Bugfix for tdb transactions

[Rusty Russell]

On Mon, 1 Feb 2010 08:39:48 pm Volker Lendecke wrote:
> On Mon, Feb 01, 2010 at 08:24:44PM +1030, Rusty Russell wrote:
> > On Mon, 1 Feb 2010 06:49:27 pm Rusty Russell wrote:
> > > On Mon, 1 Feb 2010 06:01:12 pm Volker Lendecke wrote:
> > > > Before that is fixed, should we commit my patch to fix the
> > > > problem that happens without the kill -9?
> > > 
> > > I prefer that.  I like it from a simplicity point of view, even though the
> > > larger fix will revert it.
> > 
> > Actually, I changed my mind.  Here is the simplest fix:
> 
> Well, to be honest, I disagree. This changes the rules which
> locks are to be taken when a recovery is done. A
> semantically minimal fix would not change this.

It's late here, but I think we want your patch anyway.  So I'm happy to go
with it.

But I think we should do both.  The recovery area should be protected by
the transaction, global and chain locks.  That way, if we acquire any of
those locks, we can check if recovery is needed; this will give us protection
against kill -9 during a transaction, which we don't have now.

Still working on that patch, turns out it's a bit tricky...

Cheers,
Rusty.