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
dsl_scan_init(dsl_pool_t * dp,uint64_t txg)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
dsl_scan_fini(dsl_pool_t * dp)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
dsl_scan_setup_check(void * arg,dmu_tx_t * tx)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
dsl_scan_setup_sync(void * arg,dmu_tx_t * tx)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
dsl_scan_done(dsl_scan_t * scn,boolean_t complete,dmu_tx_t * tx)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
dsl_scan_cancel_check(void * arg,dmu_tx_t * tx)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
dsl_scan_cancel_sync(void * arg,dmu_tx_t * tx)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
dsl_scan_cancel(dsl_pool_t * dp)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
dsl_free(dsl_pool_t * dp,uint64_t txg,const blkptr_t * bp)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
dsl_free_sync(zio_t * pio,dsl_pool_t * dp,uint64_t txg,const blkptr_t * bpp)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
dsl_scan_ds_maxtxg(dsl_dataset_t * ds)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
dsl_scan_sync_state(dsl_scan_t * scn,dmu_tx_t * tx)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
dsl_scan_check_pause(dsl_scan_t * scn,const zbookmark_phys_t * zb)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
dsl_scan_zil_block(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)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
dsl_scan_zil_record(zilog_t * zilog,lr_t * lrc,void * arg,uint64_t claim_txg)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
dsl_scan_zil(dsl_pool_t * dp,zil_header_t * zh)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
dsl_scan_prefetch(dsl_scan_t * scn,arc_buf_t * buf,blkptr_t * bp,uint64_t objset,uint64_t object,uint64_t blkid)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
dsl_scan_check_resume(dsl_scan_t * scn,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)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
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)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
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)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
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)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
dsl_scan_visit_rootbp(dsl_scan_t * scn,dsl_dataset_t * ds,blkptr_t * bp,dmu_tx_t * tx)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
dsl_scan_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)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 /*
823 * Note:
824 * - scn_cur_{min,max}_txg stays the same.
825 * - Setting the flag is not really necessary if
826 * scn_cur_max_txg == scn_max_txg, because there
827 * is nothing after this snapshot that we care
828 * about. However, we set it anyway and then
829 * ignore it when we retraverse it in
830 * dsl_scan_visitds().
831 */
832 scn->scn_phys.scn_bookmark.zb_objset =
833 dsl_dataset_phys(ds)->ds_next_snap_obj;
834 zfs_dbgmsg("destroying ds %llu; currently traversing; "
835 "reset zb_objset to %llu",
836 (u_longlong_t)ds->ds_object,
837 (u_longlong_t)dsl_dataset_phys(ds)->
838 ds_next_snap_obj);
839 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
840 } else {
841 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
842 ZB_DESTROYED_OBJSET, 0, 0, 0);
843 zfs_dbgmsg("destroying ds %llu; currently traversing; "
844 "reset bookmark to -1,0,0,0",
845 (u_longlong_t)ds->ds_object);
846 }
847 } else if (zap_lookup_int_key(dp->dp_meta_objset,
848 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
849 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
850 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
851 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
852 if (ds->ds_is_snapshot) {
853 /*
854 * We keep the same mintxg; it could be >
855 * ds_creation_txg if the previous snapshot was
856 * deleted too.
857 */
858 VERIFY(zap_add_int_key(dp->dp_meta_objset,
859 scn->scn_phys.scn_queue_obj,
860 dsl_dataset_phys(ds)->ds_next_snap_obj,
861 mintxg, tx) == 0);
862 zfs_dbgmsg("destroying ds %llu; in queue; "
863 "replacing with %llu",
864 (u_longlong_t)ds->ds_object,
865 (u_longlong_t)dsl_dataset_phys(ds)->
866 ds_next_snap_obj);
867 } else {
868 zfs_dbgmsg("destroying ds %llu; in queue; removing",
869 (u_longlong_t)ds->ds_object);
870 }
871 }
872
873 /*
874 * dsl_scan_sync() should be called after this, and should sync
875 * out our changed state, but just to be safe, do it here.
