On Thu, Mar 08, 2018 at 10:39:44AM +0100, Peter Zijlstra wrote:
> On Thu, Mar 08, 2018 at 04:09:30AM -0500, Mikulas Patocka wrote:
> >
> >
> > On Thu, 8 Mar 2018, Peter Zijlstra wrote:
> >
> > > On Thu, Mar 08, 2018 at 01:11:26AM -0500, Mikulas Patocka wrote:
> > > > The patch b5bc2231b8ad4387c9641f235ca0ad8cd300b6df ("objtool: Add
> > > > retpoline validation") broke compiling 64-bit kernel with 32-bit
> > > > compiler.
> > > >
> > > > This patch fixes the following error and a large number of "can't find
> > > > rela for retpoline_safe" errors that occur when using x32 or i386 gcc.
> > > >
> > > > You shouldn't use the type 'unsigned long' in objtool at all - because
> > > > its
> > > > size depends on the compiler and not on the kernel you are compiling.
> > >
> > > Your patch is wrong because the data field is actually a long. A correct
> > > patch is already in merged in tip.
> >
> > I'm wondering, why is objtool using 'unsigned long' at all? Why not
> > uint32_t and uint64_t? The size of 'unsigned long' is dependent on the
> > compiler, so it will lead to different behavior.
>
> Because pointers... As Josh said, I should've used the right ELF methods
> which use the object's ABI.
But there are still a few other places where objtool uses longs
unnecessarily. Those should probably be converted. Patches welcome...
--
Josh
