| Message ID | 87r0fg5ocg.fsf@somnus (mailing list archive) |
|---|---|
| State | Mainlined, archived |
| Headers | show |
| Series | [v2] PM: s2idle: Make sure CPUs will wakeup directly on resume | expand |
On Mon, Apr 08, 2024 at 09:02:23AM +0200, Anna-Maria Behnsen wrote: > s2idle works like a regular suspend with freezing processes and freezing > devices. All CPUs except the control CPU go into idle. Once this is > completed the control CPU kicks all other CPUs out of idle, so that they > reenter the idle loop and then enter s2idle state. The control CPU then > issues an swait() on the suspend state and therefore enters the idle loop > as well. > > Due to being kicked out of idle, the other CPUs leave their NOHZ states, > which means the tick is active and the corresponding hrtimer is programmed > to the next jiffie. > > On entering s2idle the CPUs shut down their local clockevent device to > prevent wakeups. The last CPU which enters s2idle shuts down its local > clockevent and freezes timekeeping. > > On resume, one of the CPUs receives the wakeup interrupt, unfreezes > timekeeping and its local clockevent and starts the resume process. At that > point all other CPUs are still in s2idle with their clockevents switched > off. They only resume when they are kicked by another CPU or after resuming > devices and then receiving a device interrupt. > > That means there is no guarantee that all CPUs will wakeup directly on > resume. As a consequence there is no guarantee that timers which are queued > on those CPUs and should expire directly after resume, are handled. Also > timer list timers which are remotely queued to one of those CPUs after > resume will not result in a reprogramming IPI as the tick is > active. Queueing a hrtimer will also not result in a reprogramming IPI > because the first hrtimer event is already in the past. > > The recent introduction of the timer pull model (7ee988770326 ("timers: > Implement the hierarchical pull model")) amplifies this problem, if the > current migrator is one of the non woken up CPUs. When a non pinned timer > list timer is queued and the queuing CPU goes idle, it relies on the still > suspended migrator CPU to expire the timer which will happen by chance. > > The problem exists since commit 8d89835b0467 ("PM: suspend: Do not pause > cpuidle in the suspend-to-idle path"). There the cpuidle_pause() call which > in turn invoked a wakeup for all idle CPUs was moved to a later point in > the resume process. This might not be reached or reached very late because > it waits on a timer of a still suspended CPU. > > Address this by kicking all CPUs out of idle after the control CPU returns > from swait() so that they resume their timers and restore consistent system > state. > > Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218641 > Fixes: 8d89835b0467 ("PM: suspend: Do not pause cpuidle in the suspend-to-idle path") > Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> > Reviewed-by: Thomas Gleixner <tglx@linutronix.de> > Tested-by: Mario Limonciello <mario.limonciello@amd.com> > Cc: stable@kernel.org Cute, Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> > --- > kernel/power/suspend.c | 6 ++++++ > 1 file changed, 6 insertions(+) > > --- a/kernel/power/suspend.c > +++ b/kernel/power/suspend.c > @@ -106,6 +106,12 @@ static void s2idle_enter(void) > swait_event_exclusive(s2idle_wait_head, > s2idle_state == S2IDLE_STATE_WAKE); > > + /* > + * Kick all CPUs to ensure that they resume their timers and restore > + * consistent system state. > + */ > + wake_up_all_idle_cpus(); > + > cpus_read_unlock(); > > raw_spin_lock_irq(&s2idle_lock);
On Mon, 8 Apr 2024 at 09:02, Anna-Maria Behnsen <anna-maria@linutronix.de> wrote: > > s2idle works like a regular suspend with freezing processes and freezing > devices. All CPUs except the control CPU go into idle. Once this is > completed the control CPU kicks all other CPUs out of idle, so that they > reenter the idle loop and then enter s2idle state. The control CPU then > issues an swait() on the suspend state and therefore enters the idle loop > as well. > > Due to being kicked out of idle, the other CPUs leave their NOHZ states, > which means the tick is active and the corresponding hrtimer is programmed > to the next jiffie. > > On entering s2idle the CPUs shut down their local clockevent device to > prevent wakeups. The last CPU which enters s2idle shuts down its local > clockevent and freezes timekeeping. > > On resume, one of the CPUs receives the wakeup interrupt, unfreezes > timekeeping and its local clockevent and starts the resume process. At that > point all other CPUs are still in s2idle with their clockevents switched > off. They only resume when they are kicked by another CPU or after resuming > devices and then receiving a device interrupt. > > That means there is no guarantee that all CPUs will wakeup directly on > resume. As a consequence there is no guarantee that timers which are queued > on those CPUs and should expire directly after resume, are handled. Also > timer list timers which are remotely queued to one of those CPUs after > resume will not result in a reprogramming IPI as the tick is > active. Queueing a hrtimer will also not result in a reprogramming IPI > because the first hrtimer event is already in the past. > > The recent introduction of the timer pull model (7ee988770326 ("timers: > Implement the hierarchical pull model")) amplifies this problem, if the > current migrator is one of the non woken up CPUs. When a non pinned timer > list timer is queued and the queuing CPU goes idle, it relies on the still > suspended migrator CPU to expire the timer which will happen by chance. > > The problem exists since commit 8d89835b0467 ("PM: suspend: Do not pause > cpuidle in the suspend-to-idle path"). There the cpuidle_pause() call which > in turn invoked a wakeup for all idle CPUs was moved to a later point in > the resume process. This might not be reached or reached very late because > it waits on a timer of a still suspended CPU. > > Address this by kicking all CPUs out of idle after the control CPU returns > from swait() so that they resume their timers and restore consistent system > state. > > Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218641 > Fixes: 8d89835b0467 ("PM: suspend: Do not pause cpuidle in the suspend-to-idle path") > Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> > Reviewed-by: Thomas Gleixner <tglx@linutronix.de> > Tested-by: Mario Limonciello <mario.limonciello@amd.com> > Cc: stable@kernel.org Thanks for the detailed commit message! Please add: Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Kind regards Uffe > --- > v2: Fix typos in commit message > --- > kernel/power/suspend.c | 6 ++++++ > 1 file changed, 6 insertions(+) > > --- a/kernel/power/suspend.c > +++ b/kernel/power/suspend.c > @@ -106,6 +106,12 @@ static void s2idle_enter(void) > swait_event_exclusive(s2idle_wait_head, > s2idle_state == S2IDLE_STATE_WAKE); > > + /* > + * Kick all CPUs to ensure that they resume their timers and restore > + * consistent system state. > + */ > + wake_up_all_idle_cpus(); > + > cpus_read_unlock(); > > raw_spin_lock_irq(&s2idle_lock);
On Mon, Apr 8, 2024 at 2:43 PM Ulf Hansson <ulf.hansson@linaro.org> wrote: > > On Mon, 8 Apr 2024 at 09:02, Anna-Maria Behnsen > <anna-maria@linutronix.de> wrote: > > > > s2idle works like a regular suspend with freezing processes and freezing > > devices. All CPUs except the control CPU go into idle. Once this is > > completed the control CPU kicks all other CPUs out of idle, so that they > > reenter the idle loop and then enter s2idle state. The control CPU then > > issues an swait() on the suspend state and therefore enters the idle loop > > as well. > > > > Due to being kicked out of idle, the other CPUs leave their NOHZ states, > > which means the tick is active and the corresponding hrtimer is programmed > > to the next jiffie. > > > > On entering s2idle the CPUs shut down their local clockevent device to > > prevent wakeups. The last CPU which enters s2idle shuts down its local > > clockevent and freezes timekeeping. > > > > On resume, one of the CPUs receives the wakeup interrupt, unfreezes > > timekeeping and its local clockevent and starts the resume process. At that > > point all other CPUs are still in s2idle with their clockevents switched > > off. They only resume when they are kicked by another CPU or after resuming > > devices and then receiving a device interrupt. > > > > That means there is no guarantee that all CPUs will wakeup directly on > > resume. As a consequence there is no guarantee that timers which are queued > > on those CPUs and should expire directly after resume, are handled. Also > > timer list timers which are remotely queued to one of those CPUs after > > resume will not result in a reprogramming IPI as the tick is > > active. Queueing a hrtimer will also not result in a reprogramming IPI > > because the first hrtimer event is already in the past. > > > > The recent introduction of the timer pull model (7ee988770326 ("timers: > > Implement the hierarchical pull model")) amplifies this problem, if the > > current migrator is one of the non woken up CPUs. When a non pinned timer > > list timer is queued and the queuing CPU goes idle, it relies on the still > > suspended migrator CPU to expire the timer which will happen by chance. > > > > The problem exists since commit 8d89835b0467 ("PM: suspend: Do not pause > > cpuidle in the suspend-to-idle path"). There the cpuidle_pause() call which > > in turn invoked a wakeup for all idle CPUs was moved to a later point in > > the resume process. This might not be reached or reached very late because > > it waits on a timer of a still suspended CPU. > > > > Address this by kicking all CPUs out of idle after the control CPU returns > > from swait() so that they resume their timers and restore consistent system > > state. > > > > Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218641 > > Fixes: 8d89835b0467 ("PM: suspend: Do not pause cpuidle in the suspend-to-idle path") > > Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> > > Reviewed-by: Thomas Gleixner <tglx@linutronix.de> > > Tested-by: Mario Limonciello <mario.limonciello@amd.com> > > Cc: stable@kernel.org > > Thanks for the detailed commit message! Please add: > > Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Applied as 6.9-rc material, many thanks to everyone involved!
--- a/kernel/power/suspend.c +++ b/kernel/power/suspend.c @@ -106,6 +106,12 @@ static void s2idle_enter(void) swait_event_exclusive(s2idle_wait_head, s2idle_state == S2IDLE_STATE_WAKE); + /* + * Kick all CPUs to ensure that they resume their timers and restore + * consistent system state. + */ + wake_up_all_idle_cpus(); + cpus_read_unlock(); raw_spin_lock_irq(&s2idle_lock);
s2idle works like a regular suspend with freezing processes and freezing devices. All CPUs except the control CPU go into idle. Once this is completed the control CPU kicks all other CPUs out of idle, so that they reenter the idle loop and then enter s2idle state. The control CPU then issues an swait() on the suspend state and therefore enters the idle loop as well. Due to being kicked out of idle, the other CPUs leave their NOHZ states, which means the tick is active and the corresponding hrtimer is programmed to the next jiffie. On entering s2idle the CPUs shut down their local clockevent device to prevent wakeups. The last CPU which enters s2idle shuts down its local clockevent and freezes timekeeping. On resume, one of the CPUs receives the wakeup interrupt, unfreezes timekeeping and its local clockevent and starts the resume process. At that point all other CPUs are still in s2idle with their clockevents switched off. They only resume when they are kicked by another CPU or after resuming devices and then receiving a device interrupt. That means there is no guarantee that all CPUs will wakeup directly on resume. As a consequence there is no guarantee that timers which are queued on those CPUs and should expire directly after resume, are handled. Also timer list timers which are remotely queued to one of those CPUs after resume will not result in a reprogramming IPI as the tick is active. Queueing a hrtimer will also not result in a reprogramming IPI because the first hrtimer event is already in the past. The recent introduction of the timer pull model (7ee988770326 ("timers: Implement the hierarchical pull model")) amplifies this problem, if the current migrator is one of the non woken up CPUs. When a non pinned timer list timer is queued and the queuing CPU goes idle, it relies on the still suspended migrator CPU to expire the timer which will happen by chance. The problem exists since commit 8d89835b0467 ("PM: suspend: Do not pause cpuidle in the suspend-to-idle path"). There the cpuidle_pause() call which in turn invoked a wakeup for all idle CPUs was moved to a later point in the resume process. This might not be reached or reached very late because it waits on a timer of a still suspended CPU. Address this by kicking all CPUs out of idle after the control CPU returns from swait() so that they resume their timers and restore consistent system state. Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218641 Fixes: 8d89835b0467 ("PM: suspend: Do not pause cpuidle in the suspend-to-idle path") Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Mario Limonciello <mario.limonciello@amd.com> Cc: stable@kernel.org --- v2: Fix typos in commit message --- kernel/power/suspend.c | 6 ++++++ 1 file changed, 6 insertions(+)