module/zfs/space_map.c

   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 2008 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #include <sys/zfs_context.h>
  27 #include <sys/spa.h>
  28 #include <sys/dmu.h>
  29 #include <sys/zio.h>
  30 #include <sys/space_map.h>
  31
  32 /*
  33  * Space map routines.
  34  * NOTE: caller is responsible for all locking.
  35  */
  36 static int
  37 space_map_seg_compare(const void *x1, const void *x2)
  38 {
  39         const space_seg_t *s1 = x1;
  40         const space_seg_t *s2 = x2;
  41
  42         if (s1->ss_start < s2->ss_start) {
  43                 if (s1->ss_end > s2->ss_start)
  44                         return (0);
  45                 return (-1);
  46         }
  47         if (s1->ss_start > s2->ss_start) {
  48                 if (s1->ss_start < s2->ss_end)
  49                         return (0);
  50                 return (1);
  51         }
  52         return (0);
  53 }
  54
  55 void
  56 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
  57         kmutex_t *lp)
  58 {
  59         bzero(sm, sizeof (*sm));
  60
  61         cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
  62
  63         avl_create(&sm->sm_root, space_map_seg_compare,
  64             sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
  65
  66         sm->sm_start = start;
  67         sm->sm_size = size;
  68         sm->sm_shift = shift;
  69         sm->sm_lock = lp;
  70 }
  71
  72 void
  73 space_map_destroy(space_map_t *sm)
  74 {
  75         ASSERT(!sm->sm_loaded && !sm->sm_loading);
  76         VERIFY3U(sm->sm_space, ==, 0);
  77         avl_destroy(&sm->sm_root);
  78         cv_destroy(&sm->sm_load_cv);
  79 }
  80
  81 void
  82 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
  83 {
  84         avl_index_t where;
  85         space_seg_t ssearch, *ss_before, *ss_after, *ss;
  86         uint64_t end = start + size;
  87         int merge_before, merge_after;
  88
  89         ASSERT(MUTEX_HELD(sm->sm_lock));
  90         VERIFY(size != 0);
  91         VERIFY3U(start, >=, sm->sm_start);
  92         VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
  93         VERIFY(sm->sm_space + size <= sm->sm_size);
  94         VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
  95         VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
  96
  97         ssearch.ss_start = start;
  98         ssearch.ss_end = end;
  99         ss = avl_find(&sm->sm_root, &ssearch, &where);
 100
 101         if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
 102                 zfs_panic_recover("zfs: allocating allocated segment"
 103                     "(offset=%llu size=%llu)\n",
 104                     (longlong_t)start, (longlong_t)size);
 105                 return;
 106         }
 107
 108         /* Make sure we don't overlap with either of our neighbors */
 109         VERIFY(ss == NULL);
 110
 111         ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
 112         ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
 113
 114         merge_before = (ss_before != NULL && ss_before->ss_end == start);
 115         merge_after = (ss_after != NULL && ss_after->ss_start == end);
 116
 117         if (merge_before && merge_after) {
 118                 avl_remove(&sm->sm_root, ss_before);
 119                 ss_after->ss_start = ss_before->ss_start;
 120                 kmem_free(ss_before, sizeof (*ss_before));
 121         } else if (merge_before) {
 122                 ss_before->ss_end = end;
 123         } else if (merge_after) {
 124                 ss_after->ss_start = start;
 125         } else {
 126                 ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
 127                 ss->ss_start = start;
 128                 ss->ss_end = end;
 129                 avl_insert(&sm->sm_root, ss, where);
 130         }
 131
 132         sm->sm_space += size;
 133 }
 134
 135 void
 136 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
 137 {
 138         avl_index_t where;
 139         space_seg_t ssearch, *ss, *newseg;
 140         uint64_t end = start + size;
 141         int left_over, right_over;
 142
 143         ASSERT(MUTEX_HELD(sm->sm_lock));
 144         VERIFY(size != 0);
 145         VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
 146         VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
 147
 148         ssearch.ss_start = start;
 149         ssearch.ss_end = end;
 150         ss = avl_find(&sm->sm_root, &ssearch, &where);
 151
 152         /* Make sure we completely overlap with someone */
 153         if (ss == NULL) {
 154                 zfs_panic_recover("zfs: freeing free segment "
 155                     "(offset=%llu size=%llu)",
 156                     (longlong_t)start, (longlong_t)size);
 157                 return;
 158         }
 159         VERIFY3U(ss->ss_start, <=, start);
 160         VERIFY3U(ss->ss_end, >=, end);
