On Wed, 2 Sep 2026 16:41:40 +0100
Adrián Larumbe <[email protected]> wrote:

> On 01.09.2026 15:08, Boris Brezillon wrote:
> > On Fri, 28 Aug 2026 21:56:48 +0100
> > Adrián Larumbe <[email protected]> wrote:
> >   
> > > Because MMU interrupts are only enabled when the device is reset, it
> > > happened that after DRM device registration, the very first job targeting
> > > the tiler heap BO would always time out. The reason is the reset sequence
> > > is only part of PM runtime resume, which is not called explicitly at 
> > > driver
> > > probe time, and an actual reset work item manually triggered after a HW
> > > error.
> > > 
> > > I have attempted a somewhat drastic solution, which is completely
> > > decoupling GPU/MMU/JM subsystem initialisation and reset from interrupt
> > > enablement, so that we can handle IRQ toggling a bit more flexibly.
> > > 
> > > To this end:
> > > - Ensure every subsystem with its own IRQ has an 'enable interrupts'
> > > method, and that it doesn't enable them anywhere else.
> > > - Force IRQ masking at MMU reset time. Up until, now, panfrost_mmu_reset()
> > > was clearing the MMU IRQ suspension bit, but at no point that is set 
> > > during
> > > the reset sequence.
> > > 
> > > Then manually enable all interrupts when the device is fully initialised 
> > > at
> > > probe time, right before DRM device registration, or after the reset
> > > sequence is complete. Also disable all interrupts at device remove time,
> > > so that their IRQs can be sync'ed right before tearing the device down.
> > > 
> > > Fixes: 635430797d3f ("drm/panfrost: Rework runtime PM initialization")
> > > Fixes: 876b15d2c88d ("drm/panfrost: Fix module unload")
> > > Signed-off-by: Adrián Larumbe <[email protected]>
> > > ---
> > >  drivers/gpu/drm/panfrost/panfrost_device.c | 40 
> > > ++++++++++++++++++++++--------
> > >  drivers/gpu/drm/panfrost/panfrost_device.h |  3 ++-
> > >  drivers/gpu/drm/panfrost/panfrost_gpu.c    | 19 ++++++++------
> > >  drivers/gpu/drm/panfrost/panfrost_gpu.h    |  2 ++
> > >  drivers/gpu/drm/panfrost/panfrost_job.c    |  7 +++---
> > >  drivers/gpu/drm/panfrost/panfrost_mmu.c    |  9 +++++--
> > >  drivers/gpu/drm/panfrost/panfrost_mmu.h    |  2 ++
> > >  7 files changed, 56 insertions(+), 26 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/panfrost/panfrost_device.c 
> > > b/drivers/gpu/drm/panfrost/panfrost_device.c
> > > index 9e02fb5f73c8..99f7da2180f9 100644
> > > --- a/drivers/gpu/drm/panfrost/panfrost_device.c
> > > +++ b/drivers/gpu/drm/panfrost/panfrost_device.c
> > > @@ -226,6 +226,27 @@ static int panfrost_pm_domain_init(struct 
> > > panfrost_device *pfdev)
> > >   return err;
> > >  }
> > >  
> > > +void panfrost_device_enable_int(struct panfrost_device *pfdev)
> > > +{
> > > + panfrost_gpu_enable_interrupts(pfdev);
> > > + panfrost_mmu_enable_interrupts(pfdev);
> > > + panfrost_jm_enable_interrupts(pfdev);
> > > +}
> > > +
> > > +static void panfrost_device_enable_hw(struct panfrost_device *pfdev)
> > > +{
> > > + panfrost_device_enable_int(pfdev);
> > > + panfrost_devfreq_resume(pfdev);
> > > +}
> > > +
> > > +static void panfrost_device_disable_hw(struct panfrost_device *pfdev)
> > > +{
> > > + panfrost_devfreq_suspend(pfdev);
> > > + panfrost_jm_suspend_irq(pfdev);
> > > + panfrost_mmu_suspend_irq(pfdev);
> > > + panfrost_gpu_suspend_irq(pfdev);  
> > 
> > Hm, I think I'd prefer if those suspend/resume_irq() were hidden in
> > some subcomponent panfrost_<subcomp>_suspend,resume() helpers. And
> > then we just have to resume/suspend component in the right order
> > instead of treating IRQs as a standalone object (enabling/disabling
> > only makes sense if the subcomponent handling those interrupts is
> > resumed/suspended).  
> 
> I thought it would only make sense to enable interupts for a given subsystem
> when all the other subsystems are also resumed or initialised. This was 
> prompted
> by Sashiko warning of the possibility of spurious interrupts causing a handler
> to be run when one of the subsystems it touches on hasn't yet been 
> initialised.

Well, in practice things tend to be well isolated, for instance, an
MMU IRQ should be processed entirely inside panfrost_mmu.c, with no
particular interaction with the other subsystems. So, if an MMU
interrupt fires before, say, the JM subsystem is up and running, that
shouldn't be a problem. In panthor, we have a few cases where events
get propagated between subsystems, and for those we have some
is_initialized checks. I'm not sure this applies to panfrost though.

The other advantage with this approach is that it's one step towards a
better subsystem isolation like we have in panthor, where subsystems
only see their internal state/data plus the general state exposed by
panthor_device, instead of having everything in panfrost_device, and
everyone having the ability to modify/check the state of other
subsystems. panfrost_device.c then just acts as a glue layer that knows
about the order things should be executed in, but doesn't have all the
internal details about subsystem initialization/teardown steps.

Reply via email to