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, 2016 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 #include <sys/abd.h> 40 41 /* 42 * Enable/disable prefetching of dedup-ed blocks which are going to be freed. 43 */ 44 int zfs_dedup_prefetch = 1; 45 46 static const ddt_ops_t *ddt_ops[DDT_TYPES] = { 47 &ddt_zap_ops, 48 }; 49 50 static const char *ddt_class_name[DDT_CLASSES] = { 51 "ditto", 52 "duplicate", 53 "unique", 54 }; 55 56 static void 57 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 58 dmu_tx_t *tx) 59 { 60 spa_t *spa = ddt->ddt_spa; 61 objset_t *os = ddt->ddt_os; 62 uint64_t *objectp = &ddt->ddt_object[type][class]; 63 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_flags & 64 ZCHECKSUM_FLAG_DEDUP; 65 char name[DDT_NAMELEN]; 66 67 ddt_object_name(ddt, type, class, name); 68 69 ASSERT(*objectp == 0); 70 VERIFY(ddt_ops[type]->ddt_op_create(os, objectp, tx, prehash) == 0); 71 ASSERT(*objectp != 0); 72 73 VERIFY(zap_add(os, DMU_POOL_DIRECTORY_OBJECT, name, 74 sizeof (uint64_t), 1, objectp, tx) == 0); 75 76 VERIFY(zap_add(os, spa->spa_ddt_stat_object, name, 77 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t), 78 &ddt->ddt_histogram[type][class], tx) == 0); 79 } 80 81 static void 82 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 83 dmu_tx_t *tx) 84 { 85 spa_t *spa = ddt->ddt_spa; 86 objset_t *os = ddt->ddt_os; 87 uint64_t *objectp = &ddt->ddt_object[type][class]; 88 char name[DDT_NAMELEN]; 89 90 ddt_object_name(ddt, type, class, name); 91 92 ASSERT(*objectp != 0); 93 ASSERT(ddt_object_count(ddt, type, class) == 0); 94 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class])); 95 VERIFY(zap_remove(os, DMU_POOL_DIRECTORY_OBJECT, name, tx) == 0); 96 VERIFY(zap_remove(os, spa->spa_ddt_stat_object, name, tx) == 0); 97 VERIFY(ddt_ops[type]->ddt_op_destroy(os, *objectp, tx) == 0); 98 bzero(&ddt->ddt_object_stats[type][class], sizeof (ddt_object_t)); 99 100 *objectp = 0; 101 } 102 103 static int 104 ddt_object_load(ddt_t *ddt, enum ddt_type type, enum ddt_class class) 105 { 106 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class]; 107 dmu_object_info_t doi; 108 char name[DDT_NAMELEN]; 109 int error; 110 111 ddt_object_name(ddt, type, class, name); 112 113 error = zap_lookup(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT, name, 114 sizeof (uint64_t), 1, &ddt->ddt_object[type][class]); 115 116 if (error != 0) 117 return (error); 118 119 VERIFY0(zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name, 120 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t), 121 &ddt->ddt_histogram[type][class])); 122 123 /* 124 * Seed the cached statistics. 125 */ 126 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0); 127 128 ddo->ddo_count = ddt_object_count(ddt, type, class); 129 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9; 130 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size; 131 132 return (0); 133 } 134 135 static void 136 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 137 dmu_tx_t *tx) 138 { 139 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class]; 140 dmu_object_info_t doi; 141 char name[DDT_NAMELEN]; 142 143 ddt_object_name(ddt, type, class, name); 144 145 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name, 146 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t), 147 &ddt->ddt_histogram[type][class], tx) == 0); 148 149 /* 150 * Cache DDT statistics; this is the only time they'll change. 