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 */
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 #include <sys/dsl_userhold.h>
47
48 int zfs_no_write_throttle = 0;
49 int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
50 int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
51
52 uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
53 uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
54 uint64_t zfs_write_limit_inflated = 0;
55 uint64_t zfs_write_limit_override = 0;
56
57 kmutex_t zfs_write_limit_lock;
58
59 static pgcnt_t old_physmem = 0;
60
61 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
62 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
63
64 int
65 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
66 {
67 uint64_t obj;
68 int err;
69
70 err = zap_lookup(dp->dp_meta_objset,
71 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
72 name, sizeof (obj), 1, &obj);
73 if (err)
74 return (err);
75
76 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
77 }
78
79 static dsl_pool_t *
80 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
81 {
82 dsl_pool_t *dp;
83 blkptr_t *bp = spa_get_rootblkptr(spa);
84
85 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
86 dp->dp_spa = spa;
87 dp->dp_meta_rootbp = *bp;
88 rrw_init(&dp->dp_config_rwlock, B_TRUE);
89 dp->dp_write_limit = zfs_write_limit_min;
90 txg_init(dp, txg);
91
92 txg_list_create(&dp->dp_dirty_datasets,
93 offsetof(dsl_dataset_t, ds_dirty_link));
94 txg_list_create(&dp->dp_dirty_zilogs,
95 offsetof(zilog_t, zl_dirty_link));
96 txg_list_create(&dp->dp_dirty_dirs,
97 offsetof(dsl_dir_t, dd_dirty_link));
98 txg_list_create(&dp->dp_sync_tasks,
99 offsetof(dsl_sync_task_t, dst_node));
100
101 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
102
103 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
104 1, 4, 0);
105
106 return (dp);
107 }
108
109 int
110 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
111 {
112 int err;
113 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
114
115 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
116 &dp->dp_meta_objset);
117 if (err != 0)
118 dsl_pool_close(dp);
119 else
120 *dpp = dp;
121
122 return (err);
123 }
124
125 int
126 dsl_pool_open(dsl_pool_t *dp)
127 {
128 int err;
129 dsl_dir_t *dd;
130 dsl_dataset_t *ds;
131 uint64_t obj;
132
133 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
134 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
135 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
136 &dp->dp_root_dir_obj);
137 if (err)
138 goto out;
139
140 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
141 NULL, dp, &dp->dp_root_dir);
142 if (err)
143 goto out;
144
145 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
146 if (err)
147 goto out;
148
149 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
150 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
151 if (err)
152 goto out;
153 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
154 FTAG, &ds);
155 if (err == 0) {
156 err = dsl_dataset_hold_obj(dp,
157 ds->ds_phys->ds_prev_snap_obj, dp,
158 &dp->dp_origin_snap);
159 dsl_dataset_rele(ds, FTAG);
160 }
161 dsl_dir_rele(dd, dp);
162 if (err)
163 goto out;
164 }
165
166 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
167 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
168 &dp->dp_free_dir);
169 if (err)
170 goto out;
171
172 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
173 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
174 if (err)
175 goto out;
176 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
177 dp->dp_meta_objset, obj));
178 }
179
180 if (spa_feature_is_active(dp->dp_spa,
181 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
182 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
183 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
184 &dp->dp_bptree_obj);
185 if (err != 0)
186 goto out;
187 }
188
189 if (spa_feature_is_active(dp->dp_spa,
190 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
191 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
192 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
193 &dp->dp_empty_bpobj);
194 if (err != 0)
195 goto out;
196 }
197
198 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
199 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
200 &dp->dp_tmp_userrefs_obj);
201 if (err == ENOENT)
202 err = 0;
203 if (err)
204 goto out;
205
206 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
207
208 out:
209 rrw_exit(&dp->dp_config_rwlock, FTAG);
210 return (err);
211 }
212
213 void
214 dsl_pool_close(dsl_pool_t *dp)
215 {
216 /* drop our references from dsl_pool_open() */
217
218 /*
219 * Since we held the origin_snap from "syncing" context (which
220 * includes pool-opening context), it actually only got a "ref"
221 * and not a hold, so just drop that here.
