#include <urcu/list.h>
#include <urcu/futex.h>
#include <urcu/tls-compat.h>
+#include <urcu/rand-compat.h>
+#include <urcu/debug.h>
#ifdef __cplusplus
extern "C" {
#define RCU_MEMBARRIER
#endif
-/*
- * RCU_MEMBARRIER is only possibly available on Linux. Fallback to RCU_MB
- * otherwise.
- */
-#if !defined(__linux__) && defined(RCU_MEMBARRIER)
-#undef RCU_MEMBARRIER
-#define RCU_MB
-#endif
-
-#ifdef RCU_MEMBARRIER
-#include <syscall.h>
-
-/* If the headers do not support SYS_membarrier, statically use RCU_MB */
-#ifdef SYS_membarrier
-# define MEMBARRIER_EXPEDITED (1 << 0)
-# define MEMBARRIER_DELAYED (1 << 1)
-# define MEMBARRIER_QUERY (1 << 16)
-# define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__)
-#else
-# undef RCU_MEMBARRIER
-# define RCU_MB
-#endif
-#endif
-
/*
* This code section can only be included in LGPL 2.1 compatible source code.
* See below for the function call wrappers which can be used in code meant to
RCU_READER_INACTIVE,
};
-#ifdef DEBUG_RCU
-#define rcu_assert(args...) assert(args)
-#else
-#define rcu_assert(args...)
-#endif
-
-#ifdef DEBUG_YIELD
-#include <sched.h>
-#include <time.h>
-#include <pthread.h>
-#include <unistd.h>
-
-#define RCU_YIELD_READ (1 << 0)
-#define RCU_YIELD_WRITE (1 << 1)
-
/*
- * Updates with RCU_SIGNAL are much slower. Account this in the delay.
- */
-#ifdef RCU_SIGNAL
-/* maximum sleep delay, in us */
-#define MAX_SLEEP 30000
-#else
-#define MAX_SLEEP 50
-#endif
-
-extern unsigned int rcu_yield_active;
-extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield);
-
-static inline void rcu_debug_yield_read(void)
-{
- if (rcu_yield_active & RCU_YIELD_READ)
- if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
- usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
-}
-
-static inline void rcu_debug_yield_write(void)
-{
- if (rcu_yield_active & RCU_YIELD_WRITE)
- if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1)
- usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP);
-}
-
-static inline void rcu_debug_yield_init(void)
-{
- URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self();
-}
-#else
-static inline void rcu_debug_yield_read(void)
-{
-}
-
-static inline void rcu_debug_yield_write(void)
-{
-}
-
-static inline void rcu_debug_yield_init(void)
-{
-
-}
-#endif
-
-/*
- * RCU memory barrier broadcast group. Currently, only broadcast to all process
- * threads is supported (group 0).
- *
* Slave barriers are only guaranteed to be ordered wrt master barriers.
*
* The pair ordering is detailed as (O: ordered, X: not ordered) :
* master O O
*/
-#define MB_GROUP_ALL 0
-#define RCU_MB_GROUP MB_GROUP_ALL
-
#ifdef RCU_MEMBARRIER
extern int rcu_has_sys_membarrier;
-static inline void smp_mb_slave(int group)
+static inline void smp_mb_slave(void)
{
if (caa_likely(rcu_has_sys_membarrier))
cmm_barrier();
#endif
#ifdef RCU_MB
-static inline void smp_mb_slave(int group)
+static inline void smp_mb_slave(void)
{
cmm_smp_mb();
}
#endif
#ifdef RCU_SIGNAL
-static inline void smp_mb_slave(int group)
+static inline void smp_mb_slave(void)
{
cmm_barrier();
}
#define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2))
#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1)
-/*
- * 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.
- */
-extern unsigned long rcu_gp_ctr;
+struct rcu_gp {
+ /*
+ * Global grace period counter.
+ * Contains the current RCU_GP_CTR_PHASE.
+ * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
+ * Written to only by writer with mutex taken.
+ * Read by both writer and readers.
+ */
+ unsigned long ctr;
+
+ int32_t futex;
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
+extern struct rcu_gp rcu_gp;
struct rcu_reader {
/* Data used by both reader and synchronize_rcu() */
/* Data used for registry */
struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
pthread_t tid;
+ /* Reader registered flag, for internal checks. */
+ unsigned int registered:1;
};
extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader);
-extern int32_t rcu_gp_futex;
-
/*
* Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
*/
static inline void wake_up_gp(void)
{
- if (caa_unlikely(uatomic_read(&rcu_gp_futex) == -1)) {
- uatomic_set(&rcu_gp_futex, 0);
- futex_async(&rcu_gp_futex, FUTEX_WAKE, 1,
- NULL, NULL, 0);
+ if (caa_unlikely(uatomic_read(&rcu_gp.futex) == -1)) {
+ uatomic_set(&rcu_gp.futex, 0);
+ /*
+ * Ignoring return value until we can make this function
+ * return something (because urcu_die() is not publicly
+ * exposed).
+ */
+ (void) futex_async(&rcu_gp.futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0);
}
}
v = CMM_LOAD_SHARED(*ctr);
if (!(v & RCU_GP_CTR_NEST_MASK))
return RCU_READER_INACTIVE;
- if (!((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE))
+ if (!((v ^ rcu_gp.ctr) & RCU_GP_CTR_PHASE))
return RCU_READER_ACTIVE_CURRENT;
return RCU_READER_ACTIVE_OLD;
}
/*
- * Helper for _rcu_read_lock(). The format of rcu_gp_ctr (as well as
+ * Helper for _rcu_read_lock(). The format of rcu_gp.ctr (as well as
* the per-thread rcu_reader.ctr) has the upper bits containing a count of
* _rcu_read_lock() nesting, and a lower-order bit that contains either zero
* or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in
static inline void _rcu_read_lock_update(unsigned long tmp)
{
if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
- _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
- smp_mb_slave(RCU_MB_GROUP);
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp.ctr));
+ smp_mb_slave();
} else
_CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT);
}
{
unsigned long tmp;
+ urcu_assert(URCU_TLS(rcu_reader).registered);
cmm_barrier();
tmp = URCU_TLS(rcu_reader).ctr;
+ urcu_assert((tmp & RCU_GP_CTR_NEST_MASK) != RCU_GP_CTR_NEST_MASK);
_rcu_read_lock_update(tmp);
}
static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp)
{
if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
- smp_mb_slave(RCU_MB_GROUP);
- _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
- smp_mb_slave(RCU_MB_GROUP);
+ smp_mb_slave();
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp - RCU_GP_COUNT);
+ smp_mb_slave();
wake_up_gp();
} else
- _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp - RCU_GP_COUNT);
}
/*
{
unsigned long tmp;
+ urcu_assert(URCU_TLS(rcu_reader).registered);
tmp = URCU_TLS(rcu_reader).ctr;
+ urcu_assert(tmp & RCU_GP_CTR_NEST_MASK);
_rcu_read_unlock_update_and_wakeup(tmp);
cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
}
projects / urcu.git / blobdiff