Re: [PATCH v3 2/4] x86/traps: Print non-canonical address on #GP

[Ingo Molnar]

* Borislav Petkov <bp@xxxxxxxxx> wrote:

> On Wed, Nov 20, 2019 at 01:30:58PM +0100, Ingo Molnar wrote:
> > 
> > * Jann Horn <jannh@xxxxxxxxxx> wrote:
> > 
> > > You mean something like this?
> > > 
> > > ========================
> > > diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
> > > index 9b23c4bda243..16a6bdaccb51 100644
> > > --- a/arch/x86/kernel/traps.c
> > > +++ b/arch/x86/kernel/traps.c
> > > @@ -516,32 +516,36 @@ dotraplinkage void do_bounds(struct pt_regs
> > > *regs, long error_code)
> > >   * On 64-bit, if an uncaught #GP occurs while dereferencing a non-canonical
> > >   * address, return that address.
> > >   */
> > > -static unsigned long get_kernel_gp_address(struct pt_regs *regs)
> > > +static bool get_kernel_gp_address(struct pt_regs *regs, unsigned long *addr,
> > > +                                          bool *non_canonical)
> > 
> > Yeah, that's pretty much the perfect end result!
> 
> Why do we need the bool thing? Can't we rely on the assumption that an
> address of 0 is the error case and use that to determine whether the
> resolving succeeded or not?

I'd rather we not trust the decoder and the execution environment so much 
that it never produces a 0 linear address in a #GP:

in get_addr_ref_32() we could get zero:

	linear_addr = (unsigned long)(eff_addr & 0xffffffff) + seg_base;

in get_addr_ref_16() we could get zero too:

	linear_addr = (unsigned long)(eff_addr & 0xffff) + seg_base;

Or in particularly exotic crashes we could get zero in get_addr_ref_64() 
as well:

        linear_addr = (unsigned long)eff_addr + seg_base;

although it's unlikely I suspect.

But the 32-bit case should be plausible enough?

It's also the simplest, most straightforward printout of the decoder 
state: we either see an error, or an (address,canonical) pair of values.

Thanks,

	Ingo