Re: [PATCH] tcp: fix syncookie regression

[Neal Cardwell <ncardwell@xxxxxxxxxx>]

On Tue, Mar 13, 2012 at 3:26 AM, Simon Kirby <sim@xxxxxxxxxxxxx> wrote:
>
> While deploying this on top of 3.2.9, we hit what seems to be the bug
> fixed by 4648dc97af9d496218a05353b0e442b3dfa6aaab in 3.3. I see 3.2.9 has
> daef52bab1fd26e24e8e9578f8fb33ba1d0cb412, so maybe this is exposed in 3.2
> now?
...
> net/ipv4/tcp_input.c:
>     3438 #if FASTRETRANS_DEBUG > 0
> ---> 3439         WARN_ON((int)tp->sacked_out < 0);
>     3440         WARN_ON((int)tp->lost_out < 0);
>     3441         WARN_ON((int)tp->retrans_out < 0);
>     3442         if (!tp->packets_out && tcp_is_sack(tp)) {
> ...
>     3057         /* D. Check consistency of the current state. */
> ---> 3058         tcp_verify_left_out(tp);

Yes, exactly. 3.2.9 and 3.2.10 have
daef52bab1fd26e24e8e9578f8fb33ba1d0cb412 but not the
4648dc97af9d496218a05353b0e442b3dfa6aaab that fixes the resulting
issue with sacked_out going negative, and these lines you've flagged
are exactly the symptoms we'd expect because of that. The fix is
queued up for the -stable series already, so it should be in 3.2.11, I
presume. The fix is already in 3.3-rc7.

> Oops two seconds later, scrolled off console, didn't get written to disk
> or remote syslog server, all we have is syslog-broadcasted Oops and Code
> lines due to broken printk priorities:

As you can imagine, the oops is probably connected to sacked_out going
negative; others have reported an oops along with sacked_out going
negative. The fix should take care of that.

neal
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html