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.
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.
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]
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2016 Gary Mills
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright 2017 Joyent, Inc.
26 * Copyright (c) 2017 Datto Inc.
29 #include <sys/dsl_scan.h>
30 #include <sys/dsl_pool.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_synctask.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/vdev_impl.h>
46 #include <sys/zil_impl.h>
47 #include <sys/zio_checksum.h>
50 #include <sys/sa_impl.h>
51 #include <sys/zfeature.h>
54 #include <sys/zfs_vfsops.h>
57 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
58 const zbookmark_phys_t *);
60 static scan_cb_t dsl_scan_scrub_cb;
61 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
62 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *);
63 static boolean_t dsl_scan_restarting(dsl_scan_t *, dmu_tx_t *);
65 unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
66 unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
67 unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
68 unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */
70 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
71 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
72 unsigned int zfs_obsolete_min_time_ms = 500; /* min millisecs to obsolete per txg */
73 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
75 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
76 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
78 SYSCTL_DECL(_vfs_zfs);
79 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RWTUN,
80 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RWTUN,
82 &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RWTUN,
84 &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
85 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RWTUN,
86 &zfs_scan_idle, 0, "Idle scan window in clock ticks");
87 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RWTUN,
88 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
89 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RWTUN,
90 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
91 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RWTUN,
92 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
93 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RWTUN,
94 &zfs_no_scrub_io, 0, "Disable scrub I/O");
95 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RWTUN,
96 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
98 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
99 /* max number of blocks to free in a single TXG */
100 uint64_t zfs_async_block_max_blocks = UINT64_MAX;
101 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
102 &zfs_async_block_max_blocks, 0, "Maximum number of blocks to free in one TXG");
105 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
106 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
107 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
109 extern int zfs_txg_timeout;
112 * Enable/disable the processing of the free_bpobj object.
114 boolean_t zfs_free_bpobj_enabled = B_TRUE;
116 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
117 &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
119 /* the order has to match pool_scan_type */
120 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
122 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
123 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
127 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
131 spa_t *spa = dp->dp_spa;
134 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
138 * It's possible that we're resuming a scan after a reboot so
139 * make sure that the scan_async_destroying flag is initialized
142 ASSERT(!scn->scn_async_destroying);
143 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
144 SPA_FEATURE_ASYNC_DESTROY);
146 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
147 "scrub_func", sizeof (uint64_t), 1, &f);
150 * There was an old-style scrub in progress. Restart a
151 * new-style scrub from the beginning.
153 scn->scn_restart_txg = txg;
154 zfs_dbgmsg("old-style scrub was in progress; "
155 "restarting new-style scrub in txg %llu",
156 scn->scn_restart_txg);
159 * Load the queue obj from the old location so that it
160 * can be freed by dsl_scan_done().
162 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
163 "scrub_queue", sizeof (uint64_t), 1,
164 &scn->scn_phys.scn_queue_obj);
166 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
167 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
174 if (scn->scn_phys.scn_state == DSS_SCANNING &&
175 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
177 * A new-type scrub was in progress on an old
178 * pool, and the pool was accessed by old
179 * software. Restart from the beginning, since
180 * the old software may have changed the pool in
183 scn->scn_restart_txg = txg;
184 zfs_dbgmsg("new-style scrub was modified "
185 "by old software; restarting in txg %llu",
186 scn->scn_restart_txg);
190 spa_scan_stat_init(spa);
195 dsl_scan_fini(dsl_pool_t *dp)
198 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
205 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
207 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
209 if (scn->scn_phys.scn_state == DSS_SCANNING)
210 return (SET_ERROR(EBUSY));
216 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
218 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
219 pool_scan_func_t *funcp = arg;
220 dmu_object_type_t ot = 0;
221 dsl_pool_t *dp = scn->scn_dp;
222 spa_t *spa = dp->dp_spa;
224 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
225 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
226 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
227 scn->scn_phys.scn_func = *funcp;
228 scn->scn_phys.scn_state = DSS_SCANNING;
229 scn->scn_phys.scn_min_txg = 0;
230 scn->scn_phys.scn_max_txg = tx->tx_txg;
231 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
232 scn->scn_phys.scn_start_time = gethrestime_sec();
233 scn->scn_phys.scn_errors = 0;
234 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
235 scn->scn_restart_txg = 0;
236 scn->scn_done_txg = 0;
237 spa_scan_stat_init(spa);
239 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
240 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
242 /* rewrite all disk labels */
243 vdev_config_dirty(spa->spa_root_vdev);
245 if (vdev_resilver_needed(spa->spa_root_vdev,
246 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
247 spa_event_notify(spa, NULL, NULL,
248 ESC_ZFS_RESILVER_START);
250 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
253 spa->spa_scrub_started = B_TRUE;
255 * If this is an incremental scrub, limit the DDT scrub phase
256 * to just the auto-ditto class (for correctness); the rest
257 * of the scrub should go faster using top-down pruning.
259 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
260 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
264 /* back to the generic stuff */
266 if (dp->dp_blkstats == NULL) {
268 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
270 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
272 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
273 ot = DMU_OT_ZAP_OTHER;
275 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
276 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
278 dsl_scan_sync_state(scn, tx);
280 spa_history_log_internal(spa, "scan setup", tx,
281 "func=%u mintxg=%llu maxtxg=%llu",
282 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
287 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
289 static const char *old_names[] = {
291 "scrub_ddt_bookmark",
292 "scrub_ddt_class_max",
301 dsl_pool_t *dp = scn->scn_dp;
302 spa_t *spa = dp->dp_spa;
305 /* Remove any remnants of an old-style scrub. */
306 for (i = 0; old_names[i]; i++) {
307 (void) zap_remove(dp->dp_meta_objset,
308 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
311 if (scn->scn_phys.scn_queue_obj != 0) {
312 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
313 scn->scn_phys.scn_queue_obj, tx));
314 scn->scn_phys.scn_queue_obj = 0;
317 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
320 * If we were "restarted" from a stopped state, don't bother
321 * with anything else.
