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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2013 by Delphix. All rights reserved.
  24  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  25  */
  26 
  27 #include <sys/zfs_context.h>
  28 #include <sys/dmu.h>
  29 #include <sys/dmu_tx.h>
  30 #include <sys/space_map.h>
  31 #include <sys/metaslab_impl.h>
  32 #include <sys/vdev_impl.h>
  33 #include <sys/zio.h>
  34 
  35 /*
  36  * Allow allocations to switch to gang blocks quickly. We do this to
  37  * avoid having to load lots of space_maps in a given txg. There are,
  38  * however, some cases where we want to avoid "fast" ganging and instead
  39  * we want to do an exhaustive search of all metaslabs on this device.
  40  * Currently we don't allow any gang, zil, or dump device related allocations
  41  * to "fast" gang.
  42  */
  43 #define CAN_FASTGANG(flags) \
  44         (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
  45         METASLAB_GANG_AVOID)))
  46 
  47 uint64_t metaslab_aliquot = 512ULL << 10;
  48 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;     /* force gang blocks */
  49 
  50 /*
  51  * The in-core space map representation is more compact than its on-disk form.
  52  * The zfs_condense_pct determines how much more compact the in-core
  53  * space_map representation must be before we compact it on-disk.
  54  * Values should be greater than or equal to 100.
  55  */
  56 int zfs_condense_pct = 200;
  57 
  58 /*
  59  * This value defines the number of allowed allocation failures per vdev.
  60  * If a device reaches this threshold in a given txg then we consider skipping
  61  * allocations on that device. The value of zfs_mg_alloc_failures is computed
  62  * in zio_init() unless it has been overridden in /etc/system.
  63  */
  64 int zfs_mg_alloc_failures = 0;
  65 
  66 /*
  67  * The zfs_mg_noalloc_threshold defines which metaslab groups should
  68  * be eligible for allocation. The value is defined as a percentage of
  69  * a free space. Metaslab groups that have more free space than
  70  * zfs_mg_noalloc_threshold are always eligible for allocations. Once
  71  * a metaslab group's free space is less than or equal to the
  72  * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
  73  * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
  74  * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
  75  * groups are allowed to accept allocations. Gang blocks are always
  76  * eligible to allocate on any metaslab group. The default value of 0 means
  77  * no metaslab group will be excluded based on this criterion.
  78  */
  79 int zfs_mg_noalloc_threshold = 0;
  80 
  81 /*
  82  * Metaslab debugging: when set, keeps all space maps in core to verify frees.
  83  */
  84 static int metaslab_debug = 0;
  85 
  86 /*
  87  * Minimum size which forces the dynamic allocator to change
  88  * it's allocation strategy.  Once the space map cannot satisfy
  89  * an allocation of this size then it switches to using more
  90  * aggressive strategy (i.e search by size rather than offset).
  91  */
  92 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
  93 
  94 /*
  95  * The minimum free space, in percent, which must be available
  96  * in a space map to continue allocations in a first-fit fashion.
  97  * Once the space_map's free space drops below this level we dynamically
  98  * switch to using best-fit allocations.
  99  */
 100 int metaslab_df_free_pct = 4;
 101 
 102 /*
 103  * A metaslab is considered "free" if it contains a contiguous
 104  * segment which is greater than metaslab_min_alloc_size.
 105  */
 106 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
 107 
 108 /*
 109  * Max number of space_maps to prefetch.
 110  */
 111 int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
 112 
 113 /*
 114  * Percentage bonus multiplier for metaslabs that are in the bonus area.
 115  */
 116 int metaslab_smo_bonus_pct = 150;
 117 
 118 /*
 119  * Should we be willing to write data to degraded vdevs?
 120  */
 121 boolean_t zfs_write_to_degraded = B_FALSE;
 122 
 123 /*
 124  * ==========================================================================
 125  * Metaslab classes
 126  * ==========================================================================
 127  */
 128 metaslab_class_t *
 129 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
 130 {
 131         metaslab_class_t *mc;
 132 
 133         mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
 134 
 135         mc->mc_spa = spa;
 136         mc->mc_rotor = NULL;
 137         mc->mc_ops = ops;
 138 
 139         return (mc);
 140 }
 141 
 142 void
 143 metaslab_class_destroy(metaslab_class_t *mc)
 144 {
 145         ASSERT(mc->mc_rotor == NULL);
 146         ASSERT(mc->mc_alloc == 0);
 147         ASSERT(mc->mc_deferred == 0);
 148         ASSERT(mc->mc_space == 0);
 149         ASSERT(mc->mc_dspace == 0);
 150 
 151         kmem_free(mc, sizeof (metaslab_class_t));
 152 }
 153 
 154 int
 155 metaslab_class_validate(metaslab_class_t *mc)
 156 {
 157         metaslab_group_t *mg;
 158         vdev_t *vd;
 159 
 160         /*
 161          * Must hold one of the spa_config locks.
 162          */
 163         ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
 164             spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
 165 
 166         if ((mg = mc->mc_rotor) == NULL)
 167                 return (0);
 168 
 169         do {
 170                 vd = mg->mg_vd;
 171                 ASSERT(vd->vdev_mg != NULL);
 172                 ASSERT3P(vd->vdev_top, ==, vd);
 173                 ASSERT3P(mg->mg_class, ==, mc);
 174                 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
 175         } while ((mg = mg->mg_next) != mc->mc_rotor);
 176 
 177         return (0);
 178 }
 179 
 180 void
 181 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
 182     int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
 183 {
 184         atomic_add_64(&mc->mc_alloc, alloc_delta);
 185         atomic_add_64(&mc->mc_deferred, defer_delta);
 186         atomic_add_64(&mc->mc_space, space_delta);
 187         atomic_add_64(&mc->mc_dspace, dspace_delta);
 188 }
 189 
 190 uint64_t
 191 metaslab_class_get_alloc(metaslab_class_t *mc)
 192 {
 193         return (mc->mc_alloc);
 194 }
 195 
 196 uint64_t
 197 metaslab_class_get_deferred(metaslab_class_t *mc)
 198 {
 199         return (mc->mc_deferred);
 200 }
 201 
 202 uint64_t
 203 metaslab_class_get_space(metaslab_class_t *mc)
 204 {
 205         return (mc->mc_space);
 206 }
 207 
 208 uint64_t
 209 metaslab_class_get_dspace(metaslab_class_t *mc)
 210 {
 211         return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
 212 }
 213 
 214 /*
 215  * ==========================================================================
 216  * Metaslab groups
 217  * ==========================================================================
 218  */
 219 static int
 220 metaslab_compare(const void *x1, const void *x2)
 221 {
 222         const metaslab_t *m1 = x1;
 223         const metaslab_t *m2 = x2;
 224 
 225         if (m1->ms_weight < m2->ms_weight)
 226                 return (1);
 227         if (m1->ms_weight > m2->ms_weight)
 228                 return (-1);
 229 
 230         /*
 231          * If the weights are identical, use the offset to force uniqueness.
 232          */
 233         if (m1->ms_map->sm_start < m2->ms_map->sm_start)
 234                 return (-1);
 235         if (m1->ms_map->sm_start > m2->ms_map->sm_start)
 236                 return (1);
 237 
 238         ASSERT3P(m1, ==, m2);
 239 
 240         return (0);
 241 }
 242 
 243 /*
 244  * Update the allocatable flag and the metaslab group's capacity.
 245  * The allocatable flag is set to true if the capacity is below
 246  * the zfs_mg_noalloc_threshold. If a metaslab group transitions
 247  * from allocatable to non-allocatable or vice versa then the metaslab
 248  * group's class is updated to reflect the transition.
 249  */
 250 static void
 251 metaslab_group_alloc_update(metaslab_group_t *mg)
 252 {
 253         vdev_t *vd = mg->mg_vd;
 254         metaslab_class_t *mc = mg->mg_class;
 255         vdev_stat_t *vs = &vd->vdev_stat;
 256         boolean_t was_allocatable;
 257 
 258         ASSERT(vd == vd->vdev_top);
 259 
 260         mutex_enter(&mg->mg_lock);
 261         was_allocatable = mg->mg_allocatable;
 262 
 263         mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
 264             (vs->vs_space + 1);
 265 
 266         mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold);
 267 
 268         /*
 269          * The mc_alloc_groups maintains a count of the number of
 270          * groups in this metaslab class that are still above the
 271          * zfs_mg_noalloc_threshold. This is used by the allocating
 272          * threads to determine if they should avoid allocations to
 273          * a given group. The allocator will avoid allocations to a group
 274          * if that group has reached or is below the zfs_mg_noalloc_threshold
 275          * and there are still other groups that are above the threshold.
