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 *);
60 static boolean_t dsl_scan_restarting(dsl_scan_t *, dmu_tx_t *);
61
62 unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
63 unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
64 unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
65 unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */
66
67 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
68 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
69 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
70 per txg */
71 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
72 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
73
74 SYSCTL_DECL(_vfs_zfs);
75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RWTUN,
76 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
77 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RWTUN,
78 &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
79 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RWTUN,
80 &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RWTUN,
82 &zfs_scan_idle, 0, "Idle scan window in clock ticks");
83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RWTUN,
84 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
85 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RWTUN,
86 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
87 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RWTUN,
88 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
89 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RWTUN,
90 &zfs_no_scrub_io, 0, "Disable scrub I/O");
91 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RWTUN,
92 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
93
94 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
95 /* max number of blocks to free in a single TXG */
96 uint64_t zfs_free_max_blocks = UINT64_MAX;
97 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
98 &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
99
100
101 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
102 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
103 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
104
105 extern int zfs_txg_timeout;
106
107 /*
108 * Enable/disable the processing of the free_bpobj object.
109 */
110 boolean_t zfs_free_bpobj_enabled = B_TRUE;
111
112 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
113 &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
114
115 /* the order has to match pool_scan_type */
116 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
117 NULL,
118 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
119 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
120 };
121
122 int
dsl_scan_init(dsl_pool_t * dp,uint64_t txg)123 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
124 {
125 int err;
126 dsl_scan_t *scn;
127 spa_t *spa = dp->dp_spa;
128 uint64_t f;
129
130 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
131 scn->scn_dp = dp;
132
133 /*
134 * It's possible that we're resuming a scan after a reboot so
135 * make sure that the scan_async_destroying flag is initialized
136 * appropriately.
137 */
138 ASSERT(!scn->scn_async_destroying);
139 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
140 SPA_FEATURE_ASYNC_DESTROY);
141
142 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
143 "scrub_func", sizeof (uint64_t), 1, &f);
144 if (err == 0) {
145 /*
146 * There was an old-style scrub in progress. Restart a
147 * new-style scrub from the beginning.
148 */
149 scn->scn_restart_txg = txg;
150 zfs_dbgmsg("old-style scrub was in progress; "
151 "restarting new-style scrub in txg %llu",
152 scn->scn_restart_txg);
153
154 /*
155 * Load the queue obj from the old location so that it
156 * can be freed by dsl_scan_done().
157 */
158 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
159 "scrub_queue", sizeof (uint64_t), 1,
160 &scn->scn_phys.scn_queue_obj);
161 } else {
162 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
163 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
164 &scn->scn_phys);
165 if (err == ENOENT)
166 return (0);
167 else if (err)
168 return (err);
169
170 if (scn->scn_phys.scn_state == DSS_SCANNING &&
171 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
172 /*
173 * A new-type scrub was in progress on an old
174 * pool, and the pool was accessed by old
175 * software. Restart from the beginning, since
176 * the old software may have changed the pool in
177 * the meantime.
178 */
179 scn->scn_restart_txg = txg;
180 zfs_dbgmsg("new-style scrub was modified "
181 "by old software; restarting in txg %llu",
182 scn->scn_restart_txg);
183 }
184 }
185
186 spa_scan_stat_init(spa);
187 return (0);
188 }
189
190 void
dsl_scan_fini(dsl_pool_t * dp)191 dsl_scan_fini(dsl_pool_t *dp)
192 {
193 if (dp->dp_scan) {
194 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
195 dp->dp_scan = NULL;
196 }
197 }
198
199 /* ARGSUSED */
200 static int
dsl_scan_setup_check(void * arg,dmu_tx_t * tx)201 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
202 {
203 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
204
205 if (scn->scn_phys.scn_state == DSS_SCANNING)
206 return (SET_ERROR(EBUSY));
207
208 return (0);
209 }
210
211 static void
dsl_scan_setup_sync(void * arg,dmu_tx_t * tx)212 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
213 {
214 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
215 pool_scan_func_t *funcp = arg;
216 dmu_object_type_t ot = 0;
217 dsl_pool_t *dp = scn->scn_dp;
218 spa_t *spa = dp->dp_spa;
219
220 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
221 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
222 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
223 scn->scn_phys.scn_func = *funcp;
224 scn->scn_phys.scn_state = DSS_SCANNING;
225 scn->scn_phys.scn_min_txg = 0;
226 scn->scn_phys.scn_max_txg = tx->tx_txg;
227 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
228 scn->scn_phys.scn_start_time = gethrestime_sec();
229 scn->scn_phys.scn_errors = 0;
230 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
231 scn->scn_restart_txg = 0;
232 scn->scn_done_txg = 0;
233 spa_scan_stat_init(spa);
234
235 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
236 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
237
238 /* rewrite all disk labels */
239 vdev_config_dirty(spa->spa_root_vdev);
240
241 if (vdev_resilver_needed(spa->spa_root_vdev,
242 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
243 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
244 } else {
245 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
246 }
247
248 spa->spa_scrub_started = B_TRUE;
249 /*
250 * If this is an incremental scrub, limit the DDT scrub phase
251 * to just the auto-ditto class (for correctness); the rest
252 * of the scrub should go faster using top-down pruning.
253 */
254 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
255 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
256
257 }
258
259 /* back to the generic stuff */
260
261 if (dp->dp_blkstats == NULL) {
262 dp->dp_blkstats =
263 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
264 }
265 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
266
267 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
268 ot = DMU_OT_ZAP_OTHER;
269
270 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
271 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
272
273 dsl_scan_sync_state(scn, tx);
274
275 spa_history_log_internal(spa, "scan setup", tx,
276 "func=%u mintxg=%llu maxtxg=%llu",
277 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
278 }
279
280 /* ARGSUSED */
281 static void
dsl_scan_done(dsl_scan_t * scn,boolean_t complete,dmu_tx_t * tx)282 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
283 {
284 static const char *old_names[] = {
285 "scrub_bookmark",
286 "scrub_ddt_bookmark",
287 "scrub_ddt_class_max",
288 "scrub_queue",
289 "scrub_min_txg",
290 "scrub_max_txg",
291 "scrub_func",
292 "scrub_errors",
293 NULL
294 };
295
296 dsl_pool_t *dp = scn->scn_dp;
297 spa_t *spa = dp->dp_spa;
298 int i;
299
300 /* Remove any remnants of an old-style scrub. */
301 for (i = 0; old_names[i]; i++) {
302 (void) zap_remove(dp->dp_meta_objset,
303 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
304 }
305
306 if (scn->scn_phys.scn_queue_obj != 0) {
307 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
308 scn->scn_phys.scn_queue_obj, tx));
309 scn->scn_phys.scn_queue_obj = 0;
310 }
311
312 /*
313 * If we were "restarted" from a stopped state, don't bother
314 * with anything else.
