[RFC] IPV4 long lasting timer function

[RFC] IPV4 long lasting timer function

[Thomas Gleixner]

Hi,

while tracking down some timer related ugliness I stumbled over the
timer driven function rt_secret_rebuild(), which does a loop over
rt_has_mask (1024 in my case) entries and possibly some subsequent
variable sized loops inside each step.

On a 300MHZ PPC system this accumulated to a worst case total of >5ms. I
could not reproduce it with this magnitude, but applying heavy
networking load is definitely triggering this behaviour.

Shouldn't this be converted to a workqueue, which gets triggered by a
timer instead of blocking the timer softirq and therefor the delivery of
other timer functions that long ?

tglx

Re: [RFC] IPV4 long lasting timer function

[David S. Miller]

From: Thomas Gleixner <tglx@linutronix.de>
Date: Wed, 17 Aug 2005 22:59:01 +0200

> Shouldn't this be converted to a workqueue, which gets triggered by a
> timer instead of blocking the timer softirq and therefor the delivery of
> other timer functions that long ?

We could, and I'd be happy to apply such a patch if someone
writes it up.

Re: [RFC] IPV4 long lasting timer function

[Lee Revell]

On Wed, 2005-08-17 at 22:59 +0200, Thomas Gleixner wrote:
> Hi,
> 
> while tracking down some timer related ugliness I stumbled over the
> timer driven function rt_secret_rebuild(), which does a loop over
> rt_has_mask (1024 in my case) entries and possibly some subsequent
> variable sized loops inside each step.
> 
> On a 300MHZ PPC system this accumulated to a worst case total of >5ms. I
> could not reproduce it with this magnitude, but applying heavy
> networking load is definitely triggering this behaviour.
> 
> Shouldn't this be converted to a workqueue, which gets triggered by a
> timer instead of blocking the timer softirq and therefor the delivery of
> other timer functions that long ?

Wow, blast from the past!  This was one of the very first problematic
code paths Ingo and I identified in the early days of the voluntary
preempt patch (using the crude ALSA xrun debug mechanism which was the
best we had before /proc/latency_trace).

IIRC I was able to trivially reproduce it by leaving gtk-gnutella
running overnight with a few active searches/downloads.

Lee