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/list.h>
  40
  41 /*
  42  * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
  43  */
  44 #define _LOAD_SHARED(p)        ACCESS_ONCE(p)
  45
  46 /*
  47  * Load a data from shared memory, doing a cache flush if required.
  48  */
  49 #define LOAD_SHARED(p)                  \
  50         ({                              \
  51                 smp_rmc();              \
  52                 _LOAD_SHARED(p);        \
  53         })
  54
  55 /*
  56  * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
  57  */
  58 #define _STORE_SHARED(x, v)     ({ ACCESS_ONCE(x) = (v); })
  59
  60 /*
  61  * Store v into x, where x is located in shared memory. Performs the required
  62  * cache flush after writing. Returns v.
  63  */
  64 #define STORE_SHARED(x, v)              \
  65         ({                              \
  66                 _STORE_SHARED(x, v);    \
  67                 smp_wmc();              \
  68                 (v);                    \
  69         })
  70
  71 /**
  72  * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
  73  * into a RCU read-side critical section. The pointer can later be safely
  74  * dereferenced within the critical section.
  75  *
  76  * This ensures that the pointer copy is invariant thorough the whole critical
  77  * section.
  78  *
  79  * Inserts memory barriers on architectures that require them (currently only
  80  * Alpha) and documents which pointers are protected by RCU.
  81  *
  82  * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
  83  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
  84  * data read before the pointer read by speculating the value of the pointer.
  85  * Correct ordering is ensured because the pointer is read as a volatile access.
  86  * This acts as a global side-effect operation, which forbids reordering of
  87  * dependent memory operations. Note that such concern about dependency-breaking
  88  * optimizations will eventually be taken care of by the "memory_order_consume"
  89  * addition to forthcoming C++ standard.
  90  *
  91  * Should match rcu_assign_pointer() or rcu_xchg_pointer().
  92  */
  93
  94 #define _rcu_dereference(p)     ({                                      \
  95                                 typeof(p) _________p1 = LOAD_SHARED(p); \
  96                                 smp_read_barrier_depends();             \
  97                                 (_________p1);                          \
  98                                 })
  99
 100 #define futex(...)              syscall(__NR_futex, __VA_ARGS__)
 101 #define FUTEX_WAIT              0
 102 #define FUTEX_WAKE              1
 103
 104 /*
 105  * This code section can only be included in LGPL 2.1 compatible source code.
 106  * See below for the function call wrappers which can be used in code meant to
 107  * be only linked with the Userspace RCU library. This comes with a small
 108  * performance degradation on the read-side due to the added function calls.
 109  * This is required to permit relinking with newer versions of the library.
 110  */
 111
 112 /*
 113  * The signal number used by the RCU library can be overridden with
 114  * -DSIGURCU= when compiling the library.
 115  */
 116 #ifndef SIGURCU
 117 #define SIGURCU SIGUSR1
 118 #endif
 119
 120 /*
 121  * If a reader is really non-cooperative and refuses to commit its
 122  * urcu_active_readers count to memory (there is no barrier in the reader
 123  * per-se), kick it after a few loops waiting for it.
 124  */
 125 #define KICK_READER_LOOPS 10000
 126
 127 /*
 128  * Active attempts to check for reader Q.S. before calling futex().
 129  */
 130 #define RCU_QS_ACTIVE_ATTEMPTS 100
 131
 132 #ifdef DEBUG_RCU
 133 #define rcu_assert(args...)     assert(args)
 134 #else
 135 #define rcu_assert(args...)
 136 #endif
 137
 138 #ifdef DEBUG_YIELD
 139 #include <sched.h>
 140 #include <time.h>
 141 #include <pthread.h>
 142 #include <unistd.h>
 143
 144 #define YIELD_READ      (1 << 0)
 145 #define YIELD_WRITE     (1 << 1)
 146
 147 /*
 148  * Updates without URCU_MB are much slower. Account this in
 149  * the delay.
 150  */
 151 #ifdef URCU_MB
 152 /* maximum sleep delay, in us */
 153 #define MAX_SLEEP 50
 154 #else
 155 #define MAX_SLEEP 30000
 156 #endif
 157
 158 extern unsigned int yield_active;
 159 extern unsigned int __thread rand_yield;
 160
 161 static inline void debug_yield_read(void)
 162 {
 163         if (yield_active & YIELD_READ)
 164                 if (rand_r(&rand_yield) & 0x1)
 165                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 166 }
 167
 168 static inline void debug_yield_write(void)
 169 {
 170         if (yield_active & YIELD_WRITE)
 171                 if (rand_r(&rand_yield) & 0x1)
 172                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 173 }
 174
 175 static inline void debug_yield_init(void)
 176 {
 177         rand_yield = time(NULL) ^ pthread_self();
 178 }
 179 #else
 180 static inline void debug_yield_read(void)
 181 {
 182 }
 183
 184 static inline void debug_yield_write(void)
 185 {
 186 }
 187
 188 static inline void debug_yield_init(void)
 189 {
 190
 191 }
 192 #endif
 193
 194 #ifdef URCU_MB
 195 static inline void reader_barrier()
 196 {
 197         smp_mb();
 198 }
 199 #else
 200 static inline void reader_barrier()
 201 {
 202         barrier();
 203 }
 204 #endif
 205
 206 /*
 207  * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
 208  * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 209  */
 210 #define RCU_GP_COUNT            (1UL << 0)
 211 /* Use the amount of bits equal to half of the architecture long size */
 212 #define RCU_GP_CTR_BIT          (1UL << (sizeof(long) << 2))
 213 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_BIT - 1)
 214
 215 /*
 216  * Global quiescent period counter with low-order bits unused.
