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 /* 23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved. 25 */ 26 27 #include <sys/zfs_context.h> 28 #include <sys/spa.h> 29 #include <sys/spa_impl.h> 30 #include <sys/zio.h> 31 #include <sys/ddt.h> 32 #include <sys/zap.h> 33 #include <sys/dmu_tx.h> 34 #include <sys/arc.h> 35 #include <sys/dsl_pool.h> 36 #include <sys/zio_checksum.h> 37 #include <sys/zio_compress.h> 38 #include <sys/dsl_scan.h> 39 40 /* 41 * Enable/disable prefetching of dedup-ed blocks which are going to be freed. 42 */ 43 int zfs_dedup_prefetch = 1; 44 45 static const ddt_ops_t *ddt_ops[DDT_TYPES] = { 46 &ddt_zap_ops, 47 }; 48 49 static const char *ddt_class_name[DDT_CLASSES] = { 50 "ditto", 51 "duplicate", 52 "unique", 53 }; 54 55 static void 56 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 57 dmu_tx_t *tx) 58 { 59 spa_t *spa = ddt->ddt_spa; 60 objset_t *os = ddt->ddt_os; 61 uint64_t *objectp = &ddt->ddt_object[type][class]; 62 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_flags & 63 ZCHECKSUM_FLAG_DEDUP; 64 char name[DDT_NAMELEN]; 65 66 ddt_object_name(ddt, type, class, name); 67 68 ASSERT(*objectp == 0); 69 VERIFY(ddt_ops[type]->ddt_op_create(os, objectp, tx, prehash) == 0); 70 ASSERT(*objectp != 0); 71 72 VERIFY(zap_add(os, DMU_POOL_DIRECTORY_OBJECT, name, 73 sizeof (uint64_t), 1, objectp, tx) == 0); 74 75 VERIFY(zap_add(os, spa->spa_ddt_stat_object, name, 76 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t), 77 &ddt->ddt_histogram[type][class], tx) == 0); 78 } 79 80 static void 81 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 82 dmu_tx_t *tx) 83 { 84 spa_t *spa = ddt->ddt_spa; 85 objset_t *os = ddt->ddt_os; 86 uint64_t *objectp = &ddt->ddt_object[type][class]; 87 char name[DDT_NAMELEN]; 88 89 ddt_object_name(ddt, type, class, name); 90 91 ASSERT(*objectp != 0); 92 ASSERT(ddt_object_count(ddt, type, class) == 0); 93 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class])); 94 VERIFY(zap_remove(os, DMU_POOL_DIRECTORY_OBJECT, name, tx) == 0); 95 VERIFY(zap_remove(os, spa->spa_ddt_stat_object, name, tx) == 0); 96 VERIFY(ddt_ops[type]->ddt_op_destroy(os, *objectp, tx) == 0); 97 bzero(&ddt->ddt_object_stats[type][class], sizeof (ddt_object_t)); 98 99 *objectp = 0; 100 } 101 102 static int 103 ddt_object_load(ddt_t *ddt, enum ddt_type type, enum ddt_class class) 104 { 105 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class]; 106 dmu_object_info_t doi; 107 char name[DDT_NAMELEN]; 108 int error; 109 110 ddt_object_name(ddt, type, class, name); 111 112 error = zap_lookup(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT, name, 113 sizeof (uint64_t), 1, &ddt->ddt_object[type][class]); 114 115 if (error != 0) 116 return (error); 117 118 VERIFY0(zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name, 119 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t), 120 &ddt->ddt_histogram[type][class])); 121 122 /* 123 * Seed the cached statistics. 124 */ 125 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0); 126 127 ddo->ddo_count = ddt_object_count(ddt, type, class); 128 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9; 129 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size; 130 131 return (0); 132 } 133 134 static void 135 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 136 dmu_tx_t *tx) 137 { 138 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class]; 139 dmu_object_info_t doi; 140 char name[DDT_NAMELEN]; 141 142 ddt_object_name(ddt, type, class, name); 143 144 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name, 145 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t), 146 &ddt->ddt_histogram[type][class], tx) == 0); 147 148 /* 149 * Cache DDT statistics; this is the only time they'll change. 150 */ 151 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0); 152 153 ddo->ddo_count = ddt_object_count(ddt, type, class); 154 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9; 155 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size; 156 } 157 158 static int 159 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 160 ddt_entry_t *dde) 161 { 162 if (!ddt_object_exists(ddt, type, class)) 163 return (SET_ERROR(ENOENT)); 164 165 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os, 166 ddt->ddt_object[type][class], dde)); 167 } 168 169 static void 170 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 171 ddt_entry_t *dde) 172 { 173 if (!ddt_object_exists(ddt, type, class)) 174 return; 175 176 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os, 177 ddt->ddt_object[type][class], dde); 178 } 179 180 int 181 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 182 ddt_entry_t *dde, dmu_tx_t *tx) 183 { 184 ASSERT(ddt_object_exists(ddt, type, class)); 185 186 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os, 187 ddt->ddt_object[type][class], dde, tx)); 188 } 189 190 static int 191 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 192 ddt_entry_t *dde, dmu_tx_t *tx) 193 { 194 ASSERT(ddt_object_exists(ddt, type, class)); 195 196 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os, 197 ddt->ddt_object[type][class], dde, tx)); 198 } 199 200 int 201 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 202 uint64_t *walk, ddt_entry_t *dde) 203 { 204 ASSERT(ddt_object_exists(ddt, type, class)); 205 206 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os, 207 ddt->ddt_object[type][class], dde, walk)); 208 } 209 210 uint64_t 211 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class) 212 { 213 ASSERT(ddt_object_exists(ddt, type, class)); 214 215 return (ddt_ops[type]->ddt_op_count(ddt->ddt_os, 216 ddt->ddt_object[type][class])); 217 } 218 219 int 220 ddt_object_info(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 221 dmu_object_info_t *doi) 222 { 223 if (!ddt_object_exists(ddt, type, class)) 224 return (SET_ERROR(ENOENT)); 225 226 return (dmu_object_info(ddt->ddt_os, ddt->ddt_object[type][class], 227 doi)); 228 } 229 230 boolean_t 231 ddt_object_exists(ddt_t *ddt, enum ddt_type type, enum ddt_class class) 232 { 233 return (!!ddt->ddt_object[type][class]); 234 } 235 236 void 237 ddt_object_name(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 238 char *name) 239 { 240 (void) sprintf(name, DMU_POOL_DDT, 241 zio_checksum_table[ddt->ddt_checksum].ci_name, 242 ddt_ops[type]->ddt_op_name, ddt_class_name[class]); 243 } 244 245 void 246 ddt_bp_fill(const ddt_phys_t *ddp, blkptr_t *bp, uint64_t txg) 247 { 248 ASSERT(txg != 0); 249 250 for (int d = 0; d < SPA_DVAS_PER_BP; d++) 251 bp->blk_dva[d] = ddp->ddp_dva[d]; 252 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth); 253 } 254 255 void 256 ddt_bp_create(enum zio_checksum checksum, 257 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp) 258 { 259 BP_ZERO(bp); 260 261 if (ddp != NULL) 262 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth); 263 264 bp->blk_cksum = ddk->ddk_cksum; 265 bp->blk_fill = 1; 266 267 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk)); 268 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk)); 269 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk)); 270 BP_SET_CHECKSUM(bp, checksum); 271 BP_SET_TYPE(bp, DMU_OT_DEDUP); 272 BP_SET_LEVEL(bp, 0); 273 BP_SET_DEDUP(bp, 0); 274 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 275 } 276 277 void 278 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp) 279 { 280 ddk->ddk_cksum = bp->blk_cksum; 281 ddk->ddk_prop = 0; 282 283 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp)); 284 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp)); 285 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp)); 286 } 287 288 void 289 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp) 290 { 291 ASSERT(ddp->ddp_phys_birth == 0); 292 293 for (int d = 0; d < SPA_DVAS_PER_BP; d++) 294 ddp->ddp_dva[d] = bp->blk_dva[d]; 295 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp); 296 } 297 298 void 299 ddt_phys_clear(ddt_phys_t *ddp) 300 { 301 bzero(ddp, sizeof (*ddp)); 302 } 303 304 void 305 ddt_phys_addref(ddt_phys_t *ddp) 306 { 307 ddp->ddp_refcnt++; 308 } 309 310 void 311 ddt_phys_decref(ddt_phys_t *ddp) 312 { 313 ASSERT((int64_t)ddp->ddp_refcnt > 0); 314 ddp->ddp_refcnt--; 315 } 316 317 void 318 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg) 319 { 320 blkptr_t blk; 321 322 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk); 323 ddt_phys_clear(ddp); 324 zio_free(ddt->ddt_spa, txg, &blk); 325 } 326 327 ddt_phys_t * 328 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp) 329 { 330 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys; 331 332 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 333 if (DVA_EQUAL(BP_IDENTITY(bp), &ddp->ddp_dva[0]) && 334 BP_PHYSICAL_BIRTH(bp) == ddp->ddp_phys_birth) 335 return (ddp); 336 } 337 return (NULL); 338 } 339 340 uint64_t 341 ddt_phys_total_refcnt(const ddt_entry_t *dde) 342 { 343 uint64_t refcnt = 0; 344 345 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) 346 refcnt += dde->dde_phys[p].ddp_refcnt; 347 348 return (refcnt); 349 } 350 351 static void 352 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds) 353 { 354 spa_t *spa = ddt->ddt_spa; 355 ddt_phys_t *ddp = dde->dde_phys; 356 ddt_key_t *ddk = &dde->dde_key; 357 uint64_t lsize = DDK_GET_LSIZE(ddk); 358 uint64_t psize = DDK_GET_PSIZE(ddk); 359 360 bzero(dds, sizeof (*dds)); 361 362 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 363 uint64_t dsize = 0; 364 uint64_t refcnt = ddp->ddp_refcnt; 365 366 if (ddp->ddp_phys_birth == 0) 367 continue; 368 369 for (int d = 0; d < SPA_DVAS_PER_BP; d++) 370 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]); 371 372 dds->dds_blocks += 1; 373 dds->dds_lsize += lsize; 374 dds->dds_psize += psize; 375 dds->dds_dsize += dsize; 376 377 dds->dds_ref_blocks += refcnt; 378 dds->dds_ref_lsize += lsize * refcnt; 379 dds->dds_ref_psize += psize * refcnt; 380 dds->dds_ref_dsize += dsize * refcnt; 381 } 382 } 383 384 void 385 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg) 386 { 387 const uint64_t *s = (const uint64_t *)src; 388 uint64_t *d = (uint64_t *)dst; 389 uint64_t *d_end = (uint64_t *)(dst + 1); 390 391 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */ 392 393 while (d < d_end) 394 *d++ += (*s++ ^ neg) - neg; 395 } 396 397 static void 398 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg) 399 { 400 ddt_stat_t dds; 401 ddt_histogram_t *ddh; 402 int bucket; 403 404 ddt_stat_generate(ddt, dde, &dds); 405 406 bucket = highbit64(dds.dds_ref_blocks) - 1; 407 ASSERT(bucket >= 0); 408 409 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class]; 410 411 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg); 412 } 413 414 void 415 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src) 416 { 417 for (int h = 0; h < 64; h++) 418 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0); 419 } 420 421 void 422 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh) 423 { 424 bzero(dds, sizeof (*dds)); 425 426 for (int h = 0; h < 64; h++) 427 ddt_stat_add(dds, &ddh->ddh_stat[h], 0); 428 } 429 430 boolean_t 431 ddt_histogram_empty(const ddt_histogram_t *ddh) 432 { 433 const uint64_t *s = (const uint64_t *)ddh; 434 const uint64_t *s_end = (const uint64_t *)(ddh + 1); 435 436 while (s < s_end) 437 if (*s++ != 0) 438 return (B_FALSE); 439 440 return (B_TRUE); 441 } 442 443 void 444 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total) 445 { 446 /* Sum the statistics we cached in ddt_object_sync(). */ 447 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 448 ddt_t *ddt = spa->spa_ddt[c]; 449 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 450 for (enum ddt_class class = 0; class < DDT_CLASSES; 451 class++) { 452 ddt_object_t *ddo = 453 &ddt->ddt_object_stats[type][class]; 454 ddo_total->ddo_count += ddo->ddo_count; 455 ddo_total->ddo_dspace += ddo->ddo_dspace; 456 ddo_total->ddo_mspace += ddo->ddo_mspace; 457 } 458 } 459 } 460 461 /* ... and compute the averages. */ 462 if (ddo_total->ddo_count != 0) { 463 ddo_total->ddo_dspace /= ddo_total->ddo_count; 464 ddo_total->ddo_mspace /= ddo_total->ddo_count; 465 } 466 } 467 468 void 469 ddt_get_dedup_histogram(spa_t *spa, ddt_histogram_t *ddh) 470 { 471 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 472 ddt_t *ddt = spa->spa_ddt[c]; 473 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 474 for (enum ddt_class class = 0; class < DDT_CLASSES; 475 class++) { 476 ddt_histogram_add(ddh, 477 &ddt->ddt_histogram_cache[type][class]); 478 } 479 } 480 } 481 } 482 483 void 484 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total) 485 { 486 ddt_histogram_t *ddh_total; 487 488 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_SLEEP); 489 ddt_get_dedup_histogram(spa, ddh_total); 490 ddt_histogram_stat(dds_total, ddh_total); 491 kmem_free(ddh_total, sizeof (ddt_histogram_t)); 492 } 493 494 uint64_t 495 ddt_get_dedup_dspace(spa_t *spa) 496 { 497 ddt_stat_t dds_total = { 0 }; 498 499 ddt_get_dedup_stats(spa, &dds_total); 500 return (dds_total.dds_ref_dsize - dds_total.dds_dsize); 501 } 502 503 uint64_t 504 ddt_get_pool_dedup_ratio(spa_t *spa) 505 { 506 ddt_stat_t dds_total = { 0 }; 507 508 ddt_get_dedup_stats(spa, &dds_total); 509 if (dds_total.dds_dsize == 0) 510 return (100); 511 512 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize); 513 } 514 515 int 516 ddt_ditto_copies_needed(ddt_t *ddt, ddt_entry_t *dde, ddt_phys_t *ddp_willref) 517 { 518 spa_t *spa = ddt->ddt_spa; 519 uint64_t total_refcnt = 0; 520 uint64_t ditto = spa->spa_dedup_ditto; 521 int total_copies = 0; 522 int desired_copies = 0; 523 524 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 525 ddt_phys_t *ddp = &dde->dde_phys[p]; 526 zio_t *zio = dde->dde_lead_zio[p]; 527 uint64_t refcnt = ddp->ddp_refcnt; /* committed refs */ 528 if (zio != NULL) 529 refcnt += zio->io_parent_count; /* pending refs */ 530 if (ddp == ddp_willref) 531 refcnt++; /* caller's ref */ 532 if (refcnt != 0) { 533 total_refcnt += refcnt; 534 total_copies += p; 535 } 536 } 537 538 if (ditto == 0 || ditto > UINT32_MAX) 539 ditto = UINT32_MAX; 540 541 if (total_refcnt >= 1) 542 desired_copies++; 543 if (total_refcnt >= ditto) 544 desired_copies++; 545 if (total_refcnt >= ditto * ditto) 546 desired_copies++; 547 548 return (MAX(desired_copies, total_copies) - total_copies); 549 } 550 551 int 552 ddt_ditto_copies_present(ddt_entry_t *dde) 553 { 554 ddt_phys_t *ddp = &dde->dde_phys[DDT_PHYS_DITTO]; 555 dva_t *dva = ddp->ddp_dva; 556 int copies = 0 - DVA_GET_GANG(dva); 557 558 for (int d = 0; d < SPA_DVAS_PER_BP; d++, dva++) 559 if (DVA_IS_VALID(dva)) 560 copies++; 561 562 ASSERT(copies >= 0 && copies < SPA_DVAS_PER_BP); 563 564 return (copies); 565 } 566 567 size_t 568 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len) 569 { 570 uchar_t *version = dst++; 571 int cpfunc = ZIO_COMPRESS_ZLE; 572 zio_compress_info_t *ci = &zio_compress_table[cpfunc]; 573 size_t c_len; 574 575 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */ 576 577 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level); 578 579 if (c_len == s_len) { 580 cpfunc = ZIO_COMPRESS_OFF; 581 bcopy(src, dst, s_len); 582 } 583 584 *version = cpfunc; 585 /* CONSTCOND */ 586 if (ZFS_HOST_BYTEORDER) 587 *version |= DDT_COMPRESS_BYTEORDER_MASK; 588 589 return (c_len + 1); 590 } 591 592 void 593 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len) 594 { 595 uchar_t version = *src++; 596 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK; 597 zio_compress_info_t *ci = &zio_compress_table[cpfunc]; 598 599 if (ci->ci_decompress != NULL) 600 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level); 601 else 602 bcopy(src, dst, d_len); 603 604 if (((version & DDT_COMPRESS_BYTEORDER_MASK) != 0) != 605 (ZFS_HOST_BYTEORDER != 0)) 606 byteswap_uint64_array(dst, d_len); 607 } 608 609 ddt_t * 610 ddt_select_by_checksum(spa_t *spa, enum zio_checksum c) 611 { 612 return (spa->spa_ddt[c]); 613 } 614 615 ddt_t * 616 ddt_select(spa_t *spa, const blkptr_t *bp) 617 { 618 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]); 619 } 620 621 void 622 ddt_enter(ddt_t *ddt) 623 { 624 mutex_enter(&ddt->ddt_lock); 625 } 626 627 void 628 ddt_exit(ddt_t *ddt) 629 { 630 mutex_exit(&ddt->ddt_lock); 631 } 632 633 static ddt_entry_t * 634 ddt_alloc(const ddt_key_t *ddk) 635 { 636 ddt_entry_t *dde; 637 638 dde = kmem_zalloc(sizeof (ddt_entry_t), KM_SLEEP); 639 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL); 640 641 dde->dde_key = *ddk; 642 643 return (dde); 644 } 645 646 static void 647 ddt_free(ddt_entry_t *dde) 648 { 649 ASSERT(!dde->dde_loading); 650 651 for (int p = 0; p < DDT_PHYS_TYPES; p++) 652 ASSERT(dde->dde_lead_zio[p] == NULL); 653 654 if (dde->dde_repair_data != NULL) 655 zio_buf_free(dde->dde_repair_data, 656 DDK_GET_PSIZE(&dde->dde_key)); 657 658 cv_destroy(&dde->dde_cv); 659 kmem_free(dde, sizeof (*dde)); 660 } 661 662 void 663 ddt_remove(ddt_t *ddt, ddt_entry_t *dde) 664 { 665 ASSERT(MUTEX_HELD(&ddt->ddt_lock)); 666 667 avl_remove(&ddt->ddt_tree, dde); 668 ddt_free(dde); 669 } 670 671 ddt_entry_t * 672 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add) 673 { 674 ddt_entry_t *dde, dde_search; 675 enum ddt_type type; 676 enum ddt_class class; 677 avl_index_t where; 678 int error; 679 680 ASSERT(MUTEX_HELD(&ddt->ddt_lock)); 681 682 ddt_key_fill(&dde_search.dde_key, bp); 683 684 dde = avl_find(&ddt->ddt_tree, &dde_search, &where); 685 if (dde == NULL) { 686 if (!add) 687 return (NULL); 688 dde = ddt_alloc(&dde_search.dde_key); 689 avl_insert(&ddt->ddt_tree, dde, where); 690 } 691 692 while (dde->dde_loading) 693 cv_wait(&dde->dde_cv, &ddt->ddt_lock); 694 695 if (dde->dde_loaded) 696 return (dde); 697 698 dde->dde_loading = B_TRUE; 699 700 ddt_exit(ddt); 701 702 error = ENOENT; 703 704 for (type = 0; type < DDT_TYPES; type++) { 705 for (class = 0; class < DDT_CLASSES; class++) { 706 error = ddt_object_lookup(ddt, type, class, dde); 707 if (error != ENOENT) 708 break; 709 } 710 if (error != ENOENT) 711 break; 712 } 713 714 ASSERT(error == 0 || error == ENOENT); 715 716 ddt_enter(ddt); 717 718 ASSERT(dde->dde_loaded == B_FALSE); 719 ASSERT(dde->dde_loading == B_TRUE); 720 721 dde->dde_type = type; /* will be DDT_TYPES if no entry found */ 722 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */ 723 dde->dde_loaded = B_TRUE; 724 dde->dde_loading = B_FALSE; 725 726 if (error == 0) 727 ddt_stat_update(ddt, dde, -1ULL); 728 729 cv_broadcast(&dde->dde_cv); 730 731 return (dde); 732 } 733 734 void 735 ddt_prefetch(spa_t *spa, const blkptr_t *bp) 736 { 737 ddt_t *ddt; 738 ddt_entry_t dde; 739 740 