1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or https://opensource.org/licenses/CDDL-1.0. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved. 25 */ 26 27 #include <sys/dmu.h> 28 #include <sys/dmu_impl.h> 29 #include <sys/dbuf.h> 30 #include <sys/dmu_tx.h> 31 #include <sys/dmu_objset.h> 32 #include <sys/dsl_dataset.h> 33 #include <sys/dsl_dir.h> 34 #include <sys/dsl_pool.h> 35 #include <sys/zap_impl.h> 36 #include <sys/spa.h> 37 #include <sys/sa.h> 38 #include <sys/sa_impl.h> 39 #include <sys/zfs_context.h> 40 #include <sys/trace_zfs.h> 41 42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn, 43 uint64_t arg1, uint64_t arg2); 44 45 dmu_tx_stats_t dmu_tx_stats = { 46 { "dmu_tx_assigned", KSTAT_DATA_UINT64 }, 47 { "dmu_tx_delay", KSTAT_DATA_UINT64 }, 48 { "dmu_tx_error", KSTAT_DATA_UINT64 }, 49 { "dmu_tx_suspended", KSTAT_DATA_UINT64 }, 50 { "dmu_tx_group", KSTAT_DATA_UINT64 }, 51 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 }, 52 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 }, 53 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 }, 54 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 }, 55 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 }, 56 { "dmu_tx_dirty_frees_delay", KSTAT_DATA_UINT64 }, 57 { "dmu_tx_wrlog_delay", KSTAT_DATA_UINT64 }, 58 { "dmu_tx_quota", KSTAT_DATA_UINT64 }, 59 }; 60 61 static kstat_t *dmu_tx_ksp; 62 63 dmu_tx_t * 64 dmu_tx_create_dd(dsl_dir_t *dd) 65 { 66 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP); 67 tx->tx_dir = dd; 68 if (dd != NULL) 69 tx->tx_pool = dd->dd_pool; 70 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t), 71 offsetof(dmu_tx_hold_t, txh_node)); 72 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t), 73 offsetof(dmu_tx_callback_t, dcb_node)); 74 tx->tx_start = gethrtime(); 75 return (tx); 76 } 77 78 dmu_tx_t * 79 dmu_tx_create(objset_t *os) 80 { 81 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir); 82 tx->tx_objset = os; 83 return (tx); 84 } 85 86 dmu_tx_t * 87 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg) 88 { 89 dmu_tx_t *tx = dmu_tx_create_dd(NULL); 90 91 TXG_VERIFY(dp->dp_spa, txg); 92 tx->tx_pool = dp; 93 tx->tx_txg = txg; 94 tx->tx_anyobj = TRUE; 95 96 return (tx); 97 } 98 99 int 100 dmu_tx_is_syncing(dmu_tx_t *tx) 101 { 102 return (tx->tx_anyobj); 103 } 104 105 int 106 dmu_tx_private_ok(dmu_tx_t *tx) 107 { 108 return (tx->tx_anyobj); 109 } 110 111 static dmu_tx_hold_t * 112 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type, 113 uint64_t arg1, uint64_t arg2) 114 { 115 dmu_tx_hold_t *txh; 116 117 if (dn != NULL) { 118 (void) zfs_refcount_add(&dn->dn_holds, tx); 119 if (tx->tx_txg != 0) { 120 mutex_enter(&dn->dn_mtx); 121 /* 122 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a 123 * problem, but there's no way for it to happen (for 124 * now, at least). 125 */ 126 ASSERT(dn->dn_assigned_txg == 0); 127 dn->dn_assigned_txg = tx->tx_txg; 128 (void) zfs_refcount_add(&dn->dn_tx_holds, tx); 129 mutex_exit(&dn->dn_mtx); 130 } 131 } 132 133 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP); 134 txh->txh_tx = tx; 135 txh->txh_dnode = dn; 136 zfs_refcount_create(&txh->txh_space_towrite); 137 zfs_refcount_create(&txh->txh_memory_tohold); 138 txh->txh_type = type; 139 txh->txh_arg1 = arg1; 140 txh->txh_arg2 = arg2; 141 list_insert_tail(&tx->tx_holds, txh); 142 143 return (txh); 144 } 145 146 static dmu_tx_hold_t * 147 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object, 148 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2) 149 { 150 dnode_t *dn = NULL; 151 dmu_tx_hold_t *txh; 152 int err; 153 154 if (object != DMU_NEW_OBJECT) { 155 err = dnode_hold(os, object, FTAG, &dn); 156 if (err != 0) { 157 tx->tx_err = err; 158 return (NULL); 159 } 160 } 161 txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2); 162 if (dn != NULL) 163 dnode_rele(dn, FTAG); 164 return (txh); 165 } 166 167 void 168 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn) 169 { 170 /* 171 * If we're syncing, they can manipulate any object anyhow, and 172 * the hold on the dnode_t can cause problems. 173 */ 174 if (!dmu_tx_is_syncing(tx)) 175 (void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0); 176 } 177 178 /* 179 * This function reads specified data from disk. The specified data will 180 * be needed to perform the transaction -- i.e, it will be read after 181 * we do dmu_tx_assign(). There are two reasons that we read the data now 182 * (before dmu_tx_assign()): 183 * 184 * 1. Reading it now has potentially better performance. The transaction 185 * has not yet been assigned, so the TXG is not held open, and also the 186 * caller typically has less locks held when calling dmu_tx_hold_*() than 187 * after the transaction has been assigned. This reduces the lock (and txg) 188 * hold times, thus reducing lock contention. 189 * 190 * 2. It is easier for callers (primarily the ZPL) to handle i/o errors 191 * that are detected before they start making changes to the DMU state 192 * (i.e. now). Once the transaction has been assigned, and some DMU 193 * state has been changed, it can be difficult to recover from an i/o 194 * error (e.g. to undo the changes already made in memory at the DMU 195 * layer). Typically code to do so does not exist in the caller -- it 196 * assumes that the data has already been cached and thus i/o errors are 197 * not possible. 198 * 199 * It has been observed that the i/o initiated here can be a performance 200 * problem, and it appears to be optional, because we don't look at the 201 * data which is read. However, removing this read would only serve to 202 * move the work elsewhere (after the dmu_tx_assign()), where it may 203 * have a greater impact on performance (in addition to the impact on 204 * fault tolerance noted above). 