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) 2013 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 void dsl_scan_cancel_sync(void *, dmu_tx_t *);
  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         /*
 129          * It's possible that we're resuming a scan after a reboot so
 130          * make sure that the scan_async_destroying flag is initialized
 131          * appropriately.
 132          */
 133         ASSERT(!scn->scn_async_destroying);
 134         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 135             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]);
 136
 137         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 138             "scrub_func", sizeof (uint64_t), 1, &f);
 139         if (err == 0) {
 140                 /*
 141                  * There was an old-style scrub in progress.  Restart a
 142                  * new-style scrub from the beginning.
 143                  */
 144                 scn->scn_restart_txg = txg;
 145                 zfs_dbgmsg("old-style scrub was in progress; "
 146                     "restarting new-style scrub in txg %llu",
 147                     scn->scn_restart_txg);
 148
 149                 /*
 150                  * Load the queue obj from the old location so that it
 151                  * can be freed by dsl_scan_done().
 152                  */
 153                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 154                     "scrub_queue", sizeof (uint64_t), 1,
 155                     &scn->scn_phys.scn_queue_obj);
 156         } else {
 157                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 158                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 159                     &scn->scn_phys);
 160                 if (err == ENOENT)
 161                         return (0);
 162                 else if (err)
 163                         return (err);
 164
 165                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 166                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 167                         /*
 168                          * A new-type scrub was in progress on an old
 169                          * pool, and the pool was accessed by old
 170                          * software.  Restart from the beginning, since
 171                          * the old software may have changed the pool in
 172                          * the meantime.
 173                          */
 174                         scn->scn_restart_txg = txg;
 175                         zfs_dbgmsg("new-style scrub was modified "
 176                             "by old software; restarting in txg %llu",
 177                             scn->scn_restart_txg);
 178                 }
 179         }
 180
 181         spa_scan_stat_init(spa);
 182         return (0);
 183 }
 184
 185 void
 186 dsl_scan_fini(dsl_pool_t *dp)
 187 {
 188         if (dp->dp_scan) {
 189                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 190                 dp->dp_scan = NULL;
 191         }
 192 }
 193
 194 /* ARGSUSED */
 195 static int
 196 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 197 {
 198         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 199
 200         if (scn->scn_phys.scn_state == DSS_SCANNING)
 201                 return (SET_ERROR(EBUSY));
 202
 203         return (0);
 204 }
 205
 206 static void
 207 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 208 {
 209         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 210         pool_scan_func_t *funcp = arg;
 211         dmu_object_type_t ot = 0;
 212         dsl_pool_t *dp = scn->scn_dp;
 213         spa_t *spa = dp->dp_spa;
 214
 215         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 216         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 217         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 218         scn->scn_phys.scn_func = *funcp;
 219         scn->scn_phys.scn_state = DSS_SCANNING;
 220         scn->scn_phys.scn_min_txg = 0;
 221         scn->scn_phys.scn_max_txg = tx->tx_txg;
 222         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 223         scn->scn_phys.scn_start_time = gethrestime_sec();
 224         scn->scn_phys.scn_errors = 0;
 225         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 226         scn->scn_restart_txg = 0;
 227         spa_scan_stat_init(spa);
 228
 229         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 230                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 231
 232                 /* rewrite all disk labels */
 233                 vdev_config_dirty(spa->spa_root_vdev);
 234
 235                 if (vdev_resilver_needed(spa->spa_root_vdev,
 236                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 237                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 238                 } else {
 239                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 240                 }
 241
 242                 spa->spa_scrub_started = B_TRUE;
 243                 /*
 244                  * If this is an incremental scrub, limit the DDT scrub phase
 245                  * to just the auto-ditto class (for correctness); the rest
 246                  * of the scrub should go faster using top-down pruning.
 247                  */
 248                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 249                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 250
 251         }
 252
 253         /* back to the generic stuff */
 254
 255         if (dp->dp_blkstats == NULL) {
 256                 dp->dp_blkstats =
 257                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 258         }
 259         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 260
 261         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 262                 ot = DMU_OT_ZAP_OTHER;
 263
 264         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 265             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 266
 267         dsl_scan_sync_state(scn, tx);
 268
 269         spa_history_log_internal(spa, "scan setup", tx,
 270             "func=%u mintxg=%llu maxtxg=%llu",
 271             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 272 }
 273
 274 /* ARGSUSED */
 275 static void
 276 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 277 {
 278         static const char *old_names[] = {
 279                 "scrub_bookmark",
 280                 "scrub_ddt_bookmark",
 281                 "scrub_ddt_class_max",
 282                 "scrub_queue",
 283                 "scrub_min_txg",
 284                 "scrub_max_txg",
 285                 "scrub_func",
 286                 "scrub_errors",
 287                 NULL
 288         };
 289
 290         dsl_pool_t *dp = scn->scn_dp;
 291         spa_t *spa = dp->dp_spa;
 292         int i;
 293
 294         /* Remove any remnants of an old-style scrub. */
 295         for (i = 0; old_names[i]; i++) {
 296                 (void) zap_remove(dp->dp_meta_objset,
 297                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 298         }
 299
 300         if (scn->scn_phys.scn_queue_obj != 0) {
 301                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 302                     scn->scn_phys.scn_queue_obj, tx));
 303                 scn->scn_phys.scn_queue_obj = 0;
 304         }
 305
 306         /*
 307          * If we were "restarted" from a stopped state, don't bother
 308          * with anything else.
