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@polymtl.ca>
  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/arch_uatomic.h>
  40 #include <urcu/list.h>
  41
  42 /*
  43  * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
  44  */
  45 #define _LOAD_SHARED(p)        ACCESS_ONCE(p)
  46
  47 /*
  48  * Load a data from shared memory, doing a cache flush if required.
  49  */
  50 #define LOAD_SHARED(p)                  \
  51         ({                              \
  52                 smp_rmc();              \
  53                 _LOAD_SHARED(p);        \
  54         })
  55
  56 /*
  57  * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
  58  */
  59 #define _STORE_SHARED(x, v)     ({ ACCESS_ONCE(x) = (v); })
  60
  61 /*
  62  * Store v into x, where x is located in shared memory. Performs the required
  63  * cache flush after writing. Returns v.
  64  */
  65 #define STORE_SHARED(x, v)              \
  66         ({                              \
  67                 _STORE_SHARED(x, v);    \
  68                 smp_wmc();              \
  69                 (v);                    \
  70         })
  71
  72 /**
  73  * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
  74  * into a RCU read-side critical section. The pointer can later be safely
  75  * dereferenced within the critical section.
  76  *
  77  * This ensures that the pointer copy is invariant thorough the whole critical
  78  * section.
  79  *
  80  * Inserts memory barriers on architectures that require them (currently only
  81  * Alpha) and documents which pointers are protected by RCU.
  82  *
  83  * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
  84  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
  85  * data read before the pointer read by speculating the value of the pointer.
  86  * Correct ordering is ensured because the pointer is read as a volatile access.
  87  * This acts as a global side-effect operation, which forbids reordering of
  88  * dependent memory operations. Note that such concern about dependency-breaking
  89  * optimizations will eventually be taken care of by the "memory_order_consume"
  90  * addition to forthcoming C++ standard.
  91  *
  92  * Should match rcu_assign_pointer() or rcu_xchg_pointer().
  93  */
  94
  95 #define _rcu_dereference(p)     ({                                      \
  96                                 typeof(p) _________p1 = LOAD_SHARED(p); \
  97                                 smp_read_barrier_depends();             \
  98                                 (_________p1);                          \
  99                                 })
 100
 101 #define futex(...)              syscall(__NR_futex, __VA_ARGS__)
 102 #define FUTEX_WAIT              0
 103 #define FUTEX_WAKE              1
 104
 105 /*
 106  * This code section can only be included in LGPL 2.1 compatible source code.
 107  * See below for the function call wrappers which can be used in code meant to
 108  * be only linked with the Userspace RCU library. This comes with a small
 109  * performance degradation on the read-side due to the added function calls.
 110  * This is required to permit relinking with newer versions of the library.
 111  */
 112
 113 /*
 114  * The signal number used by the RCU library can be overridden with
 115  * -DSIGURCU= when compiling the library.
 116  */
 117 #ifndef SIGURCU
 118 #define SIGURCU SIGUSR1
 119 #endif
 120
 121 /*
 122  * If a reader is really non-cooperative and refuses to commit its
 123  * urcu_active_readers count to memory (there is no barrier in the reader
 124  * per-se), kick it after a few loops waiting for it.
 125  */
 126 #define KICK_READER_LOOPS 10000
 127
 128 /*
 129  * Active attempts to check for reader Q.S. before calling futex().
 130  */
 131 #define RCU_QS_ACTIVE_ATTEMPTS 100
 132
 133 #ifdef DEBUG_RCU
 134 #define rcu_assert(args...)     assert(args)
 135 #else
 136 #define rcu_assert(args...)
 137 #endif
 138
 139 #ifdef DEBUG_YIELD
 140 #include <sched.h>
 141 #include <time.h>
 142 #include <pthread.h>
 143 #include <unistd.h>
 144
 145 #define YIELD_READ      (1 << 0)
 146 #define YIELD_WRITE     (1 << 1)
 147
 148 /*
 149  * Updates without URCU_MB are much slower. Account this in
 150  * the delay.
 151  */
 152 #ifdef URCU_MB
 153 /* maximum sleep delay, in us */
 154 #define MAX_SLEEP 50
 155 #else
 156 #define MAX_SLEEP 30000
 157 #endif
 158
 159 extern unsigned int yield_active;
 160 extern unsigned int __thread rand_yield;
 161
 162 static inline void debug_yield_read(void)
 163 {
 164         if (yield_active & YIELD_READ)
 165                 if (rand_r(&rand_yield) & 0x1)
 166                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 167 }
 168
 169 static inline void debug_yield_write(void)
 170 {
 171         if (yield_active & YIELD_WRITE)
 172                 if (rand_r(&rand_yield) & 0x1)
 173                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 174 }
 175
 176 static inline void debug_yield_init(void)
 177 {
 178         rand_yield = time(NULL) ^ pthread_self();
 179 }
 180 #else
 181 static inline void debug_yield_read(void)
 182 {
 183 }
 184
 185 static inline void debug_yield_write(void)
 186 {
 187 }
 188
 189 static inline void debug_yield_init(void)
 190 {
 191
 192 }
 193 #endif
 194
 195 #ifdef URCU_MB
 196 static inline void reader_barrier()
 197 {
 198         smp_mb();
 199 }
 200 #else
 201 static inline void reader_barrier()
 202 {
 203         barrier();
 204 }
 205 #endif
 206
 207 /*
 208  * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
 209  * 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_BIT          (1UL << (sizeof(long) << 2))
 214 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_BIT - 1)
 215
 216 /*
 217  * Global quiescent period counter with low-order bits unused.
