Re: [PATCH 19/21] clocksource: Improve comment explaining clocks_calc_max_nsecs()'s 50% safety margin
From: John Stultz
Date: Thu Apr 02 2015 - 14:50:22 EST
On Thu, Apr 2, 2015 at 11:43 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Thu, Apr 02, 2015 at 11:41:15AM -0700, John Stultz wrote:
>> On Thu, Apr 2, 2015 at 11:34 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> > On Thu, Apr 02, 2015 at 10:30:18AM -0700, John Stultz wrote:
>> >> > Should we make a further note that the tk_fast things rely on this
>> >> > slack since they're not strongly serialized against this? That is, they
>> >> > can end up using an older cycle_last value and therefore end up with a
>> >> > larger delta than other code.
>> >> Though, even with the tk_fast bits, we expect the update to happen
>> >> regularly, its just that for the benefit of lock-free access we are ok
>> >> with the possible slight inconsistencies (in the mono clock) that
>> >> could happen if we use a slightly stale value mid-update. So I don't
>> >> think the tk_fast bits are actually relying on the slack any more then
>> >> the normal timekeeping code relies on the slack to handle slight
>> >> delays in processing the updates. If we deal with time deltas large
>> >> enough to cause overflows, or time intervals larger then the hardware
>> >> can represent, we're sunk in either case. This 50% margin just makes
>> >> it easier to catch unexpected delays or issues.
>> > Right, so you're saying that even though the fast bits will see slightly
>> > larger deltas than the normal code, they should still not get anywhere
>> > near the 50% because we update much more frequently?
>> Well, they may get to 50% or slightly over (since 50% is the max nohz
>> idle length), but that's likely rare, and we shouldn't get anywhere
>> close to real failure edges (100% be it the mult-overflow or hardware
>> mask limit).
> Right, so the fast thing rely on there being slack. That was my point
> rather. I know they'll not get to the end of slack.
Right, but all of the time code relies on there being slack, since we
can't be sure timers will not be slightly late.
For the fast-timekeper stuff, its just that the deltas could be ever
so slightly larger since the lock-free reads can land during an
update. Its not that its ok for the update to be delayed longer for
fast-timekeepers, or that we expect them to handle larger deltas then
the normal time read logic can handle.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/