[PATCH 1/2] drm/amd/display: Avoid race between dcn10_set_drr() and dc_state_destruct()

[Christopher Snowhill]

On Mon Sep 2, 2024 at 2:40 AM PDT, tjakobi wrote:
> From: Tobias Jakobi <tjakobi at math.uni-bielefeld.de>
>
> dc_state_destruct() nulls the resource context of the DC state. The pipe
> context passed to dcn10_set_drr() is a member of this resource context.
>
> If dc_state_destruct() is called parallel to the IRQ processing (which
> calls dcn10_set_drr() at some point), we can end up using already nulled
> function callback fields of struct stream_resource.
>
> The logic in dcn10_set_drr() already tries to avoid this, by checking tg
> against NULL. But if the nulling happens exactly after the NULL check and
> before the next access, then we get a race.
>
> Avoid this by copying tg first to a local variable, and then use this
> variable for all the operations. This should work, as long as nobody
> frees the resource pool where the timing generators live.
>
> Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3142
> Fixes: 06ad7e164256 ("drm/amd/display: Destroy DC context while keeping DML and DML2")
> Signed-off-by: Tobias Jakobi <tjakobi at math.uni-bielefeld.de>
> ---
>  .../amd/display/dc/hwss/dcn10/dcn10_hwseq.c   | 20 +++++++++++--------
>  1 file changed, 12 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/display/dc/hwss/dcn10/dcn10_hwseq.c b/drivers/gpu/drm/amd/display/dc/hwss/dcn10/dcn10_hwseq.c
> index 3306684e805a..da8f2cb3c5db 100644
> --- a/drivers/gpu/drm/amd/display/dc/hwss/dcn10/dcn10_hwseq.c
> +++ b/drivers/gpu/drm/amd/display/dc/hwss/dcn10/dcn10_hwseq.c
> @@ -3223,15 +3223,19 @@ void dcn10_set_drr(struct pipe_ctx **pipe_ctx,
>  	 * as well.
>  	 */
>  	for (i = 0; i < num_pipes; i++) {
> -		if ((pipe_ctx[i]->stream_res.tg != NULL) && pipe_ctx[i]->stream_res.tg->funcs) {
> -			if (pipe_ctx[i]->stream_res.tg->funcs->set_drr)
> -				pipe_ctx[i]->stream_res.tg->funcs->set_drr(
> -					pipe_ctx[i]->stream_res.tg, &params);
> +		/* dc_state_destruct() might null the stream resources, so fetch tg
> +		 * here first to avoid a race condition. The lifetime of the pointee
> +		 * itself (the timing_generator object) is not a problem here.
> +		 */
> +		struct timing_generator *tg = pipe_ctx[i]->stream_res.tg;
> +
> +		if ((tg != NULL) && tg->funcs) {
> +			if (tg->funcs->set_drr)
> +				tg->funcs->set_drr(tg, &params);
>  			if (adjust.v_total_max != 0 && adjust.v_total_min != 0)
> -				if (pipe_ctx[i]->stream_res.tg->funcs->set_static_screen_control)
> -					pipe_ctx[i]->stream_res.tg->funcs->set_static_screen_control(
> -						pipe_ctx[i]->stream_res.tg,
> -						event_triggers, num_frames);
> +				if (tg->funcs->set_static_screen_control)
> +					tg->funcs->set_static_screen_control(
> +						tg, event_triggers, num_frames);
>  		}
>  	}
>  }

This fixes hard to trace panics with labwc VRR and Wayfire on RX 6700 XT. I had to use netconsole to arrive at the original bug report.

Tested-by: Christopher Snowhill <chris at kode54.net>