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_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         scn->scn_done_txg = 0;
 228         spa_scan_stat_init(spa);
 229
 230         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 231                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 232
 233                 /* rewrite all disk labels */
 234                 vdev_config_dirty(spa->spa_root_vdev);
 235
 236                 if (vdev_resilver_needed(spa->spa_root_vdev,
 237                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 238                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 239                 } else {
 240                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 241                 }
 242
 243                 spa->spa_scrub_started = B_TRUE;
 244                 /*
 245                  * If this is an incremental scrub, limit the DDT scrub phase
 246                  * to just the auto-ditto class (for correctness); the rest
 247                  * of the scrub should go faster using top-down pruning.
 248                  */
 249                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 250                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 251
 252         }
 253
 254         /* back to the generic stuff */
 255
 256         if (dp->dp_blkstats == NULL) {
 257                 dp->dp_blkstats =
 258                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 259         }
 260         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 261
 262         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 263                 ot = DMU_OT_ZAP_OTHER;
 264
 265         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 266             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 267
 268         dsl_scan_sync_state(scn, tx);
 269
 270         spa_history_log_internal(spa, "scan setup", tx,
 271             "func=%u mintxg=%llu maxtxg=%llu",
 272             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 273 }
 274
 275 /* ARGSUSED */
 276 static void
 277 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 278 {
 279         static const char *old_names[] = {
 280                 "scrub_bookmark",
 281                 "scrub_ddt_bookmark",
 282                 "scrub_ddt_class_max",
 283                 "scrub_queue",
 284                 "scrub_min_txg",
 285                 "scrub_max_txg",
 286                 "scrub_func",
 287                 "scrub_errors",
 288                 NULL
 289         };
 290
 291         dsl_pool_t *dp = scn->scn_dp;
 292         spa_t *spa = dp->dp_spa;
 293         int i;
 294
 295         /* Remove any remnants of an old-style scrub. */
 296         for (i = 0; old_names[i]; i++) {
 297                 (void) zap_remove(dp->dp_meta_objset,
 298                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 299         }
 300
 301         if (scn->scn_phys.scn_queue_obj != 0) {
 302                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 303                     scn->scn_phys.scn_queue_obj, tx));
 304                 scn->scn_phys.scn_queue_obj = 0;
 305         }
 306
 307         /*
 308          * If we were "restarted" from a stopped state, don't bother
 309          * with anything else.
 310          */
 311         if (scn->scn_phys.scn_state != DSS_SCANNING)
 312                 return;
 313
 314         if (complete)
 315                 scn->scn_phys.scn_state = DSS_FINISHED;
 316         else
 317                 scn->scn_phys.scn_state = DSS_CANCELED;
 318
 319         spa_history_log_internal(spa, "scan done", tx,
 320             "complete=%u", complete);
 321
 322         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 323                 mutex_enter(&spa->spa_scrub_lock);
 324                 while (spa->spa_scrub_inflight > 0) {
 325                         cv_wait(&spa->spa_scrub_io_cv,
 326                             &spa->spa_scrub_lock);
 327                 }
 328                 mutex_exit(&spa->spa_scrub_lock);
 329                 spa->spa_scrub_started = B_FALSE;
 330                 spa->spa_scrub_active = B_FALSE;
 331
 332                 /*
 333                  * If the scrub/resilver completed, update all DTLs to
 334                  * reflect this.  Whether it succeeded or not, vacate
 335                  * all temporary scrub DTLs.
 336                  */
 337                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 338                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 339                 if (complete) {
 340                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 341                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 342                 }
 343                 spa_errlog_rotate(spa);
 344
 345                 /*
 346                  * We may have finished replacing a device.
 347                  * Let the async thread assess this and handle the detach.
 348                  */
 349                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 350         }
 351
 352         scn->scn_phys.scn_end_time = gethrestime_sec();
 353 }
 354
 355 /* ARGSUSED */
 356 static int
 357 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 358 {
 359         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 360
 361         if (scn->scn_phys.scn_state != DSS_SCANNING)
 362                 return (SET_ERROR(ENOENT));
 363         return (0);
 364 }
 365
 366 /* ARGSUSED */
 367 static void
 368 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 369 {
 370         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 371
 372         dsl_scan_done(scn, B_FALSE, tx);
 373         dsl_scan_sync_state(scn, tx);
 374 }
 375
 376 int
 377 dsl_scan_cancel(dsl_pool_t *dp)
 378 {
 379         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 380             dsl_scan_cancel_sync, NULL, 3));
 381 }
 382
 383 static void dsl_scan_visitbp(blkptr_t *bp,
 384     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 385     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 386     dmu_tx_t *tx);
 387 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 388     dmu_objset_type_t ostype,
 389     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 390
 391 void
 392 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 393 {
 394         zio_free(dp->dp_spa, txg, bp);
 395 }
 396
 397 void
 398 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 399 {
 400         ASSERT(dsl_pool_sync_context(dp));
 401         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 402             pio->io_flags));
 403 }
 404
 405 static uint64_t
 406 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 407 {
 408         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 409         if (dsl_dataset_is_snapshot(ds))
 410                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 411         return (smt);
 412 }
 413
 414 static void
 415 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 416 {
 417         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 418             DMU_POOL_DIRECTORY_OBJECT,
 419             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 420             &scn->scn_phys, tx));
 421 }
 422
 423 static boolean_t
 424 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 425 {
 426         uint64_t elapsed_nanosecs;
 427         unsigned int mintime;
 428
 429         /* we never skip user/group accounting objects */
