*
* Userspace RCU header.
*
- * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking
- * dynamically with the userspace rcu library.
+ * TO BE INCLUDED ONLY IN CODE THAT IS TO BE RECOMPILED ON EACH LIBURCU
+ * RELEASE. See urcu.h for linking dynamically with the userspace rcu library.
*
* Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
* Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
#include <stdlib.h>
#include <pthread.h>
-#include <syscall.h>
#include <unistd.h>
#include <urcu/compiler.h>
#include <urcu/system.h>
#include <urcu/uatomic.h>
#include <urcu/list.h>
+#include <urcu/tls-compat.h>
+#include <urcu/debug.h>
/*
* This code section can only be included in LGPL 2.1 compatible source code.
extern "C" {
#endif
-/*
- * Active attempts to check for reader Q.S. before calling sleep().
- */
-#define RCU_QS_ACTIVE_ATTEMPTS 100
-
-#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 YIELD_READ (1 << 0)
-#define YIELD_WRITE (1 << 1)
-
-/*
- * Updates without RCU_MB are much slower. Account this in
- * the delay.
- */
-/* maximum sleep delay, in us */
-#define MAX_SLEEP 50
-
-extern unsigned int yield_active;
-extern unsigned int __thread rand_yield;
-
-static inline void debug_yield_read(void)
-{
- 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
+enum rcu_state {
+ RCU_READER_ACTIVE_CURRENT,
+ RCU_READER_ACTIVE_OLD,
+ RCU_READER_INACTIVE,
+};
/*
* The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use a
*/
extern void rcu_bp_register(void);
-/*
- * 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 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;
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
+extern struct rcu_gp rcu_gp;
struct rcu_reader {
/* Data used by both reader and synchronize_rcu() */
- long ctr;
+ unsigned long ctr;
/* Data used for registry */
struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
pthread_t tid;
* Adds a pointer dereference on the read-side, but won't require to unregister
* the reader thread.
*/
-extern struct rcu_reader __thread *rcu_reader;
+extern DECLARE_URCU_TLS(struct rcu_reader *, rcu_reader);
-static inline int rcu_old_gp_ongoing(long *value)
+extern int urcu_bp_has_sys_membarrier;
+
+static inline void urcu_bp_smp_mb_slave(void)
{
- long v;
+ if (caa_likely(urcu_bp_has_sys_membarrier))
+ cmm_barrier();
+ else
+ cmm_smp_mb();
+}
+
+static inline enum rcu_state rcu_reader_state(unsigned long *ctr)
+{
+ unsigned long v;
- if (value == NULL)
- return 0;
+ if (ctr == NULL)
+ return RCU_READER_INACTIVE;
/*
* Make sure both tests below are done on the same version of *value
* to insure consistency.
*/
- v = CMM_LOAD_SHARED(*value);
- return (v & RCU_GP_CTR_NEST_MASK) &&
- ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
+ 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))
+ return RCU_READER_ACTIVE_CURRENT;
+ return RCU_READER_ACTIVE_OLD;
}
-static inline void _rcu_read_lock(void)
+/*
+ * 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
+ * _rcu_read_lock() happen before the subsequent read-side critical section.
+ */
+static inline void _rcu_read_lock_update(unsigned long tmp)
{
- 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));
+ urcu_bp_smp_mb_slave();
+ } else
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp + RCU_GP_COUNT);
+}
- /* Check if registered */
- if (unlikely(!rcu_reader))
- rcu_bp_register();
+/*
+ * Enter an RCU read-side critical section.
+ *
+ * The first cmm_barrier() call ensures that the compiler does not reorder
+ * the body of _rcu_read_lock() with a mutex.
+ *
+ * This function and its helper are both less than 10 lines long. The
+ * intent is that this function meets the 10-line criterion in LGPL,
+ * allowing this function to be invoked directly from non-LGPL code.
+ */
+static inline void _rcu_read_lock(void)
+{
+ unsigned long tmp;
+ if (caa_unlikely(!URCU_TLS(rcu_reader)))
+ rcu_bp_register(); /* If not yet registered. */
cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
- tmp = rcu_reader->ctr;
- /*
- * rcu_gp_ctr is
- * RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE)
- */
- if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
- _CMM_STORE_SHARED(rcu_reader->ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
- /*
- * Set active readers count for outermost nesting level before
- * accessing the pointer.
- */
- cmm_smp_mb();
- } else {
- _CMM_STORE_SHARED(rcu_reader->ctr, tmp + RCU_GP_COUNT);
- }
+ tmp = URCU_TLS(rcu_reader)->ctr;
+ urcu_assert((tmp & RCU_GP_CTR_NEST_MASK) != RCU_GP_CTR_NEST_MASK);
+ _rcu_read_lock_update(tmp);
}
+/*
+ * Exit an RCU read-side critical section. This function is less than
+ * 10 lines of code, and is intended to be usable by non-LGPL code, as
+ * called out in LGPL.
+ */
static inline void _rcu_read_unlock(void)
{
- /*
- * Finish using rcu before decrementing the pointer.
- */
- cmm_smp_mb();
- _CMM_STORE_SHARED(rcu_reader->ctr, rcu_reader->ctr - RCU_GP_COUNT);
+ unsigned long tmp;
+
+ tmp = URCU_TLS(rcu_reader)->ctr;
+ urcu_assert(tmp & RCU_GP_CTR_NEST_MASK);
+ /* Finish using rcu before decrementing the pointer. */
+ urcu_bp_smp_mb_slave();
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp - RCU_GP_COUNT);
cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
}
-#ifdef __cplusplus
+/*
+ * Returns whether within a RCU read-side critical section.
+ *
+ * This function is less than 10 lines long. The intent is that this
+ * function meets the 10-line criterion for LGPL, allowing this function
+ * to be invoked directly from non-LGPL code.
+ */
+static inline int _rcu_read_ongoing(void)
+{
+ if (caa_unlikely(!URCU_TLS(rcu_reader)))
+ rcu_bp_register(); /* If not yet registered. */
+ return URCU_TLS(rcu_reader)->ctr & RCU_GP_CTR_NEST_MASK;
+}
+
+#ifdef __cplusplus
}
#endif