Print this page
4045 zfs write throttle & i/o scheduler performance work
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/dsl_pool.c
+++ new/usr/src/uts/common/fs/zfs/dsl_pool.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 25 */
26 26
27 27 #include <sys/dsl_pool.h>
28 28 #include <sys/dsl_dataset.h>
29 29 #include <sys/dsl_prop.h>
30 30 #include <sys/dsl_dir.h>
31 31 #include <sys/dsl_synctask.h>
32 32 #include <sys/dsl_scan.h>
33 33 #include <sys/dnode.h>
34 34 #include <sys/dmu_tx.h>
35 35 #include <sys/dmu_objset.h>
36 36 #include <sys/arc.h>
37 37 #include <sys/zap.h>
38 38 #include <sys/zio.h>
|
↓ open down ↓ |
38 lines elided |
↑ open up ↑ |
39 39 #include <sys/zfs_context.h>
40 40 #include <sys/fs/zfs.h>
41 41 #include <sys/zfs_znode.h>
42 42 #include <sys/spa_impl.h>
43 43 #include <sys/dsl_deadlist.h>
44 44 #include <sys/bptree.h>
45 45 #include <sys/zfeature.h>
46 46 #include <sys/zil_impl.h>
47 47 #include <sys/dsl_userhold.h>
48 48
49 -int zfs_no_write_throttle = 0;
50 -int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
51 -int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
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 + */
52 91
53 -uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
54 -uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
55 -uint64_t zfs_write_limit_inflated = 0;
56 -uint64_t zfs_write_limit_override = 0;
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;
57 99
58 -kmutex_t zfs_write_limit_lock;
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;
59 104
60 -static pgcnt_t old_physmem = 0;
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;
61 111
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 +
62 134 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
63 135 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
64 136
65 137 int
66 138 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
67 139 {
68 140 uint64_t obj;
69 141 int err;
70 142
71 143 err = zap_lookup(dp->dp_meta_objset,
72 144 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
73 145 name, sizeof (obj), 1, &obj);
74 146 if (err)
75 147 return (err);
76 148
77 149 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
78 150 }
79 151
|
↓ open down ↓ |
8 lines elided |
↑ open up ↑ |
80 152 static dsl_pool_t *
81 153 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
82 154 {
83 155 dsl_pool_t *dp;
84 156 blkptr_t *bp = spa_get_rootblkptr(spa);
85 157
86 158 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
87 159 dp->dp_spa = spa;
88 160 dp->dp_meta_rootbp = *bp;
89 161 rrw_init(&dp->dp_config_rwlock, B_TRUE);
90 - dp->dp_write_limit = zfs_write_limit_min;
91 162 txg_init(dp, txg);
92 163
93 164 txg_list_create(&dp->dp_dirty_datasets,
94 165 offsetof(dsl_dataset_t, ds_dirty_link));
95 166 txg_list_create(&dp->dp_dirty_zilogs,
96 167 offsetof(zilog_t, zl_dirty_link));
97 168 txg_list_create(&dp->dp_dirty_dirs,
98 169 offsetof(dsl_dir_t, dd_dirty_link));
99 170 txg_list_create(&dp->dp_sync_tasks,
100 171 offsetof(dsl_sync_task_t, dst_node));
101 172
102 173 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
174 + cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
103 175
104 176 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
105 177 1, 4, 0);
106 178
107 179 return (dp);
108 180 }
109 181
110 182 int
111 183 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
112 184 {
113 185 int err;
114 186 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
115 187
116 188 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
117 189 &dp->dp_meta_objset);
118 190 if (err != 0)
119 191 dsl_pool_close(dp);
120 192 else
121 193 *dpp = dp;
122 194
123 195 return (err);
124 196 }
125 197
126 198 int
127 199 dsl_pool_open(dsl_pool_t *dp)
128 200 {
129 201 int err;
130 202 dsl_dir_t *dd;
131 203 dsl_dataset_t *ds;
132 204 uint64_t obj;
133 205
134 206 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
135 207 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
136 208 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
137 209 &dp->dp_root_dir_obj);
138 210 if (err)
139 211 goto out;
140 212
141 213 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
142 214 NULL, dp, &dp->dp_root_dir);
143 215 if (err)
144 216 goto out;
145 217
146 218 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
147 219 if (err)
148 220 goto out;
149 221
150 222 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
151 223 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
152 224 if (err)
153 225 goto out;
154 226 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
155 227 FTAG, &ds);
156 228 if (err == 0) {
157 229 err = dsl_dataset_hold_obj(dp,
158 230 ds->ds_phys->ds_prev_snap_obj, dp,
159 231 &dp->dp_origin_snap);
160 232 dsl_dataset_rele(ds, FTAG);
161 233 }
162 234 dsl_dir_rele(dd, dp);
163 235 if (err)
164 236 goto out;
165 237 }
166 238
167 239 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
168 240 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
169 241 &dp->dp_free_dir);
170 242 if (err)
171 243 goto out;
172 244
173 245 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
174 246 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
175 247 if (err)
176 248 goto out;
177 249 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
