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 enum rcu_state {
 100         RCU_READER_ACTIVE_CURRENT,
 101         RCU_READER_ACTIVE_OLD,
 102         RCU_READER_INACTIVE,
 103 };
 104
 105 #ifdef DEBUG_RCU
 106 #define rcu_assert(args...)     assert(args)
 107 #else
 108 #define rcu_assert(args...)
 109 #endif
 110
 111 #ifdef DEBUG_YIELD
 112 #include <sched.h>
 113 #include <time.h>
 114 #include <pthread.h>
 115 #include <unistd.h>
 116
 117 #define RCU_YIELD_READ  (1 << 0)
 118 #define RCU_YIELD_WRITE (1 << 1)
 119
 120 /*
 121  * Updates with RCU_SIGNAL are much slower. Account this in the delay.
 122  */
 123 #ifdef RCU_SIGNAL
 124 /* maximum sleep delay, in us */
 125 #define MAX_SLEEP 30000
 126 #else
 127 #define MAX_SLEEP 50
 128 #endif
 129
 130 extern unsigned int rcu_yield_active;
 131 extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield);
 132
 133 static inline void rcu_debug_yield_read(void)
 134 {
 135         if (rcu_yield_active & RCU_YIELD_READ)
 136                 if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
 137                         usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
 138 }
 139
 140 static inline void rcu_debug_yield_write(void)
 141 {
 142         if (rcu_yield_active & RCU_YIELD_WRITE)
 143                 if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
 144                         usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
 145 }
 146
 147 static inline void rcu_debug_yield_init(void)
 148 {
 149         URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self();
 150 }
 151 #else
 152 static inline void rcu_debug_yield_read(void)
 153 {
 154 }
 155
 156 static inline void rcu_debug_yield_write(void)
 157 {
 158 }
 159
 160 static inline void rcu_debug_yield_init(void)
 161 {
 162
 163 }
 164 #endif
 165
 166 /*
 167  * RCU memory barrier broadcast group. Currently, only broadcast to all process
 168  * threads is supported (group 0).
 169  *
 170  * Slave barriers are only guaranteed to be ordered wrt master barriers.
 171  *
 172  * The pair ordering is detailed as (O: ordered, X: not ordered) :
 173  *               slave  master
 174  *        slave    X      O
 175  *        master   O      O
 176  */
 177
 178 #define MB_GROUP_ALL            0
 179 #define RCU_MB_GROUP            MB_GROUP_ALL
 180
 181 #ifdef RCU_MEMBARRIER
 182 extern int rcu_has_sys_membarrier;
 183
 184 static inline void smp_mb_slave(int group)
 185 {
 186         if (caa_likely(rcu_has_sys_membarrier))
 187                 cmm_barrier();
 188         else
 189                 cmm_smp_mb();
 190 }
 191 #endif
 192
 193 #ifdef RCU_MB
 194 static inline void smp_mb_slave(int group)
 195 {
 196         cmm_smp_mb();
 197 }
 198 #endif
 199
 200 #ifdef RCU_SIGNAL
 201 static inline void smp_mb_slave(int group)
 202 {
 203         cmm_barrier();
 204 }
 205 #endif
 206
 207 /*
 208  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
 209  * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 210  */
 211 #define RCU_GP_COUNT            (1UL << 0)
 212 /* Use the amount of bits equal to half of the architecture long size */
 213 #define RCU_GP_CTR_PHASE        (1UL << (sizeof(unsigned long) << 2))
 214 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 215
 216 struct rcu_gp {
 217         /*
 218          * Global grace period counter.
 219          * Contains the current RCU_GP_CTR_PHASE.
 220          * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
 221          * Written to only by writer with mutex taken.
 222          * Read by both writer and readers.
