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

From: Joel Fernandes
Date: Thu Dec 01 2022 - 17:14:10 EST


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