1 |
/* |
2 |
* PowerPC spinlock operations. |
3 |
* With the help of various source: Darwin, Linux kernel... |
4 |
* |
5 |
* $Id: spinlock.h.in,v 1.1 2007/05/24 12:54:10 mbuna Exp $ |
6 |
*/ |
7 |
#ifndef INCLUDED_PEAK_SPINLOCK_H_ |
8 |
#define INCLUDED_PEAK_SPINLOCK_H_ |
9 |
|
10 |
#define PEAK_SPINLOCK_ITERATIONS 1000 |
11 |
|
12 |
#ifdef HAVE_CONFIG_H |
13 |
#include "config.h" |
14 |
#endif |
15 |
#include <sys/types.h> |
16 |
#include <sys/time.h> |
17 |
#include <sys/signal.h> |
18 |
#include <pthread.h> |
19 |
#include <unistd.h> |
20 |
#ifdef HAVE_SCHED_H |
21 |
#include <sched.h> |
22 |
#endif |
23 |
#ifdef HAVE_ATOMIC_OP_H |
24 |
#include <sys/atomic_op.h> /* AIX */ |
25 |
#endif |
26 |
|
27 |
#include <peak/stdint.h> |
28 |
|
29 |
/* Erratum #77 on the 405 means we need a sync or dcbt before every stwcx. |
30 |
* The old ATOMIC_SYNC_FIX covered some but not all of this. |
31 |
*/ |
32 |
#ifdef CONFIG_IBM405_ERR77 |
33 |
#define PPC405_ERR77(ra,rb) "dcbt " #ra "," #rb ";" |
34 |
#else |
35 |
#define PPC405_ERR77(ra,rb) |
36 |
#endif |
37 |
|
38 |
/* PEAK defaults to SMP |
39 |
*/ |
40 |
#ifdef PEAK_CONFIG_UNIPROCESSOR |
41 |
#define SMP_WMB |
42 |
#define SMP_MB |
43 |
#else |
44 |
#define SMP_WMB "eieio\n" |
45 |
#define SMP_MB "\nsync" |
46 |
#endif /* PEAK_CONFIG_UNIPROCESSOR */ |
47 |
|
48 |
|
49 |
#if defined(__cplusplus) |
50 |
extern "C" { |
51 |
#endif |
52 |
|
53 |
/* PEAK INTERNAL SPIN LOCK |
54 |
* |
55 |
* _peak_spinlock_lock(lockp) |
56 |
* _peak_spinlock_lock_try(lockp) - returns 0 (busy) or 1 (got the lock) |
57 |
* _peak_spinlock_unlock(lockp) |
58 |
* |
59 |
*/ |
60 |
|
61 |
extern int _peak_is_threaded; |
62 |
|
63 |
#if defined(HAVE__SPIN_LOCK) && defined(HAVE__SPIN_LOCK_TRY) && defined(HAVE__SPIN_UNLOCK) |
64 |
#define USE_DARWIN_SPINLOCK 1 |
65 |
#else |
66 |
#define USE_DARWIN_SPINLOCK 0 |
67 |
#endif |
68 |
|
69 |
/* Type and initializer for this architecture. |
70 |
*/ |
71 |
#if USE_DARWIN_SPINLOCK |
72 |
typedef volatile long peak_spinlock_t; |
73 |
#define PEAK_SPINLOCK_INITIALIZER (0) |
74 |
#elif defined(__GNUC__) |
75 |
typedef struct { volatile unsigned long lock; } peak_spinlock_t; |
76 |
#define PEAK_SPINLOCK_INITIALIZER ((peak_spinlock_t){ (0) }) |
77 |
#elif defined(_AIX) |
78 |
typedef int peak_spinlock_t; /* AIX: typedef int *atomic_p; */ |
79 |
#define PEAK_SPINLOCK_INITIALIZER (0) |
80 |
#endif |
81 |
|
82 |
#if USE_DARWIN_SPINLOCK |
83 |
extern void _spin_lock(peak_spinlock_t *lockp); |
84 |
extern int _spin_lock_try(peak_spinlock_t *lockp); |
85 |
extern void _spin_unlock(peak_spinlock_t *lockp); |
86 |
#endif |
87 |
|
88 |
static inline void |
89 |
_peak_spinlock_lock(peak_spinlock_t *lockp) |
90 |
{ |
91 |
#if USE_DARWIN_SPINLOCK |
92 |
|
93 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
94 |
return; |
95 |
|
96 |
_spin_lock(lockp); /* Lucky, we have a system function for that. |
97 |
* I checked on Darwin 7.0 and it deals properly |
98 |
* with 32 or 64 bit, UP (always depress) or |
99 |
* MP (1000 spin tries before relinquish). |
100 |
* On Darwin 6.x, however, the kernel use generic |