 430         if (zb && (int64_t)zb->zb_object < 0)
 431                 return (B_FALSE);
 432
 433         if (scn->scn_pausing)
 434                 return (B_TRUE); /* we're already pausing */
 435
 436         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 437                 return (B_FALSE); /* we're resuming */
 438
 439         /* We only know how to resume from level-0 blocks. */
 440         if (zb && zb->zb_level != 0)
 441                 return (B_FALSE);
 442
 443         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 444             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 445         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 446         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 447             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 448             txg_sync_waiting(scn->scn_dp)) ||
 449             spa_shutting_down(scn->scn_dp->dp_spa)) {
 450                 if (zb) {
 451                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 452                             (longlong_t)zb->zb_objset,
 453                             (longlong_t)zb->zb_object,
 454                             (longlong_t)zb->zb_level,
 455                             (longlong_t)zb->zb_blkid);
 456                         scn->scn_phys.scn_bookmark = *zb;
 457                 }
 458                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 459                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 460                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 461                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 462                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 463                 scn->scn_pausing = B_TRUE;
 464                 return (B_TRUE);
 465         }
 466         return (B_FALSE);
 467 }
 468
 469 typedef struct zil_scan_arg {
 470         dsl_pool_t      *zsa_dp;
 471         zil_header_t    *zsa_zh;
 472 } zil_scan_arg_t;
 473
 474 /* ARGSUSED */
 475 static int
 476 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 477 {
 478         zil_scan_arg_t *zsa = arg;
 479         dsl_pool_t *dp = zsa->zsa_dp;
 480         dsl_scan_t *scn = dp->dp_scan;
 481         zil_header_t *zh = zsa->zsa_zh;
 482         zbookmark_t zb;
 483
 484         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 485                 return (0);
 486
 487         /*
 488          * One block ("stubby") can be allocated a long time ago; we
 489          * want to visit that one because it has been allocated
 490          * (on-disk) even if it hasn't been claimed (even though for
 491          * scrub there's nothing to do to it).
 492          */
 493         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 494                 return (0);
 495
 496         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 497             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 498
 499         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 500         return (0);
 501 }
 502
 503 /* ARGSUSED */
 504 static int
 505 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 506 {
 507         if (lrc->lrc_txtype == TX_WRITE) {
 508                 zil_scan_arg_t *zsa = arg;
 509                 dsl_pool_t *dp = zsa->zsa_dp;
 510                 dsl_scan_t *scn = dp->dp_scan;
 511                 zil_header_t *zh = zsa->zsa_zh;
 512                 lr_write_t *lr = (lr_write_t *)lrc;
 513                 blkptr_t *bp = &lr->lr_blkptr;
 514                 zbookmark_t zb;
 515
 516                 if (BP_IS_HOLE(bp) ||
 517                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 518                         return (0);
 519
 520                 /*
 521                  * birth can be < claim_txg if this record's txg is
 522                  * already txg sync'ed (but this log block contains
 523                  * other records that are not synced)
 524                  */
 525                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 526                         return (0);
 527
 528                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 529                     lr->lr_foid, ZB_ZIL_LEVEL,
 530                     lr->lr_offset / BP_GET_LSIZE(bp));
 531
 532                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 533         }
 534         return (0);
 535 }
 536
 537 static void
 538 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 539 {
 540         uint64_t claim_txg = zh->zh_claim_txg;
 541         zil_scan_arg_t zsa = { dp, zh };
 542         zilog_t *zilog;
 543
 544         /*
 545          * We only want to visit blocks that have been claimed but not yet
 546          * replayed (or, in read-only mode, blocks that *would* be claimed).
 547          */
 548         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 549                 return;
 550
 551         zilog = zil_alloc(dp->dp_meta_objset, zh);
 552
 553         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 554             claim_txg);
 555
 556         zil_free(zilog);
 557 }
 558
 559 /* ARGSUSED */
 560 static void
 561 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 562     uint64_t objset, uint64_t object, uint64_t blkid)
 563 {
 564         zbookmark_t czb;
 565         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 566
 567         if (zfs_no_scrub_prefetch)
 568                 return;
 569
 570         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 571             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 572                 return;
 573
 574         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 575
 576         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 577             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 578             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 579 }
 580
 581 static boolean_t
 582 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 583     const zbookmark_t *zb)
 584 {
 585         /*
 586          * We never skip over user/group accounting objects (obj<0)
 587          */
 588         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 589             (int64_t)zb->zb_object >= 0) {
 590                 /*
 591                  * If we already visited this bp & everything below (in
 592                  * a prior txg sync), don't bother doing it again.
 593                  */
 594                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 595                         return (B_TRUE);
 596
 597                 /*
 598                  * If we found the block we're trying to resume from, or
 599                  * we went past it to a different object, zero it out to
 600                  * indicate that it's OK to start checking for pausing
 601                  * again.
 602                  */
 603                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 604                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 605                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 606                             (longlong_t)zb->zb_objset,
 607                             (longlong_t)zb->zb_object,
 608                             (longlong_t)zb->zb_level,
 609                             (longlong_t)zb->zb_blkid);
 610                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 611                 }
 612         }
 613         return (B_FALSE);
 614 }
 615
 616 /*
 617  * Return nonzero on i/o error.
 618  * Return new buf to write out in *bufp.
 619  */
 620 static int
 621 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 622     dnode_phys_t *dnp, const blkptr_t *bp,
 623     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 624 {
 625         dsl_pool_t *dp = scn->scn_dp;
 626         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 627         int err;
 628
 629         if (BP_GET_LEVEL(bp) > 0) {
 630                 uint32_t flags = ARC_WAIT;
 631                 int i;
 632                 blkptr_t *cbp;
 633                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 634
 635                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 636                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 637                 if (err) {
 638                         scn->scn_phys.scn_errors++;
 639                         return (err);
 640                 }
 641                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 642                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 643                             zb->zb_object, zb->zb_blkid * epb + i);
 644                 }
 645                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 646                         zbookmark_t czb;
 647
 648                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 649                             zb->zb_level - 1,
 650                             zb->zb_blkid * epb + i);
 651                         dsl_scan_visitbp(cbp, &czb, dnp,
 652                             *bufp, ds, scn, ostype, tx);
 653                 }
 654         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 655                 uint32_t flags = ARC_WAIT;
 656
 657                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 658                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 659                 if (err) {
 660                         scn->scn_phys.scn_errors++;
 661                         return (err);
 662                 }
 663         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 664                 uint32_t flags = ARC_WAIT;
 665                 dnode_phys_t *cdnp;
 666                 int i, j;
 667                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 668
 669                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 670                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 671                 if (err) {
 672                         scn->scn_phys.scn_errors++;
 673                         return (err);
 674                 }
 675                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 676                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 677                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 678                                 dsl_scan_prefetch(scn, *bufp, cbp,
 679                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 680                         }
 681                 }
 682                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 683                         dsl_scan_visitdnode(scn, ds, ostype,
 684                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 685                 }
 686
 687         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 688                 uint32_t flags = ARC_WAIT;
 689                 objset_phys_t *osp;
 690
 691                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 692                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 693                 if (err) {
 694                         scn->scn_phys.scn_errors++;
 695                         return (err);
 696                 }
 697
 698                 osp = (*bufp)->b_data;
 699
 700                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 701                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 702
 703                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 704                         /*
 705                          * We also always visit user/group accounting
 706                          * objects, and never skip them, even if we are
 707                          * pausing.  This is necessary so that the space
 708                          * deltas from this txg get integrated.
 709                          */
 710                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 711                             &osp->os_groupused_dnode, *bufp,
 712                             DMU_GROUPUSED_OBJECT, tx);
 713                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 714                             &osp->os_userused_dnode, *bufp,
 715                             DMU_USERUSED_OBJECT, tx);
 716                 }
 717         }
 718
 719         return (0);
 720 }
 721
 722 static void
 723 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 724     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 725     uint64_t object, dmu_tx_t *tx)
 726 {
 727         int j;
 728
 729         for (j = 0; j < dnp->dn_nblkptr; j++) {
 730                 zbookmark_t czb;
 731
 732                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 733                     dnp->dn_nlevels - 1, j);
 734                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 735                     &czb, dnp, buf, ds, scn, ostype, tx);
 736         }
 737
 738         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 739                 zbookmark_t czb;
 740                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 741                     0, DMU_SPILL_BLKID);
 742                 dsl_scan_visitbp(&dnp->dn_spill,
 743                     &czb, dnp, buf, ds, scn, ostype, tx);
 744         }
 745 }
 746
 747 /*
 748  * The arguments are in this order because mdb can only print the
 749  * first 5; we want them to be useful.
 750  */
 751 static void
 752 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 753     dnode_phys_t *dnp, arc_buf_t *pbuf,
 754     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 755     dmu_tx_t *tx)
 756 {
 757         dsl_pool_t *dp = scn->scn_dp;
 758         arc_buf_t *buf = NULL;
 759         blkptr_t bp_toread = *bp;
 760
 761         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 762
 763         if (dsl_scan_check_pause(scn, zb))
 764                 return;
 765
 766         if (dsl_scan_check_resume(scn, dnp, zb))
 767                 return;
 768
 769         if (BP_IS_HOLE(bp))
 770                 return;
 771
 772         scn->scn_visited_this_txg++;
 773
 774         dprintf_bp(bp,
 775             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 776             ds, ds ? ds->ds_object : 0,
 777             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 778             pbuf, bp);
 779
 780         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 781                 return;
 782
 783         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 784             &buf) != 0)
 785                 return;
 786
 787         /*
 788          * If dsl_scan_ddt() has aready visited this block, it will have
 789          * already done any translations or scrubbing, so don't call the
 790          * callback again.
 791          */
 792         if (ddt_class_contains(dp->dp_spa,
 793             scn->scn_phys.scn_ddt_class_max, bp)) {
 794                 ASSERT(buf == NULL);
 795                 return;
 796         }
 797
 798         /*
 799          * If this block is from the future (after cur_max_txg), then we
 800          * are doing this on behalf of a deleted snapshot, and we will
 801          * revisit the future block on the next pass of this dataset.
 802          * Don't scan it now unless we need to because something
 803          * under it was modified.
 804          */
 805         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 806                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 807         }
 808         if (buf)
 809                 (void) arc_buf_remove_ref(buf, &buf);
 810 }
 811
 812 static void
 813 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 814     dmu_tx_t *tx)
 815 {
 816         zbookmark_t zb;
 817
 818         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 819             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 820         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 821             ds, scn, DMU_OST_NONE, tx);
 822
 823         dprintf_ds(ds, "finished scan%s", "");
 824 }
 825
 826 void
 827 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 828 {
 829         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 830         dsl_scan_t *scn = dp->dp_scan;
 831         uint64_t mintxg;
 832
 833         if (scn->scn_phys.scn_state != DSS_SCANNING)
 834                 return;
 835
 836         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 837                 if (dsl_dataset_is_snapshot(ds)) {
 838                         /* Note, scn_cur_{min,max}_txg stays the same. */
 839                         scn->scn_phys.scn_bookmark.zb_objset =
 840                             ds->ds_phys->ds_next_snap_obj;
 841                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 842                             "reset zb_objset to %llu",
 843                             (u_longlong_t)ds->ds_object,
 844                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 845                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 846                 } else {
 847                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 848                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 849                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 850                             "reset bookmark to -1,0,0,0",
 851                             (u_longlong_t)ds->ds_object);
 852                 }
 853         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 854             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 855                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 856                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 857                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 858                 if (dsl_dataset_is_snapshot(ds)) {
 859                         /*
 860                          * We keep the same mintxg; it could be >
 861                          * ds_creation_txg if the previous snapshot was
 862                          * deleted too.
