| Message ID | 20240912072231.439332-2-zhanjie9@hisilicon.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | cppc_cpufreq: Rework ->get() error handling when cores are idle | expand |
Hi Ionela, On 12/09/2024 17:43, Ionela Voinescu wrote: ... > > A possible (slimmer) alternative implementation for you to consider > (this merges patches 1 & 2): > > diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c > index bafa32dd375d..c16be9651a6f 100644 > --- a/drivers/cpufreq/cppc_cpufreq.c > +++ b/drivers/cpufreq/cppc_cpufreq.c > @@ -118,6 +118,9 @@ static void cppc_scale_freq_workfn(struct kthread_work *work) > > perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs, > &fb_ctrs); > + if (!perf) > + perf = cpu_data->perf_ctrls.desired_perf; > + I think it's better to just return here. If feedback counters are successfully read but unchanged, the following calculation and update in cppc_scale_freq_workfn() is meaningless because it won't change anything. > cppc_fi->prev_perf_fb_ctrs = fb_ctrs; > > perf <<= SCHED_CAPACITY_SHIFT; > @@ -726,7 +729,7 @@ static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data, > > /* Check to avoid divide-by zero and invalid delivered_perf */ > if (!delta_reference || !delta_delivered) > - return cpu_data->perf_ctrls.desired_perf; > + return 0; This makes sense to me. Here is probably why Patch 2 looks bulky. > > return (reference_perf * delta_delivered) / delta_reference; > } > @@ -736,7 +739,7 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) > struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0}; > struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); > struct cppc_cpudata *cpu_data; > - u64 delivered_perf; > + u64 delivered_perf = 0; > int ret; > > if (!policy) > @@ -747,19 +750,22 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) > cpufreq_cpu_put(policy); > > ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0); > - if (ret) > - return 0; > - > - udelay(2); /* 2usec delay between sampling */ > - > - ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); > - if (ret) > - return 0; > + if (!ret) { > + udelay(2); /* 2usec delay between sampling */ > + ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); > + } > + if (!ret) > + delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, > + &fb_ctrs_t1); TBH, 'if (!ret)' style looks very strange to me. We haven't done so anywhere in cppc_cpufreq, so let's keep consistency and make it easier for people to read and maintain? > + if ((ret == -EFAULT) || !delivered_perf) { > + if (cppc_get_desired_perf(cpu, &delivered_perf)) > + delivered_perf = cpu_data->perf_ctrls.desired_perf; will take this. > + } > > - delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, > - &fb_ctrs_t1); > + if (delivered_perf) > + return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); > > - return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); > + return 0; > } > > disclaimer: not fully checked so likely not "production ready" code :) > > Hope it helps, > Ionela. > >> >> static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state) >> -- >> 2.33.0 >> > How about this? merged patch 1 & 2 as well. diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c index bafa32dd375d..411303f2e8cb 100644 --- a/drivers/cpufreq/cppc_cpufreq.c +++ b/drivers/cpufreq/cppc_cpufreq.c @@ -118,6 +118,9 @@ static void cppc_scale_freq_workfn(struct kthread_work *work) perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs, &fb_ctrs); + if (!perf) + return; + cppc_fi->prev_perf_fb_ctrs = fb_ctrs; perf <<= SCHED_CAPACITY_SHIFT; @@ -726,7 +729,7 @@ static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data, /* Check to avoid divide-by zero and invalid delivered_perf */ if (!delta_reference || !delta_delivered) - return cpu_data->perf_ctrls.desired_perf; + return 0; return (reference_perf * delta_delivered) / delta_reference; } @@ -748,18 +751,32 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0); if (ret) - return 0; + goto out_err; udelay(2); /* 2usec delay between sampling */ ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); if (ret) - return 0; + goto out_err; delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, &fb_ctrs_t1); return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); + +out_err: + /* + * Feedback counters could be 0 when cores are powered down. + * Take desired perf for reflecting frequency in this case. + */ + if (ret == -EFAULT) { + if (cppc_get_desired_perf(cpu, &delivered_perf)) + delivered_perf = cpu_data->perf_ctrls.desired_perf; + + return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); + } + + return 0; } static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state) --- Thanks indeed! Jie
Hi, On Friday 13 Sep 2024 at 20:05:50 (+0800), Jie Zhan wrote: > > Hi Ionela, > > On 12/09/2024 17:43, Ionela Voinescu wrote: > > ... > > > > > A possible (slimmer) alternative implementation for you to consider > > (this merges patches 1 & 2): > > > > diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c > > index bafa32dd375d..c16be9651a6f 100644 > > --- a/drivers/cpufreq/cppc_cpufreq.c > > +++ b/drivers/cpufreq/cppc_cpufreq.c > > @@ -118,6 +118,9 @@ static void cppc_scale_freq_workfn(struct kthread_work *work) > > > > perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs, > > &fb_ctrs); > > + if (!perf) > > + perf = cpu_data->perf_ctrls.desired_perf; > > + > > I think it's better to just return here. > If feedback counters are successfully read but unchanged, the following > calculation and update in cppc_scale_freq_workfn() is meaningless because it > won't change anything. Agreed! > > > cppc_fi->prev_perf_fb_ctrs = fb_ctrs; > > > > perf <<= SCHED_CAPACITY_SHIFT; > > @@ -726,7 +729,7 @@ static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data, > > > > /* Check to avoid divide-by zero and invalid delivered_perf */ > > if (!delta_reference || !delta_delivered) > > - return cpu_data->perf_ctrls.desired_perf; > > + return 0; > > This makes sense to me. > Here is probably why Patch 2 looks bulky. > > > > > return (reference_perf * delta_delivered) / delta_reference; > > } > > @@ -736,7 +739,7 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) > > struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0}; > > struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); > > struct cppc_cpudata *cpu_data; > > - u64 delivered_perf; > > + u64 delivered_perf = 0; > > int ret; > > > > if (!policy) > > @@ -747,19 +750,22 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) > > cpufreq_cpu_put(policy); > > > > ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0); > > - if (ret) > > - return 0; > > - > > - udelay(2); /* 2usec delay between sampling */ > > - > > - ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); > > - if (ret) > > - return 0; > > + if (!ret) { > > + udelay(2); /* 2usec delay between sampling */ > > + ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); > > + } > > + if (!ret) > > + delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, > > + &fb_ctrs_t1); > > TBH, 'if (!ret)' style looks very strange to me. > We haven't done so anywhere in cppc_cpufreq, so let's keep consistency and make > it easier for people to read and maintain? I agree it's a bit of a difficult read, that's why I only sent my code as a suggestion. I did like the benefit of not having to have two different calls to cppc_perf_to_khz() and making the code below common for the error and non-error paths. But it's up to you. > > > + if ((ret == -EFAULT) || !delivered_perf) { > > + if (cppc_get_desired_perf(cpu, &delivered_perf)) > > + delivered_perf = cpu_data->perf_ctrls.desired_perf; > > will take this. > > > + } > > > > - delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, > > - &fb_ctrs_t1); > > + if (delivered_perf) > > + return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); > > > > - return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); > > + return 0; > > } > > > > disclaimer: not fully checked so likely not "production ready" code :) > > > > Hope it helps, > > Ionela. > > > >> > >> static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state) > >> -- > >> 2.33.0 > >> > > > > How about this? merged patch 1 & 2 as well. > > diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c > index bafa32dd375d..411303f2e8cb 100644 > --- a/drivers/cpufreq/cppc_cpufreq.c > +++ b/drivers/cpufreq/cppc_cpufreq.c > @@ -118,6 +118,9 @@ static void cppc_scale_freq_workfn(struct kthread_work *work) > > perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs, > &fb_ctrs); > + if (!perf) > + return; > + > cppc_fi->prev_perf_fb_ctrs = fb_ctrs; > > perf <<= SCHED_CAPACITY_SHIFT; > @@ -726,7 +729,7 @@ static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data, > > /* Check to avoid divide-by zero and invalid delivered_perf */ > if (!delta_reference || !delta_delivered) > - return cpu_data->perf_ctrls.desired_perf; > + return 0; > > return (reference_perf * delta_delivered) / delta_reference; > } > @@ -748,18 +751,32 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) > > ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0); > if (ret) > - return 0; > + goto out_err; > > udelay(2); /* 2usec delay between sampling */ > > ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); > if (ret) > - return 0; > + goto out_err; > > delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, > &fb_ctrs_t1); You need a check here for !delivered_perf (when at least one of the deltas is 0) in which case it would be good to take the same error path below. Something like: if(delivered_perf) return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); else ret = -EFAULT; That's why I did the tricky if/else dance above as we need to take the error path below for multiple cases. Thanks, Ionela. > > return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); > + > +out_err: > + /* > + * Feedback counters could be 0 when cores are powered down. > + * Take desired perf for reflecting frequency in this case. > + */ > + if (ret == -EFAULT) { > + if (cppc_get_desired_perf(cpu, &delivered_perf)) > + delivered_perf = cpu_data->perf_ctrls.desired_perf; > + > + return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); > + } > + > + return 0; > } > > static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state) > --- > > Thanks indeed! > Jie
diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c index bafa32dd375d..6aa3af56924b 100644 --- a/drivers/cpufreq/cppc_cpufreq.c +++ b/drivers/cpufreq/cppc_cpufreq.c @@ -748,18 +748,33 @@ static unsigned int cppc_cpufreq_get_rate(unsigned int cpu) ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0); if (ret) - return 0; + goto out_err; udelay(2); /* 2usec delay between sampling */ ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1); if (ret) - return 0; + goto out_err; delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0, &fb_ctrs_t1); return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); + +out_err: + /* + * Feedback counters could be 0 when cores are powered down. + * Take desired perf for reflecting frequency in this case. + */ + if (ret == -EFAULT) { + ret = cppc_get_desired_perf(cpu, &delivered_perf); + if (ret) + return 0; + + return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf); + } + + return 0; } static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
The CPPC performance feedback counters could return 0 when the target cpu is in a deep idle state, e.g. powered off, and those counters are not powered. In this case, cppc_cpufreq_get_rate() returns 0, and hence, cpufreq_online() gets a false error and doesn't generate a cpufreq policy, which happens in cpufreq_add_dev() when a new cpu device is added. Don't take it as an error and return the frequency corresponding to the desired perf when the feedback counters are 0. Fixes: 6a4fec4f6d30 ("cpufreq: cppc: cppc_cpufreq_get_rate() returns zero in all error cases.") Signed-off-by: Jie Zhan <zhanjie9@hisilicon.com> Reviewed-by: Zeng Heng <zengheng4@huawei.com> --- drivers/cpufreq/cppc_cpufreq.c | 19 +++++++++++++++++-- 1 file changed, 17 insertions(+), 2 deletions(-)