> OK, this patch fixed it. I don't know why.
Eh, odd. That effectively disables kernel preemption around
set_cpus_allowed, but preemption is already disabled by the task_rq_lock
call. Note, however, preemption is enabled by the task_rq_unlock and
thus wake_up_process is called with preemption enabled.
With your patch, preemption is now disabled across the call, and
subsequently the task_rq_unlock in try_to_wake_up will never call
preempt_schedule and your lock does not happen.
The actual problem may be elsewhere, and this just hides it. This is
pretty clear, since we would get a similar effect just wrapping
wake_up_process in preempt_disable. But, oh, try_to_wake_up disables
preempt, too ... hrm.
Hm, what if try_to_wake_up wakes up a process and then preemptively
schedules into it and it wants to acquire the req.sem semaphore, but
cannot, as it is still taken by set_cpus_allowed? The semaphore seems
to just be used in the migration code.
So we have init spinning on softirq threads to come up and then we have
a deadlock on req.sem from set_cpus_allowed and into the migration
thread?
Bleh ... Ingo?
Robert Love
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/