On Fri 01 Apr 00:14 PDT 2022, Vladimir Zapolskiy wrote:
> It's noted that dcvs interrupts are not self-clearing, thus an interrupt
> handler runs constantly, which leads to a severe regression in runtime.
> To fix the problem an explicit write to clear interrupt register is
> required.
>
> Fixes: 275157b367f4 ("cpufreq: qcom-cpufreq-hw: Add dcvs interrupt support")
> Signed-off-by: Vladimir Zapolskiy <vladimir.zapolskiy@xxxxxxxxxx>
> ---
> Changes from v1 to v2:
> * added a check for pending interrupt status before its handling,
> thanks to Bjorn for review
>
> drivers/cpufreq/qcom-cpufreq-hw.c | 14 ++++++++++++++
> 1 file changed, 14 insertions(+)
>
> diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c
> index f9d593ff4718..e17413a6f120 100644
> --- a/drivers/cpufreq/qcom-cpufreq-hw.c
> +++ b/drivers/cpufreq/qcom-cpufreq-hw.c
> @@ -24,6 +24,8 @@
> #define CLK_HW_DIV 2
> #define LUT_TURBO_IND 1
>
> +#define GT_IRQ_STATUS BIT(2)
> +
> #define HZ_PER_KHZ 1000
>
> struct qcom_cpufreq_soc_data {
> @@ -31,6 +33,8 @@ struct qcom_cpufreq_soc_data {
> u32 reg_dcvs_ctrl;
> u32 reg_freq_lut;
> u32 reg_volt_lut;
> + u32 reg_intr_clr;
> + u32 reg_intr_status;
> u32 reg_current_vote;
> u32 reg_perf_state;
> u8 lut_row_size;
> @@ -345,11 +349,19 @@ static void qcom_lmh_dcvs_poll(struct work_struct *work)
> static irqreturn_t qcom_lmh_dcvs_handle_irq(int irq, void *data)
> {
> struct qcom_cpufreq_data *c_data = data;
> + u32 val;
> +
> + val = readl_relaxed(c_data->base + c_data->soc_data->reg_intr_status);
Seems reasonable to read the INTR_STATUS register and bail early if
there's no interrupt.
> + if (!(val & GT_IRQ_STATUS))
> + return IRQ_HANDLED;
But if we in the interrupt handler realize that there's no interrupt
pending for us, shouldn't we return IRQ_NONE?
>
> /* Disable interrupt and enable polling */
> disable_irq_nosync(c_data->throttle_irq);
> schedule_delayed_work(&c_data->throttle_work, 0);
>
> + writel_relaxed(GT_IRQ_STATUS,
> + c_data->base + c_data->soc_data->reg_intr_clr);
And in OSM (i.e. not epss_soc_data), both reg_intr_status and
reg_intr_clr will be 0, so we end up reading and writing the wrong
register.
You need to do:
if (c_data->soc_data->reg_intr_clr)
writel_relaxed(..., reg_intr_clr);
But according to the downstream driver, this is supposed to be done in
the polling function, right before you do enable_irq().
Regards,
Bjorn
> +
> return IRQ_HANDLED;
> }
>
> @@ -368,6 +380,8 @@ static const struct qcom_cpufreq_soc_data epss_soc_data = {
> .reg_dcvs_ctrl = 0xb0,
> .reg_freq_lut = 0x100,
> .reg_volt_lut = 0x200,
> + .reg_intr_clr = 0x308,
> + .reg_intr_status = 0x30c,
> .reg_perf_state = 0x320,
> .lut_row_size = 4,
> };
> --
> 2.33.0
>
