bonding: discard lowest hash bit for 802.3ad layer3+4
authorHangbin Liu <liuhangbin@gmail.com>
Mon, 6 Nov 2017 01:01:57 +0000 (09:01 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 24 Nov 2017 07:33:40 +0000 (08:33 +0100)
commit6f239c0655cfe90c3e048bccb2e74ab1780cabba
tree95c4a3360bb0cb96a971e827b12b45e15231dd3d
parentafd9fa6619277fe460f9ad7949fd90d0985aec78
bonding: discard lowest hash bit for 802.3ad layer3+4

[ 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