1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * fs/ext4/mballoc.h 4 * 5 * Written by: Alex Tomas <alex@clusterfs.com> 6 * 7 */ 8 #ifndef _EXT4_MBALLOC_H 9 #define _EXT4_MBALLOC_H 10 11 #include <linux/time.h> 12 #include <linux/fs.h> 13 #include <linux/namei.h> 14 #include <linux/quotaops.h> 15 #include <linux/buffer_head.h> 16 #include <linux/module.h> 17 #include <linux/swap.h> 18 #include <linux/proc_fs.h> 19 #include <linux/pagemap.h> 20 #include <linux/seq_file.h> 21 #include <linux/blkdev.h> 22 #include <linux/mutex.h> 23 #include "ext4_jbd2.h" 24 #include "ext4.h" 25 26 /* 27 * mb_debug() dynamic printk msgs could be used to debug mballoc code. 28 */ 29 #ifdef CONFIG_EXT4_DEBUG 30 #define mb_debug(sb, fmt, ...) \ 31 pr_debug("[%s/%d] EXT4-fs (%s): (%s, %d): %s: " fmt, \ 32 current->comm, task_pid_nr(current), sb->s_id, \ 33 __FILE__, __LINE__, __func__, ##__VA_ARGS__) 34 #else 35 #define mb_debug(sb, fmt, ...) no_printk(fmt, ##__VA_ARGS__) 36 #endif 37 38 #define EXT4_MB_HISTORY_ALLOC 1 /* allocation */ 39 #define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */ 40 41 /* 42 * How long mballoc can look for a best extent (in found extents) 43 */ 44 #define MB_DEFAULT_MAX_TO_SCAN 200 45 46 /* 47 * How long mballoc must look for a best extent 48 */ 49 #define MB_DEFAULT_MIN_TO_SCAN 10 50 51 /* 52 * with 'ext4_mb_stats' allocator will collect stats that will be 53 * shown at umount. The collecting costs though! 54 */ 55 #define MB_DEFAULT_STATS 0 56 57 /* 58 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served 59 * by the stream allocator, which purpose is to pack requests 60 * as close each to other as possible to produce smooth I/O traffic 61 * We use locality group prealloc space for stream request. 62 * We can tune the same via /proc/fs/ext4/<partition>/stream_req 63 */ 64 #define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */ 65 66 /* 67 * for which requests use 2^N search using buddies 68 */ 69 #define MB_DEFAULT_ORDER2_REQS 2 70 71 /* 72 * default group prealloc size 512 blocks 73 */ 74 #define MB_DEFAULT_GROUP_PREALLOC 512 75 76 /* 77 * Number of groups to search linearly before performing group scanning 78 * optimization. 79 */ 80 #define MB_DEFAULT_LINEAR_LIMIT 4 81 82 /* 83 * Minimum number of groups that should be present in the file system to perform 84 * group scanning optimizations. 85 */ 86 #define MB_DEFAULT_LINEAR_SCAN_THRESHOLD 16 87 88 /* 89 * Number of valid buddy orders 90 */ 91 #define MB_NUM_ORDERS(sb) ((sb)->s_blocksize_bits + 2) 92 93 struct ext4_free_data { 94 /* this links the free block information from sb_info */ 95 struct list_head efd_list; 96 97 /* this links the free block information from group_info */ 98 struct rb_node efd_node; 99 100 /* group which free block extent belongs */ 101 ext4_group_t efd_group; 102 103 /* free block extent */ 104 ext4_grpblk_t efd_start_cluster; 105 ext4_grpblk_t efd_count; 106 107 /* transaction which freed this extent */ 108 tid_t efd_tid; 109 }; 110 111 struct ext4_prealloc_space { 112 union { 113 struct rb_node inode_node; /* for inode PA rbtree */ 114 struct list_head lg_list; /* for lg PAs */ 115 } pa_node; 116 struct list_head pa_group_list; 117 union { 118 struct list_head pa_tmp_list; 119 struct rcu_head pa_rcu; 120 } u; 121 spinlock_t pa_lock; 122 atomic_t pa_count; 123 unsigned pa_deleted; 124 ext4_fsblk_t pa_pstart; /* phys. block */ 125 ext4_lblk_t pa_lstart; /* log. block */ 126 ext4_grpblk_t pa_len; /* len of preallocated chunk */ 127 ext4_grpblk_t pa_free; /* how many blocks are free */ 128 unsigned short pa_type; /* pa type. inode or group */ 129 union { 130 rwlock_t *inode_lock; /* locks the rbtree holding this PA */ 131 spinlock_t *lg_lock; /* locks the lg list holding this PA */ 132 } pa_node_lock; 133 struct inode *pa_inode; /* used to get the inode during group discard */ 134 }; 135 136 enum { 137 MB_INODE_PA = 0, 138 MB_GROUP_PA = 1 139 }; 140 141 struct ext4_free_extent { 142 ext4_lblk_t fe_logical; 143 ext4_grpblk_t fe_start; /* In cluster units */ 144 ext4_group_t fe_group; 145 ext4_grpblk_t fe_len; /* In cluster units */ 146 }; 147 148 /* 149 * Locality group: 150 * we try to group all related changes together 151 * so that writeback can flush/allocate them together as well 152 * Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC 153 * (512). We store prealloc space into the hash based on the pa_free blocks 154 * order value.ie, fls(pa_free)-1; 155 */ 156 #define PREALLOC_TB_SIZE 10 157 struct ext4_locality_group { 158 /* for allocator */ 159 /* to serialize allocates */ 160 struct mutex lg_mutex; 161 /* list of preallocations */ 162 struct list_head lg_prealloc_list[PREALLOC_TB_SIZE]; 163 spinlock_t lg_prealloc_lock; 164 }; 165 166 struct ext4_allocation_context { 167 struct inode *ac_inode; 168 struct super_block *ac_sb; 169 170 /* original request */ 171 struct ext4_free_extent ac_o_ex; 172 173 /* goal request (normalized ac_o_ex) */ 174 struct ext4_free_extent ac_g_ex; 175 176 /* the best found extent */ 177 struct ext4_free_extent ac_b_ex; 178 179 /* copy of the best found extent taken before preallocation efforts */ 180 struct ext4_free_extent ac_f_ex; 181 182 __u32 ac_groups_considered; 183 __u32 ac_flags; /* allocation hints */ 184 __u16 ac_groups_scanned; 185 __u16 ac_groups_linear_remaining; 186 __u16 ac_found; 187 __u16 ac_tail; 188 __u16 ac_buddy; 189 __u8 ac_status; 190 __u8 ac_criteria; 191 __u8 ac_2order; /* if request is to allocate 2^N blocks and 192 * N > 0, the field stores N, otherwise 0 */ 193 __u8 ac_op; /* operation, for history only */ 194 struct page *ac_bitmap_page; 195 struct page *ac_buddy_page; 196 struct ext4_prealloc_space *ac_pa; 197 struct ext4_locality_group *ac_lg; 198 }; 199 200 #define AC_STATUS_CONTINUE 1 201 #define AC_STATUS_FOUND 2 202 #define AC_STATUS_BREAK 3 203 204 struct ext4_buddy { 205 struct page *bd_buddy_page; 206 void *bd_buddy; 207 struct page *bd_bitmap_page; 208 void *bd_bitmap; 209 struct ext4_group_info *bd_info; 210 struct super_block *bd_sb; 211 __u16 bd_blkbits; 212 ext4_group_t bd_group; 213 }; 214 215 static inline ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb, 216 struct ext4_free_extent *fex) 217 { 218 return ext4_group_first_block_no(sb, fex->fe_group) + 219 (fex->fe_start << EXT4_SB(sb)->s_cluster_bits); 220 } 221 222 typedef int (*ext4_mballoc_query_range_fn)( 223 struct super_block *sb, 224 ext4_group_t agno, 225 ext4_grpblk_t start, 226 ext4_grpblk_t len, 227 void *priv); 228 229 int 230 ext4_mballoc_query_range( 231 struct super_block *sb, 232 ext4_group_t agno, 233 ext4_grpblk_t start, 234 ext4_grpblk_t end, 235 ext4_mballoc_query_range_fn formatter, 236 void *priv); 237 238 #endif 239