Re: 2.6.23-rc2-mm2

Andrew Morton [off-list ref] writes:

There seems to be rather a lot of damage here.

I assume that the sysctl changes are what caused the netfilter oopses.

- nf_conntrack_init() calls nf_conntrack_expect_init() which fails due to
  sysctl problems.  

In particular sysctl_check_table finds issues with the sysctl table
so register_sysctl_table refuses to register it.

- nf_conntrack_init() bales out without calling nf_conntrack_helper_init()

  So nf_ct_helper_hsize never gets initialised.

- Later, netfilter client code calls helper_hash(), which gets a
  divide-by-zero due to nf_ct_helper_hsize==0.


yeah, that's a netfilter bug, but we're trying to get kernels tested here. 
If I'm feeling energetic I'll drop the sysctl changes and do rc2-mm3. 
Probably I won't feel energetic, but we'll need a lot of fixes here before
I can release the sysctl changes in another -mm, please.

As a cheap workaround it should be possible to disable SYSCTL support
in 2.6.23-rc2-mm2 to get around these issues.


Andrew for the moment I have just sent you fixes for all of the issues
that I am aware of.  Mostly they are cheap kill the sys_sysctl()
support patches.

Hopefully that is enough to bring the pain level down to manageable.

I hadn't anticipated subsystems failing because they could not register
their sysctl tables.  I was simply expecting things not to show up in
/proc/sys.  And more of the pain of making working sysctl tables to 
be pushed back to developers.

Eric