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 CODE THAT IS TO BE RECOMPILED ON EACH LIBURCU
  10  * RELEASE. See urcu.h for linking 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 #include <urcu/tls-compat.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 /*
  55  * RCU_MEMBARRIER is only possibly available on Linux. Fallback to RCU_MB
  56  * otherwise.
  57  */
  58 #if !defined(__linux__) && defined(RCU_MEMBARRIER)
  59 #undef RCU_MEMBARRIER
  60 #define RCU_MB
  61 #endif
  62
  63 #ifdef RCU_MEMBARRIER
  64 #include <syscall.h>
  65
  66 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
  67 #ifdef SYS_membarrier
  68 # define MEMBARRIER_EXPEDITED           (1 << 0)
  69 # define MEMBARRIER_DELAYED             (1 << 1)
  70 # define MEMBARRIER_QUERY               (1 << 16)
  71 # define membarrier(...)                syscall(SYS_membarrier, __VA_ARGS__)
  72 #else
  73 # undef RCU_MEMBARRIER
  74 # define RCU_MB
  75 #endif
  76 #endif
  77
  78 /*
  79  * This code section can only be included in LGPL 2.1 compatible source code.
  80  * See below for the function call wrappers which can be used in code meant to
  81  * be only linked with the Userspace RCU library. This comes with a small
  82  * performance degradation on the read-side due to the added function calls.
  83  * This is required to permit relinking with newer versions of the library.
  84  */
  85
  86 /*
  87  * The signal number used by the RCU library can be overridden with
  88  * -DSIGRCU= when compiling the library.
  89  * Provide backward compatibility for liburcu 0.3.x SIGURCU.
  90  */
  91 #ifdef SIGURCU
  92 #define SIGRCU SIGURCU
  93 #endif
  94
  95 #ifndef SIGRCU
  96 #define SIGRCU SIGUSR1
  97 #endif
  98
  99 #ifdef DEBUG_RCU
 100 #define rcu_assert(args...)     assert(args)
 101 #else
 102 #define rcu_assert(args...)
 103 #endif
 104
 105 #ifdef DEBUG_YIELD
 106 #include <sched.h>
 107 #include <time.h>
 108 #include <pthread.h>
 109 #include <unistd.h>
 110
 111 #define RCU_YIELD_READ  (1 << 0)
 112 #define RCU_YIELD_WRITE (1 << 1)
 113
 114 /*
 115  * Updates with RCU_SIGNAL are much slower. Account this in the delay.
 116  */
 117 #ifdef RCU_SIGNAL
 118 /* maximum sleep delay, in us */
 119 #define MAX_SLEEP 30000
 120 #else
 121 #define MAX_SLEEP 50
 122 #endif
 123
 124 extern unsigned int rcu_yield_active;
 125 extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield);
 126
 127 static inline void rcu_debug_yield_read(void)
 128 {
 129         if (rcu_yield_active & RCU_YIELD_READ)
 130                 if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
 131                         usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
 132 }
 133
 134 static inline void rcu_debug_yield_write(void)
 135 {
 136         if (rcu_yield_active & RCU_YIELD_WRITE)
 137                 if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
 138                         usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
 139 }
 140
 141 static inline void rcu_debug_yield_init(void)
 142 {
 143         URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self();
 144 }
 145 #else
 146 static inline void rcu_debug_yield_read(void)
 147 {
 148 }
 149
 150 static inline void rcu_debug_yield_write(void)
 151 {
 152 }
 153
 154 static inline void rcu_debug_yield_init(void)
 155 {
 156
 157 }
 158 #endif
 159
 160 /*
 161  * RCU memory barrier broadcast group. Currently, only broadcast to all process
 162  * threads is supported (group 0).
 163  *
 164  * Slave barriers are only guaranteed to be ordered wrt master barriers.
 165  *
 166  * The pair ordering is detailed as (O: ordered, X: not ordered) :
 167  *               slave  master
 168  *        slave    X      O
 169  *        master   O      O
 170  */
 171
 172 #define MB_GROUP_ALL            0
 173 #define RCU_MB_GROUP            MB_GROUP_ALL
 174
 175 #ifdef RCU_MEMBARRIER
 176 extern int rcu_has_sys_membarrier;
 177
 178 static inline void smp_mb_slave(int group)
 179 {
 180         if (caa_likely(rcu_has_sys_membarrier))
 181                 cmm_barrier();
 182         else
 183                 cmm_smp_mb();
 184 }
 185 #endif
 186
 187 #ifdef RCU_MB
 188 static inline void smp_mb_slave(int group)
 189 {
 190         cmm_smp_mb();
 191 }
 192 #endif
 193
 194 #ifdef RCU_SIGNAL
 195 static inline void smp_mb_slave(int group)
 196 {
 197         cmm_barrier();
 198 }
 199 #endif
 200
 201 /*
 202  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
 203  * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 204  */
 205 #define RCU_GP_COUNT            (1UL << 0)
 206 /* Use the amount of bits equal to half of the architecture long size */
 207 #define RCU_GP_CTR_PHASE        (1UL << (sizeof(unsigned long) << 2))
 208 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 209
 210 /*
 211  * Global quiescent period counter with low-order bits unused.
