Steven,
As we've discussed, the following changes break the futex_requeue_pi
testcase I use to validate the futex_requeue_pi feature introduced in
2.6.32(ish).
commit 43fa5460fe60dea5c610490a1d263415419c60f6
sched: Try not to migrate higher priority RT tasks
commit b3bc211cfe7d5fe94b310480d78e00bea96fbf2a
sched: Give CPU bound RT tasks preference
The source to this test is here:
http://git.kernel.org/?p=linux/kernel/git/dvhart/futextest.git;a=summary
branch: futex-lock
file: futextest/functional/futex_requeue_pi.c
The futex_requeue_pi test main loop runs as sched_other while all the
child threads are spawned as sched_fifo 1. This scenario results in the
parent thread spinning in futex_wait_queue_me (most likely on the hash
bucket spinlock) while attempting a pthread_join() on the rt child
threads. However, if I run with chrt, forcing the parent thread to fifo
1, it completes reliably:
$ sudo chrt -f 1 ./futex_requeue_pi
futex_requeue_pi: Test requeue functionality
Arguments: broadcast=0 locked=0 owner=0 timeout=0ns
Result: PASS
Whereas as sched_other:
$ sudo ./futex_requeue_pi
futex_requeue_pi: Test requeue functionality
Arguments: broadcast=0 locked=0 owner=0 timeout=0ns
<hang>
$ ps -eLo tid,rtprio,wchan:20,comm | grep futex
1987 - futex_wait_queue_me auditd
1988 - futex_wait_queue_me audispd
5533 - futex_wait_queue_me rsyslogd
5547 - futex_wait_queue_me futex_requeue_p
5549 1 futex_wait_queue_me futex_requeue_p
5550 1 futex_wait_queue_me futex_requeue_p
5551 1 futex_wait_queue_me futex_requeue_p
5552 1 futex_wait_queue_me futex_requeue_p
5553 1 futex_wait_queue_me futex_requeue_p
5554 1 futex_wait_queue_me futex_requeue_p
5555 1 futex_wait_queue_me futex_requeue_p
5556 1 futex_wait_queue_me futex_requeue_p
5557 1 futex_wait_queue_me futex_requeue_p
5558 1 ? futex_requeue_p
5559 1 ? futex_requeue_p
I'm still thinking through how your changes would impact this, and if
there could be some assumption in the futex code that fails with those
changes. I wanted to get the information out there in case it triggers
any insight from you.
This suggests that somehow an rt task blocked while holding a spinlock
(without issuing a warning to the console) or that somehow a SCHED_OTHER
task preempted an RT task which is not getting scheduled on another CPU.
Neither of these seem likely.
Thanks,
--
Darren Hart
Embedded Linux Kernel