Revert "net/mlx5: Arm only EQs with EQEs"
authorShay Drory <shayd@nvidia.com>
Thu, 25 Feb 2021 10:27:53 +0000 (12:27 +0200)
committerSaeed Mahameed <saeedm@nvidia.com>
Thu, 10 Jun 2021 00:20:05 +0000 (17:20 -0700)
In the scenario described below, an EQ can remain in FIRED state which
can result in missing an interrupt generation.

The scenario:

device                       mlx5_core driver
------                       ----------------
EQ1.eqe generated
EQ1.MSI-X sent
EQ1.state = FIRED
EQ2.eqe generated
                             mlx5_irq()
                               polls - eq1_eqes()
                               arm eq1
                               polls - eq2_eqes()
                               arm eq2
EQ2.MSI-X sent
EQ2.state = FIRED
                              mlx5_irq()
                              polls - eq2_eqes() -- no eqes found
                              driver skips EQ arming;

->EQ2 remains fired, misses generating interrupt.

Hence, always arm the EQ by reverting the cited commit in fixes tag.

Fixes: d894892dda25 ("net/mlx5: Arm only EQs with EQEs")
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
drivers/net/ethernet/mellanox/mlx5/core/eq.c

index 77c0ca6..9403334 100644 (file)
@@ -136,7 +136,7 @@ static int mlx5_eq_comp_int(struct notifier_block *nb,
 
        eqe = next_eqe_sw(eq);
        if (!eqe)
-               return 0;
+               goto out;
 
        do {
                struct mlx5_core_cq *cq;
@@ -161,6 +161,8 @@ static int mlx5_eq_comp_int(struct notifier_block *nb,
                ++eq->cons_index;
 
        } while ((++num_eqes < MLX5_EQ_POLLING_BUDGET) && (eqe = next_eqe_sw(eq)));
+
+out:
        eq_update_ci(eq, 1);
 
        if (cqn != -1)
@@ -248,9 +250,9 @@ static int mlx5_eq_async_int(struct notifier_block *nb,
                ++eq->cons_index;
 
        } while ((++num_eqes < MLX5_EQ_POLLING_BUDGET) && (eqe = next_eqe_sw(eq)));
-       eq_update_ci(eq, 1);
 
 out:
+       eq_update_ci(eq, 1);
        mlx5_eq_async_int_unlock(eq_async, recovery, &flags);
 
        return unlikely(recovery) ? num_eqes : 0;