Re: [PATCH net-next] tcp: add tcp_add_backlog()

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

On Fri, 2016-09-23 at 11:09 -0300, Marcelo Ricardo Leitner wrote:
> On Fri, Sep 23, 2016 at 06:42:51AM -0700, Eric Dumazet wrote:
> > On Fri, 2016-09-23 at 09:45 -0300, Marcelo Ricardo Leitner wrote:
> > 
> > > Aye. In that case, what about using tail instead of end?
> > 
> > 
> > What do you mean exactly ?
> 
> Something like:
> -skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
> +skb->truesize = SKB_TRUESIZE(skb_tail_offset(skb));

Certainly not ;)

This would be lying.
We really want a precise memory accounting to avoid OOM.

Some USB drivers use 8KB for their skb->head, you do not want to pretend
its 66+NET_SKB_PAF=F bytes just because there is no payload in the
packet.

> 
> And define skb_tail_offset() to something similar skb_end_offset(), so
> that it would account only for the data and not unused space in the
> buffer.
> 
> > 
> > >  Because
> > > accounting for something that we have to tweak the limits to accept is
> > > like adding a constant to both sides of the equation.
> > > But perhaps that would cut out too much of the fat which could be used
> > > later by the stack.
> > 
> > Are you facing a particular problem with current code ?
> > 
> 
> For TCP, no, just wondering. :-)
> 
> I'm having similar issues with SCTP: if the socket gets backlogged, the
> buffer accounting gets pretty messy. SCTP calculates the rwnd to be just
> rcvbuf/2, but this ratio is often different in backlog and it causes the
> advertized window to be too big, resulting in packet drops in the
> backlog.
> 
> SCTP has some way to identify and compensate this "extra" rwnd, via
> rwnd_press, and will shrink it if it detects that the window is bigger
> than the buffer available. But as the socket is backlogged, it's doesn't
> kick in soon enough to prevent such drops.
> 
> It's not just a matter of adjusting the overhead ratio (rcvbuf/2)
> because with SCTP the packets may have different sizes, so a packet with
> a chunk of 100 bytes will have a ratio and another with 1000 bytes will
> have another, within the same association.
> 
> So I'm leaning towards on updating truesize before adding to the
> backlog, but to account just for the actual packet, regardless of the
> buffer that was used for it. It still has the overhead ratio issue with
> different packet sizes, though, but smaller.
> 
> Note that SCTP doesn't have buffer auto-tuning yet.

Also for TCP, we might need to use sk->sk_wmem_queued instead of
sk->sk_sndbuf

This is because SACK processing can suddenly split skbs in 1-MSS pieces.

Problem is that for very large BDP, we can end up with thousands of skb
in backlog. So I am also considering to try to coalesce stupid ACK sent
by non GRO receivers or simply the verbose SACK blocks...

eg if backlog is under pressure and its tail is : 

ACK 1   <sack 4000:5000>

and the incoming packet is :

ACK 1  <sack 4000:6000>

Then we could replace the tail by the incoming packet with minimal
impact.

Since we might receive hundred of 'sequential' SACK blocks, this would
help to reduce time taken by the application to process the (now
smaller) backlog