netdev - Re: iperf performance regression since Linux 5.18

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 for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 17 Oct 2023 14:08:52 +0200
From: Eric Dumazet <edumazet@...gle.com>
To: Stefan Wahren <wahrenst@....net>
Cc: Neal Cardwell <ncardwell@...gle.com>, Jakub Kicinski <kuba@...nel.org>, 
	Fabio Estevam <festevam@...il.com>, linux-imx@....com, 
	Stefan Wahren <stefan.wahren@...rgebyte.com>, Michael Heimpold <mhei@...mpold.de>, netdev@...r.kernel.org, 
	Yuchung Cheng <ycheng@...gle.com>
Subject: Re: iperf performance regression since Linux 5.18

On Tue, Oct 17, 2023 at 11:53 AM Stefan Wahren <wahrenst@....net> wrote:
>
> Hi Eric,
>
> Am 16.10.23 um 20:47 schrieb Eric Dumazet:
> > On Mon, Oct 16, 2023 at 8:25 PM Stefan Wahren <wahrenst@....net> wrote:
> >> Hi Eric,
> >>
> >> Am 16.10.23 um 12:35 schrieb Eric Dumazet:
> >>> On Mon, Oct 16, 2023 at 11:49 AM Eric Dumazet <edumazet@...gle.com> wrote:
> >>> Speaking of TSQ, it seems an old change (commit 75eefc6c59fd "tcp:
> >>> tsq: add a shortcut in tcp_small_queue_check()")
> >>> has been accidentally removed in 2017 (75c119afe14f "tcp: implement
> >>> rb-tree based retransmit queue")
> >>>
> >>> Could you try this fix:
> >>>
> >>> diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
> >>> index 9c8c42c280b7638f0f4d94d68cd2c73e3c6c2bcc..e61a3a381d51b554ec8440928e22a290712f0b6b
> >>> 100644
> >>> --- a/net/ipv4/tcp_output.c
> >>> +++ b/net/ipv4/tcp_output.c
> >>> @@ -2542,6 +2542,18 @@ static bool tcp_pacing_check(struct sock *sk)
> >>>           return true;
> >>>    }
> >>>
> >>> +static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
> >>> +{
> >>> +       const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
> >>> +
> >>> +       /* No skb in the rtx queue. */
> >>> +       if (!node)
> >>> +               return true;
> >>> +
> >>> +       /* Only one skb in rtx queue. */
> >>> +       return !node->rb_left && !node->rb_right;
> >>> +}
> >>> +
> >>>    /* TCP Small Queues :
> >>>     * Control number of packets in qdisc/devices to two packets / or ~1 ms.
> >>>     * (These limits are doubled for retransmits)
> >>> @@ -2579,12 +2591,12 @@ static bool tcp_small_queue_check(struct sock
> >>> *sk, const struct sk_buff *skb,
> >>>                   limit += extra_bytes;
> >>>           }
> >>>           if (refcount_read(&sk->sk_wmem_alloc) > limit) {
> >>> -               /* Always send skb if rtx queue is empty.
> >>> +               /* Always send skb if rtx queue is empty or has one skb.
> >>>                    * No need to wait for TX completion to call us back,
> >>>                    * after softirq/tasklet schedule.
> >>>                    * This helps when TX completions are delayed too much.
> >>>                    */
> >>> -               if (tcp_rtx_queue_empty(sk))
> >>> +               if (tcp_rtx_queue_empty_or_single_skb(sk))
> >>>                           return false;
> >>>
> >>>                   set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
> >> This patch applied on top of Linux 6.1.49, TSO on, gso_max_size 65535,
> >> CONFIG_HZ_100=y
> >>
> >> root@...ragon:/boot# iperf -t 10 -i 1 -c 192.168.1.129
> >> ------------------------------------------------------------
> >> Client connecting to 192.168.1.129, TCP port 5001
> >> TCP window size:  192 KByte (default)
> >> ------------------------------------------------------------
> >> [  3] local 192.168.1.12 port 59714 connected with 192.168.1.129 port 5001
> >> [ ID] Interval       Transfer     Bandwidth
> >> [  3]  0.0- 1.0 sec  11.5 MBytes  96.5 Mbits/sec
> >> [  3]  1.0- 2.0 sec  11.4 MBytes  95.4 Mbits/sec
> >> [  3]  2.0- 3.0 sec  11.1 MBytes  93.3 Mbits/sec
> >> [  3]  3.0- 4.0 sec  11.2 MBytes  94.4 Mbits/sec
> >> [  3]  4.0- 5.0 sec  11.1 MBytes  93.3 Mbits/sec
> >> [  3]  5.0- 6.0 sec  11.2 MBytes  94.4 Mbits/sec
> >> [  3]  6.0- 7.0 sec  11.2 MBytes  94.4 Mbits/sec
> >> [  3]  7.0- 8.0 sec  11.1 MBytes  93.3 Mbits/sec
> >> [  3]  8.0- 9.0 sec  11.4 MBytes  95.4 Mbits/sec
> >> [  3]  9.0-10.0 sec  11.2 MBytes  94.4 Mbits/sec
> >> [  3]  0.0-10.0 sec   113 MBytes  94.4 Mbits/sec
> >>
> >> The figures are comparable to disabling TSO -> Good
> >>
> >> Thanks
> > Great. I suspect a very slow TX completion from fec then.
> >
> > Could you use the following bpftrace program while your iperf is running ?
> unfortuntely there is no bpftrace and most of its dependencies on my
> platform. I looked at some guides and it seems to have a lot of (build)
> dependencies. On a PC this won't be a problem, but on my ARM platform
> there is only 1 GB eMMC space left.
>
> Before investing a lot of time to get bpftrace running, is there an
> alternative solution?

No worries Stefan, I think we do not have to get precise numbers, I
will try to provide a debug patch (since you are able to build custom
kernels)

In the meantime, I will submit the official TCP patch, as you said it
was helping a lot.

Just to confirm, have you tried the patch on top of the latest net tree ?

Thanks.

> >
> > .bpftrace -e '
> > k:__dev_queue_xmit {
> >   $skb = (struct sk_buff *)arg0;
> >   if ($skb->fclone == 2) {
> >    @start[$skb] = nsecs;
> >   }
> > }
> > k:__kfree_skb {
> >   $skb = (struct sk_buff *)arg0;
> >   if ($skb->fclone == 2 && @start[$skb]) {
> >    @tx_compl_usecs = hist((nsecs - @start[$skb])/1000);
> >    delete(@start[$skb]);
> >   }
> > }
> > END { clear(@start); }'
>