xref: /titanic_44/usr/src/uts/common/fs/zfs/dsl_pool.c (revision 4c6ae46166796718de10c1839ffe0569daa48250)
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) 2012 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/dsl_pool.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_prop.h>
29 #include <sys/dsl_dir.h>
30 #include <sys/dsl_synctask.h>
31 #include <sys/dsl_scan.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/arc.h>
36 #include <sys/zap.h>
37 #include <sys/zio.h>
38 #include <sys/zfs_context.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/spa_impl.h>
42 #include <sys/dsl_deadlist.h>
43 #include <sys/bptree.h>
44 #include <sys/zfeature.h>
45 #include <sys/zil_impl.h>
46 
47 int zfs_no_write_throttle = 0;
48 int zfs_write_limit_shift = 3;			/* 1/8th of physical memory */
49 int zfs_txg_synctime_ms = 1000;		/* target millisecs to sync a txg */
50 
51 uint64_t zfs_write_limit_min = 32 << 20;	/* min write limit is 32MB */
52 uint64_t zfs_write_limit_max = 0;		/* max data payload per txg */
53 uint64_t zfs_write_limit_inflated = 0;
54 uint64_t zfs_write_limit_override = 0;
55 
56 kmutex_t zfs_write_limit_lock;
57 
58 static pgcnt_t old_physmem = 0;
59 
60 int
61 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
62 {
63 	uint64_t obj;
64 	int err;
65 
66 	err = zap_lookup(dp->dp_meta_objset,
67 	    dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
68 	    name, sizeof (obj), 1, &obj);
69 	if (err)
70 		return (err);
71 
72 	return (dsl_dir_open_obj(dp, obj, name, dp, ddp));
73 }
74 
75 static dsl_pool_t *
76 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
77 {
78 	dsl_pool_t *dp;
79 	blkptr_t *bp = spa_get_rootblkptr(spa);
80 
81 	dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
82 	dp->dp_spa = spa;
83 	dp->dp_meta_rootbp = *bp;
84 	rw_init(&dp->dp_config_rwlock, NULL, RW_DEFAULT, NULL);
85 	dp->dp_write_limit = zfs_write_limit_min;
86 	txg_init(dp, txg);
87 
88 	txg_list_create(&dp->dp_dirty_datasets,
89 	    offsetof(dsl_dataset_t, ds_dirty_link));
90 	txg_list_create(&dp->dp_dirty_zilogs,
91 	    offsetof(zilog_t, zl_dirty_link));
92 	txg_list_create(&dp->dp_dirty_dirs,
93 	    offsetof(dsl_dir_t, dd_dirty_link));
94 	txg_list_create(&dp->dp_sync_tasks,
95 	    offsetof(dsl_sync_task_group_t, dstg_node));
96 
97 	mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
98 
99 	dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
100 	    1, 4, 0);
101 
102 	return (dp);
103 }
104 
105 int
106 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
107 {
108 	int err;
109 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
110 
111 	err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
112 	    &dp->dp_meta_objset);
113 	if (err != 0)
114 		dsl_pool_close(dp);
115 	else
116 		*dpp = dp;
117 
118 	return (err);
119 }
120 
121 int
122 dsl_pool_open(dsl_pool_t *dp)
123 {
124 	int err;
125 	dsl_dir_t *dd;
126 	dsl_dataset_t *ds;
127 	uint64_t obj;
128 
129 	rw_enter(&dp->dp_config_rwlock, RW_WRITER);
130 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
131 	    DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
132 	    &dp->dp_root_dir_obj);
133 	if (err)
134 		goto out;
135 
136 	err = dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
137 	    NULL, dp, &dp->dp_root_dir);
138 	if (err)
139 		goto out;
140 
141 	err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
142 	if (err)
143 		goto out;
144 
145 	if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
146 		err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
147 		if (err)
148 			goto out;
149 		err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
150 		    FTAG, &ds);
151 		if (err == 0) {
152 			err = dsl_dataset_hold_obj(dp,
153 			    ds->ds_phys->ds_prev_snap_obj, dp,
154 			    &dp->dp_origin_snap);
155 			dsl_dataset_rele(ds, FTAG);
156 		}
157 		dsl_dir_close(dd, dp);
158 		if (err)
159 			goto out;
160 	}
161 
162 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
163 		err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
164 		    &dp->dp_free_dir);
165 		if (err)
166 			goto out;
167 
168 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
169 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
170 		if (err)
171 			goto out;
172 		VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
173 		    dp->dp_meta_objset, obj));
174 	}
175 
176 	if (spa_feature_is_active(dp->dp_spa,
177 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
178 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
