xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_pool.c (revision c166b69d)
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, 2015 by Delphix. All rights reserved.
24  * Copyright (c) 2013 Steven Hartland. All rights reserved.
25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26  * Copyright (c) 2014 Integros [integros.com]
27  */
28 
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_synctask.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/arc.h>
39 #include <sys/zap.h>
40 #include <sys/zio.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/dsl_deadlist.h>
46 #include <sys/bptree.h>
47 #include <sys/zfeature.h>
48 #include <sys/zil_impl.h>
49 #include <sys/dsl_userhold.h>
50 
51 /*
52  * ZFS Write Throttle
53  * ------------------
54  *
55  * ZFS must limit the rate of incoming writes to the rate at which it is able
56  * to sync data modifications to the backend storage. Throttling by too much
57  * creates an artificial limit; throttling by too little can only be sustained
58  * for short periods and would lead to highly lumpy performance. On a per-pool
59  * basis, ZFS tracks the amount of modified (dirty) data. As operations change
60  * data, the amount of dirty data increases; as ZFS syncs out data, the amount
61  * of dirty data decreases. When the amount of dirty data exceeds a
62  * predetermined threshold further modifications are blocked until the amount
63  * of dirty data decreases (as data is synced out).
64  *
65  * The limit on dirty data is tunable, and should be adjusted according to
66  * both the IO capacity and available memory of the system. The larger the
67  * window, the more ZFS is able to aggregate and amortize metadata (and data)
68  * changes. However, memory is a limited resource, and allowing for more dirty
69  * data comes at the cost of keeping other useful data in memory (for example
70  * ZFS data cached by the ARC).
71  *
72  * Implementation
73  *
74  * As buffers are modified dsl_pool_willuse_space() increments both the per-
75  * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
76  * dirty space used; dsl_pool_dirty_space() decrements those values as data
77  * is synced out from dsl_pool_sync(). While only the poolwide value is
78  * relevant, the per-txg value is useful for debugging. The tunable
79  * zfs_dirty_data_max determines the dirty space limit. Once that value is
80  * exceeded, new writes are halted until space frees up.
81  *
82  * The zfs_dirty_data_sync tunable dictates the threshold at which we
83  * ensure that there is a txg syncing (see the comment in txg.c for a full
84  * description of transaction group stages).
85  *
86  * The IO scheduler uses both the dirty space limit and current amount of
87  * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
88  * issues. See the comment in vdev_queue.c for details of the IO scheduler.
89  *
90  * The delay is also calculated based on the amount of dirty data.  See the
91  * comment above dmu_tx_delay() for details.
92  */
93 
94 /*
95  * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
96  * capped at zfs_dirty_data_max_max.  It can also be overridden in /etc/system.
97  */
98 uint64_t zfs_dirty_data_max;
99 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
100 int zfs_dirty_data_max_percent = 10;
101 
102 /*
103  * If there is at least this much dirty data, push out a txg.
104  */
105 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
106 
107 /*
108  * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
109  * and delay each transaction.
110  * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
111  */
112 int zfs_delay_min_dirty_percent = 60;
113 
114 /*
115  * This controls how quickly the delay approaches infinity.
116  * Larger values cause it to delay more for a given amount of dirty data.
117  * Therefore larger values will cause there to be less dirty data for a
118  * given throughput.
119  *
120  * For the smoothest delay, this value should be about 1 billion divided
121  * by the maximum number of operations per second.  This will smoothly
122  * handle between 10x and 1/10th this number.
123  *
124  * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
125  * multiply in dmu_tx_delay().
126  */
127 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
128 
129 
130 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
131 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
132 
133 int
134 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
135 {
136 	uint64_t obj;
137 	int err;
138 
139 	err = zap_lookup(dp->dp_meta_objset,
140 	    dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
141 	    name, sizeof (obj), 1, &obj);
142 	if (err)
143 		return (err);
144 
145 	return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
146 }
147 
148 static dsl_pool_t *
149 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
150 {
151 	dsl_pool_t *dp;
152 	blkptr_t *bp = spa_get_rootblkptr(spa);
153 
154 	dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
155 	dp->dp_spa = spa;
156 	dp->dp_meta_rootbp = *bp;
157 	rrw_init(&dp->dp_config_rwlock, B_TRUE);
158 	txg_init(dp, txg);
159 
160 	txg_list_create(&dp->dp_dirty_datasets,
161 	    offsetof(dsl_dataset_t, ds_dirty_link));
162 	txg_list_create(&dp->dp_dirty_zilogs,
163 	    offsetof(zilog_t, zl_dirty_link));
164 	txg_list_create(&dp->dp_dirty_dirs,
165 	    offsetof(dsl_dir_t, dd_dirty_link));
166 	txg_list_create(&dp->dp_sync_tasks,
167 	    offsetof(dsl_sync_task_t, dst_node));
168 
169 	mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
170 	cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
171 
172 	dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
173 	    1, 4, 0);
174 
175 	return (dp);
176 }
177 
178 int
179 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
180 {
181 	int err;
182 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
183 
184 	err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
185 	    &dp->dp_meta_objset);
186 	if (err != 0)
187 		dsl_pool_close(dp);
188 	else
189 		*dpp = dp;
190 
191 	return (err);
192 }
193 
194 int
195 dsl_pool_open(dsl_pool_t *dp)
196 {
197 	int err;
198 	dsl_dir_t *dd;
199 	dsl_dataset_t *ds;
200 	uint64_t obj;
201 
202 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
203 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
204 	    DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
205 	    &dp->dp_root_dir_obj);
206 	if (err)
207 		goto out;
208 
209 	err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
210 	    NULL, dp, &dp->dp_root_dir);
211 	if (err)
212 		goto out;
213 
214 	err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
215 	if (err)
216 		goto out;
217 
218 	if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
219 		err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
220 		if (err)
221 			goto out;
222 		err = dsl_dataset_hold_obj(dp,
223 		    dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
224 		if (err == 0) {
225 			err = dsl_dataset_hold_obj(dp,
226 			    dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
227 			    &dp->dp_origin_snap);
228 			dsl_dataset_rele(ds, FTAG);
229 		}
230 		dsl_dir_rele(dd, dp);
231 		if (err)
232 			goto out;
233 	}
234 
235 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
236 		err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
237 		    &dp->dp_free_dir);
238 		if (err)
239 			goto out;
240 
241 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
242 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
243 		if (err)
244 			goto out;
245 		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
246 		    dp->dp_meta_objset, obj));
247 	}
248 
249 	/*
250 	 * Note: errors ignored, because the leak dir will not exist if we
251 	 * have not encountered a leak yet.
252 	 */
253 	(void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
254 	    &dp->dp_leak_dir);
255 
256 	if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
257 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
258 		    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
259 		    &dp->dp_bptree_obj);
260 		if (err != 0)
261 			goto out;
262 	}
263 
264 	if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
265 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
266 		    DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
267 		    &dp->dp_empty_bpobj);
268 		if (err != 0)
269 			goto out;
270 	}
271 
272 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
273 	    DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
274 	    &dp->dp_tmp_userrefs_obj);
275 	if (err == ENOENT)
276 		err = 0;
277 	if (err)
278 		goto out;
279 
280 	err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
281 
282 out:
283 	rrw_exit(&dp->dp_config_rwlock, FTAG);
284 	return (err);
285 }
286 
287 void
288 dsl_pool_close(dsl_pool_t *dp)
289 {
290 	/*
291 	 * Drop our references from dsl_pool_open().
292 	 *
293 	 * Since we held the origin_snap from "syncing" context (which
294 	 * includes pool-opening context), it actually only got a "ref"
295 	 * and not a hold, so just drop that here.
296 	 */
297 	if (dp->dp_origin_snap)
298 		dsl_dataset_rele(dp->dp_origin_snap, dp);
299 	if (dp->dp_mos_dir)
300 		dsl_dir_rele(dp->dp_mos_dir, dp);
301 	if (dp->dp_free_dir)
302 		dsl_dir_rele(dp->dp_free_dir, dp);
303 	if (dp->dp_leak_dir)
304 		dsl_dir_rele(dp->dp_leak_dir, dp);
305 	if (dp->dp_root_dir)
306 		dsl_dir_rele(dp->dp_root_dir, dp);
307 
308 	bpobj_close(&dp->dp_free_bpobj);
309 
310 	/* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
311 	if (dp->dp_meta_objset)
312 		dmu_objset_evict(dp->dp_meta_objset);
313 
314 	txg_list_destroy(&dp->dp_dirty_datasets);
315 	txg_list_destroy(&dp->dp_dirty_zilogs);
316 	txg_list_destroy(&dp->dp_sync_tasks);
317 	txg_list_destroy(&dp->dp_dirty_dirs);
318 
319 	/*
320 	 * We can't set retry to TRUE since we're explicitly specifying
321 	 * a spa to flush. This is good enough; any missed buffers for
322 	 * this spa won't cause trouble, and they'll eventually fall
323 	 * out of the ARC just like any other unused buffer.
324 	 */
325 	arc_flush(dp->dp_spa, FALSE);
326 
327 	txg_fini(dp);
328 	dsl_scan_fini(dp);
329 	dmu_buf_user_evict_wait();
330 
331 	rrw_destroy(&dp->dp_config_rwlock);
332 	mutex_destroy(&dp->dp_lock);
333 	taskq_destroy(dp->dp_vnrele_taskq);
334 	if (dp->dp_blkstats)
335 		kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
336 	kmem_free(dp, sizeof (dsl_pool_t));
337 }
338 
339 dsl_pool_t *
340 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
341 {
342 	int err;
343 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
344 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
345 	objset_t *os;
346 	dsl_dataset_t *ds;
347 	uint64_t obj;
348 
349 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
350 
351 	/* create and open the MOS (meta-objset) */
352 	dp->dp_meta_objset = dmu_objset_create_impl(spa,
353 	    NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
354 
355 	/* create the pool directory */
356 	err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
357 	    DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
358 	ASSERT0(err);
359 
360 	/* Initialize scan structures */
361 	VERIFY0(dsl_scan_init(dp, txg));
362 
363 	/* create and open the root dir */
364 	dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
365 	VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
366 	    NULL, dp, &dp->dp_root_dir));
367 
368 	/* create and open the meta-objset dir */
369 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
370 	VERIFY0(dsl_pool_open_special_dir(dp,
371 	    MOS_DIR_NAME, &dp->dp_mos_dir));
372 
373 	if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
374 		/* create and open the free dir */
375 		(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
376 		    FREE_DIR_NAME, tx);
377 		VERIFY0(dsl_pool_open_special_dir(dp,
378 		    FREE_DIR_NAME, &dp->dp_free_dir));
379 
380 		/* create and open the free_bplist */
381 		obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
382 		VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
383 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
384 		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
385 		    dp->dp_meta_objset, obj));
386 	}
387 
388 	if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
389 		dsl_pool_create_origin(dp, tx);
390 
391 	/* create the root dataset */
392 	obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
393 
394 	/* create the root objset */
395 	VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
396 	rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
397 	os = dmu_objset_create_impl(dp->dp_spa, ds,
398 	    dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
399 	rrw_exit(&ds->ds_bp_rwlock, FTAG);
400 #ifdef _KERNEL
401 	zfs_create_fs(os, kcred, zplprops, tx);
402 #endif
403 	dsl_dataset_rele(ds, FTAG);
404 
405 	dmu_tx_commit(tx);
406 
407 	rrw_exit(&dp->dp_config_rwlock, FTAG);
408 
409 	return (dp);
410 }
411 
412 /*
413  * Account for the meta-objset space in its placeholder dsl_dir.