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp)) 741 return; 742 743 /* 744 * We only remove the DDT once all tables are empty and only 745 * prefetch dedup blocks when there are entries in the DDT. 746 * Thus no locking is required as the DDT can't disappear on us. 747 */ 748 ddt = ddt_select(spa, bp); 749 ddt_key_fill(&dde.dde_key, bp); 750 751 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 752 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 753 ddt_object_prefetch(ddt, type, class, &dde); 754 } 755 } 756 } 757 758 int 759 ddt_entry_compare(const void *x1, const void *x2) 760 { 761 const ddt_entry_t *dde1 = x1; 762 const ddt_entry_t *dde2 = x2; 763 const uint64_t *u1 = (const uint64_t *)&dde1->dde_key; 764 const uint64_t *u2 = (const uint64_t *)&dde2->dde_key; 765 766 for (int i = 0; i < DDT_KEY_WORDS; i++) { 767 if (u1[i] < u2[i]) 768 return (-1); 769 if (u1[i] > u2[i]) 770 return (1); 771 } 772 773 return (0); 774 } 775 776 static ddt_t * 777 ddt_table_alloc(spa_t *spa, enum zio_checksum c) 778 { 779 ddt_t *ddt; 780 781 ddt = kmem_zalloc(sizeof (*ddt), KM_SLEEP); 782 783 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL); 784 avl_create(&ddt->ddt_tree, ddt_entry_compare, 785 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node)); 786 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare, 787 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node)); 788 ddt->ddt_checksum = c; 789 ddt->ddt_spa = spa; 790 ddt->ddt_os = spa->spa_meta_objset; 791 792 return (ddt); 793 } 794 795 static void 796 ddt_table_free(ddt_t *ddt) 797 { 798 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0); 799 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0); 800 avl_destroy(&ddt->ddt_tree); 801 avl_destroy(&ddt->ddt_repair_tree); 802 mutex_destroy(&ddt->ddt_lock); 803 kmem_free(ddt, sizeof (*ddt)); 804 } 805 806 void 807 ddt_create(spa_t *spa) 808 { 809 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM; 810 811 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) 812 spa->spa_ddt[c] = ddt_table_alloc(spa, c); 813 } 814 815 int 816 ddt_load(spa_t *spa) 817 { 818 int error; 819 820 ddt_create(spa); 821 822 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 823 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1, 824 &spa->spa_ddt_stat_object); 825 826 if (error) 827 return (error == ENOENT ? 0 : error); 828 829 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 830 ddt_t *ddt = spa->spa_ddt[c]; 831 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 832 for (enum ddt_class class = 0; class < DDT_CLASSES; 833 class++) { 834 error = ddt_object_load(ddt, type, class); 835 if (error != 0 && error != ENOENT) 836 return (error); 837 } 838 } 839 840 /* 841 * Seed the cached histograms. 842 */ 843 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache, 844 sizeof (ddt->ddt_histogram)); 845 } 846 847 return (0); 848 } 849 850 void 851 ddt_unload(spa_t *spa) 852 { 853 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 854 if (spa->spa_ddt[c]) { 855 ddt_table_free(spa->spa_ddt[c]); 856 spa->spa_ddt[c] = NULL; 857 } 858 } 859 } 860 861 boolean_t 862 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp) 863 { 864 ddt_t *ddt; 865 ddt_entry_t dde; 866 867 if (!BP_GET_DEDUP(bp)) 868 return (B_FALSE); 869 870 if (max_class == DDT_CLASS_UNIQUE) 871 return (B_TRUE); 872 873 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)]; 874 875 ddt_key_fill(&dde.dde_key, bp); 876 877 for (enum ddt_type type = 0; type < DDT_TYPES; type++) 878 for (enum ddt_class class = 0; class <= max_class; class++) 879 if (ddt_object_lookup(ddt, type, class, &dde) == 0) 880 return (B_TRUE); 881 882 return (B_FALSE); 883 } 884 885 ddt_entry_t * 886 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp) 887 { 888 ddt_key_t ddk; 889 ddt_entry_t *dde; 890 891 ddt_key_fill(&ddk, bp); 892 893 dde = ddt_alloc(&ddk); 894 895 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 896 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 897 /* 898 * We can only do repair if there are multiple copies 899 * of the block. For anything in the UNIQUE class, 900 * there's definitely only one copy, so don't even try. 901 */ 902 if (class != DDT_CLASS_UNIQUE && 903 ddt_object_lookup(ddt, type, class, dde) == 0) 904 return (dde); 905 } 906 } 907 908 bzero(dde->dde_phys, sizeof (dde->dde_phys)); 909 910 return (dde); 911 } 912 913 void 914 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde) 915 { 916 avl_index_t where; 917 918 ddt_enter(ddt); 919 920 if (dde->dde_repair_data != NULL && spa_writeable(ddt->ddt_spa) && 921 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL) 922 avl_insert(&ddt->ddt_repair_tree, dde, where); 923 else 924 ddt_free(dde); 925 926 ddt_exit(ddt); 927 } 928 929 static void 930 ddt_repair_entry_done(zio_t *zio) 931 { 932 ddt_entry_t *rdde = zio->io_private; 933 934 ddt_free(rdde); 935 } 936 937 static void 938 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio) 939 { 940 ddt_phys_t *ddp = dde->dde_phys; 941 ddt_phys_t *rddp = rdde->dde_phys; 942 ddt_key_t *ddk = &dde->dde_key; 943 ddt_key_t *rddk = &rdde->dde_key; 944 zio_t *zio; 945 blkptr_t blk; 946 947 zio = zio_null(rio, rio->io_spa, NULL, 948 ddt_repair_entry_done, rdde, rio->io_flags); 949 950 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) { 951 if (ddp->ddp_phys_birth == 0 || 952 ddp->ddp_phys_birth != rddp->ddp_phys_birth || 953 bcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva))) 954 continue; 955 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk); 956 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk, 957 rdde->dde_repair_data, DDK_GET_PSIZE(rddk), NULL, NULL, 958 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL)); 959 } 960 961 zio_nowait(zio); 962 } 963 964 static void 965 ddt_repair_table(ddt_t *ddt, zio_t *rio) 966 { 967 spa_t *spa = ddt->ddt_spa; 968 ddt_entry_t *dde, *rdde_next, *rdde; 969 avl_tree_t *t = &ddt->ddt_repair_tree; 970 blkptr_t blk; 971 972 if (spa_sync_pass(spa) > 1) 973 return; 974 975 ddt_enter(ddt); 976 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) { 977 rdde_next = AVL_NEXT(t, rdde); 978 avl_remove(&ddt->ddt_repair_tree, rdde); 979 ddt_exit(ddt); 980 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk); 981 dde = ddt_repair_start(ddt, &blk); 982 ddt_repair_entry(ddt, dde, rdde, rio); 983 ddt_repair_done(ddt, dde); 984 ddt_enter(ddt); 985 } 986 ddt_exit(ddt); 987 } 988 989 static void 990 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg) 991 { 992 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool; 993 ddt_phys_t *ddp = dde->dde_phys; 994 ddt_key_t *ddk = &dde->dde_key; 995 enum ddt_type otype = dde->dde_type; 996 enum ddt_type ntype = DDT_TYPE_CURRENT; 997 enum ddt_class oclass = dde->dde_class; 998 enum ddt_class nclass; 999 uint64_t total_refcnt = 0; 1000 1001 ASSERT(dde->dde_loaded); 1002 ASSERT(!dde->dde_loading); 1003 1004 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 1005 ASSERT(dde->dde_lead_zio[p] == NULL); 1006 ASSERT((int64_t)ddp->ddp_refcnt >= 0); 1007 if (ddp->ddp_phys_birth == 0) { 1008 ASSERT(ddp->ddp_refcnt == 0); 1009 continue; 1010 } 1011 if (p == DDT_PHYS_DITTO) { 1012 if (ddt_ditto_copies_needed(ddt, dde, NULL) == 0) 1013 ddt_phys_free(ddt, ddk, ddp, txg); 1014 continue; 1015 } 1016 if (ddp->ddp_refcnt == 0) 1017 ddt_phys_free(ddt, ddk, ddp, txg); 1018 total_refcnt += ddp->ddp_refcnt; 1019 } 1020 1021 if (dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth != 0) 1022 nclass = DDT_CLASS_DITTO; 1023 else if (total_refcnt > 1) 1024 nclass = DDT_CLASS_DUPLICATE; 1025 else 1026 nclass = DDT_CLASS_UNIQUE; 1027 1028 if (otype != DDT_TYPES && 1029 (otype != ntype || oclass != nclass || total_refcnt == 0)) { 1030 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0); 1031 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT); 1032 } 1033 1034 if (total_refcnt != 0) { 1035 dde->dde_type = ntype; 1036 dde->dde_class = nclass; 1037 ddt_stat_update(ddt, dde, 0); 1038 if (!ddt_object_exists(ddt, ntype, nclass)) 1039 ddt_object_create(ddt, ntype, nclass, tx); 1040 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0); 1041 1042 /* 1043 * If the class changes, the order that we scan this bp 1044 * changes. If it decreases, we could miss it, so 1045 * scan it right now. (This covers both class changing 1046 * while we are doing ddt_walk(), and when we are 1047 * traversing.) 1048 */ 1049 if (nclass < oclass) { 1050 dsl_scan_ddt_entry(dp->dp_scan, 1051 ddt->ddt_checksum, dde, tx); 1052 } 1053 } 1054 } 1055 1056 static void 1057 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg) 1058 { 1059 spa_t *spa = ddt->ddt_spa; 1060 ddt_entry_t *dde; 1061 void *cookie = NULL; 1062 1063 if (avl_numnodes(&ddt->ddt_tree) == 0) 1064 return; 1065 1066 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP); 1067 1068 if (spa->spa_ddt_stat_object == 0) { 1069 spa->spa_ddt_stat_object = zap_create_link(ddt->ddt_os, 1070 DMU_OT_DDT_STATS, DMU_POOL_DIRECTORY_OBJECT, 1071 DMU_POOL_DDT_STATS, tx); 1072 } 1073 1074 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) { 1075 ddt_sync_entry(ddt, dde, tx, txg); 1076 ddt_free(dde); 1077 } 1078 1079 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 1080 uint64_t count = 0; 1081 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 1082 if (ddt_object_exists(ddt, type, class)) { 1083 ddt_object_sync(ddt, type, class, tx); 1084 count += ddt_object_count(ddt, type, class); 1085 } 1086 } 1087 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 1088 if (count == 0 && ddt_object_exists(ddt, type, class)) 1089 ddt_object_destroy(ddt, type, class, tx); 1090 } 1091 } 1092 1093 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache, 1094 sizeof (ddt->ddt_histogram)); 1095 } 1096 1097 void 1098 ddt_sync(spa_t *spa, uint64_t txg) 1099 { 1100 dmu_tx_t *tx; 1101 zio_t *rio = zio_root(spa, NULL, NULL, 1102 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE); 1103 1104 ASSERT(spa_syncing_txg(spa) == txg); 1105 1106 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 1107 1108 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 1109 ddt_t *ddt = spa->spa_ddt[c]; 1110 if (ddt == NULL) 1111 continue; 1112 ddt_sync_table(ddt, tx, txg); 1113 ddt_repair_table(ddt, rio); 1114 } 1115 1116 (void) zio_wait(rio); 1117 1118 dmu_tx_commit(tx); 1119 } 1120 1121 int 1122 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde) 1123 { 1124 do { 1125 do { 1126 do { 1127 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum]; 1128 int error = ENOENT; 1129 if (ddt_object_exists(ddt, ddb->ddb_type, 1130 ddb->ddb_class)) { 1131 error = ddt_object_walk(ddt, 1132 ddb->ddb_type, ddb->ddb_class, 1133 &ddb->ddb_cursor, dde); 1134 } 1135 dde->dde_type = ddb->ddb_type; 1136 dde->dde_class = ddb->ddb_class; 1137 if (error == 0) 1138 return (0); 1139 if (error != ENOENT) 1140 return (error); 1141 ddb->ddb_cursor = 0; 1142 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS); 1143 ddb->ddb_checksum = 0; 1144 } while (++ddb->ddb_type < DDT_TYPES); 1145 ddb->ddb_type = 0; 1146 } while (++ddb->ddb_class < DDT_CLASSES); 1147 1148 return (SET_ERROR(ENOENT)); 1149 } 1150