On Tue, 15 Dec 2020 at 23:52, Jakub Kicinski kuba@kernel.org wrote:
On Tue, 15 Dec 2020 06:45:31 -0800 Paul E. McKenney wrote:
Crash log:
# selftests: bpf: test_tc_edt.sh [ 503.796362] [ 503.797960] ============================= [ 503.802131] WARNING: suspicious RCU usage [ 503.806232] 5.9.15-rc1 #1 Tainted: G W [ 503.811358] ----------------------------- [ 503.815444] /usr/src/kernel/kernel/sched/core.c:7270 Illegal context switch in RCU-bh read-side critical section! [ 503.825858] [ 503.825858] other info that might help us debug this: [ 503.825858] [ 503.833998] [ 503.833998] rcu_scheduler_active = 2, debug_locks = 1 [ 503.840981] 3 locks held by kworker/u12:1/157: [ 503.845514] #0: ffff0009754ed538 ((wq_completion)netns){+.+.}-{0:0}, at: process_one_work+0x208/0x768 [ 503.855048] #1: ffff800013e63df0 (net_cleanup_work){+.+.}-{0:0}, at: process_one_work+0x208/0x768 [ 503.864201] #2: ffff8000129fe3f0 (pernet_ops_rwsem){++++}-{3:3}, at: cleanup_net+0x64/0x3b8 [ 503.872786] [ 503.872786] stack backtrace: [ 503.877229] CPU: 1 PID: 157 Comm: kworker/u12:1 Tainted: G W 5.9.15-rc1 #1 [ 503.885433] Hardware name: ARM Juno development board (r2) (DT) [ 503.891382] Workqueue: netns cleanup_net [ 503.895324] Call trace: [ 503.897786] dump_backtrace+0x0/0x1f8 [ 503.901464] show_stack+0x2c/0x38 [ 503.904796] dump_stack+0xec/0x158 [ 503.908215] lockdep_rcu_suspicious+0xd4/0xf8 [ 503.912591] ___might_sleep+0x1e4/0x208
You really are forbidden to invoke ___might_sleep() while in a BH-disable region of code, whether due to rcu_read_lock_bh(), local_bh_disable(), or whatever else.
I do see the cond_resched() in inet_twsk_purge(), but I don't immediately see a BH-disable region of code. Maybe someone more familiar with this code would have some ideas.
Or you could place checks for being in a BH-disable further up in the code. Or build with CONFIG_DEBUG_INFO=y to allow more precise interpretation of this stack trace.
I will try to reproduce this warning with DEBUG_INFO=y enabled kernel and get back to you with a better crash log.
My money would be on the option that whatever run on this workqueue before forgot to re-enable BH, but we already have a check for that... Naresh, do you have the full log? Is there nothing like "BUG: workqueue leaked lock" above the splat?
Yes [1] is the full test log link. But i do not see "BUG: workqueue leaked lock" in the log.
full log link, [1] https://lkft.validation.linaro.org/scheduler/job/2049484#L5979
- Naresh