Re: [Netnice-kernels] Re: inetd on OpenBSD

[Takashi O. <ta...@cs...>]

hi,

sorry for the slow reply. i will have a short business trip back to japan
this weekend and have been a bit hectic.

Scott Brumbaugh wrote:
> 
> > but, i have been thinking, in such a case, all the sockets should be
> > "disconnected", since the situation in which sockets are connected to
> > a VIF which is a parent of the process which accommodates the sockets
> > is a violation of the VIF association rule.
> >
> 
> This is not clear to me. In inetd all the sockets are inherited by
> the child but they are closed (disconnected) in the inetd parent.  All
> that I see as being incorrect is that the inherited sockets are not
> linked to the correct VIF in the child process.  I only see this as an
> inetd problem.  Maybe you can explain the error in my understanding as I
> don't see this as a problem in the netnice internals.
> 
> The behavior that I observe in both Linux and FreeBSD is that when a
> process creates sockets, the sockets are attached to a copy of that
> process's associated VIF.  When a new process is created it inherits
> the parents process VIF.  A process can change its associated VIF but
> this will not change the VIF association of any sockets currently in
> existence.  Only new sockets will be connected to the new process VIF.
> 
> If a process wants a socket to be associated with a specific VIF it
> needs to either explicitly set the symlink in Pso_dir or it must first
> attach the VIF as its process specific VIF as a symlink under Pso_list
> and then create the socket.
> 
> Looking at what you wrote, are you saying that when a process does the
> symlink operation in Pso_list that all the sockets under Pso_dir need
> to be reattached as well?

i think yes. all the sockets are connected to VIFs, which are the process-
attached VIF or its descendants. otherwise, it violates the protection rule.
to avoid sudden disconnection, it is programs responsibility, to reconnect
the sockets-VIF association first, before reconnects the process-VIF
association.

> I have added the vifctl perl script to the userspace tools under
> netnice/linux.  The changes consisted of a lookup in /etc/mtab for the
> nnfs mount point, and a lookup in /proc/<pid>/exe for the processes
> executable file name. I have also added code to have inetd create the
> symlinks under the socket fd directories as discussed previously.
> These changes are conditionally deactivated now because they are
> causing crashes when enabled.  I do not know what is wrong at this
> point.

this sounds good. thaks.

> I have also committed bug fix changes to NNFS.  One change is a work
> around that appears to prevent the intermittent crashes that where
> occurring in pvif_lookup.  This change affects the pvif returned from
> the lookup of a TCP socket.  Now a pvif will only be returned when a
> TCP socket is in the ESTABLISHED state.

ah. i have never thought about this issue. Linux may have different
semantics of socket lookup. we will check the BSD part. thanks for
the report.


best,

-- taka