1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 */
26
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/zap.h>
38 #include <sys/zio.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
48
49 /*
50 * ZFS Write Throttle
51 * ------------------
52 *
53 * ZFS must limit the rate of incoming writes to the rate at which it is able
54 * to sync data modifications to the backend storage. Throttling by too much
55 * creates an artificial limit; throttling by too little can only be sustained
56 * for short periods and would lead to highly lumpy performance. On a per-pool
57 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
58 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
59 * of dirty data decreases. When the amount of dirty data exceeds a
60 * predetermined threshold further modifications are blocked until the amount
61 * of dirty data decreases (as data is synced out).
62 *
63 * The limit on dirty data is tunable, and should be adjusted according to
64 * both the IO capacity and available memory of the system. The larger the
65 * window, the more ZFS is able to aggregate and amortize metadata (and data)
66 * changes. However, memory is a limited resource, and allowing for more dirty
67 * data comes at the cost of keeping other useful data in memory (for example
68 * ZFS data cached by the ARC).
69 *
70 * Implementation
71 *
72 * As buffers are modified dsl_pool_willuse_space() increments both the per-
73 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
74 * dirty space used; dsl_pool_dirty_space() decrements those values as data
75 * is synced out from dsl_pool_sync(). While only the poolwide value is
76 * relevant, the per-txg value is useful for debugging. The tunable
77 * zfs_dirty_data_max determines the dirty space limit. Once that value is
78 * exceeded, new writes are halted until space frees up.
79 *
80 * The zfs_dirty_data_sync tunable dictates the threshold at which we
81 * ensure that there is a txg syncing (see the comment in txg.c for a full
82 * description of transaction group stages).
83 *
84 * The IO scheduler uses both the dirty space limit and current amount of
85 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
86 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
87 *
88 * The delay is also calculated based on the amount of dirty data. See the
89 * comment above dmu_tx_delay() for details.
90 */
91
92 /*
93 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
94 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
95 */
96 uint64_t zfs_dirty_data_max;
97 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
98 int zfs_dirty_data_max_percent = 10;
99
100 /*
101 * If there is at least this much dirty data, push out a txg.
102 */
103 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
104
105 /*
106 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
107 * and delay each transaction.
108 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
109 */
110 int zfs_delay_min_dirty_percent = 60;
111
112 /*
113 * This controls how quickly the delay approaches infinity.
114 * Larger values cause it to delay less for a given amount of dirty data.
115 * Therefore larger values will cause there to be more dirty data for a
116 * given throughput.
117 *
118 * For the smoothest delay, this value should be about 1 billion divided
119 * by the maximum number of operations per second. This will smoothly
120 * handle between 10x and 1/10th this number.
121 *
122 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
123 * multiply in dmu_tx_delay().
124 */
125 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
126
127
128 /*
129 * XXX someday maybe turn these into #defines, and you have to tune it on a
130 * per-pool basis using zfs.conf.
