| Message ID | 20220311081111.159639-1-zhengzucheng@huawei.com (mailing list archive) |
|---|---|
| State | Rejected, archived |
| Headers | show |
| Series | cpufreq: fix cpufreq_get() can't get correct CPU frequency | expand |
On Fri, Mar 11, 2022 at 04:11:11PM +0800, z00314508 wrote: > From: Zucheng Zheng <zhengzucheng@huawei.com> > > On some specific platforms, the cpufreq driver does not define > cpufreq_driver.get() routine (eg:x86 intel_pstate driver), as a > result, the cpufreq_get() can't get the correct CPU frequency. > > Modern x86 processors include the hardware needed to accurately > calculate frequency over an interval -- APERF, MPERF and the TSC. > Here we use arch_freq_get_on_cpu() in preference to any driver > driver-specific cpufreq_driver.get() routine to get CPU frequency. > > Fixes: f8475cef9008 ("x86: use common aperfmperf_khz_on_cpu() to calculate KHz using APERF/MPERF") > Signed-off-by: Zucheng Zheng <zhengzucheng@huawei.com> > --- > drivers/cpufreq/cpufreq.c | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > index 80f535cc8a75..d777257b4454 100644 > --- a/drivers/cpufreq/cpufreq.c > +++ b/drivers/cpufreq/cpufreq.c > @@ -1806,10 +1806,14 @@ unsigned int cpufreq_get(unsigned int cpu) > { > struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); > unsigned int ret_freq = 0; > + unsigned int freq; > > if (policy) { > down_read(&policy->rwsem); > - if (cpufreq_driver->get) > + freq = arch_freq_get_on_cpu(policy->cpu); > + if (freq) > + ret_freq = freq; > + else if (cpufreq_driver->get) > ret_freq = __cpufreq_get(policy); > up_read(&policy->rwsem); > > -- > 2.18.0.huawei.25 > <formletter> This is not the correct way to submit patches for inclusion in the stable kernel tree. Please read: https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html for how to do this properly. </formletter>
On Fri, Mar 11, 2022 at 9:11 AM z00314508 <zhengzucheng@huawei.com> wrote: > > From: Zucheng Zheng <zhengzucheng@huawei.com> > > On some specific platforms, the cpufreq driver does not define > cpufreq_driver.get() routine (eg:x86 intel_pstate driver), as a I guess you mean the cpufreq driver ->get callback. No, intel_pstate doesn't implement it, because it cannot reliably return the current CPU frequency. > result, the cpufreq_get() can't get the correct CPU frequency. No, it can't, if intel_pstate is the driver, but what's the problem? This function is only called in one place in the kernel and not on x8 even. > Modern x86 processors include the hardware needed to accurately > calculate frequency over an interval -- APERF, MPERF and the TSC. You can compute the average frequency over an interval, but ->get is expected to return the actual current frequency at the time call time. > Here we use arch_freq_get_on_cpu() in preference to any driver > driver-specific cpufreq_driver.get() routine to get CPU frequency. > > Fixes: f8475cef9008 ("x86: use common aperfmperf_khz_on_cpu() to calculate KHz using APERF/MPERF") No kidding. > Signed-off-by: Zucheng Zheng <zhengzucheng@huawei.com> > --- > drivers/cpufreq/cpufreq.c | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > index 80f535cc8a75..d777257b4454 100644 > --- a/drivers/cpufreq/cpufreq.c > +++ b/drivers/cpufreq/cpufreq.c > @@ -1806,10 +1806,14 @@ unsigned int cpufreq_get(unsigned int cpu) > { > struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); > unsigned int ret_freq = 0; > + unsigned int freq; > > if (policy) { > down_read(&policy->rwsem); > - if (cpufreq_driver->get) > + freq = arch_freq_get_on_cpu(policy->cpu); > + if (freq) > + ret_freq = freq; > + else if (cpufreq_driver->get) Again, what problem exactly does this address? > ret_freq = __cpufreq_get(policy); > up_read(&policy->rwsem); > > --
On 2022/3/11 23:52, Rafael J. Wysocki wrote: > On Fri, Mar 11, 2022 at 9:11 AM z00314508 <zhengzucheng@huawei.com> wrote: >> From: Zucheng Zheng <zhengzucheng@huawei.com> >> >> On some specific platforms, the cpufreq driver does not define >> cpufreq_driver.get() routine (eg:x86 intel_pstate driver), as a > I guess you mean the cpufreq driver ->get callback. > > No, intel_pstate doesn't implement it, because it cannot reliably > return the current CPU frequency. > >> result, the cpufreq_get() can't get the correct CPU frequency. > No, it can't, if intel_pstate is the driver, but what's the problem? > This function is only called in one place in the kernel and not on x8 > even. > >> Modern x86 processors include the hardware needed to accurately >> calculate frequency over an interval -- APERF, MPERF and the TSC. > You can compute the average frequency over an interval, but ->get is > expected to return the actual current frequency at the time call time. > >> Here we use arch_freq_get_on_cpu() in