222 */
223 if (dp->dp_origin_snap)
224 dsl_dataset_rele(dp->dp_origin_snap, dp);
225 if (dp->dp_mos_dir)
226 dsl_dir_rele(dp->dp_mos_dir, dp);
227 if (dp->dp_free_dir)
228 dsl_dir_rele(dp->dp_free_dir, dp);
229 if (dp->dp_root_dir)
230 dsl_dir_rele(dp->dp_root_dir, dp);
231
232 bpobj_close(&dp->dp_free_bpobj);
233
234 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
235 if (dp->dp_meta_objset)
236 dmu_objset_evict(dp->dp_meta_objset);
237
238 txg_list_destroy(&dp->dp_dirty_datasets);
239 txg_list_destroy(&dp->dp_dirty_zilogs);
240 txg_list_destroy(&dp->dp_sync_tasks);
241 txg_list_destroy(&dp->dp_dirty_dirs);
242
243 arc_flush(dp->dp_spa);
244 txg_fini(dp);
245 dsl_scan_fini(dp);
246 rrw_destroy(&dp->dp_config_rwlock);
247 mutex_destroy(&dp->dp_lock);
248 taskq_destroy(dp->dp_vnrele_taskq);
249 if (dp->dp_blkstats)
250 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
251 kmem_free(dp, sizeof (dsl_pool_t));
252 }
253
254 dsl_pool_t *
255 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
256 {
257 int err;
258 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
259 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
260 objset_t *os;
261 dsl_dataset_t *ds;
262 uint64_t obj;
263
264 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
265
266 /* create and open the MOS (meta-objset) */
267 dp->dp_meta_objset = dmu_objset_create_impl(spa,
268 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
269
270 /* create the pool directory */
271 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
272 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
273 ASSERT0(err);
274
275 /* Initialize scan structures */
276 VERIFY0(dsl_scan_init(dp, txg));
277
278 /* create and open the root dir */
279 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
280 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
281 NULL, dp, &dp->dp_root_dir));
282
283 /* create and open the meta-objset dir */
284 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
285 VERIFY0(dsl_pool_open_special_dir(dp,
286 MOS_DIR_NAME, &dp->dp_mos_dir));
287
288 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
289 /* create and open the free dir */
290 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
291 FREE_DIR_NAME, tx);
292 VERIFY0(dsl_pool_open_special_dir(dp,
293 FREE_DIR_NAME, &dp->dp_free_dir));
294
295 /* create and open the free_bplist */
296 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
297 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
298 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
299 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
300 dp->dp_meta_objset, obj));
301 }
302
303 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
304 dsl_pool_create_origin(dp, tx);
305
306 /* create the root dataset */
307 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
308
309 /* create the root objset */
310 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
311 os = dmu_objset_create_impl(dp->dp_spa, ds,
312 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
313 #ifdef _KERNEL
314 zfs_create_fs(os, kcred, zplprops, tx);
315 #endif
316 dsl_dataset_rele(ds, FTAG);
317
318 dmu_tx_commit(tx);
319
320 rrw_exit(&dp->dp_config_rwlock, FTAG);
321
322 return (dp);
323 }
324
325 /*
326 * Account for the meta-objset space in its placeholder dsl_dir.
327 */
328 void
329 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
330 int64_t used, int64_t comp, int64_t uncomp)
331 {
332 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
333 mutex_enter(&dp->dp_lock);
334 dp->dp_mos_used_delta += used;
335 dp->dp_mos_compressed_delta += comp;
336 dp->dp_mos_uncompressed_delta += uncomp;
337 mutex_exit(&dp->dp_lock);
338 }
339
340 static int
341 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
342 {
343 dsl_deadlist_t *dl = arg;
344 dsl_deadlist_insert(dl, bp, tx);
345 return (0);
346 }
347
348 void
349 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
350 {
351 zio_t *zio;
352 dmu_tx_t *tx;
353 dsl_dir_t *dd;
354 dsl_dataset_t *ds;
355 objset_t *mos = dp->dp_meta_objset;
356 hrtime_t start, write_time;
357 uint64_t data_written;
358 int err;
359 list_t synced_datasets;
360
361 list_create(&synced_datasets, sizeof (dsl_dataset_t),
362 offsetof(dsl_dataset_t, ds_synced_link));
363
364 /*
365 * We need to copy dp_space_towrite() before doing
366 * dsl_sync_task_sync(), because
367 * dsl_dataset_snapshot_reserve_space() will increase
368 * dp_space_towrite but not actually write anything.
369 */
370 data_written = dp->dp_space_towrite[txg & TXG_MASK];
371
372 tx = dmu_tx_create_assigned(dp, txg);
373
374 dp->dp_read_overhead = 0;
375 start = gethrtime();
376
377 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
378 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
379 /*
380 * We must not sync any non-MOS datasets twice, because
381 * we may have taken a snapshot of them. However, we
382 * may sync newly-created datasets on pass 2.
383 */
384 ASSERT(!list_link_active(&ds->ds_synced_link));
385 list_insert_tail(&synced_datasets, ds);
386 dsl_dataset_sync(ds, zio, tx);
387 }
388 DTRACE_PROBE(pool_sync__1setup);
389 err = zio_wait(zio);
390
391 write_time = gethrtime() - start;
392 ASSERT(err == 0);
393 DTRACE_PROBE(pool_sync__2rootzio);
394
395 /*
396 * After the data blocks have been written (ensured by the zio_wait()
397 * above), update the user/group space accounting.
398 */
399 for (ds = list_head(&synced_datasets); ds;
400 ds = list_next(&synced_datasets, ds))
401 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
402
403 /*
404 * Sync the datasets again to push out the changes due to
405 * userspace updates. This must be done before we process the
406 * sync tasks, so that any snapshots will have the correct
407 * user accounting information (and we won't get confused
408 * about which blocks are part of the snapshot).
409 */
410 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
411 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
412 ASSERT(list_link_active(&ds->ds_synced_link));
413 dmu_buf_rele(ds->ds_dbuf, ds);
414 dsl_dataset_sync(ds, zio, tx);
415 }
416 err = zio_wait(zio);
417
418 /*
419 * Now that the datasets have been completely synced, we can
420 * clean up our in-memory structures accumulated while syncing:
421 *
422 * - move dead blocks from the pending deadlist to the on-disk deadlist
423 * - release hold from dsl_dataset_dirty()
424 */
425 while (ds = list_remove_head(&synced_datasets)) {
426 objset_t *os = ds->ds_objset;
427 bplist_iterate(&ds->ds_pending_deadlist,
428 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
429 ASSERT(!dmu_objset_is_dirty(os, txg));
430 dmu_buf_rele(ds->ds_dbuf, ds);
431 }
432
433 start = gethrtime();
434 while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
435 dsl_dir_sync(dd, tx);
436 write_time += gethrtime() - start;
437
438 /*
439 * The MOS's space is accounted for in the pool/$MOS
440 * (dp_mos_dir). We can't modify the mos while we're syncing
441 * it, so we remember the deltas and apply them here.
442 */
443 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
444 dp->dp_mos_uncompressed_delta != 0) {
445 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
446 dp->dp_mos_used_delta,
447 dp->dp_mos_compressed_delta,
448 dp->dp_mos_uncompressed_delta, tx);
449 dp->dp_mos_used_delta = 0;
450 dp->dp_mos_compressed_delta = 0;
451 dp->dp_mos_uncompressed_delta = 0;
452 }
453
454 start = gethrtime();
455 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
456 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
457 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
458 dmu_objset_sync(mos, zio, tx);
459 err = zio_wait(zio);
460 ASSERT(err == 0);
461 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
462 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
463 }
464 write_time += gethrtime() - start;
465 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
466 hrtime_t, dp->dp_read_overhead);
467 write_time -= dp->dp_read_overhead;
468
469 /*
470 * If we modify a dataset in the same txg that we want to destroy it,
471 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
472 * dsl_dir_destroy_check() will fail if there are unexpected holds.
473 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
474 * and clearing the hold on it) before we process the sync_tasks.
475 * The MOS data dirtied by the sync_tasks will be synced on the next
476 * pass.
477 */
478 DTRACE_PROBE(pool_sync__3task);
479 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
480 dsl_sync_task_t *dst;
481 /*
482 * No more sync tasks should have been added while we
483 * were syncing.
484 */
485 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
486 while (dst = txg_list_remove(&dp->dp_sync_tasks, txg))
487 dsl_sync_task_sync(dst, tx);
488 }
489
490 dmu_tx_commit(tx);
491
492 dp->dp_space_towrite[txg & TXG_MASK] = 0;
493 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
494
495 /*
496 * If the write limit max has not been explicitly set, set it
497 * to a fraction of available physical memory (default 1/8th).
498 * Note that we must inflate the limit because the spa
499 * inflates write sizes to account for data replication.
500 * Check this each sync phase to catch changing memory size.
501 */
502 if (physmem != old_physmem && zfs_write_limit_shift) {
503 mutex_enter(&zfs_write_limit_lock);
504 old_physmem = physmem;
505 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
506 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
507 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
508 mutex_exit(&zfs_write_limit_lock);
509 }
510
511 /*
512 * Attempt to keep the sync time consistent by adjusting the
513 * amount of write traffic allowed into each transaction group.
514 * Weight the throughput calculation towards the current value:
515 * thru = 3/4 old_thru + 1/4 new_thru
516 *
517 * Note: write_time is in nanosecs while dp_throughput is expressed in
518 * bytes per millisecond.
519 */
520 ASSERT(zfs_write_limit_min > 0);
521 if (data_written > zfs_write_limit_min / 8 &&
522 write_time > MSEC2NSEC(1)) {
523 uint64_t throughput = data_written / NSEC2MSEC(write_time);
524
525 if (dp->dp_throughput)
526 dp->dp_throughput = throughput / 4 +
527 3 * dp->dp_throughput / 4;
528 else
529 dp->dp_throughput = throughput;
530 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
531 MAX(zfs_write_limit_min,
532 dp->dp_throughput * zfs_txg_synctime_ms));
533 }
534 }
535
536 void
537 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
538 {
539 zilog_t *zilog;
540 dsl_dataset_t *ds;
541
542 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
543 ds = dmu_objset_ds(zilog->zl_os);
544 zil_clean(zilog, txg);
545 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
546 dmu_buf_rele(ds->ds_dbuf, zilog);
547 }
548 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
549 }
550
551 /*
552 * TRUE if the current thread is the tx_sync_thread or if we
553 * are being called from SPA context during pool initialization.
554 */
555 int
556 dsl_pool_sync_context(dsl_pool_t *dp)
557 {
558 return (curthread == dp->dp_tx.tx_sync_thread ||
559 spa_is_initializing(dp->dp_spa));
560 }
561
562 uint64_t
563 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
564 {
565 uint64_t space, resv;
566
567 /*
568 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
569 * efficiency.
570 * XXX The intent log is not accounted for, so it must fit
571 * within this slop.
572 *
573 * If we're trying to assess whether it's OK to do a free,
574 * cut the reservation in half to allow forward progress
575 * (e.g. make it possible to rm(1) files from a full pool).
576 */
577 space = spa_get_dspace(dp->dp_spa);
578 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
579 if (netfree)
580 resv >>= 1;
581
582 return (space - resv);
583 }
584
585 int
586 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
587 {
588 uint64_t reserved = 0;
589 uint64_t write_limit = (zfs_write_limit_override ?
590 zfs_write_limit_override : dp->dp_write_limit);
591
592 if (zfs_no_write_throttle) {
593 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
594 space);
595 return (0);
596 }
597
598 /*
599 * Check to see if we have exceeded the maximum allowed IO for
600 * this transaction group. We can do this without locks since
601 * a little slop here is ok. Note that we do the reserved check
602 * with only half the requested reserve: this is because the
603 * reserve requests are worst-case, and we really don't want to
604 * throttle based off of worst-case estimates.
605 */
606 if (write_limit > 0) {
607 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
608 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
609
610 if (reserved && reserved > write_limit)
611 return (SET_ERROR(ERESTART));
612 }
613
614 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
615
616 /*
617 * If this transaction group is over 7/8ths capacity, delay
618 * the caller 1 clock tick. This will slow down the "fill"
619 * rate until the sync process can catch up with us.
620 */
621 if (reserved && reserved > (write_limit - (write_limit >> 3))) {
622 txg_delay(dp, tx->tx_txg, zfs_throttle_delay,
623 zfs_throttle_resolution);
624 }
625
626 return (0);
627 }
628
629 void
630 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
631 {
632 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
633 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
634 }
635
636 void
637 dsl_pool_memory_pressure(dsl_pool_t *dp)
638 {
639 uint64_t space_inuse = 0;
640 int i;
641
642 if (dp->dp_write_limit == zfs_write_limit_min)
643 return;
644
645 for (i = 0; i < TXG_SIZE; i++) {
646 space_inuse += dp->dp_space_towrite[i];
647 space_inuse += dp->dp_tempreserved[i];
648 }
649 dp->dp_write_limit = MAX(zfs_write_limit_min,
650 MIN(dp->dp_write_limit, space_inuse / 4));
651 }
652
653 void
654 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
655 {
656 if (space > 0) {
657 mutex_enter(&dp->dp_lock);
658 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
659 mutex_exit(&dp->dp_lock);
660 }
661 }
662
663 /* ARGSUSED */
664 static int
665 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
666 {
667 dmu_tx_t *tx = arg;
668 dsl_dataset_t *ds, *prev = NULL;
669 int err;
670
671 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
672 if (err)
673 return (err);
674
675 while (ds->ds_phys->ds_prev_snap_obj != 0) {
676 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
677 FTAG, &prev);
678 if (err) {
679 dsl_dataset_rele(ds, FTAG);
680 return (err);
681 }
682
683 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
684 break;
685 dsl_dataset_rele(ds, FTAG);
686 ds = prev;
687 prev = NULL;
688 }
689
690 if (prev == NULL) {
691 prev = dp->dp_origin_snap;
692
693 /*
694 * The $ORIGIN can't have any data, or the accounting
695 * will be wrong.
696 */
697 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
698
699 /* The origin doesn't get attached to itself */
700 if (ds->ds_object == prev->ds_object) {
701 dsl_dataset_rele(ds, FTAG);
702 return (0);
703 }
704
705 dmu_buf_will_dirty(ds->ds_dbuf, tx);
706 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
707 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
708
709 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
710 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
711
712 dmu_buf_will_dirty(prev->ds_dbuf, tx);
713 prev->ds_phys->ds_num_children++;
714
715 if (ds->ds_phys->ds_next_snap_obj == 0) {
716 ASSERT(ds->ds_prev == NULL);
717 VERIFY0(dsl_dataset_hold_obj(dp,
718 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
719 }
720 }
721
722 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
723 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
724
725 if (prev->ds_phys->ds_next_clones_obj == 0) {
726 dmu_buf_will_dirty(prev->ds_dbuf, tx);
727 prev->ds_phys->ds_next_clones_obj =
728 zap_create(dp->dp_meta_objset,
729 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
730 }
731 VERIFY0(zap_add_int(dp->dp_meta_objset,
732 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
733
734 dsl_dataset_rele(ds, FTAG);
735 if (prev != dp->dp_origin_snap)
736 dsl_dataset_rele(prev, FTAG);
737 return (0);
738 }
739
740 void
741 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
742 {
743 ASSERT(dmu_tx_is_syncing(tx));
744 ASSERT(dp->dp_origin_snap != NULL);
745
746 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
747 tx, DS_FIND_CHILDREN));
748 }
749
750 /* ARGSUSED */
751 static int
752 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
753 {
754 dmu_tx_t *tx = arg;
755 objset_t *mos = dp->dp_meta_objset;
756
757 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
758 dsl_dataset_t *origin;
759
760 VERIFY0(dsl_dataset_hold_obj(dp,
761 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
762
763 if (origin->ds_dir->dd_phys->dd_clones == 0) {
764 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
765 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
766 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
767 }
768
769 VERIFY0(zap_add_int(dp->dp_meta_objset,
770 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
771
772 dsl_dataset_rele(origin, FTAG);
773 }
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 VERIFY0(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 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
795 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
796 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
797
798 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
799 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
800 }
801
802 void
803 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
804 {
805 uint64_t dsobj;
806 dsl_dataset_t *ds;
807
808 ASSERT(dmu_tx_is_syncing(tx));
809 ASSERT(dp->dp_origin_snap == NULL);
810 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
811
812 /* create the origin dir, ds, & snap-ds */
813 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
814 NULL, 0, kcred, tx);
815 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
816 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
817 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
818 dp, &dp->dp_origin_snap));
819 dsl_dataset_rele(ds, FTAG);
820 }
821
822 taskq_t *
823 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
824 {
825 return (dp->dp_vnrele_taskq);
826 }
827
828 /*
829 * Walk through the pool-wide zap object of temporary snapshot user holds
830 * and release them.
831 */
832 void
833 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
834 {
835 zap_attribute_t za;
836 zap_cursor_t zc;
837 objset_t *mos = dp->dp_meta_objset;
838 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
839
840 if (zapobj == 0)
841 return;
842 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
843
844 for (zap_cursor_init(&zc, mos, zapobj);
845 zap_cursor_retrieve(&zc, &za) == 0;
846 zap_cursor_advance(&zc)) {
847 char *htag;
848 uint64_t dsobj;
849
850 htag = strchr(za.za_name, '-');
851 *htag = '\0';
852 ++htag;
853 dsobj = strtonum(za.za_name, NULL);
854 dsl_dataset_user_release_tmp(dp, dsobj, htag);
855 }
856 zap_cursor_fini(&zc);
857 }
858
859 /*
860 * Create the pool-wide zap object for storing temporary snapshot holds.
861 */
862 void
863 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
864 {
865 objset_t *mos = dp->dp_meta_objset;
866
867 ASSERT(dp->dp_tmp_userrefs_obj == 0);
868 ASSERT(dmu_tx_is_syncing(tx));
869
870 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
871 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
872 }
873
874 static int
875 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
876 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
877 {
878 objset_t *mos = dp->dp_meta_objset;
879 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
880 char *name;
881 int error;
882
883 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
884 ASSERT(dmu_tx_is_syncing(tx));
885
886 /*
887 * If the pool was created prior to SPA_VERSION_USERREFS, the
888 * zap object for temporary holds might not exist yet.
889 */
890 if (zapobj == 0) {
891 if (holding) {
892 dsl_pool_user_hold_create_obj(dp, tx);
893 zapobj = dp->dp_tmp_userrefs_obj;
894 } else {
895 return (SET_ERROR(ENOENT));
896 }
897 }
898
899 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
900 if (holding)
901 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
902 else
903 error = zap_remove(mos, zapobj, name, tx);
904 strfree(name);
905
906 return (error);
907 }
908
909 /*
910 * Add a temporary hold for the given dataset object and tag.
911 */
912 int
913 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
914 uint64_t now, dmu_tx_t *tx)
915 {
916 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
917 }
918
919 /*
920 * Release a temporary hold for the given dataset object and tag.
921 */
922 int
923 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
924 dmu_tx_t *tx)
925 {
926 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
927 tx, B_FALSE));
928 }
929
930 /*
931 * DSL Pool Configuration Lock
932 *
933 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
934 * creation / destruction / rename / property setting). It must be held for
935 * read to hold a dataset or dsl_dir. I.e. you must call
936 * dsl_pool_config_enter() or dsl_pool_hold() before calling
937 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
938 * must be held continuously until all datasets and dsl_dirs are released.
939 *
940 * The only exception to this rule is that if a "long hold" is placed on
941 * a dataset, then the dp_config_rwlock may be dropped while the dataset
942 * is still held. The long hold will prevent the dataset from being
943 * destroyed -- the destroy will fail with EBUSY. A long hold can be
944 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
945 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
946 *
947 * Legitimate long-holders (including owners) should be long-running, cancelable
948 * tasks that should cause "zfs destroy" to fail. This includes DMU
949 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
950 * "zfs send", and "zfs diff". There are several other long-holders whose
951 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
952 *
953 * The usual formula for long-holding would be:
954 * dsl_pool_hold()
955 * dsl_dataset_hold()
956 * ... perform checks ...
957 * dsl_dataset_long_hold()
958 * dsl_pool_rele()
959 * ... perform long-running task ...
960 * dsl_dataset_long_rele()
961 * dsl_dataset_rele()
962 *
963 * Note that when the long hold is released, the dataset is still held but
964 * the pool is not held. The dataset may change arbitrarily during this time
965 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
966 * dataset except release it.
967 *
968 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
969 * or modifying operations.
970 *
971 * Modifying operations should generally use dsl_sync_task(). The synctask
972 * infrastructure enforces proper locking strategy with respect to the
973 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
974 *
975 * Read-only operations will manually hold the pool, then the dataset, obtain
976 * information from the dataset, then release the pool and dataset.
977 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
978 * hold/rele.
979 */
980
981 int
982 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
983 {
984 spa_t *spa;
985 int error;
986
987 error = spa_open(name, &spa, tag);
988 if (error == 0) {
989 *dp = spa_get_dsl(spa);
990 dsl_pool_config_enter(*dp, tag);
991 }
992 return (error);
993 }
994
995 void
996 dsl_pool_rele(dsl_pool_t *dp, void *tag)
997 {
998 dsl_pool_config_exit(dp, tag);
999 spa_close(dp->dp_spa, tag);
1000 }
1001
1002 void
1003 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1004 {
1005 /*
1006 * We use a "reentrant" reader-writer lock, but not reentrantly.
1007 *
1008 * The rrwlock can (with the track_all flag) track all reading threads,
1009 * which is very useful for debugging which code path failed to release
1010 * the lock, and for verifying that the *current* thread does hold
1011 * the lock.
1012 *
1013 * (Unlike a rwlock, which knows that N threads hold it for
1014 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1015 * if any thread holds it for read, even if this thread doesn't).
1016 */
1017 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1018 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1019 }
1020
1021 void
1022 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1023 {
1024 rrw_exit(&dp->dp_config_rwlock, tag);
1025 }
1026
1027 boolean_t
1028 dsl_pool_config_held(dsl_pool_t *dp)
1029 {
1030 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1031 }