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 */
26
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/zap.h>
38 #include <sys/zio.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
48
49 /*
50 * ZFS Write Throttle
51 * ------------------
52 *
53 * ZFS must limit the rate of incoming writes to the rate at which it is able
54 * to sync data modifications to the backend storage. Throttling by too much
55 * creates an artificial limit; throttling by too little can only be sustained
56 * for short periods and would lead to highly lumpy performance. On a per-pool
57 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
58 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
59 * of dirty data decreases. When the amount of dirty data exceeds a
60 * predetermined threshold further modifications are blocked until the amount
61 * of dirty data decreases (as data is synced out).
62 *
63 * The limit on dirty data is tunable, and should be adjusted according to
64 * both the IO capacity and available memory of the system. The larger the
65 * window, the more ZFS is able to aggregate and amortize metadata (and data)
66 * changes. However, memory is a limited resource, and allowing for more dirty
67 * data comes at the cost of keeping other useful data in memory (for example
68 * ZFS data cached by the ARC).
69 *
70 * Implementation
71 *
72 * As buffers are modified dsl_pool_willuse_space() increments both the per-
73 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
74 * dirty space used; dsl_pool_dirty_space() decrements those values as data
75 * is synced out from dsl_pool_sync(). While only the poolwide value is
76 * relevant, the per-txg value is useful for debugging. The tunable
77 * zfs_dirty_data_max determines the dirty space limit. Once that value is
78 * exceeded, new writes are halted until space frees up.
79 *
80 * The zfs_dirty_data_sync tunable dictates the threshold at which we
81 * ensure that there is a txg syncing (see the comment in txg.c for a full
82 * description of transaction group stages).
83 *
84 * The IO scheduler uses both the dirty space limit and current amount of
85 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
86 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
87 *
88 * The delay is also calculated based on the amount of dirty data. See the
89 * comment above dmu_tx_delay() for details.
90 */
91
92 /*
93 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
94 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
95 */
96 uint64_t zfs_dirty_data_max;
97 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
98 int zfs_dirty_data_max_percent = 10;
99
100 /*
101 * If there is at least this much dirty data, push out a txg.
102 */
103 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
104
105 /*
106 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
107 * and delay each transaction.
108 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
109 */
110 int zfs_delay_min_dirty_percent = 60;
111
112 /*
113 * This controls how quickly the delay approaches infinity.
114 * Larger values cause it to delay more for a given amount of dirty data.
115 * Therefore larger values will cause there to be less dirty data for a
116 * given throughput.
117 *
118 * For the smoothest delay, this value should be about 1 billion divided
119 * by the maximum number of operations per second. This will smoothly
120 * handle between 10x and 1/10th this number.
121 *
122 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
123 * multiply in dmu_tx_delay().
124 */
125 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
126
127
128 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
129 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
130
131 int
132 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
133 {
134 uint64_t obj;
135 int err;
136
137 err = zap_lookup(dp->dp_meta_objset,
138 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
139 name, sizeof (obj), 1, &obj);
140 if (err)
141 return (err);
142
143 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
144 }
145
146 static dsl_pool_t *
147 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
148 {
149 dsl_pool_t *dp;
150 blkptr_t *bp = spa_get_rootblkptr(spa);
151
152 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
153 dp->dp_spa = spa;
154 dp->dp_meta_rootbp = *bp;
155 rrw_init(&dp->dp_config_rwlock, B_TRUE);
156 txg_init(dp, txg);
157
158 txg_list_create(&dp->dp_dirty_datasets,
159 offsetof(dsl_dataset_t, ds_dirty_link));
160 txg_list_create(&dp->dp_dirty_zilogs,
