Re: [patch 2/2] x86, fpu: lazy allocation of FPU area - v3

[Arjan van de Ven <arjan@linux.intel.com>]

H. Peter Anvin wrote:
> Pavel Machek wrote:
>> On Thu 2008-03-06 16:51:41, Ingo Molnar wrote:
>>> * Pavel Machek <pavel@ucw.cz> wrote:
>>>
>>>>>> kmem_cache_alloc() can fail (return NULL) and not handling it is a 
>>>>>> bug.
>>>>> oops. you are correct. Will send a sigsegv in the failure case 
>>>>> then. Thanks.
>>>> You are introducing possibility of hard to debug error, where 
>>>> previous code just worked... Does not look like good idea to me.
>>> hm, how does it differ from any other allocation failure? We could fail 
>>
>> Well, we should not be sending SIGSEGV...? SIGBUS would be cleaner, or
>> SIGKILL... what happens when userland tries to catch this one?
>>
> 
> I'm confused...
> 
> Normally when we need memory for userspace and can't get it, we put the 
> process to sleep until memory is available.

that's what GFP_KERNEL does
> 
> Why is this different in any way?

this is just for handling the case where that fails
(basically near/totally OOM or the case where you get a fatal signal)

maybe we need a GFP_KILLABLE now that we have a TASK_KILLABLE...