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 TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
  76 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
  77     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
  78 TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
  79 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
  80     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
  81 TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
  82 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
  83     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
  84 TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
  85 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
  86     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
  87 TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
  88 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
  89     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
  90 TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
  91 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
  92     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
  93 TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
  94 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
  95     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
  96 TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
  97 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
  98     &zfs_no_scrub_io, 0, "Disable scrub I/O");
  99 TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
 100 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
 101     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
 102
 103 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
 104 /* max number of blocks to free in a single TXG */
 105 uint64_t zfs_free_max_blocks = UINT64_MAX;
 106 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
 107     &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
 108
 109
 110 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 111         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 112         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 113
 114 extern int zfs_txg_timeout;
 115
 116 /*
 117  * Enable/disable the processing of the free_bpobj object.
 118  */
 119 boolean_t zfs_free_bpobj_enabled = B_TRUE;
 120
 121 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
 122     &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
 123
 124 /* the order has to match pool_scan_type */
 125 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 126         NULL,
 127         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 128         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 129 };
 130
 131 int
 132 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 133 {
 134         int err;
 135         dsl_scan_t *scn;
 136         spa_t *spa = dp->dp_spa;
 137         uint64_t f;
 138
 139         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 140         scn->scn_dp = dp;
 141
 142         /*
 143          * It's possible that we're resuming a scan after a reboot so
 144          * make sure that the scan_async_destroying flag is initialized
 145          * appropriately.
 146          */
 147         ASSERT(!scn->scn_async_destroying);
 148         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 149             SPA_FEATURE_ASYNC_DESTROY);
 150
 151         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 152             "scrub_func", sizeof (uint64_t), 1, &f);
 153         if (err == 0) {
 154                 /*
 155                  * There was an old-style scrub in progress.  Restart a
 156                  * new-style scrub from the beginning.
 157                  */
 158                 scn->scn_restart_txg = txg;
 159                 zfs_dbgmsg("old-style scrub was in progress; "
 160                     "restarting new-style scrub in txg %llu",
 161                     scn->scn_restart_txg);
 162
 163                 /*
 164                  * Load the queue obj from the old location so that it
 165                  * can be freed by dsl_scan_done().
 166                  */
 167                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 168                     "scrub_queue", sizeof (uint64_t), 1,
 169                     &scn->scn_phys.scn_queue_obj);
 170         } else {
 171                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 172                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 173                     &scn->scn_phys);
 174                 if (err == ENOENT)
 175                         return (0);
 176                 else if (err)
 177                         return (err);
 178
 179                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 180                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 181                         /*
 182                          * A new-type scrub was in progress on an old
 183                          * pool, and the pool was accessed by old
 184                          * software.  Restart from the beginning, since
 185                          * the old software may have changed the pool in
 186                          * the meantime.
 187                          */
 188                         scn->scn_restart_txg = txg;
 189                         zfs_dbgmsg("new-style scrub was modified "
 190                             "by old software; restarting in txg %llu",
 191                             scn->scn_restart_txg);
 192                 }
 193         }
 194
 195         spa_scan_stat_init(spa);
 196         return (0);
 197 }
 198
 199 void
 200 dsl_scan_fini(dsl_pool_t *dp)
 201 {
 202         if (dp->dp_scan) {
 203                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 204                 dp->dp_scan = NULL;
 205         }
 206 }
 207
 208 /* ARGSUSED */
 209 static int
 210 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 211 {
 212         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 213
 214         if (scn->scn_phys.scn_state == DSS_SCANNING)
 215                 return (SET_ERROR(EBUSY));
 216
 217         return (0);
 218 }
 219
 220 static void
 221 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 222 {
 223         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 224         pool_scan_func_t *funcp = arg;
 225         dmu_object_type_t ot = 0;
 226         dsl_pool_t *dp = scn->scn_dp;
 227         spa_t *spa = dp->dp_spa;
 228
 229         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 230         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 231         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 232         scn->scn_phys.scn_func = *funcp;
 233         scn->scn_phys.scn_state = DSS_SCANNING;
 234         scn->scn_phys.scn_min_txg = 0;
 235         scn->scn_phys.scn_max_txg = tx->tx_txg;
 236         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 237         scn->scn_phys.scn_start_time = gethrestime_sec();
 238         scn->scn_phys.scn_errors = 0;
 239         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 240         scn->scn_restart_txg = 0;
 241         scn->scn_done_txg = 0;
 242         spa_scan_stat_init(spa);
 243
 244         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 245                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 246
 247                 /* rewrite all disk labels */
 248                 vdev_config_dirty(spa->spa_root_vdev);
 249
 250                 if (vdev_resilver_needed(spa->spa_root_vdev,
 251                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 252                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 253                 } else {
 254                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 255                 }
 256
 257                 spa->spa_scrub_started = B_TRUE;
 258                 /*
 259                  * If this is an incremental scrub, limit the DDT scrub phase
 260                  * to just the auto-ditto class (for correctness); the rest
 261                  * of the scrub should go faster using top-down pruning.
 262                  */
 263                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 264                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 265
 266         }
 267
 268         /* back to the generic stuff */
 269
 270         if (dp->dp_blkstats == NULL) {
 271                 dp->dp_blkstats =
 272                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 273         }
 274         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 275
 276         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 277                 ot = DMU_OT_ZAP_OTHER;
 278
 279         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 280             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 281
 282         dsl_scan_sync_state(scn, tx);
 283
 284         spa_history_log_internal(spa, "scan setup", tx,
 285             "func=%u mintxg=%llu maxtxg=%llu",
 286             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 287 }
 288
 289 /* ARGSUSED */
 290 static void
 291 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 292 {
 293         static const char *old_names[] = {
 294                 "scrub_bookmark",
 295                 "scrub_ddt_bookmark",
 296                 "scrub_ddt_class_max",
 297                 "scrub_queue",
 298                 "scrub_min_txg",
 299                 "scrub_max_txg",
 300                 "scrub_func",
 301                 "scrub_errors",
 302                 NULL
 303         };
 304
 305         dsl_pool_t *dp = scn->scn_dp;
 306         spa_t *spa = dp->dp_spa;
 307         int i;
 308
 309         /* Remove any remnants of an old-style scrub. */
 310         for (i = 0; old_names[i]; i++) {
 311                 (void) zap_remove(dp->dp_meta_objset,
 312                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 313         }
 314
 315         if (scn->scn_phys.scn_queue_obj != 0) {
 316                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 317                     scn->scn_phys.scn_queue_obj, tx));
 318                 scn->scn_phys.scn_queue_obj = 0;
 319         }
 320
 321         /*
 322          * If we were "restarted" from a stopped state, don't bother
 323          * with anything else.
