On Wed, Jan 31, 2024 at 01:58:08PM +0800, Xu Yang wrote:
> i.MX95 has a DDR PMU which is almostly same as i.MX93, it now supports
> read beat and write beat filter capabilities. This will add support for
> i.MX95 and enhance the driver to support specific filter handling for it.
>
> Usage:
>
> For read beat:
> ~# perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_rd_beat_filt2,counter=3,axi_mask=ID_MASK,axi_id=ID/
> ~# perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_rd_beat_filt1,counter=4,axi_mask=ID_MASK,axi_id=ID/
> ~# perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_rd_beat_filt0,counter=5,axi_mask=ID_MASK,axi_id=ID/
> eg: For edma2: perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_rd_beat_filt0,counter=5,axi_mask=0x00f,axi_id=0x00c/
>
> For write beat:
> ~# perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_wr_beat_filt,counter=2,axi_mask=ID_MASK,axi_id=ID/
> eg: For edma2: perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_wr_beat_filt,counter=2,axi_mask=0x00f,axi_id=0x00c/
>
> Signed-off-by: Xu Yang <xu.yang_2@xxxxxxx>
>
> ---
> Changes in v2:
> - put soc spefific axi filter events to drvdata according
> to franks suggestions.
> - adjust pmcfg axi_id and axi_mask config
> Changes in v3:
> - no changes
> Changes in v4:
> - only contain imx95 parts
> ---
> drivers/perf/fsl_imx9_ddr_perf.c | 86 +++++++++++++++++++++++++++++++-
> 1 file changed, 84 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/perf/fsl_imx9_ddr_perf.c b/drivers/perf/fsl_imx9_ddr_perf.c
> index b1a58e9e1617..85aaaef7212f 100644
> --- a/drivers/perf/fsl_imx9_ddr_perf.c
> +++ b/drivers/perf/fsl_imx9_ddr_perf.c
> @@ -17,9 +17,19 @@
> #define MX93_PMCFG1_RD_BT_FILT_EN BIT(29)
> #define MX93_PMCFG1_ID_MASK GENMASK(17, 0)
>
> +#define MX95_PMCFG1_WR_BEAT_FILT_EN BIT(31)
> +#define MX95_PMCFG1_RD_BEAT_FILT_EN BIT(30)
> +
> #define PMCFG2 0x04
> #define MX93_PMCFG2_ID GENMASK(17, 0)
>
> +#define PMCFG3 0x08
> +#define PMCFG4 0x0C
> +#define PMCFG5 0x10
> +#define PMCFG6 0x14
> +#define MX95_PMCFG_ID_MASK GENMASK(9, 0)
> +#define MX95_PMCFG_ID GENMASK(25, 16)
> +
> /* Global control register affects all counters and takes priority over local control registers */
> #define PMGC0 0x40
> /* Global control register bits */
> @@ -240,6 +250,18 @@ static struct attribute *imx93_ddr_perf_events_attrs[] = {
> NULL,
> };
>
> +static struct attribute *imx95_ddr_perf_events_attrs[] = {
> + /* counter2 specific events */
> + IMX9_DDR_PMU_EVENT_ATTR(eddrtq_pm_wr_beat_filt, 73),
> + /* counter3 specific events */
> + IMX9_DDR_PMU_EVENT_ATTR(eddrtq_pm_rd_beat_filt2, 73),
> + /* counter4 specific events */
> + IMX9_DDR_PMU_EVENT_ATTR(eddrtq_pm_rd_beat_filt1, 73),
> + /* counter5 specific events */
> + IMX9_DDR_PMU_EVENT_ATTR(eddrtq_pm_rd_beat_filt0, 73),
> + NULL,
> +};
> +
> PMU_FORMAT_ATTR(event, "config:0-7");
> PMU_FORMAT_ATTR(counter, "config:8-15");
> PMU_FORMAT_ATTR(axi_id, "config1:0-17");
> @@ -271,8 +293,14 @@ static const struct imx_ddr_devtype_data imx93_devtype_data = {
> .attrs = imx93_ddr_perf_events_attrs,
> };
>
> +static const struct imx_ddr_devtype_data imx95_devtype_data = {
> + .identifier = "imx95",
> + .attrs = imx95_ddr_perf_events_attrs,
> +};
> +
> static const struct of_device_id imx_ddr_pmu_dt_ids[] = {
> { .compatible = "fsl,imx93-ddr-pmu", .data = &imx93_devtype_data },
> + { .compatible = "fsl,imx95-ddr-pmu", .data = &imx95_devtype_data },
> { /* sentinel */ }
> };
> MODULE_DEVICE_TABLE(of, imx_ddr_pmu_dt_ids);
> @@ -410,6 +438,56 @@ static void imx93_ddr_perf_monitor_config(struct ddr_pmu *pmu, int cfg, int cfg1
> writel(pmcfg2, pmu->base + PMCFG2);
> }
>
> +static void imx95_ddr_perf_monitor_config(struct ddr_pmu *pmu, int cfg, int cfg1, int cfg2)
> +{
> + u32 pmcfg1, pmcfg, offset = 0;
> + int event, counter;
> +
> + event = cfg & 0x000000FF;
> + counter = (cfg & 0x0000FF00) >> 8;
> +
> + pmcfg1 = readl_relaxed(pmu->base + PMCFG1);
> +
> + if (counter == 2 && event == 73) {
> + pmcfg1 |= MX95_PMCFG1_WR_BEAT_FILT_EN;
> + offset = PMCFG3;
> + } else if (counter == 2 && event != 73) {
> + pmcfg1 &= ~MX95_PMCFG1_WR_BEAT_FILT_EN;
> + }
> +
> + if (counter == 3 && event == 73) {
> + pmcfg1 |= MX95_PMCFG1_RD_BEAT_FILT_EN;
> + offset = PMCFG4;
> + } else if (counter == 3 && event != 73) {
> + pmcfg1 &= ~MX95_PMCFG1_RD_BEAT_FILT_EN;
> + }
> +
> + if (counter == 4 && event == 73) {
> + pmcfg1 |= MX95_PMCFG1_RD_BEAT_FILT_EN;
> + offset = PMCFG5;
> + } else if (counter == 4 && event != 73) {
> + pmcfg1 &= ~MX95_PMCFG1_RD_BEAT_FILT_EN;
> + }
> +
> + if (counter == 5 && event == 73) {
> + pmcfg1 |= MX95_PMCFG1_RD_BEAT_FILT_EN;
> + offset = PMCFG6;
> + } else if (counter == 5 && event != 73) {
> + pmcfg1 &= ~MX95_PMCFG1_RD_BEAT_FILT_EN;
> + }
I think this would be much easier to read if you rewrote it as:
switch (counter) {
case 2:
...
break;
case 3:
etc.
> + writel(pmcfg1, pmu->base + PMCFG1);
writel_relaxed()? (since you're already using readl_relaxed()).
> +
> + if (offset) {
> + pmcfg = readl_relaxed(pmu->base + offset);
> + pmcfg &= ~(FIELD_PREP(MX95_PMCFG_ID_MASK, 0x3FF) |
> + FIELD_PREP(MX95_PMCFG_ID, 0x3FF));
> + pmcfg |= (FIELD_PREP(MX95_PMCFG_ID_MASK, cfg2) |
> + FIELD_PREP(MX95_PMCFG_ID, cfg1));
> + writel(pmcfg, pmu->base + offset);
Similarly.
Will