876 */
877 dsl_scan_sync_state(scn, tx);
878 }
879
880 void
dsl_scan_ds_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)881 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
882 {
883 dsl_pool_t *dp = ds->ds_dir->dd_pool;
884 dsl_scan_t *scn = dp->dp_scan;
885 uint64_t mintxg;
886
887 if (scn->scn_phys.scn_state != DSS_SCANNING)
888 return;
889
890 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
891
892 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
893 scn->scn_phys.scn_bookmark.zb_objset =
894 dsl_dataset_phys(ds)->ds_prev_snap_obj;
895 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
896 "reset zb_objset to %llu",
897 (u_longlong_t)ds->ds_object,
898 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
899 } else if (zap_lookup_int_key(dp->dp_meta_objset,
900 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
901 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
902 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
903 VERIFY(zap_add_int_key(dp->dp_meta_objset,
904 scn->scn_phys.scn_queue_obj,
905 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
906 zfs_dbgmsg("snapshotting ds %llu; in queue; "
907 "replacing with %llu",
908 (u_longlong_t)ds->ds_object,
909 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
910 }
911 dsl_scan_sync_state(scn, tx);
912 }
913
914 void
dsl_scan_ds_clone_swapped(dsl_dataset_t * ds1,dsl_dataset_t * ds2,dmu_tx_t * tx)915 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
916 {
917 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
918 dsl_scan_t *scn = dp->dp_scan;
919 uint64_t mintxg;
920
921 if (scn->scn_phys.scn_state != DSS_SCANNING)
922 return;
923
924 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
925 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
926 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
927 "reset zb_objset to %llu",
928 (u_longlong_t)ds1->ds_object,
929 (u_longlong_t)ds2->ds_object);
930 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
931 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
932 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
933 "reset zb_objset to %llu",
934 (u_longlong_t)ds2->ds_object,
935 (u_longlong_t)ds1->ds_object);
936 }
937
938 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
939 ds1->ds_object, &mintxg) == 0) {
940 int err;
941
942 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
943 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
944 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
945 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
946 err = zap_add_int_key(dp->dp_meta_objset,
947 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
948 VERIFY(err == 0 || err == EEXIST);
949 if (err == EEXIST) {
950 /* Both were there to begin with */
951 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
952 scn->scn_phys.scn_queue_obj,
953 ds1->ds_object, mintxg, tx));
954 }
955 zfs_dbgmsg("clone_swap ds %llu; in queue; "
956 "replacing with %llu",
957 (u_longlong_t)ds1->ds_object,
958 (u_longlong_t)ds2->ds_object);
959 } else if (zap_lookup_int_key(dp->dp_meta_objset,
960 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
961 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
962 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
963 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
964 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
965 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
966 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
967 zfs_dbgmsg("clone_swap ds %llu; in queue; "
968 "replacing with %llu",
969 (u_longlong_t)ds2->ds_object,
970 (u_longlong_t)ds1->ds_object);
971 }
972
973 dsl_scan_sync_state(scn, tx);
974 }
975
976 struct enqueue_clones_arg {
977 dmu_tx_t *tx;
978 uint64_t originobj;
979 };
980
981 /* ARGSUSED */
982 static int
enqueue_clones_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)983 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
984 {
985 struct enqueue_clones_arg *eca = arg;
986 dsl_dataset_t *ds;
987 int err;
988 dsl_scan_t *scn = dp->dp_scan;
989
990 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
991 return (0);
992
993 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
994 if (err)
995 return (err);
996
997 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
998 dsl_dataset_t *prev;
999 err = dsl_dataset_hold_obj(dp,
1000 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1001
1002 dsl_dataset_rele(ds, FTAG);
1003 if (err)
1004 return (err);
1005 ds = prev;
1006 }
1007 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1008 scn->scn_phys.scn_queue_obj, ds->ds_object,
1009 dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1010 dsl_dataset_rele(ds, FTAG);
1011 return (0);
1012 }
1013
1014 static void
dsl_scan_visitds(dsl_scan_t * scn,uint64_t dsobj,dmu_tx_t * tx)1015 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1016 {
1017 dsl_pool_t *dp = scn->scn_dp;
1018 dsl_dataset_t *ds;
1019 objset_t *os;
1020
1021 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1022
1023 if (scn->scn_phys.scn_cur_min_txg >=
1024 scn->scn_phys.scn_max_txg) {
1025 /*
1026 * This can happen if this snapshot was created after the
1027 * scan started, and we already completed a previous snapshot
1028 * that was created after the scan started. This snapshot
1029 * only references blocks with:
1030 *
1031 * birth < our ds_creation_txg
1032 * cur_min_txg is no less than ds_creation_txg.
1033 * We have already visited these blocks.
1034 * or
1035 * birth > scn_max_txg
1036 * The scan requested not to visit these blocks.
1037 *
1038 * Subsequent snapshots (and clones) can reference our
1039 * blocks, or blocks with even higher birth times.
1040 * Therefore we do not need to visit them either,
1041 * so we do not add them to the work queue.
1042 *
1043 * Note that checking for cur_min_txg >= cur_max_txg
1044 * is not sufficient, because in that case we may need to
1045 * visit subsequent snapshots. This happens when min_txg > 0,
1046 * which raises cur_min_txg. In this case we will visit
1047 * this dataset but skip all of its blocks, because the
1048 * rootbp's birth time is < cur_min_txg. Then we will
1049 * add the next snapshots/clones to the work queue.
1050 */
1051 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1052 dsl_dataset_name(ds, dsname);
1053 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1054 "cur_min_txg (%llu) >= max_txg (%llu)",
1055 dsobj, dsname,
1056 scn->scn_phys.scn_cur_min_txg,
1057 scn->scn_phys.scn_max_txg);
1058 kmem_free(dsname, MAXNAMELEN);
1059
1060 goto out;
1061 }
1062
1063 if (dmu_objset_from_ds(ds, &os))
1064 goto out;
1065
1066 /*
1067 * Only the ZIL in the head (non-snapshot) is valid. Even though
1068 * snapshots can have ZIL block pointers (which may be the same
1069 * BP as in the head), they must be ignored. So we traverse the
1070 * ZIL here, rather than in scan_recurse(), because the regular
1071 * snapshot block-sharing rules don't apply to it.
1072 */
1073 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1074 dsl_scan_zil(dp, &os->os_zil_header);
1075
1076 /*
1077 * Iterate over the bps in this ds.
1078 */
1079 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1080 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1081
1082 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
1083 dsl_dataset_name(ds, dsname);
1084 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1085 "pausing=%u",
1086 (longlong_t)dsobj, dsname,
1087 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1088 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1089 (int)scn->scn_pausing);
1090 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
1091
1092 if (scn->scn_pausing)
1093 goto out;
1094
1095 /*
1096 * We've finished this pass over this dataset.
1097 */
1098
1099 /*
1100 * If we did not completely visit this dataset, do another pass.
1101 */
1102 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1103 zfs_dbgmsg("incomplete pass; visiting again");
1104 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1105 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1106 scn->scn_phys.scn_queue_obj, ds->ds_object,
1107 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1108 goto out;
1109 }
1110
1111 /*
1112 * Add descendent datasets to work queue.
1113 */
1114 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1115 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1116 scn->scn_phys.scn_queue_obj,
1117 dsl_dataset_phys(ds)->ds_next_snap_obj,
1118 dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1119 }
1120 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1121 boolean_t usenext = B_FALSE;
1122 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1123 uint64_t count;
1124 /*
1125 * A bug in a previous version of the code could
1126 * cause upgrade_clones_cb() to not set
1127 * ds_next_snap_obj when it should, leading to a
1128 * missing entry. Therefore we can only use the
1129 * next_clones_obj when its count is correct.
1130 */
1131 int err = zap_count(dp->dp_meta_objset,
1132 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1133 if (err == 0 &&
1134 count == dsl_dataset_phys(ds)->ds_num_children - 1)
1135 usenext = B_TRUE;
1136 }
1137
1138 if (usenext) {
1139 VERIFY0(zap_join_key(dp->dp_meta_objset,
1140 dsl_dataset_phys(ds)->ds_next_clones_obj,
1141 scn->scn_phys.scn_queue_obj,
1142 dsl_dataset_phys(ds)->ds_creation_txg, tx));
1143 } else {
1144 struct enqueue_clones_arg eca;
1145 eca.tx = tx;
1146 eca.originobj = ds->ds_object;
1147
1148 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1149 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1150 }
1151 }
1152
1153 out:
1154 dsl_dataset_rele(ds, FTAG);
1155 }
1156
1157 /* ARGSUSED */
1158 static int
enqueue_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)1159 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1160 {
1161 dmu_tx_t *tx = arg;
1162 dsl_dataset_t *ds;
1163 int err;
1164 dsl_scan_t *scn = dp->dp_scan;
1165
1166 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1167 if (err)
1168 return (err);
1169
1170 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1171 dsl_dataset_t *prev;
1172 err = dsl_dataset_hold_obj(dp,
1173 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1174 if (err) {
1175 dsl_dataset_rele(ds, FTAG);
1176 return (err);
1177 }
1178
1179 /*
1180 * If this is a clone, we don't need to worry about it for now.
1181 */
1182 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1183 dsl_dataset_rele(ds, FTAG);
1184 dsl_dataset_rele(prev, FTAG);
1185 return (0);
1186 }
1187 dsl_dataset_rele(ds, FTAG);
1188 ds = prev;
1189 }
1190
1191 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1192 ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1193 dsl_dataset_rele(ds, FTAG);
1194 return (0);
1195 }
1196
1197 /*
1198 * Scrub/dedup interaction.
