Print this page
5056 ZFS deadlock on db_mtx and dn_holds
Reviewed by: Will Andrews <willa@spectralogic.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Approved by: Dan McDonald <danmcd@omniti.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/dsl_destroy.c
+++ new/usr/src/uts/common/fs/zfs/dsl_destroy.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
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) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
24 24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 25 * Copyright (c) 2013 by Joyent, Inc. All rights reserved.
26 26 */
27 27
28 28 #include <sys/zfs_context.h>
29 29 #include <sys/dsl_userhold.h>
30 30 #include <sys/dsl_dataset.h>
31 31 #include <sys/dsl_synctask.h>
32 32 #include <sys/dmu_tx.h>
33 33 #include <sys/dsl_pool.h>
34 34 #include <sys/dsl_dir.h>
35 35 #include <sys/dmu_traverse.h>
36 36 #include <sys/dsl_scan.h>
37 37 #include <sys/dmu_objset.h>
38 38 #include <sys/zap.h>
39 39 #include <sys/zfeature.h>
40 40 #include <sys/zfs_ioctl.h>
41 41 #include <sys/dsl_deleg.h>
42 42 #include <sys/dmu_impl.h>
43 43
|
↓ open down ↓ |
43 lines elided |
↑ open up ↑ |
44 44 typedef struct dmu_snapshots_destroy_arg {
45 45 nvlist_t *dsda_snaps;
46 46 nvlist_t *dsda_successful_snaps;
47 47 boolean_t dsda_defer;
48 48 nvlist_t *dsda_errlist;
49 49 } dmu_snapshots_destroy_arg_t;
50 50
51 51 int
52 52 dsl_destroy_snapshot_check_impl(dsl_dataset_t *ds, boolean_t defer)
53 53 {
54 - if (!dsl_dataset_is_snapshot(ds))
54 + if (!ds->ds_is_snapshot)
55 55 return (SET_ERROR(EINVAL));
56 56
57 57 if (dsl_dataset_long_held(ds))
58 58 return (SET_ERROR(EBUSY));
59 59
60 60 /*
61 61 * Only allow deferred destroy on pools that support it.
62 62 * NOTE: deferred destroy is only supported on snapshots.
63 63 */
64 64 if (defer) {
65 65 if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
66 66 SPA_VERSION_USERREFS)
67 67 return (SET_ERROR(ENOTSUP));
68 68 return (0);
69 69 }
70 70
71 71 /*
72 72 * If this snapshot has an elevated user reference count,
73 73 * we can't destroy it yet.
74 74 */
75 75 if (ds->ds_userrefs > 0)
76 76 return (SET_ERROR(EBUSY));
77 77
78 78 /*
79 79 * Can't delete a branch point.
80 80 */
81 81 if (dsl_dataset_phys(ds)->ds_num_children > 1)
82 82 return (SET_ERROR(EEXIST));
83 83
84 84 return (0);
85 85 }
86 86
87 87 static int
88 88 dsl_destroy_snapshot_check(void *arg, dmu_tx_t *tx)
89 89 {
90 90 dmu_snapshots_destroy_arg_t *dsda = arg;
91 91 dsl_pool_t *dp = dmu_tx_pool(tx);
92 92 nvpair_t *pair;
93 93 int error = 0;
94 94
95 95 if (!dmu_tx_is_syncing(tx))
96 96 return (0);
97 97
98 98 for (pair = nvlist_next_nvpair(dsda->dsda_snaps, NULL);
99 99 pair != NULL; pair = nvlist_next_nvpair(dsda->dsda_snaps, pair)) {
100 100 dsl_dataset_t *ds;
101 101
102 102 error = dsl_dataset_hold(dp, nvpair_name(pair),
103 103 FTAG, &ds);
104 104
105 105 /*
106 106 * If the snapshot does not exist, silently ignore it
107 107 * (it's "already destroyed").