205 */ 206 static int 207 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid) 208 { 209 int err; 210 dmu_buf_impl_t *db; 211 212 rw_enter(&dn->dn_struct_rwlock, RW_READER); 213 db = dbuf_hold_level(dn, level, blkid, FTAG); 214 rw_exit(&dn->dn_struct_rwlock); 215 if (db == NULL) 216 return (SET_ERROR(EIO)); 217 /* 218 * PARTIAL_FIRST allows caching for uncacheable blocks. It will 219 * be cleared after dmu_buf_will_dirty() call dbuf_read() again. 220 */ 221 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH | 222 (level == 0 ? DB_RF_PARTIAL_FIRST : 0)); 223 dbuf_rele(db, FTAG); 224 return (err); 225 } 226 227 static void 228 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 229 { 230 dnode_t *dn = txh->txh_dnode; 231 int err = 0; 232 233 if (len == 0) 234 return; 235 236 (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG); 237 238 if (dn == NULL) 239 return; 240 241 /* 242 * For i/o error checking, read the blocks that will be needed 243 * to perform the write: the first and last level-0 blocks (if 244 * they are not aligned, i.e. if they are partial-block writes), 245 * and all the level-1 blocks. 246 */ 247 if (dn->dn_maxblkid == 0) { 248 if (off < dn->dn_datablksz && 249 (off > 0 || len < dn->dn_datablksz)) { 250 err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 251 if (err != 0) { 252 txh->txh_tx->tx_err = err; 253 } 254 } 255 } else { 256 zio_t *zio = zio_root(dn->dn_objset->os_spa, 257 NULL, NULL, ZIO_FLAG_CANFAIL); 258 259 /* first level-0 block */ 260 uint64_t start = off >> dn->dn_datablkshift; 261 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) { 262 err = dmu_tx_check_ioerr(zio, dn, 0, start); 263 if (err != 0) { 264 txh->txh_tx->tx_err = err; 265 } 266 } 267 268 /* last level-0 block */ 269 uint64_t end = (off + len - 1) >> dn->dn_datablkshift; 270 if (end != start && end <= dn->dn_maxblkid && 271 P2PHASE(off + len, dn->dn_datablksz)) { 272 err = dmu_tx_check_ioerr(zio, dn, 0, end); 273 if (err != 0) { 274 txh->txh_tx->tx_err = err; 275 } 276 } 277 278 /* level-1 blocks */ 279 if (dn->dn_nlevels > 1) { 280 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 281 for (uint64_t i = (start >> shft) + 1; 282 i < end >> shft; i++) { 283 err = dmu_tx_check_ioerr(zio, dn, 1, i); 284 if (err != 0) { 285 txh->txh_tx->tx_err = err; 286 } 287 } 288 } 289 290 err = zio_wait(zio); 291 if (err != 0) { 292 txh->txh_tx->tx_err = err; 293 } 294 } 295 } 296 297 static void 298 dmu_tx_count_append(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 299 { 300 dnode_t *dn = txh->txh_dnode; 301 int err = 0; 302 303 if (len == 0) 304 return; 305 306 (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG); 307 308 if (dn == NULL) 309 return; 310 311 /* 312 * For i/o error checking, read the blocks that will be needed 313 * to perform the append; first level-0 block (if not aligned, i.e. 314 * if they are partial-block writes), no additional blocks are read. 315 */ 316 if (dn->dn_maxblkid == 0) { 317 if (off < dn->dn_datablksz && 318 (off > 0 || len < dn->dn_datablksz)) { 319 err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 320 if (err != 0) { 321 txh->txh_tx->tx_err = err; 322 } 323 } 324 } else { 325 zio_t *zio = zio_root(dn->dn_objset->os_spa, 326 NULL, NULL, ZIO_FLAG_CANFAIL); 327 328 /* first level-0 block */ 329 uint64_t start = off >> dn->dn_datablkshift; 330 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) { 331 err = dmu_tx_check_ioerr(zio, dn, 0, start); 332 if (err != 0) { 333 txh->txh_tx->tx_err = err; 334 } 335 } 336 337 err = zio_wait(zio); 338 if (err != 0) { 339 txh->txh_tx->tx_err = err; 340 } 341 } 342 } 343 344 static void 345 dmu_tx_count_dnode(dmu_tx_hold_t *txh) 346 { 347 (void) zfs_refcount_add_many(&txh->txh_space_towrite, 348 DNODE_MIN_SIZE, FTAG); 349 } 350 351 void 352 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len) 353 { 354 dmu_tx_hold_t *txh; 355 356 ASSERT0(tx->tx_txg); 357 ASSERT3U(len, <=, DMU_MAX_ACCESS); 358 ASSERT(len == 0 || UINT64_MAX - off >= len - 1); 359 360 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 361 object, THT_WRITE, off, len); 362 if (txh != NULL) { 363 dmu_tx_count_write(txh, off, len); 364 dmu_tx_count_dnode(txh); 365 } 366 } 367 368 void 369 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len) 370 { 371 dmu_tx_hold_t *txh; 372 373 ASSERT0(tx->tx_txg); 374 ASSERT3U(len, <=, DMU_MAX_ACCESS); 375 ASSERT(len == 0 || UINT64_MAX - off >= len - 1); 376 377 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len); 378 if (txh != NULL) { 379 dmu_tx_count_write(txh, off, len); 380 dmu_tx_count_dnode(txh); 381 } 382 } 383 384 /* 385 * Should be used when appending to an object and the exact offset is unknown. 386 * The write must occur at or beyond the specified offset. Only the L0 block 387 * at provided offset will be prefetched. 388 */ 389 void 390 dmu_tx_hold_append(dmu_tx_t *tx, uint64_t object, uint64_t off, int len) 391 { 392 dmu_tx_hold_t *txh; 393 394 ASSERT0(tx->tx_txg); 395 ASSERT3U(len, <=, DMU_MAX_ACCESS); 396 397 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 398 object, THT_APPEND, off, DMU_OBJECT_END); 399 if (txh != NULL) { 400 dmu_tx_count_append(txh, off, len); 401 dmu_tx_count_dnode(txh); 402 } 403 } 404 405 void 406 dmu_tx_hold_append_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len) 407 { 408 dmu_tx_hold_t *txh; 409 410 ASSERT0(tx->tx_txg); 411 ASSERT3U(len, <=, DMU_MAX_ACCESS); 412 413 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_APPEND, off, DMU_OBJECT_END); 414 if (txh != NULL) { 415 dmu_tx_count_append(txh, off, len); 416 dmu_tx_count_dnode(txh); 417 } 418 } 419 420 /* 421 * This function marks the transaction as being a "net free". The end 422 * result is that refquotas will be disabled for this transaction, and 423 * this transaction will be able to use half of the pool space overhead 424 * (see dsl_pool_adjustedsize()). Therefore this function should only 425 * be called for transactions that we expect will not cause a net increase 426 * in the amount of space used (but it's OK if that is occasionally not true). 