Re: [PATCH v3] x86/dom0: delay setting SMAP after dom0 build is done

[Jan Beulich <jbeulich@xxxxxxxx>]

On 01.08.2024 17:36, Roger Pau Monné wrote:
> On Thu, Aug 01, 2024 at 12:28:06PM +0200, Jan Beulich wrote:
>> On 01.08.2024 11:52, Roger Pau Monne wrote:
>>> @@ -1907,16 +1890,25 @@ void asmlinkage __init noreturn 
>>> __start_xen(unsigned long mbi_p)
>>>      if ( cpu_has_smep && opt_smep != SMEP_HVM_ONLY )
>>>          setup_force_cpu_cap(X86_FEATURE_XEN_SMEP);
>>>      if ( boot_cpu_has(X86_FEATURE_XEN_SMEP) )
>>> +    {
>>>          set_in_cr4(X86_CR4_SMEP);
>>> +        cr4_pv32_mask = mmu_cr4_features & XEN_CR4_PV32_BITS;
>>
>> Could be just "cr4_pv32_mask = X86_CR4_SMEP" now?
> 
> Yes, indeed, same below then.
> 
>>> +    }
>>>  
>>>      if ( !opt_smap )
>>>          setup_clear_cpu_cap(X86_FEATURE_SMAP);
>>>      if ( cpu_has_smap && opt_smap != SMAP_HVM_ONLY )
>>>          setup_force_cpu_cap(X86_FEATURE_XEN_SMAP);
>>>      if ( boot_cpu_has(X86_FEATURE_XEN_SMAP) )
>>> -        set_in_cr4(X86_CR4_SMAP);
>>> -
>>> -    cr4_pv32_mask = mmu_cr4_features & XEN_CR4_PV32_BITS;
>>> +        /*
>>> +         * Set SMAP on the %cr4 mask so that it's set for APs on bringup, 
>>> but
>>> +         * don't set for the BSP until domain building is done.
>>> +         *
>>> +         * Don't set it in cr4_pv32_mask either, until it's also set on the
>>> +         * BSP.  Otherwise the BUG in cr4_pv32_restore would trigger for 
>>> events
>>> +         * received on the BSP.
>>> +         */
>>> +        mmu_cr4_features |= X86_CR4_SMAP;
>>
>> Don't you put APs at risk this way of triggering the BUG in 
>> cr4_pv32_restore()?
>> They'll have the bit set in %cr4, but the bit remains clear in cr4_pv32_mask
>> until much later.
> 
> As long as the bit is set in %cr4, but not in cr4_pv32_mask the BUG in
> cr4_pv32_restore won't hit.

Hmm, you're right. Despite staring at that for quite some time, I got
it wrong. Feel free to add
Reviewed-by: Jan Beulich <jbeulich@xxxxxxxx>
then, preferably with said minor adjustment (left in context above).

Jan