 863                          */
 864                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 865                             scn->scn_phys.scn_queue_obj,
 866                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 867                         zfs_dbgmsg("destroying ds %llu; in queue; "
 868                             "replacing with %llu",
 869                             (u_longlong_t)ds->ds_object,
 870                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 871                 } else {
 872                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 873                             (u_longlong_t)ds->ds_object);
 874                 }
 875         } else {
 876                 zfs_dbgmsg("destroying ds %llu; ignoring",
 877                     (u_longlong_t)ds->ds_object);
 878         }
 879
 880         /*
 881          * dsl_scan_sync() should be called after this, and should sync
 882          * out our changed state, but just to be safe, do it here.
 883          */
 884         dsl_scan_sync_state(scn, tx);
 885 }
 886
 887 void
 888 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 889 {
 890         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 891         dsl_scan_t *scn = dp->dp_scan;
 892         uint64_t mintxg;
 893
 894         if (scn->scn_phys.scn_state != DSS_SCANNING)
 895                 return;
 896
 897         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 898
 899         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 900                 scn->scn_phys.scn_bookmark.zb_objset =
 901                     ds->ds_phys->ds_prev_snap_obj;
 902                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 903                     "reset zb_objset to %llu",
 904                     (u_longlong_t)ds->ds_object,
 905                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 906         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 907             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 908                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 909                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 910                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 911                     scn->scn_phys.scn_queue_obj,
 912                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 913                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 914                     "replacing with %llu",
 915                     (u_longlong_t)ds->ds_object,
 916                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 917         }
 918         dsl_scan_sync_state(scn, tx);
 919 }
 920
 921 void
 922 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 923 {
 924         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 925         dsl_scan_t *scn = dp->dp_scan;
 926         uint64_t mintxg;
 927
 928         if (scn->scn_phys.scn_state != DSS_SCANNING)
 929                 return;
 930
 931         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 932                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 933                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 934                     "reset zb_objset to %llu",
 935                     (u_longlong_t)ds1->ds_object,
 936                     (u_longlong_t)ds2->ds_object);
 937         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 938                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 939                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 940                     "reset zb_objset to %llu",
 941                     (u_longlong_t)ds2->ds_object,
 942                     (u_longlong_t)ds1->ds_object);
 943         }
 944
 945         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 946             ds1->ds_object, &mintxg) == 0) {
 947                 int err;
 948
 949                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 950                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 951                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 952                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 953                 err = zap_add_int_key(dp->dp_meta_objset,
 954                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 955                 VERIFY(err == 0 || err == EEXIST);
 956                 if (err == EEXIST) {
 957                         /* Both were there to begin with */
 958                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 959                             scn->scn_phys.scn_queue_obj,
 960                             ds1->ds_object, mintxg, tx));
 961                 }
 962                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 963                     "replacing with %llu",
 964                     (u_longlong_t)ds1->ds_object,
 965                     (u_longlong_t)ds2->ds_object);
 966         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 967             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 968                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 969                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 970                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 971                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 972                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 973                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 974                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 975                     "replacing with %llu",
 976                     (u_longlong_t)ds2->ds_object,
 977                     (u_longlong_t)ds1->ds_object);
 978         }
 979
 980         dsl_scan_sync_state(scn, tx);
 981 }
 982
 983 struct enqueue_clones_arg {
 984         dmu_tx_t *tx;
 985         uint64_t originobj;
 986 };
 987
 988 /* ARGSUSED */
 989 static int
 990 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 991 {
 992         struct enqueue_clones_arg *eca = arg;
 993         dsl_dataset_t *ds;
 994         int err;
 995         dsl_scan_t *scn = dp->dp_scan;
 996
 997         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
 998                 return (0);
 999
1000         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1001         if (err)
1002                 return (err);
1003
1004         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1005                 dsl_dataset_t *prev;
1006                 err = dsl_dataset_hold_obj(dp,
1007                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1008
1009                 dsl_dataset_rele(ds, FTAG);
1010                 if (err)
1011                         return (err);
1012                 ds = prev;
1013         }
1014         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1015             scn->scn_phys.scn_queue_obj, ds->ds_object,
1016             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1017         dsl_dataset_rele(ds, FTAG);
1018         return (0);
1019 }
1020
1021 static void
1022 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1023 {
1024         dsl_pool_t *dp = scn->scn_dp;
1025         dsl_dataset_t *ds;
1026         objset_t *os;
1027
1028         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1029
1030         if (dmu_objset_from_ds(ds, &os))
1031                 goto out;
1032
1033         /*
1034          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1035          * snapshots can have ZIL block pointers (which may be the same
1036          * BP as in the head), they must be ignored.  So we traverse the
1037          * ZIL here, rather than in scan_recurse(), because the regular
1038          * snapshot block-sharing rules don't apply to it.
1039          */
1040         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1041                 dsl_scan_zil(dp, &os->os_zil_header);
1042
1043         /*
1044          * Iterate over the bps in this ds.
