Message ID | 1598943859-21857-2-git-send-email-claude.yen@mediatek.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [1/1] PM: s2idle: Introduce syscore callbacks in s2idle flow | expand |
Quoting Claude Yen (2020-09-01 00:04:19) > From: "claude.yen" <claude.yen@mediatek.com> > > Adding syscore callbacks to s2idle makes the behavior of s2idle become > more similar to Suspend-to-Ram (S2R) and reduces potential porting > effort. > > tick_freeze() in s2idle flow calls sched_clock_suspend() and > timekeeping_suspend(), which both functions are also registered as > syscore callback. sched_clock_suspend() introduced in > commit 3f2552f7e9c5 ("timers/sched_clock: Prevent generic sched_clock > wrap caused by tick_freeze()") is added due to clock wrap issue. > By adding syscore callbacks in s2idle, if there is any syscore callbacks > also required in s2idle, additional porting effort could be saved. > > Besides, in both S2R and s2idle, before the system entering low power > state, there is a state when only one cpu alive with interrupt-disabled, > which is syscore callback designed for. Adding syscore callbacks in > s2idle is feasible option. > > Scenarios to call syscore callback: > S2R: one cpu alive when nonboot cpus are hotplug-ed off > s2idle: one cpu alive when other cpus have enter idle state > Do you have syscore ops that can't be converted to something else like CPU_PM notifier? At this point most syscore code that is important has been converted so I don't see much benefit for this patch. If anything, it will prevent conversions to code that works for both cases. > > Signed-off-by: claude.yen <claude.yen@mediatek.com>
On Tue, 2020-09-01 at 00:40 -0700, Stephen Boyd wrote: > Quoting Claude Yen (2020-09-01 00:04:19) > > From: "claude.yen" <claude.yen@mediatek.com> > > > > Adding syscore callbacks to s2idle makes the behavior of s2idle become > > more similar to Suspend-to-Ram (S2R) and reduces potential porting > > effort. > > > > tick_freeze() in s2idle flow calls sched_clock_suspend() and > > timekeeping_suspend(), which both functions are also registered as > > syscore callback. sched_clock_suspend() introduced in > > commit 3f2552f7e9c5 ("timers/sched_clock: Prevent generic sched_clock > > wrap caused by tick_freeze()") is added due to clock wrap issue. > > By adding syscore callbacks in s2idle, if there is any syscore callbacks > > also required in s2idle, additional porting effort could be saved. > > > > Besides, in both S2R and s2idle, before the system entering low power > > state, there is a state when only one cpu alive with interrupt-disabled, > > which is syscore callback designed for. Adding syscore callbacks in > > s2idle is feasible option. > > > > Scenarios to call syscore callback: > > S2R: one cpu alive when nonboot cpus are hotplug-ed off > > s2idle: one cpu alive when other cpus have enter idle state > > > > Do you have syscore ops that can't be converted to something else like > CPU_PM notifier? At this point most syscore code that is important has > been converted so I don't see much benefit for this patch. If anything, > it will prevent conversions to code that works for both cases. In Mediatek SoC's implementation, battery voltage and remaining capacity are tracked in syscore callback when system going to suspend. The reason to hook callback in syscore is that all devices are suspended, and thus battery output become steady and suitable for measurement. Such callback is not suitable for CPU_PM notifier. If s2idle is applied, the callback to track battery status is not triggered due to syscore is not called by current design. Regards, Claude
diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c index 8719731..be22174 100644 --- a/drivers/cpuidle/cpuidle.c +++ b/drivers/cpuidle/cpuidle.c @@ -23,6 +23,9 @@ #include <linux/suspend.h> #include <linux/tick.h> #include <trace/events/power.h> +#include <linux/cpumask.h> +#include <linux/spinlock.h> +#include <linux/syscore_ops.h> #include "cpuidle.h" @@ -35,6 +38,8 @@ static int enabled_devices; static int off __read_mostly; static int initialized __read_mostly; +static struct cpumask s2idle_cpumask; +static DEFINE_SPINLOCK(s2idle_spinlock); int cpuidle_disabled(void) { @@ -137,15 +142,27 @@ static void enter_s2idle_proper(struct cpuidle_driver *drv, struct cpuidle_device *dev, int index) { ktime_t time_start, time_end; + int cpu; time_start = ns_to_ktime(local_clock()); + cpu = smp_processor_id(); + + tick_freeze(); /* - * trace_suspend_resume() called by tick_freeze() for the last CPU + * trace_suspend_resume() called by syscore_suepnd() for the last CPU * executing it contains RCU usage regarded as invalid in the idle * context, so tell RCU about that. */ - RCU_NONIDLE(tick_freeze()); + spin_lock(&s2idle_spinlock); + cpumask_set_cpu(cpu, &s2idle_cpumask); + if (cpumask_weight(&s2idle_cpumask) == num_online_cpus()) { + system_state = SYSTEM_SUSPEND; + RCU_NONIDLE(syscore_suspend()); + } + + spin_unlock(&s2idle_spinlock); + /* * The state used here cannot be a "coupled" one, because the "coupled" * cpuidle mechanism enables interrupts and doing that with timekeeping @@ -154,12 +171,21 @@ static void enter_s2idle_proper(struct cpuidle_driver *drv, stop_critical_timings(); drv->states[index].enter_s2idle(dev, drv, index); WARN_ON(!irqs_disabled()); + /* - * timekeeping_resume() that will be called by tick_unfreeze() for the + * timekeeping_resume() that will be called by syscore_reume() for the * first CPU executing it calls functions containing RCU read-side * critical sections, so tell RCU about that. */ - RCU_NONIDLE(tick_unfreeze()); + spin_lock(&s2idle_spinlock); + if (cpumask_weight(&s2idle_cpumask) == num_online_cpus()) { + RCU_NONIDLE(syscore_resume()); + system_state = SYSTEM_RUNNING; + } + cpumask_clear_cpu(cpu, &s2idle_cpumask); + spin_unlock(&s2idle_spinlock); + + tick_unfreeze(); start_critical_timings(); time_end = ns_to_ktime(local_clock()); @@ -745,6 +771,8 @@ static int __init cpuidle_init(void) if (cpuidle_disabled()) return -ENODEV; + cpumask_clear(&s2idle_cpumask); + return cpuidle_add_interface(cpu_subsys.dev_root); } diff --git a/kernel/cpu_pm.c b/kernel/cpu_pm.c index 44a2593..6c2f5ce 100644 --- a/kernel/cpu_pm.c +++ b/kernel/cpu_pm.c @@ -12,6 +12,7 @@ #include <linux/notifier.h> #include <linux/spinlock.h> #include <linux/syscore_ops.h> +#include <linux/suspend.h> static ATOMIC_NOTIFIER_HEAD(cpu_pm_notifier_chain); @@ -172,6 +173,14 @@ static int cpu_pm_suspend(void) { int ret; +#ifdef CONFIG_SUSPEND + /* + * Skip cpu_pm_suspend in s2idle flow because cpu_pm notifier callbacks + * are triggered in idle framework + */ + if (s2idle_state != S2IDLE_STATE_NONE) + return 0; +#endif ret = cpu_pm_enter(); if (ret) return ret; @@ -182,6 +191,14 @@ static int cpu_pm_suspend(void) static void cpu_pm_resume(void) { +#ifdef CONFIG_SUSPEND + /* + * Skip cpu_pm_resume in s2idle flow because cpu_pm notifier callbacks + * are triggered in idle framework + */ + if (s2idle_state != S2IDLE_STATE_NONE) + return; +#endif cpu_cluster_pm_exit(); cpu_pm_exit(); } diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c index 6c9c342..19feeed 100644 --- a/kernel/time/tick-common.c +++ b/kernel/time/tick-common.c @@ -528,15 +528,8 @@ void tick_freeze(void) raw_spin_lock(&tick_freeze_lock); tick_freeze_depth++; - if (tick_freeze_depth == num_online_cpus()) { - trace_suspend_resume(TPS("timekeeping_freeze"), - smp_processor_id(), true); - system_state = SYSTEM_SUSPEND; - sched_clock_suspend(); - timekeeping_suspend(); - } else { + if (tick_freeze_depth < num_online_cpus()) tick_suspend_local(); - } raw_spin_unlock(&tick_freeze_lock); } @@ -554,13 +547,7 @@ void tick_unfreeze(void) { raw_spin_lock(&tick_freeze_lock); - if (tick_freeze_depth == num_online_cpus()) { - timekeeping_resume(); - sched_clock_resume(); - system_state = SYSTEM_RUNNING; - trace_suspend_resume(TPS("timekeeping_freeze"), - smp_processor_id(), false); - } else { + if (tick_freeze_depth < num_online_cpus()) { touch_softlockup_watchdog(); tick_resume_local(); }