 324          */
 325         if (scn->scn_phys.scn_state != DSS_SCANNING)
 326                 return;
 327
 328         if (complete)
 329                 scn->scn_phys.scn_state = DSS_FINISHED;
 330         else
 331                 scn->scn_phys.scn_state = DSS_CANCELED;
 332
 333         if (dsl_scan_restarting(scn, tx))
 334                 spa_history_log_internal(spa, "scan aborted, restarting", tx,
 335                     "errors=%llu", spa_get_errlog_size(spa));
 336         else if (!complete)
 337                 spa_history_log_internal(spa, "scan cancelled", tx,
 338                     "errors=%llu", spa_get_errlog_size(spa));
 339         else
 340                 spa_history_log_internal(spa, "scan done", tx,
 341                     "errors=%llu", spa_get_errlog_size(spa));
 342
 343         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 344                 mutex_enter(&spa->spa_scrub_lock);
 345                 while (spa->spa_scrub_inflight > 0) {
 346                         cv_wait(&spa->spa_scrub_io_cv,
 347                             &spa->spa_scrub_lock);
 348                 }
 349                 mutex_exit(&spa->spa_scrub_lock);
 350                 spa->spa_scrub_started = B_FALSE;
 351                 spa->spa_scrub_active = B_FALSE;
 352
 353                 /*
 354                  * If the scrub/resilver completed, update all DTLs to
 355                  * reflect this.  Whether it succeeded or not, vacate
 356                  * all temporary scrub DTLs.
 357                  */
 358                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 359                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 360                 if (complete) {
 361                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 362                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 363                 }
 364                 spa_errlog_rotate(spa);
 365
 366                 /*
 367                  * We may have finished replacing a device.
 368                  * Let the async thread assess this and handle the detach.
 369                  */
 370                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 371         }
 372
 373         scn->scn_phys.scn_end_time = gethrestime_sec();
 374 }
 375
 376 /* ARGSUSED */
 377 static int
 378 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 379 {
 380         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 381
 382         if (scn->scn_phys.scn_state != DSS_SCANNING)
 383                 return (SET_ERROR(ENOENT));
 384         return (0);
 385 }
 386
 387 /* ARGSUSED */
 388 static void
 389 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 390 {
 391         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 392
 393         dsl_scan_done(scn, B_FALSE, tx);
 394         dsl_scan_sync_state(scn, tx);
 395 }
 396
 397 int
 398 dsl_scan_cancel(dsl_pool_t *dp)
 399 {
 400         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 401             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 402 }
 403
 404 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 405     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 406     dmu_objset_type_t ostype, dmu_tx_t *tx);
 407 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 408     dmu_objset_type_t ostype,
 409     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
 410
 411 void
 412 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 413 {
 414         zio_free(dp->dp_spa, txg, bp);
 415 }
 416
 417 void
 418 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 419 {
 420         ASSERT(dsl_pool_sync_context(dp));
 421         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 422             pio->io_flags));
 423 }
 424
 425 static uint64_t
 426 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 427 {
 428         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 429         if (ds->ds_is_snapshot)
 430                 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
 431         return (smt);
 432 }
 433
 434 static void
 435 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 436 {
 437         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 438             DMU_POOL_DIRECTORY_OBJECT,
 439             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 440             &scn->scn_phys, tx));
 441 }
 442
 443 extern int zfs_vdev_async_write_active_min_dirty_percent;
 444
 445 static boolean_t
 446 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 447 {
 448         /* we never skip user/group accounting objects */
 449         if (zb && (int64_t)zb->zb_object < 0)
 450                 return (B_FALSE);
 451
 452         if (scn->scn_pausing)
 453                 return (B_TRUE); /* we're already pausing */
 454
 455         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 456                 return (B_FALSE); /* we're resuming */
 457
 458         /* We only know how to resume from level-0 blocks. */
 459         if (zb && zb->zb_level != 0)
 460                 return (B_FALSE);
 461
 462         /*
 463          * We pause if:
 464          *  - we have scanned for the maximum time: an entire txg
 465          *    timeout (default 5 sec)
 466          *  or
 467          *  - we have scanned for at least the minimum time (default 1 sec
 468          *    for scrub, 3 sec for resilver), and either we have sufficient
 469          *    dirty data that we are starting to write more quickly
 470          *    (default 30%), or someone is explicitly waiting for this txg
 471          *    to complete.
 472          *  or
 473          *  - the spa is shutting down because this pool is being exported
 474          *    or the machine is rebooting.
 475          */
 476         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 477             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 478         uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 479         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
 480         if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
 481             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 482             (txg_sync_waiting(scn->scn_dp) ||
 483             dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
 484             spa_shutting_down(scn->scn_dp->dp_spa)) {
 485                 if (zb) {
 486                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 487                             (longlong_t)zb->zb_objset,
 488                             (longlong_t)zb->zb_object,
 489                             (longlong_t)zb->zb_level,
 490                             (longlong_t)zb->zb_blkid);
 491                         scn->scn_phys.scn_bookmark = *zb;
 492                 }
 493                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 494                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 495                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 496                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 497                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 498                 scn->scn_pausing = B_TRUE;
 499                 return (B_TRUE);
 500         }
 501         return (B_FALSE);
 502 }
 503
 504 typedef struct zil_scan_arg {
 505         dsl_pool_t      *zsa_dp;
 506         zil_header_t    *zsa_zh;
 507 } zil_scan_arg_t;
 508
 509 /* ARGSUSED */
 510 static int
 511 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 512 {
 513         zil_scan_arg_t *zsa = arg;
 514         dsl_pool_t *dp = zsa->zsa_dp;
 515         dsl_scan_t *scn = dp->dp_scan;
 516         zil_header_t *zh = zsa->zsa_zh;
 517         zbookmark_phys_t zb;
 518
 519         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 520                 return (0);
 521
 522         /*
 523          * One block ("stubby") can be allocated a long time ago; we
 524          * want to visit that one because it has been allocated
 525          * (on-disk) even if it hasn't been claimed (even though for
 526          * scrub there's nothing to do to it).
 527          */
 528         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 529                 return (0);
 530
 531         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 532             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 533
 534         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 535         return (0);
 536 }
 537
 538 /* ARGSUSED */
 539 static int
 540 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 541 {
 542         if (lrc->lrc_txtype == TX_WRITE) {
 543                 zil_scan_arg_t *zsa = arg;
 544                 dsl_pool_t *dp = zsa->zsa_dp;
 545                 dsl_scan_t *scn = dp->dp_scan;
 546                 zil_header_t *zh = zsa->zsa_zh;
 547                 lr_write_t *lr = (lr_write_t *)lrc;
 548                 blkptr_t *bp = &lr->lr_blkptr;
 549                 zbookmark_phys_t zb;
 550
 551                 if (BP_IS_HOLE(bp) ||
 552                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 553                         return (0);
 554
 555                 /*
 556                  * birth can be < claim_txg if this record's txg is
 557                  * already txg sync'ed (but this log block contains
 558                  * other records that are not synced)
 559                  */
 560                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 561                         return (0);
 562
 563                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 564                     lr->lr_foid, ZB_ZIL_LEVEL,
 565                     lr->lr_offset / BP_GET_LSIZE(bp));
 566
 567                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 568         }
 569         return (0);
 570 }
 571
 572 static void
 573 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 574 {
 575         uint64_t claim_txg = zh->zh_claim_txg;
 576         zil_scan_arg_t zsa = { dp, zh };
 577         zilog_t *zilog;
 578
 579         /*
 580          * We only want to visit blocks that have been claimed but not yet
 581          * replayed (or, in read-only mode, blocks that *would* be claimed).
