Re: [PATCH v8 07/18] drm/msm/a6xx: Add a helper for software-resetting the GPU

[Akhil P Oommen]

On Thu, Jun 15, 2023 at 12:34:06PM +0200, Konrad Dybcio wrote:
> 
> On 6.06.2023 19:18, Akhil P Oommen wrote:
> > On Mon, May 29, 2023 at 03:52:26PM +0200, Konrad Dybcio wrote:
> >>
> >> Introduce a6xx_gpu_sw_reset() in preparation for adding GMU wrapper
> >> GPUs and reuse it in a6xx_gmu_force_off().
> >>
> >> This helper, contrary to the original usage in GMU code paths, adds
> >> a write memory barrier which together with the necessary delay should
> >> ensure that the reset is never deasserted too quickly due to e.g. OoO
> >> execution going crazy.
> >>
> >> Signed-off-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> >> ---
> >>  drivers/gpu/drm/msm/adreno/a6xx_gmu.c |  3 +--
> >>  drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 11 +++++++++++
> >>  drivers/gpu/drm/msm/adreno/a6xx_gpu.h |  1 +
> >>  3 files changed, 13 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> >> index b86be123ecd0..5ba8cba69383 100644
> >> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> >> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> >> @@ -899,8 +899,7 @@ static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
> >>  	a6xx_bus_clear_pending_transactions(adreno_gpu, true);
> >>  
> >>  	/* Reset GPU core blocks */
> >> -	gpu_write(gpu, REG_A6XX_RBBM_SW_RESET_CMD, 1);
> >> -	udelay(100);
> >> +	a6xx_gpu_sw_reset(gpu, true);
> >>  }
> >>  
> >>  static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
> >> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >> index e3ac3f045665..083ccb5bcb4e 100644
> >> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >> @@ -1634,6 +1634,17 @@ void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu, bool gx_
> >>  	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
> >>  }
> >>  
> >> +void a6xx_gpu_sw_reset(struct msm_gpu *gpu, bool assert)
> >> +{
> >> +	gpu_write(gpu, REG_A6XX_RBBM_SW_RESET_CMD, assert);
> >> +	/* Add a barrier to avoid bad surprises */
> > Can you please make this comment a bit more clear? Highlight that we
> > should ensure the register is posted at hw before polling.
> > 
> > I think this barrier is required only during assert.
> Generally it should not be strictly required at all, but I'm thinking
> that it'd be good to keep it in both cases, so that:
> 
> if (assert)
> 	we don't keep writing things to the GPU if it's in reset
> else
> 	we don't start writing things to the GPU becomes it comes
> 	out of reset
> 
> Also, if you squint hard enough at the commit message, you'll notice
> I intended for this so only be a wmb, but for some reason generalized
> it.. Perhaps that's another thing I should fix!
> for v9..

wmb() doesn't provide any ordering guarantee with the delay loop.
A common practice is to just read back the same register before
the loop because a readl followed by delay() is guaranteed to be ordered.

-Akhil.
> 
> Konrad
> > 
> > -Akhil.
> >> +	mb();
> >> +
> >> +	/* The reset line needs to be asserted for at least 100 us */
> >> +	if (assert)
> >> +		udelay(100);
> >> +}
> >> +
> >>  static int a6xx_pm_resume(struct msm_gpu *gpu)
> >>  {
> >>  	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> >> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> >> index 9580def06d45..aa70390ee1c6 100644
> >> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> >> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> >> @@ -89,5 +89,6 @@ struct msm_gpu_state *a6xx_gpu_state_get(struct msm_gpu *gpu);
> >>  int a6xx_gpu_state_put(struct msm_gpu_state *state);
> >>  
> >>  void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu, bool gx_off);
> >> +void a6xx_gpu_sw_reset(struct msm_gpu *gpu, bool assert);
> >>  
> >>  #endif /* __A6XX_GPU_H__ */
> >>
> >> -- 
> >> 2.40.1
> >>