Just decided to take another look:
On Mon, 2017-10-09 at 13:58 +0200, Johannes Berg wrote:
> On Mon, 2017-10-09 at 13:56 +0200, Jason A. Donenfeld wrote:
>
> > @@ -2266,16 +2266,17 @@ int __netlink_dump_start(struct sock *ssk,
> > struct sk_buff *skb,
> > cb->min_dump_alloc = control->min_dump_alloc;
> > cb->skb = skb;
> >
> > + if (cb->start) {
> > + ret = cb->start(cb);
> > + if (ret)
> > + goto error_unlock;
> > + }
> > +
> > nlk->cb_running = true;
> >
> > mutex_unlock(nlk->cb_mutex);
>
> Hmm. Now start is invoked with the mutex held, I'm not sure it
> actually _matters_, but that should probably be reviewed and
> mentioned in the commit log?
It sort of seems designed to run ->start outside the lock, otherwise we
wouldn't really have to acquire it again in netlink_dump() but could
just keep it across the call (with some locking changes in
netlink_recvmsg())?
Then again, clearly none of the (few) existing users actually care.
Btw - we should (separately) also remove "start" from struct
netlink_callback, it's only ever used within this function and we can
use control->start instead of cb->start here.
johannes
