On Friday, June 06/26/20, 2020 at 21:18:44 -0700, Jakub Kicinski wrote:

On Thu, 25 Jun 2020 17:28:41 +0530 Rahul Lakkireddy wrote:

quoted

+	if (refcount_read(&pi->vi_mirror_refcnt) > 1) {
+		refcount_dec(&pi->vi_mirror_refcnt);
+		return;
+	}

FWIW this looks very dodgy. If you know nothing changes the count
between the read and the dec here, you probably don't need atomic
refcounts at all..

Currently, all the callers accessing this refcount and its related
data is having the RTNL lock held by the stack. Perhaps this is a
false sense of security, especially if the stack API may change in
the future.

I'll add a proper lock to protect this data in v2 to be on the safer
side.

Thanks,
Rahul