Re: [PATCH 2/2] sched/debug: add sched_update_nr_running tracepoint

[Joel Fernandes <joel@joelfernandes.org>]

On Wed, Sep 04, 2019 at 11:43:33AM +0100, Qais Yousef wrote:
> On 09/04/19 00:23, Joel Fernandes wrote:
> > On Tue, Sep 03, 2019 at 05:05:47PM +0100, Valentin Schneider wrote:
> > > On 03/09/2019 16:43, Radim Krčmář wrote:
> > > > The paper "The Linux Scheduler: a Decade of Wasted Cores" used several
> > > > custom data gathering points to better understand what was going on in
> > > > the scheduler.
> > > > Red Hat adapted one of them for the tracepoint framework and created a
> > > > tool to plot a heatmap of nr_running, where the sched_update_nr_running
> > > > tracepoint is being used for fine grained monitoring of scheduling
> > > > imbalance.
> > > > The tool is available from https://github.com/jirvoz/plot-nr-running.
> > > > 
> > > > The best place for the tracepoints is inside the add/sub_nr_running,
> > > > which requires some shenanigans to make it work as they are defined
> > > > inside sched.h.
> > > > The tracepoints have to be included from sched.h, which means that
> > > > CREATE_TRACE_POINTS has to be defined for the whole header and this
> > > > might cause problems if tree-wide headers expose tracepoints in sched.h
> > > > dependencies, but I'd argue it's the other side's misuse of tracepoints.
> > > > 
> > > > Moving the import sched.h line lower would require fixes in s390 and ppc
> > > > headers, because they don't include dependecies properly and expect
> > > > sched.h to do it, so it is simpler to keep sched.h there and
> > > > preventively undefine CREATE_TRACE_POINTS right after.
> > > > 
> > > > Exports of the pelt tracepoints remain because they don't need to be
> > > > protected by CREATE_TRACE_POINTS and moving them closer would be
> > > > unsightly.
> > > > 
> > > 
> > > Pure trace events are frowned upon in scheduler world, try going with
> > > trace points. Qais did something very similar recently:
> > > 
> > > https://lore.kernel.org/lkml/20190604111459.2862-1-qais.yousef@arm.com/
> > > 
> > > You'll have to implement the associated trace events in a module, which
> > > lets you define your own event format and doesn't form an ABI :).
> > 
> > Is that really true? eBPF programs loaded from userspace can access
> > tracepoints through BPF_RAW_TRACEPOINT_OPEN, which is UAPI:
> > https://github.com/torvalds/linux/blob/master/include/uapi/linux/bpf.h#L103
> > 
> > I don't have a strong opinion about considering tracepoints as ABI / API or
> > not, but just want to get the facts straight :)
> 
> It is actually true.
>
> But you need to make the distinction between a tracepoint
> and a trace event first.

I know this distinction well.

> What Valentin is talking about here is the *bare*
> tracepoint without any event associated with them like the one I added to the
> scheduler recently. These ones are not accessible via eBPF, unless something
> has changed since I last tried.

Can this tracepoint be registered on with tracepoint_probe_register()?
Quickly looking at these new tracepoint, they can be otherwise how would they
even work right? If so, then eBPF can very well access it. Look at
__bpf_probe_register() and bpf_raw_tracepoint_open() which implement the
BPF_RAW_TRACEPOINT_OPEN.

> The current infrastructure needs to be expanded to allow eBPF to attach these
> bare tracepoints. Something similar to what I have in [1] is needed - but
> instead of creating a new macro it needs to expand the current macro. [2] might
> give full context of when I was trying to come up with alternatives to using
> trace events.
> 
> [1] https://github.com/qais-yousef/linux/commit/fb9fea29edb8af327e6b2bf3bc41469a8e66df8b
> [2] https://lore.kernel.org/lkml/20190415144945.tumeop4djyj45v6k@e107158-lin.cambridge.arm.com/


As I was mentioning, tracepoints, not "trace events" can already be opened
directly with BPF. I don't see how these new tracepoints are different.

I wonder if this distinction of "tracepoint" being non-ABI can be documented
somewhere. I would be happy to do that if there is a place for the same. I
really want some general "policy" in the kernel on where we draw a line in
the sand with respect to tracepoints and ABI :).

For instance, perhaps VFS can also start having non-ABI tracepoints for the
benefit of people tracing the VFS.

thanks,

 - Joel