 223          */
 224         unsigned long ctr;
 225
 226         int32_t futex;
 227 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 228
 229 extern struct rcu_gp rcu_gp;
 230
 231 struct rcu_reader {
 232         /* Data used by both reader and synchronize_rcu() */
 233         unsigned long ctr;
 234         char need_mb;
 235         /* Data used for registry */
 236         struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 237         pthread_t tid;
 238 };
 239
 240 extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader);
 241
 242 /*
 243  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 244  */
 245 static inline void wake_up_gp(void)
 246 {
 247         if (caa_unlikely(uatomic_read(&rcu_gp.futex) == -1)) {
 248                 uatomic_set(&rcu_gp.futex, 0);
 249                 futex_async(&rcu_gp.futex, FUTEX_WAKE, 1,
 250                       NULL, NULL, 0);
 251         }
 252 }
 253
 254 static inline enum rcu_state rcu_reader_state(unsigned long *ctr)
 255 {
 256         unsigned long v;
 257
 258         /*
 259          * Make sure both tests below are done on the same version of *value
 260          * to insure consistency.
 261          */
 262         v = CMM_LOAD_SHARED(*ctr);
 263         if (!(v & RCU_GP_CTR_NEST_MASK))
 264                 return RCU_READER_INACTIVE;
 265         if (!((v ^ rcu_gp.ctr) & RCU_GP_CTR_PHASE))
 266                 return RCU_READER_ACTIVE_CURRENT;
 267         return RCU_READER_ACTIVE_OLD;
 268 }
 269
 270 /*
 271  * Helper for _rcu_read_lock().  The format of rcu_gp.ctr (as well as
 272  * the per-thread rcu_reader.ctr) has the upper bits containing a count of
 273  * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
 274  * or RCU_GP_CTR_PHASE.  The smp_mb_slave() ensures that the accesses in
 275  * _rcu_read_lock() happen before the subsequent read-side critical section.
 276  */
 277 static inline void _rcu_read_lock_update(unsigned long tmp)
 278 {
 279         if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 280                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp.ctr));
 281                 smp_mb_slave(RCU_MB_GROUP);
 282         } else
 283                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT);
 284 }
 285
 286 /*
 287  * Enter an RCU read-side critical section.
 288  *
 289  * The first cmm_barrier() call ensures that the compiler does not reorder
 290  * the body of _rcu_read_lock() with a mutex.
 291  *
 292  * This function and its helper are both less than 10 lines long.  The
 293  * intent is that this function meets the 10-line criterion in LGPL,
 294  * allowing this function to be invoked directly from non-LGPL code.
 295  */
 296 static inline void _rcu_read_lock(void)
 297 {
 298         unsigned long tmp;
 299
 300         cmm_barrier();
 301         tmp = URCU_TLS(rcu_reader).ctr;
 302         _rcu_read_lock_update(tmp);
 303 }
 304
 305 /*
 306  * This is a helper function for _rcu_read_unlock().
 307  *
 308  * The first smp_mb_slave() call ensures that the critical section is
 309  * seen to precede the store to rcu_reader.ctr.
 310  * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr
 311  * before reading the update-side futex.
 312  */
 313 static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp)
 314 {
 315         if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 316                 smp_mb_slave(RCU_MB_GROUP);
 317                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
 318                 smp_mb_slave(RCU_MB_GROUP);
 319                 wake_up_gp();
 320         } else
 321                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
 322 }
 323
 324 /*
 325  * Exit an RCU read-side crtical section.  Both this function and its
 326  * helper are smaller than 10 lines of code, and are intended to be
 327  * usable by non-LGPL code, as called out in LGPL.
 328  */
 329 static inline void _rcu_read_unlock(void)
 330 {
 331         unsigned long tmp;
 332
 333         tmp = URCU_TLS(rcu_reader).ctr;
 334         _rcu_read_unlock_update_and_wakeup(tmp);
 335         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 336 }
 337
 338 /*
 339  * Returns whether within a RCU read-side critical section.
 340  *
 341  * This function is less than 10 lines long.  The intent is that this
 342  * function meets the 10-line criterion for LGPL, allowing this function
 343  * to be invoked directly from non-LGPL code.
 344  */
 345 static inline int _rcu_read_ongoing(void)
 346 {
 347         return URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK;
 348 }
 349
 350 #ifdef __cplusplus
 351 }
 352 #endif
 353
 354 #endif /* _URCU_STATIC_H */