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 (c) 2011, 2014 by Delphix. All rights reserved.
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>
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>
47 #include <sys/sa_impl.h>
48 #include <sys/zfeature.h>
50 #include <sys/zfs_vfsops.h>
53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
54 const zbookmark_phys_t *);
56 static scan_cb_t dsl_scan_scrub_cb;
57 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
58 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
60 unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
61 unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
62 unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
63 unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */
65 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
67 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
69 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
70 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
72 SYSCTL_DECL(_vfs_zfs);
73 TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
74 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
75 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
76 TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
77 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
78 &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
79 TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
80 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
81 &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
82 TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
84 &zfs_scan_idle, 0, "Idle scan window in clock ticks");
85 TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
86 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
87 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
88 TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
89 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
90 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
91 TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
92 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
93 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
94 TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
95 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
96 &zfs_no_scrub_io, 0, "Disable scrub I/O");
97 TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
98 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
99 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
101 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
102 /* max number of blocks to free in a single TXG */
103 uint64_t zfs_free_max_blocks = UINT64_MAX;
104 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
105 &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
108 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
109 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
110 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
112 extern int zfs_txg_timeout;
114 /* the order has to match pool_scan_type */
115 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
117 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
118 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
122 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
126 spa_t *spa = dp->dp_spa;
129 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
133 * It's possible that we're resuming a scan after a reboot so
134 * make sure that the scan_async_destroying flag is initialized
137 ASSERT(!scn->scn_async_destroying);
138 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
139 SPA_FEATURE_ASYNC_DESTROY);
141 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
142 "scrub_func", sizeof (uint64_t), 1, &f);
145 * There was an old-style scrub in progress. Restart a
146 * new-style scrub from the beginning.
148 scn->scn_restart_txg = txg;
149 zfs_dbgmsg("old-style scrub was in progress; "
150 "restarting new-style scrub in txg %llu",
151 scn->scn_restart_txg);
154 * Load the queue obj from the old location so that it
155 * can be freed by dsl_scan_done().
157 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
158 "scrub_queue", sizeof (uint64_t), 1,
159 &scn->scn_phys.scn_queue_obj);
161 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
162 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
169 if (scn->scn_phys.scn_state == DSS_SCANNING &&
170 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
172 * A new-type scrub was in progress on an old
173 * pool, and the pool was accessed by old
174 * software. Restart from the beginning, since
175 * the old software may have changed the pool in
178 scn->scn_restart_txg = txg;
179 zfs_dbgmsg("new-style scrub was modified "
180 "by old software; restarting in txg %llu",
181 scn->scn_restart_txg);
185 spa_scan_stat_init(spa);
190 dsl_scan_fini(dsl_pool_t *dp)
193 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
200 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
202 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
204 if (scn->scn_phys.scn_state == DSS_SCANNING)
205 return (SET_ERROR(EBUSY));
211 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
213 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
214 pool_scan_func_t *funcp = arg;
215 dmu_object_type_t ot = 0;
216 dsl_pool_t *dp = scn->scn_dp;
217 spa_t *spa = dp->dp_spa;
219 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
220 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
221 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
222 scn->scn_phys.scn_func = *funcp;
223 scn->scn_phys.scn_state = DSS_SCANNING;
224 scn->scn_phys.scn_min_txg = 0;
225 scn->scn_phys.scn_max_txg = tx->tx_txg;
226 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
227 scn->scn_phys.scn_start_time = gethrestime_sec();
228 scn->scn_phys.scn_errors = 0;
229 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
230 scn->scn_restart_txg = 0;
231 scn->scn_done_txg = 0;
232 spa_scan_stat_init(spa);
234 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
235 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
237 /* rewrite all disk labels */
238 vdev_config_dirty(spa->spa_root_vdev);
240 if (vdev_resilver_needed(spa->spa_root_vdev,
241 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
242 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
244 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
247 spa->spa_scrub_started = B_TRUE;
249 * If this is an incremental scrub, limit the DDT scrub phase
250 * to just the auto-ditto class (for correctness); the rest
251 * of the scrub should go faster using top-down pruning.
253 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
254 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
258 /* back to the generic stuff */
260 if (dp->dp_blkstats == NULL) {
262 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
264 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
266 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
267 ot = DMU_OT_ZAP_OTHER;
269 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
270 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
272 dsl_scan_sync_state(scn, tx);
274 spa_history_log_internal(spa, "scan setup", tx,
275 "func=%u mintxg=%llu maxtxg=%llu",
276 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
281 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
283 static const char *old_names[] = {
285 "scrub_ddt_bookmark",
286 "scrub_ddt_class_max",
295 dsl_pool_t *dp = scn->scn_dp;
296 spa_t *spa = dp->dp_spa;
299 /* Remove any remnants of an old-style scrub. */
300 for (i = 0; old_names[i]; i++) {
301 (void) zap_remove(dp->dp_meta_objset,
302 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
305 if (scn->scn_phys.scn_queue_obj != 0) {
306 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
307 scn->scn_phys.scn_queue_obj, tx));
308 scn->scn_phys.scn_queue_obj = 0;
312 * If we were "restarted" from a stopped state, don't bother
313 * with anything else.
