Re: Random panic in load_balance() with 3.16-rc

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jul 23, 2014 at 10:26:21AM -0700, Linus Torvalds wrote:
> On Wed, Jul 23, 2014 at 10:12 AM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > sched_init() definitely does _not_ allocate a cpumask_var.
> 
> Side note: another good rule of thumb for per-cpu variables is:
> 
>  - if you use __get_cpu_var() without taking the address of it, you're
> doing something wrong and stupid.
> 
> The whole - and really *only* - point of __get_cpu_var is to get the
> address of a a cpu variable. If you want to read the *value* of the
> variable, you should use "this_cpu_read()", which can use things like
> special instructions or segments to read the percpu area.

I think this code predates all the this_cpu* magic. But yes, agreed.

> I agree that the interface is not all that great, there's historical
> baggage there. We would have been better off with
> "__this_cpu_ptr(var)" instead of "&__get_cpu_var(var)". But that
> "__get_cpu_var()" is the old way of doing things (predating the new
> and better "this_cpu_read/write/ops()" stuff), which is why we have
> that odd interface with "&__get_cpu_var()".

I think there's a whole bunch of patches by Christoph Lameter, queued by
TJ that remove all __get_cpu_var usage and eventually the interface.