 309          */
 310         if (scn->scn_phys.scn_state != DSS_SCANNING)
 311                 return;
 312
 313         if (complete)
 314                 scn->scn_phys.scn_state = DSS_FINISHED;
 315         else
 316                 scn->scn_phys.scn_state = DSS_CANCELED;
 317
 318         spa_history_log_internal(spa, "scan done", tx,
 319             "complete=%u", complete);
 320
 321         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 322                 mutex_enter(&spa->spa_scrub_lock);
 323                 while (spa->spa_scrub_inflight > 0) {
 324                         cv_wait(&spa->spa_scrub_io_cv,
 325                             &spa->spa_scrub_lock);
 326                 }
 327                 mutex_exit(&spa->spa_scrub_lock);
 328                 spa->spa_scrub_started = B_FALSE;
 329                 spa->spa_scrub_active = B_FALSE;
 330
 331                 /*
 332                  * If the scrub/resilver completed, update all DTLs to
 333                  * reflect this.  Whether it succeeded or not, vacate
 334                  * all temporary scrub DTLs.
 335                  */
 336                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 337                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 338                 if (complete) {
 339                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 340                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 341                 }
 342                 spa_errlog_rotate(spa);
 343
 344                 /*
 345                  * We may have finished replacing a device.
 346                  * Let the async thread assess this and handle the detach.
 347                  */
 348                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 349         }
 350
 351         scn->scn_phys.scn_end_time = gethrestime_sec();
 352 }
 353
 354 /* ARGSUSED */
 355 static int
 356 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 357 {
 358         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 359
 360         if (scn->scn_phys.scn_state != DSS_SCANNING)
 361                 return (SET_ERROR(ENOENT));
 362         return (0);
 363 }
 364
 365 /* ARGSUSED */
 366 static void
 367 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 368 {
 369         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 370
 371         dsl_scan_done(scn, B_FALSE, tx);
 372         dsl_scan_sync_state(scn, tx);
 373 }
 374
 375 int
 376 dsl_scan_cancel(dsl_pool_t *dp)
 377 {
 378         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 379             dsl_scan_cancel_sync, NULL, 3));
 380 }
 381
 382 static void dsl_scan_visitbp(blkptr_t *bp,
 383     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 384     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 385     dmu_tx_t *tx);
 386 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 387     dmu_objset_type_t ostype,
 388     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 389
 390 void
 391 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 392 {
 393         zio_free(dp->dp_spa, txg, bp);
 394 }
 395
 396 void
 397 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 398 {
 399         ASSERT(dsl_pool_sync_context(dp));
 400         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 401             pio->io_flags));
 402 }
 403
 404 static uint64_t
 405 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 406 {
 407         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 408         if (dsl_dataset_is_snapshot(ds))
 409                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 410         return (smt);
 411 }
 412
 413 static void
 414 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 415 {
 416         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 417             DMU_POOL_DIRECTORY_OBJECT,
 418             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 419             &scn->scn_phys, tx));
 420 }
 421
 422 static boolean_t
 423 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 424 {
 425         uint64_t elapsed_nanosecs;
 426         unsigned int mintime;
 427
 428         /* we never skip user/group accounting objects */
 429         if (zb && (int64_t)zb->zb_object < 0)
 430                 return (B_FALSE);
 431
 432         if (scn->scn_pausing)
 433                 return (B_TRUE); /* we're already pausing */
 434
 435         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 436                 return (B_FALSE); /* we're resuming */
 437
 438         /* We only know how to resume from level-0 blocks. */
 439         if (zb && zb->zb_level != 0)
 440                 return (B_FALSE);
 441
 442         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 443             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 444         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 445         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 446             (elapsed_nanosecs / MICROSEC > mintime &&
 447             txg_sync_waiting(scn->scn_dp)) ||
 448             spa_shutting_down(scn->scn_dp->dp_spa)) {
 449                 if (zb) {
 450                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 451                             (longlong_t)zb->zb_objset,
 452                             (longlong_t)zb->zb_object,
 453                             (longlong_t)zb->zb_level,
 454                             (longlong_t)zb->zb_blkid);
 455                         scn->scn_phys.scn_bookmark = *zb;
 456                 }
 457                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 458                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 459                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 460                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 461                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 462                 scn->scn_pausing = B_TRUE;
 463                 return (B_TRUE);
 464         }
 465         return (B_FALSE);
 466 }
 467
 468 typedef struct zil_scan_arg {
 469         dsl_pool_t      *zsa_dp;
 470         zil_header_t    *zsa_zh;
 471 } zil_scan_arg_t;
 472
 473 /* ARGSUSED */
 474 static int
 475 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 476 {
 477         zil_scan_arg_t *zsa = arg;
 478         dsl_pool_t *dp = zsa->zsa_dp;
 479         dsl_scan_t *scn = dp->dp_scan;
 480         zil_header_t *zh = zsa->zsa_zh;
 481         zbookmark_t zb;
 482
 483         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 484                 return (0);
 485
 486         /*
 487          * One block ("stubby") can be allocated a long time ago; we
 488          * want to visit that one because it has been allocated
 489          * (on-disk) even if it hasn't been claimed (even though for
 490          * scrub there's nothing to do to it).
 491          */
 492         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 493                 return (0);
 494
 495         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 496             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 497
 498         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 499         return (0);
 500 }
 501
 502 /* ARGSUSED */
 503 static int
 504 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 505 {
 506         if (lrc->lrc_txtype == TX_WRITE) {
 507                 zil_scan_arg_t *zsa = arg;
 508                 dsl_pool_t *dp = zsa->zsa_dp;
 509                 dsl_scan_t *scn = dp->dp_scan;
 510                 zil_header_t *zh = zsa->zsa_zh;
 511                 lr_write_t *lr = (lr_write_t *)lrc;
 512                 blkptr_t *bp = &lr->lr_blkptr;
 513                 zbookmark_t zb;
 514
 515                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 516                         return (0);
 517
 518                 /*
 519                  * birth can be < claim_txg if this record's txg is
 520                  * already txg sync'ed (but this log block contains
 521                  * other records that are not synced)
 522                  */
 523                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 524                         return (0);
 525
 526                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 527                     lr->lr_foid, ZB_ZIL_LEVEL,
 528                     lr->lr_offset / BP_GET_LSIZE(bp));
 529
 530                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 531         }
 532         return (0);
 533 }
 534
 535 static void
 536 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 537 {
 538         uint64_t claim_txg = zh->zh_claim_txg;
 539         zil_scan_arg_t zsa = { dp, zh };
 540         zilog_t *zilog;
 541
 542         /*
 543          * We only want to visit blocks that have been claimed but not yet
 544          * replayed (or, in read-only mode, blocks that *would* be claimed).