323 if (scn->scn_phys.scn_state != DSS_SCANNING)
327 scn->scn_phys.scn_state = DSS_FINISHED;
329 scn->scn_phys.scn_state = DSS_CANCELED;
331 if (dsl_scan_restarting(scn, tx))
332 spa_history_log_internal(spa, "scan aborted, restarting", tx,
333 "errors=%llu", spa_get_errlog_size(spa));
335 spa_history_log_internal(spa, "scan cancelled", tx,
336 "errors=%llu", spa_get_errlog_size(spa));
338 spa_history_log_internal(spa, "scan done", tx,
339 "errors=%llu", spa_get_errlog_size(spa));
341 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
342 mutex_enter(&spa->spa_scrub_lock);
343 while (spa->spa_scrub_inflight > 0) {
344 cv_wait(&spa->spa_scrub_io_cv,
345 &spa->spa_scrub_lock);
347 mutex_exit(&spa->spa_scrub_lock);
348 spa->spa_scrub_started = B_FALSE;
349 spa->spa_scrub_active = B_FALSE;
352 * If the scrub/resilver completed, update all DTLs to
353 * reflect this. Whether it succeeded or not, vacate
354 * all temporary scrub DTLs.
356 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
357 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
359 spa_event_notify(spa, NULL, NULL,
360 scn->scn_phys.scn_min_txg ?
361 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
363 spa_errlog_rotate(spa);
366 * We may have finished replacing a device.
367 * Let the async thread assess this and handle the detach.
369 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
372 scn->scn_phys.scn_end_time = gethrestime_sec();
377 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
379 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
381 if (scn->scn_phys.scn_state != DSS_SCANNING)
382 return (SET_ERROR(ENOENT));
388 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
390 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
392 dsl_scan_done(scn, B_FALSE, tx);
393 dsl_scan_sync_state(scn, tx);
394 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
398 dsl_scan_cancel(dsl_pool_t *dp)
400 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
401 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
405 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
407 if (dsl_scan_scrubbing(scn->scn_dp) &&
408 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED)
415 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
417 pool_scrub_cmd_t *cmd = arg;
418 dsl_pool_t *dp = dmu_tx_pool(tx);
419 dsl_scan_t *scn = dp->dp_scan;
421 if (*cmd == POOL_SCRUB_PAUSE) {
422 /* can't pause a scrub when there is no in-progress scrub */
423 if (!dsl_scan_scrubbing(dp))
424 return (SET_ERROR(ENOENT));
426 /* can't pause a paused scrub */
427 if (dsl_scan_is_paused_scrub(scn))
428 return (SET_ERROR(EBUSY));
429 } else if (*cmd != POOL_SCRUB_NORMAL) {
430 return (SET_ERROR(ENOTSUP));
437 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
439 pool_scrub_cmd_t *cmd = arg;
440 dsl_pool_t *dp = dmu_tx_pool(tx);
441 spa_t *spa = dp->dp_spa;
442 dsl_scan_t *scn = dp->dp_scan;
444 if (*cmd == POOL_SCRUB_PAUSE) {
445 /* can't pause a scrub when there is no in-progress scrub */
446 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
447 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
448 dsl_scan_sync_state(scn, tx);
449 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
451 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
452 if (dsl_scan_is_paused_scrub(scn)) {
454 * We need to keep track of how much time we spend
455 * paused per pass so that we can adjust the scrub rate
456 * shown in the output of 'zpool status'
458 spa->spa_scan_pass_scrub_spent_paused +=
459 gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
460 spa->spa_scan_pass_scrub_pause = 0;
461 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
462 dsl_scan_sync_state(scn, tx);
468 * Set scrub pause/resume state if it makes sense to do so
471 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
473 return (dsl_sync_task(spa_name(dp->dp_spa),
474 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
475 ZFS_SPACE_CHECK_RESERVED));
479 dsl_scan_scrubbing(const dsl_pool_t *dp)
481 dsl_scan_t *scn = dp->dp_scan;
483 if (scn->scn_phys.scn_state == DSS_SCANNING &&
484 scn->scn_phys.scn_func == POOL_SCAN_SCRUB)
490 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
491 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
492 dmu_objset_type_t ostype, dmu_tx_t *tx);
493 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
494 dmu_objset_type_t ostype,
495 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
498 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
500 zio_free(dp->dp_spa, txg, bp);
504 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
506 ASSERT(dsl_pool_sync_context(dp));
507 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
512 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
514 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
515 if (ds->ds_is_snapshot)
516 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
521 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
523 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
524 DMU_POOL_DIRECTORY_OBJECT,
525 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
526 &scn->scn_phys, tx));
529 extern int zfs_vdev_async_write_active_min_dirty_percent;
532 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
534 /* we never skip user/group accounting objects */
535 if (zb && (int64_t)zb->zb_object < 0)
538 if (scn->scn_suspending)
539 return (B_TRUE); /* we're already suspending */
541 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
542 return (B_FALSE); /* we're resuming */
544 /* We only know how to resume from level-0 blocks. */
545 if (zb && zb->zb_level != 0)
550 * - we have scanned for the maximum time: an entire txg
551 * timeout (default 5 sec)
553 * - we have scanned for at least the minimum time (default 1 sec
554 * for scrub, 3 sec for resilver), and either we have sufficient
555 * dirty data that we are starting to write more quickly
556 * (default 30%), or someone is explicitly waiting for this txg
559 * - the spa is shutting down because this pool is being exported
560 * or the machine is rebooting.
