Print this page
3006 VERIFY[S,U,P] and ASSERT[S,U,P] frequently check if first argument is zero
| 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) 2012 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 #ifdef ZFS_DEBUG
58 58 refcount_create(&tx->tx_space_written);
59 59 refcount_create(&tx->tx_space_freed);
60 60 #endif
61 61 return (tx);
62 62 }
63 63
64 64 dmu_tx_t *
65 65 dmu_tx_create(objset_t *os)
66 66 {
67 67 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
68 68 tx->tx_objset = os;
69 69 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
70 70 return (tx);
71 71 }
72 72
73 73 dmu_tx_t *
74 74 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
75 75 {
76 76 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
77 77
78 78 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
79 79 tx->tx_pool = dp;
80 80 tx->tx_txg = txg;
81 81 tx->tx_anyobj = TRUE;
82 82
83 83 return (tx);
84 84 }
85 85
86 86 int
87 87 dmu_tx_is_syncing(dmu_tx_t *tx)
88 88 {
89 89 return (tx->tx_anyobj);
90 90 }
91 91
92 92 int
93 93 dmu_tx_private_ok(dmu_tx_t *tx)
94 94 {
95 95 return (tx->tx_anyobj);
96 96 }
97 97
98 98 static dmu_tx_hold_t *
99 99 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
100 100 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
101 101 {
102 102 dmu_tx_hold_t *txh;
103 103 dnode_t *dn = NULL;
104 104 int err;
105 105
106 106 if (object != DMU_NEW_OBJECT) {
107 107 err = dnode_hold(os, object, tx, &dn);
108 108 if (err) {
109 109 tx->tx_err = err;
110 110 return (NULL);
111 111 }
112 112
113 113 if (err == 0 && tx->tx_txg != 0) {
114 114 mutex_enter(&dn->dn_mtx);
115 115 /*
116 116 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 117 * problem, but there's no way for it to happen (for
118 118 * now, at least).
119 119 */
120 120 ASSERT(dn->dn_assigned_txg == 0);
121 121 dn->dn_assigned_txg = tx->tx_txg;
122 122 (void) refcount_add(&dn->dn_tx_holds, tx);
123 123 mutex_exit(&dn->dn_mtx);
124 124 }
125 125 }
126 126
127 127 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
128 128 txh->txh_tx = tx;
129 129 txh->txh_dnode = dn;
130 130 #ifdef ZFS_DEBUG
131 131 txh->txh_type = type;
132 132 txh->txh_arg1 = arg1;
133 133 txh->txh_arg2 = arg2;
134 134 #endif
135 135 list_insert_tail(&tx->tx_holds, txh);
136 136
137 137 return (txh);
138 138 }
139 139
140 140 void
141 141 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
142 142 {
143 143 /*
144 144 * If we're syncing, they can manipulate any object anyhow, and
145 145 * the hold on the dnode_t can cause problems.
146 146 */
147 147 if (!dmu_tx_is_syncing(tx)) {
148 148 (void) dmu_tx_hold_object_impl(tx, os,
149 149 object, THT_NEWOBJECT, 0, 0);
150 150 }
151 151 }
152 152
153 153 static int
154 154 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
155 155 {
156 156 int err;
157 157 dmu_buf_impl_t *db;
158 158
159 159 rw_enter(&dn->dn_struct_rwlock, RW_READER);
160 160 db = dbuf_hold_level(dn, level, blkid, FTAG);
161 161 rw_exit(&dn->dn_struct_rwlock);
162 162 if (db == NULL)
163 163 return (EIO);
164 164 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
165 165 dbuf_rele(db, FTAG);
166 166 return (err);
167 167 }
168 168
169 169 static void
170 170 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
171 171 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
172 172 {
173 173 objset_t *os = dn->dn_objset;
174 174 dsl_dataset_t *ds = os->os_dsl_dataset;
175 175 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
176 176 dmu_buf_impl_t *parent = NULL;
177 177 blkptr_t *bp = NULL;
178 178 uint64_t space;
179 179
180 180 if (level >= dn->dn_nlevels || history[level] == blkid)
181 181 return;
182 182
183 183 history[level] = blkid;
184 184
185 185 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
186 186
187 187 if (db == NULL || db == dn->dn_dbuf) {
188 188 ASSERT(level != 0);
189 189 db = NULL;
190 190 } else {
191 191 ASSERT(DB_DNODE(db) == dn);
192 192 ASSERT(db->db_level == level);
193 193 ASSERT(db->db.db_size == space);
194 194 ASSERT(db->db_blkid == blkid);
195 195 bp = db->db_blkptr;
196 196 parent = db->db_parent;
197 197 }
198 198
199 199 freeable = (bp && (freeable ||
200 200 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
201 201
202 202 if (freeable)
203 203 txh->txh_space_tooverwrite += space;
204 204 else
205 205 txh->txh_space_towrite += space;
206 206 if (bp)
207 207 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
208 208
209 209 dmu_tx_count_twig(txh, dn, parent, level + 1,
210 210 blkid >> epbs, freeable, history);
211 211 }
212 212
213 213 /* ARGSUSED */
214 214 static void
215 215 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
216 216 {
217 217 dnode_t *dn = txh->txh_dnode;
218 218 uint64_t start, end, i;
219 219 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
220 220 int err = 0;
221 221
222 222 if (len == 0)
223 223 return;
224 224
225 225 min_bs = SPA_MINBLOCKSHIFT;
226 226 max_bs = SPA_MAXBLOCKSHIFT;
227 227 min_ibs = DN_MIN_INDBLKSHIFT;
228 228 max_ibs = DN_MAX_INDBLKSHIFT;
229 229
230 230 if (dn) {
231 231 uint64_t history[DN_MAX_LEVELS];
232 232 int nlvls = dn->dn_nlevels;
233 233 int delta;
234 234
235 235 /*
236 236 * For i/o error checking, read the first and last level-0
237 237 * blocks (if they are not aligned), and all the level-1 blocks.
238 238 */
239 239 if (dn->dn_maxblkid == 0) {
240 240 delta = dn->dn_datablksz;
241 241 start = (off < dn->dn_datablksz) ? 0 : 1;
242 242 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
243 243 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
244 244 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
245 245 if (err)
246 246 goto out;
247 247 delta -= off;
248 248 }
249 249 } else {
250 250 zio_t *zio = zio_root(dn->dn_objset->os_spa,
251 251 NULL, NULL, ZIO_FLAG_CANFAIL);
252 252
253 253 /* first level-0 block */
254 254 start = off >> dn->dn_datablkshift;
255 255 if (P2PHASE(off, dn->dn_datablksz) ||
256 256 len < dn->dn_datablksz) {
257 257 err = dmu_tx_check_ioerr(zio, dn, 0, start);
258 258 if (err)
259 259 goto out;
260 260 }
261 261
262 262 /* last level-0 block */
263 263 end = (off+len-1) >> dn->dn_datablkshift;
264 264 if (end != start && end <= dn->dn_maxblkid &&
265 265 P2PHASE(off+len, dn->dn_datablksz)) {
266 266 err = dmu_tx_check_ioerr(zio, dn, 0, end);
267 267 if (err)
268 268 goto out;
269 269 }
270 270
271 271 /* level-1 blocks */
272 272 if (nlvls > 1) {
273 273 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
274 274 for (i = (start>>shft)+1; i < end>>shft; i++) {
275 275 err = dmu_tx_check_ioerr(zio, dn, 1, i);
276 276 if (err)
277 277 goto out;
278 278 }
279 279 }
280 280
281 281 err = zio_wait(zio);
282 282 if (err)
283 283 goto out;
284 284 delta = P2NPHASE(off, dn->dn_datablksz);
285 285 }
286 286
287 287 if (dn->dn_maxblkid > 0) {
288 288 /*
289 289 * The blocksize can't change,
290 290 * so we can make a more precise estimate.
291 291 */
292 292 ASSERT(dn->dn_datablkshift != 0);
293 293 min_bs = max_bs = dn->dn_datablkshift;
294 294 min_ibs = max_ibs = dn->dn_indblkshift;
295 295 } else if (dn->dn_indblkshift > max_ibs) {
296 296 /*
297 297 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
298 298 * the code will still work correctly on older pools.
299 299 */
300 300 min_ibs = max_ibs = dn->dn_indblkshift;
301 301 }
302 302
303 303 /*
304 304 * If this write is not off the end of the file
305 305 * we need to account for overwrites/unref.
306 306 */
307 307 if (start <= dn->dn_maxblkid) {
308 308 for (int l = 0; l < DN_MAX_LEVELS; l++)
309 309 history[l] = -1ULL;
310 310 }
311 311 while (start <= dn->dn_maxblkid) {
312 312 dmu_buf_impl_t *db;
313 313
314 314 rw_enter(&dn->dn_struct_rwlock, RW_READER);
315 315 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
316 316 rw_exit(&dn->dn_struct_rwlock);
317 317
318 318 if (err) {
319 319 txh->txh_tx->tx_err = err;
320 320 return;
321 321 }
322 322
323 323 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
324 324 history);
325 325 dbuf_rele(db, FTAG);
326 326 if (++start > end) {
327 327 /*
328 328 * Account for new indirects appearing
329 329 * before this IO gets assigned into a txg.
330 330 */
331 331 bits = 64 - min_bs;
332 332 epbs = min_ibs - SPA_BLKPTRSHIFT;
333 333 for (bits -= epbs * (nlvls - 1);
334 334 bits >= 0; bits -= epbs)
335 335 txh->txh_fudge += 1ULL << max_ibs;
336 336 goto out;
337 337 }
338 338 off += delta;
339 339 if (len >= delta)
340 340 len -= delta;
341 341 delta = dn->dn_datablksz;
342 342 }
343 343 }
344 344
345 345 /*
346 346 * 'end' is the last thing we will access, not one past.
347 347 * This way we won't overflow when accessing the last byte.
348 348 */
349 349 start = P2ALIGN(off, 1ULL << max_bs);
350 350 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
351 351 txh->txh_space_towrite += end - start + 1;
352 352
353 353 start >>= min_bs;
354 354 end >>= min_bs;
355 355
356 356 epbs = min_ibs - SPA_BLKPTRSHIFT;
357 357
358 358 /*
359 359 * The object contains at most 2^(64 - min_bs) blocks,
360 360 * and each indirect level maps 2^epbs.
361 361 */
362 362 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
363 363 start >>= epbs;
364 364 end >>= epbs;
365 365 ASSERT3U(end, >=, start);
366 366 txh->txh_space_towrite += (end - start + 1) << max_ibs;
367 367 if (start != 0) {
368 368 /*
369 369 * We also need a new blkid=0 indirect block
370 370 * to reference any existing file data.