1045          */
1046         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1047         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1048
1049         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1050         dsl_dataset_name(ds, dsname);
1051         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1052             "pausing=%u",
1053             (longlong_t)dsobj, dsname,
1054             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1055             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1056             (int)scn->scn_pausing);
1057         kmem_free(dsname, ZFS_MAXNAMELEN);
1058
1059         if (scn->scn_pausing)
1060                 goto out;
1061
1062         /*
1063          * We've finished this pass over this dataset.
1064          */
1065
1066         /*
1067          * If we did not completely visit this dataset, do another pass.
1068          */
1069         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1070                 zfs_dbgmsg("incomplete pass; visiting again");
1071                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1072                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1073                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1074                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1075                 goto out;
1076         }
1077
1078         /*
1079          * Add descendent datasets to work queue.
1080          */
1081         if (ds->ds_phys->ds_next_snap_obj != 0) {
1082                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1083                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1084                     ds->ds_phys->ds_creation_txg, tx) == 0);
1085         }
1086         if (ds->ds_phys->ds_num_children > 1) {
1087                 boolean_t usenext = B_FALSE;
1088                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1089                         uint64_t count;
1090                         /*
1091                          * A bug in a previous version of the code could
1092                          * cause upgrade_clones_cb() to not set
1093                          * ds_next_snap_obj when it should, leading to a
1094                          * missing entry.  Therefore we can only use the
1095                          * next_clones_obj when its count is correct.
1096                          */
1097                         int err = zap_count(dp->dp_meta_objset,
1098                             ds->ds_phys->ds_next_clones_obj, &count);
1099                         if (err == 0 &&
1100                             count == ds->ds_phys->ds_num_children - 1)
1101                                 usenext = B_TRUE;
1102                 }
1103
1104                 if (usenext) {
1105                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1106                             ds->ds_phys->ds_next_clones_obj,
1107                             scn->scn_phys.scn_queue_obj,
1108                             ds->ds_phys->ds_creation_txg, tx));
1109                 } else {
1110                         struct enqueue_clones_arg eca;
1111                         eca.tx = tx;
1112                         eca.originobj = ds->ds_object;
1113
1114                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1115                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1116                 }
1117         }
1118
1119 out:
1120         dsl_dataset_rele(ds, FTAG);
1121 }
1122
1123 /* ARGSUSED */
1124 static int
1125 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1126 {
1127         dmu_tx_t *tx = arg;
1128         dsl_dataset_t *ds;
1129         int err;
1130         dsl_scan_t *scn = dp->dp_scan;
1131
1132         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1133         if (err)
1134                 return (err);
1135
1136         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1137                 dsl_dataset_t *prev;
1138                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1139                     FTAG, &prev);
1140                 if (err) {
1141                         dsl_dataset_rele(ds, FTAG);
1142                         return (err);
1143                 }
1144
1145                 /*
1146                  * If this is a clone, we don't need to worry about it for now.
1147                  */
1148                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1149                         dsl_dataset_rele(ds, FTAG);
1150                         dsl_dataset_rele(prev, FTAG);
1151                         return (0);
1152                 }
1153                 dsl_dataset_rele(ds, FTAG);
1154                 ds = prev;
1155         }
1156
1157         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1158             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1159         dsl_dataset_rele(ds, FTAG);
1160         return (0);
1161 }
1162
1163 /*
1164  * Scrub/dedup interaction.
1165  *
1166  * If there are N references to a deduped block, we don't want to scrub it
1167  * N times -- ideally, we should scrub it exactly once.
1168  *
1169  * We leverage the fact that the dde's replication class (enum ddt_class)
1170  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1171  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1172  *
1173  * To prevent excess scrubbing, the scrub begins by walking the DDT
1174  * to find all blocks with refcnt > 1, and scrubs each of these once.
1175  * Since there are two replication classes which contain blocks with
1176  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1177  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1178  *
1179  * There would be nothing more to say if a block's refcnt couldn't change
1180  * during a scrub, but of course it can so we must account for changes
1181  * in a block's replication class.
1182  *
1183  * Here's an example of what can occur:
1184  *
1185  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1186  * when visited during the top-down scrub phase, it will be scrubbed twice.
1187  * This negates our scrub optimization, but is otherwise harmless.
1188  *
1189  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1190  * on each visit during the top-down scrub phase, it will never be scrubbed.
1191  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1192  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1193  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1194  * while a scrub is in progress, it scrubs the block right then.
