| Message ID | 20250212081438.1294503-1-zhanjie9@hisilicon.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v2] cpufreq: governor: Fix negative 'idle_time' handling in dbs_update() | expand |
On Wed, Feb 12, 2025 at 9:22 AM Jie Zhan <zhanjie9@hisilicon.com> wrote: > > We observed an issue that the cpu frequency can't raise up with a 100% cpu > load when nohz is off and the 'conservative' governor is selected. NOHZ > > 'idle_time' can be negative if it's obtained from get_cpu_idle_time_jiffy() > when nohz is off. This was found and explained in commit 9485e4ca0b48 > ("cpufreq: governor: Fix handling of special cases in dbs_update()"). > > However, commit 7592019634f8 ("cpufreq: governors: Fix long idle detection > logic in load calculation") introduced a comparison between 'idle_time' and > 'samling_rate' to detect a long idle interval. While 'idle_time' is > converted to int before comparison, it's actually promoted to unsigned > again when compared with an unsigned 'sampling_rate'. Hence, this leads to > wrong idle interval detection when it's in fact 100% busy and sets > policy_dbs->idle_periods to a very large value. 'conservative' adjusts the > frequency to minimum because of the large 'idle_periods', such that the > frequency can't raise up. 'Ondemand' doesn't use policy_dbs->idle_periods > so it fortunately avoids the issue. > > Correct negative 'idle_time' to 0 before any use of it in dbs_update(). > > Fixes: 7592019634f8 ("cpufreq: governors: Fix long idle detection logic in load calculation") > Signed-off-by: Jie Zhan <zhanjie9@hisilicon.com> > --- > v2: > - Avoid type conversion, compare current and previous idle time before > obtaining 'idle_time'. > - Update the explanation in comments. > > Discussions: > v1: https://lore.kernel.org/linux-pm/20250210130659.3533182-1-zhanjie9@hisilicon.com/ > --- > drivers/cpufreq/cpufreq_governor.c | 42 ++++++++++++++---------------- > 1 file changed, 19 insertions(+), 23 deletions(-) > > diff --git a/drivers/cpufreq/cpufreq_governor.c b/drivers/cpufreq/cpufreq_governor.c > index af44ee6a6430..c140e3f8d4f9 100644 > --- a/drivers/cpufreq/cpufreq_governor.c > +++ b/drivers/cpufreq/cpufreq_governor.c > @@ -145,7 +145,20 @@ unsigned int dbs_update(struct cpufreq_policy *policy) > time_elapsed = update_time - j_cdbs->prev_update_time; > j_cdbs->prev_update_time = update_time; > > - idle_time = cur_idle_time - j_cdbs->prev_cpu_idle; > + /* > + * cur_idle_time could be smaller than j_cdbs->prev_cpu_idle if > + * it's obtained from get_cpu_idle_time_jiffy() when NO_HZ is > + * off, where idle_time is calculated by the difference between > + * time elapsed in jiffies and "busy time" obtained from CPU > + * statistics. If a CPU is 100% busy, the time elapsed and busy > + * time should grow with the same amount in two consecutive > + * samples, but in practice there could be a tiny difference, > + * making the accumulated idle time decrease sometimes. Hence, > + * in this case, idle_time should be regarded as 0 in order to > + * make the further process correct. > + */ > + idle_time = cur_idle_time > j_cdbs->prev_cpu_idle ? > + cur_idle_time - j_cdbs->prev_cpu_idle : 0; Please avoid using the ternary operator in new code. > j_cdbs->prev_cpu_idle = cur_idle_time; > > if (ignore_nice) { > @@ -162,7 +175,7 @@ unsigned int dbs_update(struct cpufreq_policy *policy) > * calls, so the previous load value can be used then. > */ > load = j_cdbs->prev_load; > - } else if (unlikely((int)idle_time > 2 * sampling_rate && > + } else if (unlikely(idle_time > 2 * sampling_rate && > j_cdbs->prev_load)) { > /* > * If the CPU had gone completely idle and a task has > @@ -189,30 +202,13 @@ unsigned int dbs_update(struct cpufreq_policy *policy) > load = j_cdbs->prev_load; > j_cdbs->prev_load = 0; > } else { > - if (time_elapsed >= idle_time) { > - load = 100 * (time_elapsed - idle_time) / time_elapsed; > - } else { > - /* > - * That can happen if idle_time is returned by > - * get_cpu_idle_time_jiffy(). In that case > - * idle_time is roughly equal to the difference > - * between time_elapsed and "busy time" obtained > - * from CPU statistics. Then, the "busy time" > - * can end up being greater than time_elapsed > - * (for example, if jiffies_64 and the CPU > - * statistics are updated by different CPUs), > - * so idle_time may in fact be negative. That > - * means, though, that the CPU was busy all > - * the time (on the rough average) during the > - * last sampling interval and 100 can be > - * returned as the load. > - */ > - load = (int)idle_time < 0 ? 100 : 0; > - } > + load = time_elapsed > idle_time ? > + 100 * (time_elapsed - idle_time) / time_elapsed : > + 0; > j_cdbs->prev_load = load; > } > > - if (unlikely((int)idle_time > 2 * sampling_rate)) { > + if (unlikely(idle_time > 2 * sampling_rate)) { > unsigned int periods = idle_time / sampling_rate; > > if (periods < idle_periods) > --
On 13/02/2025 03:49, Rafael J. Wysocki wrote: > On Wed, Feb 12, 2025 at 9:22 AM Jie Zhan <zhanjie9@hisilicon.com> wrote: >> >> We observed an issue that the cpu frequency can't raise up with a 100% cpu >> load when nohz is off and the 'conservative' governor is selected. > > NOHZ Cool. > >> >> 'idle_time' can be negative if it's obtained from get_cpu_idle_time_jiffy() >> when nohz is off. This was found and explained in commit 9485e4ca0b48 >> ("cpufreq: governor: Fix handling of special cases in dbs_update()"). >> >> However, commit 7592019634f8 ("cpufreq: governors: Fix long idle detection >> logic in load calculation") introduced a comparison between 'idle_time' and >> 'samling_rate' to detect a long idle interval. While 'idle_time' is >> converted to int before comparison, it's actually promoted to unsigned >> again when compared with an unsigned 'sampling_rate'. Hence, this leads to >> wrong idle interval detection when it's in fact 100% busy and sets >> policy_dbs->idle_periods to a very large value. 'conservative' adjusts the >> frequency to minimum because of the large 'idle_periods', such that the >> frequency can't raise up. 'Ondemand' doesn't use policy_dbs->idle_periods >> so it fortunately avoids the issue. >> >> Correct negative 'idle_time' to 0 before any use of it in dbs_update(). >> >> Fixes: 7592019634f8 ("cpufreq: governors: Fix long idle detection logic in load calculation") >> Signed-off-by: Jie Zhan <zhanjie9@hisilicon.com> >> --- >> v2: >> - Avoid type conversion, compare current and previous idle time before >> obtaining 'idle_time'. >> - Update the explanation in comments. >> >> Discussions: >> v1: https://lore.kernel.org/linux-pm/20250210130659.3533182-1-zhanjie9@hisilicon.com/ >> --- >> drivers/cpufreq/cpufreq_governor.c | 42 ++++++++++++++---------------- >> 1 file changed, 19 insertions(+), 23 deletions(-) >> >> diff --git a/drivers/cpufreq/cpufreq_governor.c b/drivers/cpufreq/cpufreq_governor.c >> index af44ee6a6430..c140e3f8d4f9 100644 >> --- a/drivers/cpufreq/cpufreq_governor.c >> +++ b/drivers/cpufreq/cpufreq_governor.c >> @@ -145,7 +145,20 @@ unsigned int dbs_update(struct cpufreq_policy *policy) >> time_elapsed = update_time - j_cdbs->prev_update_time; >> j_cdbs->prev_update_time = update_time; >> >> - idle_time = cur_idle_time - j_cdbs->prev_cpu_idle; >> + /* >> + * cur_idle_time could be smaller than j_cdbs->prev_cpu_idle if >> + * it's obtained from get_cpu_idle_time_jiffy() when NO_HZ is >> + * off, where idle_time is calculated by the difference between >> + * time elapsed in jiffies and "busy time" obtained from CPU >> + * statistics. If a CPU is 100% busy, the time elapsed and busy >> + * time should grow with the same amount in two consecutive >> + * samples, but in practice there could be a tiny difference, >> + * making the accumulated idle time decrease sometimes. Hence, >> + * in this case, idle_time should be regarded as 0 in order to >> + * make the further process correct. >> + */ >> + idle_time = cur_idle_time > j_cdbs->prev_cpu_idle ? >> + cur_idle_time - j_cdbs->prev_cpu_idle : 0; > > Please avoid using the ternary operator in new code. Sure. Will fix this here and below. > >> j_cdbs->prev_cpu_idle = cur_idle_time; >> >> if (ignore_nice) { >> @@ -162,7 +175,7 @@ unsigned int dbs_update(struct cpufreq_policy *policy) >> * calls, so the previous load value can be used then. >> */ >> load = j_cdbs->prev_load; >> - } else if (unlikely((int)idle_time > 2 * sampling_rate && >> + } else if (unlikely(idle_time > 2 * sampling_rate && >> j_cdbs->prev_load)) { >> /* >> * If the CPU had gone completely idle and a task has >> @@ -189,30 +202,13 @@ unsigned int dbs_update(struct cpufreq_policy *policy) >> load = j_cdbs->prev_load; >> j_cdbs->prev_load = 0; >> } else { >> - if (time_elapsed >= idle_time) { >> - load = 100 * (time_elapsed - idle_time) / time_elapsed; >> - } else { >> - /* >> - * That can happen if idle_time is returned by >> - * get_cpu_idle_time_jiffy(). In that case >> - * idle_time is roughly equal to the difference >> - * between time_elapsed and "busy time" obtained >> - * from CPU statistics. Then, the "busy time" >> - * can end up being greater than time_elapsed >> - * (for example, if jiffies_64 and the CPU >> - * statistics are updated by different CPUs), >> - * so idle_time may in fact be negative. That >> - * means, though, that the CPU was busy all >> - * the time (on the rough average) during the >> - * last sampling interval and 100 can be >> - * returned as the load. >> - */ >> - load = (int)idle_time < 0 ? 100 : 0; >> - } >> + load = time_elapsed > idle_time ? >> + 100 * (time_elapsed - idle_time) / time_elapsed : >> + 0; >> j_cdbs->prev_load = load; >> } >> >> - if (unlikely((int)idle_time > 2 * sampling_rate)) { >> + if (unlikely(idle_time > 2 * sampling_rate)) { >> unsigned int periods = idle_time / sampling_rate; >> >> if (periods < idle_periods) >> --
diff --git a/drivers/cpufreq/cpufreq_governor.c b/drivers/cpufreq/cpufreq_governor.c index af44ee6a6430..c140e3f8d4f9 100644 --- a/drivers/cpufreq/cpufreq_governor.c +++ b/drivers/cpufreq/cpufreq_governor.c @@ -145,7 +145,20 @@ unsigned int dbs_update(struct cpufreq_policy *policy) time_elapsed = update_time - j_cdbs->prev_update_time; j_cdbs->prev_update_time = update_time; - idle_time = cur_idle_time - j_cdbs->prev_cpu_idle; + /* + * cur_idle_time could be smaller than j_cdbs->prev_cpu_idle if + * it's obtained from get_cpu_idle_time_jiffy() when NO_HZ is + * off, where idle_time is calculated by the difference between + * time elapsed in jiffies and "busy time" obtained from CPU + * statistics. If a CPU is 100% busy, the time elapsed and busy + * time should grow with the same amount in two consecutive + * samples, but in practice there could be a tiny difference, + * making the accumulated idle time decrease sometimes. Hence, + * in this case, idle_time should be regarded as 0 in order to + * make the further process correct. + */ + idle_time = cur_idle_time > j_cdbs->prev_cpu_idle ? + cur_idle_time - j_cdbs->prev_cpu_idle : 0; j_cdbs->prev_cpu_idle = cur_idle_time; if (ignore_nice) { @@ -162,7 +175,7 @@ unsigned int dbs_update(struct cpufreq_policy *policy) * calls, so the previous load value can be used then. */ load = j_cdbs->prev_load; - } else if (unlikely((int)idle_time > 2 * sampling_rate && + } else if (unlikely(idle_time > 2 * sampling_rate && j_cdbs->prev_load)) { /* * If the CPU had gone completely idle and a task has @@ -189,30 +202,13 @@ unsigned int dbs_update(struct cpufreq_policy *policy) load = j_cdbs->prev_load; j_cdbs->prev_load = 0; } else { - if (time_elapsed >= idle_time) { - load = 100 * (time_elapsed - idle_time) / time_elapsed; - } else { - /* - * That can happen if idle_time is returned by - * get_cpu_idle_time_jiffy(). In that case - * idle_time is roughly equal to the difference - * between time_elapsed and "busy time" obtained - * from CPU statistics. Then, the "busy time" - * can end up being greater than time_elapsed - * (for example, if jiffies_64 and the CPU - * statistics are updated by different CPUs), - * so idle_time may in fact be negative. That - * means, though, that the CPU was busy all - * the time (on the rough average) during the - * last sampling interval and 100 can be - * returned as the load. - */ - load = (int)idle_time < 0 ? 100 : 0; - } + load = time_elapsed > idle_time ? + 100 * (time_elapsed - idle_time) / time_elapsed : + 0; j_cdbs->prev_load = load; } - if (unlikely((int)idle_time > 2 * sampling_rate)) { + if (unlikely(idle_time > 2 * sampling_rate)) { unsigned int periods = idle_time / sampling_rate; if (periods < idle_periods)
We observed an issue that the cpu frequency can't raise up with a 100% cpu load when nohz is off and the 'conservative' governor is selected. 'idle_time' can be negative if it's obtained from get_cpu_idle_time_jiffy() when nohz is off. This was found and explained in commit 9485e4ca0b48 ("cpufreq: governor: Fix handling of special cases in dbs_update()"). However, commit 7592019634f8 ("cpufreq: governors: Fix long idle detection logic in load calculation") introduced a comparison between 'idle_time' and 'samling_rate' to detect a long idle interval. While 'idle_time' is converted to int before comparison, it's actually promoted to unsigned again when compared with an unsigned 'sampling_rate'. Hence, this leads to wrong idle interval detection when it's in fact 100% busy and sets policy_dbs->idle_periods to a very large value. 'conservative' adjusts the frequency to minimum because of the large 'idle_periods', such that the frequency can't raise up. 'Ondemand' doesn't use policy_dbs->idle_periods so it fortunately avoids the issue. Correct negative 'idle_time' to 0 before any use of it in dbs_update(). Fixes: 7592019634f8 ("cpufreq: governors: Fix long idle detection logic in load calculation") Signed-off-by: Jie Zhan <zhanjie9@hisilicon.com> --- v2: - Avoid type conversion, compare current and previous idle time before obtaining 'idle_time'. - Update the explanation in comments. Discussions: v1: https://lore.kernel.org/linux-pm/20250210130659.3533182-1-zhanjie9@hisilicon.com/ --- drivers/cpufreq/cpufreq_governor.c | 42 ++++++++++++++---------------- 1 file changed, 19 insertions(+), 23 deletions(-)