Add branch prediction, fix xchg for -mtune=core2
[urcu.git] / urcu.h
1 #ifndef _URCU_H
2 #define _URCU_H
3
4 /*
5 * urcu.h
6 *
7 * Userspace RCU header
8 *
9 * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
10 *
11 * Credits for Paul e. McKenney <paulmck@linux.vnet.ibm.com>
12 * for inspiration coming from the Linux kernel RCU and rcu-preempt.
13 *
14 * The barrier, mb, rmb, wmb, atomic_inc, smp_read_barrier_depends, ACCESS_ONCE
15 * and rcu_dereference primitives come from the Linux kernel.
16 *
17 * Distributed under GPLv2
18 */
19
20 #define __USE_GNU
21 #include <stdlib.h>
22
23 /* The "volatile" is due to gcc bugs */
24 #define barrier() __asm__ __volatile__("": : :"memory")
25
26 #define likely(x) __builtin_expect(!!(x), 1)
27 #define unlikely(x) __builtin_expect(!!(x), 0)
28
29 /* x86 32/64 specific */
30 #define mb() asm volatile("mfence":::"memory")
31 #define rmb() asm volatile("lfence":::"memory")
32 #define wmb() asm volatile("sfence" ::: "memory")
33
34 static inline void atomic_inc(int *v)
35 {
36 asm volatile("lock; incl %0"
37 : "+m" (*v));
38 }
39
40 #define xchg(ptr, v) \
41 ((__typeof__(*(ptr)))__xchg((unsigned long)(v), (ptr), sizeof(*(ptr))))
42
43 struct __xchg_dummy {
44 unsigned long a[100];
45 };
46 #define __xg(x) ((struct __xchg_dummy *)(x))
47
48 /*
49 * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
50 * Note 2: xchg has side effect, so that attribute volatile is necessary,
51 * but generally the primitive is invalid, *ptr is output argument. --ANK
52 */
53 static inline unsigned long __xchg(unsigned long x, volatile void *ptr,
54 int size)
55 {
56 switch (size) {
57 case 1:
58 asm volatile("xchgb %b0,%1"
59 : "=q" (x)
60 : "m" (*__xg(ptr)), "0" (x)
61 : "memory");
62 break;
63 case 2:
64 asm volatile("xchgw %w0,%1"
65 : "=r" (x)
66 : "m" (*__xg(ptr)), "0" (x)
67 : "memory");
68 break;
69 case 4:
70 asm volatile("xchgl %k0,%1"
71 : "=r" (x)
72 : "m" (*__xg(ptr)), "0" (x)
73 : "memory");
74 break;
75 case 8:
76 asm volatile("xchgq %0,%1"
77 : "=r" (x)
78 : "m" (*__xg(ptr)), "0" (x)
79 : "memory");
80 break;
81 }
82 return x;
83 }
84
85 /* Nop everywhere except on alpha. */
86 #define smp_read_barrier_depends()
87
88 /*
89 * Prevent the compiler from merging or refetching accesses. The compiler
90 * is also forbidden from reordering successive instances of ACCESS_ONCE(),
91 * but only when the compiler is aware of some particular ordering. One way
92 * to make the compiler aware of ordering is to put the two invocations of
93 * ACCESS_ONCE() in different C statements.
94 *
95 * This macro does absolutely -nothing- to prevent the CPU from reordering,
96 * merging, or refetching absolutely anything at any time. Its main intended
97 * use is to mediate communication between process-level code and irq/NMI
98 * handlers, all running on the same CPU.
99 */
100 #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
101
102 /**
103 * rcu_dereference - fetch an RCU-protected pointer in an
104 * RCU read-side critical section. This pointer may later
105 * be safely dereferenced.
106 *
107 * Inserts memory barriers on architectures that require them
108 * (currently only the Alpha), and, more importantly, documents
109 * exactly which pointers are protected by RCU.
