Document access ordering of rcu_dereference wrt VSS
[urcu.git] / urcu-static.h
CommitLineData
adcfce54
MD
1#ifndef _URCU_STATIC_H
2#define _URCU_STATIC_H
3
4/*
5 * urcu-static.h
6 *
d2d23035 7 * Userspace RCU header.
adcfce54 8 *
d2d23035
MD
9 * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking
10 * dynamically with the userspace rcu library.
adcfce54 11 *
d2d23035
MD
12 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
13 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
adcfce54 14 *
d2d23035
MD
15 * This library is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU Lesser General Public
17 * License as published by the Free Software Foundation; either
18 * version 2.1 of the License, or (at your option) any later version.
19 *
20 * This library is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * Lesser General Public License for more details.
24 *
25 * You should have received a copy of the GNU Lesser General Public
26 * License along with this library; if not, write to the Free Software
27 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
adcfce54
MD
28 *
29 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
30 */
31
32#include <stdlib.h>
33#include <pthread.h>
34
35#include <compiler.h>
36#include <arch.h>
37
38/*
39 * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
40 */
41#define _LOAD_SHARED(p) ACCESS_ONCE(p)
42
43/*
44 * Load a data from shared memory, doing a cache flush if required.
45 */
46#define LOAD_SHARED(p) \
47 ({ \
48 smp_rmc(); \
49 _LOAD_SHARED(p); \
50 })
51
52/*
53 * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
54 */
55#define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); })
56
57/*
58 * Store v into x, where x is located in shared memory. Performs the required
59 * cache flush after writing. Returns v.
60 */
61#define STORE_SHARED(x, v) \
62 ({ \
63 _STORE_SHARED(x, v); \
64 smp_wmc(); \
65 (v); \
66 })
67
68/**
69 * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
70 * into a RCU read-side critical section. The pointer can later be safely
71 * dereferenced within the critical section.
72 *
73 * This ensures that the pointer copy is invariant thorough the whole critical
74 * section.
75 *
76 * Inserts memory barriers on architectures that require them (currently only
77 * Alpha) and documents which pointers are protected by RCU.
78 *
809f4fde
MD
79 * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
80 * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
81 * data read before the pointer read by speculating the value of the pointer.
82 * Correct ordering is ensured because the pointer is read as a volatile access.
83 * This acts as a global side-effect operation, which forbids reordering of
84 * dependent memory operations.
85 *
adcfce54
MD
86 * Should match rcu_assign_pointer() or rcu_xchg_pointer().
87 */
88
89#define _rcu_dereference(p) ({ \
90 typeof(p) _________p1 = LOAD_SHARED(p); \
91 smp_read_barrier_depends(); \
92 (_________p1); \
93 })
94
95/*
96 * This code section can only be included in LGPL 2.1 compatible source code.
97 * See below for the function call wrappers which can be used in code meant to
98 * be only linked with the Userspace RCU library. This comes with a small
99 * performance degradation on the read-side due to the added function calls.
100 * This is required to permit relinking with newer versions of the library.
101 */
102
103/*
104 * The signal number used by the RCU library can be overridden with
105 * -DSIGURCU= when compiling the library.
106 */
107#ifndef SIGURCU
108#define SIGURCU SIGUSR1
109#endif
110
111/*
112 * If a reader is really non-cooperative and refuses to commit its
113 * urcu_active_readers count to memory (there is no barrier in the reader
114 * per-se), kick it after a few loops waiting for it.
115 */
116#define KICK_READER_LOOPS 10000
117
118#ifdef DEBUG_YIELD
119#include <sched.h>
120#include <time.h>
121#include <pthread.h>
122#include <unistd.h>
123
124#define YIELD_READ (1 << 0)
125#define YIELD_WRITE (1 << 1)
126
b4ce1526
MD
127/*
128 * Updates without CONFIG_URCU_AVOID_SIGNALS are much slower. Account this in
129 * the delay.
