On Thu, 29 Jun 2017, Madhavan Srinivasan wrote:

On Thursday 29 June 2017 01:11 AM, Thomas Gleixner wrote:
Idea is to handle multiple event session for a given core and
yes, my bad, current implementation is racy/broken.
But an alternate approach is to have a per-core mutex and
per-core ref count to handle this.

event_init path:
per-core mutex lock
  if ( per-core[refcount] == 0) {
       rc = opal call to start the engine;
       if (rc on failure) {
       per-core mutex unlock;
       log error info;
       return error;
       }
  }
  increment the per-core[refcount];
per-core mutex unlock;


event_destroy path:
per-core mutex lock
  decrement the per-core[refcount];
  if ( per-core[refcount] == 0) {
     rc = opal call to stop the engine;
     if (rc on failure) {
       per-core mutex unlock;
       log the failure;
       return error;
     }
  } else if ( per-core[refcount] < 0) {
     WARN()
     per-core[refcount] = 0;
  }
per-core mutext unlock;

Yes, that works and looks about right.

Thanks,

	tglx