[PATCH 03/13] ARM: b.L: core switcher code

[Russell King - ARM Linux]

On Mon, Jul 22, 2013 at 11:31:19PM -0400, Nicolas Pitre wrote:
> To solve both issues, we borrow some of the task space belonging to
> the init/idle task which, by its nature, is lightly used and therefore
> is unlikely to clash with our usage.  The init task is also never going
> away.

That is... quite a horrid hack - and it is a hack.  You claim that
the idle task is not likely to clash, but with 8 CPUs each of
256 bytes of stack, that's 2k of stack.  With the thread data
structure at over 512 bytes, that's 2.5k of 8k.

Don't forget that the amount of kernel stack used depends on many
things - not just what the thread does (in the case of the idle
thread, that includes cpuidle stuff) but also the function call
depth of any interrupt in the system, as well as the path used to
get into the idle thread.

If you want to do this, it would be much better to use the bottom
of the stack space for the idle thread associated with the CPU.  The
idle threads are all forked at boot for any possible CPUs before the
secondaries are brought online, and so are persistent.  That would
reduce the additional required space to 256 bytes on the stack, not
2k on a single stack.

It also means that if we see larger clusters, we won't have this
in-built limitation caused by the size of the kernel stack.