Before this patch, the TOC code used a pre-allocated stack of 16kb for
each possible CPU. That space overhead was the reason why the TOC
feature wasn't enabled by default for 32-bit kernels.

This patch rewrites the TOC code to use a per-cpu stack. That way we use
much less memory now and as such we enable the TOC feature by default on
all kernels.

Additionally the dump of the registers and the stacktrace wasn't
serialized, which led to multiple CPUs printing the stack backtrace at
once which rendered the output unreadable.
Now the backtraces are nicely serialized by a lock.
Signed-off-by: Helge Deller <deller@gmx.de>