Re: static DEFINE_PER_CPU vs. modules

[Arnd Bergmann <arnd@arndb.de>]

On Monday 03 May 2004 20:37, David Mosberger wrote:
> This was in the context of per-CPU addressing.  I don't know the s390x
> well but if the 32-bit offset is a memory-model limitation, fine.  The
> real problem then seems to lie in the way per-CPU addressing is
> imlemented on s390x, no?
 
Exactly. However, s390x is using the same code as every other architecture
aside from ia64. Both s390{,x} and x86_64 have some optimized way to get
to the per_cpu_offset, but that does not impact the problem.

> Perhaps s390x could use a similar trick as we do on ia64?  We put the
> per-CPU area in the last 64KB of the address space, so the absolute
> address is guaranteed to fit into a small signed integer constant.
> Newer compilers support the "model(small)" attribute so the compiler
> knows it can use a simple "addl" instruction to calculate the address
> of a per-CPU variable.

I don't understand how that would help. On an 4GB+ system, modules get
get loaded to a >4GB address, so a 32 bit PC-relative relocation that
can not get to a small positive address won't be enough for a small negative
address either. The only way to force absolute addressing instead of
PC-relative on s390x is to compile with -fPIC (as we do) _and_ declare
the symbol non-static.
It could work if we load the modules to the top of the address space as
well, is that what you mean?

	Arnd <><