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 static uint64_t
 401 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 402 {
 403         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 404         if (dsl_dataset_is_snapshot(ds))
 405                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 406         return (smt);
 407 }
 408
 409 static void
 410 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 411 {
 412         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 413             DMU_POOL_DIRECTORY_OBJECT,
 414             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 415             &scn->scn_phys, tx));
 416 }
 417
 418 static boolean_t
 419 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 420 {
 421         uint64_t elapsed_nanosecs;
 422         unsigned int mintime;
 423
 424         /* we never skip user/group accounting objects */
 425         if (zb && (int64_t)zb->zb_object < 0)
 426                 return (B_FALSE);
 427
 428         if (scn->scn_pausing)
 429                 return (B_TRUE); /* we're already pausing */
 430
 431         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 432                 return (B_FALSE); /* we're resuming */
 433
 434         /* We only know how to resume from level-0 blocks. */
 435         if (zb && zb->zb_level != 0)
 436                 return (B_FALSE);
 437
 438         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 439             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 440         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 441         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 442             (elapsed_nanosecs / MICROSEC > mintime &&
 443             txg_sync_waiting(scn->scn_dp)) ||
 444             spa_shutting_down(scn->scn_dp->dp_spa)) {
 445                 if (zb) {
 446                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 447                             (longlong_t)zb->zb_objset,
 448                             (longlong_t)zb->zb_object,
 449                             (longlong_t)zb->zb_level,
 450                             (longlong_t)zb->zb_blkid);
 451                         scn->scn_phys.scn_bookmark = *zb;
 452                 }
 453                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 454                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 455                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 456                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 457                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 458                 scn->scn_pausing = B_TRUE;
 459                 return (B_TRUE);
 460         }
 461         return (B_FALSE);
 462 }
 463
 464 typedef struct zil_scan_arg {
 465         dsl_pool_t      *zsa_dp;
 466         zil_header_t    *zsa_zh;
 467 } zil_scan_arg_t;
 468
 469 /* ARGSUSED */
 470 static int
 471 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 472 {
 473         zil_scan_arg_t *zsa = arg;
 474         dsl_pool_t *dp = zsa->zsa_dp;
 475         dsl_scan_t *scn = dp->dp_scan;
 476         zil_header_t *zh = zsa->zsa_zh;
 477         zbookmark_t zb;
 478
 479         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 480                 return (0);
 481
 482         /*
 483          * One block ("stubby") can be allocated a long time ago; we
 484          * want to visit that one because it has been allocated
 485          * (on-disk) even if it hasn't been claimed (even though for
 486          * scrub there's nothing to do to it).
 487          */
 488         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 489                 return (0);
 490
 491         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 492             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 493
 494         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 495         return (0);
 496 }
 497
 498 /* ARGSUSED */
 499 static int
 500 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 501 {
 502         if (lrc->lrc_txtype == TX_WRITE) {
 503                 zil_scan_arg_t *zsa = arg;
 504                 dsl_pool_t *dp = zsa->zsa_dp;
 505                 dsl_scan_t *scn = dp->dp_scan;
 506                 zil_header_t *zh = zsa->zsa_zh;
 507                 lr_write_t *lr = (lr_write_t *)lrc;
 508                 blkptr_t *bp = &lr->lr_blkptr;
 509                 zbookmark_t zb;
 510
 511                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 512                         return (0);
 513
 514                 /*
 515                  * birth can be < claim_txg if this record's txg is
 516                  * already txg sync'ed (but this log block contains
 517                  * other records that are not synced)
 518                  */
 519                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 520                         return (0);
 521
 522                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 523                     lr->lr_foid, ZB_ZIL_LEVEL,
 524                     lr->lr_offset / BP_GET_LSIZE(bp));
 525
 526                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 527         }
 528         return (0);
 529 }
 530
 531 static void
 532 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 533 {
 534         uint64_t claim_txg = zh->zh_claim_txg;
 535         zil_scan_arg_t zsa = { dp, zh };
 536         zilog_t *zilog;
 537
 538         /*
 539          * We only want to visit blocks that have been claimed but not yet
 540          * replayed (or, in read-only mode, blocks that *would* be claimed).
 541          */
 542         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 543                 return;
 544
 545         zilog = zil_alloc(dp->dp_meta_objset, zh);
 546
 547         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 548             claim_txg);
 549
 550         zil_free(zilog);
 551 }
 552
 553 /* ARGSUSED */
 554 static void
 555 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 556     uint64_t objset, uint64_t object, uint64_t blkid)
 557 {
 558         zbookmark_t czb;
 559         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 560
 561         if (zfs_no_scrub_prefetch)
 562                 return;
 563
 564         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 565             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 566                 return;
 567
 568         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 569
 570         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 571             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 572             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 573 }
 574
 575 static boolean_t
 576 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 577     const zbookmark_t *zb)
 578 {
 579         /*
 580          * We never skip over user/group accounting objects (obj<0)
 581          */
 582         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 583             (int64_t)zb->zb_object >= 0) {
 584                 /*
 585                  * If we already visited this bp & everything below (in
 586                  * a prior txg sync), don't bother doing it again.
