Re: [RFC] kbuild: bpf: Do not run pahole with -j on 32bit userspace

[Jiri Slaby]

On 21. 08. 24, 7:32, Jiri Slaby wrote:
> On 20. 08. 24, 16:33, Jiri Olsa wrote:
> > On Tue, Aug 20, 2024 at 10:59:50AM +0200, Jiri Slaby (SUSE) wrote:
> > > From: Jiri Slaby <jslaby@xxxxxxx>
> > >
> > > == WARNING ==
> > > This is only a PoC. There are deficiencies like CROSS_COMPILE or LLVM
> > > are completely unhandled.
> > >
> > > The simple version is just do there:
> > >    ifeq ($(CONFIG_64BIT,y)
> > > but it has its own deficiencies, of course.
> > >
> > > So any ideas, inputs?
> > > == WARNING ==
> > >
> > > When pahole is run with -j on 32bit userspace (32bit pahole in
> > > particular), it randomly fails with OOM:
> > > > btf_encoder__tag_kfuncs: Failed to get ELF section(62) data: out of 
> > > > memory.
> > > > btf_encoder__encode: failed to tag kfuncs!
> > >
> > > or simply SIGSEGV (failed to allocate the btf encoder).
> > >
> > > It very depends on how many threads are created.
> > >
> > > So do not invoke pahole with -j on 32bit.
> >
> > could you share more details about your setup?
> >
> > does it need to run on pure 32bit to reproduce?
>
> armv7l builds are 32bit only.
>
> > I can't reproduce when
> > doing cross build and running 32 bit pahole on x86_64..
>
> i586 is built using 64bit kernel. It is enough to have 32bit userspace.
> As written in the linked bug:
> https://bugzilla.suse.com/show_bug.cgi?id=1229450#c6
>
> FWIW, steps to reproduce locally:
> docker pull jirislaby/pahole_crash
> docker run -it jirislaby/pahole_crash
>
> The VM space of pahole is exhausted:
> process map: https://bugzilla.suse.com/attachment.cgi?id=876821
> strace of mmaps: https://bugzilla.suse.com/attachment.cgi?id=876822
>
> You need to run with large enough -j on a fast machine. Note that this 
> happens on build hosts even with -j4, but they are under heavy load, so 
> parallelism of held memory is high.

From https://bugzilla.suse.com/show_bug.cgi?id=1229450#c20:
Run on 64bit:
pahole -j32 -> 4.102 GB
pahole -j16 -> 3.895 GB
pahole -j1 -> 3.706 GB

On 32bit (the same vmlinux):
pahole -j32 -> 2.870 GB (crash)
pahole -j16 -> 2.810 GB
pahole -j1 -> 2.444 GB

Look there for full massif report.

So now I think we should disable BTF generation with 32bit pahole 
completely. Or someone debugs it and improves debug info loading to not 
eat that much.

thanks,
--
js
suse labs