1199 *
1200 * If there are N references to a deduped block, we don't want to scrub it
1201 * N times -- ideally, we should scrub it exactly once.
1202 *
1203 * We leverage the fact that the dde's replication class (enum ddt_class)
1204 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1205 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1206 *
1207 * To prevent excess scrubbing, the scrub begins by walking the DDT
1208 * to find all blocks with refcnt > 1, and scrubs each of these once.
1209 * Since there are two replication classes which contain blocks with
1210 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1211 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1212 *
1213 * There would be nothing more to say if a block's refcnt couldn't change
1214 * during a scrub, but of course it can so we must account for changes
1215 * in a block's replication class.
1216 *
1217 * Here's an example of what can occur:
1218 *
1219 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1220 * when visited during the top-down scrub phase, it will be scrubbed twice.
1221 * This negates our scrub optimization, but is otherwise harmless.
1222 *
1223 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1224 * on each visit during the top-down scrub phase, it will never be scrubbed.
1225 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1226 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1227 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1228 * while a scrub is in progress, it scrubs the block right then.
1229 */
1230 static void
dsl_scan_ddt(dsl_scan_t * scn,dmu_tx_t * tx)1231 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1232 {
1233 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1234 ddt_entry_t dde = { 0 };
1235 int error;
1236 uint64_t n = 0;
1237
1238 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1239 ddt_t *ddt;
1240
1241 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1242 break;
1243 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1244 (longlong_t)ddb->ddb_class,
1245 (longlong_t)ddb->ddb_type,
1246 (longlong_t)ddb->ddb_checksum,
1247 (longlong_t)ddb->ddb_cursor);
1248
1249 /* There should be no pending changes to the dedup table */
1250 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1251 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1252
1253 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1254 n++;
1255
1256 if (dsl_scan_check_pause(scn, NULL))
1257 break;
1258 }
1259
1260 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1261 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1262 (int)scn->scn_pausing);
1263
1264 ASSERT(error == 0 || error == ENOENT);
1265 ASSERT(error != ENOENT ||
1266 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1267 }
1268
1269 /* ARGSUSED */
1270 void
dsl_scan_ddt_entry(dsl_scan_t * scn,enum zio_checksum checksum,ddt_entry_t * dde,dmu_tx_t * tx)1271 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1272 ddt_entry_t *dde, dmu_tx_t *tx)
1273 {
1274 const ddt_key_t *ddk = &dde->dde_key;
1275 ddt_phys_t *ddp = dde->dde_phys;
1276 blkptr_t bp;
1277 zbookmark_phys_t zb = { 0 };
1278
1279 if (scn->scn_phys.scn_state != DSS_SCANNING)
1280 return;
1281
1282 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1283 if (ddp->ddp_phys_birth == 0 ||
1284 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1285 continue;
1286 ddt_bp_create(checksum, ddk, ddp, &bp);
1287
1288 scn->scn_visited_this_txg++;
1289 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1290 }
1291 }
1292
1293 static void
dsl_scan_visit(dsl_scan_t * scn,dmu_tx_t * tx)1294 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1295 {
1296 dsl_pool_t *dp = scn->scn_dp;
1297 zap_cursor_t zc;
1298 zap_attribute_t za;
1299
1300 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1301 scn->scn_phys.scn_ddt_class_max) {
1302 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1303 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1304 dsl_scan_ddt(scn, tx);
1305 if (scn->scn_pausing)
1306 return;
1307 }
1308
1309 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1310 /* First do the MOS & ORIGIN */
1311
1312 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1313 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1314 dsl_scan_visit_rootbp(scn, NULL,
1315 &dp->dp_meta_rootbp, tx);
1316 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1317 if (scn->scn_pausing)
1318 return;
1319
1320 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1321 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1322 enqueue_cb, tx, DS_FIND_CHILDREN));
1323 } else {
1324 dsl_scan_visitds(scn,
1325 dp->dp_origin_snap->ds_object, tx);
1326 }
1327 ASSERT(!scn->scn_pausing);
1328 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1329 ZB_DESTROYED_OBJSET) {
1330 /*
1331 * If we were paused, continue from here. Note if the
1332 * ds we were paused on was deleted, the zb_objset may
1333 * be -1, so we will skip this and find a new objset
1334 * below.
1335 */
1336 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1337 if (scn->scn_pausing)
1338 return;
1339 }
1340
1341 /*
1342 * In case we were paused right at the end of the ds, zero the
1343 * bookmark so we don't think that we're still trying to resume.
