Messages in this thread | | | Date | Tue, 13 Aug 2013 17:57:16 -0400 (EDT) | From | Vince Weaver <> | Subject | Re: [PATCH] RFC: perf, tools: Move gtk browser into separate perfgtk executable |
| |
On Tue, 13 Aug 2013, Ingo Molnar wrote:
> > that can only be addressed by either extending 'perf test' or by testing > libpfm et al sooner. The upstream kernel can only address regressions that > get reported.
Most of the tests in my test-suite are reactive. Meaning, I wrote them after an ABI-breaking change was reported elsewhere, and I needed a small test case for bisection purposes. Thus they are good for finding if a corner of the perf ABI re-breaks but they're not great at spotting new breakages.
Writing a complete test suite for something as complicated as the perf-event ABI is impractical. One thing you can do is require anyone submitting new functionality also provide a regression test, but I don't see that happening.
Another issue is that despite having some ABI definitions for files in /sys, these are broken with impunity. And I've yet to have an ABI-breaking changeset reverted based on my bug reports. So you can see why I'm not really motivated to even bother trying, as it seems like it would be pointless busy work at this point.
It would just be nice if we just straight out say "the ABI is whatever lets the perf tool run. Anything else is undefined behavior and shouldn't be relied on".
Vince
| |