diff options
author | nikolay@redhat.com <nikolay@redhat.com> | 2013-08-01 16:54:51 +0200 |
---|---|---|
committer | David S. Miller <davem@davemloft.net> | 2013-08-01 16:42:02 -0700 |
commit | 278b20837511776dc9d5f6ee1c7fabd5479838bb (patch) | |
tree | 6be201bb67d28154444b1860d6c1247d02adab23 /drivers/net/bonding/bond_alb.c | |
parent | 15077228cab68e5e8c3cbf26a7f6ebacfac4c829 (diff) | |
download | linux-278b20837511776dc9d5f6ee1c7fabd5479838bb.tar.gz linux-278b20837511776dc9d5f6ee1c7fabd5479838bb.tar.bz2 linux-278b20837511776dc9d5f6ee1c7fabd5479838bb.zip |
bonding: initial RCU conversion
This patch does the initial bonding conversion to RCU. After it the
following modes are protected by RCU alone: roundrobin, active-backup,
broadcast and xor. Modes ALB/TLB and 3ad still acquire bond->lock for
reading, and will be dealt with later. curr_active_slave needs to be
dereferenced via rcu in the converted modes because the only thing
protecting the slave after this patch is rcu_read_lock, so we need the
proper barrier for weakly ordered archs and to make sure we don't have
stale pointer. It's not tagged with __rcu yet because there's still work
to be done to remove the curr_slave_lock, so sparse will complain when
rcu_assign_pointer and rcu_dereference are used, but the alternative to use
rcu_dereference_protected would've created much bigger code churn which is
more difficult to test and review. That will be converted in time.
1. Active-backup mode
1.1 Perf recording while doing iperf -P 4
- old bonding: iperf spent 0.55% in bonding, system spent 0.29% CPU
in bonding
- new bonding: iperf spent 0.29% in bonding, system spent 0.15% CPU
in bonding
1.2. Bandwidth measurements
- old bonding: 16.1 gbps consistently
- new bonding: 17.5 gbps consistently
2. Round-robin mode
2.1 Perf recording while doing iperf -P 4
- old bonding: iperf spent 0.51% in bonding, system spent 0.24% CPU
in bonding
- new bonding: iperf spent 0.16% in bonding, system spent 0.11% CPU
in bonding
2.2 Bandwidth measurements
- old bonding: 8 gbps (variable due to packet reorderings)
- new bonding: 10 gbps (variable due to packet reorderings)
Of course the latency has improved in all converted modes, and moreover
while
doing enslave/release (since it doesn't affect tx anymore).
Also I've stress tested all modes doing enslave/release in a loop while
transmitting traffic.
Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'drivers/net/bonding/bond_alb.c')
-rw-r--r-- | drivers/net/bonding/bond_alb.c | 6 |
1 files changed, 4 insertions, 2 deletions
diff --git a/drivers/net/bonding/bond_alb.c b/drivers/net/bonding/bond_alb.c index 4d35196b4b5a..3a5db7b1df68 100644 --- a/drivers/net/bonding/bond_alb.c +++ b/drivers/net/bonding/bond_alb.c @@ -1337,6 +1337,7 @@ int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev) /* make sure that the curr_active_slave do not change during tx */ + read_lock(&bond->lock); read_lock(&bond->curr_slave_lock); switch (ntohs(skb->protocol)) { @@ -1441,11 +1442,12 @@ int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev) } read_unlock(&bond->curr_slave_lock); - + read_unlock(&bond->lock); if (res) { /* no suitable interface, frame not sent */ kfree_skb(skb); } + return NETDEV_TX_OK; } @@ -1663,7 +1665,7 @@ void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave } swap_slave = bond->curr_active_slave; - bond->curr_active_slave = new_slave; + rcu_assign_pointer(bond->curr_active_slave, new_slave); if (!new_slave || list_empty(&bond->slave_list)) return; |