108 108 */
109 109 if (error == ENOENT)
110 110 continue;
111 111
112 112 if (error == 0) {
113 113 error = dsl_destroy_snapshot_check_impl(ds,
114 114 dsda->dsda_defer);
115 115 dsl_dataset_rele(ds, FTAG);
116 116 }
117 117
118 118 if (error == 0) {
119 119 fnvlist_add_boolean(dsda->dsda_successful_snaps,
120 120 nvpair_name(pair));
121 121 } else {
122 122 fnvlist_add_int32(dsda->dsda_errlist,
123 123 nvpair_name(pair), error);
124 124 }
125 125 }
126 126
127 127 pair = nvlist_next_nvpair(dsda->dsda_errlist, NULL);
128 128 if (pair != NULL)
129 129 return (fnvpair_value_int32(pair));
130 130
131 131 return (0);
132 132 }
133 133
134 134 struct process_old_arg {
135 135 dsl_dataset_t *ds;
136 136 dsl_dataset_t *ds_prev;
137 137 boolean_t after_branch_point;
138 138 zio_t *pio;
139 139 uint64_t used, comp, uncomp;
140 140 };
141 141
142 142 static int
143 143 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
144 144 {
145 145 struct process_old_arg *poa = arg;
146 146 dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
147 147
148 148 ASSERT(!BP_IS_HOLE(bp));
149 149
150 150 if (bp->blk_birth <= dsl_dataset_phys(poa->ds)->ds_prev_snap_txg) {
151 151 dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
152 152 if (poa->ds_prev && !poa->after_branch_point &&
153 153 bp->blk_birth >
154 154 dsl_dataset_phys(poa->ds_prev)->ds_prev_snap_txg) {
155 155 dsl_dataset_phys(poa->ds_prev)->ds_unique_bytes +=
156 156 bp_get_dsize_sync(dp->dp_spa, bp);
157 157 }
158 158 } else {
159 159 poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
160 160 poa->comp += BP_GET_PSIZE(bp);
161 161 poa->uncomp += BP_GET_UCSIZE(bp);
162 162 dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
163 163 }
164 164 return (0);
165 165 }
166 166
167 167 static void
168 168 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
169 169 dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
170 170 {
171 171 struct process_old_arg poa = { 0 };
172 172 dsl_pool_t *dp = ds->ds_dir->dd_pool;
173 173 objset_t *mos = dp->dp_meta_objset;
174 174 uint64_t deadlist_obj;
175 175
176 176 ASSERT(ds->ds_deadlist.dl_oldfmt);
177 177 ASSERT(ds_next->ds_deadlist.dl_oldfmt);
178 178
179 179 poa.ds = ds;
180 180 poa.ds_prev = ds_prev;
181 181 poa.after_branch_point = after_branch_point;
182 182 poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
183 183 VERIFY0(bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
184 184 process_old_cb, &poa, tx));
185 185 VERIFY0(zio_wait(poa.pio));
186 186 ASSERT3U(poa.used, ==, dsl_dataset_phys(ds)->ds_unique_bytes);
187 187
188 188 /* change snapused */
189 189 dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
190 190 -poa.used, -poa.comp, -poa.uncomp, tx);
191 191
192 192 /* swap next's deadlist to our deadlist */
193 193 dsl_deadlist_close(&ds->ds_deadlist);
194 194 dsl_deadlist_close(&ds_next->ds_deadlist);
195 195 deadlist_obj = dsl_dataset_phys(ds)->ds_deadlist_obj;
196 196 dsl_dataset_phys(ds)->ds_deadlist_obj =
197 197 dsl_dataset_phys(ds_next)->ds_deadlist_obj;
198 198 dsl_dataset_phys(ds_next)->ds_deadlist_obj = deadlist_obj;
199 199 dsl_deadlist_open(&ds->ds_deadlist, mos,
200 200 dsl_dataset_phys(ds)->ds_deadlist_obj);
201 201 dsl_deadlist_open(&ds_next->ds_deadlist, mos,
202 202 dsl_dataset_phys(ds_next)->ds_deadlist_obj);
203 203 }
204 204
205 205 static void
206 206 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
207 207 {
208 208 objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
209 209 zap_cursor_t zc;
210 210 zap_attribute_t za;
211 211
212 212 /*
213 213 * If it is the old version, dd_clones doesn't exist so we can't
214 214 * find the clones, but dsl_deadlist_remove_key() is a no-op so it
215 215 * doesn't matter.
216 216 */
217 217 if (dsl_dir_phys(ds->ds_dir)->dd_clones == 0)
218 218 return;
219 219
220 220 for (zap_cursor_init(&zc, mos, dsl_dir_phys(ds->ds_dir)->dd_clones);
221 221 zap_cursor_retrieve(&zc, &za) == 0;
222 222 zap_cursor_advance(&zc)) {
223 223 dsl_dataset_t *clone;
224 224
225 225 VERIFY0(dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
226 226 za.za_first_integer, FTAG, &clone));
227 227 if (clone->ds_dir->dd_origin_txg > mintxg) {
228 228 dsl_deadlist_remove_key(&clone->ds_deadlist,
229 229 mintxg, tx);
230 230 dsl_dataset_remove_clones_key(clone, mintxg, tx);
231 231 }
232 232 dsl_dataset_rele(clone, FTAG);
233 233 }
234 234 zap_cursor_fini(&zc);
235 235 }
236 236
237 237 void
238 238 dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
239 239 {
240 240 int err;
241 241 int after_branch_point = FALSE;
242 242 dsl_pool_t *dp = ds->ds_dir->dd_pool;
243 243 objset_t *mos = dp->dp_meta_objset;
244 244 dsl_dataset_t *ds_prev = NULL;
245 245 uint64_t obj;
246 246
247 247 ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
248 248 ASSERT3U(dsl_dataset_phys(ds)->ds_bp.blk_birth, <=, tx->tx_txg);
