1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26 * Copyright (c) 2014 Integros [integros.com]
27 * Copyright 2016 Nexenta Systems, Inc. All rights reserved.
28 */
29
30 #include <sys/autosnap.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_synctask.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_scan.h>
38 #include <sys/dnode.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dmu_objset.h>
41 #include <sys/dmu_traverse.h>
42 #include <sys/arc.h>
43 #include <sys/zap.h>
44 #include <sys/zio.h>
45 #include <sys/zfs_context.h>
46 #include <sys/fs/zfs.h>
47 #include <sys/zfs_znode.h>
48 #include <sys/spa_impl.h>
49 #include <sys/dsl_deadlist.h>
50 #include <sys/bptree.h>
51 #include <sys/zfeature.h>
52 #include <sys/zil_impl.h>
53 #include <sys/dsl_userhold.h>
54
55 #include <sys/wbc.h>
56 #include <sys/time.h>
57
58 /*
59 * ZFS Write Throttle
60 * ------------------
61 *
62 * ZFS must limit the rate of incoming writes to the rate at which it is able
63 * to sync data modifications to the backend storage. Throttling by too much
64 * creates an artificial limit; throttling by too little can only be sustained
65 * for short periods and would lead to highly lumpy performance. On a per-pool
66 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
67 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
68 * of dirty data decreases. When the amount of dirty data exceeds a
69 * predetermined threshold further modifications are blocked until the amount
70 * of dirty data decreases (as data is synced out).
71 *
72 * The limit on dirty data is tunable, and should be adjusted according to
73 * both the IO capacity and available memory of the system. The larger the
74 * window, the more ZFS is able to aggregate and amortize metadata (and data)
75 * changes. However, memory is a limited resource, and allowing for more dirty
76 * data comes at the cost of keeping other useful data in memory (for example
77 * ZFS data cached by the ARC).
78 *
79 * Implementation
80 *
81 * As buffers are modified dsl_pool_willuse_space() increments both the per-
82 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
83 * dirty space used; dsl_pool_dirty_space() decrements those values as data
84 * is synced out from dsl_pool_sync(). While only the poolwide value is
85 * relevant, the per-txg value is useful for debugging. The tunable
86 * zfs_dirty_data_max determines the dirty space limit. Once that value is
87 * exceeded, new writes are halted until space frees up.
88 *
89 * The zfs_dirty_data_sync tunable dictates the threshold at which we
90 * ensure that there is a txg syncing (see the comment in txg.c for a full
91 * description of transaction group stages).
92 *
93 * The IO scheduler uses both the dirty space limit and current amount of
94 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
95 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
96 *
97 * The delay is also calculated based on the amount of dirty data. See the
98 * comment above dmu_tx_delay() for details.
99 */
100
101 /*
102 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
103 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
104 */
105 uint64_t zfs_dirty_data_max;
106 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
107 int zfs_dirty_data_max_percent = 10;
108
109 /*
110 * If there is at least this much dirty data, push out a txg.
111 */
112 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
113
114 /*
115 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
116 * and delay each transaction.
117 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
118 */
119 int zfs_delay_min_dirty_percent = 60;
120
121 /*
122 * This controls how quickly the delay approaches infinity.
123 * Larger values cause it to delay more for a given amount of dirty data.
124 * Therefore larger values will cause there to be less dirty data for a
125 * given throughput.
126 *
127 * For the smoothest delay, this value should be about 1 billion divided
128 * by the maximum number of operations per second. This will smoothly
129 * handle between 10x and 1/10th this number.
130 *
131 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
132 * multiply in dmu_tx_delay().
133 */
134 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
135
136 /*
137 * This determines the number of threads used by the dp_sync_taskq.
138 */
139 int zfs_sync_taskq_batch_pct = 75;
140
141 /*
142 * These tunables determine the behavior of how zil_itxg_clean() is
143 * called via zil_clean() in the context of spa_sync(). When an itxg
144 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching.
145 * If the dispatch fails, the call to zil_itxg_clean() will occur
146 * synchronously in the context of spa_sync(), which can negatively
147 * impact the performance of spa_sync() (e.g. in the case of the itxg
148 * list having a large number of itxs that needs to be cleaned).
149 *
150 * Thus, these tunables can be used to manipulate the behavior of the
151 * taskq used by zil_clean(); they determine the number of taskq entries
152 * that are pre-populated when the taskq is first created (via the
153 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
154 * taskq entries that are cached after an on-demand allocation (via the
155 * "zfs_zil_clean_taskq_maxalloc").
156 *
157 * The idea being, we want to try reasonably hard to ensure there will
158 * already be a taskq entry pre-allocated by the time that it is needed
159 * by zil_clean(). This way, we can avoid the possibility of an
160 * on-demand allocation of a new taskq entry from failing, which would
161 * result in zil_itxg_clean() being called synchronously from zil_clean()
162 * (which can adversely affect performance of spa_sync()).
