Re: [PATCH RFC] gcc-plugins: Handle GCC version mismatch for OOT modules

[Justin Forbes]

On Tue, Jan 26, 2021 at 11:07 AM Greg KH <greg@xxxxxxxxx> wrote:
>
> On Tue, Jan 26, 2021 at 10:19:34AM -0600, Josh Poimboeuf wrote:
> > On Tue, Jan 26, 2021 at 10:15:52AM -0600, Justin Forbes wrote:
> > > On Tue, Jan 26, 2021 at 10:05 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > >
> > > > On Tue, Jan 26, 2021 at 09:46:51AM -0600, Josh Poimboeuf wrote:
> > > > > On Tue, Jan 26, 2021 at 04:15:37PM +0100, Peter Zijlstra wrote:
> > > > > > On Tue, Jan 26, 2021 at 08:51:55AM -0600, Josh Poimboeuf wrote:
> > > > > > > User space mixes compiler versions all the time.  The C ABI is stable.
> > > > > > >
> > > > > > > What specifically is the harder issue you're referring to?
> > > > > >
> > > > > > I don't think the C ABI captures nearly enough. Imagine trying to mix a
> > > > > > compiler with and without asm-goto support (ok, we fail to build without
> > > > > > by now, but just imagine).
> > > > > >
> > > > > > No C ABI violated, but having that GCC extention vs not having it
> > > > > > radically changes the kernel ABI.
> > > > > >
> > > > > > I think I'm with Greg here, just don't do it.
> > > > >
> > > > > Ok, thank you for an actual example.  asm goto is a good one.
> > > > >
> > > > > But it's not a cut-and-dry issue.  Otherwise how could modversions
> > > > > possibly work?
> > > > >
> > > > > So yes, we should enforce GCC versions, but I still haven't seen a
> > > > > reason it should be more than just "same compiler and *major* version".
> > > >
> > > > Why bother? rebuilding the kernel and all modules is a matter of 10
> > > > minutes at most on a decently beefy build box.
> > > >
> > > > What actual problem are we trying to solve here?
> > >
> > > This is true for those of us used to working with source and building
> > > by hand. For users who want everything packaged, rebuilding a kernel
> > > package for install is considerably longer, and for distros providing
> > > builds for multiple arches, we are looking at a couple of hours at
> > > best.  From a Fedora standpoint, I am perfectly fine with it failing
> > > if someone tries to build a module on gcc10 when the kernel was built
> > > with gcc11.  It's tedious when the kernel was built with gcc11
> > > yesterday, and a new gcc11 build today means that kernel needs to be
> > > rebuilt.
> >
> > Right.  It's a problem for distro users.  The compiler and kernel are in
> > separate packages, with separate release cadences.  The latest compiler
> > version may not exactly match what was used to build the latest kernel.
>
> Given that distros _should_ be updating their kernel faster than the
> compiler updates, what's the real issue here?  You build a kernel, and
> all external modules, at the same time.  If you want to build them at
> different times, you make your build system ensure they were the same
> compiler so that you are "bug compatible".
>
> And yes, it might be a pain if gcc11 gets updated every other day, but
> as someone living with a "rolling-distro" you get used to it, otherwise
> you just "pin" the build tools and keep that from happening.
>
> This isn't a new thing, we've been doing this for decades, why is this
> surprising?

We definitely build considerably more kernels than toolchains. From a
rawhide standpoint though, a number of testers are willing to test RC
releases, but are not willing to run debug kernels, so they installed
rc4 yesterday, but will not install today's snapshot.  I will build
3-5 new kernels before they update to rc5.  We have been doing things
this way for over a decade. It has never been a problem until we
turned on CONFIG_GCC_PLUGIN_STRUCTLEAK_BYREF_ALL.  Suddenly I am
getting complaints.