On Fri, Oct 16, 2020 at 07:11:22PM +0200, Mike Galbraith wrote:
> Yeah, it's a straight up correctness issue as it sits.
Yeah, tell me about it, you don't even want to know what it looks like
when the local_softirq_pending_ref percpu mask gets corrupted. The
symptom will be that random syscalls, like clock_nanosleep, will make
user processes hang in the kernel, because their timers will appear to
never expire. You'll also see negative expiry times in /proc/timer_list.
Eventually the entire system dies. All of that, reproducible with a
simple flood ping, given enough time. Ask me how I come to know....
The 215602a8d212 ("enetc: use napi_schedule to be compatible with
PREEMPT_RT") commit is from before the lockdep warning came to be.