163 *
164 * Additionally, the number of threads used by the taskq can be
165 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
166 */
167 int zfs_zil_clean_taskq_nthr_pct = 100;
168 int zfs_zil_clean_taskq_minalloc = 1024;
169 int zfs_zil_clean_taskq_maxalloc = 1024 * 1024;
170
171 /*
172 * Tunable to control max number of tasks available for processing of
173 * deferred deletes.
174 */
175 int zfs_vn_rele_max_tasks = 256;
176
177 int
178 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
179 {
180 uint64_t obj;
181 int err;
182
183 err = zap_lookup(dp->dp_meta_objset,
184 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
185 name, sizeof (obj), 1, &obj);
186 if (err)
187 return (err);
188
189 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
190 }
191
192 static dsl_pool_t *
193 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
194 {
195 dsl_pool_t *dp;
196 blkptr_t *bp = spa_get_rootblkptr(spa);
197
198 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
199 dp->dp_spa = spa;
200 dp->dp_meta_rootbp = *bp;
201 rrw_init(&dp->dp_config_rwlock, B_TRUE);
202
203 dp->dp_sync_history[0] = dp->dp_sync_history[1] = 0;
204
205 txg_init(dp, txg);
206
207 txg_list_create(&dp->dp_dirty_datasets, spa,
208 offsetof(dsl_dataset_t, ds_dirty_link));
209 txg_list_create(&dp->dp_dirty_zilogs, spa,
210 offsetof(zilog_t, zl_dirty_link));
211 txg_list_create(&dp->dp_dirty_dirs, spa,
212 offsetof(dsl_dir_t, dd_dirty_link));
213 txg_list_create(&dp->dp_sync_tasks, spa,
214 offsetof(dsl_sync_task_t, dst_node));
215
216 dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
217 zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
218 TASKQ_THREADS_CPU_PCT);
219
220 dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq",
221 zfs_zil_clean_taskq_nthr_pct, minclsyspri,
222 zfs_zil_clean_taskq_minalloc,
223 zfs_zil_clean_taskq_maxalloc,
224 TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT);
225
226 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
227 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
228
229 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq",
230 zfs_vn_rele_max_tasks, minclsyspri,
231 1, zfs_vn_rele_max_tasks, TASKQ_DYNAMIC);
232
233 return (dp);
234 }
235
236 int
237 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
238 {
239 int err;
240 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
241
242 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
243 &dp->dp_meta_objset);
244 if (err != 0)
245 dsl_pool_close(dp);
246 else
247 *dpp = dp;
248
249 return (err);
250 }
251
252 int
253 dsl_pool_open(dsl_pool_t *dp)
254 {
255 int err;
256 dsl_dir_t *dd;
257 dsl_dataset_t *ds;
258 uint64_t obj;
259
260 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
261 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
262 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
263 &dp->dp_root_dir_obj);
264 if (err)
265 goto out;
266
267 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
268 NULL, dp, &dp->dp_root_dir);
269 if (err)
270 goto out;
271
272 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
273 if (err)
274 goto out;
275
276 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
277 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
278 if (err)
279 goto out;
280 err = dsl_dataset_hold_obj(dp,
281 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
282 if (err == 0) {
283 err = dsl_dataset_hold_obj(dp,
284 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
285 &dp->dp_origin_snap);
286 dsl_dataset_rele(ds, FTAG);
287 }
288 dsl_dir_rele(dd, dp);
289 if (err)
290 goto out;
291 }
292
293 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
294 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
295 &dp->dp_free_dir);
296 if (err)
297 goto out;
298
299 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
300 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
301 if (err)
302 goto out;
303 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
304 dp->dp_meta_objset, obj));
305 }
306
307 /*
308 * Note: errors ignored, because the leak dir will not exist if we
309 * have not encountered a leak yet.
310 */
311 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
312 &dp->dp_leak_dir);
313
314 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
315 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
316 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
317 &dp->dp_bptree_obj);
318 if (err != 0)
319 goto out;
320 }
321
322 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
323 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
324 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
325 &dp->dp_empty_bpobj);
326 if (err != 0)
327 goto out;
328 }
329
330 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
331 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
332 &dp->dp_tmp_userrefs_obj);
333 if (err == ENOENT)
334 err = 0;
335 if (err)
336 goto out;
337
338 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
339
340 out:
341 rrw_exit(&dp->dp_config_rwlock, FTAG);
342 return (err);
343 }
344
345 void
346 dsl_pool_close(dsl_pool_t *dp)
347 {
348 /*
349 * Drop our references from dsl_pool_open().
