Print this page
6537 Panic on zpool scrub with DEBUG kernel
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/dsl_scan.c
+++ new/usr/src/uts/common/fs/zfs/dsl_scan.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 + * Copyright 2016 Gary Mills
24 25 */
25 26
26 27 #include <sys/dsl_scan.h>
27 28 #include <sys/dsl_pool.h>
28 29 #include <sys/dsl_dataset.h>
29 30 #include <sys/dsl_prop.h>
30 31 #include <sys/dsl_dir.h>
31 32 #include <sys/dsl_synctask.h>
32 33 #include <sys/dnode.h>
33 34 #include <sys/dmu_tx.h>
34 35 #include <sys/dmu_objset.h>
35 36 #include <sys/arc.h>
36 37 #include <sys/zap.h>
37 38 #include <sys/zio.h>
38 39 #include <sys/zfs_context.h>
39 40 #include <sys/fs/zfs.h>
40 41 #include <sys/zfs_znode.h>
41 42 #include <sys/spa_impl.h>
42 43 #include <sys/vdev_impl.h>
43 44 #include <sys/zil_impl.h>
44 45 #include <sys/zio_checksum.h>
45 46 #include <sys/ddt.h>
46 47 #include <sys/sa.h>
47 48 #include <sys/sa_impl.h>
48 49 #include <sys/zfeature.h>
49 50 #ifdef _KERNEL
50 51 #include <sys/zfs_vfsops.h>
51 52 #endif
52 53
53 54 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
54 55 const zbookmark_phys_t *);
55 56
56 57 static scan_cb_t dsl_scan_scrub_cb;
57 58 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
58 59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
59 60
60 61 int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
61 62 int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
62 63 int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
63 64 int zfs_scan_idle = 50; /* idle window in clock ticks */
64 65
65 66 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66 67 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
67 68 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
68 69 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
69 70 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
70 71 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
71 72 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
72 73 /* max number of blocks to free in a single TXG */
73 74 uint64_t zfs_free_max_blocks = UINT64_MAX;
74 75
75 76 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
76 77 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
77 78 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
78 79
79 80 extern int zfs_txg_timeout;
80 81
81 82 /*
82 83 * Enable/disable the processing of the free_bpobj object.
83 84 */
84 85 boolean_t zfs_free_bpobj_enabled = B_TRUE;
85 86
86 87 /* the order has to match pool_scan_type */
87 88 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
88 89 NULL,
89 90 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
90 91 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
91 92 };
92 93
93 94 int
94 95 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
95 96 {
96 97 int err;
97 98 dsl_scan_t *scn;
98 99 spa_t *spa = dp->dp_spa;
99 100 uint64_t f;
100 101
101 102 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
102 103 scn->scn_dp = dp;
103 104
104 105 /*
105 106 * It's possible that we're resuming a scan after a reboot so
106 107 * make sure that the scan_async_destroying flag is initialized
107 108 * appropriately.
108 109 */
109 110 ASSERT(!scn->scn_async_destroying);
110 111 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
111 112 SPA_FEATURE_ASYNC_DESTROY);
112 113
113 114 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
114 115 "scrub_func", sizeof (uint64_t), 1, &f);
115 116 if (err == 0) {
116 117 /*
117 118 * There was an old-style scrub in progress. Restart a
118 119 * new-style scrub from the beginning.
119 120 */
120 121 scn->scn_restart_txg = txg;
121 122 zfs_dbgmsg("old-style scrub was in progress; "
122 123 "restarting new-style scrub in txg %llu",
123 124 scn->scn_restart_txg);
124 125
125 126 /*
126 127 * Load the queue obj from the old location so that it
127 128 * can be freed by dsl_scan_done().
128 129 */
129 130 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
130 131 "scrub_queue", sizeof (uint64_t), 1,
131 132 &scn->scn_phys.scn_queue_obj);
132 133 } else {
133 134 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
134 135 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
135 136 &scn->scn_phys);
136 137 if (err == ENOENT)
137 138 return (0);
138 139 else if (err)
139 140 return (err);
140 141
141 142 if (scn->scn_phys.scn_state == DSS_SCANNING &&
142 143 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
143 144 /*
144 145 * A new-type scrub was in progress on an old
145 146 * pool, and the pool was accessed by old
146 147 * software. Restart from the beginning, since
147 148 * the old software may have changed the pool in
148 149 * the meantime.
149 150 */
150 151 scn->scn_restart_txg = txg;
151 152 zfs_dbgmsg("new-style scrub was modified "
152 153 "by old software; restarting in txg %llu",
153 154 scn->scn_restart_txg);
154 155 }
155 156 }
156 157
157 158 spa_scan_stat_init(spa);
158 159 return (0);
159 160 }
160 161
161 162 void
162 163 dsl_scan_fini(dsl_pool_t *dp)
163 164 {
164 165 if (dp->dp_scan) {
165 166 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
166 167 dp->dp_scan = NULL;
167 168 }
168 169 }
169 170
170 171 /* ARGSUSED */
171 172 static int
172 173 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
173 174 {
174 175 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
175 176
176 177 if (scn->scn_phys.scn_state == DSS_SCANNING)
177 178 return (SET_ERROR(EBUSY));
178 179
179 180 return (0);
180 181 }
181 182
182 183 static void
183 184 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
184 185 {
185 186 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
186 187 pool_scan_func_t *funcp = arg;
187 188 dmu_object_type_t ot = 0;
188 189 dsl_pool_t *dp = scn->scn_dp;
189 190 spa_t *spa = dp->dp_spa;
190 191
191 192 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
192 193 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
193 194 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
194 195 scn->scn_phys.scn_func = *funcp;
195 196 scn->scn_phys.scn_state = DSS_SCANNING;
196 197 scn->scn_phys.scn_min_txg = 0;
197 198 scn->scn_phys.scn_max_txg = tx->tx_txg;
198 199 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
199 200 scn->scn_phys.scn_start_time = gethrestime_sec();
200 201 scn->scn_phys.scn_errors = 0;
201 202 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
202 203 scn->scn_restart_txg = 0;
203 204 scn->scn_done_txg = 0;
204 205 spa_scan_stat_init(spa);
205 206
206 207 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
207 208 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
208 209
209 210 /* rewrite all disk labels */
210 211 vdev_config_dirty(spa->spa_root_vdev);
211 212
212 213 if (vdev_resilver_needed(spa->spa_root_vdev,
213 214 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
214 215 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
215 216 } else {
216 217 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
217 218 }
218 219
219 220 spa->spa_scrub_started = B_TRUE;
220 221 /*
221 222 * If this is an incremental scrub, limit the DDT scrub phase
222 223 * to just the auto-ditto class (for correctness); the rest
223 224 * of the scrub should go faster using top-down pruning.
224 225 */
225 226 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
226 227 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
227 228
228 229 }
229 230
230 231 /* back to the generic stuff */
231 232
232 233 if (dp->dp_blkstats == NULL) {
233 234 dp->dp_blkstats =
234 235 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
235 236 }
236 237 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
237 238
238 239 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
239 240 ot = DMU_OT_ZAP_OTHER;
240 241
241 242 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
242 243 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
243 244
244 245 dsl_scan_sync_state(scn, tx);
245 246
246 247 spa_history_log_internal(spa, "scan setup", tx,
247 248 "func=%u mintxg=%llu maxtxg=%llu",
248 249 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
249 250 }
250 251
251 252 /* ARGSUSED */
252 253 static void
253 254 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
254 255 {
255 256 static const char *old_names[] = {
256 257 "scrub_bookmark",
257 258 "scrub_ddt_bookmark",
258 259 "scrub_ddt_class_max",
259 260 "scrub_queue",
260 261 "scrub_min_txg",
261 262 "scrub_max_txg",
262 263 "scrub_func",
263 264 "scrub_errors",
264 265 NULL
265 266 };
266 267
267 268 dsl_pool_t *dp = scn->scn_dp;
268 269 spa_t *spa = dp->dp_spa;
269 270 int i;
270 271
271 272 /* Remove any remnants of an old-style scrub. */
272 273 for (i = 0; old_names[i]; i++) {
273 274 (void) zap_remove(dp->dp_meta_objset,
274 275 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
275 276 }
276 277
277 278 if (scn->scn_phys.scn_queue_obj != 0) {
278 279 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
279 280 scn->scn_phys.scn_queue_obj, tx));
280 281 scn->scn_phys.scn_queue_obj = 0;
281 282 }
282 283
283 284 /*
284 285 * If we were "restarted" from a stopped state, don't bother
285 286 * with anything else.
