[PATCH] lib/xe/spin: Use READ_ONCE, WRITE_ONCE macros in Xe spin library

[Zbigniew Kempczyński]

On Wed, Apr 16, 2025 at 09:10:25AM -0700, Matthew Brost wrote:
> Without the READ_ONCE macro is xe_spin_started(), a compiler may choose
> to read the value of spin->start exactly once rather than every call of
> xe_spin_started(). If the initial read is before the GPU batch execution
> flips spin->start to a non-zero value, the test will hang forever. This
> was shown on Chrome.

I missed this explanation about Chrome. May you point out the affected
code? I'm really curious about how it looks like.

--
Zbigniew


> 
> For uniformity, update xe_spin_end to use WRITE_ONCE too.
> 
> Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> ---
>  lib/xe/xe_spin.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/xe/xe_spin.c b/lib/xe/xe_spin.c
> index 0de0b1f2e9..65c29284aa 100644
> --- a/lib/xe/xe_spin.c
> +++ b/lib/xe/xe_spin.c
> @@ -145,7 +145,7 @@ void xe_spin_init(struct xe_spin *spin, struct xe_spin_opts *opts)
>   */
>  bool xe_spin_started(struct xe_spin *spin)
>  {
> -	return spin->start != 0;
> +	return READ_ONCE(spin->start) != 0;
>  }
>  
>  /**
> @@ -162,7 +162,7 @@ void xe_spin_wait_started(struct xe_spin *spin)
>  
>  void xe_spin_end(struct xe_spin *spin)
>  {
> -	spin->end = 0;
> +	WRITE_ONCE(spin->end, 0);
>  }
>  
>  /**
> -- 
> 2.34.1
>