Re: [RFD] perf syscall error handling

[Ingo Molnar <mingo@kernel.org>]

* Arnaldo Carvalho de Melo <acme@kernel.org> wrote:

> Em Mon, Nov 10, 2014 at 11:27:25AM +0100, Ingo Molnar escreveu:
> > 
> > * Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
> > 
> > > Em Mon, Nov 03, 2014 at 05:50:19PM +0100, Peter Zijlstra escreveu:
> > > > On Mon, Nov 03, 2014 at 02:25:48PM -0200, Arnaldo Carvalho de Melo wrote:
> > >  
> > > > > The way that peterz suggested, i.e. returning information about which
> > > > > perf_event_attr and which of the parameters was invalid/had issues could
> > > > > help with fallbacking/capability querying, i.e. tooling may want to use
> > > > > some features if available automagically, fallbacking to something else
> > > > > when that fails.
> > >  
> > > > > We already do that to some degree in various cases, but for some if the
> > > > > only way that becomes available to disambiguate some EINVAL return is a
> > > > > string, code will start having strcmps :-\
> > > 
> > > > OK, so how about we do both, the offset+mask for the tools 
> > > > and the string for the humans?
> > > 
> > > Yeah, tooling tries to provide the best it can with the 
> > > offset+mask, and if doesn't manage to do anything smart with 
> > > it, just show the string and hope that helps the user to figure 
> > > out what is happening.
> > 
> > Almost: tooling should generally always consider the string as 
> > well, for the (not so uncommon) case where there can be multiple 
> > problems with the same field.
> > 
> > Really, I think the string will give the most bang for the buck, 
> > because it's really simple and straightforward on the kernel side 
> > (so that we have a good chance of achieving full coverage 
> > relatively quickly), and later on we could still complicate it 
> > all with offset+mask if there's really a need.
> > 
> > So lets start with an error string...
> 
> I don't have a problem with the order of introduction of new 
> error reporting mechanisms, or at least I can't think of one 
> right now.
> 
> So if we introduce strings now then tools/perf/ will trow them 
> to the user when it still don't have fallbacks or any other UI 
> indication of such an error.
> 
> I wonder tho if we have any previous experience on some other 
> project (or even in the kernel?) and how userspace ended up 
> using it, if just presenting those strings to the user or if 
> trying to parse it, etc, anybody?

I'm not aware of any such efforts in the Linux space - subsystems 
with administrative interfaces generally just tend to printk() a 
reason - that's obviously suboptimal in several ways.

Programmatic use in user-spaec is very simple - go with my 
initial example, tooling can either just display the error string 
and bail out, or do:

  if (unlikely(error)) {
	if (!strcmp(attr->error_str, "x86/bts: BTS not supported by this CPU architecture")) {
		fprintf(stderr, "x86/BTS: No hardware support falling back to branch sampling\n");
		activate_x86_bts_fallback_code();
		goto out;
	}
	if (!strcmp(attr->error_str, "x86/lbr: LBR not supported by this CPU architecture"))
		goto out_err;
  }

or it may do any number of other things, such as convert it to 
its internal error code. Note that the error messages should have 
some minimal structure (the 'x86/bts:' and 'x86/lbr' prefixes) to 
organize things nicely and to make string clashes less likely.

as this is a slowpath the performance of strcmp() doesn't matter, 
and in any case it's hardware accelerated or optimized well on 
most platforms.

Thanks,

	Ingo