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_dir.c
+++ new/usr/src/uts/common/fs/zfs/dsl_dir.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 Martin Matuska. All rights reserved.
25 25 */
26 26
27 27 #include <sys/dmu.h>
28 28 #include <sys/dmu_objset.h>
29 29 #include <sys/dmu_tx.h>
30 30 #include <sys/dsl_dataset.h>
31 31 #include <sys/dsl_dir.h>
32 32 #include <sys/dsl_prop.h>
33 33 #include <sys/dsl_synctask.h>
34 34 #include <sys/dsl_deleg.h>
35 35 #include <sys/spa.h>
36 36 #include <sys/metaslab.h>
37 37 #include <sys/zap.h>
38 38 #include <sys/zio.h>
39 39 #include <sys/arc.h>
40 40 #include <sys/sunddi.h>
41 41 #include "zfs_namecheck.h"
42 42
43 43 static uint64_t dsl_dir_space_towrite(dsl_dir_t *dd);
44 44
45 45 /* ARGSUSED */
46 46 static void
47 47 dsl_dir_evict(dmu_buf_t *db, void *arg)
48 48 {
49 49 dsl_dir_t *dd = arg;
50 50 dsl_pool_t *dp = dd->dd_pool;
51 51 int t;
52 52
53 53 for (t = 0; t < TXG_SIZE; t++) {
54 54 ASSERT(!txg_list_member(&dp->dp_dirty_dirs, dd, t));
55 55 ASSERT(dd->dd_tempreserved[t] == 0);
56 56 ASSERT(dd->dd_space_towrite[t] == 0);
57 57 }
58 58
59 59 if (dd->dd_parent)
60 60 dsl_dir_rele(dd->dd_parent, dd);
61 61
62 62 spa_close(dd->dd_pool->dp_spa, dd);
63 63
64 64 /*
65 65 * The props callback list should have been cleaned up by
66 66 * objset_evict().
67 67 */
68 68 list_destroy(&dd->dd_prop_cbs);
69 69 mutex_destroy(&dd->dd_lock);
70 70 kmem_free(dd, sizeof (dsl_dir_t));
71 71 }
72 72
73 73 int
74 74 dsl_dir_hold_obj(dsl_pool_t *dp, uint64_t ddobj,
75 75 const char *tail, void *tag, dsl_dir_t **ddp)
76 76 {
77 77 dmu_buf_t *dbuf;
78 78 dsl_dir_t *dd;
79 79 int err;
80 80
81 81 ASSERT(dsl_pool_config_held(dp));
82 82
83 83 err = dmu_bonus_hold(dp->dp_meta_objset, ddobj, tag, &dbuf);
84 84 if (err != 0)
85 85 return (err);
86 86 dd = dmu_buf_get_user(dbuf);
87 87 #ifdef ZFS_DEBUG
88 88 {
89 89 dmu_object_info_t doi;
90 90 dmu_object_info_from_db(dbuf, &doi);
91 91 ASSERT3U(doi.doi_type, ==, DMU_OT_DSL_DIR);
92 92 ASSERT3U(doi.doi_bonus_size, >=, sizeof (dsl_dir_phys_t));
93 93 }
94 94 #endif
95 95 if (dd == NULL) {
96 96 dsl_dir_t *winner;
97 97
98 98 dd = kmem_zalloc(sizeof (dsl_dir_t), KM_SLEEP);
99 99 dd->dd_object = ddobj;
100 100 dd->dd_dbuf = dbuf;
101 101 dd->dd_pool = dp;
102 102 dd->dd_phys = dbuf->db_data;
103 103 mutex_init(&dd->dd_lock, NULL, MUTEX_DEFAULT, NULL);
104 104
105 105 list_create(&dd->dd_prop_cbs, sizeof (dsl_prop_cb_record_t),
106 106 offsetof(dsl_prop_cb_record_t, cbr_node));
107 107
108 108 dsl_dir_snap_cmtime_update(dd);
109 109
110 110 if (dd->dd_phys->dd_parent_obj) {
111 111 err = dsl_dir_hold_obj(dp, dd->dd_phys->dd_parent_obj,
112 112 NULL, dd, &dd->dd_parent);
113 113 if (err != 0)
114 114 goto errout;
115 115 if (tail) {
116 116 #ifdef ZFS_DEBUG
117 117 uint64_t foundobj;
118 118
119 119 err = zap_lookup(dp->dp_meta_objset,
120 120 dd->dd_parent->dd_phys->dd_child_dir_zapobj,
121 121 tail, sizeof (foundobj), 1, &foundobj);
122 122 ASSERT(err || foundobj == ddobj);
123 123 #endif
124 124 (void) strcpy(dd->dd_myname, tail);
125 125 } else {
126 126 err = zap_value_search(dp->dp_meta_objset,
127 127 dd->dd_parent->dd_phys->dd_child_dir_zapobj,
128 128 ddobj, 0, dd->dd_myname);
129 129 }
130 130 if (err != 0)
131 131 goto errout;
132 132 } else {
133 133 (void) strcpy(dd->dd_myname, spa_name(dp->dp_spa));
134 134 }
135 135
136 136 if (dsl_dir_is_clone(dd)) {
137 137 dmu_buf_t *origin_bonus;
138 138 dsl_dataset_phys_t *origin_phys;
139 139
140 140 /*
141 141 * We can't open the origin dataset, because
142 142 * that would require opening this dsl_dir.
143 143 * Just look at its phys directly instead.
144 144 */
145 145 err = dmu_bonus_hold(dp->dp_meta_objset,
146 146 dd->dd_phys->dd_origin_obj, FTAG, &origin_bonus);
147 147 if (err != 0)
148 148 goto errout;
149 149 origin_phys = origin_bonus->db_data;
150 150 dd->dd_origin_txg =
151 151 origin_phys->ds_creation_txg;
152 152 dmu_buf_rele(origin_bonus, FTAG);
153 153 }
154 154
155 155 winner = dmu_buf_set_user_ie(dbuf, dd, &dd->dd_phys,
156 156 dsl_dir_evict);
157 157 if (winner) {
158 158 if (dd->dd_parent)
159 159 dsl_dir_rele(dd->dd_parent, dd);
160 160 mutex_destroy(&dd->dd_lock);
161 161 kmem_free(dd, sizeof (dsl_dir_t));
162 162 dd = winner;
163 163 } else {
164 164 spa_open_ref(dp->dp_spa, dd);
165 165 }
166 166 }
167 167
168 168 /*
169 169 * The dsl_dir_t has both open-to-close and instantiate-to-evict
170 170 * holds on the spa. We need the open-to-close holds because
171 171 * otherwise the spa_refcnt wouldn't change when we open a
172 172 * dir which the spa also has open, so we could incorrectly
173 173 * think it was OK to unload/export/destroy the pool. We need
174 174 * the instantiate-to-evict hold because the dsl_dir_t has a
175 175 * pointer to the dd_pool, which has a pointer to the spa_t.
