netdev - [RFC PATCH 0/3] tcp: allow user and drivers to tweak TSQ logic through sysctl knobs

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-next>] [day] [month] [year] [list]

Message-Id: <20180105113256.14835-1-natale.patriciello@gmail.com>
Date:   Fri,  5 Jan 2018 12:32:53 +0100
From:   Natale Patriciello <natale.patriciello@...il.com>
To:     netdev@...r.kernel.org
Cc:     Natale Patriciello <natale.patriciello@...il.com>,
        Eric Dumazet <edumazet@...gle.com>,
        Toke Høiland-Jørgensen <toke@...e.dk>,
        Carlo Augusto Grazia <carloaugusto.grazia@...more.it>
Subject: [RFC PATCH 0/3] tcp: allow user and drivers to tweak TSQ logic through sysctl knobs

With the commit 3a9b76fd0db9 ("tcp: allow drivers to tweak TSQ logic") a
network device driver can tune the number of bytes that TCP can queue in
the qdisc/devices. The tuning is done through changing the log scale
used to derive the TSQ credit from the current pacing rate (given in
B/s). By default, it is 10 (equivalent to 1/1024 of a second), but
drivers can select smaller values (from 1/512 to 1/64 of a second) that
reflects in a higher number of bytes that can be enqueued.

The approach has three inherent problems:
0. The values that can be chosen are on a logarithmic scale;
1. It assumes that drivers maintainer will test different values and
   select the best one, putting it in the stone of the driver source code;
2. It prevents end users unable to recompile the kernel from
   experimenting other values.

The users that rely on out-of-tree drivers should have a good knowledge
of the WiFi ecosystem, and be able to patch their driver, to benefit
from the approach: not everyone is a C expert.

This series reverts the above approach and adds sysctl knobs for the
tuning. Firstly, we added the option to disable the TSQ with the magic
value -1 on tcp_limit_output_bytes. Then, we added two sysctl knobs to
allow tuning the number of packets or the ms value that can be enqueued.
So, the user (or the drivers, assuming they can modify sysctl values)
can easily use other values without recompiling the kernel.

A driver that right now has chosen to use a value of 64 for the
shifting, with this patch should put in the sysctl_tcp_limit_output_ms
variable a value of 16 (which corresponds to ~16 ms of bytes that can be
enqueued) to obtain the same effect. The advantage of our series is that
the driver developer (or the end user) can also test (without
recompiling the kernel) values from 9 to 15 ms, which with the current
kernel approach cannot be tested.

Ref: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1670041
Cc: Eric Dumazet <edumazet@...gle.com>
Cc: Toke Høiland-Jørgensen <toke@...e.dk>
Cc: Carlo Augusto Grazia <carloaugusto.grazia@...more.it>

Natale Patriciello (3):
  Revert "tcp: allow drivers to tweak TSQ logic"
  tcp: Negative values of sysctl_tcp_limit_output_bytes disable TSQ
  tcp: Add tunable parameters for TSQ

 Documentation/networking/ip-sysctl.txt | 22 ++++++++++++++++++++++
 include/net/netns/ipv4.h               |  2 ++
 include/net/sock.h                     |  2 --
 net/core/sock.c                        |  1 -
 net/ipv4/sysctl_net_ipv4.c             | 14 ++++++++++++++
 net/ipv4/tcp_output.c                  | 10 ++++++++--
 6 files changed, 46 insertions(+), 5 deletions(-)

-- 
2.15.1