Re: [PATCH v4 bpf-next 00/11] Socket migration for SO_REUSEPORT.

[Kuniyuki Iwashima]

From:   Eric Dumazet <eric.dumazet@xxxxxxxxx>
Date:   Wed, 28 Apr 2021 16:18:30 +0200
> On 4/28/21 3:27 AM, Martin KaFai Lau wrote:
> > On Tue, Apr 27, 2021 at 12:38:58PM -0400, Jason Baron wrote:
> >>
> >>
> >> On 4/26/21 11:46 PM, Kuniyuki Iwashima wrote:
> >>> The SO_REUSEPORT option allows sockets to listen on the same port and to
> >>> accept connections evenly. However, there is a defect in the current
> >>> implementation [1]. When a SYN packet is received, the connection is tied
> >>> to a listening socket. Accordingly, when the listener is closed, in-flight
> >>> requests during the three-way handshake and child sockets in the accept
> >>> queue are dropped even if other listeners on the same port could accept
> >>> such connections.
> >>>
> >>> This situation can happen when various server management tools restart
> >>> server (such as nginx) processes. For instance, when we change nginx
> >>> configurations and restart it, it spins up new workers that respect the new
> >>> configuration and closes all listeners on the old workers, resulting in the
> >>> in-flight ACK of 3WHS is responded by RST.
> >>
> >> Hi Kuniyuki,
> >>
> >> I had implemented a different approach to this that I wanted to get your
> >> thoughts about. The idea is to use unix sockets and SCM_RIGHTS to pass the
> >> listen fd (or any other fd) around. Currently, if you have an 'old' webserver
> >> that you want to replace with a 'new' webserver, you would need a separate
> >> process to receive the listen fd and then have that process send the fd to
> >> the new webserver, if they are not running con-currently. So instead what
> >> I'm proposing is a 'delayed close' for a unix socket. That is, one could do:
> >>
> >> 1) bind unix socket with path '/sockets'
> >> 2) sendmsg() the listen fd via the unix socket
> >> 2) setsockopt() some 'timeout' on the unix socket (maybe 10 seconds or so)
> >> 3) exit/close the old webserver and the listen socket
> >> 4) start the new webserver
> >> 5) create new unix socket and bind to '/sockets' (if has MAY_WRITE file permissions)
> >> 6) recvmsg() the listen fd
> >>
> >> So the idea is that we set a timeout on the unix socket. If the new process
> >> does not start and bind to the unix socket, it simply closes, thus releasing
> >> the listen socket. However, if it does bind it can now call recvmsg() and
> >> use the listen fd as normal. It can then simply continue to use the old listen
> >> fds and/or create new ones and drain the old ones.
> >>
> >> Thus, the old and new webservers do not have to run concurrently. This doesn't
> >> involve any changes to the tcp layer and can be used to pass any type of fd.
> >> not sure if it's actually useful for anything else though.
> > We also used to do tcp-listen(/udp) fd transfer because the new process can not
> > bind to the same IP:PORT in the old kernel without SO_REUSEPORT.  Some of the
> > services listen to many different IP:PORT(s).  Transferring all of them
> > was ok-ish but the old and new process do not necessary listen to the same set
> > of IP:PORT(s) (e.g. the config may have changed during restart) and it further
> > complicates the fd transfer logic in the userspace.
> > 
> > It was then moved to SO_REUSEPORT.  The new process can create its listen fds
> > without depending on the old process.  It pretty much starts as if there is
> > no old process.  There is no need to transfer the fds, simplified the userspace
> > logic.  The old and new process can work independently.  The old and new process
> > still run concurrently for a brief time period to avoid service disruption.
> > 
> 
> 
> Note that another technique is to force syncookies during the switch of old/new servers.
> 
> echo 2 >/proc/sys/net/ipv4/tcp_syncookies
> 
> If there is interest, we could add a socket option to override the sysctl on a per-socket basis.

It can be a work-around but syncookies has its own downside. Forcing it may
lose some valuable TCP options. If there is an approach without syncookies,
it is better.