101 |
* PPC MP code, but well. --mbuna |
102 |
*/ |
103 |
|
104 |
#elif defined(__GNUC__) |
105 |
|
106 |
unsigned long tmp; |
107 |
|
108 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
109 |
return; |
110 |
|
111 |
__asm__ __volatile__( |
112 |
"b 1f # spin_lock\n\ |
113 |
2: lwzx %0,0,%1\n\ |
114 |
cmpwi 0,%0,0\n\ |
115 |
bne+ 2b\n\ |
116 |
1: lwarx %0,0,%1\n\ |
117 |
cmpwi 0,%0,0\n\ |
118 |
bne- 2b\n" |
119 |
PPC405_ERR77(0,%1) |
120 |
" stwcx. %2,0,%1\n\ |
121 |
bne- 2b\n\ |
122 |
isync" |
123 |
: "=&r"(tmp) |
124 |
: "r"(&lockp->lock), "r"(1) |
125 |
: "cr0", "memory"); |
126 |
|
127 |
#elif defined(_AIX) |
128 |
|
129 |
unsigned int tries = PEAK_SPINLOCK_ITERATIONS; |
130 |
|
131 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
132 |
return; |
133 |
|
134 |
while (!_check_lock(lockp, 0, 1)) |
135 |
{ |
136 |
if (--tries > 0) |
137 |
{ |
138 |
sched_yield(); |
139 |
tries = PEAK_SPINLOCK_ITERATIONS; |
140 |
} |
141 |
} |
142 |
|
143 |
#else |
144 |
#error _peak_spinlock_lock not supported |
145 |
#endif /* USE_DARWIN_SPINLOCK */ |
146 |
} |
147 |
|
148 |
static inline int |
149 |
_peak_spinlock_lock_try(peak_spinlock_t *lockp) |
150 |
{ |
151 |
#if USE_DARWIN_SPINLOCK |
152 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
153 |
return 1; /* always succeed */ |
154 |
|
155 |
return _spin_lock_try(lockp); |
156 |
#elif defined(__GNUC__) |
157 |
unsigned int old, t; |
158 |
unsigned int mask = 1; |
159 |
volatile unsigned int *p = ((volatile unsigned int *)lockp->lock); |
160 |
|
161 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
162 |
return 1; /* always succeed */ |
163 |
|
164 |
__asm__ __volatile__(SMP_WMB "\n\ |
165 |
1: lwarx %0,0,%4 \n\ |
166 |
or %1,%0,%3 \n" |
167 |
PPC405_ERR77(0,%4) |
168 |
" stwcx. %1,0,%4 \n\ |
169 |
bne 1b" |
170 |
SMP_MB |
171 |
: "=&r" (old), "=&r" (t), "=m" (*p) |
172 |
: "r" (mask), "r" (p), "m" (*p) |
173 |
: "cc", "memory"); |
174 |
|
175 |
return (old & mask) == 0; |
176 |
#elif defined(_AIX) |
177 |
return !_check_lock(lockp, 0, 1); |
178 |
#else |
179 |
#error _peak_spinlock_lock_try not supported |
180 |
#endif |
181 |
} |
182 |
|
183 |
static inline void |
184 |
_peak_spinlock_unlock(peak_spinlock_t *lockp) |
185 |
{ |
186 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
187 |
return; |
188 |
|
189 |
#if USE_DARWIN_SPINLOCK |
190 |
if (!_peak_is_threaded) /* set only if peak uses several threads */ |
191 |
return; |
192 |
|
193 |
_spin_unlock(lockp); |
194 |
|
195 |
#elif defined(__GNUC__) |
196 |
/* The eieio instruction is a faster sync: a barrier providing an ordering |
197 |
* (separately) for (a) cacheable stores and (b) loads and stores to |
198 |
* non-cacheable memory (e.g. I/O devices). |
199 |
*/ |
200 |
__asm__ __volatile__("eieio # spin_unlock": : :"memory"); |
201 |
lockp->lock = 0; |
202 |
#elif defined(_AIX) |
203 |
_clear_lock(lockp, 0); |
204 |
#else |
205 |
#error _peak_spinlock_unlock not supported |
206 |
#endif |
207 |
} |
208 |
|
209 |
#if defined(__cplusplus) |
210 |
} |
211 |
#endif |
212 |
|
213 |
#endif /* INCLUDED_PEAK_SPINLOCK_H_ */ |