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) 2012 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/dsl_scan.h>
  27 #include <sys/dsl_pool.h>
  28 #include <sys/dsl_dataset.h>
  29 #include <sys/dsl_prop.h>
  30 #include <sys/dsl_dir.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/dnode.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/arc.h>
  36 #include <sys/zap.h>
  37 #include <sys/zio.h>
  38 #include <sys/zfs_context.h>
  39 #include <sys/fs/zfs.h>
  40 #include <sys/zfs_znode.h>
  41 #include <sys/spa_impl.h>
  42 #include <sys/vdev_impl.h>
  43 #include <sys/zil_impl.h>
  44 #include <sys/zio_checksum.h>
  45 #include <sys/ddt.h>
  46 #include <sys/sa.h>
  47 #include <sys/sa_impl.h>
  48 #include <sys/zfeature.h>
  49 #ifdef _KERNEL
  50 #include <sys/zfs_vfsops.h>
  51 #endif
  52
  53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  54
  55 static scan_cb_t dsl_scan_defrag_cb;
  56 static scan_cb_t dsl_scan_scrub_cb;
  57 static scan_cb_t dsl_scan_remove_cb;
  58 static dsl_syncfunc_t dsl_scan_cancel_sync;
  59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  60
  61 unsigned int zfs_top_maxinflight = 32;  /* maximum I/Os per top-level */
  62 unsigned int zfs_resilver_delay = 2;    /* number of ticks to delay resilver */
  63 unsigned int zfs_scrub_delay = 4;       /* number of ticks to delay scrub */
  64 unsigned int zfs_scan_idle = 50;        /* idle window in clock ticks */
  65
  66 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  67 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  68 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
  69                                                  per txg */
  70 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  71 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
  72
  73 SYSCTL_DECL(_vfs_zfs);
  74 TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
  75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
  76     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
  77 TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
  78 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
  79     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
  80 TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
  81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
  82     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
  83 TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
  84 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
  85     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
  86 TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
  87 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
  88     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
  89 TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
  90 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
  91     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
  92 TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
  93 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
  94     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
  95 TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
  96 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
  97     &zfs_no_scrub_io, 0, "Disable scrub I/O");
  98 TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
  99 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
 100     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
 101
 102 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
 103
 104 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 105         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 106         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 107
 108 extern int zfs_txg_timeout;
 109
 110 /* the order has to match pool_scan_type */
 111 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 112         NULL,
 113         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 114         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 115 };
 116
 117 int
 118 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 119 {
 120         int err;
 121         dsl_scan_t *scn;
 122         spa_t *spa = dp->dp_spa;
 123         uint64_t f;
 124
 125         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 126         scn->scn_dp = dp;
 127
 128         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 129             "scrub_func", sizeof (uint64_t), 1, &f);
 130         if (err == 0) {
 131                 /*
 132                  * There was an old-style scrub in progress.  Restart a
 133                  * new-style scrub from the beginning.
 134                  */
 135                 scn->scn_restart_txg = txg;
 136                 zfs_dbgmsg("old-style scrub was in progress; "
 137                     "restarting new-style scrub in txg %llu",
 138                     scn->scn_restart_txg);
 139
 140                 /*
 141                  * Load the queue obj from the old location so that it
 142                  * can be freed by dsl_scan_done().
 143                  */
 144                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 145                     "scrub_queue", sizeof (uint64_t), 1,
 146                     &scn->scn_phys.scn_queue_obj);
 147         } else {
 148                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 149                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 150                     &scn->scn_phys);
 151                 if (err == ENOENT)
 152                         return (0);
 153                 else if (err)
 154                         return (err);
 155
 156                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 157                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 158                         /*
 159                          * A new-type scrub was in progress on an old
 160                          * pool, and the pool was accessed by old
 161                          * software.  Restart from the beginning, since
 162                          * the old software may have changed the pool in
 163                          * the meantime.
 164                          */
 165                         scn->scn_restart_txg = txg;
 166                         zfs_dbgmsg("new-style scrub was modified "
 167                             "by old software; restarting in txg %llu",
 168                             scn->scn_restart_txg);
 169                 }
 170         }
 171
 172         spa_scan_stat_init(spa);
 173         return (0);
 174 }
 175
 176 void
 177 dsl_scan_fini(dsl_pool_t *dp)
 178 {
 179         if (dp->dp_scan) {
 180                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 181                 dp->dp_scan = NULL;
 182         }
 183 }
 184
 185 /* ARGSUSED */
 186 static int
 187 dsl_scan_setup_check(void *arg1, void *arg2, dmu_tx_t *tx)
 188 {
 189         dsl_scan_t *scn = arg1;
 190
 191         if (scn->scn_phys.scn_state == DSS_SCANNING)
 192                 return (EBUSY);
 193
 194         return (0);
 195 }
 196
 197 /* ARGSUSED */
 198 static void
 199 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 200 {
 201         dsl_scan_t *scn = arg1;
 202         pool_scan_func_t *funcp = arg2;
 203         dmu_object_type_t ot = 0;
 204         dsl_pool_t *dp = scn->scn_dp;
 205         spa_t *spa = dp->dp_spa;
 206
 207         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 208         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 209         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 210         scn->scn_phys.scn_func = *funcp;
 211         scn->scn_phys.scn_state = DSS_SCANNING;
 212         scn->scn_phys.scn_min_txg = 0;
 213         scn->scn_phys.scn_max_txg = tx->tx_txg;
 214         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 215         scn->scn_phys.scn_start_time = gethrestime_sec();
 216         scn->scn_phys.scn_errors = 0;
 217         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 218         scn->scn_restart_txg = 0;
 219         spa_scan_stat_init(spa);
 220
 221         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 222                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 223
 224                 /* rewrite all disk labels */
 225                 vdev_config_dirty(spa->spa_root_vdev);
 226
 227                 if (vdev_resilver_needed(spa->spa_root_vdev,
 228                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 229                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 230                 } else {
 231                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 232                 }
 233
 234                 spa->spa_scrub_started = B_TRUE;
 235                 /*
 236                  * If this is an incremental scrub, limit the DDT scrub phase
 237                  * to just the auto-ditto class (for correctness); the rest
 238                  * of the scrub should go faster using top-down pruning.
