I apologise if this is a duplicate post. The first one didn't seem
to go through properly.
---

Having arch code know about scheduler locking in their context
switch stuff is somewhat ugly I think.

It was somewhat acceptable because they all used the same
implementation. However the ARM port recently has wanted a
slightly different system, and also does some clever !SMP
optimisations there.

So I propose a system where the locking implementations are
hidden in kernel/sched.c. Architectures can define
__ARCH_WANT_UNLOCKED_CTXSW or __ARCH_WANT_INTERRUPTS_ON_CTXSW.
Those that do need to define TIF_RUNNING, which I haven't done
in the patch.

Why have a locked, interrupts off context switch at all then?
Architectures which can do the switch quickly can keep it
atomic and not incur the small cost of an atomic set and clear
TIF_RUNNING per switch.

Affected architectures: ia64, mips, s390, sparc, sparc64, arm.

Comments?