Re: __i915_spin_request() sucks

From: Jens Axboe <axboe@kernel.dk>
To: Chris Wilson <chris@chris-wilson.co.uk>,
	Mike Galbraith <umgwanakikbuti@gmail.com>,
	Daniel Vetter <daniel.vetter@ffwll.ch>,
	DRI Development <dri-devel@lists.freedesktop.org>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: __i915_spin_request() sucks
Date: Fri, 13 Nov 2015 15:16:11 -0700	[thread overview]
Message-ID: <5646612B.20805@kernel.dk> (raw)
In-Reply-To: <20151113221212.GJ569@nuc-i3427.alporthouse.com>

On 11/13/2015 03:12 PM, Chris Wilson wrote:
> On Fri, Nov 13, 2015 at 09:22:52AM -0700, Jens Axboe wrote:
>> On 11/13/2015 09:13 AM, Mike Galbraith wrote:
>>> On Fri, 2015-11-13 at 08:36 -0700, Jens Axboe wrote:
>>>> Previous patch was obvious pre-coffee crap, this should work for using
>>>> ktime to spin max 1usec.
>>>
>>> 1us seems a tad low.  I doubt the little wooden gears and pulleys of my
>>> core2 Toshiba Satellite lappy can get one loop ground out in a usec :)
>>
>> Maybe it is, it's based off the original intent of the function,
>> though. See the original commit referenced.
>
> I've been looking at numbers from one laptop and I can set the timeout
> at 2us before we see a steep decline in what is more or less synchronous
> request handling (which affects a variety of rendering workloads).

Alright, at least that's a vast improvement from 10ms. If you send me 
something tested, I can try it here.

> Looking around, other busy loops seem to use local_clock() (i.e. rdstcll
> with a fair wind). Is that worth using here?

Honestly, don't think it matters too much for this case. You'd have to 
disable preempt to use local_clock(), fwiw. It is a faster variant 
though, but the RT people might hate you for 2us preempt disables :-)

-- 
Jens Axboe