[oi-dev] Updating Hipster and GUI log-in breakage with mounted ZFS datasets

[Jim Klimov]

18 февраля 2015 г. 7:53:49 CET, Nikola M <minikola at gmail.com> пишет:
>There is updating Hipster and GUI log-in breakage with mounted ZFS
>datasets.
>
>Updating Hipster is broken for me since before last 2014.1 update. 
>(Updating from August 2014.1 Hipster)
>I have /opt in separate dataset and it worked fine before, but when 
>updating to last 2014.1 (20141010) with pkg update -v ,  gdm and many 
>services do not start and X with it.
>Seems like some packages during update, insist in putting files inside 
>/opt exclusively on system dataset and it breaks booting rest of the OS
>
>right.
>It reports that /opt is non-empty and insists on it even if it is
>empty, 
>untill also dirs made for mounting sub-datasets are removed. After 
>booting with datasets and subdatasets mounted in /opt it sort of works 
>(boot to gdm) but it is unable to do so after regular update - that is
>a 
>bug.
>
>I tried emptying /opt and then mounting dataset to /opt (and child 
>datasets under /opt) and 20141010 booted and I was able to log in with 
>one administrator account through gdm, but logging with another account
>
>(one that I regularly used in August 2014 Hipster) completely locks 
>machine (X, VTs, everything)..
>This kind of behavior (not able to log in as one user or locking system
>
>on logging in) is what I experienced also one time before in Hipster 
>(Begining of 2014 I think) and I ended up re-making new account and 
>copying only needed data to it .
>
>My suspicion goes to the fact that at my setup, both /opt and
>/home/user 
>datasets have child datasets under it (snapshotting /opt and separate 
>snapshots for application data directories) and that is what I suspect 
>breaks both updating Hipster and locking system on logging in.
>
>Behavior with broken update (and gdm not starting and other services, 
>complaining that /opt is non-empty) is same with /hipster-2015 (I used 
>nwamadm from command line (ouch!) and pkg update to update)
>
>Conclusion is that people that are regularly updating Hipster to 
>ever-updating rolling releases and using separate datasets and 
>sub-datasets end up not being able to start system services, gdm nor X 
>and end up in CLI-only OI , not able to test new X.
>
>I suggest updating Hipster be fixed inside 2014.1 publisher first , to 
>ensure people having Hipster can update to it right, so they can move
>to 
>hipster-2015.
>If not they currently end up with largely unusable system (without 
>educated interventions), not able to follow Hipster updates anymore.  
>(And that breaks Hipster ever updating idea)
>
>
>_______________________________________________
>oi-dev mailing list
>oi-dev at openindiana.org
>http://openindiana.org/mailman/listinfo/oi-dev

Ooh, a victim for my split-root incantations! ;)
You did read my articles on 'Advanced' and otherwise non-stanrard setups, including one on maintaining split-root installations - which is the umbrella name I came up with to cover both versioned per-BE subdatasets under each rpool/ROOT/$BENAME (a split-off 'opt' base should be here as it contains upgradeable parts of the OS image), and shared OS-related datasets like /var/mail or /var/crash or /opt/mythirdpartypackagename that you want literally the same (with linear history) in whatever BE you boot into.

Indeed, there are many tricks big and small, so I encourage you to read my article (take note of non-GNU chmod to protect the mountpoints with immutability) and try out the referenced scripts to take care of smf root-fs mounts and of 'proper' BE cloning and upgrading (code is now on GitHub).

Indeed, these setups despite their benefits are not (yet) classy illumos citizens with core support, but my manual routines began development back on Solaris 10u6 (or was it u4?) and SXCE with LiveUpgrade, and surprisingly seamlessly got adapted to illumos and IPS BEs, and were finally largely scripted to ease the most tedious part - properly ordered boot-time mounts and regular updates.

Works for me on OI dev and Hipster, OmniOS bloody and even Sol10u8 (at least the mounting part, and mind the lack of immutable chmod).

So... have a nice reading ;)

HTH,
Jim Klimov
--
Typos courtesy of K-9 Mail on my Samsung Android