178 250 dp->dp_meta_objset, obj));
179 251 }
180 252
181 253 if (spa_feature_is_active(dp->dp_spa,
182 254 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
183 255 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
184 256 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
185 257 &dp->dp_bptree_obj);
186 258 if (err != 0)
187 259 goto out;
188 260 }
189 261
190 262 if (spa_feature_is_active(dp->dp_spa,
191 263 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
192 264 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
193 265 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
194 266 &dp->dp_empty_bpobj);
195 267 if (err != 0)
196 268 goto out;
197 269 }
198 270
199 271 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
200 272 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
201 273 &dp->dp_tmp_userrefs_obj);
202 274 if (err == ENOENT)
203 275 err = 0;
204 276 if (err)
205 277 goto out;
206 278
|
↓ open down ↓ |
94 lines elided |
↑ open up ↑ |
207 279 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
208 280
209 281 out:
210 282 rrw_exit(&dp->dp_config_rwlock, FTAG);
211 283 return (err);
212 284 }
213 285
214 286 void
215 287 dsl_pool_close(dsl_pool_t *dp)
216 288 {
217 - /* drop our references from dsl_pool_open() */
218 -
219 289 /*
290 + * Drop our references from dsl_pool_open().
291 + *
220 292 * Since we held the origin_snap from "syncing" context (which
221 293 * includes pool-opening context), it actually only got a "ref"
222 294 * and not a hold, so just drop that here.
223 295 */
224 296 if (dp->dp_origin_snap)
225 297 dsl_dataset_rele(dp->dp_origin_snap, dp);
226 298 if (dp->dp_mos_dir)
227 299 dsl_dir_rele(dp->dp_mos_dir, dp);
228 300 if (dp->dp_free_dir)
229 301 dsl_dir_rele(dp->dp_free_dir, dp);
230 302 if (dp->dp_root_dir)
231 303 dsl_dir_rele(dp->dp_root_dir, dp);
232 304
233 305 bpobj_close(&dp->dp_free_bpobj);
234 306
235 307 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
236 308 if (dp->dp_meta_objset)
237 309 dmu_objset_evict(dp->dp_meta_objset);
238 310
239 311 txg_list_destroy(&dp->dp_dirty_datasets);
240 312 txg_list_destroy(&dp->dp_dirty_zilogs);
241 313 txg_list_destroy(&dp->dp_sync_tasks);
242 314 txg_list_destroy(&dp->dp_dirty_dirs);
243 315
244 316 arc_flush(dp->dp_spa);
245 317 txg_fini(dp);
246 318 dsl_scan_fini(dp);
247 319 rrw_destroy(&dp->dp_config_rwlock);
248 320 mutex_destroy(&dp->dp_lock);
249 321 taskq_destroy(dp->dp_vnrele_taskq);
250 322 if (dp->dp_blkstats)
251 323 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
252 324 kmem_free(dp, sizeof (dsl_pool_t));
253 325 }
254 326
255 327 dsl_pool_t *
256 328 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
257 329 {
258 330 int err;
259 331 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
260 332 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
261 333 objset_t *os;
262 334 dsl_dataset_t *ds;
263 335 uint64_t obj;
264 336
265 337 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
266 338
267 339 /* create and open the MOS (meta-objset) */
268 340 dp->dp_meta_objset = dmu_objset_create_impl(spa,
269 341 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
270 342
271 343 /* create the pool directory */
272 344 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
273 345 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
274 346 ASSERT0(err);
275 347
276 348 /* Initialize scan structures */
277 349 VERIFY0(dsl_scan_init(dp, txg));
278 350
279 351 /* create and open the root dir */
280 352 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
281 353 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
282 354 NULL, dp, &dp->dp_root_dir));
283 355
284 356 /* create and open the meta-objset dir */
285 357 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
286 358 VERIFY0(dsl_pool_open_special_dir(dp,
287 359 MOS_DIR_NAME, &dp->dp_mos_dir));
288 360
289 361 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
290 362 /* create and open the free dir */
291 363 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
292 364 FREE_DIR_NAME, tx);
293 365 VERIFY0(dsl_pool_open_special_dir(dp,
294 366 FREE_DIR_NAME, &dp->dp_free_dir));
295 367
296 368 /* create and open the free_bplist */
297 369 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
298 370 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
299 371 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
300 372 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
301 373 dp->dp_meta_objset, obj));
302 374 }
303 375
304 376 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
305 377 dsl_pool_create_origin(dp, tx);
306 378
307 379 /* create the root dataset */
308 380 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
309 381
310 382 /* create the root objset */
311 383 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
312 384 os = dmu_objset_create_impl(dp->dp_spa, ds,
313 385 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
314 386 #ifdef _KERNEL
315 387 zfs_create_fs(os, kcred, zplprops, tx);
316 388 #endif
317 389 dsl_dataset_rele(ds, FTAG);
318 390
319 391 dmu_tx_commit(tx);
320 392
321 393 rrw_exit(&dp->dp_config_rwlock, FTAG);
322 394
323 395 return (dp);
324 396 }
325 397
326 398 /*
327 399 * Account for the meta-objset space in its placeholder dsl_dir.
328 400 */
329 401 void
330 402 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
331 403 int64_t used, int64_t comp, int64_t uncomp)
332 404 {
333 405 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
334 406 mutex_enter(&dp->dp_lock);
335 407 dp->dp_mos_used_delta += used;
336 408 dp->dp_mos_compressed_delta += comp;
337 409 dp->dp_mos_uncompressed_delta += uncomp;
338 410 mutex_exit(&dp->dp_lock);
|
↓ open down ↓ |
109 lines elided |
↑ open up ↑ |
339 411 }
340 412
341 413 static int
342 414 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
343 415 {
344 416 dsl_deadlist_t *dl = arg;
345 417 dsl_deadlist_insert(dl, bp, tx);
346 418 return (0);
347 419 }
348 420
421 +static void
422 +dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
423 +{
424 + zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
425 + dmu_objset_sync(dp->dp_meta_objset, zio, tx);
426 + VERIFY0(zio_wait(zio));
427 + dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
428 + spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
429 +}
430 +
431 +static void
432 +dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
433 +{
434 + ASSERT(MUTEX_HELD(&dp->dp_lock));
435 +
436 + if (delta < 0)
437 + ASSERT3U(-delta, <=, dp->dp_dirty_total);
438 +
439 + dp->dp_dirty_total += delta;
440 +
441 + /*
442 + * Note: we signal even when increasing dp_dirty_total.
443 + * This ensures forward progress -- each thread wakes the next waiter.
444 + */
445 + if (dp->dp_dirty_total <= zfs_dirty_data_max)
446 + cv_signal(&dp->dp_spaceavail_cv);
447 +}
448 +
349 449 void
350 450 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
351 451 {
352 452 zio_t *zio;
353 453 dmu_tx_t *tx;
354 454 dsl_dir_t *dd;
355 455 dsl_dataset_t *ds;
356 456 objset_t *mos = dp->dp_meta_objset;
357 - hrtime_t start, write_time;
358 - uint64_t data_written;
359 - int err;
360 457 list_t synced_datasets;
361 458
362 459 list_create(&synced_datasets, sizeof (dsl_dataset_t),
363 460 offsetof(dsl_dataset_t, ds_synced_link));
364 461
365 - /*
366 - * We need to copy dp_space_towrite() before doing
367 - * dsl_sync_task_sync(), because
368 - * dsl_dataset_snapshot_reserve_space() will increase
369 - * dp_space_towrite but not actually write anything.
370 - */
371 - data_written = dp->dp_space_towrite[txg & TXG_MASK];
372 -
373 462 tx = dmu_tx_create_assigned(dp, txg);
374 463
375 - dp->dp_read_overhead = 0;
376 - start = gethrtime();
377 -
464 + /*
465 + * Write out all dirty blocks of dirty datasets.
466 + */
378 467 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
379 - while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
468 + while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
380 469 /*
381 470 * We must not sync any non-MOS datasets twice, because
382 471 * we may have taken a snapshot of them. However, we
383 472 * may sync newly-created datasets on pass 2.
384 473 */
385 474 ASSERT(!list_link_active(&ds->ds_synced_link));
386 475 list_insert_tail(&synced_datasets, ds);
387 476 dsl_dataset_sync(ds, zio, tx);
388 477 }
389 - DTRACE_PROBE(pool_sync__1setup);
390 - err = zio_wait(zio);
478 + VERIFY0(zio_wait(zio));
391 479
392 - write_time = gethrtime() - start;
393 - ASSERT(err == 0);
394 - DTRACE_PROBE(pool_sync__2rootzio);
480 + /*
481 + * We have written all of the accounted dirty data, so our
482 + * dp_space_towrite should now be zero. However, some seldom-used
483 + * code paths do not adhere to this (e.g. dbuf_undirty(), also
484 + * rounding error in dbuf_write_physdone).
485 + * Shore up the accounting of any dirtied space now.
486 + */
487 + dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
395 488
396 489 /*
397 490 * After the data blocks have been written (ensured by the zio_wait()
398 491 * above), update the user/group space accounting.
399 492 */
400 - for (ds = list_head(&synced_datasets); ds;
401 - ds = list_next(&synced_datasets, ds))
493 + for (ds = list_head(&synced_datasets); ds != NULL;
494 + ds = list_next(&synced_datasets, ds)) {
402 495 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
496 + }
403 497
404 498 /*
405 499 * Sync the datasets again to push out the changes due to
406 500 * userspace updates. This must be done before we process the
407 501 * sync tasks, so that any snapshots will have the correct
408 502 * user accounting information (and we won't get confused
409 503 * about which blocks are part of the snapshot).
410 504 */
411 505 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
412 - while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
506 + while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
413 507 ASSERT(list_link_active(&ds->ds_synced_link));
414 508 dmu_buf_rele(ds->ds_dbuf, ds);
415 509 dsl_dataset_sync(ds, zio, tx);
416 510 }
417 - err = zio_wait(zio);
511 + VERIFY0(zio_wait(zio));
418 512
419 513 /*
420 514 * Now that the datasets have been completely synced, we can
421 515 * clean up our in-memory structures accumulated while syncing:
422 516 *
423 517 * - move dead blocks from the pending deadlist to the on-disk deadlist
424 518 * - release hold from dsl_dataset_dirty()
425 519 */
426 - while (ds = list_remove_head(&synced_datasets)) {
520 + while ((ds = list_remove_head(&synced_datasets)) != NULL) {
427 521 objset_t *os = ds->ds_objset;
428 522 bplist_iterate(&ds->ds_pending_deadlist,
429 523 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
430 524 ASSERT(!dmu_objset_is_dirty(os, txg));
431 525 dmu_buf_rele(ds->ds_dbuf, ds);
432 526 }
433 -
434 - start = gethrtime();
435 - while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
527 + while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
436 528 dsl_dir_sync(dd, tx);
437 - write_time += gethrtime() - start;
529 + }
438 530
439 531 /*
440 532 * The MOS's space is accounted for in the pool/$MOS
441 533 * (dp_mos_dir). We can't modify the mos while we're syncing
442 534 * it, so we remember the deltas and apply them here.
443 535 */
444 536 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
445 537 dp->dp_mos_uncompressed_delta != 0) {
446 538 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
447 539 dp->dp_mos_used_delta,
448 540 dp->dp_mos_compressed_delta,
449 541 dp->dp_mos_uncompressed_delta, tx);
450 542 dp->dp_mos_used_delta = 0;
451 543 dp->dp_mos_compressed_delta = 0;
452 544 dp->dp_mos_uncompressed_delta = 0;
453 545 }
454 546
455 - start = gethrtime();
456 547 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
457 548 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
458 - zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
459 - dmu_objset_sync(mos, zio, tx);
460 - err = zio_wait(zio);
461 - ASSERT(err == 0);
462 - dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
463 - spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
549 + dsl_pool_sync_mos(dp, tx);
464 550 }
465 - write_time += gethrtime() - start;
466 - DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
467 - hrtime_t, dp->dp_read_overhead);
468 - write_time -= dp->dp_read_overhead;
469 551
470 552 /*
471 553 * If we modify a dataset in the same txg that we want to destroy it,
472 554 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
473 555 * dsl_dir_destroy_check() will fail if there are unexpected holds.
474 556 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
475 557 * and clearing the hold on it) before we process the sync_tasks.
476 558 * The MOS data dirtied by the sync_tasks will be synced on the next
477 559 * pass.
478 560 */
479 - DTRACE_PROBE(pool_sync__3task);
480 561 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
481 562 dsl_sync_task_t *dst;
482 563 /*
483 564 * No more sync tasks should have been added while we
484 565 * were syncing.
485 566 */
486 - ASSERT(spa_sync_pass(dp->dp_spa) == 1);
487 - while (dst = txg_list_remove(&dp->dp_sync_tasks, txg))
567 + ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
568 + while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
488 569 dsl_sync_task_sync(dst, tx);
489 570 }
490 571
491 572 dmu_tx_commit(tx);
492 573
493 - dp->dp_space_towrite[txg & TXG_MASK] = 0;
494 - ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
495 -
496 - /*
497 - * If the write limit max has not been explicitly set, set it
498 - * to a fraction of available physical memory (default 1/8th).
499 - * Note that we must inflate the limit because the spa
500 - * inflates write sizes to account for data replication.
501 - * Check this each sync phase to catch changing memory size.
502 - */
503 - if (physmem != old_physmem && zfs_write_limit_shift) {
504 - mutex_enter(&zfs_write_limit_lock);
505 - old_physmem = physmem;
506 - zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
507 - zfs_write_limit_inflated = MAX(zfs_write_limit_min,
508 - spa_get_asize(dp->dp_spa, zfs_write_limit_max));
509 - mutex_exit(&zfs_write_limit_lock);
510 - }
511 -
512 - /*
513 - * Attempt to keep the sync time consistent by adjusting the
514 - * amount of write traffic allowed into each transaction group.
515 - * Weight the throughput calculation towards the current value:
516 - * thru = 3/4 old_thru + 1/4 new_thru
517 - *
518 - * Note: write_time is in nanosecs while dp_throughput is expressed in
519 - * bytes per millisecond.
520 - */
521 - ASSERT(zfs_write_limit_min > 0);
522 - if (data_written > zfs_write_limit_min / 8 &&
523 - write_time > MSEC2NSEC(1)) {
524 - uint64_t throughput = data_written / NSEC2MSEC(write_time);
525 -
526 - if (dp->dp_throughput)
527 - dp->dp_throughput = throughput / 4 +
528 - 3 * dp->dp_throughput / 4;
529 - else
530 - dp->dp_throughput = throughput;
531 - dp->dp_write_limit = MIN(zfs_write_limit_inflated,
532 - MAX(zfs_write_limit_min,
533 - dp->dp_throughput * zfs_txg_synctime_ms));
534 - }
574 + DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
535 575 }
536 576
537 577 void
538 578 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
539 579 {
540 580 zilog_t *zilog;
541 - dsl_dataset_t *ds;
542 581
543 582 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
544 - ds = dmu_objset_ds(zilog->zl_os);
583 + dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
545 584 zil_clean(zilog, txg);
546 585 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
547 586 dmu_buf_rele(ds->ds_dbuf, zilog);
548 587 }
549 588 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
550 589 }
551 590
552 591 /*
553 592 * TRUE if the current thread is the tx_sync_thread or if we
554 593 * are being called from SPA context during pool initialization.
555 594 */
556 595 int
557 596 dsl_pool_sync_context(dsl_pool_t *dp)
558 597 {
559 598 return (curthread == dp->dp_tx.tx_sync_thread ||
560 599 spa_is_initializing(dp->dp_spa));
561 600 }
562 601
563 602 uint64_t
564 603 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
565 604 {
566 605 uint64_t space, resv;
567 606
568 607 /*
569 608 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
570 609 * efficiency.
571 610 * XXX The intent log is not accounted for, so it must fit
572 611 * within this slop.
573 612 *
574 613 * If we're trying to assess whether it's OK to do a free,
575 614 * cut the reservation in half to allow forward progress
|
↓ open down ↓ |
21 lines elided |
↑ open up ↑ |
576 615 * (e.g. make it possible to rm(1) files from a full pool).
577 616 */
578 617 space = spa_get_dspace(dp->dp_spa);
579 618 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
580 619 if (netfree)
581 620 resv >>= 1;
582 621
583 622 return (space - resv);
584 623 }
585 624
586 -int
587 -dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
625 +boolean_t
626 +dsl_pool_need_dirty_delay(dsl_pool_t *dp)
588 627 {
589 - uint64_t reserved = 0;
590 - uint64_t write_limit = (zfs_write_limit_override ?
591 - zfs_write_limit_override : dp->dp_write_limit);
628 + uint64_t delay_min_bytes =
629 + zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
630 + boolean_t rv;
592 631
593 - if (zfs_no_write_throttle) {
594 - atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
595 - space);
596 - return (0);
597 - }
598 -
599 - /*
600 - * Check to see if we have exceeded the maximum allowed IO for
601 - * this transaction group. We can do this without locks since
602 - * a little slop here is ok. Note that we do the reserved check
603 - * with only half the requested reserve: this is because the
604 - * reserve requests are worst-case, and we really don't want to
605 - * throttle based off of worst-case estimates.
606 - */
607 - if (write_limit > 0) {
608 - reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
609 - + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
610 -
611 - if (reserved && reserved > write_limit)
612 - return (SET_ERROR(ERESTART));
613 - }
614 -
615 - atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
616 -
617 - /*
618 - * If this transaction group is over 7/8ths capacity, delay
619 - * the caller 1 clock tick. This will slow down the "fill"
620 - * rate until the sync process can catch up with us.
621 - */
622 - if (reserved && reserved > (write_limit - (write_limit >> 3))) {
623 - txg_delay(dp, tx->tx_txg, zfs_throttle_delay,
624 - zfs_throttle_resolution);
625 - }
626 -
627 - return (0);
632 + mutex_enter(&dp->dp_lock);
633 + if (dp->dp_dirty_total > zfs_dirty_data_sync)
634 + txg_kick(dp);
635 + rv = (dp->dp_dirty_total > delay_min_bytes);
636 + mutex_exit(&dp->dp_lock);
637 + return (rv);
628 638 }
629 639
630 640 void
631 -dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
641 +dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
632 642 {
633 - ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
634 - atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
643 + if (space > 0) {
644 + mutex_enter(&dp->dp_lock);
645 + dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
646 + dsl_pool_dirty_delta(dp, space);
647 + mutex_exit(&dp->dp_lock);
648 + }
635 649 }
636 650
637 651 void
638 -dsl_pool_memory_pressure(dsl_pool_t *dp)
639 -{
640 - uint64_t space_inuse = 0;
641 - int i;
642 -
643 - if (dp->dp_write_limit == zfs_write_limit_min)
652 +dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg) {
653 + ASSERT3S(space, >=, 0);
654 + if (space == 0)
644 655 return;
645 -
646 - for (i = 0; i < TXG_SIZE; i++) {
647 - space_inuse += dp->dp_space_towrite[i];
648 - space_inuse += dp->dp_tempreserved[i];
656 + mutex_enter(&dp->dp_lock);
657 + if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
658 + /* XXX writing something we didn't dirty? */
659 + space = dp->dp_dirty_pertxg[txg & TXG_MASK];
649 660 }
650 - dp->dp_write_limit = MAX(zfs_write_limit_min,
651 - MIN(dp->dp_write_limit, space_inuse / 4));
661 + ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
662 + dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
663 + ASSERT3U(dp->dp_dirty_total, >=, space);
664 + dsl_pool_dirty_delta(dp, -space);
665 + mutex_exit(&dp->dp_lock);
652 666 }
653 667
654 -void
655 -dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
656 -{
657 - if (space > 0) {
658 - mutex_enter(&dp->dp_lock);
659 - dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
660 - mutex_exit(&dp->dp_lock);
661 - }
662 -}
663 -
664 668 /* ARGSUSED */
665 669 static int
666 670 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
667 671 {
668 672 dmu_tx_t *tx = arg;
669 673 dsl_dataset_t *ds, *prev = NULL;
670 674 int err;
671 675
672 676 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
673 677 if (err)
674 678 return (err);
675 679
676 680 while (ds->ds_phys->ds_prev_snap_obj != 0) {
677 681 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
678 682 FTAG, &prev);
679 683 if (err) {
680 684 dsl_dataset_rele(ds, FTAG);
681 685 return (err);
682 686 }
683 687
684 688 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
685 689 break;
686 690 dsl_dataset_rele(ds, FTAG);
687 691 ds = prev;
688 692 prev = NULL;
689 693 }
690 694
691 695 if (prev == NULL) {
692 696 prev = dp->dp_origin_snap;
693 697
694 698 /*
695 699 * The $ORIGIN can't have any data, or the accounting
696 700 * will be wrong.
697 701 */
698 702 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
699 703
700 704 /* The origin doesn't get attached to itself */
701 705 if (ds->ds_object == prev->ds_object) {
702 706 dsl_dataset_rele(ds, FTAG);
703 707 return (0);
704 708 }
705 709
706 710 dmu_buf_will_dirty(ds->ds_dbuf, tx);
707 711 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
708 712 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
709 713
710 714 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
711 715 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
712 716
713 717 dmu_buf_will_dirty(prev->ds_dbuf, tx);
714 718 prev->ds_phys->ds_num_children++;
715 719
716 720 if (ds->ds_phys->ds_next_snap_obj == 0) {
717 721 ASSERT(ds->ds_prev == NULL);
718 722 VERIFY0(dsl_dataset_hold_obj(dp,
719 723 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
720 724 }
721 725 }
722 726
723 727 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
724 728 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
725 729
726 730 if (prev->ds_phys->ds_next_clones_obj == 0) {
727 731 dmu_buf_will_dirty(prev->ds_dbuf, tx);
728 732 prev->ds_phys->ds_next_clones_obj =
729 733 zap_create(dp->dp_meta_objset,
730 734 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
731 735 }
732 736 VERIFY0(zap_add_int(dp->dp_meta_objset,
733 737 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
734 738
735 739 dsl_dataset_rele(ds, FTAG);
736 740 if (prev != dp->dp_origin_snap)
737 741 dsl_dataset_rele(prev, FTAG);
738 742 return (0);
739 743 }
740 744
741 745 void
742 746 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
743 747 {
744 748 ASSERT(dmu_tx_is_syncing(tx));
745 749 ASSERT(dp->dp_origin_snap != NULL);
746 750
747 751 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
748 752 tx, DS_FIND_CHILDREN));
749 753 }
750 754
751 755 /* ARGSUSED */
752 756 static int
753 757 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
754 758 {
755 759 dmu_tx_t *tx = arg;
756 760 objset_t *mos = dp->dp_meta_objset;
757 761
758 762 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
759 763 dsl_dataset_t *origin;
760 764
761 765 VERIFY0(dsl_dataset_hold_obj(dp,
762 766 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
763 767
764 768 if (origin->ds_dir->dd_phys->dd_clones == 0) {
765 769 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
766 770 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
767 771 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
768 772 }
769 773
770 774 VERIFY0(zap_add_int(dp->dp_meta_objset,
771 775 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
772 776
773 777 dsl_dataset_rele(origin, FTAG);
774 778 }
775 779 return (0);
776 780 }
777 781
778 782 void
779 783 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
780 784 {
781 785 ASSERT(dmu_tx_is_syncing(tx));
782 786 uint64_t obj;
783 787
784 788 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
785 789 VERIFY0(dsl_pool_open_special_dir(dp,
786 790 FREE_DIR_NAME, &dp->dp_free_dir));
787 791
788 792 /*
789 793 * We can't use bpobj_alloc(), because spa_version() still
790 794 * returns the old version, and we need a new-version bpobj with
791 795 * subobj support. So call dmu_object_alloc() directly.
792 796 */
793 797 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
794 798 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
795 799 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
796 800 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
797 801 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
798 802
799 803 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
800 804 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
801 805 }
802 806
803 807 void
804 808 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
805 809 {
806 810 uint64_t dsobj;
807 811 dsl_dataset_t *ds;
808 812
809 813 ASSERT(dmu_tx_is_syncing(tx));
810 814 ASSERT(dp->dp_origin_snap == NULL);
811 815 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
812 816
813 817 /* create the origin dir, ds, & snap-ds */
814 818 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
815 819 NULL, 0, kcred, tx);
816 820 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
817 821 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
818 822 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
819 823 dp, &dp->dp_origin_snap));
820 824 dsl_dataset_rele(ds, FTAG);
821 825 }
822 826
823 827 taskq_t *
824 828 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
825 829 {
826 830 return (dp->dp_vnrele_taskq);
827 831 }
828 832
829 833 /*
830 834 * Walk through the pool-wide zap object of temporary snapshot user holds
831 835 * and release them.
832 836 */
833 837 void
834 838 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
835 839 {
836 840 zap_attribute_t za;
837 841 zap_cursor_t zc;
838 842 objset_t *mos = dp->dp_meta_objset;
839 843 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
840 844 nvlist_t *holds;
841 845
842 846 if (zapobj == 0)
843 847 return;
844 848 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
845 849
846 850 holds = fnvlist_alloc();
847 851
848 852 for (zap_cursor_init(&zc, mos, zapobj);
849 853 zap_cursor_retrieve(&zc, &za) == 0;
850 854 zap_cursor_advance(&zc)) {
851 855 char *htag;
852 856 nvlist_t *tags;
853 857
854 858 htag = strchr(za.za_name, '-');
855 859 *htag = '\0';
856 860 ++htag;
857 861 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
858 862 tags = fnvlist_alloc();
859 863 fnvlist_add_boolean(tags, htag);
860 864 fnvlist_add_nvlist(holds, za.za_name, tags);
861 865 fnvlist_free(tags);
862 866 } else {
863 867 fnvlist_add_boolean(tags, htag);
864 868 }
865 869 }
866 870 dsl_dataset_user_release_tmp(dp, holds);
867 871 fnvlist_free(holds);
868 872 zap_cursor_fini(&zc);
869 873 }
870 874
871 875 /*
872 876 * Create the pool-wide zap object for storing temporary snapshot holds.
873 877 */
874 878 void
875 879 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
876 880 {
877 881 objset_t *mos = dp->dp_meta_objset;
878 882
879 883 ASSERT(dp->dp_tmp_userrefs_obj == 0);
880 884 ASSERT(dmu_tx_is_syncing(tx));
881 885
882 886 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
883 887 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
884 888 }
885 889
886 890 static int
887 891 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
888 892 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
889 893 {
890 894 objset_t *mos = dp->dp_meta_objset;
891 895 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
892 896 char *name;
893 897 int error;
894 898
895 899 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
896 900 ASSERT(dmu_tx_is_syncing(tx));
897 901
898 902 /*
899 903 * If the pool was created prior to SPA_VERSION_USERREFS, the
900 904 * zap object for temporary holds might not exist yet.
901 905 */
902 906 if (zapobj == 0) {
903 907 if (holding) {
904 908 dsl_pool_user_hold_create_obj(dp, tx);
905 909 zapobj = dp->dp_tmp_userrefs_obj;
906 910 } else {
907 911 return (SET_ERROR(ENOENT));
908 912 }
909 913 }
910 914
911 915 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
912 916 if (holding)
913 917 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
914 918 else
915 919 error = zap_remove(mos, zapobj, name, tx);
916 920 strfree(name);
917 921
918 922 return (error);
919 923 }
920 924
921 925 /*
922 926 * Add a temporary hold for the given dataset object and tag.
923 927 */
924 928 int
925 929 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
926 930 uint64_t now, dmu_tx_t *tx)
927 931 {
928 932 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
929 933 }
930 934
931 935 /*
932 936 * Release a temporary hold for the given dataset object and tag.
933 937 */
934 938 int
935 939 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
936 940 dmu_tx_t *tx)
937 941 {
938 942 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
939 943 tx, B_FALSE));
940 944 }
941 945
942 946 /*
943 947 * DSL Pool Configuration Lock
944 948 *
945 949 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
946 950 * creation / destruction / rename / property setting). It must be held for
947 951 * read to hold a dataset or dsl_dir. I.e. you must call
948 952 * dsl_pool_config_enter() or dsl_pool_hold() before calling
949 953 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
950 954 * must be held continuously until all datasets and dsl_dirs are released.
951 955 *
952 956 * The only exception to this rule is that if a "long hold" is placed on
953 957 * a dataset, then the dp_config_rwlock may be dropped while the dataset
954 958 * is still held. The long hold will prevent the dataset from being
955 959 * destroyed -- the destroy will fail with EBUSY. A long hold can be
956 960 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
957 961 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
958 962 *
959 963 * Legitimate long-holders (including owners) should be long-running, cancelable
960 964 * tasks that should cause "zfs destroy" to fail. This includes DMU
961 965 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
962 966 * "zfs send", and "zfs diff". There are several other long-holders whose
963 967 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
964 968 *
965 969 * The usual formula for long-holding would be:
966 970 * dsl_pool_hold()
967 971 * dsl_dataset_hold()
968 972 * ... perform checks ...
969 973 * dsl_dataset_long_hold()
970 974 * dsl_pool_rele()
971 975 * ... perform long-running task ...
972 976 * dsl_dataset_long_rele()
973 977 * dsl_dataset_rele()
974 978 *
975 979 * Note that when the long hold is released, the dataset is still held but
976 980 * the pool is not held. The dataset may change arbitrarily during this time
977 981 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
978 982 * dataset except release it.
979 983 *
980 984 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
981 985 * or modifying operations.
982 986 *
983 987 * Modifying operations should generally use dsl_sync_task(). The synctask
984 988 * infrastructure enforces proper locking strategy with respect to the
985 989 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
986 990 *
987 991 * Read-only operations will manually hold the pool, then the dataset, obtain
988 992 * information from the dataset, then release the pool and dataset.
989 993 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
990 994 * hold/rele.
991 995 */
992 996
993 997 int
994 998 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
995 999 {
996 1000 spa_t *spa;
997 1001 int error;
998 1002
999 1003 error = spa_open(name, &spa, tag);
1000 1004 if (error == 0) {
1001 1005 *dp = spa_get_dsl(spa);
1002 1006 dsl_pool_config_enter(*dp, tag);
1003 1007 }
1004 1008 return (error);
1005 1009 }
1006 1010
1007 1011 void
1008 1012 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1009 1013 {
1010 1014 dsl_pool_config_exit(dp, tag);
1011 1015 spa_close(dp->dp_spa, tag);
1012 1016 }
1013 1017
1014 1018 void
1015 1019 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1016 1020 {
1017 1021 /*
1018 1022 * We use a "reentrant" reader-writer lock, but not reentrantly.
1019 1023 *
1020 1024 * The rrwlock can (with the track_all flag) track all reading threads,
1021 1025 * which is very useful for debugging which code path failed to release
1022 1026 * the lock, and for verifying that the *current* thread does hold
1023 1027 * the lock.
1024 1028 *
1025 1029 * (Unlike a rwlock, which knows that N threads hold it for
1026 1030 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1027 1031 * if any thread holds it for read, even if this thread doesn't).
1028 1032 */
1029 1033 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1030 1034 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1031 1035 }
1032 1036
1033 1037 void
1034 1038 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1035 1039 {
1036 1040 rrw_exit(&dp->dp_config_rwlock, tag);
1037 1041 }
1038 1042
1039 1043 boolean_t
1040 1044 dsl_pool_config_held(dsl_pool_t *dp)
1041 1045 {
1042 1046 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1043 1047 }
|
↓ open down ↓ |
370 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX