Message-ID: <20160119145640.GF1324@danjae.kornet>
Date:	Tue, 19 Jan 2016 23:56:40 +0900
From:	Namhyung Kim <namhyung@...nel.org>
To:	Ingo Molnar <mingo@...nel.org>
Cc:	Arnaldo Carvalho de Melo <acme@...nel.org>,
	Peter Zijlstra <peterz@...radead.org>,
	Stephane Eranian <eranian@...gle.com>,
	LKML <linux-kernel@...r.kernel.org>,
	Jiri Olsa <jolsa@...nel.org>,
	Adrian Hunter <adrian.hunter@...el.com>,
	"ak@...ux.intel.com" <ak@...ux.intel.com>
Subject: Re: [RFC] perf record: missing buildid for callstack modules

On Wed, Jan 13, 2016 at 10:57:38AM +0100, Ingo Molnar wrote:
> 
> * Arnaldo Carvalho de Melo <acme@...nel.org> wrote:
> 
> > Em Tue, Jan 12, 2016 at 11:39:43AM +0100, Ingo Molnar escreveu:
> > > But perf tooling cares very much: it can lead to subtle bugs and bad data if we 
> > > display a profile with the wrong DSO or binary. 'Bad' profiles resulting out of 
> > > binary mismatch can be very convincing and can send developers down the wrong path 
> > > for hours. I'd expect my tooling to not do that.
> >  
> > > Path names alone (the thing that exec() cares about) are not unique enough to 
> > > identify the binary that was profiled. So we need a content hash - hence the 
> > > build-ID.
> >  
> > > Can you suggest a better solution than a build-time calculated content hash?
> >  
> > > As for binary formats that suck and don't allow for a content hash: we do our 
> > > best, but of course the risk of data mismatch is there. We could perhaps cache the 
> > > binary inode's mtime field to at least produce a 'profile data is older than 
> > > binary/DSO modification date!' warning. (Which check won't catch all cases, like 
> > > cross-system profiling data matches.)
> > 
> > So, we could think of this as: binary formats that want to aid
> > observability tools to:
> > 
> > 1) Detect mismatches in contents for DSOs present at recording time to
> >    those to be used at analysis time.
> > 
> > 2) Find symtabs, DSO binary contents, CFI tables, present in the DSO
> >    where samples were taken.
> > 
> > Using mtime, as suggested in other messages will help with #1, but not
> > with #2.
> 
> But but ... why is #2 a problem with mtime? If we have an out of date record in 
> the perf.data, then the perf.data is uninteresting in 99% of the usecases! It's 
> out of date, most likely because the binary the developer is working on got 
> rebuilt, or the system got upgraded - in both cases the developer does not care 
> about the old records anymore...

We have 'perf diff' command which compares old and new performance
results of a same program.  People can use it to see how much improved
in the new version than the baseline.  In this case, the old binary
should be found from the old perf.data.

Thanks,
Namhyung


> 
> What matters is #1, to detect mismatches, to be a reliable tool. Once we've 
> detected that, we can inform the user and our job is mostly done.
> 
> But reliable != perfect time machine. Really, #2 is a second, third order concern 
> that should never cause slowdowns on the magnitude that Peter is complaining 
> about!
> 
> I realize that there might be special workflows (such as system-wide monitoring) 
> where collecting at recording time might be useful, but those are not the common 
> case - and they should not slow down the common case.
> 
> > Checking for inefficiencies in the current approach of
> > right-after-recording post-processing looking for PERF_RECORD_MMAPs,
> > Adrian suggested something here, also disabling the saving into
> > ~/.debug/ will help, collecting numbers would be great.
> 
> I think Peter mentioned a number: the kernel build time almost _doubles_ with 
> this. That's clearly unacceptable.
> 
> > But the mtime thing also requires traversing the whole perf.data
> > contents looking for those paths in PERF_RECORD_MMAP records.
> 
> But why? Why cannot we do it at perf report time, when we will parse them anyway?
> 
> Thanks,
> 
> 	Ingo

Powered by blists - more mailing lists