Re: enabling forcewake from userspace

[Ben Widawsky <ben@bwidawsk.net>]

On Sat, Mar 26, 2011 at 07:11:55AM +0000, Chris Wilson wrote:
> On Fri, 25 Mar 2011 19:23:58 -0700, Ben Widawsky <ben@bwidawsk.net> wrote:
> > The only other option floating around is IOCTLs to read/write the
> > registers, which previously seemed like a bad idea, but is now looking
> > like the only option.
> 
> The lazy option is to cross your fingers and do the forced-wake from
> userspace debug tools. I'm not exactly wild about the idea, but it does
> cut down on the number of blessed interfaces for touching hardware.
> 
> The other option is to continue to extend the debugfs to print out groups
> of registers of interest. Maybe this would be a good addition to the
> i915_*_ringbuffer_info?
This still leaves the reg write problem unsolved. But I think this is
better.

> 
> An ioctl does seem to be a better longterm solution. But maybe we can
> ignore the problem for a few more years? ;-)
I think I won't bother trying to upstream my solution. But I'm also
thinking we should make intel_reg_read/write print a warning/error if
it's running on GEN6? What do you think?


By the way, I do think the patch is beneficial just for kernel usage.
Right now it doesn't matter, but as more users of the interface pop up,
particularly in any time sensitive code... wouldn't it be awful if the
GPU could powerdown while we're servicing an interrupt?

(The patch probably doesn't need the poll when forceawake_ack is already
set, that was a copy/paste mistake.)

> -Chris

Ben