Re: 2.6.15-rt17

[Steven Rostedt <rostedt@goodmis.org>]

On Sat, 25 Feb 2006, Esben Nielsen wrote:

> On Thu, 23 Feb 2006, Thomas Gleixner wrote:
>
> > On Wed, 2006-02-22 at 17:50 +0100, Esben Nielsen wrote:
> > > I didn't know anyone looked at my patch! I am ofcourse happy about it :-)
> >
> > Was just delayed due to other work in progress.
> >
>
> You didn't use the "TestRTMutex" I send along with the patch :-(
>
> Since I learned to use unittesting that way I have been a big fan. It does
> catch a lot of stupid bugs without having to wait for the program to get
> there while keeping the ability to debug with gdb and fix it right away.
> And most important: you can keep the tests and check if your program still
> parses them after a rewrite. Very usefull to prevent repeating bugs.
>
> So here is the issues I have found:
>
> 1) grablockBKL.tst failes. Apparently you can "grab" the BKL now? Is this
> intended? I have updated the test to accept this.

since the BKL is now released on semaphores, I guess this is ok.

>
> 2) 2locksdeadlock.tst loops forever. It is a livelock: When two RT-tasks
> "deadlocks" on two mutexes they keep waking up each other in other. I
> quickly fixed that bug.

I actually thought about that.  But it still is an improvement, since it
doesn't deadlock the kernel. Just all processes that are of lower
priority.  This still should be fixed, but it needs to not cause any
noticable overhead.

>
> 3) While fixing that I came to think about what happens when you signal a
> task blocked on a task blocked on a task. I thus wrote
> 2lock3tasksBoostSignal.tst. Well, the priorities wasn't set back
> correctly. I fixed that too.
>
> I have attached the patch againt 2.6.17-rt17 (and therefore also
> included the previous patch) along with the updated tester and tests.
>
> Esben

I'll take a look at this tomorrow.

-- Steve

>
>
> > > That was why I had _reversed_ the lock ordering relative to normal in the
> > > original patch: First lock task->pi_lock. Assign lock. Lock
> > > lock->wait_lock. Then unlock task->pi_lock. Now it is safe to refer to
> > > lock. To avoid deadlocks I used _raw_spin_trylock() when locking the 2.
> > > lock.
> >
> > Stupid me. I messed that one up. Should show up in the next -rt
> >
> > Thanks
> >
> > 	tglx
> >
> >
>
>