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) 2013 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 less for a given amount of dirty data.
115 * Therefore larger values will cause there to be more 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 /*
129 * XXX someday maybe turn these into #defines, and you have to tune it on a
130 * per-pool basis using zfs.conf.
131 */
132
133
134 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
135 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
136
137 int
138 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
139 {
140 uint64_t obj;
141 int err;
142
143 err = zap_lookup(dp->dp_meta_objset,
144 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
145 name, sizeof (obj), 1, &obj);
146 if (err)
147 return (err);
148
149 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
150 }
151
152 static dsl_pool_t *
153 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
154 {
155 dsl_pool_t *dp;
156 blkptr_t *bp = spa_get_rootblkptr(spa);
157
158 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
159 dp->dp_spa = spa;
160 dp->dp_meta_rootbp = *bp;
161 rrw_init(&dp->dp_config_rwlock, B_TRUE);
162 txg_init(dp, txg);
163
164 txg_list_create(&dp->dp_dirty_datasets,
165 offsetof(dsl_dataset_t, ds_dirty_link));
166 txg_list_create(&dp->dp_dirty_zilogs,
167 offsetof(zilog_t, zl_dirty_link));
168 txg_list_create(&dp->dp_dirty_dirs,
169 offsetof(dsl_dir_t, dd_dirty_link));
170 txg_list_create(&dp->dp_sync_tasks,
171 offsetof(dsl_sync_task_t, dst_node));
172
173 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
174 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
175
176 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
177 1, 4, 0);
178
179 return (dp);
180 }
181
182 int
183 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
184 {
185 int err;
186 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
187
188 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
189 &dp->dp_meta_objset);
190 if (err != 0)
191 dsl_pool_close(dp);
192 else
193 *dpp = dp;
194
195 return (err);
196 }
197
198 int
199 dsl_pool_open(dsl_pool_t *dp)
200 {
201 int err;
202 dsl_dir_t *dd;
203 dsl_dataset_t *ds;
204 uint64_t obj;
205
206 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
207 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
208 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
209 &dp->dp_root_dir_obj);
210 if (err)
211 goto out;
212
213 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
214 NULL, dp, &dp->dp_root_dir);
215 if (err)
216 goto out;
217
218 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
219 if (err)
220 goto out;
221
222 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
223 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
224 if (err)
225 goto out;
226 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
227 FTAG, &ds);
228 if (err == 0) {
229 err = dsl_dataset_hold_obj(dp,
230 ds->ds_phys->ds_prev_snap_obj, dp,
231 &dp->dp_origin_snap);
232 dsl_dataset_rele(ds, FTAG);
233 }
234 dsl_dir_rele(dd, dp);
235 if (err)
236 goto out;
237 }
238
239 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
240 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
241 &dp->dp_free_dir);
242 if (err)
243 goto out;
244
245 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
246 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
247 if (err)
248 goto out;
249 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
250 dp->dp_meta_objset, obj));
251 }
252
253 /*
254 * Note: errors ignored, because the leak dir will not exist if we
255 * have not encountered a leak yet.
256 */
257 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
258 &dp->dp_leak_dir);
259
260 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
261 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
262 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
263 &dp->dp_bptree_obj);
264 if (err != 0)
265 goto out;
266 }
267
268 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
269 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
270 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
271 &dp->dp_empty_bpobj);
272 if (err != 0)
273 goto out;
274 }
275
276 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
277 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
278 &dp->dp_tmp_userrefs_obj);
279 if (err == ENOENT)
280 err = 0;
281 if (err)
282 goto out;
283
284 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
285
286 out:
287 rrw_exit(&dp->dp_config_rwlock, FTAG);
288 return (err);
289 }
290
291 void
292 dsl_pool_close(dsl_pool_t *dp)
293 {
294 /*
295 * Drop our references from dsl_pool_open().
296 *
297 * Since we held the origin_snap from "syncing" context (which
298 * includes pool-opening context), it actually only got a "ref"
299 * and not a hold, so just drop that here.
