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
  35 #include <compiler.h>
  36 #include <arch.h>
  37
  38 /*
  39  * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
  40  */
  41 #define _LOAD_SHARED(p)        ACCESS_ONCE(p)
  42
  43 /*
  44  * Load a data from shared memory, doing a cache flush if required.
  45  */
  46 #define LOAD_SHARED(p)                  \
  47         ({                              \
  48                 smp_rmc();              \
  49                 _LOAD_SHARED(p);        \
  50         })
  51
  52 /*
  53  * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
  54  */
  55 #define _STORE_SHARED(x, v)     ({ ACCESS_ONCE(x) = (v); })
  56
  57 /*
  58  * Store v into x, where x is located in shared memory. Performs the required
  59  * cache flush after writing. Returns v.
  60  */
  61 #define STORE_SHARED(x, v)              \
  62         ({                              \
  63                 _STORE_SHARED(x, v);    \
  64                 smp_wmc();              \
  65                 (v);                    \
  66         })
  67
  68 /**
  69  * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
  70  * into a RCU read-side critical section. The pointer can later be safely
  71  * dereferenced within the critical section.
  72  *
  73  * This ensures that the pointer copy is invariant thorough the whole critical
  74  * section.
  75  *
  76  * Inserts memory barriers on architectures that require them (currently only
  77  * Alpha) and documents which pointers are protected by RCU.
  78  *
  79  * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
  80  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
  81  * data read before the pointer read by speculating the value of the pointer.
  82  * Correct ordering is ensured because the pointer is read as a volatile access.
  83  * This acts as a global side-effect operation, which forbids reordering of
  84  * dependent memory operations. Note that such concern about dependency-breaking
  85  * optimizations will eventually be taken care of by the "memory_order_consume"
  86  * addition to forthcoming C++ standard.
  87  *
  88  * Should match rcu_assign_pointer() or rcu_xchg_pointer().
  89  */
  90
  91 #define _rcu_dereference(p)     ({                                      \
  92                                 typeof(p) _________p1 = LOAD_SHARED(p); \
  93                                 smp_read_barrier_depends();             \
  94                                 (_________p1);                          \
  95                                 })
  96
  97 /*
  98  * This code section can only be included in LGPL 2.1 compatible source code.
  99  * See below for the function call wrappers which can be used in code meant to
 100  * be only linked with the Userspace RCU library. This comes with a small
 101  * performance degradation on the read-side due to the added function calls.
 102  * This is required to permit relinking with newer versions of the library.
 103  */
 104
 105 /*
 106  * The signal number used by the RCU library can be overridden with
 107  * -DSIGURCU= when compiling the library.
 108  */
 109 #ifndef SIGURCU
 110 #define SIGURCU SIGUSR1
 111 #endif
 112
 113 /*
 114  * If a reader is really non-cooperative and refuses to commit its
 115  * urcu_active_readers count to memory (there is no barrier in the reader
 116  * per-se), kick it after a few loops waiting for it.
 117  */
 118 #define KICK_READER_LOOPS 10000
 119
 120 #ifdef DEBUG_RCU
 121 #define rcu_assert(args...)     assert(args)
 122 #else
 123 #define rcu_assert(args...)
 124 #endif
 125
 126 #ifdef DEBUG_YIELD
 127 #include <sched.h>
 128 #include <time.h>
 129 #include <pthread.h>
 130 #include <unistd.h>
 131
 132 #define YIELD_READ      (1 << 0)
 133 #define YIELD_WRITE     (1 << 1)
 134
 135 /*
 136  * Updates without URCU_MB are much slower. Account this in
 137  * the delay.
 138  */
 139 #ifdef URCU_MB
 140 /* maximum sleep delay, in us */
 141 #define MAX_SLEEP 50
 142 #else
 143 #define MAX_SLEEP 30000
 144 #endif
 145
 146 extern unsigned int yield_active;
 147 extern unsigned int __thread rand_yield;
 148
 149 static inline void debug_yield_read(void)
 150 {
 151         if (yield_active & YIELD_READ)
 152                 if (rand_r(&rand_yield) & 0x1)
 153                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 154 }
 155
 156 static inline void debug_yield_write(void)
 157 {
 158         if (yield_active & YIELD_WRITE)
 159                 if (rand_r(&rand_yield) & 0x1)
 160                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 161 }
 162
 163 static inline void debug_yield_init(void)
 164 {
 165         rand_yield = time(NULL) ^ pthread_self();
 166 }
 167 #else
 168 static inline void debug_yield_read(void)
 169 {
 170 }
 171
 172 static inline void debug_yield_write(void)
 173 {
 174 }
 175
 176 static inline void debug_yield_init(void)
 177 {
 178
 179 }
 180 #endif
 181
 182 #ifdef URCU_MB
 183 static inline void reader_barrier()
 184 {
 185         smp_mb();
 186 }
 187 #else
 188 static inline void reader_barrier()
 189 {
 190         barrier();
 191 }
 192 #endif
 193
 194 /*
 195  * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
 196  * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 197  */
 198 #define RCU_GP_COUNT            (1UL << 0)
 199 /* Use the amount of bits equal to half of the architecture long size */
 200 #define RCU_GP_CTR_BIT          (1UL << (sizeof(long) << 2))
 201 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_BIT - 1)
 202
 203 /*
 204  * Global quiescent period counter with low-order bits unused.