 217  * Using a int rather than a char to eliminate false register dependencies
 218  * causing stalls on some architectures.
 219  */
 220 extern long urcu_gp_ctr;
 221
 222 struct urcu_reader {
 223         /* Data used by both reader and synchronize_rcu() */
 224         long ctr;
 225         char need_mb;
 226         /* Data used for registry */
 227         struct list_head head __attribute__((aligned(CACHE_LINE_SIZE)));
 228         pthread_t tid;
 229 };
 230
 231 extern struct urcu_reader __thread urcu_reader;
 232
 233 extern int gp_futex;
 234
 235 /*
 236  * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
 237  */
 238 static inline void wake_up_gp(void)
 239 {
 240         if (unlikely(uatomic_read(&gp_futex) == -1)) {
 241                 uatomic_set(&gp_futex, 0);
 242                 futex(&gp_futex, FUTEX_WAKE, 1,
 243                       NULL, NULL, 0);
 244         }
 245 }
 246
 247 static inline int rcu_old_gp_ongoing(long *value)
 248 {
 249         long v;
 250
 251         if (value == NULL)
 252                 return 0;
 253         /*
 254          * Make sure both tests below are done on the same version of *value
 255          * to insure consistency.
 256          */
 257         v = LOAD_SHARED(*value);
 258         return (v & RCU_GP_CTR_NEST_MASK) &&
 259                  ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
 260 }
 261
 262 static inline void _rcu_read_lock(void)
 263 {
 264         long tmp;
 265
 266         tmp = urcu_reader.ctr;
 267         /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
 268         if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 269                 _STORE_SHARED(urcu_reader.ctr, _LOAD_SHARED(urcu_gp_ctr));
 270                 /*
 271                  * Set active readers count for outermost nesting level before
 272                  * accessing the pointer. See force_mb_all_threads().
 273                  */
 274                 reader_barrier();
 275         } else {
 276                 _STORE_SHARED(urcu_reader.ctr, tmp + RCU_GP_COUNT);
 277         }
 278 }
 279
 280 static inline void _rcu_read_unlock(void)
 281 {
 282         long tmp;
 283
 284         tmp = urcu_reader.ctr;
 285         /*
 286          * Finish using rcu before decrementing the pointer.
 287          * See force_mb_all_threads().
 288          */
 289         if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
 290                 reader_barrier();
 291                 _STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT);
 292                 /* write urcu_reader.ctr before read futex */
 293                 reader_barrier();
 294                 wake_up_gp();
 295         } else {
 296                 _STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT);
 297         }
 298 }
 299
 300 /**
 301  * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
 302  * meant to be read by RCU read-side critical sections. Returns the assigned
 303  * value.
 304  *
 305  * Documents which pointers will be dereferenced by RCU read-side critical
 306  * sections and adds the required memory barriers on architectures requiring
 307  * them. It also makes sure the compiler does not reorder code initializing the
 308  * data structure before its publication.
 309  *
 310  * Should match rcu_dereference_pointer().
 311  */
 312
 313 #define _rcu_assign_pointer(p, v)                       \
 314         ({                                              \
 315                 if (!__builtin_constant_p(v) ||         \
 316                     ((v) != NULL))                      \
 317                         wmb();                          \
 318                 STORE_SHARED(p, v);                     \
 319         })
 320
 321 /**
 322  * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
 323  * is as expected by "old". If succeeds, returns the previous pointer to the
 324  * data structure, which can be safely freed after waiting for a quiescent state
 325  * using synchronize_rcu(). If fails (unexpected value), returns old (which
 326  * should not be freed !).
 327  */
 328
 329 #define _rcu_cmpxchg_pointer(p, old, _new)              \
 330         ({                                              \
 331                 if (!__builtin_constant_p(_new) ||      \
 332                     ((_new) != NULL))                   \
 333                         wmb();                          \
 334                 uatomic_cmpxchg(p, old, _new);          \
 335         })
 336
 337 /**
 338  * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
 339  * pointer to the data structure, which can be safely freed after waiting for a
 340  * quiescent state using synchronize_rcu().
 341  */
 342
 343 #define _rcu_xchg_pointer(p, v)                         \
 344         ({                                              \
 345                 if (!__builtin_constant_p(v) ||         \
 346                     ((v) != NULL))                      \
 347                         wmb();                          \
 348                 uatomic_xchg(p, v);                     \
 349         })
 350
 351 /*
 352  * Exchanges the pointer and waits for quiescent state.
 353  * The pointer returned can be freed.
 354  */
 355 #define _rcu_publish_content(p, v)                      \
 356         ({                                              \
 357                 void *oldptr;                           \
 358                 oldptr = _rcu_xchg_pointer(p, v);       \
 359                 synchronize_rcu();                      \
 360                 oldptr;                                 \
 361         })
 362
 363 #endif /* _URCU_STATIC_H */