1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 */
26
27 #include <sys/dsl_scan.h>
28 #include <sys/dsl_pool.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_prop.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/zap.h>
38 #include <sys/zio.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/vdev_impl.h>
44 #include <sys/zil_impl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/ddt.h>
47 #include <sys/sa.h>
48 #include <sys/sa_impl.h>
49 #include <sys/zfeature.h>
50 #ifdef _KERNEL
51 #include <sys/zfs_vfsops.h>
52 #endif
53
54 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
55 const zbookmark_phys_t *);
56
57 static scan_cb_t dsl_scan_scrub_cb;
58 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
60
61 int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
62 int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
63 int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
64 int zfs_scan_idle = 50; /* idle window in clock ticks */
65
66 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
67 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
68 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
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 */
71 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
72 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
73 /* max number of blocks to free in a single TXG */
74 uint64_t zfs_free_max_blocks = UINT64_MAX;
75
76 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
77 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
78 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
79
80 extern int zfs_txg_timeout;
81
82 /*
83 * Enable/disable the processing of the free_bpobj object.
84 */
85 boolean_t zfs_free_bpobj_enabled = B_TRUE;
86
87 /* the order has to match pool_scan_type */
88 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
89 NULL,
90 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
91 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
92 };
93
94 int
95 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
96 {
97 int err;
98 dsl_scan_t *scn;
99 spa_t *spa = dp->dp_spa;
100 uint64_t f;
101
102 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
103 scn->scn_dp = dp;
104
105 /*
106 * It's possible that we're resuming a scan after a reboot so
107 * make sure that the scan_async_destroying flag is initialized
108 * appropriately.
109 */
110 ASSERT(!scn->scn_async_destroying);
111 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
112 SPA_FEATURE_ASYNC_DESTROY);
113
114 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
115 "scrub_func", sizeof (uint64_t), 1, &f);
116 if (err == 0) {
117 /*
118 * There was an old-style scrub in progress. Restart a
119 * new-style scrub from the beginning.
120 */
121 scn->scn_restart_txg = txg;
122 zfs_dbgmsg("old-style scrub was in progress; "
123 "restarting new-style scrub in txg %llu",
124 scn->scn_restart_txg);
125
126 /*
127 * Load the queue obj from the old location so that it
128 * can be freed by dsl_scan_done().
129 */
130 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
131 "scrub_queue", sizeof (uint64_t), 1,
132 &scn->scn_phys.scn_queue_obj);
133 } else {
134 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
135 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
136 &scn->scn_phys);
137 if (err == ENOENT)
138 return (0);
139 else if (err)
140 return (err);
141
142 if (scn->scn_phys.scn_state == DSS_SCANNING &&
143 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
144 /*
145 * A new-type scrub was in progress on an old
146 * pool, and the pool was accessed by old
147 * software. Restart from the beginning, since
148 * the old software may have changed the pool in
149 * the meantime.
150 */
151 scn->scn_restart_txg = txg;
152 zfs_dbgmsg("new-style scrub was modified "
153 "by old software; restarting in txg %llu",
154 scn->scn_restart_txg);
155 }
156 }
157
158 spa_scan_stat_init(spa);
159 return (0);
160 }
161
162 void
163 dsl_scan_fini(dsl_pool_t *dp)
164 {
165 if (dp->dp_scan) {
166 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
167 dp->dp_scan = NULL;
168 }
169 }
170
171 /* ARGSUSED */
172 static int
173 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
174 {
175 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
176
177 if (scn->scn_phys.scn_state == DSS_SCANNING)
178 return (SET_ERROR(EBUSY));
179
180 return (0);
181 }
182
183 static void
184 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
185 {
186 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
187 pool_scan_func_t *funcp = arg;
188 dmu_object_type_t ot = 0;
189 dsl_pool_t *dp = scn->scn_dp;
190 spa_t *spa = dp->dp_spa;
191
192 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
193 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
194 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
195 scn->scn_phys.scn_func = *funcp;
196 scn->scn_phys.scn_state = DSS_SCANNING;
197 scn->scn_phys.scn_min_txg = 0;
198 scn->scn_phys.scn_max_txg = tx->tx_txg;
199 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
200 scn->scn_phys.scn_start_time = gethrestime_sec();
201 scn->scn_phys.scn_errors = 0;
202 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
203 scn->scn_restart_txg = 0;
204 scn->scn_done_txg = 0;
205 spa_scan_stat_init(spa);
206
207 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
208 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
209
210 /* rewrite all disk labels */
211 vdev_config_dirty(spa->spa_root_vdev);
212
213 if (vdev_resilver_needed(spa->spa_root_vdev,
214 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
215 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
216 } else {
217 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
218 }
219
220 spa->spa_scrub_started = B_TRUE;
221 /*
222 * If this is an incremental scrub, limit the DDT scrub phase
223 * to just the auto-ditto class (for correctness); the rest
224 * of the scrub should go faster using top-down pruning.
225 */
226 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
227 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
228
229 }
230
231 /* back to the generic stuff */
232
233 if (dp->dp_blkstats == NULL) {
234 dp->dp_blkstats =
235 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
236 }
237 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
238
239 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
240 ot = DMU_OT_ZAP_OTHER;
241
242 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
243 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
244
245 dsl_scan_sync_state(scn, tx);
246
247 spa_history_log_internal(spa, "scan setup", tx,
248 "func=%u mintxg=%llu maxtxg=%llu",
249 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
250 }
251
252 /* ARGSUSED */
253 static void
254 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
255 {
256 static const char *old_names[] = {
257 "scrub_bookmark",
258 "scrub_ddt_bookmark",
259 "scrub_ddt_class_max",
260 "scrub_queue",
261 "scrub_min_txg",
262 "scrub_max_txg",
263 "scrub_func",
264 "scrub_errors",
265 NULL
266 };
267
268 dsl_pool_t *dp = scn->scn_dp;
269 spa_t *spa = dp->dp_spa;
270 int i;
271
272 /* Remove any remnants of an old-style scrub. */
273 for (i = 0; old_names[i]; i++) {
274 (void) zap_remove(dp->dp_meta_objset,
275 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
276 }
277
278 if (scn->scn_phys.scn_queue_obj != 0) {
279 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
280 scn->scn_phys.scn_queue_obj, tx));
281 scn->scn_phys.scn_queue_obj = 0;
282 }
283
284 /*
285 * If we were "restarted" from a stopped state, don't bother
286 * with anything else.
