On 07/31/2017 01:55 AM, Daniel Borkmann wrote: > On 07/30/2017 03:28 PM, Levin, Alexander (Sasha Levin) wrote: >> On Mon, Jul 17, 2017 at 09:30:02AM -0700, John Fastabend wrote: >>> @@ -341,9 +368,11 @@ static int dev_map_update_elem(struct bpf_map *map, >>> void *key, void *value, >>> * Remembering the driver side flush operation will happen before the >>> * net device is removed. >>> */ >>> + mutex_lock(&dev_map_list_mutex); >>> old_dev = xchg(&dtab->netdev_map[i], dev); >>> if (old_dev) >>> call_rcu(&old_dev->rcu, __dev_map_entry_free); >>> + mutex_unlock(&dev_map_list_mutex); >>> >>> return 0; >>> } >> >> This function gets called under rcu critical section, where we can't grab >> mutexes: > > Agree, same goes for the delete callback that mutex is not allowed > in this context. If I recall, this was for the devmap netdev notifier > in order to check whether we need to purge dev entries from the map, > so that the device can be unregistered gracefully. Given that devmap > ops like update/delete are only allowed from user space, we could > look into whether this map type actually needs to hold RCU at all > here, or other option is to try and get rid of the mutex altogether. > John, could you take a look for a fix? > > Thanks a lot, > Daniel >
I'll work up a fix today/tomorrow. Thanks.
