xen/x86: don't lose event interrupts
authorStefano Stabellini <sstabellini@kernel.org>
Sat, 16 Apr 2016 01:23:00 +0000 (18:23 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 3 May 2017 04:19:50 +0000 (21:19 -0700)
commit c06b6d70feb32d28f04ba37aa3df17973fd37b6b upstream.

On slow platforms with unreliable TSC, such as QEMU emulated machines,
it is possible for the kernel to request the next event in the past. In
that case, in the current implementation of xen_vcpuop_clockevent, we
simply return -ETIME. To be precise the Xen returns -ETIME and we pass
it on. However the result of this is a missed event, which simply causes
the kernel to hang.

Instead it is better to always ask the hypervisor for a timer event,
even if the timeout is in the past. That way there are no lost
interrupts and the kernel survives. To do that, remove the
VCPU_SSHOTTMR_future flag.

Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Acked-by: Juergen Gross <jgross@suse.com>
Cc: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/x86/xen/time.c

index f1ba6a092854c68017a49f9a4d6fd7c1106bde7f..8846257d8792ef5c19c463fbf5435625edeab6d5 100644 (file)
@@ -343,11 +343,11 @@ static int xen_vcpuop_set_next_event(unsigned long delta,
        WARN_ON(!clockevent_state_oneshot(evt));
 
        single.timeout_abs_ns = get_abs_timeout(delta);
-       single.flags = VCPU_SSHOTTMR_future;
+       /* Get an event anyway, even if the timeout is already expired */
+       single.flags = 0;
 
        ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
-
-       BUG_ON(ret != 0 && ret != -ETIME);
+       BUG_ON(ret != 0);
 
        return ret;
 }