350 *
351 * Since we held the origin_snap from "syncing" context (which
352 * includes pool-opening context), it actually only got a "ref"
353 * and not a hold, so just drop that here.
354 */
355 if (dp->dp_origin_snap)
356 dsl_dataset_rele(dp->dp_origin_snap, dp);
357 if (dp->dp_mos_dir)
358 dsl_dir_rele(dp->dp_mos_dir, dp);
359 if (dp->dp_free_dir)
360 dsl_dir_rele(dp->dp_free_dir, dp);
361 if (dp->dp_leak_dir)
362 dsl_dir_rele(dp->dp_leak_dir, dp);
363 if (dp->dp_root_dir)
364 dsl_dir_rele(dp->dp_root_dir, dp);
365
366 bpobj_close(&dp->dp_free_bpobj);
367
368 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
369 if (dp->dp_meta_objset)
370 dmu_objset_evict(dp->dp_meta_objset);
371
372 txg_list_destroy(&dp->dp_dirty_datasets);
373 txg_list_destroy(&dp->dp_dirty_zilogs);
374 txg_list_destroy(&dp->dp_sync_tasks);
375 txg_list_destroy(&dp->dp_dirty_dirs);
376
377 taskq_destroy(dp->dp_zil_clean_taskq);
378 taskq_destroy(dp->dp_sync_taskq);
379
380 /*
381 * We can't set retry to TRUE since we're explicitly specifying
382 * a spa to flush. This is good enough; any missed buffers for
383 * this spa won't cause trouble, and they'll eventually fall
384 * out of the ARC just like any other unused buffer.
385 */
386 arc_flush(dp->dp_spa, B_FALSE);
387 txg_fini(dp);
388 dsl_scan_fini(dp);
389 dmu_buf_user_evict_wait();
390
391 rrw_destroy(&dp->dp_config_rwlock);
392 mutex_destroy(&dp->dp_lock);
393 taskq_destroy(dp->dp_vnrele_taskq);
394 if (dp->dp_blkstats)
395 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
396 kmem_free(dp, sizeof (dsl_pool_t));
397 }
398
399 dsl_pool_t *
400 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
401 {
402 int err;
403 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
404 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
405 dsl_dataset_t *ds;
406 uint64_t obj;
407
408 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
409
410 /* create and open the MOS (meta-objset) */
411 dp->dp_meta_objset = dmu_objset_create_impl(spa,
412 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
413
414 /* create the pool directory */
415 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
416 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
417 ASSERT0(err);
418
419 /* Initialize scan structures */
420 VERIFY0(dsl_scan_init(dp, txg));
421
422 /* create and open the root dir */
423 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
424 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
425 NULL, dp, &dp->dp_root_dir));
426
427 /* create and open the meta-objset dir */
428 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
429 VERIFY0(dsl_pool_open_special_dir(dp,
430 MOS_DIR_NAME, &dp->dp_mos_dir));
431
432 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
433 /* create and open the free dir */
434 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
435 FREE_DIR_NAME, tx);
436 VERIFY0(dsl_pool_open_special_dir(dp,
437 FREE_DIR_NAME, &dp->dp_free_dir));
438
439 /* create and open the free_bplist */
440 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
441 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
442 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
443 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
444 dp->dp_meta_objset, obj));
445 }
446
447 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
448 dsl_pool_create_origin(dp, tx);
449
450 /* create the root dataset */
451 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
452
453 /* create the root objset */
454 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
455 #ifdef _KERNEL
456 {
457 objset_t *os;
458 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
459 os = dmu_objset_create_impl(dp->dp_spa, ds,
460 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
461 rrw_exit(&ds->ds_bp_rwlock, FTAG);
462 zfs_create_fs(os, kcred, zplprops, tx);
463 }
464 #endif
465 dsl_dataset_rele(ds, FTAG);
466
467 dmu_tx_commit(tx);
468
469 rrw_exit(&dp->dp_config_rwlock, FTAG);
470
471 return (dp);
472 }
473
474 /*
475 * Account for the meta-objset space in its placeholder dsl_dir.
476 */
477 void
478 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
479 int64_t used, int64_t comp, int64_t uncomp)
480 {
481 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
482 mutex_enter(&dp->dp_lock);
483 dp->dp_mos_used_delta += used;
484 dp->dp_mos_compressed_delta += comp;
485 dp->dp_mos_uncompressed_delta += uncomp;
486 mutex_exit(&dp->dp_lock);
487 }
488
489 static void
490 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
491 {
492 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
493 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
494 VERIFY0(zio_wait(zio));
495 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
496 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
497 }
498
499 static void
500 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
501 {
502 ASSERT(MUTEX_HELD(&dp->dp_lock));
503
504 if (delta < 0)
505 ASSERT3U(-delta, <=, dp->dp_dirty_total);
506
507 dp->dp_dirty_total += delta;
508
509 /*
510 * Note: we signal even when increasing dp_dirty_total.
