1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5ecc2d604Sbonwick * Common Development and Distribution License (the "License"). 6ecc2d604Sbonwick * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 22d6e555bdSGeorge Wilson * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 2401f55e48SGeorge Wilson */ 2501f55e48SGeorge Wilson 2601f55e48SGeorge Wilson /* 27*0f7643c7SGeorge Wilson * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 28fa9e4066Sahrens */ 29fa9e4066Sahrens 30fa9e4066Sahrens #ifndef _SYS_METASLAB_IMPL_H 31fa9e4066Sahrens #define _SYS_METASLAB_IMPL_H 32fa9e4066Sahrens 33fa9e4066Sahrens #include <sys/metaslab.h> 34fa9e4066Sahrens #include <sys/space_map.h> 350713e232SGeorge Wilson #include <sys/range_tree.h> 36fa9e4066Sahrens #include <sys/vdev.h> 37fa9e4066Sahrens #include <sys/txg.h> 38fa9e4066Sahrens #include <sys/avl.h> 39fa9e4066Sahrens 40fa9e4066Sahrens #ifdef __cplusplus 41fa9e4066Sahrens extern "C" { 42fa9e4066Sahrens #endif 43fa9e4066Sahrens 442e4c9986SGeorge Wilson /* 452e4c9986SGeorge Wilson * A metaslab class encompasses a category of allocatable top-level vdevs. 462e4c9986SGeorge Wilson * Each top-level vdev is associated with a metaslab group which defines 472e4c9986SGeorge Wilson * the allocatable region for that vdev. Examples of these categories include 482e4c9986SGeorge Wilson * "normal" for data block allocations (i.e. main pool allocations) or "log" 492e4c9986SGeorge Wilson * for allocations designated for intent log devices (i.e. slog devices). 502e4c9986SGeorge Wilson * When a block allocation is requested from the SPA it is associated with a 512e4c9986SGeorge Wilson * metaslab_class_t, and only top-level vdevs (i.e. metaslab groups) belonging 522e4c9986SGeorge Wilson * to the class can be used to satisfy that request. Allocations are done 532e4c9986SGeorge Wilson * by traversing the metaslab groups that are linked off of the mc_rotor field. 542e4c9986SGeorge Wilson * This rotor points to the next metaslab group where allocations will be 552e4c9986SGeorge Wilson * attempted. Allocating a block is a 3 step process -- select the metaslab 562e4c9986SGeorge Wilson * group, select the metaslab, and then allocate the block. The metaslab 572e4c9986SGeorge Wilson * class defines the low-level block allocator that will be used as the 582e4c9986SGeorge Wilson * final step in allocation. These allocators are pluggable allowing each class 592e4c9986SGeorge Wilson * to use a block allocator that best suits that class. 602e4c9986SGeorge Wilson */ 61fa9e4066Sahrens struct metaslab_class { 62*0f7643c7SGeorge Wilson kmutex_t mc_lock; 6388ecc943SGeorge Wilson spa_t *mc_spa; 64fa9e4066Sahrens metaslab_group_t *mc_rotor; 650713e232SGeorge Wilson metaslab_ops_t *mc_ops; 66b24ab676SJeff Bonwick uint64_t mc_aliquot; 67*0f7643c7SGeorge Wilson 68*0f7643c7SGeorge Wilson /* 69*0f7643c7SGeorge Wilson * Track the number of metaslab groups that have been initialized 70*0f7643c7SGeorge Wilson * and can accept allocations. An initialized metaslab group is 71*0f7643c7SGeorge Wilson * one has been completely added to the config (i.e. we have 72*0f7643c7SGeorge Wilson * updated the MOS config and the space has been added to the pool). 73*0f7643c7SGeorge Wilson */ 74*0f7643c7SGeorge Wilson uint64_t mc_groups; 75*0f7643c7SGeorge Wilson 76*0f7643c7SGeorge Wilson /* 77*0f7643c7SGeorge Wilson * Toggle to enable/disable the allocation throttle. 78*0f7643c7SGeorge Wilson */ 79*0f7643c7SGeorge Wilson boolean_t mc_alloc_throttle_enabled; 80*0f7643c7SGeorge Wilson 81*0f7643c7SGeorge Wilson /* 82*0f7643c7SGeorge Wilson * The allocation throttle works on a reservation system. Whenever 83*0f7643c7SGeorge Wilson * an asynchronous zio wants to perform an allocation it must 84*0f7643c7SGeorge Wilson * first reserve the number of blocks that it wants to allocate. 85*0f7643c7SGeorge Wilson * If there aren't sufficient slots available for the pending zio 86*0f7643c7SGeorge Wilson * then that I/O is throttled until more slots free up. The current 87*0f7643c7SGeorge Wilson * number of reserved allocations is maintained by the mc_alloc_slots 88*0f7643c7SGeorge Wilson * refcount. The mc_alloc_max_slots value determines the maximum 89*0f7643c7SGeorge Wilson * number of allocations that the system allows. Gang blocks are 90*0f7643c7SGeorge Wilson * allowed to reserve slots even if we've reached the maximum 91*0f7643c7SGeorge Wilson * number of allocations allowed. 92*0f7643c7SGeorge Wilson */ 93*0f7643c7SGeorge Wilson uint64_t mc_alloc_max_slots; 94*0f7643c7SGeorge Wilson refcount_t mc_alloc_slots; 95*0f7643c7SGeorge Wilson 9622e30981SGeorge Wilson uint64_t mc_alloc_groups; /* # of allocatable groups */ 97*0f7643c7SGeorge Wilson 98b24ab676SJeff Bonwick uint64_t mc_alloc; /* total allocated space */ 99b24ab676SJeff Bonwick uint64_t mc_deferred; /* total deferred frees */ 100b24ab676SJeff Bonwick uint64_t mc_space; /* total space (alloc + free) */ 101b24ab676SJeff Bonwick uint64_t mc_dspace; /* total deflated space */ 1022e4c9986SGeorge Wilson uint64_t mc_histogram[RANGE_TREE_HISTOGRAM_SIZE]; 103fa9e4066Sahrens }; 104fa9e4066Sahrens 1052e4c9986SGeorge Wilson /* 1062e4c9986SGeorge Wilson * Metaslab groups encapsulate all the allocatable regions (i.e. metaslabs) 1072e4c9986SGeorge Wilson * of a top-level vdev. They are linked togther to form a circular linked 1082e4c9986SGeorge Wilson * list and can belong to only one metaslab class. Metaslab groups may become 1092e4c9986SGeorge Wilson * ineligible for allocations for a number of reasons such as limited free 1102e4c9986SGeorge Wilson * space, fragmentation, or going offline. When this happens the allocator will 1112e4c9986SGeorge Wilson * simply find the next metaslab group in the linked list and attempt 1122e4c9986SGeorge Wilson * to allocate from that group instead. 1132e4c9986SGeorge Wilson */ 114fa9e4066Sahrens struct metaslab_group { 115fa9e4066Sahrens kmutex_t mg_lock; 116fa9e4066Sahrens avl_tree_t mg_metaslab_tree; 117fa9e4066Sahrens uint64_t mg_aliquot; 11822e30981SGeorge Wilson boolean_t mg_allocatable; /* can we allocate? */ 119*0f7643c7SGeorge Wilson 120*0f7643c7SGeorge Wilson /* 121*0f7643c7SGeorge Wilson * A metaslab group is considered to be initialized only after 122*0f7643c7SGeorge Wilson * we have updated the MOS config and added the space to the pool. 123*0f7643c7SGeorge Wilson * We only allow allocation attempts to a metaslab group if it 124*0f7643c7SGeorge Wilson * has been initialized. 125*0f7643c7SGeorge Wilson */ 126*0f7643c7SGeorge Wilson boolean_t mg_initialized; 127*0f7643c7SGeorge Wilson 12822e30981SGeorge Wilson uint64_t mg_free_capacity; /* percentage free */ 129fa9e4066Sahrens int64_t mg_bias; 130a1521560SJeff Bonwick int64_t mg_activation_count; 131fa9e4066Sahrens metaslab_class_t *mg_class; 132fa9e4066Sahrens vdev_t *mg_vd; 1330713e232SGeorge Wilson taskq_t *mg_taskq; 134fa9e4066Sahrens metaslab_group_t *mg_prev; 135fa9e4066Sahrens metaslab_group_t *mg_next; 136*0f7643c7SGeorge Wilson 137*0f7643c7SGeorge Wilson /* 138*0f7643c7SGeorge Wilson * Each metaslab group can handle mg_max_alloc_queue_depth allocations 139*0f7643c7SGeorge Wilson * which are tracked by mg_alloc_queue_depth. It's possible for a 140*0f7643c7SGeorge Wilson * metaslab group to handle more allocations than its max. This 141*0f7643c7SGeorge Wilson * can occur when gang blocks are required or when other groups 142*0f7643c7SGeorge Wilson * are unable to handle their share of allocations. 143*0f7643c7SGeorge Wilson */ 144*0f7643c7SGeorge Wilson uint64_t mg_max_alloc_queue_depth; 145*0f7643c7SGeorge Wilson refcount_t mg_alloc_queue_depth; 146*0f7643c7SGeorge Wilson 147*0f7643c7SGeorge Wilson /* 148*0f7643c7SGeorge Wilson * A metalab group that can no longer allocate the minimum block 149*0f7643c7SGeorge Wilson * size will set mg_no_free_space. Once a metaslab group is out 150*0f7643c7SGeorge Wilson * of space then its share of work must be distributed to other 151*0f7643c7SGeorge Wilson * groups. 152*0f7643c7SGeorge Wilson */ 153*0f7643c7SGeorge Wilson boolean_t mg_no_free_space; 154*0f7643c7SGeorge Wilson 155*0f7643c7SGeorge Wilson uint64_t mg_allocations; 156*0f7643c7SGeorge Wilson uint64_t mg_failed_allocations; 1572e4c9986SGeorge Wilson uint64_t mg_fragmentation; 1582e4c9986SGeorge Wilson uint64_t mg_histogram[RANGE_TREE_HISTOGRAM_SIZE]; 159fa9e4066Sahrens }; 160fa9e4066Sahrens 161fa9e4066Sahrens /* 1620713e232SGeorge Wilson * This value defines the number of elements in the ms_lbas array. The value 1632e4c9986SGeorge Wilson * of 64 was chosen as it covers all power of 2 buckets up to UINT64_MAX. 1642e4c9986SGeorge Wilson * This is the equivalent of highbit(UINT64_MAX). 1650713e232SGeorge Wilson */ 1660713e232SGeorge Wilson #define MAX_LBAS 64 1670713e232SGeorge Wilson 1680713e232SGeorge Wilson /* 1690713e232SGeorge Wilson * Each metaslab maintains a set of in-core trees to track metaslab operations. 1700713e232SGeorge Wilson * The in-core free tree (ms_tree) contains the current list of free segments. 1710713e232SGeorge Wilson * As blocks are allocated, the allocated segment are removed from the ms_tree 1720713e232SGeorge Wilson * and added to a per txg allocation tree (ms_alloctree). As blocks are freed, 1730713e232SGeorge Wilson * they are added to the per txg free tree (ms_freetree). These per txg 1740713e232SGeorge Wilson * trees allow us to process all allocations and frees in syncing context 1750713e232SGeorge Wilson * where it is safe to update the on-disk space maps. One additional in-core 1760713e232SGeorge Wilson * tree is maintained to track deferred frees (ms_defertree). Once a block 1770713e232SGeorge Wilson * is freed it will move from the ms_freetree to the ms_defertree. A deferred 1780713e232SGeorge Wilson * free means that a block has been freed but cannot be used by the pool 1790713e232SGeorge Wilson * until TXG_DEFER_SIZE transactions groups later. For example, a block 1800713e232SGeorge Wilson * that is freed in txg 50 will not be available for reallocation until 1810713e232SGeorge Wilson * txg 52 (50 + TXG_DEFER_SIZE). This provides a safety net for uberblock 1820713e232SGeorge Wilson * rollback. A pool could be safely rolled back TXG_DEFERS_SIZE 1830713e232SGeorge Wilson * transactions groups and ensure that no block has been reallocated. 1840713e232SGeorge Wilson * 1850713e232SGeorge Wilson * The simplified transition diagram looks like this: 1860713e232SGeorge Wilson * 1870713e232SGeorge Wilson * 1880713e232SGeorge Wilson * ALLOCATE 1890713e232SGeorge Wilson * | 1900713e232SGeorge Wilson * V 1910713e232SGeorge Wilson * free segment (ms_tree) --------> ms_alloctree ----> (write to space map) 1920713e232SGeorge Wilson * ^ 1930713e232SGeorge Wilson * | 1940713e232SGeorge Wilson * | ms_freetree <--- FREE 1950713e232SGeorge Wilson * | | 1960713e232SGeorge Wilson * | | 1970713e232SGeorge Wilson * | | 1980713e232SGeorge Wilson * +----------- ms_defertree <-------+---------> (write to space map) 19916a4a807SGeorge Wilson * 2000713e232SGeorge Wilson * 2010713e232SGeorge Wilson * Each metaslab's space is tracked in a single space map in the MOS, 20216a4a807SGeorge Wilson * which is only updated in syncing context. Each time we sync a txg, 2030713e232SGeorge Wilson * we append the allocs and frees from that txg to the space map. 2040713e232SGeorge Wilson * The pool space is only updated once all metaslabs have finished syncing. 20516a4a807SGeorge Wilson * 2060713e232SGeorge Wilson * To load the in-core free tree we read the space map from disk. 20716a4a807SGeorge Wilson * This object contains a series of alloc and free records that are 20816a4a807SGeorge Wilson * combined to make up the list of all free segments in this metaslab. These 2090713e232SGeorge Wilson * segments are represented in-core by the ms_tree and are stored in an 21016a4a807SGeorge Wilson * AVL tree. 21116a4a807SGeorge Wilson * 2120713e232SGeorge Wilson * As the space map grows (as a result of the appends) it will 2130713e232SGeorge Wilson * eventually become space-inefficient. When the metaslab's in-core free tree 2140713e232SGeorge Wilson * is zfs_condense_pct/100 times the size of the minimal on-disk 2150713e232SGeorge Wilson * representation, we rewrite it in its minimized form. If a metaslab 2160713e232SGeorge Wilson * needs to condense then we must set the ms_condensing flag to ensure 2170713e232SGeorge Wilson * that allocations are not performed on the metaslab that is being written. 218fa9e4066Sahrens */ 219fa9e4066Sahrens struct metaslab { 2200713e232SGeorge Wilson kmutex_t ms_lock; 2210713e232SGeorge Wilson kcondvar_t ms_load_cv; 2220713e232SGeorge Wilson space_map_t *ms_sm; 2230713e232SGeorge Wilson metaslab_ops_t *ms_ops; 2240713e232SGeorge Wilson uint64_t ms_id; 2250713e232SGeorge Wilson uint64_t ms_start; 2260713e232SGeorge Wilson uint64_t ms_size; 2272e4c9986SGeorge Wilson uint64_t ms_fragmentation; 2280713e232SGeorge Wilson 2290713e232SGeorge Wilson range_tree_t *ms_alloctree[TXG_SIZE]; 2300713e232SGeorge Wilson range_tree_t *ms_freetree[TXG_SIZE]; 2310713e232SGeorge Wilson range_tree_t *ms_defertree[TXG_DEFER_SIZE]; 2320713e232SGeorge Wilson range_tree_t *ms_tree; 2330713e232SGeorge Wilson 2340713e232SGeorge Wilson boolean_t ms_condensing; /* condensing? */ 2352e4c9986SGeorge Wilson boolean_t ms_condense_wanted; 2360713e232SGeorge Wilson boolean_t ms_loaded; 2370713e232SGeorge Wilson boolean_t ms_loading; 2380713e232SGeorge Wilson 239468c413aSTim Haley int64_t ms_deferspace; /* sum of ms_defermap[] space */ 240ecc2d604Sbonwick uint64_t ms_weight; /* weight vs. others in group */ 2410713e232SGeorge Wilson uint64_t ms_access_txg; 2420713e232SGeorge Wilson 2430713e232SGeorge Wilson /* 2440713e232SGeorge Wilson * The metaslab block allocators can optionally use a size-ordered 2450713e232SGeorge Wilson * range tree and/or an array of LBAs. Not all allocators use 2460713e232SGeorge Wilson * this functionality. The ms_size_tree should always contain the 2470713e232SGeorge Wilson * same number of segments as the ms_tree. The only difference 2480713e232SGeorge Wilson * is that the ms_size_tree is ordered by segment sizes. 2490713e232SGeorge Wilson */ 2500713e232SGeorge Wilson avl_tree_t ms_size_tree; 2510713e232SGeorge Wilson uint64_t ms_lbas[MAX_LBAS]; 2520713e232SGeorge Wilson 253ecc2d604Sbonwick metaslab_group_t *ms_group; /* metaslab group */ 254ecc2d604Sbonwick avl_node_t ms_group_node; /* node in metaslab group tree */ 255ecc2d604Sbonwick txg_node_t ms_txg_node; /* per-txg dirty metaslab links */ 256fa9e4066Sahrens }; 257fa9e4066Sahrens 258fa9e4066Sahrens #ifdef __cplusplus 259fa9e4066Sahrens } 260fa9e4066Sahrens #endif 261fa9e4066Sahrens 262fa9e4066Sahrens #endif /* _SYS_METASLAB_IMPL_H */ 263