287 */
288 if (scn->scn_phys.scn_state != DSS_SCANNING)
289 return;
290
291 if (complete)
292 scn->scn_phys.scn_state = DSS_FINISHED;
293 else
294 scn->scn_phys.scn_state = DSS_CANCELED;
295
296 spa_history_log_internal(spa, "scan done", tx,
297 "complete=%u", complete);
298
299 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
300 mutex_enter(&spa->spa_scrub_lock);
301 while (spa->spa_scrub_inflight > 0) {
302 cv_wait(&spa->spa_scrub_io_cv,
303 &spa->spa_scrub_lock);
304 }
305 mutex_exit(&spa->spa_scrub_lock);
306 spa->spa_scrub_started = B_FALSE;
307 spa->spa_scrub_active = B_FALSE;
308
309 /*
310 * If the scrub/resilver completed, update all DTLs to
311 * reflect this. Whether it succeeded or not, vacate
312 * all temporary scrub DTLs.
313 */
314 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
315 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
316 if (complete) {
317 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
318 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
319 }
320 spa_errlog_rotate(spa);
321
322 /*
323 * We may have finished replacing a device.
324 * Let the async thread assess this and handle the detach.
325 */
326 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
327 }
328
329 scn->scn_phys.scn_end_time = gethrestime_sec();
330 }
331
332 /* ARGSUSED */
333 static int
334 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
335 {
336 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
337
338 if (scn->scn_phys.scn_state != DSS_SCANNING)
339 return (SET_ERROR(ENOENT));
340 return (0);
341 }
342
343 /* ARGSUSED */
344 static void
345 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
346 {
347 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
348
349 dsl_scan_done(scn, B_FALSE, tx);
350 dsl_scan_sync_state(scn, tx);
351 }
352
353 int
354 dsl_scan_cancel(dsl_pool_t *dp)
355 {
356 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
357 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
358 }
359
360 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
361 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
362 dmu_objset_type_t ostype, dmu_tx_t *tx);
363 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
364 dmu_objset_type_t ostype,
365 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
366
367 void
368 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
369 {
370 zio_free(dp->dp_spa, txg, bp);
371 }
372
373 void
374 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
375 {
376 ASSERT(dsl_pool_sync_context(dp));
377 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
378 }
379
380 static uint64_t
381 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
382 {
383 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
384 if (ds->ds_is_snapshot)
385 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
386 return (smt);
387 }
388
389 static void
390 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
391 {
392 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
393 DMU_POOL_DIRECTORY_OBJECT,
394 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
395 &scn->scn_phys, tx));
396 }
397
398 extern int zfs_vdev_async_write_active_min_dirty_percent;
399
400 static boolean_t
401 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
402 {
403 /* we never skip user/group accounting objects */
404 if (zb && (int64_t)zb->zb_object < 0)
405 return (B_FALSE);
406
407 if (scn->scn_pausing)
408 return (B_TRUE); /* we're already pausing */
409
410 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
411 return (B_FALSE); /* we're resuming */
412
413 /* We only know how to resume from level-0 blocks. */
414 if (zb && zb->zb_level != 0)
415 return (B_FALSE);
416
417 /*
418 * We pause if:
419 * - we have scanned for the maximum time: an entire txg
420 * timeout (default 5 sec)
421 * or
422 * - we have scanned for at least the minimum time (default 1 sec
423 * for scrub, 3 sec for resilver), and either we have sufficient
424 * dirty data that we are starting to write more quickly
425 * (default 30%), or someone is explicitly waiting for this txg
426 * to complete.
427 * or
428 * - the spa is shutting down because this pool is being exported
429 * or the machine is rebooting.
430 */
431 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
432 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
433 uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
434 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
435 if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
436 (NSEC2MSEC(elapsed_nanosecs) > mintime &&
437 (txg_sync_waiting(scn->scn_dp) ||
438 dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
439 spa_shutting_down(scn->scn_dp->dp_spa)) {
440 if (zb) {
441 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
442 (longlong_t)zb->zb_objset,
443 (longlong_t)zb->zb_object,
444 (longlong_t)zb->zb_level,
445 (longlong_t)zb->zb_blkid);
446 scn->scn_phys.scn_bookmark = *zb;
447 }
448 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
449 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
450 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
451 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
452 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
453 scn->scn_pausing = B_TRUE;
454 return (B_TRUE);
455 }
456 return (B_FALSE);
457 }
458
459 typedef struct zil_scan_arg {
460 dsl_pool_t *zsa_dp;
461 zil_header_t *zsa_zh;
462 } zil_scan_arg_t;
463
464 /* ARGSUSED */
465 static int
466 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
467 {
468 zil_scan_arg_t *zsa = arg;
469 dsl_pool_t *dp = zsa->zsa_dp;
470 dsl_scan_t *scn = dp->dp_scan;
471 zil_header_t *zh = zsa->zsa_zh;
472 zbookmark_phys_t zb;
473
474 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
475 return (0);
476
477 /*
478 * One block ("stubby") can be allocated a long time ago; we
479 * want to visit that one because it has been allocated
480 * (on-disk) even if it hasn't been claimed (even though for
481 * scrub there's nothing to do to it).
482 */
483 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
484 return (0);
485
486 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
487 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
488
489 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
490 return (0);
491 }
492
493 /* ARGSUSED */
494 static int
495 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
496 {
497 if (lrc->lrc_txtype == TX_WRITE) {
498 zil_scan_arg_t *zsa = arg;
499 dsl_pool_t *dp = zsa->zsa_dp;
500 dsl_scan_t *scn = dp->dp_scan;
501 zil_header_t *zh = zsa->zsa_zh;
502 lr_write_t *lr = (lr_write_t *)lrc;
503 blkptr_t *bp = &lr->lr_blkptr;
504 zbookmark_phys_t zb;
505
506 if (BP_IS_HOLE(bp) ||
507 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
508 return (0);
509
510 /*
511 * birth can be < claim_txg if this record's txg is
512 * already txg sync'ed (but this log block contains
513 * other records that are not synced)
514 */
515 if (claim_txg == 0 || bp->blk_birth < claim_txg)
516 return (0);
517
518 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
519 lr->lr_foid, ZB_ZIL_LEVEL,
520 lr->lr_offset / BP_GET_LSIZE(bp));
521
522 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
523 }
524 return (0);
525 }
526
527 static void
528 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
529 {
530 uint64_t claim_txg = zh->zh_claim_txg;
531 zil_scan_arg_t zsa = { dp, zh };
532 zilog_t *zilog;
533
534 /*
535 * We only want to visit blocks that have been claimed but not yet
536 * replayed (or, in read-only mode, blocks that *would* be claimed).
537 */
538 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
539 return;
540
541 zilog = zil_alloc(dp->dp_meta_objset, zh);
542
543 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
544 claim_txg);
545
546 zil_free(zilog);
547 }
548
549 /* ARGSUSED */
550 static void
551 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
552 uint64_t objset, uint64_t object, uint64_t blkid)
553 {
554 zbookmark_phys_t czb;
555 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
556
557 if (zfs_no_scrub_prefetch)
558 return;
559
560 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
561 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
562 return;
563
564 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
565
566 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
567 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
568 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
569 }
570
571 static boolean_t
572 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
573 const zbookmark_phys_t *zb)
574 {
575 /*
576 * We never skip over user/group accounting objects (obj<0)
577 */
578 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
579 (int64_t)zb->zb_object >= 0) {
580 /*
581 * If we already visited this bp & everything below (in
582 * a prior txg sync), don't bother doing it again.
583 */
584 if (zbookmark_subtree_completed(dnp, zb,
585 &scn->scn_phys.scn_bookmark))
586 return (B_TRUE);
587
588 /*
589 * If we found the block we're trying to resume from, or
590 * we went past it to a different object, zero it out to
591 * indicate that it's OK to start checking for pausing
592 * again.
593 */
594 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
595 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
596 dprintf("resuming at %llx/%llx/%llx/%llx\n",
597 (longlong_t)zb->zb_objset,
598 (longlong_t)zb->zb_object,
599 (longlong_t)zb->zb_level,
600 (longlong_t)zb->zb_blkid);
601 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
602 }
603 }
604 return (B_FALSE);
605 }
606
607 /*
608 * Return nonzero on i/o error.
609 * Return new buf to write out in *bufp.
610 */
611 static int
612 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
613 dnode_phys_t *dnp, const blkptr_t *bp,
614 const zbookmark_phys_t *zb, dmu_tx_t *tx)
615 {
616 dsl_pool_t *dp = scn->scn_dp;
617 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
618 int err;
619
620 if (BP_GET_LEVEL(bp) > 0) {
621 arc_flags_t flags = ARC_FLAG_WAIT;
622 int i;
623 blkptr_t *cbp;
624 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
625 arc_buf_t *buf;
626
627 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
628 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
629 if (err) {
630 scn->scn_phys.scn_errors++;
631 return (err);
632 }
633 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
634 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
635 zb->zb_object, zb->zb_blkid * epb + i);
636 }
637 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
638 zbookmark_phys_t czb;
639
640 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
641 zb->zb_level - 1,
642 zb->zb_blkid * epb + i);
643 dsl_scan_visitbp(cbp, &czb, dnp,
644 ds, scn, ostype, tx);
645 }
646 (void) arc_buf_remove_ref(buf, &buf);
647 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
648 arc_flags_t flags = ARC_FLAG_WAIT;
649 dnode_phys_t *cdnp;
650 int i, j;
651 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
652 arc_buf_t *buf;
653
654 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
655 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
656 if (err) {
657 scn->scn_phys.scn_errors++;
658 return (err);
659 }
660 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
661 for (j = 0; j < cdnp->dn_nblkptr; j++) {
662 blkptr_t *cbp = &cdnp->dn_blkptr[j];
663 dsl_scan_prefetch(scn, buf, cbp,
664 zb->zb_objset, zb->zb_blkid * epb + i, j);
665 }
666 }
667 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
668 dsl_scan_visitdnode(scn, ds, ostype,
669 cdnp, zb->zb_blkid * epb + i, tx);
670 }
671
672 (void) arc_buf_remove_ref(buf, &buf);
673 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
674 arc_flags_t flags = ARC_FLAG_WAIT;
675 objset_phys_t *osp;
676 arc_buf_t *buf;
677
678 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
679 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
680 if (err) {
681 scn->scn_phys.scn_errors++;
682 return (err);
683 }
684
685 osp = buf->b_data;
686
687 dsl_scan_visitdnode(scn, ds, osp->os_type,
688 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
689
690 if (OBJSET_BUF_HAS_USERUSED(buf)) {
691 /*
692 * We also always visit user/group accounting
693 * objects, and never skip them, even if we are
694 * pausing. This is necessary so that the space
695 * deltas from this txg get integrated.
696 */
697 dsl_scan_visitdnode(scn, ds, osp->os_type,
698 &osp->os_groupused_dnode,
699 DMU_GROUPUSED_OBJECT, tx);
700 dsl_scan_visitdnode(scn, ds, osp->os_type,
701 &osp->os_userused_dnode,
702 DMU_USERUSED_OBJECT, tx);
703 }
704 (void) arc_buf_remove_ref(buf, &buf);
705 }
706
707 return (0);
708 }
709
710 static void
711 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
712 dmu_objset_type_t ostype, dnode_phys_t *dnp,
713 uint64_t object, dmu_tx_t *tx)
714 {
715 int j;
716
717 for (j = 0; j < dnp->dn_nblkptr; j++) {
718 zbookmark_phys_t czb;
719
720 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
721 dnp->dn_nlevels - 1, j);
722 dsl_scan_visitbp(&dnp->dn_blkptr[j],
723 &czb, dnp, ds, scn, ostype, tx);
724 }
725
726 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
727 zbookmark_phys_t czb;
728 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
729 0, DMU_SPILL_BLKID);
730 dsl_scan_visitbp(&dnp->dn_spill,
731 &czb, dnp, ds, scn, ostype, tx);
732 }
733 }
734
735 /*
736 * The arguments are in this order because mdb can only print the
737 * first 5; we want them to be useful.
738 */
739 static void
740 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
741 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
742 dmu_objset_type_t ostype, dmu_tx_t *tx)
743 {
744 dsl_pool_t *dp = scn->scn_dp;
745 arc_buf_t *buf = NULL;
746 blkptr_t bp_toread = *bp;
747
748 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
749
750 if (dsl_scan_check_pause(scn, zb))
751 return;
752
753 if (dsl_scan_check_resume(scn, dnp, zb))
754 return;
755
756 if (BP_IS_HOLE(bp))
757 return;
758
759 scn->scn_visited_this_txg++;
760
761 dprintf_bp(bp,
762 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
763 ds, ds ? ds->ds_object : 0,
764 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
765 bp);
766
767 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
768 return;
769
770 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
771 return;
772
773 /*
774 * If dsl_scan_ddt() has aready visited this block, it will have
775 * already done any translations or scrubbing, so don't call the
776 * callback again.
