udevd high cpu hogging in kernel-2.6.34 git8 to git-13

udevd high cpu hogging in kernel-2.6.34 git8 to git-13

[JD]

I noticed that since git8, udevd consumes from 50 to 80 percent of cpu.
The main udevd thread is niced to -4, and 6 or more child threads are 
niced -2.

So, I backpedaled to git7, which does not have this "feature" :)

Cheers,

JD

Re: udevd high cpu hogging in kernel-2.6.34 git8 to git-13

[Jiri Slaby]

On 05/28/2010 07:36 AM, JD wrote:
> I noticed that since git8, udevd consumes from 50 to 80 percent of cpu.
> The main udevd thread is niced to -4, and 6 or more child threads are
> niced -2.
> 
> So, I backpedaled to git7, which does not have this "feature" :)

Is it this one again?
http://lkml.org/lkml/2010/5/22/150

Could you try to revert the patch?

-- 
js

Re: udevd high cpu hogging in kernel-2.6.34 git8 to git-13

[Jiri Slaby]

Added lkml back to CC.

On 05/28/2010 08:06 AM, JD wrote:
> But here goes:
> I have seen this behaviour in all the git releases from git8
> to the current 2.6.34-git13
> 
> I have some of the analysis based on strace, and kernel stack
> trace.
> Instead of backing out the whole code for the new udevd, perhaps
> it (udevd) should sleep for some time before polling the same fd
> again.

So if you see in the strace output the poll to return immediately with
out fd being an anon_inode (check /proc/pid_of_udev/fd/), it is the
issue. AFAIK, it was not fixed upstream yet.

So could you try to revert a7cf4145b?

-- 
js

Re: udevd high cpu hogging in kernel-2.6.34 git8 to git-13

[JD]

On 05/27/2010 11:10 PM, Jiri Slaby wrote:
> Added lkml back to CC.
>
> On 05/28/2010 08:06 AM, JD wrote:
>    
>> But here goes:
>> I have seen this behaviour in all the git releases from git8
>> to the current 2.6.34-git13
>>
>> I have some of the analysis based on strace, and kernel stack
>> trace.
>> Instead of backing out the whole code for the new udevd, perhaps
>> it (udevd) should sleep for some time before polling the same fd
>> again.
>>      
> So if you see in the strace output the poll to return immediately with
> out fd being an anon_inode (check /proc/pid_of_udev/fd/), it is the
> issue. AFAIK, it was not fixed upstream yet.
>
> So could you try to revert a7cf4145b?
>
>    
As promised, I tested the patch submitted by Eric Paris
(See http://marc.info/?l=linux-kernel&m=127479018618584&w=2).
I patched kernel-2.6.34-git13 with it, without reverting a7cf4145b.
The thing is it works and works extremely well.
The behaviour is back to what it used to be: Only 3 udevd
threads, as in before git8, and they do not consume any
cpu bandwidth that I can see. It is well below the threshold
of 1% .

Also, it seems that it has been picked up by upstream. Compiled
and booted 2.6.34-git16 and all is well.