Re: [PATCH] drm/kms/plane: Avoiding redundant checks in case of EDEADLK retries

From: Maarten Lankhorst
Date: Mon Jul 30 2018 - 03:53:11 EST


Op 30-07-18 om 07:18 schreef Satendra Singh Thakur:
> 1. Currently, if modeset_lock is re-tried many times in case of EDEADLK
> error, then this will be the code flow
>
> retry:
> ret = drm_modeset_lock();
>
> if (ret)-->this is true
> goto out;
>
> out:
> if (fb)
> if (plane->old_fb)
> if (ret == -EDEADLK)
> goto retry;
> It can be observed that checks on fb, old_fb are redundant
> in retry-case.
> If we keep if (ret == -EDEADLK) right after the out label,
> that will avoid redundant checks.
> It won't affect normal scenario (non-retry case).
> 2. Moved NULL assignment inside if statement related to NULL check
>
> Signed-off-by: Satendra Singh Thakur <satendra.t@xxxxxxxxxxx>
> ---
> drivers/gpu/drm/drm_plane.c | 17 +++++++++--------
> 1 file changed, 9 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_plane.c b/drivers/gpu/drm/drm_plane.c
> index 0350544..ed42cd4 100644
> --- a/drivers/gpu/drm/drm_plane.c
> +++ b/drivers/gpu/drm/drm_plane.c
> @@ -658,9 +658,10 @@ static int __setplane_internal(struct drm_plane *plane,
> }
>
> out:
> - if (plane->old_fb)
> + if (plane->old_fb) {
> drm_framebuffer_put(plane->old_fb);
> - plane->old_fb = NULL;
> + plane->old_fb = NULL;
> + }
>
> return ret;
> }
> @@ -1097,17 +1098,17 @@ int drm_mode_page_flip_ioctl(struct drm_device *dev,
> }
>
> out:
> - if (fb)
> - drm_framebuffer_put(fb);
> - if (plane->old_fb)
> - drm_framebuffer_put(plane->old_fb);
> - plane->old_fb = NULL;
> -
> if (ret == -EDEADLK) {
> ret = drm_modeset_backoff(&ctx);
> if (!ret)
> goto retry;
> }
> + if (fb)
> + drm_framebuffer_put(fb);
> + if (plane->old_fb) {
> + drm_framebuffer_put(plane->old_fb);
> + plane->old_fb = NULL;
> + }
>
> drm_modeset_drop_locks(&ctx);
> drm_modeset_acquire_fini(&ctx);

It's not redundant. plane->old_fb is protected by the plane_mutex lock and should be
null before dropping the locks in drm_modeset_backoff(). Else other code will trip
with a WARN_ON(plane->old_fb != NULL);

Before sending any more patches I would strongly suggest setting
CONFIG_DEBUG_WW_MUTEX_SLOWPATH and CONFIG_PROVE_LOCKING in your kernel, and try to
run the IGT testsuite.

~Maarten