AIO, SMB2 and Samba 3.6

[Andrew Bartlett]

On Thu, 2012-02-23 at 09:14 +0100, Volker Lendecke wrote:
> On Thu, Feb 23, 2012 at 12:05:50AM -0800, Jeremy Allison wrote:
> > > Something is really fishy with the glibc aio implementation
> > > on your system. I would think the EPERM is a problem. If you
> > > really need this solved, we need to dig into glibc I think.
> > > Alternatively, you might try the aio_pthread module.
> > 
> > Isn't this the signal race that tridge found with
> > glibc aio users that change uid whilst an aio is
> > outstanding ? When smbd goes back to the root user
> > for the whole process then the Linux thread that finishes
> > doing the aio doesn't have permissions to send the
> > signal the io is finished.
> 
> That sounds like what we see here. Is there a patch to glibc
> available?

More particularly, is there any way to detect this and just (loudly)
disable AIO?  (having AIO silently fail in the default mode Samba ship
seems a little unfortunate)

As per your suggestion, when we put AIO into deployment on the NAS, we
will almost certainly use aio_pthread (backported if required), given
this issue and the well known performance issues.

Thanks for all the help,

Andrew Bartlett

-- 
Andrew Bartlett                                http://samba.org/~abartlet/
Authentication Developer, Samba Team           http://samba.org