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 <unistd.h>
  35 #include <stdint.h>
  36
  37 #include <urcu/compiler.h>
  38 #include <urcu/arch.h>
  39 #include <urcu/system.h>
  40 #include <urcu/uatomic.h>
  41 #include <urcu/list.h>
  42 #include <urcu/futex.h>
  43
  44 #ifdef __cplusplus
  45 extern "C" {
  46 #endif
  47
  48 /* Default is RCU_MEMBARRIER */
  49 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
  50 #define RCU_MEMBARRIER
  51 #endif
  52
  53 /*
  54  * RCU_MEMBARRIER is only possibly available on Linux. Fallback to RCU_MB
  55  * otherwise.
  56  */
  57 #if !defined(__linux__) && defined(RCU_MEMBARRIER)
  58 #undef RCU_MEMBARRIER
  59 #define RCU_MB
  60 #endif
  61
  62 #ifdef RCU_MEMBARRIER
  63 #include <syscall.h>
  64
  65 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
  66 #ifdef SYS_membarrier
  67 # define MEMBARRIER_EXPEDITED           (1 << 0)
  68 # define MEMBARRIER_DELAYED             (1 << 1)
  69 # define MEMBARRIER_QUERY               (1 << 16)
  70 # define membarrier(...)                syscall(SYS_membarrier, __VA_ARGS__)
  71 #else
  72 # undef RCU_MEMBARRIER
  73 # define RCU_MB
  74 #endif
  75 #endif
  76
  77 /*
  78  * This code section can only be included in LGPL 2.1 compatible source code.
  79  * See below for the function call wrappers which can be used in code meant to
  80  * be only linked with the Userspace RCU library. This comes with a small
  81  * performance degradation on the read-side due to the added function calls.
  82  * This is required to permit relinking with newer versions of the library.
  83  */
  84
  85 /*
  86  * The signal number used by the RCU library can be overridden with
  87  * -DSIGRCU= when compiling the library.
  88  * Provide backward compatibility for liburcu 0.3.x SIGURCU.
  89  */
  90 #ifdef SIGURCU
  91 #define SIGRCU SIGURCU
  92 #endif
  93
  94 #ifndef SIGRCU
  95 #define SIGRCU SIGUSR1
  96 #endif
  97
  98 #ifdef DEBUG_RCU
  99 #define rcu_assert(args...)     assert(args)
 100 #else
 101 #define rcu_assert(args...)
 102 #endif
 103
 104 #ifdef DEBUG_YIELD
 105 #include <sched.h>
 106 #include <time.h>
 107 #include <pthread.h>
 108 #include <unistd.h>
 109
 110 #define YIELD_READ      (1 << 0)
 111 #define YIELD_WRITE     (1 << 1)
 112
 113 /*
 114  * Updates with RCU_SIGNAL are much slower. Account this in the delay.
 115  */
 116 #ifdef RCU_SIGNAL
 117 /* maximum sleep delay, in us */
 118 #define MAX_SLEEP 30000
 119 #else
 120 #define MAX_SLEEP 50
 121 #endif
 122
 123 extern unsigned int yield_active;
 124 extern unsigned int __thread rand_yield;
 125
 126 static inline void debug_yield_read(void)
 127 {
 128         if (yield_active & YIELD_READ)
 129                 if (rand_r(&rand_yield) & 0x1)
 130                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 131 }
 132
 133 static inline void debug_yield_write(void)
 134 {
 135         if (yield_active & YIELD_WRITE)
 136                 if (rand_r(&rand_yield) & 0x1)
 137                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 138 }
 139
 140 static inline void debug_yield_init(void)
 141 {
 142         rand_yield = time(NULL) ^ (unsigned long) pthread_self();
 143 }
 144 #else
 145 static inline void debug_yield_read(void)
 146 {
 147 }
 148
 149 static inline void debug_yield_write(void)
 150 {
 151 }
 152
 153 static inline void debug_yield_init(void)
 154 {
 155
 156 }
 157 #endif
 158
 159 /*
 160  * RCU memory barrier broadcast group. Currently, only broadcast to all process
 161  * threads is supported (group 0).