562 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
563 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
564 uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
565 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
566 if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
567 (NSEC2MSEC(elapsed_nanosecs) > mintime &&
568 (txg_sync_waiting(scn->scn_dp) ||
569 dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
570 spa_shutting_down(scn->scn_dp->dp_spa)) {
572 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
573 (longlong_t)zb->zb_objset,
574 (longlong_t)zb->zb_object,
575 (longlong_t)zb->zb_level,
576 (longlong_t)zb->zb_blkid);
577 scn->scn_phys.scn_bookmark = *zb;
579 dprintf("suspending at DDT bookmark %llx/%llx/%llx/%llx\n",
580 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
581 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
582 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
583 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
584 scn->scn_suspending = B_TRUE;
590 typedef struct zil_scan_arg {
592 zil_header_t *zsa_zh;
597 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
599 zil_scan_arg_t *zsa = arg;
600 dsl_pool_t *dp = zsa->zsa_dp;
601 dsl_scan_t *scn = dp->dp_scan;
602 zil_header_t *zh = zsa->zsa_zh;
605 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
609 * One block ("stubby") can be allocated a long time ago; we
610 * want to visit that one because it has been allocated
611 * (on-disk) even if it hasn't been claimed (even though for
612 * scrub there's nothing to do to it).
614 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
617 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
618 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
620 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
626 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
628 if (lrc->lrc_txtype == TX_WRITE) {
629 zil_scan_arg_t *zsa = arg;
630 dsl_pool_t *dp = zsa->zsa_dp;
631 dsl_scan_t *scn = dp->dp_scan;
632 zil_header_t *zh = zsa->zsa_zh;
633 lr_write_t *lr = (lr_write_t *)lrc;
634 blkptr_t *bp = &lr->lr_blkptr;
637 if (BP_IS_HOLE(bp) ||
638 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
642 * birth can be < claim_txg if this record's txg is
643 * already txg sync'ed (but this log block contains
644 * other records that are not synced)
646 if (claim_txg == 0 || bp->blk_birth < claim_txg)
649 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
650 lr->lr_foid, ZB_ZIL_LEVEL,
651 lr->lr_offset / BP_GET_LSIZE(bp));
653 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
659 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
661 uint64_t claim_txg = zh->zh_claim_txg;
662 zil_scan_arg_t zsa = { dp, zh };
666 * We only want to visit blocks that have been claimed but not yet
667 * replayed (or, in read-only mode, blocks that *would* be claimed).
669 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
672 zilog = zil_alloc(dp->dp_meta_objset, zh);
674 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
682 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
683 uint64_t objset, uint64_t object, uint64_t blkid)
685 zbookmark_phys_t czb;
686 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
688 if (zfs_no_scrub_prefetch)
691 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
692 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
695 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
697 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
698 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
699 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
703 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
704 const zbookmark_phys_t *zb)
707 * We never skip over user/group accounting objects (obj<0)
709 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
710 (int64_t)zb->zb_object >= 0) {
712 * If we already visited this bp & everything below (in
713 * a prior txg sync), don't bother doing it again.
715 if (zbookmark_subtree_completed(dnp, zb,
716 &scn->scn_phys.scn_bookmark))
720 * If we found the block we're trying to resume from, or
721 * we went past it to a different object, zero it out to
722 * indicate that it's OK to start checking for suspending
725 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
726 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
727 dprintf("resuming at %llx/%llx/%llx/%llx\n",
728 (longlong_t)zb->zb_objset,
729 (longlong_t)zb->zb_object,
730 (longlong_t)zb->zb_level,
731 (longlong_t)zb->zb_blkid);
732 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
739 * Return nonzero on i/o error.
740 * Return new buf to write out in *bufp.
743 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
744 dnode_phys_t *dnp, const blkptr_t *bp,
745 const zbookmark_phys_t *zb, dmu_tx_t *tx)
747 dsl_pool_t *dp = scn->scn_dp;
748 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
751 if (BP_GET_LEVEL(bp) > 0) {
752 arc_flags_t flags = ARC_FLAG_WAIT;
755 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
758 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
759 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
761 scn->scn_phys.scn_errors++;
764 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
765 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
766 zb->zb_object, zb->zb_blkid * epb + i);
768 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
769 zbookmark_phys_t czb;
771 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
773 zb->zb_blkid * epb + i);
774 dsl_scan_visitbp(cbp, &czb, dnp,
775 ds, scn, ostype, tx);
777 arc_buf_destroy(buf, &buf);
778 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
779 arc_flags_t flags = ARC_FLAG_WAIT;
782 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
785 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
786 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
788 scn->scn_phys.scn_errors++;
791 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
792 for (j = 0; j < cdnp->dn_nblkptr; j++) {
793 blkptr_t *cbp = &cdnp->dn_blkptr[j];
794 dsl_scan_prefetch(scn, buf, cbp,
795 zb->zb_objset, zb->zb_blkid * epb + i, j);
798 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
799 dsl_scan_visitdnode(scn, ds, ostype,
800 cdnp, zb->zb_blkid * epb + i, tx);
803 arc_buf_destroy(buf, &buf);
804 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
805 arc_flags_t flags = ARC_FLAG_WAIT;
809 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
810 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
812 scn->scn_phys.scn_errors++;
818 dsl_scan_visitdnode(scn, ds, osp->os_type,
819 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
821 if (OBJSET_BUF_HAS_USERUSED(buf)) {
823 * We also always visit user/group accounting
824 * objects, and never skip them, even if we are
825 * suspending. This is necessary so that the space
826 * deltas from this txg get integrated.
828 dsl_scan_visitdnode(scn, ds, osp->os_type,
829 &osp->os_groupused_dnode,
830 DMU_GROUPUSED_OBJECT, tx);
831 dsl_scan_visitdnode(scn, ds, osp->os_type,
832 &osp->os_userused_dnode,
833 DMU_USERUSED_OBJECT, tx);
835 arc_buf_destroy(buf, &buf);
842 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
843 dmu_objset_type_t ostype, dnode_phys_t *dnp,
844 uint64_t object, dmu_tx_t *tx)
848 for (j = 0; j < dnp->dn_nblkptr; j++) {
849 zbookmark_phys_t czb;
851 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
852 dnp->dn_nlevels - 1, j);
853 dsl_scan_visitbp(&dnp->dn_blkptr[j],
854 &czb, dnp, ds, scn, ostype, tx);
857 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
858 zbookmark_phys_t czb;
859 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
861 dsl_scan_visitbp(&dnp->dn_spill,
862 &czb, dnp, ds, scn, ostype, tx);
867 * The arguments are in this order because mdb can only print the
868 * first 5; we want them to be useful.
871 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
872 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
873 dmu_objset_type_t ostype, dmu_tx_t *tx)
875 dsl_pool_t *dp = scn->scn_dp;
876 arc_buf_t *buf = NULL;
877 blkptr_t bp_toread = *bp;
879 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
881 if (dsl_scan_check_suspend(scn, zb))
884 if (dsl_scan_check_resume(scn, dnp, zb))
890 scn->scn_visited_this_txg++;
893 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
894 ds, ds ? ds->ds_object : 0,
895 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
898 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
901 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
905 * If dsl_scan_ddt() has already visited this block, it will have
906 * already done any translations or scrubbing, so don't call the
909 if (ddt_class_contains(dp->dp_spa,
910 scn->scn_phys.scn_ddt_class_max, bp)) {
916 * If this block is from the future (after cur_max_txg), then we
917 * are doing this on behalf of a deleted snapshot, and we will
918 * revisit the future block on the next pass of this dataset.
919 * Don't scan it now unless we need to because something
920 * under it was modified.
922 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
923 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
928 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
933 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
934 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
935 dsl_scan_visitbp(bp, &zb, NULL,
936 ds, scn, DMU_OST_NONE, tx);
938 dprintf_ds(ds, "finished scan%s", "");
942 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
944 dsl_pool_t *dp = ds->ds_dir->dd_pool;
945 dsl_scan_t *scn = dp->dp_scan;
948 if (scn->scn_phys.scn_state != DSS_SCANNING)
951 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
952 if (ds->ds_is_snapshot) {
955 * - scn_cur_{min,max}_txg stays the same.
956 * - Setting the flag is not really necessary if
957 * scn_cur_max_txg == scn_max_txg, because there
958 * is nothing after this snapshot that we care
959 * about. However, we set it anyway and then
960 * ignore it when we retraverse it in
961 * dsl_scan_visitds().
963 scn->scn_phys.scn_bookmark.zb_objset =
964 dsl_dataset_phys(ds)->ds_next_snap_obj;
965 zfs_dbgmsg("destroying ds %llu; currently traversing; "
966 "reset zb_objset to %llu",
967 (u_longlong_t)ds->ds_object,
968 (u_longlong_t)dsl_dataset_phys(ds)->
970 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
972 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
973 ZB_DESTROYED_OBJSET, 0, 0, 0);
974 zfs_dbgmsg("destroying ds %llu; currently traversing; "
975 "reset bookmark to -1,0,0,0",
976 (u_longlong_t)ds->ds_object);
978 } else if (zap_lookup_int_key(dp->dp_meta_objset,
979 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
980 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
981 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
982 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
983 if (ds->ds_is_snapshot) {
985 * We keep the same mintxg; it could be >
986 * ds_creation_txg if the previous snapshot was
989 VERIFY(zap_add_int_key(dp->dp_meta_objset,
990 scn->scn_phys.scn_queue_obj,
991 dsl_dataset_phys(ds)->ds_next_snap_obj,
993 zfs_dbgmsg("destroying ds %llu; in queue; "
994 "replacing with %llu",
995 (u_longlong_t)ds->ds_object,
996 (u_longlong_t)dsl_dataset_phys(ds)->
999 zfs_dbgmsg("destroying ds %llu; in queue; removing",
1000 (u_longlong_t)ds->ds_object);
1005 * dsl_scan_sync() should be called after this, and should sync
1006 * out our changed state, but just to be safe, do it here.
1008 dsl_scan_sync_state(scn, tx);
1012 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
1014 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1015 dsl_scan_t *scn = dp->dp_scan;
1018 if (scn->scn_phys.scn_state != DSS_SCANNING)
1021 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
1023 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
1024 scn->scn_phys.scn_bookmark.zb_objset =
1025 dsl_dataset_phys(ds)->ds_prev_snap_obj;
1026 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
1027 "reset zb_objset to %llu",
1028 (u_longlong_t)ds->ds_object,
1029 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
1030 } else if (zap_lookup_int_key(dp->dp_meta_objset,
1031 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
1032 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1033 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
1034 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1035 scn->scn_phys.scn_queue_obj,
1036 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
1037 zfs_dbgmsg("snapshotting ds %llu; in queue; "
1038 "replacing with %llu",
1039 (u_longlong_t)ds->ds_object,
1040 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
1042 dsl_scan_sync_state(scn, tx);
1046 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
1048 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
1049 dsl_scan_t *scn = dp->dp_scan;
1052 if (scn->scn_phys.scn_state != DSS_SCANNING)
1055 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
1056 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
1057 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
1058 "reset zb_objset to %llu",
1059 (u_longlong_t)ds1->ds_object,
1060 (u_longlong_t)ds2->ds_object);
1061 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
1062 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
1063 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
1064 "reset zb_objset to %llu",
1065 (u_longlong_t)ds2->ds_object,
1066 (u_longlong_t)ds1->ds_object);
1069 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1070 ds1->ds_object, &mintxg) == 0) {
1073 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
1074 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1075 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1076 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
1077 err = zap_add_int_key(dp->dp_meta_objset,
1078 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
1079 VERIFY(err == 0 || err == EEXIST);
1080 if (err == EEXIST) {
1081 /* Both were there to begin with */
1082 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1083 scn->scn_phys.scn_queue_obj,
1084 ds1->ds_object, mintxg, tx));
1086 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1087 "replacing with %llu",
1088 (u_longlong_t)ds1->ds_object,
1089 (u_longlong_t)ds2->ds_object);
1090 } else if (zap_lookup_int_key(dp->dp_meta_objset,
1091 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1092 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
1093 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1094 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1095 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1096 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1097 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1098 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1099 "replacing with %llu",
1100 (u_longlong_t)ds2->ds_object,
1101 (u_longlong_t)ds1->ds_object);
1104 dsl_scan_sync_state(scn, tx);
1107 struct enqueue_clones_arg {
1114 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1116 struct enqueue_clones_arg *eca = arg;
1119 dsl_scan_t *scn = dp->dp_scan;
1121 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1124 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1128 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1129 dsl_dataset_t *prev;
1130 err = dsl_dataset_hold_obj(dp,
1131 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1133 dsl_dataset_rele(ds, FTAG);
1138 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1139 scn->scn_phys.scn_queue_obj, ds->ds_object,
1140 dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1141 dsl_dataset_rele(ds, FTAG);
1146 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1148 dsl_pool_t *dp = scn->scn_dp;
1151 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1153 if (scn->scn_phys.scn_cur_min_txg >=
1154 scn->scn_phys.scn_max_txg) {
1156 * This can happen if this snapshot was created after the
1157 * scan started, and we already completed a previous snapshot
1158 * that was created after the scan started. This snapshot
1159 * only references blocks with:
1161 * birth < our ds_creation_txg
1162 * cur_min_txg is no less than ds_creation_txg.
