From: Henrik Austad <haus...@cisco.com>
Short story:
The following patches are needed on a 4.4 kernel to avoid
Oops in the scheduler when a sched_rr and sched_deadline task contends
on the same futex (with PI).
Longer story:
On one of our arm64 systems, we occasionally crash with an Oops in the
scheduler with the following backtrace.
[<ffffffc0000ee398>] enqueue_task_dl+0x1f0/0x420
[<ffffffc0000d0f14>] activate_task+0x7c/0x90
[<ffffffc0000edbdc>] push_dl_task+0x164/0x1c8
[<ffffffc0000edc60>] push_dl_tasks+0x20/0x30
[<ffffffc0000cc00c>] __balance_callback+0x44/0x68
[<ffffffc000d2c018>] __schedule+0x6f0/0x728
[<ffffffc000d2c278>] schedule+0x78/0x98
[<ffffffc000d2e76c>] __rt_mutex_slowlock+0x9c/0x108
[<ffffffc000d2e9d0>] rt_mutex_slowlock+0xd8/0x198
[<ffffffc0000f7f28>] rt_mutex_timed_futex_lock+0x30/0x40
[<ffffffc00012c1a8>] futex_lock_pi+0x200/0x3b0
[<ffffffc00012cf84>] do_futex+0x1c4/0x550
[<ffffffc00012d92c>] compat_SyS_futex+0x10c/0x138
[<ffffffc00008504c>] __sys_trace_return+0x0/0x4
This seems to be the same bug Xuneli Pang triggered and fixed in
e96a7705e7d3 "sched/rtmutex/deadline: Fix a PI crash for deadline
tasks". As noted by Peter Zijlstra in the previous attempt, this fix
requires a few other patches, most notably the FUTEX_UNLOCK_PI series
[1]
Testing this on a dual-core VM I have not been able to reproduce the
same crash, but pi_stress (part of the rt-test suite) reveals that
vanilla 4.4.162 behaves rather badly with a mix of deadline and
sched_(rr|fifo) tasks:
time pi_stress --rr --mlockall --sched
id=high,policy=deadline,runtime=100000,deadline=200000,period=200000
Starting PI Stress Test
Number of thread groups: 1
Duration of test run: infinite
Number of inversions per group: unlimited
Admin thread SCHED_RR priority 4
1 groups of 3 threads will be created
High thread SCHED_DEADLINE runtime 100000 deadline 200000 period 200000
Med thread SCHED_RR priority 2
Low thread SCHED_RR priority 1
Current Inversions: 141627
WATCHDOG triggered: group 0 is deadlocked!
reporter stopping due to watchdog event
Stopping test
Terminated
real 0m26.291s
user 0m0.148s
sys 0m18.819s
With this series applied, the test ran for ~4.5 hours and again for 129
minutes (when I remembered to time it) before crashing:
time pi_stress --rr --mlockall --sched
id=high,policy=deadline,runtime=100000,deadline=200000,period=200000
Starting PI Stress Test
Number of thread groups: 1
Duration of test run: infinite
Number of inversions per group: unlimited
Admin thread SCHED_RR priority 4
1 groups of 3 threads will be created
High thread SCHED_DEADLINE runtime 100000 deadline 200000 period 200000
Med thread SCHED_RR priority 2
Low thread SCHED_RR priority 1
Current Inversions: 51985223
WATCHDOG triggered: group 0 is deadlocked!
reporter stopping due to watchdog event
Stopping test
Terminated
real 129m38.807s
user 0m59.084s
sys 109m53.666s
So clearly not perfect, but a *lot* better.
The same series on our vendor-4.4 kernel moves pi_stress up from ~30
seconds before deadlock up to the same level as the VM (the test is
still going as of this writing).
I suspect other users of 4.4 would benefit from having these patches
backported, so tag them for stable. I assume 4.9 and 4.14 could benefit
as well, but I have not had time to look into those.
1) https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1359667.html
Peter Zijlstra (13):
futex: Cleanup variable names for futex_top_waiter()
futex: Use smp_store_release() in mark_wake_futex()
futex: Remove rt_mutex_deadlock_account_*()
futex,rt_mutex: Provide futex specific rt_mutex API
futex: Change locking rules
futex: Cleanup refcounting
futex: Rework inconsistent rt_mutex/futex_q state
futex: Pull rt_mutex_futex_unlock() out from under hb->lock
futex,rt_mutex: Introduce rt_mutex_init_waiter()
futex,rt_mutex: Restructure rt_mutex_finish_proxy_lock()
futex: Rework futex_lock_pi() to use rt_mutex_*_proxy_lock()
futex: Futex_unlock_pi() determinism
futex: Drop hb->lock before enqueueing on the rtmutex
Thomas Gleixner (2):
rtmutex: Make wait_lock irq safe
futex: Rename free_pi_state() to put_pi_state()
Xunlei Pang (2):
rtmutex: Deboost before waking up the top waiter
sched/rtmutex/deadline: Fix a PI crash for deadline tasks
include/linux/init_task.h | 1 +
include/linux/sched.h | 2 +
include/linux/sched/rt.h | 1 +
kernel/fork.c | 1 +
kernel/futex.c | 532 ++++++++++++++++++++++++++--------------
kernel/locking/rtmutex-debug.c | 9 -
kernel/locking/rtmutex-debug.h | 3 -
kernel/locking/rtmutex.c | 406 ++++++++++++++++++------------
kernel/locking/rtmutex.h | 2 -
kernel/locking/rtmutex_common.h | 24 +-
kernel/sched/core.c | 2 +
11 files changed, 620 insertions(+), 363 deletions(-)
--
2.7.4
