Messages in this thread | | | Date | Wed, 14 Jul 2010 16:50:02 -0400 | From | Mathieu Desnoyers <> | Subject | Re: [PATCHv9 2.6.35-rc4-tip 10/13] perf: Re-Add make_absolute_path |
| |
* Ingo Molnar (mingo@elte.hu) wrote: > > * Arnaldo Carvalho de Melo <acme@infradead.org> wrote: > > > Em Tue, Jul 13, 2010 at 10:49:27PM -0400, Steven Rostedt escreveu: > > > On Mon, 2010-07-12 at 13:12 -0300, Arnaldo Carvalho de Melo wrote: > > > > Em Mon, Jul 12, 2010 at 10:30:36AM -0400, Steven Rostedt escreveu: > > > > > > Well, I prefer to follow the kernel way of doing things, i.e. to > > > > propagate as much as possible up the callchain the error return value, > > > > so that the apps can handle it in any way they prefer, i.e. die() calls > > > > in tools/perf/builtin-foo.c are okayish, but not on tools/perf/util/. > > > > > Ah, yes, die is a bit strong. And I have been starting to avoid them > > > too. Although, when malloc fails, it's almost certain that the app will > > > die soon anyway ;-) > > > > The interesting thing is that years ago, when modules were being introduced > > in the kernel and panic() calls for things like out of memory conditions > > were being removed, some people made the same comments, 'if that happens, > > you're doomed anyway!' :-) > > > > I can see things like trying to load a huge perf.data file in the TUI > > interface failing and the user just being warned about it and going on with > > life loading some other file, etc. > > > > Certainly it is interesting to try to apply as much as possible of the > > mindset (and fear of criticism) present when coding for the kernel when one > > codes for userland. > > Yeah, and especially for perf the absolutely most important quality is > reliability. It's not just an app - it's a measurement tool. People rely on it > to reject or apply patches, on a daily basis. > > perf must be very reliable and very dependable (and i'm happy that we managed > to achieve that goal so far :), and if it fails it should be apparent that it > failed and that results should not be relied on. > > With other tools that are statistical i've sometimes seen a special type of > dangerous attitude of: "hey, it's just a sample, no biggie if it's lost, it's > just statistical anyway, lost in the noise" - but that's really a slippery > slope leading to a sloppy tool we cannot depend on 110%. > > Just like physicists or engineers want to be able to trust their measurement > instruments, do we want kernel hackers to be able to trust the results of > perf.
I'm glad to see we're on the exact same page here. :)
Having reliable trace analysis tools is crucially important for both kernel developers and end-users, especially when, as an example, someone try to use the results to find out which is the "one" odd-case longest/shortest interrupt handler during a day-long trace while trying to pinpoint a bug.
Thanks,
Mathieu
> > Thanks, > > Ingo
-- Mathieu Desnoyers Operating System Efficiency R&D Consultant EfficiOS Inc. http://www.efficios.com
| |