1163 * We have already visited these blocks.
1165 * birth > scn_max_txg
1166 * The scan requested not to visit these blocks.
1168 * Subsequent snapshots (and clones) can reference our
1169 * blocks, or blocks with even higher birth times.
1170 * Therefore we do not need to visit them either,
1171 * so we do not add them to the work queue.
1173 * Note that checking for cur_min_txg >= cur_max_txg
1174 * is not sufficient, because in that case we may need to
1175 * visit subsequent snapshots. This happens when min_txg > 0,
1176 * which raises cur_min_txg. In this case we will visit
1177 * this dataset but skip all of its blocks, because the
1178 * rootbp's birth time is < cur_min_txg. Then we will
1179 * add the next snapshots/clones to the work queue.
1181 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1182 dsl_dataset_name(ds, dsname);
1183 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1184 "cur_min_txg (%llu) >= max_txg (%llu)",
1186 scn->scn_phys.scn_cur_min_txg,
1187 scn->scn_phys.scn_max_txg);
1188 kmem_free(dsname, MAXNAMELEN);
1194 * Only the ZIL in the head (non-snapshot) is valid. Even though
1195 * snapshots can have ZIL block pointers (which may be the same
1196 * BP as in the head), they must be ignored. In addition, $ORIGIN
1197 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
1198 * need to look for a ZIL in it either. So we traverse the ZIL here,
1199 * rather than in scan_recurse(), because the regular snapshot
1200 * block-sharing rules don't apply to it.
1202 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds) &&
1203 ds->ds_dir != dp->dp_origin_snap->ds_dir) {
1205 if (dmu_objset_from_ds(ds, &os) != 0) {
1208 dsl_scan_zil(dp, &os->os_zil_header);
1212 * Iterate over the bps in this ds.
1214 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1215 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1216 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1217 rrw_exit(&ds->ds_bp_rwlock, FTAG);
1219 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
1220 dsl_dataset_name(ds, dsname);
1221 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1223 (longlong_t)dsobj, dsname,
1224 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1225 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1226 (int)scn->scn_suspending);
1227 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
1229 if (scn->scn_suspending)
1233 * We've finished this pass over this dataset.
1237 * If we did not completely visit this dataset, do another pass.
1239 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1240 zfs_dbgmsg("incomplete pass; visiting again");
1241 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1242 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1243 scn->scn_phys.scn_queue_obj, ds->ds_object,
1244 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1249 * Add descendent datasets to work queue.
1251 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1252 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1253 scn->scn_phys.scn_queue_obj,
1254 dsl_dataset_phys(ds)->ds_next_snap_obj,
1255 dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1257 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1258 boolean_t usenext = B_FALSE;
1259 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1262 * A bug in a previous version of the code could
1263 * cause upgrade_clones_cb() to not set
1264 * ds_next_snap_obj when it should, leading to a
1265 * missing entry. Therefore we can only use the
1266 * next_clones_obj when its count is correct.
1268 int err = zap_count(dp->dp_meta_objset,
1269 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1271 count == dsl_dataset_phys(ds)->ds_num_children - 1)
1276 VERIFY0(zap_join_key(dp->dp_meta_objset,
1277 dsl_dataset_phys(ds)->ds_next_clones_obj,
1278 scn->scn_phys.scn_queue_obj,
1279 dsl_dataset_phys(ds)->ds_creation_txg, tx));
1281 struct enqueue_clones_arg eca;
1283 eca.originobj = ds->ds_object;
1285 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1286 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1291 dsl_dataset_rele(ds, FTAG);
1296 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1301 dsl_scan_t *scn = dp->dp_scan;
1303 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1307 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1308 dsl_dataset_t *prev;
1309 err = dsl_dataset_hold_obj(dp,
1310 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1312 dsl_dataset_rele(ds, FTAG);
1317 * If this is a clone, we don't need to worry about it for now.
1319 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1320 dsl_dataset_rele(ds, FTAG);
1321 dsl_dataset_rele(prev, FTAG);
1324 dsl_dataset_rele(ds, FTAG);
1328 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1329 ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1330 dsl_dataset_rele(ds, FTAG);
1335 * Scrub/dedup interaction.
1337 * If there are N references to a deduped block, we don't want to scrub it
1338 * N times -- ideally, we should scrub it exactly once.
1340 * We leverage the fact that the dde's replication class (enum ddt_class)
1341 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1342 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1344 * To prevent excess scrubbing, the scrub begins by walking the DDT
1345 * to find all blocks with refcnt > 1, and scrubs each of these once.
1346 * Since there are two replication classes which contain blocks with
1347 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1348 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1350 * There would be nothing more to say if a block's refcnt couldn't change
1351 * during a scrub, but of course it can so we must account for changes
1352 * in a block's replication class.
1354 * Here's an example of what can occur:
1356 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1357 * when visited during the top-down scrub phase, it will be scrubbed twice.
1358 * This negates our scrub optimization, but is otherwise harmless.
1360 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1361 * on each visit during the top-down scrub phase, it will never be scrubbed.
1362 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1363 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1364 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1365 * while a scrub is in progress, it scrubs the block right then.
1368 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1370 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1371 ddt_entry_t dde = { 0 };
1375 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1378 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1380 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1381 (longlong_t)ddb->ddb_class,
1382 (longlong_t)ddb->ddb_type,
1383 (longlong_t)ddb->ddb_checksum,
1384 (longlong_t)ddb->ddb_cursor);
1386 /* There should be no pending changes to the dedup table */
1387 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1388 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1390 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1393 if (dsl_scan_check_suspend(scn, NULL))
1397 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
1398 "suspending=%u", (longlong_t)n,
1399 (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
1401 ASSERT(error == 0 || error == ENOENT);
1402 ASSERT(error != ENOENT ||
1403 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1408 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1409 ddt_entry_t *dde, dmu_tx_t *tx)
1411 const ddt_key_t *ddk = &dde->dde_key;
1412 ddt_phys_t *ddp = dde->dde_phys;
1414 zbookmark_phys_t zb = { 0 };
1416 if (scn->scn_phys.scn_state != DSS_SCANNING)
1419 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1420 if (ddp->ddp_phys_birth == 0 ||
1421 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1423 ddt_bp_create(checksum, ddk, ddp, &bp);
1425 scn->scn_visited_this_txg++;
1426 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1431 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1433 dsl_pool_t *dp = scn->scn_dp;
1437 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1438 scn->scn_phys.scn_ddt_class_max) {
1439 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1440 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1441 dsl_scan_ddt(scn, tx);
1442 if (scn->scn_suspending)
1446 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1447 /* First do the MOS & ORIGIN */
1449 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1450 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1451 dsl_scan_visit_rootbp(scn, NULL,
1452 &dp->dp_meta_rootbp, tx);
1453 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1454 if (scn->scn_suspending)
1457 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1458 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1459 enqueue_cb, tx, DS_FIND_CHILDREN));
1461 dsl_scan_visitds(scn,
1462 dp->dp_origin_snap->ds_object, tx);
1464 ASSERT(!scn->scn_suspending);
1465 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1466 ZB_DESTROYED_OBJSET) {
1468 * If we were suspended, continue from here. Note if the
1469 * ds we were suspended on was deleted, the zb_objset may
1470 * be -1, so we will skip this and find a new objset
1473 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1474 if (scn->scn_suspending)
1479 * In case we were suspended right at the end of the ds, zero the
1480 * bookmark so we don't think that we're still trying to resume.
1482 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1484 /* keep pulling things out of the zap-object-as-queue */
1485 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1486 scn->scn_phys.scn_queue_obj),
1487 zap_cursor_retrieve(&zc, &za) == 0) {
1491 dsobj = zfs_strtonum(za.za_name, NULL);
1492 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1493 scn->scn_phys.scn_queue_obj, dsobj, tx));
1495 /* Set up min/max txg */
1496 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1497 if (za.za_first_integer != 0) {
1498 scn->scn_phys.scn_cur_min_txg =
1499 MAX(scn->scn_phys.scn_min_txg,
1500 za.za_first_integer);
1502 scn->scn_phys.scn_cur_min_txg =
1503 MAX(scn->scn_phys.scn_min_txg,
1504 dsl_dataset_phys(ds)->ds_prev_snap_txg);
1506 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1507 dsl_dataset_rele(ds, FTAG);
1509 dsl_scan_visitds(scn, dsobj, tx);
1510 zap_cursor_fini(&zc);
1511 if (scn->scn_suspending)
1514 zap_cursor_fini(&zc);
1518 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
1520 uint64_t elapsed_nanosecs;
1525 if (scn->scn_visited_this_txg >= zfs_async_block_max_blocks)
1528 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1529 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1530 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
1531 txg_sync_waiting(scn->scn_dp)) ||
1532 spa_shutting_down(scn->scn_dp->dp_spa));
1536 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1538 dsl_scan_t *scn = arg;
1540 if (!scn->scn_is_bptree ||
1541 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1542 if (dsl_scan_async_block_should_pause(scn))
1543 return (SET_ERROR(ERESTART));
1546 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1547 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1548 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1549 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1550 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1551 scn->scn_visited_this_txg++;
1556 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1558 dsl_scan_t *scn = arg;
1559 const dva_t *dva = &bp->blk_dva[0];
1561 if (dsl_scan_async_block_should_pause(scn))
1562 return (SET_ERROR(ERESTART));
1564 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
1565 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
1566 DVA_GET_ASIZE(dva), tx);
1567 scn->scn_visited_this_txg++;
1572 dsl_scan_active(dsl_scan_t *scn)
1574 spa_t *spa = scn->scn_dp->dp_spa;
1575 uint64_t used = 0, comp, uncomp;
1577 if (spa->spa_load_state != SPA_LOAD_NONE)
1579 if (spa_shutting_down(spa))
1581 if ((scn->scn_phys.scn_state == DSS_SCANNING &&
1582 !dsl_scan_is_paused_scrub(scn)) ||
1583 (scn->scn_async_destroying && !scn->scn_async_stalled))
1586 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1587 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1588 &used, &comp, &uncomp);
1593 /* Called whenever a txg syncs. */
1595 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1597 dsl_scan_t *scn = dp->dp_scan;
1598 spa_t *spa = dp->dp_spa;
1602 * Check for scn_restart_txg before checking spa_load_state, so
1603 * that we can restart an old-style scan while the pool is being
1604 * imported (see dsl_scan_init).
1606 if (dsl_scan_restarting(scn, tx)) {
1607 pool_scan_func_t func = POOL_SCAN_SCRUB;
1608 dsl_scan_done(scn, B_FALSE, tx);
1609 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1610 func = POOL_SCAN_RESILVER;
1611 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1613 dsl_scan_setup_sync(&func, tx);
1617 * Only process scans in sync pass 1.
1619 if (spa_sync_pass(dp->dp_spa) > 1)
1623 * If the spa is shutting down, then stop scanning. This will
1624 * ensure that the scan does not dirty any new data during the
1627 if (spa_shutting_down(spa))
1631 * If the scan is inactive due to a stalled async destroy, try again.
1633 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1636 scn->scn_visited_this_txg = 0;
1637 scn->scn_suspending = B_FALSE;
1638 scn->scn_sync_start_time = gethrtime();
1639 spa->spa_scrub_active = B_TRUE;
1642 * First process the async destroys. If we suspend, don't do
1643 * any scrubbing or resilvering. This ensures that there are no
1644 * async destroys while we are scanning, so the scan code doesn't
1645 * have to worry about traversing it. It is also faster to free the
1646 * blocks than to scrub them.
1648 if (zfs_free_bpobj_enabled &&
1649 spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1650 scn->scn_is_bptree = B_FALSE;
1651 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
1652 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1653 NULL, ZIO_FLAG_MUSTSUCCEED);
1654 err = bpobj_iterate(&dp->dp_free_bpobj,
1655 dsl_scan_free_block_cb, scn, tx);
1656 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1658 if (err != 0 && err != ERESTART)
1659 zfs_panic_recover("error %u from bpobj_iterate()", err);
1662 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1663 ASSERT(scn->scn_async_destroying);
1664 scn->scn_is_bptree = B_TRUE;
1665 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1666 NULL, ZIO_FLAG_MUSTSUCCEED);
1667 err = bptree_iterate(dp->dp_meta_objset,
1668 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1669 VERIFY0(zio_wait(scn->scn_zio_root));
1671 if (err == EIO || err == ECKSUM) {
1673 } else if (err != 0 && err != ERESTART) {
1674 zfs_panic_recover("error %u from "
1675 "traverse_dataset_destroyed()", err);
1678 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1679 /* finished; deactivate async destroy feature */
1680 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1681 ASSERT(!spa_feature_is_active(spa,
1682 SPA_FEATURE_ASYNC_DESTROY));
1683 VERIFY0(zap_remove(dp->dp_meta_objset,
1684 DMU_POOL_DIRECTORY_OBJECT,
1685 DMU_POOL_BPTREE_OBJ, tx));
1686 VERIFY0(bptree_free(dp->dp_meta_objset,
1687 dp->dp_bptree_obj, tx));
1688 dp->dp_bptree_obj = 0;
1689 scn->scn_async_destroying = B_FALSE;
1690 scn->scn_async_stalled = B_FALSE;
1693 * If we didn't make progress, mark the async
1694 * destroy as stalled, so that we will not initiate
1695 * a spa_sync() on its behalf. Note that we only
1696 * check this if we are not finished, because if the
1697 * bptree had no blocks for us to visit, we can
1698 * finish without "making progress".
1700 scn->scn_async_stalled =
1701 (scn->scn_visited_this_txg == 0);
1704 if (scn->scn_visited_this_txg) {
1705 zfs_dbgmsg("freed %llu blocks in %llums from "
1706 "free_bpobj/bptree txg %llu; err=%d",
1707 (longlong_t)scn->scn_visited_this_txg,
1709 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1710 (longlong_t)tx->tx_txg, err);
1711 scn->scn_visited_this_txg = 0;
1714 * Write out changes to the DDT that may be required as a
1715 * result of the blocks freed. This ensures that the DDT
1716 * is clean when a scrub/resilver runs.
1718 ddt_sync(spa, tx->tx_txg);
1722 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1723 zfs_free_leak_on_eio &&
1724 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1725 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1726 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1728 * We have finished background destroying, but there is still
1729 * some space left in the dp_free_dir. Transfer this leaked
1730 * space to the dp_leak_dir.
1732 if (dp->dp_leak_dir == NULL) {
1733 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1734 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1736 VERIFY0(dsl_pool_open_special_dir(dp,
1737 LEAK_DIR_NAME, &dp->dp_leak_dir));
1738 rrw_exit(&dp->dp_config_rwlock, FTAG);
1740 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1741 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1742 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1743 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1744 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1745 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1746 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1747 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1750 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1751 /* finished; verify that space accounting went to zero */
1752 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1753 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1754 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1757 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
1758 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1759 DMU_POOL_OBSOLETE_BPOBJ));
1760 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
1761 ASSERT(spa_feature_is_active(dp->dp_spa,
1762 SPA_FEATURE_OBSOLETE_COUNTS));
1764 scn->scn_is_bptree = B_FALSE;
1765 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
1766 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
1767 dsl_scan_obsolete_block_cb, scn, tx);
1768 if (err != 0 && err != ERESTART)
1769 zfs_panic_recover("error %u from bpobj_iterate()", err);
1771 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
1772 dsl_pool_destroy_obsolete_bpobj(dp, tx);
1775 if (scn->scn_phys.scn_state != DSS_SCANNING)
1778 if (scn->scn_done_txg == tx->tx_txg) {
1779 ASSERT(!scn->scn_suspending);
1780 /* finished with scan. */
1781 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1782 dsl_scan_done(scn, B_TRUE, tx);
1783 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1784 dsl_scan_sync_state(scn, tx);
1788 if (dsl_scan_is_paused_scrub(scn))
1791 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1792 scn->scn_phys.scn_ddt_class_max) {
1793 zfs_dbgmsg("doing scan sync txg %llu; "
1794 "ddt bm=%llu/%llu/%llu/%llx",
1795 (longlong_t)tx->tx_txg,
1796 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1797 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1798 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1799 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1800 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1801 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1802 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1803 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1805 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1806 (longlong_t)tx->tx_txg,
1807 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1808 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1809 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1810 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1813 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1814 NULL, ZIO_FLAG_CANFAIL);
1815 dsl_pool_config_enter(dp, FTAG);
1816 dsl_scan_visit(scn, tx);
1817 dsl_pool_config_exit(dp, FTAG);
1818 (void) zio_wait(scn->scn_zio_root);
1819 scn->scn_zio_root = NULL;
1821 zfs_dbgmsg("visited %llu blocks in %llums",
1822 (longlong_t)scn->scn_visited_this_txg,
1823 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1825 if (!scn->scn_suspending) {
1826 scn->scn_done_txg = tx->tx_txg + 1;
1827 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1828 tx->tx_txg, scn->scn_done_txg);
1831 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1832 mutex_enter(&spa->spa_scrub_lock);
1833 while (spa->spa_scrub_inflight > 0) {
1834 cv_wait(&spa->spa_scrub_io_cv,
1835 &spa->spa_scrub_lock);
1837 mutex_exit(&spa->spa_scrub_lock);
1840 dsl_scan_sync_state(scn, tx);
1844 * This will start a new scan, or restart an existing one.
