urcu.h

   1 #ifndef _URCU_H
   2 #define _URCU_H
   3
   4 /*
   5  * urcu.h
   6  *
   7  * Userspace RCU header
   8  *
   9  * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
  10  *
  11  * Credits for Paul e. McKenney <paulmck@linux.vnet.ibm.com>
  12  * for inspiration coming from the Linux kernel RCU and rcu-preempt.
  13  *
  14  * The barrier, mb, rmb, wmb, atomic_inc, smp_read_barrier_depends, ACCESS_ONCE
  15  * and rcu_dereference primitives come from the Linux kernel.
  16  *
  17  * Distributed under GPLv2
  18  */
  19
  20 #include <stdlib.h>
  21 #include <pthread.h>
  22
  23 /* The "volatile" is due to gcc bugs */
  24 #define barrier() __asm__ __volatile__("": : :"memory")
  25
  26 #define likely(x)       __builtin_expect(!!(x), 1)
  27 #define unlikely(x)     __builtin_expect(!!(x), 0)
  28
  29 /* x86 32/64 specific */
  30 #define mb()    asm volatile("mfence":::"memory")
  31 #define rmb()   asm volatile("lfence":::"memory")
  32 #define wmb()   asm volatile("sfence" ::: "memory")
  33
  34 static inline void atomic_inc(int *v)
  35 {
  36         asm volatile("lock; incl %0"
  37                      : "+m" (*v));
  38 }
  39
  40 #define xchg(ptr, v)                                                    \
  41         ((__typeof__(*(ptr)))__xchg((unsigned long)(v), (ptr), sizeof(*(ptr))))
  42
  43 struct __xchg_dummy {
  44         unsigned long a[100];
  45 };
  46 #define __xg(x) ((struct __xchg_dummy *)(x))
  47
  48 /*
  49  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
  50  * Note 2: xchg has side effect, so that attribute volatile is necessary,
  51  *        but generally the primitive is invalid, *ptr is output argument. --ANK
  52  */
  53 static inline unsigned long __xchg(unsigned long x, volatile void *ptr,
  54                                    int size)
  55 {
  56         switch (size) {
  57         case 1:
  58                 asm volatile("xchgb %b0,%1"
  59                              : "=q" (x)
  60                              : "m" (*__xg(ptr)), "0" (x)
  61                              : "memory");
  62                 break;
  63         case 2:
  64                 asm volatile("xchgw %w0,%1"
  65                              : "=r" (x)
  66                              : "m" (*__xg(ptr)), "0" (x)
  67                              : "memory");
  68                 break;
  69         case 4:
  70                 asm volatile("xchgl %k0,%1"
  71                              : "=r" (x)
  72                              : "m" (*__xg(ptr)), "0" (x)
  73                              : "memory");
  74                 break;
  75         case 8:
  76                 asm volatile("xchgq %0,%1"
  77                              : "=r" (x)
  78                              : "m" (*__xg(ptr)), "0" (x)
  79                              : "memory");
  80                 break;
  81         }
  82         return x;
  83 }
  84
  85 /* Nop everywhere except on alpha. */
  86 #define smp_read_barrier_depends()
  87
  88 /*
  89  * Prevent the compiler from merging or refetching accesses.  The compiler
  90  * is also forbidden from reordering successive instances of ACCESS_ONCE(),
  91  * but only when the compiler is aware of some particular ordering.  One way
  92  * to make the compiler aware of ordering is to put the two invocations of
  93  * ACCESS_ONCE() in different C statements.
  94  *
  95  * This macro does absolutely -nothing- to prevent the CPU from reordering,
  96  * merging, or refetching absolutely anything at any time.  Its main intended
  97  * use is to mediate communication between process-level code and irq/NMI
  98  * handlers, all running on the same CPU.
  99  */
 100 #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
 101
 102 /**
 103  * rcu_dereference - fetch an RCU-protected pointer in an
 104  * RCU read-side critical section.  This pointer may later
 105  * be safely dereferenced.
 106  *
 107  * Inserts memory barriers on architectures that require them
 108  * (currently only the Alpha), and, more importantly, documents
 109  * exactly which pointers are protected by RCU.
 110  */
 111
 112 #define rcu_dereference(p)     ({ \
 113                                 typeof(p) _________p1 = ACCESS_ONCE(p); \
 114                                 smp_read_barrier_depends(); \
 115                                 (_________p1); \
 116                                 })
 117
 118 #define SIGURCU SIGUSR1
 119
 120 #ifdef DEBUG_YIELD
 121 #include <sched.h>
 122 #include <time.h>
 123 #include <pthread.h>
 124 #include <unistd.h>
 125
 126 #define YIELD_READ      (1 << 0)
 127 #define YIELD_WRITE     (1 << 1)
 128
 129 /* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */
 130 #ifdef DEBUG_FULL_MB
 131 /* maximum sleep delay, in us */
 132 #define MAX_SLEEP 50
 133 #else
 134 #define MAX_SLEEP 30000
 135 #endif
 136
 137 extern unsigned int yield_active;
 138 extern unsigned int __thread rand_yield;
 139
 140 static inline void debug_yield_read(void)
 141 {
 142         if (yield_active & YIELD_READ)
 143                 if (rand_r(&rand_yield) & 0x1)
 144                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 145 }
 146
 147 static inline void debug_yield_write(void)
 148 {
 149         if (yield_active & YIELD_WRITE)
 150                 if (rand_r(&rand_yield) & 0x1)
 151                         usleep(rand_r(&rand_yield) % MAX_SLEEP);
 152 }
 153
 154 static inline void debug_yield_init(void)
 155 {
 156         rand_yield = time(NULL) ^ pthread_self();
 157 }
 158 #else
 159 static inline void debug_yield_read(void)
 160 {
 161 }
 162
 163 static inline void debug_yield_write(void)
 164 {
 165 }
 166
 167 static inline void debug_yield_init(void)
 168 {
 169
 170 }
 171 #endif
 172
 173 #ifdef DEBUG_FULL_MB
 174 static inline void read_barrier()
 175 {
 176         mb();
 177 }
 178 #else
 179 static inline void read_barrier()
 180 {
 181         barrier();
 182 }
 183 #endif
 184
 185 /*
 186  * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
 187  * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
 188  */
 189 #define RCU_GP_COUNT            (1UL << 0)
 190 /* Use the amount of bits equal to half of the architecture long size */
 191 #define RCU_GP_CTR_BIT          (1UL << (sizeof(long) << 2))
 192 #define RCU_GP_CTR_NEST_MASK    (RCU_GP_CTR_BIT - 1)
 193
 194 /*
 195  * Global quiescent period counter with low-order bits unused.
 196  * Using a int rather than a char to eliminate false register dependencies
 197  * causing stalls on some architectures.
 198  */
 199 extern long urcu_gp_ctr;
 200
 201 extern long __thread urcu_active_readers;
 202
 203 static inline int rcu_old_gp_ongoing(long *value)
 204 {
 205         long v;
 206
 207         if (value == NULL)
 208                 return 0;
 209         debug_yield_write();
 210         /*
 211          * Make sure both tests below are done on the same version of *value
 212          * to insure consistency.
 213          */
 214         v = ACCESS_ONCE(*value);
 215         debug_yield_write();
 216         return (v & RCU_GP_CTR_NEST_MASK) &&
 217                  ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
 218 }
 219
 220 static inline void rcu_read_lock(void)
 221 {
 222         long tmp;
 223
 224         debug_yield_read();
 225         tmp = urcu_active_readers;
 226         debug_yield_read();
 227         /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
 228         if (likely(!(tmp & RCU_GP_CTR_NEST_MASK)))
 229                 urcu_active_readers = urcu_gp_ctr;
 230         else
 231                 urcu_active_readers = tmp + RCU_GP_COUNT;
 232         debug_yield_read();
 233         /*
 234          * Increment active readers count before accessing the pointer.
 235          * See force_mb_all_threads().
 236          */
 237         read_barrier();
 238         debug_yield_read();
 239 }
 240
 241 static inline void rcu_read_unlock(void)
 242 {
 243         debug_yield_read();
 244         read_barrier();
 245         debug_yield_read();
 246         /*
 247          * Finish using rcu before decrementing the pointer.
 248          * See force_mb_all_threads().
 249          */
 250         urcu_active_readers -= RCU_GP_COUNT;
 251         debug_yield_read();
 252 }
 253
 254 /**
 255  * rcu_assign_pointer - assign (publicize) a pointer to a newly
 256  * initialized structure that will be dereferenced by RCU read-side
 257  * critical sections.  Returns the value assigned.
 258  *
 259  * Inserts memory barriers on architectures that require them
 260  * (pretty much all of them other than x86), and also prevents
 261  * the compiler from reordering the code that initializes the
 262  * structure after the pointer assignment.  More importantly, this
 263  * call documents which pointers will be dereferenced by RCU read-side
 264  * code.
 265  */
 266
 267 #define rcu_assign_pointer(p, v) \
 268         ({ \
 269                 if (!__builtin_constant_p(v) || \
 270                     ((v) != NULL)) \
 271                         wmb(); \
 272                 (p) = (v); \
 273         })
 274
 275 #define rcu_xchg_pointer(p, v) \
 276         ({ \
 277                 if (!__builtin_constant_p(v) || \
 278                     ((v) != NULL)) \
 279                         wmb(); \
 280                 xchg(p, v); \
 281         })
 282
 283 extern void synchronize_rcu(void);
 284
 285 /*
 286  * Exchanges the pointer and waits for quiescent state.
 287  * The pointer returned can be freed.
 288  */
 289 #define urcu_publish_content(p, v) \
 290         ({ \
 291                 void *oldptr; \
 292                 debug_yield_write(); \
 293                 oldptr = rcu_xchg_pointer(p, v); \
 294                 synchronize_rcu(); \
 295                 oldptr; \
 296         })
 297
 298 /*
 299  * Reader thread registration.
 300  */
 301 extern void urcu_register_thread(void);
 302 extern void urcu_unregister_thread(void);
 303
 304 #endif /* _URCU_H */