Re: sysctls

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

Andrew Morton <akpm@linux-foundation.org> writes:

> I note that the networking tree is adding new sysctls:
>
> <<<<<<< HEAD/include/linux/sysctl.h
>         NET_IPV6_ACCEPT_SOURCE_ROUTE=25,
> =======
>         NET_IPV6_OPTIMISTIC_DAD=24,
>         NET_IPV6_ACCEPT_SOURCE_ROUTE=25,
>>>>>>>> /include/linux/sysctl.h
>
> (Well, it's trying to - there are some git rejects in net-2.6.22)
>
> But we kind-of decided a while back to stop doing that and to
> use CTL_UNNUMBERED.
>
> Frankly, I don't 100% remember the thinking - Eric, can you please remind
> us?

The thinking is this:

  Binary sysctl numbers are a problem because of patch conflicts like
  the above, and the related user space breakage they cause.

  In practice no one uses binary sysctl numbers.

So the policy should be to add new sysctl's using CTL_UNNUMBERED
(to prevent patch conflicts and user space breakage).

There may be cases where someone actually needs the binary sysctl
interface.  Once there is a demonstrated need we can go back
and very carefully add numbers for these very few cases, with
a strong review process.

Adding binary sysctl numbers should be done as carefully as and with
as much review as adding syscall numbers, and distro kernels and other
stable kernels should never get a sysctl number backport until the
number first reaches Linus's tree.  To avoid difference in meaning
between different kernels.

Given that no one except on BSD uses the binary sysctl interface
anyway my personal preference is to just freeze it and to reduce the
number of binary sysctls we support if possible.

Eric