Date: Mon, 16 Nov 2015 15:30:06 +0200
From: Ville =?iso-8859-1?Q?Syrj=E4l=E4?= <ville.syrjala@linux.intel.com>
To: Chris Wilson <chris@chris-wilson.co.uk>,
        Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>,
        Jens Axboe <axboe@kernel.dk>, intel-gfx@lists.freedesktop.org,
        linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org,
        Daniel Vetter <daniel.vetter@ffwll.ch>,
        Eero Tamminen <eero.t.tamminen@intel.com>,
        "Rantala, Valtteri" <valtteri.rantala@intel.com>,
        stable@kernel.vger.org
Subject: Re: [Intel-gfx] [PATCH 2/2] drm/i915: Limit the busy wait on
 requests to 2us not 10ms!
Message-ID: <20151116133006.GM4437@intel.com>
References: <1447594364-4206-1-git-send-email-chris@chris-wilson.co.uk>
 <1447594364-4206-2-git-send-email-chris@chris-wilson.co.uk>
 <5649AEED.9090807@linux.intel.com>
 <20151116111208.GQ569@nuc-i3427.alporthouse.com>
 <5649C728.5040109@linux.intel.com>
 <20151116125537.GS569@nuc-i3427.alporthouse.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20151116125537.GS569@nuc-i3427.alporthouse.com>
User-Agent: Mutt/1.5.23 (2014-03-12)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3709
Lines: 94

On Mon, Nov 16, 2015 at 12:55:37PM +0000, Chris Wilson wrote:
> On Mon, Nov 16, 2015 at 12:08:08PM +0000, Tvrtko Ursulin wrote:
> > 
> > On 16/11/15 11:12, Chris Wilson wrote:
> > >On Mon, Nov 16, 2015 at 10:24:45AM +0000, Tvrtko Ursulin wrote:
> > >>On 15/11/15 13:32, Chris Wilson wrote:
> > >>>+static u64 local_clock_us(unsigned *cpu)
> > >>>+{
> > >>>+	u64 t;
> > >>>+
> > >>>+	*cpu = get_cpu();
> > >>>+	t = local_clock() >> 10;
> > >>
> > >>Needs comment I think to explicitly mention the approximation, or
> > >>maybe drop the _us suffix?
> > >
> > >I did consider _approx_us but thought that was overkill. A comment along
> > >the lines of
> > >/* Approximately convert ns to us - the error is less than the
> > >  * truncation!
> > >  */
> > 
> > And the result is not used in subsequent calculations apart from
> > comparing against an approximate timeout?
> 
> Exactly, the timeout is fairly arbitrary and defined in the same units.
> That we truncate is a much bigger cause for concern in terms of spinning
> accurately for a definite length of time.
>  
> > >>>@@ -1161,7 +1183,7 @@ static int __i915_spin_request(struct drm_i915_gem_request *req, int state)
> > >>>  		if (signal_pending_state(state, current))
> > >>>  			break;
> > >>>
> > >>>-		if (time_after_eq(jiffies, timeout))
> > >>>+		if (busywait_stop(timeout, cpu))
> > >>>  			break;
> > >>>
> > >>>  		cpu_relax_lowlatency();
> > >>>
> > >>
> > >>Otherwise looks good. Not sure what would you convert to 32-bit from
> > >>your follow up reply since you need us resolution?
> > >
> > >s/u64/unsigned long/ s/time_after64/time_after/
> > >
> > >32bits of us resolution gives us 1000s before wraparound between the two
> > >samples. And I hope that a 1000s doesn't pass between loops. Or if it does,
> > >the GPU managed to complete its task.
> > 
> > Now I see that you did say low bits.. yes that sounds fine.
> > 
> > Btw while you are optimizing things maybe pick up this micro
> > optimization: http://patchwork.freedesktop.org/patch/64339/
> > 
> > Not in scope of this thread but under the normal development patch flow.
> 
> There's a different series which looks at tackling the scalabiltiy issue
> with dozens of concurrent waiters. I have an equivalent patch there and
> one to tidy up the seqno query.
>  
> > Btw2, any benchmark result changes with this?
> 
> Spinning still gives the dramatic (2x) improvement in the microbenchmarks
> (over pure interrupt driven waits), so that improvement is preserved.

Previously the spinning also increased power consumption without
offering any significant performance difference for some workloads.
IIRC on my BYT the average CPU power consumption was ~100mW higher
(as reported by RAPL) with xonotic the-big-keybench.dem (1920x1200
w/ "High" settings, IIRC) but average fps wasn't improved. Might
be interesting to know how the improved spin code stacks up on
that front.

> There are a couple of interesting swings in the macro tests (comparedt to
> the previous jiffie patch) just above the noise level which could well be
> a change in the throttling/scheduling. (And those tests are also the
> ones that correspond to the greatest gains (10-40%) using spinning.)
> -Chris
> 
> -- 
> Chris Wilson, Intel Open Source Technology Centre
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Ville Syrj?l?
Intel OTC
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/