On Fri, Jul 4, 2008 at 6:48 PM, Rene Herman <[EMAIL PROTECTED]> wrote: > On 04-07-08 02:05, Peter Teoh wrote: > >> I see....eh...or I don't see......the original problem was that the >> highest prio cannot acquire the lock, even though it has the highest prio. > > No. That is or was not the problem. This I believe is the fundamental > misunderstanding left. > > Please note how I said "All fine and well up to this point" when the > high-prioriy process went to sleep since the low-priority process was > holding the lock. > > Such is the nature of the mutex/futex locking model. Holding a lock > _overrides_ any scheduler priority, makes it irrelevant, and this is how > things are supposed to work. If we just have the high-priority and the > low-priority process, we have no priority inversion issues, no problem. > > It's mixing in that third medium priority process that's causing the > problem. It itself does _not_ grab the lock, yet it still keeps the > high-priority process from running since it keeps the low priority process > from running to the point where it releases the lock. So we have here a > situation where a process that isn't holding a lock the higher priority > process is waiting for is still keeping the higher priority process from > running -- effectively, the medium priority process has a higher priority > then the high-priority process since it has a higher priority then the > low-priority process. > > That's the problem called "priority inversion". > > I suppose once that's clear, the priority inheritance solution will also be: > just assign the priority of the high-prio process to the low-prio process > when the high prio-process blocks on the lock so that the medium-prio > process won't keep the (originally) low-prio process from running (tweak > solution to "assign the maximum of current prio of lock holder and prio of > process about to be blocked to lock holder" in the general case of more than > just 3 possible processes ofcourse). > > Rene. >
Ah....very clear and well written explanation - thank you for the effort put into the writing. And for the phenomena of "unbounded priority inversion", I found this very useful: http://lkml.org/lkml/2006/5/10/52 Thank you to Rene and Roberto for the discussion..... -- Regards, Peter Teoh -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to [EMAIL PROTECTED] Please read the FAQ at http://kernelnewbies.org/FAQ
