cpufreq: schedutil: Update next_freq when cpufreq_limits change
authorXuewen Yan <xuewen.yan@unisoc.com>
Wed, 19 Jul 2023 13:05:27 +0000 (21:05 +0800)
committerIngo Molnar <mingo@kernel.org>
Thu, 5 Oct 2023 20:09:50 +0000 (22:09 +0200)
When cpufreq's policy is 'single', there is a scenario that will
cause sg_policy's next_freq to be unable to update.

When the CPU's util is always max, the cpufreq will be max,
and then if we change the policy's scaling_max_freq to be a
lower freq, indeed, the sg_policy's next_freq need change to
be the lower freq, however, because the cpu_is_busy, the next_freq
would keep the max_freq.

For example:

The cpu7 is a single CPU:

  unisoc:/sys/devices/system/cpu/cpufreq/policy7 # while true;do done& [1] 4737
  unisoc:/sys/devices/system/cpu/cpufreq/policy7 # taskset -p 80 4737
  pid 4737's current affinity mask: ff
  pid 4737's new affinity mask: 80
  unisoc:/sys/devices/system/cpu/cpufreq/policy7 # cat scaling_max_freq
  2301000
  unisoc:/sys/devices/system/cpu/cpufreq/policy7 # cat scaling_cur_freq
  2301000
  unisoc:/sys/devices/system/cpu/cpufreq/policy7 # echo 2171000 > scaling_max_freq
  unisoc:/sys/devices/system/cpu/cpufreq/policy7 # cat scaling_max_freq
  2171000

At this time, the sg_policy's next_freq would stay at 2301000, which
is wrong.

To fix this, add a check for the ->need_freq_update flag.

[ mingo: Clarified the changelog. ]

Co-developed-by: Guohua Yan <guohua.yan@unisoc.com>
Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
Signed-off-by: Guohua Yan <guohua.yan@unisoc.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: "Rafael J. Wysocki" <rafael@kernel.org>
Link: https://lore.kernel.org/r/20230719130527.8074-1-xuewen.yan@unisoc.com
kernel/sched/cpufreq_schedutil.c

index 4492608..458d359 100644 (file)
@@ -350,7 +350,8 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * Except when the rq is capped by uclamp_max.
         */
        if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
-           sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+           sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq &&
+           !sg_policy->need_freq_update) {
                next_f = sg_policy->next_freq;
 
                /* Restore cached freq as next_freq has changed */