On Tue, Jun 30, 2020 at 07:17:46PM -0700, Eric Dumazet wrote:

The main issue of the prior code was the double read of key->keylen in
tcp_md5_hash_key(), not that few bytes could change under us.

I used smp_rmb() to ease backports, since old kernels had no
READ_ONCE()/WRITE_ONCE(), but ACCESS_ONCE() instead.

If it's the double-read that you're protecting against, you should
just use barrier() and the comment should say so too.

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