151 */ 152 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0); 153 154 ddo->ddo_count = ddt_object_count(ddt, type, class); 155 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9; 156 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size; 157 } 158 159 static int 160 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 161 ddt_entry_t *dde) 162 { 163 if (!ddt_object_exists(ddt, type, class)) 164 return (SET_ERROR(ENOENT)); 165 166 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os, 167 ddt->ddt_object[type][class], dde)); 168 } 169 170 static void 171 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 172 ddt_entry_t *dde) 173 { 174 if (!ddt_object_exists(ddt, type, class)) 175 return; 176 177 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os, 178 ddt->ddt_object[type][class], dde); 179 } 180 181 int 182 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 183 ddt_entry_t *dde, dmu_tx_t *tx) 184 { 185 ASSERT(ddt_object_exists(ddt, type, class)); 186 187 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os, 188 ddt->ddt_object[type][class], dde, tx)); 189 } 190 191 static int 192 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 193 ddt_entry_t *dde, dmu_tx_t *tx) 194 { 195 ASSERT(ddt_object_exists(ddt, type, class)); 196 197 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os, 198 ddt->ddt_object[type][class], dde, tx)); 199 } 200 201 int 202 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 203 uint64_t *walk, ddt_entry_t *dde) 204 { 205 ASSERT(ddt_object_exists(ddt, type, class)); 206 207 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os, 208 ddt->ddt_object[type][class], dde, walk)); 209 } 210 211 uint64_t 212 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class) 213 { 214 ASSERT(ddt_object_exists(ddt, type, class)); 215 216 return (ddt_ops[type]->ddt_op_count(ddt->ddt_os, 217 ddt->ddt_object[type][class])); 218 } 219 220 int 221 ddt_object_info(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 222 dmu_object_info_t *doi) 223 { 224 if (!ddt_object_exists(ddt, type, class)) 225 return (SET_ERROR(ENOENT)); 226 227 return (dmu_object_info(ddt->ddt_os, ddt->ddt_object[type][class], 228 doi)); 229 } 230 231 boolean_t 232 ddt_object_exists(ddt_t *ddt, enum ddt_type type, enum ddt_class class) 233 { 234 return (!!ddt->ddt_object[type][class]); 235 } 236 237 void 238 ddt_object_name(ddt_t *ddt, enum ddt_type type, enum ddt_class class, 239 char *name) 240 { 241 (void) sprintf(name, DMU_POOL_DDT, 242 zio_checksum_table[ddt->ddt_checksum].ci_name, 243 ddt_ops[type]->ddt_op_name, ddt_class_name[class]); 244 } 245 246 void 247 ddt_bp_fill(const ddt_phys_t *ddp, blkptr_t *bp, uint64_t txg) 248 { 249 ASSERT(txg != 0); 250 251 for (int d = 0; d < SPA_DVAS_PER_BP; d++) 252 bp->blk_dva[d] = ddp->ddp_dva[d]; 253 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth); 254 } 255 256 /* 257 * The bp created via this function may be used for repairs and scrub, but it 258 * will be missing the salt / IV required to do a full decrypting read. 259 */ 260 void 261 ddt_bp_create(enum zio_checksum checksum, 262 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp) 263 { 264 BP_ZERO(bp); 265 266 if (ddp != NULL) 267 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth); 268 269 bp->blk_cksum = ddk->ddk_cksum; 270 271 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk)); 272 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk)); 273 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk)); 274 BP_SET_CRYPT(bp, DDK_GET_CRYPT(ddk)); 275 BP_SET_FILL(bp, 1); 276 BP_SET_CHECKSUM(bp, checksum); 277 BP_SET_TYPE(bp, DMU_OT_DEDUP); 278 BP_SET_LEVEL(bp, 0); 279 BP_SET_DEDUP(bp, 0); 280 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 281 } 282 283 void 284 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp) 285 { 286 ddk->ddk_cksum = bp->blk_cksum; 287 ddk->ddk_prop = 0; 288 289 ASSERT(BP_IS_ENCRYPTED(bp) || !BP_USES_CRYPT(bp)); 290 291 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp)); 292 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp)); 293 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp)); 294 DDK_SET_CRYPT(ddk, BP_USES_CRYPT(bp)); 295 } 296 297 void 298 