 276          * When a group transitions from allocatable to non-allocatable or
 277          * vice versa we update the metaslab class to reflect that change.
 278          * When the mc_alloc_groups value drops to 0 that means that all
 279          * groups have reached the zfs_mg_noalloc_threshold making all groups
 280          * eligible for allocations. This effectively means that all devices
 281          * are balanced again.
 282          */
 283         if (was_allocatable && !mg->mg_allocatable)
 284                 mc->mc_alloc_groups--;
 285         else if (!was_allocatable && mg->mg_allocatable)
 286                 mc->mc_alloc_groups++;
 287         mutex_exit(&mg->mg_lock);
 288 }
 289 
 290 metaslab_group_t *
 291 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
 292 {
 293         metaslab_group_t *mg;
 294 
 295         mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
 296         mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
 297         avl_create(&mg->mg_metaslab_tree, metaslab_compare,
 298             sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
 299         mg->mg_vd = vd;
 300         mg->mg_class = mc;
 301         mg->mg_activation_count = 0;
 302 
 303         return (mg);
 304 }
 305 
 306 void
 307 metaslab_group_destroy(metaslab_group_t *mg)
 308 {
 309         ASSERT(mg->mg_prev == NULL);
 310         ASSERT(mg->mg_next == NULL);
 311         /*
 312          * We may have gone below zero with the activation count
 313          * either because we never activated in the first place or
 314          * because we're done, and possibly removing the vdev.
 315          */
 316         ASSERT(mg->mg_activation_count <= 0);
 317 
 318         avl_destroy(&mg->mg_metaslab_tree);
 319         mutex_destroy(&mg->mg_lock);
 320         kmem_free(mg, sizeof (metaslab_group_t));
 321 }
 322 
 323 void
 324 metaslab_group_activate(metaslab_group_t *mg)
 325 {
 326         metaslab_class_t *mc = mg->mg_class;
 327         metaslab_group_t *mgprev, *mgnext;
 328 
 329         ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
 330 
 331         ASSERT(mc->mc_rotor != mg);
 332         ASSERT(mg->mg_prev == NULL);
 333         ASSERT(mg->mg_next == NULL);
 334         ASSERT(mg->mg_activation_count <= 0);
 335 
 336         if (++mg->mg_activation_count <= 0)
 337                 return;
 338 
 339         mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
 340         metaslab_group_alloc_update(mg);
 341 
 342         if ((mgprev = mc->mc_rotor) == NULL) {
 343                 mg->mg_prev = mg;
 344                 mg->mg_next = mg;
 345         } else {
 346                 mgnext = mgprev->mg_next;
 347                 mg->mg_prev = mgprev;
 348                 mg->mg_next = mgnext;
 349                 mgprev->mg_next = mg;
 350                 mgnext->mg_prev = mg;
 351         }
 352         mc->mc_rotor = mg;
 353 }
 354 
 355 void
 356 metaslab_group_passivate(metaslab_group_t *mg)
 357 {
 358         metaslab_class_t *mc = mg->mg_class;
 359         metaslab_group_t *mgprev, *mgnext;
 360 
 361         ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
 362 
 363         if (--mg->mg_activation_count != 0) {
 364                 ASSERT(mc->mc_rotor != mg);
 365                 ASSERT(mg->mg_prev == NULL);
 366                 ASSERT(mg->mg_next == NULL);
 367                 ASSERT(mg->mg_activation_count < 0);
 368                 return;
 369         }
 370 
 371         mgprev = mg->mg_prev;
 372         mgnext = mg->mg_next;
 373 
 374         if (mg == mgnext) {
 375                 mc->mc_rotor = NULL;
 376         } else {
 377                 mc->mc_rotor = mgnext;
 378                 mgprev->mg_next = mgnext;
 379                 mgnext->mg_prev = mgprev;
 380         }
 381 
 382         mg->mg_prev = NULL;
 383         mg->mg_next = NULL;
 384 }
 385 
 386 static void
 387 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
 388 {
 389         mutex_enter(&mg->mg_lock);
 390         ASSERT(msp->ms_group == NULL);
 391         msp->ms_group = mg;
 392         msp->ms_weight = 0;
 393         avl_add(&mg->mg_metaslab_tree, msp);
 394         mutex_exit(&mg->mg_lock);
 395 }
 396 
 397 static void
 398 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
 399 {
 400         mutex_enter(&mg->mg_lock);
 401         ASSERT(msp->ms_group == mg);
 402         avl_remove(&mg->mg_metaslab_tree, msp);
 403         msp->ms_group = NULL;
 404         mutex_exit(&mg->mg_lock);
 405 }
 406 
 407 static void
 408 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
 409 {
 410         /*
 411          * Although in principle the weight can be any value, in
 412          * practice we do not use values in the range [1, 510].
 413          */
 414         ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
 415         ASSERT(MUTEX_HELD(&msp->ms_lock));
 416 
 417         mutex_enter(&mg->mg_lock);
 418         ASSERT(msp->ms_group == mg);
 419         avl_remove(&mg->mg_metaslab_tree, msp);
 420         msp->ms_weight = weight;
 421         avl_add(&mg->mg_metaslab_tree, msp);
 422         mutex_exit(&mg->mg_lock);
 423 }
 424 
 425 /*
 426  * Determine if a given metaslab group should skip allocations. A metaslab
 427  * group should avoid allocations if its used capacity has crossed the
 428  * zfs_mg_noalloc_threshold and there is at least one metaslab group
 429  * that can still handle allocations.
 430  */
 431 static boolean_t
 432 metaslab_group_allocatable(metaslab_group_t *mg)
 433 {
 434         vdev_t *vd = mg->mg_vd;
 435         spa_t *spa = vd->vdev_spa;
 436         metaslab_class_t *mc = mg->mg_class;
 437 
 438         /*
 439          * A metaslab group is considered allocatable if its free capacity
 440          * is greater than the set value of zfs_mg_noalloc_threshold, it's
 441          * associated with a slog, or there are no other metaslab groups
 442          * with free capacity greater than zfs_mg_noalloc_threshold.
 443          */
 444         return (mg->mg_free_capacity > zfs_mg_noalloc_threshold ||
 445             mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
 446 }
 447 
 448 /*
 449  * ==========================================================================
 450  * Common allocator routines
 451  * ==========================================================================
 452  */
 453 static int
 454 metaslab_segsize_compare(const void *x1, const void *x2)
 455 {
 456         const space_seg_t *s1 = x1;
 457         const space_seg_t *s2 = x2;
 458         uint64_t ss_size1 = s1->ss_end - s1->ss_start;
 459         uint64_t ss_size2 = s2->ss_end - s2->ss_start;
 460 
 461         if (ss_size1 < ss_size2)
 462                 return (-1);
 463         if (ss_size1 > ss_size2)
 464                 return (1);
 465 
 466         if (s1->ss_start < s2->ss_start)
 467                 return (-1);
 468         if (s1->ss_start > s2->ss_start)
 469                 return (1);
 470 
 471         return (0);
 472 }
 473 
 474 /*
 475  * This is a helper function that can be used by the allocator to find
 476  * a suitable block to allocate. This will search the specified AVL
 477  * tree looking for a block that matches the specified criteria.
 478  */
 479 static uint64_t
 480 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
 481     uint64_t align)
 482 {
 483         space_seg_t *ss, ssearch;
 484         avl_index_t where;
 485 
 486         ssearch.ss_start = *cursor;
 487         ssearch.ss_end = *cursor + size;
 488 
 489         ss = avl_find(t, &ssearch, &where);
 490         if (ss == NULL)
 491                 ss = avl_nearest(t, where, AVL_AFTER);
 492 
 493         while (ss != NULL) {
 494                 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
 495 
 496                 if (offset + size <= ss->ss_end) {
 497                         *cursor = offset + size;
 498                         return (offset);
 499                 }
 500                 ss = AVL_NEXT(t, ss);
 501         }
 502 
 503         /*
 504          * If we know we've searched the whole map (*cursor == 0), give up.
 505          * Otherwise, reset the cursor to the beginning and try again.
 506          */
 507         if (*cursor == 0)
 508                 return (-1ULL);
 509 
 510         *cursor = 0;
 511         return (metaslab_block_picker(t, cursor, size, align));
 512 }
 513 
 514 static void
 515 metaslab_pp_load(space_map_t *sm)
 516 {
 517         space_seg_t *ss;
 518 
 519         ASSERT(sm->sm_ppd == NULL);
 520         sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
 521 
 522         sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
 523         avl_create(sm->sm_pp_root, metaslab_segsize_compare,
 524             sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
 525 
 526         for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
 527                 avl_add(sm->sm_pp_root, ss);
 528 }
 529 
 530 static void
 531 metaslab_pp_unload(space_map_t *sm)
 532 {
 533         void *cookie = NULL;
 534 
 535         kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
 536         sm->sm_ppd = NULL;
 537 
 538         while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
 539                 /* tear down the tree */
 540         }
 541 
 542         avl_destroy(sm->sm_pp_root);
 543         kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
 544         sm->sm_pp_root = NULL;
 545 }
 546 
 547 /* ARGSUSED */
 548 static void
 549 metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
 550 {
 551         /* No need to update cursor */
 552 }
 553 
 554 /* ARGSUSED */
 555 static void
 556 metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
 557 {
 558         /* No need to update cursor */
 559 }
 560 
 561 /*
 562  * Return the maximum contiguous segment within the metaslab.
 563  */
 564 uint64_t
 565 metaslab_pp_maxsize(space_map_t *sm)
 566 {
 567         avl_tree_t *t = sm->sm_pp_root;
 568         space_seg_t *ss;
 569 
 570         if (t == NULL || (ss = avl_last(t)) == NULL)
 571                 return (0ULL);
 572 
 573         return (ss->ss_end - ss->ss_start);
 574 }
 575 
 576 /*
 577  * ==========================================================================
 578  * The first-fit block allocator
 579  * ==========================================================================
 580  */
 581 static uint64_t
 582 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
 583 {
 584         avl_tree_t *t = &sm->sm_root;
 585         uint64_t align = size & -size;
 586         uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
 587 
 588         return (metaslab_block_picker(t, cursor, size, align));
 589 }
 590 
 591 /* ARGSUSED */
 592 boolean_t
 593 metaslab_ff_fragmented(space_map_t *sm)
 594 {
 595         return (B_TRUE);
 596 }
 597 
 598 static space_map_ops_t metaslab_ff_ops = {
 599         metaslab_pp_load,
 600         metaslab_pp_unload,
 601         metaslab_ff_alloc,
 602         metaslab_pp_claim,
 603         metaslab_pp_free,
 604         metaslab_pp_maxsize,
 605         metaslab_ff_fragmented
 606 };
 607 
 608 /*
 609  * ==========================================================================
 610  * Dynamic block allocator -
 611  * Uses the first fit allocation scheme until space get low and then
 612  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
 613  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
 614  * ==========================================================================
 615  */
 616 static uint64_t
 617 metaslab_df_alloc(space_map_t *sm, uint64_t size)
 618 {
 619         avl_tree_t *t = &sm->sm_root;
 620         uint64_t align = size & -size;
 621         uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
 622         uint64_t max_size = metaslab_pp_maxsize(sm);
 623         int free_pct = sm->sm_space * 100 / sm->sm_size;
 624 
 625         ASSERT(MUTEX_HELD(sm->sm_lock));
 626         ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
 627 
 628         if (max_size < size)
 629                 return (-1ULL);
 630 
 631         /*
 632          * If we're running low on space switch to using the size
 633          * sorted AVL tree (best-fit).
 634          */
 635         if (max_size < metaslab_df_alloc_threshold ||
 636             free_pct < metaslab_df_free_pct) {
 637                 t = sm->sm_pp_root;
 638                 *cursor = 0;
 639         }
 640 
 641         return (metaslab_block_picker(t, cursor, size, 1ULL));
 642 }
 643 
 644 static boolean_t
 645 metaslab_df_fragmented(space_map_t *sm)
 646 {
 647         uint64_t max_size = metaslab_pp_maxsize(sm);
 648         int free_pct = sm->sm_space * 100 / sm->sm_size;
 649 
 650         if (max_size >= metaslab_df_alloc_threshold &&
 651             free_pct >= metaslab_df_free_pct)
 652                 return (B_FALSE);
 653 
 654         return (B_TRUE);
 655 }
 656 
 657 static space_map_ops_t metaslab_df_ops = {
 658         metaslab_pp_load,
 659         metaslab_pp_unload,
 660         metaslab_df_alloc,
 661         metaslab_pp_claim,
 662         metaslab_pp_free,
 663         metaslab_pp_maxsize,
 664         metaslab_df_fragmented
 665 };
 666 
 667 /*
 668  * ==========================================================================
 669  * Other experimental allocators
 670  * ==========================================================================
 671  */
 672 static uint64_t
 673 metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
 674 {
 675         avl_tree_t *t = &sm->sm_root;
 676         uint64_t *cursor = (uint64_t *)sm->sm_ppd;
 677         uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
 678         uint64_t max_size = metaslab_pp_maxsize(sm);
 679         uint64_t rsize = size;
 680         uint64_t offset = 0;
 681 
 682         ASSERT(MUTEX_HELD(sm->sm_lock));
 683         ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
 684 
 685         if (max_size < size)
 686                 return (-1ULL);
 687 
 688         ASSERT3U(*extent_end, >=, *cursor);
 689 
 690         /*
 691          * If we're running low on space switch to using the size
 692          * sorted AVL tree (best-fit).
 693          */
 694         if ((*cursor + size) > *extent_end) {
 695 
 696                 t = sm->sm_pp_root;
 697                 *cursor = *extent_end = 0;
 698 
 699                 if (max_size > 2 * SPA_MAXBLOCKSIZE)
 700                         rsize = MIN(metaslab_min_alloc_size, max_size);
 701                 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
 702                 if (offset != -1)
 703                         *cursor = offset + size;
 704         } else {
 705                 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
 706         }
 707         ASSERT3U(*cursor, <=, *extent_end);
 708         return (offset);
 709 }
 710 
 711 static boolean_t
 712 metaslab_cdf_fragmented(space_map_t *sm)
 713 {
 714         uint64_t max_size = metaslab_pp_maxsize(sm);
 715 
 716         if (max_size > (metaslab_min_alloc_size * 10))
 717                 return (B_FALSE);
 718         return (B_TRUE);
 719 }
 720 
 721 static space_map_ops_t metaslab_cdf_ops = {
 722         metaslab_pp_load,
 723         metaslab_pp_unload,
 724         metaslab_cdf_alloc,
 725         metaslab_pp_claim,
 726         metaslab_pp_free,
 727         metaslab_pp_maxsize,
 728         metaslab_cdf_fragmented
 729 };
 730 
 731 uint64_t metaslab_ndf_clump_shift = 4;
 732 
 733 static uint64_t
 734 metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
 735 {
 736         avl_tree_t *t = &sm->sm_root;
 737         avl_index_t where;
 738         space_seg_t *ss, ssearch;
 739         uint64_t hbit = highbit(size);
 740         uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
 741         uint64_t max_size = metaslab_pp_maxsize(sm);
 742 
 743         ASSERT(MUTEX_HELD(sm->sm_lock));
 744         ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
 745 
 746         if (max_size < size)
 747                 return (-1ULL);
 748 
 749         ssearch.ss_start = *cursor;
 750         ssearch.ss_end = *cursor + size;
 751 
 752         ss = avl_find(t, &ssearch, &where);
 753         if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
 754                 t = sm->sm_pp_root;
 755 
 756                 ssearch.ss_start = 0;
 757                 ssearch.ss_end = MIN(max_size,
 758                     1ULL << (hbit + metaslab_ndf_clump_shift));
 759                 ss = avl_find(t, &ssearch, &where);
 760                 if (ss == NULL)
 761                         ss = avl_nearest(t, where, AVL_AFTER);
 762                 ASSERT(ss != NULL);
 763         }
 764 
 765         if (ss != NULL) {
 766                 if (ss->ss_start + size <= ss->ss_end) {
 767                         *cursor = ss->ss_start + size;
 768                         return (ss->ss_start);
 769                 }
 770         }
 771         return (-1ULL);
 772 }
 773 
 774 static boolean_t
 775 metaslab_ndf_fragmented(space_map_t *sm)
 776 {
 777         uint64_t max_size = metaslab_pp_maxsize(sm);
 778 
 779         if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
 780                 return (B_FALSE);
 781         return (B_TRUE);
 782 }
 783 
 784 
 785 static space_map_ops_t metaslab_ndf_ops = {
 786         metaslab_pp_load,
 787         metaslab_pp_unload,
 788         metaslab_ndf_alloc,
 789         metaslab_pp_claim,
 790         metaslab_pp_free,
 791         metaslab_pp_maxsize,
 792         metaslab_ndf_fragmented
 793 };
 794 
 795 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
 796 
 797 /*
 798  * ==========================================================================
 799  * Metaslabs
 800  * ==========================================================================
 801  */
 802 metaslab_t *
 803 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
 804         uint64_t start, uint64_t size, uint64_t txg)
 805 {
 806         vdev_t *vd = mg->mg_vd;
 807         metaslab_t *msp;
 808 
 809         msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
 810         mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
 811 
 812         msp->ms_smo_syncing = *smo;
 813 
 814         /*
 815          * We create the main space map here, but we don't create the
 816          * allocmaps and freemaps until metaslab_sync_done().  This serves
 817          * two purposes: it allows metaslab_sync_done() to detect the
 818          * addition of new space; and for debugging, it ensures that we'd
 819          * data fault on any attempt to use this metaslab before it's ready.
