Intel graphics CPU usage - SDVO detect bogosity?

[Dave Airlie]

On Sun, 2010-08-15 at 21:01 -0700, Linus Torvalds wrote:
> On Sun, Aug 15, 2010 at 8:30 PM, Dave Airlie <airlied at redhat.com> wrote:
> >
> > At least we should replace mdelay with msleep in those functions.
> 
> How precise does the timing have to be? I think i2c is self-clocking,
> so it's ok to see big skews? Becuase msleep() can be off by quite a
> bit (mdelay can too, but it's _way_ more rare, and requires either a
> preemptible kernel or serious interrupt activity).

In this case it definitely doesn't matter, I expect msleep to be a much
nicer to the system in general idea esp in routines we can all at
runtime from userspace.

> 
> > Can you get a boot with drm.debug=4?
> 
> Sure. That results in a truncated dmesg (with a 128k buffer). The
> thing seems to spew out something every ten seconds:
> 
>   ...
>   [  232.610044] [drm:intel_sdvo_debug_write], SDVOB: W: 0B
>              (SDVO_CMD_GET_ATTACHED_DISPLAYS)
>   [  232.624504] [drm:intel_sdvo_debug_response], SDVOB: R: 01 00
>              (Success)
>   [  232.624517] [drm:intel_sdvo_detect], SDVO response 1 0
>   [  232.624524] [drm:intel_sdvo_debug_write], SDVOB: W: 7A 01
>              (SDVO_CMD_SET_CONTROL_BUS_SWITCH)
>   [  242.672044] [drm:intel_sdvo_debug_write], SDVOB: W: 0B
>              (SDVO_CMD_GET_ATTACHED_DISPLAYS)
>   [  242.686503] [drm:intel_sdvo_debug_response], SDVOB: R: 01 00
>              (Success)
>   [  242.686516] [drm:intel_sdvo_detect], SDVO response 1 0
>   [  242.686523] [drm:intel_sdvo_debug_write], SDVOB: W: 7A 01
>              (SDVO_CMD_SET_CONTROL_BUS_SWITCH)
>   [  252.750044] [drm:intel_sdvo_debug_write], SDVOB: W: 0B
>              (SDVO_CMD_GET_ATTACHED_DISPLAYS)
>   [  252.764526] [drm:intel_sdvo_debug_response], SDVOB: R: 01 00
>              (Success)
>   [  252.764539] [drm:intel_sdvo_detect], SDVO response 1 0
>   ...
> 
> It looks like it takes about 15 ms each time.  But 15 ms each 10s
> doesn't seem to be enough to account for the load average. Maybe it
> gets synchronized with the timer tick or something, causing the load
> average to look artificially inflated (it also doesn't match up with
> kworker using 1%+ CPU time).
> 
> So maybe there is something else going on. Maybe the load average
> thing comes from some interaction with the new workqueue thing.
> 
> I'll send the whole dmesg to you in a private message, I don't think
> we want 128kB of crud on lkml.
> 
> > I wonder are we picking up a bad SDVO, the insane code retries 50 times
> > with a hard loop delay.
> 
> Well, I'm not seeing any failures, but maybe those don't get printed out?

Oh wierd, so not where I thought it was, I expect then the SDVO HDMI
detection is completely insane,

intel_sdvo_hdmi_sink_detect looks to contain some really uninspiring
code. Might be worth adding some debug in there to see if it sinks a lot
of time.

Dave.