#ifndef _URCU_H
#define _URCU_H
+/*
+ * urcu.h
+ *
+ * Userspace RCU header
+ *
+ * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ *
+ * Credits for Paul e. McKenney <paulmck@linux.vnet.ibm.com>
+ * for inspiration coming from the Linux kernel RCU and rcu-preempt.
+ *
+ * The barrier, mb, rmb, wmb, atomic_inc, smp_read_barrier_depends, ACCESS_ONCE
+ * and rcu_dereference primitives come from the Linux kernel.
+ *
+ * Distributed under GPLv2
+ */
+
+#include <stdlib.h>
+#include <pthread.h>
+
/* The "volatile" is due to gcc bugs */
#define barrier() __asm__ __volatile__("": : :"memory")
-/* x86 32/64 specific */
-#define mb() asm volatile("mfence":::"memory")
-#define rmb() asm volatile("lfence":::"memory")
-#define wmb() asm volatile("sfence" ::: "memory")
+#define likely(x) __builtin_expect(!!(x), 1)
+#define unlikely(x) __builtin_expect(!!(x), 0)
+/* Assume SMP machine, given we don't have this information */
+#define CONFIG_SMP 1
-/* x86 32 */
-static inline void atomic_inc(int *v)
-{
- asm volatile("lock; incl %0"
- : "+m" (*v));
-}
+#ifdef CONFIG_SMP
+#define smp_mb() mb()
+#define smp_rmb() rmb()
+#define smp_wmb() wmb()
+#define smp_mc() mc()
+#define smp_rmc() rmc()
+#define smp_wmc() wmc()
+#else
+#define smp_mb() barrier()
+#define smp_rmb() barrier()
+#define smp_wmb() barrier()
+#define smp_mc() barrier()
+#define smp_rmc() barrier()
+#define smp_wmc() barrier()
+#endif
+
+#include "arch.h"
/* Nop everywhere except on alpha. */
#define smp_read_barrier_depends()
*/
#define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
+/*
+ * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
+ */
+#define _LOAD_SHARED(p) ACCESS_ONCE(p)
+
+/*
+ * Load a data from shared memory, doing a cache flush if required.
+ */
+#define LOAD_SHARED(p) \
+ ({ \
+ smp_rmc(); \
+ _LOAD_SHARED(p); \
+ })
+
+
+/*
+ * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
+ */
+#define _STORE_SHARED(x, v) \
+ do { \
+ (x) = (v); \
+ } while (0)
+
+/*
+ * Store v into x, where x is located in shared memory. Performs the required
+ * cache flush after writing.
+ */
+#define STORE_SHARED(x, v) \
+ do { \
+ _STORE_SHARED(x, v); \
+ smp_wmc(); \
+ } while (0)
+
/**
* rcu_dereference - fetch an RCU-protected pointer in an
* RCU read-side critical section. This pointer may later
*/
#define rcu_dereference(p) ({ \
- typeof(p) _________p1 = ACCESS_ONCE(p); \
+ typeof(p) _________p1 = LOAD_SHARED(p); \
smp_read_barrier_depends(); \
(_________p1); \
})
#define SIGURCU SIGUSR1
-/* Global quiescent period parity */
-extern int urcu_qparity;
+/*
+ * If a reader is really non-cooperative and refuses to commit its
+ * urcu_active_readers count to memory (there is no barrier in the reader
+ * per-se), kick it after a few loops waiting for it.
+ */
+#define KICK_READER_LOOPS 10000
+
+#ifdef DEBUG_YIELD
+#include <sched.h>
+#include <time.h>
+#include <pthread.h>
+#include <unistd.h>
+
+#define YIELD_READ (1 << 0)
+#define YIELD_WRITE (1 << 1)
+
+/* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */
+#ifdef DEBUG_FULL_MB
+/* maximum sleep delay, in us */
+#define MAX_SLEEP 50
+#else
+#define MAX_SLEEP 30000
+#endif
-extern int __thread urcu_active_readers[2];
+extern unsigned int yield_active;
+extern unsigned int __thread rand_yield;
-static inline int get_urcu_qparity(void)
+static inline void debug_yield_read(void)
{
- return urcu_qparity;
+ if (yield_active & YIELD_READ)
+ if (rand_r(&rand_yield) & 0x1)
+ usleep(rand_r(&rand_yield) % MAX_SLEEP);
}
+static inline void debug_yield_write(void)
+{
+ if (yield_active & YIELD_WRITE)
+ if (rand_r(&rand_yield) & 0x1)
+ usleep(rand_r(&rand_yield) % MAX_SLEEP);
+}
+
+static inline void debug_yield_init(void)
+{
+ rand_yield = time(NULL) ^ pthread_self();
+}
+#else
+static inline void debug_yield_read(void)
+{
+}
+
+static inline void debug_yield_write(void)
+{
+}
+
+static inline void debug_yield_init(void)
+{
+
+}
+#endif
+
+#ifdef DEBUG_FULL_MB
+static inline void reader_barrier()
+{
+ smp_mb();
+}
+#else
+static inline void reader_barrier()
+{
+ barrier();
+}
+#endif
+
+/*
+ * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
+ * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
+ */
+#define RCU_GP_COUNT (1UL << 0)
+/* Use the amount of bits equal to half of the architecture long size */
+#define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2))
+#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
+
/*
- * returns urcu_parity.
