Re: [PATCH 2/2] x86, mce: Add persistent MCE event

[Borislav Petkov <bp@amd64.org>]

On Sat, Mar 24, 2012 at 10:15:01AM +0100, Ingo Molnar wrote:
> * Borislav Petkov <bp@amd64.org> wrote:
> 
> > On Sat, Mar 24, 2012 at 08:37:31AM +0100, Ingo Molnar wrote:
> > > I was mainly thinking of reducing this:
> > > 
> > >  arch/x86/kernel/cpu/mcheck/mce.c |   53 ++++++++++++++++++++++++++++++++++++++
> > >  1 file changed, 53 insertions(+)
> > > 
> > > to almost nothing. There doesn't seem to be much MCE specific in 
> > > that code, right?
> > 
> > Yeah, this could be generalized even more, AFAICT.
> > 
> > > 
> > > > Btw, the more important question is are we going to need 
> > > > persistent events that much so that a generic approach is 
> > > > warranted? I guess maybe the black box events recording deal 
> > > > would be another user..
> > > 
> > > So, here's the big picture as I see it:
> > > 
> > > I think tracing could use persistent events: mark all the events 
> > > we want to trace as persistent from bootup, and recover the 
> > > bootup trace after the system has been booted up.
> > 
> > Right, but (more nasty questions):
> > 
> > Why would I do this, am I tracing the boot process? [...]
> 
> Correct, in essence the MCE persistent event is partially about 
> that: we are starting to collect events well before there's any 
> user-space available.
> 
> > [...] If so, then I need another syntax which enables those 
> > events from the kernel command line which gets parsed the 
> > moment ftrace and ring buffer get initialized.
> 
> Correct. Something really simple like:
> 
>   boot_trace=<event1>,<event2>...
> 
> ... which could be all implicit within MCE too. (So I'm not 
> suggesting some boot command trigger to provide the MCE case - 
> but for more general boot tracing it would be the right 
> solution.)
> 
> > IOW, I'd need userspace for perf otherwise but I don't have 
> > that before booting...
> 
> Correct. In the case of MCE there's no "userspace" really needed 
> - we just want to trace early enough. This model carries over to 
> later as well: there's no *specific* process we want to attach 
> the trace buffer to - we just want a persistent trace buffer 
> that essentially never loses MCE events.
> 
> > Then, after having booted, do I stop the trace? If no, then I 
> > can see the persistency in there so are you saying we want a 
> > low overhead, low ressource utilization machinery which runs 
> > all the time and traces the system? What are possible real 
> > life use cases for that? Scheduler analysis probably, 
> > long-term tracing of some stuff people are interested in how 
> > it behaves over long periods of time... MCE is one use case, 
> > definitely...
> 
> Boot tracing is a very real usecase, people use it to reduce 
> boot times. Today printk timestamps are used as a substitute. 
> (There's also a boot tracer plugin within ftrace, see the 
> bootup_tracer.)
> 
> > > But other, runtime models of tracing could use it as well: 
> > > basically the main difference that ftrace has to perf based 
> > > tracing today is a system-wide persistent buffer with no 
> > > particular owning process. (The rest is mostly UI and 
> > > analysis features and scope of tracing differences, and of 
> > > course a lot more love and detail went into ftrace so far.)
> > > 
> > > So MCE will in the end be just a minor user of such a 
> > > facility - I think you should aim for enabling *any* set of 
> > > events to have persistent recording properties, and add the 
> > > APIs to recover that information sanely. It should also be 
> > > possible for them to record into a shared mmap page in 
> > > essence - instead of having per event persistent buffers.
> > 
> > Sounds like ftrace. But we have that already, we only need to 
> > get to using it perf-side, no...? [...]
> 
> What we want is to extend the perf ring-buffer to be persistent 
> *as well*. It's an evidently useful model of collecting events.
> 
> All the remaining perf tooling can be used after that point - if 
> it's a bog-standard perf ring-buffer then it can be saved into a 
> perf.data and can be analyzed in a rich fashion, etc.
> 
> Think about it: for example we could do not just boot tracing 
> but also boot *profiling*, by using the PMU to sample into a 
> persistent buffer which after bootup can be put into a perf.data 
> and 'perf report' will do the right thing, etc...
> 
> Does it overlap with ftrace? Perf overlapped with ftrace from 
> day one on and it's starting to become a maintenance problem: we 
> want to remove that overlap not by keeping two separate entities 
> (both of which suck and rule in their own ways) but having a 
> unified facility.

Leaving all of the above for reference.

So, I spent some more nights sleeping on it :-)

Here's what I dreamt of:

* The last thing perf_event_init() does is init the persistent, per-cpu
buffers.

* there's no need for changing TRACE_EVENT: "boot_trace" parameter
parsing code enables those events the moment perf is initialized. We're
doing this anyway because we're enabling the trace_mce_record TP.

It sounds pretty simple to me but the devil is in the details,
especially making the persistent buffers, task-agnostic and generic
enough.

Ingo, Peter, thoughts?

Thanks.

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551