249 249 ASSERT(refcount_is_zero(&ds->ds_longholds));
250 250
251 251 if (defer &&
252 252 (ds->ds_userrefs > 0 ||
253 253 dsl_dataset_phys(ds)->ds_num_children > 1)) {
254 254 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
255 255 dmu_buf_will_dirty(ds->ds_dbuf, tx);
256 256 dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_DEFER_DESTROY;
257 257 spa_history_log_internal_ds(ds, "defer_destroy", tx, "");
258 258 return;
259 259 }
260 260
261 261 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
262 262
263 263 /* We need to log before removing it from the namespace. */
264 264 spa_history_log_internal_ds(ds, "destroy", tx, "");
265 265
266 266 dsl_scan_ds_destroyed(ds, tx);
267 267
268 268 obj = ds->ds_object;
269 269
270 270 if (ds->ds_large_blocks) {
271 271 ASSERT0(zap_contains(mos, obj, DS_FIELD_LARGE_BLOCKS));
272 272 spa_feature_decr(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS, tx);
273 273 }
274 274 if (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
275 275 ASSERT3P(ds->ds_prev, ==, NULL);
276 276 VERIFY0(dsl_dataset_hold_obj(dp,
277 277 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &ds_prev));
278 278 after_branch_point =
279 279 (dsl_dataset_phys(ds_prev)->ds_next_snap_obj != obj);
280 280
281 281 dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
282 282 if (after_branch_point &&
283 283 dsl_dataset_phys(ds_prev)->ds_next_clones_obj != 0) {
284 284 dsl_dataset_remove_from_next_clones(ds_prev, obj, tx);
285 285 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
286 286 VERIFY0(zap_add_int(mos,
287 287 dsl_dataset_phys(ds_prev)->
288 288 ds_next_clones_obj,
289 289 dsl_dataset_phys(ds)->ds_next_snap_obj,
290 290 tx));
291 291 }
292 292 }
293 293 if (!after_branch_point) {
294 294 dsl_dataset_phys(ds_prev)->ds_next_snap_obj =
295 295 dsl_dataset_phys(ds)->ds_next_snap_obj;
296 296 }
297 297 }
298 298
299 299 dsl_dataset_t *ds_next;
300 300 uint64_t old_unique;
301 301 uint64_t used = 0, comp = 0, uncomp = 0;
302 302
303 303 VERIFY0(dsl_dataset_hold_obj(dp,
304 304 dsl_dataset_phys(ds)->ds_next_snap_obj, FTAG, &ds_next));
305 305 ASSERT3U(dsl_dataset_phys(ds_next)->ds_prev_snap_obj, ==, obj);
306 306
307 307 old_unique = dsl_dataset_phys(ds_next)->ds_unique_bytes;
308 308
309 309 dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
310 310 dsl_dataset_phys(ds_next)->ds_prev_snap_obj =
311 311 dsl_dataset_phys(ds)->ds_prev_snap_obj;
312 312 dsl_dataset_phys(ds_next)->ds_prev_snap_txg =
313 313 dsl_dataset_phys(ds)->ds_prev_snap_txg;
314 314 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_txg, ==,
315 315 ds_prev ? dsl_dataset_phys(ds_prev)->ds_creation_txg : 0);
316 316
317 317 if (ds_next->ds_deadlist.dl_oldfmt) {
318 318 process_old_deadlist(ds, ds_prev, ds_next,
319 319 after_branch_point, tx);
320 320 } else {
321 321 /* Adjust prev's unique space. */
322 322 if (ds_prev && !after_branch_point) {
323 323 dsl_deadlist_space_range(&ds_next->ds_deadlist,
324 324 dsl_dataset_phys(ds_prev)->ds_prev_snap_txg,
325 325 dsl_dataset_phys(ds)->ds_prev_snap_txg,
326 326 &used, &comp, &uncomp);
327 327 dsl_dataset_phys(ds_prev)->ds_unique_bytes += used;
328 328 }
329 329
330 330 /* Adjust snapused. */
331 331 dsl_deadlist_space_range(&ds_next->ds_deadlist,
332 332 dsl_dataset_phys(ds)->ds_prev_snap_txg, UINT64_MAX,
333 333 &used, &comp, &uncomp);
334 334 dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
335 335 -used, -comp, -uncomp, tx);
336 336
337 337 /* Move blocks to be freed to pool's free list. */
338 338 dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
339 339 &dp->dp_free_bpobj, dsl_dataset_phys(ds)->ds_prev_snap_txg,
340 340 tx);
341 341 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
342 342 DD_USED_HEAD, used, comp, uncomp, tx);
343 343
344 344 /* Merge our deadlist into next's and free it. */
345 345 dsl_deadlist_merge(&ds_next->ds_deadlist,
346 346 dsl_dataset_phys(ds)->ds_deadlist_obj, tx);
|
↓ open down ↓ |
282 lines elided |
↑ open up ↑ |
347 347 }
348 348 dsl_deadlist_close(&ds->ds_deadlist);
349 349 dsl_deadlist_free(mos, dsl_dataset_phys(ds)->ds_deadlist_obj, tx);
350 350 dmu_buf_will_dirty(ds->ds_dbuf, tx);
351 351 dsl_dataset_phys(ds)->ds_deadlist_obj = 0;
352 352
353 353 /* Collapse range in clone heads */
354 354 dsl_dataset_remove_clones_key(ds,
355 355 dsl_dataset_phys(ds)->ds_creation_txg, tx);
356 356
357 - if (dsl_dataset_is_snapshot(ds_next)) {
357 + if (ds_next->ds_is_snapshot) {
358 358 dsl_dataset_t *ds_nextnext;
359 359
360 360 /*
361 361 * Update next's unique to include blocks which
362 362 * were previously shared by only this snapshot
363 363 * and it. Those blocks will be born after the
364 364 * prev snap and before this snap, and will have
365 365 * died after the next snap and before the one
366 366 * after that (ie. be on the snap after next's
367 367 * deadlist).
