Re: [PATCH v4 06/14] dma-buf/sync_file: Support (E)POLLPRI

[Sebastian Wick]

On Tue, Feb 28, 2023 at 11:52 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
>
> On Tue, Feb 28, 2023 at 6:30 AM Sebastian Wick
> <sebastian.wick@xxxxxxxxxx> wrote:
> >
> > On Tue, Feb 28, 2023 at 12:48 AM Rob Clark <robdclark@xxxxxxxxx> wrote:
> > >
> > > On Mon, Feb 27, 2023 at 2:44 PM Sebastian Wick
> > > <sebastian.wick@xxxxxxxxxx> wrote:
> > > >
> > > > On Mon, Feb 27, 2023 at 11:20 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, Feb 27, 2023 at 1:36 PM Rodrigo Vivi <rodrigo.vivi@xxxxxxxxx> wrote:
> > > > > >
> > > > > > On Fri, Feb 24, 2023 at 09:59:57AM -0800, Rob Clark wrote:
> > > > > > > On Fri, Feb 24, 2023 at 7:27 AM Luben Tuikov <luben.tuikov@xxxxxxx> wrote:
> > > > > > > >
> > > > > > > > On 2023-02-24 06:37, Tvrtko Ursulin wrote:
> > > > > > > > >
> > > > > > > > > On 24/02/2023 11:00, Pekka Paalanen wrote:
> > > > > > > > >> On Fri, 24 Feb 2023 10:50:51 +0000
> > > > > > > > >> Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:
> > > > > > > > >>
> > > > > > > > >>> On 24/02/2023 10:24, Pekka Paalanen wrote:
> > > > > > > > >>>> On Fri, 24 Feb 2023 09:41:46 +0000
> > > > > > > > >>>> Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:
> > > > > > > > >>>>
> > > > > > > > >>>>> On 24/02/2023 09:26, Pekka Paalanen wrote:
> > > > > > > > >>>>>> On Thu, 23 Feb 2023 10:51:48 -0800
> > > > > > > > >>>>>> Rob Clark <robdclark@xxxxxxxxx> wrote:
> > > > > > > > >>>>>>
> > > > > > > > >>>>>>> On Thu, Feb 23, 2023 at 1:38 AM Pekka Paalanen <ppaalanen@xxxxxxxxx> wrote:
> > > > > > > > >>>>>>>>
> > > > > > > > >>>>>>>> On Wed, 22 Feb 2023 07:37:26 -0800
> > > > > > > > >>>>>>>> Rob Clark <robdclark@xxxxxxxxx> wrote:
> > > > > > > > >>>>>>>>
> > > > > > > > >>>>>>>>> On Wed, Feb 22, 2023 at 1:49 AM Pekka Paalanen <ppaalanen@xxxxxxxxx> wrote:
> > > > > > > > >>>>>>
> > > > > > > > >>>>>> ...
> > > > > > > > >>>>>>
> > > > > > > > >>>>>>>>>> On another matter, if the application uses SET_DEADLINE with one
> > > > > > > > >>>>>>>>>> timestamp, and the compositor uses SET_DEADLINE on the same thing with
> > > > > > > > >>>>>>>>>> another timestamp, what should happen?
> > > > > > > > >>>>>>>>>
> > > > > > > > >>>>>>>>> The expectation is that many deadline hints can be set on a fence.
> > > > > > > > >>>>>>>>> The fence signaller should track the soonest deadline.
> > > > > > > > >>>>>>>>
> > > > > > > > >>>>>>>> You need to document that as UAPI, since it is observable to userspace.
> > > > > > > > >>>>>>>> It would be bad if drivers or subsystems would differ in behaviour.
> > > > > > > > >>>>>>>>
> > > > > > > > >>>>>>>
> > > > > > > > >>>>>>> It is in the end a hint.  It is about giving the driver more
> > > > > > > > >>>>>>> information so that it can make better choices.  But the driver is
> > > > > > > > >>>>>>> even free to ignore it.  So maybe "expectation" is too strong of a
> > > > > > > > >>>>>>> word.  Rather, any other behavior doesn't really make sense.  But it
> > > > > > > > >>>>>>> could end up being dictated by how the hw and/or fw works.
> > > > > > > > >>>>>>
> > > > > > > > >>>>>> It will stop being a hint once it has been implemented and used in the
> > > > > > > > >>>>>> wild long enough. The kernel userspace regression rules make sure of
> > > > > > > > >>>>>> that.
> > > > > > > > >>>>>
> > > > > > > > >>>>> Yeah, tricky and maybe a gray area in this case. I think we eluded
> > > > > > > > >>>>> elsewhere in the thread that renaming the thing might be an option.
> > > > > > > > >>>>>
> > > > > > > > >>>>> So maybe instead of deadline, which is a very strong word, use something
> > > > > > > > >>>>> along the lines of "present time hint", or "signalled time hint"? Maybe
> > > > > > > > >>>>> reads clumsy. Just throwing some ideas for a start.
> > > > > > > > >>>>
> > > > > > > > >>>> You can try, but I fear that if it ever changes behaviour and
> > > > > > > > >>>> someone notices that, it's labelled as a kernel regression. I don't
> > > > > > > > >>>> think documentation has ever been the authoritative definition of UABI
> > > > > > > > >>>> in Linux, it just guides drivers and userspace towards a common
> > > > > > > > >>>> understanding and common usage patterns.
> > > > > > > > >>>>
> > > > > > > > >>>> So even if the UABI contract is not documented (ugh), you need to be
> > > > > > > > >>>> prepared to set the UABI contract through kernel implementation.
> > > > > > > > >>>
> > > > > > > > >>> To be the devil's advocate it probably wouldn't be an ABI regression but
> > > > > > > > >>> just an regression. Same way as what nice(2) priorities mean hasn't
> > > > > > > > >>> always been the same over the years, I don't think there is a strict
> > > > > > > > >>> contract.
> > > > > > > > >>>
> > > > > > > > >>> Having said that, it may be different with latency sensitive stuff such
> > > > > > > > >>> as UIs though since it is very observable and can be very painful to users.
> > > > > > > > >>>
> > > > > > > > >>>> If you do not document the UABI contract, then different drivers are
> > > > > > > > >>>> likely to implement it differently, leading to differing behaviour.
> > > > > > > > >>>> Also userspace will invent wild ways to abuse the UABI if there is no
> > > > > > > > >>>> documentation guiding it on proper use. If userspace or end users
> > > > > > > > >>>> observe different behaviour, that's bad even if it's not a regression.
> > > > > > > > >>>>
> > > > > > > > >>>> I don't like the situation either, but it is what it is. UABI stability
> > > > > > > > >>>> trumps everything regardless of whether it was documented or not.
> > > > > > > > >>>>
> > > > > > > > >>>> I bet userspace is going to use this as a "make it faster, make it
> > > > > > > > >>>> hotter" button. I would not be surprised if someone wrote a LD_PRELOAD
> > > > > > > > >>>> library that stamps any and all fences with an expired deadline to
> > > > > > > > >>>> just squeeze out a little more through some weird side-effect.
> > > > > > > > >>>>
> > > > > > > > >>>> Well, that's hopefully overboard in scaring, but in the end, I would
> > > > > > > > >>>> like to see UABI documented so I can have a feeling of what it is for
> > > > > > > > >>>> and how it was intended to be used. That's all.
> > > > > > > > >>>
> > > > > > > > >>> We share the same concern. If you read elsewhere in these threads you
> > > > > > > > >>> will notice I have been calling this an "arms race". If the ability to
> > > > > > > > >>> make yourself go faster does not required additional privilege I also
> > > > > > > > >>> worry everyone will do it at which point it becomes pointless. So yes, I
> > > > > > > > >>> do share this concern about exposing any of this as an unprivileged uapi.
> > > > > > > > >>>
> > > > > > > > >>> Is it possible to limit access to only compositors in some sane way?
> > > > > > > > >>> Sounds tricky when dma-fence should be disconnected from DRM..
> > > > > > > > >>
> > > > > > > > >> Maybe it's not that bad in this particular case, because we are talking
> > > > > > > > >> only about boosting GPU clocks which benefits everyone (except
> > > > > > > > >> battery life) and it does not penalize other programs like e.g.
> > > > > > > > >> job priorities do.
> > > > > > > > >
> > > > > > > > > Apart from efficiency that you mentioned, which does not always favor
> > > > > > > > > higher clocks, sometimes thermal budget is also shared between CPU and
> > > > > > > > > GPU. So more GPU clocks can mean fewer CPU clocks. It's really hard to
> > > > > > > > > make optimal choices without the full coordination between both schedulers.
> > > > > > > > >
> > > > > > > > > But that is even not the main point, which is that if everyone sets the
> > > > > > > > > immediate deadline then having the deadline API is a bit pointless. For
> > > > > > > > > instance there is a reason negative nice needs CAP_SYS_ADMIN.
> > > > > > > > >
> > > > > > > > > However Rob has also pointed out the existence of uclamp.min via
> > > > > > > > > sched_setattr which is unprivileged and can influence frequency
> > > > > > > > > selection in the CPU world, so I conceded on that point. If CPU world
> > > > > > > > > has accepted it so can we I guess.
> > > > > > > > >
> > > > > > > > > So IMO we are back to whether we can agree defining it is a hint is good
> > > > > > > > > enough, be in via the name of the ioctl/flag itself or via documentation.
> > > > > > > > >
> > > > > > > > >> Drivers are not going to use the deadline for scheduling priorities,
> > > > > > > > >> right? I don't recall seeing any mention of that.
> > > > > > > > >>
> > > > > > > > >> ...right?
> > > > > > > > >
> > > > > > > > > I wouldn't have thought it would be beneficial to preclude that, or
> > > > > > > > > assume what drivers would do with the info to begin with.
> > > > > > > > >
> > > > > > > > > For instance in i915 we almost had a deadline based scheduler which was
> > > > > > > > > much fairer than the current priority sorted fifo and in an ideal world
> > > > > > > > > we would either revive or re-implement that idea. In which case
> > > > > > > > > considering the fence deadline would naturally slot in and give true
> > > > > > > > > integration with compositor deadlines (not just boost clocks and pray it
> > > > > > > > > helps).
> > > > > > > > How is user-space to decide whether to use ioctl(SET_DEADLINE) or
> > > > > > > > poll(POLLPRI)?
> > > > > > >
> > > > > > > Implementation of blocking gl/vk/cl APIs, like glFinish() would use
> > > > > > > poll(POLLPRI).  It could also set an immediate deadline and then call
> > > > > > > poll() without POLLPRI.
> > > > > > >
> > > > > > > Other than compositors which do frame-pacing I expect the main usage
> > > > > > > of either of these is mesa.
> > > > > >
> > > > > > Okay, so it looks like we already agreed that having a way to bump frequency
> > > > > > from userspace is acceptable. either because there are already other ways
> > > > > > that you can waste power and because this already acceptable in the CPU
> > > > > > world.
> > > > > >
> > > > > > But why we are doing this in hidden ways then?
> > > > > >
> > > > > > Why can't we have this hint per context that is getting executed?
> > > > > > (either with a boost-context flag or with some low/med/max or '-1' to '1'
> > > > > > value like the latency priority)?
> > > > > >
> > > > > > I don't like the waitboost because this heurisitic fails in some media cases.
> > > > > > I don't like the global setting because we might be alternating a top-priority
> > > > > > with low-priority cases...
> > > > > >
> > > > > > So, why not something per context in execution?
> > > > > >
> > > > >
> > > > > It needs to be finer granularity than per-context, because not all
> > > > > waits should trigger boosting.  For example, virglrenderer ends up
> > > > > with a thread polling unsignaled fences to know when to signal an
> > > > > interrupt to the guest virtgpu.  This alone shouldn't trigger
> > > > > boosting.  (We also wouldn't want to completely disable boosting for
> > > > > virglrenderer.)  Or the usermode driver could be waiting on a fence to
> > > > > know when to do some cleanup.
> > > > >
> > > > > That is not to say that there isn't room for per-context flags to
> > > > > disable/enable boosting for fences created by that context, meaning it
> > > > > could be an AND operation for i915 if it needs to be.
> > > >
> > > > First of all, I believe that the fence deadline hint is a good idea.
> > > > With that being said, I also don't think it is sufficient in a lot of
> > > > cases.
> > > >
> > > > The one thing I was alluding to before and that Pekka mentioned as
> > > > well is that mutter for example has a problem where we're missing the
> > > > deadline consistently because the clocks don't ramp up fast enough and
> > > > there is a MR which is just trying to keep the GPU busy to avoid this.
> > >
> > > the dynamic double/triple buffer thing?
> >
> > Yes
> >
> > > > It would be much better if the kernel could make sure the clocks are
> > > > all ramped up when we start submitting work. In the compositor we
> > > > actually have a lot of information that *should* influence clocks. We
> > > > know when we're going to start submitting work and when the deadline
> > > > for that work is beforehand. We know which windows are visible, and
> > > > which one should have the highest priority.
> > >
> > > This sounds like something orthogonal.. something for cgroups?  Ie.
> > > android moves visible/foreground apps to a different cgroup to given
> > > them higher priority.  Tvrtko had a patchset to add drm cgroup
> > > support..
> >
> > For the priority stuff, yes, probably. The visibility information on
> > the other hand could be used to determine if we want to ramp up the
> > GPU in the first place.
>
> Right, but I think that we could have multiple cgroup based knobs, one
> that adjusts priority and one that limits/disables deadline based
> boost?  This way the compositor could setup different policies for
> visible vs hidden apps influencing both how much time they get on the
> GPU and boost.

I'm not sure if a negative control really makes sense. There are
probably timing sensitive use cases where the result doesn't show up
on the local screen and penalizing them when they are not focused or
hidden might also not be the best idea.

> > > > We know when there are
> > > > input events which actually matter.
> > >
> > > This I see input as a different boost source for the driver.  (Ie. one
> > > boost signal is missing fence deadlines, another is input events,
> > > etc.)
> > >
> > > We end up using downstream input-handlers on the kernel side for this.
> > > Partially for the freq boost (but mostly not, UI interactive workloads
> > > like touchscreen scrolling don't generally need high GPU freqs, they
> > > are more memory bandwidth limited if they are limited by anything)..
> > > really the reason here is to get a head-start on the ~2ms that it
> > > takes to power up the GPU if it is suspended.
> >
> > Right, but one of my main points I want to make here is that we could
> > get the head-start not only in response to input events but also for
> > the GPU work the compositor submits and in the future also to GPU work
> > that clients commit. Except that we don't have a way to tell the
> > kernel about it.
> >
> > > But this is not quite perfect, since for example some keys should be
> > > handled on key-down but others on key-up.
> > >
> > > But again, this is something different from fence deadlines.  I'm
> > > interested in proposals because we do need something for this.  But I
> > > think it is something is orthogonal to this series.  For input, we
> > > want the kernel to know long before userspace is ready to submit
> > > rendering.
> >
> > We can do that in the compositor! Input events are really not
> > something you should care about in the kernel. Input itself is also
> > not the only indication of incoming animated content. Some network
> > packets arriving could equally well result in the same situation.
>
> We do use input boost not just for GPU freq, but also for CPU freq.
> It seems like triggering it from the kernel could happen somewhat
> sooner than userspace.  (But just speculation.)

Technically it has to be sooner but I doubt it makes any difference.
Getting a lot of false-positives on the other hand does make a
difference.

> I'm not sure network packets count.  Or at least compared to a touch
> interface, the user won't perceive the lag nearly as much, compared to
> whether or not the UI tracks their fingertips.

Sure, stutter in interactive cases is especially awful but stutter is
awful in general. If we can solve it in all cases we should do so.

> > > > We know when the deadline for
> > > > client work is.
> > > >
> > > > In the future we also want to make sure clients know beforehand when
> > > > they should start their work and when the deadline is but that's all
> > > > very much WIP in both wayland and vulkan.
> > > >
> > > > There are two issues:
> > > >
> > > > 1. The compositor has no way to communicate any of that information to
> > > > the kernel.
> > > > 2. The only connection to client work the compositor has is a fence to
> > > > the last bit of work that must be done before the deadline after a
> > > > wl_surface.commit.
> > >
> > > If the client isn't using multiple GPUs, a single fence should be
> > > sufficient.  And even if it is, well we still have all the dependency
> > > information on the kernel side.  Ie. drm/sched knows what fences it is
> > > waiting on if it is waiting to schedule the work associated with the
> > > last fence.  It would otherwise require drm/sched to be a bit more
> > > tricky than it is so far in this series.
> > >
> > > But I think the normal dual-gpu case, the app is only dealing with a single GPU?
> >
> > We generally don't know which GPU a client uses though. We know which
> > one we're using and tell the client that the buffer should be
> > compatible with it but that's the extent of information we have, until
> > we get a fence but that fence usually gets to the compositor pretty
> > late. Way too late for the compositor to tell the kernel to ramp up
> > the GPU and still have an impact.
>
> Are you sure about that?  I'd have expected the app to hand off
> fence+buffer to the compositor pretty quickly after rendering is
> submitted.  This is what I've seen elsewhere.

After rendering is submitted it is already too late if the GPU takes
2ms to wake up. And if there is no rendering submitted there is no
fence and if there is no fence it has no deadline. If we want to solve
that we also need some kind of 'work will be submitted to the queue
starting from this time' hint.

> > It also seems like we're moving away from tracking execution
> > dependencies with fences when we're switching to user mode fences.
>
> I suppose there are two cases..
>
> 1) dependent fences all from the same device.. no prob, the right
> driver gets the signal that it needs to clk up, and figures out the
> rest on it's own

AFAIU with user mode fences it's impossible for the kernel to figure
out what work depends on it and it might never signal. The whole
deadline on fences thing breaks down with user mode fences.

> 2) dependent fences from different devices.. I suppose if device B is
> waiting for a fence from device A before it can make forward progress,
> why not express this as a deadline hint on A's fence?  (But by the
> time you have this problem, you are probably not really caring about
> power consumption, so meh..)
>
> BR,
> -R
>
> > > > So in both cases a fence is just not the right primitive for us. We
> > > > need to be able to provide per-context/queue information for work that
> > > > will happen in the future and we need a way to refer to a
> > > > context/queue generically and over IPC to boost the clocks of the
> > > > device that a client is actually using and maybe even give priority.
> > > >
> > > > But like I said, having a per-fence deadline is probably still a good
> > > > idea and doesn't conflict with any of the more coarse information.
> > >
> > > Yeah, I think the thing is you need multiple things, and this is only
> > > one of them ;-)
> > >
> > > BR,
> > > -R
> > >
> >
>