 820          */
 821         msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
 822         space_map_create(msp->ms_map, start, size,
 823             vd->vdev_ashift, &msp->ms_lock);
 824 
 825         metaslab_group_add(mg, msp);
 826 
 827         if (metaslab_debug && smo->smo_object != 0) {
 828                 mutex_enter(&msp->ms_lock);
 829                 VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops,
 830                     SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
 831                 mutex_exit(&msp->ms_lock);
 832         }
 833 
 834         /*
 835          * If we're opening an existing pool (txg == 0) or creating
 836          * a new one (txg == TXG_INITIAL), all space is available now.
 837          * If we're adding space to an existing pool, the new space
 838          * does not become available until after this txg has synced.
 839          */
 840         if (txg <= TXG_INITIAL)
 841                 metaslab_sync_done(msp, 0);
 842 
 843         if (txg != 0) {
 844                 vdev_dirty(vd, 0, NULL, txg);
 845                 vdev_dirty(vd, VDD_METASLAB, msp, txg);
 846         }
 847 
 848         return (msp);
 849 }
 850 
 851 void
 852 metaslab_fini(metaslab_t *msp)
 853 {
 854         metaslab_group_t *mg = msp->ms_group;
 855 
 856         vdev_space_update(mg->mg_vd,
 857             -msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size);
 858 
 859         metaslab_group_remove(mg, msp);
 860 
 861         mutex_enter(&msp->ms_lock);
 862 
 863         space_map_unload(msp->ms_map);
 864         space_map_destroy(msp->ms_map);
 865         kmem_free(msp->ms_map, sizeof (*msp->ms_map));
 866 
 867         for (int t = 0; t < TXG_SIZE; t++) {
 868                 space_map_destroy(msp->ms_allocmap[t]);
 869                 space_map_destroy(msp->ms_freemap[t]);
 870                 kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t]));
 871                 kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t]));
 872         }
 873 
 874         for (int t = 0; t < TXG_DEFER_SIZE; t++) {
 875                 space_map_destroy(msp->ms_defermap[t]);
 876                 kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t]));
 877         }
 878 
 879         ASSERT0(msp->ms_deferspace);
 880 
 881         mutex_exit(&msp->ms_lock);
 882         mutex_destroy(&msp->ms_lock);
 883 
 884         kmem_free(msp, sizeof (metaslab_t));
 885 }
 886 
 887 #define METASLAB_WEIGHT_PRIMARY         (1ULL << 63)
 888 #define METASLAB_WEIGHT_SECONDARY       (1ULL << 62)
 889 #define METASLAB_ACTIVE_MASK            \
 890         (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
 891 
 892 static uint64_t
 893 metaslab_weight(metaslab_t *msp)
 894 {
 895         metaslab_group_t *mg = msp->ms_group;
 896         space_map_t *sm = msp->ms_map;
 897         space_map_obj_t *smo = &msp->ms_smo;
 898         vdev_t *vd = mg->mg_vd;
 899         uint64_t weight, space;
 900 
 901         ASSERT(MUTEX_HELD(&msp->ms_lock));
 902 
 903         /*
 904          * This vdev is in the process of being removed so there is nothing
 905          * for us to do here.
 906          */
 907         if (vd->vdev_removing) {
 908                 ASSERT0(smo->smo_alloc);
 909                 ASSERT0(vd->vdev_ms_shift);
 910                 return (0);
 911         }
 912 
 913         /*
 914          * The baseline weight is the metaslab's free space.
 915          */
 916         space = sm->sm_size - smo->smo_alloc;
 917         weight = space;
 918 
 919         /*
 920          * Modern disks have uniform bit density and constant angular velocity.
 921          * Therefore, the outer recording zones are faster (higher bandwidth)
 922          * than the inner zones by the ratio of outer to inner track diameter,
 923          * which is typically around 2:1.  We account for this by assigning
 924          * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
 925          * In effect, this means that we'll select the metaslab with the most
 926          * free bandwidth rather than simply the one with the most free space.
 927          */
 928         weight = 2 * weight -
 929             ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
 930         ASSERT(weight >= space && weight <= 2 * space);
 931 
 932         /*
 933          * For locality, assign higher weight to metaslabs which have
 934          * a lower offset than what we've already activated.
 935          */
 936         if (sm->sm_start <= mg->mg_bonus_area)
 937                 weight *= (metaslab_smo_bonus_pct / 100);
 938         ASSERT(weight >= space &&
 939             weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
 940 
 941         if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
 942                 /*
 943                  * If this metaslab is one we're actively using, adjust its
 944                  * weight to make it preferable to any inactive metaslab so
 945                  * we'll polish it off.
 946                  */
 947                 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
 948         }
 949         return (weight);
 950 }
 951 
 952 static void
 953 metaslab_prefetch(metaslab_group_t *mg)
 954 {
 955         spa_t *spa = mg->mg_vd->vdev_spa;
 956         metaslab_t *msp;
 957         avl_tree_t *t = &mg->mg_metaslab_tree;
 958         int m;
 959 
 960         mutex_enter(&mg->mg_lock);
 961 
 962         /*
 963          * Prefetch the next potential metaslabs
 964          */
 965         for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
 966                 space_map_t *sm = msp->ms_map;
 967                 space_map_obj_t *smo = &msp->ms_smo;
 968 
 969                 /* If we have reached our prefetch limit then we're done */
 970                 if (m >= metaslab_prefetch_limit)
 971                         break;
 972 
 973                 if (!sm->sm_loaded && smo->smo_object != 0) {
 974                         mutex_exit(&mg->mg_lock);
 975                         dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
 976                             0ULL, smo->smo_objsize);
 977                         mutex_enter(&mg->mg_lock);
 978                 }
 979         }
 980         mutex_exit(&mg->mg_lock);
 981 }
 982 
 983 static int
 984 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
 985 {
 986         metaslab_group_t *mg = msp->ms_group;
 987         space_map_t *sm = msp->ms_map;
 988         space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
 989 
 990         ASSERT(MUTEX_HELD(&msp->ms_lock));
 991 
 992         if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
 993                 space_map_load_wait(sm);
 994                 if (!sm->sm_loaded) {
 995                         space_map_obj_t *smo = &msp->ms_smo;
 996 
 997                         int error = space_map_load(sm, sm_ops, SM_FREE, smo,
 998                             spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
 999                         if (error)  {
1000                                 metaslab_group_sort(msp->ms_group, msp, 0);
1001                                 return (error);
1002                         }
1003                         for (int t = 0; t < TXG_DEFER_SIZE; t++)
1004                                 space_map_walk(msp->ms_defermap[t],
1005                                     space_map_claim, sm);
1006 
1007                 }
1008 
1009                 /*
1010                  * Track the bonus area as we activate new metaslabs.
1011                  */
1012                 if (sm->sm_start > mg->mg_bonus_area) {
1013                         mutex_enter(&mg->mg_lock);
1014                         mg->mg_bonus_area = sm->sm_start;
1015                         mutex_exit(&mg->mg_lock);
1016                 }
1017 
1018                 metaslab_group_sort(msp->ms_group, msp,
1019                     msp->ms_weight | activation_weight);
1020         }
1021         ASSERT(sm->sm_loaded);
1022         ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1023 
1024         return (0);
1025 }
1026 
1027 static void
1028 metaslab_passivate(metaslab_t *msp, uint64_t size)
1029 {
1030         /*
1031          * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1032          * this metaslab again.  In that case, it had better be empty,
1033          * or we would be leaving space on the table.
