Re: Types of mutexes?

I have numerous comments on this.

    * It is true that PROTECT mutexes (or priority ceiling mutexes, by
      their more common name) are easier to implement than priority
      inheritance mutexes.  It is not true that they are lower impact,
      if impact means the amount of CPU they use.  At least on CPUs with
      compare-and-swap, priority inheritance mutexes can be implemented
      without a system call on an uncontended access to the mutex.
    * The major benefit of priority ceiling mutexes is that they are
      deadlock free on a correctly implemented uniprocessor system.
       This is not true for multi-processor machines, and this benefit
      is kind of meaningless for systems of any complexity because they
      are impossible to completely characterize (and thus impossible to
      correctly implement).
    * To correctly use priority ceiling mutexes, you must know the
      highest priority thread that will claim the mutex (including
      priorities that were boosted by claiming other mutexes).  The
      complexity of determining this grows exponentially with the size
      of the application.  People will not want to do this on things
      like, say, telephone switches.
    * Some high-use mutexes (say, the memory allocation mutex in a Java
      VM) will be used by basically all threads, and thus must have
      their ceiling set to the highest priority thread in the system (if
      you have access to the mutexes at all).  It is possible to make
      every mutex in the system priority inheriting and the system
      safely function.

IMO, in general, priority ceiling mutexes are good for hard real-time 
system with limited complexity.  Priority inheritance mutexes are good 
for complex soft real-time systems.  (If you have to build a complex 
hard real-time system, you have my sympathy :-).  I'm guessing that most 
users will be in the complex soft real-time category, so I think it 
makes more sense to provide priority inheritance mutexes.  Actually 
priority ceiling is so easy to implement that it makes sense to provide 
both.

BTW, I have an implementation of priority inheritance using 
compare-and-swap that doesn't do a system call on an uncontested access. 
 It's at http://ssthreads.sf.net, if you are interested.

-Corey

Perez-Gonzalez, Inaky wrote:

>Hi All
>
>We are studying what types of POSIX mutexes to implement, out of PROTECTION
>and PRIORITY_INHERITANCE; being PROTECTION much more simpler, and
>INHERITANCE kind of difficult and with a high footprint/impact, I'd like to
>ask everybody if they have a _huge_ desperate need for INHERITANCE, and what
>is the need they envision for it or usage they have seen.
>
>Thanks  :)
>_______________________________________________
>pthreads-devel mailing list
>[email protected]
>http://www-124.ibm.com/developerworks/oss/mailman/listinfo/pthreads-devel
>  
>