On Wed, May 20, 2026 at 1:09 AM Boris Brezillon
<[email protected]> wrote:
>
> On Tue, 19 May 2026 14:04:47 -0700
> Chia-I Wu <[email protected]> wrote:
>
> > On Tue, May 19, 2026 at 1:45 PM Chia-I Wu <[email protected]> wrote:
> > >
> > > On Tue, May 19, 2026 at 11:26 AM Boris Brezillon
> > > <[email protected]> wrote:
> > > >
> > > > On Tue, 19 May 2026 10:16:26 -0700
> > > > Chia-I Wu <[email protected]> wrote:
> > > >
> > > > > On Tue, May 19, 2026 at 12:53 AM Boris Brezillon
> > > > > <[email protected]> wrote:
> > > > > >
> > > > > > On Mon, 18 May 2026 16:33:20 -0700
> > > > > > Chia-I Wu <[email protected]> wrote:
> > > > > >
> > > > > >
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > >         if (!ptdev->scheduler)
> > > > > > > > > > > >                 return;
> > > > > > > > > > > >
> > > > > > > > > > > > -       atomic_or(events, &ptdev->scheduler->fw_events);
> > > > > > > > > > > > -       sched_queue_work(ptdev->scheduler, fw_events);
> > > > > > > > > > > > +       
> > > > > > > > > > > > guard(spinlock_irqsave)(&ptdev->scheduler->events_lock);
> > > > > > > > > > > > +
> > > > > > > > > > > > +       if (events & JOB_INT_GLOBAL_IF) {
> > > > > > > > > > > > +               sched_process_global_irq_locked(ptdev);
> > > > > > > > > > > > +               events &= ~JOB_INT_GLOBAL_IF;
> > > > > > > > > > > > +       }
> > > > > > > > > > > > +
> > > > > > > > > > > > +       while (events) {
> > > > > > > > > > > > +               u32 csg_id = ffs(events) - 1;
> > > > > > > > > > > > +
> > > > > > > > > > > > +               sched_process_csg_irq_locked(ptdev, 
> > > > > > > > > > > > csg_id);
> > > > > > > > > > > > +               events &= ~BIT(csg_id);
> > > > > > > > > > > > +       }
> > > > > > > > > > > This handles all fw events in the irq context. Are there 
> > > > > > > > > > > concerns that
> > > > > > > > > > > it may take too long? I might be wrong, but it seems 
> > > > > > > > > > > possible to
> > > > > > > > > > > handle only CSG_SYNC_UPDATE and defer the rest as before.
> > > > > > > > > >
> > > > > > > > > > I started with just the SYNC_UPDATE processing done in the 
> > > > > > > > > > hard-irq
> > > > > > > > > > context, but after auditing the other stuff done in the 
> > > > > > > > > > handler, I
> > > > > > > > > > realized it's basically just deferring all actual 
> > > > > > > > > > processing to work
> > > > > > > > > > items. Yes, there's the overhead of demuxing the events 
> > > > > > > > > > from the
> > > > > > > > > > ack/req regs, but part of this is already done to get to 
> > > > > > > > > > SYNC_UPDATE
> > > > > > > > > > anyway, so at this point we're probably better off demuxing 
> > > > > > > > > > everything
> > > > > > > > > > and scheduling works for all kind of events.
> > > > > > > > > >
> > > > > > > > > > I also compared the perfs between the two approaches 
> > > > > > > > > > (though I didn't
> > > > > > > > > > do as much testing as I did with the new version, so I 
> > > > > > > > > > might have
> > > > > > > > > > missed something), and it didn't seem to matter at all, 
> > > > > > > > > > because the
> > > > > > > > > > interrupts we receive the most are SYNC_UPDATE and IDLE 
> > > > > > > > > > events, and
> > > > > > > > > > those are at the same level.
> > > > > > > > > Looking at ftrace irq events, when there is one active csg,
> > > > > > > > > panthor-job takes 6us (median) / 17us (95%) / 27us (slowest).
> > > > > > > > >
> > > > > > > > > I don't have a good sense if that's considered normal in 
> > > > > > > > > hardirq. But
> > > > > > > > > if that is ever an issue, and if the majority of the time is 
> > > > > > > > > spent in
> > > > > > > > > CSG_SYNC_UPDATE anyway, we can always revert the last patch 
> > > > > > > > > to move
> > > > > > > > > processing to threaded handler.
> > > > > > > >
> > > > > > > > Actually, the threaded -> hard transition (patch 9) is where 
> > > > > > > > the perf
> > > > > > > > gain is.
> > > > > > > hardirq is even more timely for sure. For our use case, the 
> > > > > > > threaded
> > > > > > > handler is RT and is also good enough.
> > > > > >
> > > > > > Yeah, true. I forgot you were forcing RT priority on threaded 
> > > > > > handlers.
> > > > > > Anyway, let's stick to hardirqs for now, and revisit it if it 
> > > > > > proves to
> > > > > > be too much work done in irq context.
> > > > > Just want to clarify that irq_thread calls sched_set_fifo to make the
> > > > > task RT. The behavior is universal and is not specific to any
> > > > > downstream kernel.
>
> There's a difference in what RT means depending on whether the system
> is configured with PREEMPT or PREEMPT_RT though. But I assume you're
> using PREEMPT not PREEMPT_RT.
Right. What I meant for an RT task is that the task uses an RT
scheduling policy (rt_policy).