161 offsetof(zilog_t, zl_dirty_link));
162 txg_list_create(&dp->dp_dirty_dirs,
163 offsetof(dsl_dir_t, dd_dirty_link));
164 txg_list_create(&dp->dp_sync_tasks,
165 offsetof(dsl_sync_task_t, dst_node));
166
167 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
168 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
169
170 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
171 1, 4, 0);
172
173 return (dp);
174 }
175
176 int
177 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
178 {
179 int err;
180 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
181
182 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
183 &dp->dp_meta_objset);
184 if (err != 0)
185 dsl_pool_close(dp);
186 else
187 *dpp = dp;
188
189 return (err);
190 }
191
192 int
193 dsl_pool_open(dsl_pool_t *dp)
194 {
195 int err;
196 dsl_dir_t *dd;
197 dsl_dataset_t *ds;
198 uint64_t obj;
199
200 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
201 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
202 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
203 &dp->dp_root_dir_obj);
204 if (err)
205 goto out;
206
207 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
208 NULL, dp, &dp->dp_root_dir);
209 if (err)
210 goto out;
211
212 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
213 if (err)
214 goto out;
215
216 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
217 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
218 if (err)
219 goto out;
220 err = dsl_dataset_hold_obj(dp,
221 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
222 if (err == 0) {
223 err = dsl_dataset_hold_obj(dp,
224 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
225 &dp->dp_origin_snap);
226 dsl_dataset_rele(ds, FTAG);
227 }
228 dsl_dir_rele(dd, dp);
229 if (err)
230 goto out;
231 }
232
233 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
234 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
235 &dp->dp_free_dir);
236 if (err)
237 goto out;
238
239 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
240 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
241 if (err)
242 goto out;
243 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
244 dp->dp_meta_objset, obj));
245 }
246
247 /*
248 * Note: errors ignored, because the leak dir will not exist if we
249 * have not encountered a leak yet.
250 */
251 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
252 &dp->dp_leak_dir);
253
254 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
255 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
256 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
257 &dp->dp_bptree_obj);
258 if (err != 0)
259 goto out;
260 }
261
262 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
263 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
264 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
265 &dp->dp_empty_bpobj);
266 if (err != 0)
267 goto out;
268 }
269
270 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
271 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
272 &dp->dp_tmp_userrefs_obj);
273 if (err == ENOENT)
274 err = 0;
275 if (err)
276 goto out;
277
278 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
279
280 out:
281 rrw_exit(&dp->dp_config_rwlock, FTAG);
282 return (err);
283 }
284
285 void
286 dsl_pool_close(dsl_pool_t *dp)
287 {
288 /*
289 * Drop our references from dsl_pool_open().
290 *
291 * Since we held the origin_snap from "syncing" context (which
292 * includes pool-opening context), it actually only got a "ref"
293 * and not a hold, so just drop that here.
294 */
295 if (dp->dp_origin_snap)
296 dsl_dataset_rele(dp->dp_origin_snap, dp);
297 if (dp->dp_mos_dir)
298 dsl_dir_rele(dp->dp_mos_dir, dp);
299 if (dp->dp_free_dir)
300 dsl_dir_rele(dp->dp_free_dir, dp);
301 if (dp->dp_leak_dir)
302 dsl_dir_rele(dp->dp_leak_dir, dp);
303 if (dp->dp_root_dir)
304 dsl_dir_rele(dp->dp_root_dir, dp);
305
306 bpobj_close(&dp->dp_free_bpobj);
307
308 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
309 if (dp->dp_meta_objset)
310 dmu_objset_evict(dp->dp_meta_objset);
311
312 txg_list_destroy(&dp->dp_dirty_datasets);
313 txg_list_destroy(&dp->dp_dirty_zilogs);
314 txg_list_destroy(&dp->dp_sync_tasks);
315 txg_list_destroy(&dp->dp_dirty_dirs);
316
317 arc_flush(dp->dp_spa);
318 txg_fini(dp);
319 dsl_scan_fini(dp);
320 rrw_destroy(&dp->dp_config_rwlock);
321 mutex_destroy(&dp->dp_lock);
322 taskq_destroy(dp->dp_vnrele_taskq);
323 if (dp->dp_blkstats)
324 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
325 kmem_free(dp, sizeof (dsl_pool_t));
326 }
327
328 dsl_pool_t *
329 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
330 {
331 int err;
332 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
333 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
334 objset_t *os;
335 dsl_dataset_t *ds;
336 uint64_t obj;
337
338 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
339
340 /* create and open the MOS (meta-objset) */
341 dp->dp_meta_objset = dmu_objset_create_impl(spa,
342 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
343
344 /* create the pool directory */
345 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
346 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
347 ASSERT0(err);
348
349 /* Initialize scan structures */
350 VERIFY0(dsl_scan_init(dp, txg));
351
352 /* create and open the root dir */
353 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
354 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
355 NULL, dp, &dp->dp_root_dir));
356
357 /* create and open the meta-objset dir */
358 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
359 VERIFY0(dsl_pool_open_special_dir(dp,
360 MOS_DIR_NAME, &dp->dp_mos_dir));
361
362 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
363 /* create and open the free dir */
364 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
365 FREE_DIR_NAME, tx);
366 VERIFY0(dsl_pool_open_special_dir(dp,
367 FREE_DIR_NAME, &dp->dp_free_dir));
368
369 /* create and open the free_bplist */
370 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
371 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
372 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
373 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
374 dp->dp_meta_objset, obj));
375 }
376
377 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
378 dsl_pool_create_origin(dp, tx);
379
380 /* create the root dataset */
381 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
382
383 /* create the root objset */
384 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
385 os = dmu_objset_create_impl(dp->dp_spa, ds,
386 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
387 #ifdef _KERNEL
388 zfs_create_fs(os, kcred, zplprops, tx);
389 #endif
390 dsl_dataset_rele(ds, FTAG);
391
392 dmu_tx_commit(tx);
393
394 rrw_exit(&dp->dp_config_rwlock, FTAG);
395
396 return (dp);
397 }
398
399 /*
400 * Account for the meta-objset space in its placeholder dsl_dir.
401 */
402 void
403 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
404 int64_t used, int64_t comp, int64_t uncomp)
405 {
406 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
407 mutex_enter(&dp->dp_lock);
408 dp->dp_mos_used_delta += used;
409 dp->dp_mos_compressed_delta += comp;
410 dp->dp_mos_uncompressed_delta += uncomp;
411 mutex_exit(&dp->dp_lock);
412 }
413
414 static int
415 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
416 {
417 dsl_deadlist_t *dl = arg;
418 dsl_deadlist_insert(dl, bp, tx);
419 return (0);
420 }
421
422 static void
423 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
424 {
425 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
426 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
427 VERIFY0(zio_wait(zio));
428 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
429 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
430 }
431
432 static void
433 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
434 {
435 ASSERT(MUTEX_HELD(&dp->dp_lock));
436
437 if (delta < 0)
438 ASSERT3U(-delta, <=, dp->dp_dirty_total);
439
440 dp->dp_dirty_total += delta;
441
442 /*
443 * Note: we signal even when increasing dp_dirty_total.
444 * This ensures forward progress -- each thread wakes the next waiter.
445 */
446 if (dp->dp_dirty_total <= zfs_dirty_data_max)
447 cv_signal(&dp->dp_spaceavail_cv);
448 }
449
450 void
451 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
452 {
453 zio_t *zio;
454 dmu_tx_t *tx;
455 dsl_dir_t *dd;
456 dsl_dataset_t *ds;
457 objset_t *mos = dp->dp_meta_objset;
458 list_t synced_datasets;
459
460 list_create(&synced_datasets, sizeof (dsl_dataset_t),
461 offsetof(dsl_dataset_t, ds_synced_link));
462
463 tx = dmu_tx_create_assigned(dp, txg);
464
465 /*
466 * Write out all dirty blocks of dirty datasets.
467 */
468 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
469 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
470 /*
471 * We must not sync any non-MOS datasets twice, because
472 * we may have taken a snapshot of them. However, we
473 * may sync newly-created datasets on pass 2.
474 */
475 ASSERT(!list_link_active(&ds->ds_synced_link));
476 list_insert_tail(&synced_datasets, ds);
477 dsl_dataset_sync(ds, zio, tx);
478 }
479 VERIFY0(zio_wait(zio));
480
481 /*
482 * We have written all of the accounted dirty data, so our
483 * dp_space_towrite should now be zero. However, some seldom-used
484 * code paths do not adhere to this (e.g. dbuf_undirty(), also
485 * rounding error in dbuf_write_physdone).
486 * Shore up the accounting of any dirtied space now.
487 */
488 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
489
490 /*
491 * After the data blocks have been written (ensured by the zio_wait()
492 * above), update the user/group space accounting.
493 */
494 for (ds = list_head(&synced_datasets); ds != NULL;
495 ds = list_next(&synced_datasets, ds)) {
496 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
497 }
498
499 /*
500 * Sync the datasets again to push out the changes due to
501 * userspace updates. This must be done before we process the
502 * sync tasks, so that any snapshots will have the correct
503 * user accounting information (and we won't get confused
504 * about which blocks are part of the snapshot).
505 */
506 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
507 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
508 ASSERT(list_link_active(&ds->ds_synced_link));
509 dmu_buf_rele(ds->ds_dbuf, ds);
510 dsl_dataset_sync(ds, zio, tx);
511 }
512 VERIFY0(zio_wait(zio));
513
514 /*
515 * Now that the datasets have been completely synced, we can
516 * clean up our in-memory structures accumulated while syncing:
517 *
518 * - move dead blocks from the pending deadlist to the on-disk deadlist
519 * - release hold from dsl_dataset_dirty()
520 */
521 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
522 objset_t *os = ds->ds_objset;
523 bplist_iterate(&ds->ds_pending_deadlist,
524 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
525 ASSERT(!dmu_objset_is_dirty(os, txg));
526 dmu_buf_rele(ds->ds_dbuf, ds);
527 }
528 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
529 dsl_dir_sync(dd, tx);
530 }
531
532 /*
533 * The MOS's space is accounted for in the pool/$MOS
534 * (dp_mos_dir). We can't modify the mos while we're syncing
535 * it, so we remember the deltas and apply them here.
536 */
537 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
538 dp->dp_mos_uncompressed_delta != 0) {
539 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
540 dp->dp_mos_used_delta,
541 dp->dp_mos_compressed_delta,
542 dp->dp_mos_uncompressed_delta, tx);
543 dp->dp_mos_used_delta = 0;
544 dp->dp_mos_compressed_delta = 0;
545 dp->dp_mos_uncompressed_delta = 0;
546 }
547
548 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
549 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
550 dsl_pool_sync_mos(dp, tx);
551 }
552
553 /*
554 * If we modify a dataset in the same txg that we want to destroy it,
555 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
556 * dsl_dir_destroy_check() will fail if there are unexpected holds.
557 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
558 * and clearing the hold on it) before we process the sync_tasks.
559 * The MOS data dirtied by the sync_tasks will be synced on the next
560 * pass.
561 */
562 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
563 dsl_sync_task_t *dst;
564 /*
565 * No more sync tasks should have been added while we
566 * were syncing.
567 */
568 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
569 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
570 dsl_sync_task_sync(dst, tx);
571 }
572
573 dmu_tx_commit(tx);
574
575 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
576 }
577
578 void
579 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
580 {
581 zilog_t *zilog;
582
583 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
584 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
585 zil_clean(zilog, txg);
586 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
587 dmu_buf_rele(ds->ds_dbuf, zilog);
588 }
589 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
590 }
591
592 /*
593 * TRUE if the current thread is the tx_sync_thread or if we
594 * are being called from SPA context during pool initialization.
