[RFC] Reduce idle vblank wakeups

[Mario Kleiner]

On Nov 16, 2011, at 6:11 PM, Matthew Garrett wrote:

> On Wed, Nov 16, 2011 at 06:03:15PM +0100, Michel Dänzer wrote:
>
>> I thought the main reason for the delay wasn't broken hardware but to
>> avoid constantly ping-ponging the vblank IRQ between on and off with
>> apps which regularly neeed the vblank counter value, as that could  
>> make
>> the counter unreliable. Maybe I'm misremembering though.
>
> If turning it on and off results in the counter value being wrong then
> surely that's a hardware problem? I've tested that turning it on  
> and off
> between every IRQ still gives valid counter values on sandybridge.
>
> -- 
> Matthew Garrett | mjg59 at srcf.ucam.org


> On Wed, Nov 16, 2011 at 06:03:15PM +0100, Michel Dänzer wrote:


 >Even if I'm not, lowering the delay shouldn't be a problem, so long as
 >it's long enough that at least apps which need the vblank counter  
every
 >or every other frame don't cause the IRQ to ping-pong. But that
 >shouldn't depend on the hardware.

Hi, and thanks Michel for cc'ing me,

It's not broken hardware, but fast ping-ponging it on and off can  
make the vblank counter and vblank timestamps unreliable for apps  
that need high timing precision, especially for the ones that use the  
OML_sync_control extensions for precise swap scheduling. My target  
application is vision science
  neuroscience, where (sub-)milliseconds often matter for visual  
stimulation.

I think making the vblank off delay driver specific via these patches  
is a good idea. Lowering the timeout to something like a few refresh  
cycles, maybe somewhere between 50 msecs and 100 msecs would be also  
fine by me. I still would like to keep some drm config option to  
disable or override the vblank off delay by users.

The intel and radeon kms drivers implement everything that's needed  
to make it mostly work. Except for a small race between the cpu and  
gpu in the vblank_disable_and_save() function <http://lxr.free- 
electrons.com/source/drivers/gpu/drm/drm_irq.c#L101> and  
drm_update_vblank_count(). It can cause an off-by-one error when  
reinitializing the drm vblank counter from the gpu's hardware counter  
if the enable/disable function is called at the wrong moment while  
the gpu's scanout is inside the vblank interval, see comments in the  
code. I have some sketchy idea for a patch that could detect when the  
race happens and retry hw counter queries to fix this. Without that  
patch, there's some chance between 0% and 4% of being off-by-one.

On current nouveau kms, disabling vblank irqs guarantees you wrong  
vblank counts and wrong vblank timestamps after turning them on  
again, because the kms driver doesn't implement the hook for hardware  
vblank counter query correctly. The drm vblank counter misses all  
counts during the vblank irq off period. Other timing related hooks  
are missing as well. I have a couple of patches queued up and some  
more to come for the ddx and kms driver to mostly fix this. NVidia  
gpu's only have hardware vblank counters for NV-50 and later, fixing  
this for earlier gpu's would require some emulation of a hw vblank  
counter inside the kms driver.

Apps that rely on the vblank counter being totally reliable over long  
periods of time currently would be in a bad situation with a lowered  
vblank off delay, but that's probably generally not a good  
assumption. Toolkits like mine, which are more paranoid, currently  
can work fine as long as the off delay is at least a few video  
refresh cycles. I do the following for scheduling a reliably timed swap:

1. Query current vblank counter current_msc and vblank timestamp  
current_ust.
2. Calculate a target vblank count target_msc, based on current_msc,  
current_ust and some target time from usercode.
3. Schedule bufferswap for target_msc.

As long as the vblank off delay is long enough so that vblanks don't  
get turned off between 1. and 3, everything is fine, otherwise bad  
things will happen.
Keeping a way to override the default off delay would be good to  
allow poor scientists to work around potentially broken drivers or  
gpu's in the future. @Matthew: I'm appealing here to your ex-  
Drosophila biologist heritage ;-)

thanks,
-mario


*********************************************************************
Mario Kleiner
Max Planck Institute for Biological Cybernetics
Spemannstr. 38
72076 Tuebingen
Germany

e-mail: mario.kleiner at tuebingen.mpg.de
office: +49 (0)7071/601-1623
fax:    +49 (0)7071/601-616
www:    http://www.kyb.tuebingen.mpg.de/~kleinerm
*********************************************************************
"For a successful technology, reality must take precedence
over public relations, for Nature cannot be fooled."
(Richard Feynman)