Re: [PATCH] selinux: init target class when add avc callback

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 06 Mar 2012 19:41:33 -0500
Eric Paris <eparis@parisplace.org> wrote:

> __initcall() functions aren't serialized?  I guess that would be bad and
> we would need a lock.  I wonder if there are other places I assumed
> __initcall() would be serialized (note that all of these call sites are
> built in and not modules if that makes a difference)

There's plenty of code in the kernel which assumes that initcalls
are singly-threaded.  And init/main.c:do_initcalls() is very
singly-threaded!  It's less clear when the initcall is executed by
modprobe, but presumably there's something in the module code which
prevents concurrent execution of module_init() functions.

So I think the list management code is acceptable, as long as we ensure
that the function is only ever called from initcall functions.  We can
add a comment, but a neat way of ensuring this is to mark the function
__init.  This saves memory and will cause a build-time warning if we
screw up.

My point was that given that this function is only ever called from
initcalls, that weak GFP_ATOMIC could/should be replaced with
GFP_KERNEL.