595 */
596 int
597 dsl_pool_sync_context(dsl_pool_t *dp)
598 {
599 return (curthread == dp->dp_tx.tx_sync_thread ||
600 spa_is_initializing(dp->dp_spa));
601 }
602
603 uint64_t
604 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
605 {
606 uint64_t space, resv;
607
608 /*
609 * If we're trying to assess whether it's OK to do a free,
610 * cut the reservation in half to allow forward progress
611 * (e.g. make it possible to rm(1) files from a full pool).
612 */
613 space = spa_get_dspace(dp->dp_spa);
614 resv = spa_get_slop_space(dp->dp_spa);
615 if (netfree)
616 resv >>= 1;
617
618 return (space - resv);
619 }
620
621 boolean_t
622 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
623 {
624 uint64_t delay_min_bytes =
625 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
626 boolean_t rv;
627
628 mutex_enter(&dp->dp_lock);
629 if (dp->dp_dirty_total > zfs_dirty_data_sync)
630 txg_kick(dp);
631 rv = (dp->dp_dirty_total > delay_min_bytes);
632 mutex_exit(&dp->dp_lock);
633 return (rv);
634 }
635
636 void
637 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
638 {
639 if (space > 0) {
640 mutex_enter(&dp->dp_lock);
641 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
642 dsl_pool_dirty_delta(dp, space);
643 mutex_exit(&dp->dp_lock);
644 }
645 }
646
647 void
648 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
649 {
650 ASSERT3S(space, >=, 0);
651 if (space == 0)
652 return;
653 mutex_enter(&dp->dp_lock);
654 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
655 /* XXX writing something we didn't dirty? */
656 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
657 }
658 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
659 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
660 ASSERT3U(dp->dp_dirty_total, >=, space);
661 dsl_pool_dirty_delta(dp, -space);
662 mutex_exit(&dp->dp_lock);
663 }
664
665 /* ARGSUSED */
666 static int
667 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
668 {
669 dmu_tx_t *tx = arg;
670 dsl_dataset_t *ds, *prev = NULL;
671 int err;
672
673 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
674 if (err)
675 return (err);
676
677 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
678 err = dsl_dataset_hold_obj(dp,
679 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
680 if (err) {
681 dsl_dataset_rele(ds, FTAG);
682 return (err);
683 }
684
685 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
686 break;
687 dsl_dataset_rele(ds, FTAG);
688 ds = prev;
689 prev = NULL;
690 }
691
692 if (prev == NULL) {
693 prev = dp->dp_origin_snap;
694
695 /*
696 * The $ORIGIN can't have any data, or the accounting
697 * will be wrong.
698 */
699 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
700
701 /* The origin doesn't get attached to itself */
702 if (ds->ds_object == prev->ds_object) {
703 dsl_dataset_rele(ds, FTAG);
704 return (0);
705 }
706
707 dmu_buf_will_dirty(ds->ds_dbuf, tx);
708 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
709 dsl_dataset_phys(ds)->ds_prev_snap_txg =
710 dsl_dataset_phys(prev)->ds_creation_txg;
711
712 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
713 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
714
715 dmu_buf_will_dirty(prev->ds_dbuf, tx);
716 dsl_dataset_phys(prev)->ds_num_children++;
717
718 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
719 ASSERT(ds->ds_prev == NULL);
720 VERIFY0(dsl_dataset_hold_obj(dp,
721 dsl_dataset_phys(ds)->ds_prev_snap_obj,
722 ds, &ds->ds_prev));
723 }
724 }
725
726 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
727 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
728
729 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
730 dmu_buf_will_dirty(prev->ds_dbuf, tx);
731 dsl_dataset_phys(prev)->ds_next_clones_obj =
732 zap_create(dp->dp_meta_objset,
733 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
734 }
735 VERIFY0(zap_add_int(dp->dp_meta_objset,
736 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
737
738 dsl_dataset_rele(ds, FTAG);
739 if (prev != dp->dp_origin_snap)
740 dsl_dataset_rele(prev, FTAG);
741 return (0);
742 }
743
744 void
745 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
746 {
747 ASSERT(dmu_tx_is_syncing(tx));
748 ASSERT(dp->dp_origin_snap != NULL);
749
750 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
751 tx, DS_FIND_CHILDREN));
752 }
753
754 /* ARGSUSED */
755 static int
756 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
757 {
758 dmu_tx_t *tx = arg;
759 objset_t *mos = dp->dp_meta_objset;
760
761 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
762 dsl_dataset_t *origin;
763
764 VERIFY0(dsl_dataset_hold_obj(dp,
765 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
766
767 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
768 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
769 dsl_dir_phys(origin->ds_dir)->dd_clones =
770 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
771 0, tx);
772 }
773
774 VERIFY0(zap_add_int(dp->dp_meta_objset,
775 dsl_dir_phys(origin->ds_dir)->dd_clones,
776 ds->ds_object, tx));
777
778 dsl_dataset_rele(origin, FTAG);
779 }
780 return (0);
781 }
782
783 void
784 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
785 {
786 ASSERT(dmu_tx_is_syncing(tx));
787 uint64_t obj;
788
789 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
790 VERIFY0(dsl_pool_open_special_dir(dp,
791 FREE_DIR_NAME, &dp->dp_free_dir));
792
793 /*
794 * We can't use bpobj_alloc(), because spa_version() still
795 * returns the old version, and we need a new-version bpobj with
796 * subobj support. So call dmu_object_alloc() directly.
797 */
798 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
799 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
800 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
801 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
802 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
803
804 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
805 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
806 }
807
808 void
809 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
810 {
811 uint64_t dsobj;
812 dsl_dataset_t *ds;
813
814 ASSERT(dmu_tx_is_syncing(tx));
815 ASSERT(dp->dp_origin_snap == NULL);
816 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
817
818 /* create the origin dir, ds, & snap-ds */
819 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
820 NULL, 0, kcred, tx);
821 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
822 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
823 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
824 dp, &dp->dp_origin_snap));
825 dsl_dataset_rele(ds, FTAG);
826 }
827
828 taskq_t *
829 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
830 {
831 return (dp->dp_vnrele_taskq);
832 }
833
834 /*
835 * Walk through the pool-wide zap object of temporary snapshot user holds
836 * and release them.
837 */
838 void
839 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
840 {
841 zap_attribute_t za;
842 zap_cursor_t zc;
843 objset_t *mos = dp->dp_meta_objset;
844 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
845 nvlist_t *holds;
846
847 if (zapobj == 0)
848 return;
849 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
850
851 holds = fnvlist_alloc();
852
853 for (zap_cursor_init(&zc, mos, zapobj);
854 zap_cursor_retrieve(&zc, &za) == 0;
855 zap_cursor_advance(&zc)) {
856 char *htag;
857 nvlist_t *tags;
858
859 htag = strchr(za.za_name, '-');
860 *htag = '\0';
861 ++htag;
862 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
863 tags = fnvlist_alloc();
864 fnvlist_add_boolean(tags, htag);
865 fnvlist_add_nvlist(holds, za.za_name, tags);
866 fnvlist_free(tags);
867 } else {
868 fnvlist_add_boolean(tags, htag);
869 }
870 }
871 dsl_dataset_user_release_tmp(dp, holds);
872 fnvlist_free(holds);
873 zap_cursor_fini(&zc);
874 }
875
876 /*
877 * Create the pool-wide zap object for storing temporary snapshot holds.
878 */
879 void
880 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
881 {
882 objset_t *mos = dp->dp_meta_objset;
883
884 ASSERT(dp->dp_tmp_userrefs_obj == 0);
885 ASSERT(dmu_tx_is_syncing(tx));
886
887 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
888 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
889 }
890
891 static int
892 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
893 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
894 {
895 objset_t *mos = dp->dp_meta_objset;
896 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
897 char *name;
898 int error;
899
900 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
901 ASSERT(dmu_tx_is_syncing(tx));
902
903 /*
904 * If the pool was created prior to SPA_VERSION_USERREFS, the
905 * zap object for temporary holds might not exist yet.
906 */
907 if (zapobj == 0) {
908 if (holding) {
909 dsl_pool_user_hold_create_obj(dp, tx);
910 zapobj = dp->dp_tmp_userrefs_obj;
911 } else {
912 return (SET_ERROR(ENOENT));
913 }
914 }
915
916 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
917 if (holding)
918 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
919 else
920 error = zap_remove(mos, zapobj, name, tx);
921 strfree(name);
922
923 return (error);
924 }
925
926 /*
927 * Add a temporary hold for the given dataset object and tag.
928 */
929 int
930 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
931 uint64_t now, dmu_tx_t *tx)
932 {
933 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
934 }
935
936 /*
937 * Release a temporary hold for the given dataset object and tag.
938 */
939 int
940 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
941 dmu_tx_t *tx)
942 {
943 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
944 tx, B_FALSE));
945 }
946
947 /*
948 * DSL Pool Configuration Lock
949 *
950 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
951 * creation / destruction / rename / property setting). It must be held for
952 * read to hold a dataset or dsl_dir. I.e. you must call
953 * dsl_pool_config_enter() or dsl_pool_hold() before calling
954 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
955 * must be held continuously until all datasets and dsl_dirs are released.
956 *
957 * The only exception to this rule is that if a "long hold" is placed on
958 * a dataset, then the dp_config_rwlock may be dropped while the dataset
959 * is still held. The long hold will prevent the dataset from being
960 * destroyed -- the destroy will fail with EBUSY. A long hold can be
961 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
962 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
963 *
964 * Legitimate long-holders (including owners) should be long-running, cancelable
965 * tasks that should cause "zfs destroy" to fail. This includes DMU
966 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
967 * "zfs send", and "zfs diff". There are several other long-holders whose
968 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
969 *
970 * The usual formula for long-holding would be:
971 * dsl_pool_hold()
972 * dsl_dataset_hold()
973 * ... perform checks ...
974 * dsl_dataset_long_hold()
975 * dsl_pool_rele()
976 * ... perform long-running task ...
977 * dsl_dataset_long_rele()
978 * dsl_dataset_rele()
979 *
980 * Note that when the long hold is released, the dataset is still held but
981 * the pool is not held. The dataset may change arbitrarily during this time
982 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
983 * dataset except release it.
984 *
985 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
986 * or modifying operations.
987 *
988 * Modifying operations should generally use dsl_sync_task(). The synctask
989 * infrastructure enforces proper locking strategy with respect to the
990 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
991 *
992 * Read-only operations will manually hold the pool, then the dataset, obtain
993 * information from the dataset, then release the pool and dataset.
994 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
995 * hold/rele.
996 */
997
998 int
999 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1000 {
1001 spa_t *spa;
1002 int error;
1003
1004 error = spa_open(name, &spa, tag);
1005 if (error == 0) {
1006 *dp = spa_get_dsl(spa);
1007 dsl_pool_config_enter(*dp, tag);
1008 }
1009 return (error);
1010 }
1011
1012 void
1013 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1014 {
1015 dsl_pool_config_exit(dp, tag);
1016 spa_close(dp->dp_spa, tag);
1017 }
1018
1019 void
1020 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1021 {
1022 /*
1023 * We use a "reentrant" reader-writer lock, but not reentrantly.
1024 *
1025 * The rrwlock can (with the track_all flag) track all reading threads,
1026 * which is very useful for debugging which code path failed to release
1027 * the lock, and for verifying that the *current* thread does hold
1028 * the lock.
1029 *
1030 * (Unlike a rwlock, which knows that N threads hold it for
1031 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1032 * if any thread holds it for read, even if this thread doesn't).
1033 */
1034 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1035 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1036 }
1037
1038 void
1039 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1040 {
1041 rrw_exit(&dp->dp_config_rwlock, tag);
1042 }
1043
1044 boolean_t
1045 dsl_pool_config_held(dsl_pool_t *dp)
1046 {
1047 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1048 }