Re: perf/Documentation/ABI -- add some documentation for perf_event sysfs usage

[Jiri Olsa <jolsa@redhat.com>]

On Fri, Jul 05, 2013 at 11:28:08AM -0400, Vince Weaver wrote:
> On Fri, 5 Jul 2013, Jiri Olsa wrote:
> 
> > On Thu, Jul 04, 2013 at 11:02:53AM +0200, Peter Zijlstra wrote:
> > > On Wed, Jul 03, 2013 at 11:14:40PM -0400, Vince Weaver wrote:
> 
> > > >    Documentation/ABI/testing/sysfs-bus-event_source-devices-events
> > > > 
> > > > Should the ldlat value be fixed to be hex?  Or should we ammend the ABI
> > > > document to allow decimal?
> > > 
> > > I don't see a good reason not to allow decimal as well. Jolsa?
> > 
> > yep, no technical problem with decimal
> > 
> > hum, the doc mentions 'event' term only, which IS hex only AFAICS ;-)
> > 
> > I think this docs should be updated and either describe all
> > allowed terms or be generic enough to cover all of them.
> 
> So you're saying the official kernel ABI should be "whatever the 
> userspace perf tool happens to accept"?

nope, I meant whatever term is used by kernel to describe the event

'event, umask, edge, pc, any, inv, cmask, ldat'

please check x86_event_sysfs_show function

> 
> That's not really useful, especially as perf doesn't distinguish between
> event strings read from sysfs and those passed on the perf command line.
> The ABI documentation in effect ends up being a pointer to a mostly 
> incomprehensible lex/yacc file.

the documentation is missing, no argument here :-\
'sysfs-bus-event_source-devices-events' could be more descriptive

> 
> I think we should just remove the 
>    Documentation/ABI/testing/sysfs-bus-event_source-devices-events
> file as it's misleading.  Userspace broke and no one cares.
> 
> It's not even easy to audit all places in the kernel that create the sysfs 
> event files because each architecture does it differently.  And there's no 
> sane way to unit test this on a new kernel release because the values 
> printed depend on the hardware you have, so without a full range of all 
> families of cpus for all architectures you never know when someone has 
> added a decimal value, or started depending on { characters, etc.
> 
> This is the problem with perf in the tools directory, any other user of 
> the ABI is eternally second-class.

with updated documentation I dont see the issue

jirka