 161         VERIFY(sm->sm_space - size <= sm->sm_size);
 162
 163         left_over = (ss->ss_start != start);
 164         right_over = (ss->ss_end != end);
 165
 166         if (left_over && right_over) {
 167                 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP);
 168                 newseg->ss_start = end;
 169                 newseg->ss_end = ss->ss_end;
 170                 ss->ss_end = start;
 171                 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
 172         } else if (left_over) {
 173                 ss->ss_end = start;
 174         } else if (right_over) {
 175                 ss->ss_start = end;
 176         } else {
 177                 avl_remove(&sm->sm_root, ss);
 178                 kmem_free(ss, sizeof (*ss));
 179         }
 180
 181         sm->sm_space -= size;
 182 }
 183
 184 boolean_t
 185 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
 186 {
 187         avl_index_t where;
 188         space_seg_t ssearch, *ss;
 189         uint64_t end = start + size;
 190
 191         ASSERT(MUTEX_HELD(sm->sm_lock));
 192         VERIFY(size != 0);
 193         VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
 194         VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
 195
 196         ssearch.ss_start = start;
 197         ssearch.ss_end = end;
 198         ss = avl_find(&sm->sm_root, &ssearch, &where);
 199
 200         return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
 201 }
 202
 203 void
 204 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
 205 {
 206         space_seg_t *ss;
 207         void *cookie = NULL;
 208
 209         ASSERT(MUTEX_HELD(sm->sm_lock));
 210
 211         while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
 212                 if (func != NULL)
 213                         func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
 214                 kmem_free(ss, sizeof (*ss));
 215         }
 216         sm->sm_space = 0;
 217 }
 218
 219 void
 220 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
 221 {
 222         space_seg_t *ss;
 223
 224         ASSERT(MUTEX_HELD(sm->sm_lock));
 225
 226         for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
 227                 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
 228 }
 229
 230 /*
 231  * Wait for any in-progress space_map_load() to complete.
 232  */
 233 void
 234 space_map_load_wait(space_map_t *sm)
 235 {
 236         ASSERT(MUTEX_HELD(sm->sm_lock));
 237
 238         while (sm->sm_loading)
 239                 cv_wait(&sm->sm_load_cv, sm->sm_lock);
 240 }
 241
 242 /*
 243  * Note: space_map_load() will drop sm_lock across dmu_read() calls.
 244  * The caller must be OK with this.
 245  */
 246 int
 247 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
 248         space_map_obj_t *smo, objset_t *os)
 249 {
 250         uint64_t *entry, *entry_map, *entry_map_end;
 251         uint64_t bufsize, size, offset, end, space;
 252         uint64_t mapstart = sm->sm_start;
 253         int error = 0;
 254
 255         ASSERT(MUTEX_HELD(sm->sm_lock));
 256
 257         space_map_load_wait(sm);
 258
 259         if (sm->sm_loaded)
 260                 return (0);
 261
 262         sm->sm_loading = B_TRUE;
 263         end = smo->smo_objsize;
 264         space = smo->smo_alloc;
 265
 266         ASSERT(sm->sm_ops == NULL);
 267         VERIFY3U(sm->sm_space, ==, 0);
 268
 269         if (maptype == SM_FREE) {
 270                 space_map_add(sm, sm->sm_start, sm->sm_size);
 271                 space = sm->sm_size - space;
 272         }
 273
 274         bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
 275         entry_map = zio_buf_alloc(bufsize);
 276
 277         mutex_exit(sm->sm_lock);
 278         if (end > bufsize)
 279                 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
 280         mutex_enter(sm->sm_lock);
 281
 282         for (offset = 0; offset < end; offset += bufsize) {
 283                 size = MIN(end - offset, bufsize);
 284                 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
 285                 VERIFY(size != 0);
 286
 287                 dprintf("object=%llu  offset=%llx  size=%llx\n",
 288                     smo->smo_object, offset, size);
 289
 290                 mutex_exit(sm->sm_lock);
 291                 error = dmu_read(os, smo->smo_object, offset, size, entry_map);
 292                 mutex_enter(sm->sm_lock);
 293                 if (error != 0)
 294                         break;
 295
 296                 entry_map_end = entry_map + (size / sizeof (uint64_t));
 297                 for (entry = entry_map; entry < entry_map_end; entry++) {
 298                         uint64_t e = *entry;
 299
 300                         if (SM_DEBUG_DECODE(e))         /* Skip debug entries */
 301                                 continue;
 302
 303                         (SM_TYPE_DECODE(e) == maptype ?
 304                             space_map_add : space_map_remove)(sm,
 305                             (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
 306                             SM_RUN_DECODE(e) << sm->sm_shift);
 307                 }
 308         }
 309
 310         if (error == 0) {
 311                 VERIFY3U(sm->sm_space, ==, space);
 312
 313                 sm->sm_loaded = B_TRUE;
 314                 sm->sm_ops = ops;
 315                 if (ops != NULL)
 316                         ops->smop_load(sm);
 317         } else {
 318                 space_map_vacate(sm, NULL, NULL);
 319         }
 320
 321         zio_buf_free(entry_map, bufsize);
 322
 323         sm->sm_loading = B_FALSE;
 324
 325         cv_broadcast(&sm->sm_load_cv);
 326
 327         return (error);
 328 }
 329
 330 void
 331 space_map_unload(space_map_t *sm)
 332 {
 333         ASSERT(MUTEX_HELD(sm->sm_lock));
 334
 335         if (sm->sm_loaded && sm->sm_ops != NULL)
 336                 sm->sm_ops->smop_unload(sm);
 337
 338         sm->sm_loaded = B_FALSE;
 339         sm->sm_ops = NULL;
 340
 341         space_map_vacate(sm, NULL, NULL);
 342 }
 343
 344 uint64_t
 345 space_map_alloc(space_map_t *sm, uint64_t size)
 346 {
 347         uint64_t start;
 348
 349         start = sm->sm_ops->smop_alloc(sm, size);
 350         if (start != -1ULL)
 351                 space_map_remove(sm, start, size);
 352         return (start);
 353 }
 354
 355 void
 356 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
 357 {
 358         sm->sm_ops->smop_claim(sm, start, size);
 359         space_map_remove(sm, start, size);
 360 }
 361
 362 void
 363 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
 364 {
 365         space_map_add(sm, start, size);
 366         sm->sm_ops->smop_free(sm, start, size);
 367 }
 368
 369 /*
 370  * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
 371  */
 372 void
 373 space_map_sync(space_map_t *sm, uint8_t maptype,
 374         space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
 375 {
 376         spa_t *spa = dmu_objset_spa(os);
 377         void *cookie = NULL;
 378         space_seg_t *ss;
 379         uint64_t bufsize, start, size, run_len;
 380         uint64_t *entry, *entry_map, *entry_map_end;
 381
 382         ASSERT(MUTEX_HELD(sm->sm_lock));
 383
 384         if (sm->sm_space == 0)
 385                 return;
 386
 387         dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
 388             smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
 389             maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
 390             sm->sm_space);
 391
 392         if (maptype == SM_ALLOC)
 393                 smo->smo_alloc += sm->sm_space;
 394         else
 395                 smo->smo_alloc -= sm->sm_space;
 396
 397         bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
 398         bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
 399         entry_map = zio_buf_alloc(bufsize);
 400         entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
 401         entry = entry_map;
 402
 403         *entry++ = SM_DEBUG_ENCODE(1) |
 404             SM_DEBUG_ACTION_ENCODE(maptype) |
 405             SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
 406             SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
 407
 408         while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
 409                 size = ss->ss_end - ss->ss_start;
 410                 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
 411
 412                 sm->sm_space -= size;
 413                 size >>= sm->sm_shift;
 414
 415                 while (size) {
 416                         run_len = MIN(size, SM_RUN_MAX);
 417
 418                         if (entry == entry_map_end) {
 419                                 mutex_exit(sm->sm_lock);
 420                                 dmu_write(os, smo->smo_object, smo->smo_objsize,
 421                                     bufsize, entry_map, tx);
 422                                 mutex_enter(sm->sm_lock);
 423                                 smo->smo_objsize += bufsize;
 424                                 entry = entry_map;
 425                         }
 426
 427                         *entry++ = SM_OFFSET_ENCODE(start) |
 428                             SM_TYPE_ENCODE(maptype) |
 429                             SM_RUN_ENCODE(run_len);
 430
 431                         start += run_len;
 432                         size -= run_len;
 433                 }
 434                 kmem_free(ss, sizeof (*ss));
 435         }
 436
 437         if (entry != entry_map) {
 438                 size = (entry - entry_map) * sizeof (uint64_t);
 439                 mutex_exit(sm->sm_lock);
 440                 dmu_write(os, smo->smo_object, smo->smo_objsize,
 441                     size, entry_map, tx);
 442                 mutex_enter(sm->sm_lock);
 443                 smo->smo_objsize += size;
 444         }
 445
 446         zio_buf_free(entry_map, bufsize);
 447
 448         VERIFY3U(sm->sm_space, ==, 0);
 449 }
 450
 451 void
 452 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
 453 {
 454         VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
 455
 456         smo->smo_objsize = 0;
 457         smo->smo_alloc = 0;
 458 }
 459
 460 /*
 461  * Space map reference trees.
 462  *
 463  * A space map is a collection of integers.  Every integer is either
 464  * in the map, or it's not.  A space map reference tree generalizes
 465  * the idea: it allows its members to have arbitrary reference counts,
 466  * as opposed to the implicit reference count of 0 or 1 in a space map.
 467  * This representation comes in handy when computing the union or
 468  * intersection of multiple space maps.  For example, the union of
 469  * N space maps is the subset of the reference tree with refcnt >= 1.
 470  * The intersection of N space maps is the subset with refcnt >= N.
 471  *
 472  * [It's very much like a Fourier transform.  Unions and intersections
 473  * are hard to perform in the 'space map domain', so we convert the maps
 474  * into the 'reference count domain', where it's trivial, then invert.]
 475  *
 476  * vdev_dtl_reassess() uses computations of this form to determine
 477  * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
 478  * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
 479  * has an outage wherever refcnt >= vdev_children.
 480  */
 481 static int
 482 space_map_ref_compare(const void *x1, const void *x2)
 483 {
 484         const space_ref_t *sr1 = x1;
 485         const space_ref_t *sr2 = x2;
 486
 487         if (sr1->sr_offset < sr2->sr_offset)
 488                 return (-1);
 489         if (sr1->sr_offset > sr2->sr_offset)
 490                 return (1);
 491
 492         if (sr1 < sr2)
 493                 return (-1);
 494         if (sr1 > sr2)
 495                 return (1);
 496
 497         return (0);
 498 }
 499
 500 void
 501 space_map_ref_create(avl_tree_t *t)
 502 {
 503         avl_create(t, space_map_ref_compare,
 504             sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
 505 }
 506
 507 void
 508 space_map_ref_destroy(avl_tree_t *t)
 509 {
 510         space_ref_t *sr;
 511         void *cookie = NULL;
 512
 513         while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
 514                 kmem_free(sr, sizeof (*sr));
 515
 516         avl_destroy(t);
 517 }
 518
 519 static void
 520 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
 521 {
 522         space_ref_t *sr;
 523
 524         sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
 525         sr->sr_offset = offset;
 526         sr->sr_refcnt = refcnt;
 527
 528         avl_add(t, sr);
 529 }
 530
 531 void
 532 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
 533         int64_t refcnt)
 534 {
 535         space_map_ref_add_node(t, start, refcnt);
 536         space_map_ref_add_node(t, end, -refcnt);
 537 }
 538
 539 /*
 540  * Convert (or add) a space map into a reference tree.
 541  */
 542 void
 543 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
 544 {
 545         space_seg_t *ss;
 546
 547         ASSERT(MUTEX_HELD(sm->sm_lock));
 548
 549         for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
 550                 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
 551 }
 552
 553 /*
 554  * Convert a reference tree into a space map.  The space map will contain
 555  * all members of the reference tree for which refcnt >= minref.
 556  */
 557 void
 558 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
 559 {
 560         uint64_t start = -1ULL;
 561         int64_t refcnt = 0;
 562         space_ref_t *sr;
 563
 564         ASSERT(MUTEX_HELD(sm->sm_lock));
 565
 566         space_map_vacate(sm, NULL, NULL);
 567
 568         for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
 569                 refcnt += sr->sr_refcnt;
 570                 if (refcnt >= minref) {
 571                         if (start == -1ULL) {
 572                                 start = sr->sr_offset;
 573                         }
 574                 } else {
 575                         if (start != -1ULL) {
 576                                 uint64_t end = sr->sr_offset;
 577                                 ASSERT(start <= end);
 578                                 if (end > start)
 579                                         space_map_add(sm, start, end - start);
 580                                 start = -1ULL;
 581                         }
 582                 }
 583         }
 584         ASSERT(refcnt == 0);
 585         ASSERT(start == -1ULL);
 586 }