On 30/10/2020 12:00, Jan Beulich wrote:
On 30.10.2020 11:57, Julien Grall wrote:
On 20/10/2020 15:10, Jan Beulich wrote:
--- a/xen/common/event_channel.c
+++ b/xen/common/event_channel.c
@@ -449,6 +449,13 @@ int evtchn_bind_virq(evtchn_bind_virq_t
spin_unlock_irqrestore(&chn->lock, flags);
+ /*
+ * If by any, the update of virq_to_evtchn[] would need guarding by
+ * virq_lock, but since this is the last action here, there's no strict
+ * need to acquire the lock. Hnece holding event_lock isn't helpful
s/Hnece/Hence/
+ * anymore at this point, but utilize that its unlocking acts as the
+ * otherwise necessary smp_wmb() here.
+ */
v->virq_to_evtchn[virq] = bind->port = port;
I think all access to v->virq_to_evtchn[virq] should use ACCESS_ONCE()
or {read, write}_atomic() to avoid any store/load tearing.
IOW you're suggesting this to be the subject of a separate patch?
I don't think such a conversion belongs here (nor even in this
series, seeing the much wider applicability of such a change
throughout the code base).
Or are you seeing anything here which
would require such a conversion to be done as a prereq?
Yes, your comment implies that it is fine to write to virq_to_evtchn[]
without the lock taken. However, this is *only* valid if the compiler
doesn't tear down load/store.
So this is a pre-req to get your comment valid.
Cheers,
--
Julien Grall