179 		    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
180 		    &dp->dp_bptree_obj);
181 		if (err != 0)
182 			goto out;
183 	}
184 
185 	if (spa_feature_is_active(dp->dp_spa,
186 	    &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
187 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
188 		    DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
189 		    &dp->dp_empty_bpobj);
190 		if (err != 0)
191 			goto out;
192 	}
193 
194 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
195 	    DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
196 	    &dp->dp_tmp_userrefs_obj);
197 	if (err == ENOENT)
198 		err = 0;
199 	if (err)
200 		goto out;
201 
202 	err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
203 
204 out:
205 	rw_exit(&dp->dp_config_rwlock);
206 	return (err);
207 }
208 
209 void
210 dsl_pool_close(dsl_pool_t *dp)
211 {
212 	/* drop our references from dsl_pool_open() */
213 
214 	/*
215 	 * Since we held the origin_snap from "syncing" context (which
216 	 * includes pool-opening context), it actually only got a "ref"
217 	 * and not a hold, so just drop that here.
218 	 */
219 	if (dp->dp_origin_snap)
220 		dsl_dataset_drop_ref(dp->dp_origin_snap, dp);
221 	if (dp->dp_mos_dir)
222 		dsl_dir_close(dp->dp_mos_dir, dp);
223 	if (dp->dp_free_dir)
224 		dsl_dir_close(dp->dp_free_dir, dp);
225 	if (dp->dp_root_dir)
226 		dsl_dir_close(dp->dp_root_dir, dp);
227 
228 	bpobj_close(&dp->dp_free_bpobj);
229 
230 	/* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
231 	if (dp->dp_meta_objset)
232 		dmu_objset_evict(dp->dp_meta_objset);
233 
234 	txg_list_destroy(&dp->dp_dirty_datasets);
235 	txg_list_destroy(&dp->dp_dirty_zilogs);
236 	txg_list_destroy(&dp->dp_sync_tasks);
237 	txg_list_destroy(&dp->dp_dirty_dirs);
238 
239 	arc_flush(dp->dp_spa);
240 	txg_fini(dp);
241 	dsl_scan_fini(dp);
242 	rw_destroy(&dp->dp_config_rwlock);
243 	mutex_destroy(&dp->dp_lock);
244 	taskq_destroy(dp->dp_vnrele_taskq);
245 	if (dp->dp_blkstats)
246 		kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
247 	kmem_free(dp, sizeof (dsl_pool_t));
248 }
249 
250 dsl_pool_t *
251 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
252 {
253 	int err;
254 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
255 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
256 	objset_t *os;
257 	dsl_dataset_t *ds;
258 	uint64_t obj;
259 
260 	/* create and open the MOS (meta-objset) */
261 	dp->dp_meta_objset = dmu_objset_create_impl(spa,
262 	    NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
263 
264 	/* create the pool directory */
265 	err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
266 	    DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
267 	ASSERT0(err);
268 
269 	/* Initialize scan structures */
270 	VERIFY3U(0, ==, dsl_scan_init(dp, txg));
271 
272 	/* create and open the root dir */
273 	dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
274 	VERIFY(0 == dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
275 	    NULL, dp, &dp->dp_root_dir));
276 
277 	/* create and open the meta-objset dir */
278 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
279 	VERIFY(0 == dsl_pool_open_special_dir(dp,
280 	    MOS_DIR_NAME, &dp->dp_mos_dir));
281 
282 	if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
283 		/* create and open the free dir */
284 		(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
285 		    FREE_DIR_NAME, tx);
286 		VERIFY(0 == dsl_pool_open_special_dir(dp,
287 		    FREE_DIR_NAME, &dp->dp_free_dir));
288 
289 		/* create and open the free_bplist */
290 		obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
291 		VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
292 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
293 		VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
294 		    dp->dp_meta_objset, obj));
295 	}
296 
297 	if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
298 		dsl_pool_create_origin(dp, tx);
299 
300 	/* create the root dataset */
301 	obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
302 
303 	/* create the root objset */
304 	VERIFY(0 == dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
305 	os = dmu_objset_create_impl(dp->dp_spa, ds,
306 	    dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
307 #ifdef _KERNEL
308 	zfs_create_fs(os, kcred, zplprops, tx);
309 #endif
310 	dsl_dataset_rele(ds, FTAG);
311 
312 	dmu_tx_commit(tx);
313 
314 	return (dp);
315 }
316 
317 /*
318  * Account for the meta-objset space in its placeholder dsl_dir.
