Re: panic: tcp_input: TCPS_LISTEN in netinet/tcp_input.c:1016

On Mon, 7 Feb 2005, Peter Holm wrote:

> While stress testing GENERIC HEAD from Feb 5 09:19 UTC + mpsafe_vfs = 1
> I got:
> 
> panic(c0832f7d,0,0,1,0) at panic+0x14b
> tcp_input(c27fca00,14,c27fca00,0,0) at tcp_input+0xbf6
> ip_input(c27fca00) at ip_input+0x50d
> netisr_processqueue(c0944cd8) at netisr_processqueue+0x6e
> swi_net(0) at swi_net+0xbe
> ithread_loop(c154d180,cbc90d48,c154d180,c0601f84,0) at ithread_loop+0x120
> fork_exit(c0601f84,c154d180,cbc90d48) at fork_exit+0xa4
> fork_trampoline() at fork_trampoline+0x8
> 
> Details at http://www.holm.cc/stress/log/cons115.html
> A KTR dump with KTR_LOCK|KTR_BUF is available.
> 
> Related reports:
> 
> -rw-r--r--  1 holm  users  10641 Dec 20 16:51 cons96.html
> -rw-r--r--  1 holm  users   9906 Dec 26 15:28 cons98.html
> -rw-r--r--  1 holm  users  15189 Dec 29 22:17 cons99.html

This would appear to be an inter-layer race between the socket code and
the TCP code.  In particular, it looks like a SYN has come in during the
call to listen() on another CPU (or perhaps a preempted thread), after the
TCP state has been set up for the listening tcpcb, but before the
SO_ACCEPTCONN flag is set in the socket state.  The TCP code panics
because it expects that if a tcpcb is in TCPS_LISTEN, the matching socket
should be in SO_ACCEPTCONN.  I'm working on a patch and hope to put it
together for you today.  However, this patch substantially tears up the
current listen code for several protocols, so it will need a fair amount
of testing.

If this problem should occur again, it would be very interesting to know
if ps, show threads, trace, et al, showed either a preempted thread on the
current CPU, or a thread on another CPU, in the listen() system call.

Thanks for the (as usual) excellent bug report!

Robert N M Watson