176 176 */
177 177 spa_open_ref(dp->dp_spa, tag);
178 178 ASSERT3P(dd->dd_pool, ==, dp);
179 179 ASSERT3U(dd->dd_object, ==, ddobj);
180 180 ASSERT3P(dd->dd_dbuf, ==, dbuf);
181 181 *ddp = dd;
182 182 return (0);
183 183
184 184 errout:
185 185 if (dd->dd_parent)
186 186 dsl_dir_rele(dd->dd_parent, dd);
187 187 mutex_destroy(&dd->dd_lock);
188 188 kmem_free(dd, sizeof (dsl_dir_t));
189 189 dmu_buf_rele(dbuf, tag);
190 190 return (err);
191 191 }
192 192
193 193 void
194 194 dsl_dir_rele(dsl_dir_t *dd, void *tag)
195 195 {
196 196 dprintf_dd(dd, "%s\n", "");
197 197 spa_close(dd->dd_pool->dp_spa, tag);
198 198 dmu_buf_rele(dd->dd_dbuf, tag);
199 199 }
200 200
201 201 /* buf must be long enough (MAXNAMELEN + strlen(MOS_DIR_NAME) + 1 should do) */
202 202 void
203 203 dsl_dir_name(dsl_dir_t *dd, char *buf)
204 204 {
205 205 if (dd->dd_parent) {
206 206 dsl_dir_name(dd->dd_parent, buf);
207 207 (void) strcat(buf, "/");
208 208 } else {
209 209 buf[0] = '\0';
210 210 }
211 211 if (!MUTEX_HELD(&dd->dd_lock)) {
212 212 /*
213 213 * recursive mutex so that we can use
214 214 * dprintf_dd() with dd_lock held
215 215 */
216 216 mutex_enter(&dd->dd_lock);
217 217 (void) strcat(buf, dd->dd_myname);
218 218 mutex_exit(&dd->dd_lock);
219 219 } else {
220 220 (void) strcat(buf, dd->dd_myname);
221 221 }
222 222 }
223 223
224 224 /* Calculate name length, avoiding all the strcat calls of dsl_dir_name */
225 225 int
226 226 dsl_dir_namelen(dsl_dir_t *dd)
227 227 {
228 228 int result = 0;
229 229
230 230 if (dd->dd_parent) {
231 231 /* parent's name + 1 for the "/" */
232 232 result = dsl_dir_namelen(dd->dd_parent) + 1;
233 233 }
234 234
235 235 if (!MUTEX_HELD(&dd->dd_lock)) {
236 236 /* see dsl_dir_name */
237 237 mutex_enter(&dd->dd_lock);
238 238 result += strlen(dd->dd_myname);
239 239 mutex_exit(&dd->dd_lock);
240 240 } else {
241 241 result += strlen(dd->dd_myname);
242 242 }
243 243
244 244 return (result);
245 245 }
246 246
247 247 static int
248 248 getcomponent(const char *path, char *component, const char **nextp)
249 249 {
250 250 char *p;
251 251
252 252 if ((path == NULL) || (path[0] == '\0'))
253 253 return (SET_ERROR(ENOENT));
254 254 /* This would be a good place to reserve some namespace... */
255 255 p = strpbrk(path, "/@");
256 256 if (p && (p[1] == '/' || p[1] == '@')) {
257 257 /* two separators in a row */
258 258 return (SET_ERROR(EINVAL));
259 259 }
260 260 if (p == NULL || p == path) {
261 261 /*
262 262 * if the first thing is an @ or /, it had better be an
263 263 * @ and it had better not have any more ats or slashes,
264 264 * and it had better have something after the @.
265 265 */
266 266 if (p != NULL &&
267 267 (p[0] != '@' || strpbrk(path+1, "/@") || p[1] == '\0'))
268 268 return (SET_ERROR(EINVAL));
269 269 if (strlen(path) >= MAXNAMELEN)
270 270 return (SET_ERROR(ENAMETOOLONG));
271 271 (void) strcpy(component, path);
272 272 p = NULL;
273 273 } else if (p[0] == '/') {
274 274 if (p - path >= MAXNAMELEN)
275 275 return (SET_ERROR(ENAMETOOLONG));
276 276 (void) strncpy(component, path, p - path);
277 277 component[p - path] = '\0';
278 278 p++;
279 279 } else if (p[0] == '@') {
280 280 /*
281 281 * if the next separator is an @, there better not be
282 282 * any more slashes.
283 283 */
284 284 if (strchr(path, '/'))
285 285 return (SET_ERROR(EINVAL));
286 286 if (p - path >= MAXNAMELEN)
287 287 return (SET_ERROR(ENAMETOOLONG));
288 288 (void) strncpy(component, path, p - path);
289 289 component[p - path] = '\0';
290 290 } else {
291 291 panic("invalid p=%p", (void *)p);
292 292 }
293 293 *nextp = p;
294 294 return (0);
295 295 }
296 296
297 297 /*
298 298 * Return the dsl_dir_t, and possibly the last component which couldn't
299 299 * be found in *tail. The name must be in the specified dsl_pool_t. This
300 300 * thread must hold the dp_config_rwlock for the pool. Returns NULL if the
301 301 * path is bogus, or if tail==NULL and we couldn't parse the whole name.
302 302 * (*tail)[0] == '@' means that the last component is a snapshot.
303 303 */
304 304 int
305 305 dsl_dir_hold(dsl_pool_t *dp, const char *name, void *tag,
306 306 dsl_dir_t **ddp, const char **tailp)
307 307 {
308 308 char buf[MAXNAMELEN];
309 309 const char *spaname, *next, *nextnext = NULL;
310 310 int err;
311 311 dsl_dir_t *dd;
312 312 uint64_t ddobj;
313 313
314 314 err = getcomponent(name, buf, &next);
315 315 if (err != 0)
316 316 return (err);
317 317
318 318 /* Make sure the name is in the specified pool. */
319 319 spaname = spa_name(dp->dp_spa);
320 320 if (strcmp(buf, spaname) != 0)
321 321 return (SET_ERROR(EINVAL));
322 322
323 323 ASSERT(dsl_pool_config_held(dp));
324 324
325 325 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, NULL, tag, &dd);
326 326 if (err != 0) {
327 327 return (err);
328 328 }
329 329
330 330 while (next != NULL) {
331 331 dsl_dir_t *child_ds;
332 332 err = getcomponent(next, buf, &nextnext);
333 333 if (err != 0)
334 334 break;
335 335 ASSERT(next[0] != '\0');
336 336 if (next[0] == '@')
337 337 break;
338 338 dprintf("looking up %s in obj%lld\n",
339 339 buf, dd->dd_phys->dd_child_dir_zapobj);
340 340
341 341 err = zap_lookup(dp->dp_meta_objset,
342 342 dd->dd_phys->dd_child_dir_zapobj,
343 343 buf, sizeof (ddobj), 1, &ddobj);
344 344 if (err != 0) {
345 345 if (err == ENOENT)
346 346 err = 0;
347 347 break;
348 348 }
349 349
350 350 err = dsl_dir_hold_obj(dp, ddobj, buf, tag, &child_ds);
351 351 if (err != 0)
352 352 break;
353 353 dsl_dir_rele(dd, tag);
354 354 dd = child_ds;
355 355 next = nextnext;
356 356 }
357 357
358 358 if (err != 0) {
359 359 dsl_dir_rele(dd, tag);
360 360 return (err);
361 361 }
362 362
363 363 /*
364 364 * It's an error if there's more than one component left, or
365 365 * tailp==NULL and there's any component left.
