Print this page
4188 assertion failed in dmu_tx_hold_free(): dn_datablkshift != 0
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/dmu_tx.c
+++ new/usr/src/uts/common/fs/zfs/dmu_tx.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 2011 Nexenta Systems, Inc. All rights reserved.
24 24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 25 */
26 26
27 27 #include <sys/dmu.h>
28 28 #include <sys/dmu_impl.h>
29 29 #include <sys/dbuf.h>
30 30 #include <sys/dmu_tx.h>
31 31 #include <sys/dmu_objset.h>
32 32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 34 #include <sys/dsl_pool.h>
35 35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 36 #include <sys/spa.h>
37 37 #include <sys/sa.h>
38 38 #include <sys/sa_impl.h>
39 39 #include <sys/zfs_context.h>
40 40 #include <sys/varargs.h>
41 41
42 42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43 43 uint64_t arg1, uint64_t arg2);
44 44
45 45
46 46 dmu_tx_t *
47 47 dmu_tx_create_dd(dsl_dir_t *dd)
48 48 {
49 49 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 50 tx->tx_dir = dd;
51 51 if (dd != NULL)
52 52 tx->tx_pool = dd->dd_pool;
53 53 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
54 54 offsetof(dmu_tx_hold_t, txh_node));
55 55 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
56 56 offsetof(dmu_tx_callback_t, dcb_node));
57 57 tx->tx_start = gethrtime();
58 58 #ifdef ZFS_DEBUG
59 59 refcount_create(&tx->tx_space_written);
60 60 refcount_create(&tx->tx_space_freed);
61 61 #endif
62 62 return (tx);
63 63 }
64 64
65 65 dmu_tx_t *
66 66 dmu_tx_create(objset_t *os)
67 67 {
68 68 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 69 tx->tx_objset = os;
70 70 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
71 71 return (tx);
72 72 }
73 73
74 74 dmu_tx_t *
75 75 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 76 {
77 77 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78 78
79 79 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
80 80 tx->tx_pool = dp;
81 81 tx->tx_txg = txg;
82 82 tx->tx_anyobj = TRUE;
83 83
84 84 return (tx);
85 85 }
86 86
87 87 int
88 88 dmu_tx_is_syncing(dmu_tx_t *tx)
89 89 {
90 90 return (tx->tx_anyobj);
91 91 }
92 92
93 93 int
94 94 dmu_tx_private_ok(dmu_tx_t *tx)
95 95 {
96 96 return (tx->tx_anyobj);
97 97 }
98 98
99 99 static dmu_tx_hold_t *
100 100 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
101 101 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
102 102 {
103 103 dmu_tx_hold_t *txh;
104 104 dnode_t *dn = NULL;
105 105 int err;
106 106
107 107 if (object != DMU_NEW_OBJECT) {
108 108 err = dnode_hold(os, object, tx, &dn);
109 109 if (err) {
110 110 tx->tx_err = err;
111 111 return (NULL);
112 112 }
113 113
114 114 if (err == 0 && tx->tx_txg != 0) {
115 115 mutex_enter(&dn->dn_mtx);
116 116 /*
117 117 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
118 118 * problem, but there's no way for it to happen (for
119 119 * now, at least).
120 120 */
121 121 ASSERT(dn->dn_assigned_txg == 0);
122 122 dn->dn_assigned_txg = tx->tx_txg;
123 123 (void) refcount_add(&dn->dn_tx_holds, tx);
124 124 mutex_exit(&dn->dn_mtx);
125 125 }
126 126 }
127 127
128 128 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 129 txh->txh_tx = tx;
130 130 txh->txh_dnode = dn;
131 131 #ifdef ZFS_DEBUG
132 132 txh->txh_type = type;
133 133 txh->txh_arg1 = arg1;
134 134 txh->txh_arg2 = arg2;
135 135 #endif
136 136 list_insert_tail(&tx->tx_holds, txh);
137 137
138 138 return (txh);
139 139 }
140 140
141 141 void
142 142 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
143 143 {
144 144 /*
145 145 * If we're syncing, they can manipulate any object anyhow, and
146 146 * the hold on the dnode_t can cause problems.
147 147 */
148 148 if (!dmu_tx_is_syncing(tx)) {
149 149 (void) dmu_tx_hold_object_impl(tx, os,
150 150 object, THT_NEWOBJECT, 0, 0);
151 151 }
152 152 }
153 153
154 154 static int
155 155 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
156 156 {
157 157 int err;
158 158 dmu_buf_impl_t *db;
159 159
160 160 rw_enter(&dn->dn_struct_rwlock, RW_READER);
161 161 db = dbuf_hold_level(dn, level, blkid, FTAG);
162 162 rw_exit(&dn->dn_struct_rwlock);
163 163 if (db == NULL)
164 164 return (SET_ERROR(EIO));
165 165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 166 dbuf_rele(db, FTAG);
167 167 return (err);
168 168 }
169 169
170 170 static void
171 171 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
172 172 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 173 {
174 174 objset_t *os = dn->dn_objset;
175 175 dsl_dataset_t *ds = os->os_dsl_dataset;
176 176 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
177 177 dmu_buf_impl_t *parent = NULL;
178 178 blkptr_t *bp = NULL;
179 179 uint64_t space;
180 180
181 181 if (level >= dn->dn_nlevels || history[level] == blkid)
182 182 return;
183 183
184 184 history[level] = blkid;
185 185
186 186 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187 187
188 188 if (db == NULL || db == dn->dn_dbuf) {
189 189 ASSERT(level != 0);
190 190 db = NULL;
191 191 } else {
192 192 ASSERT(DB_DNODE(db) == dn);
193 193 ASSERT(db->db_level == level);
194 194 ASSERT(db->db.db_size == space);
195 195 ASSERT(db->db_blkid == blkid);
196 196 bp = db->db_blkptr;
197 197 parent = db->db_parent;
198 198 }
199 199
200 200 freeable = (bp && (freeable ||
201 201 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
202 202
203 203 if (freeable)
204 204 txh->txh_space_tooverwrite += space;
205 205 else
206 206 txh->txh_space_towrite += space;
207 207 if (bp)
208 208 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209 209
210 210 dmu_tx_count_twig(txh, dn, parent, level + 1,
211 211 blkid >> epbs, freeable, history);
212 212 }
213 213
214 214 /* ARGSUSED */
215 215 static void
216 216 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 217 {
218 218 dnode_t *dn = txh->txh_dnode;
219 219 uint64_t start, end, i;
220 220 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
221 221 int err = 0;
222 222
223 223 if (len == 0)
224 224 return;
225 225
226 226 min_bs = SPA_MINBLOCKSHIFT;
227 227 max_bs = SPA_MAXBLOCKSHIFT;
228 228 min_ibs = DN_MIN_INDBLKSHIFT;
229 229 max_ibs = DN_MAX_INDBLKSHIFT;
230 230
231 231 if (dn) {
232 232 uint64_t history[DN_MAX_LEVELS];
233 233 int nlvls = dn->dn_nlevels;
234 234 int delta;
235 235
236 236 /*
237 237 * For i/o error checking, read the first and last level-0
238 238 * blocks (if they are not aligned), and all the level-1 blocks.
