xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_pool.c (revision 1ee13a44c225078280d7fba5012905966ff4ad92)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/dsl_pool.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_dir.h>
29 #include <sys/dsl_synctask.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/arc.h>
33 #include <sys/zap.h>
34 #include <sys/zio.h>
35 #include <sys/zfs_context.h>
36 #include <sys/fs/zfs.h>
37 #include <sys/zfs_znode.h>
38 #include <sys/spa_impl.h>
39 
40 int zfs_no_write_throttle = 0;
41 int zfs_write_limit_shift = 3;			/* 1/8th of physical memory */
42 int zfs_txg_synctime_ms = 5000;		/* target millisecs to sync a txg */
43 
44 uint64_t zfs_write_limit_min = 32 << 20;	/* min write limit is 32MB */
45 uint64_t zfs_write_limit_max = 0;		/* max data payload per txg */
46 uint64_t zfs_write_limit_inflated = 0;
47 uint64_t zfs_write_limit_override = 0;
48 
49 kmutex_t zfs_write_limit_lock;
50 
51 static pgcnt_t old_physmem = 0;
52 
53 static int
54 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
55 {
56 	uint64_t obj;
57 	int err;
58 
59 	err = zap_lookup(dp->dp_meta_objset,
60 	    dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
61 	    name, sizeof (obj), 1, &obj);
62 	if (err)
63 		return (err);
64 
65 	return (dsl_dir_open_obj(dp, obj, name, dp, ddp));
66 }
67 
68 static dsl_pool_t *
69 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
70 {
71 	dsl_pool_t *dp;
72 	blkptr_t *bp = spa_get_rootblkptr(spa);
73 
74 	dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
75 	dp->dp_spa = spa;
76 	dp->dp_meta_rootbp = *bp;
77 	rw_init(&dp->dp_config_rwlock, NULL, RW_DEFAULT, NULL);
78 	dp->dp_write_limit = zfs_write_limit_min;
79 	txg_init(dp, txg);
80 
81 	txg_list_create(&dp->dp_dirty_datasets,
82 	    offsetof(dsl_dataset_t, ds_dirty_link));
83 	txg_list_create(&dp->dp_dirty_dirs,
84 	    offsetof(dsl_dir_t, dd_dirty_link));
85 	txg_list_create(&dp->dp_sync_tasks,
86 	    offsetof(dsl_sync_task_group_t, dstg_node));
87 	list_create(&dp->dp_synced_datasets, sizeof (dsl_dataset_t),
88 	    offsetof(dsl_dataset_t, ds_synced_link));
89 
90 	mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
91 	mutex_init(&dp->dp_scrub_cancel_lock, NULL, MUTEX_DEFAULT, NULL);
92 
93 	dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
94 	    1, 4, 0);
95 
96 	return (dp);
97 }
98 
99 int
100 dsl_pool_open(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
101 {
102 	int err;
103 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
104 	dsl_dir_t *dd;
105 	dsl_dataset_t *ds;
106 
107 	rw_enter(&dp->dp_config_rwlock, RW_WRITER);
108 	err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
109 	    &dp->dp_meta_objset);
110 	if (err)
111 		goto out;
112 
113 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
114 	    DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
115 	    &dp->dp_root_dir_obj);
116 	if (err)
117 		goto out;
118 
119 	err = dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
120 	    NULL, dp, &dp->dp_root_dir);
121 	if (err)
122 		goto out;
123 
124 	err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
125 	if (err)
126 		goto out;
127 
128 	if (spa_version(spa) >= SPA_VERSION_ORIGIN) {
129 		err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
130 		if (err)
131 			goto out;
132 		err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
133 		    FTAG, &ds);
134 		if (err == 0) {
135 			err = dsl_dataset_hold_obj(dp,
136 			    ds->ds_phys->ds_prev_snap_obj, dp,
137 			    &dp->dp_origin_snap);
138 			dsl_dataset_rele(ds, FTAG);
139 		}
140 		dsl_dir_close(dd, dp);
141 		if (err)
142 			goto out;
143 	}
144 
145 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
146 	    DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
