| Message ID | 20220307153050.3392700-4-dmitry.baryshkov@linaro.org (mailing list archive) |
|---|---|
| State | New, archived |
| Delegated to: | viresh kumar |
| Headers | show |
| Series | [1/4] cpufreq: qcom-hw: drop affinity hint before freeing the IRQ | expand |
On Mon 07 Mar 07:30 PST 2022, Dmitry Baryshkov wrote: > Provide lightweight online and offline operations. This saves us from > parsing and tearing down the OPP tables each time the CPU is put online > or offline. Isn't that a slight understatement? Doesn't it also save us from e.g. ioremapping the memory, traversing DT to discover the policy's related_cpus and requesting the dcvs interrupt? I like the idea of getting these things out of the init/exit path. I do however think that we could move most of this to probe time, and thereby be able to rely on devm operations for many of these things. That said, I still like your idea of having a fast path for this... > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> > --- > drivers/cpufreq/qcom-cpufreq-hw.c | 28 ++++++++++++++++++++++++++-- > 1 file changed, 26 insertions(+), 2 deletions(-) > > diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c > index 580520215ee7..12b67f16b78f 100644 > --- a/drivers/cpufreq/qcom-cpufreq-hw.c > +++ b/drivers/cpufreq/qcom-cpufreq-hw.c > @@ -424,10 +424,26 @@ static int qcom_cpufreq_hw_lmh_init(struct cpufreq_policy *policy, int index) > return 0; > } > > -static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) > +static int qcom_cpufreq_hw_cpu_online(struct cpufreq_policy *policy) > { > + struct qcom_cpufreq_data *data = policy->driver_data; > + struct platform_device *pdev = cpufreq_get_driver_data(); > + int ret; > + For backwards compatibility reasons it's valid to not have data->throttle_irq. This will however cause irq_set_affinity_hint() to return -EINVAL and we'll get a print. So you should handle that gracefully. > + ret = irq_set_affinity_hint(data->throttle_irq, policy->cpus); > + if (ret) > + dev_err(&pdev->dev, "Failed to set CPU affinity of %s[%d]\n", > + data->irq_name, data->throttle_irq); > + > + return ret; > +} > + > +static int qcom_cpufreq_hw_cpu_offline(struct cpufreq_policy *policy) > +{ > + struct qcom_cpufreq_data *data = policy->driver_data; > + > if (data->throttle_irq <= 0) > - return; > + return 0; > > mutex_lock(&data->throttle_lock); > data->cancel_throttle = true; This will mark the throttle as cancelled, you need to clear this as you're bringing the policy online again. > @@ -435,6 +451,12 @@ static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) > > cancel_delayed_work_sync(&data->throttle_work); > irq_set_affinity_hint(data->throttle_irq, NULL); You don't disable_irq(data->throttle_irq) here. I think qcom_lmh_dcvs_notify() will be unhappy if we get thermal pressure from a policy with no cpus? Note though that you can't enable it in online(), as it will be enabled in ready()... > + > + return 0; > +} > + > +static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) > +{ > free_irq(data->throttle_irq, data); As above, you should treat throttle_irq <= 0 gracefully. Regards, Bjorn > } > > @@ -588,6 +610,8 @@ static struct cpufreq_driver cpufreq_qcom_hw_driver = { > .get = qcom_cpufreq_hw_get, > .init = qcom_cpufreq_hw_cpu_init, > .exit = qcom_cpufreq_hw_cpu_exit, > + .online = qcom_cpufreq_hw_cpu_online, > + .offline = qcom_cpufreq_hw_cpu_offline, > .register_em = cpufreq_register_em_with_opp, > .fast_switch = qcom_cpufreq_hw_fast_switch, > .name = "qcom-cpufreq-hw", > -- > 2.34.1 >
On Tue, 8 Mar 2022 at 01:38, Bjorn Andersson <bjorn.andersson@linaro.org> wrote: > > On Mon 07 Mar 07:30 PST 2022, Dmitry Baryshkov wrote: > > > Provide lightweight online and offline operations. This saves us from > > parsing and tearing down the OPP tables each time the CPU is put online > > or offline. > > Isn't that a slight understatement? Doesn't it also save us from e.g. > ioremapping the memory, traversing DT to discover the policy's > related_cpus and requesting the dcvs interrupt? > > I like the idea of getting these things out of the init/exit path. I do > however think that we could move most of this to probe time, and thereby > be able to rely on devm operations for many of these things. > > That said, I still like your idea of having a fast path for this... > > > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> > > --- > > drivers/cpufreq/qcom-cpufreq-hw.c | 28 ++++++++++++++++++++++++++-- > > 1 file changed, 26 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c > > index 580520215ee7..12b67f16b78f 100644 > > --- a/drivers/cpufreq/qcom-cpufreq-hw.c > > +++ b/drivers/cpufreq/qcom-cpufreq-hw.c > > @@ -424,10 +424,26 @@ static int qcom_cpufreq_hw_lmh_init(struct cpufreq_policy *policy, int index) > > return 0; > > } > > > > -static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) > > +static int qcom_cpufreq_hw_cpu_online(struct cpufreq_policy *policy) > > { > > + struct qcom_cpufreq_data *data = policy->driver_data; > > + struct platform_device *pdev = cpufreq_get_driver_data(); > > + int ret; > > + > > For backwards compatibility reasons it's valid to not have > data->throttle_irq. This will however cause irq_set_affinity_hint() to > return -EINVAL and we'll get a print. > > So you should handle that gracefully. Ack. > > > + ret = irq_set_affinity_hint(data->throttle_irq, policy->cpus); > > + if (ret) > > + dev_err(&pdev->dev, "Failed to set CPU affinity of %s[%d]\n", > > + data->irq_name, data->throttle_irq); > > + > > + return ret; > > +} > > + > > +static int qcom_cpufreq_hw_cpu_offline(struct cpufreq_policy *policy) > > +{ > > + struct qcom_cpufreq_data *data = policy->driver_data; > > + > > if (data->throttle_irq <= 0) > > - return; > > + return 0; > > > > mutex_lock(&data->throttle_lock); > > data->cancel_throttle = true; > > This will mark the throttle as cancelled, you need to clear this as > you're bringing the policy online again. ack. > > > @@ -435,6 +451,12 @@ static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) > > > > cancel_delayed_work_sync(&data->throttle_work); > > irq_set_affinity_hint(data->throttle_irq, NULL); > > You don't disable_irq(data->throttle_irq) here. I think > qcom_lmh_dcvs_notify() will be unhappy if we get thermal pressure from a > policy with no cpus? > > Note though that you can't enable it in online(), as it will be enabled > in ready()... And we can't just disable it here, as it might be enabled or might be disabled. I think the simplest would be to call free_irq() here and request_irq() to online(). > > > + > > + return 0; > > +} > > + > > +static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) > > +{ > > free_irq(data->throttle_irq, data); > > As above, you should treat throttle_irq <= 0 gracefully. > > Regards, > Bjorn > > > } > > > > @@ -588,6 +610,8 @@ static struct cpufreq_driver cpufreq_qcom_hw_driver = { > > .get = qcom_cpufreq_hw_get, > > .init = qcom_cpufreq_hw_cpu_init, > > .exit = qcom_cpufreq_hw_cpu_exit, > > + .online = qcom_cpufreq_hw_cpu_online, > > + .offline = qcom_cpufreq_hw_cpu_offline, > > .register_em = cpufreq_register_em_with_opp, > > .fast_switch = qcom_cpufreq_hw_fast_switch, > > .name = "qcom-cpufreq-hw", > > -- > > 2.34.1 > >
diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c index 580520215ee7..12b67f16b78f 100644 --- a/drivers/cpufreq/qcom-cpufreq-hw.c +++ b/drivers/cpufreq/qcom-cpufreq-hw.c @@ -424,10 +424,26 @@ static int qcom_cpufreq_hw_lmh_init(struct cpufreq_policy *policy, int index) return 0; } -static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) +static int qcom_cpufreq_hw_cpu_online(struct cpufreq_policy *policy) { + struct qcom_cpufreq_data *data = policy->driver_data; + struct platform_device *pdev = cpufreq_get_driver_data(); + int ret; + + ret = irq_set_affinity_hint(data->throttle_irq, policy->cpus); + if (ret) + dev_err(&pdev->dev, "Failed to set CPU affinity of %s[%d]\n", + data->irq_name, data->throttle_irq); + + return ret; +} + +static int qcom_cpufreq_hw_cpu_offline(struct cpufreq_policy *policy) +{ + struct qcom_cpufreq_data *data = policy->driver_data; + if (data->throttle_irq <= 0) - return; + return 0; mutex_lock(&data->throttle_lock); data->cancel_throttle = true; @@ -435,6 +451,12 @@ static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) cancel_delayed_work_sync(&data->throttle_work); irq_set_affinity_hint(data->throttle_irq, NULL); + + return 0; +} + +static void qcom_cpufreq_hw_lmh_exit(struct qcom_cpufreq_data *data) +{ free_irq(data->throttle_irq, data); } @@ -588,6 +610,8 @@ static struct cpufreq_driver cpufreq_qcom_hw_driver = { .get = qcom_cpufreq_hw_get, .init = qcom_cpufreq_hw_cpu_init, .exit = qcom_cpufreq_hw_cpu_exit, + .online = qcom_cpufreq_hw_cpu_online, + .offline = qcom_cpufreq_hw_cpu_offline, .register_em = cpufreq_register_em_with_opp, .fast_switch = qcom_cpufreq_hw_fast_switch, .name = "qcom-cpufreq-hw",
Provide lightweight online and offline operations. This saves us from parsing and tearing down the OPP tables each time the CPU is put online or offline. Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> --- drivers/cpufreq/qcom-cpufreq-hw.c | 28 ++++++++++++++++++++++++++-- 1 file changed, 26 insertions(+), 2 deletions(-)