Anna-Maria Behnsen <anna-maria@xxxxxxxxxxxxx> writes: [...] > Lets have a deeper look to msleep() internals: msleep() uses timer list > timers. Because of the design of the timer wheel (granularity of buckets > increases with the levels) and because of the granularity of jiffies the > sleep values will be longer as specified. Let's assume we are executing > a msleep(1) on a HZ=250 system: > > First msecs are mapped on jiffies, so this results in a 4ms timeout > value, as there is nothing shorter than 1 jiffie. Then the jiffie value > is handed over to timer code and msleep() adds another jiffie to the > timeout. The timeout is then 2 jiffies or 8ms. With this timeout a timer > list timer is queued. To make sure that timers will not fire early or > race with a concurrent incrementation of jiffie, timers are queued > always into the next bucket. As the timer will end up in the first level > of the timer wheel the granularity of the buckets is 1 jiffies. This > means that the timeout would be 3 jiffies in worst case. > > The additional jiffie in msleep() is the historical prevention that the > sleep time is at least the specified time. This is handled by timer > wheel core code itself, so this extra jiffie could be removed. I will > provide a patch for it. I missed to use the whole cc list above when sending the patch: https://lore.kernel.org/r/20240829074133.4547-1-anna-maria@xxxxxxxxxxxxx/ Thanks, Anna-Maria