286 287 */
287 288 if (scn->scn_phys.scn_state != DSS_SCANNING)
288 289 return;
289 290
290 291 if (complete)
291 292 scn->scn_phys.scn_state = DSS_FINISHED;
292 293 else
293 294 scn->scn_phys.scn_state = DSS_CANCELED;
294 295
295 296 spa_history_log_internal(spa, "scan done", tx,
296 297 "complete=%u", complete);
297 298
298 299 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
299 300 mutex_enter(&spa->spa_scrub_lock);
300 301 while (spa->spa_scrub_inflight > 0) {
301 302 cv_wait(&spa->spa_scrub_io_cv,
302 303 &spa->spa_scrub_lock);
303 304 }
304 305 mutex_exit(&spa->spa_scrub_lock);
305 306 spa->spa_scrub_started = B_FALSE;
306 307 spa->spa_scrub_active = B_FALSE;
307 308
308 309 /*
309 310 * If the scrub/resilver completed, update all DTLs to
310 311 * reflect this. Whether it succeeded or not, vacate
311 312 * all temporary scrub DTLs.
312 313 */
313 314 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
314 315 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
315 316 if (complete) {
316 317 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
317 318 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
318 319 }
319 320 spa_errlog_rotate(spa);
320 321
321 322 /*
322 323 * We may have finished replacing a device.
323 324 * Let the async thread assess this and handle the detach.
324 325 */
325 326 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
326 327 }
327 328
328 329 scn->scn_phys.scn_end_time = gethrestime_sec();
329 330 }
330 331
331 332 /* ARGSUSED */
332 333 static int
333 334 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
334 335 {
335 336 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
336 337
337 338 if (scn->scn_phys.scn_state != DSS_SCANNING)
338 339 return (SET_ERROR(ENOENT));
339 340 return (0);
340 341 }
341 342
342 343 /* ARGSUSED */
343 344 static void
344 345 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
345 346 {
346 347 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
347 348
348 349 dsl_scan_done(scn, B_FALSE, tx);
349 350 dsl_scan_sync_state(scn, tx);
350 351 }
351 352
352 353 int
353 354 dsl_scan_cancel(dsl_pool_t *dp)
354 355 {
355 356 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
356 357 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
357 358 }
358 359
359 360 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
360 361 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
361 362 dmu_objset_type_t ostype, dmu_tx_t *tx);
362 363 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
363 364 dmu_objset_type_t ostype,
364 365 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
365 366
366 367 void
367 368 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
368 369 {
369 370 zio_free(dp->dp_spa, txg, bp);
370 371 }
371 372
372 373 void
373 374 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
374 375 {
375 376 ASSERT(dsl_pool_sync_context(dp));
376 377 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
377 378 }
378 379
379 380 static uint64_t
380 381 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
381 382 {
382 383 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
383 384 if (ds->ds_is_snapshot)
384 385 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
385 386 return (smt);
386 387 }
387 388
388 389 static void
389 390 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
390 391 {
391 392 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
392 393 DMU_POOL_DIRECTORY_OBJECT,
393 394 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
394 395 &scn->scn_phys, tx));
395 396 }
396 397
397 398 extern int zfs_vdev_async_write_active_min_dirty_percent;
398 399
399 400 static boolean_t
400 401 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
401 402 {
402 403 /* we never skip user/group accounting objects */
403 404 if (zb && (int64_t)zb->zb_object < 0)
404 405 return (B_FALSE);
405 406
406 407 if (scn->scn_pausing)
407 408 return (B_TRUE); /* we're already pausing */
408 409
409 410 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
410 411 return (B_FALSE); /* we're resuming */
411 412
412 413 /* We only know how to resume from level-0 blocks. */
413 414 if (zb && zb->zb_level != 0)
414 415 return (B_FALSE);
415 416
416 417 /*
417 418 * We pause if:
418 419 * - we have scanned for the maximum time: an entire txg
419 420 * timeout (default 5 sec)
420 421 * or
421 422 * - we have scanned for at least the minimum time (default 1 sec
422 423 * for scrub, 3 sec for resilver), and either we have sufficient
423 424 * dirty data that we are starting to write more quickly
424 425 * (default 30%), or someone is explicitly waiting for this txg
425 426 * to complete.
426 427 * or
427 428 * - the spa is shutting down because this pool is being exported
428 429 * or the machine is rebooting.
429 430 */
430 431 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
431 432 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
432 433 uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
433 434 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
434 435 if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
435 436 (NSEC2MSEC(elapsed_nanosecs) > mintime &&
436 437 (txg_sync_waiting(scn->scn_dp) ||
437 438 dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
438 439 spa_shutting_down(scn->scn_dp->dp_spa)) {
439 440 if (zb) {
440 441 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
441 442 (longlong_t)zb->zb_objset,
442 443 (longlong_t)zb->zb_object,
443 444 (longlong_t)zb->zb_level,
444 445 (longlong_t)zb->zb_blkid);
445 446 scn->scn_phys.scn_bookmark = *zb;
446 447 }
447 448 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
448 449 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
449 450 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
450 451 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
451 452 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
452 453 scn->scn_pausing = B_TRUE;
453 454 return (B_TRUE);
454 455 }
455 456 return (B_FALSE);
456 457 }
457 458
458 459 typedef struct zil_scan_arg {
459 460 dsl_pool_t *zsa_dp;
460 461 zil_header_t *zsa_zh;
461 462 } zil_scan_arg_t;
462 463
463 464 /* ARGSUSED */
464 465 static int
465 466 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
466 467 {
467 468 zil_scan_arg_t *zsa = arg;
468 469 dsl_pool_t *dp = zsa->zsa_dp;
469 470 dsl_scan_t *scn = dp->dp_scan;
470 471 zil_header_t *zh = zsa->zsa_zh;
471 472 zbookmark_phys_t zb;
472 473
473 474 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
474 475 return (0);
475 476
476 477 /*
477 478 * One block ("stubby") can be allocated a long time ago; we
478 479 * want to visit that one because it has been allocated
479 480 * (on-disk) even if it hasn't been claimed (even though for
480 481 * scrub there's nothing to do to it).
481 482 */
482 483 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
483 484 return (0);
484 485
485 486 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
486 487 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
487 488
488 489 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
489 490 return (0);
490 491 }
491 492
492 493 /* ARGSUSED */
493 494 static int
494 495 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
495 496 {
496 497 if (lrc->lrc_txtype == TX_WRITE) {
497 498 zil_scan_arg_t *zsa = arg;
498 499 dsl_pool_t *dp = zsa->zsa_dp;
499 500 dsl_scan_t *scn = dp->dp_scan;
500 501 zil_header_t *zh = zsa->zsa_zh;
501 502 lr_write_t *lr = (lr_write_t *)lrc;
502 503 blkptr_t *bp = &lr->lr_blkptr;
503 504 zbookmark_phys_t zb;
504 505
505 506 if (BP_IS_HOLE(bp) ||
506 507 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
507 508 return (0);
508 509
509 510 /*
510 511 * birth can be < claim_txg if this record's txg is
511 512 * already txg sync'ed (but this log block contains
512 513 * other records that are not synced)
513 514 */
514 515 if (claim_txg == 0 || bp->blk_birth < claim_txg)
515 516 return (0);
516 517
517 518 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
518 519 lr->lr_foid, ZB_ZIL_LEVEL,
519 520 lr->lr_offset / BP_GET_LSIZE(bp));
520 521
521 522 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
522 523 }
523 524 return (0);
524 525 }
525 526
526 527 static void
527 528 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
528 529 {
529 530 uint64_t claim_txg = zh->zh_claim_txg;
530 531 zil_scan_arg_t zsa = { dp, zh };
531 532 zilog_t *zilog;
532 533
533 534 /*
534 535 * We only want to visit blocks that have been claimed but not yet
535 536 * replayed (or, in read-only mode, blocks that *would* be claimed).
