cpu_relax in busy loops question

On Wed, Feb 1, 2012 at 12:34 PM, Daniel Lezcano
[off-list ref] wrote:

Hi Linus,

While looking at the drivers/mfd/db8500-prcmu.c file I noticed:


?595 ? ? ? ? /* Grab the HW semaphore. */
?596 ? ? ? ? while ((readl(PRCM_SEM) & PRCM_SEM_PRCM_SEM) != 0)
?597 ? ? ? ? ? ? ? ? cpu_relax();


I was wondering why is cpu_relax needed here as readl does a memory barrier
? I thought the cpu_relax function was related to the x86 to consume less
power with an optimization of the "rep nop" loops or/and do a memory
barrier.

Should a busy-loop be always with cpu_relax ?

Last time we discussed it ar LAKML I think we came to the conclusion
that while cpu_relax() has very little semantic meaning on ARM, it
is still nice to have from a readability and maintenance point of view,
since it is pretty clear what we're trying to do.

But I will stand corrected...

Yours,
Linus Walleij