[dpdk-dev] net/mlx5: fix link status is always inconsist
Checks
Commit Message
Query the link status can ends up in an inconsist state where the port is
down but it is reporting speed. For that another query is sceduled for a
later time.
A race condition is possible between the scheduled call and the next
link status interrupt from the device. On some sceraios it will cause the
link to always stay on an inconsist state.
This patch addresses the race condition. Once delayed query is scheduled
no other agents beside it can query the link status.
Fixes: 198a3c339a8f ("mlx5: handle link status interrupts")
CC: stable@dpdk.org
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
---
drivers/net/mlx5/mlx5_ethdev.c | 37 ++++++++++++++++++-------------------
1 file changed, 18 insertions(+), 19 deletions(-)
@@ -1125,9 +1125,10 @@ struct priv *
* Nonzero if the callback process can be called immediately.
*/
static int
-priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
+priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev, int is_alarm)
{
struct ibv_async_event event;
+ struct rte_eth_link *link = &dev->data->dev_link;
int port_change = 0;
int ret = 0;
@@ -1144,22 +1145,20 @@ struct priv *
event.event_type, event.element.port_num);
ibv_ack_async_event(&event);
}
-
- if (port_change ^ priv->pending_alarm) {
- struct rte_eth_link *link = &dev->data->dev_link;
-
- priv->pending_alarm = 0;
- mlx5_link_update(dev, 0);
- if (((link->link_speed == 0) && link->link_status) ||
- ((link->link_speed != 0) && !link->link_status)) {
- /* Inconsistent status, check again later. */
- priv->pending_alarm = 1;
- rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
- mlx5_dev_link_status_handler,
- dev);
- } else
- ret = 1;
- }
+ if ((priv->pending_alarm && !is_alarm) ||
+ (!priv->pending_alarm && !port_change))
+ return ret;
+ priv->pending_alarm = 0;
+ mlx5_link_update(dev, 0);
+ if (((link->link_speed == 0) && link->link_status) ||
+ ((link->link_speed != 0) && !link->link_status)) {
+ /* Inconsistent status, check again later. */
+ priv->pending_alarm = 1;
+ rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
+ mlx5_dev_link_status_handler,
+ dev);
+ } else
+ ret = 1;
return ret;
}
@@ -1178,7 +1177,7 @@ struct priv *
priv_lock(priv);
assert(priv->pending_alarm == 1);
- ret = priv_dev_link_status_handler(priv, dev);
+ ret = priv_dev_link_status_handler(priv, dev, 1);
priv_unlock(priv);
if (ret)
_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
@@ -1201,7 +1200,7 @@ struct priv *
(void)intr_handle;
priv_lock(priv);
- ret = priv_dev_link_status_handler(priv, dev);
+ ret = priv_dev_link_status_handler(priv, dev, 0);
priv_unlock(priv);
if (ret)
_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);