On 4/15/19 7:36 AM, Tetsuo Handa wrote:
> I traced using debug printk() patch shown below.
> 

I find tracepoints (see attached patch) and perf are easier to use to
debug device refcnt problems.

For example, limit the stack you have to deal with via sysctl -w
kernel.perf_event_max_stack=16, and add a filter (e.g., --filter 'name
== "lo"') to limit collection to a specific device.