 545          */
 546         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 547                 return;
 548
 549         zilog = zil_alloc(dp->dp_meta_objset, zh);
 550
 551         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 552             claim_txg);
 553
 554         zil_free(zilog);
 555 }
 556
 557 /* ARGSUSED */
 558 static void
 559 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 560     uint64_t objset, uint64_t object, uint64_t blkid)
 561 {
 562         zbookmark_t czb;
 563         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 564
 565         if (zfs_no_scrub_prefetch)
 566                 return;
 567
 568         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 569             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 570                 return;
 571
 572         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 573
 574         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 575             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 576             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 577 }
 578
 579 static boolean_t
 580 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 581     const zbookmark_t *zb)
 582 {
 583         /*
 584          * We never skip over user/group accounting objects (obj<0)
 585          */
 586         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 587             (int64_t)zb->zb_object >= 0) {
 588                 /*
 589                  * If we already visited this bp & everything below (in
 590                  * a prior txg sync), don't bother doing it again.
 591                  */
 592                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 593                         return (B_TRUE);
 594
 595                 /*
 596                  * If we found the block we're trying to resume from, or
 597                  * we went past it to a different object, zero it out to
 598                  * indicate that it's OK to start checking for pausing
 599                  * again.
 600                  */
 601                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 602                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 603                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 604                             (longlong_t)zb->zb_objset,
 605                             (longlong_t)zb->zb_object,
 606                             (longlong_t)zb->zb_level,
 607                             (longlong_t)zb->zb_blkid);
 608                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 609                 }
 610         }
 611         return (B_FALSE);
 612 }
 613
 614 /*
 615  * Return nonzero on i/o error.
 616  * Return new buf to write out in *bufp.
 617  */
 618 static int
 619 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 620     dnode_phys_t *dnp, const blkptr_t *bp,
 621     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 622 {
 623         dsl_pool_t *dp = scn->scn_dp;
 624         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 625         int err;
 626
 627         if (BP_GET_LEVEL(bp) > 0) {
 628                 uint32_t flags = ARC_WAIT;
 629                 int i;
 630                 blkptr_t *cbp;
 631                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 632
 633                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 634                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 635                 if (err) {
 636                         scn->scn_phys.scn_errors++;
 637                         return (err);
 638                 }
 639                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 640                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 641                             zb->zb_object, zb->zb_blkid * epb + i);
 642                 }
 643                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 644                         zbookmark_t czb;
 645
 646                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 647                             zb->zb_level - 1,
 648                             zb->zb_blkid * epb + i);
 649                         dsl_scan_visitbp(cbp, &czb, dnp,
 650                             *bufp, ds, scn, ostype, tx);
 651                 }
 652         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 653                 uint32_t flags = ARC_WAIT;
 654
 655                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 656                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 657                 if (err) {
 658                         scn->scn_phys.scn_errors++;
 659                         return (err);
 660                 }
 661         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 662                 uint32_t flags = ARC_WAIT;
 663                 dnode_phys_t *cdnp;
 664                 int i, j;
 665                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 666
 667                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 668                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 669                 if (err) {
 670                         scn->scn_phys.scn_errors++;
 671                         return (err);
 672                 }
 673                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 674                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 675                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 676                                 dsl_scan_prefetch(scn, *bufp, cbp,
 677                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 678                         }
 679                 }
 680                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 681                         dsl_scan_visitdnode(scn, ds, ostype,
 682                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 683                 }
 684
 685         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 686                 uint32_t flags = ARC_WAIT;
 687                 objset_phys_t *osp;
 688
 689                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 690                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 691                 if (err) {
 692                         scn->scn_phys.scn_errors++;
 693                         return (err);
 694                 }
 695
 696                 osp = (*bufp)->b_data;
 697
 698                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 699                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 700
 701                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 702                         /*
 703                          * We also always visit user/group accounting
 704                          * objects, and never skip them, even if we are
 705                          * pausing.  This is necessary so that the space
 706                          * deltas from this txg get integrated.
 707                          */
 708                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 709                             &osp->os_groupused_dnode, *bufp,
 710                             DMU_GROUPUSED_OBJECT, tx);
 711                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 712                             &osp->os_userused_dnode, *bufp,
 713                             DMU_USERUSED_OBJECT, tx);
 714                 }
 715         }
 716
 717         return (0);
 718 }
 719
 720 static void
 721 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 722     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 723     uint64_t object, dmu_tx_t *tx)
 724 {
 725         int j;
 726
 727         for (j = 0; j < dnp->dn_nblkptr; j++) {
 728                 zbookmark_t czb;
 729
 730                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 731                     dnp->dn_nlevels - 1, j);
 732                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 733                     &czb, dnp, buf, ds, scn, ostype, tx);
 734         }
 735
 736         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 737                 zbookmark_t czb;
 738                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 739                     0, DMU_SPILL_BLKID);
 740                 dsl_scan_visitbp(&dnp->dn_spill,
 741                     &czb, dnp, buf, ds, scn, ostype, tx);
 742         }
 743 }
 744
 745 /*
 746  * The arguments are in this order because mdb can only print the
 747  * first 5; we want them to be useful.
