1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2016 Gary Mills
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright 2017 Joyent, Inc.
26 * Copyright (c) 2017 Datto Inc.
27 */
28
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>
38 #include <sys/arc.h>
39 #include <sys/zap.h>
40 #include <sys/zio.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>
48 #include <sys/ddt.h>
49 #include <sys/sa.h>
50 #include <sys/sa_impl.h>
51 #include <sys/zfeature.h>
52 #include <sys/abd.h>
53 #ifdef _KERNEL
54 #include <sys/zfs_vfsops.h>
55 #endif
56
57 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
58 const zbookmark_phys_t *);
59
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 *);
64
65 int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
66 int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
67 int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
68 int zfs_scan_idle = 50; /* idle window in clock ticks */
69
70 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
71 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
72 int zfs_obsolete_min_time_ms = 500; /* min millisecs to obsolete per txg */
73 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
74 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
75 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
76 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
77 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
78 /* max number of blocks to free in a single TXG */
79 uint64_t zfs_async_block_max_blocks = UINT64_MAX;
80
81 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
82 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
83 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
84
85 extern int zfs_txg_timeout;
86
87 /*
88 * Enable/disable the processing of the free_bpobj object.
89 */
90 boolean_t zfs_free_bpobj_enabled = B_TRUE;
91
92 /* the order has to match pool_scan_type */
93 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
94 NULL,
95 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
96 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
97 };
98
99 int
100 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
101 {
102 int err;
103 dsl_scan_t *scn;
104 spa_t *spa = dp->dp_spa;
105 uint64_t f;
106
107 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
108 scn->scn_dp = dp;
109
110 /*
111 * It's possible that we're resuming a scan after a reboot so
112 * make sure that the scan_async_destroying flag is initialized
113 * appropriately.
114 */
115 ASSERT(!scn->scn_async_destroying);
116 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
117 SPA_FEATURE_ASYNC_DESTROY);
118
119 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
120 "scrub_func", sizeof (uint64_t), 1, &f);
121 if (err == 0) {
122 /*
123 * There was an old-style scrub in progress. Restart a
124 * new-style scrub from the beginning.
125 */
126 scn->scn_restart_txg = txg;
127 zfs_dbgmsg("old-style scrub was in progress; "
128 "restarting new-style scrub in txg %llu",
129 scn->scn_restart_txg);
130
131 /*
132 * Load the queue obj from the old location so that it
133 * can be freed by dsl_scan_done().
134 */
135 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
136 "scrub_queue", sizeof (uint64_t), 1,
137 &scn->scn_phys.scn_queue_obj);
138 } else {
139 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
140 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
141 &scn->scn_phys);
142 if (err == ENOENT)
143 return (0);
144 else if (err)
145 return (err);
146
147 if (scn->scn_phys.scn_state == DSS_SCANNING &&
148 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
149 /*
150 * A new-type scrub was in progress on an old
151 * pool, and the pool was accessed by old
152 * software. Restart from the beginning, since
153 * the old software may have changed the pool in
154 * the meantime.
155 */
156 scn->scn_restart_txg = txg;
157 zfs_dbgmsg("new-style scrub was modified "
158 "by old software; restarting in txg %llu",
159 scn->scn_restart_txg);
160 }
161 }
162
163 spa_scan_stat_init(spa);
164 return (0);
165 }
166
167 void
168 dsl_scan_fini(dsl_pool_t *dp)
169 {
170 if (dp->dp_scan) {
171 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
172 dp->dp_scan = NULL;
173 }
174 }
175
176 /* ARGSUSED */
177 static int
178 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
179 {
180 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
181
182 if (scn->scn_phys.scn_state == DSS_SCANNING)
183 return (SET_ERROR(EBUSY));
184
185 return (0);
186 }
187
188 static void
189 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
190 {
191 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
192 pool_scan_func_t *funcp = arg;
193 dmu_object_type_t ot = 0;
194 dsl_pool_t *dp = scn->scn_dp;
195 spa_t *spa = dp->dp_spa;
196
197 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
198 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
199 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
200 scn->scn_phys.scn_func = *funcp;
201 scn->scn_phys.scn_state = DSS_SCANNING;
202 scn->scn_phys.scn_min_txg = 0;
203 scn->scn_phys.scn_max_txg = tx->tx_txg;
204 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
205 scn->scn_phys.scn_start_time = gethrestime_sec();
206 scn->scn_phys.scn_errors = 0;
207 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
208 scn->scn_restart_txg = 0;
209 scn->scn_done_txg = 0;
210 spa_scan_stat_init(spa);
211
212 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
213 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
214
215 /* rewrite all disk labels */
216 vdev_config_dirty(spa->spa_root_vdev);
217
218 if (vdev_resilver_needed(spa->spa_root_vdev,
219 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
220 spa_event_notify(spa, NULL, NULL,
221 ESC_ZFS_RESILVER_START);
222 } else {
223 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
224 }
225
226 spa->spa_scrub_started = B_TRUE;
227 /*
228 * If this is an incremental scrub, limit the DDT scrub phase
229 * to just the auto-ditto class (for correctness); the rest
230 * of the scrub should go faster using top-down pruning.
231 */
232 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
233 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
234
235 }
236
237 /* back to the generic stuff */
238
239 if (dp->dp_blkstats == NULL) {
240 dp->dp_blkstats =
241 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
242 }
243 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
244
245 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
246 ot = DMU_OT_ZAP_OTHER;
247
248 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
249 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
250
251 dsl_scan_sync_state(scn, tx);
252
253 spa_history_log_internal(spa, "scan setup", tx,
254 "func=%u mintxg=%llu maxtxg=%llu",
255 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
256 }
257
258 /* ARGSUSED */
259 static void
260 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
261 {
262 static const char *old_names[] = {
263 "scrub_bookmark",
264 "scrub_ddt_bookmark",
265 "scrub_ddt_class_max",
266 "scrub_queue",
267 "scrub_min_txg",
268 "scrub_max_txg",
269 "scrub_func",
270 "scrub_errors",
271 NULL
272 };
273
274 dsl_pool_t *dp = scn->scn_dp;
275 spa_t *spa = dp->dp_spa;
276 int i;
277
278 /* Remove any remnants of an old-style scrub. */
279 for (i = 0; old_names[i]; i++) {
280 (void) zap_remove(dp->dp_meta_objset,
281 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
282 }
283
284 if (scn->scn_phys.scn_queue_obj != 0) {
285 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
286 scn->scn_phys.scn_queue_obj, tx));
287 scn->scn_phys.scn_queue_obj = 0;
288 }
289
290 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
291
292 /*
293 * If we were "restarted" from a stopped state, don't bother
294 * with anything else.
295 */
296 if (scn->scn_phys.scn_state != DSS_SCANNING)
297 return;
298
299 if (complete)
300 scn->scn_phys.scn_state = DSS_FINISHED;
301 else
302 scn->scn_phys.scn_state = DSS_CANCELED;
303
304 if (dsl_scan_restarting(scn, tx))
305 spa_history_log_internal(spa, "scan aborted, restarting", tx,
306 "errors=%llu", spa_get_errlog_size(spa));
307 else if (!complete)
308 spa_history_log_internal(spa, "scan cancelled", tx,
309 "errors=%llu", spa_get_errlog_size(spa));
310 else
311 spa_history_log_internal(spa, "scan done", tx,
312 "errors=%llu", spa_get_errlog_size(spa));
313
314 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
315 mutex_enter(&spa->spa_scrub_lock);
316 while (spa->spa_scrub_inflight > 0) {
317 cv_wait(&spa->spa_scrub_io_cv,
318 &spa->spa_scrub_lock);
319 }
320 mutex_exit(&spa->spa_scrub_lock);
321 spa->spa_scrub_started = B_FALSE;
322 spa->spa_scrub_active = B_FALSE;
323
324 /*
325 * If the scrub/resilver completed, update all DTLs to
326 * reflect this. Whether it succeeded or not, vacate
327 * all temporary scrub DTLs.
328 */
329 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
330 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
331 if (complete) {
332 spa_event_notify(spa, NULL, NULL,
333 scn->scn_phys.scn_min_txg ?
334 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
335 }
336 spa_errlog_rotate(spa);
337
338 /*
339 * We may have finished replacing a device.
340 * Let the async thread assess this and handle the detach.
341 */
342 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
343 }
344
345 scn->scn_phys.scn_end_time = gethrestime_sec();
346 }
347
348 /* ARGSUSED */
349 static int
350 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
351 {
352 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
353
354 if (scn->scn_phys.scn_state != DSS_SCANNING)
355 return (SET_ERROR(ENOENT));
356 return (0);
357 }
358
359 /* ARGSUSED */
360 static void
361 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
362 {
363 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
364
365 dsl_scan_done(scn, B_FALSE, tx);
366 dsl_scan_sync_state(scn, tx);
367 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
368 }
369
370 int
371 dsl_scan_cancel(dsl_pool_t *dp)
372 {
373 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
374 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
375 }
376
377 boolean_t
378 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
379 {
380 if (dsl_scan_scrubbing(scn->scn_dp) &&
381 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED)
382 return (B_TRUE);
383
384 return (B_FALSE);
385 }
386
387 static int
388 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
389 {
390 pool_scrub_cmd_t *cmd = arg;
391 dsl_pool_t *dp = dmu_tx_pool(tx);
392 dsl_scan_t *scn = dp->dp_scan;
393
394 if (*cmd == POOL_SCRUB_PAUSE) {
395 /* can't pause a scrub when there is no in-progress scrub */
396 if (!dsl_scan_scrubbing(dp))
397 return (SET_ERROR(ENOENT));
398
399 /* can't pause a paused scrub */
400 if (dsl_scan_is_paused_scrub(scn))
401 return (SET_ERROR(EBUSY));
402 } else if (*cmd != POOL_SCRUB_NORMAL) {
403 return (SET_ERROR(ENOTSUP));
404 }
405
406 return (0);
407 }
408
409 static void
410 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
411 {
412 pool_scrub_cmd_t *cmd = arg;
413 dsl_pool_t *dp = dmu_tx_pool(tx);
414 spa_t *spa = dp->dp_spa;
415 dsl_scan_t *scn = dp->dp_scan;
416
417 if (*cmd == POOL_SCRUB_PAUSE) {
418 /* can't pause a scrub when there is no in-progress scrub */
419 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
420 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
421 dsl_scan_sync_state(scn, tx);
422 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
423 } else {
424 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
425 if (dsl_scan_is_paused_scrub(scn)) {
426 /*
427 * We need to keep track of how much time we spend
428 * paused per pass so that we can adjust the scrub rate
429 * shown in the output of 'zpool status'
430 */
431 spa->spa_scan_pass_scrub_spent_paused +=
432 gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
433 spa->spa_scan_pass_scrub_pause = 0;
434 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
435 dsl_scan_sync_state(scn, tx);
436 }
437 }
438 }
439
440 /*
441 * Set scrub pause/resume state if it makes sense to do so
442 */
443 int
444 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
445 {
446 return (dsl_sync_task(spa_name(dp->dp_spa),
447 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
448 ZFS_SPACE_CHECK_RESERVED));
449 }
450
451 boolean_t
452 dsl_scan_scrubbing(const dsl_pool_t *dp)
453 {
454 dsl_scan_t *scn = dp->dp_scan;
455
456 if (scn->scn_phys.scn_state == DSS_SCANNING &&
457 scn->scn_phys.scn_func == POOL_SCAN_SCRUB)
458 return (B_TRUE);
459
460 return (B_FALSE);
461 }
462
463 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
464 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
465 dmu_objset_type_t ostype, dmu_tx_t *tx);
466 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
467 dmu_objset_type_t ostype,
468 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
469
470 void
471 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
472 {
473 zio_free(dp->dp_spa, txg, bp);
474 }
475
476 void
477 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
478 {
479 ASSERT(dsl_pool_sync_context(dp));
480 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
481 }
482
483 static uint64_t
484 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
485 {
486 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
487 if (ds->ds_is_snapshot)
488 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
489 return (smt);
490 }
491
492 static void
493 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
494 {
495 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
496 DMU_POOL_DIRECTORY_OBJECT,
497 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
498 &scn->scn_phys, tx));
499 }
500
501 extern int zfs_vdev_async_write_active_min_dirty_percent;
502
503 static boolean_t
504 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
505 {
506 /* we never skip user/group accounting objects */
507 if (zb && (int64_t)zb->zb_object < 0)
508 return (B_FALSE);
509
510 if (scn->scn_suspending)
511 return (B_TRUE); /* we're already suspending */
512
513 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
514 return (B_FALSE); /* we're resuming */
515
516 /* We only know how to resume from level-0 blocks. */
517 if (zb && zb->zb_level != 0)
518 return (B_FALSE);
519
520 /*
521 * We suspend if:
522 * - we have scanned for the maximum time: an entire txg
523 * timeout (default 5 sec)
524 * or
525 * - we have scanned for at least the minimum time (default 1 sec
526 * for scrub, 3 sec for resilver), and either we have sufficient
527 * dirty data that we are starting to write more quickly
528 * (default 30%), or someone is explicitly waiting for this txg
529 * to complete.
530 * or
531 * - the spa is shutting down because this pool is being exported
532 * or the machine is rebooting.
533 */
534 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
535 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
536 uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
537 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
538 if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
539 (NSEC2MSEC(elapsed_nanosecs) > mintime &&
540 (txg_sync_waiting(scn->scn_dp) ||
541 dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
542 spa_shutting_down(scn->scn_dp->dp_spa)) {
543 if (zb) {
544 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
545 (longlong_t)zb->zb_objset,
546 (longlong_t)zb->zb_object,
547 (longlong_t)zb->zb_level,
548 (longlong_t)zb->zb_blkid);
549 scn->scn_phys.scn_bookmark = *zb;
550 }
551 dprintf("suspending at DDT bookmark %llx/%llx/%llx/%llx\n",
552 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
553 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
554 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
555 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
556 scn->scn_suspending = B_TRUE;
557 return (B_TRUE);
558 }
559 return (B_FALSE);
560 }
561
562 typedef struct zil_scan_arg {
563 dsl_pool_t *zsa_dp;
564 zil_header_t *zsa_zh;
565 } zil_scan_arg_t;
566
567 /* ARGSUSED */
568 static int
569 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
570 {
571 zil_scan_arg_t *zsa = arg;
572 dsl_pool_t *dp = zsa->zsa_dp;
573 dsl_scan_t *scn = dp->dp_scan;
574 zil_header_t *zh = zsa->zsa_zh;
575 zbookmark_phys_t zb;
576
577 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
578 return (0);
579
580 /*
581 * One block ("stubby") can be allocated a long time ago; we
582 * want to visit that one because it has been allocated
583 * (on-disk) even if it hasn't been claimed (even though for
584 * scrub there's nothing to do to it).
585 */
586 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
587 return (0);
588
589 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
590 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
591
592 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
593 return (0);
594 }
595
596 /* ARGSUSED */
597 static int
598 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
599 {
600 if (lrc->lrc_txtype == TX_WRITE) {
601 zil_scan_arg_t *zsa = arg;
602 dsl_pool_t *dp = zsa->zsa_dp;
603 dsl_scan_t *scn = dp->dp_scan;
604 zil_header_t *zh = zsa->zsa_zh;
605 lr_write_t *lr = (lr_write_t *)lrc;
606 blkptr_t *bp = &lr->lr_blkptr;
607 zbookmark_phys_t zb;
608
609 if (BP_IS_HOLE(bp) ||
610 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
611 return (0);
612
613 /*
614 * birth can be < claim_txg if this record's txg is
615 * already txg sync'ed (but this log block contains
616 * other records that are not synced)
617 */
618 if (claim_txg == 0 || bp->blk_birth < claim_txg)
619 return (0);
620
621 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
622 lr->lr_foid, ZB_ZIL_LEVEL,
623 lr->lr_offset / BP_GET_LSIZE(bp));
624
625 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
626 }
627 return (0);
628 }
629
630 static void
631 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
632 {
633 uint64_t claim_txg = zh->zh_claim_txg;
634 zil_scan_arg_t zsa = { dp, zh };
635 zilog_t *zilog;
636
637 /*
638 * We only want to visit blocks that have been claimed but not yet
639 * replayed (or, in read-only mode, blocks that *would* be claimed).
640 */
641 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
642 return;
643
644 zilog = zil_alloc(dp->dp_meta_objset, zh);
645
646 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
647 claim_txg);
648
649 zil_free(zilog);
650 }
651
652 /* ARGSUSED */
653 static void
654 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
655 uint64_t objset, uint64_t object, uint64_t blkid)
656 {
657 zbookmark_phys_t czb;
658 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
659
660 if (zfs_no_scrub_prefetch)
661 return;
662
663 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
664 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
665 return;
666
667 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
668
669 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
670 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
671 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
672 }
673
674 static boolean_t
675 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
676 const zbookmark_phys_t *zb)
677 {
678 /*
679 * We never skip over user/group accounting objects (obj<0)
680 */
681 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
682 (int64_t)zb->zb_object >= 0) {
683 /*
684 * If we already visited this bp & everything below (in
685 * a prior txg sync), don't bother doing it again.
686 */
687 if (zbookmark_subtree_completed(dnp, zb,
688 &scn->scn_phys.scn_bookmark))
689 return (B_TRUE);
690
691 /*
692 * If we found the block we're trying to resume from, or
693 * we went past it to a different object, zero it out to
694 * indicate that it's OK to start checking for suspending
695 * again.
696 */
697 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
698 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
699 dprintf("resuming at %llx/%llx/%llx/%llx\n",
700 (longlong_t)zb->zb_objset,
701 (longlong_t)zb->zb_object,
702 (longlong_t)zb->zb_level,
703 (longlong_t)zb->zb_blkid);
704 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
705 }
706 }
707 return (B_FALSE);
708 }
709
710 /*
711 * Return nonzero on i/o error.
712 * Return new buf to write out in *bufp.
713 */
714 static int
715 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
716 dnode_phys_t *dnp, const blkptr_t *bp,
717 const zbookmark_phys_t *zb, dmu_tx_t *tx)
718 {
719 dsl_pool_t *dp = scn->scn_dp;
720 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
721 int err;
722
723 if (BP_GET_LEVEL(bp) > 0) {
724 arc_flags_t flags = ARC_FLAG_WAIT;
725 int i;
726 blkptr_t *cbp;
727 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
728 arc_buf_t *buf;
729
730 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
731 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
732 if (err) {
733 scn->scn_phys.scn_errors++;
734 return (err);
735 }
736 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
737 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
738 zb->zb_object, zb->zb_blkid * epb + i);
739 }
740 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
741 zbookmark_phys_t czb;
742
743 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
744 zb->zb_level - 1,
745 zb->zb_blkid * epb + i);
746 dsl_scan_visitbp(cbp, &czb, dnp,
747 ds, scn, ostype, tx);
748 }
749 arc_buf_destroy(buf, &buf);
750 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
751 arc_flags_t flags = ARC_FLAG_WAIT;
752 dnode_phys_t *cdnp;
753 int i, j;
754 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
755 arc_buf_t *buf;
756
757 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
758 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
759 if (err) {
760 scn->scn_phys.scn_errors++;
761 return (err);
762 }
763 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
764 for (j = 0; j < cdnp->dn_nblkptr; j++) {
765 blkptr_t *cbp = &cdnp->dn_blkptr[j];
766 dsl_scan_prefetch(scn, buf, cbp,
767 zb->zb_objset, zb->zb_blkid * epb + i, j);
768 }
769 }
770 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
771 dsl_scan_visitdnode(scn, ds, ostype,
772 cdnp, zb->zb_blkid * epb + i, tx);
773 }
774
775 arc_buf_destroy(buf, &buf);
776 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
777 arc_flags_t flags = ARC_FLAG_WAIT;
778 objset_phys_t *osp;
779 arc_buf_t *buf;
780
781 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
782 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
783 if (err) {
784 scn->scn_phys.scn_errors++;
785 return (err);
786 }
787
788 osp = buf->b_data;
789
790 dsl_scan_visitdnode(scn, ds, osp->os_type,
791 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
792
793 if (OBJSET_BUF_HAS_USERUSED(buf)) {
794 /*
795 * We also always visit user/group accounting
796 * objects, and never skip them, even if we are
797 * suspending. This is necessary so that the space
798 * deltas from this txg get integrated.
799 */
800 dsl_scan_visitdnode(scn, ds, osp->os_type,
801 &osp->os_groupused_dnode,
802 DMU_GROUPUSED_OBJECT, tx);
803 dsl_scan_visitdnode(scn, ds, osp->os_type,
804 &osp->os_userused_dnode,
805 DMU_USERUSED_OBJECT, tx);
806 }
807 arc_buf_destroy(buf, &buf);
808 }
809
810 return (0);
811 }
812
813 static void
814 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
815 dmu_objset_type_t ostype, dnode_phys_t *dnp,
816 uint64_t object, dmu_tx_t *tx)
817 {
818 int j;
819
820 for (j = 0; j < dnp->dn_nblkptr; j++) {
821 zbookmark_phys_t czb;
822
823 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
824 dnp->dn_nlevels - 1, j);
825 dsl_scan_visitbp(&dnp->dn_blkptr[j],
826 &czb, dnp, ds, scn, ostype, tx);
827 }
828
829 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
830 zbookmark_phys_t czb;
831 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
832 0, DMU_SPILL_BLKID);
833 dsl_scan_visitbp(&dnp->dn_spill,
834 &czb, dnp, ds, scn, ostype, tx);
835 }
836 }
837
838 /*
839 * The arguments are in this order because mdb can only print the
840 * first 5; we want them to be useful.
841 */
842 static void
843 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
844 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
845 dmu_objset_type_t ostype, dmu_tx_t *tx)
846 {
847 dsl_pool_t *dp = scn->scn_dp;
848 arc_buf_t *buf = NULL;
849 blkptr_t bp_toread = *bp;
850
851 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
852
853 if (dsl_scan_check_suspend(scn, zb))
854 return;
855
856 if (dsl_scan_check_resume(scn, dnp, zb))
857 return;
858
859 if (BP_IS_HOLE(bp))
860 return;
861
862 scn->scn_visited_this_txg++;
863
864 dprintf_bp(bp,
865 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
866 ds, ds ? ds->ds_object : 0,
867 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
868 bp);
869
870 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
871 return;
872
873 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
874 return;
875
876 /*
877 * If dsl_scan_ddt() has already visited this block, it will have
878 * already done any translations or scrubbing, so don't call the
879 * callback again.
880 */
881 if (ddt_class_contains(dp->dp_spa,
882 scn->scn_phys.scn_ddt_class_max, bp)) {
883 ASSERT(buf == NULL);
884 return;
885 }
886
887 /*
888 * If this block is from the future (after cur_max_txg), then we
889 * are doing this on behalf of a deleted snapshot, and we will
890 * revisit the future block on the next pass of this dataset.
891 * Don't scan it now unless we need to because something
892 * under it was modified.
893 */
894 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
895 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
896 }
897 }
898
899 static void
900 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
901 dmu_tx_t *tx)
902 {
903 zbookmark_phys_t zb;
904
905 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
906 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
907 dsl_scan_visitbp(bp, &zb, NULL,
908 ds, scn, DMU_OST_NONE, tx);
909
910 dprintf_ds(ds, "finished scan%s", "");
911 }
912
913 void
914 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
915 {
916 dsl_pool_t *dp = ds->ds_dir->dd_pool;
917 dsl_scan_t *scn = dp->dp_scan;
918 uint64_t mintxg;
919
920 if (scn->scn_phys.scn_state != DSS_SCANNING)
921 return;
922
923 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
924 if (ds->ds_is_snapshot) {
925 /*
926 * Note:
927 * - scn_cur_{min,max}_txg stays the same.
928 * - Setting the flag is not really necessary if
929 * scn_cur_max_txg == scn_max_txg, because there
930 * is nothing after this snapshot that we care
931 * about. However, we set it anyway and then
932 * ignore it when we retraverse it in
933 * dsl_scan_visitds().
934 */
935 scn->scn_phys.scn_bookmark.zb_objset =
936 dsl_dataset_phys(ds)->ds_next_snap_obj;
937 zfs_dbgmsg("destroying ds %llu; currently traversing; "
938 "reset zb_objset to %llu",
939 (u_longlong_t)ds->ds_object,
940 (u_longlong_t)dsl_dataset_phys(ds)->
941 ds_next_snap_obj);
942 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
943 } else {
944 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
945 ZB_DESTROYED_OBJSET, 0, 0, 0);
946 zfs_dbgmsg("destroying ds %llu; currently traversing; "
947 "reset bookmark to -1,0,0,0",
948 (u_longlong_t)ds->ds_object);
949 }
950 } else if (zap_lookup_int_key(dp->dp_meta_objset,
951 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
952 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
953 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
954 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
955 if (ds->ds_is_snapshot) {
956 /*
957 * We keep the same mintxg; it could be >
958 * ds_creation_txg if the previous snapshot was
959 * deleted too.
960 */
961 VERIFY(zap_add_int_key(dp->dp_meta_objset,
962 scn->scn_phys.scn_queue_obj,
963 dsl_dataset_phys(ds)->ds_next_snap_obj,
964 mintxg, tx) == 0);
965 zfs_dbgmsg("destroying ds %llu; in queue; "
966 "replacing with %llu",
967 (u_longlong_t)ds->ds_object,
968 (u_longlong_t)dsl_dataset_phys(ds)->
969 ds_next_snap_obj);
970 } else {
971 zfs_dbgmsg("destroying ds %llu; in queue; removing",
972 (u_longlong_t)ds->ds_object);
973 }
974 }
975
976 /*
977 * dsl_scan_sync() should be called after this, and should sync
978 * out our changed state, but just to be safe, do it here.
979 */
980 dsl_scan_sync_state(scn, tx);
981 }
982
983 void
984 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
985 {
986 dsl_pool_t *dp = ds->ds_dir->dd_pool;
987 dsl_scan_t *scn = dp->dp_scan;
988 uint64_t mintxg;
989
990 if (scn->scn_phys.scn_state != DSS_SCANNING)
991 return;
992
993 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
994
995 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
996 scn->scn_phys.scn_bookmark.zb_objset =
997 dsl_dataset_phys(ds)->ds_prev_snap_obj;
998 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
999 "reset zb_objset to %llu",
1000 (u_longlong_t)ds->ds_object,
1001 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
1002 } else if (zap_lookup_int_key(dp->dp_meta_objset,
1003 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
1004 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1005 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
1006 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1007 scn->scn_phys.scn_queue_obj,
1008 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
1009 zfs_dbgmsg("snapshotting ds %llu; in queue; "
1010 "replacing with %llu",
1011 (u_longlong_t)ds->ds_object,
1012 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
1013 }
1014 dsl_scan_sync_state(scn, tx);
1015 }
1016
1017 void
1018 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
1019 {
1020 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
1021 dsl_scan_t *scn = dp->dp_scan;
1022 uint64_t mintxg;
1023
1024 if (scn->scn_phys.scn_state != DSS_SCANNING)
1025 return;
1026
1027 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
1028 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
1029 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
1030 "reset zb_objset to %llu",
1031 (u_longlong_t)ds1->ds_object,
1032 (u_longlong_t)ds2->ds_object);
1033 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
1034 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
1035 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
1036 "reset zb_objset to %llu",
1037 (u_longlong_t)ds2->ds_object,
1038 (u_longlong_t)ds1->ds_object);
1039 }
1040
1041 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1042 ds1->ds_object, &mintxg) == 0) {
1043 int err;
1044
1045 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
1046 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1047 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1048 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
1049 err = zap_add_int_key(dp->dp_meta_objset,
1050 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
1051 VERIFY(err == 0 || err == EEXIST);
1052 if (err == EEXIST) {
1053 /* Both were there to begin with */
1054 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1055 scn->scn_phys.scn_queue_obj,
1056 ds1->ds_object, mintxg, tx));
1057 }
1058 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1059 "replacing with %llu",
1060 (u_longlong_t)ds1->ds_object,
1061 (u_longlong_t)ds2->ds_object);
1062 } else if (zap_lookup_int_key(dp->dp_meta_objset,
1063 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1064 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
1065 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1066 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1067 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1068 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1069 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1070 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1071 "replacing with %llu",
1072 (u_longlong_t)ds2->ds_object,
1073 (u_longlong_t)ds1->ds_object);
1074 }
1075
1076 dsl_scan_sync_state(scn, tx);
1077 }
1078
1079 struct enqueue_clones_arg {
1080 dmu_tx_t *tx;
1081 uint64_t originobj;
1082 };
1083
1084 /* ARGSUSED */
1085 static int
1086 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1087 {
1088 struct enqueue_clones_arg *eca = arg;
1089 dsl_dataset_t *ds;
1090 int err;
1091 dsl_scan_t *scn = dp->dp_scan;
1092
1093 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1094 return (0);
1095
1096 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1097 if (err)
1098 return (err);
1099
1100 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1101 dsl_dataset_t *prev;
1102 err = dsl_dataset_hold_obj(dp,
1103 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1104
1105 dsl_dataset_rele(ds, FTAG);
1106 if (err)
1107 return (err);
1108 ds = prev;
1109 }
1110 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1111 scn->scn_phys.scn_queue_obj, ds->ds_object,
1112 dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1113 dsl_dataset_rele(ds, FTAG);
1114 return (0);
1115 }
1116
1117 static void
1118 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1119 {
1120 dsl_pool_t *dp = scn->scn_dp;
1121 dsl_dataset_t *ds;
1122
1123 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1124
1125 if (scn->scn_phys.scn_cur_min_txg >=
1126 scn->scn_phys.scn_max_txg) {
1127 /*
1128 * This can happen if this snapshot was created after the
1129 * scan started, and we already completed a previous snapshot
1130 * that was created after the scan started. This snapshot
1131 * only references blocks with:
1132 *
1133 * birth < our ds_creation_txg
1134 * cur_min_txg is no less than ds_creation_txg.
1135 * We have already visited these blocks.
1136 * or
1137 * birth > scn_max_txg
1138 * The scan requested not to visit these blocks.
1139 *
1140 * Subsequent snapshots (and clones) can reference our
1141 * blocks, or blocks with even higher birth times.
1142 * Therefore we do not need to visit them either,
1143 * so we do not add them to the work queue.
1144 *
1145 * Note that checking for cur_min_txg >= cur_max_txg
1146 * is not sufficient, because in that case we may need to
1147 * visit subsequent snapshots. This happens when min_txg > 0,
1148 * which raises cur_min_txg. In this case we will visit
1149 * this dataset but skip all of its blocks, because the
1150 * rootbp's birth time is < cur_min_txg. Then we will
1151 * add the next snapshots/clones to the work queue.
1152 */
1153 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1154 dsl_dataset_name(ds, dsname);
1155 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1156 "cur_min_txg (%llu) >= max_txg (%llu)",
1157 dsobj, dsname,
1158 scn->scn_phys.scn_cur_min_txg,
1159 scn->scn_phys.scn_max_txg);
1160 kmem_free(dsname, MAXNAMELEN);
1161
1162 goto out;
1163 }
1164
1165 /*
1166 * Only the ZIL in the head (non-snapshot) is valid. Even though
1167 * snapshots can have ZIL block pointers (which may be the same
1168 * BP as in the head), they must be ignored. In addition, $ORIGIN
1169 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
1170 * need to look for a ZIL in it either. So we traverse the ZIL here,
1171 * rather than in scan_recurse(), because the regular snapshot
1172 * block-sharing rules don't apply to it.
1173 */
1174 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds) &&
1175 ds->ds_dir != dp->dp_origin_snap->ds_dir) {
1176 objset_t *os;
1177 if (dmu_objset_from_ds(ds, &os) != 0) {
1178 goto out;
1179 }
1180 dsl_scan_zil(dp, &os->os_zil_header);
1181 }
1182
1183 /*
1184 * Iterate over the bps in this ds.
1185 */
1186 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1187 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1188 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1189 rrw_exit(&ds->ds_bp_rwlock, FTAG);
1190
1191 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
1192 dsl_dataset_name(ds, dsname);
1193 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1194 "suspending=%u",
1195 (longlong_t)dsobj, dsname,
1196 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1197 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1198 (int)scn->scn_suspending);
1199 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
1200
1201 if (scn->scn_suspending)
1202 goto out;
1203
1204 /*
1205 * We've finished this pass over this dataset.
1206 */
1207
1208 /*
1209 * If we did not completely visit this dataset, do another pass.
1210 */
1211 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1212 zfs_dbgmsg("incomplete pass; visiting again");
1213 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1214 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1215 scn->scn_phys.scn_queue_obj, ds->ds_object,
1216 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1217 goto out;
1218 }
1219
1220 /*
1221 * Add descendent datasets to work queue.
1222 */
1223 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1224 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1225 scn->scn_phys.scn_queue_obj,
1226 dsl_dataset_phys(ds)->ds_next_snap_obj,
1227 dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1228 }
1229 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1230 boolean_t usenext = B_FALSE;
1231 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1232 uint64_t count;
1233 /*
1234 * A bug in a previous version of the code could
1235 * cause upgrade_clones_cb() to not set
1236 * ds_next_snap_obj when it should, leading to a
1237 * missing entry. Therefore we can only use the
1238 * next_clones_obj when its count is correct.
1239 */
1240 int err = zap_count(dp->dp_meta_objset,
1241 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1242 if (err == 0 &&
1243 count == dsl_dataset_phys(ds)->ds_num_children - 1)
1244 usenext = B_TRUE;
1245 }
1246
1247 if (usenext) {
1248 VERIFY0(zap_join_key(dp->dp_meta_objset,
1249 dsl_dataset_phys(ds)->ds_next_clones_obj,
1250 scn->scn_phys.scn_queue_obj,
1251 dsl_dataset_phys(ds)->ds_creation_txg, tx));
1252 } else {
1253 struct enqueue_clones_arg eca;
1254 eca.tx = tx;
1255 eca.originobj = ds->ds_object;
1256
1257 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1258 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1259 }
1260 }
1261
1262 out:
1263 dsl_dataset_rele(ds, FTAG);
1264 }
1265
1266 /* ARGSUSED */
1267 static int
1268 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1269 {
1270 dmu_tx_t *tx = arg;
1271 dsl_dataset_t *ds;
1272 int err;
1273 dsl_scan_t *scn = dp->dp_scan;
1274
1275 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1276 if (err)
1277 return (err);
1278
1279 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1280 dsl_dataset_t *prev;
1281 err = dsl_dataset_hold_obj(dp,
1282 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1283 if (err) {
1284 dsl_dataset_rele(ds, FTAG);
1285 return (err);
1286 }
1287
1288 /*
1289 * If this is a clone, we don't need to worry about it for now.
1290 */
1291 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1292 dsl_dataset_rele(ds, FTAG);
1293 dsl_dataset_rele(prev, FTAG);
1294 return (0);
1295 }
1296 dsl_dataset_rele(ds, FTAG);
1297 ds = prev;
1298 }
1299
1300 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1301 ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1302 dsl_dataset_rele(ds, FTAG);
1303 return (0);
1304 }
1305
1306 /*
1307 * Scrub/dedup interaction.
1308 *
1309 * If there are N references to a deduped block, we don't want to scrub it
1310 * N times -- ideally, we should scrub it exactly once.
1311 *
1312 * We leverage the fact that the dde's replication class (enum ddt_class)
1313 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1314 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1315 *
1316 * To prevent excess scrubbing, the scrub begins by walking the DDT
1317 * to find all blocks with refcnt > 1, and scrubs each of these once.
1318 * Since there are two replication classes which contain blocks with
1319 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1320 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1321 *
1322 * There would be nothing more to say if a block's refcnt couldn't change
1323 * during a scrub, but of course it can so we must account for changes
1324 * in a block's replication class.
1325 *
1326 * Here's an example of what can occur:
1327 *
1328 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1329 * when visited during the top-down scrub phase, it will be scrubbed twice.
1330 * This negates our scrub optimization, but is otherwise harmless.
1331 *
1332 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1333 * on each visit during the top-down scrub phase, it will never be scrubbed.
1334 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1335 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1336 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1337 * while a scrub is in progress, it scrubs the block right then.
