| Message ID | 1553828349-8914-1-git-send-email-chang-an.chen@mediatek.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] sched/clock: Prevent generic sched_clock wrap caused by tick_freeze() | expand |
Hi, This is just a gentle ping for this patch. On Fri, 2019-03-29 at 10:59 +0800, Chang-An Chen wrote: > tick_freeze() introduced by suspend-to-idle in commit 124cf9117c5f > ("PM / sleep: Make it possible to quiesce timers during suspend-to-idle") > will use timekeeping_suspend() instead of syscore_suspend() during > suspend-to-idle. It means that generic sched_clock will keep going because > sched_clock_suspend() and sched_clock_resume() are not taken during > suspend-to-idle. This will lead to generic sched_clock wrap. > > For example: > In my arm system with suspend-to-idle enabled, sched_clock is registered > as "56 bits at 13MHz, resolution 76ns, wraps every 4398046511101ns", which > means the real wrapping duration is 8796093022202ns. > > [ 134.551779] suspend-to-idle suspend (timekeeping_suspend()) > [ 1204.912239] suspend-to-idle resume (timekeeping_resume()) > ...... > [ 1206.912239] suspend-to-idle suspend (timekeeping_suspend()) > [ 5880.502807] suspend-to-idle resume (timekeeping_resume()) > ...... > [ 6000.403724] suspend-to-idle suspend (timekeeping_suspend()) > [ 8035.753167] suspend-to-idle resume (timekeeping_resume()) > ...... > [ 8795.786684] (2)[321:charger_thread]...... > [ 8795.788387] (2)[321:charger_thread]...... > [ 0.057226] (0)[0:swapper/0]...... > [ 0.061447] (2)[0:swapper/2]...... > > Sched_clock was not stopped during suspend-to-idle, and sched_clock_poll > hrtimer was not expired because timekeeping_suspend() is taken during > suspend-to-idle. It makes sched_clock wrap at kernel time 8796s. > > To fix this issue, we add sched_clock_suspend() and sched_clock_resume() in > tick_freeze() together with timekeeping_suspend() and timekeeping_resume() > to make sure generic sched_clock wrapping will not happen. > > Signed-off-by: Chang-An Chen <chang-an.chen@mediatek.com> > Fixes: 124cf9117c5f (PM / sleep: Make it possible to quiesce timers during suspend-to-idle) > --- > kernel/time/sched_clock.c | 4 ++-- > kernel/time/tick-common.c | 2 ++ > kernel/time/timekeeping.h | 7 +++++++ > 3 files changed, 11 insertions(+), 2 deletions(-) > > diff --git a/kernel/time/sched_clock.c b/kernel/time/sched_clock.c > index 094b82c..930113b 100644 > --- a/kernel/time/sched_clock.c > +++ b/kernel/time/sched_clock.c > @@ -272,7 +272,7 @@ static u64 notrace suspended_sched_clock_read(void) > return cd.read_data[seq & 1].epoch_cyc; > } > > -static int sched_clock_suspend(void) > +int sched_clock_suspend(void) > { > struct clock_read_data *rd = &cd.read_data[0]; > > @@ -283,7 +283,7 @@ static int sched_clock_suspend(void) > return 0; > } > > -static void sched_clock_resume(void) > +void sched_clock_resume(void) > { > struct clock_read_data *rd = &cd.read_data[0]; > > diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c > index 529143b..df40146 100644 > --- a/kernel/time/tick-common.c > +++ b/kernel/time/tick-common.c > @@ -487,6 +487,7 @@ void tick_freeze(void) > trace_suspend_resume(TPS("timekeeping_freeze"), > smp_processor_id(), true); > system_state = SYSTEM_SUSPEND; > + sched_clock_suspend(); > timekeeping_suspend(); > } else { > tick_suspend_local(); > @@ -510,6 +511,7 @@ void tick_unfreeze(void) > > 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); > diff --git a/kernel/time/timekeeping.h b/kernel/time/timekeeping.h > index 7a9b4eb..141ab3a 100644 > --- a/kernel/time/timekeeping.h > +++ b/kernel/time/timekeeping.h > @@ -14,6 +14,13 @@ extern ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, > extern void timekeeping_warp_clock(void); > extern int timekeeping_suspend(void); > extern void timekeeping_resume(void); > +#ifdef CONFIG_GENERIC_SCHED_CLOCK > +extern int sched_clock_suspend(void); > +extern void sched_clock_resume(void); > +#else > +static inline int sched_clock_suspend(void) { return 0; } > +static inline void sched_clock_resume(void) { } > +#endif > > extern void do_timer(unsigned long ticks); > extern void update_wall_time(void);
