Re: the creation of boot_cpu_init() is wrong and accessing uninitialised data

[Andrew Morton <akpm@osdl.org>]

On Mon, 26 Jun 2006 21:45:13 -0500
James Bottomley <James.Bottomley@SteelEye.com> wrote:

> The basic problem with this function is that on most architectures
> smp_processor_id() is an alias for current_thread_info()->cpu which
> isn't given its correct value until setup_arch(), so adding a
> boot_cpu_init() that uses it *before* setup_arch() is called is plain
> wrong.  You manage to get away with it 99% of the time because its
> uninitialised value is zero and mostly the id of the booting CPU is
> zero ... but guess who's got a box currently booting on CPU 1 with no
> CPU 0?

preexisting bug, really.  printk() uses smp_processor_id().  Things like
spin_lock() will probably do so in certain debug modes.  We finally got
caught.

> Unfortunately, I can't think of a good solution for what you want to do.
> The thing that looks most plausible is hard_smp_processor_id() which
> reads the actual register value of the processor id.  However, on x86
> (and any other architectures that renumber their CPUs in setup_arch())
> this will ultimately turn out wrong again.

I think arch code should do it before calling start_kernel(), really.  It's
just such a basic part of the kernel framework.

A less wholesome but perhaps simpler solution would be to call the new
setup_smp_processor_id() on entry to start_kernel().