urcu-static.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
  37 #include <urcu/compiler.h>
  38 #include <urcu/arch.h>
  39 #include <urcu/system.h>
  40 #include <urcu/uatomic_arch.h>
  41 #include <urcu/list.h>
  42 #include <urcu/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 #ifdef RCU_MEMBARRIER
  54 #include <unistd.h>
  55 #include <sys/syscall.h>
  56
  57 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
  58 #ifdef SYS_membarrier
  59 # define MEMBARRIER_EXPEDITED           (1 << 0)
  60 # define MEMBARRIER_DELAYED             (1 << 1)
  61 # define MEMBARRIER_QUERY               (1 << 16)
  62 # define membarrier(...)                syscall(SYS_membarrier, __VA_ARGS__)
  63 #else
  64 # undef RCU_MEMBARRIER
  65 # define RCU_MB
  66 #endif
  67 #endif
  68
  69 /*
  70  * This code section can only be included in LGPL 2.1 compatible source code.
  71  * See below for the function call wrappers which can be used in code meant to
  72  * be only linked with the Userspace RCU library. This comes with a small
  73  * performance degradation on the read-side due to the added function calls.
  74  * This is required to permit relinking with newer versions of the library.
  75  */
  76
  77 /*
  78  * The signal number used by the RCU library can be overridden with
  79  * -DSIGRCU= when compiling the library.
  80  * Provide backward compatibility for liburcu 0.3.x SIGURCU.
  81  */
  82 #ifdef SIGURCU
  83 #define SIGRCU SIGURCU
  84 #endif
  85
  86 #ifndef SIGRCU
  87 #define SIGRCU SIGUSR1
  88 #endif
  89
  90 /*
  91  * If a reader is really non-cooperative and refuses to commit its
  92  * rcu_active_readers count to memory (there is no barrier in the reader
  93  * per-se), kick it after a few loops waiting for it.
  94  */
  95 #define KICK_READER_LOOPS 10000
  96
  97 /*
  98  * Active attempts to check for reader Q.S. before calling futex().
  99  */
 100 #define RCU_QS_ACTIVE_ATTEMPTS 100
 101
 102 #ifdef DEBUG_RCU
 103 #define rcu_assert(args...)     assert(args)
 104 #else
 105 #define rcu_assert(args...)
 106 #endif
 107
 108 #ifdef DEBUG_YIELD
 109 #include <sched.h>
 110 #include <time.h>
 111 #include <pthread.h>
 112 #include <unistd.h>
 113
 114 #define YIELD_READ      (1 << 0)
 115 #define YIELD_WRITE     (1 << 1)
 116
 117 /*
 118  * Updates with RCU_SIGNAL are much slower. Account this in the delay.
 119  */
 120 #ifdef RCU_SIGNAL
 121 /* maximum sleep delay, in us */
 122 #define MAX_SLEEP 30000
 123 #else
 124 #define MAX_SLEEP 50
 125 #endif
 126
 127 extern unsigned int yield_active;
 128 extern unsigned int __thread rand_yield;
 129
 130 static inline void debug_yield_read(void)
 131 {
 132         if (yield_active & YIELD_READ)
 133                 if (rand_r(&rand_yield) & 0x1)
 134                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 135 }
 136
 137 static inline void debug_yield_write(void)
 138 {
 139         if (yield_active & YIELD_WRITE)
 140                 if (rand_r(&rand_yield) & 0x1)
 141                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 142 }
 143
 144 static inline void debug_yield_init(void)
 145 {
 146         rand_yield = time(NULL) ^ pthread_self();
 147 }
 148 #else
 149 static inline void debug_yield_read(void)
 150 {
 151 }
 152
 153 static inline void debug_yield_write(void)
 154 {
 155 }
 156
 157 static inline void debug_yield_init(void)
 158 {
 159
 160 }
 161 #endif
 162
 163 /*
 164  * RCU memory barrier broadcast group. Currently, only broadcast to all process
 165  * threads is supported (group 0).
 166  *
 167  * Slave barriers are only guaranteed to be ordered wrt master barriers.