preference to any driver >> driver-specific cpufreq_driver.get() routine to get CPU frequency. >> >> Fixes: f8475cef9008 ("x86: use common aperfmperf_khz_on_cpu() to calculate KHz using APERF/MPERF") > No kidding. > >> Signed-off-by: Zucheng Zheng <zhengzucheng@huawei.com> >> --- >> drivers/cpufreq/cpufreq.c | 6 +++++- >> 1 file changed, 5 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c >> index 80f535cc8a75..d777257b4454 100644 >> --- a/drivers/cpufreq/cpufreq.c >> +++ b/drivers/cpufreq/cpufreq.c >> @@ -1806,10 +1806,14 @@ unsigned int cpufreq_get(unsigned int cpu) >> { >> struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); >> unsigned int ret_freq = 0; >> + unsigned int freq; >> >> if (policy) { >> down_read(&policy->rwsem); >> - if (cpufreq_driver->get) >> + freq = arch_freq_get_on_cpu(policy->cpu); >> + if (freq) >> + ret_freq = freq; >> + else if (cpufreq_driver->get) > Again, what problem exactly does this address? Thank you for review. In some scenarios, cpufreq driver ->get is not defined, some driver get the CPU frequency by calling cpufreq_get() will return 0. The modification to this problem is inspired by the implementation of the show_scaling_cur_freq(). > >> ret_freq = __cpufreq_get(policy); >> up_read(&policy->rwsem); >> >> -- > . >
On Tue, Mar 15, 2022 at 3:30 AM zhengzucheng <zhengzucheng@huawei.com> wrote: > > > > On 2022/3/11 23:52, Rafael J. Wysocki wrote: > > On Fri, Mar 11, 2022 at 9:11 AM z00314508 <zhengzucheng@huawei.com> wrote: > >> From: Zucheng Zheng <zhengzucheng@huawei.com> > >> > >> On some specific platforms, the cpufreq driver does not define > >> cpufreq_driver.get() routine (eg:x86 intel_pstate driver), as a > > I guess you mean the cpufreq driver ->get callback. > > > > No, intel_pstate doesn't implement it, because it cannot reliably > > return the current CPU frequency. > > > >> result, the cpufreq_get() can't get the correct CPU frequency. > > No, it can't, if intel_pstate is the driver, but what's the problem? > > This function is only called in one place in the kernel and not on x8 > > even. > > > >> Modern x86 processors include the hardware needed to accurately > >> calculate frequency over an interval -- APERF, MPERF and the TSC. > > You can compute the average frequency over an interval, but ->get is > > expected to return the actual current frequency at the time call time. > > > >> Here we use arch_freq_get_on_cpu() in preference to any driver > >> driver-specific cpufreq_driver.get() routine to get CPU frequency. > >> > >> Fixes: f8475cef9008 ("x86: use common aperfmperf_khz_on_cpu() to calculate KHz using APERF/MPERF") > > No kidding. > > > >> Signed-off-by: Zucheng Zheng <zhengzucheng@huawei.com> > >> --- > >> drivers/cpufreq/cpufreq.c | 6 +++++- > >> 1 file changed, 5 insertions(+), 1 deletion(-) > >> > >> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > >> index 80f535cc8a75..d777257b4454 100644 > >> --- a/drivers/cpufreq/cpufreq.c > >> +++ b/drivers/cpufreq/cpufreq.c > >> @@ -1806,10 +1806,14 @@ unsigned int cpufreq_get(unsigned int cpu) > >> { > >> struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); > >> unsigned int ret_freq = 0; > >> + unsigned int freq; > >> > >> if (policy) { > >> down_read(&policy->rwsem); > >> - if (cpufreq_driver->get) > >> + freq = arch_freq_get_on_cpu(policy->cpu); > >> + if (freq) > >> + ret_freq = freq; > >> + else if (cpufreq_driver->get) > > Again, what problem exactly does this address? > Thank you for review. > > In some scenarios, cpufreq driver ->get is not defined, > some driver get the CPU frequency by calling cpufreq_get() will return 0. Which driver? Again, there is only one calling cpufreq_get() in the kernel tree and it is not on x86. > The modification to this problem is inspired by the implementation of > the show_scaling_cur_freq(). > > > >> ret_freq = __cpufreq_get(policy); > >> up_read(&policy->rwsem); > >> > >> -- The answer to my question appears to be that you want cpufreq_get() to be consistent with show_scaling_cur_freq(). Fair enough, but in that case please make them both call the same lower-level routine implementing the desired behavior so as to avoid code duplication.
diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c index 80f535cc8a75..d777257b4454 100644 --- a/drivers/cpufreq/cpufreq.c +++ b/drivers/cpufreq/cpufreq.c @@ -1806,10 +1806,14 @@ unsigned int cpufreq_get(unsigned int cpu) { struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); unsigned int ret_freq = 0; + unsigned int freq; if (policy) { down_read(&policy->rwsem); - if (cpufreq_driver->get) + freq = arch_freq_get_on_cpu(policy->cpu); + if (freq) + ret_freq = freq; + else if (cpufreq_driver->get) ret_freq = __cpufreq_get(policy); up_read(&policy->rwsem);