The mqprio path of ice_setup_tc() takes pf->adev_mutex and the auxiliary
device lock first and pf->tc_mutex after them. The dcbnl handlers take
pf->tc_mutex first and reach pf->adev_mutex through ice_pf_dcb_cfg() and
ice_send_event_to_aux(). Both run with the RTNL lock held, so the two
orders do not deadlock against each other, but they form a circular
dependency for lockdep, and the DCB reconfiguration is about to hold
pf->adev_mutex for longer.
Take pf->tc_mutex first, which is the order the dcbnl side has used
since pf->adev_mutex was introduced.
Fixes: 4b31fd4d77ff ("ice: Fix RDMA VSI removal during queue rebuild")
Signed-off-by: Petr Oros <[email protected]>
---
v2:
- new patch, ice_setup_tc() took pf->adev_mutex before pf->tc_mutex
while the dcbnl handlers take them the other way round
---
drivers/net/ethernet/intel/ice/ice_main.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/net/ethernet/intel/ice/ice_main.c
b/drivers/net/ethernet/intel/ice/ice_main.c
index 984479f1e1f6d9..8693ca285d5f1e 100644
--- a/drivers/net/ethernet/intel/ice/ice_main.c
+++ b/drivers/net/ethernet/intel/ice/ice_main.c
@@ -9438,6 +9438,8 @@ ice_setup_tc(struct net_device *netdev, enum
tc_setup_type type,
return -EOPNOTSUPP;
}
+ mutex_lock(&pf->tc_mutex);
+
cdev = pf->cdev_info;
if (cdev) {
mutex_lock(&pf->adev_mutex);
@@ -9456,15 +9458,14 @@ ice_setup_tc(struct net_device *netdev, enum
tc_setup_type type,
}
/* setup traffic classifier for receive side */
- mutex_lock(&pf->tc_mutex);
err = ice_setup_tc_mqprio_qdisc(netdev, type_data);
- mutex_unlock(&pf->tc_mutex);
adev_unlock:
if (locked) {
device_unlock(&adev->dev);
mutex_unlock(&pf->adev_mutex);
}
+ mutex_unlock(&pf->tc_mutex);
return err;
case TC_SETUP_QDISC_ETF:
return ice_offload_txtime(netdev, type_data);
--
2.56.0
