bonding: Don't update slave->link until ready to commit
authorNithin Sujir <nsujir@tintri.com>
Thu, 25 May 2017 02:45:17 +0000 (19:45 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 25 May 2017 18:47:35 +0000 (14:47 -0400)
commit797a93647a48d6cb8a20641a86a71713a947f786
tree834604b4c92ec171f1a6a2c12e4ef33f26fa10a1
parent791caeb084c57e3a4d648cf1ee799d1f70c0ef4e
bonding: Don't update slave->link until ready to commit

In the loadbalance arp monitoring scheme, when a slave link change is
detected, the slave->link is immediately updated and slave_state_changed
is set. Later down the function, the rtnl_lock is acquired and the
changes are committed, updating the bond link state.

However, the acquisition of the rtnl_lock can fail. The next time the
monitor runs, since slave->link is already updated, it determines that
link is unchanged. This results in the bond link state permanently out
of sync with the slave link.

This patch modifies bond_loadbalance_arp_mon() to handle link changes
identical to bond_ab_arp_{inspect/commit}(). The new link state is
maintained in slave->new_link until we're ready to commit at which point
it's copied into slave->link.

NOTE: miimon_{inspect/commit}() has a more complex state machine
requiring the use of the bond_{propose,commit}_link_state() functions
which maintains the intermediate state in slave->link_new_state. The arp
monitors don't require that.

Testing: This bug is very easy to reproduce with the following steps.
1. In a loop, toggle a slave link of a bond slave interface.
2. In a separate loop, do ifconfig up/down of an unrelated interface to
create contention for rtnl_lock.
Within a few iterations, the bond link goes out of sync with the slave
link.

Signed-off-by: Nithin Nayak Sujir <nsujir@tintri.com>
Cc: Mahesh Bandewar <maheshb@google.com>
Cc: Jay Vosburgh <jay.vosburgh@canonical.com>
Acked-by: Mahesh Bandewar <maheshb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/bonding/bond_main.c