Re: [PATCH RFC] ipvs: reschedule new connections if previous was on FIN_

To: Julian Anastasov <ja@xxxxxx>
Subject: Re: [PATCH RFC] ipvs: reschedule new connections if previous was on FIN_WAIT or TIME_WAIT
Cc: lvs-devel@xxxxxxxxxxxxxxx, hannes@xxxxxxxxxx, jbrouer@xxxxxxxxxx
From: Marcelo Ricardo Leitner <mleitner@xxxxxxxxxx>
Date: Wed, 10 Dec 2014 10:34:18 -0200
On 09-12-2014 21:37, Julian Anastasov wrote:


On Mon, 8 Dec 2014, Marcelo Ricardo Leitner wrote:

Signed-off-by: Marcelo Ricardo Leitner <mleitner@xxxxxxxxxx>


     We have a report that not doing so may cause poor load balacing if
     applications reuse src port. With a patch like this, it would make
     new SYNs on a given connection to drop the old one and start a new

        People complained about UDP, such as RADIUS, etc.
I guess, for TCP it is some local client that can benefit
from balancing, it can be also a local testing tool.

So it's a new type of issue, ok. I don't have info on what their application is, just that it ends up reusing port numbers.

     One could say that this reuse can be done on purpose and carefully
     as a way to cause poor load balancing to cause a DoS.

     Thing is, I'm unsure if we really should do this, as it may end up
     doing more harm than good.

     WDYT? And if we do additional checks, like at least validating seq
     number, would it be better?

        I think, checking of SEQ will not help, tw_recycle
works by checking the timestamp option, SEQ of SYN is

I thought for time wait assassination it would have to use a bigger seq together with the timestamps. But anyway, Windows (since Vista, even 8) doesn't use timestamps by default, unfortunately. Windows is not really involved, it's just to have that in mind, as by relying on timestamps we would be limiting the solution.

        Also, not all connections can be expired, for
example controlling conns (FTP CTL), i.e. our attempt
to expire conn will not succeed if FTP DATA is still in
progress (cp->n_control != 0 check) or in some FW/TW state.

Interesting. I hadn't thought about this case.


  net/netfilter/ipvs/ip_vs_core.c | 15 ++++++++++++---
  1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/net/netfilter/ipvs/ip_vs_core.c b/net/netfilter/ipvs/ip_vs_core.c
--- a/net/netfilter/ipvs/ip_vs_core.c
+++ b/net/netfilter/ipvs/ip_vs_core.c
@@ -1036,6 +1036,14 @@ static inline bool is_new_conn(const struct sk_buff *skb,

+static inline bool is_new_conn_expected(const struct ip_vs_conn *cp)
+       if (cp->protocol != IPPROTO_TCP)
+               return false;
+       return (cp->state == IP_VS_TCP_S_TIME_WAIT) ||
+               (cp->state == IP_VS_TCP_S_FIN_WAIT);

        For conns with cp->flags & IP_VS_CONN_F_NOOUTPUT the
FIN_WAIT state is problematic, we see only packets from
client (eg. DR method) and can not tell if server has sent its
FIN packet. I.e. diverting large transfer from one real to
another will lead to sending of FIN+ACKs from client to
new place.

        So, I'm not sure if we should restrict the FW state, eg:

        return (cp->state == IP_VS_TCP_S_TIME_WAIT) ||
               (cp->state == IP_VS_TCP_S_FIN_WAIT &&
                (cp->flags & IP_VS_CONN_F_NOOUTPUT) &&
                time_after_eq(jiffies + cp->timeout,
                              cp->timer.expires + 1 * HZ));

not expired because only one side sent FIN. For

And when the other side send it too, it will be in TIME_WAIT then. Ok.

INPUT-ONLY we can allow expiration in IP_VS_TCP_S_FIN_WAIT,
for example, when FIN+ACK was not seen recently, say in
last second?

That's interesting.

        If such checks look dangerous we can try to expire

Or only enable it within some sysctl, like we have expire_nodest_conn? We could have a "(tcp_)rebalance_on_port_reuse" or something like that.

  /* Handle response packets: rewrite addresses and send away...
  static unsigned int
@@ -1642,9 +1650,10 @@ ip_vs_in(unsigned int hooknum, struct sk_buff *skb, int 
        cp = pp->conn_in_get(af, skb, &iph, 0);

-       if (unlikely(sysctl_expire_nodest_conn(ipvs)) && cp && cp->dest &&
-           unlikely(!atomic_read(&cp->dest->weight)) && !iph.fragoffs &&
-           is_new_conn(skb, &iph)) {
+       if (cp && cp->dest && !iph.fragoffs && is_new_conn(skb, &iph) &&
+           ((unlikely(sysctl_expire_nodest_conn(ipvs)) &&
+             unlikely(!atomic_read(&cp->dest->weight))) ||
+            unlikely(is_new_conn_expected(cp)))) {

        This check can be optimized:

        if (!iph.fragoffs && is_new_conn(skb, &iph) && cp &&
            ((unlikely(sysctl_expire_nodest_conn(ipvs)) && cp->dest &&
              unlikely(!atomic_read(&cp->dest->weight))) ||
             unlikely(is_new_conn_expected(cp)))) {

Yup. Okay.

I'll wait on the sysctl thing before posting as all the rest seems aligned. Thanks!


                cp = NULL;


Julian Anastasov <ja@xxxxxx>

To unsubscribe from this list: send the line "unsubscribe lvs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at

<Prev in Thread] Current Thread [Next in Thread>