A deadlock has been seen when swicthing clocksources which use PM-runtime. The call path is: change_clocksource ... write_seqcount_begin ... timekeeping_update ... sh_cmt_clocksource_enable ... rpm_resume pm_runtime_mark_last_busy ktime_get do read_seqcount_begin while read_seqcount_retry .... write_seqcount_end Although we should be safe because we haven't yet changed the clocksource at that time, we can't do that because of seqcount protection. Use ktime_get_mono_fast_ns() instead which is lock safe for such cases. With ktime_get_mono_fast_ns, the timestamp is not guaranteed to be monotonic across an update and as a result can goes backward. According to update_fast_timekeeper() description: "In the worst case, this can result is a slightly wrong timestamp (a few nanoseconds)". For PM-runtime autosuspend, this means only that the suspend decision may be slightly suboptimal.
Showing the changes to include/linux (1 of 2 files). Show all
Unified · Split
| @@ -105,7 +105,7 @@ static inline bool pm_runtime_callbacks_present(struct device *dev) · 104 unmodified lines | ||
| 105 | 105 | |
| 106 | 106 | static inline void pm_runtime_mark_last_busy(struct device *dev) |
| 107 | 107 | { |
| 108 | − WRITE_ONCE(dev->power.last_busy, ktime_to_ns(ktime_get())); | |
| 108 | + WRITE_ONCE(dev->power.last_busy, ktime_get_mono_fast_ns()); | |
| 109 | 109 | } |
| 110 | 110 | |
| 111 | 111 | static inline bool pm_runtime_is_irq_safe(struct device *dev) |