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 http://www.opensolaris.org/os/licensing. 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 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 /* 26 * Copyright (c) 2012, 2018 by Delphix. All rights reserved. 27 */ 28 29 #include <sys/zfs_context.h> 30 #include <sys/spa.h> 31 #include <sys/dmu.h> 32 #include <sys/dmu_tx.h> 33 #include <sys/dnode.h> 34 #include <sys/dsl_pool.h> 35 #include <sys/zio.h> 36 #include <sys/space_map.h> 37 #include <sys/refcount.h> 38 #include <sys/zfeature.h> 39 40 /* 41 * Note on space map block size: 42 * 43 * The data for a given space map can be kept on blocks of any size. 44 * Larger blocks entail fewer I/O operations, but they also cause the 45 * DMU to keep more data in-core, and also to waste more I/O bandwidth 46 * when only a few blocks have changed since the last transaction group. 47 */ 48 49 /* 50 * Enabled whenever we want to stress test the use of double-word 51 * space map entries. 52 */ 53 boolean_t zfs_force_some_double_word_sm_entries = B_FALSE; 54 55 /* 56 * Override the default indirect block size of 128K, instead using 16K for 57 * spacemaps (2^14 bytes). This dramatically reduces write inflation since 58 * appending to a spacemap typically has to write one data block (4KB) and one 59 * or two indirect blocks (16K-32K, rather than 128K). 60 */ 61 int space_map_ibs = 14; 62 63 boolean_t 64 sm_entry_is_debug(uint64_t e) 65 { 66 return (SM_PREFIX_DECODE(e) == SM_DEBUG_PREFIX); 67 } 68 69 boolean_t 70 sm_entry_is_single_word(uint64_t e) 71 { 72 uint8_t prefix = SM_PREFIX_DECODE(e); 73 return (prefix != SM_DEBUG_PREFIX && prefix != SM2_PREFIX); 74 } 75 76 boolean_t 77 sm_entry_is_double_word(uint64_t e) 78 { 79 return (SM_PREFIX_DECODE(e) == SM2_PREFIX); 80 } 81 82 /* 83 * Iterate through the space map, invoking the callback on each (non-debug) 84 * space map entry. Stop after reading 'end' bytes of the space map. 85 */ 86 int 87 space_map_iterate(space_map_t *sm, uint64_t end, sm_cb_t callback, void *arg) 88 { 89 uint64_t blksz = sm->sm_blksz; 90 91 ASSERT3U(blksz, !=, 0); 92 ASSERT3U(end, <=, space_map_length(sm)); 93 ASSERT0(P2PHASE(end, sizeof (uint64_t))); 94 95 dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, end, 96 ZIO_PRIORITY_SYNC_READ); 97 98 int error = 0; 99 for (uint64_t block_base = 0; block_base < end && error == 0; 100 block_base += blksz) { 101 dmu_buf_t *db; 102 error = dmu_buf_hold(sm->sm_os, space_map_object(sm), 103 block_base, FTAG, &db, DMU_READ_PREFETCH); 104 if (error != 0) 105 return (error); 106 107 uint64_t *block_start = db->db_data; 108 uint64_t block_length = MIN(end - block_base, blksz); 109 uint64_t *block_end = block_start + 110 (block_length / sizeof (uint64_t)); 111 112 VERIFY0(P2PHASE(block_length, sizeof (uint64_t))); 113 VERIFY3U(block_length, !=, 0); 114 ASSERT3U(blksz, ==, db->db_size); 115 116 for (uint64_t *block_cursor = block_start; 117 block_cursor < block_end && error == 0; block_cursor++) { 118 uint64_t e = *block_cursor; 119 120 if (sm_entry_is_debug(e)) /* Skip debug entries */ 121 continue; 122 123 uint64_t raw_offset, raw_run, vdev_id; 124 maptype_t type; 125 if (sm_entry_is_single_word(e)) { 126 type = SM_TYPE_DECODE(e); 127 vdev_id = SM_NO_VDEVID; 128 raw_offset = SM_OFFSET_DECODE(e); 129 raw_run = SM_RUN_DECODE(e); 130 } else { 131 /* it is a two-word entry */ 132 ASSERT(sm_entry_is_double_word(e)); 133 raw_run = SM2_RUN_DECODE(e); 134 vdev_id = SM2_VDEV_DECODE(e); 135 136 /* move on to the second word */ 137 block_cursor++; 138 e = *block_cursor; 139 VERIFY3P(block_cursor, <=, block_end); 140 141 type = SM2_TYPE_DECODE(e); 142 raw_offset = SM2_OFFSET_DECODE(e); 143 } 144 145 uint64_t entry_offset = (raw_offset << sm->sm_shift) + 146 sm->sm_start; 147 uint64_t entry_run = raw_run << sm->sm_shift; 148 149 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift)); 150 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift)); 151 ASSERT3U(entry_offset, >=, sm->sm_start); 152 ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size); 153 ASSERT3U(entry_run, <=, sm->sm_size); 154 ASSERT3U(entry_offset + entry_run, <=, 155 sm->sm_start + sm->sm_size); 156 157 space_map_entry_t sme = { 158 .sme_type = type, 159 .sme_vdev = vdev_id, 160 .sme_offset = entry_offset, 161 .sme_run = entry_run 162 }; 163 error = callback(&sme, arg); 164 } 165 dmu_buf_rele(db, FTAG); 166 } 167 return (error); 168 } 169 170 /* 171 * Reads the entries from the last block of the space map into 172 * buf in reverse order. Populates nwords with number of words 173 * in the last block. 174 * 175 * Refer to block comment within space_map_incremental_destroy() 176 * to understand why this function is needed. 177 */ 178 static int 179 space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf, 180 uint64_t bufsz, uint64_t *nwords) 181 { 182 int error = 0; 183 dmu_buf_t *db; 184 185 /* 186 * Find the offset of the last word in the space map and use 187 * that to read the last block of the space map with 188 * dmu_buf_hold(). 189 */ 190 uint64_t last_word_offset = 191 sm->sm_phys->smp_length - sizeof (uint64_t); 192 error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset, 193 FTAG, &db, DMU_READ_NO_PREFETCH); 194 if (error != 0) 195 return (error); 196 197 ASSERT3U(sm->sm_object, ==, db->db_object); 198 ASSERT3U(sm->sm_blksz, ==, db->db_size); 199 ASSERT3U(bufsz, >=, db->db_size); 200 ASSERT(nwords != NULL); 201 202 uint64_t *words = db->db_data; 203 *nwords = 204 (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t); 205 206 ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t)); 207 208 uint64_t n = *nwords; 209 uint64_t j = n - 1; 210 for (uint64_t i = 0; i < n; i++) { 211 uint64_t entry = words[i]; 212 if (sm_entry_is_double_word(entry)) { 213 /* 214 * Since we are populating the buffer backwards 215 * we have to be extra careful and add the two 216 * words of the double-word entry in the right 217 * order. 218 */ 219 ASSERT3U(j, >, 0); 220 buf[j - 1] = entry; 221 222 i++; 223 ASSERT3U(i, <, n); 224 entry = words[i]; 225 buf[j] = entry; 226 j -= 2; 227 } else { 228 ASSERT(sm_entry_is_debug(entry) || 229 sm_entry_is_single_word(entry)); 230 buf[j] = entry; 231 j--; 232 } 233 } 234 235 /* 236 * Assert that we wrote backwards all the 237 * way to the beginning of the buffer. 238 */ 239 ASSERT3S(j, ==, -1); 240 241 dmu_buf_rele(db, FTAG); 242 return (error); 243 } 244 245 /* 246 * Note: This function performs destructive actions - specifically 247 * it deletes entries from the end of the space map. Thus, callers 248 * should ensure that they are holding the appropriate locks for 249 * the space map that they provide. 250 */ 251 int 252 space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg, 253 dmu_tx_t *tx) 254 { 255 uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE); 256 uint64_t *buf = zio_buf_alloc(bufsz); 257 258 dmu_buf_will_dirty(sm->sm_dbuf, tx); 259 260 /* 261 * Ideally we would want to iterate from the beginning of the 262 * space map to the end in incremental steps. The issue with this 263 * approach is that we don't have any field on-disk that points 264 * us where to start between each step. We could try zeroing out 265 * entries that we've destroyed, but this doesn't work either as 266 * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]). 267 * 268 * As a result, we destroy its entries incrementally starting from 269 * the end after applying the callback to each of them. 270 * 271 * The problem with this approach is that we cannot literally 272 * iterate through the words in the space map backwards as we 273 * can't distinguish two-word space map entries from their second 274 * word. Thus we do the following: 275 * 276 * 1] We get all the entries from the last block of the space map 277 * and put them into a buffer in reverse order. This way the 278 * last entry comes first in the buffer, the second to last is 279 * second, etc. 280 * 2] We iterate through the entries in the buffer and we apply 281 * the callback to each one. As we move from entry to entry we 282 * we decrease the size of the space map, deleting effectively 283 * each entry. 