metaslab.c revision 265741
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) 2011, 2014 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#include <sys/spa_impl.h> 35 36SYSCTL_DECL(_vfs_zfs); 37SYSCTL_NODE(_vfs_zfs, OID_AUTO, metaslab, CTLFLAG_RW, 0, "ZFS metaslab"); 38 39/* 40 * Allow allocations to switch to gang blocks quickly. We do this to 41 * avoid having to load lots of space_maps in a given txg. There are, 42 * however, some cases where we want to avoid "fast" ganging and instead 43 * we want to do an exhaustive search of all metaslabs on this device. 44 * Currently we don't allow any gang, slog, or dump device related allocations 45 * to "fast" gang. 46 */ 47#define CAN_FASTGANG(flags) \ 48 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \ 49 METASLAB_GANG_AVOID))) 50 51#define METASLAB_WEIGHT_PRIMARY (1ULL << 63) 52#define METASLAB_WEIGHT_SECONDARY (1ULL << 62) 53#define METASLAB_ACTIVE_MASK \ 54 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY) 55 56uint64_t metaslab_aliquot = 512ULL << 10; 57uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */ 58TUNABLE_QUAD("vfs.zfs.metaslab.gang_bang", &metaslab_gang_bang); 59SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, gang_bang, CTLFLAG_RWTUN, 60 &metaslab_gang_bang, 0, 61 "Force gang block allocation for blocks larger than or equal to this value"); 62 63/* 64 * The in-core space map representation is more compact than its on-disk form. 65 * The zfs_condense_pct determines how much more compact the in-core 66 * space_map representation must be before we compact it on-disk. 67 * Values should be greater than or equal to 100. 68 */ 69int zfs_condense_pct = 200; 70TUNABLE_INT("vfs.zfs.condense_pct", &zfs_condense_pct); 71SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN, 72 &zfs_condense_pct, 0, 73 "Condense on-disk spacemap when it is more than this many percents" 74 " of in-memory counterpart"); 75 76/* 77 * The zfs_mg_noalloc_threshold defines which metaslab groups should 78 * be eligible for allocation. The value is defined as a percentage of 79 * a free space. Metaslab groups that have more free space than 80 * zfs_mg_noalloc_threshold are always eligible for allocations. Once 81 * a metaslab group's free space is less than or equal to the 82 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that 83 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold. 84 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all 85 * groups are allowed to accept allocations. Gang blocks are always 86 * eligible to allocate on any metaslab group. The default value of 0 means 87 * no metaslab group will be excluded based on this criterion. 88 */ 89int zfs_mg_noalloc_threshold = 0; 90TUNABLE_INT("vfs.zfs.mg_noalloc_threshold", &zfs_mg_noalloc_threshold); 91SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN, 92 &zfs_mg_noalloc_threshold, 0, 93 "Percentage of metaslab group size that should be free" 94 " to make it eligible for allocation"); 95 96/* 97 * When set will load all metaslabs when pool is first opened. 98 */ 99int metaslab_debug_load = 0; 100TUNABLE_INT("vfs.zfs.metaslab.debug_load", &metaslab_debug_load); 101SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN, 102 &metaslab_debug_load, 0, 103 "Load all metaslabs when pool is first opened"); 104 105/* 106 * When set will prevent metaslabs from being unloaded. 107 */ 108int metaslab_debug_unload = 0; 109TUNABLE_INT("vfs.zfs.metaslab.debug_unload", &metaslab_debug_unload); 110SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN, 111 &metaslab_debug_unload, 0, 112 "Prevent metaslabs from being unloaded"); 113 114/* 115 * Minimum size which forces the dynamic allocator to change 116 * it's allocation strategy. Once the space map cannot satisfy 117 * an allocation of this size then it switches to using more 118 * aggressive strategy (i.e search by size rather than offset). 119 */ 120uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE; 121TUNABLE_QUAD("vfs.zfs.metaslab.df_alloc_threshold", 122 &metaslab_df_alloc_threshold); 123SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN, 124 &metaslab_df_alloc_threshold, 0, 125 "Minimum size which forces the dynamic allocator to change it's allocation strategy"); 126 127/* 128 * The minimum free space, in percent, which must be available 129 * in a space map to continue allocations in a first-fit fashion. 130 * Once the space_map's free space drops below this level we dynamically 131 * switch to using best-fit allocations. 132 */ 133int metaslab_df_free_pct = 4; 134TUNABLE_INT("vfs.zfs.metaslab.df_free_pct", &metaslab_df_free_pct); 135SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN, 136 &metaslab_df_free_pct, 0, 137 "The minimum free space, in percent, which must be available in a space map to continue allocations in a first-fit fashion"); 138 139/* 140 * A metaslab is considered "free" if it contains a contiguous 141 * segment which is greater than metaslab_min_alloc_size. 142 */ 143uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS; 144TUNABLE_QUAD("vfs.zfs.metaslab.min_alloc_size", 145 &metaslab_min_alloc_size); 146SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN, 147 &metaslab_min_alloc_size, 0, 148 "A metaslab is considered \"free\" if it contains a contiguous segment which is greater than vfs.zfs.metaslab.min_alloc_size"); 149 150/* 151 * Percentage of all cpus that can be used by the metaslab taskq. 152 */ 153int metaslab_load_pct = 50; 154TUNABLE_INT("vfs.zfs.metaslab.load_pct", &metaslab_load_pct); 155SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN, 156 &metaslab_load_pct, 0, 157 "Percentage of cpus that can be used by the metaslab taskq"); 158 159/* 160 * Determines how many txgs a metaslab may remain loaded without having any 161 * allocations from it. As long as a metaslab continues to be used we will 162 * keep it loaded. 163 */ 164int metaslab_unload_delay = TXG_SIZE * 2; 165TUNABLE_INT("vfs.zfs.metaslab.unload_delay", &metaslab_unload_delay); 166SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN, 167 &metaslab_unload_delay, 0, 168 "Number of TXGs that an unused metaslab can be kept in memory"); 169 170/* 171 * Should we be willing to write data to degraded vdevs? 172 */ 173boolean_t zfs_write_to_degraded = B_FALSE; 174SYSCTL_INT(_vfs_zfs, OID_AUTO, write_to_degraded, CTLFLAG_RWTUN, 175 &zfs_write_to_degraded, 0, "Allow writing data to degraded vdevs"); 176TUNABLE_INT("vfs.zfs.write_to_degraded", &zfs_write_to_degraded); 177 178/* 179 * Max number of metaslabs per group to preload. 180 */ 181int metaslab_preload_limit = SPA_DVAS_PER_BP; 182TUNABLE_INT("vfs.zfs.metaslab.preload_limit", &metaslab_preload_limit); 183SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN, 184 &metaslab_preload_limit, 0, 185 "Max number of metaslabs per group to preload"); 186 187/* 188 * Enable/disable preloading of metaslab. 189 */ 190boolean_t metaslab_preload_enabled = B_TRUE; 191TUNABLE_INT("vfs.zfs.metaslab.preload_enabled", &metaslab_preload_enabled); 192SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN, 193 &metaslab_preload_enabled, 0, 194 "Max number of metaslabs per group to preload"); 195 196/* 197 * Enable/disable additional weight factor for each metaslab. 198 */ 199boolean_t metaslab_weight_factor_enable = B_FALSE; 200TUNABLE_INT("vfs.zfs.metaslab.weight_factor_enable", 201 &metaslab_weight_factor_enable); 202SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, weight_factor_enable, CTLFLAG_RWTUN, 203 &metaslab_weight_factor_enable, 0, 204 "Enable additional weight factor for each metaslab"); 205 206 207/* 208 * ========================================================================== 209 * Metaslab classes 210 * ========================================================================== 211 */ 212metaslab_class_t * 213metaslab_class_create(spa_t *spa, metaslab_ops_t *ops) 214{ 215 metaslab_class_t *mc; 216 217 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP); 218 219 mc->mc_spa = spa; 220 mc->mc_rotor = NULL; 221 mc->mc_ops = ops; 222 223 return (mc); 224} 225 226void 227metaslab_class_destroy(metaslab_class_t *mc) 228{ 229 ASSERT(mc->mc_rotor == NULL); 230 ASSERT(mc->mc_alloc == 0); 231 ASSERT(mc->mc_deferred == 0); 232 ASSERT(mc->mc_space == 0); 233 ASSERT(mc->mc_dspace == 0); 234 235 kmem_free(mc, sizeof (metaslab_class_t)); 236} 237 238int 239metaslab_class_validate(metaslab_class_t *mc) 240{ 241 metaslab_group_t *mg; 242 vdev_t *vd; 243 244 /* 245 * Must hold one of the spa_config locks. 