urcu/static/urcu.h

   1 #ifndef _URCU_STATIC_H
   2 #define _URCU_STATIC_H
   3
   4 /*
   5  * urcu-static.h
   6  *
   7  * Userspace RCU header.
   8  *
   9  * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking
  10  * dynamically with the userspace rcu library.
  11  *
  12  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  13  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
  14  *
  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
  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 #include <syscall.h>
  35 #include <unistd.h>
  36 #include <stdint.h>
  37
  38 #include <urcu/compiler.h>
  39 #include <urcu/arch.h>
  40 #include <urcu/system.h>
  41 #include <urcu/uatomic.h>
  42 #include <urcu/list.h>
  43 #include <urcu/futex.h>
  44
  45 #ifdef __cplusplus
  46 extern "C" {
  47 #endif
  48
  49 /* Default is RCU_MEMBARRIER */
  50 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
  51 #define RCU_MEMBARRIER
  52 #endif
  53
  54 #ifdef RCU_MEMBARRIER
  55 #include <unistd.h>
  56 #include <sys/syscall.h>
  57
  58 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
  59 #ifdef SYS_membarrier
  60 # define MEMBARRIER_EXPEDITED           (1 << 0)
  61 # define MEMBARRIER_DELAYED             (1 << 1)
  62 # define MEMBARRIER_QUERY               (1 << 16)
  63 # define membarrier(...)                syscall(SYS_membarrier, __VA_ARGS__)
  64 #else
  65 # undef RCU_MEMBARRIER
  66 # define RCU_MB
  67 #endif
  68 #endif
  69
  70 /*
  71  * This code section can only be included in LGPL 2.1 compatible source code.
  72  * See below for the function call wrappers which can be used in code meant to
  73  * be only linked with the Userspace RCU library. This comes with a small
  74  * performance degradation on the read-side due to the added function calls.
  75  * This is required to permit relinking with newer versions of the library.
  76  */
  77
  78 /*
  79  * The signal number used by the RCU library can be overridden with
  80  * -DSIGRCU= when compiling the library.
  81  * Provide backward compatibility for liburcu 0.3.x SIGURCU.
  82  */
  83 #ifdef SIGURCU
  84 #define SIGRCU SIGURCU
  85 #endif
  86
  87 #ifndef SIGRCU
  88 #define SIGRCU SIGUSR1
  89 #endif
  90
  91 /*
  92  * If a reader is really non-cooperative and refuses to commit its
  93  * rcu_active_readers count to memory (there is no barrier in the reader
  94  * per-se), kick it after a few loops waiting for it.
  95  */
  96 #define KICK_READER_LOOPS 10000
  97
  98 /*
  99  * Active attempts to check for reader Q.S. before calling futex().
 100  */
 101 #define RCU_QS_ACTIVE_ATTEMPTS 100
 102
 103 #ifdef DEBUG_RCU
 104 #define rcu_assert(args...)     assert(args)
 105 #else
 106 #define rcu_assert(args...)
 107 #endif
 108
 109 #ifdef DEBUG_YIELD
 110 #include <sched.h>
 111 #include <time.h>
 112 #include <pthread.h>
 113 #include <unistd.h>
 114
 115 #define YIELD_READ      (1 << 0)
 116 #define YIELD_WRITE     (1 << 1)
 117
 118 /*
 119  * Updates with RCU_SIGNAL are much slower. Account this in the delay.
 120  */
 121 #ifdef RCU_SIGNAL
 122 /* maximum sleep delay, in us */
 123 #define MAX_SLEEP 30000
 124 #else
 125 #define MAX_SLEEP 50
 126 #endif
 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                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 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                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 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  * RCU memory barrier broadcast group. Currently, only broadcast to all process
 166  * threads is supported (group 0).
 167  *
 168  * Slave barriers are only guaranteed to be ordered wrt master barriers.
