Hello, Andrew.

On Tue, Feb 07, 2017 at 02:14:20PM -0800, Andrew Morton wrote:
> >      extern __PCPU_DUMMY_ATTRS char __pcpu_unique_##name;  \
> >                                     ^
> 
> huh, yes.  The DEFINE_PER_CPU() macro is broken.

Yeah, that was the trade off I had to take with percpu vars to force
s390 and alpha to generate long references (GOT based addressing) for
percpu variables; otherwise, they generate memory deref which is too
limited to access the special percpu addresses.  It's explained in
include/linux/percpu-defs.h.

> If you do
> 
> foo()
> {
>       static DEFINE_PER_CPU(int, bar);
> }
> 
> then it won't compile, as described here.  It should.
> 
> And if you do
> 
> static DEFINE_PER_CPU(int, bar);
> 
> then you still get global symbols (__pcpu_unique_bar).
> 
> The kernel does the above thing in, umm, 466 places and afaict they're
> all broken.  If two code sites ever use the same identifier, they'll
> get linkage errors.

So, we have CONFIG_DEBUG_FORCE_WEAK_PER_CPU to catch those cases on
archs other than s390 or alpha.

> huh.  Seems hard to fix.

This was the only way I could come up with to support alpha and s390.
All the restrictions are there to ensure that.  If we can do s390 and
alpha w/o the global weak reference, neither restriction is necessary.

Thanks.

-- 
tejun

Reply via email to