Re: [RFC] net: sch_generic: fq_codel vs pfifo_fast - Marc Kleine-Budde

From: Marc Kleine-Budde <mkl@pengutronix.de>
To: Stephen Hemminger <stephen@networkplumber.org>
Cc: Jiri Pirko <jiri@resnulli.us>,
	netdev@vger.kernel.org, Dave Taht <dave.taht@gmail.com>,
	Jamal Hadi Salim <jhs@mojatatu.com>,
	kernel@pengutronix.de, Cong Wang <xiyou.wangcong@gmail.com>,
	linux-can@vger.kernel.org, davem@davemloft.net
Subject: Re: [RFC] net: sch_generic: fq_codel vs pfifo_fast
Date: Wed, 27 Mar 2019 20:24:13 +0100	[thread overview]
Message-ID: <0edf15df-b607-70f7-5fe8-2860eb663a53@pengutronix.de> (raw)
In-Reply-To: <20190327113059.6b1224ab@shemminger-XPS-13-9360>

[-- Attachment #1.1: Type: text/plain, Size: 1772 bytes --]

On 3/27/19 7:30 PM, Stephen Hemminger wrote:
>> on CAN networking hardware we (the CAN community) experience a lot silent,
>> unwanted frame drops inside the kernel. (See first patch for details.) So
>> here's a patch series to keep pfifo_fast as default scheduler for CAN hardware
>> by default.
> 
> Why do you set fq_codel as default qdisc if you know it doesn't work right for your
> environment. Is this a distro's want one value problem?

This is a many fold problem.

- Consider a random Linux developer attaching one of the mainline
supported USB-CAN adapters to his/her development laptop and doing some
random CAN test. As far a I heard all modern Linux distributions (but
not debian) enable fq_codel per kernel .config or via systemd. The user
will experiences dropped CAN frames on the first 1000 packages. This is
not a good user experience.

- From the user space point of view the sysctl net.core.default_qdisc
behaves a bit strange. Why can I enable a queuing discipline by default
that's known not to work on CAN devices. You might think the kernel
(should) know that CAN devices best work with pfifo_fast (or not
NET_SCHED at all).

- Addressing your question directly: Mixed environment. fq_codel works
great on Ethernet, but terrible on CAN. Of course I an use "tc" to
configure the way I want to. But from my point of view it would be nice,
if the kernel uses sane default...and handles the net.core.default_qdisc
default in a sane way.

Marc

-- 
Pengutronix e.K.                  | Marc Kleine-Budde           |
Industrial Linux Solutions        | Phone: +49-231-2826-924     |
Vertretung West/Dortmund          | Fax:   +49-5121-206917-5555 |
Amtsgericht Hildesheim, HRA 2686  | http://www.pengutronix.de   |

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]