[Intel-gfx] [RFC] [Patch] [DRM/I915] :Add a debufs I/F to dump the I915 register

[Jesse Barnes]

On Wed, 10 Jun 2009 09:06:47 +0800
yakui_zhao <yakui.zhao at intel.com> wrote:

> On Wed, 2009-06-10 at 06:35 +0800, Jesse Barnes wrote:
> > On Tue, 09 Jun 2009 15:16:53 -0700
> > Eric Anholt <eric at anholt.net> wrote:
> > 
> > > On Mon, 2009-06-08 at 08:52 +0800, yakui_zhao wrote:
> > > > On Fri, 2009-06-05 at 19:11 +0800, Eric Anholt wrote:
> > > > > On Fri, 2009-06-05 at 15:45 +0800, yakui_zhao wrote:
> > > > > > It is useful to get the register snapshot. 
> > > > > > Add a debugfs I/F named "i915_reg" to dump the I915 register
> > > > > > snapshot. And this is created under the dri/0/ of debugfs. 
> > > > > > The output format is similar to what we have done in UMS
> > > > > > mode.
> > > > > 
> > > > > I don't think that all the decode and formatting of these
> > > > > registers should be in the kernel.  Every time I've had to
> > > > > mess with register decode stuff for investigation, I've
> > > > > needed to extend the decode. Instead, we should expose the
> > > > > raw register values and make intel_reg_dumper in
> > > > > intel-gpu-tools that does the decode of actual meaning.
> > > > Sometimes we can see the register snapshot without using the
> > > > intel-gpu-tools. For example: in UMS mode we often get the
> > > > register snapshot several times in starting X. 
> > > > 
> > > > It will be good that we expose the raw register values and then
> > > > they are parsed by intel-gpu-tools if we need to extend the
> > > > decode.
> > > > 
> > > > How about adding two debugfs I/F? One is to dump the raw
> > > > register snapshot(without decode and format). Another is what I
> > > > have done in the patch.
> > > 
> > > No, I won't pull something that puts the decode in the kernel
> > > without a really good argument.  Expose the register names and
> > > values.
> > > 
> > 
> > Yakui, have you experimented with just dumping the whole mmio range?
> > If that works ok we could just make intel-gpu-tools programs to
> > interpret the data in a chipset specific way...
> It seems that this is a good point. 
> In this patch only the registers related with modesetting are dumped.
> 
> Is it ok that the whole mmio range is divided into several parts? 
> a. memory interface, instruction and interrupt control registers
> b. 3d debug registers
> c. modeset registers(display,pipe, etc).

Yeah that would be ok too...

Jesse