366 366 */
367 367 if (next != NULL &&
368 368 (tailp == NULL || (nextnext && nextnext[0] != '\0'))) {
369 369 /* bad path name */
370 370 dsl_dir_rele(dd, tag);
371 371 dprintf("next=%p (%s) tail=%p\n", next, next?next:"", tailp);
372 372 err = SET_ERROR(ENOENT);
373 373 }
374 374 if (tailp != NULL)
375 375 *tailp = next;
376 376 *ddp = dd;
377 377 return (err);
378 378 }
379 379
380 380 uint64_t
381 381 dsl_dir_create_sync(dsl_pool_t *dp, dsl_dir_t *pds, const char *name,
382 382 dmu_tx_t *tx)
383 383 {
384 384 objset_t *mos = dp->dp_meta_objset;
385 385 uint64_t ddobj;
386 386 dsl_dir_phys_t *ddphys;
387 387 dmu_buf_t *dbuf;
388 388
389 389 ddobj = dmu_object_alloc(mos, DMU_OT_DSL_DIR, 0,
390 390 DMU_OT_DSL_DIR, sizeof (dsl_dir_phys_t), tx);
391 391 if (pds) {
392 392 VERIFY(0 == zap_add(mos, pds->dd_phys->dd_child_dir_zapobj,
393 393 name, sizeof (uint64_t), 1, &ddobj, tx));
394 394 } else {
395 395 /* it's the root dir */
396 396 VERIFY(0 == zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
397 397 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1, &ddobj, tx));
398 398 }
399 399 VERIFY(0 == dmu_bonus_hold(mos, ddobj, FTAG, &dbuf));
400 400 dmu_buf_will_dirty(dbuf, tx);
401 401 ddphys = dbuf->db_data;
402 402
403 403 ddphys->dd_creation_time = gethrestime_sec();
404 404 if (pds)
405 405 ddphys->dd_parent_obj = pds->dd_object;
406 406 ddphys->dd_props_zapobj = zap_create(mos,
407 407 DMU_OT_DSL_PROPS, DMU_OT_NONE, 0, tx);
408 408 ddphys->dd_child_dir_zapobj = zap_create(mos,
409 409 DMU_OT_DSL_DIR_CHILD_MAP, DMU_OT_NONE, 0, tx);
410 410 if (spa_version(dp->dp_spa) >= SPA_VERSION_USED_BREAKDOWN)
411 411 ddphys->dd_flags |= DD_FLAG_USED_BREAKDOWN;
412 412 dmu_buf_rele(dbuf, FTAG);
413 413
414 414 return (ddobj);
415 415 }
416 416
417 417 boolean_t
418 418 dsl_dir_is_clone(dsl_dir_t *dd)
419 419 {
420 420 return (dd->dd_phys->dd_origin_obj &&
421 421 (dd->dd_pool->dp_origin_snap == NULL ||
422 422 dd->dd_phys->dd_origin_obj !=
423 423 dd->dd_pool->dp_origin_snap->ds_object));
424 424 }
425 425
426 426 void
427 427 dsl_dir_stats(dsl_dir_t *dd, nvlist_t *nv)
428 428 {
429 429 mutex_enter(&dd->dd_lock);
430 430 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
431 431 dd->dd_phys->dd_used_bytes);
432 432 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_QUOTA, dd->dd_phys->dd_quota);
433 433 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_RESERVATION,
434 434 dd->dd_phys->dd_reserved);
435 435 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
436 436 dd->dd_phys->dd_compressed_bytes == 0 ? 100 :
437 437 (dd->dd_phys->dd_uncompressed_bytes * 100 /
438 438 dd->dd_phys->dd_compressed_bytes));
439 439 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_LOGICALUSED,
440 440 dd->dd_phys->dd_uncompressed_bytes);
441 441 if (dd->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
442 442 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDSNAP,
443 443 dd->dd_phys->dd_used_breakdown[DD_USED_SNAP]);
444 444 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDDS,
445 445 dd->dd_phys->dd_used_breakdown[DD_USED_HEAD]);
446 446 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDREFRESERV,
447 447 dd->dd_phys->dd_used_breakdown[DD_USED_REFRSRV]);
448 448 dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDCHILD,
449 449 dd->dd_phys->dd_used_breakdown[DD_USED_CHILD] +
450 450 dd->dd_phys->dd_used_breakdown[DD_USED_CHILD_RSRV]);
451 451 }
452 452 mutex_exit(&dd->dd_lock);
453 453
454 454 if (dsl_dir_is_clone(dd)) {
455 455 dsl_dataset_t *ds;
456 456 char buf[MAXNAMELEN];
457 457
458 458 VERIFY0(dsl_dataset_hold_obj(dd->dd_pool,
459 459 dd->dd_phys->dd_origin_obj, FTAG, &ds));
460 460 dsl_dataset_name(ds, buf);
461 461 dsl_dataset_rele(ds, FTAG);
462 462 dsl_prop_nvlist_add_string(nv, ZFS_PROP_ORIGIN, buf);
463 463 }
464 464 }
465 465
466 466 void
467 467 dsl_dir_dirty(dsl_dir_t *dd, dmu_tx_t *tx)
468 468 {
469 469 dsl_pool_t *dp = dd->dd_pool;
470 470
471 471 ASSERT(dd->dd_phys);
472 472
473 473 if (txg_list_add(&dp->dp_dirty_dirs, dd, tx->tx_txg)) {
474 474 /* up the hold count until we can be written out */
475 475 dmu_buf_add_ref(dd->dd_dbuf, dd);
476 476 }
477 477 }
478 478
479 479 static int64_t
480 480 parent_delta(dsl_dir_t *dd, uint64_t used, int64_t delta)
481 481 {
482 482 uint64_t old_accounted = MAX(used, dd->dd_phys->dd_reserved);
483 483 uint64_t new_accounted = MAX(used + delta, dd->dd_phys->dd_reserved);
484 484 return (new_accounted - old_accounted);
485 485 }
486 486
487 487 void
488 488 dsl_dir_sync(dsl_dir_t *dd, dmu_tx_t *tx)
489 489 {
490 490 ASSERT(dmu_tx_is_syncing(tx));
491 491
492 492 mutex_enter(&dd->dd_lock);
493 493 ASSERT0(dd->dd_tempreserved[tx->tx_txg&TXG_MASK]);
494 494 dprintf_dd(dd, "txg=%llu towrite=%lluK\n", tx->tx_txg,
495 495 dd->dd_space_towrite[tx->tx_txg&TXG_MASK] / 1024);
496 496 dd->dd_space_towrite[tx->tx_txg&TXG_MASK] = 0;
497 497 mutex_exit(&dd->dd_lock);
498 498
499 499 /* release the hold from dsl_dir_dirty */
500 500 dmu_buf_rele(dd->dd_dbuf, dd);
501 501 }
502 502
503 503 static uint64_t
504 504 dsl_dir_space_towrite(dsl_dir_t *dd)
505 505 {
506 506 uint64_t space = 0;
507 507 int i;
508 508
509 509 ASSERT(MUTEX_HELD(&dd->dd_lock));
510 510
511 511 for (i = 0; i < TXG_SIZE; i++) {
512 512 space += dd->dd_space_towrite[i&TXG_MASK];
513 513 ASSERT3U(dd->dd_space_towrite[i&TXG_MASK], >=, 0);
514 514 }
515 515 return (space);
516 516 }
517 517
518 518 /*
519 519 * How much space would dd have available if ancestor had delta applied
520 520 * to it? If ondiskonly is set, we're only interested in what's
521 521 * on-disk, not estimated pending changes.
522 522 */
523 523 uint64_t
524 524 dsl_dir_space_available(dsl_dir_t *dd,
525 525 dsl_dir_t *ancestor, int64_t delta, int ondiskonly)
526 526 {
527 527 uint64_t parentspace, myspace, quota, used;
528 528
529 529 /*
530 530 * If there are no restrictions otherwise, assume we have
531 531 * unlimited space available.
532 532 */
533 533 quota = UINT64_MAX;
534 534 parentspace = UINT64_MAX;
535 535
536 536 if (dd->dd_parent != NULL) {
537 537 parentspace = dsl_dir_space_available(dd->dd_parent,
538 538 ancestor, delta, ondiskonly);
539 539 }
540 540
541 541 mutex_enter(&dd->dd_lock);
542 542 if (dd->dd_phys->dd_quota != 0)
543 543 quota = dd->dd_phys->dd_quota;
544 544 used = dd->dd_phys->dd_used_bytes;
545 545 if (!ondiskonly)
546 546 used += dsl_dir_space_towrite(dd);
547 547
548 548 if (dd->dd_parent == NULL) {
549 549 uint64_t poolsize = dsl_pool_adjustedsize(dd->dd_pool, FALSE);
550 550 quota = MIN(quota, poolsize);
551 551 }
552 552
553 553 if (dd->dd_phys->dd_reserved > used && parentspace != UINT64_MAX) {
554 554 /*
555 555 * We have some space reserved, in addition to what our
556 556 * parent gave us.