1034          */
1035         ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0);
1036         metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1037         ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1038 }
1039 
1040 /*
1041  * Determine if the in-core space map representation can be condensed on-disk.
1042  * We would like to use the following criteria to make our decision:
1043  *
1044  * 1. The size of the space map object should not dramatically increase as a
1045  * result of writing out our in-core free map.
1046  *
1047  * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1048  * times the size than the in-core representation (i.e. zfs_condense_pct = 110
1049  * and in-core = 1MB, minimal = 1.1.MB).
1050  *
1051  * Checking the first condition is tricky since we don't want to walk
1052  * the entire AVL tree calculating the estimated on-disk size. Instead we
1053  * use the size-ordered AVL tree in the space map and calculate the
1054  * size required for the largest segment in our in-core free map. If the
1055  * size required to represent that segment on disk is larger than the space
1056  * map object then we avoid condensing this map.
1057  *
1058  * To determine the second criterion we use a best-case estimate and assume
1059  * each segment can be represented on-disk as a single 64-bit entry. We refer
1060  * to this best-case estimate as the space map's minimal form.
1061  */
1062 static boolean_t
1063 metaslab_should_condense(metaslab_t *msp)
1064 {
1065         space_map_t *sm = msp->ms_map;
1066         space_map_obj_t *smo = &msp->ms_smo_syncing;
1067         space_seg_t *ss;
1068         uint64_t size, entries, segsz;
1069 
1070         ASSERT(MUTEX_HELD(&msp->ms_lock));
1071         ASSERT(sm->sm_loaded);
1072 
1073         /*
1074          * Use the sm_pp_root AVL tree, which is ordered by size, to obtain
1075          * the largest segment in the in-core free map. If the tree is
1076          * empty then we should condense the map.
1077          */
1078         ss = avl_last(sm->sm_pp_root);
1079         if (ss == NULL)
1080                 return (B_TRUE);
1081 
1082         /*
1083          * Calculate the number of 64-bit entries this segment would
1084          * require when written to disk. If this single segment would be
1085          * larger on-disk than the entire current on-disk structure, then
1086          * clearly condensing will increase the on-disk structure size.
1087          */
1088         size = (ss->ss_end - ss->ss_start) >> sm->sm_shift;
1089         entries = size / (MIN(size, SM_RUN_MAX));
1090         segsz = entries * sizeof (uint64_t);
1091 
1092         return (segsz <= smo->smo_objsize &&
1093             smo->smo_objsize >= (zfs_condense_pct *
1094             sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100);
1095 }
1096 
1097 /*
1098  * Condense the on-disk space map representation to its minimized form.
1099  * The minimized form consists of a small number of allocations followed by
1100  * the in-core free map.
1101  */
1102 static void
1103 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1104 {
1105         spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1106         space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK];
1107         space_map_t condense_map;
1108         space_map_t *sm = msp->ms_map;
1109         objset_t *mos = spa_meta_objset(spa);
1110         space_map_obj_t *smo = &msp->ms_smo_syncing;
1111 
1112         ASSERT(MUTEX_HELD(&msp->ms_lock));
1113         ASSERT3U(spa_sync_pass(spa), ==, 1);
1114         ASSERT(sm->sm_loaded);
1115 
1116         spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1117             "smo size %llu, segments %lu", txg,
1118             (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
1119             smo->smo_objsize, avl_numnodes(&sm->sm_root));
1120 
1121         /*
1122          * Create an map that is a 100% allocated map. We remove segments
1123          * that have been freed in this txg, any deferred frees that exist,
1124          * and any allocation in the future. Removing segments should be
1125          * a relatively inexpensive operation since we expect these maps to
1126          * a small number of nodes.
1127          */
1128         space_map_create(&condense_map, sm->sm_start, sm->sm_size,
1129             sm->sm_shift, sm->sm_lock);
1130         space_map_add(&condense_map, condense_map.sm_start,
1131             condense_map.sm_size);
1132 
1133         /*
1134          * Remove what's been freed in this txg from the condense_map.
1135          * Since we're in sync_pass 1, we know that all the frees from
1136          * this txg are in the freemap.
1137          */
1138         space_map_walk(freemap, space_map_remove, &condense_map);
1139 
1140         for (int t = 0; t < TXG_DEFER_SIZE; t++)
1141                 space_map_walk(msp->ms_defermap[t],
1142                     space_map_remove, &condense_map);
1143 
1144         for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1145                 space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK],
1146                     space_map_remove, &condense_map);
1147 
1148         /*
1149          * We're about to drop the metaslab's lock thus allowing
1150          * other consumers to change it's content. Set the
1151          * space_map's sm_condensing flag to ensure that
1152          * allocations on this metaslab do not occur while we're
1153          * in the middle of committing it to disk. This is only critical
1154          * for the ms_map as all other space_maps use per txg
1155          * views of their content.
1156          */
1157         sm->sm_condensing = B_TRUE;
1158 
1159         mutex_exit(&msp->ms_lock);
1160         space_map_truncate(smo, mos, tx);
1161         mutex_enter(&msp->ms_lock);
1162 
1163         /*
1164          * While we would ideally like to create a space_map representation
1165          * that consists only of allocation records, doing so can be
1166          * prohibitively expensive because the in-core free map can be
1167          * large, and therefore computationally expensive to subtract
1168          * from the condense_map. Instead we sync out two maps, a cheap
1169          * allocation only map followed by the in-core free map. While not
1170          * optimal, this is typically close to optimal, and much cheaper to
1171          * compute.
1172          */
1173         space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx);
1174         space_map_vacate(&condense_map, NULL, NULL);
1175         space_map_destroy(&condense_map);
1176 
1177         space_map_sync(sm, SM_FREE, smo, mos, tx);
1178         sm->sm_condensing = B_FALSE;
1179 
1180         spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, "
1181             "smo size %llu", txg,
1182             (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
1183             smo->smo_objsize);
1184 }
1185 
1186 /*
1187  * Write a metaslab to disk in the context of the specified transaction group.
1188  */
1189 void
1190 metaslab_sync(metaslab_t *msp, uint64_t txg)
1191 {
1192         vdev_t *vd = msp->ms_group->mg_vd;
1193         spa_t *spa = vd->vdev_spa;
1194         objset_t *mos = spa_meta_objset(spa);
1195         space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK];
1196         space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK];
1197         space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1198         space_map_t *sm = msp->ms_map;
1199         space_map_obj_t *smo = &msp->ms_smo_syncing;
1200         dmu_buf_t *db;
1201         dmu_tx_t *tx;
1202 
1203         ASSERT(!vd->vdev_ishole);
1204 
1205         /*
1206          * This metaslab has just been added so there's no work to do now.
1207          */
1208         if (*freemap == NULL) {
1209                 ASSERT3P(allocmap, ==, NULL);
1210                 return;
1211         }
1212 
1213         ASSERT3P(allocmap, !=, NULL);
1214         ASSERT3P(*freemap, !=, NULL);
1215         ASSERT3P(*freed_map, !=, NULL);
1216 
1217         if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0)
1218                 return;
1219 
1220         /*
1221          * The only state that can actually be changing concurrently with
1222          * metaslab_sync() is the metaslab's ms_map.  No other thread can
1223          * be modifying this txg's allocmap, freemap, freed_map, or smo.
1224          * Therefore, we only hold ms_lock to satify space_map ASSERTs.
1225          * We drop it whenever we call into the DMU, because the DMU
1226          * can call down to us (e.g. via zio_free()) at any time.
1227          */
1228 
1229         tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1230 
1231         if (smo->smo_object == 0) {
1232                 ASSERT(smo->smo_objsize == 0);
1233                 ASSERT(smo->smo_alloc == 0);
1234                 smo->smo_object = dmu_object_alloc(mos,
1235                     DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1236                     DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1237                 ASSERT(smo->smo_object != 0);
1238                 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1239                     (sm->sm_start >> vd->vdev_ms_shift),
1240                     sizeof (uint64_t), &smo->smo_object, tx);
1241         }
1242 
1243         mutex_enter(&msp->ms_lock);
1244 
1245         if (sm->sm_loaded && spa_sync_pass(spa) == 1 &&
1246             metaslab_should_condense(msp)) {
1247                 metaslab_condense(msp, txg, tx);
1248         } else {
1249                 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1250                 space_map_sync(*freemap, SM_FREE, smo, mos, tx);
1251         }
1252 
1253         space_map_vacate(allocmap, NULL, NULL);
1254 
1255         /*
1256          * For sync pass 1, we avoid walking the entire space map and
1257          * instead will just swap the pointers for freemap and
1258          * freed_map. We can safely do this since the freed_map is
1259          * guaranteed to be empty on the initial pass.
1260          */
1261         if (spa_sync_pass(spa) == 1) {
1262                 ASSERT0((*freed_map)->sm_space);
1263                 ASSERT0(avl_numnodes(&(*freed_map)->sm_root));
1264                 space_map_swap(freemap, freed_map);
1265         } else {
1266                 space_map_vacate(*freemap, space_map_add, *freed_map);
1267         }
1268 
1269         ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space);
1270         ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space);
1271 
1272         mutex_exit(&msp->ms_lock);
1273 
1274         VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1275         dmu_buf_will_dirty(db, tx);
1276         ASSERT3U(db->db_size, >=, sizeof (*smo));
1277         bcopy(smo, db->db_data, sizeof (*smo));
1278         dmu_buf_rele(db, FTAG);
1279 
1280         dmu_tx_commit(tx);
1281 }
1282 
1283 /*
1284  * Called after a transaction group has completely synced to mark
1285  * all of the metaslab's free space as usable.
1286  */
1287 void
1288 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1289 {
1290         space_map_obj_t *smo = &msp->ms_smo;
1291         space_map_obj_t *smosync = &msp->ms_smo_syncing;
1292         space_map_t *sm = msp->ms_map;
1293         space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1294         space_map_t **defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1295         metaslab_group_t *mg = msp->ms_group;
1296         vdev_t *vd = mg->mg_vd;
1297         int64_t alloc_delta, defer_delta;
1298 
1299         ASSERT(!vd->vdev_ishole);
1300 
1301         mutex_enter(&msp->ms_lock);
1302 
1303         /*
1304          * If this metaslab is just becoming available, initialize its
1305          * allocmaps, freemaps, and defermap and add its capacity to the vdev.
1306          */
1307         if (*freed_map == NULL) {
1308                 ASSERT(*defer_map == NULL);
1309                 for (int t = 0; t < TXG_SIZE; t++) {
1310                         msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t),
1311                             KM_SLEEP);
1312                         space_map_create(msp->ms_allocmap[t], sm->sm_start,
1313                             sm->sm_size, sm->sm_shift, sm->sm_lock);
1314                         msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t),
1315                             KM_SLEEP);
1316                         space_map_create(msp->ms_freemap[t], sm->sm_start,
1317                             sm->sm_size, sm->sm_shift, sm->sm_lock);
1318                 }
1319 
1320                 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1321                         msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t),
1322                             KM_SLEEP);
1323                         space_map_create(msp->ms_defermap[t], sm->sm_start,
1324                             sm->sm_size, sm->sm_shift, sm->sm_lock);
1325                 }
1326 
1327                 freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1328                 defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1329 
1330                 vdev_space_update(vd, 0, 0, sm->sm_size);
1331         }
1332 
1333         alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1334         defer_delta = (*freed_map)->sm_space - (*defer_map)->sm_space;
1335 
1336         vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1337 
1338         ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0);
1339         ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0);
1340 
1341         /*
1342          * If there's a space_map_load() in progress, wait for it to complete
1343          * so that we have a consistent view of the in-core space map.
1344          */
1345         space_map_load_wait(sm);
1346 
1347         /*
1348          * Move the frees from the defer_map to this map (if it's loaded).
1349          * Swap the freed_map and the defer_map -- this is safe to do
1350          * because we've just emptied out the defer_map.
1351          */
1352         space_map_vacate(*defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1353         ASSERT0((*defer_map)->sm_space);
1354         ASSERT0(avl_numnodes(&(*defer_map)->sm_root));
1355         space_map_swap(freed_map, defer_map);
1356 
1357         *smo = *smosync;
1358 
1359         msp->ms_deferspace += defer_delta;
1360         ASSERT3S(msp->ms_deferspace, >=, 0);
1361         ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1362         if (msp->ms_deferspace != 0) {
1363                 /*
1364                  * Keep syncing this metaslab until all deferred frees
1365                  * are back in circulation.
1366                  */
1367                 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1368         }
1369 
1370         /*
1371          * If the map is loaded but no longer active, evict it as soon as all
1372          * future allocations have synced.  (If we unloaded it now and then
1373          * loaded a moment later, the map wouldn't reflect those allocations.)
1374          */
1375         if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1376                 int evictable = 1;
1377 
1378                 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1379                         if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space)
1380                                 evictable = 0;
1381 
1382                 if (evictable && !metaslab_debug)
1383                         space_map_unload(sm);
1384         }
1385 
1386         metaslab_group_sort(mg, msp, metaslab_weight(msp));
1387 
1388         mutex_exit(&msp->ms_lock);
1389 }
1390 
1391 void
1392 metaslab_sync_reassess(metaslab_group_t *mg)
1393 {
1394         vdev_t *vd = mg->mg_vd;
1395         int64_t failures = mg->mg_alloc_failures;
1396 
1397         metaslab_group_alloc_update(mg);
1398 
1399         /*
1400          * Re-evaluate all metaslabs which have lower offsets than the
1401          * bonus area.
1402          */
1403         for (int m = 0; m < vd->vdev_ms_count; m++) {
1404                 metaslab_t *msp = vd->vdev_ms[m];
1405 
1406                 if (msp->ms_map->sm_start > mg->mg_bonus_area)
1407                         break;
1408 
1409                 mutex_enter(&msp->ms_lock);
1410                 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1411                 mutex_exit(&msp->ms_lock);
1412         }
1413 
1414         atomic_add_64(&mg->mg_alloc_failures, -failures);
1415 
1416         /*
1417          * Prefetch the next potential metaslabs
1418          */
1419         metaslab_prefetch(mg);
1420 }
1421 
1422 static uint64_t
1423 metaslab_distance(metaslab_t *msp, dva_t *dva)
1424 {
1425         uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1426         uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1427         uint64_t start = msp->ms_map->sm_start >> ms_shift;
1428 
1429         if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1430                 return (1ULL << 63);
1431 
1432         if (offset < start)
1433                 return ((start - offset) << ms_shift);
1434         if (offset > start)
1435                 return ((offset - start) << ms_shift);
1436         return (0);
1437 }
1438 
1439 static uint64_t
1440 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1441     uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1442 {
1443         spa_t *spa = mg->mg_vd->vdev_spa;
1444         metaslab_t *msp = NULL;
1445         uint64_t offset = -1ULL;
1446         avl_tree_t *t = &mg->mg_metaslab_tree;
1447         uint64_t activation_weight;
1448         uint64_t target_distance;
1449         int i;
1450 
1451         activation_weight = METASLAB_WEIGHT_PRIMARY;
1452         for (i = 0; i < d; i++) {
1453                 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1454                         activation_weight = METASLAB_WEIGHT_SECONDARY;
1455                         break;
1456                 }
1457         }
1458 
1459         for (;;) {
1460                 boolean_t was_active;
1461 
1462                 mutex_enter(&mg->mg_lock);
1463                 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1464                         if (msp->ms_weight < asize) {
1465                                 spa_dbgmsg(spa, "%s: failed to meet weight "
1466                                     "requirement: vdev %llu, txg %llu, mg %p, "
1467                                     "msp %p, psize %llu, asize %llu, "
1468                                     "failures %llu, weight %llu",
1469                                     spa_name(spa), mg->mg_vd->vdev_id, txg,
1470                                     mg, msp, psize, asize,
1471                                     mg->mg_alloc_failures, msp->ms_weight);
1472                                 mutex_exit(&mg->mg_lock);
1473                                 return (-1ULL);
1474                         }
1475 
1476                         /*
1477                          * If the selected metaslab is condensing, skip it.
1478                          */
1479                         if (msp->ms_map->sm_condensing)
1480                                 continue;
1481 
1482                         was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1483                         if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1484                                 break;
1485 
1486                         target_distance = min_distance +
1487                             (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1488 
1489                         for (i = 0; i < d; i++)
1490                                 if (metaslab_distance(msp, &dva[i]) <
1491                                     target_distance)
1492                                         break;
1493                         if (i == d)
1494                                 break;
1495                 }
1496                 mutex_exit(&mg->mg_lock);
1497                 if (msp == NULL)
1498                         return (-1ULL);
1499 
1500                 mutex_enter(&msp->ms_lock);
1501 
1502                 /*
1503                  * If we've already reached the allowable number of failed
1504                  * allocation attempts on this metaslab group then we
1505                  * consider skipping it. We skip it only if we're allowed
1506                  * to "fast" gang, the physical size is larger than
1507                  * a gang block, and we're attempting to allocate from
1508                  * the primary metaslab.
1509                  */
1510                 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1511                     CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1512                     activation_weight == METASLAB_WEIGHT_PRIMARY) {
1513                         spa_dbgmsg(spa, "%s: skipping metaslab group: "
1514                             "vdev %llu, txg %llu, mg %p, psize %llu, "
1515                             "asize %llu, failures %llu", spa_name(spa),
1516                             mg->mg_vd->vdev_id, txg, mg, psize, asize,
1517                             mg->mg_alloc_failures);
1518                         mutex_exit(&msp->ms_lock);
1519                         return (-1ULL);
1520                 }
1521 
1522                 /*
1523                  * Ensure that the metaslab we have selected is still
1524                  * capable of handling our request. It's possible that
1525                  * another thread may have changed the weight while we
1526                  * were blocked on the metaslab lock.
1527                  */
1528                 if (msp->ms_weight < asize || (was_active &&
1529                     !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1530                     activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1531                         mutex_exit(&msp->ms_lock);
1532                         continue;
1533                 }
1534 
1535                 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1536                     activation_weight == METASLAB_WEIGHT_PRIMARY) {
1537                         metaslab_passivate(msp,
1538                             msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1539                         mutex_exit(&msp->ms_lock);
1540                         continue;
1541                 }
1542 
1543                 if (metaslab_activate(msp, activation_weight) != 0) {
1544                         mutex_exit(&msp->ms_lock);
1545                         continue;
1546                 }
1547 
1548                 /*
1549                  * If this metaslab is currently condensing then pick again as
1550                  * we can't manipulate this metaslab until it's committed
1551                  * to disk.
1552                  */
1553                 if (msp->ms_map->sm_condensing) {
1554                         mutex_exit(&msp->ms_lock);
1555                         continue;
1556                 }
1557 
1558                 if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL)
1559                         break;
1560 
1561                 atomic_inc_64(&mg->mg_alloc_failures);
1562 
1563                 metaslab_passivate(msp, space_map_maxsize(msp->ms_map));
1564 
1565                 mutex_exit(&msp->ms_lock);
1566         }
1567 
1568         if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
1569                 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1570 
1571         space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1572 
1573         mutex_exit(&msp->ms_lock);
1574 
1575         return (offset);
1576 }
1577 
1578 /*
1579  * Allocate a block for the specified i/o.
1580  */
1581 static int
1582 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1583     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1584 {
1585         metaslab_group_t *mg, *rotor;
1586         vdev_t *vd;
1587         int dshift = 3;
1588         int all_zero;
1589         int zio_lock = B_FALSE;
1590         boolean_t allocatable;
1591         uint64_t offset = -1ULL;
1592         uint64_t asize;
1593         uint64_t distance;
1594 
1595         ASSERT(!DVA_IS_VALID(&dva[d]));
1596 
1597         /*
1598          * For testing, make some blocks above a certain size be gang blocks.
1599          */
1600         if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1601                 return (SET_ERROR(ENOSPC));
1602 
1603         /*
1604          * Start at the rotor and loop through all mgs until we find something.
1605          * Note that there's no locking on mc_rotor or mc_aliquot because
1606          * nothing actually breaks if we miss a few updates -- we just won't
1607          * allocate quite as evenly.  It all balances out over time.
1608          *
1609          * If we are doing ditto or log blocks, try to spread them across
1610          * consecutive vdevs.  If we're forced to reuse a vdev before we've
1611          * allocated all of our ditto blocks, then try and spread them out on
1612          * that vdev as much as possible.  If it turns out to not be possible,
1613          * gradually lower our standards until anything becomes acceptable.
1614          * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1615          * gives us hope of containing our fault domains to something we're
1616          * able to reason about.  Otherwise, any two top-level vdev failures
1617          * will guarantee the loss of data.  With consecutive allocation,
1618          * only two adjacent top-level vdev failures will result in data loss.
1619          *
1620          * If we are doing gang blocks (hintdva is non-NULL), try to keep
1621          * ourselves on the same vdev as our gang block header.  That
1622          * way, we can hope for locality in vdev_cache, plus it makes our
1623          * fault domains something tractable.
1624          */
1625         if (hintdva) {
1626                 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1627 
1628                 /*
1629                  * It's possible the vdev we're using as the hint no
1630                  * longer exists (i.e. removed). Consult the rotor when
1631                  * all else fails.
1632                  */
1633                 if (vd != NULL) {
1634                         mg = vd->vdev_mg;
1635 
1636                         if (flags & METASLAB_HINTBP_AVOID &&
1637                             mg->mg_next != NULL)
1638                                 mg = mg->mg_next;
1639                 } else {
1640                         mg = mc->mc_rotor;
1641                 }
1642         } else if (d != 0) {
1643                 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1644                 mg = vd->vdev_mg->mg_next;
1645         } else {
1646                 mg = mc->mc_rotor;
1647         }
1648 
1649         /*
1650          * If the hint put us into the wrong metaslab class, or into a
1651          * metaslab group that has been passivated, just follow the rotor.
1652          */
1653         if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1654                 mg = mc->mc_rotor;
1655 
1656         rotor = mg;
1657 top:
1658         all_zero = B_TRUE;
1659         do {
1660                 ASSERT(mg->mg_activation_count == 1);
1661 
1662                 vd = mg->mg_vd;
1663 
1664                 /*
1665                  * Don't allocate from faulted devices.
1666                  */
1667                 if (zio_lock) {
1668                         spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1669                         allocatable = vdev_allocatable(vd);
1670                         spa_config_exit(spa, SCL_ZIO, FTAG);
1671                 } else {
1672                         allocatable = vdev_allocatable(vd);
1673                 }
1674 
1675                 /*
1676                  * Determine if the selected metaslab group is eligible
1677                  * for allocations. If we're ganging or have requested
1678                  * an allocation for the smallest gang block size
1679                  * then we don't want to avoid allocating to the this
1680                  * metaslab group. If we're in this condition we should
1681                  * try to allocate from any device possible so that we
1682                  * don't inadvertently return ENOSPC and suspend the pool
1683                  * even though space is still available.
1684                  */
1685                 if (allocatable && CAN_FASTGANG(flags) &&
1686                     psize > SPA_GANGBLOCKSIZE)
1687                         allocatable = metaslab_group_allocatable(mg);
1688 
1689                 if (!allocatable)
1690                         goto next;
1691 
1692                 /*
1693                  * Avoid writing single-copy data to a failing vdev
1694                  * unless the user instructs us that it is okay.
1695                  */
1696                 if ((vd->vdev_stat.vs_write_errors > 0 ||
1697                     vd->vdev_state < VDEV_STATE_HEALTHY) &&
1698                     d == 0 && dshift == 3 &&
1699                     !(zfs_write_to_degraded && vd->vdev_state ==
1700                     VDEV_STATE_DEGRADED)) {
1701                         all_zero = B_FALSE;
1702                         goto next;
1703                 }
1704 
1705                 ASSERT(mg->mg_class == mc);
1706 
1707                 distance = vd->vdev_asize >> dshift;
1708                 if (distance <= (1ULL << vd->vdev_ms_shift))
1709                         distance = 0;
1710                 else
1711                         all_zero = B_FALSE;
1712 
1713                 asize = vdev_psize_to_asize(vd, psize);
1714                 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1715 
1716                 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1717                     dva, d, flags);
1718                 if (offset != -1ULL) {
1719                         /*
1720                          * If we've just selected this metaslab group,
1721                          * figure out whether the corresponding vdev is
1722                          * over- or under-used relative to the pool,
1723                          * and set an allocation bias to even it out.
1724                          */
1725                         if (mc->mc_aliquot == 0) {
1726                                 vdev_stat_t *vs = &vd->vdev_stat;
1727                                 int64_t vu, cu;
1728 
1729                                 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1730                                 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1731 
1732                                 /*
1733                                  * Calculate how much more or less we should
1734                                  * try to allocate from this device during
1735                                  * this iteration around the rotor.
1736                                  * For example, if a device is 80% full
1737                                  * and the pool is 20% full then we should
1738                                  * reduce allocations by 60% on this device.
1739                                  *
1740                                  * mg_bias = (20 - 80) * 512K / 100 = -307K
1741                                  *
1742                                  * This reduces allocations by 307K for this
1743                                  * iteration.
1744                                  */
1745                                 mg->mg_bias = ((cu - vu) *
1746                                     (int64_t)mg->mg_aliquot) / 100;
1747                         }
1748 
1749                         if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1750                             mg->mg_aliquot + mg->mg_bias) {
1751                                 mc->mc_rotor = mg->mg_next;
1752                                 mc->mc_aliquot = 0;
1753                         }
1754 
1755                         DVA_SET_VDEV(&dva[d], vd->vdev_id);
1756                         DVA_SET_OFFSET(&dva[d], offset);
1757                         DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1758                         DVA_SET_ASIZE(&dva[d], asize);
1759 
1760                         return (0);
1761                 }
1762 next:
1763                 mc->mc_rotor = mg->mg_next;
1764                 mc->mc_aliquot = 0;
1765         } while ((mg = mg->mg_next) != rotor);
1766 
1767         if (!all_zero) {
1768                 dshift++;
1769                 ASSERT(dshift < 64);
1770                 goto top;
1771         }
1772 
1773         if (!allocatable && !zio_lock) {
1774                 dshift = 3;
1775                 zio_lock = B_TRUE;
1776                 goto top;
1777         }
1778 
1779         bzero(&dva[d], sizeof (dva_t));
1780 
1781         return (SET_ERROR(ENOSPC));
1782 }
1783 
1784 /*
1785  * Free the block represented by DVA in the context of the specified
1786  * transaction group.
1787  */
1788 static void
1789 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1790 {
1791         uint64_t vdev = DVA_GET_VDEV(dva);
1792         uint64_t offset = DVA_GET_OFFSET(dva);
1793         uint64_t size = DVA_GET_ASIZE(dva);
1794         vdev_t *vd;
1795         metaslab_t *msp;
1796 
1797         ASSERT(DVA_IS_VALID(dva));
1798 
1799         if (txg > spa_freeze_txg(spa))
1800                 return;
1801 
1802         if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1803             (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1804                 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1805                     (u_longlong_t)vdev, (u_longlong_t)offset);
1806                 ASSERT(0);
1807                 return;
1808         }
1809 
1810         msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1811 
1812         if (DVA_GET_GANG(dva))
1813                 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1814 
1815         mutex_enter(&msp->ms_lock);
1816 
1817         if (now) {
1818                 space_map_remove(msp->ms_allocmap[txg & TXG_MASK],
1819                     offset, size);
1820                 space_map_free(msp->ms_map, offset, size);
1821         } else {
1822                 if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0)
1823                         vdev_dirty(vd, VDD_METASLAB, msp, txg);
1824                 space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size);
1825         }
1826 
1827         mutex_exit(&msp->ms_lock);
1828 }
1829 
1830 /*
1831  * Intent log support: upon opening the pool after a crash, notify the SPA
1832  * of blocks that the intent log has allocated for immediate write, but
1833  * which are still considered free by the SPA because the last transaction
1834  * group didn't commit yet.
1835  */
1836 static int
1837 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1838 {
1839         uint64_t vdev = DVA_GET_VDEV(dva);
1840         uint64_t offset = DVA_GET_OFFSET(dva);
1841         uint64_t size = DVA_GET_ASIZE(dva);
1842         vdev_t *vd;
1843         metaslab_t *msp;
1844         int error = 0;
1845 
1846         ASSERT(DVA_IS_VALID(dva));
1847 
1848         if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1849             (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1850                 return (SET_ERROR(ENXIO));
1851 
1852         msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1853 
1854         if (DVA_GET_GANG(dva))
1855                 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1856 
1857         mutex_enter(&msp->ms_lock);
1858 
1859         if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded)
1860                 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1861 
1862         if (error == 0 && !space_map_contains(msp->ms_map, offset, size))
1863                 error = SET_ERROR(ENOENT);
1864 
1865         if (error || txg == 0) {        /* txg == 0 indicates dry run */
1866                 mutex_exit(&msp->ms_lock);
1867                 return (error);
1868         }
1869 
1870         space_map_claim(msp->ms_map, offset, size);
1871 
1872         if (spa_writeable(spa)) {       /* don't dirty if we're zdb(1M) */
1873                 if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
1874                         vdev_dirty(vd, VDD_METASLAB, msp, txg);
1875                 space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size);
1876         }
1877 
1878         mutex_exit(&msp->ms_lock);
1879 
1880         return (0);
1881 }
1882 
1883 int
1884 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1885     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1886 {
1887         dva_t *dva = bp->blk_dva;
1888         dva_t *hintdva = hintbp->blk_dva;
1889         int error = 0;
1890 
1891         ASSERT(bp->blk_birth == 0);
1892         ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1893 
1894         spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1895 
1896         if (mc->mc_rotor == NULL) {  /* no vdevs in this class */
1897                 spa_config_exit(spa, SCL_ALLOC, FTAG);
1898                 return (SET_ERROR(ENOSPC));
1899         }
1900 
1901         ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1902         ASSERT(BP_GET_NDVAS(bp) == 0);
1903         ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1904 
1905         for (int d = 0; d < ndvas; d++) {
1906                 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1907                     txg, flags);
1908                 if (error) {
1909                         for (d--; d >= 0; d--) {
1910                                 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1911                                 bzero(&dva[d], sizeof (dva_t));
1912                         }
1913                         spa_config_exit(spa, SCL_ALLOC, FTAG);
1914                         return (error);
1915                 }
1916         }
1917         ASSERT(error == 0);
1918         ASSERT(BP_GET_NDVAS(bp) == ndvas);
1919 
1920         spa_config_exit(spa, SCL_ALLOC, FTAG);
1921 
1922         BP_SET_BIRTH(bp, txg, txg);
1923 
1924         return (0);
1925 }
1926 
1927 void
1928 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1929 {
1930         const dva_t *dva = bp->blk_dva;
1931         int ndvas = BP_GET_NDVAS(bp);
1932 
1933         ASSERT(!BP_IS_HOLE(bp));
1934         ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1935 
1936         spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1937 
1938         for (int d = 0; d < ndvas; d++)
1939                 metaslab_free_dva(spa, &dva[d], txg, now);
1940 
1941         spa_config_exit(spa, SCL_FREE, FTAG);
1942 }
1943 
1944 int
1945 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1946 {
1947         const dva_t *dva = bp->blk_dva;
1948         int ndvas = BP_GET_NDVAS(bp);
1949         int error = 0;
1950 
1951         ASSERT(!BP_IS_HOLE(bp));
1952 
1953         if (txg != 0) {
1954                 /*
1955                  * First do a dry run to make sure all DVAs are claimable,
1956                  * so we don't have to unwind from partial failures below.
1957                  */
1958                 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1959                         return (error);
1960         }
1961 
1962         spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1963 
1964         for (int d = 0; d < ndvas; d++)
1965                 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1966                         break;
1967 
1968         spa_config_exit(spa, SCL_ALLOC, FTAG);
1969 
1970         ASSERT(error == 0 || txg == 0);
1971 
1972         return (error);
1973 }
1974 
1975 static void
1976 checkmap(space_map_t *sm, uint64_t off, uint64_t size)
1977 {
1978         space_seg_t *ss;
1979         avl_index_t where;
1980 
1981         mutex_enter(sm->sm_lock);
1982         ss = space_map_find(sm, off, size, &where);
1983         if (ss != NULL)
1984                 panic("freeing free block; ss=%p", (void *)ss);
1985         mutex_exit(sm->sm_lock);
1986 }
1987 
1988 void
1989 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
1990 {
1991         if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
1992                 return;
1993 
1994         spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
1995         for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1996                 uint64_t vdid = DVA_GET_VDEV(&bp->blk_dva[i]);
1997                 vdev_t *vd = vdev_lookup_top(spa, vdid);
1998                 uint64_t off = DVA_GET_OFFSET(&bp->blk_dva[i]);
1999                 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2000                 metaslab_t *ms = vd->vdev_ms[off >> vd->vdev_ms_shift];
2001 
2002                 if (ms->ms_map->sm_loaded)
2003                         checkmap(ms->ms_map, off, size);
2004 
2005                 for (int j = 0; j < TXG_SIZE; j++)
2006                         checkmap(ms->ms_freemap[j], off, size);
2007                 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2008                         checkmap(ms->ms_defermap[j], off, size);
2009         }
2010         spa_config_exit(spa, SCL_VDEV, FTAG);
2011 }