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, 2014 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 */
27
28 #include <sys/dsl_pool.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_prop.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/dsl_scan.h>
34 #include <sys/dnode.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/arc.h>
38 #include <sys/zap.h>
39 #include <sys/zio.h>
40 #include <sys/zfs_context.h>
41 #include <sys/fs/zfs.h>
42 #include <sys/zfs_znode.h>
43 #include <sys/spa_impl.h>
44 #include <sys/dsl_deadlist.h>
45 #include <sys/bptree.h>
46 #include <sys/zfeature.h>
47 #include <sys/zil_impl.h>
48 #include <sys/dsl_userhold.h>
49
50 /*
51 * ZFS Write Throttle
52 * ------------------
53 *
54 * ZFS must limit the rate of incoming writes to the rate at which it is able
55 * to sync data modifications to the backend storage. Throttling by too much
56 * creates an artificial limit; throttling by too little can only be sustained
57 * for short periods and would lead to highly lumpy performance. On a per-pool
58 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
59 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
60 * of dirty data decreases. When the amount of dirty data exceeds a
61 * predetermined threshold further modifications are blocked until the amount
62 * of dirty data decreases (as data is synced out).
63 *
64 * The limit on dirty data is tunable, and should be adjusted according to
65 * both the IO capacity and available memory of the system. The larger the
66 * window, the more ZFS is able to aggregate and amortize metadata (and data)
67 * changes. However, memory is a limited resource, and allowing for more dirty
68 * data comes at the cost of keeping other useful data in memory (for example
69 * ZFS data cached by the ARC).
70 *
71 * Implementation
72 *
73 * As buffers are modified dsl_pool_willuse_space() increments both the per-
74 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
75 * dirty space used; dsl_pool_dirty_space() decrements those values as data
76 * is synced out from dsl_pool_sync(). While only the poolwide value is
77 * relevant, the per-txg value is useful for debugging. The tunable
78 * zfs_dirty_data_max determines the dirty space limit. Once that value is
79 * exceeded, new writes are halted until space frees up.
80 *
81 * The zfs_dirty_data_sync tunable dictates the threshold at which we
82 * ensure that there is a txg syncing (see the comment in txg.c for a full
83 * description of transaction group stages).
84 *
85 * The IO scheduler uses both the dirty space limit and current amount of
86 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
87 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
88 *
89 * The delay is also calculated based on the amount of dirty data. See the
90 * comment above dmu_tx_delay() for details.
91 */
92
93 /*
94 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
95 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
96 */
97 uint64_t zfs_dirty_data_max;
98 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
99 int zfs_dirty_data_max_percent = 10;
100
101 /*
102 * If there is at least this much dirty data, push out a txg.
103 */
104 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
105
106 /*
107 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
108 * and delay each transaction.
109 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
110 */
111 int zfs_delay_min_dirty_percent = 60;
112
113 /*
114 * This controls how quickly the delay approaches infinity.
115 * Larger values cause it to delay more for a given amount of dirty data.
116 * Therefore larger values will cause there to be less dirty data for a
117 * given throughput.
118 *
119 * For the smoothest delay, this value should be about 1 billion divided
120 * by the maximum number of operations per second. This will smoothly
121 * handle between 10x and 1/10th this number.
122 *
123 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
124 * multiply in dmu_tx_delay().