 162  *
 163  * Slave barriers are only guaranteed to be ordered wrt master barriers.
 164  *
 165  * The pair ordering is detailed as (O: ordered, X: not ordered) :
 166  *               slave  master
 167  *        slave    X      O
 168  *        master   O      O
 169  */
 170
 171 #define MB_GROUP_ALL            0
 172 #define RCU_MB_GROUP            MB_GROUP_ALL
 173
 174 #ifdef RCU_MEMBARRIER
 175 extern int has_sys_membarrier;
 176
 177 static inline void smp_mb_slave(int group)
 178 {
 179         if (caa_likely(has_sys_membarrier))
 180                 cmm_barrier();
 181         else
 182                 cmm_smp_mb();
 183 }
 184 #endif
 185
 186 #ifdef RCU_MB
 187 static inline void smp_mb_slave(int group)
 188 {
 189         cmm_smp_mb();
 190 }
 191 #endif
 192
 193 #ifdef RCU_SIGNAL
 194 static inline void smp_mb_slave(int group)
 195 {
 196         cmm_barrier();
 197 }
 198 #endif
 199
 200 /*
 201  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
 202  * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 203  */
 204 #define RCU_GP_COUNT            (1UL << 0)
 205 /* Use the amount of bits equal to half of the architecture long size */
 206 #define RCU_GP_CTR_PHASE        (1UL << (sizeof(unsigned long) << 2))
 207 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 208
 209 /*
 210  * Global quiescent period counter with low-order bits unused.
 211  * Using a int rather than a char to eliminate false register dependencies
 212  * causing stalls on some architectures.
 213  */
 214 extern unsigned long rcu_gp_ctr;
 215
 216 struct rcu_reader {
 217         /* Data used by both reader and synchronize_rcu() */
 218         unsigned long ctr;
 219         char need_mb;
 220         /* Data used for registry */
 221         struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 222         pthread_t tid;
 223 };
 224
 225 extern struct rcu_reader __thread rcu_reader;
 226
 227 extern int32_t gp_futex;
 228
 229 /*
 230  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 231  */
 232 static inline void wake_up_gp(void)
 233 {
 234         if (caa_unlikely(uatomic_read(&gp_futex) == -1)) {
 235                 uatomic_set(&gp_futex, 0);
 236                 futex_async(&gp_futex, FUTEX_WAKE, 1,
 237                       NULL, NULL, 0);
 238         }
 239 }
 240
 241 static inline int rcu_gp_ongoing(unsigned long *ctr)
 242 {
 243         unsigned long v;
 244
 245         /*
 246          * Make sure both tests below are done on the same version of *value
 247          * to insure consistency.
 248          */
 249         v = CMM_LOAD_SHARED(*ctr);
 250         return (v & RCU_GP_CTR_NEST_MASK) &&
 251                  ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
 252 }
 253
 254 static inline void _rcu_read_lock(void)
 255 {
 256         unsigned long tmp;
 257
 258         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 259         tmp = rcu_reader.ctr;
 260         /*
 261          * rcu_gp_ctr is
 262          *   RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE)
 263          */
 264         if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 265                 _CMM_STORE_SHARED(rcu_reader.ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
 266                 /*
 267                  * Set active readers count for outermost nesting level before
 268                  * accessing the pointer. See smp_mb_master().
 269                  */
 270                 smp_mb_slave(RCU_MB_GROUP);
 271         } else {
 272                 _CMM_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT);
 273         }
 274 }
 275
 276 static inline void _rcu_read_unlock(void)
 277 {
 278         unsigned long tmp;
 279
 280         tmp = rcu_reader.ctr;
 281         /*
 282          * Finish using rcu before decrementing the pointer.
 283          * See smp_mb_master().
 284          */
 285         if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 286                 smp_mb_slave(RCU_MB_GROUP);
 287                 _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT);
 288                 /* write rcu_reader.ctr before read futex */
 289                 smp_mb_slave(RCU_MB_GROUP);
 290                 wake_up_gp();
 291         } else {
 292                 _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT);
 293         }
 294         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 295 }
 296
 297 #ifdef __cplusplus
 298 }
 299 #endif
 300
 301 #endif /* _URCU_STATIC_H */