1338 */
1339 static void
1340 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1341 {
1342 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1343 ddt_entry_t dde = { 0 };
1344 int error;
1345 uint64_t n = 0;
1346
1347 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1348 ddt_t *ddt;
1349
1350 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1351 break;
1352 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1353 (longlong_t)ddb->ddb_class,
1354 (longlong_t)ddb->ddb_type,
1355 (longlong_t)ddb->ddb_checksum,
1356 (longlong_t)ddb->ddb_cursor);
1357
1358 /* There should be no pending changes to the dedup table */
1359 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1360 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1361
1362 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1363 n++;
1364
1365 if (dsl_scan_check_suspend(scn, NULL))
1366 break;
1367 }
1368
1369 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
1370 "suspending=%u", (longlong_t)n,
1371 (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
1372
1373 ASSERT(error == 0 || error == ENOENT);
1374 ASSERT(error != ENOENT ||
1375 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1376 }
1377
1378 /* ARGSUSED */
1379 void
1380 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1381 ddt_entry_t *dde, dmu_tx_t *tx)
1382 {
1383 const ddt_key_t *ddk = &dde->dde_key;
1384 ddt_phys_t *ddp = dde->dde_phys;
1385 blkptr_t bp;
1386 zbookmark_phys_t zb = { 0 };
1387
1388 if (scn->scn_phys.scn_state != DSS_SCANNING)
1389 return;
1390
1391 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1392 if (ddp->ddp_phys_birth == 0 ||
1393 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1394 continue;
1395 ddt_bp_create(checksum, ddk, ddp, &bp);
1396
1397 scn->scn_visited_this_txg++;
1398 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1399 }
1400 }
1401
1402 static void
1403 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1404 {
1405 dsl_pool_t *dp = scn->scn_dp;
1406 zap_cursor_t zc;
1407 zap_attribute_t za;
1408
1409 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1410 scn->scn_phys.scn_ddt_class_max) {
1411 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1412 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1413 dsl_scan_ddt(scn, tx);
1414 if (scn->scn_suspending)
1415 return;
1416 }
1417
1418 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1419 /* First do the MOS & ORIGIN */
1420
1421 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1422 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1423 dsl_scan_visit_rootbp(scn, NULL,
1424 &dp->dp_meta_rootbp, tx);
1425 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1426 if (scn->scn_suspending)
1427 return;
1428
1429 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1430 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1431 enqueue_cb, tx, DS_FIND_CHILDREN));
1432 } else {
1433 dsl_scan_visitds(scn,
1434 dp->dp_origin_snap->ds_object, tx);
1435 }
1436 ASSERT(!scn->scn_suspending);
1437 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1438 ZB_DESTROYED_OBJSET) {
1439 /*
1440 * If we were suspended, continue from here. Note if the
1441 * ds we were suspended on was deleted, the zb_objset may
1442 * be -1, so we will skip this and find a new objset
1443 * below.
1444 */
1445 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1446 if (scn->scn_suspending)
1447 return;
1448 }
1449
1450 /*
1451 * In case we were suspended right at the end of the ds, zero the
1452 * bookmark so we don't think that we're still trying to resume.
1453 */
1454 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1455
1456 /* keep pulling things out of the zap-object-as-queue */
1457 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1458 scn->scn_phys.scn_queue_obj),
1459 zap_cursor_retrieve(&zc, &za) == 0) {
1460 dsl_dataset_t *ds;
1461 uint64_t dsobj;
1462
1463 dsobj = zfs_strtonum(za.za_name, NULL);
1464 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1465 scn->scn_phys.scn_queue_obj, dsobj, tx));
1466
1467 /* Set up min/max txg */
1468 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1469 if (za.za_first_integer != 0) {
1470 scn->scn_phys.scn_cur_min_txg =
1471 MAX(scn->scn_phys.scn_min_txg,
1472 za.za_first_integer);
1473 } else {
1474 scn->scn_phys.scn_cur_min_txg =
1475 MAX(scn->scn_phys.scn_min_txg,
1476 dsl_dataset_phys(ds)->ds_prev_snap_txg);
1477 }
1478 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1479 dsl_dataset_rele(ds, FTAG);
1480
1481 dsl_scan_visitds(scn, dsobj, tx);
1482 zap_cursor_fini(&zc);
1483 if (scn->scn_suspending)
1484 return;
1485 }
1486 zap_cursor_fini(&zc);
1487 }
1488
1489 static boolean_t
1490 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
1491 {
1492 uint64_t elapsed_nanosecs;
1493
1494 if (zfs_recover)
1495 return (B_FALSE);
1496
1497 if (scn->scn_visited_this_txg >= zfs_async_block_max_blocks)
1498 return (B_TRUE);
1499
1500 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1501 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1502 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
1503 txg_sync_waiting(scn->scn_dp)) ||
1504 spa_shutting_down(scn->scn_dp->dp_spa));
1505 }
1506
1507 static int
1508 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1509 {
1510 dsl_scan_t *scn = arg;
1511
1512 if (!scn->scn_is_bptree ||
1513 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1514 if (dsl_scan_async_block_should_pause(scn))
1515 return (SET_ERROR(ERESTART));
1516 }
1517
1518 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1519 dmu_tx_get_txg(tx), bp, 0));
1520 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1521 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1522 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1523 scn->scn_visited_this_txg++;
1524 return (0);
1525 }
1526
1527 static int
1528 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1529 {
1530 dsl_scan_t *scn = arg;
1531 const dva_t *dva = &bp->blk_dva[0];
1532
1533 if (dsl_scan_async_block_should_pause(scn))
1534 return (SET_ERROR(ERESTART));
1535
1536 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
1537 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
1538 DVA_GET_ASIZE(dva), tx);
1539 scn->scn_visited_this_txg++;
1540 return (0);
1541 }
1542
1543 boolean_t
1544 dsl_scan_active(dsl_scan_t *scn)
1545 {
1546 spa_t *spa = scn->scn_dp->dp_spa;
1547 uint64_t used = 0, comp, uncomp;
1548
1549 if (spa->spa_load_state != SPA_LOAD_NONE)
1550 return (B_FALSE);
1551 if (spa_shutting_down(spa))
1552 return (B_FALSE);
1553 if ((scn->scn_phys.scn_state == DSS_SCANNING &&
1554 !dsl_scan_is_paused_scrub(scn)) ||
1555 (scn->scn_async_destroying && !scn->scn_async_stalled))
1556 return (B_TRUE);
1557
1558 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1559 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1560 &used, &comp, &uncomp);
1561 }
1562 return (used != 0);
1563 }
1564
1565 /* Called whenever a txg syncs. */
1566 void
1567 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1568 {
1569 dsl_scan_t *scn = dp->dp_scan;
1570 spa_t *spa = dp->dp_spa;
1571 int err = 0;
1572
1573 /*
1574 * Check for scn_restart_txg before checking spa_load_state, so
1575 * that we can restart an old-style scan while the pool is being
1576 * imported (see dsl_scan_init).
1577 */
1578 if (dsl_scan_restarting(scn, tx)) {
1579 pool_scan_func_t func = POOL_SCAN_SCRUB;
1580 dsl_scan_done(scn, B_FALSE, tx);
1581 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1582 func = POOL_SCAN_RESILVER;
1583 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1584 func, tx->tx_txg);
1585 dsl_scan_setup_sync(&func, tx);
1586 }
1587
1588 /*
1589 * Only process scans in sync pass 1.
1590 */
1591 if (spa_sync_pass(dp->dp_spa) > 1)
1592 return;
1593
1594 /*
1595 * If the spa is shutting down, then stop scanning. This will
1596 * ensure that the scan does not dirty any new data during the
1597 * shutdown phase.
1598 */
1599 if (spa_shutting_down(spa))
1600 return;
1601
1602 /*
1603 * If the scan is inactive due to a stalled async destroy, try again.
1604 */
1605 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1606 return;
1607
1608 scn->scn_visited_this_txg = 0;
1609 scn->scn_suspending = B_FALSE;
1610 scn->scn_sync_start_time = gethrtime();
1611 spa->spa_scrub_active = B_TRUE;
1612
1613 /*
1614 * First process the async destroys. If we suspend, don't do
1615 * any scrubbing or resilvering. This ensures that there are no
1616 * async destroys while we are scanning, so the scan code doesn't
1617 * have to worry about traversing it. It is also faster to free the
1618 * blocks than to scrub them.
1619 */
1620 if (zfs_free_bpobj_enabled &&
1621 spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1622 scn->scn_is_bptree = B_FALSE;
1623 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
1624 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1625 NULL, ZIO_FLAG_MUSTSUCCEED);
1626 err = bpobj_iterate(&dp->dp_free_bpobj,
1627 dsl_scan_free_block_cb, scn, tx);
1628 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1629
1630 if (err != 0 && err != ERESTART)
1631 zfs_panic_recover("error %u from bpobj_iterate()", err);
1632 }
1633
1634 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1635 ASSERT(scn->scn_async_destroying);
1636 scn->scn_is_bptree = B_TRUE;
1637 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1638 NULL, ZIO_FLAG_MUSTSUCCEED);
1639 err = bptree_iterate(dp->dp_meta_objset,
1640 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1641 VERIFY0(zio_wait(scn->scn_zio_root));
1642
1643 if (err == EIO || err == ECKSUM) {
1644 err = 0;
1645 } else if (err != 0 && err != ERESTART) {
1646 zfs_panic_recover("error %u from "
1647 "traverse_dataset_destroyed()", err);
1648 }
1649
1650 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1651 /* finished; deactivate async destroy feature */
1652 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1653 ASSERT(!spa_feature_is_active(spa,
1654 SPA_FEATURE_ASYNC_DESTROY));
1655 VERIFY0(zap_remove(dp->dp_meta_objset,
1656 DMU_POOL_DIRECTORY_OBJECT,
1657 DMU_POOL_BPTREE_OBJ, tx));
1658 VERIFY0(bptree_free(dp->dp_meta_objset,
1659 dp->dp_bptree_obj, tx));
1660 dp->dp_bptree_obj = 0;
1661 scn->scn_async_destroying = B_FALSE;
1662 scn->scn_async_stalled = B_FALSE;
1663 } else {
1664 /*
1665 * If we didn't make progress, mark the async
1666 * destroy as stalled, so that we will not initiate
1667 * a spa_sync() on its behalf. Note that we only
1668 * check this if we are not finished, because if the
1669 * bptree had no blocks for us to visit, we can
1670 * finish without "making progress".
1671 */
1672 scn->scn_async_stalled =
1673 (scn->scn_visited_this_txg == 0);
1674 }
1675 }
1676 if (scn->scn_visited_this_txg) {
1677 zfs_dbgmsg("freed %llu blocks in %llums from "
1678 "free_bpobj/bptree txg %llu; err=%u",
1679 (longlong_t)scn->scn_visited_this_txg,
1680 (longlong_t)
1681 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1682 (longlong_t)tx->tx_txg, err);
1683 scn->scn_visited_this_txg = 0;
1684
1685 /*
1686 * Write out changes to the DDT that may be required as a
1687 * result of the blocks freed. This ensures that the DDT
1688 * is clean when a scrub/resilver runs.
1689 */
1690 ddt_sync(spa, tx->tx_txg);
1691 }
1692 if (err != 0)
1693 return;
1694 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1695 zfs_free_leak_on_eio &&
1696 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1697 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1698 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1699 /*
1700 * We have finished background destroying, but there is still
1701 * some space left in the dp_free_dir. Transfer this leaked
1702 * space to the dp_leak_dir.
1703 */
1704 if (dp->dp_leak_dir == NULL) {
1705 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1706 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1707 LEAK_DIR_NAME, tx);
1708 VERIFY0(dsl_pool_open_special_dir(dp,
1709 LEAK_DIR_NAME, &dp->dp_leak_dir));
1710 rrw_exit(&dp->dp_config_rwlock, FTAG);
1711 }
1712 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1713 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1714 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1715 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1716 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1717 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1718 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1719 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1720 }
1721
1722 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1723 /* finished; verify that space accounting went to zero */
1724 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1725 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1726 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1727 }
1728
1729 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
1730 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1731 DMU_POOL_OBSOLETE_BPOBJ));
1732 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
1733 ASSERT(spa_feature_is_active(dp->dp_spa,
1734 SPA_FEATURE_OBSOLETE_COUNTS));
1735
1736 scn->scn_is_bptree = B_FALSE;
1737 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
1738 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
1739 dsl_scan_obsolete_block_cb, scn, tx);
1740 if (err != 0 && err != ERESTART)
1741 zfs_panic_recover("error %u from bpobj_iterate()", err);
1742
1743 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
1744 dsl_pool_destroy_obsolete_bpobj(dp, tx);
1745 }
1746
1747 if (scn->scn_phys.scn_state != DSS_SCANNING)
1748 return;
1749
1750 if (scn->scn_done_txg == tx->tx_txg) {
1751 ASSERT(!scn->scn_suspending);
1752 /* finished with scan. */
1753 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1754 dsl_scan_done(scn, B_TRUE, tx);
1755 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1756 dsl_scan_sync_state(scn, tx);
1757 return;
1758 }
1759
1760 if (dsl_scan_is_paused_scrub(scn))
1761 return;
1762
1763 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1764 scn->scn_phys.scn_ddt_class_max) {
1765 zfs_dbgmsg("doing scan sync txg %llu; "
1766 "ddt bm=%llu/%llu/%llu/%llx",
1767 (longlong_t)tx->tx_txg,
1768 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1769 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1770 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1771 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1772 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1773 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1774 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1775 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1776 } else {
1777 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1778 (longlong_t)tx->tx_txg,
1779 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1780 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1781 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1782 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1783 }
1784
1785 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1786 NULL, ZIO_FLAG_CANFAIL);
1787 dsl_pool_config_enter(dp, FTAG);
1788 dsl_scan_visit(scn, tx);
1789 dsl_pool_config_exit(dp, FTAG);
1790 (void) zio_wait(scn->scn_zio_root);
1791 scn->scn_zio_root = NULL;
1792
1793 zfs_dbgmsg("visited %llu blocks in %llums",
1794 (longlong_t)scn->scn_visited_this_txg,
1795 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1796
1797 if (!scn->scn_suspending) {
1798 scn->scn_done_txg = tx->tx_txg + 1;
1799 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1800 tx->tx_txg, scn->scn_done_txg);
1801 }
1802
1803 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1804 mutex_enter(&spa->spa_scrub_lock);
1805 while (spa->spa_scrub_inflight > 0) {
1806 cv_wait(&spa->spa_scrub_io_cv,
1807 &spa->spa_scrub_lock);
1808 }
1809 mutex_exit(&spa->spa_scrub_lock);
1810 }
1811
1812 dsl_scan_sync_state(scn, tx);
1813 }
1814
1815 /*
1816 * This will start a new scan, or restart an existing one.
1817 */
1818 void
1819 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1820 {
1821 if (txg == 0) {
1822 dmu_tx_t *tx;
1823 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1824 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1825
1826 txg = dmu_tx_get_txg(tx);
1827 dp->dp_scan->scn_restart_txg = txg;
1828 dmu_tx_commit(tx);
1829 } else {
1830 dp->dp_scan->scn_restart_txg = txg;
1831 }
1832 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1833 }
1834
1835 boolean_t
1836 dsl_scan_resilvering(dsl_pool_t *dp)
1837 {
1838 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1839 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1840 }
1841
1842 /*
1843 * scrub consumers
1844 */
1845
1846 static void
1847 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1848 {
1849 int i;
1850
1851 /*
1852 * If we resume after a reboot, zab will be NULL; don't record
1853 * incomplete stats in that case.
1854 */
1855 if (zab == NULL)
1856 return;
1857
1858 for (i = 0; i < 4; i++) {
1859 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1860 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1861 if (t & DMU_OT_NEWTYPE)
1862 t = DMU_OT_OTHER;
1863 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1864 int equal;
1865
1866 zb->zb_count++;
1867 zb->zb_asize += BP_GET_ASIZE(bp);
1868 zb->zb_lsize += BP_GET_LSIZE(bp);
1869 zb->zb_psize += BP_GET_PSIZE(bp);
1870 zb->zb_gangs += BP_COUNT_GANG(bp);
1871
1872 switch (BP_GET_NDVAS(bp)) {
1873 case 2:
1874 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1875 DVA_GET_VDEV(&bp->blk_dva[1]))
1876 zb->zb_ditto_2_of_2_samevdev++;
1877 break;
1878 case 3:
1879 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1880 DVA_GET_VDEV(&bp->blk_dva[1])) +
1881 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1882 DVA_GET_VDEV(&bp->blk_dva[2])) +
1883 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1884 DVA_GET_VDEV(&bp->blk_dva[2]));
1885 if (equal == 1)
1886 zb->zb_ditto_2_of_3_samevdev++;
1887 else if (equal == 3)
1888 zb->zb_ditto_3_of_3_samevdev++;
1889 break;
1890 }
1891 }
1892 }
1893
1894 static void
1895 dsl_scan_scrub_done(zio_t *zio)
1896 {
1897 spa_t *spa = zio->io_spa;
1898
1899 abd_free(zio->io_abd);
1900
1901 mutex_enter(&spa->spa_scrub_lock);
1902 spa->spa_scrub_inflight--;
1903 cv_broadcast(&spa->spa_scrub_io_cv);
1904
1905 if (zio->io_error && (zio->io_error != ECKSUM ||
1906 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1907 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1908 }
1909 mutex_exit(&spa->spa_scrub_lock);
1910 }
1911
1912 static int
1913 dsl_scan_scrub_cb(dsl_pool_t *dp,
1914 const blkptr_t *bp, const zbookmark_phys_t *zb)
1915 {
1916 dsl_scan_t *scn = dp->dp_scan;
1917 size_t size = BP_GET_PSIZE(bp);
1918 spa_t *spa = dp->dp_spa;
1919 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1920 boolean_t needs_io;
1921 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1922 int scan_delay = 0;
1923
1924 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1925 phys_birth >= scn->scn_phys.scn_max_txg)
1926 return (0);
1927
1928 count_block(dp->dp_blkstats, bp);
1929
1930 if (BP_IS_EMBEDDED(bp))
1931 return (0);
1932
1933 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1934 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1935 zio_flags |= ZIO_FLAG_SCRUB;
1936 needs_io = B_TRUE;
1937 scan_delay = zfs_scrub_delay;
1938 } else {
1939 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1940 zio_flags |= ZIO_FLAG_RESILVER;
1941 needs_io = B_FALSE;
1942 scan_delay = zfs_resilver_delay;
1943 }
1944
1945 /* If it's an intent log block, failure is expected. */
1946 if (zb->zb_level == ZB_ZIL_LEVEL)
1947 zio_flags |= ZIO_FLAG_SPECULATIVE;
1948
1949 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1950 vdev_t *vd = vdev_lookup_top(spa,
1951 DVA_GET_VDEV(&bp->blk_dva[d]));
1952
1953 /*
1954 * Keep track of how much data we've examined so that
1955 * zpool(1M) status can make useful progress reports.
1956 */
1957 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1958 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1959
1960 /* if it's a resilver, this may not be in the target range */
1961 if (!needs_io) {
1962 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1963 /*
1964 * Gang members may be spread across multiple
1965 * vdevs, so the best estimate we have is the
1966 * scrub range, which has already been checked.
1967 * XXX -- it would be better to change our
1968 * allocation policy to ensure that all
1969 * gang members reside on the same vdev.
1970 */
1971 needs_io = B_TRUE;
1972 } else {
1973 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1974 phys_birth, 1);
1975 }
1976 }
1977 }
1978
1979 if (needs_io && !zfs_no_scrub_io) {
1980 vdev_t *rvd = spa->spa_root_vdev;
1981 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1982
1983 mutex_enter(&spa->spa_scrub_lock);
1984 while (spa->spa_scrub_inflight >= maxinflight)
1985 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1986 spa->spa_scrub_inflight++;
1987 mutex_exit(&spa->spa_scrub_lock);
1988
1989 /*
1990 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1991 * then throttle our workload to limit the impact of a scan.
1992 */
1993 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1994 delay(scan_delay);
1995
1996 zio_nowait(zio_read(NULL, spa, bp,
1997 abd_alloc_for_io(size, B_FALSE), size, dsl_scan_scrub_done,
1998 NULL, ZIO_PRIORITY_SCRUB, zio_flags, zb));
1999 }
2000
2001 /* do not relocate this block */
2002 return (0);
2003 }
2004
2005 /*
2006 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
2007 * Can also be called to resume a paused scrub.
2008 */
2009 int
2010 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
2011 {
2012 spa_t *spa = dp->dp_spa;
2013 dsl_scan_t *scn = dp->dp_scan;
2014
2015 /*
2016 * Purge all vdev caches and probe all devices. We do this here
2017 * rather than in sync context because this requires a writer lock
2018 * on the spa_config lock, which we can't do from sync context. The
2019 * spa_scrub_reopen flag indicates that vdev_open() should not
2020 * attempt to start another scrub.
2021 */
2022 spa_vdev_state_enter(spa, SCL_NONE);
2023 spa->spa_scrub_reopen = B_TRUE;
2024 vdev_reopen(spa->spa_root_vdev);
2025 spa->spa_scrub_reopen = B_FALSE;
2026 (void) spa_vdev_state_exit(spa, NULL, 0);
2027
2028 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
2029 /* got scrub start cmd, resume paused scrub */
2030 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
2031 POOL_SCRUB_NORMAL);
2032 if (err == 0) {
2033 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
2034 return (ECANCELED);
2035 }
2036
2037 return (SET_ERROR(err));
2038 }
2039
2040 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
2041 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
2042 }
2043
2044 static boolean_t
2045 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
2046 {
2047 return (scn->scn_restart_txg != 0 &&
2048 scn->scn_restart_txg <= tx->tx_txg);
2049 }