Re: [PATCH v7 2/2] tracing: Add support for preempt and irq enable/disable events

From: Joel Fernandes
Date: Fri Oct 06 2017 - 03:28:28 EST

Hi Peter,

On Fri, Oct 6, 2017 at 12:07 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Thu, Oct 05, 2017 at 04:28:10PM -0700, Joel Fernandes wrote:
>> > lockdep implements the trace_hardirq_*() in terms of *_caller(). Would
>> > that make sense here?
>> In lockdep code, when trace_hardirqs_off is called,
>> trace_hardirqs_off_caller would pass CALLER_ADDR0 as
>> trace_hardirqs_off.
>> Because of this, the first argument passed to time_hardirqs_off would
>> always be an offset within trace_hardirqs_off:
>> time_hardirqs_off(CALLER_ADDR0, ip);
>> Is that intended? Seems to me that in the lockdep implementation of
>> trace_hardirqs_* in terms of *_caller(), we would completely miss the
>> second-last return address (CALLER_ADDR1) of trace_hardirqs_off().
>> Also for the above reasons, I don't think it doesn't make sense to use
>> this reuse logic for the tracer. Atleast I feel it might change the
>> current behavior of the preempt/irqsoff tracer which I don't intend to
>> change with my current patch set.
> Hurm.. so I've no clue. I never looked at any of this. I think Ingo
> wrote that before I came joined the fun. He might (although unlikely,
> because its been _many_ years) still have some memories.
> But given that lockdep enabled kernels never got those arguments right,
> how useful are they? I mean, nobody seems to have noticed much.

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.

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.


- Joel