319  */
320 void
321 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
322     int64_t used, int64_t comp, int64_t uncomp)
323 {
324 	ASSERT3U(comp, ==, uncomp); /* it's all metadata */
325 	mutex_enter(&dp->dp_lock);
326 	dp->dp_mos_used_delta += used;
327 	dp->dp_mos_compressed_delta += comp;
328 	dp->dp_mos_uncompressed_delta += uncomp;
329 	mutex_exit(&dp->dp_lock);
330 }
331 
332 static int
333 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
334 {
335 	dsl_deadlist_t *dl = arg;
336 	dsl_pool_t *dp = dmu_objset_pool(dl->dl_os);
337 	rw_enter(&dp->dp_config_rwlock, RW_READER);
338 	dsl_deadlist_insert(dl, bp, tx);
339 	rw_exit(&dp->dp_config_rwlock);
340 	return (0);
341 }
342 
343 void
344 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
345 {
346 	zio_t *zio;
347 	dmu_tx_t *tx;
348 	dsl_dir_t *dd;
349 	dsl_dataset_t *ds;
350 	objset_t *mos = dp->dp_meta_objset;
351 	hrtime_t start, write_time;
352 	uint64_t data_written;
353 	int err;
354 	list_t synced_datasets;
355 
356 	list_create(&synced_datasets, sizeof (dsl_dataset_t),
357 	    offsetof(dsl_dataset_t, ds_synced_link));
358 
359 	/*
360 	 * We need to copy dp_space_towrite() before doing
361 	 * dsl_sync_task_group_sync(), because
362 	 * dsl_dataset_snapshot_reserve_space() will increase
363 	 * dp_space_towrite but not actually write anything.
364 	 */
365 	data_written = dp->dp_space_towrite[txg & TXG_MASK];
366 
367 	tx = dmu_tx_create_assigned(dp, txg);
368 
369 	dp->dp_read_overhead = 0;
370 	start = gethrtime();
371 
372 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
373 	while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
374 		/*
375 		 * We must not sync any non-MOS datasets twice, because
376 		 * we may have taken a snapshot of them.  However, we
377 		 * may sync newly-created datasets on pass 2.
378 		 */
379 		ASSERT(!list_link_active(&ds->ds_synced_link));
380 		list_insert_tail(&synced_datasets, ds);
381 		dsl_dataset_sync(ds, zio, tx);
382 	}
383 	DTRACE_PROBE(pool_sync__1setup);
384 	err = zio_wait(zio);
385 
386 	write_time = gethrtime() - start;
387 	ASSERT(err == 0);
388 	DTRACE_PROBE(pool_sync__2rootzio);
389 
390 	/*
391 	 * After the data blocks have been written (ensured by the zio_wait()
392 	 * above), update the user/group space accounting.
393 	 */
394 	for (ds = list_head(&synced_datasets); ds;
395 	    ds = list_next(&synced_datasets, ds))
396 		dmu_objset_do_userquota_updates(ds->ds_objset, tx);
397 
398 	/*
399 	 * Sync the datasets again to push out the changes due to
400 	 * userspace updates.  This must be done before we process the
401 	 * sync tasks, so that any snapshots will have the correct
402 	 * user accounting information (and we won't get confused
403 	 * about which blocks are part of the snapshot).
