Re: [PATCH v4 1/3] drm/panthor: Add panthor_*_irq_mask_set helper

[Boris Brezillon]

On Wed, 17 Dec 2025 15:29:38 +0100
Nicolas Frattaroli <nicolas.frattaroli@xxxxxxxxxxxxx> wrote:

> Add a function to modify an IRQ's mask. If the IRQ is currently active,
> it will write to the register, otherwise it will only set the struct
> member.
> 
> There's no locking done to guarantee exclusion with the other two
> functions that touch the IRQ mask, and it should only be called from a
> context where the circumstances guarantee no concurrent access is
> performed.
> 
> Signed-off-by: Nicolas Frattaroli <nicolas.frattaroli@xxxxxxxxxxxxx>
> ---
>  drivers/gpu/drm/panthor/panthor_device.h | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/drivers/gpu/drm/panthor/panthor_device.h b/drivers/gpu/drm/panthor/panthor_device.h
> index f35e52b9546a..894d28b3eb02 100644
> --- a/drivers/gpu/drm/panthor/panthor_device.h
> +++ b/drivers/gpu/drm/panthor/panthor_device.h
> @@ -470,6 +470,13 @@ static int panthor_request_ ## __name ## _irq(struct panthor_device *ptdev,			\
>  					 panthor_ ## __name ## _irq_threaded_handler,		\
>  					 IRQF_SHARED, KBUILD_MODNAME "-" # __name,		\
>  					 pirq);							\
> +}												\
> +												\
> +static inline void panthor_ ## __name ## _irq_mask_set(struct panthor_irq *pirq, u32 mask)	\
> +{												\
> +	pirq->mask = mask;									\
> +	if (!atomic_read(&pirq->suspended))							\
> +		gpu_write(pirq->ptdev, __reg_prefix ## _INT_MASK, mask);			\

This is racy if called outside the (threaded) IRQ handler, which I
believe is the case since its called when a trace point is enabled:

1. _irq_raw_handler() sets INT_MASK to zero to avoid receiving
interrupts from this IRQ line until the threaded handler has processed
events

2. _irq_mask_set() sets INT_MASK to something non-zero, meaning the
interrupt will be re-enabled before events have been processed

this leads to at least one spurious interrupt being received before we
set INT_MASK to zero again. Probably not the end of the world, but if
we can avoid it, that'd be better.

Also, I'd like to see if we could re-purpose panthor_irq::mask to be
the mask of events the user wants to monitor instead of a pure proxy of
INT_MASK. If we do that, and we make panthor_irq::mask an atomic_t, we
can add panthor_xxx_irq_{enable,disable}_event() helpers that would do
the atomic_{or,and} on panthor_irq::mask, and write the new value to
_INT_MASK if:
- we're processing events in the threaded handler (we would need
  another field, or we'd need to turn suspended into a state that can
  encode more than just "suspended or not")
- the device is not suspended (that test you already have)

>  }
>  
>  extern struct workqueue_struct *panthor_cleanup_wq;
>