414  */
415 void
416 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
417     int64_t used, int64_t comp, int64_t uncomp)
418 {
419 	ASSERT3U(comp, ==, uncomp); /* it's all metadata */
420 	mutex_enter(&dp->dp_lock);
421 	dp->dp_mos_used_delta += used;
422 	dp->dp_mos_compressed_delta += comp;
423 	dp->dp_mos_uncompressed_delta += uncomp;
424 	mutex_exit(&dp->dp_lock);
425 }
426 
427 static int
428 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
429 {
430 	dsl_deadlist_t *dl = arg;
431 	dsl_deadlist_insert(dl, bp, tx);
432 	return (0);
433 }
434 
435 static void
436 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
437 {
438 	zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
439 	dmu_objset_sync(dp->dp_meta_objset, zio, tx);
440 	VERIFY0(zio_wait(zio));
441 	dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
442 	spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
443 }
444 
445 static void
446 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
447 {
448 	ASSERT(MUTEX_HELD(&dp->dp_lock));
449 
450 	if (delta < 0)
451 		ASSERT3U(-delta, <=, dp->dp_dirty_total);
452 
453 	dp->dp_dirty_total += delta;
454 
455 	/*
456 	 * Note: we signal even when increasing dp_dirty_total.
457 	 * This ensures forward progress -- each thread wakes the next waiter.
458 	 */
459 	if (dp->dp_dirty_total <= zfs_dirty_data_max)
460 		cv_signal(&dp->dp_spaceavail_cv);
461 }
462 
463 void
464 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
465 {
466 	zio_t *zio;
467 	dmu_tx_t *tx;
468 	dsl_dir_t *dd;
469 	dsl_dataset_t *ds;
470 	objset_t *mos = dp->dp_meta_objset;
471 	list_t synced_datasets;
472 
473 	list_create(&synced_datasets, sizeof (dsl_dataset_t),
474 	    offsetof(dsl_dataset_t, ds_synced_link));
475 
476 	tx = dmu_tx_create_assigned(dp, txg);
477 
478 	/*
479 	 * Write out all dirty blocks of dirty datasets.
480 	 */
481 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
482 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
483 		/*
484 		 * We must not sync any non-MOS datasets twice, because
485 		 * we may have taken a snapshot of them.  However, we
486 		 * may sync newly-created datasets on pass 2.
487 		 */
488 		ASSERT(!list_link_active(&ds->ds_synced_link));
489 		list_insert_tail(&synced_datasets, ds);
490 		dsl_dataset_sync(ds, zio, tx);
491 	}
492 	VERIFY0(zio_wait(zio));
493 
494 	/*
495 	 * We have written all of the accounted dirty data, so our
496 	 * dp_space_towrite should now be zero.  However, some seldom-used
497 	 * code paths do not adhere to this (e.g. dbuf_undirty(), also
498 	 * rounding error in dbuf_write_physdone).
499 	 * Shore up the accounting of any dirtied space now.
500 	 */
501 	dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
502 
503 	/*
504 	 * After the data blocks have been written (ensured by the zio_wait()
505 	 * above), update the user/group space accounting.
506 	 */
507 	for (ds = list_head(&synced_datasets); ds != NULL;
508 	    ds = list_next(&synced_datasets, ds)) {
509 		dmu_objset_do_userquota_updates(ds->ds_objset, tx);
510 	}
511 
512 	/*
513 	 * Sync the datasets again to push out the changes due to
514 	 * userspace updates.  This must be done before we process the
515 	 * sync tasks, so that any snapshots will have the correct
516 	 * user accounting information (and we won't get confused
517 	 * about which blocks are part of the snapshot).
518 	 */
519 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
520 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
521 		ASSERT(list_link_active(&ds->ds_synced_link));
522 		dmu_buf_rele(ds->ds_dbuf, ds);
523 		dsl_dataset_sync(ds, zio, tx);
524 	}
525 	VERIFY0(zio_wait(zio));
526 
527 	/*
528 	 * Now that the datasets have been completely synced, we can
529 	 * clean up our in-memory structures accumulated while syncing:
530 	 *
531 	 *  - move dead blocks from the pending deadlist to the on-disk deadlist
532 	 *  - release hold from dsl_dataset_dirty()
533 	 */
534 	while ((ds = list_remove_head(&synced_datasets)) != NULL) {
535 		objset_t *os = ds->ds_objset;
536 		bplist_iterate(&ds->ds_pending_deadlist,
537 		    deadlist_enqueue_cb, &ds->ds_deadlist, tx);
538 		ASSERT(!dmu_objset_is_dirty(os, txg));
539 		dmu_buf_rele(ds->ds_dbuf, ds);
540 	}
541 	while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
542 		dsl_dir_sync(dd, tx);
543 	}
544 
545 	/*
546 	 * The MOS's space is accounted for in the pool/$MOS
547 	 * (dp_mos_dir).  We can't modify the mos while we're syncing
548 	 * it, so we remember the deltas and apply them here.
549 	 */
550 	if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
551 	    dp->dp_mos_uncompressed_delta != 0) {
552 		dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
553 		    dp->dp_mos_used_delta,
554 		    dp->dp_mos_compressed_delta,
555 		    dp->dp_mos_uncompressed_delta, tx);
556 		dp->dp_mos_used_delta = 0;
557 		dp->dp_mos_compressed_delta = 0;
558 		dp->dp_mos_uncompressed_delta = 0;
559 	}
560 
561 	if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
562 	    list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
563 		dsl_pool_sync_mos(dp, tx);
564 	}
565 
566 	/*
567 	 * If we modify a dataset in the same txg that we want to destroy it,
568 	 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
569 	 * dsl_dir_destroy_check() will fail if there are unexpected holds.
570 	 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
571 	 * and clearing the hold on it) before we process the sync_tasks.
572 	 * The MOS data dirtied by the sync_tasks will be synced on the next
573 	 * pass.
574 	 */
575 	if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
576 		dsl_sync_task_t *dst;
577 		/*
578 		 * No more sync tasks should have been added while we
579 		 * were syncing.
580 		 */
581 		ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
582 		while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
583 			dsl_sync_task_sync(dst, tx);
584 	}
585 
586 	dmu_tx_commit(tx);
587 
588 	DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
589 }
590 
591 void
592 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
593 {
594 	zilog_t *zilog;
595 
596 	while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
597 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
598 		zil_clean(zilog, txg);
599 		ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
600 		dmu_buf_rele(ds->ds_dbuf, zilog);
601 	}
602 	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
603 }
604 
605 /*
606  * TRUE if the current thread is the tx_sync_thread or if we
607  * are being called from SPA context during pool initialization.
608  */
609 int
610 dsl_pool_sync_context(dsl_pool_t *dp)
611 {
612 	return (curthread == dp->dp_tx.tx_sync_thread ||
613 	    spa_is_initializing(dp->dp_spa));
614 }
615 
616 uint64_t
617 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
618 {
619 	uint64_t space, resv;
620 
621 	/*
622 	 * If we're trying to assess whether it's OK to do a free,
623 	 * cut the reservation in half to allow forward progress
624 	 * (e.g. make it possible to rm(1) files from a full pool).
625 	 */
626 	space = spa_get_dspace(dp->dp_spa);
627 	resv = spa_get_slop_space(dp->dp_spa);
628 	if (netfree)
629 		resv >>= 1;
630 
631 	return (space - resv);
632 }
633 
634 boolean_t
635 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
636 {
637 	uint64_t delay_min_bytes =
638 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
639 	boolean_t rv;
640 
641 	mutex_enter(&dp->dp_lock);
642 	if (dp->dp_dirty_total > zfs_dirty_data_sync)
643 		txg_kick(dp);
644 	rv = (dp->dp_dirty_total > delay_min_bytes);
645 	mutex_exit(&dp->dp_lock);
646 	return (rv);
647 }
648 
649 void
650 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
651 {
652 	if (space > 0) {
653 		mutex_enter(&dp->dp_lock);
654 		dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
655 		dsl_pool_dirty_delta(dp, space);
656 		mutex_exit(&dp->dp_lock);
657 	}
658 }
659 
660 void
661 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
662 {
663 	ASSERT3S(space, >=, 0);
664 	if (space == 0)
665 		return;
666 	mutex_enter(&dp->dp_lock);
667 	if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
668 		/* XXX writing something we didn't dirty? */
669 		space = dp->dp_dirty_pertxg[txg & TXG_MASK];
670 	}
671 	ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
672 	dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
673 	ASSERT3U(dp->dp_dirty_total, >=, space);
674 	dsl_pool_dirty_delta(dp, -space);
675 	mutex_exit(&dp->dp_lock);
676 }
677 
678 /* ARGSUSED */
679 static int
680 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
681 {
682 	dmu_tx_t *tx = arg;
683 	dsl_dataset_t *ds, *prev = NULL;
684 	int err;
685 
686 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
687 	if (err)
688 		return (err);
689 
690 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
691 		err = dsl_dataset_hold_obj(dp,
692 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
693 		if (err) {
694 			dsl_dataset_rele(ds, FTAG);
695 			return (err);
696 		}
697 
698 		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
699 			break;
700 		dsl_dataset_rele(ds, FTAG);
701 		ds = prev;
702 		prev = NULL;
703 	}
704 
705 	if (prev == NULL) {
706 		prev = dp->dp_origin_snap;
707 
708 		/*
709 		 * The $ORIGIN can't have any data, or the accounting
710 		 * will be wrong.
711 		 */
712 		rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
713 		ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
714 		rrw_exit(&ds->ds_bp_rwlock, FTAG);
715 
716 		/* The origin doesn't get attached to itself */
717 		if (ds->ds_object == prev->ds_object) {
718 			dsl_dataset_rele(ds, FTAG);
719 			return (0);
720 		}
721 
722 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
723 		dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
724 		dsl_dataset_phys(ds)->ds_prev_snap_txg =
725 		    dsl_dataset_phys(prev)->ds_creation_txg;
726 
727 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
728 		dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
729 
730 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
731 		dsl_dataset_phys(prev)->ds_num_children++;
732 
733 		if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
734 			ASSERT(ds->ds_prev == NULL);
735 			VERIFY0(dsl_dataset_hold_obj(dp,
736 			    dsl_dataset_phys(ds)->ds_prev_snap_obj,
737 			    ds, &ds->ds_prev));
738 		}
739 	}
740 
741 	ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
742 	ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
743 
744 	if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
745 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
746 		dsl_dataset_phys(prev)->ds_next_clones_obj =
747 		    zap_create(dp->dp_meta_objset,
748 		    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
749 	}
750 	VERIFY0(zap_add_int(dp->dp_meta_objset,
751 	    dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
752 
753 	dsl_dataset_rele(ds, FTAG);
754 	if (prev != dp->dp_origin_snap)
755 		dsl_dataset_rele(prev, FTAG);
756 	return (0);
757 }
758 
759 void
760 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
761 {
762 	ASSERT(dmu_tx_is_syncing(tx));
763 	ASSERT(dp->dp_origin_snap != NULL);
764 
765 	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
766 	    tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
767 }
768 
769 /* ARGSUSED */
770 static int
771 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
772 {
773 	dmu_tx_t *tx = arg;
774 	objset_t *mos = dp->dp_meta_objset;
775 
776 	if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
777 		dsl_dataset_t *origin;
778 
779 		VERIFY0(dsl_dataset_hold_obj(dp,
780 		    dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
781 
782 		if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
783 			dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
784 			dsl_dir_phys(origin->ds_dir)->dd_clones =
785 			    zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
786 			    0, tx);
787 		}
788 
789 		VERIFY0(zap_add_int(dp->dp_meta_objset,
790 		    dsl_dir_phys(origin->ds_dir)->dd_clones,
791 		    ds->ds_object, tx));
792 
793 		dsl_dataset_rele(origin, FTAG);
794 	}
795 	return (0);
796 }
797 
798 void
799 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
800 {
801 	ASSERT(dmu_tx_is_syncing(tx));
802 	uint64_t obj;
803 
804 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
805 	VERIFY0(dsl_pool_open_special_dir(dp,
806 	    FREE_DIR_NAME, &dp->dp_free_dir));
807 
808 	/*
809 	 * We can't use bpobj_alloc(), because spa_version() still
810 	 * returns the old version, and we need a new-version bpobj with
811 	 * subobj support.  So call dmu_object_alloc() directly.
812 	 */
813 	obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
814 	    SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
815 	VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
816 	    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
817 	VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
818 
819 	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
820 	    upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
821 }
822 
823 void
824 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
825 {
826 	uint64_t dsobj;
827 	dsl_dataset_t *ds;
828 
829 	ASSERT(dmu_tx_is_syncing(tx));
830 	ASSERT(dp->dp_origin_snap == NULL);
831 	ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
832 
833 	/* create the origin dir, ds, & snap-ds */
834 	dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
835 	    NULL, 0, kcred, tx);
836 	VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
837 	dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
838 	VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
839 	    dp, &dp->dp_origin_snap));
840 	dsl_dataset_rele(ds, FTAG);
841 }
842 
843 taskq_t *
844 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
845 {
846 	return (dp->dp_vnrele_taskq);
847 }
848 
849 /*
850  * Walk through the pool-wide zap object of temporary snapshot user holds
851  * and release them.
852  */
853 void
854 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
855 {
856 	zap_attribute_t za;
857 	zap_cursor_t zc;
858 	objset_t *mos = dp->dp_meta_objset;
859 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
860 	nvlist_t *holds;
861 
862 	if (zapobj == 0)
863 		return;
864 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
865 
866 	holds = fnvlist_alloc();
867 
868 	for (zap_cursor_init(&zc, mos, zapobj);
869 	    zap_cursor_retrieve(&zc, &za) == 0;
870 	    zap_cursor_advance(&zc)) {
871 		char *htag;
872 		nvlist_t *tags;
873 
874 		htag = strchr(za.za_name, '-');
875 		*htag = '\0';
876 		++htag;
877 		if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
878 			tags = fnvlist_alloc();
879 			fnvlist_add_boolean(tags, htag);
880 			fnvlist_add_nvlist(holds, za.za_name, tags);
881 			fnvlist_free(tags);
882 		} else {
883 			fnvlist_add_boolean(tags, htag);
884 		}
885 	}
886 	dsl_dataset_user_release_tmp(dp, holds);
887 	fnvlist_free(holds);
888 	zap_cursor_fini(&zc);
889 }
890 
891 /*
892  * Create the pool-wide zap object for storing temporary snapshot holds.
893  */
894 void
895 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
896 {
897 	objset_t *mos = dp->dp_meta_objset;
898 
899 	ASSERT(dp->dp_tmp_userrefs_obj == 0);
900 	ASSERT(dmu_tx_is_syncing(tx));
901 
902 	dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
903 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
904 }
905 
906 static int
907 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
908     const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
909 {
910 	objset_t *mos = dp->dp_meta_objset;
911 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
912 	char *name;
913 	int error;
914 
915 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
916 	ASSERT(dmu_tx_is_syncing(tx));
917 
918 	/*
919 	 * If the pool was created prior to SPA_VERSION_USERREFS, the
920 	 * zap object for temporary holds might not exist yet.
921 	 */
922 	if (zapobj == 0) {
923 		if (holding) {
924 			dsl_pool_user_hold_create_obj(dp, tx);
925 			zapobj = dp->dp_tmp_userrefs_obj;
926 		} else {
927 			return (SET_ERROR(ENOENT));
928 		}
929 	}
930 
931 	name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
932 	if (holding)
933 		error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
934 	else
935 		error = zap_remove(mos, zapobj, name, tx);
936 	strfree(name);
937 
938 	return (error);
939 }
940 
941 /*
942  * Add a temporary hold for the given dataset object and tag.
943  */
944 int
945 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
946     uint64_t now, dmu_tx_t *tx)
947 {
948 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
949 }
950 
951 /*
952  * Release a temporary hold for the given dataset object and tag.
953  */
954 int
955 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
956     dmu_tx_t *tx)
957 {
958 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
959 	    tx, B_FALSE));
960 }
961 
962 /*
963  * DSL Pool Configuration Lock
964  *
965  * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
966  * creation / destruction / rename / property setting).  It must be held for
967  * read to hold a dataset or dsl_dir.  I.e. you must call
968  * dsl_pool_config_enter() or dsl_pool_hold() before calling
969  * dsl_{dataset,dir}_hold{_obj}.  In most circumstances, the dp_config_rwlock
970  * must be held continuously until all datasets and dsl_dirs are released.
971  *
972  * The only exception to this rule is that if a "long hold" is placed on
973  * a dataset, then the dp_config_rwlock may be dropped while the dataset
974  * is still held.  The long hold will prevent the dataset from being
975  * destroyed -- the destroy will fail with EBUSY.  A long hold can be
976  * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
977  * (by calling dsl_{dataset,objset}_{try}own{_obj}).
978  *
979  * Legitimate long-holders (including owners) should be long-running, cancelable
980  * tasks that should cause "zfs destroy" to fail.  This includes DMU
981  * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
982  * "zfs send", and "zfs diff".  There are several other long-holders whose
983  * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
984  *
985  * The usual formula for long-holding would be:
986  * dsl_pool_hold()
987  * dsl_dataset_hold()
988  * ... perform checks ...
989  * dsl_dataset_long_hold()
990  * dsl_pool_rele()
991  * ... perform long-running task ...
992  * dsl_dataset_long_rele()
993  * dsl_dataset_rele()
994  *
995  * Note that when the long hold is released, the dataset is still held but
996  * the pool is not held.  The dataset may change arbitrarily during this time
997  * (e.g. it could be destroyed).  Therefore you shouldn't do anything to the
998  * dataset except release it.
999  *
1000  * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1001  * or modifying operations.
1002  *
1003  * Modifying operations should generally use dsl_sync_task().  The synctask
1004  * infrastructure enforces proper locking strategy with respect to the
1005  * dp_config_rwlock.  See the comment above dsl_sync_task() for details.
1006  *
1007  * Read-only operations will manually hold the pool, then the dataset, obtain
1008  * information from the dataset, then release the pool and dataset.
1009  * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1010  * hold/rele.
1011  */
1012 
1013 int
1014 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1015 {
1016 	spa_t *spa;
1017 	int error;
1018 
1019 	error = spa_open(name, &spa, tag);
1020 	if (error == 0) {
1021 		*dp = spa_get_dsl(spa);
1022 		dsl_pool_config_enter(*dp, tag);
1023 	}
1024 	return (error);
1025 }
1026 
1027 void
1028 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1029 {
1030 	dsl_pool_config_exit(dp, tag);
1031 	spa_close(dp->dp_spa, tag);
1032 }
1033 
1034 void
1035 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1036 {
1037 	/*
1038 	 * We use a "reentrant" reader-writer lock, but not reentrantly.
1039 	 *
1040 	 * The rrwlock can (with the track_all flag) track all reading threads,
1041 	 * which is very useful for debugging which code path failed to release
1042 	 * the lock, and for verifying that the *current* thread does hold
1043 	 * the lock.
1044 	 *
1045 	 * (Unlike a rwlock, which knows that N threads hold it for
1046 	 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1047 	 * if any thread holds it for read, even if this thread doesn't).
1048 	 */
1049 	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1050 	rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1051 }
1052 
1053 void
1054 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1055 {
1056 	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1057 	rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1058 }
1059 
1060 void
1061 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1062 {
1063 	rrw_exit(&dp->dp_config_rwlock, tag);
1064 }
1065 
1066 boolean_t
1067 dsl_pool_config_held(dsl_pool_t *dp)
1068 {
1069 	return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1070 }
1071 
1072 boolean_t
1073 dsl_pool_config_held_writer(dsl_pool_t *dp)
1074 {
1075 	return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1076 }
1077