110 */
111
112 #define rcu_dereference(p) ({ \
113 typeof(p) _________p1 = ACCESS_ONCE(p); \
114 smp_read_barrier_depends(); \
115 (_________p1); \
116 })
117
118 #define SIGURCU SIGUSR1
119
120 #ifdef DEBUG_YIELD
121 #include <sched.h>
122 #include <time.h>
123 #include <pthread.h>
124
125 #define YIELD_READ (1 << 0)
126 #define YIELD_WRITE (1 << 1)
127
128 extern unsigned int yield_active;
129 extern unsigned int __thread rand_yield;
130
131 static inline void debug_yield_read(void)
132 {
133 if (yield_active & YIELD_READ)
134 if (rand_r(&rand_yield) & 0x1)
135 sched_yield();
136 }
137
138 static inline void debug_yield_write(void)
139 {
140 if (yield_active & YIELD_WRITE)
141 if (rand_r(&rand_yield) & 0x1)
142 sched_yield();
143 }
144
145 static inline void debug_yield_init(void)
146 {
147 rand_yield = time(NULL) ^ pthread_self();
148 }
149 #else
150 static inline void debug_yield_read(void)
151 {
152 }
153
154 static inline void debug_yield_write(void)
155 {
156 }
157
158 static inline void debug_yield_init(void)
159 {
160
161 }
162 #endif
163
164 /*
165 * Limiting the nesting level to 256 to keep instructions small in the read
166 * fast-path.
167 */
168 #define RCU_GP_COUNT (1U << 0)
169 #define RCU_GP_CTR_BIT (1U << 8)
170 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
171
172 /*
173 * Global quiescent period counter with low-order bits unused.
174 * Using a int rather than a char to eliminate false register dependencies
175 * causing stalls on some architectures.
176 */
177 extern int urcu_gp_ctr;
178
179 extern int __thread urcu_active_readers;
180
181 static inline int rcu_old_gp_ongoing(int *value)
182 {
183 int v;
184
185 if (value == NULL)
186 return 0;
187 debug_yield_write();
188 v = ACCESS_ONCE(*value);
189 debug_yield_write();
190 return (v & RCU_GP_CTR_NEST_MASK) &&
191 ((v ^ ACCESS_ONCE(urcu_gp_ctr)) & RCU_GP_CTR_BIT);
192 }
193
194 static inline void rcu_read_lock(void)
195 {
196 int tmp;
197
198 debug_yield_read();
199 tmp = urcu_active_readers;
200 debug_yield_read();
201 if (likely(!(tmp & RCU_GP_CTR_NEST_MASK)))
202 urcu_active_readers = urcu_gp_ctr + RCU_GP_COUNT;
203 else
204 urcu_active_readers = tmp + RCU_GP_COUNT;
205 debug_yield_read();
206 /*
207 * Increment active readers count before accessing the pointer.
208 * See force_mb_all_threads().
209 */
210 barrier();
211 debug_yield_read();
212 }
213
214 static inline void rcu_read_unlock(void)
215 {
216 debug_yield_read();
217 barrier();
218 debug_yield_read();
219 /*
220 * Finish using rcu before decrementing the pointer.
221 * See force_mb_all_threads().
222 */
223 urcu_active_readers -= RCU_GP_COUNT;
224 debug_yield_read();
225 }
226
227 /**
228 * rcu_assign_pointer - assign (publicize) a pointer to a newly
229 * initialized structure that will be dereferenced by RCU read-side
230 * critical sections. Returns the value assigned.
231 *
232 * Inserts memory barriers on architectures that require them
233 * (pretty much all of them other than x86), and also prevents
234 * the compiler from reordering the code that initializes the
235 * structure after the pointer assignment. More importantly, this
236 * call documents which pointers will be dereferenced by RCU read-side
237 * code.
238 */
239
240 #define rcu_assign_pointer(p, v) \
241 ({ \
242 if (!__builtin_constant_p(v) || \
243 ((v) != NULL)) \
244 wmb(); \
245 (p) = (v); \
246 })
247
248 #define rcu_xchg_pointer(p, v) \
249 ({ \
250 if (!__builtin_constant_p(v) || \
251 ((v) != NULL)) \
252 wmb(); \
253 xchg(p, v); \
254 })
255
256 extern void synchronize_rcu(void);
257
258 /*
259 * Exchanges the pointer and waits for quiescent state.
260 * The pointer returned can be freed.
261 */
262 #define urcu_publish_content(p, v) \
263 ({ \
264 void *oldptr; \
265 debug_yield_write(); \
266 oldptr = rcu_xchg_pointer(p, v); \
267 synchronize_rcu(); \
268 oldptr; \
269 })
270
271 /*
272 * Reader thread registration.
273 */
274 extern void urcu_register_thread(void);
275 extern void urcu_unregister_thread(void);
276
277 #endif /* _URCU_H */
This page took 0.034682 seconds and 5 git commands to generate.