147 	    &dp->dp_tmp_userrefs_obj);
148 	if (err == ENOENT)
149 		err = 0;
150 	if (err)
151 		goto out;
152 
153 	/* get scrub status */
154 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
155 	    DMU_POOL_SCRUB_FUNC, sizeof (uint32_t), 1,
156 	    &dp->dp_scrub_func);
157 	if (err == 0) {
158 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
159 		    DMU_POOL_SCRUB_QUEUE, sizeof (uint64_t), 1,
160 		    &dp->dp_scrub_queue_obj);
161 		if (err)
162 			goto out;
163 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
164 		    DMU_POOL_SCRUB_MIN_TXG, sizeof (uint64_t), 1,
165 		    &dp->dp_scrub_min_txg);
166 		if (err)
167 			goto out;
168 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
169 		    DMU_POOL_SCRUB_MAX_TXG, sizeof (uint64_t), 1,
170 		    &dp->dp_scrub_max_txg);
171 		if (err)
172 			goto out;
173 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
174 		    DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t),
175 		    sizeof (dp->dp_scrub_bookmark) / sizeof (uint64_t),
176 		    &dp->dp_scrub_bookmark);
177 		if (err)
178 			goto out;
179 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
180 		    DMU_POOL_SCRUB_DDT_BOOKMARK, sizeof (uint64_t),
181 		    sizeof (dp->dp_scrub_ddt_bookmark) / sizeof (uint64_t),
182 		    &dp->dp_scrub_ddt_bookmark);
183 		if (err && err != ENOENT)
184 			goto out;
185 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
186 		    DMU_POOL_SCRUB_DDT_CLASS_MAX, sizeof (uint64_t), 1,
187 		    &dp->dp_scrub_ddt_class_max);
188 		if (err && err != ENOENT)
189 			goto out;
190 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
191 		    DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
192 		    &spa->spa_scrub_errors);
193 		if (err)
194 			goto out;
195 		if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) {
196 			/*
197 			 * A new-type scrub was in progress on an old
198 			 * pool.  Restart from the beginning, since the
199 			 * old software may have changed the pool in the
200 			 * meantime.
201 			 */
202 			dsl_pool_scrub_restart(dp);
203 		}
204 	} else {
205 		/*
206 		 * It's OK if there is no scrub in progress (and if
207 		 * there was an I/O error, ignore it).
208 		 */
209 		err = 0;
210 	}
211 
212 out:
213 	rw_exit(&dp->dp_config_rwlock);
214 	if (err)
215 		dsl_pool_close(dp);
216 	else
217 		*dpp = dp;
218 
219 	return (err);
220 }
221 
222 void
223 dsl_pool_close(dsl_pool_t *dp)
224 {
225 	/* drop our references from dsl_pool_open() */
226 
227 	/*
228 	 * Since we held the origin_snap from "syncing" context (which
229 	 * includes pool-opening context), it actually only got a "ref"
230 	 * and not a hold, so just drop that here.
231 	 */
232 	if (dp->dp_origin_snap)
233 		dsl_dataset_drop_ref(dp->dp_origin_snap, dp);
234 	if (dp->dp_mos_dir)
235 		dsl_dir_close(dp->dp_mos_dir, dp);
236 	if (dp->dp_root_dir)
237 		dsl_dir_close(dp->dp_root_dir, dp);
238 
239 	/* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
240 	if (dp->dp_meta_objset)
241 		dmu_objset_evict(dp->dp_meta_objset);
242 
243 	txg_list_destroy(&dp->dp_dirty_datasets);
244 	txg_list_destroy(&dp->dp_dirty_dirs);
245 	list_destroy(&dp->dp_synced_datasets);
246 
247 	arc_flush(dp->dp_spa);
248 	txg_fini(dp);
249 	rw_destroy(&dp->dp_config_rwlock);
250 	mutex_destroy(&dp->dp_lock);
251 	mutex_destroy(&dp->dp_scrub_cancel_lock);
252 	taskq_destroy(dp->dp_vnrele_taskq);
253 	if (dp->dp_blkstats)
254 		kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
255 	kmem_free(dp, sizeof (dsl_pool_t));
256 }
257 
258 dsl_pool_t *
259 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
260 {
261 	int err;
262 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
263 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
