blob: 3ccd889ba953c10f7c6e38200e8873c5ee344b2a [file] [log] [blame]
Mingming Cao8f6e39a2008-04-29 22:01:31 -04001/*
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>
Theodore Ts'o8a0aba72008-10-16 10:06:27 -040020#include <linux/blkdev.h>
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -050021#include <linux/mutex.h>
Mingming Cao8f6e39a2008-04-29 22:01:31 -040022#include "ext4_jbd2.h"
23#include "ext4.h"
Mingming Cao8f6e39a2008-04-29 22:01:31 -040024
25/*
26 * with AGGRESSIVE_CHECK allocator runs consistency checks over
27 * structures. these checks slow things down a lot
28 */
29#define AGGRESSIVE_CHECK__
30
31/*
32 * with DOUBLE_CHECK defined mballoc creates persistent in-core
33 * bitmaps, maintains and uses them to check for double allocations
34 */
35#define DOUBLE_CHECK__
36
37/*
38 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -040039#ifdef CONFIG_EXT4_DEBUG
40extern u8 mb_enable_debug;
41
42#define mb_debug(n, fmt, a...) \
43 do { \
44 if ((n) <= mb_enable_debug) { \
45 printk(KERN_DEBUG "(%s, %d): %s: ", \
46 __FILE__, __LINE__, __func__); \
47 printk(fmt, ## a); \
48 } \
49 } while (0)
Mingming Cao8f6e39a2008-04-29 22:01:31 -040050#else
Theodore Ts'o6ba495e2009-09-18 13:38:55 -040051#define mb_debug(n, fmt, a...)
Mingming Cao8f6e39a2008-04-29 22:01:31 -040052#endif
53
Mingming Cao8f6e39a2008-04-29 22:01:31 -040054#define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
55#define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
Mingming Cao8f6e39a2008-04-29 22:01:31 -040056
57/*
58 * How long mballoc can look for a best extent (in found extents)
59 */
60#define MB_DEFAULT_MAX_TO_SCAN 200
61
62/*
63 * How long mballoc must look for a best extent
64 */
65#define MB_DEFAULT_MIN_TO_SCAN 10
66
67/*
Mingming Cao8f6e39a2008-04-29 22:01:31 -040068 * with 'ext4_mb_stats' allocator will collect stats that will be
69 * shown at umount. The collecting costs though!
70 */
Theodore Ts'o90576c02009-09-29 15:51:30 -040071#define MB_DEFAULT_STATS 0
Mingming Cao8f6e39a2008-04-29 22:01:31 -040072
73/*
74 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
75 * by the stream allocator, which purpose is to pack requests
76 * as close each to other as possible to produce smooth I/O traffic
77 * We use locality group prealloc space for stream request.
78 * We can tune the same via /proc/fs/ext4/<parition>/stream_req
79 */
80#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
81
82/*
83 * for which requests use 2^N search using buddies
84 */
85#define MB_DEFAULT_ORDER2_REQS 2
86
87/*
88 * default group prealloc size 512 blocks
89 */
90#define MB_DEFAULT_GROUP_PREALLOC 512
91
Mingming Cao8f6e39a2008-04-29 22:01:31 -040092
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -040093struct ext4_free_data {
Bobi Jam18aadd42012-02-20 17:53:02 -050094 /* MUST be the first member */
95 struct ext4_journal_cb_entry efd_jce;
Mingming Cao8f6e39a2008-04-29 22:01:31 -040096
Bobi Jam18aadd42012-02-20 17:53:02 -050097 /* ext4_free_data private data starts from here */
98
99 /* this links the free block information from group_info */
100 struct rb_node efd_node;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -0400101
102 /* group which free block extent belongs */
Bobi Jam18aadd42012-02-20 17:53:02 -0500103 ext4_group_t efd_group;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -0400104
105 /* free block extent */
Bobi Jam18aadd42012-02-20 17:53:02 -0500106 ext4_grpblk_t efd_start_cluster;
107 ext4_grpblk_t efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -0400108
109 /* transaction which freed this extent */
Bobi Jam18aadd42012-02-20 17:53:02 -0500110 tid_t efd_tid;
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400111};
112
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400113struct ext4_prealloc_space {
114 struct list_head pa_inode_list;
115 struct list_head pa_group_list;
116 union {
117 struct list_head pa_tmp_list;
118 struct rcu_head pa_rcu;
119 } u;
120 spinlock_t pa_lock;
121 atomic_t pa_count;
122 unsigned pa_deleted;
123 ext4_fsblk_t pa_pstart; /* phys. block */
124 ext4_lblk_t pa_lstart; /* log. block */
Eric Sandeena36b4492009-08-25 22:36:45 -0400125 ext4_grpblk_t pa_len; /* len of preallocated chunk */
126 ext4_grpblk_t pa_free; /* how many blocks are free */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -0400127 unsigned short pa_type; /* pa type. inode or group */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400128 spinlock_t *pa_obj_lock;
129 struct inode *pa_inode; /* hack, for history only */
130};
131
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -0400132enum {
133 MB_INODE_PA = 0,
134 MB_GROUP_PA = 1
135};
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400136
137struct ext4_free_extent {
138 ext4_lblk_t fe_logical;
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400139 ext4_grpblk_t fe_start; /* In cluster units */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400140 ext4_group_t fe_group;
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400141 ext4_grpblk_t fe_len; /* In cluster units */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400142};
143
144/*
145 * Locality group:
146 * we try to group all related changes together
147 * so that writeback can flush/allocate them together as well
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -0400148 * Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC
149 * (512). We store prealloc space into the hash based on the pa_free blocks
150 * order value.ie, fls(pa_free)-1;
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400151 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -0400152#define PREALLOC_TB_SIZE 10
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400153struct ext4_locality_group {
154 /* for allocator */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -0400155 /* to serialize allocates */
156 struct mutex lg_mutex;
157 /* list of preallocations */
158 struct list_head lg_prealloc_list[PREALLOC_TB_SIZE];
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400159 spinlock_t lg_prealloc_lock;
160};
161
162struct ext4_allocation_context {
163 struct inode *ac_inode;
164 struct super_block *ac_sb;
165
166 /* original request */
167 struct ext4_free_extent ac_o_ex;
168
Coly Li58696f32011-02-24 14:10:00 -0500169 /* goal request (normalized ac_o_ex) */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400170 struct ext4_free_extent ac_g_ex;
171
172 /* the best found extent */
173 struct ext4_free_extent ac_b_ex;
174
Robin Dongff3fc172011-10-31 18:55:50 -0400175 /* copy of the best found extent taken before preallocation efforts */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400176 struct ext4_free_extent ac_f_ex;
177
178 /* number of iterations done. we have to track to limit searching */
179 unsigned long ac_ex_scanned;
180 __u16 ac_groups_scanned;
181 __u16 ac_found;
182 __u16 ac_tail;
183 __u16 ac_buddy;
184 __u16 ac_flags; /* allocation hints */
185 __u8 ac_status;
186 __u8 ac_criteria;
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400187 __u8 ac_2order; /* if request is to allocate 2^N blocks and
188 * N > 0, the field stores N, otherwise 0 */
189 __u8 ac_op; /* operation, for history only */
190 struct page *ac_bitmap_page;
191 struct page *ac_buddy_page;
192 struct ext4_prealloc_space *ac_pa;
193 struct ext4_locality_group *ac_lg;
194};
195
196#define AC_STATUS_CONTINUE 1
197#define AC_STATUS_FOUND 2
198#define AC_STATUS_BREAK 3
199
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400200struct ext4_buddy {
201 struct page *bd_buddy_page;
202 void *bd_buddy;
203 struct page *bd_bitmap_page;
204 void *bd_bitmap;
205 struct ext4_group_info *bd_info;
206 struct super_block *bd_sb;
207 __u16 bd_blkbits;
208 ext4_group_t bd_group;
209};
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400210
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500211static inline ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400212 struct ext4_free_extent *fex)
213{
Theodore Ts'o3212a802011-09-09 18:46:51 -0400214 return ext4_group_first_block_no(sb, fex->fe_group) +
215 (fex->fe_start << EXT4_SB(sb)->s_cluster_bits);
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400216}
217#endif