Hi,
I finally got around to merge my patches into a newer -rt kernel and repost
them.
To refresh:
pthread_mutex_timedlock() for PI-futexes doesn't work on a UP machine:
Task A tries to take a lock with a timeout of say 1 ms. This lock is owned by B
which is boosted to task A's priority. The timeout interrupt wakes up A, but
since B is already running at A's priority, A will not get any CPU before B
unlocks the lock anyway.
This series of patches does the following:
1) It adds an interface to the scheduler such that task A is woken up LIFO
instead of FIFO. That means that A is now preempting B even though they have
the same priority. Thus A can de-boost B and exit the pthread_mutex_timedlock()
before B is done with the lock.
2) This is a smaller update to the rt-mutex-tester scripts. Probably not
needed, I'll post it anyway.
3) This patch makes sure that A will not loose it's priority while boosting B.
In case of other PI-mutex operations touches A's priority or explicit
setscheduler() calls, A will not actually loose it's priority. That is
postponed until A leaves the lock operation (successfully or not). Thus A will
always be able to de-boost B.
4) Is a documentation update.
5) Is a fix to the PI-futex: There is (still) lacking a protection for
the rtmutex's internal state because. It might not be the cleanest way to do
this.
Esben
--
* Esben Nielsen <[email protected]> wrote:
> Hi,
> I finally got around to merge my patches into a newer -rt kernel and
> repost them.
ok - your patches look rather clean. I'll try to put them into the
2.6.19 based -rt queue that i'm working on and see what happens.
Ingo