 218  * Using a int rather than a char to eliminate false register dependencies
 219  * causing stalls on some architectures.
 220  */
 221 extern long urcu_gp_ctr;
 222
 223 struct urcu_reader {
 224         /* Data used by both reader and synchronize_rcu() */
 225         long ctr;
 226         char need_mb;
 227         /* Data used for registry */
 228         struct list_head head __attribute__((aligned(CACHE_LINE_SIZE)));
 229         pthread_t tid;
 230 };
 231
 232 extern struct urcu_reader __thread urcu_reader;
 233
 234 extern int gp_futex;
 235
 236 /*
 237  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 238  */
 239 static inline void wake_up_gp(void)
 240 {
 241         if (unlikely(uatomic_read(&gp_futex) == -1)) {
 242                 uatomic_set(&gp_futex, 0);
 243                 futex(&gp_futex, FUTEX_WAKE, 1,
 244                       NULL, NULL, 0);
 245         }
 246 }
 247
 248 static inline int rcu_old_gp_ongoing(long *value)
 249 {
 250         long v;
 251
 252         if (value == NULL)
 253                 return 0;
 254         /*
 255          * Make sure both tests below are done on the same version of *value
 256          * to insure consistency.
 257          */
 258         v = LOAD_SHARED(*value);
 259         return (v & RCU_GP_CTR_NEST_MASK) &&
 260                  ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
 261 }
 262
 263 static inline void _rcu_read_lock(void)
 264 {
 265         long tmp;
 266
 267         tmp = urcu_reader.ctr;
 268         /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
 269         if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 270                 _STORE_SHARED(urcu_reader.ctr, _LOAD_SHARED(urcu_gp_ctr));
 271                 /*
 272                  * Set active readers count for outermost nesting level before
 273                  * accessing the pointer. See force_mb_all_threads().
 274                  */
 275                 reader_barrier();
 276         } else {
 277                 _STORE_SHARED(urcu_reader.ctr, tmp + RCU_GP_COUNT);
 278         }
 279 }
 280
 281 static inline void _rcu_read_unlock(void)
 282 {
 283         long tmp;
 284
 285         tmp = urcu_reader.ctr;
 286         /*
 287          * Finish using rcu before decrementing the pointer.
 288          * See force_mb_all_threads().
 289          */
 290         if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 291                 reader_barrier();
 292                 _STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT);
 293                 /* write urcu_reader.ctr before read futex */
 294                 reader_barrier();
 295                 wake_up_gp();
 296         } else {
 297                 _STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT);
 298         }
 299 }
 300
 301 /**
 302  * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
 303  * meant to be read by RCU read-side critical sections. Returns the assigned
 304  * value.
 305  *
 306  * Documents which pointers will be dereferenced by RCU read-side critical
 307  * sections and adds the required memory barriers on architectures requiring
 308  * them. It also makes sure the compiler does not reorder code initializing the
 309  * data structure before its publication.
 310  *
 311  * Should match rcu_dereference_pointer().
 312  */
 313
 314 #define _rcu_assign_pointer(p, v)                       \
 315         ({                                              \
 316                 if (!__builtin_constant_p(v) ||         \
 317                     ((v) != NULL))                      \
 318                         wmb();                          \
 319                 STORE_SHARED(p, v);                     \
 320         })
 321
 322 /**
 323  * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
 324  * is as expected by "old". If succeeds, returns the previous pointer to the
 325  * data structure, which can be safely freed after waiting for a quiescent state
 326  * using synchronize_rcu(). If fails (unexpected value), returns old (which
 327  * should not be freed !).
 328  */
 329
 330 #define _rcu_cmpxchg_pointer(p, old, _new)              \
 331         ({                                              \
 332                 if (!__builtin_constant_p(_new) ||      \
 333                     ((_new) != NULL))                   \
 334                         wmb();                          \
 335                 uatomic_cmpxchg(p, old, _new);          \
 336         })
 337
 338 /**
 339  * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
 340  * pointer to the data structure, which can be safely freed after waiting for a
 341  * quiescent state using synchronize_rcu().
 342  */
 343
 344 #define _rcu_xchg_pointer(p, v)                         \
 345         ({                                              \
 346                 if (!__builtin_constant_p(v) ||         \
 347                     ((v) != NULL))                      \
 348                         wmb();                          \
 349                 uatomic_xchg(p, v);                     \
 350         })
 351
 352 /*
 353  * Exchanges the pointer and waits for quiescent state.
 354  * The pointer returned can be freed.
 355  */
 356 #define _rcu_publish_content(p, v)                      \
 357         ({                                              \
 358                 void *oldptr;                           \
 359                 oldptr = _rcu_xchg_pointer(p, v);       \
 360                 synchronize_rcu();                      \
 361                 oldptr;                                 \
 362         })
 363
 364 #endif /* _URCU_STATIC_H */