>>> On 14.03.16 at 08:16, <quan...@intel.com> wrote: > We confirmed with VT-d hardware team that 1ms is large enough for IOMMU > internal flush. So We can change the DMAR_OPERATION_TIMEOUT from 1000 ms > to 1 ms.
First of all the description mentions just one of the two uses of this constant. And then I'm hesitant to commit this as long as the panic()-s are still there when that timeout expires - reducing the timeout (the more this significantly) will raise the (theoretical) chance of actually invoking one of them. Hence the proper order in which I would like to see things happen is for this to be reduced only after having - eliminated the panic()-s - decoupled the device IOTLB invalidation from this path (after all right now flush_iotlb_qi() uses invalidate_sync() for both, and hence the shortened timeout now pretty certainly is too short for the device IOTLB case) Temporarily it may be sufficient to just eliminate the panic()-s, provided the timeout value would then be made dependent on ats_enabled. Jan _______________________________________________ Xen-devel mailing list Xen-devel@lists.xen.org http://lists.xen.org/xen-devel
