On Tue, Mar 02, 2021 at 09:57:27PM -0800, Raj, Ashok wrote: > > + ret = handle_mm_fault(vma, prm->addr, fault_flags, NULL); > > Should we add a trace similar to trace_page_fault_user() or kernel in > arch/x86/kernel/mm/fault.c Yes that would definitely be useful for debugging hardware and developping applications. I can send a separate patch to add tracepoints here and in the lower-level device fault path. > or maybe add a perf_sw_event() for device faults? It does seem like that would fit well alongside the existing PERF_COUNT_SW_PAGE_FAULTS, but I don't think it would be useful in practice, because we can't provide a context for the event. Since we're handling these faults remotely, the only way for a user to get IOPF events is to enable them on all CPUs and all tasks. Tracepoints can have 'pasid' and 'device' fields to identify an event, but the perf_sw_event wouldn't have any useful metadata apart from the faulting address. We could also add tracepoints on bind(), so users can get the PASID obtained with the PID they care about and filter fault events based on that. I've been wondering about associating a PASID with a PID internally, because we don't currently have anywhere to send SEGV signals for unhandled page faults. But I think it would be best to notify the device driver on unhandled fault and let them deal with it. They'll probably want that information anyway. Thanks, Jean
