Re: [PATCH net 2/2] Revert "net: advertise TCP MSS from the configured MTU, not the learned PMTU"

[David Laight <david.laight.linux@gmail.com>]

On Thu, 29 Oct 2026 18:21:22 +0800
Esko Mobius <echoowall@gmail.com> wrote:

> On Thu, 29 Oct 2026, David Laight wrote:
> > If it isn't a silly question, how to connections in the other direction
> > work?
> >
> > I'd have thought they'd need that ICMP message to get through.  
> 
> Not silly at all. The frag-needed always goes back to whoever sent the
> oversized packet, and only the client side filters inbound ICMP.
> 
> Server -> client (downloads): the server's full-size segment hits the
> 1456 hop, the router there sends frag-needed back to the server, which
> accepts it, learns the PMTU and resends smaller segments. The server's
> send side was never changed by 2640e6419594 (tcp_current_mss() still
> uses the learned PMTU), so downloads always worked.
> 
> Client -> server (uploads): the client's full-size segment hits the same
> hop, the frag-needed goes back towards the client (via the relay's
> conntrack), and the client-side firewall drops it. The client never
> learns to send smaller segments, so the upload stalls. Here the only
> thing that kept segments small was the smaller MSS the server
> advertised in its SYN-ACK, which is what the reverted commit removed.

But that relies on something having gone the other way first.
In general that won't be true.

It is almost worth implement something that retries with the shorter mtu
to detect the PPPoE mtu blackhole at the sending end.
Or even just use the shorter mtu for long RTT connections.
It won't make much difference to throughput (maybe about 0.3%).

David

> 
> Thanks,
> Yongjian
>