557 557 */
558 558 parentspace += dd->dd_phys->dd_reserved - used;
559 559 }
560 560
561 561 if (dd == ancestor) {
562 562 ASSERT(delta <= 0);
563 563 ASSERT(used >= -delta);
564 564 used += delta;
565 565 if (parentspace != UINT64_MAX)
566 566 parentspace -= delta;
567 567 }
568 568
569 569 if (used > quota) {
570 570 /* over quota */
571 571 myspace = 0;
572 572 } else {
573 573 /*
574 574 * the lesser of the space provided by our parent and
575 575 * the space left in our quota
576 576 */
|
↓ open down ↓ |
576 lines elided |
↑ open up ↑ |
577 577 myspace = MIN(parentspace, quota - used);
578 578 }
579 579
580 580 mutex_exit(&dd->dd_lock);
581 581
582 582 return (myspace);
583 583 }
584 584
585 585 struct tempreserve {
586 586 list_node_t tr_node;
587 - dsl_pool_t *tr_dp;
588 587 dsl_dir_t *tr_ds;
589 588 uint64_t tr_size;
590 589 };
591 590
592 591 static int
593 592 dsl_dir_tempreserve_impl(dsl_dir_t *dd, uint64_t asize, boolean_t netfree,
594 593 boolean_t ignorequota, boolean_t checkrefquota, list_t *tr_list,
595 594 dmu_tx_t *tx, boolean_t first)
596 595 {
597 596 uint64_t txg = tx->tx_txg;
598 597 uint64_t est_inflight, used_on_disk, quota, parent_rsrv;
599 598 uint64_t deferred = 0;
600 599 struct tempreserve *tr;
601 600 int retval = EDQUOT;
602 601 int txgidx = txg & TXG_MASK;
603 602 int i;
604 603 uint64_t ref_rsrv = 0;
605 604
606 605 ASSERT3U(txg, !=, 0);
607 606 ASSERT3S(asize, >, 0);
608 607
609 608 mutex_enter(&dd->dd_lock);
610 609
611 610 /*
612 611 * Check against the dsl_dir's quota. We don't add in the delta
613 612 * when checking for over-quota because they get one free hit.
614 613 */
615 614 est_inflight = dsl_dir_space_towrite(dd);
616 615 for (i = 0; i < TXG_SIZE; i++)
617 616 est_inflight += dd->dd_tempreserved[i];
618 617 used_on_disk = dd->dd_phys->dd_used_bytes;
619 618
620 619 /*
621 620 * On the first iteration, fetch the dataset's used-on-disk and
622 621 * refreservation values. Also, if checkrefquota is set, test if
623 622 * allocating this space would exceed the dataset's refquota.
624 623 */
625 624 if (first && tx->tx_objset) {
626 625 int error;
627 626 dsl_dataset_t *ds = tx->tx_objset->os_dsl_dataset;
628 627
629 628 error = dsl_dataset_check_quota(ds, checkrefquota,
630 629 asize, est_inflight, &used_on_disk, &ref_rsrv);
631 630 if (error) {
632 631 mutex_exit(&dd->dd_lock);
633 632 return (error);
634 633 }
635 634 }
636 635
637 636 /*
638 637 * If this transaction will result in a net free of space,
639 638 * we want to let it through.
640 639 */
641 640 if (ignorequota || netfree || dd->dd_phys->dd_quota == 0)
642 641 quota = UINT64_MAX;
643 642 else
644 643 quota = dd->dd_phys->dd_quota;
645 644
646 645 /*
647 646 * Adjust the quota against the actual pool size at the root
648 647 * minus any outstanding deferred frees.
649 648 * To ensure that it's possible to remove files from a full
650 649 * pool without inducing transient overcommits, we throttle
651 650 * netfree transactions against a quota that is slightly larger,
652 651 * but still within the pool's allocation slop. In cases where
653 652 * we're very close to full, this will allow a steady trickle of
654 653 * removes to get through.
655 654 */
656 655 if (dd->dd_parent == NULL) {
657 656 spa_t *spa = dd->dd_pool->dp_spa;
658 657 uint64_t poolsize = dsl_pool_adjustedsize(dd->dd_pool, netfree);
659 658 deferred = metaslab_class_get_deferred(spa_normal_class(spa));
660 659 if (poolsize - deferred < quota) {
661 660 quota = poolsize - deferred;
662 661 retval = ENOSPC;
663 662 }
664 663 }
665 664
666 665 /*
667 666 * If they are requesting more space, and our current estimate
668 667 * is over quota, they get to try again unless the actual
669 668 * on-disk is over quota and there are no pending changes (which
670 669 * may free up space for us).
671 670 */
672 671 if (used_on_disk + est_inflight >= quota) {
673 672 if (est_inflight > 0 || used_on_disk < quota ||
674 673 (retval == ENOSPC && used_on_disk < quota + deferred))
675 674 retval = ERESTART;
676 675 dprintf_dd(dd, "failing: used=%lluK inflight = %lluK "
677 676 "quota=%lluK tr=%lluK err=%d\n",
678 677 used_on_disk>>10, est_inflight>>10,
679 678 quota>>10, asize>>10, retval);
680 679 mutex_exit(&dd->dd_lock);
681 680 return (SET_ERROR(retval));
682 681 }
683 682
684 683 /* We need to up our estimated delta before dropping dd_lock */
685 684 dd->dd_tempreserved[txgidx] += asize;
686 685
687 686 parent_rsrv = parent_delta(dd, used_on_disk + est_inflight,
688 687 asize - ref_rsrv);
689 688 mutex_exit(&dd->dd_lock);
690 689
691 690 tr = kmem_zalloc(sizeof (struct tempreserve), KM_SLEEP);
692 691 tr->tr_ds = dd;
693 692 tr->tr_size = asize;
694 693 list_insert_tail(tr_list, tr);
695 694
696 695 /* see if it's OK with our parent */
697 696 if (dd->dd_parent && parent_rsrv) {
698 697 boolean_t ismos = (dd->dd_phys->dd_head_dataset_obj == 0);
699 698
700 699 return (dsl_dir_tempreserve_impl(dd->dd_parent,
701 700 parent_rsrv, netfree, ismos, TRUE, tr_list, tx, FALSE));
702 701 } else {
703 702 return (0);
704 703 }
705 704 }
706 705
707 706 /*
708 707 * Reserve space in this dsl_dir, to be used in this tx's txg.
709 708 * After the space has been dirtied (and dsl_dir_willuse_space()
710 709 * has been called), the reservation should be canceled, using
711 710 * dsl_dir_tempreserve_clear().
