jin yiting <jinyiting@xxxxxxxxxx> wrote:
[...]
The described issue is a race condition (in that
ad_agg_selection_logic clears agg->is_active under mode_lock, but
bond_open -> bond_update_slave_arr is inspecting agg->is_active outside
the lock). I don't see how the above change will reliably manage this;
the real issue looks to be that bond_update_slave_arr is committing
changes to the array (via bond_reset_slave_arr) based on a racy
inspection of the active aggregator state while it is in flux.
Also, the description of the issue says "The best aggregator in
ad_agg_selection_logic has not changed, no need to update slave arr,"
but the change above does the opposite, and will set update_slave_arr
when the aggregator has not changed (update_slave_arr remains false at
return of ad_agg_selection_logic).
I believe I understand the described problem, but I don't see
how the patch fixes it. I suspect (but haven't tested) that the proper
fix is to acquire mode_lock in bond_update_slave_arr while calling
bond_3ad_get_active_agg_info to avoid conflict with the state machine.
-J
---
-Jay Vosburgh, jay.vosburgh@xxxxxxxxxxxxx
.
Thank you for your reply. The last patch does have redundant
update slave arr.Thank you for your correction.
As you said, holding mode_lock in bond_update_slave_arr while
calling bond_3ad_get_active_agg_info can avoid conflictwith the state
machine. I have tested this patch, with ifdown/ifup operations for bond or
slaves.
But bond_update_slave_arr is expected to hold RTNL only and NO
other lock. And it have WARN_ON(lockdep_is_held(&bond->mode_lock)); in
bond_update_slave_arr. I'm not sure that holding mode_lock in
bond_update_slave_arr while calling bond_3ad_get_active_agg_info is a
correct action.
That WARN_ON came up in discussion recently, and my opinion is
that it's incorrect, and is trying to insure bond_update_slave_arr is
safe for a potential sleep when allocating memory.
https://lore.kernel.org/netdev/20210322123846.3024549-1-maximmi@xxxxxxxxxx/
The original authors haven't replied, so I would suggest you
remove the WARN_ON and the surrounding CONFIG_LOCKDEP ifdefs as part of
your patch and replace it with a call to might_sleep.
The other callers of bond_3ad_get_active_agg_info are generally
obtaining the state in order to report it to user space, so I think it's
safe to leave those calls not holding the mode_lock. The race is still
there, but the data returned to user space is a snapshot and so may
reflect an incomplete state during a transition. Further, having the
inspection functions acquire the mode_lock permits user space to spam
the lock with little effort.
-J
diff --git a/drivers/net/bonding/bond_main.c
b/drivers/net/bonding/bond_main.c
index 74cbbb2..db988e5 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -4406,7 +4406,9 @@ int bond_update_slave_arr(struct bonding *bond,
struct slave *skipslave)
if (BOND_MODE(bond) == BOND_MODE_8023AD) {
struct ad_info ad_info;
+ spin_lock_bh(&bond->mode_lock);
if (bond_3ad_get_active_agg_info(bond, &ad_info)) {
+ spin_unlock_bh(&bond->mode_lock);
pr_debug("bond_3ad_get_active_agg_info failed\n");
/* No active aggragator means it's not safe to use
* the previous array.
@@ -4414,6 +4416,7 @@ int bond_update_slave_arr(struct bonding *bond,
struct slave *skipslave)
bond_reset_slave_arr(bond);
goto out;
}
+ spin_unlock_bh(&bond->mode_lock);
agg_id = ad_info.aggregator_id;
}
bond_for_each_slave(bond, slave, iter) {
---
-Jay Vosburgh, jay.vosburgh@xxxxxxxxxxxxx
.