Re: [PATCH 0/6] Optimize the cpu hotplug locking -v2

[Linus Torvalds]

On Thu, Oct 10, 2013 at 11:43 AM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> There's places in the kernel that does for_each_cpu() that I'm sure you
> don't want to disable preemption for. Especially when you start having
> 4096 CPU machines!

We could _easily_ add preemption points in for_each_cpu() for the
MAX_SMP case. We can even do it in the macro itself.

So I wouldn't worry about things like that. I'd expect the "Oh damn,
we have to allocate" case to be more painful, but I think that is
fairly easily handled by having something like a "redo_cpu()" macro
that you can call inside for_each_cpu(). So the pattern for the
(nopefully-not-all-that-common) thing would be something like

   sometype *prealloc = NULL;

   for_each_cpu() {
         if (!prealloc) {
               rcu_read_unlock();  // Or whatever we want to call it
               prealloc = kmalloc(..);
               rcu_read_lock();
               redo_cpu();
          }
   }
   kfree(prealloc);

which is certainly more work than a sleeping lock is, but this looks
like a initialization pattern.

Looking at the things that are actually performance-critical, they
tend to be things like "gather all the statistics for all CPUs".

Honesty in advertizing: I only grepped a couple of users of
get_online_cpus(), and they all seemed like they'd be perfectly happy
with preemption-off. But maybe I was just lucky in my few samples: I
have anecdotes, not data.

              Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/