371 371 */
372 372 txh->txh_space_towrite += 1ULL << max_ibs;
373 373 }
374 374 }
375 375
376 376 out:
377 377 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
378 378 2 * DMU_MAX_ACCESS)
379 379 err = EFBIG;
380 380
381 381 if (err)
382 382 txh->txh_tx->tx_err = err;
383 383 }
384 384
385 385 static void
386 386 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
387 387 {
388 388 dnode_t *dn = txh->txh_dnode;
389 389 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
390 390 uint64_t space = mdn->dn_datablksz +
391 391 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
392 392
393 393 if (dn && dn->dn_dbuf->db_blkptr &&
394 394 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
395 395 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
396 396 txh->txh_space_tooverwrite += space;
397 397 txh->txh_space_tounref += space;
398 398 } else {
399 399 txh->txh_space_towrite += space;
400 400 if (dn && dn->dn_dbuf->db_blkptr)
401 401 txh->txh_space_tounref += space;
402 402 }
403 403 }
404 404
405 405 void
406 406 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
407 407 {
408 408 dmu_tx_hold_t *txh;
409 409
410 410 ASSERT(tx->tx_txg == 0);
411 411 ASSERT(len < DMU_MAX_ACCESS);
412 412 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
413 413
414 414 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
415 415 object, THT_WRITE, off, len);
416 416 if (txh == NULL)
417 417 return;
418 418
419 419 dmu_tx_count_write(txh, off, len);
420 420 dmu_tx_count_dnode(txh);
421 421 }
422 422
423 423 static void
424 424 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
425 425 {
426 426 uint64_t blkid, nblks, lastblk;
427 427 uint64_t space = 0, unref = 0, skipped = 0;
428 428 dnode_t *dn = txh->txh_dnode;
429 429 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
430 430 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
431 431 int epbs;
432 432
433 433 if (dn->dn_nlevels == 0)
434 434 return;
435 435
436 436 /*
437 437 * The struct_rwlock protects us against dn_nlevels
438 438 * changing, in case (against all odds) we manage to dirty &
439 439 * sync out the changes after we check for being dirty.
440 440 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
441 441 */
442 442 rw_enter(&dn->dn_struct_rwlock, RW_READER);
443 443 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
444 444 if (dn->dn_maxblkid == 0) {
445 445 if (off == 0 && len >= dn->dn_datablksz) {
446 446 blkid = 0;
447 447 nblks = 1;
448 448 } else {
449 449 rw_exit(&dn->dn_struct_rwlock);
450 450 return;
451 451 }
452 452 } else {
453 453 blkid = off >> dn->dn_datablkshift;
454 454 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
455 455
456 456 if (blkid >= dn->dn_maxblkid) {
457 457 rw_exit(&dn->dn_struct_rwlock);
458 458 return;
459 459 }
460 460 if (blkid + nblks > dn->dn_maxblkid)
461 461 nblks = dn->dn_maxblkid - blkid;
462 462
463 463 }
464 464 if (dn->dn_nlevels == 1) {
465 465 int i;
466 466 for (i = 0; i < nblks; i++) {
467 467 blkptr_t *bp = dn->dn_phys->dn_blkptr;
468 468 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
469 469 bp += blkid + i;
470 470 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
471 471 dprintf_bp(bp, "can free old%s", "");
472 472 space += bp_get_dsize(spa, bp);
473 473 }
474 474 unref += BP_GET_ASIZE(bp);
475 475 }
476 476 nblks = 0;
477 477 }
478 478
479 479 /*
480 480 * Add in memory requirements of higher-level indirects.
481 481 * This assumes a worst-possible scenario for dn_nlevels.
482 482 */
483 483 {
484 484 uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
485 485 int level = (dn->dn_nlevels > 1) ? 2 : 1;
486 486
487 487 while (level++ < DN_MAX_LEVELS) {
488 488 txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
489 489 blkcnt = 1 + (blkcnt >> epbs);
490 490 }
491 491 ASSERT(blkcnt <= dn->dn_nblkptr);
492 492 }
493 493
494 494 lastblk = blkid + nblks - 1;
495 495 while (nblks) {
496 496 dmu_buf_impl_t *dbuf;
497 497 uint64_t ibyte, new_blkid;
498 498 int epb = 1 << epbs;
499 499 int err, i, blkoff, tochk;
500 500 blkptr_t *bp;
501 501
502 502 ibyte = blkid << dn->dn_datablkshift;
503 503 err = dnode_next_offset(dn,
504 504 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
505 505 new_blkid = ibyte >> dn->dn_datablkshift;
506 506 if (err == ESRCH) {
507 507 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
508 508 break;
509 509 }
510 510 if (err) {
511 511 txh->txh_tx->tx_err = err;
512 512 break;
513 513 }
514 514 if (new_blkid > lastblk) {
515 515 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
516 516 break;
517 517 }
518 518
519 519 if (new_blkid > blkid) {
520 520 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
521 521 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
522 522 nblks -= new_blkid - blkid;
523 523 blkid = new_blkid;
524 524 }
525 525 blkoff = P2PHASE(blkid, epb);
526 526 tochk = MIN(epb - blkoff, nblks);
527 527
528 528 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
529 529 if (err) {
530 530 txh->txh_tx->tx_err = err;
531 531 break;
532 532 }
533 533
534 534 txh->txh_memory_tohold += dbuf->db.db_size;
535 535
536 536 /*
537 537 * We don't check memory_tohold against DMU_MAX_ACCESS because
538 538 * memory_tohold is an over-estimation (especially the >L1
539 539 * indirect blocks), so it could fail. Callers should have
540 540 * already verified that they will not be holding too much
541 541 * memory.
542 542 */
543 543
544 544 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
545 545 if (err != 0) {
546 546 txh->txh_tx->tx_err = err;
547 547 dbuf_rele(dbuf, FTAG);
548 548 break;
549 549 }
550 550
551 551 bp = dbuf->db.db_data;
552 552 bp += blkoff;
553 553
554 554 for (i = 0; i < tochk; i++) {
555 555 if (dsl_dataset_block_freeable(ds, &bp[i],
556 556 bp[i].blk_birth)) {
557 557 dprintf_bp(&bp[i], "can free old%s", "");
558 558 space += bp_get_dsize(spa, &bp[i]);
559 559 }
560 560 unref += BP_GET_ASIZE(bp);
561 561 }
562 562 dbuf_rele(dbuf, FTAG);
563 563
564 564 blkid += tochk;
565 565 nblks -= tochk;
566 566 }
567 567 rw_exit(&dn->dn_struct_rwlock);
568 568
569 569 /* account for new level 1 indirect blocks that might show up */
570 570 if (skipped > 0) {
571 571 txh->txh_fudge += skipped << dn->dn_indblkshift;
572 572 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
573 573 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
574 574 }
575 575 txh->txh_space_tofree += space;
576 576 txh->txh_space_tounref += unref;
577 577 }
578 578
579 579 void
580 580 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
581 581 {
582 582 dmu_tx_hold_t *txh;
583 583 dnode_t *dn;
584 584 uint64_t start, end, i;
585 585 int err, shift;
586 586 zio_t *zio;
587 587
588 588 ASSERT(tx->tx_txg == 0);
589 589
590 590 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
591 591 object, THT_FREE, off, len);
592 592 if (txh == NULL)
593 593 return;
594 594 dn = txh->txh_dnode;
595 595
596 596 /* first block */
597 597 if (off != 0)
598 598 dmu_tx_count_write(txh, off, 1);
599 599 /* last block */
600 600 if (len != DMU_OBJECT_END)
601 601 dmu_tx_count_write(txh, off+len, 1);
602 602
603 603 dmu_tx_count_dnode(txh);
604 604
605 605 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
606 606 return;
607 607 if (len == DMU_OBJECT_END)
608 608 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
609 609
610 610 /*
611 611 * For i/o error checking, read the first and last level-0
612 612 * blocks, and all the level-1 blocks. The above count_write's
613 613 * have already taken care of the level-0 blocks.
614 614 */
615 615 if (dn->dn_nlevels > 1) {
616 616 shift = dn->dn_datablkshift + dn->dn_indblkshift -
617 617 SPA_BLKPTRSHIFT;
618 618 start = off >> shift;
619 619 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
620 620
621 621 zio = zio_root(tx->tx_pool->dp_spa,
622 622 NULL, NULL, ZIO_FLAG_CANFAIL);
623 623 for (i = start; i <= end; i++) {
624 624 uint64_t ibyte = i << shift;
625 625 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
626 626 i = ibyte >> shift;
627 627 if (err == ESRCH)
628 628 break;
629 629 if (err) {
630 630 tx->tx_err = err;
631 631 return;
632 632 }
633 633
634 634 err = dmu_tx_check_ioerr(zio, dn, 1, i);
635 635 if (err) {
636 636 tx->tx_err = err;
637 637 return;
638 638 }
639 639 }
640 640 err = zio_wait(zio);
641 641 if (err) {
642 642 tx->tx_err = err;
643 643 return;
644 644 }
645 645 }
646 646
647 647 dmu_tx_count_free(txh, off, len);
648 648 }
649 649
650 650 void
651 651 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
652 652 {
653 653 dmu_tx_hold_t *txh;
654 654 dnode_t *dn;
655 655 uint64_t nblocks;
656 656 int epbs, err;
657 657
658 658 ASSERT(tx->tx_txg == 0);
659 659
660 660 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
661 661 object, THT_ZAP, add, (uintptr_t)name);
662 662 if (txh == NULL)
663 663 return;
664 664 dn = txh->txh_dnode;
665 665
666 666 dmu_tx_count_dnode(txh);
667 667
668 668 if (dn == NULL) {
669 669 /*
670 670 * We will be able to fit a new object's entries into one leaf
671 671 * block. So there will be at most 2 blocks total,
672 672 * including the header block.
673 673 */
674 674 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
675 675 return;
676 676 }
677 677
678 678 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
679 679
680 680 if (dn->dn_maxblkid == 0 && !add) {
681 681 blkptr_t *bp;
682 682
683 683 /*
684 684 * If there is only one block (i.e. this is a micro-zap)
685 685 * and we are not adding anything, the accounting is simple.
686 686 */
687 687 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
688 688 if (err) {
689 689 tx->tx_err = err;
690 690 return;
691 691 }
692 692
693 693 /*
694 694 * Use max block size here, since we don't know how much
695 695 * the size will change between now and the dbuf dirty call.
696 696 */
697 697 bp = &dn->dn_phys->dn_blkptr[0];
698 698 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
699 699 bp, bp->blk_birth))
700 700 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
701 701 else
702 702 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
703 703 if (!BP_IS_HOLE(bp))
704 704 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
705 705 return;
706 706 }
707 707
708 708 if (dn->dn_maxblkid > 0 && name) {
709 709 /*
710 710 * access the name in this fat-zap so that we'll check
711 711 * for i/o errors to the leaf blocks, etc.
712 712 */
713 713 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
714 714 8, 0, NULL);
715 715 if (err == EIO) {
716 716 tx->tx_err = err;
717 717 return;
718 718 }
719 719 }
720 720
721 721 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
722 722 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
723 723
724 724 /*
725 725 * If the modified blocks are scattered to the four winds,
726 726 * we'll have to modify an indirect twig for each.