 239                  */
 240                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 241                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 242
 243         }
 244
 245         /* back to the generic stuff */
 246
 247         if (dp->dp_blkstats == NULL) {
 248                 dp->dp_blkstats =
 249                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 250         }
 251         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 252
 253         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 254                 ot = DMU_OT_ZAP_OTHER;
 255
 256         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 257             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 258
 259         dsl_scan_sync_state(scn, tx);
 260
 261         spa_history_log_internal(LOG_POOL_SCAN, spa, tx,
 262             "func=%u mintxg=%llu maxtxg=%llu",
 263             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 264 }
 265
 266 /* ARGSUSED */
 267 static void
 268 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 269 {
 270         static const char *old_names[] = {
 271                 "scrub_bookmark",
 272                 "scrub_ddt_bookmark",
 273                 "scrub_ddt_class_max",
 274                 "scrub_queue",
 275                 "scrub_min_txg",
 276                 "scrub_max_txg",
 277                 "scrub_func",
 278                 "scrub_errors",
 279                 NULL
 280         };
 281
 282         dsl_pool_t *dp = scn->scn_dp;
 283         spa_t *spa = dp->dp_spa;
 284         int i;
 285
 286         /* Remove any remnants of an old-style scrub. */
 287         for (i = 0; old_names[i]; i++) {
 288                 (void) zap_remove(dp->dp_meta_objset,
 289                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 290         }
 291
 292         if (scn->scn_phys.scn_queue_obj != 0) {
 293                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 294                     scn->scn_phys.scn_queue_obj, tx));
 295                 scn->scn_phys.scn_queue_obj = 0;
 296         }
 297
 298         /*
 299          * If we were "restarted" from a stopped state, don't bother
 300          * with anything else.
 301          */
 302         if (scn->scn_phys.scn_state != DSS_SCANNING)
 303                 return;
 304
 305         if (complete)
 306                 scn->scn_phys.scn_state = DSS_FINISHED;
 307         else
 308                 scn->scn_phys.scn_state = DSS_CANCELED;
 309
 310         spa_history_log_internal(LOG_POOL_SCAN_DONE, spa, tx,
 311             "complete=%u", complete);
 312
 313         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 314                 mutex_enter(&spa->spa_scrub_lock);
 315                 while (spa->spa_scrub_inflight > 0) {
 316                         cv_wait(&spa->spa_scrub_io_cv,
 317                             &spa->spa_scrub_lock);
 318                 }
 319                 mutex_exit(&spa->spa_scrub_lock);
 320                 spa->spa_scrub_started = B_FALSE;
 321                 spa->spa_scrub_active = B_FALSE;
 322
 323                 /*
 324                  * If the scrub/resilver completed, update all DTLs to
 325                  * reflect this.  Whether it succeeded or not, vacate
 326                  * all temporary scrub DTLs.
 327                  */
 328                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 329                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 330                 if (complete) {
 331                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 332                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 333                 }
 334                 spa_errlog_rotate(spa);
 335
 336                 /*
 337                  * We may have finished replacing a device.
 338                  * Let the async thread assess this and handle the detach.
 339                  */
 340                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 341         }
 342
 343         scn->scn_phys.scn_end_time = gethrestime_sec();
 344 }
 345
 346 /* ARGSUSED */
 347 static int
 348 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
 349 {
 350         dsl_scan_t *scn = arg1;
 351
 352         if (scn->scn_phys.scn_state != DSS_SCANNING)
 353                 return (ENOENT);
 354         return (0);
 355 }
 356
 357 /* ARGSUSED */
 358 static void
 359 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 360 {
 361         dsl_scan_t *scn = arg1;
 362
 363         dsl_scan_done(scn, B_FALSE, tx);
 364         dsl_scan_sync_state(scn, tx);
 365 }
 366
 367 int
 368 dsl_scan_cancel(dsl_pool_t *dp)
 369 {
 370         boolean_t complete = B_FALSE;
 371         int err;
 372
 373         err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
 374             dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
 375         return (err);
 376 }
 377
 378 static void dsl_scan_visitbp(blkptr_t *bp,
 379     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 380     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 381     dmu_tx_t *tx);
 382 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 383     dmu_objset_type_t ostype,
 384     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 385
 386 void
 387 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 388 {
 389         zio_free(dp->dp_spa, txg, bp);
 390 }
 391
 392 void
 393 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 394 {
 395         ASSERT(dsl_pool_sync_context(dp));
 396         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 397             pio->io_flags));
 398 }
 399
 400 int
 401 dsl_read(zio_t *pio, spa_t *spa, const blkptr_t *bpp, arc_buf_t *pbuf,
 402     arc_done_func_t *done, void *private, int priority, int zio_flags,
 403     uint32_t *arc_flags, const zbookmark_t *zb)
 404 {
 405         return (arc_read(pio, spa, bpp, pbuf, done, private,
 406             priority, zio_flags, arc_flags, zb));
 407 }
 408
 409 int
 410 dsl_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bpp,
 411     arc_done_func_t *done, void *private, int priority, int zio_flags,
 412     uint32_t *arc_flags, const zbookmark_t *zb)
 413 {
 414         return (arc_read_nolock(pio, spa, bpp, done, private,
 415             priority, zio_flags, arc_flags, zb));
 416 }
 417
 418 static uint64_t
 419 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 420 {
 421         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 422         if (dsl_dataset_is_snapshot(ds))
 423                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 424         return (smt);
 425 }
 426
 427 static void
 428 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 429 {
 430         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 431             DMU_POOL_DIRECTORY_OBJECT,
 432             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 433             &scn->scn_phys, tx));
 434 }
 435
 436 static boolean_t
 437 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 438 {
 439         uint64_t elapsed_nanosecs;
 440         unsigned int mintime;
 441
 442         /* we never skip user/group accounting objects */
 443         if (zb && (int64_t)zb->zb_object < 0)
 444                 return (B_FALSE);
 445
 446         if (scn->scn_pausing)
 447                 return (B_TRUE); /* we're already pausing */
 448
 449         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 450                 return (B_FALSE); /* we're resuming */
 451
 452         /* We only know how to resume from level-0 blocks. */
 453         if (zb && zb->zb_level != 0)
 454                 return (B_FALSE);
 455
 456         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 457             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 458         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 459         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 460             (elapsed_nanosecs / MICROSEC > mintime &&
 461             txg_sync_waiting(scn->scn_dp)) ||
 462             spa_shutting_down(scn->scn_dp->dp_spa)) {
 463                 if (zb) {
 464                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 465                             (longlong_t)zb->zb_objset,
 466                             (longlong_t)zb->zb_object,
 467                             (longlong_t)zb->zb_level,
 468                             (longlong_t)zb->zb_blkid);
 469                         scn->scn_phys.scn_bookmark = *zb;
 470                 }
 471                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 472                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 473                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 474                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 475                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 476                 scn->scn_pausing = B_TRUE;
 477                 return (B_TRUE);
 478         }
 479         return (B_FALSE);
 480 }
 481
 482 typedef struct zil_scan_arg {
 483         dsl_pool_t      *zsa_dp;
 484         zil_header_t    *zsa_zh;
 485 } zil_scan_arg_t;
 486
 487 /* ARGSUSED */
 488 static int
 489 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 490 {
 491         zil_scan_arg_t *zsa = arg;
 492         dsl_pool_t *dp = zsa->zsa_dp;
 493         dsl_scan_t *scn = dp->dp_scan;
 494         zil_header_t *zh = zsa->zsa_zh;
 495         zbookmark_t zb;
 496
 497         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 498                 return (0);
 499
 500         /*
 501          * One block ("stubby") can be allocated a long time ago; we
 502          * want to visit that one because it has been allocated
 503          * (on-disk) even if it hasn't been claimed (even though for
 504          * scrub there's nothing to do to it).