315 if (scn->scn_phys.scn_state != DSS_SCANNING)
319 scn->scn_phys.scn_state = DSS_FINISHED;
321 scn->scn_phys.scn_state = DSS_CANCELED;
323 spa_history_log_internal(spa, "scan done", tx,
324 "complete=%u", complete);
326 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
327 mutex_enter(&spa->spa_scrub_lock);
328 while (spa->spa_scrub_inflight > 0) {
329 cv_wait(&spa->spa_scrub_io_cv,
330 &spa->spa_scrub_lock);
332 mutex_exit(&spa->spa_scrub_lock);
333 spa->spa_scrub_started = B_FALSE;
334 spa->spa_scrub_active = B_FALSE;
337 * If the scrub/resilver completed, update all DTLs to
338 * reflect this. Whether it succeeded or not, vacate
339 * all temporary scrub DTLs.
341 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
342 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
344 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
345 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
347 spa_errlog_rotate(spa);
350 * We may have finished replacing a device.
351 * Let the async thread assess this and handle the detach.
353 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
356 scn->scn_phys.scn_end_time = gethrestime_sec();
361 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
363 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
365 if (scn->scn_phys.scn_state != DSS_SCANNING)
366 return (SET_ERROR(ENOENT));
372 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
374 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
376 dsl_scan_done(scn, B_FALSE, tx);
377 dsl_scan_sync_state(scn, tx);
381 dsl_scan_cancel(dsl_pool_t *dp)
383 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
384 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
387 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
388 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
389 dmu_objset_type_t ostype, dmu_tx_t *tx);
390 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
391 dmu_objset_type_t ostype,
392 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
395 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
397 zio_free(dp->dp_spa, txg, bp);
401 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
403 ASSERT(dsl_pool_sync_context(dp));
404 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
409 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
411 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
412 if (dsl_dataset_is_snapshot(ds))
413 return (MIN(smt, ds->ds_phys->ds_creation_txg));
418 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
420 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
421 DMU_POOL_DIRECTORY_OBJECT,
422 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
423 &scn->scn_phys, tx));
427 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
429 uint64_t elapsed_nanosecs;
430 unsigned int mintime;
432 /* we never skip user/group accounting objects */
433 if (zb && (int64_t)zb->zb_object < 0)
436 if (scn->scn_pausing)
437 return (B_TRUE); /* we're already pausing */
439 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
440 return (B_FALSE); /* we're resuming */
442 /* We only know how to resume from level-0 blocks. */
443 if (zb && zb->zb_level != 0)
446 mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
447 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
448 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
449 if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
450 (NSEC2MSEC(elapsed_nanosecs) > mintime &&
451 txg_sync_waiting(scn->scn_dp)) ||
452 spa_shutting_down(scn->scn_dp->dp_spa)) {
454 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
455 (longlong_t)zb->zb_objset,
456 (longlong_t)zb->zb_object,
457 (longlong_t)zb->zb_level,
458 (longlong_t)zb->zb_blkid);
459 scn->scn_phys.scn_bookmark = *zb;
461 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
462 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
463 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
464 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
465 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
466 scn->scn_pausing = B_TRUE;
472 typedef struct zil_scan_arg {
474 zil_header_t *zsa_zh;
479 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
481 zil_scan_arg_t *zsa = arg;
482 dsl_pool_t *dp = zsa->zsa_dp;
483 dsl_scan_t *scn = dp->dp_scan;
484 zil_header_t *zh = zsa->zsa_zh;
487 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
491 * One block ("stubby") can be allocated a long time ago; we
492 * want to visit that one because it has been allocated
493 * (on-disk) even if it hasn't been claimed (even though for
494 * scrub there's nothing to do to it).
496 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
499 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
500 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
502 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
508 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
510 if (lrc->lrc_txtype == TX_WRITE) {
511 zil_scan_arg_t *zsa = arg;
512 dsl_pool_t *dp = zsa->zsa_dp;
513 dsl_scan_t *scn = dp->dp_scan;
514 zil_header_t *zh = zsa->zsa_zh;
515 lr_write_t *lr = (lr_write_t *)lrc;
516 blkptr_t *bp = &lr->lr_blkptr;
519 if (BP_IS_HOLE(bp) ||
520 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
524 * birth can be < claim_txg if this record's txg is
525 * already txg sync'ed (but this log block contains
526 * other records that are not synced)
528 if (claim_txg == 0 || bp->blk_birth < claim_txg)
531 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
532 lr->lr_foid, ZB_ZIL_LEVEL,
533 lr->lr_offset / BP_GET_LSIZE(bp));
535 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
541 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
543 uint64_t claim_txg = zh->zh_claim_txg;
544 zil_scan_arg_t zsa = { dp, zh };
548 * We only want to visit blocks that have been claimed but not yet
549 * replayed (or, in read-only mode, blocks that *would* be claimed).
551 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
554 zilog = zil_alloc(dp->dp_meta_objset, zh);
556 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
564 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
565 uint64_t objset, uint64_t object, uint64_t blkid)
567 zbookmark_phys_t czb;
568 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
570 if (zfs_no_scrub_prefetch)
573 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
574 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
577 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
579 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
580 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
581 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
585 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
586 const zbookmark_phys_t *zb)
589 * We never skip over user/group accounting objects (obj<0)
591 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
592 (int64_t)zb->zb_object >= 0) {
594 * If we already visited this bp & everything below (in
595 * a prior txg sync), don't bother doing it again.
597 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
601 * If we found the block we're trying to resume from, or
602 * we went past it to a different object, zero it out to
603 * indicate that it's OK to start checking for pausing
606 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
607 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
608 dprintf("resuming at %llx/%llx/%llx/%llx\n",
609 (longlong_t)zb->zb_objset,
610 (longlong_t)zb->zb_object,
611 (longlong_t)zb->zb_level,
612 (longlong_t)zb->zb_blkid);
613 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
620 * Return nonzero on i/o error.
621 * Return new buf to write out in *bufp.
624 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
625 dnode_phys_t *dnp, const blkptr_t *bp,
626 const zbookmark_phys_t *zb, dmu_tx_t *tx)
628 dsl_pool_t *dp = scn->scn_dp;
629 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
632 if (BP_GET_LEVEL(bp) > 0) {
633 uint32_t flags = ARC_WAIT;
636 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
639 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
640 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
642 scn->scn_phys.scn_errors++;
645 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
646 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
647 zb->zb_object, zb->zb_blkid * epb + i);
649 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
650 zbookmark_phys_t czb;
652 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
654 zb->zb_blkid * epb + i);
655 dsl_scan_visitbp(cbp, &czb, dnp,
656 ds, scn, ostype, tx);
658 (void) arc_buf_remove_ref(buf, &buf);
659 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
660 uint32_t flags = ARC_WAIT;
663 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
666 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
667 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
669 scn->scn_phys.scn_errors++;
672 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
673 for (j = 0; j < cdnp->dn_nblkptr; j++) {
674 blkptr_t *cbp = &cdnp->dn_blkptr[j];
675 dsl_scan_prefetch(scn, buf, cbp,
676 zb->zb_objset, zb->zb_blkid * epb + i, j);
679 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
680 dsl_scan_visitdnode(scn, ds, ostype,
681 cdnp, zb->zb_blkid * epb + i, tx);
684 (void) arc_buf_remove_ref(buf, &buf);
685 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
686 uint32_t flags = ARC_WAIT;
690 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
691 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
693 scn->scn_phys.scn_errors++;
699 dsl_scan_visitdnode(scn, ds, osp->os_type,
700 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
702 if (OBJSET_BUF_HAS_USERUSED(buf)) {
704 * We also always visit user/group accounting
705 * objects, and never skip them, even if we are
706 * pausing. This is necessary so that the space
707 * deltas from this txg get integrated.
709 dsl_scan_visitdnode(scn, ds, osp->os_type,
710 &osp->os_groupused_dnode,
711 DMU_GROUPUSED_OBJECT, tx);
712 dsl_scan_visitdnode(scn, ds, osp->os_type,
713 &osp->os_userused_dnode,
714 DMU_USERUSED_OBJECT, tx);
716 (void) arc_buf_remove_ref(buf, &buf);
723 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
724 dmu_objset_type_t ostype, dnode_phys_t *dnp,
725 uint64_t object, dmu_tx_t *tx)
729 for (j = 0; j < dnp->dn_nblkptr; j++) {
730 zbookmark_phys_t czb;
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, ds, scn, ostype, tx);
738 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
739 zbookmark_phys_t czb;
740 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
742 dsl_scan_visitbp(&dnp->dn_spill,
743 &czb, dnp, ds, scn, ostype, tx);
748 * The arguments are in this order because mdb can only print the
749 * first 5; we want them to be useful.
752 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
753 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
754 dmu_objset_type_t ostype, dmu_tx_t *tx)
756 dsl_pool_t *dp = scn->scn_dp;
757 arc_buf_t *buf = NULL;
758 blkptr_t bp_toread = *bp;
760 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
762 if (dsl_scan_check_pause(scn, zb))
765 if (dsl_scan_check_resume(scn, dnp, zb))
771 scn->scn_visited_this_txg++;
774 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
775 ds, ds ? ds->ds_object : 0,
776 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
779 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
782 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
786 * If dsl_scan_ddt() has aready visited this block, it will have
787 * already done any translations or scrubbing, so don't call the
790 if (ddt_class_contains(dp->dp_spa,
791 scn->scn_phys.scn_ddt_class_max, bp)) {
797 * If this block is from the future (after cur_max_txg), then we
798 * are doing this on behalf of a deleted snapshot, and we will
799 * revisit the future block on the next pass of this dataset.
800 * Don't scan it now unless we need to because something
801 * under it was modified.
803 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
804 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
809 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
814 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
815 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
816 dsl_scan_visitbp(bp, &zb, NULL,
817 ds, scn, DMU_OST_NONE, tx);
819 dprintf_ds(ds, "finished scan%s", "");
823 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
825 dsl_pool_t *dp = ds->ds_dir->dd_pool;
826 dsl_scan_t *scn = dp->dp_scan;
829 if (scn->scn_phys.scn_state != DSS_SCANNING)
832 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
833 if (dsl_dataset_is_snapshot(ds)) {
834 /* Note, scn_cur_{min,max}_txg stays the same. */
835 scn->scn_phys.scn_bookmark.zb_objset =
836 ds->ds_phys->ds_next_snap_obj;
837 zfs_dbgmsg("destroying ds %llu; currently traversing; "
838 "reset zb_objset to %llu",
839 (u_longlong_t)ds->ds_object,
840 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
841 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
843 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
844 ZB_DESTROYED_OBJSET, 0, 0, 0);
845 zfs_dbgmsg("destroying ds %llu; currently traversing; "
846 "reset bookmark to -1,0,0,0",
847 (u_longlong_t)ds->ds_object);
849 } else if (zap_lookup_int_key(dp->dp_meta_objset,
850 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
851 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
852 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
853 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
854 if (dsl_dataset_is_snapshot(ds)) {
856 * We keep the same mintxg; it could be >
857 * ds_creation_txg if the previous snapshot was
860 VERIFY(zap_add_int_key(dp->dp_meta_objset,
861 scn->scn_phys.scn_queue_obj,
862 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
863 zfs_dbgmsg("destroying ds %llu; in queue; "
864 "replacing with %llu",
865 (u_longlong_t)ds->ds_object,
866 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
868 zfs_dbgmsg("destroying ds %llu; in queue; removing",
869 (u_longlong_t)ds->ds_object);
872 zfs_dbgmsg("destroying ds %llu; ignoring",
873 (u_longlong_t)ds->ds_object);
877 * dsl_scan_sync() should be called after this, and should sync
878 * out our changed state, but just to be safe, do it here.
880 dsl_scan_sync_state(scn, tx);
884 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
886 dsl_pool_t *dp = ds->ds_dir->dd_pool;
887 dsl_scan_t *scn = dp->dp_scan;
890 if (scn->scn_phys.scn_state != DSS_SCANNING)
893 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
895 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
896 scn->scn_phys.scn_bookmark.zb_objset =
897 ds->ds_phys->ds_prev_snap_obj;
898 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
899 "reset zb_objset to %llu",
900 (u_longlong_t)ds->ds_object,
901 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
902 } else if (zap_lookup_int_key(dp->dp_meta_objset,
903 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
904 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
905 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
906 VERIFY(zap_add_int_key(dp->dp_meta_objset,
907 scn->scn_phys.scn_queue_obj,
908 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
909 zfs_dbgmsg("snapshotting ds %llu; in queue; "
910 "replacing with %llu",
911 (u_longlong_t)ds->ds_object,
912 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
914 dsl_scan_sync_state(scn, tx);
918 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
920 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
921 dsl_scan_t *scn = dp->dp_scan;
924 if (scn->scn_phys.scn_state != DSS_SCANNING)
927 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
928 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
929 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
930 "reset zb_objset to %llu",
931 (u_longlong_t)ds1->ds_object,
932 (u_longlong_t)ds2->ds_object);
933 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
934 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
935 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
936 "reset zb_objset to %llu",
937 (u_longlong_t)ds2->ds_object,
938 (u_longlong_t)ds1->ds_object);
941 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
942 ds1->ds_object, &mintxg) == 0) {
945 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
946 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
947 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
948 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
949 err = zap_add_int_key(dp->dp_meta_objset,
950 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
951 VERIFY(err == 0 || err == EEXIST);
953 /* Both were there to begin with */
954 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
955 scn->scn_phys.scn_queue_obj,
956 ds1->ds_object, mintxg, tx));
958 zfs_dbgmsg("clone_swap ds %llu; in queue; "
959 "replacing with %llu",
960 (u_longlong_t)ds1->ds_object,
961 (u_longlong_t)ds2->ds_object);
962 } else if (zap_lookup_int_key(dp->dp_meta_objset,
963 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
964 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
965 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
966 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
967 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
968 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
969 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
970 zfs_dbgmsg("clone_swap ds %llu; in queue; "
971 "replacing with %llu",
972 (u_longlong_t)ds2->ds_object,
973 (u_longlong_t)ds1->ds_object);
976 dsl_scan_sync_state(scn, tx);
979 struct enqueue_clones_arg {
986 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
988 struct enqueue_clones_arg *eca = arg;
991 dsl_scan_t *scn = dp->dp_scan;
993 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
996 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1000 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1001 dsl_dataset_t *prev;
1002 err = dsl_dataset_hold_obj(dp,
1003 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1005 dsl_dataset_rele(ds, FTAG);
1010 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1011 scn->scn_phys.scn_queue_obj, ds->ds_object,
1012 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1013 dsl_dataset_rele(ds, FTAG);
1018 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1020 dsl_pool_t *dp = scn->scn_dp;
1024 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1026 if (dmu_objset_from_ds(ds, &os))
1030 * Only the ZIL in the head (non-snapshot) is valid. Even though
1031 * snapshots can have ZIL block pointers (which may be the same
1032 * BP as in the head), they must be ignored. So we traverse the
1033 * ZIL here, rather than in scan_recurse(), because the regular
1034 * snapshot block-sharing rules don't apply to it.
1036 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1037 dsl_scan_zil(dp, &os->os_zil_header);
1040 * Iterate over the bps in this ds.
1042 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1043 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1045 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1046 dsl_dataset_name(ds, dsname);
1047 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1049 (longlong_t)dsobj, dsname,
1050 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1051 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1052 (int)scn->scn_pausing);
1053 kmem_free(dsname, ZFS_MAXNAMELEN);
1055 if (scn->scn_pausing)
1059 * We've finished this pass over this dataset.
1063 * If we did not completely visit this dataset, do another pass.
1065 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1066 zfs_dbgmsg("incomplete pass; visiting again");
1067 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1068 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1069 scn->scn_phys.scn_queue_obj, ds->ds_object,
1070 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1075 * Add descendent datasets to work queue.
1077 if (ds->ds_phys->ds_next_snap_obj != 0) {
1078 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1079 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1080 ds->ds_phys->ds_creation_txg, tx) == 0);
1082 if (ds->ds_phys->ds_num_children > 1) {
1083 boolean_t usenext = B_FALSE;
1084 if (ds->ds_phys->ds_next_clones_obj != 0) {
1087 * A bug in a previous version of the code could
1088 * cause upgrade_clones_cb() to not set
1089 * ds_next_snap_obj when it should, leading to a
1090 * missing entry. Therefore we can only use the
1091 * next_clones_obj when its count is correct.
1093 int err = zap_count(dp->dp_meta_objset,
1094 ds->ds_phys->ds_next_clones_obj, &count);
1096 count == ds->ds_phys->ds_num_children - 1)
1101 VERIFY0(zap_join_key(dp->dp_meta_objset,
1102 ds->ds_phys->ds_next_clones_obj,
1103 scn->scn_phys.scn_queue_obj,
1104 ds->ds_phys->ds_creation_txg, tx));
1106 struct enqueue_clones_arg eca;
1108 eca.originobj = ds->ds_object;
1110 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1111 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1116 dsl_dataset_rele(ds, FTAG);
1121 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1126 dsl_scan_t *scn = dp->dp_scan;
1128 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1132 while (ds->ds_phys->ds_prev_snap_obj != 0) {
1133 dsl_dataset_t *prev;
1134 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1137 dsl_dataset_rele(ds, FTAG);
1142 * If this is a clone, we don't need to worry about it for now.
1144 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1145 dsl_dataset_rele(ds, FTAG);
1146 dsl_dataset_rele(prev, FTAG);
1149 dsl_dataset_rele(ds, FTAG);
1153 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1154 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1155 dsl_dataset_rele(ds, FTAG);
1160 * Scrub/dedup interaction.
1162 * If there are N references to a deduped block, we don't want to scrub it
1163 * N times -- ideally, we should scrub it exactly once.
1165 * We leverage the fact that the dde's replication class (enum ddt_class)
1166 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1167 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1169 * To prevent excess scrubbing, the scrub begins by walking the DDT
1170 * to find all blocks with refcnt > 1, and scrubs each of these once.
1171 * Since there are two replication classes which contain blocks with
1172 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1173 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1175 * There would be nothing more to say if a block's refcnt couldn't change
1176 * during a scrub, but of course it can so we must account for changes
1177 * in a block's replication class.
1179 * Here's an example of what can occur:
1181 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1182 * when visited during the top-down scrub phase, it will be scrubbed twice.
1183 * This negates our scrub optimization, but is otherwise harmless.
1185 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1186 * on each visit during the top-down scrub phase, it will never be scrubbed.
1187 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1188 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1189 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1190 * while a scrub is in progress, it scrubs the block right then.
1193 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1195 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1196 ddt_entry_t dde = { 0 };
1200 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1203 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1205 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1206 (longlong_t)ddb->ddb_class,
1207 (longlong_t)ddb->ddb_type,
1208 (longlong_t)ddb->ddb_checksum,
1209 (longlong_t)ddb->ddb_cursor);
1211 /* There should be no pending changes to the dedup table */
1212 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1213 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1215 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1218 if (dsl_scan_check_pause(scn, NULL))
1222 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1223 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1224 (int)scn->scn_pausing);
1226 ASSERT(error == 0 || error == ENOENT);
1227 ASSERT(error != ENOENT ||
1228 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1233 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1234 ddt_entry_t *dde, dmu_tx_t *tx)
1236 const ddt_key_t *ddk = &dde->dde_key;
1237 ddt_phys_t *ddp = dde->dde_phys;
1239 zbookmark_phys_t zb = { 0 };
1241 if (scn->scn_phys.scn_state != DSS_SCANNING)
1244 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1245 if (ddp->ddp_phys_birth == 0 ||
1246 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1248 ddt_bp_create(checksum, ddk, ddp, &bp);
1250 scn->scn_visited_this_txg++;
1251 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1256 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1258 dsl_pool_t *dp = scn->scn_dp;
1262 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1263 scn->scn_phys.scn_ddt_class_max) {
1264 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1265 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1266 dsl_scan_ddt(scn, tx);
1267 if (scn->scn_pausing)
1271 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1272 /* First do the MOS & ORIGIN */
1274 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1275 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1276 dsl_scan_visit_rootbp(scn, NULL,
1277 &dp->dp_meta_rootbp, tx);
1278 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1279 if (scn->scn_pausing)
1282 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1283 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1284 enqueue_cb, tx, DS_FIND_CHILDREN));
1286 dsl_scan_visitds(scn,
1287 dp->dp_origin_snap->ds_object, tx);
1289 ASSERT(!scn->scn_pausing);
1290 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1291 ZB_DESTROYED_OBJSET) {
1293 * If we were paused, continue from here. Note if the
1294 * ds we were paused on was deleted, the zb_objset may
1295 * be -1, so we will skip this and find a new objset
1298 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1299 if (scn->scn_pausing)
1304 * In case we were paused right at the end of the ds, zero the
1305 * bookmark so we don't think that we're still trying to resume.
1307 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1309 /* keep pulling things out of the zap-object-as-queue */
1310 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1311 scn->scn_phys.scn_queue_obj),
1312 zap_cursor_retrieve(&zc, &za) == 0) {
1316 dsobj = strtonum(za.za_name, NULL);
1317 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1318 scn->scn_phys.scn_queue_obj, dsobj, tx));
1320 /* Set up min/max txg */
1321 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1322 if (za.za_first_integer != 0) {
1323 scn->scn_phys.scn_cur_min_txg =
1324 MAX(scn->scn_phys.scn_min_txg,
1325 za.za_first_integer);
1327 scn->scn_phys.scn_cur_min_txg =
1328 MAX(scn->scn_phys.scn_min_txg,
1329 ds->ds_phys->ds_prev_snap_txg);
1331 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1332 dsl_dataset_rele(ds, FTAG);
1334 dsl_scan_visitds(scn, dsobj, tx);
1335 zap_cursor_fini(&zc);
1336 if (scn->scn_pausing)
1339 zap_cursor_fini(&zc);
1343 dsl_scan_free_should_pause(dsl_scan_t *scn)
1345 uint64_t elapsed_nanosecs;
1350 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1353 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1354 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1355 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1356 txg_sync_waiting(scn->scn_dp)) ||
1357 spa_shutting_down(scn->scn_dp->dp_spa));
1361 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1363 dsl_scan_t *scn = arg;
1365 if (!scn->scn_is_bptree ||
1366 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1367 if (dsl_scan_free_should_pause(scn))
1368 return (SET_ERROR(ERESTART));
1371 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1372 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1373 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1374 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1375 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1376 scn->scn_visited_this_txg++;
1381 dsl_scan_active(dsl_scan_t *scn)
1383 spa_t *spa = scn->scn_dp->dp_spa;
1384 uint64_t used = 0, comp, uncomp;
1386 if (spa->spa_load_state != SPA_LOAD_NONE)
1388 if (spa_shutting_down(spa))
1390 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1391 (scn->scn_async_destroying && !scn->scn_async_stalled))
1394 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1395 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1396 &used, &comp, &uncomp);
1402 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1404 dsl_scan_t *scn = dp->dp_scan;
1405 spa_t *spa = dp->dp_spa;
1409 * Check for scn_restart_txg before checking spa_load_state, so
1410 * that we can restart an old-style scan while the pool is being
1411 * imported (see dsl_scan_init).
1413 if (scn->scn_restart_txg != 0 &&
1414 scn->scn_restart_txg <= tx->tx_txg) {
1415 pool_scan_func_t func = POOL_SCAN_SCRUB;
1416 dsl_scan_done(scn, B_FALSE, tx);
1417 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1418 func = POOL_SCAN_RESILVER;
1419 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1421 dsl_scan_setup_sync(&func, tx);
1425 * If the scan is inactive due to a stalled async destroy, try again.
1427 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1428 spa_sync_pass(dp->dp_spa) > 1)
1431 scn->scn_visited_this_txg = 0;
1432 scn->scn_pausing = B_FALSE;
1433 scn->scn_sync_start_time = gethrtime();
1434 spa->spa_scrub_active = B_TRUE;
1437 * First process the async destroys. If we pause, don't do
1438 * any scrubbing or resilvering. This ensures that there are no
1439 * async destroys while we are scanning, so the scan code doesn't
1440 * have to worry about traversing it. It is also faster to free the
1441 * blocks than to scrub them.
1443 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1444 scn->scn_is_bptree = B_FALSE;
1445 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1446 NULL, ZIO_FLAG_MUSTSUCCEED);
1447 err = bpobj_iterate(&dp->dp_free_bpobj,
1448 dsl_scan_free_block_cb, scn, tx);
1449 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1451 if (err != 0 && err != ERESTART)
1452 zfs_panic_recover("error %u from bpobj_iterate()", err);
1455 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1456 ASSERT(scn->scn_async_destroying);
1457 scn->scn_is_bptree = B_TRUE;
1458 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1459 NULL, ZIO_FLAG_MUSTSUCCEED);
1460 err = bptree_iterate(dp->dp_meta_objset,
1461 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1462 VERIFY0(zio_wait(scn->scn_zio_root));
1464 if (err == EIO || err == ECKSUM) {
1466 } else if (err != 0 && err != ERESTART) {
1467 zfs_panic_recover("error %u from "
1468 "traverse_dataset_destroyed()", err);
1471 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1472 /* finished; deactivate async destroy feature */
1473 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1474 ASSERT(!spa_feature_is_active(spa,
1475 SPA_FEATURE_ASYNC_DESTROY));
1476 VERIFY0(zap_remove(dp->dp_meta_objset,
1477 DMU_POOL_DIRECTORY_OBJECT,
1478 DMU_POOL_BPTREE_OBJ, tx));
1479 VERIFY0(bptree_free(dp->dp_meta_objset,
1480 dp->dp_bptree_obj, tx));
1481 dp->dp_bptree_obj = 0;
1482 scn->scn_async_destroying = B_FALSE;
1485 * If we didn't make progress, mark the async destroy as
1486 * stalled, so that we will not initiate a spa_sync() on
1489 scn->scn_async_stalled =
1490 (scn->scn_visited_this_txg == 0);
1493 if (scn->scn_visited_this_txg) {
1494 zfs_dbgmsg("freed %llu blocks in %llums from "
1495 "free_bpobj/bptree txg %llu; err=%d",
1496 (longlong_t)scn->scn_visited_this_txg,
1498 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1499 (longlong_t)tx->tx_txg, err);
1500 scn->scn_visited_this_txg = 0;
1503 * Write out changes to the DDT that may be required as a
1504 * result of the blocks freed. This ensures that the DDT
1505 * is clean when a scrub/resilver runs.
1507 ddt_sync(spa, tx->tx_txg);
1511 if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1512 (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 ||
1513 dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 ||
1514 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) {
1516 * We have finished background destroying, but there is still
1517 * some space left in the dp_free_dir. Transfer this leaked
1518 * space to the dp_leak_dir.
1520 if (dp->dp_leak_dir == NULL) {
1521 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1522 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1524 VERIFY0(dsl_pool_open_special_dir(dp,
1525 LEAK_DIR_NAME, &dp->dp_leak_dir));
1526 rrw_exit(&dp->dp_config_rwlock, FTAG);
1528 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1529 dp->dp_free_dir->dd_phys->dd_used_bytes,
1530 dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1531 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1532 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1533 -dp->dp_free_dir->dd_phys->dd_used_bytes,
1534 -dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1535 -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1537 if (!scn->scn_async_destroying) {
1538 /* finished; verify that space accounting went to zero */
1539 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
1540 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
1541 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1544 if (scn->scn_phys.scn_state != DSS_SCANNING)
1547 if (scn->scn_done_txg == tx->tx_txg) {
1548 ASSERT(!scn->scn_pausing);
1549 /* finished with scan. */
1550 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1551 dsl_scan_done(scn, B_TRUE, tx);
1552 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1553 dsl_scan_sync_state(scn, tx);
1557 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1558 scn->scn_phys.scn_ddt_class_max) {
1559 zfs_dbgmsg("doing scan sync txg %llu; "
1560 "ddt bm=%llu/%llu/%llu/%llx",
1561 (longlong_t)tx->tx_txg,
1562 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1563 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1564 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1565 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1566 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1567 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1568 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1569 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1571 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1572 (longlong_t)tx->tx_txg,
1573 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1574 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1575 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1576 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1579 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1580 NULL, ZIO_FLAG_CANFAIL);
1581 dsl_pool_config_enter(dp, FTAG);
1582 dsl_scan_visit(scn, tx);
1583 dsl_pool_config_exit(dp, FTAG);
1584 (void) zio_wait(scn->scn_zio_root);
1585 scn->scn_zio_root = NULL;
1587 zfs_dbgmsg("visited %llu blocks in %llums",
1588 (longlong_t)scn->scn_visited_this_txg,
1589 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1591 if (!scn->scn_pausing) {
1592 scn->scn_done_txg = tx->tx_txg + 1;
1593 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1594 tx->tx_txg, scn->scn_done_txg);
1597 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1598 mutex_enter(&spa->spa_scrub_lock);
1599 while (spa->spa_scrub_inflight > 0) {
1600 cv_wait(&spa->spa_scrub_io_cv,
1601 &spa->spa_scrub_lock);
1603 mutex_exit(&spa->spa_scrub_lock);
1606 dsl_scan_sync_state(scn, tx);
1610 * This will start a new scan, or restart an existing one.
