Re: [REPOST PATCH 4/8] android: convert sync to fence api, v5

[Daniel Vetter]

On Thu, Jun 19, 2014 at 08:35:29AM -0700, Colin Cross wrote:
> On Thu, Jun 19, 2014 at 5:28 AM, Daniel Vetter <daniel@xxxxxxxx> wrote:
> > On Thu, Jun 19, 2014 at 1:48 PM, Thierry Reding
> > <thierry.reding@xxxxxxxxx> wrote:
> >>> > With these changes, can we pull the android sync logic out of
> >>> > drivers/staging/ now?
> >>>
> >>> Afaik the google guys never really looked at this and acked it. So I'm not
> >>> sure whether they'll follow along. The other issue I have as the
> >>> maintainer of gfx driver is that I don't want to implement support for two
> >>> different sync object primitives (once for dma-buf and once for android
> >>> syncpts), and my impression thus far has been that even with this we're
> >>> not there.
> >>>
> >>> I'm trying to get our own android guys to upstream their i915 syncpts
> >>> support, but thus far I haven't managed to convince them to throw people's
> >>> time at this.
> >>
> >> This has been discussed a fair bit internally recently and some of our
> >> GPU experts have raised concerns that this may result in seriously
> >> degraded performance in our proprietary graphics stack. Now I don't care
> >> very much for the proprietary graphics stack, but by extension I would
> >> assume that the same restrictions are relevant for any open-source
> >> driver as well.
> >>
> >> I'm still trying to fully understand all the implications and at the
> >> same time get some of the people who raised concerns to join in this
> >> discussion. As I understand it the concern is mostly about explicit vs.
> >> implicit synchronization and having this mechanism in the kernel will
> >> implicitly synchronize all accesses to these buffers even in cases where
> >> it's not needed (read vs. write locks, etc.). In one particular instance
> >> it was even mentioned that this kind of implicit synchronization can
> >> lead to deadlocks in some use-cases (this was mentioned for Android
> >> compositing, but I suspect that the same may happen for Wayland or X
> >> compositors).
> >
> > Well the implicit fences here actually can't deadlock. That's the
> > entire point behind using ww mutexes. I've also heard tons of
> > complaints about implicit enforced syncing (especially from opencl
> > people), but in the end drivers and always expose unsynchronized
> > access for specific cases. We do that in i915 for upload buffers and
> > other fun stuff. This is about shared stuff across different drivers
> > and different processes.
> >
> > I also expect that i915 will loose implicit syncing in a few upcoming
> > hw modes because explicit syncing is a more natural fit there.
> >
> > All that isn't about the discussion at hand imo since no matter what
> > i915 needs to have on internal representation for a bit of gpu work,
> > and afaics right now we don't have that. With this patch android
> > syncpts use Maarten's fences internally, but I can't freely exchange
> > one for the other. So in i915 I still expect to get stuck with both of
> > them, which is one too many.
> >
> > The other issue (and I haven't dug into details that much) I have with
> > syncpts are some of the interface choices. Apparently you can commit a
> > fence after creation (or at least the hw composer interface works like
> > that) which means userspace can construct deadlocks with android
> > syncpts if I'm not super careful in my driver. I haven't seen any
> > generic code to do that, so I presume everyone just blindly trusts
> > surface-flinger to not do that. Speaks of the average quality of an
> > android gfx driver if the kernel is less trusted than the compositor
> > in userspace ...
> 
> Android sync is designed not to allow userspace to deadlock the
> kernel, a sync_pt should only get created by the kernel when it has
> received a chunk of work that it expects to complete in the near
> future.  The CONFIG_SW_SYNC_USER driver violates that by allowing
> userspace to create and signal arbitrary sync points, but that is
> intended only for testing sync.

Ok, that makes sense. As long as we sufficiently taint the kernel and hide
the sw_sync framework we should be good. I was confused by the hw composer
interface spec which seemed to suggest that the fences for a screen update
completion should be attached before surfaceflinger commits the state. But
I never looked at an implemention so guess that impression is wrong.

> > There's a few other things like exposing timestamps (which are tricky
> > to do right, our driver is littered with wrong attempts) and other
> > details.
> 
> Timestamps are necessary for vsync synchronization to reduce the frame latency.

I'm not against timestamps (we have them for drm vblank events too after
all), just would like for them to be optional. And we need to give
userspace very clear indication which hw clock the timestamp was based on
(or whether we're using the clock_monotonic system clock) to make sure
debug and profiling tools can properly align things. Because hw clocks
always get out of sync. Last time I've looked at the syncpt ioctls that
part was missing.

> > Finally I've never seen anyone from google or any android product guy
> > push a real driver enabling for syncpts to upstream, and google itself
> > has a bit a history of constantly exchanging their gfx framework for
> > the next best thing. So I really doubt this is worthwhile to pursue in
> > upstream with our essentially eternal api guarantees. At least until
> > we see serious uptake from vendors and gfx driver guys. Unfortunately
> > the Intel android folks are no exception here and haven't pushed
> > anything like this in my direction yet at all. Despite multiple pokes
> > from my side.
> 
> As far as I know, every SoC vendor that supports android is using sync
> now, but none of them have succeeded in pushing their drivers upstream
> for a variety of other reasons (interfaces only used by closed source
> userspaces, KMS/DRM vs ADF, ION, etc.).

Yeah I know, and it double-frustrates me that I haven't yet managed to
drag our own team into the public. Except for a few bolt-on hacks to make
deadlines the android driver is the upstream one ...

> > So from my side I think we should move ahead with Maarten's work and
> > figure the android side out once there's real interest.
> 
> As long as that doesn't involve removing the Android sync interfaces
> from staging until dma fence fds are supported, that's fine with me.

Nah, definitely not asking for that. I'd just like to have a real
implemention in a real driver merged upstream and the userpsace/kernel ABI
reviewed a bit before I want to sign up for something we'll need to keep
working forever. E.g. I also think we should expose the actual waiting
through poll and friends for neater integration with the usual display
toolkit event loops. egl shys away from such platform specific stuff, but
we can do it.

Like I've said I want to have to deal with one fence primitive in the
lower levels of my driver for scheduling and synchronization and all that.
The fence proposal fits the bill from my pov since the implicit
synchronization with dma-bufs is optional. But the integration with
android syncpts seems to not yet be there really.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/