300 */
301 if (dp->dp_origin_snap)
302 dsl_dataset_rele(dp->dp_origin_snap, dp);
303 if (dp->dp_mos_dir)
304 dsl_dir_rele(dp->dp_mos_dir, dp);
305 if (dp->dp_free_dir)
306 dsl_dir_rele(dp->dp_free_dir, dp);
307 if (dp->dp_leak_dir)
308 dsl_dir_rele(dp->dp_leak_dir, dp);
309 if (dp->dp_root_dir)
310 dsl_dir_rele(dp->dp_root_dir, dp);
311
312 bpobj_close(&dp->dp_free_bpobj);
313
314 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
315 if (dp->dp_meta_objset)
316 dmu_objset_evict(dp->dp_meta_objset);
317
318 txg_list_destroy(&dp->dp_dirty_datasets);
319 txg_list_destroy(&dp->dp_dirty_zilogs);
320 txg_list_destroy(&dp->dp_sync_tasks);
321 txg_list_destroy(&dp->dp_dirty_dirs);
322
323 arc_flush(dp->dp_spa);
324 txg_fini(dp);
325 dsl_scan_fini(dp);
326 rrw_destroy(&dp->dp_config_rwlock);
327 mutex_destroy(&dp->dp_lock);
328 taskq_destroy(dp->dp_vnrele_taskq);
329 if (dp->dp_blkstats)
330 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
331 kmem_free(dp, sizeof (dsl_pool_t));
332 }
333
334 dsl_pool_t *
335 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
336 {
337 int err;
338 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
339 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
340 objset_t *os;
341 dsl_dataset_t *ds;
342 uint64_t obj;
343
344 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
345
346 /* create and open the MOS (meta-objset) */
347 dp->dp_meta_objset = dmu_objset_create_impl(spa,
348 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
349
350 /* create the pool directory */
351 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
352 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
353 ASSERT0(err);
354
355 /* Initialize scan structures */
356 VERIFY0(dsl_scan_init(dp, txg));
357
358 /* create and open the root dir */
359 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
360 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
361 NULL, dp, &dp->dp_root_dir));
362
363 /* create and open the meta-objset dir */
364 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
365 VERIFY0(dsl_pool_open_special_dir(dp,
366 MOS_DIR_NAME, &dp->dp_mos_dir));
367
368 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
369 /* create and open the free dir */
370 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
371 FREE_DIR_NAME, tx);
372 VERIFY0(dsl_pool_open_special_dir(dp,
373 FREE_DIR_NAME, &dp->dp_free_dir));
374
375 /* create and open the free_bplist */
376 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
377 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
378 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
379 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
380 dp->dp_meta_objset, obj));
381 }
382
383 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
384 dsl_pool_create_origin(dp, tx);
385
386 /* create the root dataset */
387 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
388
389 /* create the root objset */
390 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
391 os = dmu_objset_create_impl(dp->dp_spa, ds,
392 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
393 #ifdef _KERNEL
394 zfs_create_fs(os, kcred, zplprops, tx);
395 #endif
396 dsl_dataset_rele(ds, FTAG);
397
398 dmu_tx_commit(tx);
399
400 rrw_exit(&dp->dp_config_rwlock, FTAG);
401
402 return (dp);
403 }
404
405 /*
406 * Account for the meta-objset space in its placeholder dsl_dir.
407 */
408 void
409 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
410 int64_t used, int64_t comp, int64_t uncomp)
411 {
412 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
413 mutex_enter(&dp->dp_lock);
414 dp->dp_mos_used_delta += used;
415 dp->dp_mos_compressed_delta += comp;
416 dp->dp_mos_uncompressed_delta += uncomp;
417 mutex_exit(&dp->dp_lock);
418 }
419
420 static int
421 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
422 {
423 dsl_deadlist_t *dl = arg;
424 dsl_deadlist_insert(dl, bp, tx);
425 return (0);
426 }
427
428 static void
429 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
430 {
431 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
432 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
433 VERIFY0(zio_wait(zio));
434 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
435 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
436 }
437
438 static void
439 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
440 {
441 ASSERT(MUTEX_HELD(&dp->dp_lock));
442
443 if (delta < 0)
444 ASSERT3U(-delta, <=, dp->dp_dirty_total);
445
446 dp->dp_dirty_total += delta;
447
448 /*
449 * Note: we signal even when increasing dp_dirty_total.