427 */ 428 void 429 dmu_tx_mark_netfree(dmu_tx_t *tx) 430 { 431 tx->tx_netfree = B_TRUE; 432 } 433 434 static void 435 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 436 { 437 dmu_tx_t *tx = txh->txh_tx; 438 dnode_t *dn = txh->txh_dnode; 439 int err; 440 441 ASSERT(tx->tx_txg == 0); 442 443 if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz) 444 return; 445 if (len == DMU_OBJECT_END) 446 len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off; 447 448 /* 449 * For i/o error checking, we read the first and last level-0 450 * blocks if they are not aligned, and all the level-1 blocks. 451 * 452 * Note: dbuf_free_range() assumes that we have not instantiated 453 * any level-0 dbufs that will be completely freed. Therefore we must 454 * exercise care to not read or count the first and last blocks 455 * if they are blocksize-aligned. 456 */ 457 if (dn->dn_datablkshift == 0) { 458 if (off != 0 || len < dn->dn_datablksz) 459 dmu_tx_count_write(txh, 0, dn->dn_datablksz); 460 } else { 461 /* first block will be modified if it is not aligned */ 462 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift)) 463 dmu_tx_count_write(txh, off, 1); 464 /* last block will be modified if it is not aligned */ 465 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift)) 466 dmu_tx_count_write(txh, off + len, 1); 467 } 468 469 /* 470 * Check level-1 blocks. 471 */ 472 if (dn->dn_nlevels > 1) { 473 int shift = dn->dn_datablkshift + dn->dn_indblkshift - 474 SPA_BLKPTRSHIFT; 475 uint64_t start = off >> shift; 476 uint64_t end = (off + len) >> shift; 477 478 ASSERT(dn->dn_indblkshift != 0); 479 480 /* 481 * dnode_reallocate() can result in an object with indirect 482 * blocks having an odd data block size. In this case, 483 * just check the single block. 484 */ 485 if (dn->dn_datablkshift == 0) 486 start = end = 0; 487 488 zio_t *zio = zio_root(tx->tx_pool->dp_spa, 489 NULL, NULL, ZIO_FLAG_CANFAIL); 490 for (uint64_t i = start; i <= end; i++) { 491 uint64_t ibyte = i << shift; 492 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0); 493 i = ibyte >> shift; 494 if (err == ESRCH || i > end) 495 break; 496 if (err != 0) { 497 tx->tx_err = err; 498 (void) zio_wait(zio); 499 return; 500 } 501 502 (void) zfs_refcount_add_many(&txh->txh_memory_tohold, 503 1 << dn->dn_indblkshift, FTAG); 504 505 err = dmu_tx_check_ioerr(zio, dn, 1, i); 506 if (err != 0) { 507 tx->tx_err = err; 508 (void) zio_wait(zio); 509 return; 510 } 511 } 512 err = zio_wait(zio); 513 if (err != 0) { 514 tx->tx_err = err; 515 return; 516 } 517 } 518 } 519 520 void 521 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len) 522 { 523 dmu_tx_hold_t *txh; 524 525 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 526 object, THT_FREE, off, len); 527 if (txh != NULL) { 528 dmu_tx_count_dnode(txh); 529 dmu_tx_count_free(txh, off, len); 530 } 531 } 532 533 void 534 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len) 535 { 536 dmu_tx_hold_t *txh; 537 538 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len); 539 if (txh != NULL) { 540 dmu_tx_count_dnode(txh); 541 dmu_tx_count_free(txh, off, len); 542 } 543 } 544 545 static void 546 dmu_tx_count_clone(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 547 { 548 549 /* 550 * Reuse dmu_tx_count_free(), it does exactly what we need for clone. 551 */ 552 dmu_tx_count_free(txh, off, len); 553 } 554 555 void 556 dmu_tx_hold_clone_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len) 557 { 558 dmu_tx_hold_t *txh; 559 560 ASSERT0(tx->tx_txg); 561 ASSERT(len == 0 || UINT64_MAX - off >= len - 1); 562 563 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_CLONE, off, len); 564 if (txh != NULL) { 565 dmu_tx_count_dnode(txh); 566 dmu_tx_count_clone(txh, off, len); 567 } 568 } 569 570 static void 571 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name) 572 { 573 dmu_tx_t *tx = txh->txh_tx; 574 dnode_t *dn = txh->txh_dnode; 575 int err; 576 extern int zap_micro_max_size; 577 578 ASSERT(tx->tx_txg == 0); 579 580 dmu_tx_count_dnode(txh); 581 582 /* 583 * Modifying a almost-full microzap is around the worst case (128KB) 584 * 585 * If it is a fat zap, the worst case would be 7*16KB=112KB: 586 * - 3 blocks overwritten: target leaf, ptrtbl block, header block 587 * - 4 new blocks written if adding: 588 * - 2 blocks for possibly split leaves, 589 * - 2 grown ptrtbl blocks 590 */ 591 (void) zfs_refcount_add_many(&txh->txh_space_towrite, 592 zap_micro_max_size, FTAG); 593 594 if (dn == NULL) 595 return; 596 597 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP); 598 599 if (dn->dn_maxblkid == 0 || name == NULL) { 600 /* 601 * This is a microzap (only one block), or we don't know 602 * the name. Check the first block for i/o errors. 603 */ 604 err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 605 if (err != 0) { 606 tx->tx_err = err; 607 } 608 } else { 609 /* 610 * Access the name so that we'll check for i/o errors to 611 * the leaf blocks, etc. We ignore ENOENT, as this name 612 * may not yet exist. 613 */ 614 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL); 615 if (err == EIO || err == ECKSUM || err == ENXIO) { 616 tx->tx_err = err; 617 } 618 } 619 } 620 621 void 622 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name) 623 { 624 dmu_tx_hold_t *txh; 625 626 ASSERT0(tx->tx_txg); 627 628 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 629 object, THT_ZAP, add, (uintptr_t)name); 630 if (txh != NULL) 631 dmu_tx_hold_zap_impl(txh, name); 632 } 633 634 void 635 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name) 636 { 637 dmu_tx_hold_t *txh; 638 639 ASSERT0(tx->tx_txg); 640 ASSERT(dn != NULL); 641 642 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name); 643 if (txh != NULL) 644 dmu_tx_hold_zap_impl(txh, name); 645 } 646 647 void 648 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object) 649 { 650 dmu_tx_hold_t *txh; 651 652 ASSERT(tx->tx_txg == 0); 653 654 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 655 object, THT_BONUS, 0, 0); 656 if (txh) 657 dmu_tx_count_dnode(txh); 658 } 659 660 void 661 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn) 662 { 663 dmu_tx_hold_t *txh; 664 665 ASSERT0(tx->tx_txg); 666 667 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0); 668 if (txh) 669 dmu_tx_count_dnode(txh); 670 } 671 672 void 673 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space) 674 { 675 dmu_tx_hold_t *txh; 676 677 ASSERT(tx->tx_txg == 0); 678 679 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 680 DMU_NEW_OBJECT, THT_SPACE, space, 0); 681 if (txh) { 682 (void) zfs_refcount_add_many( 683 &txh->txh_space_towrite, space, FTAG); 684 } 685 } 686 687 #ifdef ZFS_DEBUG 688 void 689 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db) 690 { 691 boolean_t match_object = B_FALSE; 692 boolean_t match_offset = B_FALSE; 693 694 DB_DNODE_ENTER(db); 695 dnode_t *dn = DB_DNODE(db); 696 ASSERT(tx->tx_txg != 0); 697 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset); 698 ASSERT3U(dn->dn_object, ==, db->db.db_object); 699 700 if (tx->tx_anyobj) { 701 DB_DNODE_EXIT(db); 702 return; 703 } 704 705 /* XXX No checking on the meta dnode for now */ 706 if (db->db.db_object == DMU_META_DNODE_OBJECT) { 707 DB_DNODE_EXIT(db); 708 return; 709 } 710 711 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL; 712 txh = list_next(&tx->tx_holds, txh)) { 713 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 714 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT) 715 match_object = TRUE; 716 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) { 717 int datablkshift = dn->dn_datablkshift ? 718 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT; 719 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 720 int shift = datablkshift + epbs * db->db_level; 721 uint64_t beginblk = shift >= 64 ? 0 : 722 (txh->txh_arg1 >> shift); 723 uint64_t endblk = shift >= 64 ? 0 : 724 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift); 725 uint64_t blkid = db->db_blkid; 726 727 /* XXX txh_arg2 better not be zero... */ 728 729 dprintf("found txh type %x beginblk=%llx endblk=%llx\n", 730 txh->txh_type, (u_longlong_t)beginblk, 731 (u_longlong_t)endblk); 732 733 switch (txh->txh_type) { 734 case THT_WRITE: 735 if (blkid >= beginblk && blkid <= endblk) 736 match_offset = TRUE; 737 /* 738 * We will let this hold work for the bonus 739 * or spill buffer so that we don't need to 740 * hold it when creating a new object. 741 */ 742 if (blkid == DMU_BONUS_BLKID || 743 blkid == DMU_SPILL_BLKID) 744 match_offset = TRUE; 745 /* 746 * They might have to increase nlevels, 747 * thus dirtying the new TLIBs. Or the 748 * might have to change the block size, 749 * thus dirying the new lvl=0 blk=0. 750 */ 751 if (blkid == 0) 752 match_offset = TRUE; 753 break; 754 case THT_APPEND: 755 if (blkid >= beginblk && (blkid <= endblk || 756 txh->txh_arg2 == DMU_OBJECT_END)) 757 match_offset = TRUE; 758 759 /* 760 * THT_WRITE used for bonus and spill blocks. 761 */ 762 ASSERT(blkid != DMU_BONUS_BLKID && 763 blkid != DMU_SPILL_BLKID); 764 765 /* 766 * They might have to increase nlevels, 767 * thus dirtying the new TLIBs. Or the 768 * might have to change the block size, 769 * thus dirying the new lvl=0 blk=0. 770 */ 771 if (blkid == 0) 772 match_offset = TRUE; 773 break; 774 case THT_FREE: 775 /* 776 * We will dirty all the level 1 blocks in 777 * the free range and perhaps the first and 778 * last level 0 block. 779 */ 780 if (blkid >= beginblk && (blkid <= endblk || 781 txh->txh_arg2 == DMU_OBJECT_END)) 782 match_offset = TRUE; 783 break; 784 case THT_SPILL: 785 if (blkid == DMU_SPILL_BLKID) 786 match_offset = TRUE; 787 break; 788 case THT_BONUS: 789 if (blkid == DMU_BONUS_BLKID) 790 match_offset = TRUE; 791 break; 792 case THT_ZAP: 793 match_offset = TRUE; 794 break; 795 case THT_NEWOBJECT: 796 match_object = TRUE; 797 break; 798 case THT_CLONE: 799 if (blkid >= beginblk && blkid <= endblk) 800 match_offset = TRUE; 801 break; 802 default: 803 cmn_err(CE_PANIC, "bad txh_type %d", 804 txh->txh_type); 805 } 806 } 807 if (match_object && match_offset) { 808 DB_DNODE_EXIT(db); 809 return; 810 } 811 } 812 DB_DNODE_EXIT(db); 813 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n", 814 (u_longlong_t)db->db.db_object, db->db_level, 815 (u_longlong_t)db->db_blkid); 816 } 817 #endif 818 819 /* 820 * If we can't do 10 iops, something is wrong. Let us go ahead 821 * and hit zfs_dirty_data_max. 822 */ 823 static const hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */ 824 825 /* 826 * We delay transactions when we've determined that the backend storage 827 * isn't able to accommodate the rate of incoming writes. 828 * 829 * If there is already a transaction waiting, we delay relative to when 830 * that transaction finishes waiting. This way the calculated min_time 831 * is independent of the number of threads concurrently executing 832 * transactions. 833 * 834 * If we are the only waiter, wait relative to when the transaction 835 * started, rather than the current time. This credits the transaction for 836 * "time already served", e.g. reading indirect blocks. 837 * 838 * The minimum time for a transaction to take is calculated as: 839 * min_time = scale * (dirty - min) / (max - dirty) 840 * min_time is then capped at zfs_delay_max_ns. 841 * 842 * The delay has two degrees of freedom that can be adjusted via tunables. 