On Thu, Oct 31, 2013 at 08:49:08AM -0600, Eric Blake wrote:
On 10/31/2013 08:47 AM, Michael S. Tsirkin wrote:
if (event & PVPANIC_PANICKED) { panicked_mon_event("pause"); - vm_stop(RUN_STATE_GUEST_PANICKED);
Don't you still need to halt the guest on a panic event, for management to have a chance to choose what to do about the panic?
Guest can just call hlt to do this. Most guests do this on a panic already.
On the one hand, the fact that the guest already has to inform the host means we are already trusting the guest behavior on a panic. On the other hand, assuming that the guest will ALWAYS halt after triggering a panic is putting a lot more trust in the guest, compared to qemu explicitly halting the guest so that management has a chance to choose to dump the guest's state at the moment the panic was flagged.
I wouldn't call it *a lot* more trust. And again, this is guest policy: if you want to do hlt from driver because you think it's safer, go for it.
The biggest argument for either removing all auto-pvpanic, or reverting pvpanic altogether, is that no one seems to be actively using pvpanic in the field yet. I wish we could get more feedback from Fujitsu as the original patch authors on what they are looking for in a working solution, rather than repeatedly second-guessing everything downstream and delaying the eradication of the buggy behavior even longer.
With my patch we have a benign device that merely reports io writes on the monitor. No code -> no bugs.
Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org