On 08/12/20 13:46, Qian Cai wrote: > On Mon, 2020-12-07 at 19:27 +0000, Valentin Schneider wrote: >> Ok, can reproduce this on a TX2 on next-20201207. I didn't use your config, >> I oldconfig'd my distro config and only modified it to CONFIG_PREEMPT_NONE. >> Interestingly the BUG happens on CPU127 here too... > > I think that number is totally random. For example, on this x86, it could > happen > for CPU8 or CPU111.
Actually on the TX2 it seems to *always* happen on CPU127. Your hotplug script sequentially offlines CPUs in increasing id values, so when CPU127 gets hotplugged it is the last online CPU of NUMA node 0. I've been staring at traces collected via echo 2 > /proc/sys/kernel/ftrace_dump_on_oops trace-cmd start -e 'sched:*' -e 'cpuhp:*' -e 'workqueue:*' ./hotplug.sh but it's still not entirely clear to me WTH is going on. I do see kworkers getting their affinity reset in workqueue_offline_cpu(), but for some reason there's a new one that wakes up on CPU127 sometime later. I haven't been able to figure out where it comes from - it obviously isn't part of the percpu worker pools, as it isn't handled during workqueue_offline_cpu(), but it still ends up affined to a single CPU... It looks something like this; traces are only from CPU127 cpuhp:sched_cpu_wait_empty() # Resets the affinity of some kworker/127:x<2 sched_switch(idle) sched_wakeup(kworker/127:2) # picks CPU127 sched_switch(kworker/127:2) # maybe_create_worker() -> creates kworker/127:3 sched_wakeup(kworker/127:3) # picks CPU127 sched_switch(kworker/127:3) # maybe_create_worker() -> creates kworker/127:4 sched_wakeup(kworker/127:4) # picks CPU127 sched_switch(kworker/127:4) # maybe_create_worker() -> creates kworker/127:5 sched_wakeup(kworker/127:5) # picks CPU127 sched_wakeup(migration/127) sched_switch(migration/127) cpuhp:take_cpu_down() BUG
