| Message ID | 1470179937-14698-1-git-send-email-pprakash@codeaurora.org (mailing list archive) |
|---|---|
| State | Rejected, archived |
| Headers | show |
Hi Rafael, On 8/2/2016 5:27 PM, Rafael J. Wysocki wrote: > On Tuesday, August 02, 2016 05:18:57 PM Prashanth Prakash wrote: >> cpufreq drivers such as CPPC works on contigious scale and does >> not have a static frequency table. This commit adds cpufreq_stats >> support for such drivers by creating a pseudo frequency table by >> discretizing the contigious scale. > No. > > And why? The motivation is to be able to collect the frequency distribution data via time_in_state sysfs entry as we can collect for other cpufreq drivers that uses a frequency table. Thanks, Prashanth -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tuesday, August 02, 2016 05:41:24 PM Prakash, Prashanth wrote: > Hi Rafael, > > On 8/2/2016 5:27 PM, Rafael J. Wysocki wrote: > > On Tuesday, August 02, 2016 05:18:57 PM Prashanth Prakash wrote: > >> cpufreq drivers such as CPPC works on contigious scale and does > >> not have a static frequency table. This commit adds cpufreq_stats > >> support for such drivers by creating a pseudo frequency table by > >> discretizing the contigious scale. > > No. > > > > And why? > The motivation is to be able to collect the frequency distribution data via > time_in_state sysfs entry as we can collect for other cpufreq drivers that > uses a frequency table. Which can be done by using the cpu_frequency trace point just fine, can't it? Plus, cpufreq_stats don't work with the schedutil governor and fast frequency switching anyway. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Rafael, On 8/2/2016 5:49 PM, Rafael J. Wysocki wrote: > On Tuesday, August 02, 2016 05:41:24 PM Prakash, Prashanth wrote: >> Hi Rafael, >> >> On 8/2/2016 5:27 PM, Rafael J. Wysocki wrote: >>> On Tuesday, August 02, 2016 05:18:57 PM Prashanth Prakash wrote: >>>> cpufreq drivers such as CPPC works on contigious scale and does >>>> not have a static frequency table. This commit adds cpufreq_stats >>>> support for such drivers by creating a pseudo frequency table by >>>> discretizing the contigious scale. >>> No. >>> >>> And why? >> The motivation is to be able to collect the frequency distribution data via >> time_in_state sysfs entry as we can collect for other cpufreq drivers that >> uses a frequency table. > Which can be done by using the cpu_frequency trace point just fine, can't it? Yes, but wanted to make sure existing userspace tools can be used with platforms using CPPC as well. > Plus, cpufreq_stats don't work with the schedutil governor and fast frequency > switching anyway. If we are expecting that usage of cpufreq_stats will slowly go away as we switch to schedutil and more drivers enable fast_switching, then I agree it doesn't make much sense to add support for cpufreq_stats. Thanks, Prashanth -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Aug 3, 2016 at 5:41 PM, Prakash, Prashanth <pprakash@codeaurora.org> wrote: > Hi Rafael, > > On 8/2/2016 5:49 PM, Rafael J. Wysocki wrote: >> On Tuesday, August 02, 2016 05:41:24 PM Prakash, Prashanth wrote: >>> Hi Rafael, >>> >>> On 8/2/2016 5:27 PM, Rafael J. Wysocki wrote: >>>> On Tuesday, August 02, 2016 05:18:57 PM Prashanth Prakash wrote: >>>>> cpufreq drivers such as CPPC works on contigious scale and does >>>>> not have a static frequency table. This commit adds cpufreq_stats >>>>> support for such drivers by creating a pseudo frequency table by >>>>> discretizing the contigious scale. >>>> No. >>>> >>>> And why? >>> The motivation is to be able to collect the frequency distribution data via >>> time_in_state sysfs entry as we can collect for other cpufreq drivers that >>> uses a frequency table. >> Which can be done by using the cpu_frequency trace point just fine, can't it? > > Yes, but wanted to make sure existing userspace tools can be used with > platforms using CPPC as well. powertop uses the cpu_frequency trace point already. Generally, everything that works with intel_pstate should be using it too, as intel_pstate doesn't use cpufreq_stats. >> Plus, cpufreq_stats don't work with the schedutil governor and fast frequency >> switching anyway. > > If we are expecting that usage of cpufreq_stats will slowly go away as we > switch to schedutil and more drivers enable fast_switching, then I agree > it doesn't make much sense to add support for cpufreq_stats. Regardless of the expectations about the future, let's just say that cpufreq_stats only makes sense if frequency tables are in use and only if __target_index() is used to change frequencies. And yes, I would like it to go away, because it is inefficient and clunky. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/cpufreq/cpufreq_stats.c b/drivers/cpufreq/cpufreq_stats.c index 06d3abd..4eb9c41 100644 --- a/drivers/cpufreq/cpufreq_stats.c +++ b/drivers/cpufreq/cpufreq_stats.c @@ -14,6 +14,9 @@ #include <linux/module.h> #include <linux/slab.h> #include <linux/cputime.h> +#include <linux/types.h> + +#define CPUFREQ_PSEUDO_FREQ_TABLE_COUNT (16) static DEFINE_SPINLOCK(cpufreq_stats_lock); @@ -25,6 +28,7 @@ struct cpufreq_stats { unsigned int last_index; u64 *time_in_state; unsigned int *freq_table; + bool pseudo_freq_table; #ifdef CONFIG_CPU_FREQ_STAT_DETAILS unsigned int *trans_table; #endif @@ -130,9 +134,17 @@ static struct attribute_group stats_attr_group = { static int freq_table_get_index(struct cpufreq_stats *stats, unsigned int freq) { int index; - for (index = 0; index < stats->max_state; index++) - if (stats->freq_table[index] == freq) - return index; + + if (stats->pseudo_freq_table) { + for (index = 0; index < stats->max_state; index++) + if (stats->freq_table[index] >= freq) + return index; + } else { + for (index = 0; index < stats->max_state; index++) + if (stats->freq_table[index] == freq) + return index; + } + return -1; } @@ -154,14 +166,14 @@ void cpufreq_stats_free_table(struct cpufreq_policy *policy) void cpufreq_stats_create_table(struct cpufreq_policy *policy) { - unsigned int i = 0, count = 0, ret = -ENOMEM; + unsigned int i = 0, count = 0, ret = -ENOMEM, step_size; struct cpufreq_stats *stats; unsigned int alloc_size; struct cpufreq_frequency_table *pos, *table; - /* We need cpufreq table for creating stats table */ + /* We need cpufreq table or min < max for creating stats table */ table = policy->freq_table; - if (unlikely(!table)) + if (!table && !(policy->min < policy->max)) return; /* stats already initialized */ @@ -173,8 +185,20 @@ void cpufreq_stats_create_table(struct cpufreq_policy *policy) return; /* Find total allocation size */ - cpufreq_for_each_valid_entry(pos, table) - count++; + if (table) { + cpufreq_for_each_valid_entry(pos, table) + count++; + } else { + stats->pseudo_freq_table = true; + count = CPUFREQ_PSEUDO_FREQ_TABLE_COUNT; + step_size = (policy->max - policy->min + 1) / count; + + /* count is larger than min-max range */ + if (!step_size) { + count = policy->max - policy->min + 1; + step_size = 1; + } + } alloc_size = count * sizeof(int) + count * sizeof(u64); @@ -195,10 +219,17 @@ void cpufreq_stats_create_table(struct cpufreq_policy *policy) stats->max_state = count; - /* Find valid-unique entries */ - cpufreq_for_each_valid_entry(pos, table) - if (freq_table_get_index(stats, pos->frequency) == -1) - stats->freq_table[i++] = pos->frequency; + if (table) { + /* Find valid-unique entries */ + cpufreq_for_each_valid_entry(pos, table) + if (freq_table_get_index(stats, pos->frequency) == -1) + stats->freq_table[i++] = pos->frequency; + } else { + /* Create a pseudo frequency table */ + for (i = 0 ; i < count ; i++) + stats->freq_table[i] = policy->min + i * step_size; + stats->freq_table[count-1] = policy->max; + } stats->state_num = i; stats->last_time = get_jiffies_64(); @@ -231,9 +262,15 @@ void cpufreq_stats_record_transition(struct cpufreq_policy *policy, new_index = freq_table_get_index(stats, new_freq); /* We can't do stats->time_in_state[-1]= .. */ - if (old_index == -1 || new_index == -1 || old_index == new_index) + if (old_index == -1 || new_index == -1) return; + if (old_index == new_index) { + if (stats->pseudo_freq_table) + stats->total_trans++; + return; + } + cpufreq_stats_update(stats); stats->last_index = new_index;
cpufreq drivers such as CPPC works on contigious scale and does not have a static frequency table. This commit adds cpufreq_stats support for such drivers by creating a pseudo frequency table by discretizing the contigious scale. Signed-off-by: Prashanth Prakash <pprakash@codeaurora.org> --- drivers/cpufreq/cpufreq_stats.c | 63 ++++++++++++++++++++++++++++++++--------- 1 file changed, 50 insertions(+), 13 deletions(-)