On Thu, 2014-11-27 at 22:27 +0100, Jan Kratochvil wrote:
> On Thu, 27 Nov 2014 12:04:39 +0100, Mark Wielaard wrote:
> > On Wed, 2014-11-26 at 15:32 +0100, Jan Kratochvil wrote:
> > > Therefore they are dead at the core_pattern time, they cannot be ptraced and
> > > therefore they cannot be unwound. One could only find them in the core file
> > > itself but that is outside of the scope of this eu-stack feature.
> > 
> > Why do you think that is out of scope? We get both the actual tid
> > through the command line option and the core file through stdin. Can't
> > we combine those two to get all information we want/need?
> 
> It could be extended that way but:
> 
> Currently elfutils is used only for the list of unwound PCs.  There is missing
> a big part of functionality present in GDB - displaying parameters, local
> variables, entry-value recovered values, data types pretty printing etc.
> And I have doubts unwind of non-crashed threads is useful at all.

But that not what I suggested. Although I think it would be cool to
extend eu-stack to also show arguments, locals, etc. That is an
interesting project on its own separate from this work.

> It is a new functionality on top of this patch that it has to read thread
> registers from PT_NOTE but access the memory via ptraced crashed thread.
> Fortunately core file PT_NOTE is first so it can be read quickly:
> Program Headers:
>   Type           Offset   VirtAddr           PhysAddr           FileSiz  MemSiz   Flg Align
>   NOTE           0x0004a0 0x0000000000000000 0x0000000000000000 0x0005f4 0x000000     0
> As reading the whole core file is in many times not possible (without some
> better kernel core dumping interface):
> 	-fsanitize=address locks up abrt-hook-ccpp
> 	https://bugzilla.redhat.com/show_bug.cgi?id=1164548

Is this bug being hit by elfutils/libdwfl core file parsing or a
specific issue with how abrt-hook-ccpp parses its input? I agree a nicer
interface for /proc/sys/kernel/core_pattern would be appreciated. But
that seems a kernel issue. We have to work with what we have now.

systemd has a standard core_pattern hook systemd-coredump (that also
uses elfutils/libdwfl to extract backtraces). Maybe we can work with
them on a interface/daemon to provide access to client programs that
wish to inspect the process/core? But again I think that is separate
from what we are currently working on.

> So I agree your proposal makes some sense but:
>  * I find it an incremental feature on top of this one.
>  * Given elfutils just lists the unwound PCs I do not see how it is useful.
> 
> What is your - or Martin's - opinion on such feature given these contraints?

I think getting backtraces for one or more threads of a crashing/exiting
process is an nice feature. My point is simply that I think that what
you wrote is a nice feature that we need to make into something that is
really usable. Whether that is by extending the support to something
more generic like --wait-pid that can be used independent of the
core_pattern filter or in providing in eu-stack all functionality
expected from a core_pattern filter program so that it is just a drop in
filter program doesn't really matter to me. All I am interested in is
that the eu-stack feature will actually be used and is not just some
example code that isn't really usable or used by actual users.

Cheers,

Mark