perf tools: Question about kmem and kernel symbol resolution

perf tools: Question about kmem and kernel symbol resolution

[Namhyung Kim]

Hi,

I'm looking kernel symbol mismatch issue, and found something in perf
kmem code.  The commit e727ca73f85d ("perf kmem: Resolve kernel
symbols again") added perf_session__create_kernel_maps() but I don't
know why.  Why did it miss the MMAP event?

I think if we create a kernel maps at report time, it might not match
to samples in a perf.data if it's recorded on a different kernel.
This is the main reason of the mismatch problem I'm currently chasing
IMHO.  What am I missing?

Thanks,
Namhyung

Re: perf tools: Question about kmem and kernel symbol resolution

[Arnaldo Carvalho de Melo]

Em Thu, Jul 31, 2014 at 05:35:32PM +0900, Namhyung Kim escreveu:
> I'm looking kernel symbol mismatch issue, and found something in perf
> kmem code.  The commit e727ca73f85d ("perf kmem: Resolve kernel
> symbols again") added perf_session__create_kernel_maps() but I don't
> know why.  Why did it miss the MMAP event?
 
> I think if we create a kernel maps at report time, it might not match
> to samples in a perf.data if it's recorded on a different kernel.
> This is the main reason of the mismatch problem I'm currently chasing
> IMHO.  What am I missing?

>From a quick look, nothing, i.e. we can not call
perf_session__create_kernel_maps() at that point, as it will create the
kernel maps from the running kernel and use it with events from the
kernel that was in place when the perf.data file being processed was
created.

Perhaps that problem was fixed somewhere else and we should just revert
that patch?

Have you tried just reverting it and checking that the results are the
expected ones? I.e. that there is the kernel MMAP event in perf.data
file and that it gets properly processed?

- Arnaldo

Re: perf tools: Question about kmem and kernel symbol resolution

[Namhyung Kim]

On Thu, 31 Jul 2014 11:27:11 -0300, Arnaldo Carvalho de Melo wrote:
> Em Thu, Jul 31, 2014 at 05:35:32PM +0900, Namhyung Kim escreveu:
>> I'm looking kernel symbol mismatch issue, and found something in perf
>> kmem code.  The commit e727ca73f85d ("perf kmem: Resolve kernel
>> symbols again") added perf_session__create_kernel_maps() but I don't
>> know why.  Why did it miss the MMAP event?
>  
>> I think if we create a kernel maps at report time, it might not match
>> to samples in a perf.data if it's recorded on a different kernel.
>> This is the main reason of the mismatch problem I'm currently chasing
>> IMHO.  What am I missing?
>
>> From a quick look, nothing, i.e. we can not call
> perf_session__create_kernel_maps() at that point, as it will create the
> kernel maps from the running kernel and use it with events from the
> kernel that was in place when the perf.data file being processed was
> created.
>
> Perhaps that problem was fixed somewhere else and we should just revert
> that patch?
>
> Have you tried just reverting it and checking that the results are the
> expected ones? I.e. that there is the kernel MMAP event in perf.data
> file and that it gets properly processed?

Simply reverting ended up with no symbols but it contains MMAP event for
sure.

Then I found a reason - it's simply because kmem tools doesn't register
mmap event handlers. :-/  Adding mmap[2] handlers + reverting ended up
with the expected output.

I'll send the fix soon.

Thanks,
Namhyung