368 368 */
369 369 VERIFY0(dsl_dataset_hold_obj(dp,
370 370 dsl_dataset_phys(ds_next)->ds_next_snap_obj,
371 371 FTAG, &ds_nextnext));
372 372 dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
373 373 dsl_dataset_phys(ds)->ds_prev_snap_txg,
374 374 dsl_dataset_phys(ds)->ds_creation_txg,
375 375 &used, &comp, &uncomp);
376 376 dsl_dataset_phys(ds_next)->ds_unique_bytes += used;
377 377 dsl_dataset_rele(ds_nextnext, FTAG);
378 378 ASSERT3P(ds_next->ds_prev, ==, NULL);
379 379
380 380 /* Collapse range in this head. */
381 381 dsl_dataset_t *hds;
382 382 VERIFY0(dsl_dataset_hold_obj(dp,
383 383 dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj, FTAG, &hds));
384 384 dsl_deadlist_remove_key(&hds->ds_deadlist,
385 385 dsl_dataset_phys(ds)->ds_creation_txg, tx);
386 386 dsl_dataset_rele(hds, FTAG);
387 387
388 388 } else {
389 389 ASSERT3P(ds_next->ds_prev, ==, ds);
390 390 dsl_dataset_rele(ds_next->ds_prev, ds_next);
391 391 ds_next->ds_prev = NULL;
392 392 if (ds_prev) {
393 393 VERIFY0(dsl_dataset_hold_obj(dp,
394 394 dsl_dataset_phys(ds)->ds_prev_snap_obj,
395 395 ds_next, &ds_next->ds_prev));
396 396 }
397 397
398 398 dsl_dataset_recalc_head_uniq(ds_next);
399 399
400 400 /*
401 401 * Reduce the amount of our unconsumed refreservation
402 402 * being charged to our parent by the amount of
403 403 * new unique data we have gained.
404 404 */
405 405 if (old_unique < ds_next->ds_reserved) {
406 406 int64_t mrsdelta;
407 407 uint64_t new_unique =
408 408 dsl_dataset_phys(ds_next)->ds_unique_bytes;
409 409
410 410 ASSERT(old_unique <= new_unique);
411 411 mrsdelta = MIN(new_unique - old_unique,
412 412 ds_next->ds_reserved - old_unique);
413 413 dsl_dir_diduse_space(ds->ds_dir,
414 414 DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
415 415 }
416 416 }
417 417 dsl_dataset_rele(ds_next, FTAG);
418 418
419 419 /*
420 420 * This must be done after the dsl_traverse(), because it will
421 421 * re-open the objset.
422 422 */
423 423 if (ds->ds_objset) {
424 424 dmu_objset_evict(ds->ds_objset);
425 425 ds->ds_objset = NULL;
426 426 }
427 427
428 428 /* remove from snapshot namespace */
429 429 dsl_dataset_t *ds_head;
430 430 ASSERT(dsl_dataset_phys(ds)->ds_snapnames_zapobj == 0);
431 431 VERIFY0(dsl_dataset_hold_obj(dp,
432 432 dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj, FTAG, &ds_head));
433 433 VERIFY0(dsl_dataset_get_snapname(ds));
434 434 #ifdef ZFS_DEBUG
435 435 {
436 436 uint64_t val;
437 437
438 438 err = dsl_dataset_snap_lookup(ds_head,
439 439 ds->ds_snapname, &val);
440 440 ASSERT0(err);
441 441 ASSERT3U(val, ==, obj);
442 442 }
443 443 #endif
444 444 VERIFY0(dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx, B_TRUE));
445 445 dsl_dataset_rele(ds_head, FTAG);
446 446
447 447 if (ds_prev != NULL)
448 448 dsl_dataset_rele(ds_prev, FTAG);
449 449
450 450 spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
451 451
452 452 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
453 453 uint64_t count;
454 454 ASSERT0(zap_count(mos,
455 455 dsl_dataset_phys(ds)->ds_next_clones_obj, &count) &&
456 456 count == 0);
457 457 VERIFY0(dmu_object_free(mos,
458 458 dsl_dataset_phys(ds)->ds_next_clones_obj, tx));
459 459 }
460 460 if (dsl_dataset_phys(ds)->ds_props_obj != 0)
461 461 VERIFY0(zap_destroy(mos, dsl_dataset_phys(ds)->ds_props_obj,
462 462 tx));
463 463 if (dsl_dataset_phys(ds)->ds_userrefs_obj != 0)
464 464 VERIFY0(zap_destroy(mos, dsl_dataset_phys(ds)->ds_userrefs_obj,
465 465 tx));
466 466 dsl_dir_rele(ds->ds_dir, ds);
467 467 ds->ds_dir = NULL;
468 468 dmu_object_free_zapified(mos, obj, tx);
469 469 }
470 470
471 471 static void
472 472 dsl_destroy_snapshot_sync(void *arg, dmu_tx_t *tx)
473 473 {
474 474 dmu_snapshots_destroy_arg_t *dsda = arg;
475 475 dsl_pool_t *dp = dmu_tx_pool(tx);
476 476 nvpair_t *pair;
477 477
478 478 for (pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, NULL);
479 479 pair != NULL;
480 480 pair = nvlist_next_nvpair(dsda->dsda_successful_snaps, pair)) {
481 481 dsl_dataset_t *ds;
482 482
483 483 VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds));
484 484
485 485 dsl_destroy_snapshot_sync_impl(ds, dsda->dsda_defer, tx);
486 486 dsl_dataset_rele(ds, FTAG);
487 487 }
488 488 }
489 489
490 490 /*
491 491 * The semantics of this function are described in the comment above
492 492 * lzc_destroy_snaps(). To summarize:
493 493 *
494 494 * The snapshots must all be in the same pool.
495 495 *
496 496 * Snapshots that don't exist will be silently ignored (considered to be
497 497 * "already deleted").
498 498 *
499 499 * On success, all snaps will be destroyed and this will return 0.
500 500 * On failure, no snaps will be destroyed, the errlist will be filled in,
501 501 * and this will return an errno.
502 502 */
503 503 int
504 504 dsl_destroy_snapshots_nvl(nvlist_t *snaps, boolean_t defer,
505 505 nvlist_t *errlist)
506 506 {
507 507 dmu_snapshots_destroy_arg_t dsda;
508 508 int error;
509 509 nvpair_t *pair;
510 510
511 511 pair = nvlist_next_nvpair(snaps, NULL);
512 512 if (pair == NULL)
513 513 return (0);
514 514
515 515 dsda.dsda_snaps = snaps;
516 516 dsda.dsda_successful_snaps = fnvlist_alloc();
517 517 dsda.dsda_defer = defer;
518 518 dsda.dsda_errlist = errlist;
519 519
520 520 error = dsl_sync_task(nvpair_name(pair),
521 521 dsl_destroy_snapshot_check, dsl_destroy_snapshot_sync,
522 522 &dsda, 0, ZFS_SPACE_CHECK_NONE);
523 523 fnvlist_free(dsda.dsda_successful_snaps);
524 524
525 525 return (error);
526 526 }
527 527
528 528 int
529 529 dsl_destroy_snapshot(const char *name, boolean_t defer)
530 530 {
531 531 int error;
532 532 nvlist_t *nvl = fnvlist_alloc();
533 533 nvlist_t *errlist = fnvlist_alloc();
534 534
535 535 fnvlist_add_boolean(nvl, name);
536 536 error = dsl_destroy_snapshots_nvl(nvl, defer, errlist);
537 537 fnvlist_free(errlist);
538 538 fnvlist_free(nvl);
539 539 return (error);
540 540 }
541 541
542 542 struct killarg {
543 543 dsl_dataset_t *ds;
544 544 dmu_tx_t *tx;
545 545 };
546 546
547 547 /* ARGSUSED */
548 548 static int
549 549 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
550 550 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
551 551 {
552 552 struct killarg *ka = arg;
553 553 dmu_tx_t *tx = ka->tx;
554 554
555 555 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
556 556 return (0);
557 557
558 558 if (zb->zb_level == ZB_ZIL_LEVEL) {
559 559 ASSERT(zilog != NULL);
560 560 /*
561 561 * It's a block in the intent log. It has no
562 562 * accounting, so just free it.
563 563 */
564 564 dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
565 565 } else {
566 566 ASSERT(zilog == NULL);
567 567 ASSERT3U(bp->blk_birth, >,
568 568 dsl_dataset_phys(ka->ds)->ds_prev_snap_txg);
569 569 (void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
570 570 }
571 571
572 572 return (0);
573 573 }
574 574
575 575 static void
576 576 old_synchronous_dataset_destroy(dsl_dataset_t *ds, dmu_tx_t *tx)
577 577 {
578 578 struct killarg ka;
579 579
580 580 /*
581 581 * Free everything that we point to (that's born after
582 582 * the previous snapshot, if we are a clone)
583 583 *
584 584 * NB: this should be very quick, because we already
585 585 * freed all the objects in open context.
586 586 */
587 587 ka.ds = ds;
588 588 ka.tx = tx;
589 589 VERIFY0(traverse_dataset(ds,
590 590 dsl_dataset_phys(ds)->ds_prev_snap_txg, TRAVERSE_POST,
591 591 kill_blkptr, &ka));
592 592 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
593 593 dsl_dataset_phys(ds)->ds_unique_bytes == 0);
594 594 }
595 595
596 596 typedef struct dsl_destroy_head_arg {
|
↓ open down ↓ |
229 lines elided |
↑ open up ↑ |
597 597 const char *ddha_name;
598 598 } dsl_destroy_head_arg_t;
599 599
600 600 int
601 601 dsl_destroy_head_check_impl(dsl_dataset_t *ds, int expected_holds)
602 602 {
603 603 int error;
604 604 uint64_t count;
605 605 objset_t *mos;
606 606
607 - ASSERT(!dsl_dataset_is_snapshot(ds));
608 - if (dsl_dataset_is_snapshot(ds))
607 + ASSERT(!ds->ds_is_snapshot);
608 + if (ds->ds_is_snapshot)
609 609 return (SET_ERROR(EINVAL));
610 610
611 611 if (refcount_count(&ds->ds_longholds) != expected_holds)
612 612 return (SET_ERROR(EBUSY));
613 613
614 614 mos = ds->ds_dir->dd_pool->dp_meta_objset;
615 615
616 616 /*
617 617 * Can't delete a head dataset if there are snapshots of it.
618 618 * (Except if the only snapshots are from the branch we cloned
619 619 * from.)
620 620 */
621 621 if (ds->ds_prev != NULL &&
622 622 dsl_dataset_phys(ds->ds_prev)->ds_next_snap_obj == ds->ds_object)
623 623 return (SET_ERROR(EBUSY));
624 624
625 625 /*
626 626 * Can't delete if there are children of this fs.
627 627 */
628 628 error = zap_count(mos,
629 629 dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, &count);
630 630 if (error != 0)
631 631 return (error);
632 632 if (count != 0)
633 633 return (SET_ERROR(EEXIST));
634 634
635 635 if (dsl_dir_is_clone(ds->ds_dir) && DS_IS_DEFER_DESTROY(ds->ds_prev) &&
636 636 dsl_dataset_phys(ds->ds_prev)->ds_num_children == 2 &&
637 637 ds->ds_prev->ds_userrefs == 0) {
638 638 /* We need to remove the origin snapshot as well. */
639 639 if (!refcount_is_zero(&ds->ds_prev->ds_longholds))
640 640 return (SET_ERROR(EBUSY));
641 641 }
642 642 return (0);
643 643 }
644 644
645 645 static int
646 646 dsl_destroy_head_check(void *arg, dmu_tx_t *tx)
647 647 {
648 648 dsl_destroy_head_arg_t *ddha = arg;
649 649 dsl_pool_t *dp = dmu_tx_pool(tx);
650 650 dsl_dataset_t *ds;
651 651 int error;
652 652
653 653 error = dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds);
654 654 if (error != 0)
655 655 return (error);
656 656
657 657 error = dsl_destroy_head_check_impl(ds, 0);
658 658 dsl_dataset_rele(ds, FTAG);
659 659 return (error);
660 660 }
661 661
662 662 static void
663 663 dsl_dir_destroy_sync(uint64_t ddobj, dmu_tx_t *tx)
664 664 {
665 665 dsl_dir_t *dd;
666 666 dsl_pool_t *dp = dmu_tx_pool(tx);
667 667 objset_t *mos = dp->dp_meta_objset;
668 668 dd_used_t t;
669 669
670 670 ASSERT(RRW_WRITE_HELD(&dmu_tx_pool(tx)->dp_config_rwlock));
671 671
672 672 VERIFY0(dsl_dir_hold_obj(dp, ddobj, NULL, FTAG, &dd));
673 673
674 674 ASSERT0(dsl_dir_phys(dd)->dd_head_dataset_obj);
675 675
676 676 /*
677 677 * Decrement the filesystem count for all parent filesystems.
678 678 *
679 679 * When we receive an incremental stream into a filesystem that already
680 680 * exists, a temporary clone is created. We never count this temporary
681 681 * clone, whose name begins with a '%'.
682 682 */
683 683 if (dd->dd_myname[0] != '%' && dd->dd_parent != NULL)
684 684 dsl_fs_ss_count_adjust(dd->dd_parent, -1,
685 685 DD_FIELD_FILESYSTEM_COUNT, tx);
686 686
687 687 /*
688 688 * Remove our reservation. The impl() routine avoids setting the
689 689 * actual property, which would require the (already destroyed) ds.
690 690 */
691 691 dsl_dir_set_reservation_sync_impl(dd, 0, tx);
692 692
693 693 ASSERT0(dsl_dir_phys(dd)->dd_used_bytes);
694 694 ASSERT0(dsl_dir_phys(dd)->dd_reserved);
695 695 for (t = 0; t < DD_USED_NUM; t++)
696 696 ASSERT0(dsl_dir_phys(dd)->dd_used_breakdown[t]);
697 697
698 698 VERIFY0(zap_destroy(mos, dsl_dir_phys(dd)->dd_child_dir_zapobj, tx));
699 699 VERIFY0(zap_destroy(mos, dsl_dir_phys(dd)->dd_props_zapobj, tx));
700 700 VERIFY0(dsl_deleg_destroy(mos, dsl_dir_phys(dd)->dd_deleg_zapobj, tx));
701 701 VERIFY0(zap_remove(mos,
702 702 dsl_dir_phys(dd->dd_parent)->dd_child_dir_zapobj,
703 703 dd->dd_myname, tx));
704 704
705 705 dsl_dir_rele(dd, FTAG);
706 706 dmu_object_free_zapified(mos, ddobj, tx);
707 707 }
708 708
709 709 void
710 710 dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
711 711 {
712 712 dsl_pool_t *dp = dmu_tx_pool(tx);
713 713 objset_t *mos = dp->dp_meta_objset;
714 714 uint64_t obj, ddobj, prevobj = 0;
715 715 boolean_t rmorigin;
716 716
717 717 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
718 718 ASSERT(ds->ds_prev == NULL ||
719 719 dsl_dataset_phys(ds->ds_prev)->ds_next_snap_obj != ds->ds_object);
720 720 ASSERT3U(dsl_dataset_phys(ds)->ds_bp.blk_birth, <=, tx->tx_txg);
721 721 ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock));
722 722
723 723 /* We need to log before removing it from the namespace. */
724 724 spa_history_log_internal_ds(ds, "destroy", tx, "");
725 725
726 726 rmorigin = (dsl_dir_is_clone(ds->ds_dir) &&
727 727 DS_IS_DEFER_DESTROY(ds->ds_prev) &&
728 728 dsl_dataset_phys(ds->ds_prev)->ds_num_children == 2 &&
729 729 ds->ds_prev->ds_userrefs == 0);
730 730
731 731 /* Remove our reservation. */
732 732 if (ds->ds_reserved != 0) {
733 733 dsl_dataset_set_refreservation_sync_impl(ds,
734 734 (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
735 735 0, tx);
736 736 ASSERT0(ds->ds_reserved);
737 737 }
738 738
739 739 if (ds->ds_large_blocks)
740 740 spa_feature_decr(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS, tx);
741 741
742 742 dsl_scan_ds_destroyed(ds, tx);
743 743
744 744 obj = ds->ds_object;
745 745
746 746 if (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
747 747 /* This is a clone */
748 748 ASSERT(ds->ds_prev != NULL);
749 749 ASSERT3U(dsl_dataset_phys(ds->ds_prev)->ds_next_snap_obj, !=,
750 750 obj);
751 751 ASSERT0(dsl_dataset_phys(ds)->ds_next_snap_obj);
752 752
753 753 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
754 754 if (dsl_dataset_phys(ds->ds_prev)->ds_next_clones_obj != 0) {
755 755 dsl_dataset_remove_from_next_clones(ds->ds_prev,
756 756 obj, tx);
757 757 }
758 758
759 759 ASSERT3U(dsl_dataset_phys(ds->ds_prev)->ds_num_children, >, 1);
760 760 dsl_dataset_phys(ds->ds_prev)->ds_num_children--;
761 761 }
762 762
763 763 /*
764 764 * Destroy the deadlist. Unless it's a clone, the
765 765 * deadlist should be empty. (If it's a clone, it's
766 766 * safe to ignore the deadlist contents.)
