Re: [RFC PATCH 04/10] x86/fpu: eager switch PKRU state

[Andy Lutomirski]

> On Sep 12, 2018, at 8:30 AM, Rik van Riel <riel@xxxxxxxxxxx> wrote:
> 
> On Wed, 2018-09-12 at 08:20 -0700, Andy Lutomirski wrote:
>>> 
>>> --- a/arch/x86/mm/pkeys.c
>>> +++ b/arch/x86/mm/pkeys.c
>>> @@ -18,6 +18,20 @@
>>> 
>>> #include <asm/cpufeature.h>             /* boot_cpu_has,
>>> ...            */
>>> #include <asm/mmu_context.h>            /*
>>> vma_pkey()                   */
>>> +#include <asm/fpu/internal.h>
>>> +
>>> +void write_pkru(u32 pkru)
>>> +{
>>> +    if (!boot_cpu_has(X86_FEATURE_OSPKE))
>>> +        return;
>>> +
>>> +    current->thread.fpu.pkru = pkru;
>>> +
>> 
>> I thought that the offset of PKRU in the xstate was fixed after boot.
> 
> You are right, it is.  However, that offset would need
> to be stored somewhere, and the value read every time
> we wanted to read or store the PKRU value from/to the
> floating point state.
> 
> I suspect that would not be any faster than keeping a
> copy of the PKRU value in a known location.
> 
>> Anyway, as written, this needs a lockdep assertion that weâre not
>> preemptible, an explicit preempt_disable(), or a comment explaining
>> why itâs okay if we get preempted in this function.
>> 
>>> +    __fpregs_changes_begin();
> 
> This handles the preemption disabling, see patch
> 3 of the series.

Sure, but the first write is *before* this. So we can be preempted with the two copies of PKRU being out of sync.

> 
>>> +    __fpregs_load_activate(&current->thread.fpu,
>>> smp_processor_id());
>>> +    __write_pkru(pkru);
>>> +    __fpregs_changes_end();
>>> +}
>>> 
>>> int __execute_only_pkey(struct mm_struct *mm)
>>> {
>>> -- 
>>> 2.19.0
>>> 
>> 
>> 
> -- 
> All Rights Reversed.