__i915_spin_request() sucks

[Jens Axboe]

On 11/12/2015 03:52 PM, Jens Axboe wrote:
> On 11/12/2015 03:19 PM, Chris Wilson wrote:
>>>> So today, I figured I'd try just killing that spin. If it fails, we'll
>>>> punt to normal completions, so easy change. And wow, MASSIVE
>>>> difference.
>>>> I can now scroll in chrome and not rage! It's like the laptop is 10x
>>>> faster now.
>>>>
>>>> Ran git blame, and found:
>>>>
>>>> commit 2def4ad99befa25775dd2f714fdd4d92faec6e34
>>>> Author: Chris Wilson <chris at chris-wilson.co.uk>
>>>> Date:   Tue Apr 7 16:20:41 2015 +0100
>>>>
>>>>      drm/i915: Optimistically spin for the request completion
>>>>
>>>> and read the commit message. Doesn't sound that impressive. Especially
>>>> not for something that screws up interactive performance by a LOT.
>>>>
>>>> What's the deal? Revert?
>>
>> The tests that it improved the most were the latency sensitive tests and
>> since my Broadwell xps13 behaves itself, I'd like to understand how it
>> culminates in an interactivity loss.
>>
>> 1. Maybe it is the uninterruptible nature of the polling, making X's
>> SIGIO jerky:
>
> This one still feels bad.
>
>> 2. Or maybe it is increased mutex contention:
>
> And so does this one... I had to manually apply hunks 2-3, and after
> doing seat-of-the-pants testing for both variants, I confirmed with perf
> that we're still seeing a ton of time in __i915_wait_request() for both
> of them.

I don't see how #2 could make any difference, you're passing in 0x3 hard 
coded for most call sites, so we poll. The ones that don't, pass a bool 
(?!).

I should note that with the basic patch of just never spinning, I don't 
see __i915_wait_request() in the profiles. At all.

-- 
Jens Axboe