diff --git a/kernel/time/sched_clock.c b/kernel/time/sched_clock.c index 094b82c..930113b 100644 --- a/kernel/time/sched_clock.c +++ b/kernel/time/sched_clock.c @@ -272,7 +272,7 @@ static u64 notrace suspended_sched_clock_read(void) return cd.read_data[seq & 1].epoch_cyc; } -static int sched_clock_suspend(void) +int sched_clock_suspend(void) { struct clock_read_data *rd = &cd.read_data[0]; @@ -283,7 +283,7 @@ static int sched_clock_suspend(void) return 0; } -static void sched_clock_resume(void) +void sched_clock_resume(void) { struct clock_read_data *rd = &cd.read_data[0]; diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c index 529143b..df40146 100644 --- a/kernel/time/tick-common.c +++ b/kernel/time/tick-common.c @@ -487,6 +487,7 @@ void tick_freeze(void) trace_suspend_resume(TPS("timekeeping_freeze"), smp_processor_id(), true); system_state = SYSTEM_SUSPEND; + sched_clock_suspend(); timekeeping_suspend(); } else { tick_suspend_local(); @@ -510,6 +511,7 @@ void tick_unfreeze(void) 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); diff --git a/kernel/time/timekeeping.h b/kernel/time/timekeeping.h index 7a9b4eb..141ab3a 100644 --- a/kernel/time/timekeeping.h +++ b/kernel/time/timekeeping.h @@ -14,6 +14,13 @@ extern ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, extern void timekeeping_warp_clock(void); extern int timekeeping_suspend(void); extern void timekeeping_resume(void); +#ifdef CONFIG_GENERIC_SCHED_CLOCK +extern int sched_clock_suspend(void); +extern void sched_clock_resume(void); +#else +static inline int sched_clock_suspend(void) { return 0; } +static inline void sched_clock_resume(void) { } +#endif extern void do_timer(unsigned long ticks); extern void update_wall_time(void);
tick_freeze() introduced by suspend-to-idle in commit 124cf9117c5f ("PM / sleep: Make it possible to quiesce timers during suspend-to-idle") will use timekeeping_suspend() instead of syscore_suspend() during suspend-to-idle. It means that generic sched_clock will keep going because sched_clock_suspend() and sched_clock_resume() are not taken during suspend-to-idle. This will lead to generic sched_clock wrap. For example: In my arm system with suspend-to-idle enabled, sched_clock is registered as "56 bits at 13MHz, resolution 76ns, wraps every 4398046511101ns", which means the real wrapping duration is 8796093022202ns. [ 134.551779] suspend-to-idle suspend (timekeeping_suspend()) [ 1204.912239] suspend-to-idle resume (timekeeping_resume()) ...... [ 1206.912239] suspend-to-idle suspend (timekeeping_suspend()) [ 5880.502807] suspend-to-idle resume (timekeeping_resume()) ...... [ 6000.403724] suspend-to-idle suspend (timekeeping_suspend()) [ 8035.753167] suspend-to-idle resume (timekeeping_resume()) ...... [ 8795.786684] (2)[321:charger_thread]...... [ 8795.788387] (2)[321:charger_thread]...... [ 0.057226] (0)[0:swapper/0]...... [ 0.061447] (2)[0:swapper/2]...... Sched_clock was not stopped during suspend-to-idle, and sched_clock_poll hrtimer was not expired because timekeeping_suspend() is taken during suspend-to-idle. It makes sched_clock wrap at kernel time 8796s. To fix this issue, we add sched_clock_suspend() and sched_clock_resume() in tick_freeze() together with timekeeping_suspend() and timekeeping_resume() to make sure generic sched_clock wrapping will not happen. Signed-off-by: Chang-An Chen <chang-an.chen@mediatek.com> Fixes: 124cf9117c5f (PM / sleep: Make it possible to quiesce timers during suspend-to-idle) --- kernel/time/sched_clock.c | 4 ++-- kernel/time/tick-common.c | 2 ++ kernel/time/timekeeping.h | 7 +++++++ 3 files changed, 11 insertions(+), 2 deletions(-)