536 537 */
537 538 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
538 539 return;
539 540
540 541 zilog = zil_alloc(dp->dp_meta_objset, zh);
541 542
542 543 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
543 544 claim_txg);
544 545
545 546 zil_free(zilog);
546 547 }
547 548
548 549 /* ARGSUSED */
549 550 static void
550 551 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
551 552 uint64_t objset, uint64_t object, uint64_t blkid)
552 553 {
553 554 zbookmark_phys_t czb;
554 555 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
555 556
556 557 if (zfs_no_scrub_prefetch)
557 558 return;
558 559
559 560 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
560 561 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
561 562 return;
562 563
563 564 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
564 565
565 566 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
566 567 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
567 568 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
568 569 }
569 570
570 571 static boolean_t
571 572 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
572 573 const zbookmark_phys_t *zb)
573 574 {
574 575 /*
575 576 * We never skip over user/group accounting objects (obj<0)
576 577 */
577 578 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
578 579 (int64_t)zb->zb_object >= 0) {
579 580 /*
580 581 * If we already visited this bp & everything below (in
581 582 * a prior txg sync), don't bother doing it again.
582 583 */
583 584 if (zbookmark_subtree_completed(dnp, zb,
584 585 &scn->scn_phys.scn_bookmark))
585 586 return (B_TRUE);
586 587
587 588 /*
588 589 * If we found the block we're trying to resume from, or
589 590 * we went past it to a different object, zero it out to
590 591 * indicate that it's OK to start checking for pausing
591 592 * again.
592 593 */
593 594 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
594 595 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
595 596 dprintf("resuming at %llx/%llx/%llx/%llx\n",
596 597 (longlong_t)zb->zb_objset,
597 598 (longlong_t)zb->zb_object,
598 599 (longlong_t)zb->zb_level,
599 600 (longlong_t)zb->zb_blkid);
600 601 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
601 602 }
602 603 }
603 604 return (B_FALSE);
604 605 }
605 606
606 607 /*
607 608 * Return nonzero on i/o error.
608 609 * Return new buf to write out in *bufp.
609 610 */
610 611 static int
611 612 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
612 613 dnode_phys_t *dnp, const blkptr_t *bp,
613 614 const zbookmark_phys_t *zb, dmu_tx_t *tx)
614 615 {
615 616 dsl_pool_t *dp = scn->scn_dp;
616 617 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
617 618 int err;
618 619
619 620 if (BP_GET_LEVEL(bp) > 0) {
620 621 arc_flags_t flags = ARC_FLAG_WAIT;
621 622 int i;
622 623 blkptr_t *cbp;
623 624 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
624 625 arc_buf_t *buf;
625 626
626 627 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
627 628 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
628 629 if (err) {
629 630 scn->scn_phys.scn_errors++;
630 631 return (err);
631 632 }
632 633 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
633 634 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
634 635 zb->zb_object, zb->zb_blkid * epb + i);
635 636 }
636 637 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
637 638 zbookmark_phys_t czb;
638 639
639 640 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
640 641 zb->zb_level - 1,
641 642 zb->zb_blkid * epb + i);
642 643 dsl_scan_visitbp(cbp, &czb, dnp,
643 644 ds, scn, ostype, tx);
644 645 }
645 646 (void) arc_buf_remove_ref(buf, &buf);
646 647 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
647 648 arc_flags_t flags = ARC_FLAG_WAIT;
648 649 dnode_phys_t *cdnp;
649 650 int i, j;
650 651 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
651 652 arc_buf_t *buf;
652 653
653 654 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
654 655 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
655 656 if (err) {
656 657 scn->scn_phys.scn_errors++;
657 658 return (err);
658 659 }
659 660 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
660 661 for (j = 0; j < cdnp->dn_nblkptr; j++) {
661 662 blkptr_t *cbp = &cdnp->dn_blkptr[j];
662 663 dsl_scan_prefetch(scn, buf, cbp,
663 664 zb->zb_objset, zb->zb_blkid * epb + i, j);
664 665 }
665 666 }
666 667 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
667 668 dsl_scan_visitdnode(scn, ds, ostype,
668 669 cdnp, zb->zb_blkid * epb + i, tx);
669 670 }
670 671
671 672 (void) arc_buf_remove_ref(buf, &buf);
672 673 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
673 674 arc_flags_t flags = ARC_FLAG_WAIT;
674 675 objset_phys_t *osp;
675 676 arc_buf_t *buf;
676 677
677 678 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
678 679 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
679 680 if (err) {
680 681 scn->scn_phys.scn_errors++;
681 682 return (err);
682 683 }
683 684
684 685 osp = buf->b_data;
685 686
686 687 dsl_scan_visitdnode(scn, ds, osp->os_type,
687 688 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
688 689
689 690 if (OBJSET_BUF_HAS_USERUSED(buf)) {
690 691 /*
691 692 * We also always visit user/group accounting
692 693 * objects, and never skip them, even if we are
693 694 * pausing. This is necessary so that the space
694 695 * deltas from this txg get integrated.
695 696 */
696 697 dsl_scan_visitdnode(scn, ds, osp->os_type,
697 698 &osp->os_groupused_dnode,
698 699 DMU_GROUPUSED_OBJECT, tx);
699 700 dsl_scan_visitdnode(scn, ds, osp->os_type,
700 701 &osp->os_userused_dnode,
701 702 DMU_USERUSED_OBJECT, tx);
702 703 }
703 704 (void) arc_buf_remove_ref(buf, &buf);
704 705 }
705 706
706 707 return (0);
707 708 }
708 709
709 710 static void
710 711 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
711 712 dmu_objset_type_t ostype, dnode_phys_t *dnp,
712 713 uint64_t object, dmu_tx_t *tx)
713 714 {
714 715 int j;
715 716
716 717 for (j = 0; j < dnp->dn_nblkptr; j++) {
717 718 zbookmark_phys_t czb;
718 719
719 720 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
720 721 dnp->dn_nlevels - 1, j);
721 722 dsl_scan_visitbp(&dnp->dn_blkptr[j],
722 723 &czb, dnp, ds, scn, ostype, tx);
723 724 }
724 725
725 726 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
726 727 zbookmark_phys_t czb;
727 728 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
728 729 0, DMU_SPILL_BLKID);
729 730 dsl_scan_visitbp(&dnp->dn_spill,
730 731 &czb, dnp, ds, scn, ostype, tx);
731 732 }
732 733 }
733 734
734 735 /*
735 736 * The arguments are in this order because mdb can only print the
736 737 * first 5; we want them to be useful.
737 738 */
738 739 static void
739 740 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
740 741 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
741 742 dmu_objset_type_t ostype, dmu_tx_t *tx)
742 743 {
743 744 dsl_pool_t *dp = scn->scn_dp;
744 745 arc_buf_t *buf = NULL;
745 746 blkptr_t bp_toread = *bp;
746 747
747 748 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
748 749
749 750 if (dsl_scan_check_pause(scn, zb))
750 751 return;
751 752
752 753 if (dsl_scan_check_resume(scn, dnp, zb))
753 754 return;
754 755
755 756 if (BP_IS_HOLE(bp))
756 757 return;
757 758
758 759 scn->scn_visited_this_txg++;
759 760
760 761 dprintf_bp(bp,
761 762 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
762 763 ds, ds ? ds->ds_object : 0,
763 764 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
764 765 bp);
765 766
766 767 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
767 768 return;
768 769
769 770 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
770 771 return;
771 772
772 773 /*
773 774 * If dsl_scan_ddt() has aready visited this block, it will have
774 775 * already done any translations or scrubbing, so don't call the
775 776 * callback again.