>
> > > >
> > > > Hm, interesting. In my testing, any of the changes before patch 9
> > > > didn't make a huge difference in term of perf, patch 9 is where the perf
> > > > gains happen. For the record, patch 6 is where we get rid of the
> > > > threaded -> work round-trip for job completion/fence signaling, and it
> > > > didn't seem to reflect in the benchmark results, but I'll do another
> > > > round of tests before posting v3, just to confirm.
> > > We care the most about signaling latency for this series.
>
> Yes, I know. It's just that it also seemed to help the throughput, which
> I initially checked to make sure we were not regressing perfs
> significantly by interrupting the system aggressively. I guess the
> reason for that is that, by reducing the latency, we also unleash the
> job submitter (if you get signaled early, and jobs tend to be
> serialized because of deps, you can submit more).
Yeah, that's likely the case.

Although, if the benchmark is supposed to be GPU-bound and is not
poorly-written, we might also want to check if the userspace driver
adds unncecessary deps between jobs. A trace should give us some
insight.

> > > I collected
> > > some numbers with baseline, with this series, and with patch 9
> > > reverted at 
> > > https://gitlab.freedesktop.org/panfrost/linux/-/work_items/85#note_3481308.
> > > Reposting the numbers here for reference
> > >
> > > |                    | baseline | entire series | patch 9 reverted |
> > > | -                  | -        | -             | -                |
> > > | frag job median    | 2.8ms    | 2.2ms         | 2.2ms            |
> > > | frag job 95%       | 4.5ms    | 2.8ms         | 2.8ms            |
> > > | frag job 99%       | 4.9ms    | 2.8ms         | 2.8ms            |
> > > | panthor-job median | 0.8us    | 6.2us         | 0.9us            |
> > > | panthor-job 95%    | 1.5us    | 16.6us        | 1.5us            |
> > > | panthor-job 99%    | 1.6us    | 28.0us        | 1.8us            |
> >
> > panthor-job rows are the durations of the raw irq handlers, collected
> > from irq/irq_handler_{entry,exit}.
> >
> > frag job rows are the durations from frag jobs, collected from
> > gpu_scheduler/drm_sched_job_{run,done}.
> >
> > The fence signaling paths of them are
> >
> >  - baseline: raw handler -> rt threaded handler -> wq job -> wq job ->
> > fence signal
> >  - entire series: raw handler -> fence signal
> >  - patch 9 reverted: raw handler -> rt threaded handler -> fence signal
>
> Just did another set of throughput tests, and I confirm the gains are
> noticeable only with patch 9 applied (that's on rk3588, which embeds a
> G610, so not the exact same setup). As an example, on
> gfxbench/gl_manhattan, I get the following score bump 2391 -> 2457.
>
> Now I need to set things up to measure latency like you did and make
> sure I'm observing the same thing: threaded handlers providing roughly
> the same latency as hardirq handlers. If not it probably has to do with
> some config options that differ and change the preemptability of the
> system.
>
> I'll hold off on the submission of v3 until this is done, because if
> threaded handlers are roughly as efficient as hardirq ones, we probably
> want to stick to threaded handlers.

Reply via email to