Print this page
2619 asynchronous destruction of ZFS file systems
2747 SPA versioning with zfs feature flags
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <gwilson@delphix.com>
Reviewed by: Richard Lowe <richlowe@richlowe.net>
Reviewed by: Dan Kruchinin <dan.kruchinin@gmail.com>
Approved by: Dan McDonald <danmcd@nexenta.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 *
|
↓ open down ↓ |
12 lines elided |
↑ open up ↑ |
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 - */
24 -/*
25 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 + * Copyright (c) 2012 by Delphix. All rights reserved.
26 25 */
27 26
28 27 #include <sys/dmu.h>
29 28 #include <sys/dmu_impl.h>
30 29 #include <sys/dbuf.h>
31 30 #include <sys/dmu_tx.h>
32 31 #include <sys/dmu_objset.h>
33 32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
34 33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
35 34 #include <sys/dsl_pool.h>
36 35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
37 36 #include <sys/spa.h>
38 37 #include <sys/sa.h>
39 38 #include <sys/sa_impl.h>
40 39 #include <sys/zfs_context.h>
41 40 #include <sys/varargs.h>
42 41
43 42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44 43 uint64_t arg1, uint64_t arg2);
45 44
46 45
47 46 dmu_tx_t *
48 47 dmu_tx_create_dd(dsl_dir_t *dd)
49 48 {
50 49 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
51 50 tx->tx_dir = dd;
52 51 if (dd)
53 52 tx->tx_pool = dd->dd_pool;
54 53 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
55 54 offsetof(dmu_tx_hold_t, txh_node));
56 55 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
57 56 offsetof(dmu_tx_callback_t, dcb_node));
58 57 #ifdef ZFS_DEBUG
59 58 refcount_create(&tx->tx_space_written);
60 59 refcount_create(&tx->tx_space_freed);
61 60 #endif
62 61 return (tx);
63 62 }
64 63
65 64 dmu_tx_t *
66 65 dmu_tx_create(objset_t *os)
67 66 {
68 67 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 68 tx->tx_objset = os;
70 69 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
71 70 return (tx);
72 71 }
73 72
74 73 dmu_tx_t *
75 74 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 75 {
77 76 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78 77
79 78 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
80 79 tx->tx_pool = dp;
81 80 tx->tx_txg = txg;
82 81 tx->tx_anyobj = TRUE;
83 82
84 83 return (tx);
85 84 }
86 85
87 86 int
88 87 dmu_tx_is_syncing(dmu_tx_t *tx)
89 88 {
90 89 return (tx->tx_anyobj);
91 90 }
92 91
93 92 int
94 93 dmu_tx_private_ok(dmu_tx_t *tx)
95 94 {
96 95 return (tx->tx_anyobj);
97 96 }
98 97
99 98 static dmu_tx_hold_t *
100 99 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
101 100 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
102 101 {
103 102 dmu_tx_hold_t *txh;
104 103 dnode_t *dn = NULL;
105 104 int err;
106 105
107 106 if (object != DMU_NEW_OBJECT) {
108 107 err = dnode_hold(os, object, tx, &dn);
109 108 if (err) {
110 109 tx->tx_err = err;
111 110 return (NULL);
112 111 }
113 112
114 113 if (err == 0 && tx->tx_txg != 0) {
115 114 mutex_enter(&dn->dn_mtx);
116 115 /*
117 116 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
118 117 * problem, but there's no way for it to happen (for
119 118 * now, at least).
120 119 */
121 120 ASSERT(dn->dn_assigned_txg == 0);
122 121 dn->dn_assigned_txg = tx->tx_txg;
123 122 (void) refcount_add(&dn->dn_tx_holds, tx);
124 123 mutex_exit(&dn->dn_mtx);
125 124 }
126 125 }
127 126
128 127 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 128 txh->txh_tx = tx;
130 129 txh->txh_dnode = dn;
131 130 #ifdef ZFS_DEBUG
132 131 txh->txh_type = type;
133 132 txh->txh_arg1 = arg1;
134 133 txh->txh_arg2 = arg2;
135 134 #endif
136 135 list_insert_tail(&tx->tx_holds, txh);
137 136
138 137 return (txh);
139 138 }
140 139
141 140 void
142 141 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
143 142 {
144 143 /*
145 144 * If we're syncing, they can manipulate any object anyhow, and
146 145 * the hold on the dnode_t can cause problems.
147 146 */
148 147 if (!dmu_tx_is_syncing(tx)) {
149 148 (void) dmu_tx_hold_object_impl(tx, os,
150 149 object, THT_NEWOBJECT, 0, 0);
151 150 }
152 151 }
153 152
154 153 static int
155 154 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
156 155 {
157 156 int err;
158 157 dmu_buf_impl_t *db;
159 158
160 159 rw_enter(&dn->dn_struct_rwlock, RW_READER);
161 160 db = dbuf_hold_level(dn, level, blkid, FTAG);
162 161 rw_exit(&dn->dn_struct_rwlock);
163 162 if (db == NULL)
164 163 return (EIO);
165 164 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 165 dbuf_rele(db, FTAG);
167 166 return (err);
168 167 }
169 168
170 169 static void
171 170 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
172 171 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 172 {
174 173 objset_t *os = dn->dn_objset;
175 174 dsl_dataset_t *ds = os->os_dsl_dataset;
176 175 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
177 176 dmu_buf_impl_t *parent = NULL;
178 177 blkptr_t *bp = NULL;
179 178 uint64_t space;
180 179
181 180 if (level >= dn->dn_nlevels || history[level] == blkid)
182 181 return;
183 182
184 183 history[level] = blkid;
185 184
186 185 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187 186
188 187 if (db == NULL || db == dn->dn_dbuf) {
189 188 ASSERT(level != 0);
190 189 db = NULL;
191 190 } else {
192 191 ASSERT(DB_DNODE(db) == dn);
193 192 ASSERT(db->db_level == level);
194 193 ASSERT(db->db.db_size == space);
195 194 ASSERT(db->db_blkid == blkid);
196 195 bp = db->db_blkptr;
197 196 parent = db->db_parent;
198 197 }
199 198
200 199 freeable = (bp && (freeable ||
201 200 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
202 201
203 202 if (freeable)
204 203 txh->txh_space_tooverwrite += space;
205 204 else
206 205 txh->txh_space_towrite += space;
207 206 if (bp)
208 207 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209 208
210 209 dmu_tx_count_twig(txh, dn, parent, level + 1,
211 210 blkid >> epbs, freeable, history);
212 211 }
213 212
214 213 /* ARGSUSED */
215 214 static void
216 215 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 216 {
218 217 dnode_t *dn = txh->txh_dnode;
219 218 uint64_t start, end, i;
220 219 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
221 220 int err = 0;
222 221
223 222 if (len == 0)
224 223 return;
225 224
226 225 min_bs = SPA_MINBLOCKSHIFT;
227 226 max_bs = SPA_MAXBLOCKSHIFT;
228 227 min_ibs = DN_MIN_INDBLKSHIFT;
229 228 max_ibs = DN_MAX_INDBLKSHIFT;
230 229
231 230 if (dn) {
232 231 uint64_t history[DN_MAX_LEVELS];
233 232 int nlvls = dn->dn_nlevels;
234 233 int delta;
235 234
236 235 /*
237 236 * For i/o error checking, read the first and last level-0
238 237 * blocks (if they are not aligned), and all the level-1 blocks.
239 238 */
240 239 if (dn->dn_maxblkid == 0) {
241 240 delta = dn->dn_datablksz;
242 241 start = (off < dn->dn_datablksz) ? 0 : 1;
243 242 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
244 243 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
245 244 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
246 245 if (err)
247 246 goto out;
248 247 delta -= off;
249 248 }
250 249 } else {
251 250 zio_t *zio = zio_root(dn->dn_objset->os_spa,
252 251 NULL, NULL, ZIO_FLAG_CANFAIL);
253 252
254 253 /* first level-0 block */
255 254 start = off >> dn->dn_datablkshift;
256 255 if (P2PHASE(off, dn->dn_datablksz) ||
257 256 len < dn->dn_datablksz) {
258 257 err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 258 if (err)
260 259 goto out;
261 260 }
262 261
263 262 /* last level-0 block */
264 263 end = (off+len-1) >> dn->dn_datablkshift;
265 264 if (end != start && end <= dn->dn_maxblkid &&
266 265 P2PHASE(off+len, dn->dn_datablksz)) {
267 266 err = dmu_tx_check_ioerr(zio, dn, 0, end);
268 267 if (err)
269 268 goto out;
270 269 }
271 270
272 271 /* level-1 blocks */
273 272 if (nlvls > 1) {
274 273 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
275 274 for (i = (start>>shft)+1; i < end>>shft; i++) {
276 275 err = dmu_tx_check_ioerr(zio, dn, 1, i);
277 276 if (err)
278 277 goto out;
279 278 }
280 279 }
281 280
282 281 err = zio_wait(zio);
283 282 if (err)
284 283 goto out;
285 284 delta = P2NPHASE(off, dn->dn_datablksz);
286 285 }
287 286
288 287 if (dn->dn_maxblkid > 0) {
289 288 /*
290 289 * The blocksize can't change,
291 290 * so we can make a more precise estimate.
292 291 */
293 292 ASSERT(dn->dn_datablkshift != 0);
294 293 min_bs = max_bs = dn->dn_datablkshift;
295 294 min_ibs = max_ibs = dn->dn_indblkshift;
296 295 } else if (dn->dn_indblkshift > max_ibs) {
297 296 /*
298 297 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
299 298 * the code will still work correctly on older pools.
300 299 */
301 300 min_ibs = max_ibs = dn->dn_indblkshift;
302 301 }
303 302
304 303 /*
305 304 * If this write is not off the end of the file
306 305 * we need to account for overwrites/unref.
307 306 */
308 307 if (start <= dn->dn_maxblkid) {
309 308 for (int l = 0; l < DN_MAX_LEVELS; l++)
310 309 history[l] = -1ULL;
311 310 }
312 311 while (start <= dn->dn_maxblkid) {
313 312 dmu_buf_impl_t *db;
314 313
315 314 rw_enter(&dn->dn_struct_rwlock, RW_READER);
316 315 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
317 316 rw_exit(&dn->dn_struct_rwlock);
318 317
319 318 if (err) {
320 319 txh->txh_tx->tx_err = err;
321 320 return;
322 321 }
323 322
324 323 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
325 324 history);
326 325 dbuf_rele(db, FTAG);
327 326 if (++start > end) {
328 327 /*
329 328 * Account for new indirects appearing
330 329 * before this IO gets assigned into a txg.
331 330 */
332 331 bits = 64 - min_bs;
333 332 epbs = min_ibs - SPA_BLKPTRSHIFT;
334 333 for (bits -= epbs * (nlvls - 1);
335 334 bits >= 0; bits -= epbs)
336 335 txh->txh_fudge += 1ULL << max_ibs;
337 336 goto out;
338 337 }
339 338 off += delta;
340 339 if (len >= delta)
341 340 len -= delta;
342 341 delta = dn->dn_datablksz;
343 342 }
344 343 }
345 344
346 345 /*
347 346 * 'end' is the last thing we will access, not one past.
348 347 * This way we won't overflow when accessing the last byte.