727 727 */
728 728 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
729 729 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
730 730 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
731 731 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
732 732 else
733 733 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
734 734 }
735 735
736 736 void
737 737 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
738 738 {
739 739 dmu_tx_hold_t *txh;
740 740
741 741 ASSERT(tx->tx_txg == 0);
742 742
743 743 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
744 744 object, THT_BONUS, 0, 0);
745 745 if (txh)
746 746 dmu_tx_count_dnode(txh);
747 747 }
748 748
749 749 void
750 750 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
751 751 {
752 752 dmu_tx_hold_t *txh;
753 753 ASSERT(tx->tx_txg == 0);
754 754
755 755 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
756 756 DMU_NEW_OBJECT, THT_SPACE, space, 0);
757 757
758 758 txh->txh_space_towrite += space;
759 759 }
760 760
761 761 int
762 762 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
763 763 {
764 764 dmu_tx_hold_t *txh;
765 765 int holds = 0;
766 766
767 767 /*
768 768 * By asserting that the tx is assigned, we're counting the
769 769 * number of dn_tx_holds, which is the same as the number of
770 770 * dn_holds. Otherwise, we'd be counting dn_holds, but
771 771 * dn_tx_holds could be 0.
772 772 */
773 773 ASSERT(tx->tx_txg != 0);
774 774
775 775 /* if (tx->tx_anyobj == TRUE) */
776 776 /* return (0); */
777 777
778 778 for (txh = list_head(&tx->tx_holds); txh;
779 779 txh = list_next(&tx->tx_holds, txh)) {
780 780 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
781 781 holds++;
782 782 }
783 783
784 784 return (holds);
785 785 }
786 786
787 787 #ifdef ZFS_DEBUG
788 788 void
789 789 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
790 790 {
791 791 dmu_tx_hold_t *txh;
792 792 int match_object = FALSE, match_offset = FALSE;
793 793 dnode_t *dn;
794 794
795 795 DB_DNODE_ENTER(db);
796 796 dn = DB_DNODE(db);
797 797 ASSERT(tx->tx_txg != 0);
798 798 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
799 799 ASSERT3U(dn->dn_object, ==, db->db.db_object);
800 800
801 801 if (tx->tx_anyobj) {
802 802 DB_DNODE_EXIT(db);
803 803 return;
804 804 }
805 805
806 806 /* XXX No checking on the meta dnode for now */
807 807 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
808 808 DB_DNODE_EXIT(db);
809 809 return;
810 810 }
811 811
812 812 for (txh = list_head(&tx->tx_holds); txh;
813 813 txh = list_next(&tx->tx_holds, txh)) {
814 814 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
815 815 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
816 816 match_object = TRUE;
817 817 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
818 818 int datablkshift = dn->dn_datablkshift ?
819 819 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
820 820 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
821 821 int shift = datablkshift + epbs * db->db_level;
822 822 uint64_t beginblk = shift >= 64 ? 0 :
823 823 (txh->txh_arg1 >> shift);
824 824 uint64_t endblk = shift >= 64 ? 0 :
825 825 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
826 826 uint64_t blkid = db->db_blkid;
827 827
828 828 /* XXX txh_arg2 better not be zero... */
829 829
830 830 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
831 831 txh->txh_type, beginblk, endblk);
832 832
833 833 switch (txh->txh_type) {
834 834 case THT_WRITE:
835 835 if (blkid >= beginblk && blkid <= endblk)
836 836 match_offset = TRUE;
837 837 /*
838 838 * We will let this hold work for the bonus
839 839 * or spill buffer so that we don't need to
840 840 * hold it when creating a new object.
841 841 */
842 842 if (blkid == DMU_BONUS_BLKID ||
843 843 blkid == DMU_SPILL_BLKID)
844 844 match_offset = TRUE;
845 845 /*
846 846 * They might have to increase nlevels,
847 847 * thus dirtying the new TLIBs. Or the
848 848 * might have to change the block size,
849 849 * thus dirying the new lvl=0 blk=0.
850 850 */
851 851 if (blkid == 0)
852 852 match_offset = TRUE;
853 853 break;
854 854 case THT_FREE:
855 855 /*
856 856 * We will dirty all the level 1 blocks in
857 857 * the free range and perhaps the first and
858 858 * last level 0 block.
859 859 */
860 860 if (blkid >= beginblk && (blkid <= endblk ||
861 861 txh->txh_arg2 == DMU_OBJECT_END))
862 862 match_offset = TRUE;
863 863 break;
864 864 case THT_SPILL:
865 865 if (blkid == DMU_SPILL_BLKID)
866 866 match_offset = TRUE;
867 867 break;
868 868 case THT_BONUS:
869 869 if (blkid == DMU_BONUS_BLKID)
870 870 match_offset = TRUE;
871 871 break;
872 872 case THT_ZAP:
873 873 match_offset = TRUE;
874 874 break;
875 875 case THT_NEWOBJECT:
876 876 match_object = TRUE;
877 877 break;
878 878 default:
879 879 ASSERT(!"bad txh_type");
880 880 }
881 881 }
882 882 if (match_object && match_offset) {
883 883 DB_DNODE_EXIT(db);
884 884 return;
885 885 }
886 886 }
887 887 DB_DNODE_EXIT(db);
888 888 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
889 889 (u_longlong_t)db->db.db_object, db->db_level,
890 890 (u_longlong_t)db->db_blkid);
891 891 }
|
↓ open down ↓ |
891 lines elided |
↑ open up ↑ |
892 892 #endif
893 893
894 894 static int
895 895 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
896 896 {
897 897 dmu_tx_hold_t *txh;
898 898 spa_t *spa = tx->tx_pool->dp_spa;
899 899 uint64_t memory, asize, fsize, usize;
900 900 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
901 901
902 - ASSERT3U(tx->tx_txg, ==, 0);
902 + ASSERT0(tx->tx_txg);
903 903
904 904 if (tx->tx_err)
905 905 return (tx->tx_err);
906 906
907 907 if (spa_suspended(spa)) {
908 908 /*
909 909 * If the user has indicated a blocking failure mode
910 910 * then return ERESTART which will block in dmu_tx_wait().
