Re: [PATCH] drm/i915: read/write IOCTLs

[Ben Widawsky <ben@bwidawsk.net>]

On Sat, Apr 02, 2011 at 07:46:31AM +0100, Chris Wilson wrote:
> 
> But perhaps we do need to reconsider the performance aspect. intel_gpu_top
> samples the ring HEAD and TAIL at around 10KHz and forcing gt-wake is
> about 50 microseconds... I hope I'm mistaken, because even batched that is
> doomed. Ben, do you mind checking that thought experiment with a little
> hard fact?

I can get some numbers for it... but I'd like to further the discussion
a bit since I have to go to the office to get a SNB, and typing is
easier than doing that right now :).

I too think we might be doomed. At the very least we have the
POSTING_READ, which is expensive. The udelays may or may not actually
occur (I'll find out). Let's not forget too that we do fix other tools,
not just intel_gpu_top, and those tools don't poll, and don't care about
timing (granted they're mostly less interesting too).

I think what we really need to try to defer the forcewake_put, as well
as something like the last patch I sent
<1301105269-23970-2-git-send-email-ben@bwidawsk.net> to remove the need
to protect force_wake_get with struct_mutex, and keep a refcount.  I'd
rather get the current stuff accepted, port the tools, and then make
improvements to the performance. However if you feel the order must be
another way, I can work with that.

> -Chris
> 

Thanks.
Ben