Results

[Sebastian Huber]

On 26/06/15 02:33, Saurabh Gadia wrote:
> @joel: It essentially means finding out the highest priority thread 
> (or ceiling) of any remaining mutex held, right?
> If you are talking this in context of nested mutex problem --->
> We never go and find the highest priority thread waiting on held 
> mutex. Inversely highest priority thread will recursively make changes 
> to top of stack of held mutexes by the holder. We don't search.
>
> Thanks,
>
> Saurabh Gadia
>
> On Thu, Jun 25, 2015 at 5:27 PM, Joel Sherrill 
> <joel.sherrill at oarcorp.com <mailto:joel.sherrill at oarcorp.com>> wrote:
>
>
>
>     On June 25, 2015 7:23:06 PM CDT, Cyrille Artho
>     <cyrille.artho at gmail.com <mailto:cyrille.artho at gmail.com>> wrote:
>
[...]
>
>     >(2) Is a fix using a global lock acceptable? Correctness-wise
>     it's the
>     >safest thing, but of course performance-wise it's a bottleneck.
>
[...]

If you use the global lock for the uncontested mutex case, then this is 
not acceptable. If you use a global lock once you must block, then it 
might be acceptable, but this solution is not scalable. I am not sure 
what you are actually working on, it would be nice to discuss and 
present more on this mailing list. Maybe even open a ticket, e.g. 
something similar to https://devel.rtems.org/ticket/2271. This way it is 
possible to relate commits to a general concept discussion.

For SMP there is a ready to use locking protocol capable to replace the 
priority inheritance mutexes:

@inproceedings{OMIP,
   author="Brandenburg, Björn B.",
   title="A Fully Preemptive Multiprocessor Semaphore Protocol for 
Latency-Sensitive Real-Time Applications",
   booktitle="Proceedings of the 25th Euromicro Conference on Real-Time 
Systems (ECRTS 2013)",
   pages="292-302",
   year="2013",
   url="http://www.mpi-sws.org/~bbb/papers/pdf/ecrts13b.pdf",
}

Implementing support for nested resources is very complex if you don't 
simplify like it was done for the current priority inheritance protocol 
in RTEMS. If you want it more general, then you have to pay for this in 
terms of additional memory space and time. The big issue is that you 
have to merge and split dependency trees. A tree merge can be done in 
O(log(n)) with a binomial heap, but the split operation will be probably 
not better than O(n). On SMP with partitioned/clustered scheduling you 
have to keep in mind that priority values of different partitions are 
not comparable.

-- 
Sebastian Huber, embedded brains GmbH

Address : Dornierstr. 4, D-82178 Puchheim, Germany
Phone   : +49 89 189 47 41-16
Fax     : +49 89 189 47 41-09
E-Mail  : sebastian.huber at embedded-brains.de
PGP     : Public key available on request.

Diese Nachricht ist keine geschäftliche Mitteilung im Sinne des EHUG.