On Fri, 6 Oct 2017 00:28:21 -0700 Joel Fernandes <joe...@google.com> wrote:
> Oh ok. So I traced this down to the original patch that added > time_hardirqs_off to lockdep. I *think* it was added just to keep the > irqsoff tracer working while lockdep is enabled, so that both lockdep > and the tracer would work at the same time. So I guess nobody noticed > any issues because nobody noticed or cared to check if irqsoff tracer > showed the same results with lockdep enabled vs disabled. As one of the main users of preemptirqsoff tracer, I can tell you that we didn't care ;-) Yes, it was added to keep them both working. It's dangerous to use CALLER_ADDR1, as lockdep can be called by the top frame, and ADDR1 would go off the stack frame and cause a fault. I just put in ADDR0 as filler as it required something to be added. We added a stack dump on new max latencies to figure out where the problem occurred. -- Steve > > I think the correct way this should be rewritten is lockdep should use > register_trace_* to register callbacks onto the new tracepoints I'm > adding. I am working on a patch that does something like this and gets > rid of the time_* functions. I will try to have an RFC out by the > weekend if I hopefully don't hit any roadblocks. > > thanks! > > - Joel