125 */
126 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
127
128
129 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
130 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
131
132 int
133 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
134 {
135 uint64_t obj;
136 int err;
137
138 err = zap_lookup(dp->dp_meta_objset,
139 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
140 name, sizeof (obj), 1, &obj);
141 if (err)
142 return (err);
143
144 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
145 }
146
147 static dsl_pool_t *
148 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
149 {
150 dsl_pool_t *dp;
151 blkptr_t *bp = spa_get_rootblkptr(spa);
152
153 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
154 dp->dp_spa = spa;
155 dp->dp_meta_rootbp = *bp;
156 rrw_init(&dp->dp_config_rwlock, B_TRUE);
157 txg_init(dp, txg);
158
159 txg_list_create(&dp->dp_dirty_datasets,
160 offsetof(dsl_dataset_t, ds_dirty_link));
161 txg_list_create(&dp->dp_dirty_zilogs,
162 offsetof(zilog_t, zl_dirty_link));
163 txg_list_create(&dp->dp_dirty_dirs,
164 offsetof(dsl_dir_t, dd_dirty_link));
165 txg_list_create(&dp->dp_sync_tasks,
166 offsetof(dsl_sync_task_t, dst_node));
167
168 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
169 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
170
171 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
172 1, 4, 0);
173
174 return (dp);
175 }
176
177 int
178 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
179 {
180 int err;
181 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
182
183 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
184 &dp->dp_meta_objset);
185 if (err != 0)
186 dsl_pool_close(dp);
187 else
188 *dpp = dp;
189
190 return (err);
191 }
192
193 int
194 dsl_pool_open(dsl_pool_t *dp)
195 {
196 int err;
197 dsl_dir_t *dd;
198 dsl_dataset_t *ds;
199 uint64_t obj;
200
201 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
202 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
203 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
204 &dp->dp_root_dir_obj);
205 if (err)
206 goto out;
207
208 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
209 NULL, dp, &dp->dp_root_dir);
210 if (err)
211 goto out;
212
213 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
214 if (err)
215 goto out;
216
217 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
218 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
219 if (err)
220 goto out;
221 err = dsl_dataset_hold_obj(dp,
222 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
223 if (err == 0) {
224 err = dsl_dataset_hold_obj(dp,
225 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
226 &dp->dp_origin_snap);
227 dsl_dataset_rele(ds, FTAG);
228 }
229 dsl_dir_rele(dd, dp);
230 if (err)
231 goto out;
232 }
233
234 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
235 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
236 &dp->dp_free_dir);
237 if (err)
238 goto out;
239
240 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
241 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
242 if (err)
243 goto out;
244 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
245 dp->dp_meta_objset, obj));
246 }
247
248 /*
249 * Note: errors ignored, because the leak dir will not exist if we
250 * have not encountered a leak yet.
251 */
252 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
253 &dp->dp_leak_dir);
254
255 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
256 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
257 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
258 &dp->dp_bptree_obj);
259 if (err != 0)
260 goto out;
261 }
262
263 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
264 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
265 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
266 &dp->dp_empty_bpobj);
267 if (err != 0)
268 goto out;
269 }
270
271 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
272 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
273 &dp->dp_tmp_userrefs_obj);
274 if (err == ENOENT)
275 err = 0;
276 if (err)
277 goto out;
278
279 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
280
281 out:
282 rrw_exit(&dp->dp_config_rwlock, FTAG);
283 return (err);
284 }
285
286 void
287 dsl_pool_close(dsl_pool_t *dp)
288 {
289 /*
290 * Drop our references from dsl_pool_open().
291 *
292 * Since we held the origin_snap from "syncing" context (which
293 * includes pool-opening context), it actually only got a "ref"
294 * and not a hold, so just drop that here.
295 */
296 if (dp->dp_origin_snap)
297 dsl_dataset_rele(dp->dp_origin_snap, dp);
298 if (dp->dp_mos_dir)
299 dsl_dir_rele(dp->dp_mos_dir, dp);
300 if (dp->dp_free_dir)
301 dsl_dir_rele(dp->dp_free_dir, dp);
302 if (dp->dp_leak_dir)
303 dsl_dir_rele(dp->dp_leak_dir, dp);
304 if (dp->dp_root_dir)
305 dsl_dir_rele(dp->dp_root_dir, dp);
306
307 bpobj_close(&dp->dp_free_bpobj);
308
309 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
310 if (dp->dp_meta_objset)
311 dmu_objset_evict(dp->dp_meta_objset);
312
313 txg_list_destroy(&dp->dp_dirty_datasets);
314 txg_list_destroy(&dp->dp_dirty_zilogs);
315 txg_list_destroy(&dp->dp_sync_tasks);
316 txg_list_destroy(&dp->dp_dirty_dirs);
317
318 arc_flush(dp->dp_spa);
319 txg_fini(dp);
320 dsl_scan_fini(dp);
321 dmu_buf_user_evict_wait();
322
323 rrw_destroy(&dp->dp_config_rwlock);
324 mutex_destroy(&dp->dp_lock);
325 taskq_destroy(dp->dp_vnrele_taskq);
326 if (dp->dp_blkstats)
327 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
328 kmem_free(dp, sizeof (dsl_pool_t));
329 }
330
331 dsl_pool_t *
332 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
333 {
334 int err;
335 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
336 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
337 objset_t *os;
338 dsl_dataset_t *ds;
339 uint64_t obj;
340
341 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
342
343 /* create and open the MOS (meta-objset) */
344 dp->dp_meta_objset = dmu_objset_create_impl(spa,
345 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
346
347 /* create the pool directory */
348 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
349 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
350 ASSERT0(err);
351
352 /* Initialize scan structures */
353 VERIFY0(dsl_scan_init(dp, txg));
354
355 /* create and open the root dir */
356 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
357 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
358 NULL, dp, &dp->dp_root_dir));
359
360 /* create and open the meta-objset dir */
361 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
362 VERIFY0(dsl_pool_open_special_dir(dp,
363 MOS_DIR_NAME, &dp->dp_mos_dir));
364
365 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
366 /* create and open the free dir */
367 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
368 FREE_DIR_NAME, tx);
369 VERIFY0(dsl_pool_open_special_dir(dp,
370 FREE_DIR_NAME, &dp->dp_free_dir));
371
372 /* create and open the free_bplist */
373 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
374 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
375 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
376 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
377 dp->dp_meta_objset, obj));
378 }
379
380 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
381 dsl_pool_create_origin(dp, tx);
382
383 /* create the root dataset */
384 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
385
386 /* create the root objset */
387 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
388 os = dmu_objset_create_impl(dp->dp_spa, ds,
389 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
390 #ifdef _KERNEL
391 zfs_create_fs(os, kcred, zplprops, tx);
392 #endif
393 dsl_dataset_rele(ds, FTAG);
394
395 dmu_tx_commit(tx);
396
397 rrw_exit(&dp->dp_config_rwlock, FTAG);
398
399 return (dp);
400 }
401
402 /*
403 * Account for the meta-objset space in its placeholder dsl_dir.
404 */
405 void
406 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
407 int64_t used, int64_t comp, int64_t uncomp)
408 {
409 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
410 mutex_enter(&dp->dp_lock);
411 dp->dp_mos_used_delta += used;
412 dp->dp_mos_compressed_delta += comp;
413 dp->dp_mos_uncompressed_delta += uncomp;
414 mutex_exit(&dp->dp_lock);
415 }
416
417 static int
418 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
419 {
420 dsl_deadlist_t *dl = arg;
421 dsl_deadlist_insert(dl, bp, tx);
422 return (0);
423 }
424
425 static void
426 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
427 {
428 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
429 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
430 VERIFY0(zio_wait(zio));
431 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
432 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
433 }
434
435 static void
436 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
437 {
438 ASSERT(MUTEX_HELD(&dp->dp_lock));
439
440 if (delta < 0)
441 ASSERT3U(-delta, <=, dp->dp_dirty_total);
442
443 dp->dp_dirty_total += delta;
444
445 /*
446 * Note: we signal even when increasing dp_dirty_total.
447 * This ensures forward progress -- each thread wakes the next waiter.
448 */
449 if (dp->dp_dirty_total <= zfs_dirty_data_max)
450 cv_signal(&dp->dp_spaceavail_cv);
451 }
452
453 void
454 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
455 {
456 zio_t *zio;
457 dmu_tx_t *tx;
458 dsl_dir_t *dd;
459 dsl_dataset_t *ds;
460 objset_t *mos = dp->dp_meta_objset;
461 list_t synced_datasets;
462
463 list_create(&synced_datasets, sizeof (dsl_dataset_t),
464 offsetof(dsl_dataset_t, ds_synced_link));
465
466 tx = dmu_tx_create_assigned(dp, txg);
467
468 /*
469 * Write out all dirty blocks of dirty datasets.
470 */
471 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
472 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
473 /*
474 * We must not sync any non-MOS datasets twice, because
475 * we may have taken a snapshot of them. However, we
476 * may sync newly-created datasets on pass 2.
477 */
478 ASSERT(!list_link_active(&ds->ds_synced_link));
479 list_insert_tail(&synced_datasets, ds);
480 dsl_dataset_sync(ds, zio, tx);
481 }
482 VERIFY0(zio_wait(zio));
483
484 /*
485 * We have written all of the accounted dirty data, so our
486 * dp_space_towrite should now be zero. However, some seldom-used
487 * code paths do not adhere to this (e.g. dbuf_undirty(), also
488 * rounding error in dbuf_write_physdone).
489 * Shore up the accounting of any dirtied space now.
490 */
491 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
492
493 /*
494 * After the data blocks have been written (ensured by the zio_wait()
495 * above), update the user/group space accounting.
496 */
497 for (ds = list_head(&synced_datasets); ds != NULL;
498 ds = list_next(&synced_datasets, ds)) {
499 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
500 }
501
502 /*
503 * Sync the datasets again to push out the changes due to
504 * userspace updates. This must be done before we process the
505 * sync tasks, so that any snapshots will have the correct
506 * user accounting information (and we won't get confused
507 * about which blocks are part of the snapshot).
508 */
509 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
510 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
511 ASSERT(list_link_active(&ds->ds_synced_link));
512 dmu_buf_rele(ds->ds_dbuf, ds);
513 dsl_dataset_sync(ds, zio, tx);
514 }
515 VERIFY0(zio_wait(zio));
516
517 /*
518 * Now that the datasets have been completely synced, we can
519 * clean up our in-memory structures accumulated while syncing:
520 *
521 * - move dead blocks from the pending deadlist to the on-disk deadlist
522 * - release hold from dsl_dataset_dirty()
523 */
524 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
525 objset_t *os = ds->ds_objset;
526 bplist_iterate(&ds->ds_pending_deadlist,
527 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
528 ASSERT(!dmu_objset_is_dirty(os, txg));
529 dmu_buf_rele(ds->ds_dbuf, ds);
530 }
531 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
532 dsl_dir_sync(dd, tx);
533 }
534
535 /*
536 * The MOS's space is accounted for in the pool/$MOS
537 * (dp_mos_dir). We can't modify the mos while we're syncing
538 * it, so we remember the deltas and apply them here.
539 */
540 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
541 dp->dp_mos_uncompressed_delta != 0) {
542 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
543 dp->dp_mos_used_delta,
544 dp->dp_mos_compressed_delta,
545 dp->dp_mos_uncompressed_delta, tx);
546 dp->dp_mos_used_delta = 0;
547 dp->dp_mos_compressed_delta = 0;
548 dp->dp_mos_uncompressed_delta = 0;
549 }
550
551 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
552 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
553 dsl_pool_sync_mos(dp, tx);
554 }
555
556 /*
557 * If we modify a dataset in the same txg that we want to destroy it,
558 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
559 * dsl_dir_destroy_check() will fail if there are unexpected holds.
560 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
561 * and clearing the hold on it) before we process the sync_tasks.
562 * The MOS data dirtied by the sync_tasks will be synced on the next
563 * pass.
564 */
565 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
566 dsl_sync_task_t *dst;
567 /*
568 * No more sync tasks should have been added while we
569 * were syncing.
570 */
571 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
572 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
573 dsl_sync_task_sync(dst, tx);
574 }
575
576 dmu_tx_commit(tx);
577
578 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
579 }
580
581 void
582 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
583 {
584 zilog_t *zilog;
585
586 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
587 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
588 zil_clean(zilog, txg);
589 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
590 dmu_buf_rele(ds->ds_dbuf, zilog);
591 }
592 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
593 }
594
595 /*
596 * TRUE if the current thread is the tx_sync_thread or if we
597 * are being called from SPA context during pool initialization.
598 */
599 int
600 dsl_pool_sync_context(dsl_pool_t *dp)
601 {
602 return (curthread == dp->dp_tx.tx_sync_thread ||
603 spa_is_initializing(dp->dp_spa));
604 }
605
606 uint64_t
607 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
608 {
609 uint64_t space, resv;
610
611 /*
612 * If we're trying to assess whether it's OK to do a free,
613 * cut the reservation in half to allow forward progress
614 * (e.g. make it possible to rm(1) files from a full pool).
615 */
616 space = spa_get_dspace(dp->dp_spa);
617 resv = spa_get_slop_space(dp->dp_spa);
618 if (netfree)
619 resv >>= 1;
620
621 return (space - resv);
622 }
623
624 boolean_t
625 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
626 {
627 uint64_t delay_min_bytes =
628 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
629 boolean_t rv;
630
631 mutex_enter(&dp->dp_lock);
632 if (dp->dp_dirty_total > zfs_dirty_data_sync)
633 txg_kick(dp);
634 rv = (dp->dp_dirty_total > delay_min_bytes);
635 mutex_exit(&dp->dp_lock);
636 return (rv);
637 }
638
639 void
640 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
641 {
642 if (space > 0) {
643 mutex_enter(&dp->dp_lock);
644 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
645 dsl_pool_dirty_delta(dp, space);
646 mutex_exit(&dp->dp_lock);
647 }
648 }
649
650 void
651 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
652 {
653 ASSERT3S(space, >=, 0);
654 if (space == 0)
655 return;
656 mutex_enter(&dp->dp_lock);
657 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
658 /* XXX writing something we didn't dirty? */
659 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
660 }
661 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
662 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
663 ASSERT3U(dp->dp_dirty_total, >=, space);
664 dsl_pool_dirty_delta(dp, -space);
665 mutex_exit(&dp->dp_lock);
666 }
667
668 /* ARGSUSED */
669 static int
670 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
671 {
672 dmu_tx_t *tx = arg;
673 dsl_dataset_t *ds, *prev = NULL;
674 int err;
675
676 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
677 if (err)
678 return (err);
679
680 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
681 err = dsl_dataset_hold_obj(dp,
682 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
683 if (err) {
684 dsl_dataset_rele(ds, FTAG);
685 return (err);
686 }
687
688 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
689 break;
690 dsl_dataset_rele(ds, FTAG);
691 ds = prev;
692 prev = NULL;
693 }
694
695 if (prev == NULL) {
696 prev = dp->dp_origin_snap;
697
698 /*
699 * The $ORIGIN can't have any data, or the accounting
700 * will be wrong.
701 */
702 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
703
704 /* The origin doesn't get attached to itself */
705 if (ds->ds_object == prev->ds_object) {
706 dsl_dataset_rele(ds, FTAG);
707 return (0);
708 }
709
710 dmu_buf_will_dirty(ds->ds_dbuf, tx);
711 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
712 dsl_dataset_phys(ds)->ds_prev_snap_txg =
713 dsl_dataset_phys(prev)->ds_creation_txg;
714
715 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
716 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
717
718 dmu_buf_will_dirty(prev->ds_dbuf, tx);
719 dsl_dataset_phys(prev)->ds_num_children++;
720
721 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
722 ASSERT(ds->ds_prev == NULL);
723 VERIFY0(dsl_dataset_hold_obj(dp,
724 dsl_dataset_phys(ds)->ds_prev_snap_obj,
725 ds, &ds->ds_prev));
726 }
727 }
728
729 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
730 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
731
732 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
733 dmu_buf_will_dirty(prev->ds_dbuf, tx);
734 dsl_dataset_phys(prev)->ds_next_clones_obj =
735 zap_create(dp->dp_meta_objset,
736 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
737 }
738 VERIFY0(zap_add_int(dp->dp_meta_objset,
739 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
740
741 dsl_dataset_rele(ds, FTAG);
742 if (prev != dp->dp_origin_snap)
743 dsl_dataset_rele(prev, FTAG);
744 return (0);
745 }
746
747 void
748 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
749 {
750 ASSERT(dmu_tx_is_syncing(tx));
751 ASSERT(dp->dp_origin_snap != NULL);
752
753 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
754 tx, DS_FIND_CHILDREN));
755 }
756
757 /* ARGSUSED */
758 static int