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp) 299 { 300 ASSERT(ddp->ddp_phys_birth == 0); 301 302 for (int d = 0; d < SPA_DVAS_PER_BP; d++) 303 ddp->ddp_dva[d] = bp->blk_dva[d]; 304 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp); 305 } 306 307 void 308 ddt_phys_clear(ddt_phys_t *ddp) 309 { 310 bzero(ddp, sizeof (*ddp)); 311 } 312 313 void 314 ddt_phys_addref(ddt_phys_t *ddp) 315 { 316 ddp->ddp_refcnt++; 317 } 318 319 void 320 ddt_phys_decref(ddt_phys_t *ddp) 321 { 322 ASSERT((int64_t)ddp->ddp_refcnt > 0); 323 ddp->ddp_refcnt--; 324 } 325 326 void 327 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg) 328 { 329 blkptr_t blk; 330 331 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk); 332 ddt_phys_clear(ddp); 333 zio_free(ddt->ddt_spa, txg, &blk); 334 } 335 336 ddt_phys_t * 337 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp) 338 { 339 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys; 340 341 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 342 if (DVA_EQUAL(BP_IDENTITY(bp), &ddp->ddp_dva[0]) && 343 BP_PHYSICAL_BIRTH(bp) == ddp->ddp_phys_birth) 344 return (ddp); 345 } 346 return (NULL); 347 } 348 349 uint64_t 350 ddt_phys_total_refcnt(const ddt_entry_t *dde) 351 { 352 uint64_t refcnt = 0; 353 354 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) 355 refcnt += dde->dde_phys[p].ddp_refcnt; 356 357 return (refcnt); 358 } 359 360 static void 361 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds) 362 { 363 spa_t *spa = ddt->ddt_spa; 364 ddt_phys_t *ddp = dde->dde_phys; 365 ddt_key_t *ddk = &dde->dde_key; 366 uint64_t lsize = DDK_GET_LSIZE(ddk); 367 uint64_t psize = DDK_GET_PSIZE(ddk); 368 369 bzero(dds, sizeof (*dds)); 370 371 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 372 uint64_t dsize = 0; 373 uint64_t refcnt = ddp->ddp_refcnt; 374 375 if (ddp->ddp_phys_birth == 0) 376 continue; 377 378 for (int d = 0; d < DDE_GET_NDVAS(dde); d++) 379 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]); 380 381 dds->dds_blocks += 1; 382 dds->dds_lsize += lsize; 383 dds->dds_psize += psize; 384 dds->dds_dsize += dsize; 385 386 dds->dds_ref_blocks += refcnt; 387 dds->dds_ref_lsize += lsize * refcnt; 388 dds->dds_ref_psize += psize * refcnt; 389 dds->dds_ref_dsize += dsize * refcnt; 390 } 391 } 392 393 void 394 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg) 395 { 396 const uint64_t *s = (const uint64_t *)src; 397 uint64_t *d = (uint64_t *)dst; 398 uint64_t *d_end = (uint64_t *)(dst + 1); 399 400 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */ 401 402 while (d < d_end) 403 *d++ += (*s++ ^ neg) - neg; 404 } 405 406 static void 407 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg) 408 { 409 ddt_stat_t dds; 410 ddt_histogram_t *ddh; 411 int bucket; 412 413 ddt_stat_generate(ddt, dde, &dds); 414 415 bucket = highbit64(dds.dds_ref_blocks) - 1; 416 ASSERT(bucket >= 0); 417 418 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class]; 419 420 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg); 421 } 422 423 void 424 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src) 425 { 426 for (int h = 0; h < 64; h++) 427 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0); 428 } 429 430 void 431 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh) 432 { 433 bzero(dds, sizeof (*dds)); 434 435 for (int h = 0; h < 64; h++) 436 ddt_stat_add(dds, &ddh->ddh_stat[h], 0); 437 } 438 439 boolean_t 440 ddt_histogram_empty(const ddt_histogram_t *ddh) 441 { 442 const uint64_t *s = (const uint64_t *)ddh; 443 const uint64_t *s_end = (const uint64_t *)(ddh + 1); 444 445 while (s < s_end) 446 if (*s++ != 0) 447 return (B_FALSE); 448 449 return (B_TRUE); 450 } 451 452 void 453 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total) 454 { 455 /* Sum the statistics we cached in ddt_object_sync(). */ 456 