1847 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1851 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1852 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1854 txg = dmu_tx_get_txg(tx);
1855 dp->dp_scan->scn_restart_txg = txg;
1858 dp->dp_scan->scn_restart_txg = txg;
1860 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1864 dsl_scan_resilvering(dsl_pool_t *dp)
1866 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1867 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1875 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1880 * If we resume after a reboot, zab will be NULL; don't record
1881 * incomplete stats in that case.
1886 for (i = 0; i < 4; i++) {
1887 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1888 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1889 if (t & DMU_OT_NEWTYPE)
1891 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1895 zb->zb_asize += BP_GET_ASIZE(bp);
1896 zb->zb_lsize += BP_GET_LSIZE(bp);
1897 zb->zb_psize += BP_GET_PSIZE(bp);
1898 zb->zb_gangs += BP_COUNT_GANG(bp);
1900 switch (BP_GET_NDVAS(bp)) {
1902 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1903 DVA_GET_VDEV(&bp->blk_dva[1]))
1904 zb->zb_ditto_2_of_2_samevdev++;
1907 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1908 DVA_GET_VDEV(&bp->blk_dva[1])) +
1909 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1910 DVA_GET_VDEV(&bp->blk_dva[2])) +
1911 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1912 DVA_GET_VDEV(&bp->blk_dva[2]));
1914 zb->zb_ditto_2_of_3_samevdev++;
1915 else if (equal == 3)
1916 zb->zb_ditto_3_of_3_samevdev++;
1923 dsl_scan_scrub_done(zio_t *zio)
1925 spa_t *spa = zio->io_spa;
1927 abd_free(zio->io_abd);
1929 mutex_enter(&spa->spa_scrub_lock);
1930 spa->spa_scrub_inflight--;
1931 cv_broadcast(&spa->spa_scrub_io_cv);
1933 if (zio->io_error && (zio->io_error != ECKSUM ||
1934 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1935 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1937 mutex_exit(&spa->spa_scrub_lock);
1941 dsl_scan_scrub_cb(dsl_pool_t *dp,
1942 const blkptr_t *bp, const zbookmark_phys_t *zb)
1944 dsl_scan_t *scn = dp->dp_scan;
1945 size_t size = BP_GET_PSIZE(bp);
1946 spa_t *spa = dp->dp_spa;
1947 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1949 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1950 unsigned int scan_delay = 0;
1952 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1953 phys_birth >= scn->scn_phys.scn_max_txg)
1956 count_block(dp->dp_blkstats, bp);
1958 if (BP_IS_EMBEDDED(bp))
1961 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1962 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1963 zio_flags |= ZIO_FLAG_SCRUB;
1965 scan_delay = zfs_scrub_delay;
1967 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1968 zio_flags |= ZIO_FLAG_RESILVER;
1970 scan_delay = zfs_resilver_delay;
1973 /* If it's an intent log block, failure is expected. */
1974 if (zb->zb_level == ZB_ZIL_LEVEL)
1975 zio_flags |= ZIO_FLAG_SPECULATIVE;
1977 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1978 vdev_t *vd = vdev_lookup_top(spa,
1979 DVA_GET_VDEV(&bp->blk_dva[d]));
1982 * Keep track of how much data we've examined so that
1983 * zpool(1M) status can make useful progress reports.
1985 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1986 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1988 /* if it's a resilver, this may not be in the target range */
1990 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1992 * Gang members may be spread across multiple
1993 * vdevs, so the best estimate we have is the
1994 * scrub range, which has already been checked.
1995 * XXX -- it would be better to change our
1996 * allocation policy to ensure that all
1997 * gang members reside on the same vdev.
2001 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
2007 if (needs_io && !zfs_no_scrub_io) {
2008 vdev_t *rvd = spa->spa_root_vdev;
2009 uint64_t maxinflight = rvd->vdev_children *
2010 MAX(zfs_top_maxinflight, 1);
2012 mutex_enter(&spa->spa_scrub_lock);
2013 while (spa->spa_scrub_inflight >= maxinflight)
2014 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2015 spa->spa_scrub_inflight++;
2016 mutex_exit(&spa->spa_scrub_lock);
2019 * If we're seeing recent (zfs_scan_idle) "important" I/Os
2020 * then throttle our workload to limit the impact of a scan.
2022 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
2023 delay(MAX((int)scan_delay, 0));
2025 zio_nowait(zio_read(NULL, spa, bp,
2026 abd_alloc_for_io(size, B_FALSE), size, dsl_scan_scrub_done,
2027 NULL, ZIO_PRIORITY_SCRUB, zio_flags, zb));
2030 /* do not relocate this block */
2035 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
2036 * Can also be called to resume a paused scrub.
2039 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
2041 spa_t *spa = dp->dp_spa;
2042 dsl_scan_t *scn = dp->dp_scan;
2045 * Purge all vdev caches and probe all devices. We do this here
2046 * rather than in sync context because this requires a writer lock
2047 * on the spa_config lock, which we can't do from sync context. The
2048 * spa_scrub_reopen flag indicates that vdev_open() should not
2049 * attempt to start another scrub.
2051 spa_vdev_state_enter(spa, SCL_NONE);
2052 spa->spa_scrub_reopen = B_TRUE;
2053 vdev_reopen(spa->spa_root_vdev);
2054 spa->spa_scrub_reopen = B_FALSE;
2055 (void) spa_vdev_state_exit(spa, NULL, 0);
2057 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
2058 /* got scrub start cmd, resume paused scrub */
2059 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
2062 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
2066 return (SET_ERROR(err));
2069 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
2070 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
2074 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
2076 return (scn->scn_restart_txg != 0 &&
2077 scn->scn_restart_txg <= tx->tx_txg);