Re: [RFC PATCH 04/11] x86,fpu: defer FPU restore until return to userspace

[Andy Lutomirski]

On Tue, Jan 13, 2015 at 9:44 AM, Rik van Riel <riel@xxxxxxxxxx> wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 01/13/2015 12:18 PM, Andy Lutomirski wrote:
>> On Tue, Jan 13, 2015 at 9:11 AM, Oleg Nesterov <oleg@xxxxxxxxxx>
>> wrote:
>>> On 01/11, riel@xxxxxxxxxx wrote:
>>>>
>>>> Defer restoring the FPU state, if so desired, until the task
>>>> returns to userspace.
>>>
>>> And I have another concern.
>>>
>>> Afaocs with this patch the idle threads will run with
>>> TIF_LOAD_FPU set but without fpu.has_fpu.
>>
>> Yuck.  IMO there are still too many possible states.
>>
>> AFAICS the sensible states for a task are:
>>
>> - Task is current on some cpu and FPU is loaded into that cpu. -
>> Task is current on some cpu and FPU is in memory. - Task is current
>> on some cpu and FPU is loaded into a different cpu.
>
> When switching a task out, in prepare_fpu_switch() we will always save
> the FPU state to memory. The third case can be treated identical to the
> second case.

Then what's last_cpu for?

>
>> - Task is not current and FPU is in memory. - Task is not current
>> and FPU is loaded into some cpu.
>
> Same for this one. When the task is not current, the FPU state
> will have been saved to memory. If we try running the task
> somewhere else, it devolves to "FPU is in memory".
>

Isn't there a case where the FPU is in memory *and* in the cpu regs?
Isn't that how you can skip reloading the FPU after going idle and
returning?  Is this what fpu_lazy_restore is for?  Confused.

>> Am I missing anything?  (In lazy mode, there are a few more
>> involving CR0.TS.)
>>
>> That's five states, plus an optional cpu number.  But we have tons
>> of state variable that can express all kinds of nonsense things.
>>
>> If we asserted that we were in a sensible state and fixed the
>> things that exited the sensible states, maybe this would be easier
>> to understand and debug.
>
> Lets see what things we could test, at different points.
>
> 1) At context switch time, we need to make sure that the previous task
>    will no longer have __thread_has_fpu()
>
> 2) When loading the FPU state, we need to make sure that the current
>    task does not have __thread_has_fpu()

Examples, any of which may be wrong:

If !current, then !TIF_LOAD_FPU

If switching out a task with TIF_LOAD_FPU set, then !has_fpu

If last_cpu == smp_processor_id(), then fpu_owner == fpu.

If has_fpu, then the task must be current somewhere and last_cpu must
be the cpu on which it's current.

At the very least, asking these questions may help fill in holes in my
understanding :)

--Andy

>
> 3) ... what else?
>
> - --
> All rights reversed
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1
>
> iQEcBAEBAgAGBQJUtVmPAAoJEM553pKExN6Dc0gIAI0JZE4VpNlJElnuGbFKzM8t
> r9dSH3XLeEX5/JARsQMNuwUL7HjjBbYO9PrdpejVAEUq2XltVyq3+RuSYMBprlpV
> aCWKrw9jn7H43++LSziXh0LP+t+zqL9+/5WUm3oVWkTXJEBu9CdNgYYP7luf87nX
> +mxqDsp5YjmmYJRW2LncK/SuEpItTAtHI5bORjsbWNQOZ2pl31cr1ZEbdrRaZqje
> 5wONfs0f2mor0Ms7FSzVJCNmps3TGJnfpkx37vzqRiPkisdWVcM16RwvM6jJTFZ+
> VmM2Jj4xWMhPYvqpZ8oZWAb8xNCl+DijjZFEGW1j9Ji5gcPQ2HLvFEWNp3TpIeM=
> =OU/0
> -----END PGP SIGNATURE-----



-- 
Andy Lutomirski
AMA Capital Management, LLC
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/