Re: [PATCH 5/5] futex: fix miss ordered wakeups

[Daniel Walker <dwalker@mvista.com>]

On Thu, 2008-06-12 at 17:24 +0200, Thomas Gleixner wrote:
> > Just because you don't use it, doesn't make it useless .. At least
> > there's enough people asking for this that it warrants me writing it..
> 
> Which is not really a good technical reason to merge such a
> patch. Your handwaving about "enough people" is just irrelevant. Are
> you going to implement a root hole as well when enough people ask for
> it ?

People asking for something is a very good reason to merge "features" ..
You can like or dislike implementations , but that doesn't reflect on
the nature of the feature.

> But there is also a Real Good technical reason why these patches are
> going nowhere else than into /dev/null:
> 
>  your approach of hijacking blocked_on is fundamentaly wrong as it
>  mixes kernel internal state with user space state.

It mixes kernel state with kernel state, not to mention each state is
isolated from the others.

>  It will break in preempt-rt at the point when this state is set and
>  the code blocks on a spinlock, which uses the rtmutex based sleeping
>  spinlock implementation and overwrites blocked_on.

That's an intersting point, however "preempt-rt" is out of tree, so it's
certainly not going be a reason to reject mainline changes.

> If there would be a real good technical reason to fix this priority
> ordering it could be done with less than 20 lines of code without
> extra locking and wreckage waiting left and right, but I have not yet
> seen a single convincing technical argument or a relevant use case
> which might justify that.

The technical reasons for including this are the same technical reasons
why we want the waiters queued in priority order .. It's a requirement
of posix, where many calls need the ordering and ultimately feed into
the futex interface. So we have every reason to do the ordering
correctly..

If you have a 20 line fix for this then great tell me what it is..

Daniel