netdev - Re: [PATCH net] tcp: fix potential xmit stalls caused by TCP_NOTSENT

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] [day] [month] [year] [list]

Date:   Mon, 25 Apr 2022 08:05:39 -0700
From:   Eric Dumazet <edumazet@...gle.com>
To:     Dave Taht <dave.taht@...il.com>
Cc:     Eric Dumazet <eric.dumazet@...il.com>,
        "David S . Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        netdev <netdev@...r.kernel.org>, Doug Porter <dsp@...com>,
        Soheil Hassas Yeganeh <soheil@...gle.com>,
        Neal Cardwell <ncardwell@...gle.com>
Subject: Re: [PATCH net] tcp: fix potential xmit stalls caused by TCP_NOTSENT_LOWAT

On Mon, Apr 25, 2022 at 6:16 AM Dave Taht <dave.taht@...il.com> wrote:
>

>
> Thx! We have been having very good results with TCP_NOTSENT_LOWAT set
> to 32k or less behind an apache traffic server... and had some really
> puzzling ones at geosync RTTs. Now I gotta go retest.
>
> Side question: Is there a guide/set of recommendations to setting this
> value more appropriately, under what circumstances? Could it
> autoconfigure?

It is a tradeoff between memory usage in the kernel (storing data in
the socket transmit queue),
and the number of times the application is woken up to let it push
another chunk of data.

32k really means : No more than one skb at a time.  (An skb can
usually store about 64KB of payload)

At Google we have been using 2MB, but I suspect this was to work
around the bug I just fixed.
We probably could use a smaller value like 1MB, leading to one
EPOLLOUT for 1/2 MB.

Precise values also depend on how much work is needed in
tcp_sendmsg(), zerocopy can play a role here.

autoconfigure ? Not sure how. Once you have provisioned a server for a
given workload,
you are all set. No need to tune the value as a function of the load.

>
> net.ipv4.tcp_notsent_lowat = 32768

This probably has caused many stalls on long rtt / lossy links...

>
> --
> FQ World Domination pending: https://blog.cerowrt.org/post/state_of_fq_codel/
> Dave Täht CEO, TekLibre, LLC