Merge branch 'bonding_rcu'
Ding Tianhong says: ==================== bonding: rebuild the lock use for bond monitor Now the bond slave list is not protected by bond lock, only by RTNL, but the monitor still use the bond lock to protect the slave list, it is useless, according to the Veaceslav's opinion, there were three way to fix the protect problem: 1. add bond_master_upper_dev_link() and bond_upper_dev_unlink() in bond->lock, but it is unsafe to call call_netdevice_notifiers() in write lock. 2. remove unused bond->lock for monitor function, only use the exist rtnl lock(), it will take performance loss in fast path. 3. use RCU to protect the slave list, of course, performance is better, but in slow path, it is ignored. obviously the solution 1 is not fit here, I will consider the 2 and 3 solution. My principle is simple, if in fast path, RCU is better, otherwise in slow path, both is well, but according to the Jay Vosburgh's opinion, the monitor will loss performace if use RTNL to protect the all slave list, so remove the bond lock and replace with RCU. The second problem is the curr_slave_lock for bond, it is too old and unwanted in many place, because the curr_active_slave would only be changed in 3 place: 1. enslave slave. 2. release slave. 3. change active slave. all above were already holding bond lock, RTNL and curr_slave_lock together, it is tedious and no need to add so mach lock, when change the curr_active_slave, you have to hold the RTNL and curr_slave_lock together, and when you read the curr_active_slave, RTNL or curr_slave_lock, any one of them is no problem. for the stability, I did not change the logic for the monitor, all change is clear and simple, I have test the patch set for lockdep, it work well and stability. v2. accept the Jay Vosburgh's opinion, remove the RTNL and replace with RCU, also add some rcu function for bond use, so the patch set reach 10. v3. accept the Nikolay Aleksandrov's opinion, remove no needed bond_has_slave_rcu(), add protection for several 3ad mode handler functions and current_arp_slave. rebuild the bond_first_slave_rcu(), make it more clear. v4. because the struct netdev_adjacent should not be exist in netdevice.h, so I have to make a new function to support micro bond_first_slave_rcu(). also add a new patch to simplify the bond_resend_igmp_join_requests_delayed(). v5. according the Jay Vosburgh's opinion, in patch 2 and 6, the calling of notify peer is hardly to happen with the bond_xxx_commit() when the monitoring is running, so the performance impact about make two round trips to one trip on RTNL is minimal, no need to do that,the reason is very clear, so modify the patch 2 and 6, recover the notify peer in RTNL alone. ==================== Signed-off-by: Jay Vosburgh <fubar@us.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Showing
Please register or sign in to comment