284 * 3] If there are no more entries in the space map or the callback 285 * returns a value other than 0, we stop iterating over the 286 * space map. If there are entries remaining and the callback 287 * returned 0, we go back to step [1]. 288 */ 289 int error = 0; 290 while (space_map_length(sm) > 0 && error == 0) { 291 uint64_t nwords = 0; 292 error = space_map_reversed_last_block_entries(sm, buf, bufsz, 293 &nwords); 294 if (error != 0) 295 break; 296 297 ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t)); 298 299 for (uint64_t i = 0; i < nwords; i++) { 300 uint64_t e = buf[i]; 301 302 if (sm_entry_is_debug(e)) { 303 sm->sm_phys->smp_length -= sizeof (uint64_t); 304 continue; 305 } 306 307 int words = 1; 308 uint64_t raw_offset, raw_run, vdev_id; 309 maptype_t type; 310 if (sm_entry_is_single_word(e)) { 311 type = SM_TYPE_DECODE(e); 312 vdev_id = SM_NO_VDEVID; 313 raw_offset = SM_OFFSET_DECODE(e); 314 raw_run = SM_RUN_DECODE(e); 315 } else { 316 ASSERT(sm_entry_is_double_word(e)); 317 words = 2; 318 319 raw_run = SM2_RUN_DECODE(e); 320 vdev_id = SM2_VDEV_DECODE(e); 321 322 /* move to the second word */ 323 i++; 324 e = buf[i]; 325 326 ASSERT3P(i, <=, nwords); 327 328 type = SM2_TYPE_DECODE(e); 329 raw_offset = SM2_OFFSET_DECODE(e); 330 } 331 332 uint64_t entry_offset = 333 (raw_offset << sm->sm_shift) + sm->sm_start; 334 uint64_t entry_run = raw_run << sm->sm_shift; 335 336 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift)); 337 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift)); 338 VERIFY3U(entry_offset, >=, sm->sm_start); 339 VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size); 340 VERIFY3U(entry_run, <=, sm->sm_size); 341 VERIFY3U(entry_offset + entry_run, <=, 342 sm->sm_start + sm->sm_size); 343 344 space_map_entry_t sme = { 345 .sme_type = type, 346 .sme_vdev = vdev_id, 347 .sme_offset = entry_offset, 348 .sme_run = entry_run 349 }; 350 error = callback(&sme, arg); 351 if (error != 0) 352 break; 353 354 if (type == SM_ALLOC) 355 sm->sm_phys->smp_alloc -= entry_run; 356 else 357 sm->sm_phys->smp_alloc += entry_run; 358 sm->sm_phys->smp_length -= words * sizeof (uint64_t); 359 } 360 } 361 362 if (space_map_length(sm) == 0) { 363 ASSERT0(error); 364 ASSERT0(space_map_allocated(sm)); 365 } 366 367 zio_buf_free(buf, bufsz); 368 return (error); 369 } 370 371 typedef struct space_map_load_arg { 372 space_map_t *smla_sm; 373 range_tree_t *smla_rt; 374 maptype_t smla_type; 375 } space_map_load_arg_t; 376 377 static int 378 space_map_load_callback(space_map_entry_t *sme, void *arg) 379 { 380 space_map_load_arg_t *smla = arg; 381 if (sme->sme_type == smla->smla_type) { 382 VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=, 383 smla->smla_sm->sm_size); 384 range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run); 385 } else { 386 range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run); 387 } 388 389 return (0); 390 } 391 392 /* 393 * Load the spacemap into the rangetree, like space_map_load. But only 394 * read the first 'length' bytes of the spacemap. 395 */ 396 int 397 space_map_load_length(space_map_t *sm, range_tree_t *rt, maptype_t maptype, 398 uint64_t length) 399 { 400 space_map_load_arg_t smla; 401 402 VERIFY0(range_tree_space(rt)); 403 404 if (maptype == SM_FREE) 405 range_tree_add(rt, sm->sm_start, sm->sm_size); 406 407 smla.smla_rt = rt; 408 smla.smla_sm = sm; 409 smla.smla_type = maptype; 410 int err = space_map_iterate(sm, length, 411 space_map_load_callback, &smla); 412 413 if (err != 0) 414 range_tree_vacate(rt, NULL, NULL); 415 416 return (err); 417 } 418 419 /* 420 * Load the space map disk into the specified range tree. Segments of maptype 421 * are added to the range tree, other segment types are removed. 