 748  */
 749 static void
 750 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 751     dnode_phys_t *dnp, arc_buf_t *pbuf,
 752     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 753     dmu_tx_t *tx)
 754 {
 755         dsl_pool_t *dp = scn->scn_dp;
 756         arc_buf_t *buf = NULL;
 757         blkptr_t bp_toread = *bp;
 758
 759         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 760
 761         if (dsl_scan_check_pause(scn, zb))
 762                 return;
 763
 764         if (dsl_scan_check_resume(scn, dnp, zb))
 765                 return;
 766
 767         if (bp->blk_birth == 0)
 768                 return;
 769
 770         scn->scn_visited_this_txg++;
 771
 772         dprintf_bp(bp,
 773             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 774             ds, ds ? ds->ds_object : 0,
 775             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 776             pbuf, bp);
 777
 778         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 779                 return;
 780
 781         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 782             &buf) != 0)
 783                 return;
 784
 785         /*
 786          * If dsl_scan_ddt() has aready visited this block, it will have
 787          * already done any translations or scrubbing, so don't call the
 788          * callback again.
 789          */
 790         if (ddt_class_contains(dp->dp_spa,
 791             scn->scn_phys.scn_ddt_class_max, bp)) {
 792                 ASSERT(buf == NULL);
 793                 return;
 794         }
 795
 796         /*
 797          * If this block is from the future (after cur_max_txg), then we
 798          * are doing this on behalf of a deleted snapshot, and we will
 799          * revisit the future block on the next pass of this dataset.
 800          * Don't scan it now unless we need to because something
 801          * under it was modified.
 802          */
 803         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 804                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 805         }
 806         if (buf)
 807                 (void) arc_buf_remove_ref(buf, &buf);
 808 }
 809
 810 static void
 811 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 812     dmu_tx_t *tx)
 813 {
 814         zbookmark_t zb;
 815
 816         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 817             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 818         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 819             ds, scn, DMU_OST_NONE, tx);
 820
 821         dprintf_ds(ds, "finished scan%s", "");
 822 }
 823
 824 void
 825 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 826 {
 827         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 828         dsl_scan_t *scn = dp->dp_scan;
 829         uint64_t mintxg;
 830
 831         if (scn->scn_phys.scn_state != DSS_SCANNING)
 832                 return;
 833
 834         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 835                 if (dsl_dataset_is_snapshot(ds)) {
 836                         /* Note, scn_cur_{min,max}_txg stays the same. */
 837                         scn->scn_phys.scn_bookmark.zb_objset =
 838                             ds->ds_phys->ds_next_snap_obj;
 839                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 840                             "reset zb_objset to %llu",
 841                             (u_longlong_t)ds->ds_object,
 842                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 843                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 844                 } else {
 845                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 846                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 847                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 848                             "reset bookmark to -1,0,0,0",
 849                             (u_longlong_t)ds->ds_object);
 850                 }
 851         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 852             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 853                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 854                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 855                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 856                 if (dsl_dataset_is_snapshot(ds)) {
 857                         /*
 858                          * We keep the same mintxg; it could be >
 859                          * ds_creation_txg if the previous snapshot was
 860                          * deleted too.
 861                          */
 862                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 863                             scn->scn_phys.scn_queue_obj,
 864                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 865                         zfs_dbgmsg("destroying ds %llu; in queue; "
 866                             "replacing with %llu",
 867                             (u_longlong_t)ds->ds_object,
 868                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 869                 } else {
 870                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 871                             (u_longlong_t)ds->ds_object);
 872                 }
 873         } else {
 874                 zfs_dbgmsg("destroying ds %llu; ignoring",
 875                     (u_longlong_t)ds->ds_object);
 876         }
 877
 878         /*
 879          * dsl_scan_sync() should be called after this, and should sync
 880          * out our changed state, but just to be safe, do it here.
 881          */
 882         dsl_scan_sync_state(scn, tx);
 883 }
 884
 885 void
 886 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 887 {
 888         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 889         dsl_scan_t *scn = dp->dp_scan;
 890         uint64_t mintxg;
 891
 892         if (scn->scn_phys.scn_state != DSS_SCANNING)
 893                 return;
 894
 895         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 896
 897         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 898                 scn->scn_phys.scn_bookmark.zb_objset =
 899                     ds->ds_phys->ds_prev_snap_obj;
 900                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 901                     "reset zb_objset to %llu",
 902                     (u_longlong_t)ds->ds_object,
 903                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 904         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 905             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 906                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 907                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 908                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 909                     scn->scn_phys.scn_queue_obj,
 910                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 911                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 912                     "replacing with %llu",
 913                     (u_longlong_t)ds->ds_object,
 914                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 915         }
 916         dsl_scan_sync_state(scn, tx);
 917 }
 918
 919 void
 920 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 921 {
 922         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 923         dsl_scan_t *scn = dp->dp_scan;
 924         uint64_t mintxg;
 925
 926         if (scn->scn_phys.scn_state != DSS_SCANNING)
 927                 return;
 928
 929         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 930                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 931                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 932                     "reset zb_objset to %llu",
 933                     (u_longlong_t)ds1->ds_object,
 934                     (u_longlong_t)ds2->ds_object);
 935         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 936                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 937                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 938                     "reset zb_objset to %llu",
 939                     (u_longlong_t)ds2->ds_object,
 940                     (u_longlong_t)ds1->ds_object);
 941         }
 942
 943         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 944             ds1->ds_object, &mintxg) == 0) {
 945                 int err;
 946
 947                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 948                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 949                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 950                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 951                 err = zap_add_int_key(dp->dp_meta_objset,