712 711 */
713 712 int
714 713 dsl_dir_tempreserve_space(dsl_dir_t *dd, uint64_t lsize, uint64_t asize,
715 714 uint64_t fsize, uint64_t usize, void **tr_cookiep, dmu_tx_t *tx)
716 715 {
717 716 int err;
718 717 list_t *tr_list;
719 718
720 719 if (asize == 0) {
721 720 *tr_cookiep = NULL;
722 721 return (0);
723 722 }
724 723
725 724 tr_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
726 725 list_create(tr_list, sizeof (struct tempreserve),
727 726 offsetof(struct tempreserve, tr_node));
|
↓ open down ↓ |
130 lines elided |
↑ open up ↑ |
728 727 ASSERT3S(asize, >, 0);
729 728 ASSERT3S(fsize, >=, 0);
730 729
731 730 err = arc_tempreserve_space(lsize, tx->tx_txg);
732 731 if (err == 0) {
733 732 struct tempreserve *tr;
734 733
735 734 tr = kmem_zalloc(sizeof (struct tempreserve), KM_SLEEP);
736 735 tr->tr_size = lsize;
737 736 list_insert_tail(tr_list, tr);
738 -
739 - err = dsl_pool_tempreserve_space(dd->dd_pool, asize, tx);
740 737 } else {
741 738 if (err == EAGAIN) {
739 + /*
740 + * If arc_memory_throttle() detected that pageout
741 + * is running and we are low on memory, we delay new
742 + * non-pageout transactions to give pageout an
743 + * advantage.
744 + *
745 + * It is unfortunate to be delaying while the caller's
746 + * locks are held.
747 + */
742 748 txg_delay(dd->dd_pool, tx->tx_txg,
743 749 MSEC2NSEC(10), MSEC2NSEC(10));
744 750 err = SET_ERROR(ERESTART);
745 751 }
746 - dsl_pool_memory_pressure(dd->dd_pool);
747 752 }
748 753
749 754 if (err == 0) {
750 - struct tempreserve *tr;
751 -
752 - tr = kmem_zalloc(sizeof (struct tempreserve), KM_SLEEP);
753 - tr->tr_dp = dd->dd_pool;
754 - tr->tr_size = asize;
755 - list_insert_tail(tr_list, tr);
756 -
757 755 err = dsl_dir_tempreserve_impl(dd, asize, fsize >= asize,
758 756 FALSE, asize > usize, tr_list, tx, TRUE);
759 757 }
760 758
761 759 if (err != 0)
762 760 dsl_dir_tempreserve_clear(tr_list, tx);
763 761 else
764 762 *tr_cookiep = tr_list;
765 763
766 764 return (err);
767 765 }
768 766
769 767 /*
770 768 * Clear a temporary reservation that we previously made with
771 769 * dsl_dir_tempreserve_space().
772 770 */
773 771 void
774 772 dsl_dir_tempreserve_clear(void *tr_cookie, dmu_tx_t *tx)
|
↓ open down ↓ |
8 lines elided |
↑ open up ↑ |
775 773 {
776 774 int txgidx = tx->tx_txg & TXG_MASK;
777 775 list_t *tr_list = tr_cookie;
778 776 struct tempreserve *tr;
779 777
780 778 ASSERT3U(tx->tx_txg, !=, 0);
781 779
782 780 if (tr_cookie == NULL)
783 781 return;
784 782
785 - while (tr = list_head(tr_list)) {
786 - if (tr->tr_dp) {
787 - dsl_pool_tempreserve_clear(tr->tr_dp, tr->tr_size, tx);
788 - } else if (tr->tr_ds) {
783 + while ((tr = list_head(tr_list)) != NULL) {
784 + if (tr->tr_ds) {
789 785 mutex_enter(&tr->tr_ds->dd_lock);
790 786 ASSERT3U(tr->tr_ds->dd_tempreserved[txgidx], >=,
791 787 tr->tr_size);
792 788 tr->tr_ds->dd_tempreserved[txgidx] -= tr->tr_size;
793 789 mutex_exit(&tr->tr_ds->dd_lock);
794 790 } else {
795 791 arc_tempreserve_clear(tr->tr_size);
796 792 }
797 793 list_remove(tr_list, tr);
798 794 kmem_free(tr, sizeof (struct tempreserve));
799 795 }
800 796
801 797 kmem_free(tr_list, sizeof (list_t));
802 798 }
803 799
804 -static void
805 -dsl_dir_willuse_space_impl(dsl_dir_t *dd, int64_t space, dmu_tx_t *tx)
800 +/*
801 + * This should be called from open context when we think we're going to write
802 + * or free space, for example when dirtying data. Be conservative; it's okay
803 + * to write less space or free more, but we don't want to write more or free
804 + * less than the amount specified.
805 + */
806 +void
807 +dsl_dir_willuse_space(dsl_dir_t *dd, int64_t space, dmu_tx_t *tx)
806 808 {
807 809 int64_t parent_space;
808 810 uint64_t est_used;
809 811
810 812 mutex_enter(&dd->dd_lock);
811 813 if (space > 0)
812 814 dd->dd_space_towrite[tx->tx_txg & TXG_MASK] += space;
813 815
814 816 est_used = dsl_dir_space_towrite(dd) + dd->dd_phys->dd_used_bytes;
815 817 parent_space = parent_delta(dd, est_used, space);
816 818 mutex_exit(&dd->dd_lock);
817 819
818 820 /* Make sure that we clean up dd_space_to* */
819 821 dsl_dir_dirty(dd, tx);
820 822
821 823 /* XXX this is potentially expensive and unnecessary... */
822 824 if (parent_space && dd->dd_parent)
823 - dsl_dir_willuse_space_impl(dd->dd_parent, parent_space, tx);
825 + dsl_dir_willuse_space(dd->dd_parent, parent_space, tx);
824 826 }
825 827
826 -/*
827 - * Call in open context when we think we're going to write/free space,
828 - * eg. when dirtying data. Be conservative (ie. OK to write less than
829 - * this or free more than this, but don't write more or free less).
830 - */
831 -void
832 -dsl_dir_willuse_space(dsl_dir_t *dd, int64_t space, dmu_tx_t *tx)
833 -{
834 - dsl_pool_willuse_space(dd->dd_pool, space, tx);
835 - dsl_dir_willuse_space_impl(dd, space, tx);
836 -}
837 -
838 828 /* call from syncing context when we actually write/free space for this dd */
839 829 void
840 830 dsl_dir_diduse_space(dsl_dir_t *dd, dd_used_t type,
841 831 int64_t used, int64_t compressed, int64_t uncompressed, dmu_tx_t *tx)
842 832 {
843 833 int64_t accounted_delta;
844 834
845 835 /*
846 836 * dsl_dataset_set_refreservation_sync_impl() calls this with
847 837 * dd_lock held, so that it can atomically update
848 838 * ds->ds_reserved and the dsl_dir accounting, so that
849 839 * dsl_dataset_check_quota() can see dataset and dir accounting
850 840 * consistently.
