Re: [PATCH 2/2] i387: support lazy restore of FPU state

[H. Peter Anvin]

On 02/19/2012 02:44 PM, H. Peter Anvin wrote:
> On 02/19/2012 02:37 PM, Linus Torvalds wrote:
>>
>>  - on *every* task switch from task A, we write A->thread.fpu.last_cpu, 
>>    whether we owned the FPU or not. And we only write a real CPU number in 
>>    the case where we owned it, and the FPU save left the state untouched 
>>    in the FPU.
>>
>>  - so when we switch into task A next time, comparing the current CPU 
>>    number with that 'last_cpu' field inarguably says "when I last switched 
>>    out, I really saved it on this CPU"
>>
>>    That, together with verifying that the per-cpu "fpu_owner_task" matches 
>>    "task A", guarantees that the state is really valid. Because we will 
>>    clear (or set to another task) fpu_owner_task if it ever gets 
>>    switched to anything else.
>>
>> But somebody should really validate this. Think through all the 
>> kernel_fpu_begin() etc cases. I think it looks pretty obvious, and it 
>> really does seem to work and improve task switching, but...
>>
> 
> I think your logic is correct but suboptimal.
> 
> What would make more sense to me is that we write last_cpu when we
> *load* the state.  After all, if you didn't load the state you couldn't
> have modified it.  In kernel_fpu_begin, *if* we end up flushing the
> state, we should set last_cpu to -1 indicating that *no* CPU currently
> owns the state -- after all, even on this CPU we would now have to
> reload the state from memory.
> 

This is obviously wrong for kernel_fpu_begin... what we should do there
is to just set fpu_owner_task to NULL as we no longer have any task's
content in the fpu; no need to much with last_cpu though.

	-hpa

--
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/