This patch fixes a lost wake-up bug in the futex-based serial lock
implementation. The previous code lacked some pieces of synchronization
(a seq_cst barrier and resetting a flag that is checked by the
respective futex_wait()), which led to those lost wake-ups when waking
and waiting were racing against each other.
Tested on ppc64 with microbenchmarks/STAMP and gl_wt.
OK for trunk?
commit a04fadca7c74f35d6d1d4f12d745c0ab44cc315a
Author: Torvald Riegel <trie...@redhat.com>
Date: Tue Jan 24 15:05:42 2012 +0100
libitm: Fix wake-up of readers in futex-based serial lock.
libitm/
* config/linux/rwlock.cc (GTM::gtm_rwlock::write_unlock): Fix reader
wake-up.
diff --git a/libitm/config/linux/rwlock.cc b/libitm/config/linux/rwlock.cc
index 24e7042..ad1b042 100644
--- a/libitm/config/linux/rwlock.cc
+++ b/libitm/config/linux/rwlock.cc
@@ -231,10 +231,13 @@ gtm_rwlock::write_unlock ()
// last writer (this can happen because write_lock_generic()
// exchanges 0 or 1 to 2 and thus might go to contended mode even if
// no other thread holds the write lock currently). Therefore, we
- // have to wake up readers here as well.
- futex_wake(&readers, INT_MAX);
+ // have to wake up readers here as well. Execute a barrier after
+ // the previous relaxed reset of writers (Dekker-style), and fall
+ // through to the normal reader wake-up code.
+ atomic_thread_fence (memory_order_seq_cst);
}
- return;
+ else
+ return;
}
// No waiting writers, so wake up all waiting readers.
// Because the fetch_and_sub is a full barrier already, we don't need
