On Tue, Apr 18, 2017 at 03:53:29PM +0300, Slawa Olhovchenkov wrote: > now_ms posible wrap (by design). > As result delta_ms calcultaed as `TICKS_TO_MS(tick_remain(now_ms, > exp)) + 1` can be negative. This caused negative > timeout.tv_sec/timeout.tv_nsec, failed kevent() (EINVAL) and infinite > loop.
That's an interesting case that we've contemplated already a few weeks ago and didn't see how to make it fail. The only case where it could happen would be when there's a timeout larger than 24.3 days (2147483647 ms to be precise) in the configuration so that a task is scheduled further in the future. But even with such an exact value in the config, the issue would last a single millisecond, as on the next millisecond the difference is reduced by one millisecond and doesn't overflow anymore. We still need to keep it in mind as a corner case to reproduce the bug though. Willy