239 239 */
240 240 if (dn->dn_maxblkid == 0) {
241 241 delta = dn->dn_datablksz;
242 242 start = (off < dn->dn_datablksz) ? 0 : 1;
243 243 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
244 244 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
245 245 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
246 246 if (err)
247 247 goto out;
248 248 delta -= off;
249 249 }
250 250 } else {
251 251 zio_t *zio = zio_root(dn->dn_objset->os_spa,
252 252 NULL, NULL, ZIO_FLAG_CANFAIL);
253 253
254 254 /* first level-0 block */
255 255 start = off >> dn->dn_datablkshift;
256 256 if (P2PHASE(off, dn->dn_datablksz) ||
257 257 len < dn->dn_datablksz) {
258 258 err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 259 if (err)
260 260 goto out;
261 261 }
262 262
263 263 /* last level-0 block */
264 264 end = (off+len-1) >> dn->dn_datablkshift;
265 265 if (end != start && end <= dn->dn_maxblkid &&
266 266 P2PHASE(off+len, dn->dn_datablksz)) {
267 267 err = dmu_tx_check_ioerr(zio, dn, 0, end);
268 268 if (err)
269 269 goto out;
270 270 }
271 271
272 272 /* level-1 blocks */
273 273 if (nlvls > 1) {
274 274 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
275 275 for (i = (start>>shft)+1; i < end>>shft; i++) {
276 276 err = dmu_tx_check_ioerr(zio, dn, 1, i);
277 277 if (err)
278 278 goto out;
279 279 }
280 280 }
281 281
282 282 err = zio_wait(zio);
283 283 if (err)
284 284 goto out;
285 285 delta = P2NPHASE(off, dn->dn_datablksz);
286 286 }
287 287
288 288 min_ibs = max_ibs = dn->dn_indblkshift;
289 289 if (dn->dn_maxblkid > 0) {
290 290 /*
291 291 * The blocksize can't change,
292 292 * so we can make a more precise estimate.
293 293 */
294 294 ASSERT(dn->dn_datablkshift != 0);
295 295 min_bs = max_bs = dn->dn_datablkshift;
296 296 }
297 297
298 298 /*
299 299 * If this write is not off the end of the file
300 300 * we need to account for overwrites/unref.
301 301 */
302 302 if (start <= dn->dn_maxblkid) {
303 303 for (int l = 0; l < DN_MAX_LEVELS; l++)
304 304 history[l] = -1ULL;
305 305 }
306 306 while (start <= dn->dn_maxblkid) {
307 307 dmu_buf_impl_t *db;
308 308
309 309 rw_enter(&dn->dn_struct_rwlock, RW_READER);
310 310 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
311 311 rw_exit(&dn->dn_struct_rwlock);
312 312
313 313 if (err) {
314 314 txh->txh_tx->tx_err = err;
315 315 return;
316 316 }
317 317
318 318 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
319 319 history);
320 320 dbuf_rele(db, FTAG);
321 321 if (++start > end) {
322 322 /*
323 323 * Account for new indirects appearing
324 324 * before this IO gets assigned into a txg.
325 325 */
326 326 bits = 64 - min_bs;
327 327 epbs = min_ibs - SPA_BLKPTRSHIFT;
328 328 for (bits -= epbs * (nlvls - 1);
329 329 bits >= 0; bits -= epbs)
330 330 txh->txh_fudge += 1ULL << max_ibs;
331 331 goto out;
332 332 }
333 333 off += delta;
334 334 if (len >= delta)
335 335 len -= delta;
336 336 delta = dn->dn_datablksz;
337 337 }
338 338 }
339 339
340 340 /*
341 341 * 'end' is the last thing we will access, not one past.
342 342 * This way we won't overflow when accessing the last byte.
343 343 */
344 344 start = P2ALIGN(off, 1ULL << max_bs);
345 345 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
346 346 txh->txh_space_towrite += end - start + 1;
347 347
348 348 start >>= min_bs;
349 349 end >>= min_bs;
350 350
351 351 epbs = min_ibs - SPA_BLKPTRSHIFT;
352 352
353 353 /*
354 354 * The object contains at most 2^(64 - min_bs) blocks,
355 355 * and each indirect level maps 2^epbs.
356 356 */
357 357 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
358 358 start >>= epbs;
359 359 end >>= epbs;
360 360 ASSERT3U(end, >=, start);
361 361 txh->txh_space_towrite += (end - start + 1) << max_ibs;
362 362 if (start != 0) {
363 363 /*
364 364 * We also need a new blkid=0 indirect block
365 365 * to reference any existing file data.
366 366 */
367 367 txh->txh_space_towrite += 1ULL << max_ibs;
368 368 }
369 369 }
370 370
371 371 out:
372 372 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
373 373 2 * DMU_MAX_ACCESS)
374 374 err = SET_ERROR(EFBIG);
375 375
376 376 if (err)
377 377 txh->txh_tx->tx_err = err;
378 378 }
379 379
380 380 static void
381 381 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
382 382 {
383 383 dnode_t *dn = txh->txh_dnode;
384 384 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
385 385 uint64_t space = mdn->dn_datablksz +
386 386 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
387 387
388 388 if (dn && dn->dn_dbuf->db_blkptr &&
389 389 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
390 390 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
391 391 txh->txh_space_tooverwrite += space;
392 392 txh->txh_space_tounref += space;
393 393 } else {
394 394 txh->txh_space_towrite += space;
395 395 if (dn && dn->dn_dbuf->db_blkptr)
396 396 txh->txh_space_tounref += space;
397 397 }
398 398 }
399 399
400 400 void
401 401 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
402 402 {
403 403 dmu_tx_hold_t *txh;
404 404
405 405 ASSERT(tx->tx_txg == 0);
406 406 ASSERT(len < DMU_MAX_ACCESS);
407 407 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
408 408
409 409 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
410 410 object, THT_WRITE, off, len);
411 411 if (txh == NULL)
412 412 return;
413 413
414 414 dmu_tx_count_write(txh, off, len);
415 415 dmu_tx_count_dnode(txh);
416 416 }
417 417
418 418 static void
419 419 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
420 420 {
421 421 uint64_t blkid, nblks, lastblk;
422 422 uint64_t space = 0, unref = 0, skipped = 0;
423 423 dnode_t *dn = txh->txh_dnode;
424 424 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
425 425 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
426 426 int epbs;
427 427 uint64_t l0span = 0, nl1blks = 0;
428 428
429 429 if (dn->dn_nlevels == 0)
430 430 return;
431 431
432 432 /*
433 433 * The struct_rwlock protects us against dn_nlevels
434 434 * changing, in case (against all odds) we manage to dirty &
435 435 * sync out the changes after we check for being dirty.
436 436 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
437 437 */
438 438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
439 439 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
440 440 if (dn->dn_maxblkid == 0) {
441 441 if (off == 0 && len >= dn->dn_datablksz) {
442 442 blkid = 0;
443 443 nblks = 1;
444 444 } else {
445 445 rw_exit(&dn->dn_struct_rwlock);
446 446 return;
447 447 }
448 448 } else {
449 449 blkid = off >> dn->dn_datablkshift;
450 450 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
451 451
452 452 if (blkid > dn->dn_maxblkid) {
453 453 rw_exit(&dn->dn_struct_rwlock);
454 454 return;
455 455 }
456 456 if (blkid + nblks > dn->dn_maxblkid)
457 457 nblks = dn->dn_maxblkid - blkid + 1;
458 458
459 459 }
460 460 l0span = nblks; /* save for later use to calc level > 1 overhead */
461 461 if (dn->dn_nlevels == 1) {
462 462 int i;
463 463 for (i = 0; i < nblks; i++) {
464 464 blkptr_t *bp = dn->dn_phys->dn_blkptr;
465 465 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
466 466 bp += blkid + i;
467 467 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
468 468 dprintf_bp(bp, "can free old%s", "");
469 469 space += bp_get_dsize(spa, bp);
470 470 }
471 471 unref += BP_GET_ASIZE(bp);
472 472 }
473 473 nl1blks = 1;
474 474 nblks = 0;
475 475 }
476 476
477 477 lastblk = blkid + nblks - 1;
478 478 while (nblks) {
479 479 dmu_buf_impl_t *dbuf;
480 480 uint64_t ibyte, new_blkid;
481 481 int epb = 1 << epbs;
482 482 int err, i, blkoff, tochk;
483 483 blkptr_t *bp;
484 484
485 485 ibyte = blkid << dn->dn_datablkshift;
486 486 err = dnode_next_offset(dn,
487 487 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
488 488 new_blkid = ibyte >> dn->dn_datablkshift;
489 489 if (err == ESRCH) {
490 490 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
491 491 break;
492 492 }
493 493 if (err) {
494 494 txh->txh_tx->tx_err = err;
495 495 break;
496 496 }
497 497 if (new_blkid > lastblk) {
498 498 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
499 499 break;
500 500 }
501 501
502 502 if (new_blkid > blkid) {
503 503 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
504 504 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
505 505 nblks -= new_blkid - blkid;
506 506 blkid = new_blkid;
507 507 }
508 508 blkoff = P2PHASE(blkid, epb);
509 509 tochk = MIN(epb - blkoff, nblks);
510 510
511 511 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
512 512 if (err) {
513 513 txh->txh_tx->tx_err = err;
514 514 break;
515 515 }
516 516
517 517 txh->txh_memory_tohold += dbuf->db.db_size;
518 518
519 519 /*
520 520 * We don't check memory_tohold against DMU_MAX_ACCESS because
521 521 * memory_tohold is an over-estimation (especially the >L1
522 522 * indirect blocks), so it could fail. Callers should have
523 523 * already verified that they will not be holding too much
524 524 * memory.
525 525 */
526 526
527 527 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
528 528 if (err != 0) {
529 529 txh->txh_tx->tx_err = err;
530 530 dbuf_rele(dbuf, FTAG);
531 531 break;
532 532 }
533 533
534 534 bp = dbuf->db.db_data;
535 535 bp += blkoff;
536 536
537 537 for (i = 0; i < tochk; i++) {
538 538 if (dsl_dataset_block_freeable(ds, &bp[i],
539 539 bp[i].blk_birth)) {
540 540 dprintf_bp(&bp[i], "can free old%s", "");
541 541 space += bp_get_dsize(spa, &bp[i]);
542 542 }
543 543 unref += BP_GET_ASIZE(bp);
544 544 }
545 545 dbuf_rele(dbuf, FTAG);
546 546
547 547 ++nl1blks;
548 548 blkid += tochk;
549 549 nblks -= tochk;
550 550 }
551 551 rw_exit(&dn->dn_struct_rwlock);
552 552
553 553 /*
554 554 * Add in memory requirements of higher-level indirects.
555 555 * This assumes a worst-possible scenario for dn_nlevels and a
556 556 * worst-possible distribution of l1-blocks over the region to free.
557 557 */
558 558 {
559 559 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
560 560 int level = 2;
561 561 /*
562 562 * Here we don't use DN_MAX_LEVEL, but calculate it with the
563 563 * given datablkshift and indblkshift. This makes the
564 564 * difference between 19 and 8 on large files.
565 565 */
566 566 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
567 567 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
568 568
569 569 while (level++ < maxlevel) {
570 570 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
571 571 << dn->dn_indblkshift;
572 572 blkcnt = 1 + (blkcnt >> epbs);
573 573 }
574 574 }
575 575
576 576 /* account for new level 1 indirect blocks that might show up */
577 577 if (skipped > 0) {
578 578 txh->txh_fudge += skipped << dn->dn_indblkshift;
579 579 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
580 580 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
581 581 }
582 582 txh->txh_space_tofree += space;
583 583 txh->txh_space_tounref += unref;
584 584 }
585 585
586 586 void
587 587 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
588 588 {
589 589 dmu_tx_hold_t *txh;
590 590 dnode_t *dn;
591 591 int err;
592 592 zio_t *zio;
593 593
594 594 ASSERT(tx->tx_txg == 0);
595 595
596 596 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
597 597 object, THT_FREE, off, len);
598 598 if (txh == NULL)
599 599 return;
600 600 dn = txh->txh_dnode;
601 601 dmu_tx_count_dnode(txh);
602 602
603 603 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
604 604 return;
605 605 if (len == DMU_OBJECT_END)
606 606 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
607 607
608 608 /*
609 609 * For i/o error checking, we read the first and last level-0
610 610 * blocks if they are not aligned, and all the level-1 blocks.
611 611 *
612 612 * Note: dbuf_free_range() assumes that we have not instantiated
613 613 * any level-0 dbufs that will be completely freed. Therefore we must
614 614 * exercise care to not read or count the first and last blocks
615 615 * if they are blocksize-aligned.
616 616 */
617 617 if (dn->dn_datablkshift == 0) {
618 618 if (off != 0 || len < dn->dn_datablksz)
619 619 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
620 620 } else {
621 621 /* first block will be modified if it is not aligned */
622 622 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
623 623 dmu_tx_count_write(txh, off, 1);
624 624 /* last block will be modified if it is not aligned */
625 625 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
626 626 dmu_tx_count_write(txh, off+len, 1);
627 627 }
|
↓ open down ↓ |
627 lines elided |
↑ open up ↑ |
628 628
629 629 /*
630 630 * Check level-1 blocks.
631 631 */
632 632 if (dn->dn_nlevels > 1) {
633 633 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
634 634 SPA_BLKPTRSHIFT;
635 635 uint64_t start = off >> shift;
636 636 uint64_t end = (off + len) >> shift;
637 637
638 - ASSERT(dn->dn_datablkshift != 0);
639 638 ASSERT(dn->dn_indblkshift != 0);
640 639
640 + /*
641 + * dnode_reallocate() can result in an object with indirect
642 + * blocks having an odd data block size. In this case,
643 + * just check the single block.
644 + */
645 + if (dn->dn_datablkshift == 0)
646 + start = end = 0;
647 +
641 648 zio = zio_root(tx->tx_pool->dp_spa,
642 649 NULL, NULL, ZIO_FLAG_CANFAIL);
643 650 for (uint64_t i = start; i <= end; i++) {
644 651 uint64_t ibyte = i << shift;
645 652 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
646 653 i = ibyte >> shift;
647 654 if (err == ESRCH)
648 655 break;
649 656 if (err) {
650 657 tx->tx_err = err;
651 658 return;
652 659 }
653 660
654 661 err = dmu_tx_check_ioerr(zio, dn, 1, i);
655 662 if (err) {
656 663 tx->tx_err = err;
657 664 return;
658 665 }
659 666 }
660 667 err = zio_wait(zio);
661 668 if (err) {
662 669 tx->tx_err = err;
663 670 return;
664 671 }
665 672 }
666 673
667 674 dmu_tx_count_free(txh, off, len);
668 675 }
669 676
670 677 void
671 678 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
672 679 {
673 680 dmu_tx_hold_t *txh;
674 681 dnode_t *dn;
675 682 uint64_t nblocks;
676 683 int epbs, err;
677 684
678 685 ASSERT(tx->tx_txg == 0);
679 686
680 687 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
681 688 object, THT_ZAP, add, (uintptr_t)name);
682 689 if (txh == NULL)
683 690 return;
684 691 dn = txh->txh_dnode;
685 692
686 693 dmu_tx_count_dnode(txh);
687 694
688 695 if (dn == NULL) {
689 696 /*
690 697 * We will be able to fit a new object's entries into one leaf
691 698 * block. So there will be at most 2 blocks total,
692 699 * including the header block.
693 700 */
694 701 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
695 702 return;
696 703 }
697 704
698 705 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
699 706
700 707 if (dn->dn_maxblkid == 0 && !add) {
701 708 blkptr_t *bp;
702 709
703 710 /*
704 711 * If there is only one block (i.e. this is a micro-zap)
705 712 * and we are not adding anything, the accounting is simple.
706 713 */
707 714 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
708 715 if (err) {
709 716 tx->tx_err = err;
710 717 return;
711 718 }
712 719
713 720 /*
714 721 * Use max block size here, since we don't know how much
715 722 * the size will change between now and the dbuf dirty call.
716 723 */
717 724 bp = &dn->dn_phys->dn_blkptr[0];
718 725 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
719 726 bp, bp->blk_birth))
720 727 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
721 728 else
722 729 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
723 730 if (!BP_IS_HOLE(bp))
724 731 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
725 732 return;
726 733 }
727 734
728 735 if (dn->dn_maxblkid > 0 && name) {
729 736 /*
730 737 * access the name in this fat-zap so that we'll check
731 738 * for i/o errors to the leaf blocks, etc.
732 739 */
733 740 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
734 741 8, 0, NULL);
735 742 if (err == EIO) {
736 743 tx->tx_err = err;
737 744 return;
738 745 }
739 746 }
740 747
741 748 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
742 749 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
743 750
744 751 /*
745 752 * If the modified blocks are scattered to the four winds,
746 753 * we'll have to modify an indirect twig for each.
747 754 */
748 755 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
749 756 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
750 757 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
751 758 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
752 759 else
753 760 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
754 761 }
755 762
756 763 void
757 764 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
758 765 {
759 766 dmu_tx_hold_t *txh;
760 767
761 768 ASSERT(tx->tx_txg == 0);
762 769
763 770 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
764 771 object, THT_BONUS, 0, 0);
765 772 if (txh)
766 773 dmu_tx_count_dnode(txh);
767 774 }
768 775
769 776 void
770 777 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
771 778 {
772 779 dmu_tx_hold_t *txh;
773 780 ASSERT(tx->tx_txg == 0);
774 781
775 782 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
776 783 DMU_NEW_OBJECT, THT_SPACE, space, 0);
777 784
778 785 txh->txh_space_towrite += space;
779 786 }
780 787
781 788 int
782 789 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
783 790 {
784 791 dmu_tx_hold_t *txh;
785 792 int holds = 0;
786 793
787 794 /*
788 795 * By asserting that the tx is assigned, we're counting the
789 796 * number of dn_tx_holds, which is the same as the number of
790 797 * dn_holds. Otherwise, we'd be counting dn_holds, but
791 798 * dn_tx_holds could be 0.
792 799 */
793 800 ASSERT(tx->tx_txg != 0);
794 801
795 802 /* if (tx->tx_anyobj == TRUE) */
796 803 /* return (0); */
797 804
798 805 for (txh = list_head(&tx->tx_holds); txh;
799 806 txh = list_next(&tx->tx_holds, txh)) {
800 807 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
801 808 holds++;
802 809 }
803 810
804 811 return (holds);
805 812 }
806 813
807 814 #ifdef ZFS_DEBUG
808 815 void
809 816 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
810 817 {
811 818 dmu_tx_hold_t *txh;
812 819 int match_object = FALSE, match_offset = FALSE;
813 820 dnode_t *dn;
814 821
815 822 DB_DNODE_ENTER(db);
816 823 dn = DB_DNODE(db);
817 824 ASSERT(tx->tx_txg != 0);
818 825 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
819 826 ASSERT3U(dn->dn_object, ==, db->db.db_object);
820 827
821 828 if (tx->tx_anyobj) {
822 829 DB_DNODE_EXIT(db);
823 830 return;
824 831 }
825 832
826 833 /* XXX No checking on the meta dnode for now */
827 834 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
828 835 DB_DNODE_EXIT(db);
829 836 return;
830 837 }
831 838
832 839 for (txh = list_head(&tx->tx_holds); txh;
833 840 txh = list_next(&tx->tx_holds, txh)) {
834 841 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
835 842 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
836 843 match_object = TRUE;
837 844 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
838 845 int datablkshift = dn->dn_datablkshift ?
839 846 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
840 847 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
841 848 int shift = datablkshift + epbs * db->db_level;
842 849 uint64_t beginblk = shift >= 64 ? 0 :
843 850 (txh->txh_arg1 >> shift);
844 851 uint64_t endblk = shift >= 64 ? 0 :
845 852 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
846 853 uint64_t blkid = db->db_blkid;
847 854
848 855 /* XXX txh_arg2 better not be zero... */
849 856
850 857 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
851 858 txh->txh_type, beginblk, endblk);
852 859
853 860 switch (txh->txh_type) {
854 861 case THT_WRITE:
855 862 if (blkid >= beginblk && blkid <= endblk)
856 863 match_offset = TRUE;
857 864 /*
858 865 * We will let this hold work for the bonus
859 866 * or spill buffer so that we don't need to
860 867 * hold it when creating a new object.