315 */
316 if (scn->scn_phys.scn_state != DSS_SCANNING)
317 return;
318
319 if (complete)
320 scn->scn_phys.scn_state = DSS_FINISHED;
321 else
322 scn->scn_phys.scn_state = DSS_CANCELED;
323
324 if (dsl_scan_restarting(scn, tx))
325 spa_history_log_internal(spa, "scan aborted, restarting", tx,
326 "errors=%llu", spa_get_errlog_size(spa));
327 else if (!complete)
328 spa_history_log_internal(spa, "scan cancelled", tx,
329 "errors=%llu", spa_get_errlog_size(spa));
330 else
331 spa_history_log_internal(spa, "scan done", tx,
332 "errors=%llu", spa_get_errlog_size(spa));
333
334 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
335 mutex_enter(&spa->spa_scrub_lock);
336 while (spa->spa_scrub_inflight > 0) {
337 cv_wait(&spa->spa_scrub_io_cv,
338 &spa->spa_scrub_lock);
339 }
340 mutex_exit(&spa->spa_scrub_lock);
341 spa->spa_scrub_started = B_FALSE;
342 spa->spa_scrub_active = B_FALSE;
343
344 /*
345 * If the scrub/resilver completed, update all DTLs to
346 * reflect this. Whether it succeeded or not, vacate
347 * all temporary scrub DTLs.
348 */
349 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
350 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
351 if (complete) {
352 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
353 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
354 }
355 spa_errlog_rotate(spa);
356
357 /*
358 * We may have finished replacing a device.
359 * Let the async thread assess this and handle the detach.
360 */
361 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
362 }
363
364 scn->scn_phys.scn_end_time = gethrestime_sec();
365 }
366
367 /* ARGSUSED */
368 static int
dsl_scan_cancel_check(void * arg,dmu_tx_t * tx)369 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
370 {
371 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
372
373 if (scn->scn_phys.scn_state != DSS_SCANNING)
374 return (SET_ERROR(ENOENT));
375 return (0);
376 }
377
378 /* ARGSUSED */
379 static void
dsl_scan_cancel_sync(void * arg,dmu_tx_t * tx)380 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
381 {
382 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
383
384 dsl_scan_done(scn, B_FALSE, tx);
385 dsl_scan_sync_state(scn, tx);
386 }
387
388 int
dsl_scan_cancel(dsl_pool_t * dp)389 dsl_scan_cancel(dsl_pool_t *dp)
390 {
391 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
392 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
393 }
394
395 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
396 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
397 dmu_objset_type_t ostype, dmu_tx_t *tx);
398 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
399 dmu_objset_type_t ostype,
400 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx, zbookmark_phys_t *);
401
402 void
dsl_free(dsl_pool_t * dp,uint64_t txg,const blkptr_t * bp)403 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
404 {
405 zio_free(dp->dp_spa, txg, bp);
406 }
407
408 void
dsl_free_sync(zio_t * pio,dsl_pool_t * dp,uint64_t txg,const blkptr_t * bpp)409 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
410 {
411 ASSERT(dsl_pool_sync_context(dp));
412 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
413 pio->io_flags));
414 }
415
416 static uint64_t
dsl_scan_ds_maxtxg(dsl_dataset_t * ds)417 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
418 {
419 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
420 if (ds->ds_is_snapshot)
421 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
422 return (smt);
423 }
424
425 static void
dsl_scan_sync_state(dsl_scan_t * scn,dmu_tx_t * tx)426 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
427 {
428 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
429 DMU_POOL_DIRECTORY_OBJECT,
430 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
431 &scn->scn_phys, tx));
432 }
433
434 extern int zfs_vdev_async_write_active_min_dirty_percent;
435
436 static boolean_t
dsl_scan_check_pause(dsl_scan_t * scn,const zbookmark_phys_t * zb)437 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
438 {
439 /* we never skip user/group accounting objects */
440 if (zb && (int64_t)zb->zb_object < 0)
441 return (B_FALSE);
442
443 if (scn->scn_pausing)
444 return (B_TRUE); /* we're already pausing */
445
446 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
447 return (B_FALSE); /* we're resuming */
448
449 /* We only know how to resume from level-0 blocks. */
450 if (zb && zb->zb_level != 0)
451 return (B_FALSE);
452
453 /*
454 * We pause if:
455 * - we have scanned for the maximum time: an entire txg
456 * timeout (default 5 sec)
457 * or
458 * - we have scanned for at least the minimum time (default 1 sec
459 * for scrub, 3 sec for resilver), and either we have sufficient
460 * dirty data that we are starting to write more quickly
461 * (default 30%), or someone is explicitly waiting for this txg
462 * to complete.
463 * or
464 * - the spa is shutting down because this pool is being exported
465 * or the machine is rebooting.
466 */
467 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
468 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
469 uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
470 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
471 if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
472 (NSEC2MSEC(elapsed_nanosecs) > mintime &&
473 (txg_sync_waiting(scn->scn_dp) ||
474 dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
475 spa_shutting_down(scn->scn_dp->dp_spa)) {
476 if (zb) {
477 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
478 (longlong_t)zb->zb_objset,
479 (longlong_t)zb->zb_object,
480 (longlong_t)zb->zb_level,
481 (longlong_t)zb->zb_blkid);
482 scn->scn_phys.scn_bookmark = *zb;
483 }
484 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
485 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
486 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
487 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
488 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
489 scn->scn_pausing = B_TRUE;
490 return (B_TRUE);
491 }
492 return (B_FALSE);
493 }
494
495 typedef struct zil_scan_arg {
496 dsl_pool_t *zsa_dp;
497 zil_header_t *zsa_zh;
498 } zil_scan_arg_t;
499
500 /* ARGSUSED */
501 static int
dsl_scan_zil_block(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)502 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
503 {
504 zil_scan_arg_t *zsa = arg;
505 dsl_pool_t *dp = zsa->zsa_dp;
506 dsl_scan_t *scn = dp->dp_scan;
507 zil_header_t *zh = zsa->zsa_zh;
508 zbookmark_phys_t zb;
509
510 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
511 return (0);
512
513 /*
514 * One block ("stubby") can be allocated a long time ago; we
515 * want to visit that one because it has been allocated
516 * (on-disk) even if it hasn't been claimed (even though for
517 * scrub there's nothing to do to it).
518 */
519 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
520 return (0);
521
522 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
523 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
524
525 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
526 return (0);
527 }
528
529 /* ARGSUSED */
530 static int
dsl_scan_zil_record(zilog_t * zilog,lr_t * lrc,void * arg,uint64_t claim_txg)531 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
532 {
533 if (lrc->lrc_txtype == TX_WRITE) {
534 zil_scan_arg_t *zsa = arg;
535 dsl_pool_t *dp = zsa->zsa_dp;
536 dsl_scan_t *scn = dp->dp_scan;
537 zil_header_t *zh = zsa->zsa_zh;
538 lr_write_t *lr = (lr_write_t *)lrc;
539 blkptr_t *bp = &lr->lr_blkptr;
540 zbookmark_phys_t zb;
541
542 if (BP_IS_HOLE(bp) ||
543 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
544 return (0);
545
546 /*
547 * birth can be < claim_txg if this record's txg is
548 * already txg sync'ed (but this log block contains
549 * other records that are not synced)
550 */
551 if (claim_txg == 0 || bp->blk_birth < claim_txg)
552 return (0);
553
554 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
555 lr->lr_foid, ZB_ZIL_LEVEL,
556 lr->lr_offset / BP_GET_LSIZE(bp));
557
558 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
559 }
560 return (0);
561 }
562
563 static void
dsl_scan_zil(dsl_pool_t * dp,zil_header_t * zh)564 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
565 {
566 uint64_t claim_txg = zh->zh_claim_txg;
567 zil_scan_arg_t zsa = { dp, zh };
568 zilog_t *zilog;
569
570 /*
571 * We only want to visit blocks that have been claimed but not yet
572 * replayed (or, in read-only mode, blocks that *would* be claimed).
