Storing pipe opens in the share_mode db ?

[Jeremy Allison]

On Thu, Jun 08, 2006 at 09:51:02PM +0200, Volker Lendecke wrote:
> On Thu, Jun 08, 2006 at 12:09:32PM -0700, Jeremy Allison wrote:
> > Is it worth trying both with a large netbench/Samba4 torture run and
> > seeing if we can detect a problem ?
> 
> Sure....
> 
> But I see only one real contention problem: When tons of
> locks exist on the locking.tdb we might run into the
> O(#oflocks) complexity on fcntl in the Kernel. Everything
> else should in theory be relievable the hash chain length.
> The kernel inefficiency is what I wanted to easy by having
> multiple locking.tdb files, thus trading file descriptors
> for locks per file.

It's not actually the hash chain length I'm worried about,
it's contention on the free list lock. That's the big
bottleneck in large numbers of tdb addition activity.

> But before measuring it this is all hand-waving.

Also very true.

Jeremy.