1 // SPDX-License-Identifier: GPL-2.0
5 * Written by: Alex Tomas <alex@clusterfs.com>
8 #ifndef _EXT4_MBALLOC_H
9 #define _EXT4_MBALLOC_H
11 #include <linux/time.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"
27 * mb_debug() dynamic printk msgs could be used to debug mballoc code.
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__)
35 #define mb_debug(sb, fmt, ...) no_printk(fmt, ##__VA_ARGS__)
38 #define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
39 #define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
42 * How long mballoc can look for a best extent (in found extents)
44 #define MB_DEFAULT_MAX_TO_SCAN 200
47 * How long mballoc must look for a best extent
49 #define MB_DEFAULT_MIN_TO_SCAN 10
52 * with 'ext4_mb_stats' allocator will collect stats that will be
53 * shown at umount. The collecting costs though!
55 #define MB_DEFAULT_STATS 0
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
64 #define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
67 * for which requests use 2^N search using buddies
69 #define MB_DEFAULT_ORDER2_REQS 2
72 * default group prealloc size 512 blocks
74 #define MB_DEFAULT_GROUP_PREALLOC 512
77 * maximum length of inode prealloc list
79 #define MB_DEFAULT_MAX_INODE_PREALLOC 512
82 * Number of groups to search linearly before performing group scanning
85 #define MB_DEFAULT_LINEAR_LIMIT 4
88 * Minimum number of groups that should be present in the file system to perform
89 * group scanning optimizations.
91 #define MB_DEFAULT_LINEAR_SCAN_THRESHOLD 16
94 * Number of valid buddy orders
96 #define MB_NUM_ORDERS(sb) ((sb)->s_blocksize_bits + 2)
98 struct ext4_free_data
{
99 /* this links the free block information from sb_info */
100 struct list_head efd_list
;
102 /* this links the free block information from group_info */
103 struct rb_node efd_node
;
105 /* group which free block extent belongs */
106 ext4_group_t efd_group
;
108 /* free block extent */
109 ext4_grpblk_t efd_start_cluster
;
110 ext4_grpblk_t efd_count
;
112 /* transaction which freed this extent */
116 struct ext4_prealloc_space
{
117 struct list_head pa_inode_list
;
118 struct list_head pa_group_list
;
120 struct list_head pa_tmp_list
;
121 struct rcu_head pa_rcu
;
126 ext4_fsblk_t pa_pstart
; /* phys. block */
127 ext4_lblk_t pa_lstart
; /* log. block */
128 ext4_grpblk_t pa_len
; /* len of preallocated chunk */
129 ext4_grpblk_t pa_free
; /* how many blocks are free */
130 unsigned short pa_type
; /* pa type. inode or group */
131 spinlock_t
*pa_obj_lock
;
132 struct inode
*pa_inode
; /* hack, for history only */
140 struct ext4_free_extent
{
141 ext4_lblk_t fe_logical
;
142 ext4_grpblk_t fe_start
; /* In cluster units */
143 ext4_group_t fe_group
;
144 ext4_grpblk_t fe_len
; /* In cluster units */
149 * we try to group all related changes together
150 * so that writeback can flush/allocate them together as well
151 * Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC
152 * (512). We store prealloc space into the hash based on the pa_free blocks
153 * order value.ie, fls(pa_free)-1;
155 #define PREALLOC_TB_SIZE 10
156 struct ext4_locality_group
{
158 /* to serialize allocates */
159 struct mutex lg_mutex
;
160 /* list of preallocations */
161 struct list_head lg_prealloc_list
[PREALLOC_TB_SIZE
];
162 spinlock_t lg_prealloc_lock
;
165 struct ext4_allocation_context
{
166 struct inode
*ac_inode
;
167 struct super_block
*ac_sb
;
169 /* original request */
170 struct ext4_free_extent ac_o_ex
;
172 /* goal request (normalized ac_o_ex) */
173 struct ext4_free_extent ac_g_ex
;
175 /* the best found extent */
176 struct ext4_free_extent ac_b_ex
;
178 /* copy of the best found extent taken before preallocation efforts */
179 struct ext4_free_extent ac_f_ex
;
181 ext4_group_t ac_last_optimal_group
;
182 __u32 ac_groups_considered
;
183 __u32 ac_flags
; /* allocation hints */
184 __u16 ac_groups_scanned
;
185 __u16 ac_groups_linear_remaining
;
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
;
200 #define AC_STATUS_CONTINUE 1
201 #define AC_STATUS_FOUND 2
202 #define AC_STATUS_BREAK 3
205 struct page
*bd_buddy_page
;
207 struct page
*bd_bitmap_page
;
209 struct ext4_group_info
*bd_info
;
210 struct super_block
*bd_sb
;
212 ext4_group_t bd_group
;
215 static inline ext4_fsblk_t
ext4_grp_offs_to_block(struct super_block
*sb
,
216 struct ext4_free_extent
*fex
)
218 return ext4_group_first_block_no(sb
, fex
->fe_group
) +
219 (fex
->fe_start
<< EXT4_SB(sb
)->s_cluster_bits
);
222 typedef int (*ext4_mballoc_query_range_fn
)(
223 struct super_block
*sb
,
230 ext4_mballoc_query_range(
231 struct super_block
*sb
,
235 ext4_mballoc_query_range_fn formatter
,