1*fa9e4066Sahrens /* 2*fa9e4066Sahrens * CDDL HEADER START 3*fa9e4066Sahrens * 4*fa9e4066Sahrens * The contents of this file are subject to the terms of the 5*fa9e4066Sahrens * Common Development and Distribution License, Version 1.0 only 6*fa9e4066Sahrens * (the "License"). You may not use this file except in compliance 7*fa9e4066Sahrens * with the License. 8*fa9e4066Sahrens * 9*fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10*fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 11*fa9e4066Sahrens * See the License for the specific language governing permissions 12*fa9e4066Sahrens * and limitations under the License. 13*fa9e4066Sahrens * 14*fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 15*fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16*fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 17*fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 18*fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 19*fa9e4066Sahrens * 20*fa9e4066Sahrens * CDDL HEADER END 21*fa9e4066Sahrens */ 22*fa9e4066Sahrens /* 23*fa9e4066Sahrens * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24*fa9e4066Sahrens * Use is subject to license terms. 25*fa9e4066Sahrens */ 26*fa9e4066Sahrens 27*fa9e4066Sahrens #ifndef _SYS_METASLAB_IMPL_H 28*fa9e4066Sahrens #define _SYS_METASLAB_IMPL_H 29*fa9e4066Sahrens 30*fa9e4066Sahrens #pragma ident "%Z%%M% %I% %E% SMI" 31*fa9e4066Sahrens 32*fa9e4066Sahrens #include <sys/metaslab.h> 33*fa9e4066Sahrens #include <sys/space_map.h> 34*fa9e4066Sahrens #include <sys/vdev.h> 35*fa9e4066Sahrens #include <sys/txg.h> 36*fa9e4066Sahrens #include <sys/avl.h> 37*fa9e4066Sahrens 38*fa9e4066Sahrens #ifdef __cplusplus 39*fa9e4066Sahrens extern "C" { 40*fa9e4066Sahrens #endif 41*fa9e4066Sahrens 42*fa9e4066Sahrens struct metaslab_class { 43*fa9e4066Sahrens metaslab_group_t *mc_rotor; 44*fa9e4066Sahrens uint64_t mc_allocated; 45*fa9e4066Sahrens }; 46*fa9e4066Sahrens 47*fa9e4066Sahrens struct metaslab_group { 48*fa9e4066Sahrens kmutex_t mg_lock; 49*fa9e4066Sahrens avl_tree_t mg_metaslab_tree; 50*fa9e4066Sahrens uint64_t mg_aliquot; 51*fa9e4066Sahrens int64_t mg_bias; 52*fa9e4066Sahrens metaslab_class_t *mg_class; 53*fa9e4066Sahrens vdev_t *mg_vd; 54*fa9e4066Sahrens metaslab_group_t *mg_prev; 55*fa9e4066Sahrens metaslab_group_t *mg_next; 56*fa9e4066Sahrens }; 57*fa9e4066Sahrens 58*fa9e4066Sahrens /* 59*fa9e4066Sahrens * Each metaslab's free block list is kept in its own DMU object in the 60*fa9e4066Sahrens * metaslab freelist dataset. To minimize space consumption, the list 61*fa9e4066Sahrens * is circular. 62*fa9e4066Sahrens * 63*fa9e4066Sahrens * Allocations and frees can happen in multiple transaction groups at 64*fa9e4066Sahrens * the same time, which makes it a bit challening to keep the metaslab 65*fa9e4066Sahrens * consistent. For example, we cannot allow frees from different 66*fa9e4066Sahrens * transaction groups to be interleaved in the metaslab's free block list. 67*fa9e4066Sahrens * 68*fa9e4066Sahrens * We address this in several ways: 69*fa9e4066Sahrens * 70*fa9e4066Sahrens * We don't allow allocations from the same metaslab in concurrent 71*fa9e4066Sahrens * transaction groups. metaslab_alloc() enforces this by checking 72*fa9e4066Sahrens * the ms_last_alloc field, which specifies the last txg in which 73*fa9e4066Sahrens * the metaslab was used for allocations. 74*fa9e4066Sahrens * 75*fa9e4066Sahrens * We can't segregate frees this way because we can't choose which 76*fa9e4066Sahrens * DVAs someone wants to free. So we keep separate in-core freelists 77*fa9e4066Sahrens * for each active transaction group. This in-core data is only 78*fa9e4066Sahrens * written to the metaslab's on-disk freelist in metaslab_sync(), 79*fa9e4066Sahrens * which solves the interleave problem: we only append frees from 80*fa9e4066Sahrens * the syncing txg to the on-disk freelist, so the appends all occur 81*fa9e4066Sahrens * in txg order. 82*fa9e4066Sahrens * 83*fa9e4066Sahrens * We cannot allow a block which was freed in a given txg to be 84*fa9e4066Sahrens * allocated again until that txg has closed; otherwise, if we 85*fa9e4066Sahrens * failed to sync that txg and had to roll back to txg - 1, 86*fa9e4066Sahrens * changes in txg + 1 could have overwritten the data. Therefore, 87*fa9e4066Sahrens * we partition the free blocks into "available" and "limbo" states. 88*fa9e4066Sahrens * A block is available if the txg in which it was freed has closed; 89*fa9e4066Sahrens * until then, the block is in limbo. Each time metaslab_sync() runs, 90*fa9e4066Sahrens * if first adds any limbo blocks to the avail list, clears the limbo 91*fa9e4066Sahrens * list, and starts writing the new limbo blocks (i.e. the ones that 92*fa9e4066Sahrens * were freed in the syncing txg). 93*fa9e4066Sahrens */ 94*fa9e4066Sahrens 95*fa9e4066Sahrens struct metaslab { 96*fa9e4066Sahrens kmutex_t ms_lock; /* metaslab lock */ 97*fa9e4066Sahrens space_map_obj_t *ms_smo; /* space map object */ 98*fa9e4066Sahrens uint64_t ms_last_alloc; /* txg of last alloc */ 99*fa9e4066Sahrens uint64_t ms_usable_end; /* end of free_obj at last sync */ 100*fa9e4066Sahrens uint64_t ms_usable_space; /* usable space at last sync */ 101*fa9e4066Sahrens metaslab_group_t *ms_group; /* metaslab group */ 102*fa9e4066Sahrens avl_node_t ms_group_node; /* node in metaslab group tree */ 103*fa9e4066Sahrens uint64_t ms_weight; /* weight vs. others in group */ 104*fa9e4066Sahrens uint8_t ms_dirty[TXG_SIZE]; /* per-txg dirty flags */ 105*fa9e4066Sahrens space_map_t ms_allocmap[TXG_SIZE]; /* allocated this txg */ 106*fa9e4066Sahrens space_map_t ms_freemap[TXG_SIZE]; /* freed this txg */ 107*fa9e4066Sahrens txg_node_t ms_txg_node; /* per-txg dirty metaslab links */ 108*fa9e4066Sahrens space_map_t ms_map; /* in-core free space map */ 109*fa9e4066Sahrens uint8_t ms_map_incore; /* space map contents are valid */ 110*fa9e4066Sahrens uint64_t ms_map_cursor[SPA_ASIZEBITS]; /* XXX -- PPD */ 111*fa9e4066Sahrens }; 112*fa9e4066Sahrens 113*fa9e4066Sahrens /* 114*fa9e4066Sahrens * ms_dirty[] flags 115*fa9e4066Sahrens */ 116*fa9e4066Sahrens #define MSD_ALLOC 0x01 /* allocated from in this txg */ 117*fa9e4066Sahrens #define MSD_FREE 0x02 /* freed to in this txg */ 118*fa9e4066Sahrens #define MSD_ADD 0x04 /* added to the pool in this txg */ 119*fa9e4066Sahrens #define MSD_CONDENSE 0x08 /* condensed in this txg */ 120*fa9e4066Sahrens 121*fa9e4066Sahrens #ifdef __cplusplus 122*fa9e4066Sahrens } 123*fa9e4066Sahrens #endif 124*fa9e4066Sahrens 125*fa9e4066Sahrens #endif /* _SYS_METASLAB_IMPL_H */ 126