Fuzzing perf_events [LWN.net]

Fuzzing perf_events

Posted Aug 11, 2015 4:02 UTC (Tue) by graydon (guest, #5009)
In reply to: Fuzzing perf_events by flussence
Parent article: Fuzzing perf_events

I am not a kernel hacker and I use perf about a dozen times a day. It's the best profiler I've ever worked with. But YMMV.

to post comments

Fuzzing perf_events

Posted Aug 11, 2015 11:23 UTC (Tue) by andresfreund (subscriber, #69562) [Link]

Yea, agreed. It's far from perfect, but it's pretty unbeatable in how quickly you can get a fairly accurate picture of performance problems.

My biggest annoyance is that call graphs are only usable if you a) recompile software using frame pointers b) use dwarf mode which generates huge data files and isn't available on many existing systems c) use lbr mode which requires a new cpu, kernel and perf version... I hate whoever made the decision to default gcc to not using frame pointers in newer gcc versions. I kinda see the point on x86-32 which is really register starved, but it's not that bad on 64bits and it makes debugging and profiling so much harder.
Second to that comes the far too generic and thus hard to search name ;)