 587                  */
 588                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 589                         return (B_TRUE);
 590
 591                 /*
 592                  * If we found the block we're trying to resume from, or
 593                  * we went past it to a different object, zero it out to
 594                  * indicate that it's OK to start checking for pausing
 595                  * again.
 596                  */
 597                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 598                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 599                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 600                             (longlong_t)zb->zb_objset,
 601                             (longlong_t)zb->zb_object,
 602                             (longlong_t)zb->zb_level,
 603                             (longlong_t)zb->zb_blkid);
 604                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 605                 }
 606         }
 607         return (B_FALSE);
 608 }
 609
 610 /*
 611  * Return nonzero on i/o error.
 612  * Return new buf to write out in *bufp.
 613  */
 614 static int
 615 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 616     dnode_phys_t *dnp, const blkptr_t *bp,
 617     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 618 {
 619         dsl_pool_t *dp = scn->scn_dp;
 620         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 621         int err;
 622
 623         if (BP_GET_LEVEL(bp) > 0) {
 624                 uint32_t flags = ARC_WAIT;
 625                 int i;
 626                 blkptr_t *cbp;
 627                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 628
 629                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 630                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 631                 if (err) {
 632                         scn->scn_phys.scn_errors++;
 633                         return (err);
 634                 }
 635                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 636                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 637                             zb->zb_object, zb->zb_blkid * epb + i);
 638                 }
 639                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 640                         zbookmark_t czb;
 641
 642                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 643                             zb->zb_level - 1,
 644                             zb->zb_blkid * epb + i);
 645                         dsl_scan_visitbp(cbp, &czb, dnp,
 646                             *bufp, ds, scn, ostype, tx);
 647                 }
 648         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 649                 uint32_t flags = ARC_WAIT;
 650
 651                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 652                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 653                 if (err) {
 654                         scn->scn_phys.scn_errors++;
 655                         return (err);
 656                 }
 657         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 658                 uint32_t flags = ARC_WAIT;
 659                 dnode_phys_t *cdnp;
 660                 int i, j;
 661                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 662
 663                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 664                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 665                 if (err) {
 666                         scn->scn_phys.scn_errors++;
 667                         return (err);
 668                 }
 669                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 670                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 671                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 672                                 dsl_scan_prefetch(scn, *bufp, cbp,
 673                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 674                         }
 675                 }
 676                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 677                         dsl_scan_visitdnode(scn, ds, ostype,
 678                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 679                 }
 680
 681         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 682                 uint32_t flags = ARC_WAIT;
 683                 objset_phys_t *osp;
 684
 685                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 686                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 687                 if (err) {
 688                         scn->scn_phys.scn_errors++;
 689                         return (err);
 690                 }
 691
 692                 osp = (*bufp)->b_data;
 693
 694                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 695                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 696
 697                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 698                         /*
 699                          * We also always visit user/group accounting
 700                          * objects, and never skip them, even if we are
 701                          * pausing.  This is necessary so that the space
 702                          * deltas from this txg get integrated.
 703                          */
 704                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 705                             &osp->os_groupused_dnode, *bufp,
 706                             DMU_GROUPUSED_OBJECT, tx);
 707                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 708                             &osp->os_userused_dnode, *bufp,
 709                             DMU_USERUSED_OBJECT, tx);
 710                 }
 711         }
 712
 713         return (0);
 714 }
 715
 716 static void
 717 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 718     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 719     uint64_t object, dmu_tx_t *tx)
 720 {
 721         int j;
 722
 723         for (j = 0; j < dnp->dn_nblkptr; j++) {
 724                 zbookmark_t czb;
 725
 726                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 727                     dnp->dn_nlevels - 1, j);
 728                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 729                     &czb, dnp, buf, ds, scn, ostype, tx);
 730         }
 731
 732         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 733                 zbookmark_t czb;
 734                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 735                     0, DMU_SPILL_BLKID);
 736                 dsl_scan_visitbp(&dnp->dn_spill,
 737                     &czb, dnp, buf, ds, scn, ostype, tx);
 738         }
 739 }
 740
 741 /*
 742  * The arguments are in this order because mdb can only print the
 743  * first 5; we want them to be useful.
 744  */
 745 static void
 746 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 747     dnode_phys_t *dnp, arc_buf_t *pbuf,
 748     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 749     dmu_tx_t *tx)
 750 {
 751         dsl_pool_t *dp = scn->scn_dp;
 752         arc_buf_t *buf = NULL;
 753         blkptr_t bp_toread = *bp;
 754
 755         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 756
 757         if (dsl_scan_check_pause(scn, zb))
 758                 return;
 759
 760         if (dsl_scan_check_resume(scn, dnp, zb))
 761                 return;
 762
 763         if (bp->blk_birth == 0)
 764                 return;
 765
 766         scn->scn_visited_this_txg++;
 767
 768         dprintf_bp(bp,
 769             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 770             ds, ds ? ds->ds_object : 0,
 771             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 772             pbuf, bp);
 773
 774         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 775                 return;
 776
 777         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 778             &buf) != 0)
 779                 return;
 780
 781         /*
 782          * If dsl_scan_ddt() has aready visited this block, it will have
 783          * already done any translations or scrubbing, so don't call the
 784          * callback again.