246 */ 247 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) || 248 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER)); 249 250 if ((mg = mc->mc_rotor) == NULL) 251 return (0); 252 253 do { 254 vd = mg->mg_vd; 255 ASSERT(vd->vdev_mg != NULL); 256 ASSERT3P(vd->vdev_top, ==, vd); 257 ASSERT3P(mg->mg_class, ==, mc); 258 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops); 259 } while ((mg = mg->mg_next) != mc->mc_rotor); 260 261 return (0); 262} 263 264void 265metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta, 266 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta) 267{ 268 atomic_add_64(&mc->mc_alloc, alloc_delta); 269 atomic_add_64(&mc->mc_deferred, defer_delta); 270 atomic_add_64(&mc->mc_space, space_delta); 271 atomic_add_64(&mc->mc_dspace, dspace_delta); 272} 273 274void 275metaslab_class_minblocksize_update(metaslab_class_t *mc) 276{ 277 metaslab_group_t *mg; 278 vdev_t *vd; 279 uint64_t minashift = UINT64_MAX; 280 281 if ((mg = mc->mc_rotor) == NULL) { 282 mc->mc_minblocksize = SPA_MINBLOCKSIZE; 283 return; 284 } 285 286 do { 287 vd = mg->mg_vd; 288 if (vd->vdev_ashift < minashift) 289 minashift = vd->vdev_ashift; 290 } while ((mg = mg->mg_next) != mc->mc_rotor); 291 292 mc->mc_minblocksize = 1ULL << minashift; 293} 294 295uint64_t 296metaslab_class_get_alloc(metaslab_class_t *mc) 297{ 298 return (mc->mc_alloc); 299} 300 301uint64_t 302metaslab_class_get_deferred(metaslab_class_t *mc) 303{ 304 return (mc->mc_deferred); 305} 306 307uint64_t 308metaslab_class_get_space(metaslab_class_t *mc) 309{ 310 return (mc->mc_space); 311} 312 313uint64_t 314metaslab_class_get_dspace(metaslab_class_t *mc) 315{ 316 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space); 317} 318 319uint64_t 320metaslab_class_get_minblocksize(metaslab_class_t *mc) 321{ 322 return (mc->mc_minblocksize); 323} 324 325/* 326 * ========================================================================== 327 * Metaslab groups 328 * ========================================================================== 329 */ 330static int 331metaslab_compare(const void *x1, const void *x2) 332{ 333 const metaslab_t *m1 = x1; 334 const metaslab_t *m2 = x2; 335 336 if (m1->ms_weight < m2->ms_weight) 337 return (1); 338 if (m1->ms_weight > m2->ms_weight) 339 return (-1); 340 341 /* 342 * If the weights are identical, use the offset to force uniqueness. 343 */ 344 if (m1->ms_start < m2->ms_start) 345 return (-1); 346 if (m1->ms_start > m2->ms_start) 347 return (1); 348 349 ASSERT3P(m1, ==, m2); 350 351 return (0); 352} 353 354/* 355 * Update the allocatable flag and the metaslab group's capacity. 356 * The allocatable flag is set to true if the capacity is below 357 * the zfs_mg_noalloc_threshold. If a metaslab group transitions 358 * from allocatable to non-allocatable or vice versa then the metaslab 359 * group's class is updated to reflect the transition. 360 */ 361static void 362metaslab_group_alloc_update(metaslab_group_t *mg) 363{ 364 vdev_t *vd = mg->mg_vd; 365 metaslab_class_t *mc = mg->mg_class; 366 vdev_stat_t *vs = &vd->vdev_stat; 367 boolean_t was_allocatable; 368 369 ASSERT(vd == vd->vdev_top); 370 371 mutex_enter(&mg->mg_lock); 372 was_allocatable = mg->mg_allocatable; 373 374 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) / 375 (vs->vs_space + 1); 376 377 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold); 378 379 /* 380 * The mc_alloc_groups maintains a count of the number of 381 * groups in this metaslab class that are still above the 382 * zfs_mg_noalloc_threshold. This is used by the allocating 383 * threads to determine if they should avoid allocations to 384 * a given group. The allocator will avoid allocations to a group 385 * if that group has reached or is below the zfs_mg_noalloc_threshold 386 * and there are still other groups that are above the threshold. 387 * When a group transitions from allocatable to non-allocatable or 388 * vice versa we update the metaslab class to reflect that change. 389 * When the mc_alloc_groups value drops to 0 that means that all 390 * groups have reached the zfs_mg_noalloc_threshold making all groups 391 * eligible for allocations. This effectively means that all devices 392 * are balanced again. 393 */ 394 if (was_allocatable && !mg->mg_allocatable) 395 mc->mc_alloc_groups--; 396 else if (!was_allocatable && mg->mg_allocatable) 397 mc->mc_alloc_groups++; 398 mutex_exit(&mg->mg_lock); 399} 400 401metaslab_group_t * 402metaslab_group_create(metaslab_class_t *mc, vdev_t *vd) 403{ 404 metaslab_group_t *mg; 405 406 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP); 407 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL); 408 avl_create(&mg->mg_metaslab_tree, metaslab_compare, 409 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node)); 410 mg->mg_vd = vd; 411 mg->mg_class = mc; 412 mg->mg_activation_count = 0; 413 414 mg->mg_taskq = taskq_create("metaslab_group_tasksq", metaslab_load_pct, 415 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT); 416 417 return (mg); 418} 419 420void 421metaslab_group_destroy(metaslab_group_t *mg) 422{ 423 ASSERT(mg->mg_prev == NULL); 424 ASSERT(mg->mg_next == NULL); 425 /* 426 * We may have gone below zero with the activation count 427 * either because we never activated in the first place or 428 * because we're done, and possibly removing the vdev. 429 */ 430 ASSERT(mg->mg_activation_count <= 0); 431 432 avl_destroy(&mg->mg_metaslab_tree); 433 mutex_destroy(&mg->mg_lock); 434 kmem_free(mg, sizeof (metaslab_group_t)); 435} 436 437void 438metaslab_group_activate(metaslab_group_t *mg) 439{ 440 metaslab_class_t *mc = mg->mg_class; 441 metaslab_group_t *mgprev, *mgnext; 442 443 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER)); 444 445 ASSERT(mc->mc_rotor != mg); 446 ASSERT(mg->mg_prev == NULL); 447 ASSERT(mg->mg_next == NULL); 448 ASSERT(mg->mg_activation_count <= 0); 449 450 if (++mg->mg_activation_count <= 0) 451 return; 452 453 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children); 454 metaslab_group_alloc_update(mg); 455 456 if ((mgprev = mc->mc_rotor) == NULL) { 457 mg->mg_prev = mg; 458 mg->mg_next = mg; 459 } else { 460 mgnext = mgprev->mg_next; 461 mg->mg_prev = mgprev; 462 mg->mg_next = mgnext; 463 mgprev->mg_next = mg; 464 mgnext->mg_prev = mg; 465 } 466 mc->mc_rotor = mg; 467 metaslab_class_minblocksize_update(mc); 468} 469 470void 471metaslab_group_passivate(metaslab_group_t *mg) 472{ 473 metaslab_class_t *mc = mg->mg_class; 474 metaslab_group_t *mgprev, *mgnext; 475 476 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER)); 477 478 if (--mg->mg_activation_count != 0) { 479 ASSERT(mc->mc_rotor != mg); 480 ASSERT(mg->mg_prev == NULL); 481 ASSERT(mg->mg_next == NULL); 482 ASSERT(mg->mg_activation_count < 0); 483 return; 484 } 485 486 taskq_wait(mg->mg_taskq); 487 488 mgprev = mg->mg_prev; 489 mgnext = mg->mg_next; 490 491 if (mg == mgnext) { 492 mc->mc_rotor = NULL; 493 } else { 494 mc->mc_rotor = mgnext; 495 mgprev->mg_next = mgnext; 496 mgnext->mg_prev = mgprev; 497 } 498 499 mg->mg_prev = NULL; 500 mg->mg_next = NULL; 501 metaslab_class_minblocksize_update(mc); 502} 503 504static void 505metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp) 506{ 507 mutex_enter(&mg->mg_lock); 508 ASSERT(msp->ms_group == NULL); 509 msp->ms_group = mg; 510 msp->ms_weight = 0; 511 avl_add(&mg->mg_metaslab_tree, msp); 512 mutex_exit(&mg->mg_lock); 513} 514 515static void 516metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp) 517{ 518 mutex_enter(&mg->mg_lock); 519 ASSERT(msp->ms_group == mg); 520 avl_remove(&mg->mg_metaslab_tree, msp); 521 msp->ms_group = NULL; 522 mutex_exit(&mg->mg_lock); 523} 524 525static void 526metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight) 527{ 528 /* 529 * Although in principle the weight can be any value, in 530 * practice we do not use values in the range [1, 510]. 531 */ 532 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0); 533 ASSERT(MUTEX_HELD(&msp->ms_lock)); 534 535 mutex_enter(&mg->mg_lock); 536 ASSERT(msp->ms_group == mg); 537 avl_remove(&mg->mg_metaslab_tree, msp); 538 msp->ms_weight = weight; 539 avl_add(&mg->mg_metaslab_tree, msp); 540 mutex_exit(&mg->mg_lock); 541} 542 543/* 544 * Determine if a given metaslab group should skip allocations. A metaslab 545 * group should avoid allocations if its used capacity has crossed the 546 * zfs_mg_noalloc_threshold and there is at least one metaslab group 547 * that can still handle allocations. 548 */ 549static boolean_t 550metaslab_group_allocatable(metaslab_group_t *mg) 551{ 552 vdev_t *vd = mg->mg_vd; 553 spa_t *spa = vd->vdev_spa; 554 metaslab_class_t *mc = mg->mg_class; 555 556 /* 557 * A metaslab group is considered allocatable if its free capacity 558 * is greater than the set value of zfs_mg_noalloc_threshold, it's 559 * associated with a slog, or there are no other metaslab groups 560 * with free capacity greater than zfs_mg_noalloc_threshold. 