On 2020/2/13 21:58, Tejun Heo wrote:
Hello,
On Thu, Feb 13, 2020 at 03:51:40PM +0800, Yufen Yu wrote:
If we destroy the device asynchronously by call_rcu(), we may need to
add a new member 'rcu_head' into struct backing_dev_info. Right?
Yes.
The code may be like:
bdi_unregister()
{
...
if (bdi->dev) {
...
device_get(bdi->dev);
device_unregister(bdi->dev);
bdi->dev = NULL; //XXX
bdi_get(bdi); //avoiding bdi to be freed before calling bdi_release_device
call_rcu(&bdi->rcu_head, bdi_release_device);
}
...
}
bdi_release_device()
{
...
put_device(bdi->dev);//XXX
bdi_put(bdi);
}
But, the problem is how do we get 'bdi->dev' in bdi_release_device().
If we do not set bdi->dev as 'NULL', re-registration bdi may cannot work well.
So, unregistering can leave ->dev along and re-registering can test
whether it's NULL and if not put the existing one and put a new one
there. Wouldn't that work?
Do you mean set bdi->dev as 'NULL' in call_rcu() callback function
(i.e. bdi_release_device()) and test 'bdi->dev' in bdi_register_va()?
I think that may do not work.
We cannot make sure the order of rcu callback function and re-registering.
Then bdi_release_device() may put the new allocated device by re-registering.
Thanks,
Yufen
