On 09/15/2017 01:53 PM, Paolo Bonzini wrote:
> On 13/09/2017 22:08, Davidlohr Bueso wrote:
>> The following patches fix and/or justify (in baby steps) some of the
>> callers. The main exception is s390, which I didn't follow how ->valid_wakeup
>> can get hoisted as kvm_vcpu_block does not use that in the wait loop.
>
> valid_wakeup is just an optimization, so it's not a problem.
>
> There seems to be always an atomic_or or set_bit before
> kvm_s390_vcpu_wakeup is called (except kvm_s390_idle_wakeup which has no
> store at all and doesn't need any serialization). So my suggestion is
> to add an smp__mb_after_atomic in kvm_s390_vcpu_wakeup; I'll let the
> s390 guys do it.
I will queue something like this
diff --git a/arch/s390/kvm/interrupt.c b/arch/s390/kvm/interrupt.c
index a832ad0..44239b5 100644
--- a/arch/s390/kvm/interrupt.c
+++ b/arch/s390/kvm/interrupt.c
@@ -1074,6 +1074,12 @@ void kvm_s390_vcpu_wakeup(struct kvm_vcpu *vcpu)
* in kvm_vcpu_block without having the waitqueue set (polling)
*/
vcpu->valid_wakeup = true;
+ /*
+ * This is mostly to document, that the read in swait_active could
+ * be moved before other stores, leading to subtle races.
+ * All current users do not store or use an atomic like update
+ */
+ __smp_mb__after_atomic();
if (swait_active(&vcpu->wq)) {
/*
* The vcpu gave up the cpu voluntarily, mark it as a good
but I am asking myself if it is "safer" to make this function use swq_has_sleepers
in case we add in a distant future another user to kvm_s390_vcpu_wakeup that does
use a normal store and everybody has already forgotten this?
