4.4-stable review patch. If anyone has any objections, please let me know.
------------------
From: Hangbin Liu liuhangbin@gmail.com
[ Upstream commit b5f862180d7011d9575d0499fa37f0f25b423b12 ]
After commit 07f4c90062f8 ("tcp/dccp: try to not exhaust ip_local_port_range in connect()"), we will try to use even ports for connect(). Then if an application (seen clearly with iperf) opens multiple streams to the same destination IP and port, each stream will be given an even source port.
So the bonding driver's simple xmit_hash_policy based on layer3+4 addressing will always hash all these streams to the same interface. And the total throughput will limited to a single slave.
Change the tcp code will impact the whole tcp behavior, only for bonding usage. Paolo Abeni suggested fix this by changing the bonding code only, which should be more reasonable, and less impact.
Fix this by discarding the lowest hash bit because it contains little entropy. After the fix we can re-balance between slaves.
Signed-off-by: Paolo Abeni pabeni@redhat.com Signed-off-by: Hangbin Liu liuhangbin@gmail.com Signed-off-by: David S. Miller davem@davemloft.net Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org --- drivers/net/bonding/bond_main.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
--- a/drivers/net/bonding/bond_main.c +++ b/drivers/net/bonding/bond_main.c @@ -3166,7 +3166,7 @@ u32 bond_xmit_hash(struct bonding *bond, hash ^= (hash >> 16); hash ^= (hash >> 8);
- return hash; + return hash >> 1; }
/*-------------------------- Device entry points ----------------------------*/