lkml.org 
[lkml]   [2018]   [Dec]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v6 06/10] soc: qcom: rpmhpd: Add RPMh power domain driver
From
Date
Quoting Rajendra Nayak (2018-12-11 01:49:34)
> The RPMh power domain driver aggregates the corner votes from various
> consumers for the ARC resources and communicates it to RPMh.
>
> With RPMh we use 2 different numbering space for corners, one used
> by the clients to express their performance needs, and another used
> to communicate to RPMh hardware.
>
> The clients express their performance requirements using a sparse
> numbering space which are mapped to meaningful levels like RET, SVS,
> NOMINAL, TURBO etc which then get mapped to another number space
> between 0 and 15 which is communicated to RPMh. The sparse number space,
> also referred to as vlvl is mapped to the continuous number space of 0
> to 15, also referred to as hlvl, using command DB.
>
> Some power domain clients could request a performance state only while
> the CPU is active, while some others could request for a certain
> performance state all the time regardless of the state of the CPU.
> We handle this by internally aggregating the votes from both type of
> clients and then send the aggregated votes to RPMh.
>
> There are also 3 different types of Votes that are comunicated to RPMh

Why capitalize vote?

> for every resource.
> 1. ACTIVE_ONLY: This specifies the requirement for the resource when the
> CPU is active
> 2. SLEEP: This specifies the requirement for the resource when the CPU
> is going to sleep
> 3. WAKE_ONLY: This specifies the requirement for the resource when the
> CPU is coming out of sleep to active state

Can you tab these in?

>
> We add data for all power domains on sdm845 SoC as part of the patch.
> The driver can be extended to support other SoCs which support RPMh
>
> Signed-off-by: Rajendra Nayak <rnayak@codeaurora.org>
> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>

Just minor nitpicks ahead and a warning.

> diff --git a/drivers/soc/qcom/Makefile b/drivers/soc/qcom/Makefile
> index f1b25fdcf2ad..dd6ca92985ee 100644
> --- a/drivers/soc/qcom/Makefile
> +++ b/drivers/soc/qcom/Makefile
> @@ -22,3 +22,4 @@ obj-$(CONFIG_QCOM_APR) += apr.o
> obj-$(CONFIG_QCOM_LLCC) += llcc-slice.o
> obj-$(CONFIG_QCOM_SDM845_LLCC) += llcc-sdm845.o
> obj-$(CONFIG_QCOM_RPMPD) += rpmpd.o
> +obj-$(CONFIG_QCOM_RPMHPD) += rpmhpd.o

Put this before RPMPD? At least it would be semi-sorted then.

> diff --git a/drivers/soc/qcom/rpmhpd.c b/drivers/soc/qcom/rpmhpd.c
> new file mode 100644
> index 000000000000..f993a86be48c
> --- /dev/null
> +++ b/drivers/soc/qcom/rpmhpd.c
> @@ -0,0 +1,417 @@
[..]
> +
> +static int rpmhpd_update_level_mapping(struct rpmhpd *rpmhpd)
> +{
> + u8 buf[RPMH_ARC_MAX_LEVELS * RPMH_ARC_LEVEL_SIZE];
> + int i, j, len, ret;
> +
> + len = cmd_db_read_aux_data_len(rpmhpd->res_name);
> + if (len <= 0)
> + return len;
> + else if (len > RPMH_ARC_MAX_LEVELS * RPMH_ARC_LEVEL_SIZE)
> + return -EINVAL;
> +
> + ret = cmd_db_read_aux_data(rpmhpd->res_name, buf, len);
> + if (ret < 0)
> + return ret;

I've changed cmd_db_read_aux_data() and that change is winding through
the arm-soc tree. This will break in the compilation phase in
linux-next.

> +
> + rpmhpd->level_count = len / RPMH_ARC_LEVEL_SIZE;
> +
> + for (i = 0; i < rpmhpd->level_count; i++) {
> + rpmhpd->level[i] = 0;
> + for (j = 0; j < RPMH_ARC_LEVEL_SIZE; j++)
> + rpmhpd->level[i] |=
> + buf[i * RPMH_ARC_LEVEL_SIZE + j] << (8 * j);
> +
> + /*
> + * The AUX data may be zero padded. These 0 valued entries at
> + * the end of the map must be ignored.
> + */
> + if (i > 0 && rpmhpd->level[i] == 0) {
> + rpmhpd->level_count = i;
> + break;
> + }
> + pr_debug("%s: ARC hlvl=%2d --> vlvl=%4u\n", rpmhpd->res_name, i,
> + rpmhpd->level[i]);
> + }
> +
> + return 0;
> +}
> +
> +static int rpmhpd_probe(struct platform_device *pdev)
> +{
> + int i, ret;
> + size_t num_pds;
> + struct device *dev = &pdev->dev;
> + struct genpd_onecell_data *data;
> + struct rpmhpd **rpmhpds;
> + const struct rpmhpd_desc *desc;
> +
> + desc = of_device_get_match_data(dev);
> + if (!desc)
> + return -EINVAL;
> +
> + rpmhpds = desc->rpmhpds;
> + num_pds = desc->num_pds;
> +
> + data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
> + if (!data)
> + return -ENOMEM;
> +
> + data->domains = devm_kcalloc(dev, num_pds, sizeof(*data->domains),
> + GFP_KERNEL);
> + if (!data->domains)
> + return -ENOMEM;
> +
> + data->num_domains = num_pds;
> +
> + ret = cmd_db_ready();

Does this matter? I thought we forced cmd_db_ready() in the rpmh driver
probe, and that was the parent of all rpmh devices so we should be fine
to not need to check it again here?

> + if (ret) {
> + if (ret != -EPROBE_DEFER)
> + dev_err(dev, "Command DB unavailable, ret=%d\n", ret);
> + return ret;
> + }
> +

\
 
 \ /
  Last update: 2018-12-11 23:01    [W:0.422 / U:0.456 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site