 785          */
 786         if (ddt_class_contains(dp->dp_spa,
 787             scn->scn_phys.scn_ddt_class_max, bp)) {
 788                 ASSERT(buf == NULL);
 789                 return;
 790         }
 791
 792         /*
 793          * If this block is from the future (after cur_max_txg), then we
 794          * are doing this on behalf of a deleted snapshot, and we will
 795          * revisit the future block on the next pass of this dataset.
 796          * Don't scan it now unless we need to because something
 797          * under it was modified.
 798          */
 799         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 800                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 801         }
 802         if (buf)
 803                 (void) arc_buf_remove_ref(buf, &buf);
 804 }
 805
 806 static void
 807 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 808     dmu_tx_t *tx)
 809 {
 810         zbookmark_t zb;
 811
 812         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 813             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 814         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 815             ds, scn, DMU_OST_NONE, tx);
 816
 817         dprintf_ds(ds, "finished scan%s", "");
 818 }
 819
 820 void
 821 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 822 {
 823         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 824         dsl_scan_t *scn = dp->dp_scan;
 825         uint64_t mintxg;
 826
 827         if (scn->scn_phys.scn_state != DSS_SCANNING)
 828                 return;
 829
 830         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 831                 if (dsl_dataset_is_snapshot(ds)) {
 832                         /* Note, scn_cur_{min,max}_txg stays the same. */
 833                         scn->scn_phys.scn_bookmark.zb_objset =
 834                             ds->ds_phys->ds_next_snap_obj;
 835                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 836                             "reset zb_objset to %llu",
 837                             (u_longlong_t)ds->ds_object,
 838                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 839                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 840                 } else {
 841                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 842                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 843                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 844                             "reset bookmark to -1,0,0,0",
 845                             (u_longlong_t)ds->ds_object);
 846                 }
 847         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 848             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 849                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 850                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 851                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 852                 if (dsl_dataset_is_snapshot(ds)) {
 853                         /*
 854                          * We keep the same mintxg; it could be >
 855                          * ds_creation_txg if the previous snapshot was
 856                          * deleted too.
 857                          */
 858                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 859                             scn->scn_phys.scn_queue_obj,
 860                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 861                         zfs_dbgmsg("destroying ds %llu; in queue; "
 862                             "replacing with %llu",
 863                             (u_longlong_t)ds->ds_object,
 864                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 865                 } else {
 866                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 867                             (u_longlong_t)ds->ds_object);
 868                 }
 869         } else {
 870                 zfs_dbgmsg("destroying ds %llu; ignoring",
 871                     (u_longlong_t)ds->ds_object);
 872         }
 873
 874         /*
 875          * dsl_scan_sync() should be called after this, and should sync
 876          * out our changed state, but just to be safe, do it here.
 877          */
 878         dsl_scan_sync_state(scn, tx);
 879 }
 880
 881 void
 882 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 883 {
 884         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 885         dsl_scan_t *scn = dp->dp_scan;
 886         uint64_t mintxg;
 887
 888         if (scn->scn_phys.scn_state != DSS_SCANNING)
 889                 return;
 890
 891         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 892
 893         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 894                 scn->scn_phys.scn_bookmark.zb_objset =
 895                     ds->ds_phys->ds_prev_snap_obj;
 896                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 897                     "reset zb_objset to %llu",
 898                     (u_longlong_t)ds->ds_object,
 899                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 900         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 901             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 902                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 903                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 904                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 905                     scn->scn_phys.scn_queue_obj,
 906                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 907                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 908                     "replacing with %llu",
 909                     (u_longlong_t)ds->ds_object,
 910                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 911         }
 912         dsl_scan_sync_state(scn, tx);
 913 }
 914
 915 void
 916 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 917 {
 918         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 919         dsl_scan_t *scn = dp->dp_scan;
 920         uint64_t mintxg;
 921
 922         if (scn->scn_phys.scn_state != DSS_SCANNING)
 923                 return;
 924
 925         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 926                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 927                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 928                     "reset zb_objset to %llu",
 929                     (u_longlong_t)ds1->ds_object,
 930                     (u_longlong_t)ds2->ds_object);
 931         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 932                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 933                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 934                     "reset zb_objset to %llu",
 935                     (u_longlong_t)ds2->ds_object,
 936                     (u_longlong_t)ds1->ds_object);
 937         }
 938
 939         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 940             ds1->ds_object, &mintxg) == 0) {
 941                 int err;
 942
 943                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 944                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 945                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 946                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 947                 err = zap_add_int_key(dp->dp_meta_objset,
 948                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 949                 VERIFY(err == 0 || err == EEXIST);
 950                 if (err == EEXIST) {
 951                         /* Both were there to begin with */
 952                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 953                             scn->scn_phys.scn_queue_obj,