777 */
778 if (ddt_class_contains(dp->dp_spa,
779 scn->scn_phys.scn_ddt_class_max, bp)) {
780 ASSERT(buf == NULL);
781 return;
782 }
783
784 /*
785 * If this block is from the future (after cur_max_txg), then we
786 * are doing this on behalf of a deleted snapshot, and we will
787 * revisit the future block on the next pass of this dataset.
788 * Don't scan it now unless we need to because something
789 * under it was modified.
790 */
791 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
792 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
793 }
794 }
795
796 static void
797 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
798 dmu_tx_t *tx)
799 {
800 zbookmark_phys_t zb;
801
802 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
803 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
804 dsl_scan_visitbp(bp, &zb, NULL,
805 ds, scn, DMU_OST_NONE, tx);
806
807 dprintf_ds(ds, "finished scan%s", "");
808 }
809
810 void
811 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
812 {
813 dsl_pool_t *dp = ds->ds_dir->dd_pool;
814 dsl_scan_t *scn = dp->dp_scan;
815 uint64_t mintxg;
816
817 if (scn->scn_phys.scn_state != DSS_SCANNING)
818 return;
819
820 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
821 if (ds->ds_is_snapshot) {
822 /* Note, scn_cur_{min,max}_txg stays the same. */
823 scn->scn_phys.scn_bookmark.zb_objset =
824 dsl_dataset_phys(ds)->ds_next_snap_obj;
825 zfs_dbgmsg("destroying ds %llu; currently traversing; "
826 "reset zb_objset to %llu",
827 (u_longlong_t)ds->ds_object,
828 (u_longlong_t)dsl_dataset_phys(ds)->
829 ds_next_snap_obj);
830 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
831 } else {
832 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
833 ZB_DESTROYED_OBJSET, 0, 0, 0);
834 zfs_dbgmsg("destroying ds %llu; currently traversing; "
835 "reset bookmark to -1,0,0,0",
836 (u_longlong_t)ds->ds_object);
837 }
838 } else if (zap_lookup_int_key(dp->dp_meta_objset,
839 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
840 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
841 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
842 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
843 if (ds->ds_is_snapshot) {
844 /*
845 * We keep the same mintxg; it could be >
846 * ds_creation_txg if the previous snapshot was
847 * deleted too.
848 */
849 VERIFY(zap_add_int_key(dp->dp_meta_objset,
850 scn->scn_phys.scn_queue_obj,
851 dsl_dataset_phys(ds)->ds_next_snap_obj,
852 mintxg, tx) == 0);
853 zfs_dbgmsg("destroying ds %llu; in queue; "
854 "replacing with %llu",
855 (u_longlong_t)ds->ds_object,
856 (u_longlong_t)dsl_dataset_phys(ds)->
857 ds_next_snap_obj);
858 } else {
859 zfs_dbgmsg("destroying ds %llu; in queue; removing",
860 (u_longlong_t)ds->ds_object);
861 }
862 } else {
863 zfs_dbgmsg("destroying ds %llu; ignoring",
864 (u_longlong_t)ds->ds_object);
865 }
866
867 /*
868 * dsl_scan_sync() should be called after this, and should sync
869 * out our changed state, but just to be safe, do it here.
870 */
871 dsl_scan_sync_state(scn, tx);
872 }
873
874 void
875 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
876 {
877 dsl_pool_t *dp = ds->ds_dir->dd_pool;
878 dsl_scan_t *scn = dp->dp_scan;
879 uint64_t mintxg;
880
881 if (scn->scn_phys.scn_state != DSS_SCANNING)
882 return;
883
884 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
885
886 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
887 scn->scn_phys.scn_bookmark.zb_objset =
888 dsl_dataset_phys(ds)->ds_prev_snap_obj;
889 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
890 "reset zb_objset to %llu",
891 (u_longlong_t)ds->ds_object,
892 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
893 } else if (zap_lookup_int_key(dp->dp_meta_objset,
894 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
895 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
896 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
897 VERIFY(zap_add_int_key(dp->dp_meta_objset,
898 scn->scn_phys.scn_queue_obj,
899 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
900 zfs_dbgmsg("snapshotting ds %llu; in queue; "
901 "replacing with %llu",
902 (u_longlong_t)ds->ds_object,
903 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
904 }
905 dsl_scan_sync_state(scn, tx);
906 }
907
908 void
909 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
910 {
911 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
912 dsl_scan_t *scn = dp->dp_scan;
913 uint64_t mintxg;
914
915 if (scn->scn_phys.scn_state != DSS_SCANNING)
916 return;
917
918 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
919 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
920 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
921 "reset zb_objset to %llu",
922 (u_longlong_t)ds1->ds_object,
923 (u_longlong_t)ds2->ds_object);
924 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
925 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
926 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
927 "reset zb_objset to %llu",
928 (u_longlong_t)ds2->ds_object,
929 (u_longlong_t)ds1->ds_object);
930 }
931
932 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
933 ds1->ds_object, &mintxg) == 0) {
934 int err;
935
936 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
937 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
938 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
939 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
940 err = zap_add_int_key(dp->dp_meta_objset,
941 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
942 VERIFY(err == 0 || err == EEXIST);
943 if (err == EEXIST) {
944 /* Both were there to begin with */
945 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
946 scn->scn_phys.scn_queue_obj,
947 ds1->ds_object, mintxg, tx));
948 }
949 zfs_dbgmsg("clone_swap ds %llu; in queue; "
950 "replacing with %llu",
951 (u_longlong_t)ds1->ds_object,
952 (u_longlong_t)ds2->ds_object);
953 } else if (zap_lookup_int_key(dp->dp_meta_objset,
954 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
955 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
956 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
957 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
958 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
959 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
960 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
961 zfs_dbgmsg("clone_swap ds %llu; in queue; "
962 "replacing with %llu",
963 (u_longlong_t)ds2->ds_object,
964 (u_longlong_t)ds1->ds_object);
965 }
966
967 dsl_scan_sync_state(scn, tx);
968 }
969
970 struct enqueue_clones_arg {
971 dmu_tx_t *tx;
972 uint64_t originobj;
973 };
974
975 /* ARGSUSED */
976 static int
977 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
978 {
979 struct enqueue_clones_arg *eca = arg;
980 dsl_dataset_t *ds;
981 int err;
982 dsl_scan_t *scn = dp->dp_scan;
983
984 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
985 return (0);
986
987 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
988 if (err)
989 return (err);
990
991 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
992 dsl_dataset_t *prev;
993 err = dsl_dataset_hold_obj(dp,
994 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
995
996 dsl_dataset_rele(ds, FTAG);
997 if (err)
998 return (err);
999 ds = prev;
1000 }
1001 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1002 scn->scn_phys.scn_queue_obj, ds->ds_object,
1003 dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1004 dsl_dataset_rele(ds, FTAG);
1005 return (0);
1006 }
1007
1008 static void
1009 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1010 {
1011 dsl_pool_t *dp = scn->scn_dp;
1012 dsl_dataset_t *ds;
1013 objset_t *os;
1014
1015 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1016
1017 if (dmu_objset_from_ds(ds, &os))
1018 goto out;
1019
1020 /*
1021 * Only the ZIL in the head (non-snapshot) is valid. Even though
1022 * snapshots can have ZIL block pointers (which may be the same
1023 * BP as in the head), they must be ignored. So we traverse the
1024 * ZIL here, rather than in scan_recurse(), because the regular
1025 * snapshot block-sharing rules don't apply to it.