776 777 */
777 778 if (ddt_class_contains(dp->dp_spa,
778 779 scn->scn_phys.scn_ddt_class_max, bp)) {
779 780 ASSERT(buf == NULL);
780 781 return;
781 782 }
782 783
783 784 /*
784 785 * If this block is from the future (after cur_max_txg), then we
785 786 * are doing this on behalf of a deleted snapshot, and we will
786 787 * revisit the future block on the next pass of this dataset.
787 788 * Don't scan it now unless we need to because something
788 789 * under it was modified.
789 790 */
790 791 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
791 792 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
792 793 }
793 794 }
794 795
795 796 static void
796 797 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
797 798 dmu_tx_t *tx)
798 799 {
799 800 zbookmark_phys_t zb;
800 801
801 802 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
802 803 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
803 804 dsl_scan_visitbp(bp, &zb, NULL,
804 805 ds, scn, DMU_OST_NONE, tx);
805 806
806 807 dprintf_ds(ds, "finished scan%s", "");
807 808 }
808 809
809 810 void
810 811 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
811 812 {
812 813 dsl_pool_t *dp = ds->ds_dir->dd_pool;
813 814 dsl_scan_t *scn = dp->dp_scan;
814 815 uint64_t mintxg;
815 816
816 817 if (scn->scn_phys.scn_state != DSS_SCANNING)
817 818 return;
818 819
819 820 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
820 821 if (ds->ds_is_snapshot) {
821 822 /* Note, scn_cur_{min,max}_txg stays the same. */
822 823 scn->scn_phys.scn_bookmark.zb_objset =
823 824 dsl_dataset_phys(ds)->ds_next_snap_obj;
824 825 zfs_dbgmsg("destroying ds %llu; currently traversing; "
825 826 "reset zb_objset to %llu",
826 827 (u_longlong_t)ds->ds_object,
827 828 (u_longlong_t)dsl_dataset_phys(ds)->
828 829 ds_next_snap_obj);
829 830 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
830 831 } else {
831 832 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
832 833 ZB_DESTROYED_OBJSET, 0, 0, 0);
833 834 zfs_dbgmsg("destroying ds %llu; currently traversing; "
834 835 "reset bookmark to -1,0,0,0",
835 836 (u_longlong_t)ds->ds_object);
836 837 }
837 838 } else if (zap_lookup_int_key(dp->dp_meta_objset,
838 839 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
839 840 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
840 841 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
841 842 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
842 843 if (ds->ds_is_snapshot) {
843 844 /*
844 845 * We keep the same mintxg; it could be >
845 846 * ds_creation_txg if the previous snapshot was
846 847 * deleted too.
847 848 */
848 849 VERIFY(zap_add_int_key(dp->dp_meta_objset,
849 850 scn->scn_phys.scn_queue_obj,
850 851 dsl_dataset_phys(ds)->ds_next_snap_obj,
851 852 mintxg, tx) == 0);
852 853 zfs_dbgmsg("destroying ds %llu; in queue; "
853 854 "replacing with %llu",
854 855 (u_longlong_t)ds->ds_object,
855 856 (u_longlong_t)dsl_dataset_phys(ds)->
856 857 ds_next_snap_obj);
857 858 } else {
858 859 zfs_dbgmsg("destroying ds %llu; in queue; removing",
859 860 (u_longlong_t)ds->ds_object);
860 861 }
861 862 } else {
862 863 zfs_dbgmsg("destroying ds %llu; ignoring",
863 864 (u_longlong_t)ds->ds_object);
864 865 }
865 866
866 867 /*
867 868 * dsl_scan_sync() should be called after this, and should sync
868 869 * out our changed state, but just to be safe, do it here.
869 870 */
870 871 dsl_scan_sync_state(scn, tx);
871 872 }
872 873
873 874 void
874 875 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
875 876 {
876 877 dsl_pool_t *dp = ds->ds_dir->dd_pool;
877 878 dsl_scan_t *scn = dp->dp_scan;
878 879 uint64_t mintxg;
879 880
880 881 if (scn->scn_phys.scn_state != DSS_SCANNING)
881 882 return;
882 883
883 884 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
884 885
885 886 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
886 887 scn->scn_phys.scn_bookmark.zb_objset =
887 888 dsl_dataset_phys(ds)->ds_prev_snap_obj;
888 889 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
889 890 "reset zb_objset to %llu",
890 891 (u_longlong_t)ds->ds_object,
891 892 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
892 893 } else if (zap_lookup_int_key(dp->dp_meta_objset,
893 894 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
894 895 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
895 896 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
896 897 VERIFY(zap_add_int_key(dp->dp_meta_objset,
897 898 scn->scn_phys.scn_queue_obj,
898 899 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
899 900 zfs_dbgmsg("snapshotting ds %llu; in queue; "
900 901 "replacing with %llu",
901 902 (u_longlong_t)ds->ds_object,
902 903 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
903 904 }
904 905 dsl_scan_sync_state(scn, tx);
905 906 }
906 907
907 908 void
908 909 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
909 910 {
910 911 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
911 912 dsl_scan_t *scn = dp->dp_scan;
912 913 uint64_t mintxg;
913 914
914 915 if (scn->scn_phys.scn_state != DSS_SCANNING)
915 916 return;
916 917
917 918 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
918 919 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
919 920 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
920 921 "reset zb_objset to %llu",
921 922 (u_longlong_t)ds1->ds_object,
922 923 (u_longlong_t)ds2->ds_object);
923 924 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
924 925 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
925 926 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
926 927 "reset zb_objset to %llu",
927 928 (u_longlong_t)ds2->ds_object,
928 929 (u_longlong_t)ds1->ds_object);
929 930 }
930 931
931 932 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
932 933 ds1->ds_object, &mintxg) == 0) {
933 934 int err;
934 935
935 936 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
936 937 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
937 938 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
938 939 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
939 940 err = zap_add_int_key(dp->dp_meta_objset,
940 941 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
941 942 VERIFY(err == 0 || err == EEXIST);
942 943 if (err == EEXIST) {
943 944 /* Both were there to begin with */
944 945 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
945 946 scn->scn_phys.scn_queue_obj,
946 947 ds1->ds_object, mintxg, tx));
947 948 }
948 949 zfs_dbgmsg("clone_swap ds %llu; in queue; "
949 950 "replacing with %llu",
950 951 (u_longlong_t)ds1->ds_object,
951 952 (u_longlong_t)ds2->ds_object);
952 953 } else if (zap_lookup_int_key(dp->dp_meta_objset,
953 954 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
954 955 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
955 956 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
956 957 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
957 958 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
958 959 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
959 960 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
960 961 zfs_dbgmsg("clone_swap ds %llu; in queue; "
961 962 "replacing with %llu",
962 963 (u_longlong_t)ds2->ds_object,
963 964 (u_longlong_t)ds1->ds_object);
964 965 }
965 966
966 967 dsl_scan_sync_state(scn, tx);
967 968 }
968 969
969 970 struct enqueue_clones_arg {
970 971 dmu_tx_t *tx;
971 972 uint64_t originobj;
972 973 };
973 974
974 975 /* ARGSUSED */
975 976 static int
976 977 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
977 978 {
978 979 struct enqueue_clones_arg *eca = arg;
979 980 dsl_dataset_t *ds;
980 981 int err;
981 982 dsl_scan_t *scn = dp->dp_scan;
982 983
983 984 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
984 985 return (0);
985 986
986 987 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
987 988 if (err)
988 989 return (err);
989 990
990 991 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
991 992 dsl_dataset_t *prev;
992 993 err = dsl_dataset_hold_obj(dp,
993 994 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
994 995
995 996 dsl_dataset_rele(ds, FTAG);
996 997 if (err)
997 998 return (err);
998 999 ds = prev;
999 1000 }
1000 1001 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1001 1002 scn->scn_phys.scn_queue_obj, ds->ds_object,
1002 1003 dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1003 1004 dsl_dataset_rele(ds, FTAG);
1004 1005 return (0);
1005 1006 }
1006 1007
1007 1008 static void
1008 1009 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1009 1010 {
1010 1011 dsl_pool_t *dp = scn->scn_dp;
1011 1012 dsl_dataset_t *ds;
1012 1013 objset_t *os;
1013 1014
1014 1015 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1015 1016
1016 1017 if (dmu_objset_from_ds(ds, &os))
1017 1018 goto out;
1018 1019
1019 1020 /*
1020 1021 * Only the ZIL in the head (non-snapshot) is valid. Even though
1021 1022 * snapshots can have ZIL block pointers (which may be the same
1022 1023 * BP as in the head), they must be ignored. So we traverse the
1023 1024 * ZIL here, rather than in scan_recurse(), because the regular
1024 1025 * snapshot block-sharing rules don't apply to it.