843 * The percentage of dirty data at which we start to delay is defined by 844 * zfs_delay_min_dirty_percent. This should typically be at or above 845 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to 846 * delay after writing at full speed has failed to keep up with the incoming 847 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly 848 * speaking, this variable determines the amount of delay at the midpoint of 849 * the curve. 850 * 851 * delay 852 * 10ms +-------------------------------------------------------------*+ 853 * | *| 854 * 9ms + *+ 855 * | *| 856 * 8ms + *+ 857 * | * | 858 * 7ms + * + 859 * | * | 860 * 6ms + * + 861 * | * | 862 * 5ms + * + 863 * | * | 864 * 4ms + * + 865 * | * | 866 * 3ms + * + 867 * | * | 868 * 2ms + (midpoint) * + 869 * | | ** | 870 * 1ms + v *** + 871 * | zfs_delay_scale ----------> ******** | 872 * 0 +-------------------------------------*********----------------+ 873 * 0% <- zfs_dirty_data_max -> 100% 874 * 875 * Note that since the delay is added to the outstanding time remaining on the 876 * most recent transaction, the delay is effectively the inverse of IOPS. 877 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve 878 * was chosen such that small changes in the amount of accumulated dirty data 879 * in the first 3/4 of the curve yield relatively small differences in the 880 * amount of delay. 881 * 882 * The effects can be easier to understand when the amount of delay is 883 * represented on a log scale: 884 * 885 * delay 886 * 100ms +-------------------------------------------------------------++ 887 * + + 888 * | | 889 * + *+ 890 * 10ms + *+ 891 * + ** + 892 * | (midpoint) ** | 893 * + | ** + 894 * 1ms + v **** + 895 * + zfs_delay_scale ----------> ***** + 896 * | **** | 897 * + **** + 898 * 100us + ** + 899 * + * + 900 * | * | 901 * + * + 902 * 10us + * + 903 * + + 904 * | | 905 * + + 906 * +--------------------------------------------------------------+ 907 * 0% <- zfs_dirty_data_max -> 100% 908 * 909 * Note here that only as the amount of dirty data approaches its limit does 910 * the delay start to increase rapidly. The goal of a properly tuned system 911 * should be to keep the amount of dirty data out of that range by first 912 * ensuring that the appropriate limits are set for the I/O scheduler to reach 913 * optimal throughput on the backend storage, and then by changing the value 914 * of zfs_delay_scale to increase the steepness of the curve. 915 */ 916 static void 917 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty) 918 { 919 dsl_pool_t *dp = tx->tx_pool; 920 uint64_t delay_min_bytes, wrlog; 921 hrtime_t wakeup, tx_time = 0, now; 922 923 /* Calculate minimum transaction time for the dirty data amount. */ 924 delay_min_bytes = 925 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 926 if (dirty > delay_min_bytes) { 927 /* 928 * The caller has already waited until we are under the max. 929 * We make them pass us the amount of dirty data so we don't 930 * have to handle the case of it being >= the max, which 931 * could cause a divide-by-zero if it's == the max. 932 */ 933 ASSERT3U(dirty, <, zfs_dirty_data_max); 934 935 tx_time = zfs_delay_scale * (dirty - delay_min_bytes) / 936 (zfs_dirty_data_max - dirty); 937 } 938 939 /* Calculate minimum transaction time for the TX_WRITE log size. */ 940 wrlog = aggsum_upper_bound(&dp->dp_wrlog_total); 941 delay_min_bytes = 942 zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100; 943 if (wrlog >= zfs_wrlog_data_max) { 944 tx_time = zfs_delay_max_ns; 945 } else if (wrlog > delay_min_bytes) { 946 tx_time = MAX(zfs_delay_scale * (wrlog - delay_min_bytes) / 947 (zfs_wrlog_data_max - wrlog), tx_time); 948 } 949 950 if (tx_time == 0) 951 return; 952 953 tx_time = MIN(tx_time, zfs_delay_max_ns); 954 now = gethrtime(); 955 if (now > tx->tx_start + tx_time) 956 return; 957 958 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty, 959 uint64_t, tx_time); 960 961 mutex_enter(&dp->dp_lock); 962 wakeup = MAX(tx->tx_start + tx_time, dp->dp_last_wakeup + tx_time); 963 dp->dp_last_wakeup = wakeup; 964 mutex_exit(&dp->dp_lock); 965 966 zfs_sleep_until(wakeup); 967 } 968 969 /* 970 * This routine attempts to assign the transaction to a transaction group. 971 * To do so, we must determine if there is sufficient free space on disk. 972 * 973 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree() 974 * on it), then it is assumed that there is sufficient free space, 975 * unless there's insufficient slop space in the pool (see the comment 976 * above spa_slop_shift in spa_misc.c). 977 * 978 * If it is not a "netfree" transaction, then if the data already on disk 979 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or 980 * ENOSPC. Otherwise, if the current rough estimate of pending changes, 981 * plus the rough estimate of this transaction's changes, may exceed the 982 * allowed usage, then this will fail with ERESTART, which will cause the 983 * caller to wait for the pending changes to be written to disk (by waiting 984 * for the next TXG to open), and then check the space usage again. 985 * 986 * The rough estimate of pending changes is comprised of the sum of: 987 * 988 * - this transaction's holds' txh_space_towrite 989 * 990 * - dd_tempreserved[], which is the sum of in-flight transactions' 991 * holds' txh_space_towrite (i.e. those transactions that have called 992 * dmu_tx_assign() but not yet called dmu_tx_commit()). 993 * 994 * - dd_space_towrite[], which is the amount of dirtied dbufs. 995 * 996 * Note that all of these values are inflated by spa_get_worst_case_asize(), 997 * which means that we may get ERESTART well before we are actually in danger 998 * of running out of space, but this also mitigates any small inaccuracies 999 * in the rough estimate (e.g. txh_space_towrite doesn't take into account 1000 * indirect blocks, and dd_space_towrite[] doesn't take into account changes 1001 * to the MOS). 