 952                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 953                 VERIFY(err == 0 || err == EEXIST);
 954                 if (err == EEXIST) {
 955                         /* Both were there to begin with */
 956                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 957                             scn->scn_phys.scn_queue_obj,
 958                             ds1->ds_object, mintxg, tx));
 959                 }
 960                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 961                     "replacing with %llu",
 962                     (u_longlong_t)ds1->ds_object,
 963                     (u_longlong_t)ds2->ds_object);
 964         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 965             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 966                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 967                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 968                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 969                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 970                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 971                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 972                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 973                     "replacing with %llu",
 974                     (u_longlong_t)ds2->ds_object,
 975                     (u_longlong_t)ds1->ds_object);
 976         }
 977
 978         dsl_scan_sync_state(scn, tx);
 979 }
 980
 981 struct enqueue_clones_arg {
 982         dmu_tx_t *tx;
 983         uint64_t originobj;
 984 };
 985
 986 /* ARGSUSED */
 987 static int
 988 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 989 {
 990         struct enqueue_clones_arg *eca = arg;
 991         dsl_dataset_t *ds;
 992         int err;
 993         dsl_scan_t *scn = dp->dp_scan;
 994
 995         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
 996                 return (0);
 997
 998         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
 999         if (err)
1000                 return (err);
1001
1002         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1003                 dsl_dataset_t *prev;
1004                 err = dsl_dataset_hold_obj(dp,
1005                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1006
1007                 dsl_dataset_rele(ds, FTAG);
1008                 if (err)
1009                         return (err);
1010                 ds = prev;
1011         }
1012         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1013             scn->scn_phys.scn_queue_obj, ds->ds_object,
1014             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1015         dsl_dataset_rele(ds, FTAG);
1016         return (0);
1017 }
1018
1019 static void
1020 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1021 {
1022         dsl_pool_t *dp = scn->scn_dp;
1023         dsl_dataset_t *ds;
1024         objset_t *os;
1025
1026         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1027
1028         if (dmu_objset_from_ds(ds, &os))
1029                 goto out;
1030
1031         /*
1032          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1033          * snapshots can have ZIL block pointers (which may be the same
1034          * BP as in the head), they must be ignored.  So we traverse the
1035          * ZIL here, rather than in scan_recurse(), because the regular
1036          * snapshot block-sharing rules don't apply to it.
1037          */
1038         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1039                 dsl_scan_zil(dp, &os->os_zil_header);
1040
1041         /*
1042          * Iterate over the bps in this ds.
1043          */
1044         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1045         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1046
1047         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1048         dsl_dataset_name(ds, dsname);
1049         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1050             "pausing=%u",
1051             (longlong_t)dsobj, dsname,
1052             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1053             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1054             (int)scn->scn_pausing);
1055         kmem_free(dsname, ZFS_MAXNAMELEN);
1056
1057         if (scn->scn_pausing)
1058                 goto out;
1059
1060         /*
1061          * We've finished this pass over this dataset.
1062          */
1063
1064         /*
1065          * If we did not completely visit this dataset, do another pass.
1066          */
1067         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1068                 zfs_dbgmsg("incomplete pass; visiting again");
1069                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1070                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1071                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1072                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1073                 goto out;
1074         }
1075
1076         /*
1077          * Add descendent datasets to work queue.
1078          */
1079         if (ds->ds_phys->ds_next_snap_obj != 0) {
1080                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1081                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1082                     ds->ds_phys->ds_creation_txg, tx) == 0);
1083         }
1084         if (ds->ds_phys->ds_num_children > 1) {
1085                 boolean_t usenext = B_FALSE;
1086                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1087                         uint64_t count;
1088                         /*
1089                          * A bug in a previous version of the code could
1090                          * cause upgrade_clones_cb() to not set
1091                          * ds_next_snap_obj when it should, leading to a
1092                          * missing entry.  Therefore we can only use the
1093                          * next_clones_obj when its count is correct.
1094                          */
1095                         int err = zap_count(dp->dp_meta_objset,
1096                             ds->ds_phys->ds_next_clones_obj, &count);
1097                         if (err == 0 &&
1098                             count == ds->ds_phys->ds_num_children - 1)
1099                                 usenext = B_TRUE;
1100                 }
1101
1102                 if (usenext) {
1103                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1104                             ds->ds_phys->ds_next_clones_obj,
1105                             scn->scn_phys.scn_queue_obj,
1106                             ds->ds_phys->ds_creation_txg, tx));
1107                 } else {
1108                         struct enqueue_clones_arg eca;
1109                         eca.tx = tx;
1110                         eca.originobj = ds->ds_object;
1111
1112                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1113                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1114                 }
1115         }
1116
1117 out:
1118         dsl_dataset_rele(ds, FTAG);
1119 }
1120
1121 /* ARGSUSED */
1122 static int
1123 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1124 {
1125         dmu_tx_t *tx = arg;
1126         dsl_dataset_t *ds;
1127         int err;
1128         dsl_scan_t *scn = dp->dp_scan;
1129
1130         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1131         if (err)
1132                 return (err);
1133
1134         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1135                 dsl_dataset_t *prev;
1136                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1137                     FTAG, &prev);
1138                 if (err) {
1139                         dsl_dataset_rele(ds, FTAG);
1140                         return (err);
1141                 }
1142
1143                 /*
1144                  * If this is a clone, we don't need to worry about it for now.
1145                  */
1146                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1147                         dsl_dataset_rele(ds, FTAG);
1148                         dsl_dataset_rele(prev, FTAG);
1149                         return (0);
1150                 }
1151                 dsl_dataset_rele(ds, FTAG);
1152                 ds = prev;
1153         }
1154
1155         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1156             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1157         dsl_dataset_rele(ds, FTAG);
1158         return (0);
1159 }
1160
1161 /*
1162  * Scrub/dedup interaction.
