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 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/zio.h> 33 #include <sys/space_map.h> 34 35 /* 36 * Space map routines. 37 * NOTE: caller is responsible for all locking. 38 */ 39 static int 40 space_map_seg_compare(const void *x1, const void *x2) 41 { 42 const space_seg_t *s1 = x1; 43 const space_seg_t *s2 = x2; 44 45 if (s1->ss_start < s2->ss_start) { 46 if (s1->ss_end > s2->ss_start) 47 return (0); 48 return (-1); 49 } 50 if (s1->ss_start > s2->ss_start) { 51 if (s1->ss_start < s2->ss_end) 52 return (0); 53 return (1); 54 } 55 return (0); 56 } 57 58 void 59 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift, 60 kmutex_t *lp) 61 { 62 bzero(sm, sizeof (*sm)); 63 64 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL); 65 66 avl_create(&sm->sm_root, space_map_seg_compare, 67 sizeof (space_seg_t), offsetof(struct space_seg, ss_node)); 68 69 sm->sm_start = start; 70 sm->sm_size = size; 71 sm->sm_shift = shift; 72 sm->sm_lock = lp; 73 } 74 75 void 76 space_map_destroy(space_map_t *sm) 77 { 78 ASSERT(!sm->sm_loaded && !sm->sm_loading); 79 VERIFY0(sm->sm_space); 80 avl_destroy(&sm->sm_root); 81 cv_destroy(&sm->sm_load_cv); 82 } 83 84 void 85 space_map_add(space_map_t *sm, uint64_t start, uint64_t size) 86 { 87 avl_index_t where; 88 space_seg_t ssearch, *ss_before, *ss_after, *ss; 89 uint64_t end = start + size; 90 int merge_before, merge_after; 91 92 ASSERT(MUTEX_HELD(sm->sm_lock)); 93 VERIFY(size != 0); 94 VERIFY3U(start, >=, sm->sm_start); 95 VERIFY3U(end, <=, sm->sm_start + sm->sm_size); 96 VERIFY(sm->sm_space + size <= sm->sm_size); 97 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0); 98 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0); 99 100 ssearch.ss_start = start; 101 ssearch.ss_end = end; 102 ss = avl_find(&sm->sm_root, &ssearch, &where); 103 104 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) { 105 zfs_panic_recover("zfs: allocating allocated segment" 106 "(offset=%llu size=%llu)\n", 107 (longlong_t)start, (longlong_t)size); 108 return; 109 } 110 111 /* Make sure we don't overlap with either of our neighbors */ 112 VERIFY(ss == NULL); 113 114 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE); 115 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER); 116 117 merge_before = (ss_before != NULL && ss_before->ss_end == start); 118 merge_after = (ss_after != NULL && ss_after->ss_start == end); 119 120 if (merge_before && merge_after) { 121 avl_remove(&sm->sm_root, ss_before); 122 if (sm->sm_pp_root) { 123 avl_remove(sm->sm_pp_root, ss_before); 124 avl_remove(sm->sm_pp_root, ss_after); 125 } 126 ss_after->ss_start = ss_before->ss_start; 127 kmem_free(ss_before, sizeof (*ss_before)); 128 ss = ss_after; 129 } else if (merge_before) { 130 ss_before->ss_end = end; 131 if (sm->sm_pp_root) 132 avl_remove(sm->sm_pp_root, ss_before); 133 ss = ss_before; 134 } else if (merge_after) { 135 ss_after->ss_start = start; 136 if (sm->sm_pp_root) 137 avl_remove(sm->sm_pp_root, ss_after); 138 ss = ss_after; 139 } else { 140 ss = kmem_alloc(sizeof (*ss), KM_SLEEP); 141 ss->ss_start = start; 142 ss->ss_end = end; 143 avl_insert(&sm->sm_root, ss, where); 144 } 145 146 if (sm->sm_pp_root) 147 avl_add(sm->sm_pp_root, ss); 148 149 sm->sm_space += size; 150 } 151 152 void 153 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size) 154 { 155 avl_index_t where; 156 space_seg_t ssearch, *ss, *newseg; 157 uint64_t end = start + size; 158 int left_over, right_over; 159 160 ASSERT(MUTEX_HELD(sm->sm_lock)); 161 VERIFY(size != 0); 162 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0); 163 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0); 164 165 ssearch.ss_start = start; 166 ssearch.ss_end = end; 167 ss = avl_find(&sm->sm_root, &ssearch, &where); 168 169 /* Make sure we completely overlap with someone */ 170 if (ss == NULL) { 171 zfs_panic_recover("zfs: freeing free segment " 172 "(offset=%llu size=%llu)", 173 (longlong_t)start, (longlong_t)size); 174 return; 175 } 176 VERIFY3U(ss->ss_start, <=, start); 177 VERIFY3U(ss->ss_end, >=, end); 178 VERIFY(sm->sm_space - size <= sm->sm_size); 179 180 left_over = (ss->ss_start != start); 181 right_over = (ss->ss_end != end); 182 183 if (sm->sm_pp_root) 184 avl_remove(sm->sm_pp_root, ss); 185 186 if (left_over && right_over) { 187 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP); 188 newseg->ss_start = end; 189 newseg->ss_end = ss->ss_end; 190 ss->ss_end = start; 191 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER); 192 if (sm->sm_pp_root) 193 avl_add(sm->sm_pp_root, newseg); 194 } else if (left_over) { 195 ss->ss_end = start; 196 } else if (right_over) { 197 ss->ss_start = end; 198 } else { 199 avl_remove(&sm->sm_root, ss); 200 kmem_free(ss, sizeof (*ss)); 201 ss = NULL; 202 } 203 204 if (sm->sm_pp_root && ss != NULL) 205 avl_add(sm->sm_pp_root, ss); 206 207 sm->sm_space -= size; 208 } 209 210 boolean_t 211 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size) 212 { 213 avl_index_t where; 214 space_seg_t ssearch, *ss; 215 uint64_t end = start + size; 216 217 ASSERT(MUTEX_HELD(sm->sm_lock)); 218 VERIFY(size != 0); 219 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0); 220 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0); 221 222 ssearch.ss_start = start; 223 ssearch.ss_end = end; 224 ss = avl_find(&sm->sm_root, &ssearch, &where); 225 226 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end); 227 } 228 229 void 230 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest) 231 { 232 space_seg_t *ss; 233 void *cookie = NULL; 234 235 ASSERT(MUTEX_HELD(sm->sm_lock)); 236 237 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) { 238 if (func != NULL) 239 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start); 240 kmem_free(ss, sizeof (*ss)); 241 } 242 sm->sm_space = 0; 243 } 244 245 void 246 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest) 247 { 248 space_seg_t *ss; 249 250 ASSERT(MUTEX_HELD(sm->sm_lock)); 251 252 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss)) 253 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start); 254 } 255 256 /* 257 * Wait for any in-progress space_map_load() to complete. 258 */ 259 void 260 space_map_load_wait(space_map_t *sm) 261 { 262 ASSERT(MUTEX_HELD(sm->sm_lock)); 263 264 while (sm->sm_loading) { 265 ASSERT(!sm->sm_loaded); 266 cv_wait(&sm->sm_load_cv, sm->sm_lock); 267 } 268 } 269 270 /* 271 * Note: space_map_load() will drop sm_lock across dmu_read() calls. 272 * The caller must be OK with this. 273 */ 274 int 275 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype, 276 space_map_obj_t *smo, objset_t *os) 277 { 278 uint64_t *entry, *entry_map, *entry_map_end; 279 uint64_t bufsize, size, offset, end, space; 280 uint64_t mapstart = sm->sm_start; 281 int error = 0; 282 283 ASSERT(MUTEX_HELD(sm->sm_lock)); 284 ASSERT(!sm->sm_loaded); 285 ASSERT(!sm->sm_loading); 286 287 sm->sm_loading = B_TRUE; 288 end = smo->smo_objsize; 289 space = smo->smo_alloc; 290 291 ASSERT(sm->sm_ops == NULL); 292 VERIFY0(sm->sm_space); 293 294 if (maptype == SM_FREE) { 295 space_map_add(sm, sm->sm_start, sm->sm_size); 296 space = sm->sm_size - space; 297 } 298 299 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT; 300 entry_map = zio_buf_alloc(bufsize); 301 302 mutex_exit(sm->sm_lock); 303 if (end > bufsize) 304 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize); 305 mutex_enter(sm->sm_lock); 306 307 for (offset = 0; offset < end; offset += bufsize) { 308 size = MIN(end - offset, bufsize); 309 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0); 310 VERIFY(size != 0); 311 312 dprintf("object=%llu offset=%llx size=%llx\n", 313 smo->smo_object, offset, size); 314 315 mutex_exit(sm->sm_lock); 316 error = dmu_read(os, smo->smo_object, offset, size, entry_map, 317 DMU_READ_PREFETCH); 318 mutex_enter(sm->sm_lock); 319 if (error != 0) 320 break; 321 322 entry_map_end = entry_map + (size / sizeof (uint64_t)); 323 for (entry = entry_map; entry < entry_map_end; entry++) { 324 uint64_t e = *entry; 325 326 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */ 327 continue; 328 329 (SM_TYPE_DECODE(e) == maptype ? 330 space_map_add : space_map_remove)(sm, 331 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart, 332 SM_RUN_DECODE(e) << sm->sm_shift); 333 } 334 } 335 336 if (error == 0) { 337 VERIFY3U(sm->sm_space, ==, space); 338 339 sm->sm_loaded = B_TRUE; 340 sm->sm_ops = ops; 341 if (ops != NULL) 342 ops->smop_load(sm); 343 } else { 344 space_map_vacate(sm, NULL, NULL); 345 } 346 347 zio_buf_free(entry_map, bufsize); 348 349 sm->sm_loading = B_FALSE; 350 351 cv_broadcast(&sm->sm_load_cv); 352 353 return (error); 354 } 355 356 void 357 space_map_unload(space_map_t *sm) 358 { 359 ASSERT(MUTEX_HELD(sm->sm_lock)); 360 361 if (sm->sm_loaded && sm->sm_ops != NULL) 362 sm->sm_ops->smop_unload(sm); 363 364 sm->sm_loaded = B_FALSE; 365 sm->sm_ops = NULL; 366 367 space_map_vacate(sm, NULL, NULL); 368 } 369 370 uint64_t 371 space_map_maxsize(space_map_t *sm) 372 { 373 ASSERT(sm->sm_ops != NULL); 374 return (sm->sm_ops->smop_max(sm)); 375 } 376 377 uint64_t 378 space_map_alloc(space_map_t *sm, uint64_t size) 379 { 380 uint64_t start; 381 382 start = sm->sm_ops->smop_alloc(sm, size); 383 if (start != -1ULL) 384 space_map_remove(sm, start, size); 385 return (start); 386 } 387 388 void 389 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size) 390 { 391 sm->sm_ops->smop_claim(sm, start, size); 392 space_map_remove(sm, start, size); 393 } 394 395 void 396 space_map_free(space_map_t *sm, uint64_t start, uint64_t size) 397 { 398 space_map_add(sm, start, size); 399 sm->sm_ops->smop_free(sm, start, size); 400 } 401 402 /* 403 * Note: space_map_sync() will drop sm_lock across dmu_write() calls. 404 */ 405 void 406 space_map_sync(space_map_t *sm, uint8_t maptype, 407 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx) 408 { 409 spa_t *spa = dmu_objset_spa(os); 410 void *cookie = NULL; 411 space_seg_t *ss; 412 uint64_t bufsize, start, size, run_len; 413 uint64_t *entry, *entry_map, *entry_map_end; 414 415 ASSERT(MUTEX_HELD(sm->sm_lock)); 416 417 if (sm->sm_space == 0) 418 return; 419 420 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n", 421 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa), 422 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root), 423 sm->sm_space); 424 425 if (maptype == SM_ALLOC) 426 smo->smo_alloc += sm->sm_space; 427 else 428 smo->smo_alloc -= sm->sm_space; 429 430 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t); 431 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT); 432 entry_map = zio_buf_alloc(bufsize); 433 entry_map_end = entry_map + (bufsize / sizeof (uint64_t)); 434 entry = entry_map; 435 436 *entry++ = SM_DEBUG_ENCODE(1) | 437 SM_DEBUG_ACTION_ENCODE(maptype) | 438 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) | 439 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx)); 440 441 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) { 442 size = ss->ss_end - ss->ss_start; 443 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift; 444 445 sm->sm_space -= size; 446 size >>= sm->sm_shift; 447 448 while (size) { 449 run_len = MIN(size, SM_RUN_MAX); 450 451 if (entry == entry_map_end) { 452 mutex_exit(sm->sm_lock); 453 dmu_write(os, smo->smo_object, smo->smo_objsize, 454 bufsize, entry_map, tx); 455 mutex_enter(sm->sm_lock); 456 smo->smo_objsize += bufsize; 457 entry = entry_map; 458 } 459 460 *entry++ = SM_OFFSET_ENCODE(start) | 461 SM_TYPE_ENCODE(maptype) | 462 SM_RUN_ENCODE(run_len); 463 464 start += run_len; 465 size -= run_len; 466 } 467 kmem_free(ss, sizeof (*ss)); 468 } 469 470 if (entry != entry_map) { 471 size = (entry - entry_map) * sizeof (uint64_t); 472 mutex_exit(sm->sm_lock); 473 dmu_write(os, smo->smo_object, smo->smo_objsize, 474 size, entry_map, tx); 475 mutex_enter(sm->sm_lock); 476 smo->smo_objsize += size; 477 } 478 479 zio_buf_free(entry_map, bufsize); 480 481 VERIFY0(sm->sm_space); 482 } 483 484 void 485 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx) 486 { 487 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0); 488 489 smo->smo_objsize = 0; 490 smo->smo_alloc = 0; 491 } 492 493 /* 494 * Space map reference trees. 495 * 496 * A space map is a collection of integers. Every integer is either 497 * in the map, or it's not. A space map reference tree generalizes 498 * the idea: it allows its members to have arbitrary reference counts, 499 * as opposed to the implicit reference count of 0 or 1 in a space map. 500 * This representation comes in handy when computing the union or 501 * intersection of multiple space maps. For example, the union of 502 * N space maps is the subset of the reference tree with refcnt >= 1. 503 * The intersection of N space maps is the subset with refcnt >= N. 504 * 505 * [It's very much like a Fourier transform. Unions and intersections 506 * are hard to perform in the 'space map domain', so we convert the maps 507 * into the 'reference count domain', where it's trivial, then invert.] 508 * 509 * vdev_dtl_reassess() uses computations of this form to determine 510 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev 511 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev 512 * has an outage wherever refcnt >= vdev_children. 513 */ 514 static int 515 space_map_ref_compare(const void *x1, const void *x2) 516 { 517 const space_ref_t *sr1 = x1; 518 const space_ref_t *sr2 = x2; 519 520 if (sr1->sr_offset < sr2->sr_offset) 521 return (-1); 522 if (sr1->sr_offset > sr2->sr_offset) 523 return (1); 524 525 if (sr1 < sr2) 526 return (-1); 527 if (sr1 > sr2) 528 return (1); 529 530 return (0); 531 } 532 533 void 534 space_map_ref_create(avl_tree_t *t) 535 { 536 avl_create(t, space_map_ref_compare, 537 sizeof (space_ref_t), offsetof(space_ref_t, sr_node)); 538 } 539 540 void 541 space_map_ref_destroy(avl_tree_t *t) 542 { 543 space_ref_t *sr; 544 void *cookie = NULL; 545 546 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL) 547 kmem_free(sr, sizeof (*sr)); 548 549 avl_destroy(t); 550 } 551 552 static void 553 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt) 554 { 555 space_ref_t *sr; 556 557 sr = kmem_alloc(sizeof (*sr), KM_SLEEP); 558 sr->sr_offset = offset; 559 sr->sr_refcnt = refcnt; 560 561 avl_add(t, sr); 562 } 563 564 void 565 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end, 566 int64_t refcnt) 567 { 568 space_map_ref_add_node(t, start, refcnt); 569 space_map_ref_add_node(t, end, -refcnt); 570 } 571 572 /* 573 * Convert (or add) a space map into a reference tree. 574 */ 575 void 576 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt) 577 { 578 space_seg_t *ss; 579 580 ASSERT(MUTEX_HELD(sm->sm_lock)); 581 582 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss)) 583 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt); 584 } 585 586 /* 587 * Convert a reference tree into a space map. The space map will contain 588 * all members of the reference tree for which refcnt >= minref. 589 */ 590 void 591 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref) 592 { 593 uint64_t start = -1ULL; 594 int64_t refcnt = 0; 595 space_ref_t *sr; 596 597 ASSERT(MUTEX_HELD(sm->sm_lock)); 598 599 space_map_vacate(sm, NULL, NULL); 600 601 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) { 602 refcnt += sr->sr_refcnt; 603 if (refcnt >= minref) { 604 if (start == -1ULL) { 605 start = sr->sr_offset; 606 } 607 } else { 608 if (start != -1ULL) { 609 uint64_t end = sr->sr_offset; 610 ASSERT(start <= end); 611 if (end > start) 612 space_map_add(sm, start, end - start); 613 start = -1ULL; 614 } 615 } 616 } 617 ASSERT(refcnt == 0); 618 ASSERT(start == -1ULL); 619 } 620