In short, it is illegal to call __pa() on an address holding
a percpu variable. The times when this actually matters are
pretty obscure (certain 32-bit NUMA systems), but it _does_
happen. It is important to keep KVM guests working on these
systems because the real hardware is getting harder and
harder to find.
This bug manifested first by me seeing a plain hang at boot
after this message:
CPU 0 irqstacks, hard=f3018000 soft=f301a000
or, sometimes, it would actually make it out to the console:
[ 0.000000] BUG: unable to handle kernel paging request at ffffffff
I eventually traced it down to the KVM async pagefault code.
This can be worked around by disabling that code either at
compile-time, or on the kernel command-line.