759 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
760 {
761 dmu_tx_t *tx = arg;
762 objset_t *mos = dp->dp_meta_objset;
763
764 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
765 dsl_dataset_t *origin;
766
767 VERIFY0(dsl_dataset_hold_obj(dp,
768 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
769
770 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
771 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
772 dsl_dir_phys(origin->ds_dir)->dd_clones =
773 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
774 0, tx);
775 }
776
777 VERIFY0(zap_add_int(dp->dp_meta_objset,
778 dsl_dir_phys(origin->ds_dir)->dd_clones,
779 ds->ds_object, tx));
780
781 dsl_dataset_rele(origin, FTAG);
782 }
783 return (0);
784 }
785
786 void
787 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
788 {
789 ASSERT(dmu_tx_is_syncing(tx));
790 uint64_t obj;
791
792 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
793 VERIFY0(dsl_pool_open_special_dir(dp,
794 FREE_DIR_NAME, &dp->dp_free_dir));
795
796 /*
797 * We can't use bpobj_alloc(), because spa_version() still
798 * returns the old version, and we need a new-version bpobj with
799 * subobj support. So call dmu_object_alloc() directly.
800 */
801 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
802 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
803 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
804 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
805 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
806
807 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
808 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
809 }
810
811 void
812 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
813 {
814 uint64_t dsobj;
815 dsl_dataset_t *ds;
816
817 ASSERT(dmu_tx_is_syncing(tx));
818 ASSERT(dp->dp_origin_snap == NULL);
819 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
820
821 /* create the origin dir, ds, & snap-ds */
822 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
823 NULL, 0, kcred, tx);
824 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
825 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
826 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
827 dp, &dp->dp_origin_snap));
828 dsl_dataset_rele(ds, FTAG);
829 }
830
831 taskq_t *
832 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
833 {
834 return (dp->dp_vnrele_taskq);
835 }
836
837 /*
838 * Walk through the pool-wide zap object of temporary snapshot user holds
839 * and release them.
840 */
841 void
842 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
843 {
844 zap_attribute_t za;
845 zap_cursor_t zc;
846 objset_t *mos = dp->dp_meta_objset;
847 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
848 nvlist_t *holds;
849
850 if (zapobj == 0)
851 return;
852 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
853
854 holds = fnvlist_alloc();
855
856 for (zap_cursor_init(&zc, mos, zapobj);
857 zap_cursor_retrieve(&zc, &za) == 0;
858 zap_cursor_advance(&zc)) {
859 char *htag;
860 nvlist_t *tags;
861
862 htag = strchr(za.za_name, '-');
863 *htag = '\0';
864 ++htag;
865 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
866 tags = fnvlist_alloc();
867 fnvlist_add_boolean(tags, htag);
868 fnvlist_add_nvlist(holds, za.za_name, tags);
869 fnvlist_free(tags);
870 } else {
871 fnvlist_add_boolean(tags, htag);
872 }
873 }
874 dsl_dataset_user_release_tmp(dp, holds);
875 fnvlist_free(holds);
876 zap_cursor_fini(&zc);
877 }
878
879 /*
880 * Create the pool-wide zap object for storing temporary snapshot holds.
881 */
882 void
883 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
884 {
885 objset_t *mos = dp->dp_meta_objset;
886
887 ASSERT(dp->dp_tmp_userrefs_obj == 0);
888 ASSERT(dmu_tx_is_syncing(tx));
889
890 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
891 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
892 }
893
894 static int
895 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
896 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
897 {
898 objset_t *mos = dp->dp_meta_objset;
899 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
900 char *name;
901 int error;
902
903 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
904 ASSERT(dmu_tx_is_syncing(tx));
905
906 /*
907 * If the pool was created prior to SPA_VERSION_USERREFS, the
908 * zap object for temporary holds might not exist yet.
909 */
910 if (zapobj == 0) {
911 if (holding) {
912 dsl_pool_user_hold_create_obj(dp, tx);
913 zapobj = dp->dp_tmp_userrefs_obj;
914 } else {
915 return (SET_ERROR(ENOENT));
916 }
917 }
918
919 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
920 if (holding)
921 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
922 else
923 error = zap_remove(mos, zapobj, name, tx);
924 strfree(name);
925
926 return (error);
927 }
928
929 /*
930 * Add a temporary hold for the given dataset object and tag.
931 */
932 int
933 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
934 uint64_t now, dmu_tx_t *tx)
935 {
936 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
937 }
938
939 /*
940 * Release a temporary hold for the given dataset object and tag.
941 */
942 int
943 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
944 dmu_tx_t *tx)
945 {
946 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
947 tx, B_FALSE));
948 }
949
950 /*
951 * DSL Pool Configuration Lock
952 *
953 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
954 * creation / destruction / rename / property setting). It must be held for
955 * read to hold a dataset or dsl_dir. I.e. you must call
956 * dsl_pool_config_enter() or dsl_pool_hold() before calling
957 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
958 * must be held continuously until all datasets and dsl_dirs are released.
959 *
960 * The only exception to this rule is that if a "long hold" is placed on
961 * a dataset, then the dp_config_rwlock may be dropped while the dataset
962 * is still held. The long hold will prevent the dataset from being
963 * destroyed -- the destroy will fail with EBUSY. A long hold can be
964 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
965 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
966 *
967 * Legitimate long-holders (including owners) should be long-running, cancelable
968 * tasks that should cause "zfs destroy" to fail. This includes DMU
969 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
970 * "zfs send", and "zfs diff". There are several other long-holders whose
971 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
972 *
973 * The usual formula for long-holding would be:
974 * dsl_pool_hold()
975 * dsl_dataset_hold()
976 * ... perform checks ...
977 * dsl_dataset_long_hold()
978 * dsl_pool_rele()
979 * ... perform long-running task ...
980 * dsl_dataset_long_rele()
981 * dsl_dataset_rele()
982 *
983 * Note that when the long hold is released, the dataset is still held but
984 * the pool is not held. The dataset may change arbitrarily during this time
985 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
986 * dataset except release it.
987 *
988 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
989 * or modifying operations.
990 *
991 * Modifying operations should generally use dsl_sync_task(). The synctask
992 * infrastructure enforces proper locking strategy with respect to the
993 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
994 *
995 * Read-only operations will manually hold the pool, then the dataset, obtain
996 * information from the dataset, then release the pool and dataset.
997 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
998 * hold/rele.
999 */
1000
1001 int
1002 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1003 {
1004 spa_t *spa;
1005 int error;
1006
1007 error = spa_open(name, &spa, tag);
1008 if (error == 0) {
1009 *dp = spa_get_dsl(spa);
1010 dsl_pool_config_enter(*dp, tag);
1011 }
1012 return (error);
1013 }
1014
1015 void
1016 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1017 {
1018 dsl_pool_config_exit(dp, tag);
1019 spa_close(dp->dp_spa, tag);
1020 }
1021
1022 void
1023 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1024 {
1025 /*
1026 * We use a "reentrant" reader-writer lock, but not reentrantly.
1027 *
1028 * The rrwlock can (with the track_all flag) track all reading threads,
1029 * which is very useful for debugging which code path failed to release
1030 * the lock, and for verifying that the *current* thread does hold
1031 * the lock.
1032 *
1033 * (Unlike a rwlock, which knows that N threads hold it for
1034 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1035 * if any thread holds it for read, even if this thread doesn't).
1036 */
1037 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1038 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1039 }
1040
1041 void
1042 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1043 {
1044 rrw_exit(&dp->dp_config_rwlock, tag);
1045 }
1046
1047 boolean_t
1048 dsl_pool_config_held(dsl_pool_t *dp)
1049 {
1050 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1051 }