511 * This ensures forward progress -- each thread wakes the next waiter.
512 */
513 if (dp->dp_dirty_total < zfs_dirty_data_max)
514 cv_signal(&dp->dp_spaceavail_cv);
515 }
516
517 void
518 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
519 {
520 zio_t *zio;
521 dmu_tx_t *tx;
522 dsl_dir_t *dd;
523 dsl_dataset_t *ds;
524 objset_t *mos = dp->dp_meta_objset;
525 spa_t *spa = dp->dp_spa;
526 list_t synced_datasets;
527 dsl_sync_task_t *iter;
528 boolean_t wbc_skip_txg = B_FALSE;
529 boolean_t sync_ops = B_FALSE;
530 boolean_t user_snap = B_FALSE;
531 zfs_autosnap_t *autosnap = spa_get_autosnap(spa);
532 boolean_t autosnap_initialized = autosnap->initialized;
533 char snap[ZFS_MAX_DATASET_NAME_LEN];
534
535 /* check if there are ny sync ops in the txg */
536 if (txg_list_head(&dp->dp_sync_tasks, txg) != NULL)
537 sync_ops = B_TRUE;
538
539 /* check if there are user snaps in the txg */
540 for (iter = txg_list_head(&dp->dp_sync_tasks, txg);
541 iter != NULL;
542 iter = txg_list_next(&dp->dp_sync_tasks, iter, txg)) {
543 if (iter->dst_syncfunc == dsl_dataset_snapshot_sync) {
544 user_snap = B_TRUE;
545 break;
546 }
547 }
548
549
550 list_create(&synced_datasets, sizeof (dsl_dataset_t),
551 offsetof(dsl_dataset_t, ds_synced_link));
552
553 tx = dmu_tx_create_assigned(dp, txg);
554
555 (void) sprintf(snap, "%s%llu", AUTOSNAP_PREFIX,
556 (unsigned long long int) txg);
557
558 if (autosnap_initialized && spa->spa_sync_pass == 1) {
559 autosnap_zone_t *azone;
560
561 rrw_enter(&dp->dp_config_rwlock, RW_READER, FTAG);
562 mutex_enter(&autosnap->autosnap_lock);
563
564 /*
565 * WBC: the mechanism to ensure all WBC-ed dirty datasets
566 * are synchronously auto-snapshotted
567 * within (or by) the same TXG sync
568 * The "synchronicity" of the rightmost boundary of the WBC
569 * window is important to avoid used-space leakages
570 * on special vdev.
571 * Note that we skip here the WBC-ed datasets that are
572 * already fully migrated and don't have data on special
573 */
574
575 for (ds = txg_list_head(&dp->dp_dirty_datasets, txg);
576 ds != NULL;
577 ds = txg_list_next(&dp->dp_dirty_datasets, ds, txg)) {
578 char ds_name[ZFS_MAX_DATASET_NAME_LEN];
579 boolean_t wbc_azone;
580
581 dsl_dataset_name(ds, ds_name);
582
583 azone = autosnap_find_zone(autosnap, ds_name, B_TRUE);
584 if (azone == NULL)
585 continue;
586
587 if ((azone->flags & AUTOSNAP_CREATOR) == 0)
588 continue;
589
590 if (azone->created)
591 continue;
592
593 azone->delayed = B_TRUE;
594 azone->dirty = B_TRUE;
595 wbc_azone = (azone->flags & AUTOSNAP_WBC) != 0;
596
597 if (autosnap_confirm_snap(azone, txg)) {
598 if (!wbc_azone && !user_snap && !sync_ops) {
599 autosnap_create_snapshot(azone,
600 snap, dp, txg, tx);
601 }
602 } else if (wbc_azone) {
603 wbc_skip_txg = B_TRUE;
604 }
605 }
606
607 azone = list_head(&autosnap->autosnap_zones);
608 while (azone != NULL) {
609 boolean_t wbc_azone =
610 ((azone->flags & AUTOSNAP_WBC) != 0);
611
612 if (user_snap) {
613 azone->delayed = B_TRUE;
614 } else if (!azone->dirty && azone->delayed) {
615 if (autosnap_confirm_snap(azone, txg)) {
616 if (!wbc_azone && !user_snap &&
617 !sync_ops) {
618 autosnap_create_snapshot(azone,
619 snap, dp, txg, tx);
620 }
621 } else if (wbc_azone) {
622 wbc_skip_txg = B_TRUE;
623 }
624 }
625
626 azone = list_next(&autosnap->autosnap_zones, azone);
627 }
628
629 mutex_exit(&autosnap->autosnap_lock);
630 rrw_exit(&dp->dp_config_rwlock, FTAG);
631 }
632
633
634 /*
635 * Write out all dirty blocks of dirty datasets.