 505          */
 506         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 507                 return (0);
 508
 509         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 510             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 511
 512         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 513         return (0);
 514 }
 515
 516 /* ARGSUSED */
 517 static int
 518 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 519 {
 520         if (lrc->lrc_txtype == TX_WRITE) {
 521                 zil_scan_arg_t *zsa = arg;
 522                 dsl_pool_t *dp = zsa->zsa_dp;
 523                 dsl_scan_t *scn = dp->dp_scan;
 524                 zil_header_t *zh = zsa->zsa_zh;
 525                 lr_write_t *lr = (lr_write_t *)lrc;
 526                 blkptr_t *bp = &lr->lr_blkptr;
 527                 zbookmark_t zb;
 528
 529                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 530                         return (0);
 531
 532                 /*
 533                  * birth can be < claim_txg if this record's txg is
 534                  * already txg sync'ed (but this log block contains
 535                  * other records that are not synced)
 536                  */
 537                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 538                         return (0);
 539
 540                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 541                     lr->lr_foid, ZB_ZIL_LEVEL,
 542                     lr->lr_offset / BP_GET_LSIZE(bp));
 543
 544                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 545         }
 546         return (0);
 547 }
 548
 549 static void
 550 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 551 {
 552         uint64_t claim_txg = zh->zh_claim_txg;
 553         zil_scan_arg_t zsa = { dp, zh };
 554         zilog_t *zilog;
 555
 556         /*
 557          * We only want to visit blocks that have been claimed but not yet
 558          * replayed (or, in read-only mode, blocks that *would* be claimed).
 559          */
 560         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 561                 return;
 562
 563         zilog = zil_alloc(dp->dp_meta_objset, zh);
 564
 565         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 566             claim_txg);
 567
 568         zil_free(zilog);
 569 }
 570
 571 /* ARGSUSED */
 572 static void
 573 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 574     uint64_t objset, uint64_t object, uint64_t blkid)
 575 {
 576         zbookmark_t czb;
 577         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 578
 579         if (zfs_no_scrub_prefetch)
 580                 return;
 581
 582         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 583             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 584                 return;
 585
 586         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 587
 588         /*
 589          * XXX need to make sure all of these arc_read() prefetches are
 590          * done before setting xlateall (similar to dsl_read())
 591          */
 592         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 593             buf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 594             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 595 }
 596
 597 static boolean_t
 598 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 599     const zbookmark_t *zb)
 600 {
 601         /*
 602          * We never skip over user/group accounting objects (obj<0)
 603          */
 604         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 605             (int64_t)zb->zb_object >= 0) {
 606                 /*
 607                  * If we already visited this bp & everything below (in
 608                  * a prior txg sync), don't bother doing it again.
 609                  */
 610                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 611                         return (B_TRUE);
 612
 613                 /*
 614                  * If we found the block we're trying to resume from, or
 615                  * we went past it to a different object, zero it out to
 616                  * indicate that it's OK to start checking for pausing
 617                  * again.
 618                  */
 619                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 620                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 621                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 622                             (longlong_t)zb->zb_objset,
 623                             (longlong_t)zb->zb_object,
 624                             (longlong_t)zb->zb_level,
 625                             (longlong_t)zb->zb_blkid);
 626                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 627                 }
 628         }
 629         return (B_FALSE);
 630 }
 631
 632 /*
 633  * Return nonzero on i/o error.
 634  * Return new buf to write out in *bufp.
 635  */
 636 static int
 637 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 638     dnode_phys_t *dnp, const blkptr_t *bp,
 639     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 640 {
 641         dsl_pool_t *dp = scn->scn_dp;
 642         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 643         int err;
 644
 645         if (BP_GET_LEVEL(bp) > 0) {
 646                 uint32_t flags = ARC_WAIT;
 647                 int i;
 648                 blkptr_t *cbp;
 649                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 650
 651                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 652                     arc_getbuf_func, bufp,
 653                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 654                 if (err) {
 655                         scn->scn_phys.scn_errors++;
 656                         return (err);
 657                 }
 658                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 659                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 660                             zb->zb_object, zb->zb_blkid * epb + i);
 661                 }
 662                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 663                         zbookmark_t czb;
 664
 665                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 666                             zb->zb_level - 1,
 667                             zb->zb_blkid * epb + i);
 668                         dsl_scan_visitbp(cbp, &czb, dnp,
 669                             *bufp, ds, scn, ostype, tx);
 670                 }
 671         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 672                 uint32_t flags = ARC_WAIT;
 673
 674                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 675                     arc_getbuf_func, bufp,
 676                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 677                 if (err) {
 678                         scn->scn_phys.scn_errors++;
 679                         return (err);
 680                 }
 681         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 682                 uint32_t flags = ARC_WAIT;
 683                 dnode_phys_t *cdnp;
 684                 int i, j;
 685                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 686
 687                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 688                     arc_getbuf_func, bufp,
 689                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 690                 if (err) {
 691                         scn->scn_phys.scn_errors++;
 692                         return (err);
 693                 }
 694                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 695                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 696                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 697                                 dsl_scan_prefetch(scn, *bufp, cbp,
 698                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 699                         }
 700                 }
 701                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 702                         dsl_scan_visitdnode(scn, ds, ostype,
 703                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 704                 }
 705
 706         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 707                 uint32_t flags = ARC_WAIT;
 708                 objset_phys_t *osp;
 709
 710                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 711                     arc_getbuf_func, bufp,
 712                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 713                 if (err) {
 714                         scn->scn_phys.scn_errors++;
 715                         return (err);
 716                 }
 717
 718                 osp = (*bufp)->b_data;
 719
 720                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 721                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 722
 723                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 724                         /*
 725                          * We also always visit user/group accounting
 726                          * objects, and never skip them, even if we are
 727                          * pausing.  This is necessary so that the space
 728                          * deltas from this txg get integrated.