573 */
574 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
575 return;
576
577 zilog = zil_alloc(dp->dp_meta_objset, zh);
578
579 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
580 claim_txg);
581
582 zil_free(zilog);
583 }
584
585 /* ARGSUSED */
586 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,zbookmark_phys_t * czb)587 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
588 uint64_t objset, uint64_t object, uint64_t blkid, zbookmark_phys_t *czb)
589 {
590 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
591
592 if (zfs_no_scrub_prefetch)
593 return;
594
595 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
596 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
597 return;
598
599 SET_BOOKMARK(czb, objset, object, BP_GET_LEVEL(bp), blkid);
600
601 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
602 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
603 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, czb);
604 }
605
606 static boolean_t
dsl_scan_check_resume(dsl_scan_t * scn,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)607 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
608 const zbookmark_phys_t *zb)
609 {
610 /*
611 * We never skip over user/group accounting objects (obj<0)
612 */
613 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
614 (int64_t)zb->zb_object >= 0) {
615 /*
616 * If we already visited this bp & everything below (in
617 * a prior txg sync), don't bother doing it again.
618 */
619 if (zbookmark_subtree_completed(dnp, zb,
620 &scn->scn_phys.scn_bookmark))
621 return (B_TRUE);
622
623 /*
624 * If we found the block we're trying to resume from, or
625 * we went past it to a different object, zero it out to
626 * indicate that it's OK to start checking for pausing
627 * again.
628 */
629 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
630 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
631 dprintf("resuming at %llx/%llx/%llx/%llx\n",
632 (longlong_t)zb->zb_objset,
633 (longlong_t)zb->zb_object,
634 (longlong_t)zb->zb_level,
635 (longlong_t)zb->zb_blkid);
636 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
637 }
638 }
639 return (B_FALSE);
640 }
641
642 /*
643 * Return nonzero on i/o error.
644 * Return new buf to write out in *bufp.
645 */
646 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)647 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
648 dnode_phys_t *dnp, const blkptr_t *bp,
649 const zbookmark_phys_t *zb, dmu_tx_t *tx)
650 {
651 dsl_pool_t *dp = scn->scn_dp;
652 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
653 int err;
654
655 if (BP_GET_LEVEL(bp) > 0) {
656 arc_flags_t flags = ARC_FLAG_WAIT;
657 int i;
658 blkptr_t *cbp;
659 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
660 arc_buf_t *buf;
661 zbookmark_phys_t *czb;
662
663 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
664 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
665 if (err) {
666 scn->scn_phys.scn_errors++;
667 return (err);
668 }
669 czb = kmem_alloc(sizeof (*czb), KM_SLEEP);
670 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
671 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
672 zb->zb_object, zb->zb_blkid * epb + i, czb);
673 }
674 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
675 SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
676 zb->zb_level - 1,
677 zb->zb_blkid * epb + i);
678 dsl_scan_visitbp(cbp, czb, dnp,
679 ds, scn, ostype, tx);
680 }
681 kmem_free(czb, sizeof (*czb));
682 arc_buf_destroy(buf, &buf);
683 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
684 arc_flags_t flags = ARC_FLAG_WAIT;
685 dnode_phys_t *cdnp;
686 int i, j;
687 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
688 arc_buf_t *buf;
689 zbookmark_phys_t *czb;
690
691 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
692 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
693 if (err) {
694 scn->scn_phys.scn_errors++;
695 return (err);
696 }
697 czb = kmem_alloc(sizeof (*czb), KM_SLEEP);
698 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
699 for (j = 0; j < cdnp->dn_nblkptr; j++) {
700 blkptr_t *cbp = &cdnp->dn_blkptr[j];
701 dsl_scan_prefetch(scn, buf, cbp,
702 zb->zb_objset, zb->zb_blkid * epb + i, j,
703 czb);
704 }
705 }
706 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
707 dsl_scan_visitdnode(scn, ds, ostype,
708 cdnp, zb->zb_blkid * epb + i, tx, czb);
709 }
710 kmem_free(czb, sizeof (*czb));
711
712 arc_buf_destroy(buf, &buf);
713 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
714 arc_flags_t flags = ARC_FLAG_WAIT;
715 objset_phys_t *osp;
716 arc_buf_t *buf;
717 zbookmark_phys_t *czb;
718
719 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
720 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
721 if (err) {
722 scn->scn_phys.scn_errors++;
723 return (err);
724 }
725
726 osp = buf->b_data;
727
728 czb = kmem_alloc(sizeof (*czb), KM_SLEEP);
729 dsl_scan_visitdnode(scn, ds, osp->os_type,
730 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx, czb);
731
732 if (OBJSET_BUF_HAS_USERUSED(buf)) {
733 /*
734 * We also always visit user/group accounting
735 * objects, and never skip them, even if we are
736 * pausing. This is necessary so that the space
737 * deltas from this txg get integrated.
738 */
739 dsl_scan_visitdnode(scn, ds, osp->os_type,
740 &osp->os_groupused_dnode,
741 DMU_GROUPUSED_OBJECT, tx, czb);
742 dsl_scan_visitdnode(scn, ds, osp->os_type,
743 &osp->os_userused_dnode,
744 DMU_USERUSED_OBJECT, tx, czb);
745 }
746 kmem_free(czb, sizeof (*czb));
747 arc_buf_destroy(buf, &buf);
748 }
749
750 return (0);
751 }
752
753 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,zbookmark_phys_t * czb)754 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
755 dmu_objset_type_t ostype, dnode_phys_t *dnp,
756 uint64_t object, dmu_tx_t *tx, zbookmark_phys_t *czb)
757 {
758 int j;
759
760 for (j = 0; j < dnp->dn_nblkptr; j++) {
761 SET_BOOKMARK(czb, ds ? ds->ds_object : 0, object,
762 dnp->dn_nlevels - 1, j);
763 dsl_scan_visitbp(&dnp->dn_blkptr[j],
764 czb, dnp, ds, scn, ostype, tx);
765 }
766
767 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
768 SET_BOOKMARK(czb, ds ? ds->ds_object : 0, object,
769 0, DMU_SPILL_BLKID);
770 dsl_scan_visitbp(&dnp->dn_spill,
771 czb, dnp, ds, scn, ostype, tx);
772 }
773 }
774
775 /*
776 * The arguments are in this order because mdb can only print the
777 * first 5; we want them to be useful.
778 */
779 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)780 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
781 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
782 dmu_objset_type_t ostype, dmu_tx_t *tx)
783 {
784 dsl_pool_t *dp = scn->scn_dp;
785 blkptr_t *bp_toread = NULL;
786
787 if (dsl_scan_check_pause(scn, zb))
788 return;
789
790 if (dsl_scan_check_resume(scn, dnp, zb))
791 return;
792
793 if (BP_IS_HOLE(bp))
794 return;
795
796 scn->scn_visited_this_txg++;
797
798 dprintf_bp(bp,
799 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
800 ds, ds ? ds->ds_object : 0,
801 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
802 bp);
803
804 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
805 return;
806
807 bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
808 *bp_toread = *bp;
809
810 if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
811 goto out;
812
813 /*
814 * If dsl_scan_ddt() has aready visited this block, it will have
815 * already done any translations or scrubbing, so don't call the
816 * callback again.