422 */ 423 int 424 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype) 425 { 426 return (space_map_load_length(sm, rt, maptype, space_map_length(sm))); 427 } 428 429 void 430 space_map_histogram_clear(space_map_t *sm) 431 { 432 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) 433 return; 434 435 bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram)); 436 } 437 438 boolean_t 439 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt) 440 { 441 /* 442 * Verify that the in-core range tree does not have any 443 * ranges smaller than our sm_shift size. 444 */ 445 for (int i = 0; i < sm->sm_shift; i++) { 446 if (rt->rt_histogram[i] != 0) 447 return (B_FALSE); 448 } 449 return (B_TRUE); 450 } 451 452 void 453 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx) 454 { 455 int idx = 0; 456 457 ASSERT(dmu_tx_is_syncing(tx)); 458 VERIFY3U(space_map_object(sm), !=, 0); 459 460 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) 461 return; 462 463 dmu_buf_will_dirty(sm->sm_dbuf, tx); 464 465 ASSERT(space_map_histogram_verify(sm, rt)); 466 /* 467 * Transfer the content of the range tree histogram to the space 468 * map histogram. The space map histogram contains 32 buckets ranging 469 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree, 470 * however, can represent ranges from 2^0 to 2^63. Since the space 471 * map only cares about allocatable blocks (minimum of sm_shift) we 472 * can safely ignore all ranges in the range tree smaller than sm_shift. 473 */ 474 for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) { 475 476 /* 477 * Since the largest histogram bucket in the space map is 478 * 2^(32+sm_shift-1), we need to normalize the values in 479 * the range tree for any bucket larger than that size. For 480 * example given an sm_shift of 9, ranges larger than 2^40 481 * would get normalized as if they were 1TB ranges. Assume 482 * the range tree had a count of 5 in the 2^44 (16TB) bucket, 483 * the calculation below would normalize this to 5 * 2^4 (16). 484 */ 485 ASSERT3U(i, >=, idx + sm->sm_shift); 486 sm->sm_phys->smp_histogram[idx] += 487 rt->rt_histogram[i] << (i - idx - sm->sm_shift); 488 489 /* 490 * Increment the space map's index as long as we haven't 491 * reached the maximum bucket size. Accumulate all ranges 492 * larger than the max bucket size into the last bucket. 493 */ 494 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) { 495 ASSERT3U(idx + sm->sm_shift, ==, i); 496 idx++; 497 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE); 498 } 499 } 500 } 501 502 static void 503 space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx) 504 { 505 dmu_buf_will_dirty(sm->sm_dbuf, tx); 506 507 uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) | 508 SM_DEBUG_ACTION_ENCODE(maptype) | 509 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) | 510 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx)); 511 512 dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_length, 513 sizeof (dentry), &dentry, tx); 514 515 sm->sm_phys->smp_length += sizeof (dentry); 516 } 517 518 /* 519 * Writes one or more entries given a segment. 520 * 521 * Note: The function may release the dbuf from the pointer initially 522 * passed to it, and return a different dbuf. Also, the space map's 523 * dbuf must be dirty for the changes in sm_phys to take effect. 524 */ 525 static void 526 space_map_write_seg(space_map_t *sm, range_seg_t *rs, maptype_t maptype, 527 uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp, void *tag, dmu_tx_t *tx) 528 { 529 ASSERT3U(words, !=, 0); 530 ASSERT3U(words, <=, 2); 531 532 /* ensure the vdev_id can be represented by the space map */ 533 ASSERT3U(vdev_id, <=, SM_NO_VDEVID); 534 535 /* 536 * if this is a single word entry, ensure that no vdev was 537 * specified. 538 */ 539 IMPLY(words == 1, vdev_id == SM_NO_VDEVID); 540 541 dmu_buf_t *db = *dbp; 542 ASSERT3U(db->db_size, ==, sm->sm_blksz); 543 544 uint64_t *block_base = db->db_data; 545 uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t)); 546 uint64_t *block_cursor = block_base + 547 (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t); 548 549 ASSERT3P(block_cursor, <=, block_end); 550 551 uint64_t size = (rs->rs_end - rs->rs_start) >> sm->sm_shift; 552 uint64_t start = (rs->rs_start - sm->sm_start) >> sm->sm_shift; 553 uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX; 554 555 ASSERT3U(rs->rs_start, >=, sm->sm_start); 556 ASSERT3U(rs->rs_start, <, sm->sm_start + sm->sm_size); 557 ASSERT3U(rs->rs_end - rs->rs_start, <=, sm->sm_size); 558 ASSERT3U(rs->rs_end, <=, sm->sm_start + sm->sm_size); 559 560 while (size != 0) { 561 ASSERT3P(block_cursor, <=, block_end); 562 563 /* 564 * If we are at the end of this block, flush it and start 565 * writing again from the beginning. 