Hello, On Thu, Feb 13, 2020 at 10:46:34AM +0800, Yufen Yu wrote: > For each time of register, bdi_register() will try to create a new 'dev'. > > bdi_register > bdi_register_va > if (bdi->dev) // if bdi->dev is not NULL, return directly > return 0; > dev = device_create_vargs()... > > So, I think freeing bdi->dev until bdi itself does may be a problem > for drivers that supported re-registration bdi, such as: Ugh, thanks for noticing that. I guess the right thing to do is then going full RCU. What do you think about expanding your previous patch so that ->dev has __rcu annotation, users use the RCU accessors and the device is destroyed asynchronously through call_rcu()? Thanks. -- tejun