861 868 */
862 869 if (blkid == DMU_BONUS_BLKID ||
863 870 blkid == DMU_SPILL_BLKID)
864 871 match_offset = TRUE;
865 872 /*
866 873 * They might have to increase nlevels,
867 874 * thus dirtying the new TLIBs. Or the
868 875 * might have to change the block size,
869 876 * thus dirying the new lvl=0 blk=0.
870 877 */
871 878 if (blkid == 0)
872 879 match_offset = TRUE;
873 880 break;
874 881 case THT_FREE:
875 882 /*
876 883 * We will dirty all the level 1 blocks in
877 884 * the free range and perhaps the first and
878 885 * last level 0 block.
879 886 */
880 887 if (blkid >= beginblk && (blkid <= endblk ||
881 888 txh->txh_arg2 == DMU_OBJECT_END))
882 889 match_offset = TRUE;
883 890 break;
884 891 case THT_SPILL:
885 892 if (blkid == DMU_SPILL_BLKID)
886 893 match_offset = TRUE;
887 894 break;
888 895 case THT_BONUS:
889 896 if (blkid == DMU_BONUS_BLKID)
890 897 match_offset = TRUE;
891 898 break;
892 899 case THT_ZAP:
893 900 match_offset = TRUE;
894 901 break;
895 902 case THT_NEWOBJECT:
896 903 match_object = TRUE;
897 904 break;
898 905 default:
899 906 ASSERT(!"bad txh_type");
900 907 }
901 908 }
902 909 if (match_object && match_offset) {
903 910 DB_DNODE_EXIT(db);
904 911 return;
905 912 }
906 913 }
907 914 DB_DNODE_EXIT(db);
908 915 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
909 916 (u_longlong_t)db->db.db_object, db->db_level,
910 917 (u_longlong_t)db->db_blkid);
911 918 }
912 919 #endif
913 920
914 921 /*
915 922 * If we can't do 10 iops, something is wrong. Let us go ahead
916 923 * and hit zfs_dirty_data_max.
917 924 */
918 925 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
919 926 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
920 927
921 928 /*
922 929 * We delay transactions when we've determined that the backend storage
923 930 * isn't able to accommodate the rate of incoming writes.
924 931 *
925 932 * If there is already a transaction waiting, we delay relative to when
926 933 * that transaction finishes waiting. This way the calculated min_time
927 934 * is independent of the number of threads concurrently executing
928 935 * transactions.
929 936 *
930 937 * If we are the only waiter, wait relative to when the transaction
931 938 * started, rather than the current time. This credits the transaction for
932 939 * "time already served", e.g. reading indirect blocks.
933 940 *
934 941 * The minimum time for a transaction to take is calculated as:
935 942 * min_time = scale * (dirty - min) / (max - dirty)
936 943 * min_time is then capped at zfs_delay_max_ns.
937 944 *
938 945 * The delay has two degrees of freedom that can be adjusted via tunables.
939 946 * The percentage of dirty data at which we start to delay is defined by
940 947 * zfs_delay_min_dirty_percent. This should typically be at or above
941 948 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
942 949 * delay after writing at full speed has failed to keep up with the incoming
943 950 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
944 951 * speaking, this variable determines the amount of delay at the midpoint of
945 952 * the curve.
946 953 *
947 954 * delay
948 955 * 10ms +-------------------------------------------------------------*+
949 956 * | *|
950 957 * 9ms + *+
951 958 * | *|
952 959 * 8ms + *+
953 960 * | * |
954 961 * 7ms + * +
955 962 * | * |
956 963 * 6ms + * +
957 964 * | * |
958 965 * 5ms + * +
959 966 * | * |
960 967 * 4ms + * +
961 968 * | * |
962 969 * 3ms + * +
963 970 * | * |
964 971 * 2ms + (midpoint) * +
965 972 * | | ** |
966 973 * 1ms + v *** +
967 974 * | zfs_delay_scale ----------> ******** |
968 975 * 0 +-------------------------------------*********----------------+
969 976 * 0% <- zfs_dirty_data_max -> 100%
970 977 *
971 978 * Note that since the delay is added to the outstanding time remaining on the
972 979 * most recent transaction, the delay is effectively the inverse of IOPS.
973 980 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
974 981 * was chosen such that small changes in the amount of accumulated dirty data
975 982 * in the first 3/4 of the curve yield relatively small differences in the
976 983 * amount of delay.
977 984 *
978 985 * The effects can be easier to understand when the amount of delay is
979 986 * represented on a log scale:
980 987 *
981 988 * delay
982 989 * 100ms +-------------------------------------------------------------++
983 990 * + +
984 991 * | |
985 992 * + *+
986 993 * 10ms + *+
987 994 * + ** +
988 995 * | (midpoint) ** |
989 996 * + | ** +
990 997 * 1ms + v **** +
991 998 * + zfs_delay_scale ----------> ***** +
992 999 * | **** |
993 1000 * + **** +
994 1001 * 100us + ** +
995 1002 * + * +
996 1003 * | * |
997 1004 * + * +
998 1005 * 10us + * +
999 1006 * + +
1000 1007 * | |
1001 1008 * + +
1002 1009 * +--------------------------------------------------------------+
1003 1010 * 0% <- zfs_dirty_data_max -> 100%
1004 1011 *
1005 1012 * Note here that only as the amount of dirty data approaches its limit does
1006 1013 * the delay start to increase rapidly. The goal of a properly tuned system
1007 1014 * should be to keep the amount of dirty data out of that range by first
1008 1015 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1009 1016 * optimal throughput on the backend storage, and then by changing the value
1010 1017 * of zfs_delay_scale to increase the steepness of the curve.
1011 1018 */
1012 1019 static void
1013 1020 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1014 1021 {
1015 1022 dsl_pool_t *dp = tx->tx_pool;
1016 1023 uint64_t delay_min_bytes =
1017 1024 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1018 1025 hrtime_t wakeup, min_tx_time, now;
1019 1026
1020 1027 if (dirty <= delay_min_bytes)
1021 1028 return;
1022 1029
1023 1030 /*
1024 1031 * The caller has already waited until we are under the max.
1025 1032 * We make them pass us the amount of dirty data so we don't
1026 1033 * have to handle the case of it being >= the max, which could
1027 1034 * cause a divide-by-zero if it's == the max.
