Re: [tip: core/objtool] objtool: Fail the kernel build on fatal errors

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Fri, Feb 14, 2020 at 01:10:26AM +0100, Thomas Gleixner wrote:
> Josh Poimboeuf <jpoimboe@redhat.com> writes:
> > On Tue, Feb 11, 2020 at 12:47:38PM -0000, tip-bot2 for Josh Poimboeuf wrote:
> >> The following commit has been merged into the core/objtool branch of tip:
> >> 
> >> Commit-ID:     644592d328370af4b3e027b7b1ae9f81613782d8
> >> Gitweb:        https://git.kernel.org/tip/644592d328370af4b3e027b7b1ae9f81613782d8
> >> Author:        Josh Poimboeuf <jpoimboe@redhat.com>
> >> AuthorDate:    Mon, 10 Feb 2020 12:32:38 -06:00
> >> Committer:     Borislav Petkov <bp@suse.de>
> >> CommitterDate: Tue, 11 Feb 2020 13:27:03 +01:00
> >> 
> >> objtool: Fail the kernel build on fatal errors
> >> 
> >> When objtool encounters a fatal error, it usually means the binary is
> >> corrupt or otherwise broken in some way.  Up until now, such errors were
> >> just treated as warnings which didn't fail the kernel build.
> >> 
> >> However, objtool is now stable enough that if a fatal error is
> >> discovered, it most likely means something is seriously wrong and it
> >> should fail the kernel build.
> >> 
> >> Note that this doesn't apply to "normal" objtool warnings; only fatal
> >> ones.
> >
> > Clang still has some toolchain issues which need to be sorted out, so
> > upgrading the fatal errors is causing their CI to fail.
> 
> Good. Last time we made it fail they just fixed their stuff.
> 
> > So I think we need to drop this one for now.
> 
> Why? It's our decision to define which level of toolchain brokeness is
> tolerable.
> 
> > Boris, are you able to just drop it or should I send a revert?
> 
> I really want to see a revert which has a proper justification why the
> issues of clang are tolerable along with a clear statement when this
> fatal error will come back. And 'when' means a date, not 'when clang is
> fixed'.

Fair enough.  The root cause was actually a bug in binutils which gets
triggered by a new clang feature.  So instead of reverting the above
patch, I think I've figured out a way to work around the binutils bug,
while also improving objtool at the same time (win-win).

The binutils bug will be fixed in binutils 2.35.

BTW, to be fair, this was less "Clang has issues" and more "Josh is
lazy".  I didn't test the patch with Clang -- I tend to rely on 0-day
bot reports because I don't have the bandwidth to test the
kernel/config/toolchain combinations.  Nick tells me Clang will soon be
integrated with the 0-day bot, which should help prevent this type of
thing in the future.

-- 
Josh