1026 */
1027 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1028 dsl_scan_zil(dp, &os->os_zil_header);
1029
1030 /*
1031 * Iterate over the bps in this ds.
1032 */
1033 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1034 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1035
1036 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1037 dsl_dataset_name(ds, dsname);
1038 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1039 "pausing=%u",
1040 (longlong_t)dsobj, dsname,
1041 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1042 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1043 (int)scn->scn_pausing);
1044 kmem_free(dsname, ZFS_MAXNAMELEN);
1045
1046 if (scn->scn_pausing)
1047 goto out;
1048
1049 /*
1050 * We've finished this pass over this dataset.
1051 */
1052
1053 /*
1054 * If we did not completely visit this dataset, do another pass.
1055 */
1056 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1057 zfs_dbgmsg("incomplete pass; visiting again");
1058 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1059 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1060 scn->scn_phys.scn_queue_obj, ds->ds_object,
1061 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1062 goto out;
1063 }
1064
1065 /*
1066 * Add descendent datasets to work queue.
1067 */
1068 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1069 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1070 scn->scn_phys.scn_queue_obj,
1071 dsl_dataset_phys(ds)->ds_next_snap_obj,
1072 dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1073 }
1074 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1075 boolean_t usenext = B_FALSE;
1076 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1077 uint64_t count;
1078 /*
1079 * A bug in a previous version of the code could
1080 * cause upgrade_clones_cb() to not set
1081 * ds_next_snap_obj when it should, leading to a
1082 * missing entry. Therefore we can only use the
1083 * next_clones_obj when its count is correct.
1084 */
1085 int err = zap_count(dp->dp_meta_objset,
1086 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1087 if (err == 0 &&
1088 count == dsl_dataset_phys(ds)->ds_num_children - 1)
1089 usenext = B_TRUE;
1090 }
1091
1092 if (usenext) {
1093 VERIFY0(zap_join_key(dp->dp_meta_objset,
1094 dsl_dataset_phys(ds)->ds_next_clones_obj,
1095 scn->scn_phys.scn_queue_obj,
1096 dsl_dataset_phys(ds)->ds_creation_txg, tx));
1097 } else {
1098 struct enqueue_clones_arg eca;
1099 eca.tx = tx;
1100 eca.originobj = ds->ds_object;
1101
1102 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1103 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1104 }
1105 }
1106
1107 out:
1108 dsl_dataset_rele(ds, FTAG);
1109 }
1110
1111 /* ARGSUSED */
1112 static int
1113 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1114 {
1115 dmu_tx_t *tx = arg;
1116 dsl_dataset_t *ds;
1117 int err;
1118 dsl_scan_t *scn = dp->dp_scan;
1119
1120 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1121 if (err)
1122 return (err);
1123
1124 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1125 dsl_dataset_t *prev;
1126 err = dsl_dataset_hold_obj(dp,
1127 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1128 if (err) {
1129 dsl_dataset_rele(ds, FTAG);
1130 return (err);
1131 }
1132
1133 /*
1134 * If this is a clone, we don't need to worry about it for now.
1135 */
1136 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1137 dsl_dataset_rele(ds, FTAG);
1138 dsl_dataset_rele(prev, FTAG);
1139 return (0);
1140 }
1141 dsl_dataset_rele(ds, FTAG);
1142 ds = prev;
1143 }
1144
1145 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1146 ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1147 dsl_dataset_rele(ds, FTAG);
1148 return (0);
1149 }
1150
1151 /*
1152 * Scrub/dedup interaction.
1153 *
1154 * If there are N references to a deduped block, we don't want to scrub it
1155 * N times -- ideally, we should scrub it exactly once.
1156 *
1157 * We leverage the fact that the dde's replication class (enum ddt_class)
1158 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1159 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1160 *
1161 * To prevent excess scrubbing, the scrub begins by walking the DDT
1162 * to find all blocks with refcnt > 1, and scrubs each of these once.
1163 * Since there are two replication classes which contain blocks with
1164 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1165 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1166 *
1167 * There would be nothing more to say if a block's refcnt couldn't change
1168 * during a scrub, but of course it can so we must account for changes
1169 * in a block's replication class.
1170 *
1171 * Here's an example of what can occur:
1172 *
1173 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1174 * when visited during the top-down scrub phase, it will be scrubbed twice.
1175 * This negates our scrub optimization, but is otherwise harmless.
1176 *
1177 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1178 * on each visit during the top-down scrub phase, it will never be scrubbed.
1179 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1180 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1181 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1182 * while a scrub is in progress, it scrubs the block right then.