636 */
637 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
638 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
639
640 /*
641 * We must not sync any non-MOS datasets twice, because
642 * we may have taken a snapshot of them. However, we
643 * may sync newly-created datasets on pass 2.
644 */
645 ASSERT(!list_link_active(&ds->ds_synced_link));
646 list_insert_tail(&synced_datasets, ds);
647 dsl_dataset_sync(ds, zio, tx);
648 }
649
650 VERIFY0(zio_wait(zio));
651
652 if (autosnap_initialized && spa->spa_sync_pass == 1 &&
653 !user_snap) {
654 autosnap_zone_t *azone;
655
656 rrw_enter(&dp->dp_config_rwlock, RW_READER, FTAG);
657 mutex_enter(&autosnap->autosnap_lock);
658
659 /*
660 * At this stage we are walking over all delayed zones
661 * to create autosnaps
662 */
663
664 azone = list_head(&autosnap->autosnap_zones);
665 while (azone != NULL) {
666 boolean_t skip_zone =
667 ((azone->flags & AUTOSNAP_CREATOR) == 0);
668
669 if (azone->delayed && !skip_zone) {
670 boolean_t wbc_azone =
671 ((azone->flags & AUTOSNAP_WBC) != 0);
672
673 if ((!wbc_azone || !wbc_skip_txg) &&
674 autosnap_confirm_snap(azone, txg)) {
675 autosnap_create_snapshot(azone,
676 snap, dp, txg, tx);
677 }
678 }
679
680 if (skip_zone)
681 azone->delayed = B_FALSE;
682
683 azone = list_next(&autosnap->autosnap_zones, azone);
684 }
685
686 mutex_exit(&autosnap->autosnap_lock);
687 rrw_exit(&dp->dp_config_rwlock, FTAG);
688 }
689
690 /*
691 * We have written all of the accounted dirty data, so our
692 * dp_space_towrite should now be zero. However, some seldom-used
693 * code paths do not adhere to this (e.g. dbuf_undirty(), also
694 * rounding error in dbuf_write_physdone).
695 * Shore up the accounting of any dirtied space now.
696 */
697 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
698
699 /*
700 * Update the long range free counter after
701 * we're done syncing user data
702 */
703 mutex_enter(&dp->dp_lock);
704 ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
705 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
706 dp->dp_long_freeing_total -=
707 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
708 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
709 mutex_exit(&dp->dp_lock);
710
711 /*
712 * After the data blocks have been written (ensured by the zio_wait()
713 * above), update the user/group space accounting. This happens
714 * in tasks dispatched to dp_sync_taskq, so wait for them before
715 * continuing.
716 */
717 for (ds = list_head(&synced_datasets); ds != NULL;
718 ds = list_next(&synced_datasets, ds)) {
719 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
720 }
721 taskq_wait(dp->dp_sync_taskq);
722
723 /*
724 * Sync the datasets again to push out the changes due to
725 * userspace updates. This must be done before we process the
726 * sync tasks, so that any snapshots will have the correct
727 * user accounting information (and we won't get confused
728 * about which blocks are part of the snapshot).
729 */
730
731 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
732 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
733 ASSERT(list_link_active(&ds->ds_synced_link));
734 dmu_buf_rele(ds->ds_dbuf, ds);
735 dsl_dataset_sync(ds, zio, tx);
736 }
737 VERIFY0(zio_wait(zio));
738
739 /*
740 * Now that the datasets have been completely synced, we can
741 * clean up our in-memory structures accumulated while syncing:
742 *
743 * - move dead blocks from the pending deadlist to the on-disk deadlist
744 * - release hold from dsl_dataset_dirty()
745 */
746 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
747 dsl_dataset_sync_done(ds, tx);
748 }
749 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
750 dsl_dir_sync(dd, tx);
751 }
752
753 /*
754 * The MOS's space is accounted for in the pool/$MOS
755 * (dp_mos_dir). We can't modify the mos while we're syncing
756 * it, so we remember the deltas and apply them here.
757 */
758 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
759 dp->dp_mos_uncompressed_delta != 0) {
760 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
761 dp->dp_mos_used_delta,
762 dp->dp_mos_compressed_delta,
763 dp->dp_mos_uncompressed_delta, tx);
764 dp->dp_mos_used_delta = 0;
765 dp->dp_mos_compressed_delta = 0;
766 dp->dp_mos_uncompressed_delta = 0;
767 }
768
769 if (!multilist_is_empty(mos->os_dirty_dnodes[txg & TXG_MASK])) {
770 dsl_pool_sync_mos(dp, tx);
771 }
772
773 /*
774 * If we modify a dataset in the same txg that we want to destroy it,
775 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
776 * dsl_dir_destroy_check() will fail if there are unexpected holds.
777 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
778 * and clearing the hold on it) before we process the sync_tasks.
779 * The MOS data dirtied by the sync_tasks will be synced on the next
780 * pass.
781 */
782
783 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
784 dsl_sync_task_t *dst;
785 /*
786 * No more sync tasks should have been added while we
787 * were syncing.
788 */
789 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
790 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
791 dsl_sync_task_sync(dst, tx);
792 }
793
794 if (spa_feature_is_active(spa, SPA_FEATURE_WBC)) {
795 wbc_trigger_wbcthread(dp->dp_spa,
796 ((dp->dp_sync_history[0] + dp->dp_sync_history[1]) / 2));
797 }
798
799 dmu_tx_commit(tx);
800
801 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
802 }
803
804 void
805 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
806 {
807 zilog_t *zilog;
808
809 while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
810 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
811 /*
812 * We don't remove the zilog from the dp_dirty_zilogs
813 * list until after we've cleaned it. This ensures that
814 * callers of zilog_is_dirty() receive an accurate
815 * answer when they are racing with the spa sync thread.
816 */
817 zil_clean(zilog, txg);
818 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
819 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
820 dmu_buf_rele(ds->ds_dbuf, zilog);
821 }
822 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
823 }
824
825 /*
826 * TRUE if the current thread is the tx_sync_thread or if we
827 * are being called from SPA context during pool initialization.
828 */
829 int
830 dsl_pool_sync_context(dsl_pool_t *dp)
831 {
832 return (curthread == dp->dp_tx.tx_sync_thread ||
833 spa_is_initializing(dp->dp_spa) ||
834 taskq_member(dp->dp_sync_taskq, curthread));
835 }
836
837 uint64_t
838 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
839 {
840 uint64_t space, resv;
841
842 /*
843 * If we're trying to assess whether it's OK to do a free,
844 * cut the reservation in half to allow forward progress
845 * (e.g. make it possible to rm(1) files from a full pool).
846 */
847 space = spa_get_dspace(dp->dp_spa);
848 resv = spa_get_slop_space(dp->dp_spa);
849 if (netfree)
850 resv >>= 1;
851
852 return (space - resv);
853 }
854
855 boolean_t
856 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
857 {
858 uint64_t delay_min_bytes =
859 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
860 boolean_t rv;
861
862 if (dp->dp_dirty_total > zfs_dirty_data_sync)
863 txg_kick(dp);
864 rv = (dp->dp_dirty_total > delay_min_bytes);
865
866 return (rv);
867 }
868
869 void
870 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
871 {
872 if (space > 0) {
873 mutex_enter(&dp->dp_lock);
874 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
875 dsl_pool_dirty_delta(dp, space);
876 mutex_exit(&dp->dp_lock);
877 }
878 }
879
880 void
881 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
882 {
883 ASSERT3S(space, >=, 0);
884 if (space == 0)
885 return;
886 mutex_enter(&dp->dp_lock);
887 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
888 /* XXX writing something we didn't dirty? */
889 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
890 }
891 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
892 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
893 ASSERT3U(dp->dp_dirty_total, >=, space);
894 dsl_pool_dirty_delta(dp, -space);
895 mutex_exit(&dp->dp_lock);
896 }
897
898 /* ARGSUSED */
899 static int
900 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
901 {
902 dmu_tx_t *tx = arg;
903 dsl_dataset_t *ds, *prev = NULL;
904 int err;
905
906 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
907 if (err)
908 return (err);
909
910 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
911 err = dsl_dataset_hold_obj(dp,
912 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
913 if (err) {
914 dsl_dataset_rele(ds, FTAG);
915 return (err);
916 }
917
918 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
919 break;
920 dsl_dataset_rele(ds, FTAG);
921 ds = prev;
922 prev = NULL;
923 }
924
925 if (prev == NULL) {
926 prev = dp->dp_origin_snap;
927
928 /*
929 * The $ORIGIN can't have any data, or the accounting
930 * will be wrong.
931 */
932 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
933 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
934 rrw_exit(&ds->ds_bp_rwlock, FTAG);
935
936 /* The origin doesn't get attached to itself */
937 if (ds->ds_object == prev->ds_object) {
938 dsl_dataset_rele(ds, FTAG);
939 return (0);
940 }
941
942 dmu_buf_will_dirty(ds->ds_dbuf, tx);
943 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
944 dsl_dataset_phys(ds)->ds_prev_snap_txg =
945 dsl_dataset_phys(prev)->ds_creation_txg;
946
947 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
948 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
949
950 dmu_buf_will_dirty(prev->ds_dbuf, tx);
951 dsl_dataset_phys(prev)->ds_num_children++;
952
953 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
954 ASSERT(ds->ds_prev == NULL);
955 VERIFY0(dsl_dataset_hold_obj(dp,
956 dsl_dataset_phys(ds)->ds_prev_snap_obj,
957 ds, &ds->ds_prev));
958 }
959 }
960
961 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
962 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
963
964 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
965 dmu_buf_will_dirty(prev->ds_dbuf, tx);
966 dsl_dataset_phys(prev)->ds_next_clones_obj =
967 zap_create(dp->dp_meta_objset,
968 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
969 }
970 VERIFY0(zap_add_int(dp->dp_meta_objset,
971 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
972
973 dsl_dataset_rele(ds, FTAG);
974 if (prev != dp->dp_origin_snap)
975 dsl_dataset_rele(prev, FTAG);
976 return (0);
977 }
978
979 void
980 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
981 {
982 ASSERT(dmu_tx_is_syncing(tx));
983 ASSERT(dp->dp_origin_snap != NULL);
984
985 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
986 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
987 }
988
989 /* ARGSUSED */
990 static int
991 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
992 {
993 dmu_tx_t *tx = arg;
994 objset_t *mos = dp->dp_meta_objset;
995
996 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
997 dsl_dataset_t *origin;
998
999 VERIFY0(dsl_dataset_hold_obj(dp,
1000 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
1001
1002 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
1003 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
1004 dsl_dir_phys(origin->ds_dir)->dd_clones =
1005 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
1006 0, tx);
1007 }
1008
1009 VERIFY0(zap_add_int(dp->dp_meta_objset,
1010 dsl_dir_phys(origin->ds_dir)->dd_clones,
1011 ds->ds_object, tx));
1012
1013 dsl_dataset_rele(origin, FTAG);
1014 }
1015 return (0);
1016 }
1017
1018 void
1019 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
1020 {
1021 ASSERT(dmu_tx_is_syncing(tx));
1022 uint64_t obj;
1023
1024 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
1025 VERIFY0(dsl_pool_open_special_dir(dp,
1026 FREE_DIR_NAME, &dp->dp_free_dir));
1027
1028 /*
1029 * We can't use bpobj_alloc(), because spa_version() still
1030 * returns the old version, and we need a new-version bpobj with
1031 * subobj support. So call dmu_object_alloc() directly.
1032 */
1033 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
1034 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
1035 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1036 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
1037 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
1038
1039 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1040 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
1041 }
1042
1043 void
1044 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
1045 {
1046 uint64_t dsobj;
1047 dsl_dataset_t *ds;
1048
1049 ASSERT(dmu_tx_is_syncing(tx));
1050 ASSERT(dp->dp_origin_snap == NULL);
1051 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
1052
1053 /* create the origin dir, ds, & snap-ds */
1054 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
1055 NULL, 0, kcred, tx);
1056 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1057 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
1058 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
1059 dp, &dp->dp_origin_snap));
1060 dsl_dataset_rele(ds, FTAG);
1061 }
1062
1063 taskq_t *
1064 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
1065 {
1066 return (dp->dp_vnrele_taskq);
1067 }
1068
1069 /*
1070 * Walk through the pool-wide zap object of temporary snapshot user holds
1071 * and release them.
1072 */
1073 void
1074 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
1075 {
1076 zap_attribute_t za;
1077 zap_cursor_t zc;
1078 objset_t *mos = dp->dp_meta_objset;
1079 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1080 nvlist_t *holds;
1081
1082 if (zapobj == 0)
1083 return;
1084 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1085
1086 holds = fnvlist_alloc();
1087
1088 for (zap_cursor_init(&zc, mos, zapobj);
1089 zap_cursor_retrieve(&zc, &za) == 0;
1090 zap_cursor_advance(&zc)) {
1091 char *htag;
1092 nvlist_t *tags;
1093
1094 htag = strchr(za.za_name, '-');
1095 *htag = '\0';
1096 ++htag;
1097 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
1098 tags = fnvlist_alloc();
1099 fnvlist_add_boolean(tags, htag);
1100 fnvlist_add_nvlist(holds, za.za_name, tags);
1101 fnvlist_free(tags);
1102 } else {
1103 fnvlist_add_boolean(tags, htag);
1104 }
1105 }
1106 dsl_dataset_user_release_tmp(dp, holds);
1107 fnvlist_free(holds);
1108 zap_cursor_fini(&zc);
1109 }
1110
1111 /*
1112 * Create the pool-wide zap object for storing temporary snapshot holds.
1113 */
1114 void
1115 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
1116 {
1117 objset_t *mos = dp->dp_meta_objset;
1118
1119 ASSERT(dp->dp_tmp_userrefs_obj == 0);
1120 ASSERT(dmu_tx_is_syncing(tx));
1121
1122 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
1123 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
1124 }
1125
1126 static int
1127 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1128 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1129 {
1130 objset_t *mos = dp->dp_meta_objset;
1131 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1132 char *name;
1133 int error;
1134
1135 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1136 ASSERT(dmu_tx_is_syncing(tx));
1137
1138 /*
1139 * If the pool was created prior to SPA_VERSION_USERREFS, the
1140 * zap object for temporary holds might not exist yet.
1141 */
1142 if (zapobj == 0) {
1143 if (holding) {
1144 dsl_pool_user_hold_create_obj(dp, tx);
1145 zapobj = dp->dp_tmp_userrefs_obj;
1146 } else {
1147 return (SET_ERROR(ENOENT));
1148 }
1149 }
1150
1151 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1152 if (holding)
1153 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1154 else
1155 error = zap_remove(mos, zapobj, name, tx);
1156 strfree(name);
1157
1158 return (error);
1159 }
1160
1161 /*
1162 * Add a temporary hold for the given dataset object and tag.
1163 */
1164 int
1165 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1166 uint64_t now, dmu_tx_t *tx)
1167 {
1168 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1169 }
1170
1171 /*
1172 * Release a temporary hold for the given dataset object and tag.
1173 */
1174 int
1175 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1176 dmu_tx_t *tx)
1177 {
1178 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
1179 tx, B_FALSE));
1180 }
1181
1182 /*
1183 * DSL Pool Configuration Lock
1184 *
1185 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1186 * creation / destruction / rename / property setting). It must be held for
1187 * read to hold a dataset or dsl_dir. I.e. you must call
1188 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1189 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1190 * must be held continuously until all datasets and dsl_dirs are released.
1191 *
1192 * The only exception to this rule is that if a "long hold" is placed on
1193 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1194 * is still held. The long hold will prevent the dataset from being
1195 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1196 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1197 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1198 *
1199 * Legitimate long-holders (including owners) should be long-running, cancelable
1200 * tasks that should cause "zfs destroy" to fail. This includes DMU
1201 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1202 * "zfs send", and "zfs diff". There are several other long-holders whose
1203 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1204 *
1205 * The usual formula for long-holding would be:
1206 * dsl_pool_hold()
1207 * dsl_dataset_hold()
1208 * ... perform checks ...
1209 * dsl_dataset_long_hold()
1210 * dsl_pool_rele()
1211 * ... perform long-running task ...
1212 * dsl_dataset_long_rele()
1213 * dsl_dataset_rele()
1214 *
1215 * Note that when the long hold is released, the dataset is still held but
1216 * the pool is not held. The dataset may change arbitrarily during this time
1217 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1218 * dataset except release it.
1219 *
1220 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1221 * or modifying operations.
1222 *
1223 * Modifying operations should generally use dsl_sync_task(). The synctask
1224 * infrastructure enforces proper locking strategy with respect to the
1225 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1226 *
1227 * Read-only operations will manually hold the pool, then the dataset, obtain
1228 * information from the dataset, then release the pool and dataset.
1229 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1230 * hold/rele.
1231 */
1232
1233 int
1234 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1235 {
1236 spa_t *spa;
1237 int error;
1238
1239 error = spa_open(name, &spa, tag);
1240 if (error == 0) {
1241 *dp = spa_get_dsl(spa);
1242 dsl_pool_config_enter(*dp, tag);
1243 }
1244 return (error);
1245 }
1246
1247 void
1248 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1249 {
1250 dsl_pool_config_exit(dp, tag);
1251 spa_close(dp->dp_spa, tag);
1252 }
1253
1254 void
1255 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1256 {
1257 /*
1258 * We use a "reentrant" reader-writer lock, but not reentrantly.
1259 *
1260 * The rrwlock can (with the track_all flag) track all reading threads,
1261 * which is very useful for debugging which code path failed to release
1262 * the lock, and for verifying that the *current* thread does hold
1263 * the lock.
1264 *
1265 * (Unlike a rwlock, which knows that N threads hold it for
1266 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1267 * if any thread holds it for read, even if this thread doesn't).
1268 */
1269 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1270 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1271 }
1272
1273 void
1274 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1275 {
1276 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1277 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1278 }
1279
1280 void
1281 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1282 {
1283 rrw_exit(&dp->dp_config_rwlock, tag);
1284 }
1285
1286 boolean_t
1287 dsl_pool_config_held(dsl_pool_t *dp)
1288 {
1289 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1290 }
1291
1292 boolean_t
1293 dsl_pool_config_held_writer(dsl_pool_t *dp)
1294 {
1295 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1296 }