1025 1026 */
1026 1027 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1027 1028 dsl_scan_zil(dp, &os->os_zil_header);
1028 1029
1029 1030 /*
1030 1031 * Iterate over the bps in this ds.
1031 1032 */
1032 1033 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1033 1034 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1034 1035
1035 1036 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1036 1037 dsl_dataset_name(ds, dsname);
1037 1038 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1038 1039 "pausing=%u",
1039 1040 (longlong_t)dsobj, dsname,
1040 1041 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1041 1042 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1042 1043 (int)scn->scn_pausing);
1043 1044 kmem_free(dsname, ZFS_MAXNAMELEN);
1044 1045
1045 1046 if (scn->scn_pausing)
1046 1047 goto out;
1047 1048
1048 1049 /*
1049 1050 * We've finished this pass over this dataset.
1050 1051 */
1051 1052
1052 1053 /*
1053 1054 * If we did not completely visit this dataset, do another pass.
1054 1055 */
1055 1056 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1056 1057 zfs_dbgmsg("incomplete pass; visiting again");
1057 1058 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1058 1059 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1059 1060 scn->scn_phys.scn_queue_obj, ds->ds_object,
1060 1061 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1061 1062 goto out;
1062 1063 }
1063 1064
1064 1065 /*
1065 1066 * Add descendent datasets to work queue.
1066 1067 */
1067 1068 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1068 1069 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1069 1070 scn->scn_phys.scn_queue_obj,
1070 1071 dsl_dataset_phys(ds)->ds_next_snap_obj,
1071 1072 dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1072 1073 }
1073 1074 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1074 1075 boolean_t usenext = B_FALSE;
1075 1076 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1076 1077 uint64_t count;
1077 1078 /*
1078 1079 * A bug in a previous version of the code could
1079 1080 * cause upgrade_clones_cb() to not set
1080 1081 * ds_next_snap_obj when it should, leading to a
1081 1082 * missing entry. Therefore we can only use the
1082 1083 * next_clones_obj when its count is correct.
1083 1084 */
1084 1085 int err = zap_count(dp->dp_meta_objset,
1085 1086 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1086 1087 if (err == 0 &&
1087 1088 count == dsl_dataset_phys(ds)->ds_num_children - 1)
1088 1089 usenext = B_TRUE;
1089 1090 }
1090 1091
1091 1092 if (usenext) {
1092 1093 VERIFY0(zap_join_key(dp->dp_meta_objset,
1093 1094 dsl_dataset_phys(ds)->ds_next_clones_obj,
1094 1095 scn->scn_phys.scn_queue_obj,
1095 1096 dsl_dataset_phys(ds)->ds_creation_txg, tx));
1096 1097 } else {
1097 1098 struct enqueue_clones_arg eca;
1098 1099 eca.tx = tx;
1099 1100 eca.originobj = ds->ds_object;
1100 1101
1101 1102 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1102 1103 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1103 1104 }
1104 1105 }
1105 1106
1106 1107 out:
1107 1108 dsl_dataset_rele(ds, FTAG);
1108 1109 }
1109 1110
1110 1111 /* ARGSUSED */
1111 1112 static int
1112 1113 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1113 1114 {
1114 1115 dmu_tx_t *tx = arg;
1115 1116 dsl_dataset_t *ds;
1116 1117 int err;
1117 1118 dsl_scan_t *scn = dp->dp_scan;
1118 1119
1119 1120 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1120 1121 if (err)
1121 1122 return (err);
1122 1123
1123 1124 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1124 1125 dsl_dataset_t *prev;
1125 1126 err = dsl_dataset_hold_obj(dp,
1126 1127 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1127 1128 if (err) {
1128 1129 dsl_dataset_rele(ds, FTAG);
1129 1130 return (err);
1130 1131 }
1131 1132
1132 1133 /*
1133 1134 * If this is a clone, we don't need to worry about it for now.
1134 1135 */
1135 1136 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1136 1137 dsl_dataset_rele(ds, FTAG);
1137 1138 dsl_dataset_rele(prev, FTAG);
1138 1139 return (0);
1139 1140 }
1140 1141 dsl_dataset_rele(ds, FTAG);
1141 1142 ds = prev;
1142 1143 }
1143 1144
1144 1145 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1145 1146 ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1146 1147 dsl_dataset_rele(ds, FTAG);
1147 1148 return (0);
1148 1149 }
1149 1150
1150 1151 /*
1151 1152 * Scrub/dedup interaction.
1152 1153 *
1153 1154 * If there are N references to a deduped block, we don't want to scrub it
1154 1155 * N times -- ideally, we should scrub it exactly once.
1155 1156 *
1156 1157 * We leverage the fact that the dde's replication class (enum ddt_class)
1157 1158 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1158 1159 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1159 1160 *
1160 1161 * To prevent excess scrubbing, the scrub begins by walking the DDT
1161 1162 * to find all blocks with refcnt > 1, and scrubs each of these once.
1162 1163 * Since there are two replication classes which contain blocks with
1163 1164 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1164 1165 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1165 1166 *
1166 1167 * There would be nothing more to say if a block's refcnt couldn't change
1167 1168 * during a scrub, but of course it can so we must account for changes
1168 1169 * in a block's replication class.
1169 1170 *
1170 1171 * Here's an example of what can occur:
1171 1172 *
1172 1173 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1173 1174 * when visited during the top-down scrub phase, it will be scrubbed twice.
1174 1175 * This negates our scrub optimization, but is otherwise harmless.
1175 1176 *
1176 1177 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1177 1178 * on each visit during the top-down scrub phase, it will never be scrubbed.
1178 1179 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1179 1180 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1180 1181 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1181 1182 * while a scrub is in progress, it scrubs the block right then.