349 348 */
350 349 start = P2ALIGN(off, 1ULL << max_bs);
351 350 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
352 351 txh->txh_space_towrite += end - start + 1;
353 352
354 353 start >>= min_bs;
355 354 end >>= min_bs;
356 355
357 356 epbs = min_ibs - SPA_BLKPTRSHIFT;
358 357
359 358 /*
360 359 * The object contains at most 2^(64 - min_bs) blocks,
361 360 * and each indirect level maps 2^epbs.
362 361 */
363 362 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
364 363 start >>= epbs;
365 364 end >>= epbs;
366 365 ASSERT3U(end, >=, start);
367 366 txh->txh_space_towrite += (end - start + 1) << max_ibs;
368 367 if (start != 0) {
369 368 /*
370 369 * We also need a new blkid=0 indirect block
371 370 * to reference any existing file data.
372 371 */
373 372 txh->txh_space_towrite += 1ULL << max_ibs;
374 373 }
375 374 }
376 375
377 376 out:
378 377 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
379 378 2 * DMU_MAX_ACCESS)
380 379 err = EFBIG;
381 380
382 381 if (err)
383 382 txh->txh_tx->tx_err = err;
384 383 }
385 384
386 385 static void
387 386 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
388 387 {
389 388 dnode_t *dn = txh->txh_dnode;
390 389 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
391 390 uint64_t space = mdn->dn_datablksz +
392 391 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
393 392
394 393 if (dn && dn->dn_dbuf->db_blkptr &&
395 394 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
396 395 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
397 396 txh->txh_space_tooverwrite += space;
398 397 txh->txh_space_tounref += space;
399 398 } else {
400 399 txh->txh_space_towrite += space;
401 400 if (dn && dn->dn_dbuf->db_blkptr)
402 401 txh->txh_space_tounref += space;
403 402 }
404 403 }
405 404
406 405 void
407 406 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
408 407 {
409 408 dmu_tx_hold_t *txh;
410 409
411 410 ASSERT(tx->tx_txg == 0);
412 411 ASSERT(len < DMU_MAX_ACCESS);
413 412 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
414 413
415 414 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
416 415 object, THT_WRITE, off, len);
417 416 if (txh == NULL)
418 417 return;
419 418
420 419 dmu_tx_count_write(txh, off, len);
421 420 dmu_tx_count_dnode(txh);
422 421 }
423 422
424 423 static void
425 424 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
426 425 {
427 426 uint64_t blkid, nblks, lastblk;
428 427 uint64_t space = 0, unref = 0, skipped = 0;
429 428 dnode_t *dn = txh->txh_dnode;
430 429 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
431 430 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
432 431 int epbs;
433 432
434 433 if (dn->dn_nlevels == 0)
435 434 return;
436 435
437 436 /*
438 437 * The struct_rwlock protects us against dn_nlevels
439 438 * changing, in case (against all odds) we manage to dirty &
440 439 * sync out the changes after we check for being dirty.
441 440 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
442 441 */
443 442 rw_enter(&dn->dn_struct_rwlock, RW_READER);
444 443 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
445 444 if (dn->dn_maxblkid == 0) {
446 445 if (off == 0 && len >= dn->dn_datablksz) {
447 446 blkid = 0;
448 447 nblks = 1;
449 448 } else {
450 449 rw_exit(&dn->dn_struct_rwlock);
451 450 return;
452 451 }
453 452 } else {
454 453 blkid = off >> dn->dn_datablkshift;
455 454 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
456 455
457 456 if (blkid >= dn->dn_maxblkid) {
458 457 rw_exit(&dn->dn_struct_rwlock);
459 458 return;
460 459 }
461 460 if (blkid + nblks > dn->dn_maxblkid)
462 461 nblks = dn->dn_maxblkid - blkid;
463 462
464 463 }
465 464 if (dn->dn_nlevels == 1) {
466 465 int i;
467 466 for (i = 0; i < nblks; i++) {
468 467 blkptr_t *bp = dn->dn_phys->dn_blkptr;
469 468 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
470 469 bp += blkid + i;
471 470 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
472 471 dprintf_bp(bp, "can free old%s", "");
473 472 space += bp_get_dsize(spa, bp);
474 473 }
475 474 unref += BP_GET_ASIZE(bp);
476 475 }
477 476 nblks = 0;
478 477 }
479 478
480 479 /*
481 480 * Add in memory requirements of higher-level indirects.
482 481 * This assumes a worst-possible scenario for dn_nlevels.
483 482 */
484 483 {
485 484 uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
486 485 int level = (dn->dn_nlevels > 1) ? 2 : 1;
487 486
488 487 while (level++ < DN_MAX_LEVELS) {
489 488 txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
490 489 blkcnt = 1 + (blkcnt >> epbs);
491 490 }
492 491 ASSERT(blkcnt <= dn->dn_nblkptr);
493 492 }
494 493
495 494 lastblk = blkid + nblks - 1;
496 495 while (nblks) {
497 496 dmu_buf_impl_t *dbuf;
498 497 uint64_t ibyte, new_blkid;
499 498 int epb = 1 << epbs;
500 499 int err, i, blkoff, tochk;
501 500 blkptr_t *bp;
502 501
503 502 ibyte = blkid << dn->dn_datablkshift;
504 503 err = dnode_next_offset(dn,
505 504 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
506 505 new_blkid = ibyte >> dn->dn_datablkshift;
507 506 if (err == ESRCH) {
508 507 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
509 508 break;
510 509 }
511 510 if (err) {
512 511 txh->txh_tx->tx_err = err;
513 512 break;
514 513 }
515 514 if (new_blkid > lastblk) {
516 515 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
517 516 break;
518 517 }
519 518
520 519 if (new_blkid > blkid) {
521 520 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
522 521 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
523 522 nblks -= new_blkid - blkid;
524 523 blkid = new_blkid;
525 524 }
526 525 blkoff = P2PHASE(blkid, epb);
527 526 tochk = MIN(epb - blkoff, nblks);
528 527
529 528 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
530 529 if (err) {
531 530 txh->txh_tx->tx_err = err;
532 531 break;
533 532 }
534 533
535 534 txh->txh_memory_tohold += dbuf->db.db_size;
536 535
537 536 /*
538 537 * We don't check memory_tohold against DMU_MAX_ACCESS because
539 538 * memory_tohold is an over-estimation (especially the >L1
540 539 * indirect blocks), so it could fail. Callers should have
541 540 * already verified that they will not be holding too much
542 541 * memory.
543 542 */
544 543
545 544 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
546 545 if (err != 0) {
547 546 txh->txh_tx->tx_err = err;
548 547 dbuf_rele(dbuf, FTAG);
549 548 break;
550 549 }
551 550
552 551 bp = dbuf->db.db_data;
553 552 bp += blkoff;
554 553
555 554 for (i = 0; i < tochk; i++) {
556 555 if (dsl_dataset_block_freeable(ds, &bp[i],
557 556 bp[i].blk_birth)) {
558 557 dprintf_bp(&bp[i], "can free old%s", "");
559 558 space += bp_get_dsize(spa, &bp[i]);
560 559 }
561 560 unref += BP_GET_ASIZE(bp);
562 561 }
563 562 dbuf_rele(dbuf, FTAG);
564 563
565 564 blkid += tochk;
566 565 nblks -= tochk;
567 566 }
568 567 rw_exit(&dn->dn_struct_rwlock);
569 568
570 569 /* account for new level 1 indirect blocks that might show up */
571 570 if (skipped > 0) {
572 571 txh->txh_fudge += skipped << dn->dn_indblkshift;
573 572 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
574 573 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
575 574 }
576 575 txh->txh_space_tofree += space;
577 576 txh->txh_space_tounref += unref;
578 577 }
579 578
580 579 void
581 580 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
582 581 {
583 582 dmu_tx_hold_t *txh;
584 583 dnode_t *dn;
585 584 uint64_t start, end, i;
586 585 int err, shift;
587 586 zio_t *zio;
588 587
589 588 ASSERT(tx->tx_txg == 0);
590 589
591 590 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
592 591 object, THT_FREE, off, len);
593 592 if (txh == NULL)
594 593 return;
595 594 dn = txh->txh_dnode;
596 595
597 596 /* first block */
598 597 if (off != 0)
599 598 dmu_tx_count_write(txh, off, 1);
600 599 /* last block */
601 600 if (len != DMU_OBJECT_END)
602 601 dmu_tx_count_write(txh, off+len, 1);
603 602
604 603 dmu_tx_count_dnode(txh);
605 604
606 605 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
607 606 return;
608 607 if (len == DMU_OBJECT_END)
609 608 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
610 609
611 610 /*
612 611 * For i/o error checking, read the first and last level-0
613 612 * blocks, and all the level-1 blocks. The above count_write's
614 613 * have already taken care of the level-0 blocks.
615 614 */
616 615 if (dn->dn_nlevels > 1) {
617 616 shift = dn->dn_datablkshift + dn->dn_indblkshift -
618 617 SPA_BLKPTRSHIFT;
619 618 start = off >> shift;
620 619 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
621 620
622 621 zio = zio_root(tx->tx_pool->dp_spa,
623 622 NULL, NULL, ZIO_FLAG_CANFAIL);
624 623 for (i = start; i <= end; i++) {
625 624 uint64_t ibyte = i << shift;
626 625 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
627 626 i = ibyte >> shift;
628 627 if (err == ESRCH)
629 628 break;
630 629 if (err) {
631 630 tx->tx_err = err;
632 631 return;
633 632 }
634 633
635 634 err = dmu_tx_check_ioerr(zio, dn, 1, i);
636 635 if (err) {
637 636 tx->tx_err = err;
638 637 return;
639 638 }
640 639 }
641 640 err = zio_wait(zio);
642 641 if (err) {
643 642 tx->tx_err = err;
644 643 return;
645 644 }
646 645 }
647 646
648 647 dmu_tx_count_free(txh, off, len);
649 648 }
650 649
651 650 void
652 651 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
653 652 {
654 653 dmu_tx_hold_t *txh;
655 654 dnode_t *dn;
656 655 uint64_t nblocks;
657 656 int epbs, err;
658 657
659 658 ASSERT(tx->tx_txg == 0);
660 659
661 660 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
662 661 object, THT_ZAP, add, (uintptr_t)name);
663 662 if (txh == NULL)
664 663 return;
665 664 dn = txh->txh_dnode;
666 665
667 666 dmu_tx_count_dnode(txh);
668 667
|
↓ open down ↓ |
633 lines elided |
↑ open up ↑ |
669 668 if (dn == NULL) {
670 669 /*
671 670 * We will be able to fit a new object's entries into one leaf
672 671 * block. So there will be at most 2 blocks total,
673 672 * including the header block.
674 673 */
675 674 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
676 675 return;
677 676 }
678 677
679 - ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
678 + ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
680 679
681 680 if (dn->dn_maxblkid == 0 && !add) {
682 681 blkptr_t *bp;
683 682
684 683 /*
685 684 * If there is only one block (i.e. this is a micro-zap)
686 685 * and we are not adding anything, the accounting is simple.
687 686 */
688 687 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
689 688 if (err) {
690 689 tx->tx_err = err;
691 690 return;
692 691 }
693 692
694 693 /*
695 694 * Use max block size here, since we don't know how much
696 695 * the size will change between now and the dbuf dirty call.
697 696 */
698 697 bp = &dn->dn_phys->dn_blkptr[0];
699 698 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
700 699 bp, bp->blk_birth))
701 700 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
702 701 else
703 702 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
704 703 if (!BP_IS_HOLE(bp))
705 704 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
706 705 return;
707 706 }
708 707
709 708 if (dn->dn_maxblkid > 0 && name) {
710 709 /*
711 710 * access the name in this fat-zap so that we'll check
712 711 * for i/o errors to the leaf blocks, etc.
713 712 */
714 713 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
715 714 8, 0, NULL);
716 715 if (err == EIO) {
717 716 tx->tx_err = err;
718 717 return;
719 718 }
720 719 }
721 720
722 721 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
723 722 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
724 723
725 724 /*
726 725 * If the modified blocks are scattered to the four winds,
727 726 * we'll have to modify an indirect twig for each.
728 727 */
729 728 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
730 729 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
731 730 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
732 731 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
733 732 else
734 733 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
735 734 }
736 735
737 736 void
738 737 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
739 738 {
740 739 dmu_tx_hold_t *txh;
741 740
742 741 ASSERT(tx->tx_txg == 0);
743 742
744 743 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
745 744 object, THT_BONUS, 0, 0);
746 745 if (txh)
747 746 dmu_tx_count_dnode(txh);
748 747 }
749 748
750 749 void
751 750 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
752 751 {
753 752 dmu_tx_hold_t *txh;
754 753 ASSERT(tx->tx_txg == 0);
755 754
756 755 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
757 756 DMU_NEW_OBJECT, THT_SPACE, space, 0);
758 757
759 758 txh->txh_space_towrite += space;
760 759 }
761 760
762 761 int
763 762 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
764 763 {
765 764 dmu_tx_hold_t *txh;
766 765 int holds = 0;
767 766
768 767 /*
769 768 * By asserting that the tx is assigned, we're counting the
770 769 * number of dn_tx_holds, which is the same as the number of
771 770 * dn_holds. Otherwise, we'd be counting dn_holds, but
772 771 * dn_tx_holds could be 0.
773 772 */
774 773 ASSERT(tx->tx_txg != 0);
775 774
776 775 /* if (tx->tx_anyobj == TRUE) */
777 776 /* return (0); */
778 777
779 778 for (txh = list_head(&tx->tx_holds); txh;
780 779 txh = list_next(&tx->tx_holds, txh)) {
781 780 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
782 781 holds++;
783 782 }
784 783
785 784 return (holds);
786 785 }
787 786
788 787 #ifdef ZFS_DEBUG
789 788 void
790 789 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
791 790 {
792 791 dmu_tx_hold_t *txh;
793 792 int match_object = FALSE, match_offset = FALSE;
794 793 dnode_t *dn;
795 794
796 795 DB_DNODE_ENTER(db);
797 796 dn = DB_DNODE(db);
798 797 ASSERT(tx->tx_txg != 0);
799 798 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
800 799 ASSERT3U(dn->dn_object, ==, db->db.db_object);
801 800
802 801 if (tx->tx_anyobj) {
803 802 DB_DNODE_EXIT(db);
804 803 return;
805 804 }
806 805
807 806 /* XXX No checking on the meta dnode for now */
808 807 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
809 808 DB_DNODE_EXIT(db);
810 809 return;
811 810 }
812 811
813 812 for (txh = list_head(&tx->tx_holds); txh;
814 813 txh = list_next(&tx->tx_holds, txh)) {
815 814 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
816 815 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
817 816 match_object = TRUE;
818 817 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
819 818 int datablkshift = dn->dn_datablkshift ?
820 819 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
821 820 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
822 821 int shift = datablkshift + epbs * db->db_level;
823 822 uint64_t beginblk = shift >= 64 ? 0 :
824 823 (txh->txh_arg1 >> shift);
825 824 uint64_t endblk = shift >= 64 ? 0 :
826 825 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
827 826 uint64_t blkid = db->db_blkid;
828 827
829 828 /* XXX txh_arg2 better not be zero... */
830 829
831 830 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
832 831 txh->txh_type, beginblk, endblk);
833 832
834 833 switch (txh->txh_type) {
835 834 case THT_WRITE:
836 835 if (blkid >= beginblk && blkid <= endblk)
837 836 match_offset = TRUE;
838 837 /*
839 838 * We will let this hold work for the bonus
840 839 * or spill buffer so that we don't need to
841 840 * hold it when creating a new object.
842 841 */
843 842 if (blkid == DMU_BONUS_BLKID ||
844 843 blkid == DMU_SPILL_BLKID)
845 844 match_offset = TRUE;
846 845 /*
847 846 * They might have to increase nlevels,
848 847 * thus dirtying the new TLIBs. Or the
849 848 * might have to change the block size,
850 849 * thus dirying the new lvl=0 blk=0.
851 850 */
852 851 if (blkid == 0)
853 852 match_offset = TRUE;
854 853 break;
855 854 case THT_FREE:
856 855 /*
857 856 * We will dirty all the level 1 blocks in
858 857 * the free range and perhaps the first and
859 858 * last level 0 block.
860 859 */
861 860 if (blkid >= beginblk && (blkid <= endblk ||
862 861 txh->txh_arg2 == DMU_OBJECT_END))
863 862 match_offset = TRUE;
864 863 break;
865 864 case THT_SPILL:
866 865 if (blkid == DMU_SPILL_BLKID)
867 866 match_offset = TRUE;
868 867 break;
869 868 case THT_BONUS:
870 869 if (blkid == DMU_BONUS_BLKID)
871 870 match_offset = TRUE;
872 871 break;
873 872 case THT_ZAP:
874 873 match_offset = TRUE;
875 874 break;
876 875 case THT_NEWOBJECT:
877 876 match_object = TRUE;
878 877 break;
879 878 default:
880 879 ASSERT(!"bad txh_type");
881 880 }
882 881 }
883 882 if (match_object && match_offset) {
884 883 DB_DNODE_EXIT(db);
885 884 return;
886 885 }
887 886 }
888 887 DB_DNODE_EXIT(db);
889 888 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
890 889 (u_longlong_t)db->db.db_object, db->db_level,
891 890 (u_longlong_t)db->db_blkid);
892 891 }
893 892 #endif
894 893
895 894 static int
896 895 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
897 896 {
898 897 dmu_tx_hold_t *txh;
899 898 spa_t *spa = tx->tx_pool->dp_spa;
900 899 uint64_t memory, asize, fsize, usize;
901 900 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
902 901
903 902 ASSERT3U(tx->tx_txg, ==, 0);
904 903
905 904 if (tx->tx_err)
906 905 return (tx->tx_err);
907 906
908 907 if (spa_suspended(spa)) {
909 908 /*
910 909 * If the user has indicated a blocking failure mode
911 910 * then return ERESTART which will block in dmu_tx_wait().
912 911 * Otherwise, return EIO so that an error can get
913 912 * propagated back to the VOP calls.
914 913 *
915 914 * Note that we always honor the txg_how flag regardless
916 915 * of the failuremode setting.
917 916 */
918 917 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
919 918 txg_how != TXG_WAIT)
920 919 return (EIO);
921 920
922 921 return (ERESTART);
923 922 }
924 923
925 924 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
926 925 tx->tx_needassign_txh = NULL;
927 926
928 927 /*
929 928 * NB: No error returns are allowed after txg_hold_open, but
930 929 * before processing the dnode holds, due to the
931 930 * dmu_tx_unassign() logic.
932 931 */
933 932
934 933 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
935 934 for (txh = list_head(&tx->tx_holds); txh;
936 935 txh = list_next(&tx->tx_holds, txh)) {
937 936 dnode_t *dn = txh->txh_dnode;
938 937 if (dn != NULL) {
939 938 mutex_enter(&dn->dn_mtx);
940 939 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
941 940 mutex_exit(&dn->dn_mtx);
942 941 tx->tx_needassign_txh = txh;
943 942 return (ERESTART);
944 943 }
945 944 if (dn->dn_assigned_txg == 0)
946 945 dn->dn_assigned_txg = tx->tx_txg;
947 946 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
948 947 (void) refcount_add(&dn->dn_tx_holds, tx);
949 948 mutex_exit(&dn->dn_mtx);
950 949 }
951 950 towrite += txh->txh_space_towrite;
952 951 tofree += txh->txh_space_tofree;
953 952 tooverwrite += txh->txh_space_tooverwrite;
954 953 tounref += txh->txh_space_tounref;
955 954 tohold += txh->txh_memory_tohold;
956 955 fudge += txh->txh_fudge;
957 956 }
958 957
959 958 /*
960 959 * NB: This check must be after we've held the dnodes, so that
961 960 * the dmu_tx_unassign() logic will work properly
962 961 */
963 962 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
964 963 return (ERESTART);
965 964
966 965 /*
967 966 * If a snapshot has been taken since we made our estimates,
968 967 * assume that we won't be able to free or overwrite anything.
969 968 */
970 969 if (tx->tx_objset &&
971 970 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
972 971 tx->tx_lastsnap_txg) {
973 972 towrite += tooverwrite;
974 973 tooverwrite = tofree = 0;
975 974 }
976 975
977 976 /* needed allocation: worst-case estimate of write space */
978 977 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
979 978 /* freed space estimate: worst-case overwrite + free estimate */
980 979 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
981 980 /* convert unrefd space to worst-case estimate */
982 981 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
983 982 /* calculate memory footprint estimate */
984 983 memory = towrite + tooverwrite + tohold;
985 984
986 985 #ifdef ZFS_DEBUG
987 986 /*
988 987 * Add in 'tohold' to account for our dirty holds on this memory
989 988 * XXX - the "fudge" factor is to account for skipped blocks that
990 989 * we missed because dnode_next_offset() misses in-core-only blocks.
991 990 */
992 991 tx->tx_space_towrite = asize +
993 992 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
994 993 tx->tx_space_tofree = tofree;
995 994 tx->tx_space_tooverwrite = tooverwrite;
996 995 tx->tx_space_tounref = tounref;
997 996 #endif
998 997
999 998 if (tx->tx_dir && asize != 0) {
1000 999 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1001 1000 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1002 1001 if (err)
1003 1002 return (err);
1004 1003 }
1005 1004
1006 1005 return (0);
1007 1006 }
1008 1007
1009 1008 static void
1010 1009 dmu_tx_unassign(dmu_tx_t *tx)
1011 1010 {
1012 1011 dmu_tx_hold_t *txh;
1013 1012
1014 1013 if (tx->tx_txg == 0)
1015 1014 return;
1016 1015
1017 1016 txg_rele_to_quiesce(&tx->tx_txgh);
1018 1017
1019 1018 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1020 1019 txh = list_next(&tx->tx_holds, txh)) {
1021 1020 dnode_t *dn = txh->txh_dnode;
1022 1021
1023 1022 if (dn == NULL)
1024 1023 continue;
1025 1024 mutex_enter(&dn->dn_mtx);
1026 1025 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1027 1026
1028 1027 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1029 1028 dn->dn_assigned_txg = 0;
1030 1029 cv_broadcast(&dn->dn_notxholds);
1031 1030 }
1032 1031 mutex_exit(&dn->dn_mtx);
1033 1032 }
1034 1033
1035 1034 txg_rele_to_sync(&tx->tx_txgh);
1036 1035
1037 1036 tx->tx_lasttried_txg = tx->tx_txg;
1038 1037 tx->tx_txg = 0;
1039 1038 }
1040 1039
1041 1040 /*
1042 1041 * Assign tx to a transaction group. txg_how can be one of:
1043 1042 *
1044 1043 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1045 1044 * a new one. This should be used when you're not holding locks.
1046 1045 * If will only fail if we're truly out of space (or over quota).
1047 1046 *
1048 1047 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1049 1048 * blocking, returns immediately with ERESTART. This should be used
1050 1049 * whenever you're holding locks. On an ERESTART error, the caller
1051 1050 * should drop locks, do a dmu_tx_wait(tx), and try again.
1052 1051 *
1053 1052 * (3) A specific txg. Use this if you need to ensure that multiple
1054 1053 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1055 1054 * returns ERESTART if it can't assign you into the requested txg.
1056 1055 */
1057 1056 int
1058 1057 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1059 1058 {
1060 1059 int err;
1061 1060
1062 1061 ASSERT(tx->tx_txg == 0);
1063 1062 ASSERT(txg_how != 0);
1064 1063 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1065 1064
1066 1065 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1067 1066 dmu_tx_unassign(tx);
1068 1067
1069 1068 if (err != ERESTART || txg_how != TXG_WAIT)
1070 1069 return (err);
1071 1070
1072 1071 dmu_tx_wait(tx);
1073 1072 }
1074 1073
1075 1074 txg_rele_to_quiesce(&tx->tx_txgh);
1076 1075
1077 1076 return (0);
1078 1077 }
1079 1078
1080 1079 void
1081 1080 dmu_tx_wait(dmu_tx_t *tx)
1082 1081 {
1083 1082 spa_t *spa = tx->tx_pool->dp_spa;
1084 1083
1085 1084 ASSERT(tx->tx_txg == 0);
1086 1085
1087 1086 /*
1088 1087 * It's possible that the pool has become active after this thread
1089 1088 * has tried to obtain a tx. If that's the case then his
1090 1089 * tx_lasttried_txg would not have been assigned.
1091 1090 */
1092 1091 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1093 1092 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1094 1093 } else if (tx->tx_needassign_txh) {
1095 1094 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1096 1095
1097 1096 mutex_enter(&dn->dn_mtx);
1098 1097 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1099 1098 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1100 1099 mutex_exit(&dn->dn_mtx);
1101 1100 tx->tx_needassign_txh = NULL;
1102 1101 } else {
1103 1102 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1104 1103 }
1105 1104 }
1106 1105
1107 1106 void
1108 1107 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1109 1108 {
1110 1109 #ifdef ZFS_DEBUG
1111 1110 if (tx->tx_dir == NULL || delta == 0)
1112 1111 return;
1113 1112
1114 1113 if (delta > 0) {
1115 1114 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1116 1115 tx->tx_space_towrite);
1117 1116 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1118 1117 } else {
1119 1118 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1120 1119 }
1121 1120 #endif
1122 1121 }
1123 1122
1124 1123 void
1125 1124 dmu_tx_commit(dmu_tx_t *tx)
1126 1125 {
1127 1126 dmu_tx_hold_t *txh;
1128 1127
1129 1128 ASSERT(tx->tx_txg != 0);
1130 1129
1131 1130 while (txh = list_head(&tx->tx_holds)) {
1132 1131 dnode_t *dn = txh->txh_dnode;
1133 1132
1134 1133 list_remove(&tx->tx_holds, txh);
1135 1134 kmem_free(txh, sizeof (dmu_tx_hold_t));
1136 1135 if (dn == NULL)
1137 1136 continue;
1138 1137 mutex_enter(&dn->dn_mtx);
1139 1138 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1140 1139
1141 1140 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1142 1141 dn->dn_assigned_txg = 0;
1143 1142 cv_broadcast(&dn->dn_notxholds);
1144 1143 }
1145 1144 mutex_exit(&dn->dn_mtx);
1146 1145 dnode_rele(dn, tx);
1147 1146 }
1148 1147
1149 1148 if (tx->tx_tempreserve_cookie)
1150 1149 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1151 1150
1152 1151 if (!list_is_empty(&tx->tx_callbacks))
1153 1152 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1154 1153
1155 1154 if (tx->tx_anyobj == FALSE)
1156 1155 txg_rele_to_sync(&tx->tx_txgh);
1157 1156
1158 1157 list_destroy(&tx->tx_callbacks);
1159 1158 list_destroy(&tx->tx_holds);
1160 1159 #ifdef ZFS_DEBUG
1161 1160 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1162 1161 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1163 1162 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1164 1163 refcount_destroy_many(&tx->tx_space_written,
1165 1164 refcount_count(&tx->tx_space_written));
1166 1165 refcount_destroy_many(&tx->tx_space_freed,
1167 1166 refcount_count(&tx->tx_space_freed));
1168 1167 #endif
1169 1168 kmem_free(tx, sizeof (dmu_tx_t));
1170 1169 }
1171 1170
1172 1171 void
1173 1172 dmu_tx_abort(dmu_tx_t *tx)
1174 1173 {
1175 1174 dmu_tx_hold_t *txh;
1176 1175
1177 1176 ASSERT(tx->tx_txg == 0);
1178 1177
1179 1178 while (txh = list_head(&tx->tx_holds)) {
1180 1179 dnode_t *dn = txh->txh_dnode;
1181 1180
1182 1181 list_remove(&tx->tx_holds, txh);
1183 1182 kmem_free(txh, sizeof (dmu_tx_hold_t));
1184 1183 if (dn != NULL)
1185 1184 dnode_rele(dn, tx);
1186 1185 }
1187 1186
1188 1187 /*
1189 1188 * Call any registered callbacks with an error code.
1190 1189 */
1191 1190 if (!list_is_empty(&tx->tx_callbacks))
1192 1191 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1193 1192
1194 1193 list_destroy(&tx->tx_callbacks);
1195 1194 list_destroy(&tx->tx_holds);
1196 1195 #ifdef ZFS_DEBUG
1197 1196 refcount_destroy_many(&tx->tx_space_written,
1198 1197 refcount_count(&tx->tx_space_written));
1199 1198 refcount_destroy_many(&tx->tx_space_freed,
1200 1199 refcount_count(&tx->tx_space_freed));
1201 1200 #endif
1202 1201 kmem_free(tx, sizeof (dmu_tx_t));
1203 1202 }
1204 1203
1205 1204 uint64_t
1206 1205 dmu_tx_get_txg(dmu_tx_t *tx)
1207 1206 {
1208 1207 ASSERT(tx->tx_txg != 0);
1209 1208 return (tx->tx_txg);
1210 1209 }
1211 1210
1212 1211 void
1213 1212 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1214 1213 {
1215 1214 dmu_tx_callback_t *dcb;
1216 1215
1217 1216 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1218 1217
1219 1218 dcb->dcb_func = func;
1220 1219 dcb->dcb_data = data;
1221 1220
1222 1221 list_insert_tail(&tx->tx_callbacks, dcb);
1223 1222 }
1224 1223
1225 1224 /*
1226 1225 * Call all the commit callbacks on a list, with a given error code.
1227 1226 */
1228 1227 void
1229 1228 dmu_tx_do_callbacks(list_t *cb_list, int error)
1230 1229 {
1231 1230 dmu_tx_callback_t *dcb;
1232 1231
1233 1232 while (dcb = list_head(cb_list)) {
1234 1233 list_remove(cb_list, dcb);
1235 1234 dcb->dcb_func(dcb->dcb_data, error);
1236 1235 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1237 1236 }
1238 1237 }
1239 1238
1240 1239 /*
1241 1240 * Interface to hold a bunch of attributes.
1242 1241 * used for creating new files.
1243 1242 * attrsize is the total size of all attributes
1244 1243 * to be added during object creation
1245 1244 *
1246 1245 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1247 1246 */
1248 1247
1249 1248 /*
1250 1249 * hold necessary attribute name for attribute registration.
1251 1250 * should be a very rare case where this is needed. If it does
1252 1251 * happen it would only happen on the first write to the file system.
1253 1252 */
1254 1253 static void
1255 1254 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1256 1255 {
1257 1256 int i;
1258 1257
1259 1258 if (!sa->sa_need_attr_registration)
1260 1259 return;
1261 1260
1262 1261 for (i = 0; i != sa->sa_num_attrs; i++) {
1263 1262 if (!sa->sa_attr_table[i].sa_registered) {
1264 1263 if (sa->sa_reg_attr_obj)
1265 1264 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1266 1265 B_TRUE, sa->sa_attr_table[i].sa_name);
1267 1266 else
1268 1267 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1269 1268 B_TRUE, sa->sa_attr_table[i].sa_name);
1270 1269 }
1271 1270 }
1272 1271 }
1273 1272
1274 1273
1275 1274 void
1276 1275 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1277 1276 {
1278 1277 dnode_t *dn;
1279 1278 dmu_tx_hold_t *txh;
1280 1279
1281 1280 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1282 1281 THT_SPILL, 0, 0);
1283 1282
1284 1283 dn = txh->txh_dnode;
1285 1284
1286 1285 if (dn == NULL)
1287 1286 return;
1288 1287
1289 1288 /* If blkptr doesn't exist then add space to towrite */
1290 1289 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1291 1290 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1292 1291 } else {
1293 1292 blkptr_t *bp;
1294 1293
1295 1294 bp = &dn->dn_phys->dn_spill;
1296 1295 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1297 1296 bp, bp->blk_birth))
1298 1297 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1299 1298 else
1300 1299 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1301 1300 if (!BP_IS_HOLE(bp))
1302 1301 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1303 1302 }
1304 1303 }
1305 1304
1306 1305 void
1307 1306 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1308 1307 {
1309 1308 sa_os_t *sa = tx->tx_objset->os_sa;
1310 1309
1311 1310 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1312 1311
1313 1312 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1314 1313 return;
1315 1314
1316 1315 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1317 1316 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1318 1317 else {
1319 1318 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1320 1319 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1321 1320 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 1321 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1323 1322 }
1324 1323
1325 1324 dmu_tx_sa_registration_hold(sa, tx);
1326 1325
1327 1326 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1328 1327 return;
1329 1328
1330 1329 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1331 1330 THT_SPILL, 0, 0);
1332 1331 }
1333 1332
1334 1333 /*
1335 1334 * Hold SA attribute
1336 1335 *
1337 1336 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1338 1337 *
1339 1338 * variable_size is the total size of all variable sized attributes
1340 1339 * passed to this function. It is not the total size of all
1341 1340 * variable size attributes that *may* exist on this object.
1342 1341 */
1343 1342 void
1344 1343 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1345 1344 {
1346 1345 uint64_t object;
1347 1346 sa_os_t *sa = tx->tx_objset->os_sa;
1348 1347
1349 1348 ASSERT(hdl != NULL);
1350 1349
1351 1350 object = sa_handle_object(hdl);
1352 1351
1353 1352 dmu_tx_hold_bonus(tx, object);
1354 1353
1355 1354 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1356 1355 return;
1357 1356
1358 1357 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1359 1358 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1360 1359 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1361 1360 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1362 1361 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 1362 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1364 1363 }
1365 1364
1366 1365 dmu_tx_sa_registration_hold(sa, tx);
1367 1366
1368 1367 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1369 1368 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1370 1369
1371 1370 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1372 1371 ASSERT(tx->tx_txg == 0);
1373 1372 dmu_tx_hold_spill(tx, object);
1374 1373 } else {
1375 1374 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1376 1375 dnode_t *dn;
1377 1376
1378 1377 DB_DNODE_ENTER(db);
1379 1378 dn = DB_DNODE(db);
1380 1379 if (dn->dn_have_spill) {
1381 1380 ASSERT(tx->tx_txg == 0);
1382 1381 dmu_tx_hold_spill(tx, object);
1383 1382 }
1384 1383 DB_DNODE_EXIT(db);
1385 1384 }
1386 1385 }
|
↓ open down ↓ |
697 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX