On Thu, Oct 29, 2015 at 6:21 PM, Neal P. Murphy
[off-list ref] wrote:

On Thu, 29 Oct 2015 17:01:24 -0700
Ani Sinha [off-list ref] wrote:

quoted

On Wed, Oct 28, 2015 at 11:40 PM, Neal P. Murphy
[off-list ref] wrote:

quoted

On Wed, 28 Oct 2015 02:36:50 -0400
"Neal P. Murphy" [off-list ref] wrote:

quoted

On Mon, 26 Oct 2015 21:06:33 +0100
Pablo Neira Ayuso [off-list ref] wrote:

quoted

Hi,

On Mon, Oct 26, 2015 at 11:55:39AM -0700, Ani Sinha wrote:

quoted

netfilter: nf_conntrack: fix RCU race in nf_conntrack_find_get

Please, no need to Cc everyone here. Please, submit your Netfilter
patches to netfilter-devel@vger.kernel.org.

Moreover, it would be great if the subject includes something
descriptive on what you need, for this I'd suggest:

[PATCH -stable 3.4,backport] netfilter: nf_conntrack: fix RCU race in nf_conntrack_find_get

I'm including Neal P. Murphy, he said he would help testing these
backports, getting a Tested-by: tag usually speeds up things too.

I've probably done about as much seat-of-the-pants testing as I can. All opening/closing the same destination IP/port.

Host: Debian Jessie, 8-core Vishera 8350 at 4.4 GHz, 16GiB RAM at (I think) 2100MHz.

Traffic generator 1: 6-CPU KVM running 64-bit Smoothwall Express 3.1 (linux 3.4.109 without these patches), with 8GiB RAM and 9GiB swap. Packets sent across PURPLE (to bypass NAT and firewall).

Traffic generator 2: 32-bit KVM running Smoothwall Express 3.1 (linux 3.4.110 with these patches), 3GiB RAM and minimal swap.

In the first set of tests, generator 1's traffic passed through Generator 2 as a NATting firewall, to the host's web server. In the second set of tests, generator 2's traffic went through NAT to the host's web server.

The load tests:
  - 2500 processes using 2500 addresses and random src ports
  - 2500 processes using 2500 addresses and the same src port
  - 2500 processes using the same src address and port

I also tested using stock NF timeouts and using 1 second timeouts.

Bandwidth used got as high as 16Mb/s for some tests.

Conntracks got up to 200 000 or so or bounced between 1 and 2, depending on the test and the timeouts.

I did not reproduce the problem these patches solve. But more importantly, I saw no problems at all. Each time I terminated a test, RAM usage returned to about that of post-boot; so there were no apparent memory leaks. No kernel messages and no netfilter messages appeared during the tests.

If I have time, I suppose I could run another set of tests: 2500 source processes using 2500 addresses times 200 ports to connect to 2500 addresses times 200 ports on a destination system. Each process opens 200 sockets, then closes them. And repeats ad infinitum. But I might have to be clever since I can't run 500 000 processes; but I could run 20 VMs; that would get it down to about 12 000 processes per VM. And I might have to figure out how to allow allow processes on the destination system to open hundreds or thousands of sockets.

Should I resend the patch with a Tested-by: tag?

... Oh, wait. Not yet. The dawn just broke over ol' Marblehead here. I only tested TCP; I need to hammer UDP, too.

Can I set the timeouts to zero? Or is one as low as I can go?

I don't see any assertion or check against 0 sec timeouts. You can
try. Your conntrack entries will be constantly flushing.

N