Re: [PATCH] x86/hvmloader: fix usage of NULL with cpuid_count()

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Fri, Apr 25, 2025 at 01:23:30PM +0100, Alejandro Vallejo wrote:
> On Thu Apr 24, 2025 at 1:58 PM BST, Roger Pau Monne wrote:
> > The commit that added support for retrieving the APIC IDs from the APs
> > introduced several usages of cpuid() with NULL parameters, which is not
> > handled by the underlying implementation.  For GCC I expect this results in
> > writes to the physical address at 0, however for Clang the generated code
> > in smp.o is:
> >
> > tools/firmware/hvmloader/smp.o: file format elf32-i386
> >
> > Disassembly of section .text:
> >
> > 00000000 <smp_initialise>:
> >        0: 55                            pushl   %ebp
> >        1: 89 e5                         movl    %esp, %ebp
> >        3: 53                            pushl   %ebx
> >        4: 31 c0                         xorl    %eax, %eax
> >        6: 31 c9                         xorl    %ecx, %ecx
> >        8: 0f a2                         cpuid
> >
> > Showing the usage of a NULL pointer results in undefined behavior, and
> > clang refusing to generate further code after it.
> >
> > Fix by using a temporary variable in cpuid_count() in place for any NULL
> > parameter.
> >
> > Fixes: 9ad0db58c7e2 ('tools/hvmloader: Retrieve APIC IDs from the APs 
> > themselves')
> > Signed-off-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> 
> Ugh, that's on me. I was sure I saw the pattern in Xen (from where the
> code came from), but clearly I hallucinated.
> 
> > ---
> > Could also be fixed by using the temporary variable in the call sites,
> > however that's more code in the call sites at the expense of less checking.
> > I don't think the extra NULL check logic in cpuid_count() is that bad.
> >
> > Overall the solution proposed in this patch is safer going forward, as it
> > prevent issues like this from being introduced in the first place.
> 
> Might be worth moving this same extra checks onto Xen's cpuid. There's
> no shortage of `junk` variables at the callsites.
> 
> > ---
> >  tools/firmware/hvmloader/util.h | 11 +++++++++++
> >  1 file changed, 11 insertions(+)
> >
> > diff --git a/tools/firmware/hvmloader/util.h 
> > b/tools/firmware/hvmloader/util.h
> > index 644450c51ceb..765a013ddd9e 100644
> > --- a/tools/firmware/hvmloader/util.h
> > +++ b/tools/firmware/hvmloader/util.h
> > @@ -190,6 +190,17 @@ static inline void cpuid_count(
> >      uint32_t *ecx,
> >      uint32_t *edx)
> >  {
> > +    uint32_t tmp;
> > +
> > +    if ( !eax )
> > +        eax = &tmp;
> > +    if ( !ebx )
> > +        ebx = &tmp;
> > +    if ( !ecx )
> > +        ecx = &tmp;
> > +    if ( !edx )
> > +        edx = &tmp;
> > +
> 
> A somewhat more compact alternative that doesn't require tmp would be:
> 
>   eax = eax ?: &leaf;
>   ebx = ebx ?: &leaf;
>   ecx = ecx ?: &leaf;
>   edx = edx ?: &leaf;

But that performs the write unconditionally?  It might be more compact
code-wise, but might incur in an unneeded store?

> It clobbers `leaf`, but only after it's no longer relevant.

My preference was to use a explicitly tmp variable, but I could switch
to using leaf if that's the preference.  Given that Andrew has already
Acked the current version I'm tempted to just go with what has already
been Acked.

Thanks, Roger.