 582          */
 583         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 584                 return;
 585
 586         zilog = zil_alloc(dp->dp_meta_objset, zh);
 587
 588         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 589             claim_txg);
 590
 591         zil_free(zilog);
 592 }
 593
 594 /* ARGSUSED */
 595 static void
 596 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 597     uint64_t objset, uint64_t object, uint64_t blkid)
 598 {
 599         zbookmark_phys_t czb;
 600         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 601
 602         if (zfs_no_scrub_prefetch)
 603                 return;
 604
 605         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 606             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 607                 return;
 608
 609         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 610
 611         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 612             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 613             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 614 }
 615
 616 static boolean_t
 617 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 618     const zbookmark_phys_t *zb)
 619 {
 620         /*
 621          * We never skip over user/group accounting objects (obj<0)
 622          */
 623         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 624             (int64_t)zb->zb_object >= 0) {
 625                 /*
 626                  * If we already visited this bp & everything below (in
 627                  * a prior txg sync), don't bother doing it again.
 628                  */
 629                 if (zbookmark_subtree_completed(dnp, zb,
 630                     &scn->scn_phys.scn_bookmark))
 631                         return (B_TRUE);
 632
 633                 /*
 634                  * If we found the block we're trying to resume from, or
 635                  * we went past it to a different object, zero it out to
 636                  * indicate that it's OK to start checking for pausing
 637                  * again.
 638                  */
 639                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 640                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 641                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 642                             (longlong_t)zb->zb_objset,
 643                             (longlong_t)zb->zb_object,
 644                             (longlong_t)zb->zb_level,
 645                             (longlong_t)zb->zb_blkid);
 646                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 647                 }
 648         }
 649         return (B_FALSE);
 650 }
 651
 652 /*
 653  * Return nonzero on i/o error.
 654  * Return new buf to write out in *bufp.
 655  */
 656 static int
 657 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 658     dnode_phys_t *dnp, const blkptr_t *bp,
 659     const zbookmark_phys_t *zb, dmu_tx_t *tx)
 660 {
 661         dsl_pool_t *dp = scn->scn_dp;
 662         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 663         int err;
 664
 665         if (BP_GET_LEVEL(bp) > 0) {
 666                 arc_flags_t flags = ARC_FLAG_WAIT;
 667                 int i;
 668                 blkptr_t *cbp;
 669                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 670                 arc_buf_t *buf;
 671
 672                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 673                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 674                 if (err) {
 675                         scn->scn_phys.scn_errors++;
 676                         return (err);
 677                 }
 678                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 679                         dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
 680                             zb->zb_object, zb->zb_blkid * epb + i);
 681                 }
 682                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 683                         zbookmark_phys_t czb;
 684
 685                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 686                             zb->zb_level - 1,
 687                             zb->zb_blkid * epb + i);
 688                         dsl_scan_visitbp(cbp, &czb, dnp,
 689                             ds, scn, ostype, tx);
 690                 }
 691                 arc_buf_destroy(buf, &buf);
 692         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 693                 arc_flags_t flags = ARC_FLAG_WAIT;
 694                 dnode_phys_t *cdnp;
 695                 int i, j;
 696                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 697                 arc_buf_t *buf;
 698
 699                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 700                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 701                 if (err) {
 702                         scn->scn_phys.scn_errors++;
 703                         return (err);
 704                 }
 705                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 706                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 707                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 708                                 dsl_scan_prefetch(scn, buf, cbp,
 709                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 710                         }
 711                 }
 712                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 713                         dsl_scan_visitdnode(scn, ds, ostype,
 714                             cdnp, zb->zb_blkid * epb + i, tx);
 715                 }
 716
 717                 arc_buf_destroy(buf, &buf);
 718         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 719                 arc_flags_t flags = ARC_FLAG_WAIT;
 720                 objset_phys_t *osp;
 721                 arc_buf_t *buf;
 722
 723                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 724                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 725                 if (err) {
 726                         scn->scn_phys.scn_errors++;
 727                         return (err);
 728                 }
 729
 730                 osp = buf->b_data;
 731
 732                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 733                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
 734
 735                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 736                         /*
 737                          * We also always visit user/group accounting
 738                          * objects, and never skip them, even if we are
 739                          * pausing.  This is necessary so that the space
 740                          * deltas from this txg get integrated.
 741                          */
 742                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 743                             &osp->os_groupused_dnode,
 744                             DMU_GROUPUSED_OBJECT, tx);
 745                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 746                             &osp->os_userused_dnode,
 747                             DMU_USERUSED_OBJECT, tx);
 748                 }
 749                 arc_buf_destroy(buf, &buf);
 750         }
 751
 752         return (0);
 753 }
 754
 755 static void
 756 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 757     dmu_objset_type_t ostype, dnode_phys_t *dnp,
 758     uint64_t object, dmu_tx_t *tx)
 759 {
 760         int j;
 761
 762         for (j = 0; j < dnp->dn_nblkptr; j++) {
 763                 zbookmark_phys_t czb;
 764
 765                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 766                     dnp->dn_nlevels - 1, j);
 767                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 768                     &czb, dnp, ds, scn, ostype, tx);
 769         }
 770
 771         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 772                 zbookmark_phys_t czb;
 773                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 774                     0, DMU_SPILL_BLKID);
 775                 dsl_scan_visitbp(&dnp->dn_spill,
 776                     &czb, dnp, ds, scn, ostype, tx);
 777         }
 778 }
 779
 780 /*
 781  * The arguments are in this order because mdb can only print the
 782  * first 5; we want them to be useful.
 783  */
 784 static void
 785 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 786     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 787     dmu_objset_type_t ostype, dmu_tx_t *tx)
 788 {
 789         dsl_pool_t *dp = scn->scn_dp;
 790         arc_buf_t *buf = NULL;
 791         blkptr_t bp_toread = *bp;
 792
 793         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 794
 795         if (dsl_scan_check_pause(scn, zb))
 796                 return;
 797
 798         if (dsl_scan_check_resume(scn, dnp, zb))
 799                 return;
 800
 801         if (BP_IS_HOLE(bp))
 802                 return;
 803
 804         scn->scn_visited_this_txg++;
 805
 806         dprintf_bp(bp,
 807             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
 808             ds, ds ? ds->ds_object : 0,
 809             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 810             bp);
 811
 812         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 813                 return;
 814
 815         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
 816                 return;
 817
 818         /*
 819          * If dsl_scan_ddt() has aready visited this block, it will have
 820          * already done any translations or scrubbing, so don't call the
 821          * callback again.
