Re: TCPBacklogDrops during aggressive bursts of traffic

[Eric Dumazet <eric.dumazet@gmail.com>]

On Tue, 2012-05-15 at 17:47 +0100, Ben Hutchings wrote:
> On Tue, 2012-05-15 at 18:34 +0200, Eric Dumazet wrote:
> > On Tue, 2012-05-15 at 17:29 +0100, Kieran Mansley wrote:
> > > On Tue, 2012-05-15 at 16:56 +0200, Eric Dumazet wrote:
> > > > 
> > > > Please try latest kernels, this is probably 'fixed'
> > > 
> > > I've just tried with 3.4.0-rc7 and the problem is still reproducible.
> > > It's perhaps harder to reproduce than on 3.3.6 but still there.
> > > 
> > > > What network driver are you using ? 
> > > 
> > > The receiver is using the sfc driver that is included in the kernel
> > > build, together with an SFC 9020 NIC. 
> > > 
> > > Kieran
> > > 
> > 
> > MTU ?
> 
> 1500
> 
> > What is typical skb->truesize of skb given to stack in RX path ?
> >
> > If drivers use PAGE_SIZE fragments, then you are more likely to hit
> > limit.
> 
> We're passing page fragments into GRO.


Yes, I can see drivers/net/ethernet/sfc/rx.c is even lying about
truesize. Thats explain why you trigger the backlogdrop even on 2.6
kernels.

skb->len = rx_buf->len;
skb->data_len = rx_buf->len;
skb->truesize += rx_buf->len; // instead of real frag size

So skb->truesize are rather small.

napi_get_frags() could probably updated in net-next to use the first
frag as skb->head to save 512 bytes per skb.

You could try setting tcp_adv_win_scale to -2