urcu/static/urcu-bp.h

   1 #ifndef _URCU_BP_STATIC_H
   2 #define _URCU_BP_STATIC_H
   3
   4 /*
   5  * urcu-bp-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
  36 #include <urcu/compiler.h>
  37 #include <urcu/arch.h>
  38 #include <urcu/system.h>
  39 #include <urcu/uatomic.h>
  40 #include <urcu/list.h>
  41 #include <urcu/tls-compat.h>
  42
  43 /*
  44  * This code section can only be included in LGPL 2.1 compatible source code.
  45  * See below for the function call wrappers which can be used in code meant to
  46  * be only linked with the Userspace RCU library. This comes with a small
  47  * performance degradation on the read-side due to the added function calls.
  48  * This is required to permit relinking with newer versions of the library.
  49  */
  50
  51 #ifdef __cplusplus
  52 extern "C" {
  53 #endif
  54
  55 #ifdef DEBUG_RCU
  56 #define rcu_assert(args...)     assert(args)
  57 #else
  58 #define rcu_assert(args...)
  59 #endif
  60
  61 #ifdef DEBUG_YIELD
  62 #include <sched.h>
  63 #include <time.h>
  64 #include <pthread.h>
  65 #include <unistd.h>
  66
  67 #define RCU_YIELD_READ  (1 << 0)
  68 #define RCU_YIELD_WRITE (1 << 1)
  69
  70 /*
  71  * Updates without RCU_MB are much slower. Account this in
  72  * the delay.
  73  */
  74 /* maximum sleep delay, in us */
  75 #define MAX_SLEEP 50
  76
  77 extern unsigned int rcu_yield_active;
  78 extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield);
  79
  80 static inline void rcu_debug_yield_read(void)
  81 {
  82         if (rcu_yield_active & RCU_YIELD_READ)
  83                 if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
  84                         usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
  85 }
  86
  87 static inline void rcu_debug_yield_write(void)
  88 {
  89         if (rcu_yield_active & RCU_YIELD_WRITE)
  90                 if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
  91                         usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
  92 }
  93
  94 static inline void rcu_debug_yield_init(void)
  95 {
  96         URCU_TLS(rcu_rand_yield) = time(NULL) ^ pthread_self();
  97 }
  98 #else
  99 static inline void rcu_debug_yield_read(void)
 100 {
 101 }
 102
 103 static inline void rcu_debug_yield_write(void)
 104 {
 105 }
 106
 107 static inline void rcu_debug_yield_init(void)
 108 {
 109
 110 }
 111 #endif
 112
 113 /*
 114  * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use a
 115  * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 116  */
 117 #define RCU_GP_COUNT            (1UL << 0)
 118 /* Use the amount of bits equal to half of the architecture long size */
 119 #define RCU_GP_CTR_PHASE                (1UL << (sizeof(long) << 2))
 120 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_PHASE - 1)
 121
 122 /*
 123  * Used internally by _rcu_read_lock.
 124  */
 125 extern void rcu_bp_register(void);
 126
 127 /*
 128  * Global quiescent period counter with low-order bits unused.
 129  * Using a int rather than a char to eliminate false register dependencies
 130  * causing stalls on some architectures.
 131  */
 132 extern long rcu_gp_ctr;
 133
 134 struct rcu_reader {
 135         /* Data used by both reader and synchronize_rcu() */
 136         long ctr;
 137         /* Data used for registry */
 138         struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 139         pthread_t tid;
 140         int alloc;      /* registry entry allocated */
 141 };
 142
 143 /*
 144  * Bulletproof version keeps a pointer to a registry not part of the TLS.
 145  * Adds a pointer dereference on the read-side, but won't require to unregister
 146  * the reader thread.
 147  */
 148 extern DECLARE_URCU_TLS(struct rcu_reader *, rcu_reader);
 149
 150 static inline int rcu_old_gp_ongoing(long *value)
 151 {
 152         long v;
 153
 154         if (value == NULL)
 155                 return 0;
 156         /*
 157          * Make sure both tests below are done on the same version of *value
 158          * to insure consistency.
 159          */
 160         v = CMM_LOAD_SHARED(*value);
 161         return (v & RCU_GP_CTR_NEST_MASK) &&
 162                  ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
 163 }
 164
 165 /*
 166  * Helper for _rcu_read_lock().  The format of rcu_gp_ctr (as well as
 167  * the per-thread rcu_reader.ctr) has the upper bits containing a count of
 168  * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
 169  * or RCU_GP_CTR_PHASE.  The smp_mb_slave() ensures that the accesses in
 170  * _rcu_read_lock() happen before the subsequent read-side critical section.
 171  */
 172 static inline void _rcu_read_lock_update(unsigned long tmp)
 173 {
 174         if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 175                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
 176                 cmm_smp_mb();
 177         } else
 178                 _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp + RCU_GP_COUNT);
 179 }
 180
 181 /*
 182  * Enter an RCU read-side critical section.
 183  *
 184  * The first cmm_barrier() call ensures that the compiler does not reorder
 185  * the body of _rcu_read_lock() with a mutex.
 186  *
 187  * This function and its helper are both less than 10 lines long.  The
 188  * intent is that this function meets the 10-line criterion in LGPL,
 189  * allowing this function to be invoked directly from non-LGPL code.
 190  */
 191 static inline void _rcu_read_lock(void)
 192 {
 193         long tmp;
 194
 195         if (caa_unlikely(!URCU_TLS(rcu_reader)))
 196                 rcu_bp_register(); /* If not yet registered. */
 197         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 198         tmp = URCU_TLS(rcu_reader)->ctr;
 199         _rcu_read_lock_update(tmp);
 200 }
 201
 202 /*
 203  * Exit an RCU read-side critical section.  This function is less than
 204  * 10 lines of code, and is intended to be usable by non-LGPL code, as
 205  * called out in LGPL.
 206  */
 207 static inline void _rcu_read_unlock(void)
 208 {
 209         /*
 210          * Finish using rcu before decrementing the pointer.
 211          */
 212         cmm_smp_mb();
 213         _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, URCU_TLS(rcu_reader)->ctr - RCU_GP_COUNT);
 214         cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 215 }
 216
 217 #ifdef __cplusplus
 218 }
 219 #endif
 220
 221 #endif /* _URCU_BP_STATIC_H */