1163  *
1164  * If there are N references to a deduped block, we don't want to scrub it
1165  * N times -- ideally, we should scrub it exactly once.
1166  *
1167  * We leverage the fact that the dde's replication class (enum ddt_class)
1168  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1169  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1170  *
1171  * To prevent excess scrubbing, the scrub begins by walking the DDT
1172  * to find all blocks with refcnt > 1, and scrubs each of these once.
1173  * Since there are two replication classes which contain blocks with
1174  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1175  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1176  *
1177  * There would be nothing more to say if a block's refcnt couldn't change
1178  * during a scrub, but of course it can so we must account for changes
1179  * in a block's replication class.
1180  *
1181  * Here's an example of what can occur:
1182  *
1183  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1184  * when visited during the top-down scrub phase, it will be scrubbed twice.
1185  * This negates our scrub optimization, but is otherwise harmless.
1186  *
1187  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1188  * on each visit during the top-down scrub phase, it will never be scrubbed.
1189  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1190  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1191  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1192  * while a scrub is in progress, it scrubs the block right then.
1193  */
1194 static void
1195 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1196 {
1197         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1198         ddt_entry_t dde = { 0 };
1199         int error;
1200         uint64_t n = 0;
1201
1202         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1203                 ddt_t *ddt;
1204
1205                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1206                         break;
1207                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1208                     (longlong_t)ddb->ddb_class,
1209                     (longlong_t)ddb->ddb_type,
1210                     (longlong_t)ddb->ddb_checksum,
1211                     (longlong_t)ddb->ddb_cursor);
1212
1213                 /* There should be no pending changes to the dedup table */
1214                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1215                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1216
1217                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1218                 n++;
1219
1220                 if (dsl_scan_check_pause(scn, NULL))
1221                         break;
1222         }
1223
1224         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1225             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1226             (int)scn->scn_pausing);
1227
1228         ASSERT(error == 0 || error == ENOENT);
1229         ASSERT(error != ENOENT ||
1230             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1231 }
1232
1233 /* ARGSUSED */
1234 void
1235 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1236     ddt_entry_t *dde, dmu_tx_t *tx)
1237 {
1238         const ddt_key_t *ddk = &dde->dde_key;
1239         ddt_phys_t *ddp = dde->dde_phys;
1240         blkptr_t bp;
1241         zbookmark_t zb = { 0 };
1242
1243         if (scn->scn_phys.scn_state != DSS_SCANNING)
1244                 return;
1245
1246         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1247                 if (ddp->ddp_phys_birth == 0 ||
1248                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1249                         continue;
1250                 ddt_bp_create(checksum, ddk, ddp, &bp);
1251
1252                 scn->scn_visited_this_txg++;
1253                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1254         }
1255 }
1256
1257 static void
1258 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1259 {
1260         dsl_pool_t *dp = scn->scn_dp;
1261         zap_cursor_t zc;
1262         zap_attribute_t za;
1263
1264         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1265             scn->scn_phys.scn_ddt_class_max) {
1266                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1267                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1268                 dsl_scan_ddt(scn, tx);
1269                 if (scn->scn_pausing)
1270                         return;
1271         }
1272
1273         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1274                 /* First do the MOS & ORIGIN */
1275
1276                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1277                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1278                 dsl_scan_visit_rootbp(scn, NULL,
1279                     &dp->dp_meta_rootbp, tx);
1280                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1281                 if (scn->scn_pausing)
1282                         return;
1283
1284                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1285                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1286                             enqueue_cb, tx, DS_FIND_CHILDREN));
1287                 } else {
1288                         dsl_scan_visitds(scn,
1289                             dp->dp_origin_snap->ds_object, tx);
1290                 }
1291                 ASSERT(!scn->scn_pausing);
1292         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1293             ZB_DESTROYED_OBJSET) {
1294                 /*
1295                  * If we were paused, continue from here.  Note if the
1296                  * ds we were paused on was deleted, the zb_objset may
1297                  * be -1, so we will skip this and find a new objset
1298                  * below.
1299                  */
1300                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1301                 if (scn->scn_pausing)
1302                         return;
1303         }
1304
1305         /*
1306          * In case we were paused right at the end of the ds, zero the
1307          * bookmark so we don't think that we're still trying to resume.