131 */
132
133
134 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
135 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
136
137 int
138 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
139 {
140 uint64_t obj;
141 int err;
142
143 err = zap_lookup(dp->dp_meta_objset,
144 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
145 name, sizeof (obj), 1, &obj);
146 if (err)
147 return (err);
148
149 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
150 }
151
152 static dsl_pool_t *
153 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
154 {
155 dsl_pool_t *dp;
156 blkptr_t *bp = spa_get_rootblkptr(spa);
157
158 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
159 dp->dp_spa = spa;
160 dp->dp_meta_rootbp = *bp;
161 rrw_init(&dp->dp_config_rwlock, B_TRUE);
162 txg_init(dp, txg);
163
164 txg_list_create(&dp->dp_dirty_datasets,
165 offsetof(dsl_dataset_t, ds_dirty_link));
166 txg_list_create(&dp->dp_dirty_zilogs,
167 offsetof(zilog_t, zl_dirty_link));
168 txg_list_create(&dp->dp_dirty_dirs,
169 offsetof(dsl_dir_t, dd_dirty_link));
170 txg_list_create(&dp->dp_sync_tasks,
171 offsetof(dsl_sync_task_t, dst_node));
172
173 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
174 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
175
176 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
177 1, 4, 0);
178
179 return (dp);
180 }
181
182 int
183 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
184 {
185 int err;
186 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
187
188 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
189 &dp->dp_meta_objset);
190 if (err != 0)
191 dsl_pool_close(dp);
192 else
193 *dpp = dp;
194
195 return (err);
196 }
197
198 int
199 dsl_pool_open(dsl_pool_t *dp)
200 {
201 int err;
202 dsl_dir_t *dd;
203 dsl_dataset_t *ds;
204 uint64_t obj;
205
206 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
207 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
208 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
209 &dp->dp_root_dir_obj);
210 if (err)
211 goto out;
212
213 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
214 NULL, dp, &dp->dp_root_dir);
215 if (err)
216 goto out;
217
218 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
219 if (err)
220 goto out;
221
222 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
223 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
224 if (err)
225 goto out;
226 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
227 FTAG, &ds);
228 if (err == 0) {
229 err = dsl_dataset_hold_obj(dp,
230 ds->ds_phys->ds_prev_snap_obj, dp,
231 &dp->dp_origin_snap);
232 dsl_dataset_rele(ds, FTAG);
233 }
234 dsl_dir_rele(dd, dp);
235 if (err)
236 goto out;
237 }
238
239 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
240 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
241 &dp->dp_free_dir);
242 if (err)
243 goto out;
244
245 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
246 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
247 if (err)
248 goto out;
249 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
250 dp->dp_meta_objset, obj));
251 }
252
253 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
254 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
255 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
256 &dp->dp_bptree_obj);
257 if (err != 0)
258 goto out;
259 }
260
261 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
262 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
263 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
264 &dp->dp_empty_bpobj);
265 if (err != 0)
266 goto out;
267 }
268
269 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
270 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
271 &dp->dp_tmp_userrefs_obj);
272 if (err == ENOENT)
273 err = 0;
274 if (err)
275 goto out;
276
277 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
278
279 out:
280 rrw_exit(&dp->dp_config_rwlock, FTAG);
281 return (err);
282 }
283
284 void
285 dsl_pool_close(dsl_pool_t *dp)
286 {
287 /*
288 * Drop our references from dsl_pool_open().
289 *
290 * Since we held the origin_snap from "syncing" context (which
291 * includes pool-opening context), it actually only got a "ref"
292 * and not a hold, so just drop that here.
293 */
294 if (dp->dp_origin_snap)
295 dsl_dataset_rele(dp->dp_origin_snap, dp);
296 if (dp->dp_mos_dir)
297 dsl_dir_rele(dp->dp_mos_dir, dp);
298 if (dp->dp_free_dir)
299 dsl_dir_rele(dp->dp_free_dir, dp);
300 if (dp->dp_root_dir)
301 dsl_dir_rele(dp->dp_root_dir, dp);
302
303 bpobj_close(&dp->dp_free_bpobj);
304
305 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
306 if (dp->dp_meta_objset)
307 dmu_objset_evict(dp->dp_meta_objset);
308
309 txg_list_destroy(&dp->dp_dirty_datasets);
310 txg_list_destroy(&dp->dp_dirty_zilogs);
311 txg_list_destroy(&dp->dp_sync_tasks);