404 	 */
405 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
406 	while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
407 		ASSERT(list_link_active(&ds->ds_synced_link));
408 		dmu_buf_rele(ds->ds_dbuf, ds);
409 		dsl_dataset_sync(ds, zio, tx);
410 	}
411 	err = zio_wait(zio);
412 
413 	/*
414 	 * Now that the datasets have been completely synced, we can
415 	 * clean up our in-memory structures accumulated while syncing:
416 	 *
417 	 *  - move dead blocks from the pending deadlist to the on-disk deadlist
418 	 *  - clean up zil records
419 	 *  - release hold from dsl_dataset_dirty()
420 	 */
421 	while (ds = list_remove_head(&synced_datasets)) {
422 		objset_t *os = ds->ds_objset;
423 		bplist_iterate(&ds->ds_pending_deadlist,
424 		    deadlist_enqueue_cb, &ds->ds_deadlist, tx);
425 		ASSERT(!dmu_objset_is_dirty(os, txg));
426 		dmu_buf_rele(ds->ds_dbuf, ds);
427 	}
428 
429 	start = gethrtime();
430 	while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
431 		dsl_dir_sync(dd, tx);
432 	write_time += gethrtime() - start;
433 
434 	/*
435 	 * The MOS's space is accounted for in the pool/$MOS
436 	 * (dp_mos_dir).  We can't modify the mos while we're syncing
437 	 * it, so we remember the deltas and apply them here.
438 	 */
439 	if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
440 	    dp->dp_mos_uncompressed_delta != 0) {
441 		dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
442 		    dp->dp_mos_used_delta,
443 		    dp->dp_mos_compressed_delta,
444 		    dp->dp_mos_uncompressed_delta, tx);
445 		dp->dp_mos_used_delta = 0;
446 		dp->dp_mos_compressed_delta = 0;
447 		dp->dp_mos_uncompressed_delta = 0;
448 	}
449 
450 	start = gethrtime();
451 	if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
452 	    list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
453 		zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
454 		dmu_objset_sync(mos, zio, tx);
455 		err = zio_wait(zio);
456 		ASSERT(err == 0);
457 		dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
458 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
459 	}
460 	write_time += gethrtime() - start;
461 	DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
462 	    hrtime_t, dp->dp_read_overhead);
463 	write_time -= dp->dp_read_overhead;
464 
465 	/*
466 	 * If we modify a dataset in the same txg that we want to destroy it,
467 	 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
468 	 * dsl_dir_destroy_check() will fail if there are unexpected holds.
469 	 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
470 	 * and clearing the hold on it) before we process the sync_tasks.
471 	 * The MOS data dirtied by the sync_tasks will be synced on the next
472 	 * pass.
473 	 */
474 	DTRACE_PROBE(pool_sync__3task);
475 	if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
476 		dsl_sync_task_group_t *dstg;
477 		/*
478 		 * No more sync tasks should have been added while we
479 		 * were syncing.
480 		 */
481 		ASSERT(spa_sync_pass(dp->dp_spa) == 1);
482 		while (dstg = txg_list_remove(&dp->dp_sync_tasks, txg))
483 			dsl_sync_task_group_sync(dstg, tx);
484 	}
485 
486 	dmu_tx_commit(tx);
487 
488 	dp->dp_space_towrite[txg & TXG_MASK] = 0;
489 	ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
490 
491 	/*
492 	 * If the write limit max has not been explicitly set, set it
493 	 * to a fraction of available physical memory (default 1/8th).
494 	 * Note that we must inflate the limit because the spa
495 	 * inflates write sizes to account for data replication.
496 	 * Check this each sync phase to catch changing memory size.
497 	 */
498 	if (physmem != old_physmem && zfs_write_limit_shift) {
499 		mutex_enter(&zfs_write_limit_lock);
500 		old_physmem = physmem;
501 		zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
502 		zfs_write_limit_inflated = MAX(zfs_write_limit_min,
503 		    spa_get_asize(dp->dp_spa, zfs_write_limit_max));
504 		mutex_exit(&zfs_write_limit_lock);
505 	}
506 
507 	/*
508 	 * Attempt to keep the sync time consistent by adjusting the
509 	 * amount of write traffic allowed into each transaction group.
510 	 * Weight the throughput calculation towards the current value:
511 	 * 	thru = 3/4 old_thru + 1/4 new_thru
512 	 *
513 	 * Note: write_time is in nanosecs, so write_time/MICROSEC
514 	 * yields millisecs
515 	 */
516 	ASSERT(zfs_write_limit_min > 0);
517 	if (data_written > zfs_write_limit_min / 8 && write_time > MICROSEC) {
518 		uint64_t throughput = data_written / (write_time / MICROSEC);
519 
520 		if (dp->dp_throughput)
521 			dp->dp_throughput = throughput / 4 +
522 			    3 * dp->dp_throughput / 4;
523 		else
524 			dp->dp_throughput = throughput;
525 		dp->dp_write_limit = MIN(zfs_write_limit_inflated,
526 		    MAX(zfs_write_limit_min,
527 		    dp->dp_throughput * zfs_txg_synctime_ms));
528 	}
529 }
530 
531 void
532 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
533 {
534 	zilog_t *zilog;
535 	dsl_dataset_t *ds;
536 
537 	while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
538 		ds = dmu_objset_ds(zilog->zl_os);
539 		zil_clean(zilog, txg);
540 		ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
541 		dmu_buf_rele(ds->ds_dbuf, zilog);
542 	}
543 	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
544 }
545 
546 /*
547  * TRUE if the current thread is the tx_sync_thread or if we
548  * are being called from SPA context during pool initialization.
549  */
550 int
551 dsl_pool_sync_context(dsl_pool_t *dp)
552 {
553 	return (curthread == dp->dp_tx.tx_sync_thread ||
554 	    spa_is_initializing(dp->dp_spa));
555 }
556 
557 uint64_t
558 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
559 {
560 	uint64_t space, resv;
561 
562 	/*
563 	 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
564 	 * efficiency.
565 	 * XXX The intent log is not accounted for, so it must fit
566 	 * within this slop.
567 	 *
568 	 * If we're trying to assess whether it's OK to do a free,
569 	 * cut the reservation in half to allow forward progress
570 	 * (e.g. make it possible to rm(1) files from a full pool).
571 	 */
572 	space = spa_get_dspace(dp->dp_spa);
573 	resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
574 	if (netfree)
575 		resv >>= 1;
576 
577 	return (space - resv);
578 }
579 
580 int
581 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
582 {
583 	uint64_t reserved = 0;
584 	uint64_t write_limit = (zfs_write_limit_override ?
585 	    zfs_write_limit_override : dp->dp_write_limit);
586 
587 	if (zfs_no_write_throttle) {
588 		atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
589 		    space);
590 		return (0);
591 	}
592 
593 	/*
594 	 * Check to see if we have exceeded the maximum allowed IO for
595 	 * this transaction group.  We can do this without locks since
596 	 * a little slop here is ok.  Note that we do the reserved check
597 	 * with only half the requested reserve: this is because the
598 	 * reserve requests are worst-case, and we really don't want to
599 	 * throttle based off of worst-case estimates.
600 	 */
601 	if (write_limit > 0) {
602 		reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
603 		    + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
604 
605 		if (reserved && reserved > write_limit)
606 			return (ERESTART);
607 	}
608 
609 	atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
610 
611 	/*
612 	 * If this transaction group is over 7/8ths capacity, delay
613 	 * the caller 1 clock tick.  This will slow down the "fill"
614 	 * rate until the sync process can catch up with us.
615 	 */
616 	if (reserved && reserved > (write_limit - (write_limit >> 3)))
617 		txg_delay(dp, tx->tx_txg, 1);
618 
619 	return (0);
620 }
621 
622 void
623 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
624 {
625 	ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
626 	atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
627 }
628 
629 void
630 dsl_pool_memory_pressure(dsl_pool_t *dp)
631 {
632 	uint64_t space_inuse = 0;
633 	int i;
634 
635 	if (dp->dp_write_limit == zfs_write_limit_min)
636 		return;
637 
638 	for (i = 0; i < TXG_SIZE; i++) {
639 		space_inuse += dp->dp_space_towrite[i];
640 		space_inuse += dp->dp_tempreserved[i];
641 	}
642 	dp->dp_write_limit = MAX(zfs_write_limit_min,
643 	    MIN(dp->dp_write_limit, space_inuse / 4));
644 }
645 
646 void
647 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
648 {
649 	if (space > 0) {
650 		mutex_enter(&dp->dp_lock);
651 		dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
652 		mutex_exit(&dp->dp_lock);
653 	}
654 }
655 
656 /* ARGSUSED */
657 static int
658 upgrade_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
659 {
660 	dmu_tx_t *tx = arg;
661 	dsl_dataset_t *ds, *prev = NULL;
662 	int err;
663 	dsl_pool_t *dp = spa_get_dsl(spa);
664 
665 	err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
666 	if (err)
667 		return (err);
668 
669 	while (ds->ds_phys->ds_prev_snap_obj != 0) {
670 		err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
671 		    FTAG, &prev);
672 		if (err) {
673 			dsl_dataset_rele(ds, FTAG);
674 			return (err);
675 		}
676 
677 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
678 			break;
679 		dsl_dataset_rele(ds, FTAG);
680 		ds = prev;
681 		prev = NULL;
682 	}
683 
684 	if (prev == NULL) {
685 		prev = dp->dp_origin_snap;
686 
687 		/*
688 		 * The $ORIGIN can't have any data, or the accounting
689 		 * will be wrong.
690 		 */
691 		ASSERT(prev->ds_phys->ds_bp.blk_birth == 0);
692 
693 		/* The origin doesn't get attached to itself */
694 		if (ds->ds_object == prev->ds_object) {
695 			dsl_dataset_rele(ds, FTAG);
696 			return (0);
697 		}
698 
699 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
700 		ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
701 		ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
702 
703 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
704 		ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
705 
706 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
707 		prev->ds_phys->ds_num_children++;
708 
709 		if (ds->ds_phys->ds_next_snap_obj == 0) {
710 			ASSERT(ds->ds_prev == NULL);
711 			VERIFY(0 == dsl_dataset_hold_obj(dp,
712 			    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
713 		}
714 	}
715 
716 	ASSERT(ds->ds_dir->dd_phys->dd_origin_obj == prev->ds_object);
717 	ASSERT(ds->ds_phys->ds_prev_snap_obj == prev->ds_object);
718 
719 	if (prev->ds_phys->ds_next_clones_obj == 0) {
720 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
721 		prev->ds_phys->ds_next_clones_obj =
722 		    zap_create(dp->dp_meta_objset,
723 		    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
724 	}
725 	VERIFY(0 == zap_add_int(dp->dp_meta_objset,
726 	    prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
727 
728 	dsl_dataset_rele(ds, FTAG);
729 	if (prev != dp->dp_origin_snap)
730 		dsl_dataset_rele(prev, FTAG);
731 	return (0);
732 }
733 
734 void
735 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
736 {
737 	ASSERT(dmu_tx_is_syncing(tx));
738 	ASSERT(dp->dp_origin_snap != NULL);
739 
740 	VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL, upgrade_clones_cb,
741 	    tx, DS_FIND_CHILDREN));
742 }
743 
744 /* ARGSUSED */
745 static int
746 upgrade_dir_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
747 {
748 	dmu_tx_t *tx = arg;
749 	dsl_dataset_t *ds;
750 	dsl_pool_t *dp = spa_get_dsl(spa);
751 	objset_t *mos = dp->dp_meta_objset;
752 
753 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
754 
755 	if (ds->ds_dir->dd_phys->dd_origin_obj) {
756 		dsl_dataset_t *origin;
757 
758 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
759 		    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
760 
761 		if (origin->ds_dir->dd_phys->dd_clones == 0) {
762 			dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
763 			origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
764 			    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
765 		}
766 
767 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
768 		    origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
769 
770 		dsl_dataset_rele(origin, FTAG);
771 	}
772 
773 	dsl_dataset_rele(ds, FTAG);
774 	return (0);
775 }
776 
777 void
778 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
779 {
780 	ASSERT(dmu_tx_is_syncing(tx));
781 	uint64_t obj;
782 
783 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
784 	VERIFY(0 == dsl_pool_open_special_dir(dp,
785 	    FREE_DIR_NAME, &dp->dp_free_dir));
786 
787 	/*
788 	 * We can't use bpobj_alloc(), because spa_version() still
789 	 * returns the old version, and we need a new-version bpobj with
790 	 * subobj support.  So call dmu_object_alloc() directly.
791 	 */
792 	obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
793 	    SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
794 	VERIFY3U(0, ==, zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
795 	    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
796 	VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
797 	    dp->dp_meta_objset, obj));
798 
799 	VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL,
800 	    upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
801 }
802 
803 void
804 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
805 {
806 	uint64_t dsobj;
807 	dsl_dataset_t *ds;
808 
809 	ASSERT(dmu_tx_is_syncing(tx));
810 	ASSERT(dp->dp_origin_snap == NULL);
811 
812 	/* create the origin dir, ds, & snap-ds */
813 	rw_enter(&dp->dp_config_rwlock, RW_WRITER);
814 	dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
815 	    NULL, 0, kcred, tx);
816 	VERIFY(0 == dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
817 	dsl_dataset_snapshot_sync(ds, ORIGIN_DIR_NAME, tx);
818 	VERIFY(0 == dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
819 	    dp, &dp->dp_origin_snap));
820 	dsl_dataset_rele(ds, FTAG);
821 	rw_exit(&dp->dp_config_rwlock);
822 }
823 
824 taskq_t *
825 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
826 {
827 	return (dp->dp_vnrele_taskq);
828 }
829 
830 /*
831  * Walk through the pool-wide zap object of temporary snapshot user holds
832  * and release them.
833  */
834 void
835 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
836 {
837 	zap_attribute_t za;
838 	zap_cursor_t zc;
839 	objset_t *mos = dp->dp_meta_objset;
840 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
841 
842 	if (zapobj == 0)
843 		return;
844 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
845 
846 	for (zap_cursor_init(&zc, mos, zapobj);
847 	    zap_cursor_retrieve(&zc, &za) == 0;
848 	    zap_cursor_advance(&zc)) {
849 		char *htag;
850 		uint64_t dsobj;
851 
852 		htag = strchr(za.za_name, '-');
853 		*htag = '\0';
854 		++htag;
855 		dsobj = strtonum(za.za_name, NULL);
856 		(void) dsl_dataset_user_release_tmp(dp, dsobj, htag, B_FALSE);
857 	}
858 	zap_cursor_fini(&zc);
859 }
860 
861 /*
862  * Create the pool-wide zap object for storing temporary snapshot holds.
863  */
864 void
865 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
866 {
867 	objset_t *mos = dp->dp_meta_objset;
868 
869 	ASSERT(dp->dp_tmp_userrefs_obj == 0);
870 	ASSERT(dmu_tx_is_syncing(tx));
871 
872 	dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
873 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
874 }
875 
876 static int
877 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
878     const char *tag, uint64_t *now, dmu_tx_t *tx, boolean_t holding)
879 {
880 	objset_t *mos = dp->dp_meta_objset;
881 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
882 	char *name;
883 	int error;
884 
885 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
886 	ASSERT(dmu_tx_is_syncing(tx));
887 
888 	/*
889 	 * If the pool was created prior to SPA_VERSION_USERREFS, the
890 	 * zap object for temporary holds might not exist yet.
891 	 */
892 	if (zapobj == 0) {
893 		if (holding) {
894 			dsl_pool_user_hold_create_obj(dp, tx);
895 			zapobj = dp->dp_tmp_userrefs_obj;
896 		} else {
897 			return (ENOENT);
898 		}
899 	}
900 
901 	name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
902 	if (holding)
903 		error = zap_add(mos, zapobj, name, 8, 1, now, tx);
904 	else
905 		error = zap_remove(mos, zapobj, name, tx);
906 	strfree(name);
907 
908 	return (error);
909 }
910 
911 /*
912  * Add a temporary hold for the given dataset object and tag.
913  */
914 int
915 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
916     uint64_t *now, dmu_tx_t *tx)
917 {
918 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
919 }
920 
921 /*
922  * Release a temporary hold for the given dataset object and tag.
923  */
924 int
925 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
926     dmu_tx_t *tx)
927 {
928 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
929 	    tx, B_FALSE));
930 }
931