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 457 ddt_t *ddt = spa->spa_ddt[c]; 458 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 459 for (enum ddt_class class = 0; class < DDT_CLASSES; 460 class++) { 461 ddt_object_t *ddo = 462 &ddt->ddt_object_stats[type][class]; 463 ddo_total->ddo_count += ddo->ddo_count; 464 ddo_total->ddo_dspace += ddo->ddo_dspace; 465 ddo_total->ddo_mspace += ddo->ddo_mspace; 466 } 467 } 468 } 469 470 /* ... and compute the averages. */ 471 if (ddo_total->ddo_count != 0) { 472 ddo_total->ddo_dspace /= ddo_total->ddo_count; 473 ddo_total->ddo_mspace /= ddo_total->ddo_count; 474 } 475 } 476 477 void 478 ddt_get_dedup_histogram(spa_t *spa, ddt_histogram_t *ddh) 479 { 480 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 481 ddt_t *ddt = spa->spa_ddt[c]; 482 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 483 for (enum ddt_class class = 0; class < DDT_CLASSES; 484 class++) { 485 ddt_histogram_add(ddh, 486 &ddt->ddt_histogram_cache[type][class]); 487 } 488 } 489 } 490 } 491 492 void 493 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total) 494 { 495 ddt_histogram_t *ddh_total; 496 497 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_SLEEP); 498 ddt_get_dedup_histogram(spa, ddh_total); 499 ddt_histogram_stat(dds_total, ddh_total); 500 kmem_free(ddh_total, sizeof (ddt_histogram_t)); 501 } 502 503 uint64_t 504 ddt_get_dedup_dspace(spa_t *spa) 505 { 506 ddt_stat_t dds_total = { 0 }; 507 508 ddt_get_dedup_stats(spa, &dds_total); 509 return (dds_total.dds_ref_dsize - dds_total.dds_dsize); 510 } 511 512 uint64_t 513 ddt_get_pool_dedup_ratio(spa_t *spa) 514 { 515 ddt_stat_t dds_total = { 0 }; 516 517 ddt_get_dedup_stats(spa, &dds_total); 518 if (dds_total.dds_dsize == 0) 519 return (100); 520 521 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize); 522 } 523 524 int 525 ddt_ditto_copies_needed(ddt_t *ddt, ddt_entry_t *dde, ddt_phys_t *ddp_willref) 526 { 527 spa_t *spa = ddt->ddt_spa; 528 uint64_t total_refcnt = 0; 529 uint64_t ditto = spa->spa_dedup_ditto; 530 int total_copies = 0; 531 int desired_copies = 0; 532 int copies_needed = 0; 533 534 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 535 ddt_phys_t *ddp = &dde->dde_phys[p]; 536 zio_t *zio = dde->dde_lead_zio[p]; 537 uint64_t refcnt = ddp->ddp_refcnt; /* committed refs */ 538 if (zio != NULL) 539 refcnt += zio->io_parent_count; /* pending refs */ 540 if (ddp == ddp_willref) 541 refcnt++; /* caller's ref */ 542 if (refcnt != 0) { 543 total_refcnt += refcnt; 544 total_copies += p; 545 } 546 } 547 548 if (ditto == 0 || ditto > UINT32_MAX) 549 ditto = UINT32_MAX; 550 551 if (total_refcnt >= 1) 552 desired_copies++; 553 if (total_refcnt >= ditto) 554 desired_copies++; 555 if (total_refcnt >= ditto * ditto) 556 desired_copies++; 557 558 copies_needed = MAX(desired_copies, total_copies) - total_copies; 559 560 /* encrypted blocks store their IV in DVA[2] */ 561 if (DDK_GET_CRYPT(&dde->dde_key)) 562 copies_needed = MIN(copies_needed, SPA_DVAS_PER_BP - 1); 563 564 return (copies_needed); 565 } 566 567 int 568 ddt_ditto_copies_present(ddt_entry_t *dde) 569 { 570 ddt_phys_t *ddp = &dde->dde_phys[DDT_PHYS_DITTO]; 571 dva_t *dva = ddp->ddp_dva; 572 int copies = 0 - DVA_GET_GANG(dva); 573 574 for (int d = 0; d < DDE_GET_NDVAS(dde); d++, dva++) 575 if (DVA_IS_VALID(dva)) 576 copies++; 577 578 ASSERT(copies >= 0 && copies < SPA_DVAS_PER_BP); 579 580 return (copies); 581 } 582 583 size_t 584 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len) 585 { 586 uchar_t *version = dst++; 587 int cpfunc = ZIO_COMPRESS_ZLE; 588 zio_compress_info_t *ci = &zio_compress_table[cpfunc]; 589 size_t c_len; 590 591 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */ 592 593 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level); 594 595 if (c_len == s_len) { 596 cpfunc = ZIO_COMPRESS_OFF; 597 