Re: [PATCH bpf v2] bpf: Add config for skipping BTF enum64s

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Sat, Sep 03, 2022 at 07:26:58AM +0200, Greg KH escreveu:
> On Fri, Sep 02, 2022 at 06:51:00PM +0200, Jiri Olsa wrote:
> > On Sun, Aug 28, 2022 at 08:33:17PM -0300, Martin Rodriguez Reboredo wrote:
> > > After the release of pahole 1.24 some people in the dwarves mailing list
> > > notified issues related to building the kernel with the BTF_DEBUG_INFO
> > > option toggled. They seem to be happenning due to the kernel and
> > > resolve_btfids interpreting btf types erroneously. In the dwarves list
> > > I've proposed a change to the scripts that I've written while testing
> > > the Rust kernel, it simply passes the --skip_encoding_btf_enum64 to
> > > pahole if it has version 1.24.
> > > 
> > > v1 -> v2:
> > > - Switch to off by default and remove the config option.
> > > - Send it to stable instead.
> > 
> > hi,
> > we have change that needs to go to stable kernels but does not have the
> > equivalent fix in Linus tree
> 
> Why isn't it also relevant in Linus's tree?

See below.
 
> > what would be the best way to submit it?
> 
> Submit it here and document the heck out of why this isn't in Linus's
> tree, what changes instead fixed it there, and so on.  Look in the
> archives for examples of how this is done, one recent one that I can
> think of is here:
> 	https://lore.kernel.org/r/20220831191348.3388208-1-jannh@google.com
> 
> > the issue is that new 'pahole' will generate BTF data that are not supported
> > by older kernels, so we need to add --skip_encoding_btf_enum64 option to
> > stable kernel's scripts/pahole-flags.sh to generate proper BTF data
> > 
> > we got complains that after upgrading to latest pahole the stable kernel
> > compilation fails
> 
> And what is happening in Linus's tree for this same issue?

So, BTF_KIND_ENUM64 is a new BTF tag, one that is not accepted by older
kernels, but is accepted by the BPF verifier on Linus' tree.

Its about avoiding having a pahole command line with lots of
--enable-new-feature-foo for new stuff with the default producing the
most recent BTF spec.

One way to documenting it: if you update pahole, then please use
--skip_encoding_FOO for these new FOO features on kernels where those
aren't supported.

So this isn't a backport from a fix on Linus' tree, as both the older
pahole that doesn't encode BTF_KIND_ENUM64 and the new one, that encodes
it by default, work with Linus' tree.

Does this violates the stable@ rules?

- Arnaldo