 168  *
 169  * The pair ordering is detailed as (O: ordered, X: not ordered) :
 170  *               slave  master
 171  *        slave    X      O
 172  *        master   O      O
 173  */
 174
 175 #define MB_GROUP_ALL            0
 176 #define RCU_MB_GROUP            MB_GROUP_ALL
 177
 178 #ifdef RCU_MEMBARRIER
 179 extern int has_sys_membarrier;
 180
 181 static inline void smp_mb_slave(int group)
 182 {
 183         if (likely(has_sys_membarrier))
 184                 cmm_barrier();
 185         else
 186                 cmm_smp_mb();
 187 }
 188 #endif
 189
 190 #ifdef RCU_MB
 191 static inline void smp_mb_slave(int group)
 192 {
 193         cmm_smp_mb();
 194 }
 195 #endif
 196
 197 #ifdef RCU_SIGNAL
 198 static inline void smp_mb_slave(int group)
 199 {
 200         cmm_barrier();
 201 }
 202 #endif
 203
 204 /*
 205  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
 206  * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 207  */
 208 #define RCU_GP_COUNT            (1UL << 0)
 209 /* Use the amount of bits equal to half of the architecture long size */
 210 #define RCU_GP_CTR_PHASE        (1UL << (sizeof(unsigned long) << 2))
 211 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 212
 213 /*
 214  * Global quiescent period counter with low-order bits unused.
 215  * Using a int rather than a char to eliminate false register dependencies
 216  * causing stalls on some architectures.
 217  */
 218 extern unsigned long rcu_gp_ctr;
 219
 220 struct rcu_reader {
 221         /* Data used by both reader and synchronize_rcu() */
 222         unsigned long ctr;
 223         char need_mb;
 224         /* Data used for registry */
 225         struct list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 226         pthread_t tid;
 227 };
 228
 229 extern struct rcu_reader __thread rcu_reader;
 230
 231 extern int gp_futex;
 232
 233 /*
 234  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 235  */
 236 static inline void wake_up_gp(void)
 237 {
 238         if (unlikely(uatomic_read(&gp_futex) == -1)) {
 239                 uatomic_set(&gp_futex, 0);
 240                 futex_async(&gp_futex, FUTEX_WAKE, 1,
 241                       NULL, NULL, 0);
 242         }
 243 }
 244
 245 static inline int rcu_gp_ongoing(unsigned long *ctr)
 246 {
 247         unsigned long v;
 248
 249         /*
 250          * Make sure both tests below are done on the same version of *value
 251          * to insure consistency.
 252          */
 253         v = CAA_LOAD_SHARED(*ctr);
 254         return (v & RCU_GP_CTR_NEST_MASK) &&
 255                  ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
 256 }
 257
 258 static inline void _rcu_read_lock(void)
 259 {
 260         unsigned long tmp;
 261
 262         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 263         tmp = rcu_reader.ctr;
 264         /*
 265          * rcu_gp_ctr is
 266          *   RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE)
 267          */
 268         if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 269                 _CAA_STORE_SHARED(rcu_reader.ctr, _CAA_LOAD_SHARED(rcu_gp_ctr));
 270                 /*
 271                  * Set active readers count for outermost nesting level before
 272                  * accessing the pointer. See smp_mb_master().
 273                  */
 274                 smp_mb_slave(RCU_MB_GROUP);
 275         } else {
 276                 _CAA_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT);
 277         }
 278 }
 279
 280 static inline void _rcu_read_unlock(void)
 281 {
 282         unsigned long tmp;
 283
 284         tmp = rcu_reader.ctr;
 285         /*
 286          * Finish using rcu before decrementing the pointer.
 287          * See smp_mb_master().
 288          */
 289         if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 290                 smp_mb_slave(RCU_MB_GROUP);
 291                 _CAA_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT);
 292                 /* write rcu_reader.ctr before read futex */
 293                 smp_mb_slave(RCU_MB_GROUP);
 294                 wake_up_gp();
 295         } else {
 296                 _CAA_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT);
 297         }
 298         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 299 }
 300
 301 #ifdef __cplusplus
 302 }
 303 #endif
 304
 305 #endif /* _URCU_STATIC_H */