On Fri, May 15, 2015 at 01:02:19PM -0400, David Miller wrote:

Actually, Herbert, I think this lock_sock() is needed.

Otherwise nothing protects nlk_sk(sk)->portid, upon which we
perform a non-atomic test-and-set operation here.

If you remove the lock_sock(), two parallel bind/inserts are
possible on the same socket, potentially resulting in socket
state corruption.

You're quite right.  I forgot about that case.

However, the code as is still buggy because it fails to reset
portid when our auto-allocated portid collides with somebody
else, resulting in a bogus bind failure.

I'll fix that up too in a new patch.

Cheers,
-- 
Email: Herbert Xu [off-list ref]
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt