Hi Thomas, On 25 January 2015 at 01:07, Thomas Gleixner <t...@linutronix.de> wrote: > On Sat, 24 Jan 2015, Xunlei Pang wrote: > >> Before this, I tried to add some code to catch such problem at the >> time of registering the clocksource, like using the >> CLOCKSOURCE_MASK(), for example 64bit counter will never wrap for >> us. But there may be other values like CLOCKSOURCE_MASK(56), I just >> can't figure out exactly how to do this judge. > > I don't think there is a good way to do so. Registration time is the > wrong place anyway because the problem depends on: > > - The width of the counter > - The frequency of the counter > > The frequency of the counter might even change after registration. Now > add the unknown duration of the suspend to the picture and you're > completely lost. > > All we can do is provide information about the actual wraparound time, > if the CLOCK_SOURCE_SUSPEND_NONSTOP flag is set and the wraparound > time is less than some reasonable margin. >
Yes, we can only deal with it approximately. How about this? 1) Add a new member about reference wraparound time(max system suspend period allowed) to struct clocksource. In __clocksource_updatefreq_scale(), we can use "sec" which already applys 12.5% margin as its value. 2) Add a new tuneable sysctl threshold with a default time period value(for example, 365 days) We can also printk its value when booting or changing its value to notice people about this. 3) then, in timekeeping_resume(), we can compare the reference wraparound of the nonstop clocksource with the sysctl threshold to decide if it is dependable to use. Thanks, Xunlei > > > > -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/