Date:   Mon, 21 May 2018 09:48:57 -0400
From:   Neil Horman <nhorman@...driver.com>
To:     Michael Tuexen <tuexen@...muenster.de>
Cc:     Marcelo Ricardo Leitner <marcelo.leitner@...il.com>,
        Xin Long <lucien.xin@...il.com>,
        network dev <netdev@...r.kernel.org>,
        linux-sctp@...r.kernel.org, davem@...emloft.net
Subject: Re: [PATCH net-next] sctp: add support for SCTP_REUSE_PORT sockopt

On Mon, May 21, 2018 at 02:16:56PM +0200, Michael Tuexen wrote:
> > On 21. May 2018, at 13:39, Neil Horman <nhorman@...driver.com> wrote:
> > 
> > On Sun, May 20, 2018 at 10:54:04PM -0300, Marcelo Ricardo Leitner wrote:
> >> On Sun, May 20, 2018 at 08:50:59PM -0400, Neil Horman wrote:
> >>> On Sat, May 19, 2018 at 03:44:40PM +0800, Xin Long wrote:
> >>>> This feature is actually already supported by sk->sk_reuse which can be
> >>>> set by SO_REUSEADDR. But it's not working exactly as RFC6458 demands in
> >>>> section 8.1.27, like:
> >>>> 
> >>>>  - This option only supports one-to-one style SCTP sockets
> >>>>  - This socket option must not be used after calling bind()
> >>>>    or sctp_bindx().
> >>>> 
> >>>> Besides, SCTP_REUSE_PORT sockopt should be provided for user's programs.
> >>>> Otherwise, the programs with SCTP_REUSE_PORT from other systems will not
> >>>> work in linux.
> >>>> 
> >>>> This patch reuses sk->sk_reuse and works pretty much as SO_REUSEADDR,
> >>>> just with some extra setup limitations that are neeeded when it is being
> >>>> enabled.
> >>>> 
> >>>> "It should be noted that the behavior of the socket-level socket option
> >>>> to reuse ports and/or addresses for SCTP sockets is unspecified", so it
> >>>> leaves SO_REUSEADDR as is for the compatibility.
> >>>> 
> >>>> Signed-off-by: Xin Long <lucien.xin@...il.com>
> >>>> ---
> >>>> include/uapi/linux/sctp.h |  1 +
> >>>> net/sctp/socket.c         | 48 +++++++++++++++++++++++++++++++++++++++++++++++
> >>>> 2 files changed, 49 insertions(+)
> >>>> 
> >>> A few things:
> >>> 
> >>> 1) I agree with Tom, this feature is a complete duplication of the SK_REUSEPORT
> >>> socket option.  I understand that this is an implementation of the option in the
> >>> RFC, but its definately a duplication of a feature, which makes several things
> >>> really messy.
> >>> 
> >>> 2) The overloading of the sk_reuse opeion is a bad idea, for several reasons.
> >>> Chief among them is the behavioral interference between this patch and the
> >>> SO_REUSEADDR socket level option, that also sets this feature.  If you set
> >>> sk_reuse via SO_REUSEADDR, you will set the SCTP port reuse feature regardless
> >>> of the bind or 1:1/1:m state of the socket.  Vice versa, if you set this socket
> >>> option via the SCTP_PORT_REUSE option you will inadvertently turn on address
> >>> reuse for the socket.  We can't do that.
> >> 
> >> Given your comments, going a bit further here, one other big
> >> implication is that a port would never be able to be considered to
> >> fully meet SCTP standards regarding reuse because a rogue application
> >> may always abuse of the socket level opt to gain access to the port.
> >> 
> >> IOW, the patch allows the application to use such restrictions against
> >> itself and nothing else, which undermines the patch idea.
> >> 
> > Agreed.
> > 
> >> I lack the knowledge on why the SCTP option was proposed in the RFC. I
> >> guess they had a good reason to add the restriction on 1:1/1:m style.
> >> Does the usage of the current imply in any risk to SCTP sockets? If
> >> yes, that would give some grounds for going forward with the SCTP
> >> option.
> >> 
> > I'm also not privy to why the sctp option was proposed, though I expect that the
> > lack of standardization of SO_REUSEPORT probably had something to do with it.
> > As for the reasoning behind restriction to only 1:1 sockets, if I had to guess,
> > I would say it likely because it creates ordering difficulty at the application
> > level.
> > 
> > CC-ing Michael Tuxen, who I believe had some input on this RFC.  Hopefully he
> > can shed some light on this.
> Dear all,
> 
> the reason this was added is to have a specified way to allow a system to
> behave like a client and server making use of the INIT collision.
> 
> For 1-to-many style sockets you can do this by creating a socket, binding it,
> calling listen on it and trying to connect to the peer.
> 
> For 1-to-1 style sockets you need two sockets for it. One listener and one
> you use to connect (and close it in case of failure, open a new one...).
> 
> It was not clear if one can achieve this with SO_REUSEPORT and/or SO_REUSEADDR
> on all platforms. We left that unspecified.
> 
> I hope this makes the intention clearer.
> 
I think it makes the intention clearer yes, but it unfortunately does nothing in
my mind to clarify how the implementation should best handle the potential
overlap in functionality.  What I see here is that we have two functional paths
(the SO_REUSEPORT path and the SCTP_PORT_REUSE path), which may or may not
(depending on the OS implementation achieve the same functional goal (allowing
multiple sockets to share a port while allowing one socket to listen and the
other connect to a remote peer).  If both implementations do the same thing on a
given platform, we can either just alias one to another and be done, but if they
don't then we either have to implement both paths, and ensure that the
SO_REUSEPORT path is a no-op/error return for SCTP sockets, or that each path
implements a distinct feature set that is cleaarly documented.

That said, I think we may be in luck.  Looking at the connect and listen paths,
it appears to me that:

1) Sockets ignore SO_REUSEPORT in the connect and listen paths (save for any
autobinding) so it would appear that the intent of the SCTP rfc can be honored
via SO_REUSEPORT on linux.  

2) SO_REUSEPORT prevents changing state after a bind has occured, so we can honr
that part of the SCTP RFC.

The only missing part is the restriction that SCTP_REUSE_PORT has which is
unaccounted for is that 1:M sockets aren't allowed to enable port reuse.
However, I think the implication from Michaels description above is that port
reuse on a 1:M socket is implicit because a single socket can connect and listen
in that use case, rather than there being a danger to doing so.

As such, I would propose that we implement this socket option by simply setting
the sk->sk_reuseport field in the sock structure, and document the fact that
linux does not restrict port reuse from 1:M sockets.

Thoughts?
Neil

Powered by blists - more mailing lists