On Sun, Dec 23, 2012 at 11:25:40AM +0000, Will Deacon wrote:

On Fri, Dec 21, 2012 at 09:23:17PM +0000, Stephen Warren wrote:

quoted

What exactly did you mean by "what we currently have"; did you mean that
cpu_kill() should work without cpu_die() having executed on the target
CPU itself first, so Tegra should simply implement cpu_kill()? As Joseph
says above, I'm not sure that will work. Any more detailed thoughts you
have here would be greatly appreciated. Thanks.

I guess this comes down to the different between cpu_die/cpu_kill and
whether you actually need cpu_die if you're not planning on returning to
the same kernel that you left. If you really need to call cpu_die, then we
need to figure out a way to do that whilst ensuring that only one CPU is
left standing at the time we start copying the new kernel into place.

I took a closer look at the code to try and figure out a way to solve
this...

First of all, disable_nonboot_cpus *does* actually wait for the secondaries
to get out of the way, because it calls __cpu_die and we end up waiting on a
completion. The problem is that we IPI the secondaries in smp_send_stop,
which puts then into a while(1) cpu_relax(); loop with interrupts disabled,
so that completion won't be able to finish (the cores won't pass through
idle and call cpu_die() themselves).

This means we must call disable_nonboot_cpus before sending the IPI. My
first attempt was to make a hotplug-aware smp_send_stop, using
disable_nonboot_cpus and smp_kill_cpus but this is also problematic because
it's called by panic, where the simpler IPI-based code is probably a lot
better.

The best bet is probably to call disable_nonboot_cpus in machine shutdown,
before sending the stop and subsequent kill. It's very similar to your
initial suggestion, but without the #elif.

Does that work for you?

Will