On Tue, Dec 08, 2020 at 03:57:44PM -0800, Jakub Kicinski wrote: > On Mon, 7 Dec 2020 01:00:40 +0000 Vladimir Oltean wrote: > > - ensuring through convention that user space always takes > > net->netdev_lists_lock when calling dev_get_stats, and documenting > > that, and therefore making it unnecessary to lock in bonding. > > This seems like the better option to me. Makes the locking rules pretty > clear.
It is non-obvious to me that top-level callers of dev_get_stats should hold a lock as specific as the one protecting the lists of network interfaces. In the vast majority of implementations of dev_get_stats, that lock would not actually protect anything, which would lead into just one more lock that is used for more than it should be. In my tree I had actually already switched over to mutex_lock_nested. Nonetheless, I am still open if you want to make the case that simplicity should prevail over specificity. But in that case, maybe we should just keep on using the RTNL mutex.
