Use xchg in publish content
[urcu.git] / urcu.h
CommitLineData
27b012e2
MD
1#ifndef _URCU_H
2#define _URCU_H
3
b257a10b
MD
4/*
5 * urcu.h
6 *
7 * Userspace RCU header
8 *
9 * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
10 *
5e7e64b9
MD
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 *
b257a10b
MD
17 * Distributed under GPLv2
18 */
19
1430ee0b
MD
20#include <stdlib.h>
21
27b012e2
MD
22/* The "volatile" is due to gcc bugs */
23#define barrier() __asm__ __volatile__("": : :"memory")
24
25/* x86 32/64 specific */
26#define mb() asm volatile("mfence":::"memory")
27#define rmb() asm volatile("lfence":::"memory")
28#define wmb() asm volatile("sfence" ::: "memory")
29
27b012e2
MD
30static inline void atomic_inc(int *v)
31{
32 asm volatile("lock; incl %0"
f69f195a 33 : "+m" (*v));
27b012e2
MD
34}
35
f4a486ac
MD
36#define xchg(ptr, v) \
37 ((__typeof__(*(ptr)))__xchg((unsigned long)(v), (ptr), sizeof(*(ptr))))
38
39struct __xchg_dummy {
40 unsigned long a[100];
41};
42#define __xg(x) ((struct __xchg_dummy *)(x))
43
44/*
45 * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
46 * Note 2: xchg has side effect, so that attribute volatile is necessary,
47 * but generally the primitive is invalid, *ptr is output argument. --ANK
48 */
49static inline unsigned long __xchg(unsigned long x, volatile void *ptr,
50 int size)
51{
52 switch (size) {
53 case 1:
54 asm volatile("xchgb %b0,%1"
55 : "=q" (x)
56 : "m" (*__xg(ptr)), "0" (x)
57 : "memory");
58 break;
59 case 2:
60 asm volatile("xchgw %w0,%1"
61 : "=r" (x)
62 : "m" (*__xg(ptr)), "0" (x)
63 : "memory");
64 break;
65 case 4:
66 asm volatile("xchgl %0,%1"
67 : "=r" (x)
68 : "m" (*__xg(ptr)), "0" (x)
69 : "memory");
70 break;
71 }
72 return x;
73}
74
27b012e2
MD
75/* Nop everywhere except on alpha. */
76#define smp_read_barrier_depends()
77
41718ff9
MD
78/*
79 * Prevent the compiler from merging or refetching accesses. The compiler
80 * is also forbidden from reordering successive instances of ACCESS_ONCE(),
81 * but only when the compiler is aware of some particular ordering. One way
82 * to make the compiler aware of ordering is to put the two invocations of
83 * ACCESS_ONCE() in different C statements.
84 *
85 * This macro does absolutely -nothing- to prevent the CPU from reordering,
86 * merging, or refetching absolutely anything at any time. Its main intended
87 * use is to mediate communication between process-level code and irq/NMI
88 * handlers, all running on the same CPU.
89 */
90#define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
91
92/**
93 * rcu_dereference - fetch an RCU-protected pointer in an
94 * RCU read-side critical section. This pointer may later
95 * be safely dereferenced.
96 *
97 * Inserts memory barriers on architectures that require them
98 * (currently only the Alpha), and, more importantly, documents
99 * exactly which pointers are protected by RCU.
100 */
101
102#define rcu_dereference(p) ({ \
103 typeof(p) _________p1 = ACCESS_ONCE(p); \
104 smp_read_barrier_depends(); \
105 (_________p1); \
106 })
107
27b012e2
MD
108#define SIGURCU SIGUSR1
109
cf380c2f
MD
110#ifdef DEBUG_YIELD
111#include <sched.h>
112
113#define YIELD_READ (1 << 0)
114#define YIELD_WRITE (1 << 1)
115
9d335088
MD
116extern unsigned int yield_active;
117extern unsigned int __thread rand_yield;
cf380c2f
MD
118
119static inline void debug_yield_read(void)
120{
121 if (yield_active & YIELD_READ)
9d335088
MD
122 if (rand_r(&rand_yield) & 0x1)
123 sched_yield();
cf380c2f
MD
124}
125
126static inline void debug_yield_write(void)
127{
128 if (yield_active & YIELD_WRITE)
9d335088
MD
129 if (rand_r(&rand_yield) & 0x1)
130 sched_yield();
131}
132
133static inline void debug_yield_init(void)
134{
135 rand_yield = time(NULL) ^ pthread_self();
cf380c2f
MD
136}
137#else
138static inline void debug_yield_read(void)
139{
140}
141
142static inline void debug_yield_write(void)
143{
9d335088
MD
144}
145
146static inline void debug_yield_init(void)
147{
148
cf380c2f
MD
149}
150#endif
151
1430ee0b
MD
152/*
153 * Limiting the nesting level to 256 to keep instructions small in the read
154 * fast-path.
