Re: [RFC 00/15] x86_64: Optimize percpu accesses

[Jeremy Fitzhardinge]

Arjan van de Ven wrote:
> what's wrong with zero based btw?
>   

Nothing in princple.  In practice it's triggering an amazing variety of 
toolchain bugs.

> do they stop us from using gcc's __thread keyword for per cpu variables
> or something? (*that* would be a nice feature)
>   

The powerpc guys tried it, and it doesn't work.  per-cpu is not 
semantically equivalent to per-thread.  If you have a function in which 
you refer to a percpu variable and then have a preemptable section in 
the middle followed by another reference to the same percpu variable, 
it's hard to stop gcc from caching a reference to the old tls variable, 
even though we may have switched cpus in the meantime.

Also, we explicitly use the other segment register in kernel mode, to 
avoid segment register switches where possible.  Even with 
-mcmodel=kernel, gcc generates %fs references to tls variables.

   J
--
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/