911 911 * Otherwise, return EIO so that an error can get
912 912 * propagated back to the VOP calls.
913 913 *
914 914 * Note that we always honor the txg_how flag regardless
915 915 * of the failuremode setting.
916 916 */
917 917 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
918 918 txg_how != TXG_WAIT)
919 919 return (EIO);
920 920
921 921 return (ERESTART);
922 922 }
923 923
924 924 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
925 925 tx->tx_needassign_txh = NULL;
926 926
927 927 /*
928 928 * NB: No error returns are allowed after txg_hold_open, but
929 929 * before processing the dnode holds, due to the
930 930 * dmu_tx_unassign() logic.
931 931 */
932 932
933 933 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
934 934 for (txh = list_head(&tx->tx_holds); txh;
935 935 txh = list_next(&tx->tx_holds, txh)) {
936 936 dnode_t *dn = txh->txh_dnode;
937 937 if (dn != NULL) {
938 938 mutex_enter(&dn->dn_mtx);
939 939 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
940 940 mutex_exit(&dn->dn_mtx);
941 941 tx->tx_needassign_txh = txh;
942 942 return (ERESTART);
943 943 }
944 944 if (dn->dn_assigned_txg == 0)
945 945 dn->dn_assigned_txg = tx->tx_txg;
946 946 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
947 947 (void) refcount_add(&dn->dn_tx_holds, tx);
948 948 mutex_exit(&dn->dn_mtx);
949 949 }
950 950 towrite += txh->txh_space_towrite;
951 951 tofree += txh->txh_space_tofree;
952 952 tooverwrite += txh->txh_space_tooverwrite;
953 953 tounref += txh->txh_space_tounref;
954 954 tohold += txh->txh_memory_tohold;
955 955 fudge += txh->txh_fudge;
956 956 }
957 957
958 958 /*
959 959 * NB: This check must be after we've held the dnodes, so that
960 960 * the dmu_tx_unassign() logic will work properly
961 961 */
962 962 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
963 963 return (ERESTART);
964 964
965 965 /*
966 966 * If a snapshot has been taken since we made our estimates,
967 967 * assume that we won't be able to free or overwrite anything.
968 968 */
969 969 if (tx->tx_objset &&
970 970 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
971 971 tx->tx_lastsnap_txg) {
972 972 towrite += tooverwrite;
973 973 tooverwrite = tofree = 0;
974 974 }
975 975
976 976 /* needed allocation: worst-case estimate of write space */
977 977 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
978 978 /* freed space estimate: worst-case overwrite + free estimate */
979 979 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
980 980 /* convert unrefd space to worst-case estimate */
981 981 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
982 982 /* calculate memory footprint estimate */
983 983 memory = towrite + tooverwrite + tohold;
984 984
985 985 #ifdef ZFS_DEBUG
986 986 /*
987 987 * Add in 'tohold' to account for our dirty holds on this memory
988 988 * XXX - the "fudge" factor is to account for skipped blocks that
989 989 * we missed because dnode_next_offset() misses in-core-only blocks.
990 990 */
991 991 tx->tx_space_towrite = asize +
992 992 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
993 993 tx->tx_space_tofree = tofree;
994 994 tx->tx_space_tooverwrite = tooverwrite;
995 995 tx->tx_space_tounref = tounref;
996 996 #endif
997 997
998 998 if (tx->tx_dir && asize != 0) {
999 999 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1000 1000 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1001 1001 if (err)
1002 1002 return (err);
1003 1003 }
1004 1004
1005 1005 return (0);
1006 1006 }
1007 1007
1008 1008 static void
1009 1009 dmu_tx_unassign(dmu_tx_t *tx)
1010 1010 {
1011 1011 dmu_tx_hold_t *txh;
1012 1012
1013 1013 if (tx->tx_txg == 0)
1014 1014 return;
1015 1015
1016 1016 txg_rele_to_quiesce(&tx->tx_txgh);
1017 1017
1018 1018 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1019 1019 txh = list_next(&tx->tx_holds, txh)) {
1020 1020 dnode_t *dn = txh->txh_dnode;
1021 1021
1022 1022 if (dn == NULL)
1023 1023 continue;
1024 1024 mutex_enter(&dn->dn_mtx);
1025 1025 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1026 1026
1027 1027 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1028 1028 dn->dn_assigned_txg = 0;
1029 1029 cv_broadcast(&dn->dn_notxholds);
1030 1030 }
1031 1031 mutex_exit(&dn->dn_mtx);
1032 1032 }
1033 1033
1034 1034 txg_rele_to_sync(&tx->tx_txgh);
1035 1035
1036 1036 tx->tx_lasttried_txg = tx->tx_txg;
1037 1037 tx->tx_txg = 0;
1038 1038 }
1039 1039
1040 1040 /*
1041 1041 * Assign tx to a transaction group. txg_how can be one of:
1042 1042 *
1043 1043 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1044 1044 * a new one. This should be used when you're not holding locks.
1045 1045 * If will only fail if we're truly out of space (or over quota).
1046 1046 *
1047 1047 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1048 1048 * blocking, returns immediately with ERESTART. This should be used
1049 1049 * whenever you're holding locks. On an ERESTART error, the caller
1050 1050 * should drop locks, do a dmu_tx_wait(tx), and try again.
1051 1051 *
1052 1052 * (3) A specific txg. Use this if you need to ensure that multiple
1053 1053 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1054 1054 * returns ERESTART if it can't assign you into the requested txg.
1055 1055 */
1056 1056 int
1057 1057 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1058 1058 {
1059 1059 int err;
1060 1060
1061 1061 ASSERT(tx->tx_txg == 0);
1062 1062 ASSERT(txg_how != 0);
1063 1063 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1064 1064
1065 1065 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1066 1066 dmu_tx_unassign(tx);
1067 1067
1068 1068 if (err != ERESTART || txg_how != TXG_WAIT)
1069 1069 return (err);
1070 1070
1071 1071 dmu_tx_wait(tx);
1072 1072 }
1073 1073
1074 1074 txg_rele_to_quiesce(&tx->tx_txgh);
1075 1075
1076 1076 return (0);
1077 1077 }
1078 1078
1079 1079 void
1080 1080 dmu_tx_wait(dmu_tx_t *tx)
1081 1081 {
1082 1082 spa_t *spa = tx->tx_pool->dp_spa;
1083 1083
1084 1084 ASSERT(tx->tx_txg == 0);
1085 1085
1086 1086 /*
1087 1087 * It's possible that the pool has become active after this thread
1088 1088 * has tried to obtain a tx. If that's the case then his
1089 1089 * tx_lasttried_txg would not have been assigned.
1090 1090 */
1091 1091 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1092 1092 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1093 1093 } else if (tx->tx_needassign_txh) {
1094 1094 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1095 1095
1096 1096 mutex_enter(&dn->dn_mtx);
1097 1097 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1098 1098 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1099 1099 mutex_exit(&dn->dn_mtx);
1100 1100 tx->tx_needassign_txh = NULL;
1101 1101 } else {
1102 1102 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1103 1103 }
1104 1104 }
1105 1105
1106 1106 void
1107 1107 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1108 1108 {
1109 1109 #ifdef ZFS_DEBUG
1110 1110 if (tx->tx_dir == NULL || delta == 0)
1111 1111 return;
1112 1112
1113 1113 if (delta > 0) {
1114 1114 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1115 1115 tx->tx_space_towrite);
1116 1116 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1117 1117 } else {
1118 1118 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1119 1119 }
1120 1120 #endif
1121 1121 }
1122 1122
1123 1123 void
1124 1124 dmu_tx_commit(dmu_tx_t *tx)
1125 1125 {
1126 1126 dmu_tx_hold_t *txh;
1127 1127
1128 1128 ASSERT(tx->tx_txg != 0);
1129 1129
1130 1130 while (txh = list_head(&tx->tx_holds)) {
1131 1131 dnode_t *dn = txh->txh_dnode;
1132 1132
1133 1133 list_remove(&tx->tx_holds, txh);
1134 1134 kmem_free(txh, sizeof (dmu_tx_hold_t));
1135 1135 if (dn == NULL)
1136 1136 continue;
1137 1137 mutex_enter(&dn->dn_mtx);
1138 1138 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1139 1139
1140 1140 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1141 1141 dn->dn_assigned_txg = 0;
1142 1142 cv_broadcast(&dn->dn_notxholds);
1143 1143 }
1144 1144 mutex_exit(&dn->dn_mtx);
1145 1145 dnode_rele(dn, tx);
1146 1146 }
1147 1147
1148 1148 if (tx->tx_tempreserve_cookie)
1149 1149 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1150 1150
1151 1151 if (!list_is_empty(&tx->tx_callbacks))
1152 1152 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1153 1153
1154 1154 if (tx->tx_anyobj == FALSE)
1155 1155 txg_rele_to_sync(&tx->tx_txgh);
1156 1156
1157 1157 list_destroy(&tx->tx_callbacks);
1158 1158 list_destroy(&tx->tx_holds);
1159 1159 #ifdef ZFS_DEBUG
1160 1160 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1161 1161 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1162 1162 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1163 1163 refcount_destroy_many(&tx->tx_space_written,
1164 1164 refcount_count(&tx->tx_space_written));
1165 1165 refcount_destroy_many(&tx->tx_space_freed,
1166 1166 refcount_count(&tx->tx_space_freed));
1167 1167 #endif
1168 1168 kmem_free(tx, sizeof (dmu_tx_t));
1169 1169 }
1170 1170
1171 1171 void
1172 1172 dmu_tx_abort(dmu_tx_t *tx)
1173 1173 {
1174 1174 dmu_tx_hold_t *txh;
1175 1175
1176 1176 ASSERT(tx->tx_txg == 0);
1177 1177
1178 1178 while (txh = list_head(&tx->tx_holds)) {
1179 1179 dnode_t *dn = txh->txh_dnode;
1180 1180
1181 1181 list_remove(&tx->tx_holds, txh);
1182 1182 kmem_free(txh, sizeof (dmu_tx_hold_t));
1183 1183 if (dn != NULL)
1184 1184 dnode_rele(dn, tx);
1185 1185 }
1186 1186
1187 1187 /*
1188 1188 * Call any registered callbacks with an error code.
1189 1189 */
1190 1190 if (!list_is_empty(&tx->tx_callbacks))
1191 1191 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1192 1192
1193 1193 list_destroy(&tx->tx_callbacks);
1194 1194 list_destroy(&tx->tx_holds);
1195 1195 #ifdef ZFS_DEBUG
1196 1196 refcount_destroy_many(&tx->tx_space_written,
1197 1197 refcount_count(&tx->tx_space_written));
1198 1198 refcount_destroy_many(&tx->tx_space_freed,
1199 1199 refcount_count(&tx->tx_space_freed));
1200 1200 #endif
1201 1201 kmem_free(tx, sizeof (dmu_tx_t));
1202 1202 }
1203 1203
1204 1204 uint64_t
1205 1205 dmu_tx_get_txg(dmu_tx_t *tx)
1206 1206 {
1207 1207 ASSERT(tx->tx_txg != 0);
1208 1208 return (tx->tx_txg);
1209 1209 }
1210 1210
1211 1211 void
1212 1212 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1213 1213 {
1214 1214 dmu_tx_callback_t *dcb;
1215 1215
1216 1216 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1217 1217
1218 1218 dcb->dcb_func = func;
1219 1219 dcb->dcb_data = data;
1220 1220
1221 1221 list_insert_tail(&tx->tx_callbacks, dcb);
1222 1222 }
1223 1223
1224 1224 /*
1225 1225 * Call all the commit callbacks on a list, with a given error code.
