xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_pool.c (revision bfaed0b91e57062c38bc16b4f89db3c8f0052a9b)
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, 2016 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 void
428 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
429 {
430 	zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
431 	dmu_objset_sync(dp->dp_meta_objset, zio, tx);
432 	VERIFY0(zio_wait(zio));
433 	dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
434 	spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
435 }
436 
437 static void
438 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
439 {
440 	ASSERT(MUTEX_HELD(&dp->dp_lock));
441 
442 	if (delta < 0)
443 		ASSERT3U(-delta, <=, dp->dp_dirty_total);
444 
445 	dp->dp_dirty_total += delta;
446 
447 	/*
448 	 * Note: we signal even when increasing dp_dirty_total.
449 	 * This ensures forward progress -- each thread wakes the next waiter.
450 	 */
451 	if (dp->dp_dirty_total <= zfs_dirty_data_max)
452 		cv_signal(&dp->dp_spaceavail_cv);
453 }
454 
455 void
456 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
457 {
458 	zio_t *zio;
459 	dmu_tx_t *tx;
460 	dsl_dir_t *dd;
461 	dsl_dataset_t *ds;
462 	objset_t *mos = dp->dp_meta_objset;
463 	list_t synced_datasets;
464 
465 	list_create(&synced_datasets, sizeof (dsl_dataset_t),
466 	    offsetof(dsl_dataset_t, ds_synced_link));
467 
468 	tx = dmu_tx_create_assigned(dp, txg);
469 
470 	/*
471 	 * Write out all dirty blocks of dirty datasets.
472 	 */
473 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
474 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
475 		/*
476 		 * We must not sync any non-MOS datasets twice, because
477 		 * we may have taken a snapshot of them.  However, we
478 		 * may sync newly-created datasets on pass 2.
479 		 */
480 		ASSERT(!list_link_active(&ds->ds_synced_link));
481 		list_insert_tail(&synced_datasets, ds);
482 		dsl_dataset_sync(ds, zio, tx);
483 	}
484 	VERIFY0(zio_wait(zio));
485 
486 	/*
487 	 * We have written all of the accounted dirty data, so our
488 	 * dp_space_towrite should now be zero.  However, some seldom-used
489 	 * code paths do not adhere to this (e.g. dbuf_undirty(), also
490 	 * rounding error in dbuf_write_physdone).
491 	 * Shore up the accounting of any dirtied space now.
492 	 */
493 	dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
494 
495 	/*
496 	 * After the data blocks have been written (ensured by the zio_wait()
497 	 * above), update the user/group space accounting.
498 	 */
499 	for (ds = list_head(&synced_datasets); ds != NULL;
500 	    ds = list_next(&synced_datasets, ds)) {
501 		dmu_objset_do_userquota_updates(ds->ds_objset, tx);
502 	}
503 
504 	/*
505 	 * Sync the datasets again to push out the changes due to
506 	 * userspace updates.  This must be done before we process the
507 	 * sync tasks, so that any snapshots will have the correct
508 	 * user accounting information (and we won't get confused
509 	 * about which blocks are part of the snapshot).
510 	 */
511 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
512 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
513 		ASSERT(list_link_active(&ds->ds_synced_link));
514 		dmu_buf_rele(ds->ds_dbuf, ds);
515 		dsl_dataset_sync(ds, zio, tx);
516 	}
517 	VERIFY0(zio_wait(zio));
518 
519 	/*
520 	 * Now that the datasets have been completely synced, we can
521 	 * clean up our in-memory structures accumulated while syncing:
522 	 *
523 	 *  - move dead blocks from the pending deadlist to the on-disk deadlist
524 	 *  - release hold from dsl_dataset_dirty()
525 	 */
526 	while ((ds = list_remove_head(&synced_datasets)) != NULL) {
527 		dsl_dataset_sync_done(ds, tx);
528 	}
529 	while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
530 		dsl_dir_sync(dd, tx);
531 	}
532 
533 	/*
534 	 * The MOS's space is accounted for in the pool/$MOS
535 	 * (dp_mos_dir).  We can't modify the mos while we're syncing
536 	 * it, so we remember the deltas and apply them here.
