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