Hi Paul,

On 10/21/2015 01:13 PM, Peter Zijlstra wrote:

On Wed, Oct 21, 2015 at 08:24:11PM +1100, Paul Mackerras wrote:

quoted

On Tue, Oct 20, 2015 at 04:00:31PM +0200, Peter Zijlstra wrote:

quoted

Should we not take this opportunity to get rid of these open-coded wait
loops?


Does this work?

No, on Book3S HV (POWER8) the VM hangs immediately after the kernel
brings up all the secondary vCPUs, and is then unkillable.  I'm not
sure what's wrong, although I wonder why you have TASK_IDLE rather
than TASK_INTERRUPTIBLE in the ___wait_event call.

This was under the assumption that INTERRUPTIBLE was because you wanted
to avoid increasing load. Which was based on the lack of
signal_pending() tests near there (although there might have been in the
outermost loop which I overlooked).

If it does rely on signals, then this was obviously false and TASK_IDLE
is indeed wrong.

If I get this right, the current patch is okay but you are going to
redactor the code? Maybe I can cherry pick your patch then and update
this patch accordingly.

cheers,
daniel