1028 1035 */
1029 1036 ASSERT3U(dirty, <, zfs_dirty_data_max);
1030 1037
1031 1038 now = gethrtime();
1032 1039 min_tx_time = zfs_delay_scale *
1033 1040 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1034 1041 if (now > tx->tx_start + min_tx_time)
1035 1042 return;
1036 1043
1037 1044 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1038 1045
1039 1046 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1040 1047 uint64_t, min_tx_time);
1041 1048
1042 1049 mutex_enter(&dp->dp_lock);
1043 1050 wakeup = MAX(tx->tx_start + min_tx_time,
1044 1051 dp->dp_last_wakeup + min_tx_time);
1045 1052 dp->dp_last_wakeup = wakeup;
1046 1053 mutex_exit(&dp->dp_lock);
1047 1054
1048 1055 #ifdef _KERNEL
1049 1056 mutex_enter(&curthread->t_delay_lock);
1050 1057 while (cv_timedwait_hires(&curthread->t_delay_cv,
1051 1058 &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1052 1059 CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1053 1060 continue;
1054 1061 mutex_exit(&curthread->t_delay_lock);
1055 1062 #else
1056 1063 hrtime_t delta = wakeup - gethrtime();
1057 1064 struct timespec ts;
1058 1065 ts.tv_sec = delta / NANOSEC;
1059 1066 ts.tv_nsec = delta % NANOSEC;
1060 1067 (void) nanosleep(&ts, NULL);
1061 1068 #endif
1062 1069 }
1063 1070
1064 1071 static int
1065 1072 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1066 1073 {
1067 1074 dmu_tx_hold_t *txh;
1068 1075 spa_t *spa = tx->tx_pool->dp_spa;
1069 1076 uint64_t memory, asize, fsize, usize;
1070 1077 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1071 1078
1072 1079 ASSERT0(tx->tx_txg);
1073 1080
1074 1081 if (tx->tx_err)
1075 1082 return (tx->tx_err);
1076 1083
1077 1084 if (spa_suspended(spa)) {
1078 1085 /*
1079 1086 * If the user has indicated a blocking failure mode
1080 1087 * then return ERESTART which will block in dmu_tx_wait().
1081 1088 * Otherwise, return EIO so that an error can get
1082 1089 * propagated back to the VOP calls.
1083 1090 *
1084 1091 * Note that we always honor the txg_how flag regardless
1085 1092 * of the failuremode setting.
1086 1093 */
1087 1094 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1088 1095 txg_how != TXG_WAIT)
1089 1096 return (SET_ERROR(EIO));
1090 1097
1091 1098 return (SET_ERROR(ERESTART));
1092 1099 }
1093 1100
1094 1101 if (!tx->tx_waited &&
1095 1102 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1096 1103 tx->tx_wait_dirty = B_TRUE;
1097 1104 return (SET_ERROR(ERESTART));
1098 1105 }
1099 1106
1100 1107 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1101 1108 tx->tx_needassign_txh = NULL;
1102 1109
1103 1110 /*
1104 1111 * NB: No error returns are allowed after txg_hold_open, but
1105 1112 * before processing the dnode holds, due to the
1106 1113 * dmu_tx_unassign() logic.
1107 1114 */
1108 1115
1109 1116 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1110 1117 for (txh = list_head(&tx->tx_holds); txh;
1111 1118 txh = list_next(&tx->tx_holds, txh)) {
1112 1119 dnode_t *dn = txh->txh_dnode;
1113 1120 if (dn != NULL) {
1114 1121 mutex_enter(&dn->dn_mtx);
1115 1122 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1116 1123 mutex_exit(&dn->dn_mtx);
1117 1124 tx->tx_needassign_txh = txh;
1118 1125 return (SET_ERROR(ERESTART));
1119 1126 }
1120 1127 if (dn->dn_assigned_txg == 0)
1121 1128 dn->dn_assigned_txg = tx->tx_txg;
1122 1129 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1123 1130 (void) refcount_add(&dn->dn_tx_holds, tx);
1124 1131 mutex_exit(&dn->dn_mtx);
1125 1132 }
1126 1133 towrite += txh->txh_space_towrite;
1127 1134 tofree += txh->txh_space_tofree;
1128 1135 tooverwrite += txh->txh_space_tooverwrite;
1129 1136 tounref += txh->txh_space_tounref;
1130 1137 tohold += txh->txh_memory_tohold;
1131 1138 fudge += txh->txh_fudge;
1132 1139 }
1133 1140
1134 1141 /*
1135 1142 * If a snapshot has been taken since we made our estimates,
1136 1143 * assume that we won't be able to free or overwrite anything.
1137 1144 */
1138 1145 if (tx->tx_objset &&
1139 1146 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1140 1147 tx->tx_lastsnap_txg) {
1141 1148 towrite += tooverwrite;
1142 1149 tooverwrite = tofree = 0;
1143 1150 }
1144 1151
1145 1152 /* needed allocation: worst-case estimate of write space */
1146 1153 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1147 1154 /* freed space estimate: worst-case overwrite + free estimate */
1148 1155 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1149 1156 /* convert unrefd space to worst-case estimate */
1150 1157 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1151 1158 /* calculate memory footprint estimate */
1152 1159 memory = towrite + tooverwrite + tohold;
1153 1160
1154 1161 #ifdef ZFS_DEBUG
1155 1162 /*
1156 1163 * Add in 'tohold' to account for our dirty holds on this memory
1157 1164 * XXX - the "fudge" factor is to account for skipped blocks that
1158 1165 * we missed because dnode_next_offset() misses in-core-only blocks.
1159 1166 */
1160 1167 tx->tx_space_towrite = asize +
1161 1168 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1162 1169 tx->tx_space_tofree = tofree;
1163 1170 tx->tx_space_tooverwrite = tooverwrite;
1164 1171 tx->tx_space_tounref = tounref;
1165 1172 #endif
1166 1173
1167 1174 if (tx->tx_dir && asize != 0) {
1168 1175 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1169 1176 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1170 1177 if (err)
1171 1178 return (err);
1172 1179 }
1173 1180
1174 1181 return (0);
1175 1182 }
1176 1183
1177 1184 static void
1178 1185 dmu_tx_unassign(dmu_tx_t *tx)
1179 1186 {
1180 1187 dmu_tx_hold_t *txh;
1181 1188
1182 1189 if (tx->tx_txg == 0)
1183 1190 return;
1184 1191
1185 1192 txg_rele_to_quiesce(&tx->tx_txgh);
1186 1193
1187 1194 /*
1188 1195 * Walk the transaction's hold list, removing the hold on the
1189 1196 * associated dnode, and notifying waiters if the refcount drops to 0.
1190 1197 */
1191 1198 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1192 1199 txh = list_next(&tx->tx_holds, txh)) {
1193 1200 dnode_t *dn = txh->txh_dnode;
1194 1201
1195 1202 if (dn == NULL)
1196 1203 continue;
1197 1204 mutex_enter(&dn->dn_mtx);
1198 1205 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1199 1206
1200 1207 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1201 1208 dn->dn_assigned_txg = 0;
1202 1209 cv_broadcast(&dn->dn_notxholds);
1203 1210 }
1204 1211 mutex_exit(&dn->dn_mtx);
1205 1212 }
1206 1213
1207 1214 txg_rele_to_sync(&tx->tx_txgh);
1208 1215
1209 1216 tx->tx_lasttried_txg = tx->tx_txg;
1210 1217 tx->tx_txg = 0;
1211 1218 }
1212 1219
1213 1220 /*
1214 1221 * Assign tx to a transaction group. txg_how can be one of:
1215 1222 *
1216 1223 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1217 1224 * a new one. This should be used when you're not holding locks.
