On Mon, Sep 23, 2013 at 5:10 PM, Benjamin Herrenschmidt
<benh@kernel.crashing.org> wrote:
>
> BTW, that boils down to a choice between using r13 as either a TLS for
> current or current_thread_info, or as a per-cpu pointer, which one is
> the most performance critical ?

I think you can tune most of the architecture setup to best suit your needs.

For example, on x86, we don't have much choice: the per-cpu accessors
are going to be faster than the alternatives, and there are patches
afoot to tune the preempt and rcu-readside counters to use the percpu
area (and then save/restore things at task switch time). But having
the counters natively in the thread_info struct is fine too and is
what we do now.

Generally, we've put the performance-critical stuff into
"current_thread_info" as opposed to "current", so it's likely that if
the choice is between those two, then you might want to pick %r13
pointing to the thread-info rather than the "struct task_struct" (ie
things like low-level thread flags). But which is better probably
depends on load, and again, some of it you can tweak by just making
per-architecture structure choices and making the macros point at one
or the other.

There's a few things that really depend on per-cpu areas, but I don't
think it's a huge performance issue if you have to indirect off memory
to get that. Most of the performance issues with per-cpu stuff is
about avoiding cachelines ping-ponging back and forth, not so much
about fast direct access. Of course, if some load really uses a *lot*
of percpu accesses, you get both.

The advantage of having %r13 point to thread data (which is "stable"
as far as the compiler is concerned) as opposed to having it be a
per-cpu pointer (which can change randomly due to task switching) is
that from a correctness standpoint I really do think that either
thread-info or current is *much* easier to handle than using it for
the per-cpu base pointer.

             Linus