Re: [PATCH] drm/i915: Enforce TYPESAFE_BY_RCU vs refcount mb on reinitialisation

[Mika Kuoppala <mika.kuoppala@linux.intel.com>]

Chris Wilson <chris@chris-wilson.co.uk> writes:

> By using TYPESAFE_BY_RCU, we accept that requests may be swapped out from
> underneath us, even when using rcu_read_lock(). We use a strong barrier
> on acquiring the refcount during lookup, but this needs to be paired
> with a barrier on re-initialising it. Currently we call dma_fence_init,
> which ultimately does a plain atomic_set(1) on the refcount, not
> providing any memory barriers. As we inspect some state before even
> acquiring the refcount in the lookup (by arguing that we can detect
> inconsistent requests), that state should be initialised before the
> refcount.
>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_request.c | 7 +++++++
>  1 file changed, 7 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
> index 5c2c93cbab12..04a0b8e75533 100644
> --- a/drivers/gpu/drm/i915/i915_request.c
> +++ b/drivers/gpu/drm/i915/i915_request.c
> @@ -768,6 +768,13 @@ i915_request_alloc(struct intel_engine_cs *engine, struct i915_gem_context *ctx)
>  	rq->timeline = ce->ring->timeline;
>  	GEM_BUG_ON(rq->timeline == &engine->timeline);
>  
> +	/*
> +	 * In order to coordinate with our RCU lookup,
> +	 * __i915_gem_active_get_rcu(), we need to ensure that the change
> +	 * to rq->engine is visible before acquring the refcount in the lookup.
> +	 */
> +	smp_wmb();
> +

There is quite a lot going on here as we try to get a reference
into a shapeshifting request.

By looking at the code acquiring it, dma_fence_get_rcu
and dma_fence_init and then the precheck of the request,
should memory barrier be:

smb_mb_before_atomic()?

Admittedly that would be uglier as fence_init hides the atomic_set,
but it is atomic on we are serializing. Especially
as there is no atomic in callsight.

Further, as engine and the kref are tightly bound,
should we initialize everything not related first, then
do engine init, wmb, fence init in a tight proximity?

Thanks,
-Mika

>  	spin_lock_init(&rq->lock);
>  	dma_fence_init(&rq->fence,
>  		       &i915_fence_ops,
> -- 
> 2.18.0
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx