It turns out the kernel relies on barrier() to force a reload of the
percpu offset value.  Since we can't easily modify the definition of
barrier() to include "tp" as an output register, we instead provide a
definition of __my_cpu_offset as extended assembly that includes a fake
stack read to hazard against barrier(), forcing gcc to know that it
must reread "tp" and recompute anything based on "tp" after a barrier.

This fixes observed hangs in the slub allocator when we are looping
on a percpu cmpxchg_double.

A similar fix for ARMv7 was made in June in change 509eb76e.

Cc: stable@vger.kernel.org
Signed-off-by: NChris Metcalf <cmetcalf@tilera.com>

This change is the core kernel support for TILEPro and TILE64 chips.
No driver support (except the console driver) is included yet.

This includes the relevant Linux headers in asm/; the low-level
low-level "Tile architecture" headers in arch/, which are
shared with the hypervisor, etc., and are build-system agnostic;
and the relevant hypervisor headers in hv/.
Signed-off-by: NChris Metcalf <cmetcalf@tilera.com>
Acked-by: NArnd Bergmann <arnd@arndb.de>
Acked-by: NFUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Reviewed-by: NPaul Mundt <lethal@linux-sh.org>