 954                             ds1->ds_object, mintxg, tx));
 955                 }
 956                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 957                     "replacing with %llu",
 958                     (u_longlong_t)ds1->ds_object,
 959                     (u_longlong_t)ds2->ds_object);
 960         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 961             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 962                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 963                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 964                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 965                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 966                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 967                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 968                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 969                     "replacing with %llu",
 970                     (u_longlong_t)ds2->ds_object,
 971                     (u_longlong_t)ds1->ds_object);
 972         }
 973
 974         dsl_scan_sync_state(scn, tx);
 975 }
 976
 977 struct enqueue_clones_arg {
 978         dmu_tx_t *tx;
 979         uint64_t originobj;
 980 };
 981
 982 /* ARGSUSED */
 983 static int
 984 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 985 {
 986         struct enqueue_clones_arg *eca = arg;
 987         dsl_dataset_t *ds;
 988         int err;
 989         dsl_pool_t *dp = spa->spa_dsl_pool;
 990         dsl_scan_t *scn = dp->dp_scan;
 991
 992         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
 993         if (err)
 994                 return (err);
 995
 996         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 997                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 998                         dsl_dataset_t *prev;
 999                         err = dsl_dataset_hold_obj(dp,
1000                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1001
1002                         dsl_dataset_rele(ds, FTAG);
1003                         if (err)
1004                                 return (err);
1005                         ds = prev;
1006                 }
1007                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1008                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1009                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1010         }
1011         dsl_dataset_rele(ds, FTAG);
1012         return (0);
1013 }
1014
1015 static void
1016 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1017 {
1018         dsl_pool_t *dp = scn->scn_dp;
1019         dsl_dataset_t *ds;
1020         objset_t *os;
1021
1022         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1023
1024         if (dmu_objset_from_ds(ds, &os))
1025                 goto out;
1026
1027         /*
1028          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1029          * snapshots can have ZIL block pointers (which may be the same
1030          * BP as in the head), they must be ignored.  So we traverse the
1031          * ZIL here, rather than in scan_recurse(), because the regular
1032          * snapshot block-sharing rules don't apply to it.
1033          */
1034         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1035                 dsl_scan_zil(dp, &os->os_zil_header);
1036
1037         /*
1038          * Iterate over the bps in this ds.
1039          */
1040         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1041         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1042
1043         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1044         dsl_dataset_name(ds, dsname);
1045         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1046             "pausing=%u",
1047             (longlong_t)dsobj, dsname,
1048             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1049             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1050             (int)scn->scn_pausing);
1051         kmem_free(dsname, ZFS_MAXNAMELEN);
1052
1053         if (scn->scn_pausing)
1054                 goto out;
1055
1056         /*
1057          * We've finished this pass over this dataset.
1058          */
1059
1060         /*
1061          * If we did not completely visit this dataset, do another pass.
1062          */
1063         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1064                 zfs_dbgmsg("incomplete pass; visiting again");
1065                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1066                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1067                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1068                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1069                 goto out;
1070         }
1071
1072         /*
1073          * Add descendent datasets to work queue.
1074          */
1075         if (ds->ds_phys->ds_next_snap_obj != 0) {
1076                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1077                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1078                     ds->ds_phys->ds_creation_txg, tx) == 0);
1079         }
1080         if (ds->ds_phys->ds_num_children > 1) {
1081                 boolean_t usenext = B_FALSE;
1082                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1083                         uint64_t count;
1084                         /*
1085                          * A bug in a previous version of the code could
1086                          * cause upgrade_clones_cb() to not set
1087                          * ds_next_snap_obj when it should, leading to a
1088                          * missing entry.  Therefore we can only use the
1089                          * next_clones_obj when its count is correct.
1090                          */
1091                         int err = zap_count(dp->dp_meta_objset,
1092                             ds->ds_phys->ds_next_clones_obj, &count);
1093                         if (err == 0 &&
1094                             count == ds->ds_phys->ds_num_children - 1)
1095                                 usenext = B_TRUE;
1096                 }
1097
1098                 if (usenext) {
1099                         VERIFY(zap_join_key(dp->dp_meta_objset,
1100                             ds->ds_phys->ds_next_clones_obj,
1101                             scn->scn_phys.scn_queue_obj,
1102                             ds->ds_phys->ds_creation_txg, tx) == 0);
1103                 } else {
1104                         struct enqueue_clones_arg eca;
1105                         eca.tx = tx;
1106                         eca.originobj = ds->ds_object;
1107
1108                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1109                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1110                 }
1111         }
1112
1113 out:
1114         dsl_dataset_rele(ds, FTAG);
1115 }
1116
1117 /* ARGSUSED */
1118 static int
1119 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1120 {
1121         dmu_tx_t *tx = arg;
1122         dsl_dataset_t *ds;
1123         int err;
1124         dsl_pool_t *dp = spa->spa_dsl_pool;
1125         dsl_scan_t *scn = dp->dp_scan;
1126
1127         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1128         if (err)
1129                 return (err);
1130
1131         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1132                 dsl_dataset_t *prev;
1133                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1134                     FTAG, &prev);
1135                 if (err) {
1136                         dsl_dataset_rele(ds, FTAG);
1137                         return (err);
1138                 }
1139
1140                 /*
1141                  * If this is a clone, we don't need to worry about it for now.
