netdev - Re: "TCP: eth0: Driver has suspect GRO implementation, TCP performance may be compromised." message with "ethtool -K eth0 gro off"

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <1486128246.21871.83.camel@edumazet-glaptop3.roam.corp.google.com>
Date:   Fri, 03 Feb 2017 05:24:06 -0800
From:   Eric Dumazet <eric.dumazet@...il.com>
To:     Marcelo Ricardo Leitner <marcelo.leitner@...il.com>
Cc:     Markus Trippelsdorf <markus@...ppelsdorf.de>,
        netdev@...r.kernel.org
Subject: Re: "TCP: eth0: Driver has suspect GRO implementation, TCP
 performance may be compromised." message with "ethtool -K eth0 gro off"

On Fri, 2017-02-03 at 09:54 -0200, Marcelo Ricardo Leitner wrote:
> On Thu, Feb 02, 2017 at 05:59:24AM -0800, Eric Dumazet wrote:
> > On Thu, 2017-02-02 at 05:31 -0800, Eric Dumazet wrote:
> > 
> > > Anyway, I suspect the test is simply buggy ;)
> > > 
> > > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> > > index 41dcbd568cbe2403f2a9e659669afe462a42e228..5394a39fcce964a7fe7075b1531a8a1e05550a54 100644
> > > --- a/net/ipv4/tcp_input.c
> > > +++ b/net/ipv4/tcp_input.c
> > > @@ -164,7 +164,7 @@ static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
> > >  	if (len >= icsk->icsk_ack.rcv_mss) {
> > >  		icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
> > >  					       tcp_sk(sk)->advmss);
> > > -		if (unlikely(icsk->icsk_ack.rcv_mss != len))
> > > +		if (unlikely(icsk->icsk_ack.rcv_mss != len && skb_is_gso(skb)))
> > >  			tcp_gro_dev_warn(sk, skb);
> > >  	} else {
> > >  		/* Otherwise, we make more careful check taking into account,
> > 
> > This wont really help.
> > 
> > Our tcp_sk(sk)->advmss can be lower than the MSS used by the remote
> > peer.
> > 
> > ip ro add .... advmss 512
> 
> I don't follow. With a good driver, how can advmss be smaller than the
> MSS used by the remote peer? Even with the route entry above, I get
> segments just up to advmss, and no warning.
> 

A TCP flow has two ends.

Common MTU = 1500

One can have advmss 500, the other one no advmss (or the standard 1460
one)

So if we compare apple and orange, result might be shocking ;)

If you want to reproduce this use the "ip ro add .... advmss 512" hint,
and/or play with sysctl_tcp_mtu_probing