bcopy(src, dst, s_len); 598 } 599 600 *version = cpfunc; 601 /* CONSTCOND */ 602 if (ZFS_HOST_BYTEORDER) 603 *version |= DDT_COMPRESS_BYTEORDER_MASK; 604 605 return (c_len + 1); 606 } 607 608 void 609 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len) 610 { 611 uchar_t version = *src++; 612 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK; 613 zio_compress_info_t *ci = &zio_compress_table[cpfunc]; 614 615 if (ci->ci_decompress != NULL) 616 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level); 617 else 618 bcopy(src, dst, d_len); 619 620 if (((version & DDT_COMPRESS_BYTEORDER_MASK) != 0) != 621 (ZFS_HOST_BYTEORDER != 0)) 622 byteswap_uint64_array(dst, d_len); 623 } 624 625 ddt_t * 626 ddt_select_by_checksum(spa_t *spa, enum zio_checksum c) 627 { 628 return (spa->spa_ddt[c]); 629 } 630 631 ddt_t * 632 ddt_select(spa_t *spa, const blkptr_t *bp) 633 { 634 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]); 635 } 636 637 void 638 ddt_enter(ddt_t *ddt) 639 { 640 mutex_enter(&ddt->ddt_lock); 641 } 642 643 void 644 ddt_exit(ddt_t *ddt) 645 { 646 mutex_exit(&ddt->ddt_lock); 647 } 648 649 static ddt_entry_t * 650 ddt_alloc(const ddt_key_t *ddk) 651 { 652 ddt_entry_t *dde; 653 654 dde = kmem_zalloc(sizeof (ddt_entry_t), KM_SLEEP); 655 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL); 656 657 dde->dde_key = *ddk; 658 659 return (dde); 660 } 661 662 static void 663 ddt_free(ddt_entry_t *dde) 664 { 665 ASSERT(!dde->dde_loading); 666 667 for (int p = 0; p < DDT_PHYS_TYPES; p++) 668 ASSERT(dde->dde_lead_zio[p] == NULL); 669 670 if (dde->dde_repair_abd != NULL) 671 abd_free(dde->dde_repair_abd); 672 673 cv_destroy(&dde->dde_cv); 674 kmem_free(dde, sizeof (*dde)); 675 } 676 677 void 678 ddt_remove(ddt_t *ddt, ddt_entry_t *dde) 679 { 680 ASSERT(MUTEX_HELD(&ddt->ddt_lock)); 681 682 avl_remove(&ddt->ddt_tree, dde); 683 ddt_free(dde); 684 } 685 686 ddt_entry_t * 687 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add) 688 { 689 ddt_entry_t *dde, dde_search; 690 enum ddt_type type; 691 enum ddt_class class; 692 avl_index_t where; 693 int error; 694 695 ASSERT(MUTEX_HELD(&ddt->ddt_lock)); 696 697 ddt_key_fill(&dde_search.dde_key, bp); 698 699 dde = avl_find(&ddt->ddt_tree, &dde_search, &where); 700 if (dde == NULL) { 701 if (!add) 702 return (NULL); 703 dde = ddt_alloc(&dde_search.dde_key); 704 avl_insert(&ddt->ddt_tree, dde, where); 705 } 706 707 while (dde->dde_loading) 708 cv_wait(&dde->dde_cv, &ddt->ddt_lock); 709 710 if (dde->dde_loaded) 711 return (dde); 712 713 dde->dde_loading = B_TRUE; 714 715 ddt_exit(ddt); 716 717 error = ENOENT; 718 719 for (type = 0; type < DDT_TYPES; type++) { 720 for (class = 0; class < DDT_CLASSES; class++) { 721 error = ddt_object_lookup(ddt, type, class, dde); 722 if (error != ENOENT) { 723 ASSERT0(error); 724 break; 725 } 726 } 727 if (error != ENOENT) 728 break; 729 } 730 731 ddt_enter(ddt); 732 733 ASSERT(dde->dde_loaded == B_FALSE); 734 ASSERT(dde->dde_loading == B_TRUE); 735 736 dde->dde_type = type; /* will be DDT_TYPES if no entry found */ 737 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */ 738 dde->dde_loaded = B_TRUE; 739 dde->dde_loading = B_FALSE; 740 741 if (error == 0) 742 ddt_stat_update(ddt, dde, -1ULL); 743 744 cv_broadcast(&dde->dde_cv); 745 746 return (dde); 747 } 748 749 void 750 ddt_prefetch(spa_t *spa, const blkptr_t *bp) 751 { 752 ddt_t *ddt; 753 ddt_entry_t dde; 754 755 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp)) 756 return; 757 758 /* 759 * We only remove the DDT once all tables are empty and only 760 * prefetch dedup blocks when there are entries in the DDT. 761 * Thus no locking is required as the DDT can't disappear on us. 762 */ 763 ddt = ddt_select(spa, bp); 764 ddt_key_fill(&dde.dde_key, bp); 765 766 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 767 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 768 ddt_object_prefetch(ddt, type, class, &dde); 769 } 770 } 771 } 772 773 /* 774 * Opaque struct used for ddt_key comparison 775 */ 776 #define DDT_KEY_CMP_LEN (sizeof (ddt_key_t) / sizeof (uint16_t)) 777 778 typedef struct ddt_key_cmp { 779 uint16_t u16[DDT_KEY_CMP_LEN]; 780 } ddt_key_cmp_t; 781 782 int 783 ddt_entry_compare(const void *x1, const void *x2) 784 { 785 const ddt_entry_t *dde1 = x1; 786 const ddt_entry_t *dde2 = x2; 787 const ddt_key_cmp_t *k1 = (const ddt_key_cmp_t *)&dde1->dde_key; 788 const ddt_key_cmp_t *k2 = (const ddt_key_cmp_t *)&dde2->dde_key; 789 int32_t cmp = 0; 790 791 for (int i = 0; i < DDT_KEY_CMP_LEN; i++) { 792 cmp = (int32_t)k1->u16[i] - (int32_t)k2->u16[i]; 793 if (likely(cmp)) 794 break; 795 } 796 797 return (AVL_ISIGN(cmp)); 798 } 799 800 static ddt_t * 801 ddt_table_alloc(spa_t *spa, enum zio_checksum c) 802 { 803 ddt_t *ddt; 804 805 ddt = kmem_zalloc(sizeof (*ddt), KM_SLEEP); 806 807 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL); 808 avl_create(&ddt->ddt_tree, ddt_entry_compare, 809 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node)); 810 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare, 811 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node)); 812 ddt->ddt_checksum = c; 813 ddt->ddt_spa = spa; 814 ddt->ddt_os = spa->spa_meta_objset; 815 816 return (ddt); 817 } 818 819 static void 820 ddt_table_free(ddt_t *ddt) 821 { 822 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0); 823 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0); 824 avl_destroy(&ddt->ddt_tree); 825 avl_destroy(&ddt->ddt_repair_tree); 826 mutex_destroy(&ddt->ddt_lock); 827 kmem_free(ddt, sizeof (*ddt)); 828 } 829 830 void 831 ddt_create(spa_t *spa) 832 { 833 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM; 834 835 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) 836 spa->spa_ddt[c] = ddt_table_alloc(spa, c); 837 } 838 839 int 840 ddt_load(spa_t *spa) 841 { 842 int error; 843 844 ddt_create(spa); 845 846 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 847 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1, 848 &spa->spa_ddt_stat_object); 849 850 if (error) 851 return (error == ENOENT ? 0 : error); 852 853 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 854 ddt_t *ddt = spa->spa_ddt[c]; 855 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 856 for (enum ddt_class class = 0; class < DDT_CLASSES; 857 class++) { 858 error = ddt_object_load(ddt, type, class); 859 if (error != 0 && error != ENOENT) 860 return (error); 861 } 862 } 863 864 /* 865 * Seed the cached histograms. 866 */ 867 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache, 868 sizeof (ddt->ddt_histogram)); 869 } 870 871 return (0); 872 } 873 874 void 875 ddt_unload(spa_t *spa) 876 { 877 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 878 if (spa->spa_ddt[c]) { 879 ddt_table_free(spa->spa_ddt[c]); 880 spa->spa_ddt[c] = NULL; 881 } 882 } 883 } 884 885 boolean_t 886 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp) 887 { 888 ddt_t *ddt; 889 ddt_entry_t dde; 890 891 if (!BP_GET_DEDUP(bp)) 892 return (B_FALSE); 893 894 if (max_class == DDT_CLASS_UNIQUE) 895 return (B_TRUE); 896 897 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)]; 898 899 ddt_key_fill(&dde.dde_key, bp); 900 901 for (enum ddt_type type = 0; type < DDT_TYPES; type++) 902 for (enum ddt_class class = 0; class <= max_class; class++) 903 if (ddt_object_lookup(ddt, type, class, &dde) == 0) 904 return (B_TRUE); 905 906 return (B_FALSE); 907 } 908 909 ddt_entry_t * 910 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp) 911 { 912 ddt_key_t ddk; 913 ddt_entry_t *dde; 914 915 ddt_key_fill(&ddk, bp); 916 917 dde = ddt_alloc(&ddk); 918 919 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 920 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 921 /* 922 * We can only do repair if there are multiple copies 923 * of the block. For anything in the UNIQUE class, 924 * there's definitely only one copy, so don't even try. 925 */ 926 if (class != DDT_CLASS_UNIQUE && 927 ddt_object_lookup(ddt, type, class, dde) == 0) 928 return (dde); 929 } 930 } 931 932 bzero(dde->dde_phys, sizeof (dde->dde_phys)); 933 934 return (dde); 935 } 936 937 void 938 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde) 939 { 940 avl_index_t where; 941 942 ddt_enter(ddt); 943 944 if (dde->dde_repair_abd != NULL && spa_writeable(ddt->ddt_spa) && 945 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL) 946 avl_insert(&ddt->ddt_repair_tree, dde, where); 947 else 948 ddt_free(dde); 949 950 ddt_exit(ddt); 951 } 952 953 static void 954 ddt_repair_entry_done(zio_t *zio) 955 { 956 ddt_entry_t *rdde = zio->io_private; 957 958 ddt_free(rdde); 959 } 960 961 static void 962 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio) 963 { 964 ddt_phys_t *ddp = dde->dde_phys; 965 ddt_phys_t *rddp = rdde->dde_phys; 966 ddt_key_t *ddk = &dde->dde_key; 967 ddt_key_t *rddk = &rdde->dde_key; 968 zio_t *zio; 969 blkptr_t blk; 970 971 zio = zio_null(rio, rio->io_spa, NULL, 972 ddt_repair_entry_done, rdde, rio->io_flags); 973 974 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) { 975 if (ddp->ddp_phys_birth == 0 || 976 ddp->ddp_phys_birth != rddp->ddp_phys_birth || 977 bcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva))) 978 continue; 979 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk); 980 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk, 981 rdde->dde_repair_abd, DDK_GET_PSIZE(rddk), NULL, NULL, 982 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL)); 983 } 984 985 zio_nowait(zio); 986 } 987 988 static void 989 ddt_repair_table(ddt_t *ddt, zio_t *rio) 990 { 991 spa_t *spa = ddt->ddt_spa; 992 ddt_entry_t *dde, *rdde_next, *rdde; 993 avl_tree_t *t = &ddt->ddt_repair_tree; 994 blkptr_t blk; 995 996 if (spa_sync_pass(spa) > 1) 997 return; 998 999 ddt_enter(ddt); 1000 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) { 1001 rdde_next = AVL_NEXT(t, rdde); 1002 avl_remove(&ddt->ddt_repair_tree, rdde); 1003 ddt_exit(ddt); 1004 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk); 1005 dde = ddt_repair_start(ddt, &blk); 1006 ddt_repair_entry(ddt, dde, rdde, rio); 1007 ddt_repair_done(ddt, dde); 1008 ddt_enter(ddt); 1009 } 1010 ddt_exit(ddt); 1011 } 1012 1013 static void 1014 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg) 1015 { 1016 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool; 1017 ddt_phys_t *ddp = dde->dde_phys; 1018 ddt_key_t *ddk = &dde->dde_key; 1019 enum ddt_type otype = dde->dde_type; 1020 enum ddt_type ntype = DDT_TYPE_CURRENT; 1021 enum ddt_class oclass = dde->dde_class; 1022 enum ddt_class nclass; 1023 uint64_t total_refcnt = 0; 1024 1025 ASSERT(dde->dde_loaded); 1026 ASSERT(!dde->dde_loading); 1027 1028 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 1029 ASSERT(dde->dde_lead_zio[p] == NULL); 1030 ASSERT((int64_t)ddp->ddp_refcnt >= 0); 1031 if (ddp->ddp_phys_birth == 0) { 1032 ASSERT(ddp->ddp_refcnt == 0); 1033 continue; 1034 } 1035 if (p == DDT_PHYS_DITTO) { 1036 if (ddt_ditto_copies_needed(ddt, dde, NULL) == 0) 1037 ddt_phys_free(ddt, ddk, ddp, txg); 1038 continue; 1039 } 1040 if (ddp->ddp_refcnt == 0) 1041 ddt_phys_free(ddt, ddk, ddp, txg); 1042 total_refcnt += ddp->ddp_refcnt; 1043 } 1044 1045 if (dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth != 0) 1046 nclass = DDT_CLASS_DITTO; 1047 else if (total_refcnt > 1) 1048 nclass = DDT_CLASS_DUPLICATE; 1049 else 1050 nclass = DDT_CLASS_UNIQUE; 1051 1052 if (otype != DDT_TYPES && 1053 (otype != ntype || oclass != nclass || total_refcnt == 0)) { 1054 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0); 1055 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT); 1056 } 1057 1058 if (total_refcnt != 0) { 1059 dde->dde_type = ntype; 1060 dde->dde_class = nclass; 1061 ddt_stat_update(ddt, dde, 0); 1062 if (!ddt_object_exists(ddt, ntype, nclass)) 1063 ddt_object_create(ddt, ntype, nclass, tx); 1064 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0); 1065 1066 /* 1067 * If the class changes, the order that we scan this bp 1068 * changes. If it decreases, we could miss it, so 1069 * scan it right now. (This covers both class changing 1070 * while we are doing ddt_walk(), and when we are 1071 * traversing.) 1072 */ 1073 if (nclass < oclass) { 1074 dsl_scan_ddt_entry(dp->dp_scan, 1075 ddt->ddt_checksum, dde, tx); 1076 } 1077 } 1078 } 1079 1080 static void 1081 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg) 1082 { 1083 spa_t *spa = ddt->ddt_spa; 1084 ddt_entry_t *dde; 1085 void *cookie = NULL; 1086 1087 if (avl_numnodes(&ddt->ddt_tree) == 0) 1088 return; 1089 1090 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP); 1091 1092 if (spa->spa_ddt_stat_object == 0) { 1093 spa->spa_ddt_stat_object = zap_create_link(ddt->ddt_os, 1094 DMU_OT_DDT_STATS, DMU_POOL_DIRECTORY_OBJECT, 1095 DMU_POOL_DDT_STATS, tx); 1096 } 1097 1098 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) { 1099 ddt_sync_entry(ddt, dde, tx, txg); 1100 ddt_free(dde); 1101 } 1102 1103 for (enum ddt_type type = 0; type < DDT_TYPES; type++) { 1104 uint64_t count = 0; 1105 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 1106 if (ddt_object_exists(ddt, type, class)) { 1107 ddt_object_sync(ddt, type, class, tx); 1108 count += ddt_object_count(ddt, type, class); 1109 } 1110 } 1111 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) { 1112 if (count == 0 && ddt_object_exists(ddt, type, class)) 1113 ddt_object_destroy(ddt, type, class, tx); 1114 } 1115 } 1116 1117 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache, 1118 sizeof (ddt->ddt_histogram)); 1119 } 1120 1121 void 1122 ddt_sync(spa_t *spa, uint64_t txg) 1123 { 1124 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan; 1125 dmu_tx_t *tx; 1126 zio_t *rio; 1127 1128 ASSERT(spa_syncing_txg(spa) == txg); 1129 1130 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 1131 1132 rio = zio_root(spa, NULL, NULL, 1133 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SELF_HEAL); 1134 1135 /* 1136 * This function may cause an immediate scan of ddt blocks (see 1137 * the comment above dsl_scan_ddt() for details). We set the 1138 * scan's root zio here so that we can wait for any scan IOs in 1139 * addition to the regular ddt IOs. 1140 */ 1141 ASSERT3P(scn->scn_zio_root, ==, NULL); 1142 scn->scn_zio_root = rio; 1143 1144 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) { 1145 ddt_t *ddt = spa->spa_ddt[c]; 1146 if (ddt == NULL) 1147 continue; 1148 ddt_sync_table(ddt, tx, txg); 1149 ddt_repair_table(ddt, rio); 1150 } 1151 1152 (void) zio_wait(rio); 1153 scn->scn_zio_root = NULL; 1154 1155 dmu_tx_commit(tx); 1156 } 1157 1158 int 1159 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde) 1160 { 1161 do { 1162 do { 1163 do { 1164 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum]; 1165 int error = ENOENT; 1166 if (ddt_object_exists(ddt, ddb->ddb_type, 1167 ddb->ddb_class)) { 1168 error = ddt_object_walk(ddt, 1169 ddb->ddb_type, ddb->ddb_class, 1170 &ddb->ddb_cursor, dde); 1171 } 1172 dde->dde_type = ddb->ddb_type; 1173 dde->dde_class = ddb->ddb_class; 1174 if (error == 0) 1175 return (0); 1176 if (error != ENOENT) 1177 return (error); 1178 ddb->ddb_cursor = 0; 1179 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS); 1180 ddb->ddb_checksum = 0; 1181 } while (++ddb->ddb_type < DDT_TYPES); 1182 ddb->ddb_type = 0; 1183 } while (++ddb->ddb_class < DDT_CLASSES); 1184 1185 return (SET_ERROR(ENOENT)); 1186 } 1187