Re: [PATCH 4/4] interconnect: qcom: icc-rpm: Remodel how QoS settings are stored

From: Stephan Gerhold
Date: Tue Mar 26 2024 - 16:57:21 EST


On Tue, Mar 26, 2024 at 08:42:35PM +0100, Konrad Dybcio wrote:
> Currently, the QoS settings are stored in the node data, even though
> they're a property of the bus/provider instead. Moreover, they are only
> needed during the probe step, so they can be easily moved into struct
> qcom_icc_desc.
>
> Reshuffle things around to make it anywhere near readable & comparable
> with a reference. As a nice bonus, a lot of bytes are shaved off and
> a few miliseconds are shaved off here and there.
>
> As an example, bloat-o-meter reports this on sm6115.o:
> Total: Before=14799, After=13263, chg -10.38%
>
> Signed-off-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> ---
> drivers/interconnect/qcom/icc-rpm.c | 123 +++++----
> drivers/interconnect/qcom/icc-rpm.h | 13 +-
> drivers/interconnect/qcom/msm8909.c | 268 ++++++++++---------
> drivers/interconnect/qcom/msm8916.c | 153 ++++++-----
> drivers/interconnect/qcom/msm8939.c | 157 ++++++-----
> drivers/interconnect/qcom/msm8996.c | 517 +++++++++++++++++-------------------
> drivers/interconnect/qcom/qcm2290.c | 416 +++++++++++++++++------------
> drivers/interconnect/qcom/sdm660.c | 393 +++++++++++++--------------
> drivers/interconnect/qcom/sm6115.c | 239 ++++++++++++-----
> 9 files changed, 1224 insertions(+), 1055 deletions(-)
>
> [...]
> @@ -70,20 +68,18 @@ struct qcom_icc_provider {
> };
>
> /**
> - * struct qcom_icc_qos - Qualcomm specific interconnect QoS parameters
> + * struct qcom_icc_qos_data - Qualcomm specific interconnect QoS parameters
> * @areq_prio: node requests priority
> * @prio_level: priority level for bus communication
> * @limit_commands: activate/deactivate limiter mode during runtime
> - * @ap_owned: indicates if the node is owned by the AP or by the RPM
> * @qos_mode: default qos mode for this node
> * @qos_port: qos port number for finding qos registers of this node
> * @urg_fwd_en: enable urgent forwarding
> */
> -struct qcom_icc_qos {
> +struct qcom_icc_qos_data {
> u32 areq_prio;
> u32 prio_level;
> bool limit_commands;
> - bool ap_owned;
> int qos_mode;
> int qos_port;
> bool urg_fwd_en;

Side note: There is a potential for more micro-optimization here: You
could save 4 bytes of padding if you move all bools together at the end
of the struct. :D

> [...]
> @@ -134,6 +131,8 @@ struct qcom_icc_desc {
> bool keep_alive;
> enum qcom_icc_type type;
> const struct regmap_config *regmap_cfg;
> + const struct qcom_icc_qos_data * const qos_data;
> + const u16 qos_data_num;
> unsigned int qos_offset;

Nitpick: Why is the u16 const when the other (non-pointer) members are
not? The u16 also feels a bit like overkill here. The struct would have
exactly the same size with a full unsigned int because of padding.

Alternatively, you could consider using an empty last entry as sentinel
instead of adding the count (i.e. with NOC_QOS_MODE_INVALID = 0). Not
sure what is cleaner here.

I haven't looked closely at the actual conversion of the definitions in
the drivers. What is the chance that you made an accidental mistake in
there? Or was it scripted? :D

Thanks,
Stephan