Re: [PATCH v2] x86: Disable kernel stack offset randomization for !TSC

["Maciej W. Rozycki" <macro@orcam.me.uk>]

Jason,

 Would you mind commenting on the below?

On Mon, 9 Jan 2023, Ingo Molnar wrote:

> > For x86 kernel stack offset randomization uses the RDTSC instruction, 
> > which causes an invalid opcode exception with hardware that does not 
> > implement this instruction:
> 
> > @@ -85,7 +86,8 @@ static inline void arch_exit_to_user_mod
> >  	 * Therefore, final stack offset entropy will be 5 (x86_64) or
> >  	 * 6 (ia32) bits.
> >  	 */
> > -	choose_random_kstack_offset(rdtsc() & 0xFF);
> > +	if (cpu_feature_enabled(X86_FEATURE_TSC))
> > +		choose_random_kstack_offset(rdtsc() & 0xFF);
> >  }
> 
> While this is an obscure corner case, falling back to 0 offset silently 
> feels a bit wrong - could we at least attempt to generate some 
> unpredictability in this case?
> 
> It's not genuine entropy, but we could pass in a value that varies from 
> task to task and which is not an 'obviously known' constant value like the 
> 0 fallback?
> 
> For example the lowest 8 bits of the virtual page number of the current 
> task plus the lowest 8 bits of jiffies should vary from task to task, has 
> some time dependence and is cheap to compute:
> 
> 	(((unsigned long)current >> 12) + jiffies) & 0xFF
> 
> This combined with the per-CPU forward storage of previous offsets:
> 
> #define choose_random_kstack_offset(rand) do {                          \
>         if (static_branch_maybe(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT, \
>                                 &randomize_kstack_offset)) {            \
>                 u32 offset = raw_cpu_read(kstack_offset);               \
>                 offset ^= (rand);                                       \
>                 raw_cpu_write(kstack_offset, offset);                   \
>         }                                                               \
> 
> Should make this reasonably hard to guess for long-running tasks even if 
> there's no TSC - and make it hard to guess even for tasks whose creation an 
> attacker controls, unless there's an info-leak to rely on.

 Sure, I'm fine implementing it, even in such a way so as not to cause a 
code size/performance regression for X86_TSC configurations.  But is the 
calculation really unpredictable enough?  I don't feel competent enough to 
decide.  Jason, what do you think?

  Maciej