net/mlx5: Delay events till ib registration ends
When mlx5_ib registers itself to mlx5_core as an interface, it will
call mlx5_add_device which will call mlx5_ib interface add callback,
in case the latter successfully returns, only then mlx5_core will add
it to the interface list and async events will be forwarded to mlx5_ib.
Between mlx5_ib interface add callback and mlx5_core adding the mlx5_ib
interface to its devices list, arriving mlx5_core events can be missed
by the new mlx5_ib registering interface.
In other words:
thread 1: mlx5_ib: mlx5_register_interface(dev)
thread 1: mlx5_core: mlx5_add_device(dev)
thread 1: mlx5_core: ctx = dev->add => (mlx5_ib)->mlx5_ib_add
thread 2: mlx5_core_event: **new event arrives, forward to dev_list
thread 1: mlx5_core: add_ctx_to_dev_list(ctx)
/* previous event was missed by the new interface.*/
It is ok to miss events before dev->add (mlx5_ib)->mlx5_ib_add_device
but not after.
We fix this race by accumulating the events that come between the
ib_register_device (inside mlx5_add_device->(dev->add)) till the adding
to the list completes and fire them to the new registering interface
after that.
Fixes: f1ee87fe55c8 ("net/mlx5: Organize device list API in one place")
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h
index 88d6eb5..d26f18b 100644
--- a/include/linux/mlx5/driver.h
+++ b/include/linux/mlx5/driver.h
@@ -647,6 +647,9 @@ struct mlx5_priv {
struct list_head ctx_list;
spinlock_t ctx_lock;
+ struct list_head waiting_events_list;
+ bool is_accum_events;
+
struct mlx5_flow_steering *steering;
struct mlx5_mpfs *mpfs;
struct mlx5_eswitch *eswitch;