 729                          */
 730                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 731                             &osp->os_groupused_dnode, *bufp,
 732                             DMU_GROUPUSED_OBJECT, tx);
 733                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 734                             &osp->os_userused_dnode, *bufp,
 735                             DMU_USERUSED_OBJECT, tx);
 736                 }
 737         }
 738
 739         return (0);
 740 }
 741
 742 static void
 743 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 744     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 745     uint64_t object, dmu_tx_t *tx)
 746 {
 747         int j;
 748
 749         for (j = 0; j < dnp->dn_nblkptr; j++) {
 750                 zbookmark_t czb;
 751
 752                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 753                     dnp->dn_nlevels - 1, j);
 754                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 755                     &czb, dnp, buf, ds, scn, ostype, tx);
 756         }
 757
 758         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 759                 zbookmark_t czb;
 760                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 761                     0, DMU_SPILL_BLKID);
 762                 dsl_scan_visitbp(&dnp->dn_spill,
 763                     &czb, dnp, buf, ds, scn, ostype, tx);
 764         }
 765 }
 766
 767 /*
 768  * The arguments are in this order because mdb can only print the
 769  * first 5; we want them to be useful.
 770  */
 771 static void
 772 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 773     dnode_phys_t *dnp, arc_buf_t *pbuf,
 774     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 775     dmu_tx_t *tx)
 776 {
 777         dsl_pool_t *dp = scn->scn_dp;
 778         arc_buf_t *buf = NULL;
 779         blkptr_t bp_toread = *bp;
 780
 781         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 782
 783         if (dsl_scan_check_pause(scn, zb))
 784                 return;
 785
 786         if (dsl_scan_check_resume(scn, dnp, zb))
 787                 return;
 788
 789         if (bp->blk_birth == 0)
 790                 return;
 791
 792         scn->scn_visited_this_txg++;
 793
 794         dprintf_bp(bp,
 795             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 796             ds, ds ? ds->ds_object : 0,
 797             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 798             pbuf, bp);
 799
 800         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 801                 return;
 802
 803         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 804             &buf) != 0)
 805                 return;
 806
 807         /*
 808          * If dsl_scan_ddt() has aready visited this block, it will have
 809          * already done any translations or scrubbing, so don't call the
 810          * callback again.
 811          */
 812         if (ddt_class_contains(dp->dp_spa,
 813             scn->scn_phys.scn_ddt_class_max, bp)) {
 814                 ASSERT(buf == NULL);
 815                 return;
 816         }
 817
 818         /*
 819          * If this block is from the future (after cur_max_txg), then we
 820          * are doing this on behalf of a deleted snapshot, and we will
 821          * revisit the future block on the next pass of this dataset.
 822          * Don't scan it now unless we need to because something
 823          * under it was modified.
 824          */
 825         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 826                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 827         }
 828         if (buf)
 829                 (void) arc_buf_remove_ref(buf, &buf);
 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_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, 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 (dsl_dataset_is_snapshot(ds)) {
 858                         /* Note, scn_cur_{min,max}_txg stays the same. */
 859                         scn->scn_phys.scn_bookmark.zb_objset =
 860                             ds->ds_phys->ds_next_snap_obj;
 861                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 862                             "reset zb_objset to %llu",
 863                             (u_longlong_t)ds->ds_object,
 864                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 865                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 866                 } else {
 867                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 868                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 869                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 870                             "reset bookmark to -1,0,0,0",
 871                             (u_longlong_t)ds->ds_object);
 872                 }
 873         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 874             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 875                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 876                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 877                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 878                 if (dsl_dataset_is_snapshot(ds)) {
 879                         /*
 880                          * We keep the same mintxg; it could be >
 881                          * ds_creation_txg if the previous snapshot was
 882                          * deleted too.
 883                          */
 884                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 885                             scn->scn_phys.scn_queue_obj,
 886                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 887                         zfs_dbgmsg("destroying ds %llu; in queue; "
 888                             "replacing with %llu",
 889                             (u_longlong_t)ds->ds_object,
 890                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 891                 } else {
 892                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 893                             (u_longlong_t)ds->ds_object);
 894                 }
 895         } else {
 896                 zfs_dbgmsg("destroying ds %llu; ignoring",
 897                     (u_longlong_t)ds->ds_object);
 898         }
 899
 900         /*
 901          * dsl_scan_sync() should be called after this, and should sync
 902          * out our changed state, but just to be safe, do it here.
 903          */
 904         dsl_scan_sync_state(scn, tx);
 905 }
 906
 907 void
 908 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 909 {
 910         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 911         dsl_scan_t *scn = dp->dp_scan;
 912         uint64_t mintxg;
 913
 914         if (scn->scn_phys.scn_state != DSS_SCANNING)
 915                 return;
 916
 917         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 918
 919         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 920                 scn->scn_phys.scn_bookmark.zb_objset =
 921                     ds->ds_phys->ds_prev_snap_obj;
 922                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 923                     "reset zb_objset to %llu",
 924                     (u_longlong_t)ds->ds_object,
 925                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 926         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 927             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 928                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 929                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 930                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 931                     scn->scn_phys.scn_queue_obj,
 932                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 933                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 934                     "replacing with %llu",
 935                     (u_longlong_t)ds->ds_object,
 936                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 937         }
 938         dsl_scan_sync_state(scn, tx);
 939 }
 940
 941 void
 942 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 943 {
 944         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 945         dsl_scan_t *scn = dp->dp_scan;
 946         uint64_t mintxg;
 947
 948         if (scn->scn_phys.scn_state != DSS_SCANNING)
 949                 return;
 950
 951         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 952                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 953                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 954                     "reset zb_objset to %llu",
 955                     (u_longlong_t)ds1->ds_object,
 956                     (u_longlong_t)ds2->ds_object);
 957         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 958                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 959                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 960                     "reset zb_objset to %llu",
 961                     (u_longlong_t)ds2->ds_object,
 962                     (u_longlong_t)ds1->ds_object);
 963         }
 964
 965         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 966             ds1->ds_object, &mintxg) == 0) {
 967                 int err;
 968
 969                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 970                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 971                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 972                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 973                 err = zap_add_int_key(dp->dp_meta_objset,
 974                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 975                 VERIFY(err == 0 || err == EEXIST);
 976                 if (err == EEXIST) {
 977                         /* Both were there to begin with */
 978                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 979                             scn->scn_phys.scn_queue_obj,
 980                             ds1->ds_object, mintxg, tx));
 981                 }
 982                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 983                     "replacing with %llu",
 984                     (u_longlong_t)ds1->ds_object,
 985                     (u_longlong_t)ds2->ds_object);
 986         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 987             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 988                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 989                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 990                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 991                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 992                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 993                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 994                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 995                     "replacing with %llu",
 996                     (u_longlong_t)ds2->ds_object,
 997                     (u_longlong_t)ds1->ds_object);
 998         }
 999
1000         dsl_scan_sync_state(scn, tx);
1001 }
1002
1003 struct enqueue_clones_arg {
1004         dmu_tx_t *tx;
1005         uint64_t originobj;
1006 };
1007
1008 /* ARGSUSED */
1009 static int
1010 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1011 {
1012         struct enqueue_clones_arg *eca = arg;
1013         dsl_dataset_t *ds;
1014         int err;
1015         dsl_pool_t *dp = spa->spa_dsl_pool;
1016         dsl_scan_t *scn = dp->dp_scan;
1017
1018         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1019         if (err)
1020                 return (err);
1021
1022         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
1023                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1024                         dsl_dataset_t *prev;
1025                         err = dsl_dataset_hold_obj(dp,
1026                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1027
1028                         dsl_dataset_rele(ds, FTAG);
1029                         if (err)
1030                                 return (err);
1031                         ds = prev;
1032                 }
1033                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1034                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1035                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1036         }
1037         dsl_dataset_rele(ds, FTAG);
1038         return (0);
1039 }
1040
1041 static void
1042 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1043 {
1044         dsl_pool_t *dp = scn->scn_dp;
1045         dsl_dataset_t *ds;
1046         objset_t *os;
1047
1048         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1049
1050         if (dmu_objset_from_ds(ds, &os))
1051                 goto out;
1052
1053         /*
1054          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1055          * snapshots can have ZIL block pointers (which may be the same
1056          * BP as in the head), they must be ignored.  So we traverse the
1057          * ZIL here, rather than in scan_recurse(), because the regular
1058          * snapshot block-sharing rules don't apply to it.
1059          */
1060         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1061                 dsl_scan_zil(dp, &os->os_zil_header);
1062
1063         /*
1064          * Iterate over the bps in this ds.
1065          */
1066         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1067         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1068
1069         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1070         dsl_dataset_name(ds, dsname);
1071         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1072             "pausing=%u",
1073             (longlong_t)dsobj, dsname,
1074             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1075             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1076             (int)scn->scn_pausing);
1077         kmem_free(dsname, ZFS_MAXNAMELEN);
1078
1079         if (scn->scn_pausing)
1080                 goto out;
1081
1082         /*
1083          * We've finished this pass over this dataset.
1084          */
1085
1086         /*
1087          * If we did not completely visit this dataset, do another pass.
1088          */
1089         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1090                 zfs_dbgmsg("incomplete pass; visiting again");
1091                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1092                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1093                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1094                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1095                 goto out;
1096         }
1097
1098         /*
1099          * Add descendent datasets to work queue.
1100          */
1101         if (ds->ds_phys->ds_next_snap_obj != 0) {
1102                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1103                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1104                     ds->ds_phys->ds_creation_txg, tx) == 0);
1105         }
1106         if (ds->ds_phys->ds_num_children > 1) {
1107                 boolean_t usenext = B_FALSE;
1108                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1109                         uint64_t count;
1110                         /*
1111                          * A bug in a previous version of the code could
1112                          * cause upgrade_clones_cb() to not set
1113                          * ds_next_snap_obj when it should, leading to a
1114                          * missing entry.  Therefore we can only use the
1115                          * next_clones_obj when its count is correct.
1116                          */
1117                         int err = zap_count(dp->dp_meta_objset,
1118                             ds->ds_phys->ds_next_clones_obj, &count);
1119                         if (err == 0 &&
1120                             count == ds->ds_phys->ds_num_children - 1)
1121                                 usenext = B_TRUE;
1122                 }
1123
1124                 if (usenext) {
1125                         VERIFY(zap_join_key(dp->dp_meta_objset,
1126                             ds->ds_phys->ds_next_clones_obj,
1127                             scn->scn_phys.scn_queue_obj,
1128                             ds->ds_phys->ds_creation_txg, tx) == 0);
1129                 } else {
1130                         struct enqueue_clones_arg eca;
1131                         eca.tx = tx;
1132                         eca.originobj = ds->ds_object;
1133
1134                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1135                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1136                 }
1137         }
1138
1139 out:
1140         dsl_dataset_rele(ds, FTAG);
1141 }
1142
1143 /* ARGSUSED */
1144 static int
1145 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1146 {
1147         dmu_tx_t *tx = arg;
1148         dsl_dataset_t *ds;
1149         int err;
1150         dsl_pool_t *dp = spa->spa_dsl_pool;
1151         dsl_scan_t *scn = dp->dp_scan;
1152
1153         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1154         if (err)
1155                 return (err);
1156
1157         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1158                 dsl_dataset_t *prev;
1159                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1160                     FTAG, &prev);
1161                 if (err) {
1162                         dsl_dataset_rele(ds, FTAG);
1163                         return (err);
1164                 }
1165
1166                 /*
1167                  * If this is a clone, we don't need to worry about it for now.
1168                  */
1169                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1170                         dsl_dataset_rele(ds, FTAG);
1171                         dsl_dataset_rele(prev, FTAG);
1172                         return (0);
1173                 }
1174                 dsl_dataset_rele(ds, FTAG);
1175                 ds = prev;
1176         }
1177
1178         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1179             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1180         dsl_dataset_rele(ds, FTAG);
1181         return (0);
1182 }
1183
1184 /*
1185  * Scrub/dedup interaction.
1186  *
1187  * If there are N references to a deduped block, we don't want to scrub it
1188  * N times -- ideally, we should scrub it exactly once.
1189  *
1190  * We leverage the fact that the dde's replication class (enum ddt_class)
1191  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1192  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1193  *
1194  * To prevent excess scrubbing, the scrub begins by walking the DDT
1195  * to find all blocks with refcnt > 1, and scrubs each of these once.
1196  * Since there are two replication classes which contain blocks with
1197  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1198  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1199  *
1200  * There would be nothing more to say if a block's refcnt couldn't change
1201  * during a scrub, but of course it can so we must account for changes
1202  * in a block's replication class.
1203  *
1204  * Here's an example of what can occur:
1205  *
1206  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1207  * when visited during the top-down scrub phase, it will be scrubbed twice.
1208  * This negates our scrub optimization, but is otherwise harmless.
1209  *
1210  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1211  * on each visit during the top-down scrub phase, it will never be scrubbed.
1212  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1213  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1214  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1215  * while a scrub is in progress, it scrubs the block right then.
1216  */
1217 static void
1218 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1219 {
1220         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1221         ddt_entry_t dde = { 0 };
1222         int error;
1223         uint64_t n = 0;
1224
1225         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1226                 ddt_t *ddt;
1227
1228                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1229                         break;
1230                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1231                     (longlong_t)ddb->ddb_class,
1232                     (longlong_t)ddb->ddb_type,
1233                     (longlong_t)ddb->ddb_checksum,
1234                     (longlong_t)ddb->ddb_cursor);
1235
1236                 /* There should be no pending changes to the dedup table */
1237                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1238                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1239
1240                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1241                 n++;
1242
1243                 if (dsl_scan_check_pause(scn, NULL))
1244                         break;
1245         }
1246
1247         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1248             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1249             (int)scn->scn_pausing);
1250
1251         ASSERT(error == 0 || error == ENOENT);
1252         ASSERT(error != ENOENT ||
1253             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1254 }
1255
1256 /* ARGSUSED */
1257 void
1258 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1259     ddt_entry_t *dde, dmu_tx_t *tx)
1260 {
1261         const ddt_key_t *ddk = &dde->dde_key;
1262         ddt_phys_t *ddp = dde->dde_phys;
1263         blkptr_t bp;
1264         zbookmark_t zb = { 0 };
1265
1266         if (scn->scn_phys.scn_state != DSS_SCANNING)
1267                 return;
1268
1269         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1270                 if (ddp->ddp_phys_birth == 0 ||
1271                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1272                         continue;
1273                 ddt_bp_create(checksum, ddk, ddp, &bp);
1274
1275                 scn->scn_visited_this_txg++;
1276                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1277         }
1278 }
1279
1280 static void
1281 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1282 {
1283         dsl_pool_t *dp = scn->scn_dp;
1284         zap_cursor_t zc;
1285         zap_attribute_t za;
1286
1287         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1288             scn->scn_phys.scn_ddt_class_max) {
1289                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1290                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1291                 dsl_scan_ddt(scn, tx);
1292                 if (scn->scn_pausing)
1293                         return;
1294         }
1295
1296         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1297                 /* First do the MOS & ORIGIN */
1298
1299                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1300                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1301                 dsl_scan_visit_rootbp(scn, NULL,
1302                     &dp->dp_meta_rootbp, tx);
1303                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1304                 if (scn->scn_pausing)
1305                         return;
1306
1307                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1308                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1309                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1310                 } else {
1311                         dsl_scan_visitds(scn,
1312                             dp->dp_origin_snap->ds_object, tx);
1313                 }
1314                 ASSERT(!scn->scn_pausing);
1315         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1316             ZB_DESTROYED_OBJSET) {
1317                 /*
1318                  * If we were paused, continue from here.  Note if the
1319                  * ds we were paused on was deleted, the zb_objset may
1320                  * be -1, so we will skip this and find a new objset
1321                  * below.
1322                  */
1323                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1324                 if (scn->scn_pausing)
1325                         return;
1326         }
1327
1328         /*
1329          * In case we were paused right at the end of the ds, zero the
1330          * bookmark so we don't think that we're still trying to resume.
1331          */
1332         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1333
1334         /* keep pulling things out of the zap-object-as-queue */
1335         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1336             scn->scn_phys.scn_queue_obj),
1337             zap_cursor_retrieve(&zc, &za) == 0) {
1338                 dsl_dataset_t *ds;
1339                 uint64_t dsobj;
1340
1341                 dsobj = strtonum(za.za_name, NULL);
1342                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1343                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1344
1345                 /* Set up min/max txg */
1346                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1347                 if (za.za_first_integer != 0) {
1348                         scn->scn_phys.scn_cur_min_txg =
1349                             MAX(scn->scn_phys.scn_min_txg,
1350                             za.za_first_integer);
1351                 } else {
1352                         scn->scn_phys.scn_cur_min_txg =
1353                             MAX(scn->scn_phys.scn_min_txg,
1354                             ds->ds_phys->ds_prev_snap_txg);
1355                 }
1356                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1357                 dsl_dataset_rele(ds, FTAG);
1358
1359                 dsl_scan_visitds(scn, dsobj, tx);
1360                 zap_cursor_fini(&zc);
1361                 if (scn->scn_pausing)
1362                         return;
1363         }
1364         zap_cursor_fini(&zc);
1365 }
1366
1367 static boolean_t
1368 dsl_scan_free_should_pause(dsl_scan_t *scn)
1369 {
1370         uint64_t elapsed_nanosecs;
1371
1372         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1373         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1374             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1375             txg_sync_waiting(scn->scn_dp)) ||
1376             spa_shutting_down(scn->scn_dp->dp_spa));
1377 }
1378
1379 static int
1380 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1381 {
1382         dsl_scan_t *scn = arg;
1383
1384         if (!scn->scn_is_bptree ||
1385             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1386                 if (dsl_scan_free_should_pause(scn))
1387                         return (ERESTART);
1388         }
1389
1390         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1391             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1392         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1393             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1394             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1395         scn->scn_visited_this_txg++;
1396         return (0);
1397 }
1398
1399 boolean_t
1400 dsl_scan_active(dsl_scan_t *scn)
1401 {
1402         spa_t *spa = scn->scn_dp->dp_spa;
1403         uint64_t used = 0, comp, uncomp;
1404
1405         if (spa->spa_load_state != SPA_LOAD_NONE)
1406                 return (B_FALSE);
1407         if (spa_shutting_down(spa))
1408                 return (B_FALSE);
1409
1410         if (scn->scn_phys.scn_state == DSS_SCANNING)
1411                 return (B_TRUE);
1412
1413         if (spa_feature_is_active(spa,
1414             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1415                 return (B_TRUE);
1416         }
1417         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1418                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1419                     &used, &comp, &uncomp);
1420         }
1421         return (used != 0);
1422 }
1423
1424 void
1425 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1426 {
1427         dsl_scan_t *scn = dp->dp_scan;
1428         spa_t *spa = dp->dp_spa;
1429         int err;
1430
1431         /*
1432          * Check for scn_restart_txg before checking spa_load_state, so
1433          * that we can restart an old-style scan while the pool is being
1434          * imported (see dsl_scan_init).