1182 1183 */
1183 1184 static void
1184 1185 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1185 1186 {
1186 1187 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1187 1188 ddt_entry_t dde = { 0 };
1188 1189 int error;
1189 1190 uint64_t n = 0;
1190 1191
1191 1192 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1192 1193 ddt_t *ddt;
1193 1194
1194 1195 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1195 1196 break;
1196 1197 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1197 1198 (longlong_t)ddb->ddb_class,
1198 1199 (longlong_t)ddb->ddb_type,
1199 1200 (longlong_t)ddb->ddb_checksum,
1200 1201 (longlong_t)ddb->ddb_cursor);
1201 1202
1202 1203 /* There should be no pending changes to the dedup table */
1203 1204 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1204 1205 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1205 1206
1206 1207 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1207 1208 n++;
1208 1209
1209 1210 if (dsl_scan_check_pause(scn, NULL))
1210 1211 break;
1211 1212 }
1212 1213
1213 1214 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1214 1215 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1215 1216 (int)scn->scn_pausing);
1216 1217
1217 1218 ASSERT(error == 0 || error == ENOENT);
1218 1219 ASSERT(error != ENOENT ||
1219 1220 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1220 1221 }
1221 1222
1222 1223 /* ARGSUSED */
1223 1224 void
1224 1225 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1225 1226 ddt_entry_t *dde, dmu_tx_t *tx)
1226 1227 {
1227 1228 const ddt_key_t *ddk = &dde->dde_key;
1228 1229 ddt_phys_t *ddp = dde->dde_phys;
1229 1230 blkptr_t bp;
1230 1231 zbookmark_phys_t zb = { 0 };
1231 1232
1232 1233 if (scn->scn_phys.scn_state != DSS_SCANNING)
1233 1234 return;
1234 1235
1235 1236 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1236 1237 if (ddp->ddp_phys_birth == 0 ||
1237 1238 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1238 1239 continue;
1239 1240 ddt_bp_create(checksum, ddk, ddp, &bp);
1240 1241
1241 1242 scn->scn_visited_this_txg++;
1242 1243 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1243 1244 }
1244 1245 }
1245 1246
1246 1247 static void
1247 1248 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1248 1249 {
1249 1250 dsl_pool_t *dp = scn->scn_dp;
1250 1251 zap_cursor_t zc;
1251 1252 zap_attribute_t za;
1252 1253
1253 1254 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1254 1255 scn->scn_phys.scn_ddt_class_max) {
1255 1256 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1256 1257 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1257 1258 dsl_scan_ddt(scn, tx);
1258 1259 if (scn->scn_pausing)
1259 1260 return;
1260 1261 }
1261 1262
1262 1263 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1263 1264 /* First do the MOS & ORIGIN */
1264 1265
1265 1266 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1266 1267 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1267 1268 dsl_scan_visit_rootbp(scn, NULL,
1268 1269 &dp->dp_meta_rootbp, tx);
1269 1270 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1270 1271 if (scn->scn_pausing)
1271 1272 return;
1272 1273
1273 1274 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1274 1275 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1275 1276 enqueue_cb, tx, DS_FIND_CHILDREN));
1276 1277 } else {
1277 1278 dsl_scan_visitds(scn,
1278 1279 dp->dp_origin_snap->ds_object, tx);
1279 1280 }
1280 1281 ASSERT(!scn->scn_pausing);
1281 1282 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1282 1283 ZB_DESTROYED_OBJSET) {
1283 1284 /*
1284 1285 * If we were paused, continue from here. Note if the
1285 1286 * ds we were paused on was deleted, the zb_objset may
1286 1287 * be -1, so we will skip this and find a new objset
1287 1288 * below.
1288 1289 */
1289 1290 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1290 1291 if (scn->scn_pausing)
1291 1292 return;
1292 1293 }
1293 1294
1294 1295 /*
1295 1296 * In case we were paused right at the end of the ds, zero the
1296 1297 * bookmark so we don't think that we're still trying to resume.
1297 1298 */
1298 1299 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1299 1300
1300 1301 /* keep pulling things out of the zap-object-as-queue */
1301 1302 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1302 1303 scn->scn_phys.scn_queue_obj),
1303 1304 zap_cursor_retrieve(&zc, &za) == 0) {
1304 1305 dsl_dataset_t *ds;
1305 1306 uint64_t dsobj;
1306 1307
1307 1308 dsobj = strtonum(za.za_name, NULL);
1308 1309 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1309 1310 scn->scn_phys.scn_queue_obj, dsobj, tx));
1310 1311
1311 1312 /* Set up min/max txg */
1312 1313 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1313 1314 if (za.za_first_integer != 0) {
1314 1315 scn->scn_phys.scn_cur_min_txg =
1315 1316 MAX(scn->scn_phys.scn_min_txg,
1316 1317 za.za_first_integer);
1317 1318 } else {
1318 1319 scn->scn_phys.scn_cur_min_txg =
1319 1320 MAX(scn->scn_phys.scn_min_txg,
1320 1321 dsl_dataset_phys(ds)->ds_prev_snap_txg);
1321 1322 }
1322 1323 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1323 1324 dsl_dataset_rele(ds, FTAG);
1324 1325
1325 1326 dsl_scan_visitds(scn, dsobj, tx);
1326 1327 zap_cursor_fini(&zc);
1327 1328 if (scn->scn_pausing)
1328 1329 return;
1329 1330 }
1330 1331 zap_cursor_fini(&zc);
1331 1332 }
1332 1333
1333 1334 static boolean_t
1334 1335 dsl_scan_free_should_pause(dsl_scan_t *scn)
1335 1336 {
1336 1337 uint64_t elapsed_nanosecs;
1337 1338
1338 1339 if (zfs_recover)
1339 1340 return (B_FALSE);
1340 1341
1341 1342 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1342 1343 return (B_TRUE);
1343 1344
1344 1345 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1345 1346 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1346 1347 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1347 1348 txg_sync_waiting(scn->scn_dp)) ||
1348 1349 spa_shutting_down(scn->scn_dp->dp_spa));
1349 1350 }
1350 1351
1351 1352 static int
1352 1353 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1353 1354 {
1354 1355 dsl_scan_t *scn = arg;
1355 1356
1356 1357 if (!scn->scn_is_bptree ||
1357 1358 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1358 1359 if (dsl_scan_free_should_pause(scn))
1359 1360 return (SET_ERROR(ERESTART));
1360 1361 }
1361 1362
1362 1363 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1363 1364 dmu_tx_get_txg(tx), bp, 0));
1364 1365 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1365 1366 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1366 1367 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1367 1368 scn->scn_visited_this_txg++;
1368 1369 return (0);
1369 1370 }
1370 1371
1371 1372 boolean_t
1372 1373 dsl_scan_active(dsl_scan_t *scn)
1373 1374 {
1374 1375 spa_t *spa = scn->scn_dp->dp_spa;
1375 1376 uint64_t used = 0, comp, uncomp;
1376 1377
1377 1378 if (spa->spa_load_state != SPA_LOAD_NONE)
1378 1379 return (B_FALSE);
1379 1380 if (spa_shutting_down(spa))
1380 1381 return (B_FALSE);
1381 1382 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1382 1383 (scn->scn_async_destroying && !scn->scn_async_stalled))
1383 1384 return (B_TRUE);
1384 1385
1385 1386 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1386 1387 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1387 1388 &used, &comp, &uncomp);
1388 1389 }
1389 1390 return (used != 0);
1390 1391 }
1391 1392
1392 1393 void
1393 1394 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1394 1395 {
1395 1396 dsl_scan_t *scn = dp->dp_scan;
1396 1397 spa_t *spa = dp->dp_spa;
1397 1398 int err = 0;
1398 1399
1399 1400 /*
1400 1401 * Check for scn_restart_txg before checking spa_load_state, so
1401 1402 * that we can restart an old-style scan while the pool is being
1402 1403 * imported (see dsl_scan_init).
1403 1404 */
1404 1405 if (scn->scn_restart_txg != 0 &&
1405 1406 scn->scn_restart_txg <= tx->tx_txg) {
1406 1407 pool_scan_func_t func = POOL_SCAN_SCRUB;
1407 1408 dsl_scan_done(scn, B_FALSE, tx);
1408 1409 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1409 1410 func = POOL_SCAN_RESILVER;
1410 1411 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1411 1412 func, tx->tx_txg);
1412 1413 dsl_scan_setup_sync(&func, tx);
1413 1414 }
1414 1415
1415 1416 /*
1416 1417 * If the scan is inactive due to a stalled async destroy, try again.
1417 1418 */
1418 1419 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1419 1420 spa_sync_pass(dp->dp_spa) > 1)
1420 1421 return;
1421 1422
1422 1423 scn->scn_visited_this_txg = 0;
1423 1424 scn->scn_pausing = B_FALSE;
1424 1425 scn->scn_sync_start_time = gethrtime();
1425 1426 spa->spa_scrub_active = B_TRUE;
1426 1427
1427 1428 /*
1428 1429 * First process the async destroys. If we pause, don't do
1429 1430 * any scrubbing or resilvering. This ensures that there are no
1430 1431 * async destroys while we are scanning, so the scan code doesn't
1431 1432 * have to worry about traversing it. It is also faster to free the
1432 1433 * blocks than to scrub them.