 212  * Using a int rather than a char to eliminate false register dependencies
 213  * causing stalls on some architectures.
 214  */
 215 extern unsigned long rcu_gp_ctr;
 216
 217 struct rcu_reader {
 218         /* Data used by both reader and synchronize_rcu() */
 219         unsigned long ctr;
 220         char need_mb;
 221         /* Data used for registry */
 222         struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 223         pthread_t tid;
 224 };
 225
 226 extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader);
 227
 228 extern int32_t rcu_gp_futex;
 229
 230 /*
 231  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 232  */
 233 static inline void wake_up_gp(void)
 234 {
 235         if (caa_unlikely(uatomic_read(&rcu_gp_futex) == -1)) {
 236                 uatomic_set(&rcu_gp_futex, 0);
 237                 futex_async(&rcu_gp_futex, FUTEX_WAKE, 1,
 238                       NULL, NULL, 0);
 239         }
 240 }
 241
 242 static inline int rcu_gp_ongoing(unsigned long *ctr)
 243 {
 244         unsigned long v;
 245
 246         /*
 247          * Make sure both tests below are done on the same version of *value
 248          * to insure consistency.
 249          */
 250         v = CMM_LOAD_SHARED(*ctr);
 251         return (v & RCU_GP_CTR_NEST_MASK) &&
 252                  ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
 253 }
 254
 255 /*
 256  * Helper for _rcu_read_lock().  The format of rcu_gp_ctr (as well as
 257  * the per-thread rcu_reader.ctr) has the upper bits containing a count of
 258  * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
 259  * or RCU_GP_CTR_PHASE.  The smp_mb_slave() ensures that the accesses in
 260  * _rcu_read_lock() happen before the subsequent read-side critical section.
 261  */
 262 static inline void _rcu_read_lock_update(unsigned long tmp)
 263 {
 264         if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 265                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
 266                 smp_mb_slave(RCU_MB_GROUP);
 267         } else
 268                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT);
 269 }
 270
 271 /*
 272  * Enter an RCU read-side critical section.
 273  *
 274  * The first cmm_barrier() call ensures that the compiler does not reorder
 275  * the body of _rcu_read_lock() with a mutex.
 276  *
 277  * This function and its helper are both less than 10 lines long.  The
 278  * intent is that this function meets the 10-line criterion in LGPL,
 279  * allowing this function to be invoked directly from non-LGPL code.
 280  */
 281 static inline void _rcu_read_lock(void)
 282 {
 283         unsigned long tmp;
 284
 285         cmm_barrier();
 286         tmp = URCU_TLS(rcu_reader).ctr;
 287         _rcu_read_lock_update(tmp);
 288 }
 289
 290 /*
 291  * This is a helper function for _rcu_read_unlock().
 292  *
 293  * The first smp_mb_slave() call ensures that the critical section is
 294  * seen to precede the store to rcu_reader.ctr.
 295  * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr
 296  * before reading the update-side futex.
 297  */
 298 static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp)
 299 {
 300         if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 301                 smp_mb_slave(RCU_MB_GROUP);
 302                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
 303                 smp_mb_slave(RCU_MB_GROUP);
 304                 wake_up_gp();
 305         } else
 306                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
 307 }
 308
 309 /*
 310  * Exit an RCU read-side crtical section.  Both this function and its
 311  * helper are smaller than 10 lines of code, and are intended to be
 312  * usable by non-LGPL code, as called out in LGPL.
 313  */
 314 static inline void _rcu_read_unlock(void)
 315 {
 316         unsigned long tmp;
 317
 318         tmp = URCU_TLS(rcu_reader).ctr;
 319         _rcu_read_unlock_update_and_wakeup(tmp);
 320         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 321 }
 322
 323 #ifdef __cplusplus
 324 }
 325 #endif
 326
 327 #endif /* _URCU_STATIC_H */