Print this page
3740 Poor ZFS send / receive performance due to snapshot hold / release processing
Submitted by: Steven Hartland <steven.hartland@multiplay.co.uk>
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
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
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 + * Copyright (c) 2013 Steven Hartland. All rights reserved.
24 25 */
25 26
26 27 #include <sys/dsl_pool.h>
27 28 #include <sys/dsl_dataset.h>
28 29 #include <sys/dsl_prop.h>
29 30 #include <sys/dsl_dir.h>
30 31 #include <sys/dsl_synctask.h>
31 32 #include <sys/dsl_scan.h>
32 33 #include <sys/dnode.h>
33 34 #include <sys/dmu_tx.h>
34 35 #include <sys/dmu_objset.h>
35 36 #include <sys/arc.h>
36 37 #include <sys/zap.h>
37 38 #include <sys/zio.h>
38 39 #include <sys/zfs_context.h>
39 40 #include <sys/fs/zfs.h>
40 41 #include <sys/zfs_znode.h>
41 42 #include <sys/spa_impl.h>
42 43 #include <sys/dsl_deadlist.h>
43 44 #include <sys/bptree.h>
44 45 #include <sys/zfeature.h>
45 46 #include <sys/zil_impl.h>
46 47 #include <sys/dsl_userhold.h>
47 48
48 49 int zfs_no_write_throttle = 0;
49 50 int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
50 51 int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
51 52
52 53 uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
53 54 uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
54 55 uint64_t zfs_write_limit_inflated = 0;
55 56 uint64_t zfs_write_limit_override = 0;
56 57
57 58 kmutex_t zfs_write_limit_lock;
58 59
59 60 static pgcnt_t old_physmem = 0;
60 61
61 62 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
62 63 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
63 64
64 65 int
65 66 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
66 67 {
67 68 uint64_t obj;
68 69 int err;
69 70
70 71 err = zap_lookup(dp->dp_meta_objset,
71 72 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
72 73 name, sizeof (obj), 1, &obj);
73 74 if (err)
74 75 return (err);
75 76
76 77 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
77 78 }
78 79
79 80 static dsl_pool_t *
80 81 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
81 82 {
82 83 dsl_pool_t *dp;
83 84 blkptr_t *bp = spa_get_rootblkptr(spa);
84 85
85 86 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
86 87 dp->dp_spa = spa;
87 88 dp->dp_meta_rootbp = *bp;
88 89 rrw_init(&dp->dp_config_rwlock, B_TRUE);
89 90 dp->dp_write_limit = zfs_write_limit_min;
90 91 txg_init(dp, txg);
91 92
92 93 txg_list_create(&dp->dp_dirty_datasets,
93 94 offsetof(dsl_dataset_t, ds_dirty_link));
94 95 txg_list_create(&dp->dp_dirty_zilogs,
95 96 offsetof(zilog_t, zl_dirty_link));
96 97 txg_list_create(&dp->dp_dirty_dirs,
97 98 offsetof(dsl_dir_t, dd_dirty_link));
98 99 txg_list_create(&dp->dp_sync_tasks,
99 100 offsetof(dsl_sync_task_t, dst_node));
100 101
101 102 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
102 103
103 104 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
104 105 1, 4, 0);
105 106
106 107 return (dp);
107 108 }
108 109
109 110 int
110 111 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
111 112 {
112 113 int err;
113 114 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
114 115
115 116 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
116 117 &dp->dp_meta_objset);
117 118 if (err != 0)
118 119 dsl_pool_close(dp);
119 120 else
120 121 *dpp = dp;
121 122
122 123 return (err);
123 124 }
124 125
125 126 int
126 127 dsl_pool_open(dsl_pool_t *dp)
127 128 {
128 129 int err;
129 130 dsl_dir_t *dd;
130 131 dsl_dataset_t *ds;
131 132 uint64_t obj;
132 133
133 134 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
134 135 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
135 136 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
136 137 &dp->dp_root_dir_obj);
137 138 if (err)
138 139 goto out;
139 140
140 141 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
141 142 NULL, dp, &dp->dp_root_dir);
142 143 if (err)
143 144 goto out;
144 145
145 146 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
146 147 if (err)
147 148 goto out;
148 149
149 150 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
150 151 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
151 152 if (err)
152 153 goto out;
153 154 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
154 155 FTAG, &ds);
155 156 if (err == 0) {
