[PATCH 40/45] powerpc, irq: Use GFP_ATOMIC allocations in atomic context

[Srivatsa S. Bhat]

On 06/25/2013 08:43 AM, Benjamin Herrenschmidt wrote:
> On Tue, 2013-06-25 at 12:58 +1000, Michael Ellerman wrote:
>> On Tue, Jun 25, 2013 at 12:13:04PM +1000, Benjamin Herrenschmidt wrote:
>>> On Tue, 2013-06-25 at 12:08 +1000, Michael Ellerman wrote:
>>>> We're not checking for allocation failure, which we should be.
>>>>
>>>> But this code is only used on powermac and 85xx, so it should probably
>>>> just be a TODO to fix this up to handle the failure.
>>>
>>> And what can we do if they fail ?
>>
>> Fail up the chain and not unplug the CPU presumably.
> 
> BTW. Isn't Srivatsa series removing the need to stop_machine() for
> unplug ? 

Yes.

That should mean we should be able to use GFP_KERNEL no ?

No, because whatever code was being executed in stop_machine() context
would still be executed with interrupts disabled. So allocations that
can sleep would continue to be forbidden in this path.

In the CPU unplug sequence, the CPU_DYING notifications (and the surrounding
code) is guaranteed to be run:
a. _on_ the CPU going offline
b. with interrupts disabled on that CPU.

My patchset will retain these guarantees even after removing stop_machine().
And these are required for the correct execution of the code in this path,
since they rely on these semantics.

So I guess I'll retain the patch as it is. Thank you!

Regards,
Srivatsa S. Bhat