851 841 */
852 842 boolean_t needlock = !MUTEX_HELD(&dd->dd_lock);
853 843
854 844 ASSERT(dmu_tx_is_syncing(tx));
855 845 ASSERT(type < DD_USED_NUM);
856 846
857 847 dmu_buf_will_dirty(dd->dd_dbuf, tx);
858 848
859 849 if (needlock)
860 850 mutex_enter(&dd->dd_lock);
861 851 accounted_delta = parent_delta(dd, dd->dd_phys->dd_used_bytes, used);
862 852 ASSERT(used >= 0 || dd->dd_phys->dd_used_bytes >= -used);
863 853 ASSERT(compressed >= 0 ||
864 854 dd->dd_phys->dd_compressed_bytes >= -compressed);
865 855 ASSERT(uncompressed >= 0 ||
866 856 dd->dd_phys->dd_uncompressed_bytes >= -uncompressed);
867 857 dd->dd_phys->dd_used_bytes += used;
868 858 dd->dd_phys->dd_uncompressed_bytes += uncompressed;
869 859 dd->dd_phys->dd_compressed_bytes += compressed;
870 860
871 861 if (dd->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
872 862 ASSERT(used > 0 ||
873 863 dd->dd_phys->dd_used_breakdown[type] >= -used);
874 864 dd->dd_phys->dd_used_breakdown[type] += used;
875 865 #ifdef DEBUG
876 866 dd_used_t t;
877 867 uint64_t u = 0;
878 868 for (t = 0; t < DD_USED_NUM; t++)
879 869 u += dd->dd_phys->dd_used_breakdown[t];
880 870 ASSERT3U(u, ==, dd->dd_phys->dd_used_bytes);
881 871 #endif
882 872 }
883 873 if (needlock)
884 874 mutex_exit(&dd->dd_lock);
885 875
886 876 if (dd->dd_parent != NULL) {
887 877 dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD,
888 878 accounted_delta, compressed, uncompressed, tx);
889 879 dsl_dir_transfer_space(dd->dd_parent,
890 880 used - accounted_delta,
891 881 DD_USED_CHILD_RSRV, DD_USED_CHILD, tx);
892 882 }
893 883 }
894 884
895 885 void
896 886 dsl_dir_transfer_space(dsl_dir_t *dd, int64_t delta,
897 887 dd_used_t oldtype, dd_used_t newtype, dmu_tx_t *tx)
898 888 {
899 889 ASSERT(dmu_tx_is_syncing(tx));
900 890 ASSERT(oldtype < DD_USED_NUM);
901 891 ASSERT(newtype < DD_USED_NUM);
902 892
903 893 if (delta == 0 || !(dd->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN))
904 894 return;
905 895
906 896 dmu_buf_will_dirty(dd->dd_dbuf, tx);
907 897 mutex_enter(&dd->dd_lock);
908 898 ASSERT(delta > 0 ?
909 899 dd->dd_phys->dd_used_breakdown[oldtype] >= delta :
910 900 dd->dd_phys->dd_used_breakdown[newtype] >= -delta);
911 901 ASSERT(dd->dd_phys->dd_used_bytes >= ABS(delta));
912 902 dd->dd_phys->dd_used_breakdown[oldtype] -= delta;
913 903 dd->dd_phys->dd_used_breakdown[newtype] += delta;
914 904 mutex_exit(&dd->dd_lock);
915 905 }
916 906
917 907 typedef struct dsl_dir_set_qr_arg {
918 908 const char *ddsqra_name;
919 909 zprop_source_t ddsqra_source;
920 910 uint64_t ddsqra_value;
921 911 } dsl_dir_set_qr_arg_t;
922 912
923 913 static int
924 914 dsl_dir_set_quota_check(void *arg, dmu_tx_t *tx)
925 915 {
926 916 dsl_dir_set_qr_arg_t *ddsqra = arg;
927 917 dsl_pool_t *dp = dmu_tx_pool(tx);
928 918 dsl_dataset_t *ds;
929 919 int error;
930 920 uint64_t towrite, newval;
931 921
932 922 error = dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds);
933 923 if (error != 0)
934 924 return (error);
935 925
936 926 error = dsl_prop_predict(ds->ds_dir, "quota",
937 927 ddsqra->ddsqra_source, ddsqra->ddsqra_value, &newval);
938 928 if (error != 0) {
939 929 dsl_dataset_rele(ds, FTAG);
940 930 return (error);
941 931 }
942 932
943 933 if (newval == 0) {
944 934 dsl_dataset_rele(ds, FTAG);
945 935 return (0);
946 936 }
947 937
948 938 mutex_enter(&ds->ds_dir->dd_lock);
949 939 /*
950 940 * If we are doing the preliminary check in open context, and
951 941 * there are pending changes, then don't fail it, since the
952 942 * pending changes could under-estimate the amount of space to be
953 943 * freed up.
954 944 */
955 945 towrite = dsl_dir_space_towrite(ds->ds_dir);
956 946 if ((dmu_tx_is_syncing(tx) || towrite == 0) &&
957 947 (newval < ds->ds_dir->dd_phys->dd_reserved ||
958 948 newval < ds->ds_dir->dd_phys->dd_used_bytes + towrite)) {
959 949 error = SET_ERROR(ENOSPC);
960 950 }
961 951 mutex_exit(&ds->ds_dir->dd_lock);
962 952 dsl_dataset_rele(ds, FTAG);
963 953 return (error);
964 954 }
965 955
966 956 static void
967 957 dsl_dir_set_quota_sync(void *arg, dmu_tx_t *tx)
968 958 {
969 959 dsl_dir_set_qr_arg_t *ddsqra = arg;
970 960 dsl_pool_t *dp = dmu_tx_pool(tx);
971 961 dsl_dataset_t *ds;
972 962 uint64_t newval;
973 963
974 964 VERIFY0(dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds));
975 965
976 966 if (spa_version(dp->dp_spa) >= SPA_VERSION_RECVD_PROPS) {
977 967 dsl_prop_set_sync_impl(ds, zfs_prop_to_name(ZFS_PROP_QUOTA),
978 968 ddsqra->ddsqra_source, sizeof (ddsqra->ddsqra_value), 1,
979 969 &ddsqra->ddsqra_value, tx);
980 970
981 971 VERIFY0(dsl_prop_get_int_ds(ds,
982 972 zfs_prop_to_name(ZFS_PROP_QUOTA), &newval));
983 973 } else {
984 974 newval = ddsqra->ddsqra_value;
985 975 spa_history_log_internal_ds(ds, "set", tx, "%s=%lld",
986 976 zfs_prop_to_name(ZFS_PROP_QUOTA), (longlong_t)newval);
987 977 }
988 978
989 979 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
990 980 mutex_enter(&ds->ds_dir->dd_lock);
991 981 ds->ds_dir->dd_phys->dd_quota = newval;
992 982 mutex_exit(&ds->ds_dir->dd_lock);
993 983 dsl_dataset_rele(ds, FTAG);
994 984 }
995 985
996 986 int
997 987 dsl_dir_set_quota(const char *ddname, zprop_source_t source, uint64_t quota)
998 988 {
999 989 dsl_dir_set_qr_arg_t ddsqra;
1000 990
1001 991 ddsqra.ddsqra_name = ddname;
1002 992 ddsqra.ddsqra_source = source;
1003 993 ddsqra.ddsqra_value = quota;
1004 994
1005 995 return (dsl_sync_task(ddname, dsl_dir_set_quota_check,
1006 996 dsl_dir_set_quota_sync, &ddsqra, 0));
1007 997 }
1008 998
1009 999 int
1010 1000 dsl_dir_set_reservation_check(void *arg, dmu_tx_t *tx)
1011 1001 {
1012 1002 dsl_dir_set_qr_arg_t *ddsqra = arg;
1013 1003 dsl_pool_t *dp = dmu_tx_pool(tx);
1014 1004 dsl_dataset_t *ds;
1015 1005 dsl_dir_t *dd;
1016 1006 uint64_t newval, used, avail;
1017 1007 int error;
1018 1008
1019 1009 error = dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds);
1020 1010 if (error != 0)
1021 1011 return (error);
1022 1012 dd = ds->ds_dir;
1023 1013
1024 1014 /*
1025 1015 * If we are doing the preliminary check in open context, the
1026 1016 * space estimates may be inaccurate.
1027 1017 */
1028 1018 if (!dmu_tx_is_syncing(tx)) {
1029 1019 dsl_dataset_rele(ds, FTAG);
1030 1020 return (0);
1031 1021 }
1032 1022
1033 1023 error = dsl_prop_predict(ds->ds_dir,
1034 1024 zfs_prop_to_name(ZFS_PROP_RESERVATION),
1035 1025 ddsqra->ddsqra_source, ddsqra->ddsqra_value, &newval);
1036 1026 if (error != 0) {
1037 1027 dsl_dataset_rele(ds, FTAG);
1038 1028 return (error);
1039 1029 }
1040 1030
1041 1031 mutex_enter(&dd->dd_lock);
1042 1032 used = dd->dd_phys->dd_used_bytes;
1043 1033 mutex_exit(&dd->dd_lock);
1044 1034
1045 1035 if (dd->dd_parent) {
1046 1036 avail = dsl_dir_space_available(dd->dd_parent,
1047 1037 NULL, 0, FALSE);
1048 1038 } else {
1049 1039 avail = dsl_pool_adjustedsize(dd->dd_pool, B_FALSE) - used;
1050 1040 }
1051 1041
1052 1042 if (MAX(used, newval) > MAX(used, dd->dd_phys->dd_reserved)) {
1053 1043 uint64_t delta = MAX(used, newval) -
1054 1044 MAX(used, dd->dd_phys->dd_reserved);
1055 1045
1056 1046 if (delta > avail ||
1057 1047 (dd->dd_phys->dd_quota > 0 &&
1058 1048 newval > dd->dd_phys->dd_quota))
1059 1049 error = SET_ERROR(ENOSPC);
1060 1050 }
1061 1051
1062 1052 dsl_dataset_rele(ds, FTAG);
1063 1053 return (error);
1064 1054 }
1065 1055
1066 1056 void
1067 1057 dsl_dir_set_reservation_sync_impl(dsl_dir_t *dd, uint64_t value, dmu_tx_t *tx)
1068 1058 {
1069 1059 uint64_t used;
1070 1060 int64_t delta;
1071 1061
1072 1062 dmu_buf_will_dirty(dd->dd_dbuf, tx);
1073 1063
1074 1064 mutex_enter(&dd->dd_lock);
1075 1065 used = dd->dd_phys->dd_used_bytes;
1076 1066 delta = MAX(used, value) - MAX(used, dd->dd_phys->dd_reserved);
1077 1067 dd->dd_phys->dd_reserved = value;
1078 1068
1079 1069 if (dd->dd_parent != NULL) {
1080 1070 /* Roll up this additional usage into our ancestors */
1081 1071 dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD_RSRV,
1082 1072 delta, 0, 0, tx);
1083 1073 }
1084 1074 mutex_exit(&dd->dd_lock);
1085 1075 }
1086 1076
1087 1077
1088 1078 static void
1089 1079 dsl_dir_set_reservation_sync(void *arg, dmu_tx_t *tx)
1090 1080 {
1091 1081 dsl_dir_set_qr_arg_t *ddsqra = arg;
1092 1082 dsl_pool_t *dp = dmu_tx_pool(tx);
1093 1083 dsl_dataset_t *ds;
1094 1084 uint64_t newval;
1095 1085
1096 1086 VERIFY0(dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds));
1097 1087
1098 1088 if (spa_version(dp->dp_spa) >= SPA_VERSION_RECVD_PROPS) {
1099 1089 dsl_prop_set_sync_impl(ds,
1100 1090 zfs_prop_to_name(ZFS_PROP_RESERVATION),
1101 1091 ddsqra->ddsqra_source, sizeof (ddsqra->ddsqra_value), 1,
1102 1092 &ddsqra->ddsqra_value, tx);
1103 1093
1104 1094 VERIFY0(dsl_prop_get_int_ds(ds,
1105 1095 zfs_prop_to_name(ZFS_PROP_RESERVATION), &newval));
1106 1096 } else {
1107 1097 newval = ddsqra->ddsqra_value;
1108 1098 spa_history_log_internal_ds(ds, "set", tx, "%s=%lld",
1109 1099 zfs_prop_to_name(ZFS_PROP_RESERVATION),
1110 1100 (longlong_t)newval);
1111 1101 }
1112 1102
1113 1103 dsl_dir_set_reservation_sync_impl(ds->ds_dir, newval, tx);
1114 1104 dsl_dataset_rele(ds, FTAG);
1115 1105 }
1116 1106
1117 1107 int
1118 1108 dsl_dir_set_reservation(const char *ddname, zprop_source_t source,
1119 1109 uint64_t reservation)
1120 1110 {
1121 1111 dsl_dir_set_qr_arg_t ddsqra;
1122 1112
1123 1113 ddsqra.ddsqra_name = ddname;
1124 1114 ddsqra.ddsqra_source = source;
1125 1115 ddsqra.ddsqra_value = reservation;
1126 1116
1127 1117 return (dsl_sync_task(ddname, dsl_dir_set_reservation_check,
1128 1118 dsl_dir_set_reservation_sync, &ddsqra, 0));
1129 1119 }
1130 1120
1131 1121 static dsl_dir_t *
1132 1122 closest_common_ancestor(dsl_dir_t *ds1, dsl_dir_t *ds2)
1133 1123 {
1134 1124 for (; ds1; ds1 = ds1->dd_parent) {
1135 1125 dsl_dir_t *dd;
1136 1126 for (dd = ds2; dd; dd = dd->dd_parent) {
1137 1127 if (ds1 == dd)
1138 1128 return (dd);
1139 1129 }
1140 1130 }
1141 1131 return (NULL);
1142 1132 }
1143 1133
1144 1134 /*
1145 1135 * If delta is applied to dd, how much of that delta would be applied to
1146 1136 * ancestor? Syncing context only.
1147 1137 */
1148 1138 static int64_t
1149 1139 would_change(dsl_dir_t *dd, int64_t delta, dsl_dir_t *ancestor)
1150 1140 {
1151 1141 if (dd == ancestor)
1152 1142 return (delta);
1153 1143
1154 1144 mutex_enter(&dd->dd_lock);
1155 1145 delta = parent_delta(dd, dd->dd_phys->dd_used_bytes, delta);
1156 1146 mutex_exit(&dd->dd_lock);
1157 1147 return (would_change(dd->dd_parent, delta, ancestor));
1158 1148 }
1159 1149
1160 1150 typedef struct dsl_dir_rename_arg {
1161 1151 const char *ddra_oldname;
1162 1152 const char *ddra_newname;
1163 1153 } dsl_dir_rename_arg_t;
1164 1154
1165 1155 /* ARGSUSED */
1166 1156 static int
1167 1157 dsl_valid_rename(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
1168 1158 {
1169 1159 int *deltap = arg;
1170 1160 char namebuf[MAXNAMELEN];
1171 1161
1172 1162 dsl_dataset_name(ds, namebuf);
1173 1163
1174 1164 if (strlen(namebuf) + *deltap >= MAXNAMELEN)
1175 1165 return (SET_ERROR(ENAMETOOLONG));
1176 1166 return (0);
1177 1167 }
1178 1168
1179 1169 static int
1180 1170 dsl_dir_rename_check(void *arg, dmu_tx_t *tx)
1181 1171 {
1182 1172 dsl_dir_rename_arg_t *ddra = arg;
1183 1173 dsl_pool_t *dp = dmu_tx_pool(tx);
1184 1174 dsl_dir_t *dd, *newparent;
1185 1175 const char *mynewname;
1186 1176 int error;
1187 1177 int delta = strlen(ddra->ddra_newname) - strlen(ddra->ddra_oldname);
1188 1178
1189 1179 /* target dir should exist */
1190 1180 error = dsl_dir_hold(dp, ddra->ddra_oldname, FTAG, &dd, NULL);
1191 1181 if (error != 0)
1192 1182 return (error);
1193 1183
1194 1184 /* new parent should exist */
1195 1185 error = dsl_dir_hold(dp, ddra->ddra_newname, FTAG,
1196 1186 &newparent, &mynewname);
1197 1187 if (error != 0) {
1198 1188 dsl_dir_rele(dd, FTAG);
1199 1189 return (error);
1200 1190 }
1201 1191
1202 1192 /* can't rename to different pool */
1203 1193 if (dd->dd_pool != newparent->dd_pool) {
1204 1194 dsl_dir_rele(newparent, FTAG);
1205 1195 dsl_dir_rele(dd, FTAG);
1206 1196 return (SET_ERROR(ENXIO));
1207 1197 }
1208 1198
1209 1199 /* new name should not already exist */
1210 1200 if (mynewname == NULL) {
1211 1201 dsl_dir_rele(newparent, FTAG);
1212 1202 dsl_dir_rele(dd, FTAG);
1213 1203 return (SET_ERROR(EEXIST));
1214 1204 }
1215 1205
1216 1206 /* if the name length is growing, validate child name lengths */
1217 1207 if (delta > 0) {
1218 1208 error = dmu_objset_find_dp(dp, dd->dd_object, dsl_valid_rename,
1219 1209 &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
1220 1210 if (error != 0) {
1221 1211 dsl_dir_rele(newparent, FTAG);
1222 1212 dsl_dir_rele(dd, FTAG);
1223 1213 return (error);
1224 1214 }
1225 1215 }
1226 1216
1227 1217 if (newparent != dd->dd_parent) {
1228 1218 /* is there enough space? */
1229 1219 uint64_t myspace =
1230 1220 MAX(dd->dd_phys->dd_used_bytes, dd->dd_phys->dd_reserved);
1231 1221
1232 1222 /* no rename into our descendant */
1233 1223 if (closest_common_ancestor(dd, newparent) == dd) {
1234 1224 dsl_dir_rele(newparent, FTAG);
1235 1225 dsl_dir_rele(dd, FTAG);
1236 1226 return (SET_ERROR(EINVAL));
1237 1227 }
1238 1228
1239 1229 error = dsl_dir_transfer_possible(dd->dd_parent,
1240 1230 newparent, myspace);
1241 1231 if (error != 0) {
1242 1232 dsl_dir_rele(newparent, FTAG);
1243 1233 dsl_dir_rele(dd, FTAG);
1244 1234 return (error);
1245 1235 }
1246 1236 }
1247 1237
1248 1238 dsl_dir_rele(newparent, FTAG);
1249 1239 dsl_dir_rele(dd, FTAG);
1250 1240 return (0);
1251 1241 }
1252 1242
1253 1243 static void
1254 1244 dsl_dir_rename_sync(void *arg, dmu_tx_t *tx)
1255 1245 {
1256 1246 dsl_dir_rename_arg_t *ddra = arg;
1257 1247 dsl_pool_t *dp = dmu_tx_pool(tx);
1258 1248 dsl_dir_t *dd, *newparent;
1259 1249 const char *mynewname;
1260 1250 int error;
1261 1251 objset_t *mos = dp->dp_meta_objset;
1262 1252
1263 1253 VERIFY0(dsl_dir_hold(dp, ddra->ddra_oldname, FTAG, &dd, NULL));
1264 1254 VERIFY0(dsl_dir_hold(dp, ddra->ddra_newname, FTAG, &newparent,
1265 1255 &mynewname));
1266 1256
1267 1257 /* Log this before we change the name. */
1268 1258 spa_history_log_internal_dd(dd, "rename", tx,
1269 1259 "-> %s", ddra->ddra_newname);
1270 1260
1271 1261 if (newparent != dd->dd_parent) {
1272 1262 dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD,
1273 1263 -dd->dd_phys->dd_used_bytes,
1274 1264 -dd->dd_phys->dd_compressed_bytes,
1275 1265 -dd->dd_phys->dd_uncompressed_bytes, tx);
1276 1266 dsl_dir_diduse_space(newparent, DD_USED_CHILD,
1277 1267 dd->dd_phys->dd_used_bytes,
1278 1268 dd->dd_phys->dd_compressed_bytes,
1279 1269 dd->dd_phys->dd_uncompressed_bytes, tx);
1280 1270
1281 1271 if (dd->dd_phys->dd_reserved > dd->dd_phys->dd_used_bytes) {
1282 1272 uint64_t unused_rsrv = dd->dd_phys->dd_reserved -
1283 1273 dd->dd_phys->dd_used_bytes;
1284 1274
1285 1275 dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD_RSRV,
1286 1276 -unused_rsrv, 0, 0, tx);
1287 1277 dsl_dir_diduse_space(newparent, DD_USED_CHILD_RSRV,
1288 1278 unused_rsrv, 0, 0, tx);
1289 1279 }
1290 1280 }
1291 1281
1292 1282 dmu_buf_will_dirty(dd->dd_dbuf, tx);
1293 1283
1294 1284 /* remove from old parent zapobj */
1295 1285 error = zap_remove(mos, dd->dd_parent->dd_phys->dd_child_dir_zapobj,
1296 1286 dd->dd_myname, tx);
1297 1287 ASSERT0(error);
1298 1288
1299 1289 (void) strcpy(dd->dd_myname, mynewname);
1300 1290 dsl_dir_rele(dd->dd_parent, dd);
1301 1291 dd->dd_phys->dd_parent_obj = newparent->dd_object;
1302 1292 VERIFY0(dsl_dir_hold_obj(dp,
1303 1293 newparent->dd_object, NULL, dd, &dd->dd_parent));
1304 1294
1305 1295 /* add to new parent zapobj */
1306 1296 VERIFY0(zap_add(mos, newparent->dd_phys->dd_child_dir_zapobj,
1307 1297 dd->dd_myname, 8, 1, &dd->dd_object, tx));
1308 1298
1309 1299 dsl_prop_notify_all(dd);
1310 1300
1311 1301 dsl_dir_rele(newparent, FTAG);
1312 1302 dsl_dir_rele(dd, FTAG);
1313 1303 }
1314 1304
1315 1305 int
1316 1306 dsl_dir_rename(const char *oldname, const char *newname)
1317 1307 {
1318 1308 dsl_dir_rename_arg_t ddra;
1319 1309
1320 1310 ddra.ddra_oldname = oldname;
1321 1311 ddra.ddra_newname = newname;
1322 1312
1323 1313 return (dsl_sync_task(oldname,
1324 1314 dsl_dir_rename_check, dsl_dir_rename_sync, &ddra, 3));
1325 1315 }
1326 1316
1327 1317 int
1328 1318 dsl_dir_transfer_possible(dsl_dir_t *sdd, dsl_dir_t *tdd, uint64_t space)
1329 1319 {
1330 1320 dsl_dir_t *ancestor;
1331 1321 int64_t adelta;
1332 1322 uint64_t avail;
1333 1323
1334 1324 ancestor = closest_common_ancestor(sdd, tdd);
1335 1325 adelta = would_change(sdd, -space, ancestor);
1336 1326 avail = dsl_dir_space_available(tdd, ancestor, adelta, FALSE);
1337 1327 if (avail < space)
1338 1328 return (SET_ERROR(ENOSPC));
1339 1329
1340 1330 return (0);
1341 1331 }
1342 1332
1343 1333 timestruc_t
1344 1334 dsl_dir_snap_cmtime(dsl_dir_t *dd)
1345 1335 {
1346 1336 timestruc_t t;
1347 1337
1348 1338 mutex_enter(&dd->dd_lock);
1349 1339 t = dd->dd_snap_cmtime;
1350 1340 mutex_exit(&dd->dd_lock);
1351 1341
1352 1342 return (t);
1353 1343 }
1354 1344
1355 1345 void
1356 1346 dsl_dir_snap_cmtime_update(dsl_dir_t *dd)
1357 1347 {
1358 1348 timestruc_t t;
1359 1349
1360 1350 gethrestime(&t);
1361 1351 mutex_enter(&dd->dd_lock);
1362 1352 dd->dd_snap_cmtime = t;
1363 1353 mutex_exit(&dd->dd_lock);
1364 1354 }
|
↓ open down ↓ |
517 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX