aboutsummaryrefslogtreecommitdiffstats
path: root/include/linux
diff options
context:
space:
mode:
authorEric Dumazet <dada1@cosmosbay.com>2008-01-03 23:41:28 -0500
committerDavid S. Miller <davem@davemloft.net>2008-01-28 18:00:35 -0500
commit571e7682026fd0e25833d103a3eeb74be29bf199 (patch)
tree3707c7fb4ea4384a163cddc9fac76e9d8860a109 /include/linux
parent789675e216617b1331875c42a81f58227a06df91 (diff)
[LIB] pcounter : unline too big functions
Before pushing pcounter to Linus tree, I would like to make some adjustments. Goal is to reduce kernel text size, by unlining too big functions. When a pcounter is bound to a statically defined per_cpu variable, we define two small helpers functions. (No more folding function using the fat for_each_possible_cpu(cpu) ... ) static DEFINE_PER_CPU(int, NAME##_pcounter_values); static void NAME##_pcounter_add(struct pcounter *self, int val) { __get_cpu_var(NAME##_pcounter_values) += val; } static int NAME##_pcounter_getval(const struct pcounter *self, int cpu) { return per_cpu(NAME##_pcounter_values, cpu); } Fast path is therefore unchanged, while folding/alloc/free is now unlined. This saves 228 bytes on i386 Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'include/linux')
-rw-r--r--include/linux/pcounter.h80
1 files changed, 26 insertions, 54 deletions
diff --git a/include/linux/pcounter.h b/include/linux/pcounter.h
index 9c4760a328f3..a82d9f2628ca 100644
--- a/include/linux/pcounter.h
+++ b/include/linux/pcounter.h
@@ -1,41 +1,39 @@
1#ifndef __LINUX_PCOUNTER_H 1#ifndef __LINUX_PCOUNTER_H
2#define __LINUX_PCOUNTER_H 2#define __LINUX_PCOUNTER_H
3 3/*
4 * Using a dynamic percpu 'int' variable has a cost :
5 * 1) Extra dereference
6 * Current per_cpu_ptr() implementation uses an array per 'percpu variable'.
7 * 2) memory cost of NR_CPUS*(32+sizeof(void *)) instead of num_possible_cpus()*4
8 *
9 * This pcounter implementation is an abstraction to be able to use
10 * either a static or a dynamic per cpu variable.
11 * One dynamic per cpu variable gets a fast & cheap implementation, we can
12 * change pcounter implementation too.
13 */
4struct pcounter { 14struct pcounter {
5#ifdef CONFIG_SMP 15#ifdef CONFIG_SMP
6 void (*add)(struct pcounter *self, int inc); 16 void (*add)(struct pcounter *self, int inc);
7 int (*getval)(const struct pcounter *self); 17 int (*getval)(const struct pcounter *self, int cpu);
8 int *per_cpu_values; 18 int *per_cpu_values;
9#else 19#else
10 int val; 20 int val;
11#endif 21#endif
12}; 22};
13 23
14/*
15 * Special macros to let pcounters use a fast version of {getvalue|add}
16 * using a static percpu variable per pcounter instead of an allocated one,
17 * saving one dereference.
18 * This might be changed if/when dynamic percpu vars become fast.
19 */
20#ifdef CONFIG_SMP 24#ifdef CONFIG_SMP
21#include <linux/cpumask.h>
22#include <linux/percpu.h> 25#include <linux/percpu.h>
23 26
24#define DEFINE_PCOUNTER(NAME) \ 27#define DEFINE_PCOUNTER(NAME) \
25static DEFINE_PER_CPU(int, NAME##_pcounter_values); \ 28static DEFINE_PER_CPU(int, NAME##_pcounter_values); \
26static void NAME##_pcounter_add(struct pcounter *self, int inc) \ 29static void NAME##_pcounter_add(struct pcounter *self, int val) \
27{ \ 30{ \
28 __get_cpu_var(NAME##_pcounter_values) += inc; \ 31 __get_cpu_var(NAME##_pcounter_values) += val; \
29} \ 32} \
30 \ 33static int NAME##_pcounter_getval(const struct pcounter *self, int cpu) \
31static int NAME##_pcounter_getval(const struct pcounter *self) \ 34{ \
32{ \ 35 return per_cpu(NAME##_pcounter_values, cpu); \
33 int res = 0, cpu; \ 36} \
34 \
35 for_each_possible_cpu(cpu) \
36 res += per_cpu(NAME##_pcounter_values, cpu); \
37 return res; \
38}
39 37
40#define PCOUNTER_MEMBER_INITIALIZER(NAME, MEMBER) \ 38#define PCOUNTER_MEMBER_INITIALIZER(NAME, MEMBER) \
41 MEMBER = { \ 39 MEMBER = { \
@@ -43,42 +41,16 @@ static int NAME##_pcounter_getval(const struct pcounter *self) \
43 .getval = NAME##_pcounter_getval, \ 41 .getval = NAME##_pcounter_getval, \
44 } 42 }
45 43
46extern void pcounter_def_add(struct pcounter *self, int inc);
47extern int pcounter_def_getval(const struct pcounter *self);
48
49static inline int pcounter_alloc(struct pcounter *self)
50{
51 int rc = 0;
52 if (self->add == NULL) {
53 self->per_cpu_values = alloc_percpu(int);
54 if (self->per_cpu_values != NULL) {
55 self->add = pcounter_def_add;
56 self->getval = pcounter_def_getval;
57 } else
58 rc = 1;
59 }
60 return rc;
61}
62
63static inline void pcounter_free(struct pcounter *self)
64{
65 if (self->per_cpu_values != NULL) {
66 free_percpu(self->per_cpu_values);
67 self->per_cpu_values = NULL;
68 self->getval = NULL;
69 self->add = NULL;
70 }
71}
72 44
73static inline void pcounter_add(struct pcounter *self, int inc) 45static inline void pcounter_add(struct pcounter *self, int inc)
74{ 46{
75 self->add(self, inc); 47 self->add(self, inc);
76} 48}
77 49
78static inline int pcounter_getval(const struct pcounter *self) 50extern int pcounter_getval(const struct pcounter *self);
79{ 51extern int pcounter_alloc(struct pcounter *self);
80 return self->getval(self); 52extern void pcounter_free(struct pcounter *self);
81} 53
82 54
83#else /* CONFIG_SMP */ 55#else /* CONFIG_SMP */
84 56