blob: 3ef1df6ae9ec6f67102c52cf9624a317e321d19b [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/*
Mingming Cao8f6e39a2008-04-29 22:01:31 -040026 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -040027#ifdef CONFIG_EXT4_DEBUG
Theodore Ts'oa0b30c12013-02-09 16:28:20 -050028extern ushort ext4_mballoc_debug;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -040029
30#define mb_debug(n, fmt, a...) \
31 do { \
Theodore Ts'oa0b30c12013-02-09 16:28:20 -050032 if ((n) <= ext4_mballoc_debug) { \
Theodore Ts'o6ba495e2009-09-18 13:38:55 -040033 printk(KERN_DEBUG "(%s, %d): %s: ", \
34 __FILE__, __LINE__, __func__); \
35 printk(fmt, ## a); \
36 } \
37 } while (0)
Mingming Cao8f6e39a2008-04-29 22:01:31 -040038#else
Patrick Palka024949e2014-02-17 20:50:59 -050039#define mb_debug(n, fmt, a...) no_printk(fmt, ## a)
Mingming Cao8f6e39a2008-04-29 22:01:31 -040040#endif
41
Mingming Cao8f6e39a2008-04-29 22:01:31 -040042#define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
43#define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
Mingming Cao8f6e39a2008-04-29 22:01:31 -040044
45/*
46 * How long mballoc can look for a best extent (in found extents)
47 */
48#define MB_DEFAULT_MAX_TO_SCAN 200
49
50/*
51 * How long mballoc must look for a best extent
52 */
53#define MB_DEFAULT_MIN_TO_SCAN 10
54
55/*
Mingming Cao8f6e39a2008-04-29 22:01:31 -040056 * with 'ext4_mb_stats' allocator will collect stats that will be
57 * shown at umount. The collecting costs though!
58 */
Theodore Ts'o90576c02009-09-29 15:51:30 -040059#define MB_DEFAULT_STATS 0
Mingming Cao8f6e39a2008-04-29 22:01:31 -040060
61/*
62 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
63 * by the stream allocator, which purpose is to pack requests
64 * as close each to other as possible to produce smooth I/O traffic
65 * We use locality group prealloc space for stream request.
66 * We can tune the same via /proc/fs/ext4/<parition>/stream_req
67 */
68#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
69
70/*
71 * for which requests use 2^N search using buddies
72 */
73#define MB_DEFAULT_ORDER2_REQS 2
74
75/*
76 * default group prealloc size 512 blocks
77 */
78#define MB_DEFAULT_GROUP_PREALLOC 512
79
Mingming Cao8f6e39a2008-04-29 22:01:31 -040080
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -040081struct ext4_free_data {
Bobi Jam18aadd42012-02-20 17:53:02 -050082 /* MUST be the first member */
83 struct ext4_journal_cb_entry efd_jce;
Mingming Cao8f6e39a2008-04-29 22:01:31 -040084
Bobi Jam18aadd42012-02-20 17:53:02 -050085 /* ext4_free_data private data starts from here */
86
87 /* this links the free block information from group_info */
88 struct rb_node efd_node;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -040089
90 /* group which free block extent belongs */
Bobi Jam18aadd42012-02-20 17:53:02 -050091 ext4_group_t efd_group;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -040092
93 /* free block extent */
Bobi Jam18aadd42012-02-20 17:53:02 -050094 ext4_grpblk_t efd_start_cluster;
95 ext4_grpblk_t efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -040096
97 /* transaction which freed this extent */
Bobi Jam18aadd42012-02-20 17:53:02 -050098 tid_t efd_tid;
Mingming Cao8f6e39a2008-04-29 22:01:31 -040099};
100
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400101struct ext4_prealloc_space {
102 struct list_head pa_inode_list;
103 struct list_head pa_group_list;
104 union {
105 struct list_head pa_tmp_list;
106 struct rcu_head pa_rcu;
107 } u;
108 spinlock_t pa_lock;
109 atomic_t pa_count;
110 unsigned pa_deleted;
111 ext4_fsblk_t pa_pstart; /* phys. block */
112 ext4_lblk_t pa_lstart; /* log. block */
Eric Sandeena36b4492009-08-25 22:36:45 -0400113 ext4_grpblk_t pa_len; /* len of preallocated chunk */
114 ext4_grpblk_t pa_free; /* how many blocks are free */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -0400115 unsigned short pa_type; /* pa type. inode or group */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400116 spinlock_t *pa_obj_lock;
117 struct inode *pa_inode; /* hack, for history only */
118};
119
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -0400120enum {
121 MB_INODE_PA = 0,
122 MB_GROUP_PA = 1
123};
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400124
125struct ext4_free_extent {
126 ext4_lblk_t fe_logical;
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400127 ext4_grpblk_t fe_start; /* In cluster units */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400128 ext4_group_t fe_group;
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400129 ext4_grpblk_t fe_len; /* In cluster units */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400130};
131
132/*
133 * Locality group:
134 * we try to group all related changes together
135 * so that writeback can flush/allocate them together as well
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -0400136 * Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC
137 * (512). We store prealloc space into the hash based on the pa_free blocks
138 * order value.ie, fls(pa_free)-1;
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400139 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -0400140#define PREALLOC_TB_SIZE 10
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400141struct ext4_locality_group {
142 /* for allocator */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -0400143 /* to serialize allocates */
144 struct mutex lg_mutex;
145 /* list of preallocations */
146 struct list_head lg_prealloc_list[PREALLOC_TB_SIZE];
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400147 spinlock_t lg_prealloc_lock;
148};
149
150struct ext4_allocation_context {
151 struct inode *ac_inode;
152 struct super_block *ac_sb;
153
154 /* original request */
155 struct ext4_free_extent ac_o_ex;
156
Coly Li58696f32011-02-24 14:10:00 -0500157 /* goal request (normalized ac_o_ex) */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400158 struct ext4_free_extent ac_g_ex;
159
160 /* the best found extent */
161 struct ext4_free_extent ac_b_ex;
162
Robin Dongff3fc172011-10-31 18:55:50 -0400163 /* copy of the best found extent taken before preallocation efforts */
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400164 struct ext4_free_extent ac_f_ex;
165
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400166 __u16 ac_groups_scanned;
167 __u16 ac_found;
168 __u16 ac_tail;
169 __u16 ac_buddy;
170 __u16 ac_flags; /* allocation hints */
171 __u8 ac_status;
172 __u8 ac_criteria;
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400173 __u8 ac_2order; /* if request is to allocate 2^N blocks and
174 * N > 0, the field stores N, otherwise 0 */
175 __u8 ac_op; /* operation, for history only */
176 struct page *ac_bitmap_page;
177 struct page *ac_buddy_page;
178 struct ext4_prealloc_space *ac_pa;
179 struct ext4_locality_group *ac_lg;
180};
181
182#define AC_STATUS_CONTINUE 1
183#define AC_STATUS_FOUND 2
184#define AC_STATUS_BREAK 3
185
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400186struct ext4_buddy {
187 struct page *bd_buddy_page;
188 void *bd_buddy;
189 struct page *bd_bitmap_page;
190 void *bd_bitmap;
191 struct ext4_group_info *bd_info;
192 struct super_block *bd_sb;
193 __u16 bd_blkbits;
194 ext4_group_t bd_group;
195};
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400196
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500197static inline ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400198 struct ext4_free_extent *fex)
199{
Theodore Ts'o3212a802011-09-09 18:46:51 -0400200 return ext4_group_first_block_no(sb, fex->fe_group) +
201 (fex->fe_start << EXT4_SB(sb)->s_cluster_bits);
Mingming Cao8f6e39a2008-04-29 22:01:31 -0400202}
203#endif