1195  */
1196 static void
1197 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1198 {
1199         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1200         ddt_entry_t dde = { 0 };
1201         int error;
1202         uint64_t n = 0;
1203
1204         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1205                 ddt_t *ddt;
1206
1207                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1208                         break;
1209                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1210                     (longlong_t)ddb->ddb_class,
1211                     (longlong_t)ddb->ddb_type,
1212                     (longlong_t)ddb->ddb_checksum,
1213                     (longlong_t)ddb->ddb_cursor);
1214
1215                 /* There should be no pending changes to the dedup table */
1216                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1217                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1218
1219                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1220                 n++;
1221
1222                 if (dsl_scan_check_pause(scn, NULL))
1223                         break;
1224         }
1225
1226         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1227             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1228             (int)scn->scn_pausing);
1229
1230         ASSERT(error == 0 || error == ENOENT);
1231         ASSERT(error != ENOENT ||
1232             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1233 }
1234
1235 /* ARGSUSED */
1236 void
1237 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1238     ddt_entry_t *dde, dmu_tx_t *tx)
1239 {
1240         const ddt_key_t *ddk = &dde->dde_key;
1241         ddt_phys_t *ddp = dde->dde_phys;
1242         blkptr_t bp;
1243         zbookmark_t zb = { 0 };
1244
1245         if (scn->scn_phys.scn_state != DSS_SCANNING)
1246                 return;
1247
1248         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1249                 if (ddp->ddp_phys_birth == 0 ||
1250                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1251                         continue;
1252                 ddt_bp_create(checksum, ddk, ddp, &bp);
1253
1254                 scn->scn_visited_this_txg++;
1255                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1256         }
1257 }
1258
1259 static void
1260 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1261 {
1262         dsl_pool_t *dp = scn->scn_dp;
1263         zap_cursor_t zc;
1264         zap_attribute_t za;
1265
1266         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1267             scn->scn_phys.scn_ddt_class_max) {
1268                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1269                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1270                 dsl_scan_ddt(scn, tx);
1271                 if (scn->scn_pausing)
1272                         return;
1273         }
1274
1275         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1276                 /* First do the MOS & ORIGIN */
1277
1278                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1279                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1280                 dsl_scan_visit_rootbp(scn, NULL,
1281                     &dp->dp_meta_rootbp, tx);
1282                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1283                 if (scn->scn_pausing)
1284                         return;
1285
1286                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1287                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1288                             enqueue_cb, tx, DS_FIND_CHILDREN));
1289                 } else {
1290                         dsl_scan_visitds(scn,
1291                             dp->dp_origin_snap->ds_object, tx);
1292                 }
1293                 ASSERT(!scn->scn_pausing);
1294         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1295             ZB_DESTROYED_OBJSET) {
1296                 /*
1297                  * If we were paused, continue from here.  Note if the
1298                  * ds we were paused on was deleted, the zb_objset may
1299                  * be -1, so we will skip this and find a new objset
1300                  * below.
1301                  */
1302                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1303                 if (scn->scn_pausing)
1304                         return;
1305         }
1306
1307         /*
1308          * In case we were paused right at the end of the ds, zero the
1309          * bookmark so we don't think that we're still trying to resume.
1310          */
1311         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1312
1313         /* keep pulling things out of the zap-object-as-queue */
1314         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1315             scn->scn_phys.scn_queue_obj),
1316             zap_cursor_retrieve(&zc, &za) == 0) {
1317                 dsl_dataset_t *ds;
1318                 uint64_t dsobj;
1319
1320                 dsobj = strtonum(za.za_name, NULL);
1321                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1322                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1323
1324                 /* Set up min/max txg */
1325                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1326                 if (za.za_first_integer != 0) {
1327                         scn->scn_phys.scn_cur_min_txg =
1328                             MAX(scn->scn_phys.scn_min_txg,
1329                             za.za_first_integer);
1330                 } else {
1331                         scn->scn_phys.scn_cur_min_txg =
1332                             MAX(scn->scn_phys.scn_min_txg,
1333                             ds->ds_phys->ds_prev_snap_txg);
1334                 }
1335                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1336                 dsl_dataset_rele(ds, FTAG);
1337
1338                 dsl_scan_visitds(scn, dsobj, tx);
1339                 zap_cursor_fini(&zc);
1340                 if (scn->scn_pausing)
1341                         return;
1342         }
1343         zap_cursor_fini(&zc);
1344 }
1345
1346 static boolean_t
1347 dsl_scan_free_should_pause(dsl_scan_t *scn)
1348 {
1349         uint64_t elapsed_nanosecs;
1350
1351         if (zfs_recover)
1352                 return (B_FALSE);
1353
1354         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1355         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1356             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1357             txg_sync_waiting(scn->scn_dp)) ||
1358             spa_shutting_down(scn->scn_dp->dp_spa));
1359 }
1360
1361 static int
1362 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1363 {
1364         dsl_scan_t *scn = arg;
1365
1366         if (!scn->scn_is_bptree ||
1367             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1368                 if (dsl_scan_free_should_pause(scn))
1369                         return (SET_ERROR(ERESTART));
1370         }
1371
1372         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1373             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1374         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1375             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1376             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1377         scn->scn_visited_this_txg++;
1378         return (0);
1379 }
1380
1381 boolean_t
1382 dsl_scan_active(dsl_scan_t *scn)
1383 {
1384         spa_t *spa = scn->scn_dp->dp_spa;
1385         uint64_t used = 0, comp, uncomp;
1386
1387         if (spa->spa_load_state != SPA_LOAD_NONE)
1388                 return (B_FALSE);
1389         if (spa_shutting_down(spa))
1390                 return (B_FALSE);
1391         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1392             scn->scn_async_destroying)
1393                 return (B_TRUE);
1394
1395         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1396                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1397                     &used, &comp, &uncomp);
1398         }
1399         return (used != 0);
1400 }
1401
1402 void
1403 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1404 {
1405         dsl_scan_t *scn = dp->dp_scan;
1406         spa_t *spa = dp->dp_spa;
1407         int err;
1408
1409         /*
1410          * Check for scn_restart_txg before checking spa_load_state, so
1411          * that we can restart an old-style scan while the pool is being
1412          * imported (see dsl_scan_init).
1413          */
1414         if (scn->scn_restart_txg != 0 &&
1415             scn->scn_restart_txg <= tx->tx_txg) {
1416                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1417                 dsl_scan_done(scn, B_FALSE, tx);
1418                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1419                         func = POOL_SCAN_RESILVER;
1420                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1421                     func, tx->tx_txg);
1422                 dsl_scan_setup_sync(&func, tx);
1423         }
1424
1425         if (!dsl_scan_active(scn) ||
1426             spa_sync_pass(dp->dp_spa) > 1)
1427                 return;
1428
1429         scn->scn_visited_this_txg = 0;
1430         scn->scn_pausing = B_FALSE;
1431         scn->scn_sync_start_time = gethrtime();
1432         spa->spa_scrub_active = B_TRUE;
1433
1434         /*
1435          * First process the free list.  If we pause the free, don't do
1436          * any scanning.  This ensures that there is no free list when
1437          * we are scanning, so the scan code doesn't have to worry about
1438          * traversing it.
1439          */
1440         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1441                 scn->scn_is_bptree = B_FALSE;
1442                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1443                     NULL, ZIO_FLAG_MUSTSUCCEED);
1444                 err = bpobj_iterate(&dp->dp_free_bpobj,
1445                     dsl_scan_free_block_cb, scn, tx);
1446                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1447
1448                 if (err == 0 && spa_feature_is_active(spa,
1449                     SPA_FEATURE_ASYNC_DESTROY)) {
1450                         ASSERT(scn->scn_async_destroying);
1451                         scn->scn_is_bptree = B_TRUE;
1452                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1453                             NULL, ZIO_FLAG_MUSTSUCCEED);
1454                         err = bptree_iterate(dp->dp_meta_objset,
1455                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1456                             scn, tx);
1457                         VERIFY0(zio_wait(scn->scn_zio_root));
1458
1459                         if (err == 0) {
1460                                 /* finished; deactivate async destroy feature */
1461                                 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY,
1462                                     tx);
1463                                 ASSERT(!spa_feature_is_active(spa,
1464                                     SPA_FEATURE_ASYNC_DESTROY));
1465                                 VERIFY0(zap_remove(dp->dp_meta_objset,
1466                                     DMU_POOL_DIRECTORY_OBJECT,
1467                                     DMU_POOL_BPTREE_OBJ, tx));
1468                                 VERIFY0(bptree_free(dp->dp_meta_objset,
1469                                     dp->dp_bptree_obj, tx));
1470                                 dp->dp_bptree_obj = 0;
1471                                 scn->scn_async_destroying = B_FALSE;
1472                         }
1473                 }
1474                 if (scn->scn_visited_this_txg) {
1475                         zfs_dbgmsg("freed %llu blocks in %llums from "
1476                             "free_bpobj/bptree txg %llu",
1477                             (longlong_t)scn->scn_visited_this_txg,
1478                             (longlong_t)
1479                             NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1480                             (longlong_t)tx->tx_txg);
1481                         scn->scn_visited_this_txg = 0;
1482                         /*
1483                          * Re-sync the ddt so that we can further modify
1484                          * it when doing bprewrite.
1485                          */
1486                         ddt_sync(spa, tx->tx_txg);
1487                 }
1488                 if (err == ERESTART)
1489                         return;
1490         }
1491
1492         if (scn->scn_phys.scn_state != DSS_SCANNING)
1493                 return;
1494
1495         if (scn->scn_done_txg == tx->tx_txg) {
1496                 ASSERT(!scn->scn_pausing);
1497                 /* finished with scan. */
1498                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1499                 dsl_scan_done(scn, B_TRUE, tx);
1500                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1501                 dsl_scan_sync_state(scn, tx);
1502                 return;
1503         }
1504
1505         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1506             scn->scn_phys.scn_ddt_class_max) {
1507                 zfs_dbgmsg("doing scan sync txg %llu; "
1508                     "ddt bm=%llu/%llu/%llu/%llx",
1509                     (longlong_t)tx->tx_txg,
1510                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1511                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1512                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1513                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1514                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1515                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1516                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1517                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1518         } else {
1519                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1520                     (longlong_t)tx->tx_txg,
1521                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1522                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1523                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1524                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1525         }
1526
1527         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1528             NULL, ZIO_FLAG_CANFAIL);
1529         dsl_pool_config_enter(dp, FTAG);
1530         dsl_scan_visit(scn, tx);
1531         dsl_pool_config_exit(dp, FTAG);
1532         (void) zio_wait(scn->scn_zio_root);
1533         scn->scn_zio_root = NULL;
1534
1535         zfs_dbgmsg("visited %llu blocks in %llums",
1536             (longlong_t)scn->scn_visited_this_txg,
1537             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1538
1539         if (!scn->scn_pausing) {
1540                 scn->scn_done_txg = tx->tx_txg + 1;
1541                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1542                     tx->tx_txg, scn->scn_done_txg);
1543         }
1544
1545         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1546                 mutex_enter(&spa->spa_scrub_lock);
1547                 while (spa->spa_scrub_inflight > 0) {
1548                         cv_wait(&spa->spa_scrub_io_cv,
1549                             &spa->spa_scrub_lock);
1550                 }
1551                 mutex_exit(&spa->spa_scrub_lock);
1552         }
1553
1554         dsl_scan_sync_state(scn, tx);
1555 }
1556
1557 /*
1558  * This will start a new scan, or restart an existing one.
1559  */
1560 void
1561 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1562 {
1563         if (txg == 0) {
1564                 dmu_tx_t *tx;
1565                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1566                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1567
1568                 txg = dmu_tx_get_txg(tx);
1569                 dp->dp_scan->scn_restart_txg = txg;
1570                 dmu_tx_commit(tx);
1571         } else {
1572                 dp->dp_scan->scn_restart_txg = txg;
1573         }
1574         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1575 }
1576
1577 boolean_t
1578 dsl_scan_resilvering(dsl_pool_t *dp)
1579 {
1580         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1581             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1582 }
1583
1584 /*
1585  * scrub consumers
1586  */
1587
1588 static void
1589 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1590 {
1591         int i;
1592
1593         /*
1594          * If we resume after a reboot, zab will be NULL; don't record
1595          * incomplete stats in that case.
1596          */
1597         if (zab == NULL)
1598                 return;
1599
1600         for (i = 0; i < 4; i++) {
1601                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1602                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1603                 if (t & DMU_OT_NEWTYPE)
1604                         t = DMU_OT_OTHER;
1605                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1606                 int equal;
1607
1608                 zb->zb_count++;
1609                 zb->zb_asize += BP_GET_ASIZE(bp);
1610                 zb->zb_lsize += BP_GET_LSIZE(bp);
1611                 zb->zb_psize += BP_GET_PSIZE(bp);
1612                 zb->zb_gangs += BP_COUNT_GANG(bp);
1613
1614                 switch (BP_GET_NDVAS(bp)) {
1615                 case 2:
1616                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1617                             DVA_GET_VDEV(&bp->blk_dva[1]))
1618                                 zb->zb_ditto_2_of_2_samevdev++;
1619                         break;
1620                 case 3:
1621                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1622                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1623                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1624                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1625                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1626                             DVA_GET_VDEV(&bp->blk_dva[2]));
1627                         if (equal == 1)
1628                                 zb->zb_ditto_2_of_3_samevdev++;
1629                         else if (equal == 3)
1630                                 zb->zb_ditto_3_of_3_samevdev++;
1631                         break;
1632                 }
1633         }
1634 }
1635
1636 static void
1637 dsl_scan_scrub_done(zio_t *zio)
1638 {
1639         spa_t *spa = zio->io_spa;
1640
1641         zio_data_buf_free(zio->io_data, zio->io_size);
1642
1643         mutex_enter(&spa->spa_scrub_lock);
1644         spa->spa_scrub_inflight--;
1645         cv_broadcast(&spa->spa_scrub_io_cv);
1646
1647         if (zio->io_error && (zio->io_error != ECKSUM ||
1648             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1649                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1650         }
1651         mutex_exit(&spa->spa_scrub_lock);
1652 }
1653
1654 static int
1655 dsl_scan_scrub_cb(dsl_pool_t *dp,
1656     const blkptr_t *bp, const zbookmark_t *zb)
1657 {
1658         dsl_scan_t *scn = dp->dp_scan;
1659         size_t size = BP_GET_PSIZE(bp);
1660         spa_t *spa = dp->dp_spa;
1661         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1662         boolean_t needs_io;
1663         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1664         unsigned int scan_delay = 0;
1665
1666         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1667             phys_birth >= scn->scn_phys.scn_max_txg)
1668                 return (0);
1669
1670         count_block(dp->dp_blkstats, bp);
1671
1672         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1673         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1674                 zio_flags |= ZIO_FLAG_SCRUB;
1675                 needs_io = B_TRUE;
1676                 scan_delay = zfs_scrub_delay;
1677         } else {
1678                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1679                 zio_flags |= ZIO_FLAG_RESILVER;
1680                 needs_io = B_FALSE;
1681                 scan_delay = zfs_resilver_delay;
1682         }
1683
1684         /* If it's an intent log block, failure is expected. */
1685         if (zb->zb_level == ZB_ZIL_LEVEL)
1686                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1687
1688         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1689                 vdev_t *vd = vdev_lookup_top(spa,
1690                     DVA_GET_VDEV(&bp->blk_dva[d]));
1691
1692                 /*
1693                  * Keep track of how much data we've examined so that
1694                  * zpool(1M) status can make useful progress reports.
1695                  */
1696                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1697                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1698
1699                 /* if it's a resilver, this may not be in the target range */
1700                 if (!needs_io) {
1701                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1702                                 /*
1703                                  * Gang members may be spread across multiple
1704                                  * vdevs, so the best estimate we have is the
1705                                  * scrub range, which has already been checked.
1706                                  * XXX -- it would be better to change our
1707                                  * allocation policy to ensure that all
1708                                  * gang members reside on the same vdev.
1709                                  */
1710                                 needs_io = B_TRUE;
1711                         } else {
1712                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1713                                     phys_birth, 1);
1714                         }
1715                 }
1716         }
1717
1718         if (needs_io && !zfs_no_scrub_io) {
1719                 vdev_t *rvd = spa->spa_root_vdev;
1720                 uint64_t maxinflight = rvd->vdev_children *
1721                     MAX(zfs_top_maxinflight, 1);
1722                 void *data = zio_data_buf_alloc(size);
1723
1724                 mutex_enter(&spa->spa_scrub_lock);
1725                 while (spa->spa_scrub_inflight >= maxinflight)
1726                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1727                 spa->spa_scrub_inflight++;
1728                 mutex_exit(&spa->spa_scrub_lock);
1729
1730                 /*
1731                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1732                  * then throttle our workload to limit the impact of a scan.
1733                  */
1734                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1735                         delay(MAX((int)scan_delay, 0));
1736
1737                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1738                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1739                     zio_flags, zb));
1740         }
1741
1742         /* do not relocate this block */
1743         return (0);
1744 }
1745
1746 int
1747 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1748 {
1749         spa_t *spa = dp->dp_spa;
1750
1751         /*
1752          * Purge all vdev caches and probe all devices.  We do this here
1753          * rather than in sync context because this requires a writer lock
1754          * on the spa_config lock, which we can't do from sync context.  The
1755          * spa_scrub_reopen flag indicates that vdev_open() should not
1756          * attempt to start another scrub.
1757          */
1758         spa_vdev_state_enter(spa, SCL_NONE);
1759         spa->spa_scrub_reopen = B_TRUE;
1760         vdev_reopen(spa->spa_root_vdev);
1761         spa->spa_scrub_reopen = B_FALSE;
1762         (void) spa_vdev_state_exit(spa, NULL, 0);
1763
1764         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1765             dsl_scan_setup_sync, &func, 0));
1766 }