On Thu, Sep 04, 2008 at 10:53:45PM +0200, Andi Kleen wrote:
The way it was explained to me months ago, is that it can't. Currently it
works by making sure the nmi watchdog is disabled on the box (either
compiled off, or nmi_watchdog=0). Then it registers itself on the DIE
chain. Because it is on the top of the chain, it just so happens to
be the first handler to process the NMIs, so HP is happy. Kdump works
because it registers later on top of the HP handler thus it is called
before HP can panic the box. However, I suspect if one were to run
oprofile in nmi mode, HP would panic the box on the first profile point.
Hence our attempt at a solution by creating a registration for a new
nmi_handler, which in hindsight seemed to miss a couple of cases.