[Intel-gfx] [PATCH] drm: Aggressively disable vblanks

[Mario Kleiner]

On Dec 26, 2010, at 3:53 PM, Andrew Lutomirski wrote:

> On Wed, Dec 22, 2010 at 4:06 PM, Mario Kleiner
> <mario.kleiner at tuebingen.mpg.de> wrote:
>>
>> There's a new drm module parameter for selecting the timeout: echo  
>> 50 >
>> /sys/module/drm/parameters/vblankoffdelay
>> would set the timeout to 50 msecs. A setting of zero will disable  
>> the timer,
>> so vblank irq's would stay on all the time.
>>
>> The default setting is still 5000 msecs as before, but reducing  
>> this to 100
>> msecs wouldn't be a real problem imho. At least i didn't observe any
>> miscounting during extensive testing with 100 msecs.
>>
>> The patches in drm-next fix a couple of races that i observed on  
>> intel and
>> radeon during testing and a few that i didn't see but that i could  
>> imagine
>> happening. It tries to make sure that the saved final count at  
>> vblank irq
>> disable of the software vblank_count and the gpu counter are  
>> consistent - no
>> off by one errors. They also try to detect and filter out spurious  
>> vblank
>> interrupts at vblank enable time, e.g., on the radeon.
>>
>> There's still one possible race in the disable path which i will  
>> try to fix:
>> We don't know when exactly the hardware counter increments wrt. the
>> processing of the vblank interrupt - it could increment a few
>> (dozen/hundred) microseconds before or after the irq handler runs,  
>> so if you
>> happen to query the hardware counter while the gpu is inside the  
>> vblank you
>> can't be sure if you picked up the old count or the new count for  
>> that
>> vblank.
>
> That's disgusting.  Does this affect many GPUs?  (I can't imagine why
> any sensible implementation wouldn't guarantee that the counter
> increments just before the IRQ.)

;-). I don't know, but at least on the tested R500 and R600 class  
Radeon's, this was the case, so i assume it's at least this way on  
many radeon gpu's (probably all avivo parts?) out there. We don't  
have any evergreen gpu's yet in our lab so i don't know how the more  
recent parts behave. Also it doesn't matter

I guess it's also a matter of definition when a new video frame  
starts? Leading edge / trailing edge of vblank? Start of vsync?  
Something else?

>>
>> This only matters during vblank disable. For that reason it's not  
>> such a
>> good idea to disable vblank irq's from within the vblank irq  
>> handler. I
>> tried that and it didn't work well --> When doing it from within  
>> irq you
>> basically maximize the chance of hitting the time window when the  
>> race can
>> happen. Delaying within the irq handler by a millisecond would fix  
>> that, but
>> that's not what we want.
>>
>> Having the disable in a function triggered by a timer like now is  
>> the most
>> simple solution i could come up with. There we can burn a few dozen
>> microseconds if neccessary to remove this race.
>
> Maybe I'm missing something, but other than the race above (which
> seems like it shouldn't exist on sane hardware), the new code seems
> more complicated than necessary.

I don't think it's more complicated than necessary for what it tries  
to achieve, but of course i'm a bit biased. It also started off more  
simple and grew a bit when i found new issues with the tested gpu's.

The aim is to fix a couple of real races and to make vblank counts  
and timestamps as trustworthy and oml_sync_control spec-conformant  
(see http://www.opengl.org/registry/specs/OML/glx_sync_control.txt)  
as possible. It only consumes a few additional bytes of memory  
(approx. 40 bytes) per crtc, doesn't use excessive time inside the  
irq handler and tries to avoid taking locks  that are shared between  
irq and non-irq context to avoid delays in irq execution, also if  
used with a kernel with the preempt_rt patch applied (important for  
my use case and other hard realtime apps). It's pretty self-contained  
and because most of it is driver-independent it can handle similar  
issues on different gpu's and kms drivers without the need for us to  
check each single gpu + driver combo if it has such issues or not.

1. There's the hardware vblank counter race, and it's there on lots  
of existing hardware, regardless if this is sane or not, so it needs  
to be handled.

2. There are gpu's firing spurious vblank irq's as soon as you enable  
irq's -- you need to filter those out, otherwise counts and  
timestamps will be wrong for at least one refresh cycle after vblank  
irq enable. You don't know when these redundant ones get delivered or  
if they get delivered at all because a real vblank irq enable gets  
triggered by some userspace calls, not locked to the video refresh  
cycle and because the enable code itself holds spin_lock_irqsave  
locks and may or may not (depending on number of cores and irq  
routing) prevent delivery of the vblank irq's concurrent to its own  
execution -> a possible race between the drm_handle_vblank() routine  
and the drm_update_vblank_count() routine each time you call  
drm_vblank_get().

3. There's gpu's sometimes, but not on other times, firing the irq  
too early (1-3 scanlines observed on radeon's) so you get your irq  
before the associated vblank interval and need to do at least all  
timestamping as if you are already in that vblank. This may be  
dependent on both video mode and on the dispatch delay of the vblank  
irq handler (e.g., due to some other unrelated code holding off  
irq's, e.g., by holding spin_lock_irqsave() locks).

4. In the old code it could happen that the vblank counter gets  
incremented after it was "disabled", e.g., because vblank irq's were  
turned off in the gpu, but there was still a vblank irq pending  
(e.g., due to some spin_lock_irqsave on the relevant core),  
incrementing the counter after it was supposedly frozen. With the new  
oml_sync_control timestamping in place, such off-by-ones would be  
worse as they would also corrupt timestamps.

There were some more issues which i can't remember from the top of my  
head which get handled by the current code (Note to myself: Take more  
notes!).

>
> Why not do nothing at all on vblank disable and, on enable, do this:
>
> Call get_vblank_counter and declare the return value to be correct.

Because declaring it to be correct isn't the same as it being  
correct. Also the code needs to handle wraparound of the hardware  
counter and for that it needs correct start values from the vblank  
disable routine, which is why the disable routine needs to work as  
race-free as possible. The vblank count for a frame also needs to be  
consistent with the vblank timestamp for that frame at all times,  
otherwise the oml_sync_control extension becomes too unreliable to be  
trustworthy and useful for serious applications.

> On each vblank IRQ (or maybe just the first one after enable), read
> the counter again and if it matches the cached value, then assume we
> raced on enable (i.e. we enabled right at the beginning of the vblank
> interval, and this interrupt is redundant).  In that case, do nothing.
>  Otherwise increment the cached value by one.
>
> On hardware with the silly race where get_vblank_counter in the IRQ
> can return a number one too low, use the scanout pos to fix it (i.e.
> if the scanout position is at the end of the frame, assume we hit that
> race and increment by 1).
>

See the other races above. Iirc i tried something similar already and  
they made it fail/unreliable. The current patch uses the vblank  
timestamp instead of the hardware vblank counter to detect and filter  
redundant irq's, because with the timestamping patches at least on  
intel and ati gpu's (and hopefully nouveau/nvidia and others asap)  
these are well defined and accurate (to define the end of a vblank  
interval) so they can serve as a reference point. The tbd fix for  
race condition #1 will also use scanout position as a fixed reference.

The sample client code (below) for scheduling accurately timed  
bufferswaps needs precise and trustworthy return values from  
glXGetSyncValuesOML() for it to work. If vblank's are disabled at  
time of invocation of glXGetSyncValuesOML() then that function will  
trigger the real vblank irq enable path and use its return values for  
swap scheduling - i.e. unless it is called within or close to a  
vblank, it uses the vblank count and timestamp computed in  
drm_update_vblank_count(), usually before the vblank irq had a chance  
to run. For that reason it is important for my kind of applications  
that it really delivers the right counts and timestamps especially in  
the enable path.

This is the context in case you're interested why i'm so protective  
of the current implementation: The toolkit i'm developing is probably  
one of the currently most demanding clients of the new dri2 swap &  
sync bits and it is used for neuroscience research. Many of the  
experiments there require very precise visual timing and often sub- 
millisecond accurate timestamping. Too many unhandled races in the  
wrong places could really spoil the work of the scientists that are  
my users. As long as we have a disable timeout of 5 seconds, vblank  
irq's probably won't disable at all during most experiment sessions  
and even if they do, the frequency of possible screwups is probably  
small enough to be annoying but manageable with statistics (detecting/ 
discarding outliers in experimental results etc.). At a disable  
timeout of around 50 msecs, the error rate would be unbearable.  
That's why i would like to make sure that the implementation can  
handle at least the already known quirks of a large number of the  
gpu's out there if we go down to 50 msecs. But i assume races in that  
code would affect the quality of "normal" media players as well, once  
we choose such low timeouts.

Thanks and belated happy x-mas,
-mario

> This means that it would be safe to disable vblanks from any context
> at all, because it has no effect other than turning off the interupt.
>
> --Andy
>
>>
>> There could be other races that i couldn't think of and that i  
>> didn't see
>> during my testing with my 2 radeons and 1 intel gpu. Therefore i  
>> think we
>> should keep vblank irq's enabled for at least 2 or maybe 3 refresh  
>> cycles if
>> they aren't used by anyone. Apps that want to schedule swaps very  
>> precisely
>> and the ddx/drm/kms-driver itself do multiple queries in quick  
>> succession
>> for a typical swapbuffers call, i.e., drm_vblank_get() -> query ->
>> drm_vblank_put(), so on an otherwise idle graphics system the  
>> refcount will
>> toggle between zero and one multiple times during a swap, usually  
>> within a
>> few milliseconds. If the timeout is big enough so that irq's don't  
>> get
>> disabled within such a sequence of toggles, even if there's a bit of
>> scheduling delay for the x-server or client, then a client will  
>> see at least
>> consistent vblank count during a swap, even if there are still  
>> some races
>> hiding somewhere that we don't handle properly. That should be  
>> good enough,
>> and paranoid clients can always increase the timeout value or set  
>> it to
>> infinite.
>>
>> E.g., my toolkit schedules a swap for a specific system time like  
>> this:
>>
>> 1. glXGetSyncValuesOML(... &base_msc, &base_ust);
>> 2. calculate target_msc based on user provided swap deadline t and
>> (base_msc, base_ust) as a baseline.
>> 3. glXSwapBuffersMscOML(...., target_msc,...);
>> 4. glXWaitForSbcOML() or use Intel_swap_events for retrieving the  
>> true msc
>> and ust of swap completion.
>>
>> => Doesn't matter if there would be an off-by-one error in vblank  
>> counting
>> due to an unknown race, as long as it doesn't happen between 1.  
>> and 4. As
>> long as there aren't any client/x-server scheduling delays between  
>> step 1
>> and 3 of more than /sys/module/drm/parameters/vblankoffdelay  
>> msecs, nothing
>> can go wrong even if there are race conditions left in that area.
>>
>> => 50-100 msecs as new default would be probably good enough and  
>> at the same
>> time prevent the "blinking cursor keeps vblank irq's on all the time"
>> problem.
>>
>> I didn't reduce the timeout in the current patches because the  
>> filtering for
>> race-conditions and other gpu funkyness relies on somewhat precise  
>> vblank
>> timestamps and the timestamping hooks aren't yet implemented in  
>> the nouveau
>> kms. Maybe i manage to get this working over christmas. Patches to  
>> nouveau
>> would be simple, i just don't know the mmio register addresses for  
>> crtc
>> scanout position on nvidia gpu's.
>>
>> -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)
>>
>>

*********************************************************************
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)