Re: Priority Ceiling for SMP?

On Thu, 8 Nov 2007, Bernhard Kuhn wrote:

Luís Henriques wrote:

quoted

Maybe I didn't completely understood your question, but... if there is a
problem with SMP efficiency (I do not really know, but believe you :) ) with
priority ceiling (PC) implementation, this problem should have been already
identified (and possibly solved) in the current PI implementation, right?

To make it clear: I'm not aware of any efficiency issues with
the current PI implementation :-)


The (potential) efficiency issue with PC on SMP derives from the
dead-lock avoidance mechanism that you mentioned: If a task
locks a resource, then it's priority is raised to the ceiling
(which is typically the highest priority of all task sharing
that resource). Once that happens, no other task (sharing the
same resource) could then lock another resource that potentially
leads to a dead-lock situation.

Example:

* A low priority Task A shall share resources X and Y
  with a high priority Task B.

* Task A shall execute the following fraction of code:

    1. lock(X)
    2. lock(Y)

* Task B shall execute the following fraction of code:

    1. lock(Y)
    2. lock(X)

Classic ABBA deadlock :) Priority ceiling is not designed to allow
that. It's just a coincidence that it does not lead to a deadlock on
UP, but thats not a semantical property.

Thanks,

	tglx