 822          */
 823         if (ddt_class_contains(dp->dp_spa,
 824             scn->scn_phys.scn_ddt_class_max, bp)) {
 825                 ASSERT(buf == NULL);
 826                 return;
 827         }
 828
 829         /*
 830          * If this block is from the future (after cur_max_txg), then we
 831          * are doing this on behalf of a deleted snapshot, and we will
 832          * revisit the future block on the next pass of this dataset.
 833          * Don't scan it now unless we need to because something
 834          * under it was modified.
 835          */
 836         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 837                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 838         }
 839 }
 840
 841 static void
 842 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 843     dmu_tx_t *tx)
 844 {
 845         zbookmark_phys_t zb;
 846
 847         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 848             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 849         dsl_scan_visitbp(bp, &zb, NULL,
 850             ds, scn, DMU_OST_NONE, tx);
 851
 852         dprintf_ds(ds, "finished scan%s", "");
 853 }
 854
 855 void
 856 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 857 {
 858         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 859         dsl_scan_t *scn = dp->dp_scan;
 860         uint64_t mintxg;
 861
 862         if (scn->scn_phys.scn_state != DSS_SCANNING)
 863                 return;
 864
 865         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 866                 if (ds->ds_is_snapshot) {
 867                         /*
 868                          * Note:
 869                          *  - scn_cur_{min,max}_txg stays the same.
 870                          *  - Setting the flag is not really necessary if
 871                          *    scn_cur_max_txg == scn_max_txg, because there
 872                          *    is nothing after this snapshot that we care
 873                          *    about.  However, we set it anyway and then
 874                          *    ignore it when we retraverse it in
 875                          *    dsl_scan_visitds().
 876                          */
 877                         scn->scn_phys.scn_bookmark.zb_objset =
 878                             dsl_dataset_phys(ds)->ds_next_snap_obj;
 879                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 880                             "reset zb_objset to %llu",
 881                             (u_longlong_t)ds->ds_object,
 882                             (u_longlong_t)dsl_dataset_phys(ds)->
 883                             ds_next_snap_obj);
 884                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 885                 } else {
 886                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 887                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 888                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 889                             "reset bookmark to -1,0,0,0",
 890                             (u_longlong_t)ds->ds_object);
 891                 }
 892         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 893             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 894                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
 895                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 896                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 897                 if (ds->ds_is_snapshot) {
 898                         /*
 899                          * We keep the same mintxg; it could be >
 900                          * ds_creation_txg if the previous snapshot was
 901                          * deleted too.
 902                          */
 903                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 904                             scn->scn_phys.scn_queue_obj,
 905                             dsl_dataset_phys(ds)->ds_next_snap_obj,
 906                             mintxg, tx) == 0);
 907                         zfs_dbgmsg("destroying ds %llu; in queue; "
 908                             "replacing with %llu",
 909                             (u_longlong_t)ds->ds_object,
 910                             (u_longlong_t)dsl_dataset_phys(ds)->
 911                             ds_next_snap_obj);
 912                 } else {
 913                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 914                             (u_longlong_t)ds->ds_object);
 915                 }
 916         }
 917
 918         /*
 919          * dsl_scan_sync() should be called after this, and should sync
 920          * out our changed state, but just to be safe, do it here.
 921          */
 922         dsl_scan_sync_state(scn, tx);
 923 }
 924
 925 void
 926 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 927 {
 928         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 929         dsl_scan_t *scn = dp->dp_scan;
 930         uint64_t mintxg;
 931
 932         if (scn->scn_phys.scn_state != DSS_SCANNING)
 933                 return;
 934
 935         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 936
 937         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 938                 scn->scn_phys.scn_bookmark.zb_objset =
 939                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
 940                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 941                     "reset zb_objset to %llu",
 942                     (u_longlong_t)ds->ds_object,
 943                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 944         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 945             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 946                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 947                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 948                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 949                     scn->scn_phys.scn_queue_obj,
 950                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
 951                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 952                     "replacing with %llu",
 953                     (u_longlong_t)ds->ds_object,
 954                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 955         }
 956         dsl_scan_sync_state(scn, tx);
 957 }
 958
 959 void
 960 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 961 {
 962         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 963         dsl_scan_t *scn = dp->dp_scan;
 964         uint64_t mintxg;
 965
 966         if (scn->scn_phys.scn_state != DSS_SCANNING)
 967                 return;
 968
 969         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 970                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 971                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 972                     "reset zb_objset to %llu",
 973                     (u_longlong_t)ds1->ds_object,
 974                     (u_longlong_t)ds2->ds_object);
 975         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 976                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 977                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 978                     "reset zb_objset to %llu",
 979                     (u_longlong_t)ds2->ds_object,
 980                     (u_longlong_t)ds1->ds_object);
 981         }
 982
 983         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 984             ds1->ds_object, &mintxg) == 0) {
 985                 int err;
 986
 987                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 988                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 989                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 990                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 991                 err = zap_add_int_key(dp->dp_meta_objset,
 992                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 993                 VERIFY(err == 0 || err == EEXIST);
 994                 if (err == EEXIST) {
 995                         /* Both were there to begin with */
 996                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 997                             scn->scn_phys.scn_queue_obj,
 998                             ds1->ds_object, mintxg, tx));
 999                 }
1000                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1001                     "replacing with %llu",
1002                     (u_longlong_t)ds1->ds_object,
1003                     (u_longlong_t)ds2->ds_object);
1004         } else if (zap_lookup_int_key(dp->dp_meta_objset,
1005             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1006                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
1007                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1008                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1009                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1010                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1011                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1012                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1013                     "replacing with %llu",
1014                     (u_longlong_t)ds2->ds_object,
1015                     (u_longlong_t)ds1->ds_object);
1016         }
1017
1018         dsl_scan_sync_state(scn, tx);
1019 }
1020
1021 struct enqueue_clones_arg {
1022         dmu_tx_t *tx;
1023         uint64_t originobj;
1024 };
1025
1026 /* ARGSUSED */
1027 static int
1028 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1029 {
1030         struct enqueue_clones_arg *eca = arg;
1031         dsl_dataset_t *ds;
1032         int err;
1033         dsl_scan_t *scn = dp->dp_scan;
1034
1035         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1036                 return (0);
1037
1038         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1039         if (err)
1040                 return (err);
1041
1042         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1043                 dsl_dataset_t *prev;
1044                 err = dsl_dataset_hold_obj(dp,
1045                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1046
1047                 dsl_dataset_rele(ds, FTAG);
1048                 if (err)
1049                         return (err);
1050                 ds = prev;
1051         }
1052         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1053             scn->scn_phys.scn_queue_obj, ds->ds_object,
1054             dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1055         dsl_dataset_rele(ds, FTAG);
1056         return (0);
1057 }
1058
1059 static void
1060 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1061 {
1062         dsl_pool_t *dp = scn->scn_dp;
1063         dsl_dataset_t *ds;
1064         objset_t *os;
1065
1066         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1067
1068         if (scn->scn_phys.scn_cur_min_txg >=
1069             scn->scn_phys.scn_max_txg) {
1070                 /*
1071                  * This can happen if this snapshot was created after the
1072                  * scan started, and we already completed a previous snapshot
1073                  * that was created after the scan started.  This snapshot
1074                  * only references blocks with:
1075                  *
1076                  *      birth < our ds_creation_txg
1077                  *      cur_min_txg is no less than ds_creation_txg.
