Messages in this thread | | | Date | Tue, 26 Feb 2013 13:17:07 +0100 (CET) | From | Thomas Gleixner <> | Subject | Re: [RFC patch 7/8] timekeeping: Implement a shadow timekeeper |
| |
On Fri, 22 Feb 2013, John Stultz wrote:
> On 02/21/2013 02:51 PM, Thomas Gleixner wrote: > > Use the shadow timekeeper to do the update_wall_time() adjustments and > > then copy it over to the real timekeeper. > > > > Keep the shadow timekeeper in sync when updating stuff outside of > > update_wall_time(). > > > > This allows us to limit the timekeeper_seq hold time to the update of > > the real timekeeper and the vsyscall data in the next patch. > > > > Signed-off-by: Thomas Gleixner<tglx@linutronix.de> > > --- > > So up to here it all looks ok to me (and not so different from my earlier > attempts at the same). > > The only gotcha here that I realized with my earlier patches, is that in order > to do the shadow copy update properly, we are also going to need to merge the > NTP state data into the timekeeper. Otherwise, we could run into odd cases > where as we update the shadow copy, we change the NTP state which then would > affect the non-shadow timekeeping state that is about to be updated. One > example: A the leap second lands, and the tai offset gets bumped in the ntp > state, while we do a similar counter adjustment to the shadow-copy. Then > before the real/active timekeeper is updated, someone gets the tai offset and > applies it to that pre-update timekeeper state, and gets an invalid tai time. > > The down side is that the NTP state data is fairly large, and so adding it to > the timekeeper will cause the memcopys to be a bit more painful. > > I'm looking at the NTP code now to try to see if we can bound where the NTP > state is accessed, so we can maybe thin out what ntp state is linked to > timekeeper updates, and only move that data over to the timekeeper.
Hmm. Can we block the NTP data readout while we are doing the update ?
Thanks,
tglx
| |