817 */
818 if (ddt_class_contains(dp->dp_spa,
819 scn->scn_phys.scn_ddt_class_max, bp)) {
820 goto out;
821 }
822
823 /*
824 * If this block is from the future (after cur_max_txg), then we
825 * are doing this on behalf of a deleted snapshot, and we will
826 * revisit the future block on the next pass of this dataset.
827 * Don't scan it now unless we need to because something
828 * under it was modified.
829 */
830 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
831 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
832 }
833
834 out:
835 kmem_free(bp_toread, sizeof (blkptr_t));
836 }
837
838 static void
dsl_scan_visit_rootbp(dsl_scan_t * scn,dsl_dataset_t * ds,blkptr_t * bp,dmu_tx_t * tx)839 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
840 dmu_tx_t *tx)
841 {
842 zbookmark_phys_t zb;
843
844 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
845 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
846 dsl_scan_visitbp(bp, &zb, NULL,
847 ds, scn, DMU_OST_NONE, tx);
848
849 dprintf_ds(ds, "finished scan%s", "");
850 }
851
852 void
dsl_scan_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)853 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
854 {
855 dsl_pool_t *dp = ds->ds_dir->dd_pool;
856 dsl_scan_t *scn = dp->dp_scan;
857 uint64_t mintxg;
858
859 if (scn->scn_phys.scn_state != DSS_SCANNING)
860 return;
861
862 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
863 if (ds->ds_is_snapshot) {
864 /*
865 * Note:
866 * - scn_cur_{min,max}_txg stays the same.
867 * - Setting the flag is not really necessary if
868 * scn_cur_max_txg == scn_max_txg, because there
869 * is nothing after this snapshot that we care
870 * about. However, we set it anyway and then
871 * ignore it when we retraverse it in
872 * dsl_scan_visitds().
873 */
874 scn->scn_phys.scn_bookmark.zb_objset =
875 dsl_dataset_phys(ds)->ds_next_snap_obj;
876 zfs_dbgmsg("destroying ds %llu; currently traversing; "
877 "reset zb_objset to %llu",
878 (u_longlong_t)ds->ds_object,
879 (u_longlong_t)dsl_dataset_phys(ds)->
880 ds_next_snap_obj);
881 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
882 } else {
883 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
884 ZB_DESTROYED_OBJSET, 0, 0, 0);
885 zfs_dbgmsg("destroying ds %llu; currently traversing; "
886 "reset bookmark to -1,0,0,0",
887 (u_longlong_t)ds->ds_object);
888 }
889 } else if (zap_lookup_int_key(dp->dp_meta_objset,
890 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
891 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
892 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
893 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
894 if (ds->ds_is_snapshot) {
895 /*
896 * We keep the same mintxg; it could be >
897 * ds_creation_txg if the previous snapshot was
898 * deleted too.
899 */
900 VERIFY(zap_add_int_key(dp->dp_meta_objset,
901 scn->scn_phys.scn_queue_obj,
902 dsl_dataset_phys(ds)->ds_next_snap_obj,
903 mintxg, tx) == 0);
904 zfs_dbgmsg("destroying ds %llu; in queue; "
905 "replacing with %llu",
906 (u_longlong_t)ds->ds_object,
907 (u_longlong_t)dsl_dataset_phys(ds)->
908 ds_next_snap_obj);
909 } else {
910 zfs_dbgmsg("destroying ds %llu; in queue; removing",
911 (u_longlong_t)ds->ds_object);
912 }
913 }
914
915 /*
916 * dsl_scan_sync() should be called after this, and should sync
917 * out our changed state, but just to be safe, do it here.
918 */
919 dsl_scan_sync_state(scn, tx);
920 }
921
922 void
dsl_scan_ds_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)923 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
924 {
925 dsl_pool_t *dp = ds->ds_dir->dd_pool;
926 dsl_scan_t *scn = dp->dp_scan;
927 uint64_t mintxg;
928
929 if (scn->scn_phys.scn_state != DSS_SCANNING)
930 return;
931
932 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
933
934 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
935 scn->scn_phys.scn_bookmark.zb_objset =
936 dsl_dataset_phys(ds)->ds_prev_snap_obj;
937 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
938 "reset zb_objset to %llu",
939 (u_longlong_t)ds->ds_object,
940 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
941 } else if (zap_lookup_int_key(dp->dp_meta_objset,
942 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
943 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
944 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
945 VERIFY(zap_add_int_key(dp->dp_meta_objset,
946 scn->scn_phys.scn_queue_obj,
947 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
948 zfs_dbgmsg("snapshotting ds %llu; in queue; "
949 "replacing with %llu",
950 (u_longlong_t)ds->ds_object,
951 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
952 }
953 dsl_scan_sync_state(scn, tx);
954 }
955
956 void
dsl_scan_ds_clone_swapped(dsl_dataset_t * ds1,dsl_dataset_t * ds2,dmu_tx_t * tx)957 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
958 {
959 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
960 dsl_scan_t *scn = dp->dp_scan;
961 uint64_t mintxg;
962
963 if (scn->scn_phys.scn_state != DSS_SCANNING)
964 return;
965
966 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
967 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
968 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
969 "reset zb_objset to %llu",
970 (u_longlong_t)ds1->ds_object,
971 (u_longlong_t)ds2->ds_object);
972 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
973 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
974 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
975 "reset zb_objset to %llu",
976 (u_longlong_t)ds2->ds_object,
977 (u_longlong_t)ds1->ds_object);
978 }
979
980 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
981 ds1->ds_object, &mintxg) == 0) {
982 int err;
983
984 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
985 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
986 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
987 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
988 err = zap_add_int_key(dp->dp_meta_objset,
989 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
990 VERIFY(err == 0 || err == EEXIST);
991 if (err == EEXIST) {
992 /* Both were there to begin with */
993 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
994 scn->scn_phys.scn_queue_obj,
995 ds1->ds_object, mintxg, tx));
996 }
997 zfs_dbgmsg("clone_swap ds %llu; in queue; "
998 "replacing with %llu",
999 (u_longlong_t)ds1->ds_object,
1000 (u_longlong_t)ds2->ds_object);
1001 } else if (zap_lookup_int_key(dp->dp_meta_objset,
1002 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1003 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
1004 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1005 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1006 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1007 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1008 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1009 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1010 "replacing with %llu",
1011 (u_longlong_t)ds2->ds_object,
1012 (u_longlong_t)ds1->ds_object);
1013 }
1014
1015 dsl_scan_sync_state(scn, tx);
1016 }
1017
1018 struct enqueue_clones_arg {
1019 dmu_tx_t *tx;
1020 uint64_t originobj;
1021 };
1022
1023 /* ARGSUSED */
1024 static int
enqueue_clones_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)1025 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1026 {
1027 struct enqueue_clones_arg *eca = arg;
1028 dsl_dataset_t *ds;
1029 int err;
1030 dsl_scan_t *scn = dp->dp_scan;
1031
1032 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1033 return (0);
1034
1035 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1036 if (err)
1037 return (err);
1038
1039 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1040 dsl_dataset_t *prev;
1041 err = dsl_dataset_hold_obj(dp,
1042 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1043
1044 dsl_dataset_rele(ds, FTAG);
1045 if (err)
1046 return (err);
1047 ds = prev;
1048 }
1049 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1050 scn->scn_phys.scn_queue_obj, ds->ds_object,
1051 dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1052 dsl_dataset_rele(ds, FTAG);
1053 return (0);
1054 }
1055
1056 static void
dsl_scan_visitds(dsl_scan_t * scn,uint64_t dsobj,dmu_tx_t * tx)1057 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1058 {
1059 dsl_pool_t *dp = scn->scn_dp;
1060 dsl_dataset_t *ds;
1061 objset_t *os;
1062
1063 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1064
1065 if (scn->scn_phys.scn_cur_min_txg >=
1066 scn->scn_phys.scn_max_txg) {
1067 /*
1068 * This can happen if this snapshot was created after the
1069 * scan started, and we already completed a previous snapshot
1070 * that was created after the scan started. This snapshot
1071 * only references blocks with:
1072 *
1073 * birth < our ds_creation_txg
1074 * cur_min_txg is no less than ds_creation_txg.