1142                  */
1143                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1144                         dsl_dataset_rele(ds, FTAG);
1145                         dsl_dataset_rele(prev, FTAG);
1146                         return (0);
1147                 }
1148                 dsl_dataset_rele(ds, FTAG);
1149                 ds = prev;
1150         }
1151
1152         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1153             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1154         dsl_dataset_rele(ds, FTAG);
1155         return (0);
1156 }
1157
1158 /*
1159  * Scrub/dedup interaction.
1160  *
1161  * If there are N references to a deduped block, we don't want to scrub it
1162  * N times -- ideally, we should scrub it exactly once.
1163  *
1164  * We leverage the fact that the dde's replication class (enum ddt_class)
1165  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1166  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1167  *
1168  * To prevent excess scrubbing, the scrub begins by walking the DDT
1169  * to find all blocks with refcnt > 1, and scrubs each of these once.
1170  * Since there are two replication classes which contain blocks with
1171  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1172  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1173  *
1174  * There would be nothing more to say if a block's refcnt couldn't change
1175  * during a scrub, but of course it can so we must account for changes
1176  * in a block's replication class.
1177  *
1178  * Here's an example of what can occur:
1179  *
1180  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1181  * when visited during the top-down scrub phase, it will be scrubbed twice.
1182  * This negates our scrub optimization, but is otherwise harmless.
1183  *
1184  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1185  * on each visit during the top-down scrub phase, it will never be scrubbed.
1186  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1187  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1188  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1189  * while a scrub is in progress, it scrubs the block right then.
1190  */
1191 static void
1192 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1193 {
1194         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1195         ddt_entry_t dde = { 0 };
1196         int error;
1197         uint64_t n = 0;
1198
1199         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1200                 ddt_t *ddt;
1201
1202                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1203                         break;
1204                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1205                     (longlong_t)ddb->ddb_class,
1206                     (longlong_t)ddb->ddb_type,
1207                     (longlong_t)ddb->ddb_checksum,
1208                     (longlong_t)ddb->ddb_cursor);
1209
1210                 /* There should be no pending changes to the dedup table */
1211                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1212                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1213
1214                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1215                 n++;
1216
1217                 if (dsl_scan_check_pause(scn, NULL))
1218                         break;
1219         }
1220
1221         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1222             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1223             (int)scn->scn_pausing);
1224
1225         ASSERT(error == 0 || error == ENOENT);
1226         ASSERT(error != ENOENT ||
1227             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1228 }
1229
1230 /* ARGSUSED */
1231 void
1232 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1233     ddt_entry_t *dde, dmu_tx_t *tx)
1234 {
1235         const ddt_key_t *ddk = &dde->dde_key;
1236         ddt_phys_t *ddp = dde->dde_phys;
1237         blkptr_t bp;
1238         zbookmark_t zb = { 0 };
1239
1240         if (scn->scn_phys.scn_state != DSS_SCANNING)
1241                 return;
1242
1243         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1244                 if (ddp->ddp_phys_birth == 0 ||
1245                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1246                         continue;
1247                 ddt_bp_create(checksum, ddk, ddp, &bp);
1248
1249                 scn->scn_visited_this_txg++;
1250                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1251         }
1252 }
1253
1254 static void
1255 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1256 {
1257         dsl_pool_t *dp = scn->scn_dp;
1258         zap_cursor_t zc;
1259         zap_attribute_t za;
1260
1261         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1262             scn->scn_phys.scn_ddt_class_max) {
1263                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1264                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1265                 dsl_scan_ddt(scn, tx);
1266                 if (scn->scn_pausing)
1267                         return;
1268         }
1269
1270         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1271                 /* First do the MOS & ORIGIN */
1272
1273                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1274                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1275                 dsl_scan_visit_rootbp(scn, NULL,
1276                     &dp->dp_meta_rootbp, tx);
1277                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1278                 if (scn->scn_pausing)
1279                         return;
1280
1281                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1282                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1283                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1284                 } else {
1285                         dsl_scan_visitds(scn,
1286                             dp->dp_origin_snap->ds_object, tx);
1287                 }
1288                 ASSERT(!scn->scn_pausing);
1289         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1290             ZB_DESTROYED_OBJSET) {
1291                 /*
1292                  * If we were paused, continue from here.  Note if the
1293                  * ds we were paused on was deleted, the zb_objset may
1294                  * be -1, so we will skip this and find a new objset
1295                  * below.
