On 29.05.2013 09:08, Eric Dumazet wrote:

On Tue, 2013-05-28 at 18:31 -0700, Paul E. McKenney wrote:

quoted

On Tue, May 28, 2013 at 05:34:53PM -0700, Eric Dumazet wrote:

quoted

On Tue, 2013-05-28 at 13:10 +0400, Roman Gushchin wrote:

quoted

On 28.05.2013 04:12, Eric Dumazet wrote:

About your earlier question, I really don't know why compiler would
cache a memory read if we explicitly use barrier() to prevent this from
happening.

I agree.

BTW Roman patch generates a double load as in :

2cb1:       49 8b 07                mov    (%r15),%rax
2cb4:       49 8b 07                mov    (%r15),%rax


...
2ea2:       e8 f9 dc ff ff          callq  ba0 <sock_put>
2ea7:       8b 0c 24                mov    (%rsp),%ecx
2eaa:       e9 02 fe ff ff          jmpq   2cb1 <udp4_lib_lookup2+0x91>

because of ACCESS_ONCE() used twice, once explicitly in
hlist_nulls_for_each_entry_rcu(), and once in rcu_dereference_raw()

While barrier();ptr = rcu_dereference(X); does generate a single load.

It's true.

Unfortunately, using barrier() can also prevent gcc from making some
(acceptable) code optimizations, because barrier() has a global effect,
and everything we want to reload is the (head->first) pointer.
So, to be absolutely precise, we have to introduce and use
the ACCESS_FIELD_ONCE() macro.

In any case, it doesn't look like a big problem.
In my mind, the best solution is to use the ACCESS_FIELD_ONCE() macro,
but using barrier() is also an acceptable solution.

Regards,
Roman