Re: [PATCH v2 4/8] drm: zynqmp_dp: Rearrange zynqmp_dp for better padding

[Tomi Valkeinen]

On 20/03/2024 00:51, Sean Anderson wrote:
> Sort the members of struct zynqmp_dp to reduce padding necessary for
> alignment.
>
> Signed-off-by: Sean Anderson <sean.anderson@xxxxxxxxx>
> ---
>
> Changes in v2:
> - New
>
>   drivers/gpu/drm/xlnx/zynqmp_dp.c | 28 ++++++++++++++--------------
>   1 file changed, 14 insertions(+), 14 deletions(-)
>
> diff --git a/drivers/gpu/drm/xlnx/zynqmp_dp.c b/drivers/gpu/drm/xlnx/zynqmp_dp.c
> index 8635b5673386..f1834c8e3c02 100644
> --- a/drivers/gpu/drm/xlnx/zynqmp_dp.c
> +++ b/drivers/gpu/drm/xlnx/zynqmp_dp.c
> @@ -255,10 +255,10 @@ struct zynqmp_dp_link_config {
>    * @fmt: format identifier string
>    */
>   struct zynqmp_dp_mode {
> -	u8 bw_code;
> -	u8 lane_cnt;
> -	int pclock;
>   	const char *fmt;
> +	int pclock;
> +	u8 bw_code;
> +	u8 lane_cnt;
>   };
>   
>   /**
> @@ -295,27 +295,27 @@ struct zynqmp_dp_config {
>    * @train_set: set of training data
>    */
>   struct zynqmp_dp {
> +	struct drm_dp_aux aux;
> +	struct drm_bridge bridge;
> +	struct delayed_work hpd_work;
> +
> +	struct drm_bridge *next_bridge;
>   	struct device *dev;
>   	struct zynqmp_dpsub *dpsub;
>   	void __iomem *iomem;
>   	struct reset_control *reset;
> -	int irq;
> -
> -	struct drm_bridge bridge;
> -	struct drm_bridge *next_bridge;
> -
> -	struct zynqmp_dp_config config;
> -	struct drm_dp_aux aux;
>   	struct phy *phy[ZYNQMP_DP_MAX_LANES];
> -	u8 num_lanes;
> -	struct delayed_work hpd_work;
> +
>   	enum drm_connector_status status;
> +	int irq;
>   	bool enabled;
>   
> -	u8 dpcd[DP_RECEIVER_CAP_SIZE];
> -	struct zynqmp_dp_link_config link_config;
>   	struct zynqmp_dp_mode mode;
> +	struct zynqmp_dp_link_config link_config;
> +	struct zynqmp_dp_config config;
> +	u8 dpcd[DP_RECEIVER_CAP_SIZE];
>   	u8 train_set[ZYNQMP_DP_MAX_LANES];
> +	u8 num_lanes;
>   };
>   
>   static inline struct zynqmp_dp *bridge_to_dp(struct drm_bridge *bridge)

If you change the order of the fields, you should change the order in 
the kernel doc accordingly.

To be honest, I'm not sure if I like this patch. We have usually one 
instance of these structs allocated. How many bytes do we save?

I'm fine with getting easy savings by changing the field order in some 
cases, but I think the "human" side of the order is important too: 
usually the fields are grouped in some way, and ordered so that the more 
base or generic ones are first, and fields for some specific feature are 
later. And fields protected by a lock should be grouped together, with 
their lock being first/last in that group.

Looking at the zynqmp_dp struct with this patch, I get an urge to start 
moving things around: dev, dpsub, iomem, etc first, hpd somewhere later. 
Base config fields like config, num_lanes, irq would be grouped 
together. Etc.

 Tomi