1296                  */
1297                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1298                 if (scn->scn_pausing)
1299                         return;
1300         }
1301
1302         /*
1303          * In case we were paused right at the end of the ds, zero the
1304          * bookmark so we don't think that we're still trying to resume.
1305          */
1306         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1307
1308         /* keep pulling things out of the zap-object-as-queue */
1309         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1310             scn->scn_phys.scn_queue_obj),
1311             zap_cursor_retrieve(&zc, &za) == 0) {
1312                 dsl_dataset_t *ds;
1313                 uint64_t dsobj;
1314
1315                 dsobj = strtonum(za.za_name, NULL);
1316                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1317                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1318
1319                 /* Set up min/max txg */
1320                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1321                 if (za.za_first_integer != 0) {
1322                         scn->scn_phys.scn_cur_min_txg =
1323                             MAX(scn->scn_phys.scn_min_txg,
1324                             za.za_first_integer);
1325                 } else {
1326                         scn->scn_phys.scn_cur_min_txg =
1327                             MAX(scn->scn_phys.scn_min_txg,
1328                             ds->ds_phys->ds_prev_snap_txg);
1329                 }
1330                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1331                 dsl_dataset_rele(ds, FTAG);
1332
1333                 dsl_scan_visitds(scn, dsobj, tx);
1334                 zap_cursor_fini(&zc);
1335                 if (scn->scn_pausing)
1336                         return;
1337         }
1338         zap_cursor_fini(&zc);
1339 }
1340
1341 static boolean_t
1342 dsl_scan_free_should_pause(dsl_scan_t *scn)
1343 {
1344         uint64_t elapsed_nanosecs;
1345
1346         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1347         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1348             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1349             txg_sync_waiting(scn->scn_dp)) ||
1350             spa_shutting_down(scn->scn_dp->dp_spa));
1351 }
1352
1353 static int
1354 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1355 {
1356         dsl_scan_t *scn = arg;
1357
1358         if (!scn->scn_is_bptree ||
1359             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1360                 if (dsl_scan_free_should_pause(scn))
1361                         return (ERESTART);
1362         }
1363
1364         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1365             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1366         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1367             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1368             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1369         scn->scn_visited_this_txg++;
1370         return (0);
1371 }
1372
1373 boolean_t
1374 dsl_scan_active(dsl_scan_t *scn)
1375 {
1376         spa_t *spa = scn->scn_dp->dp_spa;
1377         uint64_t used = 0, comp, uncomp;
1378
1379         if (spa->spa_load_state != SPA_LOAD_NONE)
1380                 return (B_FALSE);
1381         if (spa_shutting_down(spa))
1382                 return (B_FALSE);
1383
1384         if (scn->scn_phys.scn_state == DSS_SCANNING)
1385                 return (B_TRUE);
1386
1387         if (spa_feature_is_active(spa,
1388             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1389                 return (B_TRUE);
1390         }
1391         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1392                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1393                     &used, &comp, &uncomp);
1394         }
1395         return (used != 0);
1396 }
1397
1398 void
1399 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1400 {
1401         dsl_scan_t *scn = dp->dp_scan;
1402         spa_t *spa = dp->dp_spa;
1403         int err;
1404
1405         /*
1406          * Check for scn_restart_txg before checking spa_load_state, so
1407          * that we can restart an old-style scan while the pool is being
1408          * imported (see dsl_scan_init).
1409          */
1410         if (scn->scn_restart_txg != 0 &&
1411             scn->scn_restart_txg <= tx->tx_txg) {
1412                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1413                 dsl_scan_done(scn, B_FALSE, tx);
1414                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1415                         func = POOL_SCAN_RESILVER;
1416                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1417                     func, tx->tx_txg);
1418                 dsl_scan_setup_sync(scn, &func, tx);
1419         }
1420
1421         if (!dsl_scan_active(scn) ||
1422             spa_sync_pass(dp->dp_spa) > 1)
1423                 return;
1424
1425         scn->scn_visited_this_txg = 0;
1426         scn->scn_pausing = B_FALSE;
1427         scn->scn_sync_start_time = gethrtime();
1428         spa->spa_scrub_active = B_TRUE;
1429
1430         /*
1431          * First process the free list.  If we pause the free, don't do
1432          * any scanning.  This ensures that there is no free list when
1433          * we are scanning, so the scan code doesn't have to worry about
1434          * traversing it.
1435          */
1436         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1437                 scn->scn_is_bptree = B_FALSE;
1438                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1439                     NULL, ZIO_FLAG_MUSTSUCCEED);
1440                 err = bpobj_iterate(&dp->dp_free_bpobj,
1441                     dsl_scan_free_block_cb, scn, tx);
1442                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1443
1444                 if (err == 0 && spa_feature_is_active(spa,
1445                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1446                         scn->scn_is_bptree = B_TRUE;
1447                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1448                             NULL, ZIO_FLAG_MUSTSUCCEED);
1449                         err = bptree_iterate(dp->dp_meta_objset,
1450                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1451                             scn, tx);
1452                         VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1453                         if (err != 0)
1454                                 return;
1455
1456                         /* disable async destroy feature */
1457                         spa_feature_decr(spa,
1458                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY], tx);
1459                         ASSERT(!spa_feature_is_active(spa,
1460                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]));
1461                         VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
1462                             DMU_POOL_DIRECTORY_OBJECT,
1463                             DMU_POOL_BPTREE_OBJ, tx));
1464                         VERIFY3U(0, ==, bptree_free(dp->dp_meta_objset,
1465                             dp->dp_bptree_obj, tx));
1466                         dp->dp_bptree_obj = 0;
1467                 }
1468                 if (scn->scn_visited_this_txg) {
1469                         zfs_dbgmsg("freed %llu blocks in %llums from "
1470                             "free_bpobj/bptree txg %llu",
1471                             (longlong_t)scn->scn_visited_this_txg,
1472                             (longlong_t)
1473                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1474                             (longlong_t)tx->tx_txg);
1475                         scn->scn_visited_this_txg = 0;
1476                         /*
1477                          * Re-sync the ddt so that we can further modify
1478                          * it when doing bprewrite.
1479                          */
1480                         ddt_sync(spa, tx->tx_txg);
1481                 }
1482                 if (err == ERESTART)
1483                         return;
1484         }
1485
1486         if (scn->scn_phys.scn_state != DSS_SCANNING)
1487                 return;
1488
1489         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1490             scn->scn_phys.scn_ddt_class_max) {
1491                 zfs_dbgmsg("doing scan sync txg %llu; "
1492                     "ddt bm=%llu/%llu/%llu/%llx",
1493                     (longlong_t)tx->tx_txg,
1494                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1495                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1496                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1497                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1498                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1499                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1500                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1501                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1502         } else {
1503                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1504                     (longlong_t)tx->tx_txg,
1505                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1506                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1507                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1508                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1509         }
1510
1511         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1512             NULL, ZIO_FLAG_CANFAIL);
1513         dsl_scan_visit(scn, tx);
1514         (void) zio_wait(scn->scn_zio_root);
1515         scn->scn_zio_root = NULL;
1516
1517         zfs_dbgmsg("visited %llu blocks in %llums",
1518             (longlong_t)scn->scn_visited_this_txg,
1519             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1520
1521         if (!scn->scn_pausing) {
1522                 /* finished with scan. */
1523                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1524                 dsl_scan_done(scn, B_TRUE, tx);
1525         }
1526
1527         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1528                 mutex_enter(&spa->spa_scrub_lock);
1529                 while (spa->spa_scrub_inflight > 0) {
1530                         cv_wait(&spa->spa_scrub_io_cv,
1531                             &spa->spa_scrub_lock);
1532                 }
1533                 mutex_exit(&spa->spa_scrub_lock);
1534         }
1535
1536         dsl_scan_sync_state(scn, tx);
1537 }
1538
1539 /*
1540  * This will start a new scan, or restart an existing one.
1541  */
1542 void
1543 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1544 {
1545         if (txg == 0) {
1546                 dmu_tx_t *tx;
1547                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1548                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1549
1550                 txg = dmu_tx_get_txg(tx);
1551                 dp->dp_scan->scn_restart_txg = txg;
1552                 dmu_tx_commit(tx);
1553         } else {
1554                 dp->dp_scan->scn_restart_txg = txg;
1555         }
1556         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1557 }
1558
1559 boolean_t
1560 dsl_scan_resilvering(dsl_pool_t *dp)
1561 {
1562         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1563             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1564 }
1565
1566 /*
1567  * scrub consumers
1568  */
1569
1570 static void
1571 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1572 {
1573         int i;
1574
1575         /*
1576          * If we resume after a reboot, zab will be NULL; don't record
1577          * incomplete stats in that case.
1578          */
1579         if (zab == NULL)
1580                 return;
1581
1582         for (i = 0; i < 4; i++) {
1583                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1584                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1585                 if (t & DMU_OT_NEWTYPE)
1586                         t = DMU_OT_OTHER;
1587                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1588                 int equal;
1589
1590                 zb->zb_count++;
1591                 zb->zb_asize += BP_GET_ASIZE(bp);
1592                 zb->zb_lsize += BP_GET_LSIZE(bp);
1593                 zb->zb_psize += BP_GET_PSIZE(bp);
1594                 zb->zb_gangs += BP_COUNT_GANG(bp);
1595
1596                 switch (BP_GET_NDVAS(bp)) {
1597                 case 2:
1598                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1599                             DVA_GET_VDEV(&bp->blk_dva[1]))
1600                                 zb->zb_ditto_2_of_2_samevdev++;
1601                         break;
1602                 case 3:
1603                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1604                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1605                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1606                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1607                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1608                             DVA_GET_VDEV(&bp->blk_dva[2]));
1609                         if (equal == 1)
1610                                 zb->zb_ditto_2_of_3_samevdev++;
1611                         else if (equal == 3)
1612                                 zb->zb_ditto_3_of_3_samevdev++;
1613                         break;
1614                 }
1615         }
1616 }
1617
1618 static void
1619 dsl_scan_scrub_done(zio_t *zio)
1620 {
1621         spa_t *spa = zio->io_spa;
1622
1623         zio_data_buf_free(zio->io_data, zio->io_size);
1624
1625         mutex_enter(&spa->spa_scrub_lock);
1626         spa->spa_scrub_inflight--;
1627         cv_broadcast(&spa->spa_scrub_io_cv);
1628
1629         if (zio->io_error && (zio->io_error != ECKSUM ||
1630             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1631                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1632         }
1633         mutex_exit(&spa->spa_scrub_lock);
1634 }
1635
1636 static int
1637 dsl_scan_scrub_cb(dsl_pool_t *dp,
1638     const blkptr_t *bp, const zbookmark_t *zb)
1639 {
1640         dsl_scan_t *scn = dp->dp_scan;
1641         size_t size = BP_GET_PSIZE(bp);
1642         spa_t *spa = dp->dp_spa;
1643         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1644         boolean_t needs_io;
1645         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1646         int zio_priority;
1647         unsigned int scan_delay = 0;
1648
1649         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1650             phys_birth >= scn->scn_phys.scn_max_txg)
1651                 return (0);
1652
1653         count_block(dp->dp_blkstats, bp);
1654
1655         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1656         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1657                 zio_flags |= ZIO_FLAG_SCRUB;
1658                 zio_priority = ZIO_PRIORITY_SCRUB;
1659                 needs_io = B_TRUE;
1660                 scan_delay = zfs_scrub_delay;
1661         } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1662                 zio_flags |= ZIO_FLAG_RESILVER;
1663                 zio_priority = ZIO_PRIORITY_RESILVER;
1664                 needs_io = B_FALSE;
1665                 scan_delay = zfs_resilver_delay;
1666         }
1667
1668         /* If it's an intent log block, failure is expected. */
1669         if (zb->zb_level == ZB_ZIL_LEVEL)
1670                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1671
1672         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1673                 vdev_t *vd = vdev_lookup_top(spa,
1674                     DVA_GET_VDEV(&bp->blk_dva[d]));
1675
1676                 /*
1677                  * Keep track of how much data we've examined so that
1678                  * zpool(1M) status can make useful progress reports.
1679                  */
1680                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1681                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1682
1683                 /* if it's a resilver, this may not be in the target range */
1684                 if (!needs_io) {
1685                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1686                                 /*
1687                                  * Gang members may be spread across multiple
1688                                  * vdevs, so the best estimate we have is the
1689                                  * scrub range, which has already been checked.
1690                                  * XXX -- it would be better to change our
1691                                  * allocation policy to ensure that all
1692                                  * gang members reside on the same vdev.
1693                                  */
1694                                 needs_io = B_TRUE;
1695                         } else {
1696                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1697                                     phys_birth, 1);
1698                         }
1699                 }
1700         }
1701
1702         if (needs_io && !zfs_no_scrub_io) {
1703                 vdev_t *rvd = spa->spa_root_vdev;
1704                 uint64_t maxinflight = rvd->vdev_children *
1705                     MAX(zfs_top_maxinflight, 1);
1706                 void *data = zio_data_buf_alloc(size);
1707
1708                 mutex_enter(&spa->spa_scrub_lock);
1709                 while (spa->spa_scrub_inflight >= maxinflight)
1710                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1711                 spa->spa_scrub_inflight++;
1712                 mutex_exit(&spa->spa_scrub_lock);
1713
1714                 /*
1715                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1716                  * then throttle our workload to limit the impact of a scan.
1717                  */
1718                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1719                         delay(MAX((int)scan_delay, 0));
1720
1721                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1722                     dsl_scan_scrub_done, NULL, zio_priority,
1723                     zio_flags, zb));
1724         }
1725
1726         /* do not relocate this block */
1727         return (0);
1728 }
1729
1730 int
1731 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1732 {
1733         spa_t *spa = dp->dp_spa;
1734
1735         /*
1736          * Purge all vdev caches and probe all devices.  We do this here
1737          * rather than in sync context because this requires a writer lock
1738          * on the spa_config lock, which we can't do from sync context.  The
1739          * spa_scrub_reopen flag indicates that vdev_open() should not
1740          * attempt to start another scrub.
1741          */
1742         spa_vdev_state_enter(spa, SCL_NONE);
1743         spa->spa_scrub_reopen = B_TRUE;
1744         vdev_reopen(spa->spa_root_vdev);
1745         spa->spa_scrub_reopen = B_FALSE;
1746         (void) spa_vdev_state_exit(spa, NULL, 0);
1747
1748         return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1749             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1750 }