urcu-qsbr: repeat sched_yield() loop, yield at offline rather than online.

[urcu.git] / urcu-qsbr-static.h

diff --git a/urcu-qsbr-static.h b/urcu-qsbr-static.h
index 86ae7d4aca91a6dd0ac998e5bf425b054a152716..9cc5db493ca1362d1781b44507a10e562623df1f 100644 (file)
--- a/urcu-qsbr-static.h
+++ b/urcu-qsbr-static.h
@@ -32,6 +32,8 @@
 #include <stdlib.h>
 #include <pthread.h>
 #include <assert.h>
+#include <limits.h>
+#include <sched.h>
 
 #include <compiler.h>
 #include <arch.h>
@@ -96,11 +98,16 @@
 
 /*
  * If a reader is really non-cooperative and refuses to commit its
- * rcu_reader_qs_gp count to memory (there is no barrier in the reader
+ * rcu_reader qs_gp 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
 
+/*
+ * Active attempts to check for reader Q.S. before calling sched_yield().
+ */
+#define RCU_QS_ACTIVE_ATTEMPTS 100
+
 #ifdef DEBUG_RCU
 #define rcu_assert(args...)    assert(args)
 #else
@@ -160,26 +167,49 @@ static inline void reader_barrier()
        smp_mb();
 }
 
+#define RCU_GP_ONLINE          (1UL << 0)
+#define RCU_GP_ONGOING         (1UL << 1)
+#define RCU_GP_CTR             (1UL << 2)
+
 /*
  * 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 urcu_gp_ctr;
+extern unsigned long urcu_gp_ctr;
 
-extern long __thread rcu_reader_qs_gp;
+struct urcu_reader_status {
+       unsigned long qs_gp;
+       unsigned long gp_waiting;
+};
 
-static inline int rcu_gp_ongoing(long *value)
+extern struct urcu_reader_status __thread urcu_reader_status;
+
+#if (BITS_PER_LONG < 64)
+static inline int rcu_gp_ongoing(unsigned long *value)
 {
+       unsigned long reader_gp;
+
        if (value == NULL)
                return 0;
+       reader_gp = LOAD_SHARED(*value);
+       return reader_gp && ((reader_gp ^ urcu_gp_ctr) & RCU_GP_CTR);
+}
+#else /* !(BITS_PER_LONG < 64) */
+static inline int rcu_gp_ongoing(unsigned long *value)
+{
+       unsigned long reader_gp;
 
-       return LOAD_SHARED(*value) & 1;
+       if (value == NULL)
+               return 0;
+       reader_gp = LOAD_SHARED(*value);
+       return reader_gp && (reader_gp - urcu_gp_ctr > ULONG_MAX / 2);
 }
+#endif  /* !(BITS_PER_LONG < 64) */
 
 static inline void _rcu_read_lock(void)
 {
-       rcu_assert(rcu_reader_qs_gp & 1);
+       rcu_assert(urcu_reader_status.qs_gp);
 }
 
 static inline void _rcu_read_unlock(void)
@@ -188,20 +218,36 @@ static inline void _rcu_read_unlock(void)
 
 static inline void _rcu_quiescent_state(void)
 {
-       smp_mb();       
-       rcu_reader_qs_gp = ACCESS_ONCE(urcu_gp_ctr) + 1;
+       long gp_ctr;
+
+       smp_mb();
+       /*
+        * volatile accesses can be reordered by the compiler when put in the
+        * same expression.
+        */
+       if (unlikely((gp_ctr = LOAD_SHARED(urcu_gp_ctr)) & RCU_GP_ONGOING) &&
+           unlikely(urcu_reader_status.gp_waiting)) {
+               _STORE_SHARED(urcu_reader_status.qs_gp, gp_ctr);
+               sched_yield();
+       } else {
+               _STORE_SHARED(urcu_reader_status.qs_gp, gp_ctr);
+       }
        smp_mb();
 }
 
 static inline void _rcu_thread_offline(void)
 {
        smp_mb();
-       rcu_reader_qs_gp = 0;
+       STORE_SHARED(urcu_reader_status.qs_gp, 0);
+       if (unlikely(LOAD_SHARED(urcu_gp_ctr) & RCU_GP_ONGOING) &&
+           unlikely(urcu_reader_status.gp_waiting)) {
+               sched_yield();
+       }
 }
 
 static inline void _rcu_thread_online(void)
 {
-       rcu_reader_qs_gp = ACCESS_ONCE(urcu_gp_ctr) + 1;
+       _STORE_SHARED(urcu_reader_status.qs_gp, LOAD_SHARED(urcu_gp_ctr));
        smp_mb();
 }
 
@@ -226,6 +272,22 @@ static inline void _rcu_thread_online(void)
                STORE_SHARED(p, v);                     \
        })
 
+/**
+ * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
+ * is as expected by "old". If succeeds, returns the previous pointer to the
+ * data structure, which can be safely freed after waiting for a quiescent state
+ * using synchronize_rcu(). If fails (unexpected value), returns old (which
+ * should not be freed !).
+ */
+
+#define _rcu_cmpxchg_pointer(p, old, _new)             \
+       ({                                              \
+               if (!__builtin_constant_p(_new) ||      \
+                   ((_new) != NULL))                   \
+                       wmb();                          \
+               cmpxchg(p, old, _new);                  \
+       })
+
 /**
  * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
  * pointer to the data structure, which can be safely freed after waiting for a