1075 * We have already visited these blocks.
1076 * or
1077 * birth > scn_max_txg
1078 * The scan requested not to visit these blocks.
1079 *
1080 * Subsequent snapshots (and clones) can reference our
1081 * blocks, or blocks with even higher birth times.
1082 * Therefore we do not need to visit them either,
1083 * so we do not add them to the work queue.
1084 *
1085 * Note that checking for cur_min_txg >= cur_max_txg
1086 * is not sufficient, because in that case we may need to
1087 * visit subsequent snapshots. This happens when min_txg > 0,
1088 * which raises cur_min_txg. In this case we will visit
1089 * this dataset but skip all of its blocks, because the
1090 * rootbp's birth time is < cur_min_txg. Then we will
1091 * add the next snapshots/clones to the work queue.
1092 */
1093 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1094 dsl_dataset_name(ds, dsname);
1095 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1096 "cur_min_txg (%llu) >= max_txg (%llu)",
1097 dsobj, dsname,
1098 scn->scn_phys.scn_cur_min_txg,
1099 scn->scn_phys.scn_max_txg);
1100 kmem_free(dsname, MAXNAMELEN);
1101
1102 goto out;
1103 }
1104
1105 if (dmu_objset_from_ds(ds, &os))
1106 goto out;
1107
1108 /*
1109 * Only the ZIL in the head (non-snapshot) is valid. Even though
1110 * snapshots can have ZIL block pointers (which may be the same
1111 * BP as in the head), they must be ignored. So we traverse the
1112 * ZIL here, rather than in scan_recurse(), because the regular
1113 * snapshot block-sharing rules don't apply to it.
1114 */
1115 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1116 dsl_scan_zil(dp, &os->os_zil_header);
1117
1118 /*
1119 * Iterate over the bps in this ds.
1120 */
1121 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1122 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1123 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1124 rrw_exit(&ds->ds_bp_rwlock, FTAG);
1125
1126 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
1127 dsl_dataset_name(ds, dsname);
1128 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1129 "pausing=%u",
1130 (longlong_t)dsobj, dsname,
1131 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1132 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1133 (int)scn->scn_pausing);
1134 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
1135
1136 if (scn->scn_pausing)
1137 goto out;
1138
1139 /*
1140 * We've finished this pass over this dataset.
1141 */
1142
1143 /*
1144 * If we did not completely visit this dataset, do another pass.
1145 */
1146 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1147 zfs_dbgmsg("incomplete pass; visiting again");
1148 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1149 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1150 scn->scn_phys.scn_queue_obj, ds->ds_object,
1151 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1152 goto out;
1153 }
1154
1155 /*
1156 * Add descendent datasets to work queue.
1157 */
1158 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1159 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1160 scn->scn_phys.scn_queue_obj,
1161 dsl_dataset_phys(ds)->ds_next_snap_obj,
1162 dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1163 }
1164 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1165 boolean_t usenext = B_FALSE;
1166 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1167 uint64_t count;
1168 /*
1169 * A bug in a previous version of the code could
1170 * cause upgrade_clones_cb() to not set
1171 * ds_next_snap_obj when it should, leading to a
1172 * missing entry. Therefore we can only use the
1173 * next_clones_obj when its count is correct.
1174 */
1175 int err = zap_count(dp->dp_meta_objset,
1176 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1177 if (err == 0 &&
1178 count == dsl_dataset_phys(ds)->ds_num_children - 1)
1179 usenext = B_TRUE;
1180 }
1181
1182 if (usenext) {
1183 VERIFY0(zap_join_key(dp->dp_meta_objset,
1184 dsl_dataset_phys(ds)->ds_next_clones_obj,
1185 scn->scn_phys.scn_queue_obj,
1186 dsl_dataset_phys(ds)->ds_creation_txg, tx));
1187 } else {
1188 struct enqueue_clones_arg eca;
1189 eca.tx = tx;
1190 eca.originobj = ds->ds_object;
1191
1192 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1193 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1194 }
1195 }
1196
1197 out:
1198 dsl_dataset_rele(ds, FTAG);
1199 }
1200
1201 /* ARGSUSED */
1202 static int
enqueue_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)1203 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1204 {
1205 dmu_tx_t *tx = arg;
1206 dsl_dataset_t *ds;
1207 int err;
1208 dsl_scan_t *scn = dp->dp_scan;
1209
1210 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1211 if (err)
1212 return (err);
1213
1214 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1215 dsl_dataset_t *prev;
1216 err = dsl_dataset_hold_obj(dp,
1217 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1218 if (err) {
1219 dsl_dataset_rele(ds, FTAG);
1220 return (err);
1221 }
1222
1223 /*
1224 * If this is a clone, we don't need to worry about it for now.
1225 */
1226 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1227 dsl_dataset_rele(ds, FTAG);
1228 dsl_dataset_rele(prev, FTAG);
1229 return (0);
1230 }
1231 dsl_dataset_rele(ds, FTAG);
1232 ds = prev;
1233 }
1234
1235 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1236 ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1237 dsl_dataset_rele(ds, FTAG);
1238 return (0);
1239 }
1240
1241 /*
1242 * Scrub/dedup interaction.
1243 *
1244 * If there are N references to a deduped block, we don't want to scrub it
1245 * N times -- ideally, we should scrub it exactly once.
1246 *
1247 * We leverage the fact that the dde's replication class (enum ddt_class)
1248 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1249 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1250 *
1251 * To prevent excess scrubbing, the scrub begins by walking the DDT
1252 * to find all blocks with refcnt > 1, and scrubs each of these once.
1253 * Since there are two replication classes which contain blocks with
1254 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1255 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1256 *
1257 * There would be nothing more to say if a block's refcnt couldn't change
1258 * during a scrub, but of course it can so we must account for changes
1259 * in a block's replication class.
1260 *
1261 * Here's an example of what can occur:
1262 *
1263 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1264 * when visited during the top-down scrub phase, it will be scrubbed twice.
1265 * This negates our scrub optimization, but is otherwise harmless.