1433 1434 */
1434 1435 if (zfs_free_bpobj_enabled &&
1435 1436 spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1436 1437 scn->scn_is_bptree = B_FALSE;
1437 1438 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1438 1439 NULL, ZIO_FLAG_MUSTSUCCEED);
1439 1440 err = bpobj_iterate(&dp->dp_free_bpobj,
1440 1441 dsl_scan_free_block_cb, scn, tx);
1441 1442 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1442 1443
1443 1444 if (err != 0 && err != ERESTART)
1444 1445 zfs_panic_recover("error %u from bpobj_iterate()", err);
1445 1446 }
1446 1447
1447 1448 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1448 1449 ASSERT(scn->scn_async_destroying);
1449 1450 scn->scn_is_bptree = B_TRUE;
1450 1451 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1451 1452 NULL, ZIO_FLAG_MUSTSUCCEED);
1452 1453 err = bptree_iterate(dp->dp_meta_objset,
1453 1454 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1454 1455 VERIFY0(zio_wait(scn->scn_zio_root));
1455 1456
1456 1457 if (err == EIO || err == ECKSUM) {
1457 1458 err = 0;
1458 1459 } else if (err != 0 && err != ERESTART) {
1459 1460 zfs_panic_recover("error %u from "
1460 1461 "traverse_dataset_destroyed()", err);
1461 1462 }
1462 1463
1463 1464 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1464 1465 /* finished; deactivate async destroy feature */
1465 1466 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1466 1467 ASSERT(!spa_feature_is_active(spa,
1467 1468 SPA_FEATURE_ASYNC_DESTROY));
1468 1469 VERIFY0(zap_remove(dp->dp_meta_objset,
1469 1470 DMU_POOL_DIRECTORY_OBJECT,
1470 1471 DMU_POOL_BPTREE_OBJ, tx));
1471 1472 VERIFY0(bptree_free(dp->dp_meta_objset,
1472 1473 dp->dp_bptree_obj, tx));
1473 1474 dp->dp_bptree_obj = 0;
1474 1475 scn->scn_async_destroying = B_FALSE;
1475 1476 scn->scn_async_stalled = B_FALSE;
1476 1477 } else {
1477 1478 /*
1478 1479 * If we didn't make progress, mark the async
1479 1480 * destroy as stalled, so that we will not initiate
1480 1481 * a spa_sync() on its behalf. Note that we only
1481 1482 * check this if we are not finished, because if the
1482 1483 * bptree had no blocks for us to visit, we can
1483 1484 * finish without "making progress".
1484 1485 */
1485 1486 scn->scn_async_stalled =
1486 1487 (scn->scn_visited_this_txg == 0);
1487 1488 }
1488 1489 }
1489 1490 if (scn->scn_visited_this_txg) {
1490 1491 zfs_dbgmsg("freed %llu blocks in %llums from "
1491 1492 "free_bpobj/bptree txg %llu; err=%u",
1492 1493 (longlong_t)scn->scn_visited_this_txg,
1493 1494 (longlong_t)
1494 1495 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1495 1496 (longlong_t)tx->tx_txg, err);
1496 1497 scn->scn_visited_this_txg = 0;
|
↓ open down ↓ |
1463 lines elided |
↑ open up ↑ |
1497 1498
1498 1499 /*
1499 1500 * Write out changes to the DDT that may be required as a
1500 1501 * result of the blocks freed. This ensures that the DDT
1501 1502 * is clean when a scrub/resilver runs.
1502 1503 */
1503 1504 ddt_sync(spa, tx->tx_txg);
1504 1505 }
1505 1506 if (err != 0)
1506 1507 return;
1507 - if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1508 + if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1509 + zfs_free_leak_on_eio &&
1508 1510 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1509 1511 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1510 1512 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1511 1513 /*
1512 1514 * We have finished background destroying, but there is still
1513 1515 * some space left in the dp_free_dir. Transfer this leaked
1514 1516 * space to the dp_leak_dir.
1515 1517 */
1516 1518 if (dp->dp_leak_dir == NULL) {
1517 1519 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1518 1520 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1519 1521 LEAK_DIR_NAME, tx);
1520 1522 VERIFY0(dsl_pool_open_special_dir(dp,
1521 1523 LEAK_DIR_NAME, &dp->dp_leak_dir));
1522 1524 rrw_exit(&dp->dp_config_rwlock, FTAG);
|
↓ open down ↓ |
5 lines elided |
↑ open up ↑ |
1523 1525 }
1524 1526 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1525 1527 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1526 1528 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1527 1529 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1528 1530 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1529 1531 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1530 1532 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1531 1533 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1532 1534 }
1533 - if (!scn->scn_async_destroying) {
1535 + if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1534 1536 /* finished; verify that space accounting went to zero */
1535 1537 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1536 1538 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1537 1539 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1538 1540 }
1539 1541
1540 1542 if (scn->scn_phys.scn_state != DSS_SCANNING)
1541 1543 return;
1542 1544
1543 1545 if (scn->scn_done_txg == tx->tx_txg) {
1544 1546 ASSERT(!scn->scn_pausing);
1545 1547 /* finished with scan. */
1546 1548 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1547 1549 dsl_scan_done(scn, B_TRUE, tx);
1548 1550 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1549 1551 dsl_scan_sync_state(scn, tx);
1550 1552 return;
1551 1553 }
1552 1554
1553 1555 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1554 1556 scn->scn_phys.scn_ddt_class_max) {
1555 1557 zfs_dbgmsg("doing scan sync txg %llu; "
1556 1558 "ddt bm=%llu/%llu/%llu/%llx",
1557 1559 (longlong_t)tx->tx_txg,
1558 1560 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1559 1561 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1560 1562 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1561 1563 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1562 1564 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1563 1565 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1564 1566 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1565 1567 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1566 1568 } else {
1567 1569 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1568 1570 (longlong_t)tx->tx_txg,
1569 1571 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1570 1572 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1571 1573 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1572 1574 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1573 1575 }
1574 1576
1575 1577 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1576 1578 NULL, ZIO_FLAG_CANFAIL);
1577 1579 dsl_pool_config_enter(dp, FTAG);
1578 1580 dsl_scan_visit(scn, tx);
1579 1581 dsl_pool_config_exit(dp, FTAG);
1580 1582 (void) zio_wait(scn->scn_zio_root);
1581 1583 scn->scn_zio_root = NULL;
1582 1584
1583 1585 zfs_dbgmsg("visited %llu blocks in %llums",
1584 1586 (longlong_t)scn->scn_visited_this_txg,
1585 1587 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1586 1588
1587 1589 if (!scn->scn_pausing) {
1588 1590 scn->scn_done_txg = tx->tx_txg + 1;
1589 1591 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1590 1592 tx->tx_txg, scn->scn_done_txg);
1591 1593 }
1592 1594
1593 1595 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1594 1596 mutex_enter(&spa->spa_scrub_lock);
1595 1597 while (spa->spa_scrub_inflight > 0) {
1596 1598 cv_wait(&spa->spa_scrub_io_cv,
1597 1599 &spa->spa_scrub_lock);
1598 1600 }
1599 1601 mutex_exit(&spa->spa_scrub_lock);
1600 1602 }
1601 1603
1602 1604 dsl_scan_sync_state(scn, tx);
1603 1605 }
1604 1606
1605 1607 /*
1606 1608 * This will start a new scan, or restart an existing one.
