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  * This code section can only be included in LGPL 2.1 compatible source code.
  56  * See below for the function call wrappers which can be used in code meant to
  57  * be only linked with the Userspace RCU library. This comes with a small
  58  * performance degradation on the read-side due to the added function calls.
  59  * This is required to permit relinking with newer versions of the library.
  60  */
  61
  62 /*
  63  * The signal number used by the RCU library can be overridden with
  64  * -DSIGRCU= when compiling the library.
  65  * Provide backward compatibility for liburcu 0.3.x SIGURCU.
  66  */
  67 #ifdef SIGURCU
  68 #define SIGRCU SIGURCU
  69 #endif
  70
  71 #ifndef SIGRCU
  72 #define SIGRCU SIGUSR1
  73 #endif
  74
  75 enum rcu_state {
  76         RCU_READER_ACTIVE_CURRENT,
  77         RCU_READER_ACTIVE_OLD,
  78         RCU_READER_INACTIVE,
  79 };
  80
  81 #ifdef DEBUG_RCU
  82 #define rcu_assert(args...)     assert(args)
  83 #else
  84 #define rcu_assert(args...)
  85 #endif
  86
  87 /*
  88  * RCU memory barrier broadcast group. Currently, only broadcast to all process
  89  * threads is supported (group 0).
  90  *
  91  * Slave barriers are only guaranteed to be ordered wrt master barriers.
  92  *
  93  * The pair ordering is detailed as (O: ordered, X: not ordered) :
  94  *               slave  master
  95  *        slave    X      O
  96  *        master   O      O
  97  */
  98
  99 #define MB_GROUP_ALL            0
 100 #define RCU_MB_GROUP            MB_GROUP_ALL
 101
 102 #ifdef RCU_MEMBARRIER
 103 extern int rcu_has_sys_membarrier;
 104
 105 static inline void smp_mb_slave(int group)
 106 {
 107         if (caa_likely(rcu_has_sys_membarrier))
 108                 cmm_barrier();
 109         else
 110                 cmm_smp_mb();
 111 }
 112 #endif
 113
 114 #ifdef RCU_MB
 115 static inline void smp_mb_slave(int group)
 116 {
 117         cmm_smp_mb();
 118 }
 119 #endif
 120
 121 #ifdef RCU_SIGNAL
 122 static inline void smp_mb_slave(int group)
 123 {
 124         cmm_barrier();
 125 }
 126 #endif
 127
 128 /*
 129  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
 130  * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 131  */
 132 #define RCU_GP_COUNT            (1UL << 0)
 133 /* Use the amount of bits equal to half of the architecture long size */
 134 #define RCU_GP_CTR_PHASE        (1UL << (sizeof(unsigned long) << 2))
 135 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 136
 137 struct rcu_gp {
 138         /*
 139          * Global grace period counter.
 140          * Contains the current RCU_GP_CTR_PHASE.
 141          * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
 142          * Written to only by writer with mutex taken.
 143          * Read by both writer and readers.
 144          */
 145         unsigned long ctr;
 146
 147         int32_t futex;
 148 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 149
 150 extern struct rcu_gp rcu_gp;
 151
 152 struct rcu_reader {
 153         /* Data used by both reader and synchronize_rcu() */
 154         unsigned long ctr;
 155         char need_mb;
 156         /* Data used for registry */
 157         struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 158         pthread_t tid;
 159 };
 160
 161 extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader);
 162
 163 /*
 164  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 165  */
 166 static inline void wake_up_gp(void)
 167 {
 168         if (caa_unlikely(uatomic_read(&rcu_gp.futex) == -1)) {
 169                 uatomic_set(&rcu_gp.futex, 0);
 170                 /*
 171                  * Ignoring return value until we can make this function
 172                  * return something (because urcu_die() is not publicly
 173                  * exposed).
 174                  */
 175                 (void) futex_async(&rcu_gp.futex, FUTEX_WAKE, 1,
 176                                 NULL, NULL, 0);
 177         }
 178 }
 179
 180 static inline enum rcu_state rcu_reader_state(unsigned long *ctr)
 181 {
 182         unsigned long v;
 183
 184         /*
 185          * Make sure both tests below are done on the same version of *value
 186          * to insure consistency.
 187          */
 188         v = CMM_LOAD_SHARED(*ctr);
 189         if (!(v & RCU_GP_CTR_NEST_MASK))
 190                 return RCU_READER_INACTIVE;
 191         if (!((v ^ rcu_gp.ctr) & RCU_GP_CTR_PHASE))
 192                 return RCU_READER_ACTIVE_CURRENT;
 193         return RCU_READER_ACTIVE_OLD;
 194 }
 195
 196 /*
 197  * Helper for _rcu_read_lock().  The format of rcu_gp.ctr (as well as
 198  * the per-thread rcu_reader.ctr) has the upper bits containing a count of
 199  * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
 200  * or RCU_GP_CTR_PHASE.  The smp_mb_slave() ensures that the accesses in
 201  * _rcu_read_lock() happen before the subsequent read-side critical section.
 202  */
 203 static inline void _rcu_read_lock_update(unsigned long tmp)
 204 {
 205         if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 206                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp.ctr));
 207                 smp_mb_slave(RCU_MB_GROUP);
 208         } else
 209                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT);
 210 }
 211
 212 /*
 213  * Enter an RCU read-side critical section.
 214  *
 215  * The first cmm_barrier() call ensures that the compiler does not reorder
 216  * the body of _rcu_read_lock() with a mutex.
 217  *
 218  * This function and its helper are both less than 10 lines long.  The
 219  * intent is that this function meets the 10-line criterion in LGPL,
 220  * allowing this function to be invoked directly from non-LGPL code.
 221  */
 222 static inline void _rcu_read_lock(void)
 223 {
 224         unsigned long tmp;
 225
 226         cmm_barrier();
 227         tmp = URCU_TLS(rcu_reader).ctr;
 228         _rcu_read_lock_update(tmp);
 229 }
 230
 231 /*
 232  * This is a helper function for _rcu_read_unlock().
 233  *
 234  * The first smp_mb_slave() call ensures that the critical section is
 235  * seen to precede the store to rcu_reader.ctr.
 236  * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr
 237  * before reading the update-side futex.
 238  */
 239 static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp)
 240 {
 241         if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 242                 smp_mb_slave(RCU_MB_GROUP);
 243                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
 244                 smp_mb_slave(RCU_MB_GROUP);
 245                 wake_up_gp();
 246         } else
 247                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
 248 }
 249
 250 /*
 251  * Exit an RCU read-side crtical section.  Both this function and its
 252  * helper are smaller than 10 lines of code, and are intended to be
 253  * usable by non-LGPL code, as called out in LGPL.
 254  */
 255 static inline void _rcu_read_unlock(void)
 256 {
 257         unsigned long tmp;
 258
 259         tmp = URCU_TLS(rcu_reader).ctr;
 260         _rcu_read_unlock_update_and_wakeup(tmp);
 261         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 262 }
 263
 264 /*
 265  * Returns whether within a RCU read-side critical section.
 266  *
 267  * This function is less than 10 lines long.  The intent is that this
 268  * function meets the 10-line criterion for LGPL, allowing this function
 269  * to be invoked directly from non-LGPL code.
 270  */
 271 static inline int _rcu_read_ongoing(void)
 272 {
 273         return URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK;
 274 }
 275
 276 #ifdef __cplusplus
 277 }
 278 #endif
 279
 280 #endif /* _URCU_STATIC_H */