Re: [RFC] net: remove erroneous sk null assignment in timestamping

[Richard Cochran <richardcochran@gmail.com>]

On Sat, Oct 08, 2011 at 12:32:15PM +0200, Johannes Berg wrote:
> On Sat, 2011-10-08 at 10:57 +0200, Eric Dumazet wrote:
> > Check following commit changelog to get some information on this.
> > 
> > commit 2b85a34e911bf483c27cfdd124aeb1605145dc80
> > Author: Eric Dumazet <eric.dumazet@gmail.com>
> > Date:   Thu Jun 11 02:55:43 2009 -0700
> > 
> >     net: No more expensive sock_hold()/sock_put() on each tx
...
> There's one thing I still miss though: It seems to me that if you have a
> reference to a socket that has been sk_free()'ed (which is possible
> since it might still have sk_wmem_alloc > 0) you can't sock_hold() that
> socket. That feels a bit unexpected -- and might happen in the code
> Richard just suggested.

Yes, I have been trying to see how to solve this, but it looks like I
am out of luck. 

Even if I use skb_set_owner_w() in skb_clone_tx_timestamp(), still the
sock might go away during skb_orphan() in sock_queue_err_skb().

It is no good to take sock_hold() in skb_complete_tx_timestamp(),
since, as you point out, it might not be safe to call.

So, I wonder, when is it safe to call sock_hold?

Are the 101 odd callers protected against the situation where the last
sock_out() has already happened?

Thanks,

Richard