Commit | Line | Data |
---|---|---|
8f6e39a7 MC |
1 | /* |
2 | * fs/ext4/mballoc.h | |
3 | * | |
4 | * Written by: Alex Tomas <alex@clusterfs.com> | |
5 | * | |
6 | */ | |
7 | #ifndef _EXT4_MBALLOC_H | |
8 | #define _EXT4_MBALLOC_H | |
9 | ||
10 | #include <linux/time.h> | |
11 | #include <linux/fs.h> | |
12 | #include <linux/namei.h> | |
13 | #include <linux/quotaops.h> | |
14 | #include <linux/buffer_head.h> | |
15 | #include <linux/module.h> | |
16 | #include <linux/swap.h> | |
17 | #include <linux/proc_fs.h> | |
18 | #include <linux/pagemap.h> | |
19 | #include <linux/seq_file.h> | |
8a0aba73 | 20 | #include <linux/blkdev.h> |
920313a7 | 21 | #include <linux/mutex.h> |
8f6e39a7 MC |
22 | #include "ext4_jbd2.h" |
23 | #include "ext4.h" | |
8f6e39a7 | 24 | |
8f6e39a7 MC |
25 | /* |
26 | */ | |
6ba495e9 | 27 | #ifdef CONFIG_EXT4_DEBUG |
a0b30c12 | 28 | extern ushort ext4_mballoc_debug; |
6ba495e9 | 29 | |
d74f3d25 JP |
30 | #define mb_debug(n, fmt, ...) \ |
31 | do { \ | |
32 | if ((n) <= ext4_mballoc_debug) { \ | |
33 | printk(KERN_DEBUG "(%s, %d): %s: " fmt, \ | |
34 | __FILE__, __LINE__, __func__, ##__VA_ARGS__); \ | |
35 | } \ | |
36 | } while (0) | |
8f6e39a7 | 37 | #else |
d74f3d25 | 38 | #define mb_debug(n, fmt, ...) no_printk(fmt, ##__VA_ARGS__) |
8f6e39a7 MC |
39 | #endif |
40 | ||
8f6e39a7 MC |
41 | #define EXT4_MB_HISTORY_ALLOC 1 /* allocation */ |
42 | #define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */ | |
8f6e39a7 MC |
43 | |
44 | /* | |
45 | * How long mballoc can look for a best extent (in found extents) | |
46 | */ | |
47 | #define MB_DEFAULT_MAX_TO_SCAN 200 | |
48 | ||
49 | /* | |
50 | * How long mballoc must look for a best extent | |
51 | */ | |
52 | #define MB_DEFAULT_MIN_TO_SCAN 10 | |
53 | ||
8f6e39a7 MC |
54 | /* |
55 | * with 'ext4_mb_stats' allocator will collect stats that will be | |
56 | * shown at umount. The collecting costs though! | |
57 | */ | |
90576c0b | 58 | #define MB_DEFAULT_STATS 0 |
8f6e39a7 MC |
59 | |
60 | /* | |
61 | * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served | |
62 | * by the stream allocator, which purpose is to pack requests | |
63 | * as close each to other as possible to produce smooth I/O traffic | |
64 | * We use locality group prealloc space for stream request. | |
65 | * We can tune the same via /proc/fs/ext4/<parition>/stream_req | |
66 | */ | |
67 | #define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */ | |
68 | ||
69 | /* | |
70 | * for which requests use 2^N search using buddies | |
71 | */ | |
72 | #define MB_DEFAULT_ORDER2_REQS 2 | |
73 | ||
74 | /* | |
75 | * default group prealloc size 512 blocks | |
76 | */ | |
77 | #define MB_DEFAULT_GROUP_PREALLOC 512 | |
78 | ||
8f6e39a7 | 79 | |
c894058d | 80 | struct ext4_free_data { |
a0154344 DJ |
81 | /* this links the free block information from sb_info */ |
82 | struct list_head efd_list; | |
8f6e39a7 | 83 | |
18aadd47 BJ |
84 | /* this links the free block information from group_info */ |
85 | struct rb_node efd_node; | |
c894058d AK |
86 | |
87 | /* group which free block extent belongs */ | |
18aadd47 | 88 | ext4_group_t efd_group; |
c894058d AK |
89 | |
90 | /* free block extent */ | |
18aadd47 BJ |
91 | ext4_grpblk_t efd_start_cluster; |
92 | ext4_grpblk_t efd_count; | |
c894058d AK |
93 | |
94 | /* transaction which freed this extent */ | |
18aadd47 | 95 | tid_t efd_tid; |
8f6e39a7 MC |
96 | }; |
97 | ||
8f6e39a7 MC |
98 | struct ext4_prealloc_space { |
99 | struct list_head pa_inode_list; | |
100 | struct list_head pa_group_list; | |
101 | union { | |
102 | struct list_head pa_tmp_list; | |
103 | struct rcu_head pa_rcu; | |
104 | } u; | |
105 | spinlock_t pa_lock; | |
106 | atomic_t pa_count; | |
107 | unsigned pa_deleted; | |
108 | ext4_fsblk_t pa_pstart; /* phys. block */ | |
109 | ext4_lblk_t pa_lstart; /* log. block */ | |
a36b4498 ES |
110 | ext4_grpblk_t pa_len; /* len of preallocated chunk */ |
111 | ext4_grpblk_t pa_free; /* how many blocks are free */ | |
cc0fb9ad | 112 | unsigned short pa_type; /* pa type. inode or group */ |
8f6e39a7 MC |
113 | spinlock_t *pa_obj_lock; |
114 | struct inode *pa_inode; /* hack, for history only */ | |
115 | }; | |
116 | ||
cc0fb9ad AK |
117 | enum { |
118 | MB_INODE_PA = 0, | |
119 | MB_GROUP_PA = 1 | |
120 | }; | |
8f6e39a7 MC |
121 | |
122 | struct ext4_free_extent { | |
123 | ext4_lblk_t fe_logical; | |
53accfa9 | 124 | ext4_grpblk_t fe_start; /* In cluster units */ |
8f6e39a7 | 125 | ext4_group_t fe_group; |
53accfa9 | 126 | ext4_grpblk_t fe_len; /* In cluster units */ |
8f6e39a7 MC |
127 | }; |
128 | ||
129 | /* | |
130 | * Locality group: | |
131 | * we try to group all related changes together | |
132 | * so that writeback can flush/allocate them together as well | |
6be2ded1 AK |
133 | * Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC |
134 | * (512). We store prealloc space into the hash based on the pa_free blocks | |
135 | * order value.ie, fls(pa_free)-1; | |
8f6e39a7 | 136 | */ |
6be2ded1 | 137 | #define PREALLOC_TB_SIZE 10 |
8f6e39a7 MC |
138 | struct ext4_locality_group { |
139 | /* for allocator */ | |
6be2ded1 AK |
140 | /* to serialize allocates */ |
141 | struct mutex lg_mutex; | |
142 | /* list of preallocations */ | |
143 | struct list_head lg_prealloc_list[PREALLOC_TB_SIZE]; | |
8f6e39a7 MC |
144 | spinlock_t lg_prealloc_lock; |
145 | }; | |
146 | ||
147 | struct ext4_allocation_context { | |
148 | struct inode *ac_inode; | |
149 | struct super_block *ac_sb; | |
150 | ||
151 | /* original request */ | |
152 | struct ext4_free_extent ac_o_ex; | |
153 | ||
58696f3a | 154 | /* goal request (normalized ac_o_ex) */ |
8f6e39a7 MC |
155 | struct ext4_free_extent ac_g_ex; |
156 | ||
157 | /* the best found extent */ | |
158 | struct ext4_free_extent ac_b_ex; | |
159 | ||
ff3fc173 | 160 | /* copy of the best found extent taken before preallocation efforts */ |
8f6e39a7 MC |
161 | struct ext4_free_extent ac_f_ex; |
162 | ||
8f6e39a7 MC |
163 | __u16 ac_groups_scanned; |
164 | __u16 ac_found; | |
165 | __u16 ac_tail; | |
166 | __u16 ac_buddy; | |
167 | __u16 ac_flags; /* allocation hints */ | |
168 | __u8 ac_status; | |
169 | __u8 ac_criteria; | |
8f6e39a7 MC |
170 | __u8 ac_2order; /* if request is to allocate 2^N blocks and |
171 | * N > 0, the field stores N, otherwise 0 */ | |
172 | __u8 ac_op; /* operation, for history only */ | |
173 | struct page *ac_bitmap_page; | |
174 | struct page *ac_buddy_page; | |
175 | struct ext4_prealloc_space *ac_pa; | |
176 | struct ext4_locality_group *ac_lg; | |
177 | }; | |
178 | ||
179 | #define AC_STATUS_CONTINUE 1 | |
180 | #define AC_STATUS_FOUND 2 | |
181 | #define AC_STATUS_BREAK 3 | |
182 | ||
8f6e39a7 MC |
183 | struct ext4_buddy { |
184 | struct page *bd_buddy_page; | |
185 | void *bd_buddy; | |
186 | struct page *bd_bitmap_page; | |
187 | void *bd_bitmap; | |
188 | struct ext4_group_info *bd_info; | |
189 | struct super_block *bd_sb; | |
190 | __u16 bd_blkbits; | |
191 | ext4_group_t bd_group; | |
192 | }; | |
8f6e39a7 | 193 | |
c3a326a6 | 194 | static inline ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb, |
8f6e39a7 MC |
195 | struct ext4_free_extent *fex) |
196 | { | |
3212a80a TT |
197 | return ext4_group_first_block_no(sb, fex->fe_group) + |
198 | (fex->fe_start << EXT4_SB(sb)->s_cluster_bits); | |
8f6e39a7 | 199 | } |
0c9ec4be DW |
200 | |
201 | typedef int (*ext4_mballoc_query_range_fn)( | |
202 | struct super_block *sb, | |
203 | ext4_group_t agno, | |
204 | ext4_grpblk_t start, | |
205 | ext4_grpblk_t len, | |
206 | void *priv); | |
207 | ||
208 | int | |
209 | ext4_mballoc_query_range( | |
210 | struct super_block *sb, | |
211 | ext4_group_t agno, | |
212 | ext4_grpblk_t start, | |
213 | ext4_grpblk_t end, | |
214 | ext4_mballoc_query_range_fn formatter, | |
215 | void *priv); | |
216 | ||
8f6e39a7 | 217 | #endif |