1266 *
1267 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1268 * on each visit during the top-down scrub phase, it will never be scrubbed.
1269 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1270 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1271 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1272 * while a scrub is in progress, it scrubs the block right then.
1273 */
1274 static void
dsl_scan_ddt(dsl_scan_t * scn,dmu_tx_t * tx)1275 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1276 {
1277 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1278 ddt_entry_t dde = { 0 };
1279 int error;
1280 uint64_t n = 0;
1281
1282 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1283 ddt_t *ddt;
1284
1285 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1286 break;
1287 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1288 (longlong_t)ddb->ddb_class,
1289 (longlong_t)ddb->ddb_type,
1290 (longlong_t)ddb->ddb_checksum,
1291 (longlong_t)ddb->ddb_cursor);
1292
1293 /* There should be no pending changes to the dedup table */
1294 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1295 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1296
1297 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1298 n++;
1299
1300 if (dsl_scan_check_pause(scn, NULL))
1301 break;
1302 }
1303
1304 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1305 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1306 (int)scn->scn_pausing);
1307
1308 ASSERT(error == 0 || error == ENOENT);
1309 ASSERT(error != ENOENT ||
1310 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1311 }
1312
1313 /* ARGSUSED */
1314 void
dsl_scan_ddt_entry(dsl_scan_t * scn,enum zio_checksum checksum,ddt_entry_t * dde,dmu_tx_t * tx)1315 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1316 ddt_entry_t *dde, dmu_tx_t *tx)
1317 {
1318 const ddt_key_t *ddk = &dde->dde_key;
1319 ddt_phys_t *ddp = dde->dde_phys;
1320 blkptr_t bp;
1321 zbookmark_phys_t zb = { 0 };
1322
1323 if (scn->scn_phys.scn_state != DSS_SCANNING)
1324 return;
1325
1326 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1327 if (ddp->ddp_phys_birth == 0 ||
1328 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1329 continue;
1330 ddt_bp_create(checksum, ddk, ddp, &bp);
1331
1332 scn->scn_visited_this_txg++;
1333 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1334 }
1335 }
1336
1337 static void
dsl_scan_visit(dsl_scan_t * scn,dmu_tx_t * tx)1338 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1339 {
1340 dsl_pool_t *dp = scn->scn_dp;
1341 zap_cursor_t zc;
1342 zap_attribute_t za;
1343
1344 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1345 scn->scn_phys.scn_ddt_class_max) {
1346 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1347 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1348 dsl_scan_ddt(scn, tx);
1349 if (scn->scn_pausing)
1350 return;
1351 }
1352
1353 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1354 /* First do the MOS & ORIGIN */
1355
1356 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1357 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1358 dsl_scan_visit_rootbp(scn, NULL,
1359 &dp->dp_meta_rootbp, tx);
1360 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1361 if (scn->scn_pausing)
1362 return;
1363
1364 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1365 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1366 enqueue_cb, tx, DS_FIND_CHILDREN));
1367 } else {
1368 dsl_scan_visitds(scn,
1369 dp->dp_origin_snap->ds_object, tx);
1370 }
1371 ASSERT(!scn->scn_pausing);
1372 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1373 ZB_DESTROYED_OBJSET) {
1374 /*
1375 * If we were paused, continue from here. Note if the
1376 * ds we were paused on was deleted, the zb_objset may
1377 * be -1, so we will skip this and find a new objset
1378 * below.
1379 */
1380 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1381 if (scn->scn_pausing)
1382 return;
1383 }
1384
1385 /*
1386 * In case we were paused right at the end of the ds, zero the
1387 * bookmark so we don't think that we're still trying to resume.
1388 */
1389 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1390
1391 /* keep pulling things out of the zap-object-as-queue */
1392 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1393 scn->scn_phys.scn_queue_obj),
1394 zap_cursor_retrieve(&zc, &za) == 0) {
1395 dsl_dataset_t *ds;
1396 uint64_t dsobj;
1397
1398 dsobj = strtonum(za.za_name, NULL);
1399 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1400 scn->scn_phys.scn_queue_obj, dsobj, tx));
1401
1402 /* Set up min/max txg */
1403 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1404 if (za.za_first_integer != 0) {
1405 scn->scn_phys.scn_cur_min_txg =
1406 MAX(scn->scn_phys.scn_min_txg,
1407 za.za_first_integer);
1408 } else {
1409 scn->scn_phys.scn_cur_min_txg =
1410 MAX(scn->scn_phys.scn_min_txg,
1411 dsl_dataset_phys(ds)->ds_prev_snap_txg);
1412 }
1413 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1414 dsl_dataset_rele(ds, FTAG);
1415
1416 dsl_scan_visitds(scn, dsobj, tx);
1417 zap_cursor_fini(&zc);
1418 if (scn->scn_pausing)
1419 return;
1420 }
1421 zap_cursor_fini(&zc);
1422 }
1423
1424 static boolean_t
dsl_scan_free_should_pause(dsl_scan_t * scn)1425 dsl_scan_free_should_pause(dsl_scan_t *scn)
1426 {
1427 uint64_t elapsed_nanosecs;
1428
1429 if (zfs_recover)
1430 return (B_FALSE);
1431
1432 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1433 return (B_TRUE);
1434
1435 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1436 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1437 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1438 txg_sync_waiting(scn->scn_dp)) ||
1439 spa_shutting_down(scn->scn_dp->dp_spa));
1440 }
1441
1442 static int
dsl_scan_free_block_cb(void * arg,const blkptr_t * bp,dmu_tx_t * tx)1443 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1444 {
1445 dsl_scan_t *scn = arg;
1446
1447 if (!scn->scn_is_bptree ||
1448 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1449 if (dsl_scan_free_should_pause(scn))
1450 return (SET_ERROR(ERESTART));
1451 }
1452
1453 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1454 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1455 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1456 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1457 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1458 scn->scn_visited_this_txg++;
1459 return (0);
1460 }
1461
1462 boolean_t
dsl_scan_active(dsl_scan_t * scn)1463 dsl_scan_active(dsl_scan_t *scn)
1464 {
1465 spa_t *spa = scn->scn_dp->dp_spa;
1466 uint64_t used = 0, comp, uncomp;
1467
1468 if (spa->spa_load_state != SPA_LOAD_NONE)
1469 return (B_FALSE);
1470 if (spa_shutting_down(spa))
1471 return (B_FALSE);
1472 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1473 (scn->scn_async_destroying && !scn->scn_async_stalled))
1474 return (B_TRUE);
1475
1476 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1477 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1478 &used, &comp, &uncomp);
1479 }
1480 return (used != 0);
1481 }
1482
1483 void
dsl_scan_sync(dsl_pool_t * dp,dmu_tx_t * tx)1484 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1485 {
1486 dsl_scan_t *scn = dp->dp_scan;
1487 spa_t *spa = dp->dp_spa;
1488 int err = 0;
1489
1490 /*
1491 * Check for scn_restart_txg before checking spa_load_state, so
1492 * that we can restart an old-style scan while the pool is being
1493 * imported (see dsl_scan_init).