1308          */
1309         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1310
1311         /* keep pulling things out of the zap-object-as-queue */
1312         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1313             scn->scn_phys.scn_queue_obj),
1314             zap_cursor_retrieve(&zc, &za) == 0) {
1315                 dsl_dataset_t *ds;
1316                 uint64_t dsobj;
1317
1318                 dsobj = strtonum(za.za_name, NULL);
1319                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1320                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1321
1322                 /* Set up min/max txg */
1323                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1324                 if (za.za_first_integer != 0) {
1325                         scn->scn_phys.scn_cur_min_txg =
1326                             MAX(scn->scn_phys.scn_min_txg,
1327                             za.za_first_integer);
1328                 } else {
1329                         scn->scn_phys.scn_cur_min_txg =
1330                             MAX(scn->scn_phys.scn_min_txg,
1331                             ds->ds_phys->ds_prev_snap_txg);
1332                 }
1333                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1334                 dsl_dataset_rele(ds, FTAG);
1335
1336                 dsl_scan_visitds(scn, dsobj, tx);
1337                 zap_cursor_fini(&zc);
1338                 if (scn->scn_pausing)
1339                         return;
1340         }
1341         zap_cursor_fini(&zc);
1342 }
1343
1344 static boolean_t
1345 dsl_scan_free_should_pause(dsl_scan_t *scn)
1346 {
1347         uint64_t elapsed_nanosecs;
1348
1349         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1350         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1351             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1352             txg_sync_waiting(scn->scn_dp)) ||
1353             spa_shutting_down(scn->scn_dp->dp_spa));
1354 }
1355
1356 static int
1357 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1358 {
1359         dsl_scan_t *scn = arg;
1360
1361         if (!scn->scn_is_bptree ||
1362             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1363                 if (dsl_scan_free_should_pause(scn))
1364                         return (SET_ERROR(ERESTART));
1365         }
1366
1367         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1368             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1369         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1370             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1371             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1372         scn->scn_visited_this_txg++;
1373         return (0);
1374 }
1375
1376 boolean_t
1377 dsl_scan_active(dsl_scan_t *scn)
1378 {
1379         spa_t *spa = scn->scn_dp->dp_spa;
1380         uint64_t used = 0, comp, uncomp;
1381
1382         if (spa->spa_load_state != SPA_LOAD_NONE)
1383                 return (B_FALSE);
1384         if (spa_shutting_down(spa))
1385                 return (B_FALSE);
1386
1387         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1388             scn->scn_async_destroying)
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(&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                         ASSERT(scn->scn_async_destroying);
1447                         scn->scn_is_bptree = B_TRUE;
1448                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1449                             NULL, ZIO_FLAG_MUSTSUCCEED);
1450                         err = bptree_iterate(dp->dp_meta_objset,
1451                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1452                             scn, tx);
1453                         VERIFY0(zio_wait(scn->scn_zio_root));
1454
1455                         if (err == 0) {
1456                                 zfeature_info_t *feat = &spa_feature_table
1457                                     [SPA_FEATURE_ASYNC_DESTROY];
1458                                 /* finished; deactivate async destroy feature */
1459                                 spa_feature_decr(spa, feat, tx);
1460                                 ASSERT(!spa_feature_is_active(spa, feat));
1461                                 VERIFY0(zap_remove(dp->dp_meta_objset,
1462                                     DMU_POOL_DIRECTORY_OBJECT,
1463                                     DMU_POOL_BPTREE_OBJ, tx));
1464                                 VERIFY0(bptree_free(dp->dp_meta_objset,
1465                                     dp->dp_bptree_obj, tx));
1466                                 dp->dp_bptree_obj = 0;
1467                                 scn->scn_async_destroying = B_FALSE;
1468                         }
1469                 }
1470                 if (scn->scn_visited_this_txg) {
1471                         zfs_dbgmsg("freed %llu blocks in %llums from "
1472                             "free_bpobj/bptree txg %llu",
1473                             (longlong_t)scn->scn_visited_this_txg,
1474                             (longlong_t)
1475                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1476                             (longlong_t)tx->tx_txg);
1477                         scn->scn_visited_this_txg = 0;
1478                         /*
1479                          * Re-sync the ddt so that we can further modify
1480                          * it when doing bprewrite.
1481                          */
1482                         ddt_sync(spa, tx->tx_txg);
1483                 }
1484                 if (err == ERESTART)
1485                         return;
1486         }
1487
1488         if (scn->scn_phys.scn_state != DSS_SCANNING)
1489                 return;
1490
1491         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1492             scn->scn_phys.scn_ddt_class_max) {
1493                 zfs_dbgmsg("doing scan sync txg %llu; "
1494                     "ddt bm=%llu/%llu/%llu/%llx",
1495                     (longlong_t)tx->tx_txg,
1496                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1497                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1498                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1499                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1500                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1501                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1502                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1503                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1504         } else {
1505                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1506                     (longlong_t)tx->tx_txg,
1507                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1508                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1509                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1510                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1511         }
1512
1513         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1514             NULL, ZIO_FLAG_CANFAIL);
1515         dsl_pool_config_enter(dp, FTAG);
1516         dsl_scan_visit(scn, tx);
1517         dsl_pool_config_exit(dp, FTAG);
1518         (void) zio_wait(scn->scn_zio_root);
1519         scn->scn_zio_root = NULL;
1520
1521         zfs_dbgmsg("visited %llu blocks in %llums",
1522             (longlong_t)scn->scn_visited_this_txg,
1523             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1524
1525         if (!scn->scn_pausing) {
1526                 /* finished with scan. */
1527                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1528                 dsl_scan_done(scn, B_TRUE, tx);
1529         }
1530
1531         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1532                 mutex_enter(&spa->spa_scrub_lock);
1533                 while (spa->spa_scrub_inflight > 0) {
1534                         cv_wait(&spa->spa_scrub_io_cv,
1535                             &spa->spa_scrub_lock);
1536                 }
1537                 mutex_exit(&spa->spa_scrub_lock);
1538         }
1539
1540         dsl_scan_sync_state(scn, tx);
1541 }
1542
1543 /*
1544  * This will start a new scan, or restart an existing one.
1545  */
1546 void
1547 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1548 {
1549         if (txg == 0) {
1550                 dmu_tx_t *tx;
1551                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1552                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1553
1554                 txg = dmu_tx_get_txg(tx);
1555                 dp->dp_scan->scn_restart_txg = txg;
1556                 dmu_tx_commit(tx);
1557         } else {
1558                 dp->dp_scan->scn_restart_txg = txg;
1559         }
1560         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1561 }
1562
1563 boolean_t
1564 dsl_scan_resilvering(dsl_pool_t *dp)
1565 {
1566         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1567             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1568 }
1569
1570 /*
1571  * scrub consumers
1572  */
1573
1574 static void
1575 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1576 {
1577         int i;
1578
1579         /*
1580          * If we resume after a reboot, zab will be NULL; don't record
1581          * incomplete stats in that case.
