1117         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1118         rrw_exit(&ds->ds_bp_rwlock, FTAG);
1119
1120         char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
1121         dsl_dataset_name(ds, dsname);
1122         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1123             "pausing=%u",
1124             (longlong_t)dsobj, dsname,
1125             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1126             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1127             (int)scn->scn_pausing);
1128         kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
1129
1130         if (scn->scn_pausing)
1131                 goto out;
1132
1133         /*
1134          * We've finished this pass over this dataset.
1135          */
1136
1137         /*
1138          * If we did not completely visit this dataset, do another pass.
1139          */
1140         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1141                 zfs_dbgmsg("incomplete pass; visiting again");
1142                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1143                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1144                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1145                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1146                 goto out;
1147         }
1148
1149         /*
1150          * Add descendent datasets to work queue.
1151          */
1152         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1153                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1154                     scn->scn_phys.scn_queue_obj,
1155                     dsl_dataset_phys(ds)->ds_next_snap_obj,
1156                     dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1157         }
1158         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1159                 boolean_t usenext = B_FALSE;
1160                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1161                         uint64_t count;
1162                         /*
1163                          * A bug in a previous version of the code could
1164                          * cause upgrade_clones_cb() to not set
1165                          * ds_next_snap_obj when it should, leading to a
1166                          * missing entry.  Therefore we can only use the
1167                          * next_clones_obj when its count is correct.
1168                          */
1169                         int err = zap_count(dp->dp_meta_objset,
1170                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1171                         if (err == 0 &&
1172                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
1173                                 usenext = B_TRUE;
1174                 }
1175
1176                 if (usenext) {
1177                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1178                             dsl_dataset_phys(ds)->ds_next_clones_obj,
1179                             scn->scn_phys.scn_queue_obj,
1180                             dsl_dataset_phys(ds)->ds_creation_txg, tx));
1181                 } else {
1182                         struct enqueue_clones_arg eca;
1183                         eca.tx = tx;
1184                         eca.originobj = ds->ds_object;
1185
1186                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1187                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1188                 }
1189         }
1190
1191 out:
1192         dsl_dataset_rele(ds, FTAG);
1193 }
1194
1195 /* ARGSUSED */
1196 static int
1197 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1198 {
1199         dmu_tx_t *tx = arg;
1200         dsl_dataset_t *ds;
1201         int err;
1202         dsl_scan_t *scn = dp->dp_scan;
1203
1204         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1205         if (err)
1206                 return (err);
1207
1208         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1209                 dsl_dataset_t *prev;
1210                 err = dsl_dataset_hold_obj(dp,
1211                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1212                 if (err) {
1213                         dsl_dataset_rele(ds, FTAG);
1214                         return (err);
1215                 }
1216
1217                 /*
1218                  * If this is a clone, we don't need to worry about it for now.
1219                  */
1220                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1221                         dsl_dataset_rele(ds, FTAG);
1222                         dsl_dataset_rele(prev, FTAG);
1223                         return (0);
1224                 }
1225                 dsl_dataset_rele(ds, FTAG);
1226                 ds = prev;
1227         }
1228
1229         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1230             ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1231         dsl_dataset_rele(ds, FTAG);
1232         return (0);
1233 }
1234
1235 /*
1236  * Scrub/dedup interaction.
1237  *
1238  * If there are N references to a deduped block, we don't want to scrub it
1239  * N times -- ideally, we should scrub it exactly once.
1240  *
1241  * We leverage the fact that the dde's replication class (enum ddt_class)
1242  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1243  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1244  *
1245  * To prevent excess scrubbing, the scrub begins by walking the DDT
1246  * to find all blocks with refcnt > 1, and scrubs each of these once.
1247  * Since there are two replication classes which contain blocks with
1248  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1249  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1250  *
1251  * There would be nothing more to say if a block's refcnt couldn't change
1252  * during a scrub, but of course it can so we must account for changes
1253  * in a block's replication class.
1254  *
1255  * Here's an example of what can occur:
1256  *
1257  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1258  * when visited during the top-down scrub phase, it will be scrubbed twice.
1259  * This negates our scrub optimization, but is otherwise harmless.
1260  *
1261  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1262  * on each visit during the top-down scrub phase, it will never be scrubbed.
1263  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1264  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1265  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1266  * while a scrub is in progress, it scrubs the block right then.
1267  */
1268 static void
1269 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1270 {
1271         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1272         ddt_entry_t dde = { 0 };
1273         int error;
1274         uint64_t n = 0;
1275
1276         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1277                 ddt_t *ddt;
1278
1279                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1280                         break;
1281                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1282                     (longlong_t)ddb->ddb_class,
1283                     (longlong_t)ddb->ddb_type,
1284                     (longlong_t)ddb->ddb_checksum,
1285                     (longlong_t)ddb->ddb_cursor);
1286
1287                 /* There should be no pending changes to the dedup table */
1288                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1289                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1290
1291                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1292                 n++;
1293
1294                 if (dsl_scan_check_pause(scn, NULL))
1295                         break;
1296         }
1297
1298         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1299             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1300             (int)scn->scn_pausing);
1301
1302         ASSERT(error == 0 || error == ENOENT);
1303         ASSERT(error != ENOENT ||
1304             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1305 }
1306
1307 /* ARGSUSED */
1308 void
1309 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1310     ddt_entry_t *dde, dmu_tx_t *tx)
1311 {
1312         const ddt_key_t *ddk = &dde->dde_key;
1313         ddt_phys_t *ddp = dde->dde_phys;
1314         blkptr_t bp;
1315         zbookmark_phys_t zb = { 0 };
1316
1317         if (scn->scn_phys.scn_state != DSS_SCANNING)
1318                 return;
1319
1320         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1321                 if (ddp->ddp_phys_birth == 0 ||
1322                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1323                         continue;
1324                 ddt_bp_create(checksum, ddk, ddp, &bp);
1325
1326                 scn->scn_visited_this_txg++;
1327                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1328         }
1329 }
1330
1331 static void
1332 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1333 {
1334         dsl_pool_t *dp = scn->scn_dp;
1335         zap_cursor_t zc;
1336         zap_attribute_t za;
1337
1338         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1339             scn->scn_phys.scn_ddt_class_max) {
1340                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1341                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1342                 dsl_scan_ddt(scn, tx);
1343                 if (scn->scn_pausing)
1344                         return;
1345         }
1346
1347         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1348                 /* First do the MOS & ORIGIN */
1349
1350                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1351                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1352                 dsl_scan_visit_rootbp(scn, NULL,
1353                     &dp->dp_meta_rootbp, tx);
1354                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1355                 if (scn->scn_pausing)
1356                         return;
1357
1358                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1359                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1360                             enqueue_cb, tx, DS_FIND_CHILDREN));
1361                 } else {
1362                         dsl_scan_visitds(scn,
1363                             dp->dp_origin_snap->ds_object, tx);
1364                 }
1365                 ASSERT(!scn->scn_pausing);
1366         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1367             ZB_DESTROYED_OBJSET) {
1368                 /*
1369                  * If we were paused, continue from here.  Note if the
1370                  * ds we were paused on was deleted, the zb_objset may
1371                  * be -1, so we will skip this and find a new objset
1372                  * below.
1373                  */
1374                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1375                 if (scn->scn_pausing)
1376                         return;
1377         }
1378
1379         /*
1380          * In case we were paused right at the end of the ds, zero the
1381          * bookmark so we don't think that we're still trying to resume.