130 */
131#ifdef CONFIG_URCU_AVOID_SIGNALS
adcfce54
MD
132/* maximum sleep delay, in us */
133#define MAX_SLEEP 50
134#else
135#define MAX_SLEEP 30000
136#endif
137
138extern unsigned int yield_active;
139extern unsigned int __thread rand_yield;
140
141static inline void debug_yield_read(void)
142{
143 if (yield_active & YIELD_READ)
144 if (rand_r(&rand_yield) & 0x1)
145 usleep(rand_r(&rand_yield) % MAX_SLEEP);
146}
147
148static inline void debug_yield_write(void)
149{
150 if (yield_active & YIELD_WRITE)
151 if (rand_r(&rand_yield) & 0x1)
152 usleep(rand_r(&rand_yield) % MAX_SLEEP);
153}
154
155static inline void debug_yield_init(void)
156{
157 rand_yield = time(NULL) ^ pthread_self();
158}
159#else
160static inline void debug_yield_read(void)
161{
162}
163
164static inline void debug_yield_write(void)
165{
166}
167
168static inline void debug_yield_init(void)
169{
170
171}
172#endif
173
b4ce1526 174#ifdef CONFIG_URCU_AVOID_SIGNALS
adcfce54
MD
175static inline void reader_barrier()
176{
177 smp_mb();
178}
179#else
180static inline void reader_barrier()
181{
182 barrier();
183}
184#endif
185
186/*
187 * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
188 * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
189 */
190#define RCU_GP_COUNT (1UL << 0)
191/* Use the amount of bits equal to half of the architecture long size */
192#define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2))
193#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
194
195/*
196 * Global quiescent period counter with low-order bits unused.
197 * Using a int rather than a char to eliminate false register dependencies
198 * causing stalls on some architectures.
199 */
200extern long urcu_gp_ctr;
201
202extern long __thread urcu_active_readers;
203
204static inline int rcu_old_gp_ongoing(long *value)
205{
206 long v;
207
208 if (value == NULL)
209 return 0;
210 /*
211 * Make sure both tests below are done on the same version of *value
212 * to insure consistency.
213 */
214 v = LOAD_SHARED(*value);
215 return (v & RCU_GP_CTR_NEST_MASK) &&
216 ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
217}
218
219static inline void _rcu_read_lock(void)
220{
221 long tmp;
222
223 tmp = urcu_active_readers;
224 /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
67ef1a2c 225 if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
adcfce54 226 _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr));
67ef1a2c
MD
227 /*
228 * Set active readers count for outermost nesting level before
229 * accessing the pointer. See force_mb_all_threads().
230 */
231 reader_barrier();
232 } else {
adcfce54 233 _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT);
67ef1a2c 234 }
adcfce54
MD
235}
236
237static inline void _rcu_read_unlock(void)
238{
239 reader_barrier();
240 /*
241 * Finish using rcu before decrementing the pointer.
242 * See force_mb_all_threads().
67ef1a2c
MD
243 * Formally only needed for outermost nesting level, but leave barrier
244 * in place for nested unlocks to remove a branch from the common case
245 * (no nesting).
adcfce54
MD
246 */
247 _STORE_SHARED(urcu_active_readers, urcu_active_readers - RCU_GP_COUNT);
248}
249
250/**
251 * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
252 * meant to be read by RCU read-side critical sections. Returns the assigned
253 * value.
254 *
255 * Documents which pointers will be dereferenced by RCU read-side critical
256 * sections and adds the required memory barriers on architectures requiring
257 * them. It also makes sure the compiler does not reorder code initializing the
258 * data structure before its publication.
259 *
260 * Should match rcu_dereference_pointer().
261 */
262
263#define _rcu_assign_pointer(p, v) \
264 ({ \
265 if (!__builtin_constant_p(v) || \
266 ((v) != NULL)) \
267 wmb(); \
268 STORE_SHARED(p, v); \
269 })
270
271/**
272 * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
67ef1a2c 273 * pointer to the data structure, which can be safely freed after waiting for a
adcfce54
MD
274 * quiescent state using synchronize_rcu().
275 */
276
277#define _rcu_xchg_pointer(p, v) \
278 ({ \
279 if (!__builtin_constant_p(v) || \
280 ((v) != NULL)) \
281 wmb(); \
282 xchg(p, v); \
283 })
284
285/*
286 * Exchanges the pointer and waits for quiescent state.
287 * The pointer returned can be freed.
288 */
289#define _rcu_publish_content(p, v) \
290 ({ \
291 void *oldptr; \
292 oldptr = _rcu_xchg_pointer(p, v); \
293 synchronize_rcu(); \
294 oldptr; \
295 })
296
297#endif /* _URCU_STATIC_H */
This page took 0.034363 seconds and 4 git commands to generate.