155 */
156#define RCU_GP_COUNT (1U << 0)
157#define RCU_GP_CTR_BIT (1U << 8)
158#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
159
160/* Global quiescent period counter with low-order bits unused. */
161extern int urcu_gp_ctr;
27b012e2 162
1430ee0b 163extern int __thread urcu_active_readers;
27b012e2 164
1430ee0b 165static inline int rcu_old_gp_ongoing(int *value)
27b012e2 166{
1430ee0b
MD
167 int v;
168
169 if (value == NULL)
170 return 0;
171 debug_yield_write();
172 v = ACCESS_ONCE(*value);
173 debug_yield_write();
174 return (v & RCU_GP_CTR_NEST_MASK) &&
175 ((v ^ ACCESS_ONCE(urcu_gp_ctr)) & RCU_GP_CTR_BIT);
27b012e2
MD
176}
177
1430ee0b 178static inline void rcu_read_lock(void)
27b012e2 179{
1430ee0b
MD
180 int tmp;
181
cf380c2f 182 debug_yield_read();
1430ee0b 183 tmp = urcu_active_readers;
cf380c2f 184 debug_yield_read();
1430ee0b
MD
185 if (!(tmp & RCU_GP_CTR_NEST_MASK))
186 urcu_active_readers = urcu_gp_ctr + RCU_GP_COUNT;
187 else
188 urcu_active_readers = tmp + RCU_GP_COUNT;
cf380c2f 189 debug_yield_read();
27b012e2
MD
190 /*
191 * Increment active readers count before accessing the pointer.
192 * See force_mb_all_threads().
193 */
194 barrier();
cf380c2f 195 debug_yield_read();
27b012e2
MD
196}
197
1430ee0b 198static inline void rcu_read_unlock(void)
27b012e2 199{
cf380c2f 200 debug_yield_read();
27b012e2 201 barrier();
cf380c2f 202 debug_yield_read();
27b012e2
MD
203 /*
204 * Finish using rcu before decrementing the pointer.
205 * See force_mb_all_threads().
206 */
1430ee0b 207 urcu_active_readers -= RCU_GP_COUNT;
cf380c2f 208 debug_yield_read();
27b012e2
MD
209}
210
e462817e
MD
211/**
212 * rcu_assign_pointer - assign (publicize) a pointer to a newly
213 * initialized structure that will be dereferenced by RCU read-side
214 * critical sections. Returns the value assigned.
215 *
216 * Inserts memory barriers on architectures that require them
217 * (pretty much all of them other than x86), and also prevents
218 * the compiler from reordering the code that initializes the
219 * structure after the pointer assignment. More importantly, this
220 * call documents which pointers will be dereferenced by RCU read-side
221 * code.
222 */
223
224#define rcu_assign_pointer(p, v) \
225 ({ \
226 if (!__builtin_constant_p(v) || \
227 ((v) != NULL)) \
228 wmb(); \
229 (p) = (v); \
230 })
231
f4a486ac
MD
232#define rcu_xchg_pointer(p, v) \
233 ({ \
234 if (!__builtin_constant_p(v) || \
235 ((v) != NULL)) \
236 wmb(); \
237 xchg(p, v); \
238 })
239
e462817e 240extern void synchronize_rcu(void);
27b012e2 241
f4a486ac
MD
242/*
243 * Exchanges the pointer and waits for quiescent state.
244 * The pointer returned can be freed.
245 */
246#define urcu_publish_content(p, v) \
247 ({ \
248 void *oldptr; \
249 debug_yield_write(); \
250 oldptr = rcu_xchg_pointer(p, v); \
251 synchronize_rcu(); \
252 oldptr; \
253 })
254
27b012e2
MD
255/*
256 * Reader thread registration.
257 */
258extern void urcu_register_thread(void);
5e7e64b9 259extern void urcu_unregister_thread(void);
27b012e2
MD
260
261#endif /* _URCU_H */
This page took 0.032693 seconds and 4 git commands to generate.