1183 */
1184 static void
1185 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1186 {
1187 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1188 ddt_entry_t dde = { 0 };
1189 int error;
1190 uint64_t n = 0;
1191
1192 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1193 ddt_t *ddt;
1194
1195 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1196 break;
1197 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1198 (longlong_t)ddb->ddb_class,
1199 (longlong_t)ddb->ddb_type,
1200 (longlong_t)ddb->ddb_checksum,
1201 (longlong_t)ddb->ddb_cursor);
1202
1203 /* There should be no pending changes to the dedup table */
1204 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1205 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1206
1207 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1208 n++;
1209
1210 if (dsl_scan_check_pause(scn, NULL))
1211 break;
1212 }
1213
1214 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1215 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1216 (int)scn->scn_pausing);
1217
1218 ASSERT(error == 0 || error == ENOENT);
1219 ASSERT(error != ENOENT ||
1220 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1221 }
1222
1223 /* ARGSUSED */
1224 void
1225 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1226 ddt_entry_t *dde, dmu_tx_t *tx)
1227 {
1228 const ddt_key_t *ddk = &dde->dde_key;
1229 ddt_phys_t *ddp = dde->dde_phys;
1230 blkptr_t bp;
1231 zbookmark_phys_t zb = { 0 };
1232
1233 if (scn->scn_phys.scn_state != DSS_SCANNING)
1234 return;
1235
1236 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1237 if (ddp->ddp_phys_birth == 0 ||
1238 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1239 continue;
1240 ddt_bp_create(checksum, ddk, ddp, &bp);
1241
1242 scn->scn_visited_this_txg++;
1243 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1244 }
1245 }
1246
1247 static void
1248 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1249 {
1250 dsl_pool_t *dp = scn->scn_dp;
1251 zap_cursor_t zc;
1252 zap_attribute_t za;
1253
1254 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1255 scn->scn_phys.scn_ddt_class_max) {
1256 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1257 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1258 dsl_scan_ddt(scn, tx);
1259 if (scn->scn_pausing)
1260 return;
1261 }
1262
1263 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1264 /* First do the MOS & ORIGIN */
1265
1266 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1267 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1268 dsl_scan_visit_rootbp(scn, NULL,
1269 &dp->dp_meta_rootbp, tx);
1270 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1271 if (scn->scn_pausing)
1272 return;
1273
1274 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1275 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1276 enqueue_cb, tx, DS_FIND_CHILDREN));
1277 } else {
1278 dsl_scan_visitds(scn,
1279 dp->dp_origin_snap->ds_object, tx);
1280 }
1281 ASSERT(!scn->scn_pausing);
1282 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1283 ZB_DESTROYED_OBJSET) {
1284 /*
1285 * If we were paused, continue from here. Note if the
1286 * ds we were paused on was deleted, the zb_objset may
1287 * be -1, so we will skip this and find a new objset
1288 * below.
1289 */
1290 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1291 if (scn->scn_pausing)
1292 return;
1293 }
1294
1295 /*
1296 * In case we were paused right at the end of the ds, zero the
1297 * bookmark so we don't think that we're still trying to resume.
1298 */
1299 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1300
1301 /* keep pulling things out of the zap-object-as-queue */
1302 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1303 scn->scn_phys.scn_queue_obj),
1304 zap_cursor_retrieve(&zc, &za) == 0) {
1305 dsl_dataset_t *ds;
1306 uint64_t dsobj;
1307
1308 dsobj = strtonum(za.za_name, NULL);
1309 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1310 scn->scn_phys.scn_queue_obj, dsobj, tx));
1311
1312 /* Set up min/max txg */
1313 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1314 if (za.za_first_integer != 0) {
1315 scn->scn_phys.scn_cur_min_txg =
1316 MAX(scn->scn_phys.scn_min_txg,
1317 za.za_first_integer);
1318 } else {
1319 scn->scn_phys.scn_cur_min_txg =
1320 MAX(scn->scn_phys.scn_min_txg,
1321 dsl_dataset_phys(ds)->ds_prev_snap_txg);
1322 }
1323 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1324 dsl_dataset_rele(ds, FTAG);
1325
1326 dsl_scan_visitds(scn, dsobj, tx);
1327 zap_cursor_fini(&zc);
1328 if (scn->scn_pausing)
1329 return;
1330 }
1331 zap_cursor_fini(&zc);
1332 }
1333
1334 static boolean_t
1335 dsl_scan_free_should_pause(dsl_scan_t *scn)
1336 {
1337 uint64_t elapsed_nanosecs;
1338
1339 if (zfs_recover)
1340 return (B_FALSE);
1341
1342 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1343 return (B_TRUE);
1344
1345 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1346 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1347 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1348 txg_sync_waiting(scn->scn_dp)) ||
1349 spa_shutting_down(scn->scn_dp->dp_spa));
1350 }
1351
1352 static int
1353 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1354 {
1355 dsl_scan_t *scn = arg;
1356
1357 if (!scn->scn_is_bptree ||
1358 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1359 if (dsl_scan_free_should_pause(scn))
1360 return (SET_ERROR(ERESTART));
1361 }
1362
1363 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1364 dmu_tx_get_txg(tx), bp, 0));
1365 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1366 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1367 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1368 scn->scn_visited_this_txg++;
1369 return (0);
1370 }
1371
1372 boolean_t
1373 dsl_scan_active(dsl_scan_t *scn)
1374 {
1375 spa_t *spa = scn->scn_dp->dp_spa;
1376 uint64_t used = 0, comp, uncomp;
1377
1378 if (spa->spa_load_state != SPA_LOAD_NONE)
1379 return (B_FALSE);
1380 if (spa_shutting_down(spa))
1381 return (B_FALSE);
1382 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1383 (scn->scn_async_destroying && !scn->scn_async_stalled))
1384 return (B_TRUE);
1385
1386 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1387 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1388 &used, &comp, &uncomp);
1389 }
1390 return (used != 0);
1391 }
1392
1393 void
1394 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1395 {
1396 dsl_scan_t *scn = dp->dp_scan;
1397 spa_t *spa = dp->dp_spa;
1398 int err = 0;
1399
1400 /*
1401 * Check for scn_restart_txg before checking spa_load_state, so
1402 * that we can restart an old-style scan while the pool is being
1403 * imported (see dsl_scan_init).
1404 */
1405 if (scn->scn_restart_txg != 0 &&
1406 scn->scn_restart_txg <= tx->tx_txg) {
1407 pool_scan_func_t func = POOL_SCAN_SCRUB;
1408 dsl_scan_done(scn, B_FALSE, tx);
1409 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1410 func = POOL_SCAN_RESILVER;
1411 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1412 func, tx->tx_txg);
1413 dsl_scan_setup_sync(&func, tx);
1414 }
1415
1416 /*
1417 * If the scan is inactive due to a stalled async destroy, try again.
1418 */
1419 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1420 spa_sync_pass(dp->dp_spa) > 1)
1421 return;
1422
1423 scn->scn_visited_this_txg = 0;
1424 scn->scn_pausing = B_FALSE;
1425 scn->scn_sync_start_time = gethrtime();
1426 spa->spa_scrub_active = B_TRUE;
1427
1428 /*
1429 * First process the async destroys. If we pause, don't do
1430 * any scrubbing or resilvering. This ensures that there are no
1431 * async destroys while we are scanning, so the scan code doesn't
1432 * have to worry about traversing it. It is also faster to free the
1433 * blocks than to scrub them.
1434 */
1435 if (zfs_free_bpobj_enabled &&
1436 spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1437 scn->scn_is_bptree = B_FALSE;
1438 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1439 NULL, ZIO_FLAG_MUSTSUCCEED);
1440 err = bpobj_iterate(&dp->dp_free_bpobj,
1441 dsl_scan_free_block_cb, scn, tx);
1442 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1443
1444 if (err != 0 && err != ERESTART)
1445 zfs_panic_recover("error %u from bpobj_iterate()", err);
1446 }
1447
1448 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1449 ASSERT(scn->scn_async_destroying);
1450 scn->scn_is_bptree = B_TRUE;
1451 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1452 NULL, ZIO_FLAG_MUSTSUCCEED);
1453 err = bptree_iterate(dp->dp_meta_objset,
1454 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1455 VERIFY0(zio_wait(scn->scn_zio_root));
1456
1457 if (err == EIO || err == ECKSUM) {
1458 err = 0;
1459 } else if (err != 0 && err != ERESTART) {
1460 zfs_panic_recover("error %u from "
1461 "traverse_dataset_destroyed()", err);
1462 }
1463
1464 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1465 /* finished; deactivate async destroy feature */
1466 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1467 ASSERT(!spa_feature_is_active(spa,
1468 SPA_FEATURE_ASYNC_DESTROY));
1469 VERIFY0(zap_remove(dp->dp_meta_objset,
1470 DMU_POOL_DIRECTORY_OBJECT,
1471 DMU_POOL_BPTREE_OBJ, tx));
1472 VERIFY0(bptree_free(dp->dp_meta_objset,
1473 dp->dp_bptree_obj, tx));
1474 dp->dp_bptree_obj = 0;
1475 scn->scn_async_destroying = B_FALSE;
1476 scn->scn_async_stalled = B_FALSE;
1477 } else {
1478 /*
1479 * If we didn't make progress, mark the async
1480 * destroy as stalled, so that we will not initiate
1481 * a spa_sync() on its behalf. Note that we only
1482 * check this if we are not finished, because if the
1483 * bptree had no blocks for us to visit, we can
1484 * finish without "making progress".
1485 */
1486 scn->scn_async_stalled =
1487 (scn->scn_visited_this_txg == 0);
1488 }
1489 }
1490 if (scn->scn_visited_this_txg) {
1491 zfs_dbgmsg("freed %llu blocks in %llums from "
1492 "free_bpobj/bptree txg %llu; err=%u",
1493 (longlong_t)scn->scn_visited_this_txg,
1494 (longlong_t)
1495 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1496 (longlong_t)tx->tx_txg, err);
1497 scn->scn_visited_this_txg = 0;
1498
1499 /*
1500 * Write out changes to the DDT that may be required as a
1501 * result of the blocks freed. This ensures that the DDT
1502 * is clean when a scrub/resilver runs.
1503 */
1504 ddt_sync(spa, tx->tx_txg);
1505 }
1506 if (err != 0)
1507 return;
1508 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1509 zfs_free_leak_on_eio &&
1510 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1511 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1512 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1513 /*
1514 * We have finished background destroying, but there is still
1515 * some space left in the dp_free_dir. Transfer this leaked
1516 * space to the dp_leak_dir.
1517 */
1518 if (dp->dp_leak_dir == NULL) {
1519 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1520 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1521 LEAK_DIR_NAME, tx);
1522 VERIFY0(dsl_pool_open_special_dir(dp,
1523 LEAK_DIR_NAME, &dp->dp_leak_dir));
1524 rrw_exit(&dp->dp_config_rwlock, FTAG);
1525 }
1526 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1527 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1528 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1529 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1530 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1531 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1532 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1533 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1534 }
1535 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1536 /* finished; verify that space accounting went to zero */
1537 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1538 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1539 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1540 }
1541
1542 if (scn->scn_phys.scn_state != DSS_SCANNING)
1543 return;
1544
1545 if (scn->scn_done_txg == tx->tx_txg) {
1546 ASSERT(!scn->scn_pausing);
1547 /* finished with scan. */
1548 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1549 dsl_scan_done(scn, B_TRUE, tx);
1550 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1551 dsl_scan_sync_state(scn, tx);
1552 return;
1553 }
1554
1555 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1556 scn->scn_phys.scn_ddt_class_max) {
1557 zfs_dbgmsg("doing scan sync txg %llu; "
1558 "ddt bm=%llu/%llu/%llu/%llx",
1559 (longlong_t)tx->tx_txg,
1560 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1561 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1562 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1563 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1564 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1565 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1566 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1567 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1568 } else {
1569 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1570 (longlong_t)tx->tx_txg,
1571 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1572 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1573 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1574 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1575 }
1576
1577 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1578 NULL, ZIO_FLAG_CANFAIL);
1579 dsl_pool_config_enter(dp, FTAG);
1580 dsl_scan_visit(scn, tx);
1581 dsl_pool_config_exit(dp, FTAG);
1582 (void) zio_wait(scn->scn_zio_root);
1583 scn->scn_zio_root = NULL;
1584
1585 zfs_dbgmsg("visited %llu blocks in %llums",
1586 (longlong_t)scn->scn_visited_this_txg,
1587 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1588
1589 if (!scn->scn_pausing) {
1590 scn->scn_done_txg = tx->tx_txg + 1;
1591 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1592 tx->tx_txg, scn->scn_done_txg);
1593 }
1594
1595 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1596 mutex_enter(&spa->spa_scrub_lock);
1597 while (spa->spa_scrub_inflight > 0) {
1598 cv_wait(&spa->spa_scrub_io_cv,
1599 &spa->spa_scrub_lock);
1600 }
1601 mutex_exit(&spa->spa_scrub_lock);
1602 }
1603
1604 dsl_scan_sync_state(scn, tx);
1605 }
1606
1607 /*
1608 * This will start a new scan, or restart an existing one.
