Forgot to Cc lkml.
On Mon, 14 Oct 2019, Davidlohr Bueso wrote:
>Because it is not a Rmw operation, atomic_set() is never serialized,
>and therefore the 'upgradable' smp_mb__{before,after}_atomic() calls
>that order the write to preempt_state are completely bogus.
>
>This patch replaces these with smp_mb(), which seems like the
>original intent of when the code was written.
>
>Signed-off-by: Davidlohr Bueso <[email protected]>
>---
> drivers/gpu/drm/msm/adreno/a5xx_preempt.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
>diff --git a/drivers/gpu/drm/msm/adreno/a5xx_preempt.c b/drivers/gpu/drm/msm/adreno/a5xx_preempt.c
>index 9cf9353a7ff1..d27d8d3208c6 100644
>--- a/drivers/gpu/drm/msm/adreno/a5xx_preempt.c
>+++ b/drivers/gpu/drm/msm/adreno/a5xx_preempt.c
>@@ -30,10 +30,10 @@ static inline void set_preempt_state(struct a5xx_gpu *gpu,
> * preemption or in the interrupt handler so barriers are needed
> * before...
> */
>- smp_mb__before_atomic();
>+ smp_mb();
> atomic_set(&gpu->preempt_state, new);
> /* ... and after*/
>- smp_mb__after_atomic();
>+ smp_mb();
> }
>
> /* Write the most recent wptr for the given ring into the hardware */
>--
>2.16.4
>