<div dir="ltr"><div><br></div><div><pre><span class="gmail-q">> From: Tobias Jakobi <<a href="mailto:tjakobi@math.uni-bielefeld.de">tjakobi@math.uni-bielefeld.de</a>>
> 
> 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: <a href="https://gitlab.freedesktop.org/drm/amd/-/issues/3142">https://gitlab.freedesktop.org/drm/amd/-/issues/3142</a>
> Fixes: 06ad7e164256 ("drm/amd/display: Destroy DC context while
> keeping DML and DML2")
> Signed-off-by: Tobias Jakobi <<a href="mailto:tjakobi@math.uni-bielefeld.de">tjakobi@math.uni-bielefeld.de</a>>
> ---
>  .../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]-><a href="http://stream_res.tg">stream_res.tg</a> != 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]-><a href="http://stream_res.tg">stream_res.tg</a>,
> &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]-
> ><a href="http://stream_res.tg">stream_res.tg</a>;
> +
> +          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]-
> ><a href="http://stream_res.tg">stream_res.tg</a>,
> -                                          event_triggers,
> num_frames);
> +                          if (tg->funcs-
> >set_static_screen_control)
> +                                  tg->funcs-
> >set_static_screen_control(
> +                                          tg, event_triggers,
> num_frames);
>           }
>   }
>  }<br><br></span></pre><pre><span class="gmail-q">This fixes full system freezes when taking screenshots at low framerates with VRR enabled on an RX 7900 XTX.<br></span><br>Tested-by: Raoul van Rüschen <<a href="mailto:raoul.van.rueschen@gmail.com">raoul.van.rueschen@gmail.com</a>></pre></div></div>