537 	 */
538 	if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
539 	    dp->dp_mos_uncompressed_delta != 0) {
540 		dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
541 		    dp->dp_mos_used_delta,
542 		    dp->dp_mos_compressed_delta,
543 		    dp->dp_mos_uncompressed_delta, tx);
544 		dp->dp_mos_used_delta = 0;
545 		dp->dp_mos_compressed_delta = 0;
546 		dp->dp_mos_uncompressed_delta = 0;
547 	}
548 
549 	if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
550 	    list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
551 		dsl_pool_sync_mos(dp, tx);
552 	}
553 
554 	/*
555 	 * If we modify a dataset in the same txg that we want to destroy it,
556 	 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
557 	 * dsl_dir_destroy_check() will fail if there are unexpected holds.
558 	 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
559 	 * and clearing the hold on it) before we process the sync_tasks.
560 	 * The MOS data dirtied by the sync_tasks will be synced on the next
561 	 * pass.
562 	 */
563 	if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
564 		dsl_sync_task_t *dst;
565 		/*
566 		 * No more sync tasks should have been added while we
567 		 * were syncing.
568 		 */
569 		ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
570 		while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
571 			dsl_sync_task_sync(dst, tx);
572 	}
573 
574 	dmu_tx_commit(tx);
575 
576 	DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
577 }
578 
579 void
580 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
581 {
582 	zilog_t *zilog;
583 
584 	while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
585 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
586 		/*
587 		 * We don't remove the zilog from the dp_dirty_zilogs
588 		 * list until after we've cleaned it. This ensures that
589 		 * callers of zilog_is_dirty() receive an accurate
590 		 * answer when they are racing with the spa sync thread.
591 		 */
592 		zil_clean(zilog, txg);
593 		(void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
594 		ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
595 		dmu_buf_rele(ds->ds_dbuf, zilog);
596 	}
597 	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
598 }
599 
600 /*
601  * TRUE if the current thread is the tx_sync_thread or if we
602  * are being called from SPA context during pool initialization.
603  */
604 int
605 dsl_pool_sync_context(dsl_pool_t *dp)
606 {
607 	return (curthread == dp->dp_tx.tx_sync_thread ||
608 	    spa_is_initializing(dp->dp_spa));
609 }
610 
611 uint64_t
612 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
613 {
614 	uint64_t space, resv;
615 
616 	/*
617 	 * If we're trying to assess whether it's OK to do a free,
618 	 * cut the reservation in half to allow forward progress
619 	 * (e.g. make it possible to rm(1) files from a full pool).
620 	 */
621 	space = spa_get_dspace(dp->dp_spa);
622 	resv = spa_get_slop_space(dp->dp_spa);
623 	if (netfree)
624 		resv >>= 1;
625 
626 	return (space - resv);
627 }
628 
629 boolean_t
630 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
631 {
632 	uint64_t delay_min_bytes =
633 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
634 	boolean_t rv;
635 
636 	mutex_enter(&dp->dp_lock);
637 	if (dp->dp_dirty_total > zfs_dirty_data_sync)
638 		txg_kick(dp);
639 	rv = (dp->dp_dirty_total > delay_min_bytes);
640 	mutex_exit(&dp->dp_lock);
641 	return (rv);
642 }
643 
644 void
645 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
646 {
647 	if (space > 0) {
648 		mutex_enter(&dp->dp_lock);
649 		dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
650 		dsl_pool_dirty_delta(dp, space);
651 		mutex_exit(&dp->dp_lock);
652 	}
653 }
654 
655 void
656 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
657 {
658 	ASSERT3S(space, >=, 0);
659 	if (space == 0)
660 		return;
661 	mutex_enter(&dp->dp_lock);
662 	if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
663 		/* XXX writing something we didn't dirty? */
664 		space = dp->dp_dirty_pertxg[txg & TXG_MASK];
665 	}
666 	ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
667 	dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
668 	ASSERT3U(dp->dp_dirty_total, >=, space);
669 	dsl_pool_dirty_delta(dp, -space);
670 	mutex_exit(&dp->dp_lock);
671 }
672 
673 /* ARGSUSED */
674 static int
675 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
676 {
677 	dmu_tx_t *tx = arg;
678 	dsl_dataset_t *ds, *prev = NULL;
679 	int err;
680 
681 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
682 	if (err)
683 		return (err);
684 
685 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
686 		err = dsl_dataset_hold_obj(dp,
687 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
688 		if (err) {
689 			dsl_dataset_rele(ds, FTAG);
690 			return (err);
691 		}
692 
693 		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
694 			break;
695 		dsl_dataset_rele(ds, FTAG);
696 		ds = prev;
697 		prev = NULL;
698 	}
699 
700 	if (prev == NULL) {
701 		prev = dp->dp_origin_snap;
702 
703 		/*
704 		 * The $ORIGIN can't have any data, or the accounting
705 		 * will be wrong.
706 		 */
707 		rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
708 		ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
709 		rrw_exit(&ds->ds_bp_rwlock, FTAG);
710 
711 		/* The origin doesn't get attached to itself */
712 		if (ds->ds_object == prev->ds_object) {
713 			dsl_dataset_rele(ds, FTAG);
714 			return (0);
715 		}
716 
717 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
718 		dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
719 		dsl_dataset_phys(ds)->ds_prev_snap_txg =
720 		    dsl_dataset_phys(prev)->ds_creation_txg;
721 
722 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
723 		dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
724 
725 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
726 		dsl_dataset_phys(prev)->ds_num_children++;
727 
728 		if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
729 			ASSERT(ds->ds_prev == NULL);
730 			VERIFY0(dsl_dataset_hold_obj(dp,
731 			    dsl_dataset_phys(ds)->ds_prev_snap_obj,
732 			    ds, &ds->ds_prev));
733 		}
734 	}
735 
736 	ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
737 	ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
738 
739 	if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
740 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
741 		dsl_dataset_phys(prev)->ds_next_clones_obj =
742 		    zap_create(dp->dp_meta_objset,
743 		    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
744 	}
745 	VERIFY0(zap_add_int(dp->dp_meta_objset,
746 	    dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
747 
748 	dsl_dataset_rele(ds, FTAG);
749 	if (prev != dp->dp_origin_snap)
750 		dsl_dataset_rele(prev, FTAG);
751 	return (0);
752 }
753 
754 void
755 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
756 {
757 	ASSERT(dmu_tx_is_syncing(tx));
758 	ASSERT(dp->dp_origin_snap != NULL);
759 
760 	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
761 	    tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
762 }
763 
764 /* ARGSUSED */
765 static int
766 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
767 {
768 	dmu_tx_t *tx = arg;
769 	objset_t *mos = dp->dp_meta_objset;
770 
771 	if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
772 		dsl_dataset_t *origin;
773 
774 		VERIFY0(dsl_dataset_hold_obj(dp,
775 		    dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
776 
777 		if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
778 			dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
779 			dsl_dir_phys(origin->ds_dir)->dd_clones =
780 			    zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
781 			    0, tx);
782 		}
783 
784 		VERIFY0(zap_add_int(dp->dp_meta_objset,
785 		    dsl_dir_phys(origin->ds_dir)->dd_clones,
786 		    ds->ds_object, tx));
787 
788 		dsl_dataset_rele(origin, FTAG);
789 	}
790 	return (0);
791 }
792 
793 void
794 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
795 {
796 	ASSERT(dmu_tx_is_syncing(tx));
797 	uint64_t obj;
798 
799 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
800 	VERIFY0(dsl_pool_open_special_dir(dp,
801 	    FREE_DIR_NAME, &dp->dp_free_dir));
802 
803 	/*
804 	 * We can't use bpobj_alloc(), because spa_version() still
805 	 * returns the old version, and we need a new-version bpobj with
806 	 * subobj support.  So call dmu_object_alloc() directly.
807 	 */
808 	obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
809 	    SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
810 	VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
811 	    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
812 	VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
813 
814 	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
815 	    upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
816 }
817 
818 void
819 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
820 {
821 	uint64_t dsobj;
822 	dsl_dataset_t *ds;
823 
824 	ASSERT(dmu_tx_is_syncing(tx));
825 	ASSERT(dp->dp_origin_snap == NULL);
826 	ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
827 
828 	/* create the origin dir, ds, & snap-ds */
829 	dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
830 	    NULL, 0, kcred, tx);
831 	VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
832 	dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
833 	VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
834 	    dp, &dp->dp_origin_snap));
835 	dsl_dataset_rele(ds, FTAG);
836 }
837 
838 taskq_t *
839 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
840 {
841 	return (dp->dp_vnrele_taskq);
842 }
843 
844 /*
845  * Walk through the pool-wide zap object of temporary snapshot user holds
846  * and release them.
847  */
848 void
849 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
850 {
851 	zap_attribute_t za;
852 	zap_cursor_t zc;
853 	objset_t *mos = dp->dp_meta_objset;
854 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
855 	nvlist_t *holds;
856 
857 	if (zapobj == 0)
858 		return;
859 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
860 
861 	holds = fnvlist_alloc();
862 
863 	for (zap_cursor_init(&zc, mos, zapobj);
864 	    zap_cursor_retrieve(&zc, &za) == 0;
865 	    zap_cursor_advance(&zc)) {
866 		char *htag;
867 		nvlist_t *tags;
868 
869 		htag = strchr(za.za_name, '-');
870 		*htag = '\0';
871 		++htag;
872 		if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
873 			tags = fnvlist_alloc();
874 			fnvlist_add_boolean(tags, htag);
875 			fnvlist_add_nvlist(holds, za.za_name, tags);
876 			fnvlist_free(tags);
877 		} else {
878 			fnvlist_add_boolean(tags, htag);
879 		}
880 	}
881 	dsl_dataset_user_release_tmp(dp, holds);
882 	fnvlist_free(holds);
883 	zap_cursor_fini(&zc);
884 }
885 
886 /*
887  * Create the pool-wide zap object for storing temporary snapshot holds.
888  */
889 void
890 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
891 {
892 	objset_t *mos = dp->dp_meta_objset;
893 
894 	ASSERT(dp->dp_tmp_userrefs_obj == 0);
895 	ASSERT(dmu_tx_is_syncing(tx));
896 
897 	dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
898 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
899 }
900 
901 static int
902 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
903     const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
904 {
905 	objset_t *mos = dp->dp_meta_objset;
906 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
907 	char *name;
908 	int error;
909 
910 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
911 	ASSERT(dmu_tx_is_syncing(tx));
912 
913 	/*
914 	 * If the pool was created prior to SPA_VERSION_USERREFS, the
915 	 * zap object for temporary holds might not exist yet.
916 	 */
917 	if (zapobj == 0) {
918 		if (holding) {
919 			dsl_pool_user_hold_create_obj(dp, tx);
920 			zapobj = dp->dp_tmp_userrefs_obj;
921 		} else {
922 			return (SET_ERROR(ENOENT));
923 		}
924 	}
925 
926 	name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
927 	if (holding)
928 		error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
929 	else
930 		error = zap_remove(mos, zapobj, name, tx);
931 	strfree(name);
932 
933 	return (error);
934 }
935 
936 /*
937  * Add a temporary hold for the given dataset object and tag.
938  */
939 int
940 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
941     uint64_t now, dmu_tx_t *tx)
942 {
943 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
944 }
945 
946 /*
947  * Release a temporary hold for the given dataset object and tag.
948  */
949 int
950 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
951     dmu_tx_t *tx)
952 {
953 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
954 	    tx, B_FALSE));
955 }
956 
957 /*
958  * DSL Pool Configuration Lock
959  *
960  * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
961  * creation / destruction / rename / property setting).  It must be held for
962  * read to hold a dataset or dsl_dir.  I.e. you must call
963  * dsl_pool_config_enter() or dsl_pool_hold() before calling
964  * dsl_{dataset,dir}_hold{_obj}.  In most circumstances, the dp_config_rwlock
965  * must be held continuously until all datasets and dsl_dirs are released.
966  *
967  * The only exception to this rule is that if a "long hold" is placed on
968  * a dataset, then the dp_config_rwlock may be dropped while the dataset
969  * is still held.  The long hold will prevent the dataset from being
970  * destroyed -- the destroy will fail with EBUSY.  A long hold can be
971  * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
972  * (by calling dsl_{dataset,objset}_{try}own{_obj}).
973  *
974  * Legitimate long-holders (including owners) should be long-running, cancelable
975  * tasks that should cause "zfs destroy" to fail.  This includes DMU
976  * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
977  * "zfs send", and "zfs diff".  There are several other long-holders whose
978  * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
979  *
980  * The usual formula for long-holding would be:
981  * dsl_pool_hold()
982  * dsl_dataset_hold()
983  * ... perform checks ...
984  * dsl_dataset_long_hold()
985  * dsl_pool_rele()
986  * ... perform long-running task ...
987  * dsl_dataset_long_rele()
988  * dsl_dataset_rele()
989  *
990  * Note that when the long hold is released, the dataset is still held but
991  * the pool is not held.  The dataset may change arbitrarily during this time
992  * (e.g. it could be destroyed).  Therefore you shouldn't do anything to the
993  * dataset except release it.
994  *
995  * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
996  * or modifying operations.
997  *
998  * Modifying operations should generally use dsl_sync_task().  The synctask
999  * infrastructure enforces proper locking strategy with respect to the
1000  * dp_config_rwlock.  See the comment above dsl_sync_task() for details.
1001  *
1002  * Read-only operations will manually hold the pool, then the dataset, obtain
1003  * information from the dataset, then release the pool and dataset.
1004  * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1005  * hold/rele.
1006  */
1007 
1008 int
1009 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1010 {
1011 	spa_t *spa;
1012 	int error;
1013 
1014 	error = spa_open(name, &spa, tag);
1015 	if (error == 0) {
1016 		*dp = spa_get_dsl(spa);
1017 		dsl_pool_config_enter(*dp, tag);
1018 	}
1019 	return (error);
1020 }
1021 
1022 void
1023 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1024 {
1025 	dsl_pool_config_exit(dp, tag);
1026 	spa_close(dp->dp_spa, tag);
1027 }
1028 
1029 void
1030 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1031 {
1032 	/*
1033 	 * We use a "reentrant" reader-writer lock, but not reentrantly.
1034 	 *
1035 	 * The rrwlock can (with the track_all flag) track all reading threads,
1036 	 * which is very useful for debugging which code path failed to release
1037 	 * the lock, and for verifying that the *current* thread does hold
1038 	 * the lock.
1039 	 *
1040 	 * (Unlike a rwlock, which knows that N threads hold it for
1041 	 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1042 	 * if any thread holds it for read, even if this thread doesn't).
1043 	 */
1044 	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1045 	rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1046 }
1047 
1048 void
1049 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1050 {
1051 	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1052 	rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1053 }
1054 
1055 void
1056 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1057 {
1058 	rrw_exit(&dp->dp_config_rwlock, tag);
1059 }
1060 
1061 boolean_t
1062 dsl_pool_config_held(dsl_pool_t *dp)
1063 {
1064 	return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1065 }
1066 
1067 boolean_t
1068 dsl_pool_config_held_writer(dsl_pool_t *dp)
1069 {
1070 	return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1071 }
1072