From: Maulik Shah maulik.shah@oss.qualcomm.com
[ Upstream commit 500ba33284416255b9a5b50ace24470b6fe77ea5 ]
pm_domain_cpu_gov is selecting a cluster idle state but does not consider latency tolerance of child CPUs. This results in deeper cluster idle state whose latency does not meet latency tolerance requirement.
Select deeper idle state only if global and device latency tolerance of all child CPUs meet.
Test results on SM8750 with 300 usec PM-QoS on CPU0 which is less than domain idle state entry (2150) + exit (1983) usec latency mentioned in devicetree, demonstrate the issue.
# echo 300 > /sys/devices/system/cpu/cpu0/power/pm_qos_resume_latency_us
Before: (Usage is incrementing) ====== # cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states State Time Spent(ms) Usage Rejected Above Below S0 29817 537 8 270 0
# cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states State Time Spent(ms) Usage Rejected Above Below S0 30348 542 8 271 0
After: (Usage is not incrementing due to latency tolerance) ====== # cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states State Time Spent(ms) Usage Rejected Above Below S0 39319 626 14 307 0
# cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states State Time Spent(ms) Usage Rejected Above Below S0 39319 626 14 307 0
Signed-off-by: Maulik Shah maulik.shah@oss.qualcomm.com Fixes: e94999688e3a ("PM / Domains: Add genpd governor for CPUs") Cc: stable@vger.kernel.org Link: https://lore.kernel.org/r/20250709-pmdomain_qos-v2-1-976b12257899@oss.qualco... Signed-off-by: Ulf Hansson ulf.hansson@linaro.org [ replaced cpu_latency_qos_limit() with pm_qos_request(PM_QOS_CPU_DMA_LATENCY) ] Signed-off-by: Sasha Levin sashal@kernel.org --- drivers/base/power/domain_governor.c | 18 ++++++++++++++++-- 1 file changed, 16 insertions(+), 2 deletions(-)
diff --git a/drivers/base/power/domain_governor.c b/drivers/base/power/domain_governor.c index 490ed7deb99a7..35f2e961756b2 100644 --- a/drivers/base/power/domain_governor.c +++ b/drivers/base/power/domain_governor.c @@ -8,6 +8,7 @@ #include <linux/pm_domain.h> #include <linux/pm_qos.h> #include <linux/hrtimer.h> +#include <linux/cpu.h> #include <linux/cpuidle.h> #include <linux/cpumask.h> #include <linux/ktime.h> @@ -254,6 +255,8 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd) struct generic_pm_domain *genpd = pd_to_genpd(pd); struct cpuidle_device *dev; ktime_t domain_wakeup, next_hrtimer; + struct device *cpu_dev; + s64 cpu_constraint, global_constraint; s64 idle_duration_ns; int cpu, i;
@@ -264,6 +267,7 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd) if (!(genpd->flags & GENPD_FLAG_CPU_DOMAIN)) return true;
+ global_constraint = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); /* * Find the next wakeup for any of the online CPUs within the PM domain * and its subdomains. Note, we only need the genpd->cpus, as it already @@ -277,8 +281,16 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd) if (ktime_before(next_hrtimer, domain_wakeup)) domain_wakeup = next_hrtimer; } + + cpu_dev = get_cpu_device(cpu); + if (cpu_dev) { + cpu_constraint = dev_pm_qos_raw_resume_latency(cpu_dev); + if (cpu_constraint < global_constraint) + global_constraint = cpu_constraint; + } }
+ global_constraint *= NSEC_PER_USEC; /* The minimum idle duration is from now - until the next wakeup. */ idle_duration_ns = ktime_to_ns(ktime_sub(domain_wakeup, ktime_get())); if (idle_duration_ns <= 0) @@ -291,8 +303,10 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd) */ i = genpd->state_idx; do { - if (idle_duration_ns >= (genpd->states[i].residency_ns + - genpd->states[i].power_off_latency_ns)) { + if ((idle_duration_ns >= (genpd->states[i].residency_ns + + genpd->states[i].power_off_latency_ns)) && + (global_constraint >= (genpd->states[i].power_on_latency_ns + + genpd->states[i].power_off_latency_ns))) { genpd->state_idx = i; return true; }