450 * This ensures forward progress -- each thread wakes the next waiter.
451 */
452 if (dp->dp_dirty_total <= zfs_dirty_data_max)
453 cv_signal(&dp->dp_spaceavail_cv);
454 }
455
456 void
457 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
458 {
459 zio_t *zio;
460 dmu_tx_t *tx;
461 dsl_dir_t *dd;
462 dsl_dataset_t *ds;
463 objset_t *mos = dp->dp_meta_objset;
464 list_t synced_datasets;
465
466 list_create(&synced_datasets, sizeof (dsl_dataset_t),
467 offsetof(dsl_dataset_t, ds_synced_link));
468
469 tx = dmu_tx_create_assigned(dp, txg);
470
471 /*
472 * Write out all dirty blocks of dirty datasets.
473 */
474 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
475 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
476 /*
477 * We must not sync any non-MOS datasets twice, because
478 * we may have taken a snapshot of them. However, we
479 * may sync newly-created datasets on pass 2.
480 */
481 ASSERT(!list_link_active(&ds->ds_synced_link));
482 list_insert_tail(&synced_datasets, ds);
483 dsl_dataset_sync(ds, zio, tx);
484 }
485 VERIFY0(zio_wait(zio));
486
487 /*
488 * We have written all of the accounted dirty data, so our
489 * dp_space_towrite should now be zero. However, some seldom-used
490 * code paths do not adhere to this (e.g. dbuf_undirty(), also
491 * rounding error in dbuf_write_physdone).
492 * Shore up the accounting of any dirtied space now.
493 */
494 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
495
496 /*
497 * After the data blocks have been written (ensured by the zio_wait()
498 * above), update the user/group space accounting.
499 */
500 for (ds = list_head(&synced_datasets); ds != NULL;
501 ds = list_next(&synced_datasets, ds)) {
502 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
503 }
504
505 /*
506 * Sync the datasets again to push out the changes due to
507 * userspace updates. This must be done before we process the
508 * sync tasks, so that any snapshots will have the correct
509 * user accounting information (and we won't get confused
510 * about which blocks are part of the snapshot).
511 */
512 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
513 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
514 ASSERT(list_link_active(&ds->ds_synced_link));
515 dmu_buf_rele(ds->ds_dbuf, ds);
516 dsl_dataset_sync(ds, zio, tx);
517 }
518 VERIFY0(zio_wait(zio));
519
520 /*
521 * Now that the datasets have been completely synced, we can
522 * clean up our in-memory structures accumulated while syncing:
523 *
524 * - move dead blocks from the pending deadlist to the on-disk deadlist
525 * - release hold from dsl_dataset_dirty()
526 */
527 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
528 objset_t *os = ds->ds_objset;
529 bplist_iterate(&ds->ds_pending_deadlist,
530 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
531 ASSERT(!dmu_objset_is_dirty(os, txg));
532 dmu_buf_rele(ds->ds_dbuf, ds);
533 }
534 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
535 dsl_dir_sync(dd, tx);
536 }
537
538 /*
539 * The MOS's space is accounted for in the pool/$MOS
540 * (dp_mos_dir). We can't modify the mos while we're syncing
541 * it, so we remember the deltas and apply them here.
542 */
543 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
544 dp->dp_mos_uncompressed_delta != 0) {
545 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
546 dp->dp_mos_used_delta,
547 dp->dp_mos_compressed_delta,
548 dp->dp_mos_uncompressed_delta, tx);
549 dp->dp_mos_used_delta = 0;
550 dp->dp_mos_compressed_delta = 0;
551 dp->dp_mos_uncompressed_delta = 0;
552 }
553
554 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
555 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
556 dsl_pool_sync_mos(dp, tx);
557 }
558
559 /*
560 * If we modify a dataset in the same txg that we want to destroy it,
561 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
562 * dsl_dir_destroy_check() will fail if there are unexpected holds.
563 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
564 * and clearing the hold on it) before we process the sync_tasks.
565 * The MOS data dirtied by the sync_tasks will be synced on the next
566 * pass.
567 */
568 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
569 dsl_sync_task_t *dst;
570 /*
571 * No more sync tasks should have been added while we
572 * were syncing.
573 */
574 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
575 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
576 dsl_sync_task_sync(dst, tx);
577 }
578
579 dmu_tx_commit(tx);
580
581 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
582 }
583
584 void
585 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
586 {
587 zilog_t *zilog;
588
589 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
590 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
591 zil_clean(zilog, txg);
592 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
593 dmu_buf_rele(ds->ds_dbuf, zilog);
594 }
595 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
596 }
597
598 /*
599 * TRUE if the current thread is the tx_sync_thread or if we
600 * are being called from SPA context during pool initialization.
601 */
602 int
603 dsl_pool_sync_context(dsl_pool_t *dp)
604 {
605 return (curthread == dp->dp_tx.tx_sync_thread ||
606 spa_is_initializing(dp->dp_spa));
607 }
608
609 uint64_t
610 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
611 {
612 uint64_t space, resv;
613
614 /*
615 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
616 * efficiency.
617 * XXX The intent log is not accounted for, so it must fit
618 * within this slop.
619 *
620 * If we're trying to assess whether it's OK to do a free,
621 * cut the reservation in half to allow forward progress
622 * (e.g. make it possible to rm(1) files from a full pool).
623 */
624 space = spa_get_dspace(dp->dp_spa);
625 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
626 if (netfree)
627 resv >>= 1;
628
629 return (space - resv);
630 }
631
632 boolean_t
633 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
634 {
635 uint64_t delay_min_bytes =
636 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
637 boolean_t rv;
638
639 mutex_enter(&dp->dp_lock);
640 if (dp->dp_dirty_total > zfs_dirty_data_sync)
641 txg_kick(dp);
642 rv = (dp->dp_dirty_total > delay_min_bytes);
643 mutex_exit(&dp->dp_lock);
644 return (rv);
645 }
646
647 void
648 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
649 {
650 if (space > 0) {
651 mutex_enter(&dp->dp_lock);
652 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
653 dsl_pool_dirty_delta(dp, space);
654 mutex_exit(&dp->dp_lock);
655 }
656 }
657
658 void
659 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
660 {
661 ASSERT3S(space, >=, 0);
662 if (space == 0)
663 return;
664 mutex_enter(&dp->dp_lock);
665 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
666 /* XXX writing something we didn't dirty? */
667 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
668 }
669 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
670 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
671 ASSERT3U(dp->dp_dirty_total, >=, space);
672 dsl_pool_dirty_delta(dp, -space);
673 mutex_exit(&dp->dp_lock);
674 }
675
676 /* ARGSUSED */
677 static int
678 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
679 {
680 dmu_tx_t *tx = arg;
681 dsl_dataset_t *ds, *prev = NULL;
682 int err;
683
684 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
685 if (err)
686 return (err);
687
688 while (ds->ds_phys->ds_prev_snap_obj != 0) {
689 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
690 FTAG, &prev);
691 if (err) {
692 dsl_dataset_rele(ds, FTAG);
693 return (err);
694 }
695
696 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
697 break;
698 dsl_dataset_rele(ds, FTAG);
699 ds = prev;
700 prev = NULL;
701 }
702
703 if (prev == NULL) {
704 prev = dp->dp_origin_snap;
705
706 /*
707 * The $ORIGIN can't have any data, or the accounting
708 * will be wrong.
709 */
710 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
711
712 /* The origin doesn't get attached to itself */
713 if (ds->ds_object == prev->ds_object) {
714 dsl_dataset_rele(ds, FTAG);
715 return (0);
716 }
717
718 dmu_buf_will_dirty(ds->ds_dbuf, tx);
719 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
720 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
721
722 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
723 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
724
725 dmu_buf_will_dirty(prev->ds_dbuf, tx);
726 prev->ds_phys->ds_num_children++;
727
728 if (ds->ds_phys->ds_next_snap_obj == 0) {
729 ASSERT(ds->ds_prev == NULL);
730 VERIFY0(dsl_dataset_hold_obj(dp,
731 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
732 }
733 }
734
735 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
736 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
737
738 if (prev->ds_phys->ds_next_clones_obj == 0) {
739 dmu_buf_will_dirty(prev->ds_dbuf, tx);
740 prev->ds_phys->ds_next_clones_obj =
741 zap_create(dp->dp_meta_objset,
742 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
743 }
744 VERIFY0(zap_add_int(dp->dp_meta_objset,
745 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
746
747 dsl_dataset_rele(ds, FTAG);
748 if (prev != dp->dp_origin_snap)
749 dsl_dataset_rele(prev, FTAG);
750 return (0);
751 }
752
753 void
754 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
755 {
756 ASSERT(dmu_tx_is_syncing(tx));
757 ASSERT(dp->dp_origin_snap != NULL);
758
759 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
760 tx, DS_FIND_CHILDREN));
761 }
762
763 /* ARGSUSED */
764 static int
765 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
766 {
767 dmu_tx_t *tx = arg;
768 objset_t *mos = dp->dp_meta_objset;
769
770 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
771 dsl_dataset_t *origin;
772
773 VERIFY0(dsl_dataset_hold_obj(dp,
774 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
775
776 if (origin->ds_dir->dd_phys->dd_clones == 0) {
777 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
778 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
779 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
780 }
781
782 VERIFY0(zap_add_int(dp->dp_meta_objset,
783 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
784
785 dsl_dataset_rele(origin, FTAG);
786 }
787 return (0);
788 }
789
790 void
791 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
792 {
793 ASSERT(dmu_tx_is_syncing(tx));
794 uint64_t obj;
795
796 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
797 VERIFY0(dsl_pool_open_special_dir(dp,
798 FREE_DIR_NAME, &dp->dp_free_dir));
799
800 /*
801 * We can't use bpobj_alloc(), because spa_version() still
802 * returns the old version, and we need a new-version bpobj with
803 * subobj support. So call dmu_object_alloc() directly.
804 */
805 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
806 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
807 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
808 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
809 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
810
811 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
812 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
813 }
814
815 void
816 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
817 {
818 uint64_t dsobj;
819 dsl_dataset_t *ds;
820
821 ASSERT(dmu_tx_is_syncing(tx));
822 ASSERT(dp->dp_origin_snap == NULL);
823 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
824
825 /* create the origin dir, ds, & snap-ds */
826 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
827 NULL, 0, kcred, tx);
828 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
829 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
830 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
831 dp, &dp->dp_origin_snap));
832 dsl_dataset_rele(ds, FTAG);
833 }
834
835 taskq_t *
836 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
837 {
838 return (dp->dp_vnrele_taskq);
839 }
840
841 /*
842 * Walk through the pool-wide zap object of temporary snapshot user holds
843 * and release them.
844 */
845 void
846 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
847 {
848 zap_attribute_t za;
849 zap_cursor_t zc;
850 objset_t *mos = dp->dp_meta_objset;
851 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
852 nvlist_t *holds;
853
854 if (zapobj == 0)
855 return;
856 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
857
858 holds = fnvlist_alloc();
859
860 for (zap_cursor_init(&zc, mos, zapobj);
861 zap_cursor_retrieve(&zc, &za) == 0;
862 zap_cursor_advance(&zc)) {
863 char *htag;
864 nvlist_t *tags;
865
866 htag = strchr(za.za_name, '-');
867 *htag = '\0';
868 ++htag;
869 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
870 tags = fnvlist_alloc();
871 fnvlist_add_boolean(tags, htag);
872 fnvlist_add_nvlist(holds, za.za_name, tags);
873 fnvlist_free(tags);
874 } else {
875 fnvlist_add_boolean(tags, htag);
876 }
877 }
878 dsl_dataset_user_release_tmp(dp, holds);
879 fnvlist_free(holds);
880 zap_cursor_fini(&zc);
881 }
882
883 /*
884 * Create the pool-wide zap object for storing temporary snapshot holds.
885 */
886 void
887 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
888 {
889 objset_t *mos = dp->dp_meta_objset;
890
891 ASSERT(dp->dp_tmp_userrefs_obj == 0);
892 ASSERT(dmu_tx_is_syncing(tx));
893
894 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
895 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
896 }
897
898 static int
899 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
900 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
901 {
902 objset_t *mos = dp->dp_meta_objset;
903 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
904 char *name;
905 int error;
906
907 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
908 ASSERT(dmu_tx_is_syncing(tx));
909
910 /*
911 * If the pool was created prior to SPA_VERSION_USERREFS, the
912 * zap object for temporary holds might not exist yet.
913 */
914 if (zapobj == 0) {
915 if (holding) {
916 dsl_pool_user_hold_create_obj(dp, tx);
917 zapobj = dp->dp_tmp_userrefs_obj;
918 } else {
919 return (SET_ERROR(ENOENT));
920 }
921 }
922
923 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
924 if (holding)
925 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
926 else
927 error = zap_remove(mos, zapobj, name, tx);
928 strfree(name);
929
930 return (error);
931 }
932
933 /*
934 * Add a temporary hold for the given dataset object and tag.
935 */
936 int
937 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
938 uint64_t now, dmu_tx_t *tx)
939 {
940 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
941 }
942
943 /*
944 * Release a temporary hold for the given dataset object and tag.
945 */
946 int
947 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
948 dmu_tx_t *tx)
949 {
950 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
951 tx, B_FALSE));
952 }
953
954 /*
955 * DSL Pool Configuration Lock
956 *
957 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
958 * creation / destruction / rename / property setting). It must be held for
959 * read to hold a dataset or dsl_dir. I.e. you must call
960 * dsl_pool_config_enter() or dsl_pool_hold() before calling
961 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
962 * must be held continuously until all datasets and dsl_dirs are released.
963 *
964 * The only exception to this rule is that if a "long hold" is placed on
965 * a dataset, then the dp_config_rwlock may be dropped while the dataset
966 * is still held. The long hold will prevent the dataset from being
967 * destroyed -- the destroy will fail with EBUSY. A long hold can be
968 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
969 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
970 *
971 * Legitimate long-holders (including owners) should be long-running, cancelable
972 * tasks that should cause "zfs destroy" to fail. This includes DMU
973 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
974 * "zfs send", and "zfs diff". There are several other long-holders whose
975 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
976 *
977 * The usual formula for long-holding would be:
978 * dsl_pool_hold()
979 * dsl_dataset_hold()
980 * ... perform checks ...
981 * dsl_dataset_long_hold()
982 * dsl_pool_rele()
983 * ... perform long-running task ...
984 * dsl_dataset_long_rele()
985 * dsl_dataset_rele()
986 *
987 * Note that when the long hold is released, the dataset is still held but
988 * the pool is not held. The dataset may change arbitrarily during this time
989 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
990 * dataset except release it.
991 *
992 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
993 * or modifying operations.
994 *
995 * Modifying operations should generally use dsl_sync_task(). The synctask
996 * infrastructure enforces proper locking strategy with respect to the
997 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
998 *
999 * Read-only operations will manually hold the pool, then the dataset, obtain
1000 * information from the dataset, then release the pool and dataset.
1001 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1002 * hold/rele.
1003 */
1004
1005 int
1006 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1007 {
1008 spa_t *spa;
1009 int error;
1010
1011 error = spa_open(name, &spa, tag);
1012 if (error == 0) {
1013 *dp = spa_get_dsl(spa);
1014 dsl_pool_config_enter(*dp, tag);
1015 }
1016 return (error);
1017 }
1018
1019 void
1020 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1021 {
1022 dsl_pool_config_exit(dp, tag);
1023 spa_close(dp->dp_spa, tag);
1024 }
1025
1026 void
1027 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1028 {
1029 /*
1030 * We use a "reentrant" reader-writer lock, but not reentrantly.
1031 *
1032 * The rrwlock can (with the track_all flag) track all reading threads,
1033 * which is very useful for debugging which code path failed to release
1034 * the lock, and for verifying that the *current* thread does hold
1035 * the lock.
1036 *
1037 * (Unlike a rwlock, which knows that N threads hold it for
1038 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1039 * if any thread holds it for read, even if this thread doesn't).
1040 */
1041 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1042 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1043 }
1044
1045 void
1046 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1047 {
1048 rrw_exit(&dp->dp_config_rwlock, tag);
1049 }
1050
1051 boolean_t
1052 dsl_pool_config_held(dsl_pool_t *dp)
1053 {
1054 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1055 }