566 */ 567 if (block_cursor == block_end) { 568 dmu_buf_rele(db, tag); 569 570 uint64_t next_word_offset = sm->sm_phys->smp_length; 571 VERIFY0(dmu_buf_hold(sm->sm_os, 572 space_map_object(sm), next_word_offset, 573 tag, &db, DMU_READ_PREFETCH)); 574 dmu_buf_will_dirty(db, tx); 575 576 /* update caller's dbuf */ 577 *dbp = db; 578 579 ASSERT3U(db->db_size, ==, sm->sm_blksz); 580 581 block_base = db->db_data; 582 block_cursor = block_base; 583 block_end = block_base + 584 (db->db_size / sizeof (uint64_t)); 585 } 586 587 /* 588 * If we are writing a two-word entry and we only have one 589 * word left on this block, just pad it with an empty debug 590 * entry and write the two-word entry in the next block. 591 */ 592 uint64_t *next_entry = block_cursor + 1; 593 if (next_entry == block_end && words > 1) { 594 ASSERT3U(words, ==, 2); 595 *block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) | 596 SM_DEBUG_ACTION_ENCODE(0) | 597 SM_DEBUG_SYNCPASS_ENCODE(0) | 598 SM_DEBUG_TXG_ENCODE(0); 599 block_cursor++; 600 sm->sm_phys->smp_length += sizeof (uint64_t); 601 ASSERT3P(block_cursor, ==, block_end); 602 continue; 603 } 604 605 uint64_t run_len = MIN(size, run_max); 606 switch (words) { 607 case 1: 608 *block_cursor = SM_OFFSET_ENCODE(start) | 609 SM_TYPE_ENCODE(maptype) | 610 SM_RUN_ENCODE(run_len); 611 block_cursor++; 612 break; 613 case 2: 614 /* write the first word of the entry */ 615 *block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) | 616 SM2_RUN_ENCODE(run_len) | 617 SM2_VDEV_ENCODE(vdev_id); 618 block_cursor++; 619 620 /* move on to the second word of the entry */ 621 ASSERT3P(block_cursor, <, block_end); 622 *block_cursor = SM2_TYPE_ENCODE(maptype) | 623 SM2_OFFSET_ENCODE(start); 624 block_cursor++; 625 break; 626 default: 627 panic("%d-word space map entries are not supported", 628 words); 629 break; 630 } 631 sm->sm_phys->smp_length += words * sizeof (uint64_t); 632 633 start += run_len; 634 size -= run_len; 635 } 636 ASSERT0(size); 637 638 } 639 640 /* 641 * Note: The space map's dbuf must be dirty for the changes in sm_phys to 642 * take effect. 643 */ 644 static void 645 space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype, 646 uint64_t vdev_id, dmu_tx_t *tx) 647 { 648 spa_t *spa = tx->tx_pool->dp_spa; 649 dmu_buf_t *db; 650 651 space_map_write_intro_debug(sm, maptype, tx); 652 653 #ifdef DEBUG 654 /* 655 * We do this right after we write the intro debug entry 656 * because the estimate does not take it into account. 657 */ 658 uint64_t initial_objsize = sm->sm_phys->smp_length; 659 uint64_t estimated_growth = 660 space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID); 661 uint64_t estimated_final_objsize = initial_objsize + estimated_growth; 662 #endif 663 664 /* 665 * Find the offset right after the last word in the space map 666 * and use that to get a hold of the last block, so we can 667 * start appending to it. 668 */ 669 uint64_t next_word_offset = sm->sm_phys->smp_length; 670 VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm), 671 next_word_offset, FTAG, &db, DMU_READ_PREFETCH)); 672 ASSERT3U(db->db_size, ==, sm->sm_blksz); 673 674 dmu_buf_will_dirty(db, tx); 675 676 avl_tree_t *t = &rt->rt_root; 677 for (range_seg_t *rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) { 678 uint64_t offset = (rs->rs_start - sm->sm_start) >> sm->sm_shift; 679 uint64_t length = (rs->rs_end - rs->rs_start) >> sm->sm_shift; 680 uint8_t words = 1; 681 682 /* 683 * We only write two-word entries when both of the following 684 * are true: 685 * 686 * [1] The feature is enabled. 687 * [2] The offset or run is too big for a single-word entry, 688 * or the vdev_id is set (meaning not equal to 689 * SM_NO_VDEVID). 690 * 691 * Note that for purposes of testing we've added the case that 692 * we write two-word entries occasionally when the feature is 693 * enabled and zfs_force_some_double_word_sm_entries has been 694 * set. 695 */ 696 if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) && 697 (offset >= (1ULL << SM_OFFSET_BITS) || 698 length > SM_RUN_MAX || 699 vdev_id != SM_NO_VDEVID || 700 (zfs_force_some_double_word_sm_entries && 701 spa_get_random(100) == 0))) 702 words = 2; 703 704 space_map_write_seg(sm, rs, maptype, vdev_id, words, 705 &db, FTAG, tx); 706 } 707 708 dmu_buf_rele(db, FTAG); 709 710 #ifdef DEBUG 711 /* 712 * We expect our estimation to be based on the worst case 713 * scenario [see comment in space_map_estimate_optimal_size()]. 714 * Therefore we expect the actual objsize to be equal or less 715 * than whatever we estimated it to be. 716 */ 717 ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_length); 718 #endif 719 } 720 721 /* 722 * Note: This function manipulates the state of the given space map but 723 * does not hold any locks implicitly. Thus the caller is responsible 724 * for synchronizing writes to the space map. 725 */ 726 void 727 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype, 728 uint64_t vdev_id, dmu_tx_t *tx) 729 { 730 objset_t *os = sm->sm_os; 731 732 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 733 VERIFY3U(space_map_object(sm), !=, 0); 734 735 dmu_buf_will_dirty(sm->sm_dbuf, tx); 736 737 /* 738 * This field is no longer necessary since the in-core space map 739 * now contains the object number but is maintained for backwards 740 * compatibility. 741 */ 742 sm->sm_phys->smp_object = sm->sm_object; 743 744 if (range_tree_is_empty(rt)) { 745 VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object); 746 return; 747 } 748 749 if (maptype == SM_ALLOC) 750 sm->sm_phys->smp_alloc += range_tree_space(rt); 751 else 752 sm->sm_phys->smp_alloc -= range_tree_space(rt); 753 754 uint64_t nodes = avl_numnodes(&rt->rt_root); 755 uint64_t rt_space = range_tree_space(rt); 756 757 space_map_write_impl(sm, rt, maptype, vdev_id, tx); 758 759 /* 760 * Ensure that the space_map's accounting wasn't changed 761 * while we were in the middle of writing it out. 762 */ 763 VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root)); 764 VERIFY3U(range_tree_space(rt), ==, rt_space); 765 } 766 767 static int 768 space_map_open_impl(space_map_t *sm) 769 { 770 int error; 771 u_longlong_t blocks; 772 773 error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf); 774 if (error) 775 return (error); 776 777 dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks); 778 sm->sm_phys = sm->sm_dbuf->db_data; 779 return (0); 780 } 781 782 int 783 space_map_open(space_map_t **smp, objset_t *os, uint64_t object, 784 uint64_t start, uint64_t size, uint8_t shift) 785 { 786 space_map_t *sm; 787 int error; 788 789 ASSERT(*smp == NULL); 790 ASSERT(os != NULL); 791 ASSERT(object != 0); 792 793 sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP); 794 795 sm->sm_start = start; 796 sm->sm_size = size; 797 sm->sm_shift = shift; 798 sm->sm_os = os; 799 sm->sm_object = object; 800 801 error = space_map_open_impl(sm); 802 if (error != 0) { 803 space_map_close(sm); 804 return (error); 805 } 806 *smp = sm; 807 808 return (0); 809 } 810 811 void 812 space_map_close(space_map_t *sm) 813 { 814 if (sm == NULL) 815 return; 816 817 if (sm->sm_dbuf != NULL) 818 dmu_buf_rele(sm->sm_dbuf, sm); 819 sm->sm_dbuf = NULL; 820 sm->sm_phys = NULL; 821 822 kmem_free(sm, sizeof (*sm)); 823 } 824 825 void 826 space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx) 827 { 828 objset_t *os = sm->sm_os; 829 spa_t *spa = dmu_objset_spa(os); 830 dmu_object_info_t doi; 831 832 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 833 ASSERT(dmu_tx_is_syncing(tx)); 834 VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa)); 835 836 dmu_object_info_from_db(sm->sm_dbuf, &doi); 837 838 /* 839 * If the space map has the wrong bonus size (because 840 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or 841 * the wrong block size (because space_map_blksz has changed), 842 * free and re-allocate its object with the updated sizes. 843 * 844 * Otherwise, just truncate the current object. 845 */ 846 if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) && 847 doi.doi_bonus_size != sizeof (space_map_phys_t)) || 848 doi.doi_data_block_size != blocksize || 849 doi.doi_metadata_block_size != 1 << space_map_ibs) { 850 zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating " 851 "object[%llu]: old bonus %u, old blocksz %u", 852 dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object, 853 doi.doi_bonus_size, doi.doi_data_block_size); 854 855 space_map_free(sm, tx); 856 dmu_buf_rele(sm->sm_dbuf, sm); 857 858 sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx); 859 VERIFY0(space_map_open_impl(sm)); 860 } else { 861 VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx)); 862 863 /* 864 * If the spacemap is reallocated, its histogram 865 * will be reset. Do the same in the common case so that 866 * bugs related to the uncommon case do not go unnoticed. 867 */ 868 bzero(sm->sm_phys->smp_histogram, 869 sizeof (sm->sm_phys->smp_histogram)); 870 } 871 872 dmu_buf_will_dirty(sm->sm_dbuf, tx); 873 sm->sm_phys->smp_length = 0; 874 sm->sm_phys->smp_alloc = 0; 875 } 876 877 uint64_t 878 space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx) 879 { 880 spa_t *spa = dmu_objset_spa(os); 881 uint64_t object; 882 int bonuslen; 883 884 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) { 885 spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx); 886 bonuslen = sizeof (space_map_phys_t); 887 ASSERT3U(bonuslen, <=, dmu_bonus_max()); 888 } else { 889 bonuslen = SPACE_MAP_SIZE_V0; 890 } 891 892 object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize, 893 space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx); 894 895 return (object); 896 } 897 898 void 899 space_map_free_obj(objset_t *os, uint64_t smobj, dmu_tx_t *tx) 900 { 901 spa_t *spa = dmu_objset_spa(os); 902 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) { 903 dmu_object_info_t doi; 904 905 VERIFY0(dmu_object_info(os, smobj, &doi)); 906 if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) { 907 spa_feature_decr(spa, 908 SPA_FEATURE_SPACEMAP_HISTOGRAM, tx); 909 } 910 } 911 912 VERIFY0(dmu_object_free(os, smobj, tx)); 913 } 914 915 void 916 space_map_free(space_map_t *sm, dmu_tx_t *tx) 917 { 918 if (sm == NULL) 919 return; 920 921 space_map_free_obj(sm->sm_os, space_map_object(sm), tx); 922 sm->sm_object = 0; 923 } 924 925 /* 926 * Given a range tree, it makes a worst-case estimate of how much 927 * space would the tree's segments take if they were written to 928 * the given space map. 929 */ 930 uint64_t 931 space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt, 932 uint64_t vdev_id) 933 { 934 spa_t *spa = dmu_objset_spa(sm->sm_os); 935 uint64_t shift = sm->sm_shift; 936 uint64_t *histogram = rt->rt_histogram; 937 uint64_t entries_for_seg = 0; 938 939 /* 940 * In order to get a quick estimate of the optimal size that this 941 * range tree would have on-disk as a space map, we iterate through 942 * its histogram buckets instead of iterating through its nodes. 943 * 944 * Note that this is a highest-bound/worst-case estimate for the 945 * following reasons: 946 * 947 * 1] We assume that we always add a debug padding for each block 948 * we write and we also assume that we start at the last word 949 * of a block attempting to write a two-word entry. 950 * 2] Rounding up errors due to the way segments are distributed 951 * in the buckets of the range tree's histogram. 952 * 3] The activation of zfs_force_some_double_word_sm_entries 953 * (tunable) when testing. 954 * 955 * = Math and Rounding Errors = 956 * 957 * rt_histogram[i] bucket of a range tree represents the number 958 * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given 959 * that, we want to divide the buckets into groups: Buckets that 960 * can be represented using a single-word entry, ones that can 961 * be represented with a double-word entry, and ones that can 962 * only be represented with multiple two-word entries. 963 * 964 * [Note that if the new encoding feature is not enabled there 965 * are only two groups: single-word entry buckets and multiple 966 * single-word entry buckets. The information below assumes 967 * two-word entries enabled, but it can easily applied when 968 * the feature is not enabled] 969 * 970 * To find the highest bucket that can be represented with a 971 * single-word entry we look at the maximum run that such entry 972 * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that 973 * the run of a space map entry is shifted by sm_shift, thus we 974 * add it to the exponent]. This way, excluding the value of the 975 * maximum run that can be represented by a single-word entry, 976 * all runs that are smaller exist in buckets 0 to 977 * SM_RUN_BITS + shift - 1. 978 * 979 * To find the highest bucket that can be represented with a 980 * double-word entry, we follow the same approach. Finally, any 981 * bucket higher than that are represented with multiple two-word 982 * entries. To be more specific, if the highest bucket whose 983 * segments can be represented with a single two-word entry is X, 984 * then bucket X+1 will need 2 two-word entries for each of its 985 * segments, X+2 will need 4, X+3 will need 8, ...etc. 986 * 987 * With all of the above we make our estimation based on bucket 988 * groups. There is a rounding error though. As we mentioned in 989 * the example with the one-word entry, the maximum run that can 990 * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is 991 * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of 992 * that length fall into the next bucket (and bucket group) where 993 * we start counting two-word entries and this is one more reason 994 * why the estimated size may end up being bigger than the actual 995 * size written. 996 */ 997 uint64_t size = 0; 998 uint64_t idx = 0; 999 1000 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) || 1001 (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) { 1002 1003 /* 1004 * If we are trying to force some double word entries just 1005 * assume the worst-case of every single word entry being 1006 * written as a double word entry. 1007 */ 1008 uint64_t entry_size = 1009 (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) && 1010 zfs_force_some_double_word_sm_entries) ? 1011 (2 * sizeof (uint64_t)) : sizeof (uint64_t); 1012 1013 uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1; 1014 for (; idx <= single_entry_max_bucket; idx++) 1015 size += histogram[idx] * entry_size; 1016 1017 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) { 1018 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) { 1019 ASSERT3U(idx, >=, single_entry_max_bucket); 1020 entries_for_seg = 1021 1ULL << (idx - single_entry_max_bucket); 1022 size += histogram[idx] * 1023 entries_for_seg * entry_size; 1024 } 1025 return (size); 1026 } 1027 } 1028 1029 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)); 1030 1031 uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1; 1032 for (; idx <= double_entry_max_bucket; idx++) 1033 size += histogram[idx] * 2 * sizeof (uint64_t); 1034 1035 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) { 1036 ASSERT3U(idx, >=, double_entry_max_bucket); 1037 entries_for_seg = 1ULL << (idx - double_entry_max_bucket); 1038 size += histogram[idx] * 1039 entries_for_seg * 2 * sizeof (uint64_t); 1040 } 1041 1042 /* 1043 * Assume the worst case where we start with the padding at the end 1044 * of the current block and we add an extra padding entry at the end 1045 * of all subsequent blocks. 1046 */ 1047 size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t); 1048 1049 return (size); 1050 } 1051 1052 uint64_t 1053 space_map_object(space_map_t *sm) 1054 { 1055 return (sm != NULL ? sm->sm_object : 0); 1056 } 1057 1058 int64_t 1059 space_map_allocated(space_map_t *sm) 1060 { 1061 return (sm != NULL ? sm->sm_phys->smp_alloc : 0); 1062 } 1063 1064 uint64_t 1065 space_map_length(space_map_t *sm) 1066 { 1067 return (sm != NULL ? sm->sm_phys->smp_length : 0); 1068 } 1069