 205  * Using a int rather than a char to eliminate false register dependencies
 206  * causing stalls on some architectures.
 207  */
 208 extern long urcu_gp_ctr;
 209
 210 extern long __thread urcu_active_readers;
 211
 212 static inline int rcu_old_gp_ongoing(long *value)
 213 {
 214         long v;
 215
 216         if (value == NULL)
 217                 return 0;
 218         /*
 219          * Make sure both tests below are done on the same version of *value
 220          * to insure consistency.
 221          */
 222         v = LOAD_SHARED(*value);
 223         return (v & RCU_GP_CTR_NEST_MASK) &&
 224                  ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
 225 }
 226
 227 static inline void _rcu_read_lock(void)
 228 {
 229         long tmp;
 230
 231         tmp = urcu_active_readers;
 232         /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
 233         if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
 234                 _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr));
 235                 /*
 236                  * Set active readers count for outermost nesting level before
 237                  * accessing the pointer. See force_mb_all_threads().
 238                  */
 239                 reader_barrier();
 240         } else {
 241                 _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT);
 242         }
 243 }
 244
 245 static inline void _rcu_read_unlock(void)
 246 {
 247         reader_barrier();
 248         /*
 249          * Finish using rcu before decrementing the pointer.
 250          * See force_mb_all_threads().
 251          * Formally only needed for outermost nesting level, but leave barrier
 252          * in place for nested unlocks to remove a branch from the common case
 253          * (no nesting).
 254          */
 255         _STORE_SHARED(urcu_active_readers, urcu_active_readers - RCU_GP_COUNT);
 256 }
 257
 258 /**
 259  * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
 260  * meant to be read by RCU read-side critical sections. Returns the assigned
 261  * value.
 262  *
 263  * Documents which pointers will be dereferenced by RCU read-side critical
 264  * sections and adds the required memory barriers on architectures requiring
 265  * them. It also makes sure the compiler does not reorder code initializing the
 266  * data structure before its publication.
 267  *
 268  * Should match rcu_dereference_pointer().
 269  */
 270
 271 #define _rcu_assign_pointer(p, v)                       \
 272         ({                                              \
 273                 if (!__builtin_constant_p(v) ||         \
 274                     ((v) != NULL))                      \
 275                         wmb();                          \
 276                 STORE_SHARED(p, v);                     \
 277         })
 278
 279 /**
 280  * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
 281  * is as expected by "old". If succeeds, returns the previous pointer to the
 282  * data structure, which can be safely freed after waiting for a quiescent state
 283  * using synchronize_rcu(). If fails (unexpected value), returns old (which
 284  * should not be freed !).
 285  */
 286
 287 #define _rcu_cmpxchg_pointer(p, old, _new)              \
 288         ({                                              \
 289                 if (!__builtin_constant_p(_new) ||      \
 290                     ((_new) != NULL))                   \
 291                         wmb();                          \
 292                 cmpxchg(p, old, _new);                  \
 293         })
 294
 295 /**
 296  * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
 297  * pointer to the data structure, which can be safely freed after waiting for a
 298  * quiescent state using synchronize_rcu().
 299  */
 300
 301 #define _rcu_xchg_pointer(p, v)                         \
 302         ({                                              \
 303                 if (!__builtin_constant_p(v) ||         \
 304                     ((v) != NULL))                      \
 305                         wmb();                          \
 306                 xchg(p, v);                             \
 307         })
 308
 309 /*
 310  * Exchanges the pointer and waits for quiescent state.
 311  * The pointer returned can be freed.
 312  */
 313 #define _rcu_publish_content(p, v)                      \
 314         ({                                              \
 315                 void *oldptr;                           \
 316                 oldptr = _rcu_xchg_pointer(p, v);       \
 317                 synchronize_rcu();                      \
 318                 oldptr;                                 \
 319         })
 320
 321 #endif /* _URCU_STATIC_H */