1582          */
1583         if (zab == NULL)
1584                 return;
1585
1586         for (i = 0; i < 4; i++) {
1587                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1588                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1589                 if (t & DMU_OT_NEWTYPE)
1590                         t = DMU_OT_OTHER;
1591                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1592                 int equal;
1593
1594                 zb->zb_count++;
1595                 zb->zb_asize += BP_GET_ASIZE(bp);
1596                 zb->zb_lsize += BP_GET_LSIZE(bp);
1597                 zb->zb_psize += BP_GET_PSIZE(bp);
1598                 zb->zb_gangs += BP_COUNT_GANG(bp);
1599
1600                 switch (BP_GET_NDVAS(bp)) {
1601                 case 2:
1602                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1603                             DVA_GET_VDEV(&bp->blk_dva[1]))
1604                                 zb->zb_ditto_2_of_2_samevdev++;
1605                         break;
1606                 case 3:
1607                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1608                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1609                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1610                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1611                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1612                             DVA_GET_VDEV(&bp->blk_dva[2]));
1613                         if (equal == 1)
1614                                 zb->zb_ditto_2_of_3_samevdev++;
1615                         else if (equal == 3)
1616                                 zb->zb_ditto_3_of_3_samevdev++;
1617                         break;
1618                 }
1619         }
1620 }
1621
1622 static void
1623 dsl_scan_scrub_done(zio_t *zio)
1624 {
1625         spa_t *spa = zio->io_spa;
1626
1627         zio_data_buf_free(zio->io_data, zio->io_size);
1628
1629         mutex_enter(&spa->spa_scrub_lock);
1630         spa->spa_scrub_inflight--;
1631         cv_broadcast(&spa->spa_scrub_io_cv);
1632
1633         if (zio->io_error && (zio->io_error != ECKSUM ||
1634             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1635                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1636         }
1637         mutex_exit(&spa->spa_scrub_lock);
1638 }
1639
1640 static int
1641 dsl_scan_scrub_cb(dsl_pool_t *dp,
1642     const blkptr_t *bp, const zbookmark_t *zb)
1643 {
1644         dsl_scan_t *scn = dp->dp_scan;
1645         size_t size = BP_GET_PSIZE(bp);
1646         spa_t *spa = dp->dp_spa;
1647         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1648         boolean_t needs_io;
1649         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1650         int zio_priority;
1651         unsigned int scan_delay = 0;
1652
1653         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1654             phys_birth >= scn->scn_phys.scn_max_txg)
1655                 return (0);
1656
1657         count_block(dp->dp_blkstats, bp);
1658
1659         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1660         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1661                 zio_flags |= ZIO_FLAG_SCRUB;
1662                 zio_priority = ZIO_PRIORITY_SCRUB;
1663                 needs_io = B_TRUE;
1664                 scan_delay = zfs_scrub_delay;
1665         } else {
1666                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1667                 zio_flags |= ZIO_FLAG_RESILVER;
1668                 zio_priority = ZIO_PRIORITY_RESILVER;
1669                 needs_io = B_FALSE;
1670                 scan_delay = zfs_resilver_delay;
1671         }
1672
1673         /* If it's an intent log block, failure is expected. */
1674         if (zb->zb_level == ZB_ZIL_LEVEL)
1675                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1676
1677         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1678                 vdev_t *vd = vdev_lookup_top(spa,
1679                     DVA_GET_VDEV(&bp->blk_dva[d]));
1680
1681                 /*
1682                  * Keep track of how much data we've examined so that
1683                  * zpool(1M) status can make useful progress reports.
1684                  */
1685                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1686                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1687
1688                 /* if it's a resilver, this may not be in the target range */
1689                 if (!needs_io) {
1690                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1691                                 /*
1692                                  * Gang members may be spread across multiple
1693                                  * vdevs, so the best estimate we have is the
1694                                  * scrub range, which has already been checked.
1695                                  * XXX -- it would be better to change our
1696                                  * allocation policy to ensure that all
1697                                  * gang members reside on the same vdev.
1698                                  */
1699                                 needs_io = B_TRUE;
1700                         } else {
1701                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1702                                     phys_birth, 1);
1703                         }
1704                 }
1705         }
1706
1707         if (needs_io && !zfs_no_scrub_io) {
1708                 vdev_t *rvd = spa->spa_root_vdev;
1709                 uint64_t maxinflight = rvd->vdev_children *
1710                     MAX(zfs_top_maxinflight, 1);
1711                 void *data = zio_data_buf_alloc(size);
1712
1713                 mutex_enter(&spa->spa_scrub_lock);
1714                 while (spa->spa_scrub_inflight >= maxinflight)
1715                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1716                 spa->spa_scrub_inflight++;
1717                 mutex_exit(&spa->spa_scrub_lock);
1718
1719                 /*
1720                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1721                  * then throttle our workload to limit the impact of a scan.
1722                  */
1723                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1724                         delay(MAX((int)scan_delay, 0));
1725
1726                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1727                     dsl_scan_scrub_done, NULL, zio_priority,
1728                     zio_flags, zb));
1729         }
1730
1731         /* do not relocate this block */
1732         return (0);
1733 }
1734
1735 int
1736 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1737 {
1738         spa_t *spa = dp->dp_spa;
1739
1740         /*
1741          * Purge all vdev caches and probe all devices.  We do this here
1742          * rather than in sync context because this requires a writer lock
1743          * on the spa_config lock, which we can't do from sync context.  The
1744          * spa_scrub_reopen flag indicates that vdev_open() should not
1745          * attempt to start another scrub.
1746          */
1747         spa_vdev_state_enter(spa, SCL_NONE);
1748         spa->spa_scrub_reopen = B_TRUE;
1749         vdev_reopen(spa->spa_root_vdev);
1750         spa->spa_scrub_reopen = B_FALSE;
1751         (void) spa_vdev_state_exit(spa, NULL, 0);
1752
1753         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1754             dsl_scan_setup_sync, &func, 0));
1755 }