561 */ 562 return (mg->mg_free_capacity > zfs_mg_noalloc_threshold || 563 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0); 564} 565 566/* 567 * ========================================================================== 568 * Range tree callbacks 569 * ========================================================================== 570 */ 571 572/* 573 * Comparison function for the private size-ordered tree. Tree is sorted 574 * by size, larger sizes at the end of the tree. 575 */ 576static int 577metaslab_rangesize_compare(const void *x1, const void *x2) 578{ 579 const range_seg_t *r1 = x1; 580 const range_seg_t *r2 = x2; 581 uint64_t rs_size1 = r1->rs_end - r1->rs_start; 582 uint64_t rs_size2 = r2->rs_end - r2->rs_start; 583 584 if (rs_size1 < rs_size2) 585 return (-1); 586 if (rs_size1 > rs_size2) 587 return (1); 588 589 if (r1->rs_start < r2->rs_start) 590 return (-1); 591 592 if (r1->rs_start > r2->rs_start) 593 return (1); 594 595 return (0); 596} 597 598/* 599 * Create any block allocator specific components. The current allocators 600 * rely on using both a size-ordered range_tree_t and an array of uint64_t's. 601 */ 602static void 603metaslab_rt_create(range_tree_t *rt, void *arg) 604{ 605 metaslab_t *msp = arg; 606 607 ASSERT3P(rt->rt_arg, ==, msp); 608 ASSERT(msp->ms_tree == NULL); 609 610 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare, 611 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node)); 612} 613 614/* 615 * Destroy the block allocator specific components. 616 */ 617static void 618metaslab_rt_destroy(range_tree_t *rt, void *arg) 619{ 620 metaslab_t *msp = arg; 621 622 ASSERT3P(rt->rt_arg, ==, msp); 623 ASSERT3P(msp->ms_tree, ==, rt); 624 ASSERT0(avl_numnodes(&msp->ms_size_tree)); 625 626 avl_destroy(&msp->ms_size_tree); 627} 628 629static void 630metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg) 631{ 632 metaslab_t *msp = arg; 633 634 ASSERT3P(rt->rt_arg, ==, msp); 635 ASSERT3P(msp->ms_tree, ==, rt); 636 VERIFY(!msp->ms_condensing); 637 avl_add(&msp->ms_size_tree, rs); 638} 639 640static void 641metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg) 642{ 643 metaslab_t *msp = arg; 644 645 ASSERT3P(rt->rt_arg, ==, msp); 646 ASSERT3P(msp->ms_tree, ==, rt); 647 VERIFY(!msp->ms_condensing); 648 avl_remove(&msp->ms_size_tree, rs); 649} 650 651static void 652metaslab_rt_vacate(range_tree_t *rt, void *arg) 653{ 654 metaslab_t *msp = arg; 655 656 ASSERT3P(rt->rt_arg, ==, msp); 657 ASSERT3P(msp->ms_tree, ==, rt); 658 659 /* 660 * Normally one would walk the tree freeing nodes along the way. 661 * Since the nodes are shared with the range trees we can avoid 662 * walking all nodes and just reinitialize the avl tree. The nodes 663 * will be freed by the range tree, so we don't want to free them here. 664 */ 665 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare, 666 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node)); 667} 668 669static range_tree_ops_t metaslab_rt_ops = { 670 metaslab_rt_create, 671 metaslab_rt_destroy, 672 metaslab_rt_add, 673 metaslab_rt_remove, 674 metaslab_rt_vacate 675}; 676 677/* 678 * ========================================================================== 679 * Metaslab block operations 680 * ========================================================================== 681 */ 682 683/* 684 * Return the maximum contiguous segment within the metaslab. 685 */ 686uint64_t 687metaslab_block_maxsize(metaslab_t *msp) 688{ 689 avl_tree_t *t = &msp->ms_size_tree; 690 range_seg_t *rs; 691 692 if (t == NULL || (rs = avl_last(t)) == NULL) 693 return (0ULL); 694 695 return (rs->rs_end - rs->rs_start); 696} 697 698uint64_t 699metaslab_block_alloc(metaslab_t *msp, uint64_t size) 700{ 701 uint64_t start; 702 range_tree_t *rt = msp->ms_tree; 703 704 VERIFY(!msp->ms_condensing); 705 706 start = msp->ms_ops->msop_alloc(msp, size); 707 if (start != -1ULL) { 708 vdev_t *vd = msp->ms_group->mg_vd; 709 710 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift)); 711 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); 712 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size); 713 range_tree_remove(rt, start, size); 714 } 715 return (start); 716} 717 718/* 719 * ========================================================================== 720 * Common allocator routines 721 * ========================================================================== 722 */ 723 724/* 725 * This is a helper function that can be used by the allocator to find 726 * a suitable block to allocate. This will search the specified AVL 727 * tree looking for a block that matches the specified criteria. 728 */ 729static uint64_t 730metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size, 731 uint64_t align) 732{ 733 range_seg_t *rs, rsearch; 734 avl_index_t where; 735 736 rsearch.rs_start = *cursor; 737 rsearch.rs_end = *cursor + size; 738 739 rs = avl_find(t, &rsearch, &where); 740 if (rs == NULL) 741 rs = avl_nearest(t, where, AVL_AFTER); 742 743 while (rs != NULL) { 744 uint64_t offset = P2ROUNDUP(rs->rs_start, align); 745 746 if (offset + size <= rs->rs_end) { 747 *cursor = offset + size; 748 return (offset); 749 } 750 rs = AVL_NEXT(t, rs); 751 } 752 753 /* 754 * If we know we've searched the whole map (*cursor == 0), give up. 755 * Otherwise, reset the cursor to the beginning and try again. 756 */ 757 if (*cursor == 0) 758 return (-1ULL); 759 760 *cursor = 0; 761 return (metaslab_block_picker(t, cursor, size, align)); 762} 763 764/* 765 * ========================================================================== 766 * The first-fit block allocator 767 * ========================================================================== 768 */ 769static uint64_t 770metaslab_ff_alloc(metaslab_t *msp, uint64_t size) 771{ 772 /* 773 * Find the largest power of 2 block size that evenly divides the 774 * requested size. This is used to try to allocate blocks with similar 775 * alignment from the same area of the metaslab (i.e. same cursor 776 * bucket) but it does not guarantee that other allocations sizes 777 * may exist in the same region. 778 */ 779 uint64_t align = size & -size; 780 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1]; 781 avl_tree_t *t = &msp->ms_tree->rt_root; 782 783 return (metaslab_block_picker(t, cursor, size, align)); 784} 785 786/* ARGSUSED */ 787static boolean_t 788metaslab_ff_fragmented(metaslab_t *msp) 789{ 790 return (B_TRUE); 791} 792 793static metaslab_ops_t metaslab_ff_ops = { 794 metaslab_ff_alloc, 795 metaslab_ff_fragmented 796}; 797 798/* 799 * ========================================================================== 800 * Dynamic block allocator - 801 * Uses the first fit allocation scheme until space get low and then 802 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold 803 * and metaslab_df_free_pct to determine when to switch the allocation scheme. 804 * ========================================================================== 805 */ 806static uint64_t 807metaslab_df_alloc(metaslab_t *msp, uint64_t size) 808{ 809 /* 810 * Find the largest power of 2 block size that evenly divides the 811 * requested size. This is used to try to allocate blocks with similar 812 * alignment from the same area of the metaslab (i.e. same cursor 813 * bucket) but it does not guarantee that other allocations sizes 814 * may exist in the same region. 815 */ 816 uint64_t align = size & -size; 817 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1]; 818 range_tree_t *rt = msp->ms_tree; 819 avl_tree_t *t = &rt->rt_root; 820 uint64_t max_size = metaslab_block_maxsize(msp); 821 int free_pct = range_tree_space(rt) * 100 / msp->ms_size; 822 823 ASSERT(MUTEX_HELD(&msp->ms_lock)); 824 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree)); 825 826 if (max_size < size) 827 return (-1ULL); 828 829 /* 830 * If we're running low on space switch to using the size 831 * sorted AVL tree (best-fit). 832 */ 833 if (max_size < metaslab_df_alloc_threshold || 834 free_pct < metaslab_df_free_pct) { 835 t = &msp->ms_size_tree; 836 *cursor = 0; 837 } 838 839 return (metaslab_block_picker(t, cursor, size, 1ULL)); 840} 841 842static boolean_t 843metaslab_df_fragmented(metaslab_t *msp) 844{ 845 range_tree_t *rt = msp->ms_tree; 846 uint64_t max_size = metaslab_block_maxsize(msp); 847 int free_pct = range_tree_space(rt) * 100 / msp->ms_size; 848 849 if (max_size >= metaslab_df_alloc_threshold && 850 free_pct >= metaslab_df_free_pct) 851 return (B_FALSE); 852 853 return (B_TRUE); 854} 855 856static metaslab_ops_t metaslab_df_ops = { 857 metaslab_df_alloc, 858 metaslab_df_fragmented 859}; 860 861/* 862 * ========================================================================== 863 * Cursor fit block allocator - 864 * Select the largest region in the metaslab, set the cursor to the beginning 865 * of the range and the cursor_end to the end of the range. As allocations 866 * are made advance the cursor. Continue allocating from the cursor until 867 * the range is exhausted and then find a new range. 868 * ========================================================================== 869 */ 870static uint64_t 871metaslab_cf_alloc(metaslab_t *msp, uint64_t size) 872{ 873 range_tree_t *rt = msp->ms_tree; 874 avl_tree_t *t = &msp->ms_size_tree; 875 uint64_t *cursor = &msp->ms_lbas[0]; 876 uint64_t *cursor_end = &msp->ms_lbas[1]; 877 uint64_t offset = 0; 878 879 ASSERT(MUTEX_HELD(&msp->ms_lock)); 880 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root)); 881 882 ASSERT3U(*cursor_end, >=, *cursor); 883 884 if ((*cursor + size) > *cursor_end) { 885 range_seg_t *rs; 886 887 rs = avl_last(&msp->ms_size_tree); 888 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) 889 return (-1ULL); 890 891 *cursor = rs->rs_start; 892 *cursor_end = rs->rs_end; 893 } 894 895 offset = *cursor; 896 *cursor += size; 897 898 return (offset); 899} 900 901static boolean_t 902metaslab_cf_fragmented(metaslab_t *msp) 903{ 904 return (metaslab_block_maxsize(msp) < metaslab_min_alloc_size); 905} 906 907static metaslab_ops_t metaslab_cf_ops = { 908 metaslab_cf_alloc, 909 metaslab_cf_fragmented 910}; 911 912/* 913 * ========================================================================== 914 * New dynamic fit allocator - 915 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift 916 * contiguous blocks. If no region is found then just use the largest segment 917 * that remains. 918 * ========================================================================== 919 */ 920 921/* 922 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift) 923 * to request from the allocator. 924 */ 925uint64_t metaslab_ndf_clump_shift = 4; 926 927static uint64_t 928metaslab_ndf_alloc(metaslab_t *msp, uint64_t size) 929{ 930 avl_tree_t *t = &msp->ms_tree->rt_root; 931 avl_index_t where; 932 range_seg_t *rs, rsearch; 933 uint64_t hbit = highbit64(size); 934 uint64_t *cursor = &msp->ms_lbas[hbit - 1]; 935 uint64_t max_size = metaslab_block_maxsize(msp); 936 937 ASSERT(MUTEX_HELD(&msp->ms_lock)); 938 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree)); 939 940 if (max_size < size) 941 return (-1ULL); 942 943 rsearch.rs_start = *cursor; 944 rsearch.rs_end = *cursor + size; 945 946 rs = avl_find(t, &rsearch, &where); 947 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) { 948 t = &msp->ms_size_tree; 949 950 rsearch.rs_start = 0; 951 rsearch.rs_end = MIN(max_size, 952 1ULL << (hbit + metaslab_ndf_clump_shift)); 953 rs = avl_find(t, &rsearch, &where); 954 if (rs == NULL) 955 rs = avl_nearest(t, where, AVL_AFTER); 956 ASSERT(rs != NULL); 957 } 958 959 if ((rs->rs_end - rs->rs_start) >= size) { 960 *cursor = rs->rs_start + size; 961 return (rs->rs_start); 962 } 963 return (-1ULL); 964} 965 966static boolean_t 967metaslab_ndf_fragmented(metaslab_t *msp) 968{ 969 return (metaslab_block_maxsize(msp) <= 970 (metaslab_min_alloc_size << metaslab_ndf_clump_shift)); 971} 972 973static metaslab_ops_t metaslab_ndf_ops = { 974 metaslab_ndf_alloc, 975 metaslab_ndf_fragmented 976}; 977 978metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops; 979 980/* 981 * ========================================================================== 982 * Metaslabs 983 * ========================================================================== 984 */ 985 986/* 987 * Wait for any in-progress metaslab loads to complete. 988 */ 989void 990metaslab_load_wait(metaslab_t *msp) 991{ 992 ASSERT(MUTEX_HELD(&msp->ms_lock)); 993 994 while (msp->ms_loading) { 995 ASSERT(!msp->ms_loaded); 996 cv_wait(&msp->ms_load_cv, &msp->ms_lock); 997 } 998} 999 1000int 1001metaslab_load(metaslab_t *msp) 1002{ 1003 int error = 0; 1004 1005 ASSERT(MUTEX_HELD(&msp->ms_lock)); 1006 ASSERT(!msp->ms_loaded); 1007 ASSERT(!msp->ms_loading); 1008 1009 msp->ms_loading = B_TRUE; 1010 1011 /* 1012 * If the space map has not been allocated yet, then treat 1013 * all the space in the metaslab as free and add it to the 1014 * ms_tree. 1015 */ 1016 if (msp->ms_sm != NULL) 1017 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE); 1018 else 1019 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size); 1020 1021 msp->ms_loaded = (error == 0); 1022 msp->ms_loading = B_FALSE; 1023 1024 if (msp->ms_loaded) { 1025 for (int t = 0; t < TXG_DEFER_SIZE; t++) { 1026 range_tree_walk(msp->ms_defertree[t], 1027 range_tree_remove, msp->ms_tree); 1028 } 1029 } 1030 cv_broadcast(&msp->ms_load_cv); 1031 return (error); 1032} 1033 1034void 1035metaslab_unload(metaslab_t *msp) 1036{ 1037 ASSERT(MUTEX_HELD(&msp->ms_lock)); 1038 range_tree_vacate(msp->ms_tree, NULL, NULL); 1039 msp->ms_loaded = B_FALSE; 1040 msp->ms_weight &= ~METASLAB_ACTIVE_MASK; 1041} 1042 1043metaslab_t * 1044metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg) 1045{ 1046 vdev_t *vd = mg->mg_vd; 1047 objset_t *mos = vd->vdev_spa->spa_meta_objset; 1048 metaslab_t *msp; 1049 1050 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP); 1051 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL); 1052 cv_init(&msp->ms_load_cv, NULL, CV_DEFAULT, NULL); 1053 msp->ms_id = id; 1054 msp->ms_start = id << vd->vdev_ms_shift; 1055 msp->ms_size = 1ULL << vd->vdev_ms_shift; 1056 1057 /* 1058 * We only open space map objects that already exist. All others 1059 * will be opened when we finally allocate an object for it. 1060 */ 1061 if (object != 0) { 1062 VERIFY0(space_map_open(&msp->ms_sm, mos, object, msp->ms_start, 1063 msp->ms_size, vd->vdev_ashift, &msp->ms_lock)); 1064 ASSERT(msp->ms_sm != NULL); 1065 } 1066 1067 /* 1068 * We create the main range tree here, but we don't create the 1069 * alloctree and freetree until metaslab_sync_done(). This serves 1070 * two purposes: it allows metaslab_sync_done() to detect the 1071 * addition of new space; and for debugging, it ensures that we'd 1072 * data fault on any attempt to use this metaslab before it's ready. 1073 */ 1074 msp->ms_tree = range_tree_create(&metaslab_rt_ops, msp, &msp->ms_lock); 1075 metaslab_group_add(mg, msp); 1076 1077 msp->ms_ops = mg->mg_class->mc_ops; 1078 1079 /* 1080 * If we're opening an existing pool (txg == 0) or creating 1081 * a new one (txg == TXG_INITIAL), all space is available now. 1082 * If we're adding space to an existing pool, the new space 1083 * does not become available until after this txg has synced. 1084 */ 1085 if (txg <= TXG_INITIAL) 1086 metaslab_sync_done(msp, 0); 1087 1088 /* 1089 * If metaslab_debug_load is set and we're initializing a metaslab 1090 * that has an allocated space_map object then load the its space 1091 * map so that can verify frees. 1092 */ 1093 if (metaslab_debug_load && msp->ms_sm != NULL) { 1094 mutex_enter(&msp->ms_lock); 1095 VERIFY0(metaslab_load(msp)); 1096 mutex_exit(&msp->ms_lock); 1097 } 1098 1099 if (txg != 0) { 1100 vdev_dirty(vd, 0, NULL, txg); 1101 vdev_dirty(vd, VDD_METASLAB, msp, txg); 1102 } 1103 1104 return (msp); 1105} 1106 1107void 1108metaslab_fini(metaslab_t *msp) 1109{ 1110 metaslab_group_t *mg = msp->ms_group; 1111 1112 metaslab_group_remove(mg, msp); 1113 1114 mutex_enter(&msp->ms_lock); 1115 1116 VERIFY(msp->ms_group == NULL); 1117 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm), 1118 0, -msp->ms_size); 1119 space_map_close(msp->ms_sm); 1120 1121 metaslab_unload(msp); 1122 range_tree_destroy(msp->ms_tree); 1123 1124 for (int t = 0; t < TXG_SIZE; t++) { 1125 range_tree_destroy(msp->ms_alloctree[t]); 1126 range_tree_destroy(msp->ms_freetree[t]); 1127 } 1128 1129 for (int t = 0; t < TXG_DEFER_SIZE; t++) { 1130 range_tree_destroy(msp->ms_defertree[t]); 1131 } 1132 1133 ASSERT0(msp->ms_deferspace); 1134 1135 mutex_exit(&msp->ms_lock); 1136 cv_destroy(&msp->ms_load_cv); 1137 mutex_destroy(&msp->ms_lock); 1138 1139 kmem_free(msp, sizeof (metaslab_t)); 1140} 1141 1142/* 1143 * Apply a weighting factor based on the histogram information for this 1144 * metaslab. The current weighting factor is somewhat arbitrary and requires 1145 * additional investigation. The implementation provides a measure of 1146 * "weighted" free space and gives a higher weighting for larger contiguous 1147 * regions. The weighting factor is determined by counting the number of 1148 * sm_shift sectors that exist in each region represented by the histogram. 1149 * That value is then multiplied by the power of 2 exponent and the sm_shift 1150 * value. 1151 * 1152 * For example, assume the 2^21 histogram bucket has 4 2MB regions and the 1153 * metaslab has an sm_shift value of 9 (512B): 1154 * 1155 * 1) calculate the number of sm_shift sectors in the region: 1156 * 2^21 / 2^9 = 2^12 = 4096 * 4 (number of regions) = 16384 1157 * 2) multiply by the power of 2 exponent and the sm_shift value: 1158 * 16384 * 21 * 9 = 3096576 1159 * This value will be added to the weighting of the metaslab. 