+ * Global quiescent period counter with low-order bits unused.
+ * Using a int rather than a char to eliminate false register dependencies
+ * causing stalls on some architectures.
*/
-static inline int rcu_read_lock(void)
+extern long urcu_gp_ctr;
+
+extern long __thread urcu_active_readers;
+
+static inline int rcu_old_gp_ongoing(long *value)
+{
+ long v;
+
+ if (value == NULL)
+ return 0;
+ /*
+ * Make sure both tests below are done on the same version of *value
+ * to insure consistency.
+ */
+ v = LOAD_SHARED(*value);
+ return (v & RCU_GP_CTR_NEST_MASK) &&
+ ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
+}
+
+static inline void rcu_read_lock(void)
{
- int urcu_parity = get_urcu_qparity();
- urcu_active_readers[urcu_parity]++;
+ long tmp;
+
+ tmp = urcu_active_readers;
+ /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
+ /*
+ * The data dependency "read urcu_gp_ctr, write urcu_active_readers",
+ * serializes those two memory operations. The memory barrier in the
+ * signal handler ensures we receive the proper memory commit barriers
+ * required by _STORE_SHARED and _LOAD_SHARED whenever communication
+ * with the writer is needed.
+ */
+ if (likely(!(tmp & RCU_GP_CTR_NEST_MASK)))
+ _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr));
+ else
+ _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT);
/*
* Increment active readers count before accessing the pointer.
* See force_mb_all_threads().
*/
- barrier();
- return urcu_parity;
+ reader_barrier();
}
-static inline void rcu_read_unlock(int urcu_parity)
+static inline void rcu_read_unlock(void)
{
- barrier();
+ reader_barrier();
/*
* Finish using rcu before decrementing the pointer.
* See force_mb_all_threads().
*/
- urcu_active_readers[urcu_parity]--;
+ _STORE_SHARED(urcu_active_readers, urcu_active_readers - RCU_GP_COUNT);
}
-extern void rcu_write_lock(void);
-extern void rcu_write_unlock(void);
+/**
+ * rcu_assign_pointer - assign (publicize) a pointer to a newly
+ * initialized structure that will be dereferenced by RCU read-side
+ * critical sections. Returns the value assigned.
+ *
+ * Inserts memory barriers on architectures that require them
+ * (pretty much all of them other than x86), and also prevents
+ * the compiler from reordering the code that initializes the
+ * structure after the pointer assignment. More importantly, this
+ * call documents which pointers will be dereferenced by RCU read-side
+ * code.
+ */
+
+#define rcu_assign_pointer(p, v) \
+ ({ \
+ if (!__builtin_constant_p(v) || \
+ ((v) != NULL)) \
+ wmb(); \
+ STORE_SHARED(p, v); \
+ })
+
+#define rcu_xchg_pointer(p, v) \
+ ({ \
+ if (!__builtin_constant_p(v) || \
+ ((v) != NULL)) \
+ wmb(); \
+ xchg(p, v); \
+ })
-extern void *urcu_publish_content(void **ptr, void *new);
+extern void synchronize_rcu(void);
+
+/*
+ * Exchanges the pointer and waits for quiescent state.
+ * The pointer returned can be freed.
+ */
+#define urcu_publish_content(p, v) \
+ ({ \
+ void *oldptr; \
+ oldptr = rcu_xchg_pointer(p, v); \
+ synchronize_rcu(); \
+ oldptr; \
+ })
/*
* Reader thread registration.
*/
extern void urcu_register_thread(void);
-extern void urcu_register_thread(void);
+extern void urcu_unregister_thread(void);
#endif /* _URCU_H */