Re: [PATCH v2 3/3] devfreq: move statistics to separate struct

From: Kamil Konieczny
Date: Thu Dec 05 2019 - 09:31:01 EST

Next message: Thomas Hellstrom: "Re: [PATCH v15 00/23] Generic page walk and ptdump"
Previous message: Alexander Shishkin: "[PATCH 2/2] perf/x86/intel/bts: Fix the use of page_private()"
In reply to: Chanwoo Choi: "Re: [PATCH v2 3/3] devfreq: move statistics to separate struct"
Next in thread: Kamil Konieczny: "[PATCH v2 2/3] devfreq: add clearing transitions stats"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi Chanwoo,

On 05.12.2019 02:46, Chanwoo Choi wrote:
> On 12/5/19 12:00 AM, Kamil Konieczny wrote:
>> Count time and transitions between devfreq frequencies in separate struct
>> for improved code readability and maintenance.
>>
>> Signed-off-by: Kamil Konieczny <k.konieczny@xxxxxxxxxxx>
>> ---
>> Changes in v2:
>> squash three patches into one, do not modify devfreq_profile and separate stats
>> into devfreq_stats
>>
>> drivers/devfreq/devfreq.c | 68 +++++++++++++++++++++++----------------
>> include/linux/devfreq.h | 31 ++++++++++++------
>> 2 files changed, 62 insertions(+), 37 deletions(-)
>>
>> diff --git a/drivers/devfreq/devfreq.c b/drivers/devfreq/devfreq.c
>> index 901af3b66a76..4d50c8f10bd2 100644
>> --- a/drivers/devfreq/devfreq.c
>> +++ b/drivers/devfreq/devfreq.c
>> @@ -198,6 +198,7 @@ static int set_freq_table(struct devfreq *devfreq)
>> */
>> int devfreq_update_status(struct devfreq *devfreq, unsigned long freq)
>> {
>> + struct devfreq_stats *stats = devfreq->stats;
>> int lev, prev_lev, ret = 0;
>> unsigned long long cur_time;
>>
>> @@ -214,8 +215,8 @@ int devfreq_update_status(struct devfreq *devfreq, unsigned long freq)
>> goto out;
>> }
>>
>> - devfreq->time_in_state[prev_lev] +=
>> - cur_time - devfreq->last_stat_updated;
>> + stats->time_in_state[prev_lev] +=
>> + cur_time - stats->last_stat_updated;
>>
>> lev = devfreq_get_freq_level(devfreq, freq);
>> if (lev < 0) {
>> @@ -224,13 +225,12 @@ int devfreq_update_status(struct devfreq *devfreq, unsigned long freq)
>> }
>>
>> if (lev != prev_lev) {
>> - devfreq->trans_table[(prev_lev *
>> - devfreq->profile->max_state) + lev]++;
>> - devfreq->total_trans++;
>> + stats->trans_table[(prev_lev * stats->max_state) + lev]++;
>> + stats->total_trans++;
>> }
>>
>> out:
>> - devfreq->last_stat_updated = cur_time;
>> + stats->last_stat_updated = cur_time;
>> return ret;
>> }
>> EXPORT_SYMBOL(devfreq_update_status);
>> @@ -525,7 +525,7 @@ void devfreq_monitor_resume(struct devfreq *devfreq)
>> msecs_to_jiffies(devfreq->profile->polling_ms));
>>
>> out_update:
>> - devfreq->last_stat_updated = get_jiffies_64();
>> + devfreq->stats->last_stat_updated = get_jiffies_64();
>> devfreq->stop_polling = false;
>>
>> if (devfreq->profile->get_cur_freq &&
>> @@ -735,28 +735,38 @@ struct devfreq *devfreq_add_device(struct device *dev,
>> goto err_out;
>> }
>>
>> - devfreq->trans_table = devm_kzalloc(&devfreq->dev,
>> + devfreq->stats = devm_kzalloc(&devfreq->dev, sizeof(*devfreq->stats),
>> + GFP_KERNEL);
>
> Each devfreq has only one stats structure always. Don't need to define
> it with pointer type. It is enough to define as following without pointer type:
>
> struct devfreq_stats stats;
>
Good point, I will change this.

>> + if (!devfreq->stats) {
>> + mutex_unlock(&devfreq->lock);
>> + err = -ENOMEM;
>> + goto err_devfreq;
>> + }
>> +
>> + devfreq->stats->freq_table = devfreq->profile->freq_table;
>> + devfreq->stats->max_state = devfreq->profile->max_state;
>> + devfreq->stats->trans_table = devm_kzalloc(&devfreq->dev,
>> array3_size(sizeof(unsigned int),
>> - devfreq->profile->max_state,
>> - devfreq->profile->max_state),
>> + devfreq->stats->max_state,
>> + devfreq->stats->max_state),
>> GFP_KERNEL);
>> - if (!devfreq->trans_table) {
>> + if (!devfreq->stats->trans_table) {
>> mutex_unlock(&devfreq->lock);
>> err = -ENOMEM;
>> goto err_devfreq;
>> }
>>
>> - devfreq->time_in_state = devm_kcalloc(&devfreq->dev,
>> - devfreq->profile->max_state,
>> - sizeof(*devfreq->time_in_state),
>> + devfreq->stats->time_in_state = devm_kcalloc(&devfreq->dev,
>> + devfreq->stats->max_state,
>> + sizeof(*devfreq->stats->time_in_state),
>> GFP_KERNEL);
>
> Actually, this patch will be conflict with patch[1].
> [1] https://www.spinics.net/lists/arm-kernel/msg768822.html
>
> First of all, I have to merge patches[1][2] for the devfreq pm-qos
> during v5.5-rc period. It requires the linux-pm's maintainer.
> [2] https://www.spinics.net/lists/arm-kernel/msg769761.html
>
> After finishing the job[1][2], I'll merge this patch
> if finished the review of this patch. Just share the possible merge
> conflict to you.
>

Ok, I will rebase and send when you will need this,
please remind me when it happens.

>
>> - if (!devfreq->time_in_state) {
>> + if (!devfreq->stats->time_in_state) {
>> mutex_unlock(&devfreq->lock);
>> err = -ENOMEM;
>> goto err_devfreq;
>> }
>>
>> - devfreq->last_stat_updated = get_jiffies_64();
>> + devfreq->stats->last_stat_updated = get_jiffies_64();
>>
>> srcu_init_notifier_head(&devfreq->transition_notifier_list);
>>
>> @@ -1435,9 +1445,10 @@ static ssize_t trans_stat_show(struct device *dev,
>> struct device_attribute *attr, char *buf)
>> {
>> struct devfreq *devfreq = to_devfreq(dev);
>> + struct devfreq_stats *stats = devfreq->stats;
>> + unsigned int max_state = stats->max_state;
>> ssize_t len;
>> int i, j;
>> - unsigned int max_state = devfreq->profile->max_state;
>>
>> if (max_state == 0)
>> return sprintf(buf, "Not Supported.\n");
>> @@ -1454,28 +1465,28 @@ static ssize_t trans_stat_show(struct device *dev,
>> len += sprintf(buf + len, " :");
>> for (i = 0; i < max_state; i++)
>> len += sprintf(buf + len, "%10lu",
>> - devfreq->profile->freq_table[i]);
>> + stats->freq_table[i]);
>>
>> len += sprintf(buf + len, " time(ms)\n");
>>
>> for (i = 0; i < max_state; i++) {
>> - if (devfreq->profile->freq_table[i]
>> + if (stats->freq_table[i]
>> == devfreq->previous_freq) {
>> len += sprintf(buf + len, "*");
>> } else {
>> len += sprintf(buf + len, " ");
>> }
>> len += sprintf(buf + len, "%10lu:",
>> - devfreq->profile->freq_table[i]);
>> + stats->freq_table[i]);
>> for (j = 0; j < max_state; j++)
>> len += sprintf(buf + len, "%10u",
>> - devfreq->trans_table[(i * max_state) + j]);
>> + stats->trans_table[(i * max_state) + j]);
>> len += sprintf(buf + len, "%10llu\n", (u64)
>> - jiffies64_to_msecs(devfreq->time_in_state[i]));
>> + jiffies64_to_msecs(stats->time_in_state[i]));
>> }
>>
>> len += sprintf(buf + len, "Total transition : %u\n",
>> - devfreq->total_trans);
>> + stats->total_trans);
>> return len;
>> }
>>
>> @@ -1484,16 +1495,17 @@ static ssize_t trans_stat_store(struct device *dev,
>> const char *buf, size_t count)
>> {
>> struct devfreq *df = to_devfreq(dev);
>> - unsigned int cnt = df->profile->max_state;
>> + struct devfreq_stats *stats = df->stats;
>> + unsigned int cnt = stats->max_state;
>>
>> if (cnt == 0)
>> return count;
>>
>> mutex_lock(&df->lock);
>> - memset(df->time_in_state, 0, cnt * sizeof(u64));
>> - memset(df->trans_table, 0, cnt * cnt * sizeof(int));
>> - df->last_stat_updated = get_jiffies_64();
>> - df->total_trans = 0;
>> + memset(stats->time_in_state, 0, cnt * sizeof(u64));
>> + memset(stats->trans_table, 0, cnt * cnt * sizeof(int));
>> + stats->last_stat_updated = get_jiffies_64();
>> + stats->total_trans = 0;
>> mutex_unlock(&df->lock);
>>
>> return count;
>> diff --git a/include/linux/devfreq.h b/include/linux/devfreq.h
>> index b81a86e47fb9..2715719924e7 100644
>> --- a/include/linux/devfreq.h
>> +++ b/include/linux/devfreq.h
>> @@ -106,6 +106,25 @@ struct devfreq_dev_profile {
>> unsigned int max_state;
>> };
>>
>> +/**
>> + * struct devfreq_stats - Devfreq's transitions stats counters
>
> Devfreq's transitions stats counters -> Statistics of devfreq device behavior

Thanks, changed.

>> + * @freq_table: List of frequencies in ascending order.
>> + * @max_state: The size of freq_table.
>> + * @total_trans: Number of devfreq transitions.
>> + * @trans_table: Statistics of devfreq transitions.
>> + * @time_in_state: Statistics of devfreq states.
>> + * @last_stat_updated: The last time stats were updated.
>> + */
>> +struct devfreq_stats {
>> + unsigned long *freq_table;
>> + unsigned int max_state;
>
> Acutally, I'm sorry I has not yet completely agreed to move
> 'freq_table and 'max_state' from struct devfreq to struct devfreq_stats.
> It has not any critical benefit and any problem.
> So, don't move 'freq_table' and 'max_state'.

I removed these two from devfreq_stats and use devfreq->profile as in original
devfreq code.

>
>> +
>> + unsigned int total_trans;
>> + unsigned int *trans_table;
>> + u64 *time_in_state;
>> + unsigned long long last_stat_updated;
>> +};
>> +
>> /**
>> * struct devfreq - Device devfreq structure
>> * @node: list node - contains the devices with devfreq that have been
>> @@ -131,10 +150,7 @@ struct devfreq_dev_profile {
>> * @suspend_freq: frequency of a device set during suspend phase.
>> * @resume_freq: frequency of a device set in resume phase.
>> * @suspend_count: suspend requests counter for a device.
>> - * @total_trans: Number of devfreq transitions
>> - * @trans_table: Statistics of devfreq transitions
>> - * @time_in_state: Statistics of devfreq states
>> - * @last_stat_updated: The last time stat updated
>> + * @stats: Statistics of devfreq transitions and states times
>
> Statistics of devfreq transitions and states times
> -> Statistics of devfreq device behavior
>

Ok.

>> * @transition_notifier_list: list head of DEVFREQ_TRANSITION_NOTIFIER notifier
>> *
>> * This structure stores the devfreq information for a give device.
>> @@ -171,11 +187,8 @@ struct devfreq {
>> unsigned long resume_freq;
>> atomic_t suspend_count;
>>
>> - /* information for device frequency transition */
>> - unsigned int total_trans;
>> - unsigned int *trans_table;
>> - u64 *time_in_state;
>> - unsigned long long last_stat_updated;
>> + /* information for device frequency transitions */
>> + struct devfreq_stats *stats;
>
> Each devfreq has only one stats structure always. Don't need to define
> it with pointer type. It is enough to define as following without pointer type:
>
> struct devfreq_stats stats;
>
>>
>> struct srcu_notifier_head transition_notifier_list;
>> };
>>
--
Best regards,
Kamil Konieczny
Samsung R&D Institute Poland

Next message: Thomas Hellstrom: "Re: [PATCH v15 00/23] Generic page walk and ptdump"
Previous message: Alexander Shishkin: "[PATCH 2/2] perf/x86/intel/bts: Fix the use of page_private()"
In reply to: Chanwoo Choi: "Re: [PATCH v2 3/3] devfreq: move statistics to separate struct"
Next in thread: Kamil Konieczny: "[PATCH v2 2/3] devfreq: add clearing transitions stats"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]