1607 1609 */
1608 1610 void
1609 1611 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1610 1612 {
1611 1613 if (txg == 0) {
1612 1614 dmu_tx_t *tx;
1613 1615 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1614 1616 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1615 1617
1616 1618 txg = dmu_tx_get_txg(tx);
1617 1619 dp->dp_scan->scn_restart_txg = txg;
1618 1620 dmu_tx_commit(tx);
1619 1621 } else {
1620 1622 dp->dp_scan->scn_restart_txg = txg;
1621 1623 }
1622 1624 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1623 1625 }
1624 1626
1625 1627 boolean_t
1626 1628 dsl_scan_resilvering(dsl_pool_t *dp)
1627 1629 {
1628 1630 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1629 1631 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1630 1632 }
1631 1633
1632 1634 /*
1633 1635 * scrub consumers
1634 1636 */
1635 1637
1636 1638 static void
1637 1639 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1638 1640 {
1639 1641 int i;
1640 1642
1641 1643 /*
1642 1644 * If we resume after a reboot, zab will be NULL; don't record
1643 1645 * incomplete stats in that case.
1644 1646 */
1645 1647 if (zab == NULL)
1646 1648 return;
1647 1649
1648 1650 for (i = 0; i < 4; i++) {
1649 1651 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1650 1652 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1651 1653 if (t & DMU_OT_NEWTYPE)
1652 1654 t = DMU_OT_OTHER;
1653 1655 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1654 1656 int equal;
1655 1657
1656 1658 zb->zb_count++;
1657 1659 zb->zb_asize += BP_GET_ASIZE(bp);
1658 1660 zb->zb_lsize += BP_GET_LSIZE(bp);
1659 1661 zb->zb_psize += BP_GET_PSIZE(bp);
1660 1662 zb->zb_gangs += BP_COUNT_GANG(bp);
1661 1663
1662 1664 switch (BP_GET_NDVAS(bp)) {
1663 1665 case 2:
1664 1666 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1665 1667 DVA_GET_VDEV(&bp->blk_dva[1]))
1666 1668 zb->zb_ditto_2_of_2_samevdev++;
1667 1669 break;
1668 1670 case 3:
1669 1671 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1670 1672 DVA_GET_VDEV(&bp->blk_dva[1])) +
1671 1673 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1672 1674 DVA_GET_VDEV(&bp->blk_dva[2])) +
1673 1675 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1674 1676 DVA_GET_VDEV(&bp->blk_dva[2]));
1675 1677 if (equal == 1)
1676 1678 zb->zb_ditto_2_of_3_samevdev++;
1677 1679 else if (equal == 3)
1678 1680 zb->zb_ditto_3_of_3_samevdev++;
1679 1681 break;
1680 1682 }
1681 1683 }
1682 1684 }
1683 1685
1684 1686 static void
1685 1687 dsl_scan_scrub_done(zio_t *zio)
1686 1688 {
1687 1689 spa_t *spa = zio->io_spa;
1688 1690
1689 1691 zio_data_buf_free(zio->io_data, zio->io_size);
1690 1692
1691 1693 mutex_enter(&spa->spa_scrub_lock);
1692 1694 spa->spa_scrub_inflight--;
1693 1695 cv_broadcast(&spa->spa_scrub_io_cv);
1694 1696
1695 1697 if (zio->io_error && (zio->io_error != ECKSUM ||
1696 1698 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1697 1699 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1698 1700 }
1699 1701 mutex_exit(&spa->spa_scrub_lock);
1700 1702 }
1701 1703
1702 1704 static int
1703 1705 dsl_scan_scrub_cb(dsl_pool_t *dp,
1704 1706 const blkptr_t *bp, const zbookmark_phys_t *zb)
1705 1707 {
1706 1708 dsl_scan_t *scn = dp->dp_scan;
1707 1709 size_t size = BP_GET_PSIZE(bp);
1708 1710 spa_t *spa = dp->dp_spa;
1709 1711 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1710 1712 boolean_t needs_io;
1711 1713 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1712 1714 int scan_delay = 0;
1713 1715
1714 1716 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1715 1717 phys_birth >= scn->scn_phys.scn_max_txg)
1716 1718 return (0);
1717 1719
1718 1720 count_block(dp->dp_blkstats, bp);
1719 1721
1720 1722 if (BP_IS_EMBEDDED(bp))
1721 1723 return (0);
1722 1724
1723 1725 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1724 1726 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1725 1727 zio_flags |= ZIO_FLAG_SCRUB;
1726 1728 needs_io = B_TRUE;
1727 1729 scan_delay = zfs_scrub_delay;
1728 1730 } else {
1729 1731 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1730 1732 zio_flags |= ZIO_FLAG_RESILVER;
1731 1733 needs_io = B_FALSE;
1732 1734 scan_delay = zfs_resilver_delay;
1733 1735 }
1734 1736
1735 1737 /* If it's an intent log block, failure is expected. */
1736 1738 if (zb->zb_level == ZB_ZIL_LEVEL)
1737 1739 zio_flags |= ZIO_FLAG_SPECULATIVE;
1738 1740
1739 1741 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1740 1742 vdev_t *vd = vdev_lookup_top(spa,
1741 1743 DVA_GET_VDEV(&bp->blk_dva[d]));
1742 1744
1743 1745 /*
1744 1746 * Keep track of how much data we've examined so that
1745 1747 * zpool(1M) status can make useful progress reports.
1746 1748 */
1747 1749 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1748 1750 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1749 1751
1750 1752 /* if it's a resilver, this may not be in the target range */
1751 1753 if (!needs_io) {
1752 1754 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1753 1755 /*
1754 1756 * Gang members may be spread across multiple
1755 1757 * vdevs, so the best estimate we have is the
1756 1758 * scrub range, which has already been checked.
1757 1759 * XXX -- it would be better to change our
1758 1760 * allocation policy to ensure that all
1759 1761 * gang members reside on the same vdev.
1760 1762 */
1761 1763 needs_io = B_TRUE;
1762 1764 } else {
1763 1765 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1764 1766 phys_birth, 1);
1765 1767 }
1766 1768 }
1767 1769 }
1768 1770
1769 1771 if (needs_io && !zfs_no_scrub_io) {
1770 1772 vdev_t *rvd = spa->spa_root_vdev;
1771 1773 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1772 1774 void *data = zio_data_buf_alloc(size);
1773 1775
1774 1776 mutex_enter(&spa->spa_scrub_lock);
1775 1777 while (spa->spa_scrub_inflight >= maxinflight)
1776 1778 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1777 1779 spa->spa_scrub_inflight++;
1778 1780 mutex_exit(&spa->spa_scrub_lock);
1779 1781
1780 1782 /*
1781 1783 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1782 1784 * then throttle our workload to limit the impact of a scan.
1783 1785 */
1784 1786 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1785 1787 delay(scan_delay);
1786 1788
1787 1789 zio_nowait(zio_read(NULL, spa, bp, data, size,
1788 1790 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1789 1791 zio_flags, zb));
1790 1792 }
1791 1793
1792 1794 /* do not relocate this block */
1793 1795 return (0);
1794 1796 }
1795 1797
1796 1798 int
1797 1799 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1798 1800 {
1799 1801 spa_t *spa = dp->dp_spa;
1800 1802
1801 1803 /*
1802 1804 * Purge all vdev caches and probe all devices. We do this here
1803 1805 * rather than in sync context because this requires a writer lock
1804 1806 * on the spa_config lock, which we can't do from sync context. The
1805 1807 * spa_scrub_reopen flag indicates that vdev_open() should not
1806 1808 * attempt to start another scrub.
1807 1809 */
1808 1810 spa_vdev_state_enter(spa, SCL_NONE);
1809 1811 spa->spa_scrub_reopen = B_TRUE;
1810 1812 vdev_reopen(spa->spa_root_vdev);
1811 1813 spa->spa_scrub_reopen = B_FALSE;
1812 1814 (void) spa_vdev_state_exit(spa, NULL, 0);
1813 1815
1814 1816 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1815 1817 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1816 1818 }
|
↓ open down ↓ |
273 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX