[Intel-gfx] [PATCH] intel: Delay testing for userptr until first use

[Tvrtko Ursulin]

On 11/04/2014 06:44 PM, Chris Wilson wrote:
> On Tue, Nov 04, 2014 at 03:14:04PM +0000, Tvrtko Ursulin wrote:
>>
>> On 11/04/2014 02:31 PM, Chris Wilson wrote:
>>> Running __mmu_notifier_register() is surprisingly expensive, so let's
>>> not do that unless we have to.
>>
>> Affects some program startup or what? What is the cost? I would add
>> some notes in the commit for future reference.
>
> It was affecting a badly behavely igt test, that was provoking the
> linear walk in mm_take_all_locks() and then causing each one to restart
> due to signals... Performance fell off a cliff. This was extreme, but
> one can suppose that on a similarly stressed system startup performance
> will also degrade.

Oh, so restarting the ioctl based on EINTR which propagates all the way 
up from mm_take_all_locks? Can't really imagine a signal torture similar 
to IGT in real life but check on first use makes sense anyway.

Regards,

Tvrtko