 169  *
 170  * The pair ordering is detailed as (O: ordered, X: not ordered) :
 171  *               slave  master
 172  *        slave    X      O
 173  *        master   O      O
 174  */
 175
 176 #define MB_GROUP_ALL            0
 177 #define RCU_MB_GROUP            MB_GROUP_ALL
 178
 179 #ifdef RCU_MEMBARRIER
 180 extern int has_sys_membarrier;
 181
 182 static inline void smp_mb_slave(int group)
 183 {
 184         if (likely(has_sys_membarrier))
 185                 cmm_barrier();
 186         else
 187                 cmm_smp_mb();
 188 }
 189 #endif
 190
 191 #ifdef RCU_MB
 192 static inline void smp_mb_slave(int group)
 193 {
 194         cmm_smp_mb();
 195 }
 196 #endif
 197
 198 #ifdef RCU_SIGNAL
 199 static inline void smp_mb_slave(int group)
 200 {
 201         cmm_barrier();
 202 }
 203 #endif
 204
 205 /*
 206  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
 207  * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 208  */
 209 #define RCU_GP_COUNT            (1UL << 0)
 210 /* Use the amount of bits equal to half of the architecture long size */
 211 #define RCU_GP_CTR_PHASE        (1UL << (sizeof(unsigned long) << 2))
 212 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 213
 214 /*
 215  * Global quiescent period counter with low-order bits unused.
 216  * Using a int rather than a char to eliminate false register dependencies
 217  * causing stalls on some architectures.
 218  */
 219 extern unsigned long rcu_gp_ctr;
 220
 221 struct rcu_reader {
 222         /* Data used by both reader and synchronize_rcu() */
 223         unsigned long ctr;
 224         char need_mb;
 225         /* Data used for registry */
 226         struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 227         pthread_t tid;
 228 };
 229
 230 extern struct rcu_reader __thread rcu_reader;
 231
 232 extern int32_t gp_futex;
 233
 234 /*
 235  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 236  */
 237 static inline void wake_up_gp(void)
 238 {
 239         if (unlikely(uatomic_read(&gp_futex) == -1)) {
 240                 uatomic_set(&gp_futex, 0);
 241                 futex_async(&gp_futex, FUTEX_WAKE, 1,
 242                       NULL, NULL, 0);
 243         }
 244 }
 245
 246 static inline int rcu_gp_ongoing(unsigned long *ctr)
 247 {
 248         unsigned long v;
 249
 250         /*
 251          * Make sure both tests below are done on the same version of *value
 252          * to insure consistency.
 253          */
 254         v = CMM_LOAD_SHARED(*ctr);
 255         return (v & RCU_GP_CTR_NEST_MASK) &&
 256                  ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
 257 }
 258
 259 static inline void _rcu_read_lock(void)
 260 {
 261         unsigned long tmp;
 262
 263         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 264         tmp = rcu_reader.ctr;
 265         /*
 266          * rcu_gp_ctr is
 267          *   RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE)
 268          */
 269         if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 270                 _CMM_STORE_SHARED(rcu_reader.ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
 271                 /*
 272                  * Set active readers count for outermost nesting level before
 273                  * accessing the pointer. See smp_mb_master().
 274                  */
 275                 smp_mb_slave(RCU_MB_GROUP);
 276         } else {
 277                 _CMM_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT);
 278         }
 279 }
 280
 281 static inline void _rcu_read_unlock(void)
 282 {
 283         unsigned long tmp;
 284
 285         tmp = rcu_reader.ctr;
 286         /*
 287          * Finish using rcu before decrementing the pointer.
 288          * See smp_mb_master().
 289          */
 290         if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 291                 smp_mb_slave(RCU_MB_GROUP);
 292                 _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT);
 293                 /* write rcu_reader.ctr before read futex */
 294                 smp_mb_slave(RCU_MB_GROUP);
 295                 wake_up_gp();
 296         } else {
 297                 _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT);
 298         }
 299         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 300 }
 301
 302 #ifdef __cplusplus
 303 }
 304 #endif
 305
 306 #endif /* _URCU_STATIC_H */