1344 */
1345 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1346
1347 /* keep pulling things out of the zap-object-as-queue */
1348 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1349 scn->scn_phys.scn_queue_obj),
1350 zap_cursor_retrieve(&zc, &za) == 0) {
1351 dsl_dataset_t *ds;
1352 uint64_t dsobj;
1353
1354 dsobj = strtonum(za.za_name, NULL);
1355 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1356 scn->scn_phys.scn_queue_obj, dsobj, tx));
1357
1358 /* Set up min/max txg */
1359 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1360 if (za.za_first_integer != 0) {
1361 scn->scn_phys.scn_cur_min_txg =
1362 MAX(scn->scn_phys.scn_min_txg,
1363 za.za_first_integer);
1364 } else {
1365 scn->scn_phys.scn_cur_min_txg =
1366 MAX(scn->scn_phys.scn_min_txg,
1367 dsl_dataset_phys(ds)->ds_prev_snap_txg);
1368 }
1369 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1370 dsl_dataset_rele(ds, FTAG);
1371
1372 dsl_scan_visitds(scn, dsobj, tx);
1373 zap_cursor_fini(&zc);
1374 if (scn->scn_pausing)
1375 return;
1376 }
1377 zap_cursor_fini(&zc);
1378 }
1379
1380 static boolean_t
dsl_scan_free_should_pause(dsl_scan_t * scn)1381 dsl_scan_free_should_pause(dsl_scan_t *scn)
1382 {
1383 uint64_t elapsed_nanosecs;
1384
1385 if (zfs_recover)
1386 return (B_FALSE);
1387
1388 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1389 return (B_TRUE);
1390
1391 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1392 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1393 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1394 txg_sync_waiting(scn->scn_dp)) ||
1395 spa_shutting_down(scn->scn_dp->dp_spa));
1396 }
1397
1398 static int
dsl_scan_free_block_cb(void * arg,const blkptr_t * bp,dmu_tx_t * tx)1399 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1400 {
1401 dsl_scan_t *scn = arg;
1402
1403 if (!scn->scn_is_bptree ||
1404 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1405 if (dsl_scan_free_should_pause(scn))
1406 return (SET_ERROR(ERESTART));
1407 }
1408
1409 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1410 dmu_tx_get_txg(tx), bp, 0));
1411 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1412 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1413 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1414 scn->scn_visited_this_txg++;
1415 return (0);
1416 }
1417
1418 boolean_t
dsl_scan_active(dsl_scan_t * scn)1419 dsl_scan_active(dsl_scan_t *scn)
1420 {
1421 spa_t *spa = scn->scn_dp->dp_spa;
1422 uint64_t used = 0, comp, uncomp;
1423
1424 if (spa->spa_load_state != SPA_LOAD_NONE)
1425 return (B_FALSE);
1426 if (spa_shutting_down(spa))
1427 return (B_FALSE);
1428 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1429 (scn->scn_async_destroying && !scn->scn_async_stalled))
1430 return (B_TRUE);
1431
1432 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1433 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1434 &used, &comp, &uncomp);
1435 }
1436 return (used != 0);
1437 }
1438
1439 void
dsl_scan_sync(dsl_pool_t * dp,dmu_tx_t * tx)1440 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1441 {
1442 dsl_scan_t *scn = dp->dp_scan;
1443 spa_t *spa = dp->dp_spa;
1444 int err = 0;
1445
1446 /*
1447 * Check for scn_restart_txg before checking spa_load_state, so
1448 * that we can restart an old-style scan while the pool is being
1449 * imported (see dsl_scan_init).
1450 */
1451 if (scn->scn_restart_txg != 0 &&
1452 scn->scn_restart_txg <= tx->tx_txg) {
1453 pool_scan_func_t func = POOL_SCAN_SCRUB;
1454 dsl_scan_done(scn, B_FALSE, tx);
1455 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1456 func = POOL_SCAN_RESILVER;
1457 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1458 func, tx->tx_txg);
1459 dsl_scan_setup_sync(&func, tx);
1460 }
1461
1462 /*
1463 * Only process scans in sync pass 1.
1464 */
1465 if (spa_sync_pass(dp->dp_spa) > 1)
1466 return;
1467
1468 /*
1469 * If the spa is shutting down, then stop scanning. This will
1470 * ensure that the scan does not dirty any new data during the
1471 * shutdown phase.