1494 */
1495 if (dsl_scan_restarting(scn, tx)) {
1496 pool_scan_func_t func = POOL_SCAN_SCRUB;
1497 dsl_scan_done(scn, B_FALSE, tx);
1498 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1499 func = POOL_SCAN_RESILVER;
1500 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1501 func, tx->tx_txg);
1502 dsl_scan_setup_sync(&func, tx);
1503 }
1504
1505 /*
1506 * Only process scans in sync pass 1.
1507 */
1508 if (spa_sync_pass(dp->dp_spa) > 1)
1509 return;
1510
1511 /*
1512 * If the spa is shutting down, then stop scanning. This will
1513 * ensure that the scan does not dirty any new data during the
1514 * shutdown phase.
1515 */
1516 if (spa_shutting_down(spa))
1517 return;
1518
1519 /*
1520 * If the scan is inactive due to a stalled async destroy, try again.
1521 */
1522 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1523 return;
1524
1525 scn->scn_visited_this_txg = 0;
1526 scn->scn_pausing = B_FALSE;
1527 scn->scn_sync_start_time = gethrtime();
1528 spa->spa_scrub_active = B_TRUE;
1529
1530 /*
1531 * First process the async destroys. If we pause, don't do
1532 * any scrubbing or resilvering. This ensures that there are no
1533 * async destroys while we are scanning, so the scan code doesn't
1534 * have to worry about traversing it. It is also faster to free the
1535 * blocks than to scrub them.
1536 */
1537 if (zfs_free_bpobj_enabled &&
1538 spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1539 scn->scn_is_bptree = B_FALSE;
1540 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1541 NULL, ZIO_FLAG_MUSTSUCCEED);
1542 err = bpobj_iterate(&dp->dp_free_bpobj,
1543 dsl_scan_free_block_cb, scn, tx);
1544 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1545
1546 if (err != 0 && err != ERESTART)
1547 zfs_panic_recover("error %u from bpobj_iterate()", err);
1548 }
1549
1550 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1551 ASSERT(scn->scn_async_destroying);
1552 scn->scn_is_bptree = B_TRUE;
1553 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1554 NULL, ZIO_FLAG_MUSTSUCCEED);
1555 err = bptree_iterate(dp->dp_meta_objset,
1556 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1557 VERIFY0(zio_wait(scn->scn_zio_root));
1558
1559 if (err == EIO || err == ECKSUM) {
1560 err = 0;
1561 } else if (err != 0 && err != ERESTART) {
1562 zfs_panic_recover("error %u from "
1563 "traverse_dataset_destroyed()", err);
1564 }
1565
1566 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1567 /* finished; deactivate async destroy feature */
1568 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1569 ASSERT(!spa_feature_is_active(spa,
1570 SPA_FEATURE_ASYNC_DESTROY));
1571 VERIFY0(zap_remove(dp->dp_meta_objset,
1572 DMU_POOL_DIRECTORY_OBJECT,
1573 DMU_POOL_BPTREE_OBJ, tx));
1574 VERIFY0(bptree_free(dp->dp_meta_objset,
1575 dp->dp_bptree_obj, tx));
1576 dp->dp_bptree_obj = 0;
1577 scn->scn_async_destroying = B_FALSE;
1578 scn->scn_async_stalled = B_FALSE;
1579 } else {
1580 /*
1581 * If we didn't make progress, mark the async
1582 * destroy as stalled, so that we will not initiate
1583 * a spa_sync() on its behalf. Note that we only
1584 * check this if we are not finished, because if the
1585 * bptree had no blocks for us to visit, we can
1586 * finish without "making progress".
1587 */
1588 scn->scn_async_stalled =
1589 (scn->scn_visited_this_txg == 0);
1590 }
1591 }
1592 if (scn->scn_visited_this_txg) {
1593 zfs_dbgmsg("freed %llu blocks in %llums from "
1594 "free_bpobj/bptree txg %llu; err=%d",
1595 (longlong_t)scn->scn_visited_this_txg,
1596 (longlong_t)
1597 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1598 (longlong_t)tx->tx_txg, err);
1599 scn->scn_visited_this_txg = 0;
1600
1601 /*
1602 * Write out changes to the DDT that may be required as a
1603 * result of the blocks freed. This ensures that the DDT
1604 * is clean when a scrub/resilver runs.
1605 */
1606 ddt_sync(spa, tx->tx_txg);
1607 }
1608 if (err != 0)
1609 return;
1610 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1611 zfs_free_leak_on_eio &&
1612 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1613 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1614 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1615 /*
1616 * We have finished background destroying, but there is still
1617 * some space left in the dp_free_dir. Transfer this leaked
1618 * space to the dp_leak_dir.
1619 */
1620 if (dp->dp_leak_dir == NULL) {
1621 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1622 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1623 LEAK_DIR_NAME, tx);
1624 VERIFY0(dsl_pool_open_special_dir(dp,
1625 LEAK_DIR_NAME, &dp->dp_leak_dir));
1626 rrw_exit(&dp->dp_config_rwlock, FTAG);
1627 }
1628 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1629 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1630 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1631 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1632 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1633 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1634 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1635 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1636 }
1637 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1638 /* finished; verify that space accounting went to zero */
1639 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1640 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1641 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1642 }
1643
1644 if (scn->scn_phys.scn_state != DSS_SCANNING)
1645 return;
1646
1647 if (scn->scn_done_txg == tx->tx_txg) {
1648 ASSERT(!scn->scn_pausing);
1649 /* finished with scan. */
1650 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1651 dsl_scan_done(scn, B_TRUE, tx);
1652 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1653 dsl_scan_sync_state(scn, tx);
1654 return;
1655 }
1656
1657 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1658 scn->scn_phys.scn_ddt_class_max) {
1659 zfs_dbgmsg("doing scan sync txg %llu; "
1660 "ddt bm=%llu/%llu/%llu/%llx",
1661 (longlong_t)tx->tx_txg,
1662 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1663 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1664 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1665 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1666 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1667 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1668 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1669 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1670 } else {
1671 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1672 (longlong_t)tx->tx_txg,
1673 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1674 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1675 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1676 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1677 }
1678
1679 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1680 NULL, ZIO_FLAG_CANFAIL);
1681 dsl_pool_config_enter(dp, FTAG);
1682 dsl_scan_visit(scn, tx);
1683 dsl_pool_config_exit(dp, FTAG);
1684 (void) zio_wait(scn->scn_zio_root);
1685 scn->scn_zio_root = NULL;
1686
1687 zfs_dbgmsg("visited %llu blocks in %llums",
1688 (longlong_t)scn->scn_visited_this_txg,
1689 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1690
1691 if (!scn->scn_pausing) {
1692 scn->scn_done_txg = tx->tx_txg + 1;
1693 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1694 tx->tx_txg, scn->scn_done_txg);
1695 }
1696
1697 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1698 mutex_enter(&spa->spa_scrub_lock);
1699 while (spa->spa_scrub_inflight > 0) {
1700 cv_wait(&spa->spa_scrub_io_cv,
1701 &spa->spa_scrub_lock);
1702 }
1703 mutex_exit(&spa->spa_scrub_lock);
1704 }
1705
1706 dsl_scan_sync_state(scn, tx);
1707 }
1708
1709 /*
1710 * This will start a new scan, or restart an existing one.
