Thanks Petr promptly response.
On Fri 2020-10-16 11:51:09, Joseph Jang wrote:
> From: josephjang <josephj...@google.com>
>
> Add suspend timeout handler to prevent device stuck during suspend/
> resume process. Suspend timeout handler will dump disk sleep task
> at first round timeout and trigger kernel panic at second round timeout.
> The default timer for each round is 30 seconds.
A better solution would be to resume instead of panic().
[Joseph] suspend_timeout() will trigger kernel panic() only when
suspend thread stuck (deadlock/hang) for 2*30 seconds.
At that moment, I don't know how to resume the suspend thread. So I
just could trigger panic to reboot system.
If you have better suggestions, I am willing to study it.
> Note: Can use following command to simulate suspend hang for testing.
> adb shell echo 1 > /sys/power/pm_hang
This looks dangerous. It adds a simple way to panic() the system.
First, it should get enabled separately. e.g.
CONFIG_TEST_PM_SLEEP_MONITOR.
Second, I would add it as a module that might get loaded
and unloaded.
[Joseph] Agree to enable new compile flag for test module.
I think it is better to create separate patch for the new test module right?
> diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
> index 8b1bb5ee7e5d..6f2679cfd9d1 100644
> --- a/kernel/power/suspend.c
> +++ b/kernel/power/suspend.c
Using kthread looks like an overkill to me. I wonder how this actually
works when the kthreads get freezed. It might be enough to implement
just a timer callback. Start the timer in start_suspend_mon() and
delete it in stop_suspend_mon(). Or do I miss anything?
Anyway, the kthread implementation looks a but hairy. If you really
need to use kthread, I suggest to use kthread_worker API. You would
need to run an init work to setup the RT scheduling. Then you
could just call kthread_queue_delayed_work(()
and kthread_cancel_delayed_work_sync() to start and stop
the monitor.
[Joseph]
Actually, I had ever think we just need to use
add_timer()/del_timer_sync() for start_suspend_mon()/stop_suspend_mon() before.
But I am not sure if add_timer() may cause any performance impact in
suspend thread or not.
So I try to create a suspend monitor kthread and just flip the flag in
suspend thread.
Thank you,
Joseph.
> @@ -114,6 +251,10 @@ static void s2idle_enter(void)
> s2idle_state = S2IDLE_STATE_NONE;
> raw_spin_unlock_irq(&s2idle_lock);
>
> +#ifdef CONFIG_PM_SLEEP_MONITOR
> + start_suspend_mon();
> +#endif
It is better to solve this by defining start_suspend_mon() as empty
function when the config option is disabled. For example, see
how vgacon_text_force() is defined in console.h.
[Joseph] Thank you for good suggestions.
May I know if I could use IS_ENABLED() ?
if (IS_ENABLED(CONFIG_PM_SLEEP_MONITOR))
start_suspend_mon();
Best Regards,
Petr
Thank you,
Joseph.
Petr Mladek <pmla...@suse.com> 於 2020年10月16日 週五 下午5:01寫道：
>
> On Fri 2020-10-16 11:51:09, Joseph Jang wrote:
> > From: josephjang <josephj...@google.com>
> >
> > Add suspend timeout handler to prevent device stuck during suspend/
> > resume process. Suspend timeout handler will dump disk sleep task
> > at first round timeout and trigger kernel panic at second round timeout.
> > The default timer for each round is 30 seconds.
>
> A better solution would be to resume instead of panic().
>
> > Note: Can use following command to simulate suspend hang for testing.
> > adb shell echo 1 > /sys/power/pm_hang
>
> This looks dangerous. It adds a simple way to panic() the system.
>
> First, it should get enabled separately. e.g.
> CONFIG_TEST_PM_SLEEP_MONITOR.
>
> Second, I would add it as a module that might get loaded
> and unloaded.
>
> > diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
> > index 8b1bb5ee7e5d..6f2679cfd9d1 100644
> > --- a/kernel/power/suspend.c
> > +++ b/kernel/power/suspend.c
> > +static int suspend_monitor_kthread(void *arg)
> > +{
> > + long err;
> > + struct sched_param param = {.sched_priority
> > + = MAX_RT_PRIO-1};
> > + static int timeout_count;
> > + static long timeout;
> > +
> > + pr_info("Init ksuspend_mon thread\n");
> > +
> > + sched_setscheduler(current, SCHED_FIFO, ¶m);
> > +
> > + timeout_count = 0;
> > + timeout = MAX_SCHEDULE_TIMEOUT;
> > +
> > + do {
> > + /* Wait suspend timer timeout */
> > + err = wait_event_interruptible_timeout(
> > + power_suspend_waitqueue,
> > + (suspend_mon_toggle != TOGGLE_NONE),
> > + timeout);
> > +
> > + mutex_lock(&suspend_mon_lock);
> > + /* suspend monitor state change */
> > + if (suspend_mon_toggle != TOGGLE_NONE) {
> > + if (suspend_mon_toggle == TOGGLE_START) {
> > + timeout = msecs_to_jiffies(
> > + SUSPEND_TIMER_TIMEOUT_MS);
> > + pr_info("Start suspend monitor\n");
> > + } else if (suspend_mon_toggle == TOGGLE_STOP) {
> > + timeout = MAX_SCHEDULE_TIMEOUT;
> > + timeout_count = 0;
> > + pr_info("Stop suspend monitor\n");
> > + }
> > + suspend_mon_toggle = TOGGLE_NONE;
> > + mutex_unlock(&suspend_mon_lock);
> > + continue;
> > + }
> > + mutex_unlock(&suspend_mon_lock);
> > +
> > + /* suspend monitor event handler */
> > + if (err == 0) {
> > + timeout_count++;
> > + suspend_timeout(timeout_count);
> > + } else if (err == -ERESTARTSYS) {
> > + pr_info("Exit ksuspend_mon!");
> > + break;
> > + }
> > + } while (1);
> > +
> > + return 0;
> > +}
>
> Using kthread looks like an overkill to me. I wonder how this actually
> works when the kthreads get freezed. It might be enough to implement
> just a timer callback. Start the timer in start_suspend_mon() and
> delete it in stop_suspend_mon(). Or do I miss anything?
>
> Anyway, the kthread implementation looks a but hairy. If you really
> need to use kthread, I suggest to use kthread_worker API. You would
> need to run an init work to setup the RT scheduling. Then you
> could just call kthread_queue_delayed_work(()
> and kthread_cancel_delayed_work_sync() to start and stop
> the monitor.
>
>
> > @@ -114,6 +251,10 @@ static void s2idle_enter(void)
> > s2idle_state = S2IDLE_STATE_NONE;
> > raw_spin_unlock_irq(&s2idle_lock);
> >
> > +#ifdef CONFIG_PM_SLEEP_MONITOR
> > + start_suspend_mon();
> > +#endif
>
> It is better to solve this by defining start_suspend_mon() as empty
> function when the config option is disabled. For example, see
> how vgacon_text_force() is defined in console.h.
>
> Best Regards,
> Petr
--
Embedded Software engineer