1472 */
1473 if (spa_shutting_down(spa))
1474 return;
1475
1476 /*
1477 * If the scan is inactive due to a stalled async destroy, try again.
1478 */
1479 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1480 return;
1481
1482 scn->scn_visited_this_txg = 0;
1483 scn->scn_pausing = B_FALSE;
1484 scn->scn_sync_start_time = gethrtime();
1485 spa->spa_scrub_active = B_TRUE;
1486
1487 /*
1488 * First process the async destroys. If we pause, don't do
1489 * any scrubbing or resilvering. This ensures that there are no
1490 * async destroys while we are scanning, so the scan code doesn't
1491 * have to worry about traversing it. It is also faster to free the
1492 * blocks than to scrub them.
1493 */
1494 if (zfs_free_bpobj_enabled &&
1495 spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1496 scn->scn_is_bptree = B_FALSE;
1497 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1498 NULL, ZIO_FLAG_MUSTSUCCEED);
1499 err = bpobj_iterate(&dp->dp_free_bpobj,
1500 dsl_scan_free_block_cb, scn, tx);
1501 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1502
1503 if (err != 0 && err != ERESTART)
1504 zfs_panic_recover("error %u from bpobj_iterate()", err);
1505 }
1506
1507 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1508 ASSERT(scn->scn_async_destroying);
1509 scn->scn_is_bptree = B_TRUE;
1510 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1511 NULL, ZIO_FLAG_MUSTSUCCEED);
1512 err = bptree_iterate(dp->dp_meta_objset,
1513 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1514 VERIFY0(zio_wait(scn->scn_zio_root));
1515
1516 if (err == EIO || err == ECKSUM) {
1517 err = 0;
1518 } else if (err != 0 && err != ERESTART) {
1519 zfs_panic_recover("error %u from "
1520 "traverse_dataset_destroyed()", err);
1521 }
1522
1523 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1524 /* finished; deactivate async destroy feature */
1525 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1526 ASSERT(!spa_feature_is_active(spa,
1527 SPA_FEATURE_ASYNC_DESTROY));
1528 VERIFY0(zap_remove(dp->dp_meta_objset,
1529 DMU_POOL_DIRECTORY_OBJECT,
1530 DMU_POOL_BPTREE_OBJ, tx));
1531 VERIFY0(bptree_free(dp->dp_meta_objset,
1532 dp->dp_bptree_obj, tx));
1533 dp->dp_bptree_obj = 0;
1534 scn->scn_async_destroying = B_FALSE;
1535 scn->scn_async_stalled = B_FALSE;
1536 } else {
1537 /*
1538 * If we didn't make progress, mark the async
1539 * destroy as stalled, so that we will not initiate
1540 * a spa_sync() on its behalf. Note that we only
1541 * check this if we are not finished, because if the
1542 * bptree had no blocks for us to visit, we can
1543 * finish without "making progress".
1544 */
1545 scn->scn_async_stalled =
1546 (scn->scn_visited_this_txg == 0);
1547 }
1548 }
1549 if (scn->scn_visited_this_txg) {
1550 zfs_dbgmsg("freed %llu blocks in %llums from "
1551 "free_bpobj/bptree txg %llu; err=%u",
1552 (longlong_t)scn->scn_visited_this_txg,
1553 (longlong_t)
1554 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1555 (longlong_t)tx->tx_txg, err);
1556 scn->scn_visited_this_txg = 0;
1557
1558 /*
1559 * Write out changes to the DDT that may be required as a
1560 * result of the blocks freed. This ensures that the DDT
1561 * is clean when a scrub/resilver runs.
1562 */
1563 ddt_sync(spa, tx->tx_txg);
1564 }
1565 if (err != 0)
1566 return;
1567 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1568 zfs_free_leak_on_eio &&
1569 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1570 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1571 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1572 /*
1573 * We have finished background destroying, but there is still
1574 * some space left in the dp_free_dir. Transfer this leaked
1575 * space to the dp_leak_dir.