1435          */
1436         if (scn->scn_restart_txg != 0 &&
1437             scn->scn_restart_txg <= tx->tx_txg) {
1438                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1439                 dsl_scan_done(scn, B_FALSE, tx);
1440                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1441                         func = POOL_SCAN_RESILVER;
1442                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1443                     func, tx->tx_txg);
1444                 dsl_scan_setup_sync(scn, &func, tx);
1445         }
1446
1447         if (!dsl_scan_active(scn) ||
1448             spa_sync_pass(dp->dp_spa) > 1)
1449                 return;
1450
1451         scn->scn_visited_this_txg = 0;
1452         scn->scn_pausing = B_FALSE;
1453         scn->scn_sync_start_time = gethrtime();
1454         spa->spa_scrub_active = B_TRUE;
1455
1456         /*
1457          * First process the free list.  If we pause the free, don't do
1458          * any scanning.  This ensures that there is no free list when
1459          * we are scanning, so the scan code doesn't have to worry about
1460          * traversing it.
1461          */
1462         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1463                 scn->scn_is_bptree = B_FALSE;
1464                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1465                     NULL, ZIO_FLAG_MUSTSUCCEED);
1466                 err = bpobj_iterate(&dp->dp_free_bpobj,
1467                     dsl_scan_free_block_cb, scn, tx);
1468                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1469
1470                 if (err == 0 && spa_feature_is_active(spa,
1471                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1472                         scn->scn_is_bptree = B_TRUE;
1473                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1474                             NULL, ZIO_FLAG_MUSTSUCCEED);
1475                         err = bptree_iterate(dp->dp_meta_objset,
1476                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1477                             scn, tx);
1478                         VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1479                         if (err != 0)
1480                                 return;
1481
1482                         /* disable async destroy feature */
1483                         spa_feature_decr(spa,
1484                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY], tx);
1485                         ASSERT(!spa_feature_is_active(spa,
1486                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]));
1487                         VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
1488                             DMU_POOL_DIRECTORY_OBJECT,
1489                             DMU_POOL_BPTREE_OBJ, tx));
1490                         VERIFY3U(0, ==, bptree_free(dp->dp_meta_objset,
1491                             dp->dp_bptree_obj, tx));
1492                         dp->dp_bptree_obj = 0;
1493                 }
1494                 if (scn->scn_visited_this_txg) {
1495                         zfs_dbgmsg("freed %llu blocks in %llums from "
1496                             "free_bpobj/bptree txg %llu",
1497                             (longlong_t)scn->scn_visited_this_txg,
1498                             (longlong_t)
1499                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1500                             (longlong_t)tx->tx_txg);
1501                         scn->scn_visited_this_txg = 0;
1502                         /*
1503                          * Re-sync the ddt so that we can further modify
1504                          * it when doing bprewrite.
1505                          */
1506                         ddt_sync(spa, tx->tx_txg);
1507                 }
1508                 if (err == ERESTART)
1509                         return;
1510         }
1511
1512         if (scn->scn_phys.scn_state != DSS_SCANNING)
1513                 return;
1514
1515         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1516             scn->scn_phys.scn_ddt_class_max) {
1517                 zfs_dbgmsg("doing scan sync txg %llu; "
1518                     "ddt bm=%llu/%llu/%llu/%llx",
1519                     (longlong_t)tx->tx_txg,
1520                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1521                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1522                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1523                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1524                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1525                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1526                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1527                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1528         } else {
1529                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1530                     (longlong_t)tx->tx_txg,
1531                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1532                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1533                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1534                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1535         }
1536
1537         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1538             NULL, ZIO_FLAG_CANFAIL);
1539         dsl_scan_visit(scn, tx);
1540         (void) zio_wait(scn->scn_zio_root);
1541         scn->scn_zio_root = NULL;
1542
1543         zfs_dbgmsg("visited %llu blocks in %llums",
1544             (longlong_t)scn->scn_visited_this_txg,
1545             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1546
1547         if (!scn->scn_pausing) {
1548                 /* finished with scan. */
1549                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1550                 dsl_scan_done(scn, B_TRUE, tx);
1551         }
1552
1553         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1554                 mutex_enter(&spa->spa_scrub_lock);
1555                 while (spa->spa_scrub_inflight > 0) {
1556                         cv_wait(&spa->spa_scrub_io_cv,
1557                             &spa->spa_scrub_lock);
1558                 }
1559                 mutex_exit(&spa->spa_scrub_lock);
1560         }
1561
1562         dsl_scan_sync_state(scn, tx);
1563 }
1564
1565 /*
1566  * This will start a new scan, or restart an existing one.
