[Intel-gfx] Responsiveness Changes to i915 Driver

Chris Wilson chris at chris-wilson.co.uk
Fri Aug 15 08:26:28 CEST 2014


On Thu, Aug 14, 2014 at 11:52:47PM +0000, Wilde, Martin wrote:
> Greetings,
> 
> I am submitting the below changes to i915 Gfx driver to support resume time Responsiveness measurements.  These changes parallel the work already done in the IVB Windows Gfx driver.  These changes in addition to other OTS scripts (suspend_resume) allow tracking of what is referred to as the “B2I” or “Button To Image” time of the platform.  The shorter this time, the more responsiveness the platform is viewed by the end user.  Panel selection is an important factor in providing a more responsive system.  Note there is no dependency on other scripts.  The changes are standalone.
> 
> 
>   *   Display the current T1_T3 value.  This is used to verify that the timing set in the VBT is correct.  We have seen many instances where the value is not set correctly for the panel and the resume time is longer than necessary (e.g. 500ms T3 versus 200ms T3).
>   *   Print the time when the first page flip occurs.  This is when the user first sees the desktop displayed from resume.  While this measurement could be done by other methods, this is the actual time that the desktop manager/framebuffer makes the driver request and the Gfx driver performs the action.  Thus any layering software added can be correlated to increases in this time.
> 
> To support the latter (first page flip), I added a new ftrace called “trace_i915_resume”.  I looked at the existing page flip trace message and that one is designed for every page flip.  I did not want to convolute it with the one time flip trace on resume.  I used a trace message instead of a printk to reduce any performance impacts of using a printk.  Additionally printk is not reliable of when the message actually appears in the kernel log.

I am sorry, you are complaining that there is a tracepoint that gives
you exactly what you want already, only that userspace needs to do some
filtering? Which by the way, does not give you what you say you want
anyway - the scanout is already active long before the first flip is
handled, and in many, many cass that flip is just a figment of your
imagination. Maybe what you mean is to B2UR rather than
button-to-static-image.
-Chris


-- 
Chris Wilson, Intel Open Source Technology Centre



More information about the Intel-gfx mailing list