1002 * 1003 * Note that due to this algorithm, it is possible to exceed the allowed 1004 * usage by one transaction. Also, as we approach the allowed usage, 1005 * we will allow a very limited amount of changes into each TXG, thus 1006 * decreasing performance. 1007 */ 1008 static int 1009 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how) 1010 { 1011 spa_t *spa = tx->tx_pool->dp_spa; 1012 1013 ASSERT0(tx->tx_txg); 1014 1015 if (tx->tx_err) { 1016 DMU_TX_STAT_BUMP(dmu_tx_error); 1017 return (tx->tx_err); 1018 } 1019 1020 if (spa_suspended(spa)) { 1021 DMU_TX_STAT_BUMP(dmu_tx_suspended); 1022 1023 /* 1024 * If the user has indicated a blocking failure mode 1025 * then return ERESTART which will block in dmu_tx_wait(). 1026 * Otherwise, return EIO so that an error can get 1027 * propagated back to the VOP calls. 1028 * 1029 * Note that we always honor the txg_how flag regardless 1030 * of the failuremode setting. 1031 */ 1032 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE && 1033 !(txg_how & TXG_WAIT)) 1034 return (SET_ERROR(EIO)); 1035 1036 return (SET_ERROR(ERESTART)); 1037 } 1038 1039 if (!tx->tx_dirty_delayed && 1040 dsl_pool_need_wrlog_delay(tx->tx_pool)) { 1041 tx->tx_wait_dirty = B_TRUE; 1042 DMU_TX_STAT_BUMP(dmu_tx_wrlog_delay); 1043 return (SET_ERROR(ERESTART)); 1044 } 1045 1046 if (!tx->tx_dirty_delayed && 1047 dsl_pool_need_dirty_delay(tx->tx_pool)) { 1048 tx->tx_wait_dirty = B_TRUE; 1049 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay); 1050 return (SET_ERROR(ERESTART)); 1051 } 1052 1053 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh); 1054 tx->tx_needassign_txh = NULL; 1055 1056 /* 1057 * NB: No error returns are allowed after txg_hold_open, but 1058 * before processing the dnode holds, due to the 1059 * dmu_tx_unassign() logic. 1060 */ 1061 1062 uint64_t towrite = 0; 1063 uint64_t tohold = 0; 1064 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL; 1065 txh = list_next(&tx->tx_holds, txh)) { 1066 dnode_t *dn = txh->txh_dnode; 1067 if (dn != NULL) { 1068 /* 1069 * This thread can't hold the dn_struct_rwlock 1070 * while assigning the tx, because this can lead to 1071 * deadlock. Specifically, if this dnode is already 1072 * assigned to an earlier txg, this thread may need 1073 * to wait for that txg to sync (the ERESTART case 1074 * below). The other thread that has assigned this 1075 * dnode to an earlier txg prevents this txg from 1076 * syncing until its tx can complete (calling 1077 * dmu_tx_commit()), but it may need to acquire the 1078 * dn_struct_rwlock to do so (e.g. via 1079 * dmu_buf_hold*()). 1080 * 1081 * Note that this thread can't hold the lock for 1082 * read either, but the rwlock doesn't record 1083 * enough information to make that assertion. 1084 */ 1085 ASSERT(!RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1086 1087 mutex_enter(&dn->dn_mtx); 1088 if (dn->dn_assigned_txg == tx->tx_txg - 1) { 1089 mutex_exit(&dn->dn_mtx); 1090 tx->tx_needassign_txh = txh; 1091 DMU_TX_STAT_BUMP(dmu_tx_group); 1092 return (SET_ERROR(ERESTART)); 1093 } 1094 if (dn->dn_assigned_txg == 0) 1095 dn->dn_assigned_txg = tx->tx_txg; 1096 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1097 (void) zfs_refcount_add(&dn->dn_tx_holds, tx); 1098 mutex_exit(&dn->dn_mtx); 1099 } 1100 towrite += zfs_refcount_count(&txh->txh_space_towrite); 1101 tohold += zfs_refcount_count(&txh->txh_memory_tohold); 1102 } 1103 1104 /* needed allocation: worst-case estimate of write space */ 1105 uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite); 1106 /* calculate memory footprint estimate */ 1107 uint64_t memory = towrite + tohold; 1108 1109 if (tx->tx_dir != NULL && asize != 0) { 1110 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory, 1111 asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx); 1112 if (err != 0) 1113 return (err); 1114 } 1115 1116 DMU_TX_STAT_BUMP(dmu_tx_assigned); 1117 1118 return (0); 1119 } 1120 1121 static void 1122 dmu_tx_unassign(dmu_tx_t *tx) 1123 { 1124 if (tx->tx_txg == 0) 1125 return; 1126 1127 txg_rele_to_quiesce(&tx->tx_txgh); 1128 1129 /* 1130 * Walk the transaction's hold list, removing the hold on the 1131 * associated dnode, and notifying waiters if the refcount drops to 0. 1132 */ 1133 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); 1134 txh && txh != tx->tx_needassign_txh; 1135 txh = list_next(&tx->tx_holds, txh)) { 1136 dnode_t *dn = txh->txh_dnode; 1137 1138 if (dn == NULL) 1139 continue; 1140 mutex_enter(&dn->dn_mtx); 1141 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1142 1143 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) { 1144 dn->dn_assigned_txg = 0; 1145 cv_broadcast(&dn->dn_notxholds); 1146 } 1147 mutex_exit(&dn->dn_mtx); 1148 } 1149 1150 txg_rele_to_sync(&tx->tx_txgh); 1151 1152 tx->tx_lasttried_txg = tx->tx_txg; 1153 tx->tx_txg = 0; 1154 } 1155 1156 /* 1157 * Assign tx to a transaction group; txg_how is a bitmask: 1158 * 1159 * If TXG_WAIT is set and the currently open txg is full, this function 1160 * will wait until there's a new txg. This should be used when no locks 1161 * are being held. With this bit set, this function will only fail if 1162 * we're truly out of space (or over quota). 1163 * 1164 * If TXG_WAIT is *not* set and we can't assign into the currently open 1165 * txg without blocking, this function will return immediately with 1166 * ERESTART. This should be used whenever locks are being held. On an 1167 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(), 1168 * and try again. 1169 * 1170 * If TXG_NOTHROTTLE is set, this indicates that this tx should not be 1171 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for 1172 * details on the throttle). This is used by the VFS operations, after 1173 * they have already called dmu_tx_wait() (though most likely on a 1174 * different tx). 1175 * 1176 * It is guaranteed that subsequent successful calls to dmu_tx_assign() 1177 * will assign the tx to monotonically increasing txgs. Of course this is 1178 * not strong monotonicity, because the same txg can be returned multiple 1179 * times in a row. This guarantee holds both for subsequent calls from 1180 * one thread and for multiple threads. For example, it is impossible to 1181 * observe the following sequence of events: 1182 * 1183 * Thread 1 Thread 2 1184 * 1185 * dmu_tx_assign(T1, ...) 1186 * 1 <- dmu_tx_get_txg(T1) 1187 * dmu_tx_assign(T2, ...) 1188 * 2 <- dmu_tx_get_txg(T2) 1189 * dmu_tx_assign(T3, ...) 1190 * 1 <- dmu_tx_get_txg(T3) 1191 */ 1192 int 1193 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how) 1194 { 1195 int err; 1196 1197 ASSERT(tx->tx_txg == 0); 1198 ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE)); 1199 ASSERT(!dsl_pool_sync_context(tx->tx_pool)); 1200 1201 /* If we might wait, we must not hold the config lock. */ 1202 IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool)); 1203 1204 if ((txg_how & TXG_NOTHROTTLE)) 1205 tx->tx_dirty_delayed = B_TRUE; 1206 1207 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) { 1208 dmu_tx_unassign(tx); 1209 1210 if (err != ERESTART || !(txg_how & TXG_WAIT)) 1211 return (err); 1212 1213 dmu_tx_wait(tx); 1214 } 1215 1216 txg_rele_to_quiesce(&tx->tx_txgh); 1217 1218 return (0); 1219 } 1220 1221 void 1222 dmu_tx_wait(dmu_tx_t *tx) 1223 { 1224 spa_t *spa = tx->tx_pool->dp_spa; 1225 dsl_pool_t *dp = tx->tx_pool; 1226 hrtime_t before; 1227 1228 ASSERT(tx->tx_txg == 0); 1229 ASSERT(!dsl_pool_config_held(tx->tx_pool)); 1230 1231 before = gethrtime(); 1232 1233 if (tx->tx_wait_dirty) { 1234 uint64_t dirty; 1235 1236 /* 1237 * dmu_tx_try_assign() has determined that we need to wait 1238 * because we've consumed much or all of the dirty buffer 1239 * space. 1240 */ 1241 mutex_enter(&dp->dp_lock); 1242 if (dp->dp_dirty_total >= zfs_dirty_data_max) 1243 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max); 1244 while (dp->dp_dirty_total >= zfs_dirty_data_max) 1245 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock); 1246 dirty = dp->dp_dirty_total; 1247 mutex_exit(&dp->dp_lock); 1248 1249 dmu_tx_delay(tx, dirty); 1250 1251 tx->tx_wait_dirty = B_FALSE; 1252 1253 /* 1254 * Note: setting tx_dirty_delayed only has effect if the 1255 * caller used TX_WAIT. Otherwise they are going to 1256 * destroy this tx and try again. The common case, 1257 * zfs_write(), uses TX_WAIT. 1258 */ 1259 tx->tx_dirty_delayed = B_TRUE; 1260 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) { 1261 /* 1262 * If the pool is suspended we need to wait until it 1263 * is resumed. Note that it's possible that the pool 1264 * has become active after this thread has tried to 1265 * obtain a tx. If that's the case then tx_lasttried_txg 1266 * would not have been set. 1267 */ 1268 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1); 1269 } else if (tx->tx_needassign_txh) { 1270 dnode_t *dn = tx->tx_needassign_txh->txh_dnode; 1271 1272 mutex_enter(&dn->dn_mtx); 1273 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1) 1274 cv_wait(&dn->dn_notxholds, &dn->dn_mtx); 1275 mutex_exit(&dn->dn_mtx); 1276 tx->tx_needassign_txh = NULL; 1277 } else { 1278 /* 1279 * If we have a lot of dirty data just wait until we sync 1280 * out a TXG at which point we'll hopefully have synced 1281 * a portion of the changes. 1282 */ 1283 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1); 1284 } 1285 1286 spa_tx_assign_add_nsecs(spa, gethrtime() - before); 1287 } 1288 1289 static void 1290 dmu_tx_destroy(dmu_tx_t *tx) 1291 { 1292 dmu_tx_hold_t *txh; 1293 1294 while ((txh = list_head(&tx->tx_holds)) != NULL) { 1295 dnode_t *dn = txh->txh_dnode; 1296 1297 list_remove(&tx->tx_holds, txh); 1298 zfs_refcount_destroy_many(&txh->txh_space_towrite, 1299 zfs_refcount_count(&txh->txh_space_towrite)); 1300 zfs_refcount_destroy_many(&txh->txh_memory_tohold, 1301 zfs_refcount_count(&txh->txh_memory_tohold)); 1302 kmem_free(txh, sizeof (dmu_tx_hold_t)); 1303 if (dn != NULL) 1304 dnode_rele(dn, tx); 1305 } 1306 1307 list_destroy(&tx->tx_callbacks); 1308 list_destroy(&tx->tx_holds); 1309 kmem_free(tx, sizeof (dmu_tx_t)); 1310 } 1311 1312 void 1313 dmu_tx_commit(dmu_tx_t *tx) 1314 { 1315 ASSERT(tx->tx_txg != 0); 1316 1317 /* 1318 * Go through the transaction's hold list and remove holds on 1319 * associated dnodes, notifying waiters if no holds remain. 1320 */ 1321 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL; 1322 txh = list_next(&tx->tx_holds, txh)) { 1323 dnode_t *dn = txh->txh_dnode; 1324 1325 if (dn == NULL) 1326 continue; 1327 1328 mutex_enter(&dn->dn_mtx); 1329 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1330 1331 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) { 1332 dn->dn_assigned_txg = 0; 1333 cv_broadcast(&dn->dn_notxholds); 1334 } 1335 mutex_exit(&dn->dn_mtx); 1336 } 1337 1338 if (tx->tx_tempreserve_cookie) 1339 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx); 1340 1341 if (!list_is_empty(&tx->tx_callbacks)) 1342 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks); 1343 1344 if (tx->tx_anyobj == FALSE) 1345 txg_rele_to_sync(&tx->tx_txgh); 1346 1347 dmu_tx_destroy(tx); 1348 } 1349 1350 void 1351 dmu_tx_abort(dmu_tx_t *tx) 1352 { 1353 ASSERT(tx->tx_txg == 0); 1354 1355 /* 1356 * Call any registered callbacks with an error code. 1357 */ 1358 if (!list_is_empty(&tx->tx_callbacks)) 1359 dmu_tx_do_callbacks(&tx->tx_callbacks, SET_ERROR(ECANCELED)); 1360 1361 dmu_tx_destroy(tx); 1362 } 1363 1364 uint64_t 1365 dmu_tx_get_txg(dmu_tx_t *tx) 1366 { 1367 ASSERT(tx->tx_txg != 0); 1368 return (tx->tx_txg); 1369 } 1370 1371 dsl_pool_t * 1372 dmu_tx_pool(dmu_tx_t *tx) 1373 { 1374 ASSERT(tx->tx_pool != NULL); 1375 return (tx->tx_pool); 1376 } 1377 1378 void 1379 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data) 1380 { 1381 dmu_tx_callback_t *dcb; 1382 1383 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP); 1384 1385 dcb->dcb_func = func; 1386 dcb->dcb_data = data; 1387 1388 list_insert_tail(&tx->tx_callbacks, dcb); 1389 } 1390 1391 /* 1392 * Call all the commit callbacks on a list, with a given error code. 1393 */ 1394 void 1395 dmu_tx_do_callbacks(list_t *cb_list, int error) 1396 { 1397 dmu_tx_callback_t *dcb; 1398 1399 while ((dcb = list_remove_tail(cb_list)) != NULL) { 1400 dcb->dcb_func(dcb->dcb_data, error); 1401 kmem_free(dcb, sizeof (dmu_tx_callback_t)); 1402 } 1403 } 1404 1405 /* 1406 * Interface to hold a bunch of attributes. 1407 * used for creating new files. 1408 * attrsize is the total size of all attributes 1409 * to be added during object creation 1410 * 1411 * For updating/adding a single attribute dmu_tx_hold_sa() should be used. 1412 */ 1413 1414 /* 1415 * hold necessary attribute name for attribute registration. 1416 * should be a very rare case where this is needed. If it does 1417 * happen it would only happen on the first write to the file system. 1418 */ 1419 static void 1420 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx) 1421 { 1422 if (!sa->sa_need_attr_registration) 1423 return; 1424 1425 for (int i = 0; i != sa->sa_num_attrs; i++) { 1426 if (!sa->sa_attr_table[i].sa_registered) { 1427 if (sa->sa_reg_attr_obj) 1428 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj, 1429 B_TRUE, sa->sa_attr_table[i].sa_name); 1430 else 1431 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, 1432 B_TRUE, sa->sa_attr_table[i].sa_name); 1433 } 1434 } 1435 } 1436 1437 void 1438 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object) 1439 { 1440 dmu_tx_hold_t *txh; 1441 1442 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object, 1443 THT_SPILL, 0, 0); 1444 if (txh != NULL) 1445 (void) zfs_refcount_add_many(&txh->txh_space_towrite, 1446 SPA_OLD_MAXBLOCKSIZE, FTAG); 1447 } 1448 1449 void 1450 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize) 1451 { 1452 sa_os_t *sa = tx->tx_objset->os_sa; 1453 1454 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT); 1455 1456 if (tx->tx_objset->os_sa->sa_master_obj == 0) 1457 return; 1458 1459 if (tx->tx_objset->os_sa->sa_layout_attr_obj) { 1460 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL); 1461 } else { 1462 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS); 1463 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY); 1464 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1465 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1466 } 1467 1468 dmu_tx_sa_registration_hold(sa, tx); 1469 1470 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill) 1471 return; 1472 1473 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT, 1474 THT_SPILL, 0, 0); 1475 } 1476 1477 /* 1478 * Hold SA attribute 1479 * 1480 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size) 1481 * 1482 * variable_size is the total size of all variable sized attributes 1483 * passed to this function. It is not the total size of all 1484 * variable size attributes that *may* exist on this object. 1485 */ 1486 void 1487 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow) 1488 { 1489 uint64_t object; 1490 sa_os_t *sa = tx->tx_objset->os_sa; 1491 1492 ASSERT(hdl != NULL); 1493 1494 object = sa_handle_object(hdl); 1495 1496 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus; 1497 DB_DNODE_ENTER(db); 1498 dmu_tx_hold_bonus_by_dnode(tx, DB_DNODE(db)); 1499 DB_DNODE_EXIT(db); 1500 1501 if (tx->tx_objset->os_sa->sa_master_obj == 0) 1502 return; 1503 1504 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 || 1505 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) { 1506 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS); 1507 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY); 1508 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1509 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1510 } 1511 1512 dmu_tx_sa_registration_hold(sa, tx); 1513 1514 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj) 1515 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL); 1516 1517 if (sa->sa_force_spill || may_grow || hdl->sa_spill) { 1518 ASSERT(tx->tx_txg == 0); 1519 dmu_tx_hold_spill(tx, object); 1520 } else { 1521 dnode_t *dn; 1522 1523 DB_DNODE_ENTER(db); 1524 dn = DB_DNODE(db); 1525 if (dn->dn_have_spill) { 1526 ASSERT(tx->tx_txg == 0); 1527 dmu_tx_hold_spill(tx, object); 1528 } 1529 DB_DNODE_EXIT(db); 1530 } 1531 } 1532 1533 void 1534 dmu_tx_init(void) 1535 { 1536 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc", 1537 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t), 1538 KSTAT_FLAG_VIRTUAL); 1539 1540 if (dmu_tx_ksp != NULL) { 1541 dmu_tx_ksp->ks_data = &dmu_tx_stats; 1542 kstat_install(dmu_tx_ksp); 1543 } 1544 } 1545 1546 void 1547 dmu_tx_fini(void) 1548 { 1549 if (dmu_tx_ksp != NULL) { 1550 kstat_delete(dmu_tx_ksp); 1551 dmu_tx_ksp = NULL; 1552 } 1553 } 1554 1555 #if defined(_KERNEL) 1556 EXPORT_SYMBOL(dmu_tx_create); 1557 EXPORT_SYMBOL(dmu_tx_hold_write); 1558 EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode); 1559 EXPORT_SYMBOL(dmu_tx_hold_append); 1560 EXPORT_SYMBOL(dmu_tx_hold_append_by_dnode); 1561 EXPORT_SYMBOL(dmu_tx_hold_free); 1562 EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode); 1563 EXPORT_SYMBOL(dmu_tx_hold_zap); 1564 EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode); 1565 EXPORT_SYMBOL(dmu_tx_hold_bonus); 1566 EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode); 1567 EXPORT_SYMBOL(dmu_tx_abort); 1568 EXPORT_SYMBOL(dmu_tx_assign); 1569 EXPORT_SYMBOL(dmu_tx_wait); 1570 EXPORT_SYMBOL(dmu_tx_commit); 1571 EXPORT_SYMBOL(dmu_tx_mark_netfree); 1572 EXPORT_SYMBOL(dmu_tx_get_txg); 1573 EXPORT_SYMBOL(dmu_tx_callback_register); 1574 EXPORT_SYMBOL(dmu_tx_do_callbacks); 1575 EXPORT_SYMBOL(dmu_tx_hold_spill); 1576 EXPORT_SYMBOL(dmu_tx_hold_sa_create); 1577 EXPORT_SYMBOL(dmu_tx_hold_sa); 1578 #endif 1579