Second, bond_change_active_slave() is called on few occasions under write_lock_bh()
which means it runs in atomic context. I was thinking that it might be better to
invoke call_netdevice_notifiers() from thread context,

Sorry if I wasn't clear. As seen from the trace, current code indeed delivers the event
from a thread but the context is made atomic as of the _bh locking, so my question is whether
the context should be made not atomic before delivering the event (eg through queuing a work
struct to be executed later by the kernel work queue thread, etc)

Or

Call Trace:
 [<ffffffff804777d7>] schedule+0x98/0x57b
 [<ffffffff80277836>] dbg_redzone1+0x16/0x1f
 [<ffffffffa0106f22>] :ib_ipoib:ipoib_start_xmit+0x445/0x459
 [<ffffffff802799c2>] kmem_cache_alloc_node+0x147/0x177
 [<ffffffff8040a939>] __alloc_skb+0x35/0x12b
 [<ffffffff8022c99b>] __cond_resched+0x1c/0x43
 [<ffffffff80477e11>] _cond_resched+0x2d/0x38
 [<ffffffff802798a0>] kmem_cache_alloc_node+0x25/0x177
 [<ffffffff8040a939>] __alloc_skb+0x35/0x12b
 [<ffffffff8041825e>] rtmsg_ifinfo+0x3a/0xd4
 [<ffffffff80418335>] rtnetlink_event+0x3d/0x41
 [<ffffffff8047b925>] notifier_call_chain+0x30/0x54
 [<ffffffffa00a3d4b>] :bonding:bond_select_active_slave+0xb9/0xe8
 [<ffffffffa00a495e>] :bonding:__bond_mii_monitor+0x43a/0x464
 [<ffffffffa00a49e6>] :bonding:bond_mii_monitor+0x5e/0xaa
 [<ffffffffa00a4988>] :bonding:bond_mii_monitor+0x0/0xaa
 [<ffffffff8023d6fa>] run_workqueue+0x7f/0x107
 [<ffffffff8023d782>] worker_thread+0x0/0xef
 [<ffffffff8023d867>] worker_thread+0xe5/0xef
 [<ffffffff8024088f>] autoremove_wake_function+0x0/0x2e
 [<ffffffff8024088f>] autoremove_wake_function+0x0/0x2e
 [<ffffffff8024055a>] kthread+0x3d/0x63
 [<ffffffff8020c068>] child_rip+0xa/0x12
 [<ffffffff8024051d>] kthread+0x0/0x63
 [<ffffffff8020c05e>] child_rip+0x0/0x12