Nikolay Aleksandrov nikolay@nvidia.com wrote:
On 28/07/2021 09:19, Yufeng Mo wrote:
Some time ago, I reported a calltrace issue "did not find a suitable aggregator", please see[1]. After a period of analysis and reproduction, I find that this problem is caused by concurrency.
Before the problem occurs, the bond structure is like follows:
bond0 - slaver0(eth0) - agg0.lag_ports -> port0 - port1 \ port0 \ slaver1(eth1) - agg1.lag_ports -> NULL \ port1
If we run 'ifenslave bond0 -d eth1', the process is like below:
excuting __bond_release_one() | bond_upper_dev_unlink()[step1] | | | | | bond_3ad_lacpdu_recv() | | ->bond_3ad_rx_indication() | | spin_lock_bh() | | ->ad_rx_machine() | | ->__record_pdu()[step2] | | spin_unlock_bh() | | | | bond_3ad_state_machine_handler() | spin_lock_bh() | ->ad_port_selection_logic() | ->try to find free aggregator[step3] | ->try to find suitable aggregator[step4] | ->did not find a suitable aggregator[step5] | spin_unlock_bh() | | | | bond_3ad_unbind_slave() | spin_lock_bh() spin_unlock_bh()
step1: already removed slaver1(eth1) from list, but port1 remains step2: receive a lacpdu and update port0 step3: port0 will be removed from agg0.lag_ports. The struct is "agg0.lag_ports -> port1" now, and agg0 is not free. At the same time, slaver1/agg1 has been removed from the list by step1. So we can't find a free aggregator now. step4: can't find suitable aggregator because of step2 step5: cause a calltrace since port->aggregator is NULL
To solve this concurrency problem, the range of bond->mode_lock is extended from only bond_3ad_unbind_slave() to both bond_upper_dev_unlink() and bond_3ad_unbind_slave().
[1]https://lore.kernel.org/netdev/10374.1611947473@famine/
Signed-off-by: Yufeng Mo moyufeng@huawei.com Acked-by: Jay Vosburgh jay.vosburgh@canonical.com
drivers/net/bonding/bond_3ad.c | 7 +------ drivers/net/bonding/bond_main.c | 6 +++++- 2 files changed, 6 insertions(+), 7 deletions(-)
[snip]
/** diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c index 0ff7567..deb019e 100644 --- a/drivers/net/bonding/bond_main.c +++ b/drivers/net/bonding/bond_main.c @@ -2129,14 +2129,18 @@ static int __bond_release_one(struct net_device *bond_dev, /* recompute stats just before removing the slave */ bond_get_stats(bond->dev, &bond->bond_stats);
- bond_upper_dev_unlink(bond, slave); /* unregister rx_handler early so bond_handle_frame wouldn't be called
*/ netdev_rx_handler_unregister(slave_dev);
- for this slave anymore.
- /* Sync against bond_3ad_state_machine_handler() */
- spin_lock_bh(&bond->mode_lock);
- bond_upper_dev_unlink(bond, slave);
this calls netdev_upper_dev_unlink() which calls call_netdevice_notifiers_info() for NETDEV_PRECHANGEUPPER and NETDEV_CHANGEUPPER, both of which are allowed to sleep so you cannot hold the mode lock
Indeed it does, I missed that the callbacks can sleep.
after netdev_rx_handler_unregister() the bond's recv_probe cannot be executed so you don't really need to unlink it under mode_lock or move mode_lock at all
I don't think moving the call to netdev_rx_handler_unregister is sufficient to close the race. If it's moved above the call to bond_upper_dev_unlink, the probe won't be called afterwards, but the LACPDU could have arrived just prior to the unregister and changed the port state in the bond_3ad_lacpdu_recv call sequence ("step 2", something in the LACPDU causes AD_PORT_SELECTED to be cleared). Later, bond_3ad_state_machine_handler runs in a separate work queue context, and could process the effect of the LACPDU after the rx_handler unregister, and still race with the upper_dev_unlink.
I suspect the solution is to rework ad_port_selection_logic to correctly handle the situation where no aggregator is available. Off the top of my head, I think something along the lines of:
diff --git a/drivers/net/bonding/bond_3ad.c b/drivers/net/bonding/bond_3ad.c index 6908822d9773..eb6223e4510e 100644 --- a/drivers/net/bonding/bond_3ad.c +++ b/drivers/net/bonding/bond_3ad.c @@ -1537,6 +1537,10 @@ static void ad_port_selection_logic(struct port *port, bool *update_slave_arr) slave_err(bond->dev, port->slave->dev, "Port %d did not find a suitable aggregator\n", port->actor_port_number); + aggregator = __get_first_agg(port); + ad_agg_selection_logic(aggregator, update_slave_arr); + + return; } } /* if all aggregator's ports are READY_N == TRUE, set ready=TRUE
I've not compiled or tested this, but the theory is that it will reselect a new aggregator for the bond (which happens anyway later in the function), then returns, leaving "port" as not AD_PORT_SELECTED. The next run of the state machine should attempt to select it again, and presumably succeed at that time.
This may leave the bond with no active ports for one interval between runs of the state machine, unfortunately, but it should eliminate the panic.
Another possibility might be netdev_rx_handler_unregister, then bond_3ad_unbind_slave, and finally bond_upper_dev_unlink, but I'm not sure right off if that would have other side effects.
Yufeng, would you be able to test the above and see if it resolves the issue in your test?
-J
if (BOND_MODE(bond) == BOND_MODE_8023AD) bond_3ad_unbind_slave(slave);
spin_unlock_bh(&bond->mode_lock);
if (bond_mode_can_use_xmit_hash(bond)) bond_update_slave_arr(bond, slave);
--- -Jay Vosburgh, jay.vosburgh@canonical.com