1711 */
1712 void
dsl_resilver_restart(dsl_pool_t * dp,uint64_t txg)1713 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1714 {
1715 if (txg == 0) {
1716 dmu_tx_t *tx;
1717 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1718 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1719
1720 txg = dmu_tx_get_txg(tx);
1721 dp->dp_scan->scn_restart_txg = txg;
1722 dmu_tx_commit(tx);
1723 } else {
1724 dp->dp_scan->scn_restart_txg = txg;
1725 }
1726 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1727 }
1728
1729 boolean_t
dsl_scan_resilvering(dsl_pool_t * dp)1730 dsl_scan_resilvering(dsl_pool_t *dp)
1731 {
1732 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1733 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1734 }
1735
1736 /*
1737 * scrub consumers
1738 */
1739
1740 static void
count_block(zfs_all_blkstats_t * zab,const blkptr_t * bp)1741 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1742 {
1743 int i;
1744
1745 /*
1746 * If we resume after a reboot, zab will be NULL; don't record
1747 * incomplete stats in that case.
1748 */
1749 if (zab == NULL)
1750 return;
1751
1752 for (i = 0; i < 4; i++) {
1753 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1754 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1755 if (t & DMU_OT_NEWTYPE)
1756 t = DMU_OT_OTHER;
1757 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1758 int equal;
1759
1760 zb->zb_count++;
1761 zb->zb_asize += BP_GET_ASIZE(bp);
1762 zb->zb_lsize += BP_GET_LSIZE(bp);
1763 zb->zb_psize += BP_GET_PSIZE(bp);
1764 zb->zb_gangs += BP_COUNT_GANG(bp);
1765
1766 switch (BP_GET_NDVAS(bp)) {
1767 case 2:
1768 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1769 DVA_GET_VDEV(&bp->blk_dva[1]))
1770 zb->zb_ditto_2_of_2_samevdev++;
1771 break;
1772 case 3:
1773 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1774 DVA_GET_VDEV(&bp->blk_dva[1])) +
1775 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1776 DVA_GET_VDEV(&bp->blk_dva[2])) +
1777 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1778 DVA_GET_VDEV(&bp->blk_dva[2]));
1779 if (equal == 1)
1780 zb->zb_ditto_2_of_3_samevdev++;
1781 else if (equal == 3)
1782 zb->zb_ditto_3_of_3_samevdev++;
1783 break;
1784 }
1785 }
1786 }
1787
1788 static void
dsl_scan_scrub_done(zio_t * zio)1789 dsl_scan_scrub_done(zio_t *zio)
1790 {
1791 spa_t *spa = zio->io_spa;
1792
1793 zio_data_buf_free(zio->io_data, zio->io_size);
1794
1795 mutex_enter(&spa->spa_scrub_lock);
1796 spa->spa_scrub_inflight--;
1797 cv_broadcast(&spa->spa_scrub_io_cv);
1798
1799 if (zio->io_error && (zio->io_error != ECKSUM ||
1800 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1801 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1802 }
1803 mutex_exit(&spa->spa_scrub_lock);
1804 }
1805
1806 static int
dsl_scan_scrub_cb(dsl_pool_t * dp,const blkptr_t * bp,const zbookmark_phys_t * zb)1807 dsl_scan_scrub_cb(dsl_pool_t *dp,
1808 const blkptr_t *bp, const zbookmark_phys_t *zb)
1809 {
1810 dsl_scan_t *scn = dp->dp_scan;
1811 size_t size = BP_GET_PSIZE(bp);
1812 spa_t *spa = dp->dp_spa;
1813 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1814 boolean_t needs_io;
1815 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1816 unsigned int scan_delay = 0;
1817
1818 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1819 phys_birth >= scn->scn_phys.scn_max_txg)
1820 return (0);
1821
1822 count_block(dp->dp_blkstats, bp);
1823
1824 if (BP_IS_EMBEDDED(bp))
1825 return (0);
1826
1827 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1828 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1829 zio_flags |= ZIO_FLAG_SCRUB;
1830 needs_io = B_TRUE;
1831 scan_delay = zfs_scrub_delay;
1832 } else {
1833 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1834 zio_flags |= ZIO_FLAG_RESILVER;
1835 needs_io = B_FALSE;
1836 scan_delay = zfs_resilver_delay;
1837 }
1838
1839 /* If it's an intent log block, failure is expected. */
1840 if (zb->zb_level == ZB_ZIL_LEVEL)
1841 zio_flags |= ZIO_FLAG_SPECULATIVE;
1842
1843 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1844 vdev_t *vd = vdev_lookup_top(spa,
1845 DVA_GET_VDEV(&bp->blk_dva[d]));
1846
1847 /*
1848 * Keep track of how much data we've examined so that
1849 * zpool(1M) status can make useful progress reports.
1850 */
1851 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1852 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1853
1854 /* if it's a resilver, this may not be in the target range */
1855 if (!needs_io) {
1856 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1857 /*
1858 * Gang members may be spread across multiple
1859 * vdevs, so the best estimate we have is the
1860 * scrub range, which has already been checked.
1861 * XXX -- it would be better to change our
1862 * allocation policy to ensure that all
1863 * gang members reside on the same vdev.
1864 */
1865 needs_io = B_TRUE;
1866 } else {
1867 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1868 phys_birth, 1);
1869 }
1870 }
1871 }
1872
1873 if (needs_io && !zfs_no_scrub_io) {
1874 vdev_t *rvd = spa->spa_root_vdev;
1875 uint64_t maxinflight = rvd->vdev_children *
1876 MAX(zfs_top_maxinflight, 1);
1877 void *data = zio_data_buf_alloc(size);
1878
1879 mutex_enter(&spa->spa_scrub_lock);
1880 while (spa->spa_scrub_inflight >= maxinflight)
1881 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1882 spa->spa_scrub_inflight++;
1883 mutex_exit(&spa->spa_scrub_lock);
1884
1885 /*
1886 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1887 * then throttle our workload to limit the impact of a scan.
1888 */
1889 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1890 delay(MAX((int)scan_delay, 0));
1891
1892 zio_nowait(zio_read(NULL, spa, bp, data, size,
1893 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1894 zio_flags, zb));
1895 }
1896
1897 /* do not relocate this block */
1898 return (0);
1899 }
1900
1901 int
dsl_scan(dsl_pool_t * dp,pool_scan_func_t func)1902 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1903 {
1904 spa_t *spa = dp->dp_spa;
1905
1906 /*
1907 * Purge all vdev caches and probe all devices. We do this here
1908 * rather than in sync context because this requires a writer lock
1909 * on the spa_config lock, which we can't do from sync context. The
1910 * spa_scrub_reopen flag indicates that vdev_open() should not
1911 * attempt to start another scrub.
1912 */
1913 spa_vdev_state_enter(spa, SCL_NONE);
1914 spa->spa_scrub_reopen = B_TRUE;
1915 vdev_reopen(spa->spa_root_vdev);
1916 spa->spa_scrub_reopen = B_FALSE;
1917 (void) spa_vdev_state_exit(spa, NULL, 0);
1918
1919 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1920 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1921 }
1922
1923 static boolean_t
dsl_scan_restarting(dsl_scan_t * scn,dmu_tx_t * tx)1924 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
1925 {
1926 return (scn->scn_restart_txg != 0 &&
1927 scn->scn_restart_txg <= tx->tx_txg);
1928 }
1929