On Thu, May 28, 2015 at 10:41 PM, Andrey Korolyov [off-list ref] wrote:

Hi,

I am currently playing with SYNPROXY target to optimize SYN filtering
performance and by occasion found that TCP SYN packets containing port
0 can result in a soft lockup when conntrack is enabled just by
itself, given high packet ratio (I`ve reached 450kpps so far with 60b
packets on a /32<->/32 flood with enabled flow control at the media
level and middle-level E3 Xeon on receiver side). Same flood with port

quoted

0 going just well, producing same ceil numbers but without visible

lockups in kernel log. I`ve tested the issue on a broad range of 3.x
kernels and all of them are seemingly affected. Fast and dirty grep
revealed special conditions for port 0 only for protocol-specific
helpers, but there are none of them.

Please find both same captures and traceback below.

Attached trace without GSO, at its presence can be somehow confusing
in a previous sample. The testbed using net.nf_conntrack_max =
2000000, forgot to mention that previously.