Liane Praza wrote: > Steve Peng wrote: >> Hi All, >> >> During the code review phase for the following smf performance >> improvement: >> >> 6351623 Initial manifest-import is slow >> >> http://cr.opensolaris.org/~stevep/6351623-2 >> >> One of David's recent comments from the code review brings up my >> intention of sending this email and soliciting your opinion. >> >> The question is what's the best thing to do if we fail copying the tmpfs >> repository back to the root file system after a successful services >> import. The current implementation is to switch the repository back >> to the /etc/svc/repository.db (may be stale) if we fail the copy. >> The original reason for doing this is if we don't switch back, we may >> stay on the tmpfs repository for a long period of time and may suffer >> more loss of customization in case of the system crash. So basically >> I am >> looking for the best solution to address this. There are couple of >> known alternatives: >> >> (1) exit svc.configd and put the system in maintenance mode >> (2) retry copy operation but this really doesn't address the problem >> IMO. > > I prefer that for now we go into maintenance mode. When we do full > early-manifest-import, it'll be time to think harder about increasing > support for the read-only root scenario, but there's no value in that > for this putback which only moves to tmpfs during the (late) > manifest-import. > > That said, exiting svc.configd isn't going to be sufficient since > startd will restart it. I'm guessing you're suggesting something like > failing the switch operation, having manifest-import exit with > SMF_EXIT_ERR_FATAL and printing an appropriate recovery message to the > screen to help the admin, and having configd restarted by startd to > flush the cache. But, can you be more detailed about your plans for #1? Liane,
Your guess is correct. If we fail to copy back from tmpfs to the root file system then there is a good chance that retry may fail as well and admin's intervene is required to fix the root problem and restart everything. Thanks Steve > > thanks, > liane
