Re: [PATCH bpf-next] kbuild, bpf: Enable reproducible BTF generation

["Thomas Weißschuh" <linux@weissschuh.net>]

On 2024-12-12 14:54:36-0800, Andrii Nakryiko wrote:
> On Thu, Dec 12, 2024 at 1:07 PM Thomas Weißschuh <linux@weissschuh.net> wrote:
> >
> > Hi Andrii,
> >
> > On 2024-12-12 11:23:03-0800, Andrii Nakryiko wrote:
> > > On Tue, Dec 10, 2024 at 10:24 PM Thomas Weißschuh <linux@weissschuh.net> wrote:
> > > > On 2024-12-11 00:17:02+0000, Ihor Solodrai wrote:
> > > > > On Tuesday, December 10th, 2024 at 3:23 PM, Thomas Weißschuh <linux@weissschuh.net> wrote:
> > > > >
> > > > > >
> > > > > >
> > > > > > Pahole v1.27 added a new BTF generation feature to support
> > > > > > reproducibility in the face of multithreading.
> > > > > > Enable it if supported and reproducible builds are requested.
> > > > > >
> > > > > > As unknown --btf_features are ignored, avoid the test for the pahole
> > > > > > version to keep the line readable.
> > > > > >
> > > > > > Fixes: b4f72786429c ("scripts/pahole-flags.sh: Parse DWARF and generate BTF with multithreading.")
> > > > > > Fixes: 72d091846de9 ("kbuild: avoid too many execution of scripts/pahole-flags.sh")
> > > > > > Link: https://lore.kernel.org/lkml/4154d202-5c72-493e-bf3f-bce882a296c6@gentoo.org/
> > > > > > Link: https://lore.kernel.org/lkml/20240322-pahole-reprodicible-v1-1-3eaafb1842da@weissschuh.net/
> > > > > > Signed-off-by: Thomas Weißschuh linux@weissschuh.net
> > > > > >
> > > > > > ---
> > > > > > scripts/Makefile.btf | 1 +
> > > > > > 1 file changed, 1 insertion(+)
> > > > > >
> > > > > > diff --git a/scripts/Makefile.btf b/scripts/Makefile.btf
> > > > > > index c3cbeb13de503555adcf00029a0b328e74381f13..da23265bc8b3cf43c0a1c89fbc4f53815a290e13 100644
> > > > > > --- a/scripts/Makefile.btf
> > > > > > +++ b/scripts/Makefile.btf
> > > > > > @@ -22,6 +22,7 @@ else
> > > > > >
> > > > > > # Switch to using --btf_features for v1.26 and later.
> > > > > > pahole-flags-$(call test-ge, $(pahole-ver), 126) = -j$(JOBS) --btf_features=encode_force,var,float,enum64,decl_tag,type_tag,optimized_func,consistent_func,decl_tag_kfuncs
> > > > > > +pahole-flags-$(if $(KBUILD_BUILD_TIMESTAMP),y) += --btf_features=reproducible_build
> > > > >
> > > > > Hi Thomas,
> > > > >
> > > > > There are a couple of issues with reproducible_build flag which I
> > > > > think are worth mentioning here. I don't know all the reasons behind
> > > > > adding this now, and it's optional too, so feel free to discard my
> > > > > comments.
> > > > >
> > > > > Currently with this flag, the BTF output is deterministic for a given
> > > > > order of DWARF compilation units. So the BTF will be the same for the
> > > > > same vmlinux binary. However, if the vmlinux is rebuilt due to an
> > > > > incremental change in a source code, my understanding is that there is
> > > > > no guarantee that DWARF CUs will be in the same order in the binary.
> > > >
> > > > The goal behind reproducible builds is to produce bit-by-bit idential
> > > > binaries. If the CUs are in a different order then that requirement
> > > > would have been broken there already.
> > >
> > > I'm curious, how do we guarantee that we get bit-by-bit identical
> > > DWARF? Do we enforce the order of linking of .o files into the final
> > > vmlinux? Is this described anywhere?
> >
> > The CU order has to be fixed, otherwise the non-debugging parts of the
> > binary would not be reproducible either.
> > For docs is Documentation/kbuild/reproducible-builds.rst, the linked
> > reproducible-builds.org project has much more information.
> >
> > Also besides reproducible builds, lots of kernel components rely
> > (accidentally or intentionally) on a stable initialization order, which
> > is also defined by linking order.
> >
> > From Documentation/kbuild/makefiles.rst:
> >
> >         Link order is significant, because certain functions
> >         (module_init() / __initcall) will be called during boot in the
> >         order they appear. So keep in mind that changing the link
> >         order may e.g. change the order in which your SCSI
> >         controllers are detected, and thus your disks are renumbered.
> >
> > > > For an incremental build a full relink with *all* CUs is done, not only
> > > > the changed once, so the order should always be the same.
> > >
> > > The concern here is whether linker guarantees that CUs' DWARF data
> > > will be appended in exactly the same order in such case?
> >
> > Otherwise it wouldn't be reproducible in general.
> > The pahole developers specifically implemented
> > --btf_features=reproducible_build for use in the kernel; after I sent
> > a precursor patch to this one (also linked in the patch):
> >
> > https://lore.kernel.org/lkml/20240322-pahole-reprodicible-v1-1-3eaafb1842da@weissschuh.net/
> >
> > In general the kernel already supports reproducible builds.
> 
> Great, thanks for all the info!
> 
> I do agree with Ihor that KBUILD_BUILD_TIMESTAMP is a non-obvious and
> surprising way to enable this behavior, but if that's what's used for
> other aspects of kernel build I guess it's fine by me.

Agreed. So far KBUILD_BUILD_TIMESTAMP is really only used for timestamp
related configuration. While it's not a perfect fit, adding yet another
switch that needs to be specified can't be the answer, either.

Maybe the Kbuild maintainers have some preference?

> Ihor's work on making BTF generation more deterministic w.r.t. CU
> order would automatically benefit --btf_features=reproducible_build in
> the end and might make it unnecessary, but there is no need to block
> on a completion of that work.

Sounds good.

[..]