problem with noop and pfifo_fast Qdisc switching

Hi,

I have a situation where one of important message (phonet pipe message) gets dropped because of 'noop' Qdisc.
This situation mostly happens when coming out of sleep and there is a message to send.
What I have observed is this.
In normal case the qdisc used id pfifo_fast. When CPU goes to sleep, it changes the qdisc to noop.
When CPU wakes up (because of a message from lower layers), phonet tries to send a control packet to driver and finds qdisc is still 'noop'. 
Because the concerned message is a control message, this packet drop is creating deadlock situation.

How the switch between pfifo_fast and noop is handled? Is this something solved in 3.0?
We are running a older version of kernel (2.6.35 based).

Thanks,
Hemant
 

Le jeudi 19 janvier 2012 à 13:47 +0100, Hemant-vilas RAMDASI a écrit :

Hi,

I have a situation where one of important message (phonet pipe message) gets dropped because of 'noop' Qdisc.
This situation mostly happens when coming out of sleep and there is a message to send.
What I have observed is this.
In normal case the qdisc used id pfifo_fast. When CPU goes to sleep, it changes the qdisc to noop.
When CPU wakes up (because of a message from lower layers), phonet tries to send a control packet to driver and finds qdisc is still 'noop'. 
Because the concerned message is a control message, this packet drop is creating deadlock situation.

How the switch between pfifo_fast and noop is handled? Is this something solved in 3.0?
We are running a older version of kernel (2.6.35 based).

If message is important, you should check return code from xmit
(NET_XMIT_CN is returned by noop_enqueue), and maybe resend it later.

Not sure by what you mean by 'cpu goes to sleep' or 'cpu wakes up' ?

Is it the machine going to suspend mode ?

Hi,

-----Original Message-----
From: Eric Dumazet [mailto:eric.dumazet@gmail.com]
Sent: Thursday, January 19, 2012 6:46 PM
To: Hemant-vilas RAMDASI
Cc: netdev@vger.kernel.org; netdev-owner@vger.kernel.org
Subject: Re: problem with noop and pfifo_fast Qdisc switching

Le jeudi 19 janvier 2012 à 13:47 +0100, Hemant-vilas RAMDASI a écrit :

quoted

Hi,

I have a situation where one of important message (phonet pipe

message) gets dropped because of 'noop' Qdisc.

quoted

This situation mostly happens when coming out of sleep and there is a

message to send.

quoted

What I have observed is this.
In normal case the qdisc used id pfifo_fast. When CPU goes to sleep,

it changes the qdisc to noop.

quoted

When CPU wakes up (because of a message from lower layers), phonet

tries to send a control packet to driver and finds qdisc is still
'noop'.

quoted

Because the concerned message is a control message, this packet drop

is creating deadlock situation.

quoted

How the switch between pfifo_fast and noop is handled? Is this

something solved in 3.0?

quoted

We are running a older version of kernel (2.6.35 based).

If message is important, you should check return code from xmit
(NET_XMIT_CN is returned by noop_enqueue), and maybe resend it later.

Not sure by what you mean by 'cpu goes to sleep' or 'cpu wakes up' ?

Is it the machine going to suspend mode ?

Yes, I mean machine going to suspend..
The problem with this message is that it can be retried only a couple of time.
We saw a code where dev_deactivate is called which changes the Qdisc to 'noop'. 
I hope it will restore it in dev_activate() again when machine resumes. We are checking this.
Is there any race between when dev_queue_xmit is called and dev_activate is still not complete?

Thanks,
Hemant

Le jeudi 19 janvier 2012 à 14:44 +0100, Hemant-vilas RAMDASI a écrit :

Yes, I mean machine going to suspend..
The problem with this message is that it can be retried only a couple of time.
We saw a code where dev_deactivate is called which changes the Qdisc to 'noop'. 
I hope it will restore it in dev_activate() again when machine resumes. We are checking this.
Is there any race between when dev_queue_xmit is called and dev_activate is still not complete?

No race, just dev_queue_xmit() returns an error ;)

There is no 'queue' at this level.

IP has sort of a (small) queue because of ARP resolution.