| Message ID | 2435090.1mJ0fSsrDY@kreacher (mailing list archive) |
|---|---|
| State | Mainlined, archived |
| Headers | show |
| Series | ACPI: processor: Add QoS requests for all CPUs | expand |
On 25-10-19, 02:41, Rafael J. Wysocki wrote: > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > The _PPC change notifications from the platform firmware are per-CPU, > so acpi_processor_ppc_init() needs to add a frequency QoS request > for each CPU covered by a cpufreq policy to take all of them into > account. > > Even though ACPI thermal control of CPUs sets frequency limits > per processor package, it also needs a frequency QoS request for each > CPU in a cpufreq policy in case some of them are taken offline and > the frequency limit needs to be set through the remaining online > ones (this is slightly excessive, because all CPUs covered by one > cpufreq policy will set the same frequency limit through their QoS > requests, but it is not incorrect). > > Modify the code in accordance with the above observations. I am not sure if I understood everything you just said, but I don't see how things can break with the current code we have. Both acpi_thermal_cpufreq_init() and acpi_processor_ppc_init() are called from acpi_processor_notifier() which is registered as a policy notifier and is called when a policy is created or removed. Even if some CPUs of a policy go offline, it won't matter as the request for the policy stays and it will be dropped only when all the CPUs of a policy go offline. What am I missing ?
On Fri, Oct 25, 2019 at 4:53 AM Viresh Kumar <viresh.kumar@linaro.org> wrote: > > On 25-10-19, 02:41, Rafael J. Wysocki wrote: > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > The _PPC change notifications from the platform firmware are per-CPU, > > so acpi_processor_ppc_init() needs to add a frequency QoS request > > for each CPU covered by a cpufreq policy to take all of them into > > account. > > > > Even though ACPI thermal control of CPUs sets frequency limits > > per processor package, it also needs a frequency QoS request for each > > CPU in a cpufreq policy in case some of them are taken offline and > > the frequency limit needs to be set through the remaining online > > ones (this is slightly excessive, because all CPUs covered by one > > cpufreq policy will set the same frequency limit through their QoS > > requests, but it is not incorrect). > > > > Modify the code in accordance with the above observations. > > I am not sure if I understood everything you just said, but I don't > see how things can break with the current code we have. > > Both acpi_thermal_cpufreq_init() and acpi_processor_ppc_init() are > called from acpi_processor_notifier() which is registered as a policy > notifier and is called when a policy is created or removed. Even if > some CPUs of a policy go offline, it won't matter as the request for > the policy stays and it will be dropped only when all the CPUs of a > policy go offline. > > What am I missing ? The way the request is used. Say there are two CPUs, A and B, in the same policy. A is policy->cpu, so acpi_processor_ppc_init() adds a QoS request for A only (note that the B's QoS request, B->perflib_req, remains inactive). Now, some time later, the platform firmware notifies the OS of a _PPC change for B. That means acpi_processor_notify() is called and it calls acpi_processor_ppc_has_changed(B) and that invokes acpi_processor_get_platform_limit(B), which in turn looks at the B's QoS request (B->perflib_req) and sees that it is inactive, so 0 is returned without doing anything. However, *some* QoS request should be updated then. Would it be correct to update the A's QoS request in that case? No, because the _PPC limit for A may be different that the _PPC limit for B in principle. The thermal case is not completely analogous, because cpufreq_set_cur_state() finds online CPUs in the same package as the target one and tries to update the QoS request for each of them, which will include the original policy->cpu, whose QoS request has been registered by acpi_thermal_cpufreq_init(), as long as it is online. If it is offline, it will be skipped and there is no easy way to find a "previous policy->cpu". It is possible to do that, but IMO it is more straightforward to have a request for each CPU added.
On Fri, Oct 25, 2019 at 10:17 AM Rafael J. Wysocki <rafael@kernel.org> wrote: > > On Fri, Oct 25, 2019 at 4:53 AM Viresh Kumar <viresh.kumar@linaro.org> wrote: > > > > On 25-10-19, 02:41, Rafael J. Wysocki wrote: > > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > > > The _PPC change notifications from the platform firmware are per-CPU, > > > so acpi_processor_ppc_init() needs to add a frequency QoS request > > > for each CPU covered by a cpufreq policy to take all of them into > > > account. > > > > > > Even though ACPI thermal control of CPUs sets frequency limits > > > per processor package, it also needs a frequency QoS request for each > > > CPU in a cpufreq policy in case some of them are taken offline and > > > the frequency limit needs to be set through the remaining online > > > ones (this is slightly excessive, because all CPUs covered by one > > > cpufreq policy will set the same frequency limit through their QoS > > > requests, but it is not incorrect). > > > > > > Modify the code in accordance with the above observations. > > > > I am not sure if I understood everything you just said, but I don't > > see how things can break with the current code we have. > > > > Both acpi_thermal_cpufreq_init() and acpi_processor_ppc_init() are > > called from acpi_processor_notifier() which is registered as a policy > > notifier and is called when a policy is created or removed. Even if > > some CPUs of a policy go offline, it won't matter as the request for > > the policy stays and it will be dropped only when all the CPUs of a > > policy go offline. > > > > What am I missing ? > > The way the request is used. > > Say there are two CPUs, A and B, in the same policy. A is > policy->cpu, so acpi_processor_ppc_init() adds a QoS request for A > only (note that the B's QoS request, B->perflib_req, remains > inactive). > > Now, some time later, the platform firmware notifies the OS of a _PPC > change for B. That means acpi_processor_notify() is called and it > calls acpi_processor_ppc_has_changed(B) and that invokes > acpi_processor_get_platform_limit(B), which in turn looks at the B's > QoS request (B->perflib_req) and sees that it is inactive, so 0 is > returned without doing anything. However, *some* QoS request should > be updated then. > > Would it be correct to update the A's QoS request in that case? No, > because the _PPC limit for A may be different that the _PPC limit for > B in principle. > > The thermal case is not completely analogous, because > cpufreq_set_cur_state() finds online CPUs in the same package as the > target one and tries to update the QoS request for each of them, which > will include the original policy->cpu, whose QoS request has been > registered by acpi_thermal_cpufreq_init(), as long as it is online. > If it is offline, it will be skipped and there is no easy way to find > a "previous policy->cpu". It is possible to do that, but IMO it is > more straightforward to have a request for each CPU added. BTW, IMO processor_thremal can be changed to use one frequency QoS request per policy on top of this, but I'd rather take one step at a time. :-)
On 25-10-19, 02:41, Rafael J. Wysocki wrote: > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > The _PPC change notifications from the platform firmware are per-CPU, > so acpi_processor_ppc_init() needs to add a frequency QoS request > for each CPU covered by a cpufreq policy to take all of them into > account. > > Even though ACPI thermal control of CPUs sets frequency limits > per processor package, it also needs a frequency QoS request for each > CPU in a cpufreq policy in case some of them are taken offline and > the frequency limit needs to be set through the remaining online > ones (this is slightly excessive, because all CPUs covered by one > cpufreq policy will set the same frequency limit through their QoS > requests, but it is not incorrect). > > Modify the code in accordance with the above observations. > > Fixes: d15ce412737a ("ACPI: cpufreq: Switch to QoS requests instead of cpufreq notifier") > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > --- > drivers/acpi/processor_perflib.c | 38 +++++++++++++++++++++++--------------- > drivers/acpi/processor_thermal.c | 38 +++++++++++++++++++++++--------------- > 2 files changed, 46 insertions(+), 30 deletions(-) Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
On 25-10-19, 10:17, Rafael J. Wysocki wrote: > On Fri, Oct 25, 2019 at 4:53 AM Viresh Kumar <viresh.kumar@linaro.org> wrote: > > > > On 25-10-19, 02:41, Rafael J. Wysocki wrote: > > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > > > The _PPC change notifications from the platform firmware are per-CPU, > > > so acpi_processor_ppc_init() needs to add a frequency QoS request > > > for each CPU covered by a cpufreq policy to take all of them into > > > account. > > > > > > Even though ACPI thermal control of CPUs sets frequency limits > > > per processor package, it also needs a frequency QoS request for each > > > CPU in a cpufreq policy in case some of them are taken offline and > > > the frequency limit needs to be set through the remaining online > > > ones (this is slightly excessive, because all CPUs covered by one > > > cpufreq policy will set the same frequency limit through their QoS > > > requests, but it is not incorrect). > > > > > > Modify the code in accordance with the above observations. > > > > I am not sure if I understood everything you just said, but I don't > > see how things can break with the current code we have. > > > > Both acpi_thermal_cpufreq_init() and acpi_processor_ppc_init() are > > called from acpi_processor_notifier() which is registered as a policy > > notifier and is called when a policy is created or removed. Even if > > some CPUs of a policy go offline, it won't matter as the request for > > the policy stays and it will be dropped only when all the CPUs of a > > policy go offline. > > > > What am I missing ? > > The way the request is used. Yes, I missed the point :)
Index: linux-pm/drivers/acpi/processor_thermal.c =================================================================== --- linux-pm.orig/drivers/acpi/processor_thermal.c +++ linux-pm/drivers/acpi/processor_thermal.c @@ -127,26 +127,34 @@ static int cpufreq_set_cur_state(unsigne void acpi_thermal_cpufreq_init(struct cpufreq_policy *policy) { - int cpu = policy->cpu; - struct acpi_processor *pr = per_cpu(processors, cpu); - int ret; - - if (!pr) - return; - - ret = freq_qos_add_request(&policy->constraints, &pr->thermal_req, - FREQ_QOS_MAX, INT_MAX); - if (ret < 0) - pr_err("Failed to add freq constraint for CPU%d (%d)\n", cpu, - ret); + unsigned int cpu; + + for_each_cpu(cpu, policy->related_cpus) { + struct acpi_processor *pr = per_cpu(processors, cpu); + int ret; + + if (!pr) + continue; + + ret = freq_qos_add_request(&policy->constraints, + &pr->thermal_req, + FREQ_QOS_MAX, INT_MAX); + if (ret < 0) + pr_err("Failed to add freq constraint for CPU%d (%d)\n", + cpu, ret); + } } void acpi_thermal_cpufreq_exit(struct cpufreq_policy *policy) { - struct acpi_processor *pr = per_cpu(processors, policy->cpu); + unsigned int cpu; + + for_each_cpu(cpu, policy->related_cpus) { + struct acpi_processor *pr = per_cpu(processors, policy->cpu); - if (pr) - freq_qos_remove_request(&pr->thermal_req); + if (pr) + freq_qos_remove_request(&pr->thermal_req); + } } #else /* ! CONFIG_CPU_FREQ */ static int cpufreq_get_max_state(unsigned int cpu) Index: linux-pm/drivers/acpi/processor_perflib.c =================================================================== --- linux-pm.orig/drivers/acpi/processor_perflib.c +++ linux-pm/drivers/acpi/processor_perflib.c @@ -159,26 +159,34 @@ void acpi_processor_ignore_ppc_init(void void acpi_processor_ppc_init(struct cpufreq_policy *policy) { - int cpu = policy->cpu; - struct acpi_processor *pr = per_cpu(processors, cpu); - int ret; - - if (!pr) - return; - - ret = freq_qos_add_request(&policy->constraints, &pr->perflib_req, - FREQ_QOS_MAX, INT_MAX); - if (ret < 0) - pr_err("Failed to add freq constraint for CPU%d (%d)\n", cpu, - ret); + unsigned int cpu; + + for_each_cpu(cpu, policy->related_cpus) { + struct acpi_processor *pr = per_cpu(processors, cpu); + int ret; + + if (!pr) + continue; + + ret = freq_qos_add_request(&policy->constraints, + &pr->perflib_req, + FREQ_QOS_MAX, INT_MAX); + if (ret < 0) + pr_err("Failed to add freq constraint for CPU%d (%d)\n", + cpu, ret); + } } void acpi_processor_ppc_exit(struct cpufreq_policy *policy) { - struct acpi_processor *pr = per_cpu(processors, policy->cpu); + unsigned int cpu; + + for_each_cpu(cpu, policy->related_cpus) { + struct acpi_processor *pr = per_cpu(processors, cpu); - if (pr) - freq_qos_remove_request(&pr->perflib_req); + if (pr) + freq_qos_remove_request(&pr->perflib_req); + } } static int acpi_processor_get_performance_control(struct acpi_processor *pr)