1160 */ 1161static uint64_t 1162metaslab_weight_factor(metaslab_t *msp) 1163{ 1164 uint64_t factor = 0; 1165 uint64_t sectors; 1166 int i; 1167 1168 /* 1169 * A null space map means that the entire metaslab is free, 1170 * calculate a weight factor that spans the entire size of the 1171 * metaslab. 1172 */ 1173 if (msp->ms_sm == NULL) { 1174 vdev_t *vd = msp->ms_group->mg_vd; 1175 1176 i = highbit64(msp->ms_size) - 1; 1177 sectors = msp->ms_size >> vd->vdev_ashift; 1178 return (sectors * i * vd->vdev_ashift); 1179 } 1180 1181 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) 1182 return (0); 1183 1184 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE(msp->ms_sm); i++) { 1185 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0) 1186 continue; 1187 1188 /* 1189 * Determine the number of sm_shift sectors in the region 1190 * indicated by the histogram. For example, given an 1191 * sm_shift value of 9 (512 bytes) and i = 4 then we know 1192 * that we're looking at an 8K region in the histogram 1193 * (i.e. 9 + 4 = 13, 2^13 = 8192). To figure out the 1194 * number of sm_shift sectors (512 bytes in this example), 1195 * we would take 8192 / 512 = 16. Since the histogram 1196 * is offset by sm_shift we can simply use the value of 1197 * of i to calculate this (i.e. 2^i = 16 where i = 4). 1198 */ 1199 sectors = msp->ms_sm->sm_phys->smp_histogram[i] << i; 1200 factor += (i + msp->ms_sm->sm_shift) * sectors; 1201 } 1202 return (factor * msp->ms_sm->sm_shift); 1203} 1204 1205static uint64_t 1206metaslab_weight(metaslab_t *msp) 1207{ 1208 metaslab_group_t *mg = msp->ms_group; 1209 vdev_t *vd = mg->mg_vd; 1210 uint64_t weight, space; 1211 1212 ASSERT(MUTEX_HELD(&msp->ms_lock)); 1213 1214 /* 1215 * This vdev is in the process of being removed so there is nothing 1216 * for us to do here. 1217 */ 1218 if (vd->vdev_removing) { 1219 ASSERT0(space_map_allocated(msp->ms_sm)); 1220 ASSERT0(vd->vdev_ms_shift); 1221 return (0); 1222 } 1223 1224 /* 1225 * The baseline weight is the metaslab's free space. 1226 */ 1227 space = msp->ms_size - space_map_allocated(msp->ms_sm); 1228 weight = space; 1229 1230 /* 1231 * Modern disks have uniform bit density and constant angular velocity. 1232 * Therefore, the outer recording zones are faster (higher bandwidth) 1233 * than the inner zones by the ratio of outer to inner track diameter, 1234 * which is typically around 2:1. We account for this by assigning 1235 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x). 1236 * In effect, this means that we'll select the metaslab with the most 1237 * free bandwidth rather than simply the one with the most free space. 1238 */ 1239 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count; 1240 ASSERT(weight >= space && weight <= 2 * space); 1241 1242 msp->ms_factor = metaslab_weight_factor(msp); 1243 if (metaslab_weight_factor_enable) 1244 weight += msp->ms_factor; 1245 1246 if (msp->ms_loaded && !msp->ms_ops->msop_fragmented(msp)) { 1247 /* 1248 * If this metaslab is one we're actively using, adjust its 1249 * weight to make it preferable to any inactive metaslab so 1250 * we'll polish it off. 1251 */ 1252 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK); 1253 } 1254 1255 return (weight); 1256} 1257 1258static int 1259metaslab_activate(metaslab_t *msp, uint64_t activation_weight) 1260{ 1261 ASSERT(MUTEX_HELD(&msp->ms_lock)); 1262 1263 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) { 1264 metaslab_load_wait(msp); 1265 if (!msp->ms_loaded) { 1266 int error = metaslab_load(msp); 1267 if (error) { 1268 metaslab_group_sort(msp->ms_group, msp, 0); 1269 return (error); 1270 } 1271 } 1272 1273 metaslab_group_sort(msp->ms_group, msp, 1274 msp->ms_weight | activation_weight); 1275 } 1276 ASSERT(msp->ms_loaded); 1277 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); 1278 1279 return (0); 1280} 1281 1282static void 1283metaslab_passivate(metaslab_t *msp, uint64_t size) 1284{ 1285 /* 1286 * If size < SPA_MINBLOCKSIZE, then we will not allocate from 1287 * this metaslab again. In that case, it had better be empty, 1288 * or we would be leaving space on the table. 1289 */ 1290 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0); 1291 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size)); 1292 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0); 1293} 1294 1295static void 1296metaslab_preload(void *arg) 1297{ 1298 metaslab_t *msp = arg; 1299 spa_t *spa = msp->ms_group->mg_vd->vdev_spa; 1300 1301 mutex_enter(&msp->ms_lock); 1302 metaslab_load_wait(msp); 1303 if (!msp->ms_loaded) 1304 (void) metaslab_load(msp); 1305 1306 /* 1307 * Set the ms_access_txg value so that we don't unload it right away. 1308 */ 1309 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1; 1310 mutex_exit(&msp->ms_lock); 1311} 1312 1313static void 1314metaslab_group_preload(metaslab_group_t *mg) 1315{ 1316 spa_t *spa = mg->mg_vd->vdev_spa; 1317 metaslab_t *msp; 1318 avl_tree_t *t = &mg->mg_metaslab_tree; 1319 int m = 0; 1320 1321 if (spa_shutting_down(spa) || !metaslab_preload_enabled) { 1322 taskq_wait(mg->mg_taskq); 1323 return; 1324 } 1325 mutex_enter(&mg->mg_lock); 1326 1327 /* 1328 * Prefetch the next potential metaslabs 1329 */ 1330 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) { 1331 1332 /* If we have reached our preload limit then we're done */ 1333 if (++m > metaslab_preload_limit) 1334 break; 1335 1336 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload, 1337 msp, TQ_SLEEP) != 0); 1338 } 1339 mutex_exit(&mg->mg_lock); 1340} 1341 1342/* 1343 * Determine if the space map's on-disk footprint is past our tolerance 1344 * for inefficiency. We would like to use the following criteria to make 1345 * our decision: 1346 * 1347 * 1. The size of the space map object should not dramatically increase as a 1348 * result of writing out the free space range tree. 1349 * 1350 * 2. The minimal on-disk space map representation is zfs_condense_pct/100 1351 * times the size than the free space range tree representation 1352 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB). 1353 * 1354 * Checking the first condition is tricky since we don't want to walk 1355 * the entire AVL tree calculating the estimated on-disk size. Instead we 1356 * use the size-ordered range tree in the metaslab and calculate the 1357 * size required to write out the largest segment in our free tree. If the 1358 * size required to represent that segment on disk is larger than the space 1359 * map object then we avoid condensing this map. 1360 * 1361 * To determine the second criterion we use a best-case estimate and assume 1362 * each segment can be represented on-disk as a single 64-bit entry. We refer 1363 * to this best-case estimate as the space map's minimal form. 1364 */ 1365static boolean_t 1366metaslab_should_condense(metaslab_t *msp) 1367{ 1368 space_map_t *sm = msp->ms_sm; 1369 range_seg_t *rs; 1370 uint64_t size, entries, segsz; 1371 1372 ASSERT(MUTEX_HELD(&msp->ms_lock)); 1373 ASSERT(msp->ms_loaded); 1374 1375 /* 1376 * Use the ms_size_tree range tree, which is ordered by size, to 1377 * obtain the largest segment in the free tree. If the tree is empty 1378 * then we should condense the map. 1379 */ 1380 rs = avl_last(&msp->ms_size_tree); 1381 if (rs == NULL) 1382 return (B_TRUE); 1383 1384 /* 1385 * Calculate the number of 64-bit entries this segment would 1386 * require when written to disk. If this single segment would be 1387 * larger on-disk than the entire current on-disk structure, then 1388 * clearly condensing will increase the on-disk structure size. 1389 */ 1390 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift; 1391 entries = size / (MIN(size, SM_RUN_MAX)); 1392 segsz = entries * sizeof (uint64_t); 1393 1394 return (segsz <= space_map_length(msp->ms_sm) && 1395 space_map_length(msp->ms_sm) >= (zfs_condense_pct * 1396 sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root)) / 100); 1397} 1398 1399/* 1400 * Condense the on-disk space map representation to its minimized form. 1401 * The minimized form consists of a small number of allocations followed by 1402 * the entries of the free range tree. 1403 */ 1404static void 1405metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx) 1406{ 1407 spa_t *spa = msp->ms_group->mg_vd->vdev_spa; 1408 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK]; 1409 range_tree_t *condense_tree; 1410 space_map_t *sm = msp->ms_sm; 1411 1412 ASSERT(MUTEX_HELD(&msp->ms_lock)); 1413 ASSERT3U(spa_sync_pass(spa), ==, 1); 1414 ASSERT(msp->ms_loaded); 1415 1416 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, " 1417 "smp size %llu, segments %lu", txg, msp->ms_id, msp, 1418 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root)); 1419 1420 /* 1421 * Create an range tree that is 100% allocated. We remove segments 1422 * that have been freed in this txg, any deferred frees that exist, 1423 * and any allocation in the future. Removing segments should be 1424 * a relatively inexpensive operation since we expect these trees to 1425 * have a small number of nodes. 