Re: [PATCH v3 0/7] perf: Add ioctl for PMU driver configuration

[Mathieu Poirier]

On Tue, 14 Aug 2018 at 11:09, Kim Phillips <kim.phillips@xxxxxxx> wrote:
>
> On Tue, 14 Aug 2018 10:15:56 -0600
> Mathieu Poirier <mathieu.poirier@xxxxxxxxxx> wrote:
>
> > On Mon, 13 Aug 2018 at 11:46, Kim Phillips <kim.phillips@xxxxxxx> wrote:
> > > It yields success for the --per-thread case..:
> > >
> > > $ sudo taskset -c 0 ./perf record -e cs_etm/@20010000.etf/ --per-thread uname -a
> > > Linux juno 4.18.0-rc8-00011-gb82af52c4b35-dirty #147 SMP PREEMPT Thu Aug 9 11:20:37 CDT 2018 aarch64 GNU/Linux
> > > [ perf record: Woken up 0 times to write data ]
> > > Warning:
> > > AUX data lost 1 times out of 2!
> > >
> > > [ perf record: Captured and wrote 0.067 MB perf.data ]
> > > $
> > >
> > > ..but not for CPU-wide?:
> > >
> > > $ sudo taskset -c 0 ./perf record -e cs_etm/@20010000.etf/ uname -a
> > > failed to mmap with 12 (Cannot allocate memory)
> > > $ sudo taskset -c 0 ./perf record -e cs_etm/@20010000.etf/ -C 0 uname -a
> > > failed to mmap with 12 (Cannot allocate memory)
> > > $
> >
> > This patchset is getting very old and a fair amount of things have
> > changed since then.  I'm hoping to be coming out with a new one
> > shortly.  Nonetheless the above is returning an error in CPU-wide
> > scenarios while the feature is being implemented.  Isn't what you
> > requested or have I misunderstood your comment?
>
> No, sigh, I just automatically assumed the patchset would include
> CPU-wide support again.  If it were being done that way, we'd all know
> that the feature(s) this patchset adds would be doing the right thing
> for that purpose, guaranteed.

The patchset published on this list never had support for CPU-wide scenarios.

This is only a preparatory step, the first one in a few more to come.
Sending the whole thing in one go would be way too heavy and is not
realistic.

>
> The other thing that's going on here is that I'm becoming numb to the
> loathsome "failed to mmap with 12 (Cannot allocate memory)" being
> returned no matter what the error is/was. E.g., an error that would
> indicate a sense of non-implementation would be much better
> appreciated than presumably what the above is doing, i.e., returning
> -ENOMEM.  That, backed up with specific details in the form of human
> readable text in dmesg would be *most* welcome.

As part of the refactoring of the code to support CPU-wide scenarios I
intend to emit better diagnostic messages from the driver.  Modifying
rb_alloc_aux() to propagate the error message generated by the
architecture specific PMUs doesn't look hard either and I _may_ get to
it as part of this work.

Mathieu

>
> Thanks,
>
> Kim