FW: Sporadic permission denied error in Samba 4.3 and 4.4, 4.2 OK

[Jeremy Allison]

On Mon, Nov 14, 2016 at 10:54:57PM +0100, Volker Lendecke wrote:
> On Mon, Nov 14, 2016 at 01:50:01PM -0800, Jeremy Allison wrote:
> > On Mon, Nov 14, 2016 at 09:42:53PM +0000, Youzhong Yang wrote:
> > > Yes, it's illumos/SmartOS system running Samba. It's possible there's some kind of bug either in kernel or in user-land programming. If you guys can point out a few critical points to put debug statements, I would be more than happy to debug the issue here.
> > 
> > There's no magic debug that will show a kernel
> > bug. If you're reporting it right, the issue is
> > on 4.5.x on your kernel a sendmsg() call on a
> > connected UNIX domain socket is dropping the
> > message.
> > 
> > That simply should not happen. For a lease break
> > message this isn't anything fancy such as fd-passing,
> > this should be a simple datagram message.
> > 
> > Look at the test program in:
> > 
> > source3/lib/unix_msg/test_source.c
> > 
> > Ensure this works correctly on your kernel.
> > 
> > Can you run directly out of master on a test
> > system, so we know we're looking at identical
> > code ?
> 
> Master has changed vastly, unix_msg.c is gone.

Yeah I know, I was giving him advice on looking
at his current source :-). Ensuring that test
works should be task #1 IMHO.

But if there really is a problem, I'd like to
reproduce on master to fix going forward :-).

> Also, I still have some FreeBSD-specific patches for an overload
> situation: FreeBSD does not block, but instead returns ENOBUFS. We need
> to a) treat ENOBUFS the same as EWOULDBLOCK in the nonblocking path,
> and b) do send-polling in the thread until EWOULDBLOCK vanishes. I could
> imagine something similar in Solaris too.

Oh that's possible ! Youzhong, can you
use dtrace to find out if your kernel is
returning ENOBUFS to the sendmsg() in the
case where the message is being dropped ?

Jeremy.