Re: [Linaro-mm-sig] [PATCH] drm/msm: Add fence->wait() op

[Rob Clark]

On Thu, Jul 22, 2021 at 1:42 AM Christian König
<ckoenig.leichtzumerken@xxxxxxxxx> wrote:
>
> Am 21.07.21 um 21:03 schrieb Daniel Vetter:
> > On Wed, Jul 21, 2021 at 09:34:43AM -0700, Rob Clark wrote:
> >> On Wed, Jul 21, 2021 at 12:59 AM Daniel Vetter <daniel@xxxxxxxx> wrote:
> >>> On Wed, Jul 21, 2021 at 12:32 AM Rob Clark <robdclark@xxxxxxxxx> wrote:
> >>>> On Tue, Jul 20, 2021 at 1:55 PM Daniel Vetter <daniel@xxxxxxxx> wrote:
> >>>>> On Tue, Jul 20, 2021 at 8:26 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
> >>>>>> On Tue, Jul 20, 2021 at 11:03 AM Christian König
> >>>>>> <ckoenig.leichtzumerken@xxxxxxxxx> wrote:
> >>>>>>> Hi Rob,
> >>>>>>>
> >>>>>>> Am 20.07.21 um 17:07 schrieb Rob Clark:
> >>>>>>>> From: Rob Clark <robdclark@xxxxxxxxxxxx>
> >>>>>>>>
> >>>>>>>> Somehow we had neither ->wait() nor dma_fence_signal() calls, and no
> >>>>>>>> one noticed.  Oops.
> >>>>>>>
> >>>>>>> I'm not sure if that is a good idea.
> >>>>>>>
> >>>>>>> The dma_fence->wait() callback is pretty much deprecated and should not
> >>>>>>> be used any more.
> >>>>>>>
> >>>>>>> What exactly do you need that for?
> >>>>>> Well, the alternative is to track the set of fences which have
> >>>>>> signalling enabled, and then figure out which ones to signal, which
> >>>>>> seems like a lot more work, vs just re-purposing the wait
> >>>>>> implementation we already have for non-dma_fence cases ;-)
> >>>>>>
> >>>>>> Why is the ->wait() callback (pretty much) deprecated?
> >>>>> Because if you need it that means for your driver dma_fence_add_cb is
> >>>>> broken, which means a _lot_ of things don't work. Like dma_buf poll
> >>>>> (compositors have patches to start using that), and I think
> >>>>> drm/scheduler also becomes rather unhappy.
> >>>> I'm starting to page back in how this works.. fence cb's aren't broken
> >>>> (which is also why dma_fence_wait() was not completely broken),
> >>>> because in retire_submits() we call
> >>>> dma_fence_is_signaled(submit->hw_fence).
> >>>>
> >>>> But the reason that the custom wait function cleans up a tiny bit of
> >>>> jank is that the wait_queue_head_t gets signaled earlier, before we
> >>>> start iterating the submits and doing all that retire_submit() stuff
> >>>> (unpin/unref bo's, etc).  I suppose I could just split things up to
> >>>> call dma_fence_signal() earlier, and *then* do the retire_submits()
> >>>> stuff.
> >>> Yeah reducing the latency there sounds like a good idea.
> >>> -Daniel
> >>>
> >> Hmm, no, turns out that isn't the problem.. or, well, it is probably a
> >> good idea to call drm_fence_signal() earlier.  But it seems like
> >> waking up from wait_event_* is faster than wake_up_state(wait->task,
> >> TASK_NORMAL).  I suppose the wake_up_state() approach still needs for
> >> the scheduler to get around to schedule the runnable task.
>
> As far as I know wake_up_state() tries to run the thread on the CPU it
> was scheduled last, while wait_event_* makes the thread run on the CPU
> who issues the wake by default.
>
> And yes I've also noticed this already and it was one of the reason why
> I suggested to use a wait_queue instead of the hand wired dma_fence_wait
> implementation.
>
> >>
> >> So for now, I'm going back to my own wait function (plus earlier
> >> drm_fence_signal())
> >>
> >> Before removing dma_fence_opps::wait(), I guess we want to re-think
> >> dma_fence_default_wait().. but I think that would require a
> >> dma_fence_context base class (rather than just a raw integer).
> > Uh that's not great ... can't we fix this instead of papering over it in
> > drivers? Aside from maybe different wakeup flags it all is supposed to
> > work exactly the same underneath, and whether using a wait queue or not
> > really shouldn't matter.
>
> Well it would have been nicer if we used the existing infrastructure
> instead of re-inventing stuff for dma_fence, but that chance is long gone.
>
> And you don't need a dma_fence_context base class, but rather just a
> flag in the dma_fence_ops if you want to change the behavior.

Hmm, I was thinking dma_fence_context to have a place for the
wait_queue_head, but I guess that could also be per-dma_fence