1609 */
1610 void
1611 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1612 {
1613 if (txg == 0) {
1614 dmu_tx_t *tx;
1615 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1616 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1617
1618 txg = dmu_tx_get_txg(tx);
1619 dp->dp_scan->scn_restart_txg = txg;
1620 dmu_tx_commit(tx);
1621 } else {
1622 dp->dp_scan->scn_restart_txg = txg;
1623 }
1624 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1625 }
1626
1627 boolean_t
1628 dsl_scan_resilvering(dsl_pool_t *dp)
1629 {
1630 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1631 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1632 }
1633
1634 /*
1635 * scrub consumers
1636 */
1637
1638 static void
1639 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1640 {
1641 int i;
1642
1643 /*
1644 * If we resume after a reboot, zab will be NULL; don't record
1645 * incomplete stats in that case.
1646 */
1647 if (zab == NULL)
1648 return;
1649
1650 for (i = 0; i < 4; i++) {
1651 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1652 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1653 if (t & DMU_OT_NEWTYPE)
1654 t = DMU_OT_OTHER;
1655 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1656 int equal;
1657
1658 zb->zb_count++;
1659 zb->zb_asize += BP_GET_ASIZE(bp);
1660 zb->zb_lsize += BP_GET_LSIZE(bp);
1661 zb->zb_psize += BP_GET_PSIZE(bp);
1662 zb->zb_gangs += BP_COUNT_GANG(bp);
1663
1664 switch (BP_GET_NDVAS(bp)) {
1665 case 2:
1666 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1667 DVA_GET_VDEV(&bp->blk_dva[1]))
1668 zb->zb_ditto_2_of_2_samevdev++;
1669 break;
1670 case 3:
1671 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1672 DVA_GET_VDEV(&bp->blk_dva[1])) +
1673 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1674 DVA_GET_VDEV(&bp->blk_dva[2])) +
1675 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1676 DVA_GET_VDEV(&bp->blk_dva[2]));
1677 if (equal == 1)
1678 zb->zb_ditto_2_of_3_samevdev++;
1679 else if (equal == 3)
1680 zb->zb_ditto_3_of_3_samevdev++;
1681 break;
1682 }
1683 }
1684 }
1685
1686 static void
1687 dsl_scan_scrub_done(zio_t *zio)
1688 {
1689 spa_t *spa = zio->io_spa;
1690
1691 zio_data_buf_free(zio->io_data, zio->io_size);
1692
1693 mutex_enter(&spa->spa_scrub_lock);
1694 spa->spa_scrub_inflight--;
1695 cv_broadcast(&spa->spa_scrub_io_cv);
1696
1697 if (zio->io_error && (zio->io_error != ECKSUM ||
1698 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1699 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1700 }
1701 mutex_exit(&spa->spa_scrub_lock);
1702 }
1703
1704 static int
1705 dsl_scan_scrub_cb(dsl_pool_t *dp,
1706 const blkptr_t *bp, const zbookmark_phys_t *zb)
1707 {
1708 dsl_scan_t *scn = dp->dp_scan;
1709 size_t size = BP_GET_PSIZE(bp);
1710 spa_t *spa = dp->dp_spa;
1711 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1712 boolean_t needs_io;
1713 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1714 int scan_delay = 0;
1715
1716 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1717 phys_birth >= scn->scn_phys.scn_max_txg)
1718 return (0);
1719
1720 count_block(dp->dp_blkstats, bp);
1721
1722 if (BP_IS_EMBEDDED(bp))
1723 return (0);
1724
1725 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1726 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1727 zio_flags |= ZIO_FLAG_SCRUB;
1728 needs_io = B_TRUE;
1729 scan_delay = zfs_scrub_delay;
1730 } else {
1731 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1732 zio_flags |= ZIO_FLAG_RESILVER;
1733 needs_io = B_FALSE;
1734 scan_delay = zfs_resilver_delay;
1735 }
1736
1737 /* If it's an intent log block, failure is expected. */
1738 if (zb->zb_level == ZB_ZIL_LEVEL)
1739 zio_flags |= ZIO_FLAG_SPECULATIVE;
1740
1741 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1742 vdev_t *vd = vdev_lookup_top(spa,
1743 DVA_GET_VDEV(&bp->blk_dva[d]));
1744
1745 /*
1746 * Keep track of how much data we've examined so that
1747 * zpool(1M) status can make useful progress reports.
1748 */
1749 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1750 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1751
1752 /* if it's a resilver, this may not be in the target range */
1753 if (!needs_io) {
1754 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1755 /*
1756 * Gang members may be spread across multiple
1757 * vdevs, so the best estimate we have is the
1758 * scrub range, which has already been checked.
1759 * XXX -- it would be better to change our
1760 * allocation policy to ensure that all
1761 * gang members reside on the same vdev.
1762 */
1763 needs_io = B_TRUE;
1764 } else {
1765 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1766 phys_birth, 1);
1767 }
1768 }
1769 }
1770
1771 if (needs_io && !zfs_no_scrub_io) {
1772 vdev_t *rvd = spa->spa_root_vdev;
1773 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1774 void *data = zio_data_buf_alloc(size);
1775
1776 mutex_enter(&spa->spa_scrub_lock);
1777 while (spa->spa_scrub_inflight >= maxinflight)
1778 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1779 spa->spa_scrub_inflight++;
1780 mutex_exit(&spa->spa_scrub_lock);
1781
1782 /*
1783 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1784 * then throttle our workload to limit the impact of a scan.
1785 */
1786 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1787 delay(scan_delay);
1788
1789 zio_nowait(zio_read(NULL, spa, bp, data, size,
1790 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1791 zio_flags, zb));
1792 }
1793
1794 /* do not relocate this block */
1795 return (0);
1796 }
1797
1798 int
1799 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1800 {
1801 spa_t *spa = dp->dp_spa;
1802
1803 /*
1804 * Purge all vdev caches and probe all devices. We do this here
1805 * rather than in sync context because this requires a writer lock
1806 * on the spa_config lock, which we can't do from sync context. The
1807 * spa_scrub_reopen flag indicates that vdev_open() should not
1808 * attempt to start another scrub.
1809 */
1810 spa_vdev_state_enter(spa, SCL_NONE);
1811 spa->spa_scrub_reopen = B_TRUE;
1812 vdev_reopen(spa->spa_root_vdev);
1813 spa->spa_scrub_reopen = B_FALSE;
1814 (void) spa_vdev_state_exit(spa, NULL, 0);
1815
1816 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1817 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1818 }