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