1226 1226 */
1227 1227 void
1228 1228 dmu_tx_do_callbacks(list_t *cb_list, int error)
1229 1229 {
1230 1230 dmu_tx_callback_t *dcb;
1231 1231
1232 1232 while (dcb = list_head(cb_list)) {
1233 1233 list_remove(cb_list, dcb);
1234 1234 dcb->dcb_func(dcb->dcb_data, error);
1235 1235 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1236 1236 }
1237 1237 }
1238 1238
1239 1239 /*
1240 1240 * Interface to hold a bunch of attributes.
1241 1241 * used for creating new files.
1242 1242 * attrsize is the total size of all attributes
1243 1243 * to be added during object creation
1244 1244 *
1245 1245 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1246 1246 */
1247 1247
1248 1248 /*
1249 1249 * hold necessary attribute name for attribute registration.
1250 1250 * should be a very rare case where this is needed. If it does
1251 1251 * happen it would only happen on the first write to the file system.
1252 1252 */
1253 1253 static void
1254 1254 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1255 1255 {
1256 1256 int i;
1257 1257
1258 1258 if (!sa->sa_need_attr_registration)
1259 1259 return;
1260 1260
1261 1261 for (i = 0; i != sa->sa_num_attrs; i++) {
1262 1262 if (!sa->sa_attr_table[i].sa_registered) {
1263 1263 if (sa->sa_reg_attr_obj)
1264 1264 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1265 1265 B_TRUE, sa->sa_attr_table[i].sa_name);
1266 1266 else
1267 1267 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1268 1268 B_TRUE, sa->sa_attr_table[i].sa_name);
1269 1269 }
1270 1270 }
1271 1271 }
1272 1272
1273 1273
1274 1274 void
1275 1275 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1276 1276 {
1277 1277 dnode_t *dn;
1278 1278 dmu_tx_hold_t *txh;
1279 1279
1280 1280 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1281 1281 THT_SPILL, 0, 0);
1282 1282
1283 1283 dn = txh->txh_dnode;
1284 1284
1285 1285 if (dn == NULL)
1286 1286 return;
1287 1287
1288 1288 /* If blkptr doesn't exist then add space to towrite */
1289 1289 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1290 1290 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1291 1291 } else {
1292 1292 blkptr_t *bp;
1293 1293
1294 1294 bp = &dn->dn_phys->dn_spill;
1295 1295 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1296 1296 bp, bp->blk_birth))
1297 1297 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1298 1298 else
1299 1299 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1300 1300 if (!BP_IS_HOLE(bp))
1301 1301 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1302 1302 }
1303 1303 }
1304 1304
1305 1305 void
1306 1306 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1307 1307 {
1308 1308 sa_os_t *sa = tx->tx_objset->os_sa;
1309 1309
1310 1310 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1311 1311
1312 1312 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1313 1313 return;
1314 1314
1315 1315 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1316 1316 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1317 1317 else {
1318 1318 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1319 1319 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1320 1320 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1321 1321 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 1322 }
1323 1323
1324 1324 dmu_tx_sa_registration_hold(sa, tx);
1325 1325
1326 1326 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1327 1327 return;
1328 1328
1329 1329 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1330 1330 THT_SPILL, 0, 0);
1331 1331 }
1332 1332
1333 1333 /*
1334 1334 * Hold SA attribute
1335 1335 *
1336 1336 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1337 1337 *
1338 1338 * variable_size is the total size of all variable sized attributes
1339 1339 * passed to this function. It is not the total size of all
1340 1340 * variable size attributes that *may* exist on this object.
1341 1341 */
1342 1342 void
1343 1343 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1344 1344 {
1345 1345 uint64_t object;
1346 1346 sa_os_t *sa = tx->tx_objset->os_sa;
1347 1347
1348 1348 ASSERT(hdl != NULL);
1349 1349
1350 1350 object = sa_handle_object(hdl);
1351 1351
1352 1352 dmu_tx_hold_bonus(tx, object);
1353 1353
1354 1354 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1355 1355 return;
1356 1356
1357 1357 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1358 1358 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1359 1359 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1360 1360 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1361 1361 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1362 1362 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 1363 }
1364 1364
1365 1365 dmu_tx_sa_registration_hold(sa, tx);
1366 1366
1367 1367 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1368 1368 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1369 1369
1370 1370 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1371 1371 ASSERT(tx->tx_txg == 0);
1372 1372 dmu_tx_hold_spill(tx, object);
1373 1373 } else {
1374 1374 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1375 1375 dnode_t *dn;
1376 1376
1377 1377 DB_DNODE_ENTER(db);
1378 1378 dn = DB_DNODE(db);
1379 1379 if (dn->dn_have_spill) {
1380 1380 ASSERT(tx->tx_txg == 0);
1381 1381 dmu_tx_hold_spill(tx, object);
1382 1382 }
1383 1383 DB_DNODE_EXIT(db);
1384 1384 }
1385 1385 }
|
↓ open down ↓ |
473 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX