On Tue, 2005-03-29 at 13:56 +0200, Ingo Molnar wrote: > > To have a task take back the ownership, I had the stealer call > > task_blocks_on_lock on the task that it stole it from. To get this to > > work, when a task is given the pending ownership, it doesn't NULL the > > blocked_on at that point (although the waiter->task is set to NULL). > > But this gives the race condition in pi_setprio where it checks for > > p->blocked_on still exists. Reason is that I don't want the waking up > > of a process to call any more locks. To solve this, I had to (and this > > is what I don't like right now) add another flag for the process > > called PF_BLOCKED. So that this can tell the pi_setprio when to stop. > > This flag is set in task_blocks_on_lock and cleared in pick_new_owner > > where the setting of blocked_on to NULL use to be. > > which locks are affected? I'd prefer the simplest solution. If there's > more overhead with deadlock detection (which is a debugging feature), > that doesnt matter much.
I've already covered the deadlock detection problems, and that didn't add anymore overhead. That was just what kept breaking every time I changed something ;-) The overhead is caused by the pi_setprio knowing when to stop following the chain of blocked processes. Since it can't check for p->blocked_on == NULL anymore. I first had it check that or p->flags & PF_PENDOWNER, but since the process waking up can happen at anytime without grabbing any lock, this flag can be cleared as well as the blocked_on at the time of this test. So I added another flag to p->flags to tell when the process is locked and not pending ownership. I don't know yet if this works, I have to run to an appointment now and I'll find out when I get back. Oh, and I disabled the deadlock check on the stolen case. So when a pending owner goes back to blocked, right now it doesn't check the deadlock, since the code uses current as the test, and that no longer applies. So for now I have a test in task_blocks_on_lock to see if task == current otherwise, don't check for deadlocks. -- Steve - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