312 txg_list_destroy(&dp->dp_dirty_dirs);
313
314 arc_flush(dp->dp_spa);
315 txg_fini(dp);
316 dsl_scan_fini(dp);
317 rrw_destroy(&dp->dp_config_rwlock);
318 mutex_destroy(&dp->dp_lock);
319 taskq_destroy(dp->dp_vnrele_taskq);
320 if (dp->dp_blkstats)
321 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
322 kmem_free(dp, sizeof (dsl_pool_t));
323 }
324
325 dsl_pool_t *
326 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
327 {
328 int err;
329 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
330 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
331 objset_t *os;
332 dsl_dataset_t *ds;
333 uint64_t obj;
334
335 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
336
337 /* create and open the MOS (meta-objset) */
338 dp->dp_meta_objset = dmu_objset_create_impl(spa,
339 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
340
341 /* create the pool directory */
342 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
343 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
344 ASSERT0(err);
345
346 /* Initialize scan structures */
347 VERIFY0(dsl_scan_init(dp, txg));
348
349 /* create and open the root dir */
350 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
351 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
352 NULL, dp, &dp->dp_root_dir));
353
354 /* create and open the meta-objset dir */
355 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
356 VERIFY0(dsl_pool_open_special_dir(dp,
357 MOS_DIR_NAME, &dp->dp_mos_dir));
358
359 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
360 /* create and open the free dir */
361 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
362 FREE_DIR_NAME, tx);
363 VERIFY0(dsl_pool_open_special_dir(dp,
364 FREE_DIR_NAME, &dp->dp_free_dir));
365
366 /* create and open the free_bplist */
367 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
368 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
369 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
370 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
371 dp->dp_meta_objset, obj));
372 }
373
374 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
375 dsl_pool_create_origin(dp, tx);
376
377 /* create the root dataset */
378 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
379
380 /* create the root objset */
381 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
382 os = dmu_objset_create_impl(dp->dp_spa, ds,
383 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
384 #ifdef _KERNEL
385 zfs_create_fs(os, kcred, zplprops, tx);
386 #endif
387 dsl_dataset_rele(ds, FTAG);
388
389 dmu_tx_commit(tx);
390
391 rrw_exit(&dp->dp_config_rwlock, FTAG);
392
393 return (dp);
394 }
395
396 /*
397 * Account for the meta-objset space in its placeholder dsl_dir.
398 */
399 void
400 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
401 int64_t used, int64_t comp, int64_t uncomp)
402 {
403 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
404 mutex_enter(&dp->dp_lock);
405 dp->dp_mos_used_delta += used;
406 dp->dp_mos_compressed_delta += comp;
407 dp->dp_mos_uncompressed_delta += uncomp;
408 mutex_exit(&dp->dp_lock);
409 }
410
411 static int
412 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
413 {
414 dsl_deadlist_t *dl = arg;
415 dsl_deadlist_insert(dl, bp, tx);
416 return (0);
417 }
418
419 static void
420 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
421 {
422 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
423 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
424 VERIFY0(zio_wait(zio));
425 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
426 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
427 }
428
429 static void
430 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
431 {
432 ASSERT(MUTEX_HELD(&dp->dp_lock));
433
434 if (delta < 0)
435 ASSERT3U(-delta, <=, dp->dp_dirty_total);
436
437 dp->dp_dirty_total += delta;
438
439 /*
440 * Note: we signal even when increasing dp_dirty_total.
441 * This ensures forward progress -- each thread wakes the next waiter.
442 */
443 if (dp->dp_dirty_total <= zfs_dirty_data_max)
444 cv_signal(&dp->dp_spaceavail_cv);
445 }
446
447 void
448 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
449 {
450 zio_t *zio;
451 dmu_tx_t *tx;
452 dsl_dir_t *dd;
453 dsl_dataset_t *ds;
454 objset_t *mos = dp->dp_meta_objset;
455 list_t synced_datasets;
456
457 list_create(&synced_datasets, sizeof (dsl_dataset_t),
458 offsetof(dsl_dataset_t, ds_synced_link));
459
460 tx = dmu_tx_create_assigned(dp, txg);
461
462 /*
463 * Write out all dirty blocks of dirty datasets.
464 */
465 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
466 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
467 /*
468 * We must not sync any non-MOS datasets twice, because
469 * we may have taken a snapshot of them. However, we
470 * may sync newly-created datasets on pass 2.