1426 */ 1427 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock); 1428 range_tree_add(condense_tree, msp->ms_start, msp->ms_size); 1429 1430 /* 1431 * Remove what's been freed in this txg from the condense_tree. 1432 * Since we're in sync_pass 1, we know that all the frees from 1433 * this txg are in the freetree. 1434 */ 1435 range_tree_walk(freetree, range_tree_remove, condense_tree); 1436 1437 for (int t = 0; t < TXG_DEFER_SIZE; t++) { 1438 range_tree_walk(msp->ms_defertree[t], 1439 range_tree_remove, condense_tree); 1440 } 1441 1442 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) { 1443 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK], 1444 range_tree_remove, condense_tree); 1445 } 1446 1447 /* 1448 * We're about to drop the metaslab's lock thus allowing 1449 * other consumers to change it's content. Set the 1450 * metaslab's ms_condensing flag to ensure that 1451 * allocations on this metaslab do not occur while we're 1452 * in the middle of committing it to disk. This is only critical 1453 * for the ms_tree as all other range trees use per txg 1454 * views of their content. 1455 */ 1456 msp->ms_condensing = B_TRUE; 1457 1458 mutex_exit(&msp->ms_lock); 1459 space_map_truncate(sm, tx); 1460 mutex_enter(&msp->ms_lock); 1461 1462 /* 1463 * While we would ideally like to create a space_map representation 1464 * that consists only of allocation records, doing so can be 1465 * prohibitively expensive because the in-core free tree can be 1466 * large, and therefore computationally expensive to subtract 1467 * from the condense_tree. Instead we sync out two trees, a cheap 1468 * allocation only tree followed by the in-core free tree. While not 1469 * optimal, this is typically close to optimal, and much cheaper to 1470 * compute. 1471 */ 1472 space_map_write(sm, condense_tree, SM_ALLOC, tx); 1473 range_tree_vacate(condense_tree, NULL, NULL); 1474 range_tree_destroy(condense_tree); 1475 1476 space_map_write(sm, msp->ms_tree, SM_FREE, tx); 1477 msp->ms_condensing = B_FALSE; 1478} 1479 1480/* 1481 * Write a metaslab to disk in the context of the specified transaction group. 1482 */ 1483void 1484metaslab_sync(metaslab_t *msp, uint64_t txg) 1485{ 1486 metaslab_group_t *mg = msp->ms_group; 1487 vdev_t *vd = mg->mg_vd; 1488 spa_t *spa = vd->vdev_spa; 1489 objset_t *mos = spa_meta_objset(spa); 1490 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK]; 1491 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK]; 1492 range_tree_t **freed_tree = 1493 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]; 1494 dmu_tx_t *tx; 1495 uint64_t object = space_map_object(msp->ms_sm); 1496 1497 ASSERT(!vd->vdev_ishole); 1498 1499 /* 1500 * This metaslab has just been added so there's no work to do now. 1501 */ 1502 if (*freetree == NULL) { 1503 ASSERT3P(alloctree, ==, NULL); 1504 return; 1505 } 1506 1507 ASSERT3P(alloctree, !=, NULL); 1508 ASSERT3P(*freetree, !=, NULL); 1509 ASSERT3P(*freed_tree, !=, NULL); 1510 1511 if (range_tree_space(alloctree) == 0 && 1512 range_tree_space(*freetree) == 0) 1513 return; 1514 1515 /* 1516 * The only state that can actually be changing concurrently with 1517 * metaslab_sync() is the metaslab's ms_tree. No other thread can 1518 * be modifying this txg's alloctree, freetree, freed_tree, or 1519 * space_map_phys_t. Therefore, we only hold ms_lock to satify 1520 * space_map ASSERTs. We drop it whenever we call into the DMU, 1521 * because the DMU can call down to us (e.g. via zio_free()) at 1522 * any time. 1523 */ 1524 1525 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg); 1526 1527 if (msp->ms_sm == NULL) { 1528 uint64_t new_object; 1529 1530 new_object = space_map_alloc(mos, tx); 1531 VERIFY3U(new_object, !=, 0); 1532 1533 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object, 1534 msp->ms_start, msp->ms_size, vd->vdev_ashift, 1535 &msp->ms_lock)); 1536 ASSERT(msp->ms_sm != NULL); 1537 } 1538 1539 mutex_enter(&msp->ms_lock); 1540 1541 if (msp->ms_loaded && spa_sync_pass(spa) == 1 && 1542 metaslab_should_condense(msp)) { 1543 metaslab_condense(msp, txg, tx); 1544 } else { 1545 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx); 1546 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx); 1547 } 1548 1549 range_tree_vacate(alloctree, NULL, NULL); 1550 1551 if (msp->ms_loaded) { 1552 /* 1553 * When the space map is loaded, we have an accruate 1554 * histogram in the range tree. This gives us an opportunity 1555 * to bring the space map's histogram up-to-date so we clear 1556 * it first before updating it. 1557 */ 1558 space_map_histogram_clear(msp->ms_sm); 1559 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx); 1560 } else { 1561 /* 1562 * Since the space map is not loaded we simply update the 1563 * exisiting histogram with what was freed in this txg. This 1564 * means that the on-disk histogram may not have an accurate 1565 * view of the free space but it's close enough to allow 1566 * us to make allocation decisions. 1567 */ 1568 space_map_histogram_add(msp->ms_sm, *freetree, tx); 1569 } 1570 1571 /* 1572 * For sync pass 1, we avoid traversing this txg's free range tree 1573 * and instead will just swap the pointers for freetree and 1574 * freed_tree. We can safely do this since the freed_tree is 1575 * guaranteed to be empty on the initial pass. 1576 */ 1577 if (spa_sync_pass(spa) == 1) { 1578 range_tree_swap(freetree, freed_tree); 1579 } else { 1580 range_tree_vacate(*freetree, range_tree_add, *freed_tree); 1581 } 1582 1583 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK])); 1584 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK])); 1585 1586 mutex_exit(&msp->ms_lock); 1587 1588 if (object != space_map_object(msp->ms_sm)) { 1589 object = space_map_object(msp->ms_sm); 1590 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) * 1591 msp->ms_id, sizeof (uint64_t), &object, tx); 1592 } 1593 dmu_tx_commit(tx); 1594} 1595 1596/* 1597 * Called after a transaction group has completely synced to mark 1598 * all of the metaslab's free space as usable. 1599 */ 1600void 1601metaslab_sync_done(metaslab_t *msp, uint64_t txg) 1602{ 1603 metaslab_group_t *mg = msp->ms_group; 1604 vdev_t *vd = mg->mg_vd; 1605 range_tree_t **freed_tree; 1606 range_tree_t **defer_tree; 1607 int64_t alloc_delta, defer_delta; 1608 1609 ASSERT(!vd->vdev_ishole); 1610 1611 mutex_enter(&msp->ms_lock); 1612 1613 /* 1614 * If this metaslab is just becoming available, initialize its 1615 * alloctrees, freetrees, and defertree and add its capacity to 1616 * the vdev. 1617 */ 1618 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) { 1619 for (int t = 0; t < TXG_SIZE; t++) { 1620 ASSERT(msp->ms_alloctree[t] == NULL); 1621 ASSERT(msp->ms_freetree[t] == NULL); 1622 1623 msp->ms_alloctree[t] = range_tree_create(NULL, msp, 1624 &msp->ms_lock); 1625 msp->ms_freetree[t] = range_tree_create(NULL, msp, 1626 &msp->ms_lock); 1627 } 1628 1629 for (int t = 0; t < TXG_DEFER_SIZE; t++) { 1630 ASSERT(msp->ms_defertree[t] == NULL); 1631 1632 msp->ms_defertree[t] = range_tree_create(NULL, msp, 1633 &msp->ms_lock); 1634 } 1635 1636 vdev_space_update(vd, 0, 0, msp->ms_size); 1637 } 1638 1639 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]; 1640 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE]; 1641 1642 alloc_delta = space_map_alloc_delta(msp->ms_sm); 1643 defer_delta = range_tree_space(*freed_tree) - 1644 range_tree_space(*defer_tree); 1645 1646 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0); 1647 1648 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK])); 1649 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK])); 1650 1651 /* 1652 * If there's a metaslab_load() in progress, wait for it to complete 1653 * so that we have a consistent view of the in-core space map. 1654 */ 1655 metaslab_load_wait(msp); 1656 1657 /* 1658 * Move the frees from the defer_tree back to the free 1659 * range tree (if it's loaded). Swap the freed_tree and the 1660 * defer_tree -- this is safe to do because we've just emptied out 1661 * the defer_tree. 1662 */ 1663 range_tree_vacate(*defer_tree, 1664 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree); 1665 range_tree_swap(freed_tree, defer_tree); 1666 1667 space_map_update(msp->ms_sm); 1668 1669 msp->ms_deferspace += defer_delta; 1670 ASSERT3S(msp->ms_deferspace, >=, 0); 1671 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size); 1672 if (msp->ms_deferspace != 0) { 1673 /* 1674 * Keep syncing this metaslab until all deferred frees 1675 * are back in circulation. 1676 */ 1677 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1); 1678 } 1679 1680 if (msp->ms_loaded && msp->ms_access_txg < txg) { 1681 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) { 1682 VERIFY0(range_tree_space( 1683 msp->ms_alloctree[(txg + t) & TXG_MASK])); 1684 } 1685 1686 if (!metaslab_debug_unload) 1687 metaslab_unload(msp); 1688 } 1689 1690 metaslab_group_sort(mg, msp, metaslab_weight(msp)); 1691 mutex_exit(&msp->ms_lock); 1692 1693} 1694 1695void 1696metaslab_sync_reassess(metaslab_group_t *mg) 1697{ 1698 metaslab_group_alloc_update(mg); 1699 1700 /* 1701 * Preload the next potential metaslabs 1702 */ 1703 metaslab_group_preload(mg); 1704} 1705 1706static uint64_t 1707metaslab_distance(metaslab_t *msp, dva_t *dva) 1708{ 1709 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift; 1710 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift; 1711 uint64_t start = msp->ms_id; 1712 1713 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva)) 1714 return (1ULL << 63); 1715 1716 if (offset < start) 1717 return ((start - offset) << ms_shift); 1718 if (offset > start) 1719 return ((offset - start) << ms_shift); 1720 return (0); 1721} 1722 1723static uint64_t 1724metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize, 1725 uint64_t txg, uint64_t min_distance, dva_t *dva, int d) 1726{ 1727 spa_t *spa = mg->mg_vd->vdev_spa; 1728 metaslab_t *msp = NULL; 1729 uint64_t offset = -1ULL; 1730 avl_tree_t *t = &mg->mg_metaslab_tree; 1731 uint64_t activation_weight; 1732 uint64_t target_distance; 1733 int i; 1734 1735 activation_weight = METASLAB_WEIGHT_PRIMARY; 1736 for (i = 0; i < d; i++) { 1737 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) { 1738 activation_weight = METASLAB_WEIGHT_SECONDARY; 1739 break; 1740 } 1741 } 1742 1743 for (;;) { 1744 boolean_t was_active; 1745 1746 mutex_enter(&mg->mg_lock); 1747 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) { 1748 if (msp->ms_weight < asize) { 1749 spa_dbgmsg(spa, "%s: failed to meet weight " 1750 "requirement: vdev %llu, txg %llu, mg %p, " 1751 "msp %p, psize %llu, asize %llu, " 1752 "weight %llu", spa_name(spa), 1753 mg->mg_vd->vdev_id, txg, 1754 mg, msp, psize, asize, msp->ms_weight); 1755 mutex_exit(&mg->mg_lock); 1756 return (-1ULL); 1757 } 1758 1759 /* 1760 * If the selected metaslab is condensing, skip it. 1761 */ 1762 if (msp->ms_condensing) 1763 continue; 1764 1765 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK; 1766 if (activation_weight == METASLAB_WEIGHT_PRIMARY) 1767 break; 1768 1769 target_distance = min_distance + 1770 (space_map_allocated(msp->ms_sm) != 0 ? 0 : 1771 min_distance >> 1); 1772 1773 for (i = 0; i < d; i++) 1774 if (metaslab_distance(msp, &dva[i]) < 1775 target_distance) 1776 break; 1777 if (i == d) 1778 break; 1779 } 1780 mutex_exit(&mg->mg_lock); 1781 if (msp == NULL) 1782 return (-1ULL); 1783 1784 mutex_enter(&msp->ms_lock); 1785 1786 /* 1787 * Ensure that the metaslab we have selected is still 1788 * capable of handling our request. It's possible that 1789 * another thread may have changed the weight while we 1790 * were blocked on the metaslab lock. 1791 */ 1792 if (msp->ms_weight < asize || (was_active && 1793 !(msp->ms_weight & METASLAB_ACTIVE_MASK) && 1794 activation_weight == METASLAB_WEIGHT_PRIMARY)) { 1795 mutex_exit(&msp->ms_lock); 1796 continue; 1797 } 1798 1799 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) && 1800 activation_weight == METASLAB_WEIGHT_PRIMARY) { 1801 metaslab_passivate(msp, 1802 msp->ms_weight & ~METASLAB_ACTIVE_MASK); 1803 mutex_exit(&msp->ms_lock); 1804 continue; 1805 } 1806 1807 if (metaslab_activate(msp, activation_weight) != 0) { 1808 mutex_exit(&msp->ms_lock); 1809 continue; 1810 } 1811 1812 /* 1813 * If this metaslab is currently condensing then pick again as 1814 * we can't manipulate this metaslab until it's committed 1815 * to disk. 1816 */ 1817 if (msp->ms_condensing) { 1818 mutex_exit(&msp->ms_lock); 1819 continue; 1820 } 1821 1822 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL) 1823 break; 1824 1825 metaslab_passivate(msp, metaslab_block_maxsize(msp)); 1826 mutex_exit(&msp->ms_lock); 1827 } 1828 1829 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0) 1830 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg); 1831 1832 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize); 1833 msp->ms_access_txg = txg + metaslab_unload_delay; 1834 1835 mutex_exit(&msp->ms_lock); 1836 1837 return (offset); 1838} 1839 1840/* 1841 * Allocate a block for the specified i/o. 1842 */ 1843static int 1844metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize, 1845 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags) 1846{ 1847 metaslab_group_t *mg, *rotor; 1848 vdev_t *vd; 1849 int dshift = 3; 1850 int all_zero; 1851 int zio_lock = B_FALSE; 1852 boolean_t allocatable; 1853 uint64_t offset = -1ULL; 1854 uint64_t asize; 1855 uint64_t distance; 1856 1857 ASSERT(!DVA_IS_VALID(&dva[d])); 1858 1859 /* 1860 * For testing, make some blocks above a certain size be gang blocks. 1861 */ 1862 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) 1863 return (SET_ERROR(ENOSPC)); 1864 1865 /* 1866 * Start at the rotor and loop through all mgs until we find something. 1867 * Note that there's no locking on mc_rotor or mc_aliquot because 1868 * nothing actually breaks if we miss a few updates -- we just won't 1869 * allocate quite as evenly. It all balances out over time. 1870 * 1871 * If we are doing ditto or log blocks, try to spread them across 1872 * consecutive vdevs. If we're forced to reuse a vdev before we've 1873 * allocated all of our ditto blocks, then try and spread them out on 1874 * that vdev as much as possible. If it turns out to not be possible, 1875 * gradually lower our standards until anything becomes acceptable. 1876 * Also, allocating on consecutive vdevs (as opposed to random vdevs) 1877 * gives us hope of containing our fault domains to something we're 1878 * able to reason about. Otherwise, any two top-level vdev failures 1879 * will guarantee the loss of data. With consecutive allocation, 1880 * only two adjacent top-level vdev failures will result in data loss. 1881 * 1882 * If we are doing gang blocks (hintdva is non-NULL), try to keep 1883 * ourselves on the same vdev as our gang block header. That 1884 * way, we can hope for locality in vdev_cache, plus it makes our 1885 * fault domains something tractable. 1886 */ 1887 if (hintdva) { 1888 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d])); 1889 1890 /* 1891 * It's possible the vdev we're using as the hint no 1892 * longer exists (i.e. removed). Consult the rotor when 1893 * all else fails. 1894 */ 1895 if (vd != NULL) { 1896 mg = vd->vdev_mg; 1897 1898 if (flags & METASLAB_HINTBP_AVOID && 1899 mg->mg_next != NULL) 1900 mg = mg->mg_next; 1901 } else { 1902 mg = mc->mc_rotor; 1903 } 1904 } else if (d != 0) { 1905 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1])); 1906 mg = vd->vdev_mg->mg_next; 1907 } else { 1908 mg = mc->mc_rotor; 1909 } 1910 1911 /* 1912 * If the hint put us into the wrong metaslab class, or into a 1913 * metaslab group that has been passivated, just follow the rotor. 1914 */ 1915 if (mg->mg_class != mc || mg->mg_activation_count <= 0) 1916 mg = mc->mc_rotor; 1917 1918 rotor = mg; 1919top: 1920 all_zero = B_TRUE; 1921 do { 1922 ASSERT(mg->mg_activation_count == 1); 1923 1924 vd = mg->mg_vd; 1925 1926 /* 1927 * Don't allocate from faulted devices. 1928 */ 1929 if (zio_lock) { 1930 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER); 1931 allocatable = vdev_allocatable(vd); 1932 spa_config_exit(spa, SCL_ZIO, FTAG); 1933 } else { 1934 allocatable = vdev_allocatable(vd); 1935 } 1936 1937 /* 1938 * Determine if the selected metaslab group is eligible 1939 * for allocations. If we're ganging or have requested 1940 * an allocation for the smallest gang block size 1941 * then we don't want to avoid allocating to the this 1942 * metaslab group. If we're in this condition we should 1943 * try to allocate from any device possible so that we 1944 * don't inadvertently return ENOSPC and suspend the pool 1945 * even though space is still available. 1946 */ 1947 if (allocatable && CAN_FASTGANG(flags) && 1948 psize > SPA_GANGBLOCKSIZE) 1949 allocatable = metaslab_group_allocatable(mg); 1950 1951 if (!allocatable) 1952 goto next; 1953 1954 /* 1955 * Avoid writing single-copy data to a failing vdev 1956 * unless the user instructs us that it is okay. 1957 */ 1958 if ((vd->vdev_stat.vs_write_errors > 0 || 1959 vd->vdev_state < VDEV_STATE_HEALTHY) && 1960 d == 0 && dshift == 3 && 1961 !(zfs_write_to_degraded && vd->vdev_state == 1962 VDEV_STATE_DEGRADED)) { 1963 all_zero = B_FALSE; 1964 goto next; 1965 } 1966 1967 ASSERT(mg->mg_class == mc); 1968 1969 distance = vd->vdev_asize >> dshift; 1970 if (distance <= (1ULL << vd->vdev_ms_shift)) 1971 distance = 0; 1972 else 1973 all_zero = B_FALSE; 1974 1975 asize = vdev_psize_to_asize(vd, psize); 1976 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0); 1977 1978 offset = metaslab_group_alloc(mg, psize, asize, txg, distance, 1979 dva, d); 1980 if (offset != -1ULL) { 1981 /* 1982 * If we've just selected this metaslab group, 1983 * figure out whether the corresponding vdev is 1984 * over- or under-used relative to the pool, 1985 * and set an allocation bias to even it out. 1986 */ 1987 if (mc->mc_aliquot == 0) { 1988 vdev_stat_t *vs = &vd->vdev_stat; 1989 int64_t vu, cu; 1990 1991 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1); 1992 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1); 1993 1994 /* 1995 * Calculate how much more or less we should 1996 * try to allocate from this device during 1997 * this iteration around the rotor. 1998 * For example, if a device is 80% full 1999 * and the pool is 20% full then we should 2000 * reduce allocations by 60% on this device. 2001 * 2002 * mg_bias = (20 - 80) * 512K / 100 = -307K 2003 * 2004 * This reduces allocations by 307K for this 2005 * iteration. 2006 */ 2007 mg->mg_bias = ((cu - vu) * 2008 (int64_t)mg->mg_aliquot) / 100; 2009 } 2010 2011 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >= 2012 mg->mg_aliquot + mg->mg_bias) { 2013 mc->mc_rotor = mg->mg_next; 2014 mc->mc_aliquot = 0; 2015 } 2016 2017 DVA_SET_VDEV(&dva[d], vd->vdev_id); 2018 DVA_SET_OFFSET(&dva[d], offset); 2019 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER)); 2020 DVA_SET_ASIZE(&dva[d], asize); 2021 2022 return (0); 2023 } 2024next: 2025 mc->mc_rotor = mg->mg_next; 2026 mc->mc_aliquot = 0; 2027 } while ((mg = mg->mg_next) != rotor); 2028 2029 if (!all_zero) { 2030 dshift++; 2031 ASSERT(dshift < 64); 2032 goto top; 2033 } 2034 2035 if (!allocatable && !zio_lock) { 2036 dshift = 3; 2037 zio_lock = B_TRUE; 2038 goto top; 2039 } 2040 2041 bzero(&dva[d], sizeof (dva_t)); 2042 2043 return (SET_ERROR(ENOSPC)); 2044} 2045 2046/* 2047 * Free the block represented by DVA in the context of the specified 2048 * transaction group. 2049 */ 2050static void 2051metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now) 2052{ 2053 uint64_t vdev = DVA_GET_VDEV(dva); 2054 uint64_t offset = DVA_GET_OFFSET(dva); 2055 uint64_t size = DVA_GET_ASIZE(dva); 2056 vdev_t *vd; 2057 metaslab_t *msp; 2058 2059 ASSERT(DVA_IS_VALID(dva)); 2060 2061 if (txg > spa_freeze_txg(spa)) 2062 return; 2063 2064 if ((vd = vdev_lookup_top(spa, vdev)) == NULL || 2065 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) { 2066 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu", 2067 (u_longlong_t)vdev, (u_longlong_t)offset); 2068 ASSERT(0); 2069 return; 2070 } 2071 2072 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; 2073 2074 if (DVA_GET_GANG(dva)) 2075 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 2076 2077 mutex_enter(&msp->ms_lock); 2078 2079 if (now) { 2080 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK], 2081 offset, size); 2082 2083 VERIFY(!msp->ms_condensing); 2084 VERIFY3U(offset, >=, msp->ms_start); 2085 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size); 2086 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=, 2087 msp->ms_size); 2088 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); 2089 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); 2090 range_tree_add(msp->ms_tree, offset, size); 2091 } else { 2092 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0) 2093 vdev_dirty(vd, VDD_METASLAB, msp, txg); 2094 range_tree_add(msp->ms_freetree[txg & TXG_MASK], 2095 offset, size); 2096 } 2097 2098 mutex_exit(&msp->ms_lock); 2099} 2100 2101/* 2102 * Intent log support: upon opening the pool after a crash, notify the SPA 2103 * of blocks that the intent log has allocated for immediate write, but 2104 * which are still considered free by the SPA because the last transaction 2105 * group didn't commit yet. 2106 */ 2107static int 2108metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg) 2109{ 2110 uint64_t vdev = DVA_GET_VDEV(dva); 2111 uint64_t offset = DVA_GET_OFFSET(dva); 2112 uint64_t size = DVA_GET_ASIZE(dva); 2113 vdev_t *vd; 2114 metaslab_t *msp; 2115 int error = 0; 2116 2117 ASSERT(DVA_IS_VALID(dva)); 2118 2119 if ((vd = vdev_lookup_top(spa, vdev)) == NULL || 2120 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) 2121 return (SET_ERROR(ENXIO)); 2122 2123 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; 2124 2125 if (DVA_GET_GANG(dva)) 2126 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 2127 2128 mutex_enter(&msp->ms_lock); 2129 2130 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded) 2131 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY); 2132 2133 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size)) 2134 error = SET_ERROR(ENOENT); 2135 2136 if (error || txg == 0) { /* txg == 0 indicates dry run */ 2137 mutex_exit(&msp->ms_lock); 2138 return (error); 2139 } 2140 2141 VERIFY(!msp->ms_condensing); 2142 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); 2143 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); 2144 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size); 2145 range_tree_remove(msp->ms_tree, offset, size); 2146 2147 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */ 2148 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0) 2149 vdev_dirty(vd, VDD_METASLAB, msp, txg); 2150 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size); 2151 } 2152 2153 mutex_exit(&msp->ms_lock); 2154 2155 return (0); 2156} 2157 2158int 2159metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp, 2160 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags) 2161{ 2162 dva_t *dva = bp->blk_dva; 2163 dva_t *hintdva = hintbp->blk_dva; 2164 int error = 0; 2165 2166 ASSERT(bp->blk_birth == 0); 2167 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0); 2168 2169 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); 2170 2171 if (mc->mc_rotor == NULL) { /* no vdevs in this class */ 2172 spa_config_exit(spa, SCL_ALLOC, FTAG); 2173 return (SET_ERROR(ENOSPC)); 2174 } 2175 2176 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa)); 2177 ASSERT(BP_GET_NDVAS(bp) == 0); 2178 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp)); 2179 2180 for (int d = 0; d < ndvas; d++) { 2181 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva, 2182 txg, flags); 2183 if (error != 0) { 2184 for (d--; d >= 0; d--) { 2185 metaslab_free_dva(spa, &dva[d], txg, B_TRUE); 2186 bzero(&dva[d], sizeof (dva_t)); 2187 } 2188 spa_config_exit(spa, SCL_ALLOC, FTAG); 2189 return (error); 2190 } 2191 } 2192 ASSERT(error == 0); 2193 ASSERT(BP_GET_NDVAS(bp) == ndvas); 2194 2195 spa_config_exit(spa, SCL_ALLOC, FTAG); 2196 2197 BP_SET_BIRTH(bp, txg, txg); 2198 2199 return (0); 2200} 2201 2202void 2203metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now) 2204{ 2205 const dva_t *dva = bp->blk_dva; 2206 int ndvas = BP_GET_NDVAS(bp); 2207 2208 ASSERT(!BP_IS_HOLE(bp)); 2209 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa)); 2210 2211 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER); 2212 2213 for (int d = 0; d < ndvas; d++) 2214 metaslab_free_dva(spa, &dva[d], txg, now); 2215 2216 spa_config_exit(spa, SCL_FREE, FTAG); 2217} 2218 2219int 2220metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg) 2221{ 2222 const dva_t *dva = bp->blk_dva; 2223 int ndvas = BP_GET_NDVAS(bp); 2224 int error = 0; 2225 2226 ASSERT(!BP_IS_HOLE(bp)); 2227 2228 if (txg != 0) { 2229 /* 2230 * First do a dry run to make sure all DVAs are claimable, 2231 * so we don't have to unwind from partial failures below. 2232 */ 2233 if ((error = metaslab_claim(spa, bp, 0)) != 0) 2234 return (error); 2235 } 2236 2237 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); 2238 2239 for (int d = 0; d < ndvas; d++) 2240 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0) 2241 break; 2242 2243 spa_config_exit(spa, SCL_ALLOC, FTAG); 2244 2245 ASSERT(error == 0 || txg == 0); 2246 2247 return (error); 2248} 2249 2250void 2251metaslab_check_free(spa_t *spa, const blkptr_t *bp) 2252{ 2253 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0) 2254 return; 2255 2256 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); 2257 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 2258 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 2259 vdev_t *vd = vdev_lookup_top(spa, vdev); 2260 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 2261 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]); 2262 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; 2263 2264 if (msp->ms_loaded) 2265 range_tree_verify(msp->ms_tree, offset, size); 2266 2267 for (int j = 0; j < TXG_SIZE; j++) 2268 range_tree_verify(msp->ms_freetree[j], offset, size); 2269 for (int j = 0; j < TXG_DEFER_SIZE; j++) 2270 range_tree_verify(msp->ms_defertree[j], offset, size); 2271 } 2272 spa_config_exit(spa, SCL_VDEV, FTAG); 2273} 2274