Re: [PATCH 1/2] arc_timer: fix tickless idle
Mitsis, Peter <Peter.Mitsis@...>
See [Peter]toggle quoted message Show quoted text
-----Original Message-----[Peter]: Good catch on this off by one error.
First an historical note. The reason the original code only touched the count register was that it was not known if lowering the limit register to a value below that of the count register would trigger an interrupt. (I did not have access to a board with the ARCv2 timer and so was unable to confirm or deny that behavior.) That being said, if changing the limit value from a value higher than the count to a value lower than the count does not generate an interrupt, then I have no objections on that front.
Second, Using timer0_count_register_get() at this point is a no-no as it introduces a race condition that can cause us to lose track of a whole tick. Remember, the timer is always counting. If the elapsed ticks were announced just prior to a "tick boundary", the call to timer0_count_register_set() could then occur just after the "tick boundary". The result of this scenario is that nearly one whole tick may be lost when the modulus operation is performed. This is why we use the <current_count> retrieved from the start of _timer_idle_exit(). Yes, it does introduce some gradual drift, but this can be ameliorated somewhat by only modifying the count if there is more than one tick left in the tickles idle.
Hope this helps.