1078                  *      We have already visited these blocks.
1079                  * or
1080                  *      birth > scn_max_txg
1081                  *      The scan requested not to visit these blocks.
1082                  *
1083                  * Subsequent snapshots (and clones) can reference our
1084                  * blocks, or blocks with even higher birth times.
1085                  * Therefore we do not need to visit them either,
1086                  * so we do not add them to the work queue.
1087                  *
1088                  * Note that checking for cur_min_txg >= cur_max_txg
1089                  * is not sufficient, because in that case we may need to
1090                  * visit subsequent snapshots.  This happens when min_txg > 0,
1091                  * which raises cur_min_txg.  In this case we will visit
1092                  * this dataset but skip all of its blocks, because the
1093                  * rootbp's birth time is < cur_min_txg.  Then we will
1094                  * add the next snapshots/clones to the work queue.
1095                  */
1096                 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1097                 dsl_dataset_name(ds, dsname);
1098                 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1099                     "cur_min_txg (%llu) >= max_txg (%llu)",
1100                     dsobj, dsname,
1101                     scn->scn_phys.scn_cur_min_txg,
1102                     scn->scn_phys.scn_max_txg);
1103                 kmem_free(dsname, MAXNAMELEN);
1104
1105                 goto out;
1106         }
1107
1108         if (dmu_objset_from_ds(ds, &os))
1109                 goto out;
1110
1111         /*
1112          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1113          * snapshots can have ZIL block pointers (which may be the same
1114          * BP as in the head), they must be ignored.  So we traverse the
1115          * ZIL here, rather than in scan_recurse(), because the regular
1116          * snapshot block-sharing rules don't apply to it.
1117          */
1118         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1119                 dsl_scan_zil(dp, &os->os_zil_header);
1120
1121         /*
1122          * Iterate over the bps in this ds.
1123          */
1124         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1125         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1126         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1127         rrw_exit(&ds->ds_bp_rwlock, FTAG);
1128
1129         char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
1130         dsl_dataset_name(ds, dsname);
1131         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1132             "pausing=%u",
1133             (longlong_t)dsobj, dsname,
1134             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1135             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1136             (int)scn->scn_pausing);
1137         kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
1138
1139         if (scn->scn_pausing)
1140                 goto out;
1141
1142         /*
1143          * We've finished this pass over this dataset.
1144          */
1145
1146         /*
1147          * If we did not completely visit this dataset, do another pass.
1148          */
1149         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1150                 zfs_dbgmsg("incomplete pass; visiting again");
1151                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1152                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1153                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1154                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1155                 goto out;
1156         }
1157
1158         /*
1159          * Add descendent datasets to work queue.
1160          */
1161         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1162                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1163                     scn->scn_phys.scn_queue_obj,
1164                     dsl_dataset_phys(ds)->ds_next_snap_obj,
1165                     dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1166         }
1167         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1168                 boolean_t usenext = B_FALSE;
1169                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1170                         uint64_t count;
1171                         /*
1172                          * A bug in a previous version of the code could
1173                          * cause upgrade_clones_cb() to not set
1174                          * ds_next_snap_obj when it should, leading to a
1175                          * missing entry.  Therefore we can only use the
1176                          * next_clones_obj when its count is correct.
1177                          */
1178                         int err = zap_count(dp->dp_meta_objset,
1179                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1180                         if (err == 0 &&
1181                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
1182                                 usenext = B_TRUE;
1183                 }
1184
1185                 if (usenext) {
1186                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1187                             dsl_dataset_phys(ds)->ds_next_clones_obj,
1188                             scn->scn_phys.scn_queue_obj,
1189                             dsl_dataset_phys(ds)->ds_creation_txg, tx));
1190                 } else {
1191                         struct enqueue_clones_arg eca;
1192                         eca.tx = tx;
1193                         eca.originobj = ds->ds_object;
1194
1195                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1196                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1197                 }
1198         }
1199
1200 out:
1201         dsl_dataset_rele(ds, FTAG);
1202 }
1203
1204 /* ARGSUSED */
1205 static int
1206 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1207 {
1208         dmu_tx_t *tx = arg;
1209         dsl_dataset_t *ds;
1210         int err;
1211         dsl_scan_t *scn = dp->dp_scan;
1212
1213         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1214         if (err)
1215                 return (err);
1216
1217         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1218                 dsl_dataset_t *prev;
1219                 err = dsl_dataset_hold_obj(dp,
1220                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1221                 if (err) {
1222                         dsl_dataset_rele(ds, FTAG);
1223                         return (err);
1224                 }
1225
1226                 /*
1227                  * If this is a clone, we don't need to worry about it for now.
1228                  */
1229                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1230                         dsl_dataset_rele(ds, FTAG);
1231                         dsl_dataset_rele(prev, FTAG);
1232                         return (0);
1233                 }
1234                 dsl_dataset_rele(ds, FTAG);
1235                 ds = prev;
1236         }
1237
1238         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1239             ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1240         dsl_dataset_rele(ds, FTAG);
1241         return (0);
1242 }
1243
1244 /*
1245  * Scrub/dedup interaction.
1246  *
1247  * If there are N references to a deduped block, we don't want to scrub it
1248  * N times -- ideally, we should scrub it exactly once.
1249  *
1250  * We leverage the fact that the dde's replication class (enum ddt_class)
1251  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1252  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1253  *
1254  * To prevent excess scrubbing, the scrub begins by walking the DDT
1255  * to find all blocks with refcnt > 1, and scrubs each of these once.
1256  * Since there are two replication classes which contain blocks with
1257  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1258  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1259  *
1260  * There would be nothing more to say if a block's refcnt couldn't change
1261  * during a scrub, but of course it can so we must account for changes
1262  * in a block's replication class.
1263  *
1264  * Here's an example of what can occur:
1265  *
1266  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1267  * when visited during the top-down scrub phase, it will be scrubbed twice.
1268  * This negates our scrub optimization, but is otherwise harmless.
