[Cbe-oss-dev] [PATCH] spusched: extend list_mutex coverage

[Jeremy Kerr]

Christoph,

> spu->ctx is only stable while we hold the list_mutex is held.

We only update spu->ctx with both list_mutex and ctx->state mutex, so 
I'm a little unclear about how the dangling reference could happen as 
we hold the victim's state_mutex in find_victim(), and the target 
context's state_mutex in spusched_tick().

Could you add a description of the sequence to the changelog?

> @@ -659,16 +659,15 @@ static struct spu *find_victim(struct sp
>  				 * restart the search.
>  				 */
>  				mutex_unlock(&victim->state_mutex);
> +				mutex_unlock(&cbe_spu_info[node].list_mutex);
>  				victim = NULL;
>  				goto restart;
>  			}

This entire if-block should no longer be possible if we hold list_mutex

> @@ -875,8 +880,13 @@ static noinline void spusched_tick(struc
>  	struct spu_context *new = NULL;
>  	struct spu *spu = NULL;
>
> -	if (spu_acquire(ctx))
> -		BUG();	/* a kernel thread never has signals pending */
> +	/*
> +	 * If we can't lock the context give it another timeslice.
> +	 */
> +	if (!mutex_trylock(&ctx->state_mutex)) {
> +		ctx->time_slice++;

No need to increment time_slice, as this will give the context two extra 
timeslices.

Cheers,


Jeremy