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/<partition>/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 /*
82  * Number of groups to search linearly before performing group scanning
83  * optimization.
84  */
85 #define MB_DEFAULT_LINEAR_LIMIT		4
86 
87 /*
88  * Minimum number of groups that should be present in the file system to perform
89  * group scanning optimizations.
90  */
91 #define MB_DEFAULT_LINEAR_SCAN_THRESHOLD	16
92 
93 /*
94  * Number of valid buddy orders
95  */
96 #define MB_NUM_ORDERS(sb)		((sb)->s_blocksize_bits + 2)
97 
98 struct ext4_free_data {
99 	/* this links the free block information from sb_info */
100 	struct list_head		efd_list;
101 
102 	/* this links the free block information from group_info */
103 	struct rb_node			efd_node;
104 
105 	/* group which free block extent belongs */
106 	ext4_group_t			efd_group;
107 
108 	/* free block extent */
109 	ext4_grpblk_t			efd_start_cluster;
110 	ext4_grpblk_t			efd_count;
111 
112 	/* transaction which freed this extent */
113 	tid_t				efd_tid;
114 };
115 
116 struct ext4_prealloc_space {
117 	struct list_head	pa_inode_list;
118 	struct list_head	pa_group_list;
119 	union {
120 		struct list_head pa_tmp_list;
121 		struct rcu_head	pa_rcu;
122 	} u;
123 	spinlock_t		pa_lock;
124 	atomic_t		pa_count;
125 	unsigned		pa_deleted;
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 */
133 };
134 
135 enum {
136 	MB_INODE_PA = 0,
137 	MB_GROUP_PA = 1
138 };
139 
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 */
145 };
146 
147 /*
148  * Locality group:
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;
154  */
155 #define PREALLOC_TB_SIZE 10
156 struct ext4_locality_group {
157 	/* for allocator */
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;
163 };
164 
165 struct ext4_allocation_context {
166 	struct inode *ac_inode;
167 	struct super_block *ac_sb;
168 
169 	/* original request */
170 	struct ext4_free_extent ac_o_ex;
171 
172 	/* goal request (normalized ac_o_ex) */
173 	struct ext4_free_extent ac_g_ex;
174 
175 	/* the best found extent */
176 	struct ext4_free_extent ac_b_ex;
177 
178 	/* copy of the best found extent taken before preallocation efforts */
179 	struct ext4_free_extent ac_f_ex;
180 
181 	__u32 ac_groups_considered;
182 	__u32 ac_flags;		/* allocation hints */
183 	__u16 ac_groups_scanned;
184 	__u16 ac_groups_linear_remaining;
185 	__u16 ac_found;
186 	__u16 ac_tail;
187 	__u16 ac_buddy;
188 	__u8 ac_status;
189 	__u8 ac_criteria;
190 	__u8 ac_2order;		/* if request is to allocate 2^N blocks and
191 				 * N > 0, the field stores N, otherwise 0 */
192 	__u8 ac_op;		/* operation, for history only */
193 	struct page *ac_bitmap_page;
194 	struct page *ac_buddy_page;
195 	struct ext4_prealloc_space *ac_pa;
196 	struct ext4_locality_group *ac_lg;
197 };
198 
199 #define AC_STATUS_CONTINUE	1
200 #define AC_STATUS_FOUND		2
201 #define AC_STATUS_BREAK		3
202 
203 struct ext4_buddy {
204 	struct page *bd_buddy_page;
205 	void *bd_buddy;
206 	struct page *bd_bitmap_page;
207 	void *bd_bitmap;
208 	struct ext4_group_info *bd_info;
209 	struct super_block *bd_sb;
210 	__u16 bd_blkbits;
211 	ext4_group_t bd_group;
212 };
213 
ext4_grp_offs_to_block(struct super_block * sb,struct ext4_free_extent * fex)214 static inline ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
215 					struct ext4_free_extent *fex)
216 {
217 	return ext4_group_first_block_no(sb, fex->fe_group) +
218 		(fex->fe_start << EXT4_SB(sb)->s_cluster_bits);
219 }
220 
221 typedef int (*ext4_mballoc_query_range_fn)(
222 	struct super_block		*sb,
223 	ext4_group_t			agno,
224 	ext4_grpblk_t			start,
225 	ext4_grpblk_t			len,
226 	void				*priv);
227 
228 int
229 ext4_mballoc_query_range(
230 	struct super_block		*sb,
231 	ext4_group_t			agno,
232 	ext4_grpblk_t			start,
233 	ext4_grpblk_t			end,
234 	ext4_mballoc_query_range_fn	formatter,
235 	void				*priv);
236 
237 #endif
238