1576 */
1577 if (dp->dp_leak_dir == NULL) {
1578 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1579 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1580 LEAK_DIR_NAME, tx);
1581 VERIFY0(dsl_pool_open_special_dir(dp,
1582 LEAK_DIR_NAME, &dp->dp_leak_dir));
1583 rrw_exit(&dp->dp_config_rwlock, FTAG);
1584 }
1585 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1586 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1587 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1588 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1589 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1590 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1591 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1592 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1593 }
1594 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1595 /* finished; verify that space accounting went to zero */
1596 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1597 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1598 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1599 }
1600
1601 if (scn->scn_phys.scn_state != DSS_SCANNING)
1602 return;
1603
1604 if (scn->scn_done_txg == tx->tx_txg) {
1605 ASSERT(!scn->scn_pausing);
1606 /* finished with scan. */
1607 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1608 dsl_scan_done(scn, B_TRUE, tx);
1609 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1610 dsl_scan_sync_state(scn, tx);
1611 return;
1612 }
1613
1614 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1615 scn->scn_phys.scn_ddt_class_max) {
1616 zfs_dbgmsg("doing scan sync txg %llu; "
1617 "ddt bm=%llu/%llu/%llu/%llx",
1618 (longlong_t)tx->tx_txg,
1619 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1620 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1621 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1622 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1623 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1624 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1625 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1626 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1627 } else {
1628 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1629 (longlong_t)tx->tx_txg,
1630 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1631 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1632 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1633 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1634 }
1635
1636 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1637 NULL, ZIO_FLAG_CANFAIL);
1638 dsl_pool_config_enter(dp, FTAG);
1639 dsl_scan_visit(scn, tx);
1640 dsl_pool_config_exit(dp, FTAG);
1641 (void) zio_wait(scn->scn_zio_root);
1642 scn->scn_zio_root = NULL;
1643
1644 zfs_dbgmsg("visited %llu blocks in %llums",
1645 (longlong_t)scn->scn_visited_this_txg,
1646 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1647
1648 if (!scn->scn_pausing) {
1649 scn->scn_done_txg = tx->tx_txg + 1;
1650 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1651 tx->tx_txg, scn->scn_done_txg);
1652 }
1653
1654 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1655 mutex_enter(&spa->spa_scrub_lock);
1656 while (spa->spa_scrub_inflight > 0) {
1657 cv_wait(&spa->spa_scrub_io_cv,
1658 &spa->spa_scrub_lock);
1659 }
1660 mutex_exit(&spa->spa_scrub_lock);
1661 }
1662
1663 dsl_scan_sync_state(scn, tx);
1664 }
1665
1666 /*
1667 * This will start a new scan, or restart an existing one.
1668 */
1669 void
dsl_resilver_restart(dsl_pool_t * dp,uint64_t txg)1670 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1671 {
1672 if (txg == 0) {
1673 dmu_tx_t *tx;
1674 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1675 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1676
1677 txg = dmu_tx_get_txg(tx);
1678 dp->dp_scan->scn_restart_txg = txg;
1679 dmu_tx_commit(tx);
1680 } else {
1681 dp->dp_scan->scn_restart_txg = txg;
1682 }
1683 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1684 }
1685
1686 boolean_t
dsl_scan_resilvering(dsl_pool_t * dp)1687 dsl_scan_resilvering(dsl_pool_t *dp)
1688 {
1689 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1690 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1691 }
1692
1693 /*
1694 * scrub consumers
1695 */
1696
1697 static void
count_block(zfs_all_blkstats_t * zab,const blkptr_t * bp)1698 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1699 {
1700 int i;
1701
1702 /*
1703 * If we resume after a reboot, zab will be NULL; don't record
1704 * incomplete stats in that case.
1705 */
1706 if (zab == NULL)
1707 return;
1708
1709 for (i = 0; i < 4; i++) {
1710 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1711 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1712 if (t & DMU_OT_NEWTYPE)
1713 t = DMU_OT_OTHER;
1714 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1715 int equal;
1716
1717 zb->zb_count++;
1718 zb->zb_asize += BP_GET_ASIZE(bp);
1719 zb->zb_lsize += BP_GET_LSIZE(bp);
1720 zb->zb_psize += BP_GET_PSIZE(bp);
1721 zb->zb_gangs += BP_COUNT_GANG(bp);
1722
