Re: [PATCH] Replace completions with semaphores

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2008-04-15 at 09:17 +0200, Bart Van Assche wrote:
> On Tue, Apr 15, 2008 at 8:46 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > On Tue, 2008-04-15 at 08:18 +0200, Bart Van Assche wrote:
> >  > Would it really be a good idea to give a synchronization concept that
> >  > behaves exactly like a semaphore another name than "semaphore" ? The
> >  > semaphore concept is well known and is taught in every computer
> >  > science course.
> >
> >  Are the ramifications wrt priority inversion taught? Is it made clear
> >  that its hard to validate because there is no clear resource owner?
> >
> >  Afaik, non of these subjects are touched upon in the CS-101 courses and
> >  that is exactly the problem. So you can say they are not well know, they
> >  are just widely misunderstood.
> >
> >  And yes, if there are more hand a very few such users it doesn't make
> >  sense to keep them open coded.
> 
> Regarding semaphores and priority inversion: I have never recommended
> the use of semaphores over mutexes, all I recommended is to keep the
> name "semaphore" for something that behaves like a semaphore. There
> might be better ways to discourage the use of the semaphore API, e.g.
> letting the compiler print a warning every time a semaphore function
> is called unless one or another #define has been enabled.

That sounds horrible; I really prefer targeted replacements like
completions that make it clear what they're supposed to be used for.

> Regarding priority inheritance: does the above mean that you consider
> priority inheritance as an optimal solution for realizing real-time
> behavior in the kernel ? Are you aware of the fundamental problems
> associated with priority inheritance ? These issues are well explained
> in Victor Yodaiken's paper "Against priority inheritance". See also
> http://www.linuxdevices.com/files/misc/yodaiken-july02.pdf .

Priority inheritance isn't ideal, but comming from a general purpose
kernel that wasn't build from scratch to accomodate hard realtime its
basically the only option.

Also things like lockdep are a real help to a lot of developers - loads
of locking bugs never make it into the kernel because of it.