On Tue, Nov 29, 2011 at 12:12, Daniel Vetter <daniel at ffwll.ch> wrote: > On Tue, Nov 29, 2011 at 01:05:03PM +0000, Chris Wilson wrote: > > On Tue, 29 Nov 2011 10:55:05 -0200, Eugeni Dodonov < > eugeni.dodonov at intel.com> wrote: > > > This allows to enable/disable rps and rc6 from userspace. This is > > > necessary for to have predictable results from hardware counters, and > also > > > to provide a finer granularity over power control from userspace. > > > > Reading registers and counters is handled with i915_forcewake_user. So > > the question is what benefit does htis give over module reload. It looks > > far riskier... > > The observation architecture bspec says that we need to disable rc6, > otherwise we'll get garbage in the gpu perf counters. But imo I think this > should be handled in the kernel (we neeed to write tons of funny registers > to set things up anyway), at least for the time-based perf counter > sampling. For prototyping things with intel_reg_write disabling rc6 at > module load time seems sufficient to me. > The major idea behind this was to provide additional facilities for controlling power and performance-related details of the driver from userspace, without module reloading or reboot. Currently, both the GPU turbo and rc6/fbc operations are mostly autonomous, there is not much we can influence on their behavior without a reboot or module reload. So my idea was to have a way to toggle - and in some future, ideally, fine-tune their behavior from userspace during the execution. For GPU turbo, we can control it to some point from userspace via i915_max_freq. And with this patch, we could also control the rps/rc6 behavior to some point as well. For the perf counters, I thought on the following flow of execution with this patch: 1. prev_val=$(cat /sys/kernel/debug/dri/0/i915_enable_rc6) 2. echo 0 > /sys/kernel/debug/dri/0/i915_enable_rc6 3. do the perf counters aquisition and testing 4. echo $(prev_val) > /sys/kernel/debug/dri/0/i915_enable_rc6 Comments? -- Eugeni Dodonov <http://eugeni.dodonov.net/> -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.freedesktop.org/archives/intel-gfx/attachments/20111129/48da5170/attachment.htm>
