Compacting encrypted Realms fail on some devices

[kneth]

Version 0.89.0 of core fixes bugs in SharedGroup::compact() but we see that it fails on some Android devices (on when compacting encrypted Realms). I have been able to single step so I can see that it is SharedGroup::compact() that crashes the app. This is not a duplicate of #757 as we see the crash on an Android 4.0.4 devices (the device used by Jenkins).

See realm/realm-java#955

@tgoyne @finnschiermer @cmelchior

[finnschiermer]

@tgoyne: I just briefly went through SharedGroup::Compact, but didn't immediately spot the problem. Can you see any "misuse" or pieces missing in that method with regard to encryption ?

[tgoyne]

Calling rename() before end_read()/complete_detach() relies on the standard unix behavior of being able to overwrite the filesystem entry of an open file without effecting the file, so if that's not the case on some Android devices then problems wouldn't be too surprising. Otherwise there's nothing obvious.

[finnschiermer]

@tgoyne: Wasn't this fixed by PR #805 ?

[tgoyne]

No, any device suffering from the problem fixed in #805 could not even open a Realm file.

[finnschiermer]

The SharedGroup must be in the attached state (or open) when compact() is called. Has this error ever been seen outside of our java testcase?

[finnschiermer]

I'm going to make compact() assert if called in unattached state.

[kneth]

@finnschiermer Please remember that the Java binding is built with assertions disabled (it was decided a long time ago). If you think we should enable these assertions, please discuss with @timanglade and @bmunkholm.

[finnschiermer]

hrmkay - exceptions then :-)

[kneth]

Or a return value (true if compacting was successfully, false otherwise).

[finnschiermer]

connects to pr #832

[finnschiermer]

turns out, I was wrong about how the binding calls compact, so the hunt continues

[finnschiermer]

We need on-device debugging before we can make real progress.

[cmelchior]

You have it. It is called println ;), but yes the debugging capabilities on a device are horrendously bad :(

[finnschiermer]

just fyi we are stalling this until we have arranged for better debugging (which is almost there), so I'm unassigning myself.

[bmunkholm]

@fsa: Is this possible now?

[bmunkholm]

Ping @finnschiermer

[finnschiermer]

@kneth: can this be reproduced with a newer Core? I'm asking because I think it might actually have been fixed without us noticing.

[kneth]

@finnschiermer It looks like compacting a populated and encrypted Realm leave it corrupted (core version 0.93.0). I'll have to investigate a bit more to draw any conclusions.

[kneth]

Please see #1201.

[finnschiermer]

Now fixed by #1201