767 767 */
768 768 dsl_deadlist_close(&ds->ds_deadlist);
769 769 dsl_deadlist_free(mos, dsl_dataset_phys(ds)->ds_deadlist_obj, tx);
770 770 dmu_buf_will_dirty(ds->ds_dbuf, tx);
771 771 dsl_dataset_phys(ds)->ds_deadlist_obj = 0;
772 772
773 773 objset_t *os;
774 774 VERIFY0(dmu_objset_from_ds(ds, &os));
775 775
776 776 if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
777 777 old_synchronous_dataset_destroy(ds, tx);
778 778 } else {
779 779 /*
780 780 * Move the bptree into the pool's list of trees to
781 781 * clean up and update space accounting information.
782 782 */
783 783 uint64_t used, comp, uncomp;
784 784
785 785 zil_destroy_sync(dmu_objset_zil(os), tx);
786 786
787 787 if (!spa_feature_is_active(dp->dp_spa,
788 788 SPA_FEATURE_ASYNC_DESTROY)) {
789 789 dsl_scan_t *scn = dp->dp_scan;
790 790 spa_feature_incr(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY,
791 791 tx);
792 792 dp->dp_bptree_obj = bptree_alloc(mos, tx);
793 793 VERIFY0(zap_add(mos,
794 794 DMU_POOL_DIRECTORY_OBJECT,
795 795 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
796 796 &dp->dp_bptree_obj, tx));
797 797 ASSERT(!scn->scn_async_destroying);
798 798 scn->scn_async_destroying = B_TRUE;
799 799 }
800 800
801 801 used = dsl_dir_phys(ds->ds_dir)->dd_used_bytes;
802 802 comp = dsl_dir_phys(ds->ds_dir)->dd_compressed_bytes;
803 803 uncomp = dsl_dir_phys(ds->ds_dir)->dd_uncompressed_bytes;
804 804
805 805 ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
806 806 dsl_dataset_phys(ds)->ds_unique_bytes == used);
807 807
808 808 bptree_add(mos, dp->dp_bptree_obj,
809 809 &dsl_dataset_phys(ds)->ds_bp,
810 810 dsl_dataset_phys(ds)->ds_prev_snap_txg,
811 811 used, comp, uncomp, tx);
812 812 dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
813 813 -used, -comp, -uncomp, tx);
814 814 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
815 815 used, comp, uncomp, tx);
816 816 }
817 817
818 818 if (ds->ds_prev != NULL) {
819 819 if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
820 820 VERIFY0(zap_remove_int(mos,
821 821 dsl_dir_phys(ds->ds_prev->ds_dir)->dd_clones,
822 822 ds->ds_object, tx));
823 823 }
824 824 prevobj = ds->ds_prev->ds_object;
825 825 dsl_dataset_rele(ds->ds_prev, ds);
826 826 ds->ds_prev = NULL;
827 827 }
828 828
829 829 /*
830 830 * This must be done after the dsl_traverse(), because it will
831 831 * re-open the objset.
832 832 */
833 833 if (ds->ds_objset) {
834 834 dmu_objset_evict(ds->ds_objset);
835 835 ds->ds_objset = NULL;
836 836 }
837 837
838 838 /* Erase the link in the dir */
839 839 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
840 840 dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj = 0;
841 841 ddobj = ds->ds_dir->dd_object;
842 842 ASSERT(dsl_dataset_phys(ds)->ds_snapnames_zapobj != 0);
843 843 VERIFY0(zap_destroy(mos,
844 844 dsl_dataset_phys(ds)->ds_snapnames_zapobj, tx));
845 845
846 846 if (ds->ds_bookmarks != 0) {
847 847 VERIFY0(zap_destroy(mos, ds->ds_bookmarks, tx));
848 848 spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
849 849 }
850 850
851 851 spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
852 852
853 853 ASSERT0(dsl_dataset_phys(ds)->ds_next_clones_obj);
854 854 ASSERT0(dsl_dataset_phys(ds)->ds_props_obj);
855 855 ASSERT0(dsl_dataset_phys(ds)->ds_userrefs_obj);
856 856 dsl_dir_rele(ds->ds_dir, ds);
857 857 ds->ds_dir = NULL;
858 858 dmu_object_free_zapified(mos, obj, tx);
859 859
860 860 dsl_dir_destroy_sync(ddobj, tx);
861 861
862 862 if (rmorigin) {
863 863 dsl_dataset_t *prev;
864 864 VERIFY0(dsl_dataset_hold_obj(dp, prevobj, FTAG, &prev));
865 865 dsl_destroy_snapshot_sync_impl(prev, B_FALSE, tx);
866 866 dsl_dataset_rele(prev, FTAG);
867 867 }
868 868 }
869 869
870 870 static void
871 871 dsl_destroy_head_sync(void *arg, dmu_tx_t *tx)
872 872 {
873 873 dsl_destroy_head_arg_t *ddha = arg;
874 874 dsl_pool_t *dp = dmu_tx_pool(tx);
875 875 dsl_dataset_t *ds;
876 876
877 877 VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
878 878 dsl_destroy_head_sync_impl(ds, tx);
879 879 dsl_dataset_rele(ds, FTAG);
880 880 }
881 881
882 882 static void
883 883 dsl_destroy_head_begin_sync(void *arg, dmu_tx_t *tx)
884 884 {
885 885 dsl_destroy_head_arg_t *ddha = arg;
886 886 dsl_pool_t *dp = dmu_tx_pool(tx);
887 887 dsl_dataset_t *ds;
888 888
889 889 VERIFY0(dsl_dataset_hold(dp, ddha->ddha_name, FTAG, &ds));
890 890
891 891 /* Mark it as inconsistent on-disk, in case we crash */
892 892 dmu_buf_will_dirty(ds->ds_dbuf, tx);
893 893 dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
894 894
895 895 spa_history_log_internal_ds(ds, "destroy begin", tx, "");
896 896 dsl_dataset_rele(ds, FTAG);
897 897 }
898 898
899 899 int
900 900 dsl_destroy_head(const char *name)
901 901 {
902 902 dsl_destroy_head_arg_t ddha;
903 903 int error;
904 904 spa_t *spa;
905 905 boolean_t isenabled;
906 906
907 907 #ifdef _KERNEL
908 908 zfs_destroy_unmount_origin(name);
909 909 #endif
910 910
911 911 error = spa_open(name, &spa, FTAG);
912 912 if (error != 0)
913 913 return (error);
914 914 isenabled = spa_feature_is_enabled(spa, SPA_FEATURE_ASYNC_DESTROY);
915 915 spa_close(spa, FTAG);
916 916
917 917 ddha.ddha_name = name;
918 918
919 919 if (!isenabled) {
920 920 objset_t *os;
921 921
922 922 error = dsl_sync_task(name, dsl_destroy_head_check,
923 923 dsl_destroy_head_begin_sync, &ddha,
924 924 0, ZFS_SPACE_CHECK_NONE);
925 925 if (error != 0)
926 926 return (error);
927 927
928 928 /*
929 929 * Head deletion is processed in one txg on old pools;
930 930 * remove the objects from open context so that the txg sync
931 931 * is not too long.
932 932 */
933 933 error = dmu_objset_own(name, DMU_OST_ANY, B_FALSE, FTAG, &os);
934 934 if (error == 0) {
935 935 uint64_t prev_snap_txg =
936 936 dsl_dataset_phys(dmu_objset_ds(os))->
937 937 ds_prev_snap_txg;
938 938 for (uint64_t obj = 0; error == 0;
939 939 error = dmu_object_next(os, &obj, FALSE,
940 940 prev_snap_txg))
941 941 (void) dmu_free_long_object(os, obj);
942 942 /* sync out all frees */
943 943 txg_wait_synced(dmu_objset_pool(os), 0);
944 944 dmu_objset_disown(os, FTAG);
945 945 }
946 946 }
947 947
948 948 return (dsl_sync_task(name, dsl_destroy_head_check,
949 949 dsl_destroy_head_sync, &ddha, 0, ZFS_SPACE_CHECK_NONE));
950 950 }
951 951
952 952 /*
953 953 * Note, this function is used as the callback for dmu_objset_find(). We
954 954 * always return 0 so that we will continue to find and process
955 955 * inconsistent datasets, even if we encounter an error trying to
956 956 * process one of them.
957 957 */
958 958 /* ARGSUSED */
959 959 int
960 960 dsl_destroy_inconsistent(const char *dsname, void *arg)
961 961 {
962 962 objset_t *os;
963 963
964 964 if (dmu_objset_hold(dsname, FTAG, &os) == 0) {
965 965 boolean_t inconsistent = DS_IS_INCONSISTENT(dmu_objset_ds(os));
966 966 dmu_objset_rele(os, FTAG);
967 967 if (inconsistent)
968 968 (void) dsl_destroy_head(dsname);
969 969 }
970 970 return (0);
971 971 }
|
↓ open down ↓ |
353 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX