percpu: introduce pcpu_alloc_info and pcpu_group_info
[safe/jmp/linux-2.6] / include / linux / percpu.h
1 #ifndef __LINUX_PERCPU_H
2 #define __LINUX_PERCPU_H
3
4 #include <linux/preempt.h>
5 #include <linux/slab.h> /* For kmalloc() */
6 #include <linux/smp.h>
7 #include <linux/cpumask.h>
8 #include <linux/pfn.h>
9
10 #include <asm/percpu.h>
11
12 /* enough to cover all DEFINE_PER_CPUs in modules */
13 #ifdef CONFIG_MODULES
14 #define PERCPU_MODULE_RESERVE           (8 << 10)
15 #else
16 #define PERCPU_MODULE_RESERVE           0
17 #endif
18
19 #ifndef PERCPU_ENOUGH_ROOM
20 #define PERCPU_ENOUGH_ROOM                                              \
21         (ALIGN(__per_cpu_end - __per_cpu_start, SMP_CACHE_BYTES) +      \
22          PERCPU_MODULE_RESERVE)
23 #endif
24
25 /*
26  * Must be an lvalue. Since @var must be a simple identifier,
27  * we force a syntax error here if it isn't.
28  */
29 #define get_cpu_var(var) (*({                           \
30         extern int simple_identifier_##var(void);       \
31         preempt_disable();                              \
32         &__get_cpu_var(var); }))
33 #define put_cpu_var(var) preempt_enable()
34
35 #ifdef CONFIG_SMP
36
37 #ifndef CONFIG_HAVE_LEGACY_PER_CPU_AREA
38
39 /* minimum unit size, also is the maximum supported allocation size */
40 #define PCPU_MIN_UNIT_SIZE              PFN_ALIGN(64 << 10)
41
42 /*
43  * PERCPU_DYNAMIC_RESERVE indicates the amount of free area to piggy
44  * back on the first chunk for dynamic percpu allocation if arch is
45  * manually allocating and mapping it for faster access (as a part of
46  * large page mapping for example).
47  *
48  * The following values give between one and two pages of free space
49  * after typical minimal boot (2-way SMP, single disk and NIC) with
50  * both defconfig and a distro config on x86_64 and 32.  More
51  * intelligent way to determine this would be nice.
52  */
53 #if BITS_PER_LONG > 32
54 #define PERCPU_DYNAMIC_RESERVE          (20 << 10)
55 #else
56 #define PERCPU_DYNAMIC_RESERVE          (12 << 10)
57 #endif
58
59 extern void *pcpu_base_addr;
60 extern const int *pcpu_unit_map;
61
62 struct pcpu_group_info {
63         int                     nr_units;       /* aligned # of units */
64         unsigned long           base_offset;    /* base address offset */
65         unsigned int            *cpu_map;       /* unit->cpu map, empty
66                                                  * entries contain NR_CPUS */
67 };
68
69 struct pcpu_alloc_info {
70         size_t                  static_size;
71         size_t                  reserved_size;
72         size_t                  dyn_size;
73         size_t                  unit_size;
74         size_t                  atom_size;
75         size_t                  alloc_size;
76         size_t                  __ai_size;      /* internal, don't use */
77         int                     nr_groups;      /* 0 if grouping unnecessary */
78         struct pcpu_group_info  groups[];
79 };
80
81 enum pcpu_fc {
82         PCPU_FC_AUTO,
83         PCPU_FC_EMBED,
84         PCPU_FC_PAGE,
85         PCPU_FC_LPAGE,
86
87         PCPU_FC_NR,
88 };
89 extern const char *pcpu_fc_names[PCPU_FC_NR];
90
91 extern enum pcpu_fc pcpu_chosen_fc;
92
93 typedef void * (*pcpu_fc_alloc_fn_t)(unsigned int cpu, size_t size,
94                                      size_t align);
95 typedef void (*pcpu_fc_free_fn_t)(void *ptr, size_t size);
96 typedef void (*pcpu_fc_populate_pte_fn_t)(unsigned long addr);
97 typedef int (pcpu_fc_cpu_distance_fn_t)(unsigned int from, unsigned int to);
98 typedef void (*pcpu_fc_map_fn_t)(void *ptr, size_t size, void *addr);
99
100 extern struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
101                                                              int nr_units);
102 extern void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai);
103
104 extern struct pcpu_alloc_info * __init pcpu_build_alloc_info(
105                                 size_t reserved_size, ssize_t dyn_size,
106                                 size_t atom_size,
107                                 pcpu_fc_cpu_distance_fn_t cpu_distance_fn);
108
109 extern size_t __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
110                                             void *base_addr);
111
112 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
113 extern ssize_t __init pcpu_embed_first_chunk(
114                                 size_t reserved_size, ssize_t dyn_size);
115 #endif
116
117 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
118 extern ssize_t __init pcpu_page_first_chunk(
119                                 size_t reserved_size,
120                                 pcpu_fc_alloc_fn_t alloc_fn,
121                                 pcpu_fc_free_fn_t free_fn,
122                                 pcpu_fc_populate_pte_fn_t populate_pte_fn);
123 #endif
124
125 #ifdef CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK
126 extern ssize_t __init pcpu_lpage_first_chunk(
127                                 const struct pcpu_alloc_info *ai,
128                                 pcpu_fc_alloc_fn_t alloc_fn,
129                                 pcpu_fc_free_fn_t free_fn,
130                                 pcpu_fc_map_fn_t map_fn);
131
132 extern void *pcpu_lpage_remapped(void *kaddr);
133 #else
134 static inline void *pcpu_lpage_remapped(void *kaddr)
135 {
136         return NULL;
137 }
138 #endif
139
140 /*
141  * Use this to get to a cpu's version of the per-cpu object
142  * dynamically allocated. Non-atomic access to the current CPU's
143  * version should probably be combined with get_cpu()/put_cpu().
144  */
145 #define per_cpu_ptr(ptr, cpu)   SHIFT_PERCPU_PTR((ptr), per_cpu_offset((cpu)))
146
147 extern void *__alloc_reserved_percpu(size_t size, size_t align);
148
149 #else /* CONFIG_HAVE_LEGACY_PER_CPU_AREA */
150
151 struct percpu_data {
152         void *ptrs[1];
153 };
154
155 /* pointer disguising messes up the kmemleak objects tracking */
156 #ifndef CONFIG_DEBUG_KMEMLEAK
157 #define __percpu_disguise(pdata) (struct percpu_data *)~(unsigned long)(pdata)
158 #else
159 #define __percpu_disguise(pdata) (struct percpu_data *)(pdata)
160 #endif
161
162 #define per_cpu_ptr(ptr, cpu)                                           \
163 ({                                                                      \
164         struct percpu_data *__p = __percpu_disguise(ptr);               \
165         (__typeof__(ptr))__p->ptrs[(cpu)];                              \
166 })
167
168 #endif /* CONFIG_HAVE_LEGACY_PER_CPU_AREA */
169
170 extern void *__alloc_percpu(size_t size, size_t align);
171 extern void free_percpu(void *__pdata);
172
173 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
174 extern void __init setup_per_cpu_areas(void);
175 #endif
176
177 #else /* CONFIG_SMP */
178
179 #define per_cpu_ptr(ptr, cpu) ({ (void)(cpu); (ptr); })
180
181 static inline void *__alloc_percpu(size_t size, size_t align)
182 {
183         /*
184          * Can't easily make larger alignment work with kmalloc.  WARN
185          * on it.  Larger alignment should only be used for module
186          * percpu sections on SMP for which this path isn't used.
187          */
188         WARN_ON_ONCE(align > SMP_CACHE_BYTES);
189         return kzalloc(size, GFP_KERNEL);
190 }
191
192 static inline void free_percpu(void *p)
193 {
194         kfree(p);
195 }
196
197 static inline void __init setup_per_cpu_areas(void) { }
198
199 static inline void *pcpu_lpage_remapped(void *kaddr)
200 {
201         return NULL;
202 }
203
204 #endif /* CONFIG_SMP */
205
206 #define alloc_percpu(type)      (type *)__alloc_percpu(sizeof(type), \
207                                                        __alignof__(type))
208
209 /*
210  * Optional methods for optimized non-lvalue per-cpu variable access.
211  *
212  * @var can be a percpu variable or a field of it and its size should
213  * equal char, int or long.  percpu_read() evaluates to a lvalue and
214  * all others to void.
215  *
216  * These operations are guaranteed to be atomic w.r.t. preemption.
217  * The generic versions use plain get/put_cpu_var().  Archs are
218  * encouraged to implement single-instruction alternatives which don't
219  * require preemption protection.
220  */
221 #ifndef percpu_read
222 # define percpu_read(var)                                               \
223   ({                                                                    \
224         typeof(per_cpu_var(var)) __tmp_var__;                           \
225         __tmp_var__ = get_cpu_var(var);                                 \
226         put_cpu_var(var);                                               \
227         __tmp_var__;                                                    \
228   })
229 #endif
230
231 #define __percpu_generic_to_op(var, val, op)                            \
232 do {                                                                    \
233         get_cpu_var(var) op val;                                        \
234         put_cpu_var(var);                                               \
235 } while (0)
236
237 #ifndef percpu_write
238 # define percpu_write(var, val)         __percpu_generic_to_op(var, (val), =)
239 #endif
240
241 #ifndef percpu_add
242 # define percpu_add(var, val)           __percpu_generic_to_op(var, (val), +=)
243 #endif
244
245 #ifndef percpu_sub
246 # define percpu_sub(var, val)           __percpu_generic_to_op(var, (val), -=)
247 #endif
248
249 #ifndef percpu_and
250 # define percpu_and(var, val)           __percpu_generic_to_op(var, (val), &=)
251 #endif
252
253 #ifndef percpu_or
254 # define percpu_or(var, val)            __percpu_generic_to_op(var, (val), |=)
255 #endif
256
257 #ifndef percpu_xor
258 # define percpu_xor(var, val)           __percpu_generic_to_op(var, (val), ^=)
259 #endif
260
261 #endif /* __LINUX_PERCPU_H */