On 07/31/2013 08:56 PM, Stephen Hemminger wrote:

On Wed, 31 Jul 2013 20:44:49 +0200
Nikolay Aleksandrov [off-list ref] wrote:

quoted

On 07/31/2013 08:37 PM, Stephen Hemminger wrote:

quoted

On Wed, 31 Jul 2013 17:12:29 +0200
Nikolay Aleksandrov [off-list ref] wrote:

quoted

diff --git a/drivers/net/bonding/bond_3ad.c b/drivers/net/bonding/bond_3ad.c
index 390061d..80e288c 100644
--- a/drivers/net/bonding/bond_3ad.c
+++ b/drivers/net/bonding/bond_3ad.c

@@ -143,10 +143,13 @@ static inline struct bonding *__get_bond_by_port(struct port *port)
  */
 static inline struct port *__get_first_port(struct bonding *bond)
 {
+	struct slave *first_slave;
+
 	if (bond->slave_cnt == 0)
 		return NULL;
+	first_slave = bond_first_slave(bond);

As Jay said, it would be be better to have bond_first_slave return
NULL (if no slaves), and eliminate slave_cnt. It would also fix
a race here between slave_cnt and all slave's being removed.

Hi Stephen,
First off - thank you for the review.

What do you mean by eliminate slave_cnt ?
We need it for various calculations throughout the bonding.
There's no race here because read_lock(&bond->lock) is held every time this
is called and slave_cnt can change only under write_lock of the same lock,
the 3ad code is not yet converted to RCU.

Nik

I would hope the goal would be to eliminate all read_lock's and allow
it to be totally RCU based.

You could then reduce slave_cnt to being only accessed by under a spin_lock
when doing management actions.

Yes, that is the end goal. And my end-implementation does just that -
removes curr_slave_lock completely and makes bond->lock a spinlock. But for
now we need it for the parts that aren't converted, as I said in the
beginning I'm doing the conversion gradually and there will be a special
series that takes care of 3ad mode as there are many potential problems for
RCU there because of its current design.