Re: regression: sysctl_check changes in 2.6.24 are O(n) resulting in slow creation of 10000 network interfaces

[ebiederm@xmission.com (Eric W. Biederman)]

Benjamin LaHaise <bcrl@kvack.org> writes:

> Hello folks,
>
> 2.6.24-rc6 regresses on the 10000 network interface creation test relative to 
> 2.6.23.  The cause appears to be the new code in sysctl_check_lookup(), which 
> shows up as the #1 item while profiling.  Is a revert of this new code 
> possible until its scaling issues are fixed?  2.6.23 can do more than 100 new 
> network interfaces per second for the first few thousand devices, but with 
> 2.6.24-rc6 the results drop off rather dramatically to less than 10 interfaces
> per second.  The 10000 interface test is unbearable with the new sysctl_check 
> code.

Why do we need 10000 interfaces?  Why isn't network device creation a
slow path?

The problem seems to be in the data structures used by sysctl.
You are increasing the length of the linked list each time you
add a network interface.  So sysctl lookups slow down.
At 10000 entries that is a long linked list walk.

At 100000 things get even longer.  Now the numbers you report still
seem like a lot of time to me.  My guess would be that we are getting
badly into cache miss territory.

If what you describe is a real scenario where users care we need
to fix the sysctl data structures so that they scale.

Because of this bug report and another one I got earlier today
about a real bug in the parallel port code detected by the very
lookup that is slowing you down.  I am quite reluctant to contemplate
pulling this code.  It seems to be doing it's job, if in some cases
uncomfortably so.

So is this a bug report telling me that there are users with
10k or 100k interfaces that care.  So we need to fix sysctl.

Is there a specific kernel test case that is run often that having
slow sysctl performance matters for?  CONFIG_SYSCTL=n should solve
that if it is specialized.

Eric