fs/ext4/mballoc.h

   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/<parition>/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  * maximum length of inode prealloc list
  78  */
  79 #define MB_DEFAULT_MAX_INODE_PREALLOC   512
  80
  81 struct ext4_free_data {
  82         /* this links the free block information from sb_info */
  83         struct list_head                efd_list;
  84
  85         /* this links the free block information from group_info */
  86         struct rb_node                  efd_node;
  87
  88         /* group which free block extent belongs */
  89         ext4_group_t                    efd_group;
  90
  91         /* free block extent */
  92         ext4_grpblk_t                   efd_start_cluster;
  93         ext4_grpblk_t                   efd_count;
  94
  95         /* transaction which freed this extent */
  96         tid_t                           efd_tid;
  97 };
  98
  99 struct ext4_prealloc_space {
 100         struct list_head        pa_inode_list;
 101         struct list_head        pa_group_list;
 102         union {
 103                 struct list_head pa_tmp_list;
 104                 struct rcu_head pa_rcu;
 105         } u;
 106         spinlock_t              pa_lock;
 107         atomic_t                pa_count;
 108         unsigned                pa_deleted;
 109         ext4_fsblk_t            pa_pstart;      /* phys. block */
 110         ext4_lblk_t             pa_lstart;      /* log. block */
 111         ext4_grpblk_t           pa_len;         /* len of preallocated chunk */
 112         ext4_grpblk_t           pa_free;        /* how many blocks are free */
 113         unsigned short          pa_type;        /* pa type. inode or group */
 114         spinlock_t              *pa_obj_lock;
 115         struct inode            *pa_inode;      /* hack, for history only */
 116 };
 117
 118 enum {
 119         MB_INODE_PA = 0,
 120         MB_GROUP_PA = 1
 121 };
 122
 123 struct ext4_free_extent {
 124         ext4_lblk_t fe_logical;
 125         ext4_grpblk_t fe_start; /* In cluster units */
 126         ext4_group_t fe_group;
 127         ext4_grpblk_t fe_len;   /* In cluster units */
 128 };
 129
 130 /*
 131  * Locality group:
 132  *   we try to group all related changes together
 133  *   so that writeback can flush/allocate them together as well
 134  *   Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC
 135  *   (512). We store prealloc space into the hash based on the pa_free blocks
 136  *   order value.ie, fls(pa_free)-1;
 137  */
 138 #define PREALLOC_TB_SIZE 10
 139 struct ext4_locality_group {
 140         /* for allocator */
 141         /* to serialize allocates */
 142         struct mutex            lg_mutex;
 143         /* list of preallocations */
 144         struct list_head        lg_prealloc_list[PREALLOC_TB_SIZE];
 145         spinlock_t              lg_prealloc_lock;
 146 };
 147
 148 struct ext4_allocation_context {
 149         struct inode *ac_inode;
 150         struct super_block *ac_sb;
 151
 152         /* original request */
 153         struct ext4_free_extent ac_o_ex;
 154
 155         /* goal request (normalized ac_o_ex) */
 156         struct ext4_free_extent ac_g_ex;
 157
 158         /* the best found extent */
 159         struct ext4_free_extent ac_b_ex;
 160
 161         /* copy of the best found extent taken before preallocation efforts */
 162         struct ext4_free_extent ac_f_ex;
 163
 164         __u16 ac_groups_scanned;
 165         __u16 ac_found;
 166         __u16 ac_tail;
 167         __u16 ac_buddy;
 168         __u16 ac_flags;         /* allocation hints */
 169         __u8 ac_status;
 170         __u8 ac_criteria;
 171         __u8 ac_2order;         /* if request is to allocate 2^N blocks and
 172                                  * N > 0, the field stores N, otherwise 0 */
 173         __u8 ac_op;             /* operation, for history only */
 174         struct page *ac_bitmap_page;
 175         struct page *ac_buddy_page;
 176         struct ext4_prealloc_space *ac_pa;
 177         struct ext4_locality_group *ac_lg;
 178 };
 179
 180 #define AC_STATUS_CONTINUE      1
 181 #define AC_STATUS_FOUND         2
 182 #define AC_STATUS_BREAK         3
 183
 184 struct ext4_buddy {
 185         struct page *bd_buddy_page;
 186         void *bd_buddy;
 187         struct page *bd_bitmap_page;
 188         void *bd_bitmap;
 189         struct ext4_group_info *bd_info;
 190         struct super_block *bd_sb;
 191         __u16 bd_blkbits;
 192         ext4_group_t bd_group;
 193 };
 194
 195 static inline ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
 196                                         struct ext4_free_extent *fex)
 197 {
 198         return ext4_group_first_block_no(sb, fex->fe_group) +
 199                 (fex->fe_start << EXT4_SB(sb)->s_cluster_bits);
 200 }
 201
 202 typedef int (*ext4_mballoc_query_range_fn)(
 203         struct super_block              *sb,
 204         ext4_group_t                    agno,
 205         ext4_grpblk_t                   start,
 206         ext4_grpblk_t                   len,
 207         void                            *priv);
 208
 209 int
 210 ext4_mballoc_query_range(
 211         struct super_block              *sb,
 212         ext4_group_t                    agno,
 213         ext4_grpblk_t                   start,
 214         ext4_grpblk_t                   end,
 215         ext4_mballoc_query_range_fn     formatter,
 216         void                            *priv);
 217
 218 #endif