qsbr: only mark reader thread as being waited for in contended case

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

diff --git a/urcu-qsbr-static.h b/urcu-qsbr-static.h
index 87305cb694a9bcc4e175f37829736fcdb81d9686..f1300f320cce2f0605f149a806ebe1e96c5c1464 100644 (file)
--- a/urcu-qsbr-static.h
+++ b/urcu-qsbr-static.h
@@ -33,8 +33,8 @@
 #include <pthread.h>
 #include <assert.h>
 #include <limits.h>
-#include <syscall.h>
-#include <unistd.h>
+#include <sched.h>
+#include <sys/time.h>
 
 #include <compiler.h>
 #include <arch.h>
@@ -89,10 +89,6 @@
                                (_________p1);                          \
                                })
 
-#define futex(...)             syscall(__NR_futex, __VA_ARGS__)
-#define FUTEX_WAIT             0
-#define FUTEX_WAKE             1
-
 /*
  * 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
@@ -103,13 +99,13 @@
 
 /*
  * 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 futex().
+ * Active attempts to check for reader Q.S. before calling sched_yield().
  */
 #define RCU_QS_ACTIVE_ATTEMPTS 100
 
@@ -173,7 +169,8 @@ static inline void reader_barrier()
 }
 
 #define RCU_GP_ONLINE          (1UL << 0)
-#define RCU_GP_CTR             (1UL << 1)
+#define RCU_GP_ONGOING         (1UL << 1)
+#define RCU_GP_CTR             (1UL << 2)
 
 /*
  * Global quiescent period counter with low-order bits unused.
@@ -182,21 +179,14 @@ static inline void reader_barrier()
  */
 extern unsigned long urcu_gp_ctr;
 
-extern unsigned long __thread rcu_reader_qs_gp;
+struct urcu_reader_status {
+       unsigned long qs_gp;
+       unsigned long gp_waiting;
+       unsigned long qs_time_delta_usec;
+       struct timeval qs_time_last;
+};
 
-extern int gp_futex;
-
-/*
- * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
- */
-static inline void wake_up_gp(void)
-{
-       if (unlikely(atomic_read(&gp_futex) == -1)) {
-               atomic_set(&gp_futex, 0);
-               futex(&gp_futex, FUTEX_WAKE, 1,
-                     NULL, NULL, 0);
-       }
-}
+extern struct urcu_reader_status __thread urcu_reader_status;
 
 #if (BITS_PER_LONG < 64)
 static inline int rcu_gp_ongoing(unsigned long *value)
@@ -222,7 +212,7 @@ static inline int rcu_gp_ongoing(unsigned long *value)
 
 static inline void _rcu_read_lock(void)
 {
-       rcu_assert(rcu_reader_qs_gp);
+       rcu_assert(urcu_reader_status.qs_gp);
 }
 
 static inline void _rcu_read_unlock(void)
@@ -231,24 +221,56 @@ static inline void _rcu_read_unlock(void)
 
 static inline void _rcu_quiescent_state(void)
 {
-       smp_mb();       
-       _STORE_SHARED(rcu_reader_qs_gp, _LOAD_SHARED(urcu_gp_ctr));
-       smp_mb();       /* write rcu_reader_qs_gp before read futex */
-       wake_up_gp();
+       long gp_ctr;
+       struct timeval current_time, delta_time;
+
+       smp_mb();
+       gettimeofday(&current_time, NULL);
+       timersub(&current_time, &urcu_reader_status.qs_time_last, &delta_time);
+       if (delta_time.tv_sec >= 1)
+               _STORE_SHARED(urcu_reader_status.qs_time_delta_usec, 1000000);
+       else
+               _STORE_SHARED(urcu_reader_status.qs_time_delta_usec,
+                     (unsigned long)delta_time.tv_usec);
+       urcu_reader_status.qs_time_last = current_time;
+       /*
+        * 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();
-       STORE_SHARED(rcu_reader_qs_gp, 0);
-       smp_mb();       /* write rcu_reader_qs_gp before read futex */
-       wake_up_gp();
+       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)
 {
-       _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
+       struct timeval current_time, delta_time;
+
+       gettimeofday(&current_time, NULL);
+       timersub(&current_time, &urcu_reader_status.qs_time_last, &delta_time);
+       if (delta_time.tv_sec >= 1)
+               _STORE_SHARED(urcu_reader_status.qs_time_delta_usec, 1000000);
+       else
+               _STORE_SHARED(urcu_reader_status.qs_time_delta_usec,
+                     (unsigned long)delta_time.tv_usec);
+       urcu_reader_status.qs_time_last = current_time;
+       _STORE_SHARED(urcu_reader_status.qs_time_delta_usec, 0);
+       _STORE_SHARED(urcu_reader_status.qs_gp, LOAD_SHARED(urcu_gp_ctr));
        smp_mb();
 }