471 */
472 ASSERT(!list_link_active(&ds->ds_synced_link));
473 list_insert_tail(&synced_datasets, ds);
474 dsl_dataset_sync(ds, zio, tx);
475 }
476 VERIFY0(zio_wait(zio));
477
478 /*
479 * We have written all of the accounted dirty data, so our
480 * dp_space_towrite should now be zero. However, some seldom-used
481 * code paths do not adhere to this (e.g. dbuf_undirty(), also
482 * rounding error in dbuf_write_physdone).
483 * Shore up the accounting of any dirtied space now.
484 */
485 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
486
487 /*
488 * After the data blocks have been written (ensured by the zio_wait()
489 * above), update the user/group space accounting.
490 */
491 for (ds = list_head(&synced_datasets); ds != NULL;
492 ds = list_next(&synced_datasets, ds)) {
493 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
494 }
495
496 /*
497 * Sync the datasets again to push out the changes due to
498 * userspace updates. This must be done before we process the
499 * sync tasks, so that any snapshots will have the correct
500 * user accounting information (and we won't get confused
501 * about which blocks are part of the snapshot).
502 */
503 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
504 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
505 ASSERT(list_link_active(&ds->ds_synced_link));
506 dmu_buf_rele(ds->ds_dbuf, ds);
507 dsl_dataset_sync(ds, zio, tx);
508 }
509 VERIFY0(zio_wait(zio));
510
511 /*
512 * Now that the datasets have been completely synced, we can
513 * clean up our in-memory structures accumulated while syncing:
514 *
515 * - move dead blocks from the pending deadlist to the on-disk deadlist
516 * - release hold from dsl_dataset_dirty()
517 */
518 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
519 objset_t *os = ds->ds_objset;
520 bplist_iterate(&ds->ds_pending_deadlist,
521 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
522 ASSERT(!dmu_objset_is_dirty(os, txg));
523 dmu_buf_rele(ds->ds_dbuf, ds);
524 }
525 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
526 dsl_dir_sync(dd, tx);
527 }
528
529 /*
530 * The MOS's space is accounted for in the pool/$MOS
531 * (dp_mos_dir). We can't modify the mos while we're syncing
532 * it, so we remember the deltas and apply them here.
533 */
534 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
535 dp->dp_mos_uncompressed_delta != 0) {
536 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
537 dp->dp_mos_used_delta,
538 dp->dp_mos_compressed_delta,
539 dp->dp_mos_uncompressed_delta, tx);
540 dp->dp_mos_used_delta = 0;
541 dp->dp_mos_compressed_delta = 0;
542 dp->dp_mos_uncompressed_delta = 0;
543 }
544
545 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
546 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
547 dsl_pool_sync_mos(dp, tx);
548 }
549
550 /*
551 * If we modify a dataset in the same txg that we want to destroy it,
552 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
553 * dsl_dir_destroy_check() will fail if there are unexpected holds.
554 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
555 * and clearing the hold on it) before we process the sync_tasks.
556 * The MOS data dirtied by the sync_tasks will be synced on the next
557 * pass.
558 */
559 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
560 dsl_sync_task_t *dst;
561 /*
562 * No more sync tasks should have been added while we
563 * were syncing.
564 */
565 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
566 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
567 dsl_sync_task_sync(dst, tx);
568 }
569
570 dmu_tx_commit(tx);
571
572 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
573 }
574
575 void
576 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
577 {
578 zilog_t *zilog;
579
580 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
581 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
582 zil_clean(zilog, txg);
583 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
584 dmu_buf_rele(ds->ds_dbuf, zilog);
585 }
586 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
587 }
588
589 /*
590 * TRUE if the current thread is the tx_sync_thread or if we
591 * are being called from SPA context during pool initialization.
592 */
593 int
594 dsl_pool_sync_context(dsl_pool_t *dp)
595 {
596 return (curthread == dp->dp_tx.tx_sync_thread ||
597 spa_is_initializing(dp->dp_spa));
598 }
599
600 uint64_t
601 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
602 {
603 uint64_t space, resv;
604
605 /*
606 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
607 * efficiency.
608 * XXX The intent log is not accounted for, so it must fit
609 * within this slop.
610 *
611 * If we're trying to assess whether it's OK to do a free,
612 * cut the reservation in half to allow forward progress
613 * (e.g. make it possible to rm(1) files from a full pool).
614 */
615 space = spa_get_dspace(dp->dp_spa);
616 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
617 if (netfree)
618 resv >>= 1;
619
620 return (space - resv);
621 }
622
623 boolean_t
624 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
625 {
626 uint64_t delay_min_bytes =
627 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
628 boolean_t rv;
629
630 mutex_enter(&dp->dp_lock);
631 if (dp->dp_dirty_total > zfs_dirty_data_sync)
632 txg_kick(dp);
633 rv = (dp->dp_dirty_total > delay_min_bytes);
634 mutex_exit(&dp->dp_lock);
635 return (rv);
636 }
637
638 void
639 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
640 {
641 if (space > 0) {
642 mutex_enter(&dp->dp_lock);
643 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
644 dsl_pool_dirty_delta(dp, space);
645 mutex_exit(&dp->dp_lock);
646 }
647 }
648
649 void
650 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
651 {
652 ASSERT3S(space, >=, 0);
653 if (space == 0)
654 return;
655 mutex_enter(&dp->dp_lock);
656 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
657 /* XXX writing something we didn't dirty? */
658 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
659 }
660 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
661 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
662 ASSERT3U(dp->dp_dirty_total, >=, space);
663 dsl_pool_dirty_delta(dp, -space);
664 mutex_exit(&dp->dp_lock);
665 }
666
667 /* ARGSUSED */
668 static int
669 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
670 {
671 dmu_tx_t *tx = arg;
672 dsl_dataset_t *ds, *prev = NULL;
673 int err;
674
675 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
676 if (err)
677 return (err);
678
679 while (ds->ds_phys->ds_prev_snap_obj != 0) {
680 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
681 FTAG, &prev);
682 if (err) {
683 dsl_dataset_rele(ds, FTAG);
684 return (err);
685 }
686
687 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
688 break;
689 dsl_dataset_rele(ds, FTAG);
690 ds = prev;
691 prev = NULL;
692 }
693
694 if (prev == NULL) {
695 prev = dp->dp_origin_snap;
696
697 /*
698 * The $ORIGIN can't have any data, or the accounting
699 * will be wrong.
700 */
701 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
702
703 /* The origin doesn't get attached to itself */
704 if (ds->ds_object == prev->ds_object) {
705 dsl_dataset_rele(ds, FTAG);
706 return (0);
707 }
708
709 dmu_buf_will_dirty(ds->ds_dbuf, tx);
710 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
711 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
712
713 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
714 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
715
716 dmu_buf_will_dirty(prev->ds_dbuf, tx);
717 prev->ds_phys->ds_num_children++;
718
719 if (ds->ds_phys->ds_next_snap_obj == 0) {
720 ASSERT(ds->ds_prev == NULL);
721 VERIFY0(dsl_dataset_hold_obj(dp,
722 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
723 }
724 }
725
726 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
727 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
728
729 if (prev->ds_phys->ds_next_clones_obj == 0) {
730 dmu_buf_will_dirty(prev->ds_dbuf, tx);
731 prev->ds_phys->ds_next_clones_obj =
732 zap_create(dp->dp_meta_objset,
733 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
734 }
735 VERIFY0(zap_add_int(dp->dp_meta_objset,
736 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
737
738 dsl_dataset_rele(ds, FTAG);
739 if (prev != dp->dp_origin_snap)
740 dsl_dataset_rele(prev, FTAG);
741 return (0);
742 }
743
744 void
745 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
746 {
747 ASSERT(dmu_tx_is_syncing(tx));
748 ASSERT(dp->dp_origin_snap != NULL);
749
750 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
751 tx, DS_FIND_CHILDREN));
752 }
753
754 /* ARGSUSED */
755 static int
756 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
757 {
758 dmu_tx_t *tx = arg;
759 objset_t *mos = dp->dp_meta_objset;
760
761 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
762 dsl_dataset_t *origin;
763
764 VERIFY0(dsl_dataset_hold_obj(dp,
765 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
766
767 if (origin->ds_dir->dd_phys->dd_clones == 0) {
768 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
769 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
770 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
771 }
772
773 VERIFY0(zap_add_int(dp->dp_meta_objset,
774 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
775
776 dsl_dataset_rele(origin, FTAG);
777 }
778 return (0);
779 }
780
781 void
782 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
783 {
784 ASSERT(dmu_tx_is_syncing(tx));
785 uint64_t obj;
786
787 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
788 VERIFY0(dsl_pool_open_special_dir(dp,
789 FREE_DIR_NAME, &dp->dp_free_dir));
790
791 /*
792 * We can't use bpobj_alloc(), because spa_version() still
793 * returns the old version, and we need a new-version bpobj with
794 * subobj support. So call dmu_object_alloc() directly.
795 */
796 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
797 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
798 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
799 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
800 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
801
802 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
803 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
804 }
805
806 void
807 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
808 {
809 uint64_t dsobj;
810 dsl_dataset_t *ds;
811
812 ASSERT(dmu_tx_is_syncing(tx));
813 ASSERT(dp->dp_origin_snap == NULL);
814 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
815
816 /* create the origin dir, ds, & snap-ds */
817 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
818 NULL, 0, kcred, tx);
819 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
820 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
821 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
822 dp, &dp->dp_origin_snap));
823 dsl_dataset_rele(ds, FTAG);
824 }
825
826 taskq_t *
827 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
828 {
829 return (dp->dp_vnrele_taskq);
830 }
831
832 /*
833 * Walk through the pool-wide zap object of temporary snapshot user holds
834 * and release them.
835 */
836 void
837 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
838 {
839 zap_attribute_t za;
840 zap_cursor_t zc;
841 objset_t *mos = dp->dp_meta_objset;
842 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
843 nvlist_t *holds;
844
845 if (zapobj == 0)
846 return;
847 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
848
849 holds = fnvlist_alloc();
850
851 for (zap_cursor_init(&zc, mos, zapobj);
852 zap_cursor_retrieve(&zc, &za) == 0;
853 zap_cursor_advance(&zc)) {
854 char *htag;
855 nvlist_t *tags;
856
857 htag = strchr(za.za_name, '-');
858 *htag = '\0';
859 ++htag;
860 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
861 tags = fnvlist_alloc();
862 fnvlist_add_boolean(tags, htag);
863 fnvlist_add_nvlist(holds, za.za_name, tags);
864 fnvlist_free(tags);
865 } else {
866 fnvlist_add_boolean(tags, htag);
867 }
868 }
869 dsl_dataset_user_release_tmp(dp, holds);
870 fnvlist_free(holds);
871 zap_cursor_fini(&zc);
872 }
873
874 /*
875 * Create the pool-wide zap object for storing temporary snapshot holds.
876 */
877 void
878 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
879 {
880 objset_t *mos = dp->dp_meta_objset;
881
882 ASSERT(dp->dp_tmp_userrefs_obj == 0);
883 ASSERT(dmu_tx_is_syncing(tx));
884
885 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
886 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
887 }
888
889 static int
890 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
891 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
892 {
893 objset_t *mos = dp->dp_meta_objset;
894 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
895 char *name;
896 int error;
897
898 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
899 ASSERT(dmu_tx_is_syncing(tx));
900
901 /*
902 * If the pool was created prior to SPA_VERSION_USERREFS, the
903 * zap object for temporary holds might not exist yet.
904 */
905 if (zapobj == 0) {
906 if (holding) {
907 dsl_pool_user_hold_create_obj(dp, tx);
908 zapobj = dp->dp_tmp_userrefs_obj;
909 } else {
910 return (SET_ERROR(ENOENT));
911 }
912 }
913
914 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
915 if (holding)
916 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
917 else
918 error = zap_remove(mos, zapobj, name, tx);
919 strfree(name);
920
921 return (error);
922 }
923
924 /*
925 * Add a temporary hold for the given dataset object and tag.
926 */
927 int
928 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
929 uint64_t now, dmu_tx_t *tx)
930 {
931 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
932 }
933
934 /*
935 * Release a temporary hold for the given dataset object and tag.
936 */
937 int
938 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
939 dmu_tx_t *tx)
940 {
941 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
942 tx, B_FALSE));
943 }
944
945 /*
946 * DSL Pool Configuration Lock
947 *
948 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
949 * creation / destruction / rename / property setting). It must be held for
950 * read to hold a dataset or dsl_dir. I.e. you must call
951 * dsl_pool_config_enter() or dsl_pool_hold() before calling
952 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
953 * must be held continuously until all datasets and dsl_dirs are released.
954 *
955 * The only exception to this rule is that if a "long hold" is placed on
956 * a dataset, then the dp_config_rwlock may be dropped while the dataset
957 * is still held. The long hold will prevent the dataset from being
958 * destroyed -- the destroy will fail with EBUSY. A long hold can be
959 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
960 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
961 *
962 * Legitimate long-holders (including owners) should be long-running, cancelable
963 * tasks that should cause "zfs destroy" to fail. This includes DMU
964 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
965 * "zfs send", and "zfs diff". There are several other long-holders whose
966 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
967 *
968 * The usual formula for long-holding would be:
969 * dsl_pool_hold()
970 * dsl_dataset_hold()
971 * ... perform checks ...
972 * dsl_dataset_long_hold()
973 * dsl_pool_rele()
974 * ... perform long-running task ...
975 * dsl_dataset_long_rele()
976 * dsl_dataset_rele()
977 *
978 * Note that when the long hold is released, the dataset is still held but
979 * the pool is not held. The dataset may change arbitrarily during this time
980 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
981 * dataset except release it.
982 *
983 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
984 * or modifying operations.
985 *
986 * Modifying operations should generally use dsl_sync_task(). The synctask
987 * infrastructure enforces proper locking strategy with respect to the
988 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
989 *
990 * Read-only operations will manually hold the pool, then the dataset, obtain
991 * information from the dataset, then release the pool and dataset.
992 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
993 * hold/rele.
994 */
995
996 int
997 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
998 {
999 spa_t *spa;
1000 int error;
1001
1002 error = spa_open(name, &spa, tag);
1003 if (error == 0) {
1004 *dp = spa_get_dsl(spa);
1005 dsl_pool_config_enter(*dp, tag);
1006 }
1007 return (error);
1008 }
1009
1010 void
1011 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1012 {
1013 dsl_pool_config_exit(dp, tag);
1014 spa_close(dp->dp_spa, tag);
1015 }
1016
1017 void
1018 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1019 {
1020 /*
1021 * We use a "reentrant" reader-writer lock, but not reentrantly.
1022 *
1023 * The rrwlock can (with the track_all flag) track all reading threads,
1024 * which is very useful for debugging which code path failed to release
1025 * the lock, and for verifying that the *current* thread does hold
1026 * the lock.
1027 *
1028 * (Unlike a rwlock, which knows that N threads hold it for
1029 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1030 * if any thread holds it for read, even if this thread doesn't).
1031 */
1032 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1033 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1034 }
1035
1036 void
1037 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1038 {
1039 rrw_exit(&dp->dp_config_rwlock, tag);
1040 }
1041
1042 boolean_t
1043 dsl_pool_config_held(dsl_pool_t *dp)
1044 {
1045 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1046 }