156 157 err = dsl_dataset_hold_obj(dp,
157 158 ds->ds_phys->ds_prev_snap_obj, dp,
158 159 &dp->dp_origin_snap);
159 160 dsl_dataset_rele(ds, FTAG);
160 161 }
161 162 dsl_dir_rele(dd, dp);
162 163 if (err)
163 164 goto out;
164 165 }
165 166
166 167 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
167 168 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
168 169 &dp->dp_free_dir);
169 170 if (err)
170 171 goto out;
171 172
172 173 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
173 174 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
174 175 if (err)
175 176 goto out;
176 177 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
177 178 dp->dp_meta_objset, obj));
178 179 }
179 180
180 181 if (spa_feature_is_active(dp->dp_spa,
181 182 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
182 183 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
183 184 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
184 185 &dp->dp_bptree_obj);
185 186 if (err != 0)
186 187 goto out;
187 188 }
188 189
189 190 if (spa_feature_is_active(dp->dp_spa,
190 191 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
191 192 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
192 193 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
193 194 &dp->dp_empty_bpobj);
194 195 if (err != 0)
195 196 goto out;
196 197 }
197 198
198 199 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
199 200 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
200 201 &dp->dp_tmp_userrefs_obj);
201 202 if (err == ENOENT)
202 203 err = 0;
203 204 if (err)
204 205 goto out;
205 206
206 207 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
207 208
208 209 out:
209 210 rrw_exit(&dp->dp_config_rwlock, FTAG);
210 211 return (err);
211 212 }
212 213
213 214 void
214 215 dsl_pool_close(dsl_pool_t *dp)
215 216 {
216 217 /* drop our references from dsl_pool_open() */
217 218
218 219 /*
219 220 * Since we held the origin_snap from "syncing" context (which
220 221 * includes pool-opening context), it actually only got a "ref"
221 222 * and not a hold, so just drop that here.
222 223 */
223 224 if (dp->dp_origin_snap)
224 225 dsl_dataset_rele(dp->dp_origin_snap, dp);
225 226 if (dp->dp_mos_dir)
226 227 dsl_dir_rele(dp->dp_mos_dir, dp);
227 228 if (dp->dp_free_dir)
228 229 dsl_dir_rele(dp->dp_free_dir, dp);
229 230 if (dp->dp_root_dir)
230 231 dsl_dir_rele(dp->dp_root_dir, dp);
231 232
232 233 bpobj_close(&dp->dp_free_bpobj);
233 234
234 235 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
235 236 if (dp->dp_meta_objset)
236 237 dmu_objset_evict(dp->dp_meta_objset);
237 238
238 239 txg_list_destroy(&dp->dp_dirty_datasets);
239 240 txg_list_destroy(&dp->dp_dirty_zilogs);
240 241 txg_list_destroy(&dp->dp_sync_tasks);
241 242 txg_list_destroy(&dp->dp_dirty_dirs);
242 243
243 244 arc_flush(dp->dp_spa);
244 245 txg_fini(dp);
245 246 dsl_scan_fini(dp);
246 247 rrw_destroy(&dp->dp_config_rwlock);
247 248 mutex_destroy(&dp->dp_lock);
248 249 taskq_destroy(dp->dp_vnrele_taskq);
249 250 if (dp->dp_blkstats)
250 251 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
251 252 kmem_free(dp, sizeof (dsl_pool_t));
252 253 }
253 254
254 255 dsl_pool_t *
255 256 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
256 257 {
257 258 int err;
258 259 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
259 260 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
260 261 objset_t *os;
261 262 dsl_dataset_t *ds;
262 263 uint64_t obj;
263 264
264 265 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
265 266
266 267 /* create and open the MOS (meta-objset) */
267 268 dp->dp_meta_objset = dmu_objset_create_impl(spa,
268 269 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
269 270
270 271 /* create the pool directory */
271 272 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
272 273 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
273 274 ASSERT0(err);
274 275
275 276 /* Initialize scan structures */
276 277 VERIFY0(dsl_scan_init(dp, txg));
277 278
278 279 /* create and open the root dir */
279 280 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
280 281 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
281 282 NULL, dp, &dp->dp_root_dir));
282 283
283 284 /* create and open the meta-objset dir */
284 285 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
285 286 VERIFY0(dsl_pool_open_special_dir(dp,
