Re: Ping? Re: [PATCH rc] kvm: Prevent compiling virt/kvm/vfio.c unless VFIO is selected

[Sean Christopherson <seanjc@google.com>]

On Wed, Nov 29, 2023, Jason Gunthorpe wrote:
> On Wed, Nov 29, 2023 at 05:07:45PM -0800, Sean Christopherson wrote:
> > On Wed, Nov 29, 2023, Michael Ellerman wrote:
> > > Sean Christopherson <seanjc@google.com> writes:
> > > > On Fri, Nov 10, 2023, Michael Ellerman wrote:
> > > >> Jason Gunthorpe <jgg@nvidia.com> writes:
> > > >> > There are a bunch of reported randconfig failures now because of this,
> > > >> > something like:
> > > >> >
> > > >> >>> arch/powerpc/kvm/../../../virt/kvm/vfio.c:89:7: warning: attribute declaration must precede definition [-Wignored-attributes]
> > > >> >            fn = symbol_get(vfio_file_iommu_group);
> > > >> >                 ^
> > > >> >    include/linux/module.h:805:60: note: expanded from macro 'symbol_get'
> > > >> >    #define symbol_get(x) ({ extern typeof(x) x __attribute__((weak,visibility("hidden"))); &(x); })
> > > >> >
> > > >> > It happens because the arch forces KVM_VFIO without knowing if VFIO is
> > > >> > even enabled.
> > > >> 
> > > >> This is still breaking some builds. Can we get this fix in please?
> > > >> 
> > > >> cheers
> > > >> 
> > > >> > Split the kconfig so the arch selects the usual HAVE_KVM_ARCH_VFIO and
> > > >> > then KVM_VFIO is only enabled if the arch wants it and VFIO is turned on.
> > > >
> > > > Heh, so I was trying to figure out why things like vfio_file_set_kvm() aren't
> > > > problematic, i.e. why the existing mess didn't cause failures.  I can't repro the
> > > > warning (requires clang-16?), but IIUC the reason only the group code is problematic
> > > > is that vfio.h creates a stub for vfio_file_iommu_group() and thus there's no symbol,
> > > > whereas vfio.h declares vfio_file_set_kvm() unconditionally.
> > > 
> > > That warning I'm unsure about.
> > 
> > Ah, it's the same warning, I just missed the CONFIG_MODULES=n requirement.
> 
> Oh, wait, doesn't that mean the approach won't work? IIRC doesn't
> symbol_get turn into just &fn when non-modular turning this into a
> link failure without the kconfig part?

Yes, but it doesn't cause linker errors.  IIUC, because the extern declaration
is tagged "weak", a dummy default is used.  E.g. on x86, this is what is generated
with VFIO=y

                fn = symbol_get(vfio_file_is_valid);
                if (!fn)
   0xffffffff810396c5 <+5>:	mov    $0xffffffff81829230,%rax
   0xffffffff810396cc <+12>:	test   %rax,%rax

whereas VFIO=n gets

                fn = symbol_get(vfio_file_is_valid);
                if (!fn)
   0xffffffff810396c5 <+5>:	mov    $0x0,%rax
   0xffffffff810396cc <+12>:	test   %rax,%rax

I have no idea if the fact that symbol_get() generates '0', i.e. the !NULL checks
work as expected, is intentional or if KVM works by sheer dumb luck.  I'm not
entirely sure I want to find out, though it's definitely extra motiviation to get
the KVM mess fixed sooner than later.