Performance issue with systemd-coredump and container process linking 2000 shared libraries.

[Mark Wielaard]

Hi Romain,

On Wed, Jun 21, 2023 at 06:14:27PM +0000, Romain GEISSLER wrote:
> > Le 21 juin 2023 à 18:24, Mark Wielaard <mark@klomp.org> a écrit :
> > 
> > Seeing those performance results I understand why you were suspecting
> > the linked list data structure used in elf_getdata_rawchunk.
> > 
> > Would you be able to test a rebuild libelf with the attached patch,
> > which replaces that datastructure with a binary search tree?
> > 
> > It didn't really show much speedup locally (in the noise, maybe 0.01
> > sec faster on ~0.25 sec run). But if there are more than 2000 calls to
> > elf_getdata_rawchunk it should make things faster.
> 
> Actually I have spent some time today to reproduce the issue for
> real, using directly systemd. The details can be found in
> https://github.com/systemd/systemd/pull/28093 which was just
> merged. This change in systemd is enough to fix the biggest part of
> the "slow" parsing of cores. Yet even with my systemd patch, we will
> still call elf_getdata_rawchunk 2000 times if the crashing process
> had 2000 shared libraries, so indeed your patch in elfutils might
> still be welcome. I will test it later when I go home, or tomorrow.

That patch looks good. It should reduce the number of calls to
elf_getdata_rawchunk a lot. Making it less urgent that function is
fast. But if you could test it that would be appreciated. I'll also
test it a bit more and will probably merge it if no issues show up
since it does seem better than keep using the linear list.

Thanks,

Mark