Re: debugging kernel during packet drops

[James King <t.james.king@gmail.com>]

On Mon, Mar 22, 2010 at 3:41 AM, Jorrit Kronjee <j.kronjee@infopact.nl> wrote:
> At around 300 kpps, the amount of packet drops is 40 kpps. For me, this
> amount is too significant to ignore. I see the load average go from a
> comfortable 0.00 to 1.78, mainly caused by ksoftirqd processes. At 200
> kpps, the average amount of packet drops is 23 kpps. At 100 kpps, it's
> still 2 kpps.
>
> When I disable the hashlimit module the packet drops disappear again.
> Now I know that hashlimit is made for more than one thing, namely
> limiting packets based on source/destination host and source/destination
> port, so it's not as efficient as it could be for my purposes. I could
> rewrite it, but before I do that, I would like to know if the module
> itself is really what's causing it, or if there's some underlying cause
> that I'm not seeing. So my question in short: how can I discover why
> it's dropping packets?

I'm not sure whether it's even related to the problem you're having,
but I had a similar problem on a bnx2 interface with high packet rates
when using l7-filter.  ifconfig reported huge numbers of dropped
packets, corresponding to rx_fw_discards from "ethtool -S ethX"
output.  I resolved this by bumping up the driver RX ring size (which
was defaulting to 100 out of a maximum possible size of 1020).