286 287 MOS_DIR_NAME, &dp->dp_mos_dir));
287 288
288 289 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
289 290 /* create and open the free dir */
290 291 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
291 292 FREE_DIR_NAME, tx);
292 293 VERIFY0(dsl_pool_open_special_dir(dp,
293 294 FREE_DIR_NAME, &dp->dp_free_dir));
294 295
295 296 /* create and open the free_bplist */
296 297 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
297 298 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
298 299 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
299 300 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
300 301 dp->dp_meta_objset, obj));
301 302 }
302 303
303 304 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
304 305 dsl_pool_create_origin(dp, tx);
305 306
306 307 /* create the root dataset */
307 308 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
308 309
309 310 /* create the root objset */
310 311 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
311 312 os = dmu_objset_create_impl(dp->dp_spa, ds,
312 313 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
313 314 #ifdef _KERNEL
314 315 zfs_create_fs(os, kcred, zplprops, tx);
315 316 #endif
316 317 dsl_dataset_rele(ds, FTAG);
317 318
318 319 dmu_tx_commit(tx);
319 320
320 321 rrw_exit(&dp->dp_config_rwlock, FTAG);
321 322
322 323 return (dp);
323 324 }
324 325
325 326 /*
326 327 * Account for the meta-objset space in its placeholder dsl_dir.
327 328 */
328 329 void
329 330 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
330 331 int64_t used, int64_t comp, int64_t uncomp)
331 332 {
332 333 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
333 334 mutex_enter(&dp->dp_lock);
334 335 dp->dp_mos_used_delta += used;
335 336 dp->dp_mos_compressed_delta += comp;
336 337 dp->dp_mos_uncompressed_delta += uncomp;
337 338 mutex_exit(&dp->dp_lock);
338 339 }
339 340
340 341 static int
341 342 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
342 343 {
343 344 dsl_deadlist_t *dl = arg;
344 345 dsl_deadlist_insert(dl, bp, tx);
345 346 return (0);
346 347 }
347 348
348 349 void
349 350 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
350 351 {
351 352 zio_t *zio;
352 353 dmu_tx_t *tx;
353 354 dsl_dir_t *dd;
354 355 dsl_dataset_t *ds;
355 356 objset_t *mos = dp->dp_meta_objset;
356 357 hrtime_t start, write_time;
357 358 uint64_t data_written;
358 359 int err;
359 360 list_t synced_datasets;
360 361
361 362 list_create(&synced_datasets, sizeof (dsl_dataset_t),
362 363 offsetof(dsl_dataset_t, ds_synced_link));
363 364
364 365 /*
365 366 * We need to copy dp_space_towrite() before doing
366 367 * dsl_sync_task_sync(), because
367 368 * dsl_dataset_snapshot_reserve_space() will increase
368 369 * dp_space_towrite but not actually write anything.
369 370 */
370 371 data_written = dp->dp_space_towrite[txg & TXG_MASK];
371 372
372 373 tx = dmu_tx_create_assigned(dp, txg);
373 374
374 375 dp->dp_read_overhead = 0;
375 376 start = gethrtime();
376 377
377 378 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
378 379 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
379 380 /*
380 381 * We must not sync any non-MOS datasets twice, because
381 382 * we may have taken a snapshot of them. However, we
382 383 * may sync newly-created datasets on pass 2.
383 384 */
384 385 ASSERT(!list_link_active(&ds->ds_synced_link));
385 386 list_insert_tail(&synced_datasets, ds);
386 387 dsl_dataset_sync(ds, zio, tx);
387 388 }
388 389 DTRACE_PROBE(pool_sync__1setup);
389 390 err = zio_wait(zio);
390 391
391 392 write_time = gethrtime() - start;
392 393 ASSERT(err == 0);
393 394 DTRACE_PROBE(pool_sync__2rootzio);
394 395
395 396 /*
396 397 * After the data blocks have been written (ensured by the zio_wait()
397 398 * above), update the user/group space accounting.
398 399 */
399 400 for (ds = list_head(&synced_datasets); ds;
400 401 ds = list_next(&synced_datasets, ds))
401 402 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
402 403
403 404 /*
404 405 * Sync the datasets again to push out the changes due to
405 406 * userspace updates. This must be done before we process the
406 407 * sync tasks, so that any snapshots will have the correct
407 408 * user accounting information (and we won't get confused
408 409 * about which blocks are part of the snapshot).
409 410 */
410 411 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
411 412 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
412 413 ASSERT(list_link_active(&ds->ds_synced_link));
413 414 dmu_buf_rele(ds->ds_dbuf, ds);
414 415 dsl_dataset_sync(ds, zio, tx);
415 416 }
416 417 err = zio_wait(zio);
417 418
418 419 /*
419 420 * Now that the datasets have been completely synced, we can
420 421 * clean up our in-memory structures accumulated while syncing:
421 422 *
422 423 * - move dead blocks from the pending deadlist to the on-disk deadlist
423 424 * - release hold from dsl_dataset_dirty()
424 425 */
425 426 while (ds = list_remove_head(&synced_datasets)) {
426 427 objset_t *os = ds->ds_objset;
427 428 bplist_iterate(&ds->ds_pending_deadlist,
428 429 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
429 430 ASSERT(!dmu_objset_is_dirty(os, txg));
430 431 dmu_buf_rele(ds->ds_dbuf, ds);
431 432 }
432 433
433 434 start = gethrtime();
434 435 while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
435 436 dsl_dir_sync(dd, tx);
436 437 write_time += gethrtime() - start;
437 438
438 439 /*
439 440 * The MOS's space is accounted for in the pool/$MOS
440 441 * (dp_mos_dir). We can't modify the mos while we're syncing
441 442 * it, so we remember the deltas and apply them here.
442 443 */
443 444 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
444 445 dp->dp_mos_uncompressed_delta != 0) {
445 446 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
446 447 dp->dp_mos_used_delta,
447 448 dp->dp_mos_compressed_delta,
448 449 dp->dp_mos_uncompressed_delta, tx);
449 450 dp->dp_mos_used_delta = 0;
450 451 dp->dp_mos_compressed_delta = 0;
451 452 dp->dp_mos_uncompressed_delta = 0;
452 453 }
453 454
454 455 start = gethrtime();
455 456 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
456 457 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
457 458 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
458 459 dmu_objset_sync(mos, zio, tx);
459 460 err = zio_wait(zio);
460 461 ASSERT(err == 0);
461 462 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
462 463 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
463 464 }
464 465 write_time += gethrtime() - start;
465 466 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
466 467 hrtime_t, dp->dp_read_overhead);
467 468 write_time -= dp->dp_read_overhead;
468 469
469 470 /*
470 471 * If we modify a dataset in the same txg that we want to destroy it,
471 472 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
472 473 * dsl_dir_destroy_check() will fail if there are unexpected holds.
473 474 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
474 475 * and clearing the hold on it) before we process the sync_tasks.
475 476 * The MOS data dirtied by the sync_tasks will be synced on the next
476 477 * pass.
477 478 */
478 479 DTRACE_PROBE(pool_sync__3task);
479 480 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
480 481 dsl_sync_task_t *dst;
481 482 /*
482 483 * No more sync tasks should have been added while we
483 484 * were syncing.
484 485 */
485 486 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
486 487 while (dst = txg_list_remove(&dp->dp_sync_tasks, txg))
487 488 dsl_sync_task_sync(dst, tx);
488 489 }
489 490
490 491 dmu_tx_commit(tx);
491 492
492 493 dp->dp_space_towrite[txg & TXG_MASK] = 0;
493 494 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
494 495
495 496 /*
496 497 * If the write limit max has not been explicitly set, set it
497 498 * to a fraction of available physical memory (default 1/8th).
498 499 * Note that we must inflate the limit because the spa
499 500 * inflates write sizes to account for data replication.
500 501 * Check this each sync phase to catch changing memory size.
501 502 */
502 503 if (physmem != old_physmem && zfs_write_limit_shift) {
503 504 mutex_enter(&zfs_write_limit_lock);
504 505 old_physmem = physmem;
505 506 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
506 507 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
507 508 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
508 509 mutex_exit(&zfs_write_limit_lock);
509 510 }
510 511
511 512 /*
512 513 * Attempt to keep the sync time consistent by adjusting the
513 514 * amount of write traffic allowed into each transaction group.
514 515 * Weight the throughput calculation towards the current value:
515 516 * thru = 3/4 old_thru + 1/4 new_thru
516 517 *
517 518 * Note: write_time is in nanosecs while dp_throughput is expressed in
518 519 * bytes per millisecond.
519 520 */
520 521 ASSERT(zfs_write_limit_min > 0);
521 522 if (data_written > zfs_write_limit_min / 8 &&
522 523 write_time > MSEC2NSEC(1)) {
523 524 uint64_t throughput = data_written / NSEC2MSEC(write_time);
524 525
525 526 if (dp->dp_throughput)
526 527 dp->dp_throughput = throughput / 4 +
527 528 3 * dp->dp_throughput / 4;
528 529 else
529 530 dp->dp_throughput = throughput;
530 531 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
531 532 MAX(zfs_write_limit_min,
532 533 dp->dp_throughput * zfs_txg_synctime_ms));
533 534 }
534 535 }
535 536
536 537 void
537 538 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
538 539 {
539 540 zilog_t *zilog;
540 541 dsl_dataset_t *ds;
541 542
542 543 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
543 544 ds = dmu_objset_ds(zilog->zl_os);
544 545 zil_clean(zilog, txg);
545 546 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
546 547 dmu_buf_rele(ds->ds_dbuf, zilog);
547 548 }
548 549 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
549 550 }
550 551
551 552 /*
552 553 * TRUE if the current thread is the tx_sync_thread or if we
553 554 * are being called from SPA context during pool initialization.
554 555 */
555 556 int
556 557 dsl_pool_sync_context(dsl_pool_t *dp)
557 558 {
558 559 return (curthread == dp->dp_tx.tx_sync_thread ||
559 560 spa_is_initializing(dp->dp_spa));
560 561 }
561 562
562 563 uint64_t
563 564 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
564 565 {
565 566 uint64_t space, resv;
566 567
567 568 /*
568 569 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
569 570 * efficiency.
570 571 * XXX The intent log is not accounted for, so it must fit
571 572 * within this slop.
572 573 *
573 574 * If we're trying to assess whether it's OK to do a free,
574 575 * cut the reservation in half to allow forward progress
575 576 * (e.g. make it possible to rm(1) files from a full pool).
576 577 */
577 578 space = spa_get_dspace(dp->dp_spa);
578 579 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
579 580 if (netfree)
580 581 resv >>= 1;
581 582
582 583 return (space - resv);
583 584 }
584 585
585 586 int
586 587 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
587 588 {
588 589 uint64_t reserved = 0;
589 590 uint64_t write_limit = (zfs_write_limit_override ?
590 591 zfs_write_limit_override : dp->dp_write_limit);
591 592
592 593 if (zfs_no_write_throttle) {
593 594 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
594 595 space);
595 596 return (0);
596 597 }
597 598
598 599 /*
599 600 * Check to see if we have exceeded the maximum allowed IO for
600 601 * this transaction group. We can do this without locks since
601 602 * a little slop here is ok. Note that we do the reserved check
602 603 * with only half the requested reserve: this is because the
603 604 * reserve requests are worst-case, and we really don't want to
604 605 * throttle based off of worst-case estimates.
605 606 */
606 607 if (write_limit > 0) {
607 608 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
608 609 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
609 610
610 611 if (reserved && reserved > write_limit)
611 612 return (SET_ERROR(ERESTART));
612 613 }
613 614
614 615 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
615 616
616 617 /*
617 618 * If this transaction group is over 7/8ths capacity, delay
618 619 * the caller 1 clock tick. This will slow down the "fill"
619 620 * rate until the sync process can catch up with us.
620 621 */
621 622 if (reserved && reserved > (write_limit - (write_limit >> 3))) {
622 623 txg_delay(dp, tx->tx_txg, zfs_throttle_delay,
623 624 zfs_throttle_resolution);
624 625 }
625 626
626 627 return (0);
627 628 }
628 629
629 630 void
630 631 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
631 632 {
632 633 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
633 634 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
634 635 }
635 636
636 637 void
637 638 dsl_pool_memory_pressure(dsl_pool_t *dp)
638 639 {
639 640 uint64_t space_inuse = 0;
640 641 int i;
641 642
642 643 if (dp->dp_write_limit == zfs_write_limit_min)
643 644 return;
644 645
645 646 for (i = 0; i < TXG_SIZE; i++) {
646 647 space_inuse += dp->dp_space_towrite[i];
647 648 space_inuse += dp->dp_tempreserved[i];
648 649 }
649 650 dp->dp_write_limit = MAX(zfs_write_limit_min,
650 651 MIN(dp->dp_write_limit, space_inuse / 4));
651 652 }
652 653
653 654 void
654 655 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
655 656 {
656 657 if (space > 0) {
657 658 mutex_enter(&dp->dp_lock);
658 659 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
659 660 mutex_exit(&dp->dp_lock);
660 661 }
661 662 }
662 663
663 664 /* ARGSUSED */
664 665 static int
665 666 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
666 667 {
667 668 dmu_tx_t *tx = arg;
668 669 dsl_dataset_t *ds, *prev = NULL;
669 670 int err;
670 671
671 672 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
672 673 if (err)
673 674 return (err);
674 675
675 676 while (ds->ds_phys->ds_prev_snap_obj != 0) {
676 677 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
677 678 FTAG, &prev);
678 679 if (err) {
679 680 dsl_dataset_rele(ds, FTAG);
680 681 return (err);
681 682 }
682 683
683 684 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
684 685 break;
685 686 dsl_dataset_rele(ds, FTAG);
686 687 ds = prev;
687 688 prev = NULL;
688 689 }
689 690
690 691 if (prev == NULL) {
691 692 prev = dp->dp_origin_snap;
692 693
693 694 /*
694 695 * The $ORIGIN can't have any data, or the accounting
695 696 * will be wrong.
696 697 */
697 698 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
698 699
699 700 /* The origin doesn't get attached to itself */
700 701 if (ds->ds_object == prev->ds_object) {
701 702 dsl_dataset_rele(ds, FTAG);
702 703 return (0);
703 704 }
704 705
705 706 dmu_buf_will_dirty(ds->ds_dbuf, tx);
706 707 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
707 708 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
708 709
709 710 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
710 711 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
711 712
712 713 dmu_buf_will_dirty(prev->ds_dbuf, tx);
713 714 prev->ds_phys->ds_num_children++;
714 715
715 716 if (ds->ds_phys->ds_next_snap_obj == 0) {
716 717 ASSERT(ds->ds_prev == NULL);
717 718 VERIFY0(dsl_dataset_hold_obj(dp,
718 719 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
719 720 }
720 721 }
721 722
722 723 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
723 724 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
724 725
725 726 if (prev->ds_phys->ds_next_clones_obj == 0) {
726 727 dmu_buf_will_dirty(prev->ds_dbuf, tx);
727 728 prev->ds_phys->ds_next_clones_obj =
728 729 zap_create(dp->dp_meta_objset,
729 730 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
730 731 }
731 732 VERIFY0(zap_add_int(dp->dp_meta_objset,
732 733 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
733 734
734 735 dsl_dataset_rele(ds, FTAG);
735 736 if (prev != dp->dp_origin_snap)
736 737 dsl_dataset_rele(prev, FTAG);
737 738 return (0);
738 739 }
739 740
740 741 void
741 742 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
742 743 {
743 744 ASSERT(dmu_tx_is_syncing(tx));
744 745 ASSERT(dp->dp_origin_snap != NULL);
745 746
746 747 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
747 748 tx, DS_FIND_CHILDREN));
748 749 }
749 750
750 751 /* ARGSUSED */
751 752 static int
752 753 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
753 754 {
754 755 dmu_tx_t *tx = arg;
755 756 objset_t *mos = dp->dp_meta_objset;
756 757
757 758 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
758 759 dsl_dataset_t *origin;
759 760
760 761 VERIFY0(dsl_dataset_hold_obj(dp,
761 762 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
762 763
763 764 if (origin->ds_dir->dd_phys->dd_clones == 0) {
764 765 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
765 766 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
766 767 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
767 768 }
768 769
769 770 VERIFY0(zap_add_int(dp->dp_meta_objset,
770 771 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
771 772
772 773 dsl_dataset_rele(origin, FTAG);
773 774 }
774 775 return (0);
775 776 }
776 777
777 778 void
778 779 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
779 780 {
780 781 ASSERT(dmu_tx_is_syncing(tx));
781 782 uint64_t obj;
782 783
783 784 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
784 785 VERIFY0(dsl_pool_open_special_dir(dp,
785 786 FREE_DIR_NAME, &dp->dp_free_dir));
786 787
787 788 /*
788 789 * We can't use bpobj_alloc(), because spa_version() still
789 790 * returns the old version, and we need a new-version bpobj with
790 791 * subobj support. So call dmu_object_alloc() directly.
791 792 */
792 793 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
793 794 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
794 795 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
795 796 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
796 797 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
797 798
798 799 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
799 800 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
800 801 }
801 802
802 803 void
803 804 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
804 805 {
805 806 uint64_t dsobj;
806 807 dsl_dataset_t *ds;
807 808
808 809 ASSERT(dmu_tx_is_syncing(tx));
809 810 ASSERT(dp->dp_origin_snap == NULL);
810 811 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
811 812
812 813 /* create the origin dir, ds, & snap-ds */
813 814 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
814 815 NULL, 0, kcred, tx);
815 816 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
816 817 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
817 818 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
818 819 dp, &dp->dp_origin_snap));
819 820 dsl_dataset_rele(ds, FTAG);
820 821 }
821 822
822 823 taskq_t *
823 824 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
824 825 {
825 826 return (dp->dp_vnrele_taskq);
826 827 }
827 828
828 829 /*
↓ open down ↓ |
795 lines elided |
↑ open up ↑ |
829 830 * Walk through the pool-wide zap object of temporary snapshot user holds
830 831 * and release them.
831 832 */
832 833 void
833 834 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
834 835 {
835 836 zap_attribute_t za;
836 837 zap_cursor_t zc;
837 838 objset_t *mos = dp->dp_meta_objset;
838 839 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
840 + nvlist_t *holds;
839 841
840 842 if (zapobj == 0)
841 843 return;
842 844 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
843 845
846 + holds = fnvlist_alloc();
847 +
844 848 for (zap_cursor_init(&zc, mos, zapobj);
845 849 zap_cursor_retrieve(&zc, &za) == 0;
846 850 zap_cursor_advance(&zc)) {
847 851 char *htag;
848 - uint64_t dsobj;
852 + nvlist_t *tags;
849 853
850 854 htag = strchr(za.za_name, '-');
851 855 *htag = '\0';
852 856 ++htag;
853 - dsobj = strtonum(za.za_name, NULL);
854 - dsl_dataset_user_release_tmp(dp, dsobj, htag);
857 + if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
858 + tags = fnvlist_alloc();
859 + fnvlist_add_boolean(tags, htag);
860 + fnvlist_add_nvlist(holds, za.za_name, tags);
861 + fnvlist_free(tags);
862 + } else {
863 + fnvlist_add_boolean(tags, htag);
864 + }
855 865 }
866 + dsl_dataset_user_release_tmp(dp, holds);
867 + fnvlist_free(holds);
856 868 zap_cursor_fini(&zc);
857 869 }
858 870
859 871 /*
860 872 * Create the pool-wide zap object for storing temporary snapshot holds.
861 873 */
862 874 void
863 875 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
864 876 {
865 877 objset_t *mos = dp->dp_meta_objset;
866 878
867 879 ASSERT(dp->dp_tmp_userrefs_obj == 0);
868 880 ASSERT(dmu_tx_is_syncing(tx));
869 881
870 882 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
871 883 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
872 884 }
873 885
874 886 static int
875 887 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
876 888 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
877 889 {
878 890 objset_t *mos = dp->dp_meta_objset;
879 891 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
880 892 char *name;
881 893 int error;
882 894
883 895 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
884 896 ASSERT(dmu_tx_is_syncing(tx));
885 897
886 898 /*
887 899 * If the pool was created prior to SPA_VERSION_USERREFS, the
888 900 * zap object for temporary holds might not exist yet.
889 901 */
890 902 if (zapobj == 0) {
891 903 if (holding) {
892 904 dsl_pool_user_hold_create_obj(dp, tx);
893 905 zapobj = dp->dp_tmp_userrefs_obj;
894 906 } else {
895 907 return (SET_ERROR(ENOENT));
896 908 }
897 909 }
898 910
899 911 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
900 912 if (holding)
901 913 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
902 914 else
903 915 error = zap_remove(mos, zapobj, name, tx);
904 916 strfree(name);
905 917
906 918 return (error);
907 919 }
908 920
909 921 /*
910 922 * Add a temporary hold for the given dataset object and tag.
911 923 */
912 924 int
913 925 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
914 926 uint64_t now, dmu_tx_t *tx)
915 927 {
916 928 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
917 929 }
918 930
919 931 /*
920 932 * Release a temporary hold for the given dataset object and tag.
921 933 */
922 934 int
923 935 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
924 936 dmu_tx_t *tx)
925 937 {
926 938 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
927 939 tx, B_FALSE));
928 940 }
929 941
930 942 /*
931 943 * DSL Pool Configuration Lock
932 944 *
933 945 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
934 946 * creation / destruction / rename / property setting). It must be held for
935 947 * read to hold a dataset or dsl_dir. I.e. you must call
936 948 * dsl_pool_config_enter() or dsl_pool_hold() before calling
937 949 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
938 950 * must be held continuously until all datasets and dsl_dirs are released.
939 951 *
940 952 * The only exception to this rule is that if a "long hold" is placed on
941 953 * a dataset, then the dp_config_rwlock may be dropped while the dataset
942 954 * is still held. The long hold will prevent the dataset from being
943 955 * destroyed -- the destroy will fail with EBUSY. A long hold can be
944 956 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
945 957 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
946 958 *
947 959 * Legitimate long-holders (including owners) should be long-running, cancelable
948 960 * tasks that should cause "zfs destroy" to fail. This includes DMU
949 961 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
950 962 * "zfs send", and "zfs diff". There are several other long-holders whose
951 963 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
952 964 *
953 965 * The usual formula for long-holding would be:
954 966 * dsl_pool_hold()
955 967 * dsl_dataset_hold()
956 968 * ... perform checks ...
957 969 * dsl_dataset_long_hold()
958 970 * dsl_pool_rele()
959 971 * ... perform long-running task ...
960 972 * dsl_dataset_long_rele()
961 973 * dsl_dataset_rele()
962 974 *
963 975 * Note that when the long hold is released, the dataset is still held but
964 976 * the pool is not held. The dataset may change arbitrarily during this time
965 977 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
966 978 * dataset except release it.
967 979 *
968 980 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
969 981 * or modifying operations.
970 982 *
971 983 * Modifying operations should generally use dsl_sync_task(). The synctask
972 984 * infrastructure enforces proper locking strategy with respect to the
973 985 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
974 986 *
975 987 * Read-only operations will manually hold the pool, then the dataset, obtain
976 988 * information from the dataset, then release the pool and dataset.
977 989 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
978 990 * hold/rele.
979 991 */
980 992
981 993 int
982 994 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
983 995 {
984 996 spa_t *spa;
985 997 int error;
986 998
987 999 error = spa_open(name, &spa, tag);
988 1000 if (error == 0) {
989 1001 *dp = spa_get_dsl(spa);
990 1002 dsl_pool_config_enter(*dp, tag);
991 1003 }
992 1004 return (error);
993 1005 }
994 1006
995 1007 void
996 1008 dsl_pool_rele(dsl_pool_t *dp, void *tag)
997 1009 {
998 1010 dsl_pool_config_exit(dp, tag);
999 1011 spa_close(dp->dp_spa, tag);
1000 1012 }
1001 1013
1002 1014 void
1003 1015 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1004 1016 {
1005 1017 /*
1006 1018 * We use a "reentrant" reader-writer lock, but not reentrantly.
1007 1019 *
1008 1020 * The rrwlock can (with the track_all flag) track all reading threads,
1009 1021 * which is very useful for debugging which code path failed to release
1010 1022 * the lock, and for verifying that the *current* thread does hold
1011 1023 * the lock.
1012 1024 *
1013 1025 * (Unlike a rwlock, which knows that N threads hold it for
1014 1026 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1015 1027 * if any thread holds it for read, even if this thread doesn't).
1016 1028 */
1017 1029 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1018 1030 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1019 1031 }
1020 1032
1021 1033 void
1022 1034 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1023 1035 {
1024 1036 rrw_exit(&dp->dp_config_rwlock, tag);
1025 1037 }
1026 1038
1027 1039 boolean_t
1028 1040 dsl_pool_config_held(dsl_pool_t *dp)
1029 1041 {
1030 1042 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1031 1043 }
↓ open down ↓ |
166 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX