[Intel-gfx] [PATCH v2] drm/i915/gt: update request engine before removing virtual GuC engine

[Andi Shyti]

Hi Andrzej,

On Thu, Jul 06, 2023 at 05:16:11PM +0200, Andrzej Hajda wrote:
> GuC virtual engines can be removed before request removal. On the other
> side driver expects rq->engine to be a valid pointer for a whole life of
> request. Setting rq->engine to an always valid engine should solve
> the issue.
> 
> The patch fixes bug detected by KASAN with following signature:
> i915 0000:00:02.0: [drm:i915_drop_caches_set [i915]] Dropping caches: 0x0000005c [0x0000005c]
> BUG: KASAN: slab-use-after-free in i915_fence_release+0x2a2/0x2f0 [i915]
> Read of size 4 at addr ffff88813ffda6e8 by task kworker/u32:10/157
> ...
> Allocated by task 1184:
> ...
> guc_create_virtual+0x4d/0x1160 [i915]
> i915_gem_create_context+0x11ee/0x18c0 [i915]
> ...
> Freed by task 151:
> ...
> intel_guc_deregister_done_process_msg+0x324/0x6d0 [i915]
> ...

so the only difference between v1 and v2 is this part of the log?

> Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/7926
> Signed-off-by: Andrzej Hajda <andrzej.hajda at intel.com>
> ---
>  drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> index a0e3ef1c65d246..2c877ea5eda6f0 100644
> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
> @@ -3461,6 +3461,8 @@ static void guc_prio_fini(struct i915_request *rq, struct intel_context *ce)
>  static void remove_from_context(struct i915_request *rq)
>  {
>  	struct intel_context *ce = request_to_scheduling_context(rq);
> +	struct intel_engine_cs *engine;
> +	intel_engine_mask_t tmp;
>  
>  	GEM_BUG_ON(intel_context_is_child(ce));
>  
> @@ -3478,6 +3480,15 @@ static void remove_from_context(struct i915_request *rq)
>  
>  	atomic_dec(&ce->guc_id.ref);
>  	i915_request_notify_execute_cb_imm(rq);
> +
> +	/*
> +	 * GuC virtual engine can disappear after this call, so let's assign
> +	 * something valid, as driver expects this to be always valid pointer.
> +	 */
> +	for_each_engine_masked(engine, rq->engine->gt, rq->execution_mask, tmp) {
> +		rq->engine = engine;

yes... here the context might lose the virtual engine... I wonder
whether this is the rigth solution, though. Maybe we should set
rq->engine = NULL; and check for NULL? Don't know.

Andi

> +		break;
> +	}
>  }
>  
>  static const struct intel_context_ops guc_context_ops = {
> -- 
> 2.34.1