1382          */
1383         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1384
1385         /* keep pulling things out of the zap-object-as-queue */
1386         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1387             scn->scn_phys.scn_queue_obj),
1388             zap_cursor_retrieve(&zc, &za) == 0) {
1389                 dsl_dataset_t *ds;
1390                 uint64_t dsobj;
1391
1392                 dsobj = strtonum(za.za_name, NULL);
1393                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1394                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1395
1396                 /* Set up min/max txg */
1397                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1398                 if (za.za_first_integer != 0) {
1399                         scn->scn_phys.scn_cur_min_txg =
1400                             MAX(scn->scn_phys.scn_min_txg,
1401                             za.za_first_integer);
1402                 } else {
1403                         scn->scn_phys.scn_cur_min_txg =
1404                             MAX(scn->scn_phys.scn_min_txg,
1405                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
1406                 }
1407                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1408                 dsl_dataset_rele(ds, FTAG);
1409
1410                 dsl_scan_visitds(scn, dsobj, tx);
1411                 zap_cursor_fini(&zc);
1412                 if (scn->scn_pausing)
1413                         return;
1414         }
1415         zap_cursor_fini(&zc);
1416 }
1417
1418 static boolean_t
1419 dsl_scan_free_should_pause(dsl_scan_t *scn)
1420 {
1421         uint64_t elapsed_nanosecs;
1422
1423         if (zfs_recover)
1424                 return (B_FALSE);
1425
1426         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1427                 return (B_TRUE);
1428
1429         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1430         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1431             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1432             txg_sync_waiting(scn->scn_dp)) ||
1433             spa_shutting_down(scn->scn_dp->dp_spa));
1434 }
1435
1436 static int
1437 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1438 {
1439         dsl_scan_t *scn = arg;
1440
1441         if (!scn->scn_is_bptree ||
1442             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1443                 if (dsl_scan_free_should_pause(scn))
1444                         return (SET_ERROR(ERESTART));
1445         }
1446
1447         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1448             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1449         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1450             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1451             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1452         scn->scn_visited_this_txg++;
1453         return (0);
1454 }
1455
1456 boolean_t
1457 dsl_scan_active(dsl_scan_t *scn)
1458 {
1459         spa_t *spa = scn->scn_dp->dp_spa;
1460         uint64_t used = 0, comp, uncomp;
1461
1462         if (spa->spa_load_state != SPA_LOAD_NONE)
1463                 return (B_FALSE);
1464         if (spa_shutting_down(spa))
1465                 return (B_FALSE);
1466         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1467             (scn->scn_async_destroying && !scn->scn_async_stalled))
1468                 return (B_TRUE);
1469
1470         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1471                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1472                     &used, &comp, &uncomp);
1473         }
1474         return (used != 0);
1475 }
1476
1477 void
1478 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1479 {
1480         dsl_scan_t *scn = dp->dp_scan;
1481         spa_t *spa = dp->dp_spa;
1482         int err = 0;
1483
1484         /*
1485          * Check for scn_restart_txg before checking spa_load_state, so
1486          * that we can restart an old-style scan while the pool is being
1487          * imported (see dsl_scan_init).
1488          */
1489         if (dsl_scan_restarting(scn, tx)) {
1490                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1491                 dsl_scan_done(scn, B_FALSE, tx);
1492                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1493                         func = POOL_SCAN_RESILVER;
1494                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1495                     func, tx->tx_txg);
1496                 dsl_scan_setup_sync(&func, tx);
1497         }
1498
1499         /*
1500          * Only process scans in sync pass 1.
1501          */
1502         if (spa_sync_pass(dp->dp_spa) > 1)
1503                 return;
1504
1505         /*
1506          * If the spa is shutting down, then stop scanning. This will
1507          * ensure that the scan does not dirty any new data during the
1508          * shutdown phase.
1509          */
1510         if (spa_shutting_down(spa))
1511                 return;
1512
1513         /*
1514          * If the scan is inactive due to a stalled async destroy, try again.
1515          */
1516         if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1517                 return;
1518
1519         scn->scn_visited_this_txg = 0;
1520         scn->scn_pausing = B_FALSE;
1521         scn->scn_sync_start_time = gethrtime();
1522         spa->spa_scrub_active = B_TRUE;
1523
1524         /*
1525          * First process the async destroys.  If we pause, don't do
1526          * any scrubbing or resilvering.  This ensures that there are no
1527          * async destroys while we are scanning, so the scan code doesn't
1528          * have to worry about traversing it.  It is also faster to free the
1529          * blocks than to scrub them.
1530          */
1531         if (zfs_free_bpobj_enabled &&
1532             spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1533                 scn->scn_is_bptree = B_FALSE;
1534                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1535                     NULL, ZIO_FLAG_MUSTSUCCEED);
1536                 err = bpobj_iterate(&dp->dp_free_bpobj,
1537                     dsl_scan_free_block_cb, scn, tx);
1538                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1539
1540                 if (err != 0 && err != ERESTART)
1541                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1542         }
1543
1544         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1545                 ASSERT(scn->scn_async_destroying);
1546                 scn->scn_is_bptree = B_TRUE;
1547                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1548                     NULL, ZIO_FLAG_MUSTSUCCEED);
1549                 err = bptree_iterate(dp->dp_meta_objset,
1550                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1551                 VERIFY0(zio_wait(scn->scn_zio_root));
1552
1553                 if (err == EIO || err == ECKSUM) {
1554                         err = 0;
1555                 } else if (err != 0 && err != ERESTART) {
1556                         zfs_panic_recover("error %u from "
1557                             "traverse_dataset_destroyed()", err);
1558                 }
1559
1560                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1561                         /* finished; deactivate async destroy feature */
1562                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1563                         ASSERT(!spa_feature_is_active(spa,
1564                             SPA_FEATURE_ASYNC_DESTROY));
1565                         VERIFY0(zap_remove(dp->dp_meta_objset,
1566                             DMU_POOL_DIRECTORY_OBJECT,
1567                             DMU_POOL_BPTREE_OBJ, tx));
1568                         VERIFY0(bptree_free(dp->dp_meta_objset,
1569                             dp->dp_bptree_obj, tx));
1570                         dp->dp_bptree_obj = 0;
1571                         scn->scn_async_destroying = B_FALSE;
1572                         scn->scn_async_stalled = B_FALSE;
1573                 } else {
1574                         /*
1575                          * If we didn't make progress, mark the async
1576                          * destroy as stalled, so that we will not initiate
1577                          * a spa_sync() on its behalf.  Note that we only
1578                          * check this if we are not finished, because if the
1579                          * bptree had no blocks for us to visit, we can
1580                          * finish without "making progress".
1581                          */
1582                         scn->scn_async_stalled =
1583                             (scn->scn_visited_this_txg == 0);
1584                 }
1585         }
1586         if (scn->scn_visited_this_txg) {
1587                 zfs_dbgmsg("freed %llu blocks in %llums from "
1588                     "free_bpobj/bptree txg %llu; err=%d",
1589                     (longlong_t)scn->scn_visited_this_txg,
1590                     (longlong_t)
1591                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1592                     (longlong_t)tx->tx_txg, err);
1593                 scn->scn_visited_this_txg = 0;
1594
1595                 /*
1596                  * Write out changes to the DDT that may be required as a
1597                  * result of the blocks freed.  This ensures that the DDT
1598                  * is clean when a scrub/resilver runs.
1599                  */
1600                 ddt_sync(spa, tx->tx_txg);
1601         }
1602         if (err != 0)
1603                 return;
1604         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1605             zfs_free_leak_on_eio &&
1606             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1607             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1608             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1609                 /*
1610                  * We have finished background destroying, but there is still
1611                  * some space left in the dp_free_dir. Transfer this leaked
1612                  * space to the dp_leak_dir.
1613                  */
1614                 if (dp->dp_leak_dir == NULL) {
1615                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1616                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1617                             LEAK_DIR_NAME, tx);
1618                         VERIFY0(dsl_pool_open_special_dir(dp,
1619                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1620                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1621                 }
1622                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1623                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1624                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1625                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1626                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1627                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1628                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1629                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1630         }
1631         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1632                 /* finished; verify that space accounting went to zero */
1633                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1634                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1635                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1636         }
1637
1638         if (scn->scn_phys.scn_state != DSS_SCANNING)
1639                 return;
1640
1641         if (scn->scn_done_txg == tx->tx_txg) {
1642                 ASSERT(!scn->scn_pausing);
1643                 /* finished with scan. */
1644                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1645                 dsl_scan_done(scn, B_TRUE, tx);
1646                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1647                 dsl_scan_sync_state(scn, tx);
1648                 return;
1649         }
1650
1651         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1652             scn->scn_phys.scn_ddt_class_max) {
1653                 zfs_dbgmsg("doing scan sync txg %llu; "
1654                     "ddt bm=%llu/%llu/%llu/%llx",
1655                     (longlong_t)tx->tx_txg,
1656                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1657                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1658                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1659                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1660                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1661                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1662                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1663                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1664         } else {
1665                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1666                     (longlong_t)tx->tx_txg,
1667                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1668                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1669                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1670                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1671         }
1672
1673         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1674             NULL, ZIO_FLAG_CANFAIL);
1675         dsl_pool_config_enter(dp, FTAG);
1676         dsl_scan_visit(scn, tx);
1677         dsl_pool_config_exit(dp, FTAG);
1678         (void) zio_wait(scn->scn_zio_root);
1679         scn->scn_zio_root = NULL;
1680
1681         zfs_dbgmsg("visited %llu blocks in %llums",
1682             (longlong_t)scn->scn_visited_this_txg,
1683             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1684
1685         if (!scn->scn_pausing) {
1686                 scn->scn_done_txg = tx->tx_txg + 1;
1687                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1688                     tx->tx_txg, scn->scn_done_txg);
1689         }
1690
1691         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1692                 mutex_enter(&spa->spa_scrub_lock);
1693                 while (spa->spa_scrub_inflight > 0) {
1694                         cv_wait(&spa->spa_scrub_io_cv,
1695                             &spa->spa_scrub_lock);
1696                 }
1697                 mutex_exit(&spa->spa_scrub_lock);
1698         }
1699
1700         dsl_scan_sync_state(scn, tx);
1701 }
1702
1703 /*
1704  * This will start a new scan, or restart an existing one.
1705  */
1706 void
1707 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1708 {
1709         if (txg == 0) {
1710                 dmu_tx_t *tx;
1711                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1712                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1713
1714                 txg = dmu_tx_get_txg(tx);
1715                 dp->dp_scan->scn_restart_txg = txg;
1716                 dmu_tx_commit(tx);
1717         } else {
1718                 dp->dp_scan->scn_restart_txg = txg;
1719         }
1720         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1721 }
1722
1723 boolean_t
1724 dsl_scan_resilvering(dsl_pool_t *dp)
1725 {
1726         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1727             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1728 }
1729
1730 /*
1731  * scrub consumers
1732  */
1733
1734 static void
1735 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1736 {
1737         int i;
1738
1739         /*
1740          * If we resume after a reboot, zab will be NULL; don't record
1741          * incomplete stats in that case.
1742          */
1743         if (zab == NULL)
1744                 return;
1745
1746         for (i = 0; i < 4; i++) {
1747                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1748                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1749                 if (t & DMU_OT_NEWTYPE)
1750                         t = DMU_OT_OTHER;
1751                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1752                 int equal;
1753
1754                 zb->zb_count++;
1755                 zb->zb_asize += BP_GET_ASIZE(bp);
1756                 zb->zb_lsize += BP_GET_LSIZE(bp);
1757                 zb->zb_psize += BP_GET_PSIZE(bp);
1758                 zb->zb_gangs += BP_COUNT_GANG(bp);
1759
1760                 switch (BP_GET_NDVAS(bp)) {
1761                 case 2:
1762                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1763                             DVA_GET_VDEV(&bp->blk_dva[1]))
1764                                 zb->zb_ditto_2_of_2_samevdev++;
1765                         break;
1766                 case 3:
1767                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1768                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1769                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1770                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1771                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1772                             DVA_GET_VDEV(&bp->blk_dva[2]));
1773                         if (equal == 1)
1774                                 zb->zb_ditto_2_of_3_samevdev++;
1775                         else if (equal == 3)
1776                                 zb->zb_ditto_3_of_3_samevdev++;
1777                         break;
1778                 }
1779         }
1780 }
1781
1782 static void
1783 dsl_scan_scrub_done(zio_t *zio)
1784 {
1785         spa_t *spa = zio->io_spa;
1786
1787         zio_data_buf_free(zio->io_data, zio->io_size);
1788
1789         mutex_enter(&spa->spa_scrub_lock);
1790         spa->spa_scrub_inflight--;
1791         cv_broadcast(&spa->spa_scrub_io_cv);
1792
1793         if (zio->io_error && (zio->io_error != ECKSUM ||
1794             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1795                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1796         }
1797         mutex_exit(&spa->spa_scrub_lock);
1798 }
1799
1800 static int
1801 dsl_scan_scrub_cb(dsl_pool_t *dp,
1802     const blkptr_t *bp, const zbookmark_phys_t *zb)
1803 {
1804         dsl_scan_t *scn = dp->dp_scan;
1805         size_t size = BP_GET_PSIZE(bp);
1806         spa_t *spa = dp->dp_spa;
1807         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1808         boolean_t needs_io;
1809         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1810         unsigned int scan_delay = 0;
1811
1812         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1813             phys_birth >= scn->scn_phys.scn_max_txg)
1814                 return (0);
1815
1816         count_block(dp->dp_blkstats, bp);
1817
1818         if (BP_IS_EMBEDDED(bp))
1819                 return (0);
1820
1821         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1822         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1823                 zio_flags |= ZIO_FLAG_SCRUB;
1824                 needs_io = B_TRUE;
1825                 scan_delay = zfs_scrub_delay;
1826         } else {
1827                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1828                 zio_flags |= ZIO_FLAG_RESILVER;
1829                 needs_io = B_FALSE;
1830                 scan_delay = zfs_resilver_delay;
1831         }
1832
1833         /* If it's an intent log block, failure is expected. */
1834         if (zb->zb_level == ZB_ZIL_LEVEL)
1835                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1836
1837         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1838                 vdev_t *vd = vdev_lookup_top(spa,
1839                     DVA_GET_VDEV(&bp->blk_dva[d]));
1840
1841                 /*
1842                  * Keep track of how much data we've examined so that
1843                  * zpool(1M) status can make useful progress reports.
1844                  */
1845                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1846                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1847
1848                 /* if it's a resilver, this may not be in the target range */
1849                 if (!needs_io) {
1850                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1851                                 /*
1852                                  * Gang members may be spread across multiple
1853                                  * vdevs, so the best estimate we have is the
1854                                  * scrub range, which has already been checked.
1855                                  * XXX -- it would be better to change our
1856                                  * allocation policy to ensure that all
1857                                  * gang members reside on the same vdev.
1858                                  */
1859                                 needs_io = B_TRUE;
1860                         } else {
1861                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1862                                     phys_birth, 1);
1863                         }
1864                 }
1865         }
1866
1867         if (needs_io && !zfs_no_scrub_io) {
1868                 vdev_t *rvd = spa->spa_root_vdev;
1869                 uint64_t maxinflight = rvd->vdev_children *
1870                     MAX(zfs_top_maxinflight, 1);
1871                 void *data = zio_data_buf_alloc(size);
1872
1873                 mutex_enter(&spa->spa_scrub_lock);
1874                 while (spa->spa_scrub_inflight >= maxinflight)
1875                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1876                 spa->spa_scrub_inflight++;
1877                 mutex_exit(&spa->spa_scrub_lock);
1878
1879                 /*
1880                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1881                  * then throttle our workload to limit the impact of a scan.
1882                  */
1883                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1884                         delay(MAX((int)scan_delay, 0));
1885
1886                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1887                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1888                     zio_flags, zb));
1889         }
1890
1891         /* do not relocate this block */
1892         return (0);
1893 }
1894
1895 int
1896 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1897 {
1898         spa_t *spa = dp->dp_spa;
1899
1900         /*
1901          * Purge all vdev caches and probe all devices.  We do this here
1902          * rather than in sync context because this requires a writer lock
1903          * on the spa_config lock, which we can't do from sync context.  The
1904          * spa_scrub_reopen flag indicates that vdev_open() should not
1905          * attempt to start another scrub.
1906          */
1907         spa_vdev_state_enter(spa, SCL_NONE);
1908         spa->spa_scrub_reopen = B_TRUE;
1909         vdev_reopen(spa->spa_root_vdev);
1910         spa->spa_scrub_reopen = B_FALSE;
1911         (void) spa_vdev_state_exit(spa, NULL, 0);
1912
1913         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1914             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1915 }
1916
1917 static boolean_t
1918 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
1919 {
1920         return (scn->scn_restart_txg != 0 &&
1921             scn->scn_restart_txg <= tx->tx_txg);
1922 }