From: Jeremy Fitzhardinge <jeremy.fitzhardinge@xxxxxxxxxx>
If we don't put a real memory barrier between the unlocking increment
of the queue head and the check for lock waiters, we can end up with a
deadlock as a result of the unload write being reordered with respect
to the waiters read. In effect, the check on the waiters count could
happen before the unlock, which leads to a deadlock:
unlocker locker
check waiters (none)
check lock head (timeout)
add to waiters
block
release lock => deadlock
Adding a mb() barrier prevents the unlocker's waiter check from happening
before the lock release.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@xxxxxxxxxx>
---
arch/x86/include/asm/spinlock.h | 7 +++++++
1 files changed, 7 insertions(+), 0 deletions(-)
diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/asm/spinlock.h
index 3deabca..b9a1aae 100644
--- a/arch/x86/include/asm/spinlock.h
+++ b/arch/x86/include/asm/spinlock.h
@@ -88,6 +88,13 @@ static inline void __ticket_sub_waiting(struct arch_spinlock *lock)
static __always_inline bool __ticket_lock_waiters(const struct arch_spinlock *lock)
{
+ /*
+ * Make sure everyone sees the unlock write before we check the
+ * waiting count. The processor ordering doesn't guarantee this
+ * because they're different memory locations.
+ */
+ mb();
+
return ACCESS_ONCE(lock->waiting) != 0;
}
#endif /* CONFIG_PARAVIRT_SPINLOCKS */
--
1.7.2.3
_______________________________________________
Virtualization mailing list
Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/virtualization