1723 switch (BP_GET_NDVAS(bp)) {
1724 case 2:
1725 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1726 DVA_GET_VDEV(&bp->blk_dva[1]))
1727 zb->zb_ditto_2_of_2_samevdev++;
1728 break;
1729 case 3:
1730 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1731 DVA_GET_VDEV(&bp->blk_dva[1])) +
1732 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1733 DVA_GET_VDEV(&bp->blk_dva[2])) +
1734 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1735 DVA_GET_VDEV(&bp->blk_dva[2]));
1736 if (equal == 1)
1737 zb->zb_ditto_2_of_3_samevdev++;
1738 else if (equal == 3)
1739 zb->zb_ditto_3_of_3_samevdev++;
1740 break;
1741 }
1742 }
1743 }
1744
1745 static void
dsl_scan_scrub_done(zio_t * zio)1746 dsl_scan_scrub_done(zio_t *zio)
1747 {
1748 spa_t *spa = zio->io_spa;
1749
1750 zio_data_buf_free(zio->io_data, zio->io_size);
1751
1752 mutex_enter(&spa->spa_scrub_lock);
1753 spa->spa_scrub_inflight--;
1754 cv_broadcast(&spa->spa_scrub_io_cv);
1755
1756 if (zio->io_error && (zio->io_error != ECKSUM ||
1757 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1758 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1759 }
1760 mutex_exit(&spa->spa_scrub_lock);
1761 }
1762
1763 static int
dsl_scan_scrub_cb(dsl_pool_t * dp,const blkptr_t * bp,const zbookmark_phys_t * zb)1764 dsl_scan_scrub_cb(dsl_pool_t *dp,
1765 const blkptr_t *bp, const zbookmark_phys_t *zb)
1766 {
1767 dsl_scan_t *scn = dp->dp_scan;
1768 size_t size = BP_GET_PSIZE(bp);
1769 spa_t *spa = dp->dp_spa;
1770 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1771 boolean_t needs_io;
1772 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1773 int scan_delay = 0;
1774
1775 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1776 phys_birth >= scn->scn_phys.scn_max_txg)
1777 return (0);
1778
1779 count_block(dp->dp_blkstats, bp);
1780
1781 if (BP_IS_EMBEDDED(bp))
1782 return (0);
1783
1784 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1785 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1786 zio_flags |= ZIO_FLAG_SCRUB;
1787 needs_io = B_TRUE;
1788 scan_delay = zfs_scrub_delay;
1789 } else {
1790 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1791 zio_flags |= ZIO_FLAG_RESILVER;
1792 needs_io = B_FALSE;
1793 scan_delay = zfs_resilver_delay;
1794 }
1795
1796 /* If it's an intent log block, failure is expected. */
1797 if (zb->zb_level == ZB_ZIL_LEVEL)
1798 zio_flags |= ZIO_FLAG_SPECULATIVE;
1799
1800 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1801 vdev_t *vd = vdev_lookup_top(spa,
1802 DVA_GET_VDEV(&bp->blk_dva[d]));
1803
1804 /*
1805 * Keep track of how much data we've examined so that
1806 * zpool(1M) status can make useful progress reports.
1807 */
1808 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1809 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1810
1811 /* if it's a resilver, this may not be in the target range */
1812 if (!needs_io) {
1813 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1814 /*
1815 * Gang members may be spread across multiple
1816 * vdevs, so the best estimate we have is the
1817 * scrub range, which has already been checked.
1818 * XXX -- it would be better to change our
1819 * allocation policy to ensure that all
1820 * gang members reside on the same vdev.
1821 */
1822 needs_io = B_TRUE;
1823 } else {
1824 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1825 phys_birth, 1);
1826 }
1827 }
1828 }
1829
1830 if (needs_io && !zfs_no_scrub_io) {
1831 vdev_t *rvd = spa->spa_root_vdev;
1832 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1833 void *data = zio_data_buf_alloc(size);
1834
1835 mutex_enter(&spa->spa_scrub_lock);
1836 while (spa->spa_scrub_inflight >= maxinflight)
1837 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1838 spa->spa_scrub_inflight++;
1839 mutex_exit(&spa->spa_scrub_lock);
1840
1841 /*
1842 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1843 * then throttle our workload to limit the impact of a scan.
1844 */
1845 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1846 delay(scan_delay);
1847
1848 zio_nowait(zio_read(NULL, spa, bp, data, size,
1849 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1850 zio_flags, zb));
1851 }
1852
1853 /* do not relocate this block */
1854 return (0);
1855 }
1856
1857 int
dsl_scan(dsl_pool_t * dp,pool_scan_func_t func)1858 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1859 {
1860 spa_t *spa = dp->dp_spa;
1861
1862 /*
1863 * Purge all vdev caches and probe all devices. We do this here
1864 * rather than in sync context because this requires a writer lock
1865 * on the spa_config lock, which we can't do from sync context. The
1866 * spa_scrub_reopen flag indicates that vdev_open() should not
1867 * attempt to start another scrub.
1868 */
1869 spa_vdev_state_enter(spa, SCL_NONE);
1870 spa->spa_scrub_reopen = B_TRUE;
1871 vdev_reopen(spa->spa_root_vdev);
1872 spa->spa_scrub_reopen = B_FALSE;
1873 (void) spa_vdev_state_exit(spa, NULL, 0);
1874
1875 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1876 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1877 }
1878