1567  */
1568 void
1569 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1570 {
1571         if (txg == 0) {
1572                 dmu_tx_t *tx;
1573                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1574                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1575
1576                 txg = dmu_tx_get_txg(tx);
1577                 dp->dp_scan->scn_restart_txg = txg;
1578                 dmu_tx_commit(tx);
1579         } else {
1580                 dp->dp_scan->scn_restart_txg = txg;
1581         }
1582         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1583 }
1584
1585 boolean_t
1586 dsl_scan_resilvering(dsl_pool_t *dp)
1587 {
1588         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1589             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1590 }
1591
1592 /*
1593  * scrub consumers
1594  */
1595
1596 static void
1597 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1598 {
1599         int i;
1600
1601         /*
1602          * If we resume after a reboot, zab will be NULL; don't record
1603          * incomplete stats in that case.
1604          */
1605         if (zab == NULL)
1606                 return;
1607
1608         for (i = 0; i < 4; i++) {
1609                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1610                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1611                 if (t & DMU_OT_NEWTYPE)
1612                         t = DMU_OT_OTHER;
1613                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1614                 int equal;
1615
1616                 zb->zb_count++;
1617                 zb->zb_asize += BP_GET_ASIZE(bp);
1618                 zb->zb_lsize += BP_GET_LSIZE(bp);
1619                 zb->zb_psize += BP_GET_PSIZE(bp);
1620                 zb->zb_gangs += BP_COUNT_GANG(bp);
1621
1622                 switch (BP_GET_NDVAS(bp)) {
1623                 case 2:
1624                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1625                             DVA_GET_VDEV(&bp->blk_dva[1]))
1626                                 zb->zb_ditto_2_of_2_samevdev++;
1627                         break;
1628                 case 3:
1629                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1630                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1631                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1632                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1633                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1634                             DVA_GET_VDEV(&bp->blk_dva[2]));
1635                         if (equal == 1)
1636                                 zb->zb_ditto_2_of_3_samevdev++;
1637                         else if (equal == 3)
1638                                 zb->zb_ditto_3_of_3_samevdev++;
1639                         break;
1640                 }
1641         }
1642 }
1643
1644 static void
1645 dsl_scan_scrub_done(zio_t *zio)
1646 {
1647         spa_t *spa = zio->io_spa;
1648
1649         zio_data_buf_free(zio->io_data, zio->io_size);
1650
1651         mutex_enter(&spa->spa_scrub_lock);
1652         spa->spa_scrub_inflight--;
1653         cv_broadcast(&spa->spa_scrub_io_cv);
1654
1655         if (zio->io_error && (zio->io_error != ECKSUM ||
1656             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1657                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1658         }
1659         mutex_exit(&spa->spa_scrub_lock);
1660 }
1661
1662 static int
1663 dsl_scan_scrub_cb(dsl_pool_t *dp,
1664     const blkptr_t *bp, const zbookmark_t *zb)
1665 {
1666         dsl_scan_t *scn = dp->dp_scan;
1667         size_t size = BP_GET_PSIZE(bp);
1668         spa_t *spa = dp->dp_spa;
1669         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1670         boolean_t needs_io;
1671         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1672         int zio_priority;
1673         unsigned int scan_delay = 0;
1674
1675         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1676             phys_birth >= scn->scn_phys.scn_max_txg)
1677                 return (0);
1678
1679         count_block(dp->dp_blkstats, bp);
1680
1681         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1682         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1683                 zio_flags |= ZIO_FLAG_SCRUB;
1684                 zio_priority = ZIO_PRIORITY_SCRUB;
1685                 needs_io = B_TRUE;
1686                 scan_delay = zfs_scrub_delay;
1687         } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1688                 zio_flags |= ZIO_FLAG_RESILVER;
1689                 zio_priority = ZIO_PRIORITY_RESILVER;
1690                 needs_io = B_FALSE;
1691                 scan_delay = zfs_resilver_delay;
1692         }
1693
1694         /* If it's an intent log block, failure is expected. */
1695         if (zb->zb_level == ZB_ZIL_LEVEL)
1696                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1697
1698         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1699                 vdev_t *vd = vdev_lookup_top(spa,
1700                     DVA_GET_VDEV(&bp->blk_dva[d]));
1701
1702                 /*
1703                  * Keep track of how much data we've examined so that
1704                  * zpool(1M) status can make useful progress reports.
1705                  */
1706                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1707                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1708
1709                 /* if it's a resilver, this may not be in the target range */
1710                 if (!needs_io) {
1711                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1712                                 /*
1713                                  * Gang members may be spread across multiple
1714                                  * vdevs, so the best estimate we have is the
1715                                  * scrub range, which has already been checked.
1716                                  * XXX -- it would be better to change our
1717                                  * allocation policy to ensure that all
1718                                  * gang members reside on the same vdev.
1719                                  */
1720                                 needs_io = B_TRUE;
1721                         } else {
1722                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1723                                     phys_birth, 1);
1724                         }
1725                 }
1726         }
1727
1728         if (needs_io && !zfs_no_scrub_io) {
1729                 vdev_t *rvd = spa->spa_root_vdev;
1730                 uint64_t maxinflight = rvd->vdev_children *
1731                     MAX(zfs_top_maxinflight, 1);
1732                 void *data = zio_data_buf_alloc(size);
1733
1734                 mutex_enter(&spa->spa_scrub_lock);
1735                 while (spa->spa_scrub_inflight >= maxinflight)
1736                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1737                 spa->spa_scrub_inflight++;
1738                 mutex_exit(&spa->spa_scrub_lock);
1739
1740                 /*
1741                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1742                  * then throttle our workload to limit the impact of a scan.
1743                  */
1744                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1745                         delay(MAX((int)scan_delay, 0));
1746
1747                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1748                     dsl_scan_scrub_done, NULL, zio_priority,
1749                     zio_flags, zb));
1750         }
1751
1752         /* do not relocate this block */
1753         return (0);
1754 }
1755
1756 int
1757 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1758 {
1759         spa_t *spa = dp->dp_spa;
1760
1761         /*
1762          * Purge all vdev caches and probe all devices.  We do this here
1763          * rather than in sync context because this requires a writer lock
1764          * on the spa_config lock, which we can't do from sync context.  The
1765          * spa_scrub_reopen flag indicates that vdev_open() should not
1766          * attempt to start another scrub.
1767          */
1768         spa_vdev_state_enter(spa, SCL_NONE);
1769         spa->spa_scrub_reopen = B_TRUE;
1770         vdev_reopen(spa->spa_root_vdev);
1771         spa->spa_scrub_reopen = B_FALSE;
1772         (void) spa_vdev_state_exit(spa, NULL, 0);
1773
1774         return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1775             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1776 }