1613 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1617 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1618 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1620 txg = dmu_tx_get_txg(tx);
1621 dp->dp_scan->scn_restart_txg = txg;
1624 dp->dp_scan->scn_restart_txg = txg;
1626 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1630 dsl_scan_resilvering(dsl_pool_t *dp)
1632 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1633 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1641 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1646 * If we resume after a reboot, zab will be NULL; don't record
1647 * incomplete stats in that case.
1652 for (i = 0; i < 4; i++) {
1653 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1654 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1655 if (t & DMU_OT_NEWTYPE)
1657 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1661 zb->zb_asize += BP_GET_ASIZE(bp);
1662 zb->zb_lsize += BP_GET_LSIZE(bp);
1663 zb->zb_psize += BP_GET_PSIZE(bp);
1664 zb->zb_gangs += BP_COUNT_GANG(bp);
1666 switch (BP_GET_NDVAS(bp)) {
1668 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1669 DVA_GET_VDEV(&bp->blk_dva[1]))
1670 zb->zb_ditto_2_of_2_samevdev++;
1673 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1674 DVA_GET_VDEV(&bp->blk_dva[1])) +
1675 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1676 DVA_GET_VDEV(&bp->blk_dva[2])) +
1677 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1678 DVA_GET_VDEV(&bp->blk_dva[2]));
1680 zb->zb_ditto_2_of_3_samevdev++;
1681 else if (equal == 3)
1682 zb->zb_ditto_3_of_3_samevdev++;
1689 dsl_scan_scrub_done(zio_t *zio)
1691 spa_t *spa = zio->io_spa;
1693 zio_data_buf_free(zio->io_data, zio->io_size);
1695 mutex_enter(&spa->spa_scrub_lock);
1696 spa->spa_scrub_inflight--;
1697 cv_broadcast(&spa->spa_scrub_io_cv);
1699 if (zio->io_error && (zio->io_error != ECKSUM ||
1700 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1701 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1703 mutex_exit(&spa->spa_scrub_lock);
1707 dsl_scan_scrub_cb(dsl_pool_t *dp,
1708 const blkptr_t *bp, const zbookmark_phys_t *zb)
1710 dsl_scan_t *scn = dp->dp_scan;
1711 size_t size = BP_GET_PSIZE(bp);
1712 spa_t *spa = dp->dp_spa;
1713 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1715 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1716 unsigned int scan_delay = 0;
1718 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1719 phys_birth >= scn->scn_phys.scn_max_txg)
1722 count_block(dp->dp_blkstats, bp);
1724 if (BP_IS_EMBEDDED(bp))
1727 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1728 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1729 zio_flags |= ZIO_FLAG_SCRUB;
1731 scan_delay = zfs_scrub_delay;
1733 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1734 zio_flags |= ZIO_FLAG_RESILVER;
1736 scan_delay = zfs_resilver_delay;
1739 /* If it's an intent log block, failure is expected. */
1740 if (zb->zb_level == ZB_ZIL_LEVEL)
1741 zio_flags |= ZIO_FLAG_SPECULATIVE;
1743 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1744 vdev_t *vd = vdev_lookup_top(spa,
1745 DVA_GET_VDEV(&bp->blk_dva[d]));
1748 * Keep track of how much data we've examined so that
1749 * zpool(1M) status can make useful progress reports.
1751 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1752 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1754 /* if it's a resilver, this may not be in the target range */
1756 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1758 * Gang members may be spread across multiple
1759 * vdevs, so the best estimate we have is the
1760 * scrub range, which has already been checked.
1761 * XXX -- it would be better to change our
1762 * allocation policy to ensure that all
1763 * gang members reside on the same vdev.
1767 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1773 if (needs_io && !zfs_no_scrub_io) {
1774 vdev_t *rvd = spa->spa_root_vdev;
1775 uint64_t maxinflight = rvd->vdev_children *
1776 MAX(zfs_top_maxinflight, 1);
1777 void *data = zio_data_buf_alloc(size);
1779 mutex_enter(&spa->spa_scrub_lock);
1780 while (spa->spa_scrub_inflight >= maxinflight)
1781 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1782 spa->spa_scrub_inflight++;
1783 mutex_exit(&spa->spa_scrub_lock);
1786 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1787 * then throttle our workload to limit the impact of a scan.
1789 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1790 delay(MAX((int)scan_delay, 0));
1792 zio_nowait(zio_read(NULL, spa, bp, data, size,
1793 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1797 /* do not relocate this block */
1802 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1804 spa_t *spa = dp->dp_spa;
1807 * Purge all vdev caches and probe all devices. We do this here
1808 * rather than in sync context because this requires a writer lock
1809 * on the spa_config lock, which we can't do from sync context. The
1810 * spa_scrub_reopen flag indicates that vdev_open() should not
1811 * attempt to start another scrub.
1813 spa_vdev_state_enter(spa, SCL_NONE);
1814 spa->spa_scrub_reopen = B_TRUE;
1815 vdev_reopen(spa->spa_root_vdev);
1816 spa->spa_scrub_reopen = B_FALSE;
1817 (void) spa_vdev_state_exit(spa, NULL, 0);
1819 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1820 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));