264 	objset_t *os;
265 	dsl_dataset_t *ds;
266 	uint64_t dsobj;
267 
268 	/* create and open the MOS (meta-objset) */
269 	dp->dp_meta_objset = dmu_objset_create_impl(spa,
270 	    NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
271 
272 	/* create the pool directory */
273 	err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
274 	    DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
275 	ASSERT3U(err, ==, 0);
276 
277 	/* create and open the root dir */
278 	dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
279 	VERIFY(0 == dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
280 	    NULL, dp, &dp->dp_root_dir));
281 
282 	/* create and open the meta-objset dir */
283 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
284 	VERIFY(0 == dsl_pool_open_special_dir(dp,
285 	    MOS_DIR_NAME, &dp->dp_mos_dir));
286 
287 	if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
288 		dsl_pool_create_origin(dp, tx);
289 
290 	/* create the root dataset */
291 	dsobj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
292 
293 	/* create the root objset */
294 	VERIFY(0 == dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
295 	os = dmu_objset_create_impl(dp->dp_spa, ds,
296 	    dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
297 #ifdef _KERNEL
298 	zfs_create_fs(os, kcred, zplprops, tx);
299 #endif
300 	dsl_dataset_rele(ds, FTAG);
301 
302 	dmu_tx_commit(tx);
303 
304 	return (dp);
305 }
306 
307 void
308 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
309 {
310 	zio_t *zio;
311 	dmu_tx_t *tx;
312 	dsl_dir_t *dd;
313 	dsl_dataset_t *ds;
314 	dsl_sync_task_group_t *dstg;
315 	objset_t *mos = dp->dp_meta_objset;
316 	hrtime_t start, write_time;
317 	uint64_t data_written;
318 	int err;
319 
320 	tx = dmu_tx_create_assigned(dp, txg);
321 
322 	dp->dp_read_overhead = 0;
323 	start = gethrtime();
324 
325 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
326 	while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
327 		/*
328 		 * We must not sync any non-MOS datasets twice, because
329 		 * we may have taken a snapshot of them.  However, we
330 		 * may sync newly-created datasets on pass 2.
331 		 */
332 		ASSERT(!list_link_active(&ds->ds_synced_link));
333 		list_insert_tail(&dp->dp_synced_datasets, ds);
334 		dsl_dataset_sync(ds, zio, tx);
335 	}
336 	DTRACE_PROBE(pool_sync__1setup);
337 	err = zio_wait(zio);
338 
339 	write_time = gethrtime() - start;
340 	ASSERT(err == 0);
341 	DTRACE_PROBE(pool_sync__2rootzio);
342 
343 	for (ds = list_head(&dp->dp_synced_datasets); ds;
344 	    ds = list_next(&dp->dp_synced_datasets, ds))
345 		dmu_objset_do_userquota_callbacks(ds->ds_objset, tx);
346 
347 	/*
348 	 * Sync the datasets again to push out the changes due to
349 	 * userquota updates.  This must be done before we process the
350 	 * sync tasks, because that could cause a snapshot of a dataset
351 	 * whose ds_bp will be rewritten when we do this 2nd sync.
352 	 */
353 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
354 	while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
355 		ASSERT(list_link_active(&ds->ds_synced_link));
356 		dmu_buf_rele(ds->ds_dbuf, ds);
357 		dsl_dataset_sync(ds, zio, tx);
358 	}
359 	err = zio_wait(zio);
360 
361 	/*
362 	 * If anything was added to a deadlist during a zio done callback,
363 	 * it had to be put on the deferred queue.  Enqueue it for real now.
364 	 */
365 	for (ds = list_head(&dp->dp_synced_datasets); ds;
366 	    ds = list_next(&dp->dp_synced_datasets, ds))
367 		bplist_sync(&ds->ds_deadlist,
368 		    bplist_enqueue_cb, &ds->ds_deadlist, tx);
369 
370 	while (dstg = txg_list_remove(&dp->dp_sync_tasks, txg)) {
371 		/*
372 		 * No more sync tasks should have been added while we
373 		 * were syncing.
374 		 */
375 		ASSERT(spa_sync_pass(dp->dp_spa) == 1);
376 		dsl_sync_task_group_sync(dstg, tx);
377 	}
378 	DTRACE_PROBE(pool_sync__3task);
379 
380 	start = gethrtime();
381 	while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
382 		dsl_dir_sync(dd, tx);
383 	write_time += gethrtime() - start;
384 
385 	if (spa_sync_pass(dp->dp_spa) == 1) {
386 		dp->dp_scrub_prefetch_zio_root = zio_root(dp->dp_spa, NULL,
387 		    NULL, ZIO_FLAG_CANFAIL);
388 		dsl_pool_scrub_sync(dp, tx);
389 		(void) zio_wait(dp->dp_scrub_prefetch_zio_root);
390 	}
391 
392 	start = gethrtime();
393 	if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
394 	    list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
395 		zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
396 		dmu_objset_sync(mos, zio, tx);
397 		err = zio_wait(zio);
398 		ASSERT(err == 0);
399 		dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
400 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
401 	}
402 	write_time += gethrtime() - start;
403 	DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
404 	    hrtime_t, dp->dp_read_overhead);
405 	write_time -= dp->dp_read_overhead;
406 
407 	dmu_tx_commit(tx);
408 
409 	data_written = dp->dp_space_towrite[txg & TXG_MASK];
410 	dp->dp_space_towrite[txg & TXG_MASK] = 0;
411 	ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
412 
413 	/*
414 	 * If the write limit max has not been explicitly set, set it
415 	 * to a fraction of available physical memory (default 1/8th).
416 	 * Note that we must inflate the limit because the spa
417 	 * inflates write sizes to account for data replication.
418 	 * Check this each sync phase to catch changing memory size.
419 	 */
420 	if (physmem != old_physmem && zfs_write_limit_shift) {
421 		mutex_enter(&zfs_write_limit_lock);
422 		old_physmem = physmem;
423 		zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
424 		zfs_write_limit_inflated = MAX(zfs_write_limit_min,
425 		    spa_get_asize(dp->dp_spa, zfs_write_limit_max));
426 		mutex_exit(&zfs_write_limit_lock);
427 	}
428 
429 	/*
430 	 * Attempt to keep the sync time consistent by adjusting the
431 	 * amount of write traffic allowed into each transaction group.
432 	 * Weight the throughput calculation towards the current value:
433 	 * 	thru = 3/4 old_thru + 1/4 new_thru
434 	 *
435 	 * Note: write_time is in nanosecs, so write_time/MICROSEC
436 	 * yields millisecs
437 	 */
438 	ASSERT(zfs_write_limit_min > 0);
439 	if (data_written > zfs_write_limit_min / 8 && write_time > MICROSEC) {
440 		uint64_t throughput = data_written / (write_time / MICROSEC);
441 
442 		if (dp->dp_throughput)
443 			dp->dp_throughput = throughput / 4 +
444 			    3 * dp->dp_throughput / 4;
445 		else
446 			dp->dp_throughput = throughput;
447 		dp->dp_write_limit = MIN(zfs_write_limit_inflated,
448 		    MAX(zfs_write_limit_min,
449 		    dp->dp_throughput * zfs_txg_synctime_ms));
450 	}
451 }
452 
453 void
454 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
455 {
456 	dsl_dataset_t *ds;
457 	objset_t *os;
458 
459 	while (ds = list_head(&dp->dp_synced_datasets)) {
460 		list_remove(&dp->dp_synced_datasets, ds);
461 		os = ds->ds_objset;
462 		zil_clean(os->os_zil);
463 		ASSERT(!dmu_objset_is_dirty(os, txg));
464 		dmu_buf_rele(ds->ds_dbuf, ds);
465 	}
466 	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
467 }
468 
469 /*
470  * TRUE if the current thread is the tx_sync_thread or if we
471  * are being called from SPA context during pool initialization.
472  */
473 int
474 dsl_pool_sync_context(dsl_pool_t *dp)
475 {
476 	return (curthread == dp->dp_tx.tx_sync_thread ||
477 	    spa_get_dsl(dp->dp_spa) == NULL);
478 }
479 
480 uint64_t
481 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
482 {
483 	uint64_t space, resv;
484 
485 	/*
486 	 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
487 	 * efficiency.
488 	 * XXX The intent log is not accounted for, so it must fit
489 	 * within this slop.
490 	 *
491 	 * If we're trying to assess whether it's OK to do a free,
492 	 * cut the reservation in half to allow forward progress
493 	 * (e.g. make it possible to rm(1) files from a full pool).
494 	 */
495 	space = spa_get_dspace(dp->dp_spa);
496 	resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
497 	if (netfree)
498 		resv >>= 1;
499 
500 	return (space - resv);
501 }
502 
503 int
504 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
505 {
506 	uint64_t reserved = 0;
507 	uint64_t write_limit = (zfs_write_limit_override ?
508 	    zfs_write_limit_override : dp->dp_write_limit);
509 
510 	if (zfs_no_write_throttle) {
511 		atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
512 		    space);
513 		return (0);
514 	}
515 
516 	/*
517 	 * Check to see if we have exceeded the maximum allowed IO for
518 	 * this transaction group.  We can do this without locks since
519 	 * a little slop here is ok.  Note that we do the reserved check
520 	 * with only half the requested reserve: this is because the
521 	 * reserve requests are worst-case, and we really don't want to
522 	 * throttle based off of worst-case estimates.
523 	 */
524 	if (write_limit > 0) {
525 		reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
526 		    + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
527 
528 		if (reserved && reserved > write_limit)
529 			return (ERESTART);
530 	}
531 
532 	atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
533 
534 	/*
535 	 * If this transaction group is over 7/8ths capacity, delay
536 	 * the caller 1 clock tick.  This will slow down the "fill"
537 	 * rate until the sync process can catch up with us.
538 	 */
539 	if (reserved && reserved > (write_limit - (write_limit >> 3)))
540 		txg_delay(dp, tx->tx_txg, 1);
541 
542 	return (0);
543 }
544 
545 void
546 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
547 {
548 	ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
549 	atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
550 }
551 
552 void
553 dsl_pool_memory_pressure(dsl_pool_t *dp)
554 {
555 	uint64_t space_inuse = 0;
556 	int i;
557 
558 	if (dp->dp_write_limit == zfs_write_limit_min)
559 		return;
560 
561 	for (i = 0; i < TXG_SIZE; i++) {
562 		space_inuse += dp->dp_space_towrite[i];
563 		space_inuse += dp->dp_tempreserved[i];
564 	}
565 	dp->dp_write_limit = MAX(zfs_write_limit_min,
566 	    MIN(dp->dp_write_limit, space_inuse / 4));
567 }
568 
569 void
570 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
571 {
572 	if (space > 0) {
573 		mutex_enter(&dp->dp_lock);
574 		dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
575 		mutex_exit(&dp->dp_lock);
576 	}
577 }
578 
579 /* ARGSUSED */
580 static int
581 upgrade_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
582 {
583 	dmu_tx_t *tx = arg;
584 	dsl_dataset_t *ds, *prev = NULL;
585 	int err;
586 	dsl_pool_t *dp = spa_get_dsl(spa);
587 
588 	err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
589 	if (err)
590 		return (err);
591 
592 	while (ds->ds_phys->ds_prev_snap_obj != 0) {
593 		err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
594 		    FTAG, &prev);
595 		if (err) {
596 			dsl_dataset_rele(ds, FTAG);
597 			return (err);
598 		}
599 
600 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
601 			break;
602 		dsl_dataset_rele(ds, FTAG);
603 		ds = prev;
604 		prev = NULL;
605 	}
606 
607 	if (prev == NULL) {
608 		prev = dp->dp_origin_snap;
609 
610 		/*
611 		 * The $ORIGIN can't have any data, or the accounting
612 		 * will be wrong.
613 		 */
614 		ASSERT(prev->ds_phys->ds_bp.blk_birth == 0);
615 
616 		/* The origin doesn't get attached to itself */
617 		if (ds->ds_object == prev->ds_object) {
618 			dsl_dataset_rele(ds, FTAG);
619 			return (0);
620 		}
621 
622 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
623 		ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
624 		ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
625 
626 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
627 		ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
628 
629 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
630 		prev->ds_phys->ds_num_children++;
631 
632 		if (ds->ds_phys->ds_next_snap_obj == 0) {
633 			ASSERT(ds->ds_prev == NULL);
634 			VERIFY(0 == dsl_dataset_hold_obj(dp,
635 			    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
636 		}
637 	}
638 
639 	ASSERT(ds->ds_dir->dd_phys->dd_origin_obj == prev->ds_object);
640 	ASSERT(ds->ds_phys->ds_prev_snap_obj == prev->ds_object);
641 
642 	if (prev->ds_phys->ds_next_clones_obj == 0) {
643 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
644 		prev->ds_phys->ds_next_clones_obj =
645 		    zap_create(dp->dp_meta_objset,
646 		    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
647 	}
648 	VERIFY(0 == zap_add_int(dp->dp_meta_objset,
649 	    prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
650 
651 	dsl_dataset_rele(ds, FTAG);
652 	if (prev != dp->dp_origin_snap)
653 		dsl_dataset_rele(prev, FTAG);
654 	return (0);
655 }
656 
657 void
658 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
659 {
660 	ASSERT(dmu_tx_is_syncing(tx));
661 	ASSERT(dp->dp_origin_snap != NULL);
662 
663 	VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL, upgrade_clones_cb,
664 	    tx, DS_FIND_CHILDREN));
665 }
666 
667 void
668 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
669 {
670 	uint64_t dsobj;
671 	dsl_dataset_t *ds;
672 
673 	ASSERT(dmu_tx_is_syncing(tx));
674 	ASSERT(dp->dp_origin_snap == NULL);
675 
676 	/* create the origin dir, ds, & snap-ds */
677 	rw_enter(&dp->dp_config_rwlock, RW_WRITER);
678 	dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
679 	    NULL, 0, kcred, tx);
680 	VERIFY(0 == dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
681 	dsl_dataset_snapshot_sync(ds, ORIGIN_DIR_NAME, kcred, tx);
682 	VERIFY(0 == dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
683 	    dp, &dp->dp_origin_snap));
684 	dsl_dataset_rele(ds, FTAG);
685 	rw_exit(&dp->dp_config_rwlock);
686 }
687 
688 taskq_t *
689 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
690 {
691 	return (dp->dp_vnrele_taskq);
692 }
693 
694 /*
695  * Walk through the pool-wide zap object of temporary snapshot user holds
696  * and release them.
697  */
698 void
699 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
700 {
701 	zap_attribute_t za;
702 	zap_cursor_t zc;
703 	objset_t *mos = dp->dp_meta_objset;
704 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
705 
706 	if (zapobj == 0)
707 		return;
708 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
709 
710 	for (zap_cursor_init(&zc, mos, zapobj);
711 	    zap_cursor_retrieve(&zc, &za) == 0;
712 	    zap_cursor_advance(&zc)) {
713 		char *htag;
714 		uint64_t dsobj;
715 
716 		htag = strchr(za.za_name, '-');
717 		*htag = '\0';
718 		++htag;
719 		dsobj = strtonum(za.za_name, NULL);
720 		(void) dsl_dataset_user_release_tmp(dp, dsobj, htag);
721 	}
722 	zap_cursor_fini(&zc);
723 }
724 
725 /*
726  * Create the pool-wide zap object for storing temporary snapshot holds.
727  */
728 void
729 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
730 {
731 	objset_t *mos = dp->dp_meta_objset;
732 
733 	ASSERT(dp->dp_tmp_userrefs_obj == 0);
734 	ASSERT(dmu_tx_is_syncing(tx));
735 
736 	dp->dp_tmp_userrefs_obj = zap_create(mos, DMU_OT_USERREFS,
737 	    DMU_OT_NONE, 0, tx);
738 
739 	VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS,
740 	    sizeof (uint64_t), 1, &dp->dp_tmp_userrefs_obj, tx) == 0);
741 }
742 
743 static int
744 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
745     const char *tag, uint64_t *now, dmu_tx_t *tx, boolean_t holding)
746 {
747 	objset_t *mos = dp->dp_meta_objset;
748 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
749 	char *name;
750 	int error;
751 
752 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
753 	ASSERT(dmu_tx_is_syncing(tx));
754 
755 	/*
756 	 * If the pool was created prior to SPA_VERSION_USERREFS, the
757 	 * zap object for temporary holds might not exist yet.
758 	 */
759 	if (zapobj == 0) {
760 		if (holding) {
761 			dsl_pool_user_hold_create_obj(dp, tx);
762 			zapobj = dp->dp_tmp_userrefs_obj;
763 		} else {
764 			return (ENOENT);
765 		}
766 	}
767 
768 	name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
769 	if (holding)
770 		error = zap_add(mos, zapobj, name, 8, 1, now, tx);
771 	else
772 		error = zap_remove(mos, zapobj, name, tx);
773 	strfree(name);
774 
775 	return (error);
776 }
777 
778 /*
779  * Add a temporary hold for the given dataset object and tag.
780  */
781 int
782 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
783     uint64_t *now, dmu_tx_t *tx)
784 {
785 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
786 }
787 
788 /*
789  * Release a temporary hold for the given dataset object and tag.
790  */
791 int
792 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
793     dmu_tx_t *tx)
794 {
795 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
796 	    tx, B_FALSE));
797 }
798