1218 1225 * It will only fail if we're truly out of space (or over quota).
1219 1226 *
1220 1227 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1221 1228 * blocking, returns immediately with ERESTART. This should be used
1222 1229 * whenever you're holding locks. On an ERESTART error, the caller
1223 1230 * should drop locks, do a dmu_tx_wait(tx), and try again.
1224 1231 *
1225 1232 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1226 1233 * has already been called on behalf of this operation (though
1227 1234 * most likely on a different tx).
1228 1235 */
1229 1236 int
1230 1237 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1231 1238 {
1232 1239 int err;
1233 1240
1234 1241 ASSERT(tx->tx_txg == 0);
1235 1242 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1236 1243 txg_how == TXG_WAITED);
1237 1244 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1238 1245
1239 1246 /* If we might wait, we must not hold the config lock. */
1240 1247 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1241 1248
1242 1249 if (txg_how == TXG_WAITED)
1243 1250 tx->tx_waited = B_TRUE;
1244 1251
1245 1252 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1246 1253 dmu_tx_unassign(tx);
1247 1254
1248 1255 if (err != ERESTART || txg_how != TXG_WAIT)
1249 1256 return (err);
1250 1257
1251 1258 dmu_tx_wait(tx);
1252 1259 }
1253 1260
1254 1261 txg_rele_to_quiesce(&tx->tx_txgh);
1255 1262
1256 1263 return (0);
1257 1264 }
1258 1265
1259 1266 void
1260 1267 dmu_tx_wait(dmu_tx_t *tx)
1261 1268 {
1262 1269 spa_t *spa = tx->tx_pool->dp_spa;
1263 1270 dsl_pool_t *dp = tx->tx_pool;
1264 1271
1265 1272 ASSERT(tx->tx_txg == 0);
1266 1273 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1267 1274
1268 1275 if (tx->tx_wait_dirty) {
1269 1276 /*
1270 1277 * dmu_tx_try_assign() has determined that we need to wait
1271 1278 * because we've consumed much or all of the dirty buffer
1272 1279 * space.
1273 1280 */
1274 1281 mutex_enter(&dp->dp_lock);
1275 1282 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1276 1283 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1277 1284 uint64_t dirty = dp->dp_dirty_total;
1278 1285 mutex_exit(&dp->dp_lock);
1279 1286
1280 1287 dmu_tx_delay(tx, dirty);
1281 1288
1282 1289 tx->tx_wait_dirty = B_FALSE;
1283 1290
1284 1291 /*
1285 1292 * Note: setting tx_waited only has effect if the caller
1286 1293 * used TX_WAIT. Otherwise they are going to destroy
1287 1294 * this tx and try again. The common case, zfs_write(),
1288 1295 * uses TX_WAIT.
1289 1296 */
1290 1297 tx->tx_waited = B_TRUE;
1291 1298 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1292 1299 /*
1293 1300 * If the pool is suspended we need to wait until it
1294 1301 * is resumed. Note that it's possible that the pool
1295 1302 * has become active after this thread has tried to
1296 1303 * obtain a tx. If that's the case then tx_lasttried_txg
1297 1304 * would not have been set.
1298 1305 */
1299 1306 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1300 1307 } else if (tx->tx_needassign_txh) {
1301 1308 /*
1302 1309 * A dnode is assigned to the quiescing txg. Wait for its
1303 1310 * transaction to complete.
1304 1311 */
1305 1312 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1306 1313
1307 1314 mutex_enter(&dn->dn_mtx);
1308 1315 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1309 1316 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1310 1317 mutex_exit(&dn->dn_mtx);
1311 1318 tx->tx_needassign_txh = NULL;
1312 1319 } else {
1313 1320 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1314 1321 }
1315 1322 }
1316 1323
1317 1324 void
1318 1325 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1319 1326 {
1320 1327 #ifdef ZFS_DEBUG
1321 1328 if (tx->tx_dir == NULL || delta == 0)
1322 1329 return;
1323 1330
1324 1331 if (delta > 0) {
1325 1332 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1326 1333 tx->tx_space_towrite);
1327 1334 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1328 1335 } else {
1329 1336 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1330 1337 }
1331 1338 #endif
1332 1339 }
1333 1340
1334 1341 void
1335 1342 dmu_tx_commit(dmu_tx_t *tx)
1336 1343 {
1337 1344 dmu_tx_hold_t *txh;
1338 1345
1339 1346 ASSERT(tx->tx_txg != 0);
1340 1347
1341 1348 /*
1342 1349 * Go through the transaction's hold list and remove holds on
1343 1350 * associated dnodes, notifying waiters if no holds remain.
1344 1351 */
1345 1352 while (txh = list_head(&tx->tx_holds)) {
1346 1353 dnode_t *dn = txh->txh_dnode;
1347 1354
1348 1355 list_remove(&tx->tx_holds, txh);
1349 1356 kmem_free(txh, sizeof (dmu_tx_hold_t));
1350 1357 if (dn == NULL)
1351 1358 continue;
1352 1359 mutex_enter(&dn->dn_mtx);
1353 1360 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1354 1361
1355 1362 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1356 1363 dn->dn_assigned_txg = 0;
1357 1364 cv_broadcast(&dn->dn_notxholds);
1358 1365 }
1359 1366 mutex_exit(&dn->dn_mtx);
1360 1367 dnode_rele(dn, tx);
1361 1368 }
1362 1369
1363 1370 if (tx->tx_tempreserve_cookie)
1364 1371 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1365 1372
1366 1373 if (!list_is_empty(&tx->tx_callbacks))
1367 1374 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1368 1375
1369 1376 if (tx->tx_anyobj == FALSE)
1370 1377 txg_rele_to_sync(&tx->tx_txgh);
1371 1378
1372 1379 list_destroy(&tx->tx_callbacks);
1373 1380 list_destroy(&tx->tx_holds);
1374 1381 #ifdef ZFS_DEBUG
1375 1382 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1376 1383 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1377 1384 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1378 1385 refcount_destroy_many(&tx->tx_space_written,
1379 1386 refcount_count(&tx->tx_space_written));
1380 1387 refcount_destroy_many(&tx->tx_space_freed,
1381 1388 refcount_count(&tx->tx_space_freed));
1382 1389 #endif
1383 1390 kmem_free(tx, sizeof (dmu_tx_t));
1384 1391 }
1385 1392
1386 1393 void
1387 1394 dmu_tx_abort(dmu_tx_t *tx)
1388 1395 {
1389 1396 dmu_tx_hold_t *txh;
1390 1397
1391 1398 ASSERT(tx->tx_txg == 0);
1392 1399
1393 1400 while (txh = list_head(&tx->tx_holds)) {
1394 1401 dnode_t *dn = txh->txh_dnode;
1395 1402
1396 1403 list_remove(&tx->tx_holds, txh);
1397 1404 kmem_free(txh, sizeof (dmu_tx_hold_t));
1398 1405 if (dn != NULL)
1399 1406 dnode_rele(dn, tx);
1400 1407 }
1401 1408
1402 1409 /*
1403 1410 * Call any registered callbacks with an error code.
