On Fri, 13 Nov 2026 14:47:34 +0000
[email protected] wrote:

> Thank you for your contribution! Sashiko AI review found 2 potential
> issues to consider:
> - [Medium] drm/tyr: prepared clocks leaked on error in
>   TyrPlatformDriver::probe()
> - [Low] drm/tyr: overly restrictive &'reset self lifetime in
>   ResetHandle::schedule()
> --
> 
> commit 2f1b0845121531379549eca9ca27fd1ffc58562f
> Author: Onur Özkan <[email protected]>
> 
> drm/tyr: add GPU reset infrastructure
> 
> This commit adds GPU reset infrastructure using a dedicated ordered
> workqueue and an SRCU-based hardware access gate, and drains pending
> or running reset work on teardown.
> 
> Link: https://gitlab.freedesktop.org/panfrost/linux/-/work_items/28
> 
> > diff --git a/drivers/gpu/drm/tyr/driver.rs b/drivers/gpu/drm/tyr/driver.rs
> > index 5b57bc9a1d3e7..863c8008a132b 100644
> > --- a/drivers/gpu/drm/tyr/driver.rs
> > +++ b/drivers/gpu/drm/tyr/driver.rs
> [ ... ]
> > @@ -171,8 +153,11 @@ fn probe<'bound>(
> >                  _reg <- unsafe { drm::Registration::new(
> >                      pdev.as_ref(),
> >                      unreg_dev,
> > -                    pin_init!(TyrDrmRegistrationData {
> > +                    try_pin_init!(TyrDrmRegistrationData {
> >                          pdev,
> > +                        // SAFETY: The registration is owned by platform 
> > driver data,
> > +                        // so reset work is drained before borrowed data 
> > expires.
> > +                        reset <- reset::ResetHandle::new(pdev, 
> > iomem.as_arc_borrow())?,
> >                          fw: firmware,
> >                          clks <- new_mutex!(Clocks {
> >                              core: core_clk,
> 
> [Severity: Medium]
> Does this leak the prepared and enabled clocks if ResetHandle::new() or
> its PinInit initializer fails?
> 
> Drop for Clk and OptionalClk only calls clk_put(), while
> disable_unprepare() is only called in Clocks::drop():
> 
> drivers/gpu/drm/tyr/driver.rs:Clocks::drop() {
>     fn drop(&mut self) {
>         self.core.disable_unprepare();
>         self.stacks.disable_unprepare();
>         self.coregroup.disable_unprepare();
>     }
> }
> 
> Because TyrDrmRegistrationData now initializes reset via try_pin_init!
> before constructing Clocks, a failure in Queue::build() or HwGate::new()
> (as well as earlier error paths after prepare_enable() in probe()) drops
> core_clk, stacks_clk, and coregroup_clk without disabling or unpreparing
> them, leaving the hardware clocks enabled and triggering a warning in
> clk_put().  Wrapping the clocks in Clocks immediately after
> prepare_enable() would ensure they are cleaned up on all error paths.

I can fix this by guarding each enabled clock with ScopeGuard but IIUC this will
no longer be necessary once Daniel's Clk series [1] is applied. If we plan to
land this soon I can still do it. Otherwise, it's pointless.

@Daniel thoughts ?

[1]: 
https://lore.kernel.org/rust-for-linux/[email protected]

> 
> > diff --git a/drivers/gpu/drm/tyr/reset.rs b/drivers/gpu/drm/tyr/reset.rs
> > new file mode 100644
> > index 0000000000000..640371b3104da
> > --- /dev/null
> > +++ b/drivers/gpu/drm/tyr/reset.rs
> [ ... ]
> > +pub(crate) struct ResetHandle<'reset> {
> > +    #[pin]
> > +    controller: ScopedWork<Controller<'reset>>,
> > +    wq: OwnedQueue,
> > +}
> [ ... ]
> > +    /// Schedules a GPU reset on the dedicated workqueue.
> > +    ///
> > +    /// If a reset is already pending or in progress the call is a no-op.
> > +    #[expect(dead_code)]
> > +    pub(crate) fn schedule(&'reset self) {
> > +        // TODO: Similar to `panthor_device_schedule_reset()` in Panthor, 
> > add a
> > +        // power management check once Tyr supports it.
> > +
> > +        if self
> > +            .controller
> > +            .try_transition(ResetState::Idle, ResetState::Pending)
> > +        {
> > +            let _ = self.wq.enqueue_scoped(&self.controller);
> > +        }
> > +    }
> 
> [Severity: Low]
> Can ResetHandle::schedule() take &self instead of &'reset self?
> 
> ScopedWork wraps its payload in Opaque (UnsafeCell), making
> ResetHandle<'reset> invariant over 'reset, where 'reset is the device
> binding lifetime in TyrDrmRegistrationData<'drm>.  Requiring &'reset self
> prevents callers holding a normal short-lived borrow of
> TyrDrmRegistrationData or ResetHandle from calling schedule(), and
> neither try_transition() nor enqueue_scoped() needs the reference to live
> for 'reset.
> 
> -- 
> Sashiko AI review · 
> https://sashiko.dev/#/patchset/[email protected]?part=2

Reply via email to