1269  *
1270  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1271  * on each visit during the top-down scrub phase, it will never be scrubbed.
1272  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1273  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1274  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1275  * while a scrub is in progress, it scrubs the block right then.
1276  */
1277 static void
1278 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1279 {
1280         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1281         ddt_entry_t dde = { 0 };
1282         int error;
1283         uint64_t n = 0;
1284
1285         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1286                 ddt_t *ddt;
1287
1288                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1289                         break;
1290                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1291                     (longlong_t)ddb->ddb_class,
1292                     (longlong_t)ddb->ddb_type,
1293                     (longlong_t)ddb->ddb_checksum,
1294                     (longlong_t)ddb->ddb_cursor);
1295
1296                 /* There should be no pending changes to the dedup table */
1297                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1298                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1299
1300                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1301                 n++;
1302
1303                 if (dsl_scan_check_pause(scn, NULL))
1304                         break;
1305         }
1306
1307         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1308             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1309             (int)scn->scn_pausing);
1310
1311         ASSERT(error == 0 || error == ENOENT);
1312         ASSERT(error != ENOENT ||
1313             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1314 }
1315
1316 /* ARGSUSED */
1317 void
1318 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1319     ddt_entry_t *dde, dmu_tx_t *tx)
1320 {
1321         const ddt_key_t *ddk = &dde->dde_key;
1322         ddt_phys_t *ddp = dde->dde_phys;
1323         blkptr_t bp;
1324         zbookmark_phys_t zb = { 0 };
1325
1326         if (scn->scn_phys.scn_state != DSS_SCANNING)
1327                 return;
1328
1329         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1330                 if (ddp->ddp_phys_birth == 0 ||
1331                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1332                         continue;
1333                 ddt_bp_create(checksum, ddk, ddp, &bp);
1334
1335                 scn->scn_visited_this_txg++;
1336                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1337         }
1338 }
1339
1340 static void
1341 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1342 {
1343         dsl_pool_t *dp = scn->scn_dp;
1344         zap_cursor_t zc;
1345         zap_attribute_t za;
1346
1347         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1348             scn->scn_phys.scn_ddt_class_max) {
1349                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1350                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1351                 dsl_scan_ddt(scn, tx);
1352                 if (scn->scn_pausing)
1353                         return;
1354         }
1355
1356         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1357                 /* First do the MOS & ORIGIN */
1358
1359                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1360                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1361                 dsl_scan_visit_rootbp(scn, NULL,
1362                     &dp->dp_meta_rootbp, tx);
1363                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1364                 if (scn->scn_pausing)
1365                         return;
1366
1367                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1368                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1369                             enqueue_cb, tx, DS_FIND_CHILDREN));
1370                 } else {
1371                         dsl_scan_visitds(scn,
1372                             dp->dp_origin_snap->ds_object, tx);
1373                 }
1374                 ASSERT(!scn->scn_pausing);
1375         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1376             ZB_DESTROYED_OBJSET) {
1377                 /*
1378                  * If we were paused, continue from here.  Note if the
1379                  * ds we were paused on was deleted, the zb_objset may
1380                  * be -1, so we will skip this and find a new objset
1381                  * below.
1382                  */
1383                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1384                 if (scn->scn_pausing)
1385                         return;
1386         }
1387
1388         /*
1389          * In case we were paused right at the end of the ds, zero the
1390          * bookmark so we don't think that we're still trying to resume.
1391          */
1392         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1393
1394         /* keep pulling things out of the zap-object-as-queue */
1395         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1396             scn->scn_phys.scn_queue_obj),
1397             zap_cursor_retrieve(&zc, &za) == 0) {
1398                 dsl_dataset_t *ds;
1399                 uint64_t dsobj;
1400
1401                 dsobj = strtonum(za.za_name, NULL);
1402                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1403                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1404
1405                 /* Set up min/max txg */
1406                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1407                 if (za.za_first_integer != 0) {
1408                         scn->scn_phys.scn_cur_min_txg =
1409                             MAX(scn->scn_phys.scn_min_txg,
1410                             za.za_first_integer);
1411                 } else {
1412                         scn->scn_phys.scn_cur_min_txg =
1413                             MAX(scn->scn_phys.scn_min_txg,
1414                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
1415                 }
1416                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1417                 dsl_dataset_rele(ds, FTAG);
1418
1419                 dsl_scan_visitds(scn, dsobj, tx);
1420                 zap_cursor_fini(&zc);
1421                 if (scn->scn_pausing)
1422                         return;
1423         }
1424         zap_cursor_fini(&zc);
1425 }
1426
1427 static boolean_t
1428 dsl_scan_free_should_pause(dsl_scan_t *scn)
1429 {
1430         uint64_t elapsed_nanosecs;
1431
1432         if (zfs_recover)
1433                 return (B_FALSE);
1434
1435         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1436                 return (B_TRUE);
1437
1438         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1439         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1440             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1441             txg_sync_waiting(scn->scn_dp)) ||
1442             spa_shutting_down(scn->scn_dp->dp_spa));
1443 }
1444
1445 static int
1446 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1447 {
1448         dsl_scan_t *scn = arg;
1449
1450         if (!scn->scn_is_bptree ||
1451             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1452                 if (dsl_scan_free_should_pause(scn))
1453                         return (SET_ERROR(ERESTART));
1454         }
1455
1456         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1457             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1458         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1459             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1460             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1461         scn->scn_visited_this_txg++;
1462         return (0);
1463 }
1464
1465 boolean_t
1466 dsl_scan_active(dsl_scan_t *scn)
1467 {
1468         spa_t *spa = scn->scn_dp->dp_spa;
1469         uint64_t used = 0, comp, uncomp;
1470
1471         if (spa->spa_load_state != SPA_LOAD_NONE)
1472                 return (B_FALSE);
1473         if (spa_shutting_down(spa))
1474                 return (B_FALSE);
1475         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1476             (scn->scn_async_destroying && !scn->scn_async_stalled))
1477                 return (B_TRUE);
1478
1479         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1480                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1481                     &used, &comp, &uncomp);
1482         }
1483         return (used != 0);
1484 }
1485
1486 void
1487 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1488 {
1489         dsl_scan_t *scn = dp->dp_scan;
1490         spa_t *spa = dp->dp_spa;
1491         int err = 0;
1492
1493         /*
1494          * Check for scn_restart_txg before checking spa_load_state, so
1495          * that we can restart an old-style scan while the pool is being
1496          * imported (see dsl_scan_init).
1497          */
1498         if (dsl_scan_restarting(scn, tx)) {
1499                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1500                 dsl_scan_done(scn, B_FALSE, tx);
1501                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1502                         func = POOL_SCAN_RESILVER;
1503                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1504                     func, tx->tx_txg);
1505                 dsl_scan_setup_sync(&func, tx);
1506         }
1507
1508         /*
1509          * Only process scans in sync pass 1.
1510          */
1511         if (spa_sync_pass(dp->dp_spa) > 1)
1512                 return;
1513
1514         /*
1515          * If the spa is shutting down, then stop scanning. This will
1516          * ensure that the scan does not dirty any new data during the
1517          * shutdown phase.
1518          */
1519         if (spa_shutting_down(spa))
1520                 return;
1521
1522         /*
1523          * If the scan is inactive due to a stalled async destroy, try again.
1524          */
1525         if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1526                 return;
1527
1528         scn->scn_visited_this_txg = 0;
1529         scn->scn_pausing = B_FALSE;
1530         scn->scn_sync_start_time = gethrtime();
1531         spa->spa_scrub_active = B_TRUE;
1532
1533         /*
1534          * First process the async destroys.  If we pause, don't do
1535          * any scrubbing or resilvering.  This ensures that there are no
1536          * async destroys while we are scanning, so the scan code doesn't
1537          * have to worry about traversing it.  It is also faster to free the
1538          * blocks than to scrub them.
1539          */
1540         if (zfs_free_bpobj_enabled &&
1541             spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1542                 scn->scn_is_bptree = B_FALSE;
1543                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1544                     NULL, ZIO_FLAG_MUSTSUCCEED);
1545                 err = bpobj_iterate(&dp->dp_free_bpobj,
1546                     dsl_scan_free_block_cb, scn, tx);
1547                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1548
1549                 if (err != 0 && err != ERESTART)
1550                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1551         }
1552
1553         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1554                 ASSERT(scn->scn_async_destroying);
1555                 scn->scn_is_bptree = B_TRUE;
1556                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1557                     NULL, ZIO_FLAG_MUSTSUCCEED);
1558                 err = bptree_iterate(dp->dp_meta_objset,
1559                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1560                 VERIFY0(zio_wait(scn->scn_zio_root));
1561
1562                 if (err == EIO || err == ECKSUM) {
1563                         err = 0;
1564                 } else if (err != 0 && err != ERESTART) {
1565                         zfs_panic_recover("error %u from "
1566                             "traverse_dataset_destroyed()", err);
1567                 }
1568
1569                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1570                         /* finished; deactivate async destroy feature */
1571                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1572                         ASSERT(!spa_feature_is_active(spa,
1573                             SPA_FEATURE_ASYNC_DESTROY));
1574                         VERIFY0(zap_remove(dp->dp_meta_objset,
1575                             DMU_POOL_DIRECTORY_OBJECT,
1576                             DMU_POOL_BPTREE_OBJ, tx));
1577                         VERIFY0(bptree_free(dp->dp_meta_objset,
1578                             dp->dp_bptree_obj, tx));
1579                         dp->dp_bptree_obj = 0;
1580                         scn->scn_async_destroying = B_FALSE;
1581                         scn->scn_async_stalled = B_FALSE;
1582                 } else {
1583                         /*
1584                          * If we didn't make progress, mark the async
1585                          * destroy as stalled, so that we will not initiate
1586                          * a spa_sync() on its behalf.  Note that we only
1587                          * check this if we are not finished, because if the
1588                          * bptree had no blocks for us to visit, we can
1589                          * finish without "making progress".
1590                          */
1591                         scn->scn_async_stalled =
1592                             (scn->scn_visited_this_txg == 0);
1593                 }
1594         }
1595         if (scn->scn_visited_this_txg) {
1596                 zfs_dbgmsg("freed %llu blocks in %llums from "
1597                     "free_bpobj/bptree txg %llu; err=%d",
1598                     (longlong_t)scn->scn_visited_this_txg,
1599                     (longlong_t)
1600                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1601                     (longlong_t)tx->tx_txg, err);
1602                 scn->scn_visited_this_txg = 0;
1603
1604                 /*
1605                  * Write out changes to the DDT that may be required as a
1606                  * result of the blocks freed.  This ensures that the DDT
1607                  * is clean when a scrub/resilver runs.
1608                  */
1609                 ddt_sync(spa, tx->tx_txg);
1610         }
1611         if (err != 0)
1612                 return;
1613         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1614             zfs_free_leak_on_eio &&
1615             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1616             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1617             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1618                 /*
1619                  * We have finished background destroying, but there is still
1620                  * some space left in the dp_free_dir. Transfer this leaked
1621                  * space to the dp_leak_dir.
1622                  */
1623                 if (dp->dp_leak_dir == NULL) {
1624                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1625                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1626                             LEAK_DIR_NAME, tx);
1627                         VERIFY0(dsl_pool_open_special_dir(dp,
1628                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1629                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1630                 }
1631                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1632                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1633                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1634                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1635                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1636                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1637                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1638                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1639         }
1640         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1641                 /* finished; verify that space accounting went to zero */
1642                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1643                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1644                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1645         }
1646
1647         if (scn->scn_phys.scn_state != DSS_SCANNING)
1648                 return;
1649
1650         if (scn->scn_done_txg == tx->tx_txg) {
1651                 ASSERT(!scn->scn_pausing);
1652                 /* finished with scan. */
1653                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1654                 dsl_scan_done(scn, B_TRUE, tx);
1655                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1656                 dsl_scan_sync_state(scn, tx);
1657                 return;
1658         }
1659
1660         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1661             scn->scn_phys.scn_ddt_class_max) {
1662                 zfs_dbgmsg("doing scan sync txg %llu; "
1663                     "ddt bm=%llu/%llu/%llu/%llx",
1664                     (longlong_t)tx->tx_txg,
1665                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1666                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1667                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1668                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1669                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1670                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1671                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1672                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1673         } else {
1674                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1675                     (longlong_t)tx->tx_txg,
1676                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1677                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1678                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1679                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1680         }
1681
1682         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1683             NULL, ZIO_FLAG_CANFAIL);
1684         dsl_pool_config_enter(dp, FTAG);
1685         dsl_scan_visit(scn, tx);
1686         dsl_pool_config_exit(dp, FTAG);
1687         (void) zio_wait(scn->scn_zio_root);
1688         scn->scn_zio_root = NULL;
1689
1690         zfs_dbgmsg("visited %llu blocks in %llums",
1691             (longlong_t)scn->scn_visited_this_txg,
1692             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1693
1694         if (!scn->scn_pausing) {
1695                 scn->scn_done_txg = tx->tx_txg + 1;
1696                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1697                     tx->tx_txg, scn->scn_done_txg);
1698         }
1699
1700         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1701                 mutex_enter(&spa->spa_scrub_lock);
1702                 while (spa->spa_scrub_inflight > 0) {
1703                         cv_wait(&spa->spa_scrub_io_cv,
1704                             &spa->spa_scrub_lock);
1705                 }
1706                 mutex_exit(&spa->spa_scrub_lock);
1707         }
1708
1709         dsl_scan_sync_state(scn, tx);
1710 }
1711
1712 /*
1713  * This will start a new scan, or restart an existing one.
1714  */
1715 void
1716 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1717 {
1718         if (txg == 0) {
1719                 dmu_tx_t *tx;
1720                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1721                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1722
1723                 txg = dmu_tx_get_txg(tx);
1724                 dp->dp_scan->scn_restart_txg = txg;
1725                 dmu_tx_commit(tx);
1726         } else {
1727                 dp->dp_scan->scn_restart_txg = txg;
1728         }
1729         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1730 }
1731
1732 boolean_t
1733 dsl_scan_resilvering(dsl_pool_t *dp)
1734 {
1735         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1736             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1737 }
1738
1739 /*
1740  * scrub consumers
1741  */
1742
1743 static void
1744 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1745 {
1746         int i;
1747
1748         /*
1749          * If we resume after a reboot, zab will be NULL; don't record
1750          * incomplete stats in that case.
1751          */
1752         if (zab == NULL)
1753                 return;
1754
1755         for (i = 0; i < 4; i++) {
1756                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1757                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1758                 if (t & DMU_OT_NEWTYPE)
1759                         t = DMU_OT_OTHER;
1760                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1761                 int equal;
1762
1763                 zb->zb_count++;
1764                 zb->zb_asize += BP_GET_ASIZE(bp);
1765                 zb->zb_lsize += BP_GET_LSIZE(bp);
1766                 zb->zb_psize += BP_GET_PSIZE(bp);
1767                 zb->zb_gangs += BP_COUNT_GANG(bp);
1768
1769                 switch (BP_GET_NDVAS(bp)) {
1770                 case 2:
1771                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1772                             DVA_GET_VDEV(&bp->blk_dva[1]))
1773                                 zb->zb_ditto_2_of_2_samevdev++;
1774                         break;
1775                 case 3:
1776                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1777                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1778                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1779                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1780                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1781                             DVA_GET_VDEV(&bp->blk_dva[2]));
1782                         if (equal == 1)
1783                                 zb->zb_ditto_2_of_3_samevdev++;
1784                         else if (equal == 3)
1785                                 zb->zb_ditto_3_of_3_samevdev++;
1786                         break;
1787                 }
1788         }
1789 }
1790
1791 static void
1792 dsl_scan_scrub_done(zio_t *zio)
1793 {
1794         spa_t *spa = zio->io_spa;
1795
1796         zio_data_buf_free(zio->io_data, zio->io_size);
1797
1798         mutex_enter(&spa->spa_scrub_lock);
1799         spa->spa_scrub_inflight--;
1800         cv_broadcast(&spa->spa_scrub_io_cv);
1801
1802         if (zio->io_error && (zio->io_error != ECKSUM ||
1803             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1804                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1805         }
1806         mutex_exit(&spa->spa_scrub_lock);
1807 }
1808
1809 static int
1810 dsl_scan_scrub_cb(dsl_pool_t *dp,
1811     const blkptr_t *bp, const zbookmark_phys_t *zb)
1812 {
1813         dsl_scan_t *scn = dp->dp_scan;
1814         size_t size = BP_GET_PSIZE(bp);
1815         spa_t *spa = dp->dp_spa;
1816         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1817         boolean_t needs_io;
1818         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1819         unsigned int scan_delay = 0;
1820
1821         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1822             phys_birth >= scn->scn_phys.scn_max_txg)
1823                 return (0);
1824
1825         count_block(dp->dp_blkstats, bp);
1826
1827         if (BP_IS_EMBEDDED(bp))
1828                 return (0);
1829
1830         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1831         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1832                 zio_flags |= ZIO_FLAG_SCRUB;
1833                 needs_io = B_TRUE;
1834                 scan_delay = zfs_scrub_delay;
1835         } else {
1836                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1837                 zio_flags |= ZIO_FLAG_RESILVER;
1838                 needs_io = B_FALSE;
1839                 scan_delay = zfs_resilver_delay;
1840         }
1841
1842         /* If it's an intent log block, failure is expected. */
1843         if (zb->zb_level == ZB_ZIL_LEVEL)
1844                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1845
1846         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1847                 vdev_t *vd = vdev_lookup_top(spa,
1848                     DVA_GET_VDEV(&bp->blk_dva[d]));
1849
1850                 /*
1851                  * Keep track of how much data we've examined so that
1852                  * zpool(1M) status can make useful progress reports.
1853                  */
1854                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1855                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1856
1857                 /* if it's a resilver, this may not be in the target range */
1858                 if (!needs_io) {
1859                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1860                                 /*
1861                                  * Gang members may be spread across multiple
1862                                  * vdevs, so the best estimate we have is the
1863                                  * scrub range, which has already been checked.
1864                                  * XXX -- it would be better to change our
1865                                  * allocation policy to ensure that all
1866                                  * gang members reside on the same vdev.
1867                                  */
1868                                 needs_io = B_TRUE;
1869                         } else {
1870                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1871                                     phys_birth, 1);
1872                         }
1873                 }
1874         }
1875
1876         if (needs_io && !zfs_no_scrub_io) {
1877                 vdev_t *rvd = spa->spa_root_vdev;
1878                 uint64_t maxinflight = rvd->vdev_children *
1879                     MAX(zfs_top_maxinflight, 1);
1880                 void *data = zio_data_buf_alloc(size);
1881
1882                 mutex_enter(&spa->spa_scrub_lock);
1883                 while (spa->spa_scrub_inflight >= maxinflight)
1884                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1885                 spa->spa_scrub_inflight++;
1886                 mutex_exit(&spa->spa_scrub_lock);
1887
1888                 /*
1889                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1890                  * then throttle our workload to limit the impact of a scan.
1891                  */
1892                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1893                         delay(MAX((int)scan_delay, 0));
1894
1895                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1896                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1897                     zio_flags, zb));
1898         }
1899
1900         /* do not relocate this block */
1901         return (0);
1902 }
1903
1904 int
1905 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1906 {
1907         spa_t *spa = dp->dp_spa;
1908
1909         /*
1910          * Purge all vdev caches and probe all devices.  We do this here
1911          * rather than in sync context because this requires a writer lock
1912          * on the spa_config lock, which we can't do from sync context.  The
1913          * spa_scrub_reopen flag indicates that vdev_open() should not
1914          * attempt to start another scrub.
1915          */
1916         spa_vdev_state_enter(spa, SCL_NONE);
1917         spa->spa_scrub_reopen = B_TRUE;
1918         vdev_reopen(spa->spa_root_vdev);
1919         spa->spa_scrub_reopen = B_FALSE;
1920         (void) spa_vdev_state_exit(spa, NULL, 0);
1921
1922         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1923             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1924 }
1925
1926 static boolean_t
1927 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
1928 {
1929         return (scn->scn_restart_txg != 0 &&
1930             scn->scn_restart_txg <= tx->tx_txg);
1931 }