1404 1411 */
1405 1412 if (!list_is_empty(&tx->tx_callbacks))
1406 1413 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1407 1414
1408 1415 list_destroy(&tx->tx_callbacks);
1409 1416 list_destroy(&tx->tx_holds);
1410 1417 #ifdef ZFS_DEBUG
1411 1418 refcount_destroy_many(&tx->tx_space_written,
1412 1419 refcount_count(&tx->tx_space_written));
1413 1420 refcount_destroy_many(&tx->tx_space_freed,
1414 1421 refcount_count(&tx->tx_space_freed));
1415 1422 #endif
1416 1423 kmem_free(tx, sizeof (dmu_tx_t));
1417 1424 }
1418 1425
1419 1426 uint64_t
1420 1427 dmu_tx_get_txg(dmu_tx_t *tx)
1421 1428 {
1422 1429 ASSERT(tx->tx_txg != 0);
1423 1430 return (tx->tx_txg);
1424 1431 }
1425 1432
1426 1433 dsl_pool_t *
1427 1434 dmu_tx_pool(dmu_tx_t *tx)
1428 1435 {
1429 1436 ASSERT(tx->tx_pool != NULL);
1430 1437 return (tx->tx_pool);
1431 1438 }
1432 1439
1433 1440
1434 1441 void
1435 1442 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1436 1443 {
1437 1444 dmu_tx_callback_t *dcb;
1438 1445
1439 1446 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1440 1447
1441 1448 dcb->dcb_func = func;
1442 1449 dcb->dcb_data = data;
1443 1450
1444 1451 list_insert_tail(&tx->tx_callbacks, dcb);
1445 1452 }
1446 1453
1447 1454 /*
1448 1455 * Call all the commit callbacks on a list, with a given error code.
1449 1456 */
1450 1457 void
1451 1458 dmu_tx_do_callbacks(list_t *cb_list, int error)
1452 1459 {
1453 1460 dmu_tx_callback_t *dcb;
1454 1461
1455 1462 while (dcb = list_head(cb_list)) {
1456 1463 list_remove(cb_list, dcb);
1457 1464 dcb->dcb_func(dcb->dcb_data, error);
1458 1465 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1459 1466 }
1460 1467 }
1461 1468
1462 1469 /*
1463 1470 * Interface to hold a bunch of attributes.
1464 1471 * used for creating new files.
1465 1472 * attrsize is the total size of all attributes
1466 1473 * to be added during object creation
1467 1474 *
1468 1475 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1469 1476 */
1470 1477
1471 1478 /*
1472 1479 * hold necessary attribute name for attribute registration.
1473 1480 * should be a very rare case where this is needed. If it does
1474 1481 * happen it would only happen on the first write to the file system.
1475 1482 */
1476 1483 static void
1477 1484 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1478 1485 {
1479 1486 int i;
1480 1487
1481 1488 if (!sa->sa_need_attr_registration)
1482 1489 return;
1483 1490
1484 1491 for (i = 0; i != sa->sa_num_attrs; i++) {
1485 1492 if (!sa->sa_attr_table[i].sa_registered) {
1486 1493 if (sa->sa_reg_attr_obj)
1487 1494 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1488 1495 B_TRUE, sa->sa_attr_table[i].sa_name);
1489 1496 else
1490 1497 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1491 1498 B_TRUE, sa->sa_attr_table[i].sa_name);
1492 1499 }
1493 1500 }
1494 1501 }
1495 1502
1496 1503
1497 1504 void
1498 1505 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1499 1506 {
1500 1507 dnode_t *dn;
1501 1508 dmu_tx_hold_t *txh;
1502 1509
1503 1510 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1504 1511 THT_SPILL, 0, 0);
1505 1512
1506 1513 dn = txh->txh_dnode;
1507 1514
1508 1515 if (dn == NULL)
1509 1516 return;
1510 1517
1511 1518 /* If blkptr doesn't exist then add space to towrite */
1512 1519 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1513 1520 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1514 1521 } else {
1515 1522 blkptr_t *bp;
1516 1523
1517 1524 bp = &dn->dn_phys->dn_spill;
1518 1525 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1519 1526 bp, bp->blk_birth))
1520 1527 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1521 1528 else
1522 1529 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1523 1530 if (!BP_IS_HOLE(bp))
1524 1531 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1525 1532 }
1526 1533 }
1527 1534
1528 1535 void
1529 1536 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1530 1537 {
1531 1538 sa_os_t *sa = tx->tx_objset->os_sa;
1532 1539
1533 1540 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1534 1541
1535 1542 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1536 1543 return;
1537 1544
1538 1545 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1539 1546 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1540 1547 else {
1541 1548 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1542 1549 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1543 1550 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1544 1551 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1545 1552 }
1546 1553
1547 1554 dmu_tx_sa_registration_hold(sa, tx);
1548 1555
1549 1556 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1550 1557 return;
1551 1558
1552 1559 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1553 1560 THT_SPILL, 0, 0);
1554 1561 }
1555 1562
1556 1563 /*
1557 1564 * Hold SA attribute
1558 1565 *
1559 1566 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1560 1567 *
1561 1568 * variable_size is the total size of all variable sized attributes
1562 1569 * passed to this function. It is not the total size of all
1563 1570 * variable size attributes that *may* exist on this object.
1564 1571 */
1565 1572 void
1566 1573 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1567 1574 {
1568 1575 uint64_t object;
1569 1576 sa_os_t *sa = tx->tx_objset->os_sa;
1570 1577
1571 1578 ASSERT(hdl != NULL);
1572 1579
1573 1580 object = sa_handle_object(hdl);
1574 1581
1575 1582 dmu_tx_hold_bonus(tx, object);
1576 1583
1577 1584 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1578 1585 return;
1579 1586
1580 1587 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1581 1588 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1582 1589 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1583 1590 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1584 1591 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1585 1592 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1586 1593 }
1587 1594
1588 1595 dmu_tx_sa_registration_hold(sa, tx);
1589 1596
1590 1597 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1591 1598 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1592 1599
1593 1600 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1594 1601 ASSERT(tx->tx_txg == 0);
1595 1602 dmu_tx_hold_spill(tx, object);
1596 1603 } else {
1597 1604 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1598 1605 dnode_t *dn;
1599 1606
1600 1607 DB_DNODE_ENTER(db);
1601 1608 dn = DB_DNODE(db);
1602 1609 if (dn->dn_have_spill) {
1603 1610 ASSERT(tx->tx_txg == 0);
1604 1611 dmu_tx_hold_spill(tx, object);
1605 1612 }
1606 1613 DB_DNODE_EXIT(db);
1607 1614 }
1608 1615 }
|
↓ open down ↓ |
958 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX