On Fri, Dec 6, 2019 at 8:45 AM Stephen Smalley <sds@xxxxxxxxxxxxx> wrote:
> On 12/5/19 7:50 PM, Paul Moore wrote:
> > On Thu, Dec 5, 2019 at 1:14 PM Paul Moore <paul@xxxxxxxxxxxxxx> wrote:
> >> On Thu, Dec 5, 2019 at 1:10 PM Stephen Smalley <sds@xxxxxxxxxxxxx> wrote:
> >>> On 12/5/19 12:41 PM, Paul Moore wrote:
> >>>> Hmm. I haven't done any debugging yet, but the BPF tests are failing
> >>>> (they pass with kernel-5.5.0-0.rc0.git5.1.2.secnext.fc32.x86_64):
> >
> > ...
> >
> >>> They all pass for me (with your next-queue branch, using the
> >>> selinux-testsuite defconfig fragment merged with the Fedora config).
> >>
> >> Oh goodie, I'm special :/
> >>
> >> FWIW, my current test kernel is the next-queue branch rebased on top
> >> of Linus' current tree, using the latest config from the secnext
> >> kernel builds (Fedora Rawhide + stuff for the test suite).
> >>
> >>> The error above doesn't look SELinux-related; it looks like your kernel
> >>> is rejecting the trivial bpf program used in the test code as being
> >>> invalid for some reason.
> >>
> >> That's where I'm at as well, I'm building an instrumented kernel right
> >> now to try and track down the source. I'm sure it is something silly
> >> like a messed up kernel config or something, but I'd like to
> >> understand *why*.
> >
> > I traced the "./bpf_test -p" failure down to a BTF check in the BPF
> > verifier, there is a comment in that code block which helpfully reads:
> > "Either gcc or pahole or kernel are broken.".
> >
> > :/
> >
> > The relevant commit is 8580ac9404f6 ("bpf: Process in-kernel BTF"),
> > and it appears to be new for v5.5; it isn't present in selinux/next or
> > selinux/next-queue. Recompiling with CONFIG_DEBUG_INFO_BTF disabled
> > does allow "./bpf_test -p" to succeed, but I hit other BPF test
> > failures further along. For reasons I don't understand, the secnext
> > kernel builds (which should have this code, and have
> > CONFIG_DEBUG_INFO_BTF enabled) are not hitting this problem, but that
> > may be due to differences in the build tools on the two systems
> > (although they *should* be the same).
> >
> > Given that we haven't hit -rc1 yet, and everyone else's builds are
> > working just fine, I'm going to leave this alone for now. Whatever
> > the problems may be, they definitely don't appear to be SELinux
> > related.
>
> I re-based next-queue on top of -linus, enabled CONFIG_DEBUG_INFO_BTF,
> rebuilt and booted new kernel, did a git clean -fdx in the
> selinux-testsuite directory, and built/ran the testsuite; bpf tests
> still passed for me. This was on F31.
Thanks. If I had another day to spend on this I'm reasonably
confident it would end up being the result of some oddity/difference
in the Rawhide toolchains, or ghosts. Either way, this seems to be
both unrelated to the SELinux changes and not something anyone else is
seeing despite a number of different scenarios. I still need to go
through the two patches and cleanup some of the whitespace issues we
talked about earlier, but barring anything weird happening (see the
previous ghost comment) I'll merge both of these patches into
selinux/next when the merge window closes this weekend.
--
paul moore
www.paul-moore.com
