lkml.org 
[lkml]   [2004]   [Oct]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [Ext-rt-dev] Re: [ANNOUNCE] Linux 2.6 Real Time Kernel
Date
> >
> > I think patch size is an issue, but I also think that , eventually, we
> > should change all spin_lock calls that actually lock a mutex to be more
> > distinct so it's obvious what is going on. Sven and I both agree that
> > this should be addressed. Is this a non-issue for you? What does the
> > community want? I don't find your code or ours acceptable in it's
> > current form , due to this issue.
> >
> > With the addition of PREEMPT_REALTIME it looks like you more than
> > doubled the size of voluntary preempt. I really feel that it should
> > remain as two distinct patches. They are dependent , but the scope of
> > the changes are too vast to lump it all together.
> >
>
>
> If there is the tendency to touch the concurrency controls in general
> all over the kernel, then I would suggest a script driven overhaul of
> all concurrency controls like spin_locks, mutexes and semaphores to
> general macros like
>
> enter_critical_section(TYPE, &var, &flags, whatever);
> leave_critical_section(TYPE, &var, flags, whatever);
>
> where TYPE might be SPIN_LOCK, SPIN_LOCK_IRQ, MUTEX, PMUTEX or whatever
> we have and come up with in the future.
>
> This could be done in a first step and then it is clearly identifiable
> and it gives us more flexibility to wrap different implementations and
> lets us change particular points in a more clear way.
>
> I would be willing to provide some scripted conversion aid, if there is
> enough interest to that. I started with some test files and the results
> are quite encouraging.
>



Ideally we would eventually provide some level of tunability, i.e.
if you want the spinlocks all the way around it should be possible
to have that, or one could enable degrees of enhancements,
expanding on the existing sequence starting with PREEMPT, IRQ_THREADS,
BKL, MUTEX, etc. In addition to that, once the minim set of spinlocks
necessary for RT is established, additional layers, corresponding to
the lock nesting order, could be established, making the "mutex-depth"
somewhat configurable based on the performance requirements.

The entire effort would have the side effect of making the locking and
critical sections more distinct, and reveal soft spots in concurrency
code, as well as to raise awareness of the code density inside
critical sections.

The concept of tunable foreground / background responsiveness,
based on preemptability of low priority processes comes to mind.
A lot of folks would probably not mind making UI responsiveness
a little crisper, others will want the throughput.

I realize this is an early stage to be looking at it so high end,
but I think in general this type of script would not be a bad addition
to the patch kit(s).


Sven


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 14:06    [from the cache]
©2003-2011 Jasper Spaans