Re: [PATCH 1/2] adreno: Shutdown the GPU properly

[Joel Fernandes]

On Thu, Dec 1, 2022 at 10:06 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
>
> On Thu, Dec 1, 2022 at 12:08 PM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> >
> > On Sat, Nov 12, 2022 at 6:44 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
> > >
> > > On Fri, Nov 11, 2022 at 1:08 PM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > >
> > > >
> > > > > On Nov 11, 2022, at 2:50 PM, Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > During kexec on ARM device, we notice that device_shutdown() only calls
> > > > > pm_runtime_force_suspend() while shutting down the GPU. This means the GPU
> > > > > kthread is still running and further, there maybe active submits.
> > > > >
> > > > > This causes all kinds of issues during a kexec reboot:
> > > > >
> > > > > Warning from shutdown path:
> > > > >
> > > > > [  292.509662] WARNING: CPU: 0 PID: 6304 at [...] adreno_runtime_suspend+0x3c/0x44
> > > > > [  292.509863] Hardware name: Google Lazor (rev3 - 8) with LTE (DT)
> > > > > [  292.509872] pstate: 80400009 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> > > > > [  292.509881] pc : adreno_runtime_suspend+0x3c/0x44
> > > > > [  292.509891] lr : pm_generic_runtime_suspend+0x30/0x44
> > > > > [  292.509905] sp : ffffffc014473bf0
> > > > > [...]
> > > > > [  292.510043] Call trace:
> > > > > [  292.510051]  adreno_runtime_suspend+0x3c/0x44
> > > > > [  292.510061]  pm_generic_runtime_suspend+0x30/0x44
> > > > > [  292.510071]  pm_runtime_force_suspend+0x54/0xc8
> > > > > [  292.510081]  adreno_shutdown+0x1c/0x28
> > > > > [  292.510090]  platform_shutdown+0x2c/0x38
> > > > > [  292.510104]  device_shutdown+0x158/0x210
> > > > > [  292.510119]  kernel_restart_prepare+0x40/0x4c
> > > > >
> > > > > And here from GPU kthread, an SError OOPs:
> > > > >
> > > > > [  192.648789]  el1h_64_error+0x7c/0x80
> > > > > [  192.648812]  el1_interrupt+0x20/0x58
> > > > > [  192.648833]  el1h_64_irq_handler+0x18/0x24
> > > > > [  192.648854]  el1h_64_irq+0x7c/0x80
> > > > > [  192.648873]  local_daif_inherit+0x10/0x18
> > > > > [  192.648900]  el1h_64_sync_handler+0x48/0xb4
> > > > > [  192.648921]  el1h_64_sync+0x7c/0x80
> > > > > [  192.648941]  a6xx_gmu_set_oob+0xbc/0x1fc
> > > > > [  192.648968]  a6xx_hw_init+0x44/0xe38
> > > > > [  192.648991]  msm_gpu_hw_init+0x48/0x80
> > > > > [  192.649013]  msm_gpu_submit+0x5c/0x1a8
> > > > > [  192.649034]  msm_job_run+0xb0/0x11c
> > > > > [  192.649058]  drm_sched_main+0x170/0x434
> > > > > [  192.649086]  kthread+0x134/0x300
> > > > > [  192.649114]  ret_from_fork+0x10/0x20
> > > > >
> > > > > Fix by calling adreno_system_suspend() in the device_shutdown() path.
> > > > >
> > > > > Cc: Rob Clark <robdclark@xxxxxxxxxxxx>
> > > > > Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
> > > > > Cc: Ricardo Ribalda <ribalda@xxxxxxxxxxxx>
> > > > > Cc: Ross Zwisler <zwisler@xxxxxxxxxx>
> > > > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > > > ---
> > > > > drivers/gpu/drm/msm/adreno/adreno_device.c | 5 ++++-
> > > > > 1 file changed, 4 insertions(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/drivers/gpu/drm/msm/adreno/adreno_device.c b/drivers/gpu/drm/msm/adreno/adreno_device.c
> > > > > index 24b489b6129a..f0cff62812c3 100644
> > > > > --- a/drivers/gpu/drm/msm/adreno/adreno_device.c
> > > > > +++ b/drivers/gpu/drm/msm/adreno/adreno_device.c
> > > > > @@ -607,9 +607,12 @@ static int adreno_remove(struct platform_device *pdev)
> > > > >    return 0;
> > > > > }
> > > > >
> > > > > +static int adreno_system_suspend(struct device *dev);
> > > > > static void adreno_shutdown(struct platform_device *pdev)
> > > > > {
> > > > > -    pm_runtime_force_suspend(&pdev->dev);
> > > > > +    struct msm_gpu *gpu = dev_to_gpu(&pdev->dev);
> > > > > +
> > > >
> > > > This local variable definition should go to patch 2/2. Will fix in v2.
> > > >
> > > > Thanks,
> > > >
> > > >  - Joel
> > > >
> > > >
> > > > > +    WARN_ON_ONCE(adreno_system_suspend(&pdev->dev));
> > >
> > > I think maybe adreno_unbind() needs the same treatment?  Any path
> > > where we yank out the power cord without ensuring the scheduler is
> > > parked means we'd be racing with jobs in the scheduler queue.  Ie.
> > > userspace could queue a job before it is frozen, but the drm/scheduler
> > > kthread hasn't yet called the msm_job_run() callback (which does
> > > various touching of the now powered off hw).  So I think we need to
> > > ensure that the scheduler is parked in all paths that call
> > > pm_runtime_force_suspend() (as that bypasses the runpm reference that
> > > would otherwise unsure the hw is powered before msm_job_run pokes at
> > > registers)
> >
> > a6xx_gmu_remove() calls pm_runtime_force_suspend() , would that need a
> > treatment too?
> >
> > Though, adreno_system_suspend() is a static function in adreno_device.cc
>
> Naw, you get there indirectly from adreno_unbind()

Ah gotcha, thanks.

 - Joel