net/mlx5e: Allow concurrent creation of mod_hdr entries
authorVlad Buslov <vladbu@mellanox.com>
Thu, 8 Aug 2019 13:53:15 +0000 (16:53 +0300)
committerSaeed Mahameed <saeedm@mellanox.com>
Fri, 9 Aug 2019 21:54:10 +0000 (14:54 -0700)
Mod_hdr entries creation is fully synchronized by mod_hdr_tbl->lock. In
order to allow concurrent allocation of hardware resources used to offload
header rewrite, extend mlx5e_mod_hdr_entry with 'res_ready' completion.
Move call to mlx5_modify_header_alloc() out of mod_hdr_tbl->lock critical
section. Modify code that attaches new flows to existing mh to wait for
'res_ready' completion before using the entry. Insert mh to mod_hdr table
before provisioning it to hardware and modify all users of mod_hdr table to
verify that mh was fully initialized by checking completion result for
negative value (and to wait for 'res_ready' completion, if necessary).

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c

index 0600b78..fcaf9ab 100644 (file)
@@ -199,6 +199,8 @@ struct mlx5e_mod_hdr_entry {
        u32 mod_hdr_id;
 
        refcount_t refcnt;
+       struct completion res_ready;
+       int compl_result;
 };
 
 #define MLX5_MH_ACT_SZ MLX5_UN_SZ_BYTES(set_action_in_add_action_in_auto)
@@ -326,7 +328,8 @@ static void mlx5e_mod_hdr_put(struct mlx5e_priv *priv,
        mutex_unlock(&tbl->lock);
 
        WARN_ON(!list_empty(&mh->flows));
-       mlx5_modify_header_dealloc(priv->mdev, mh->mod_hdr_id);
+       if (mh->compl_result > 0)
+               mlx5_modify_header_dealloc(priv->mdev, mh->mod_hdr_id);
 
        kfree(mh);
 }
@@ -359,13 +362,21 @@ static int mlx5e_attach_mod_hdr(struct mlx5e_priv *priv,
 
        mutex_lock(&tbl->lock);
        mh = mlx5e_mod_hdr_get(tbl, &key, hash_key);
-       if (mh)
+       if (mh) {
+               mutex_unlock(&tbl->lock);
+               wait_for_completion(&mh->res_ready);
+
+               if (mh->compl_result < 0) {
+                       err = -EREMOTEIO;
+                       goto attach_header_err;
+               }
                goto attach_flow;
+       }
 
        mh = kzalloc(sizeof(*mh) + actions_size, GFP_KERNEL);
        if (!mh) {
-               err = -ENOMEM;
-               goto out_err;
+               mutex_unlock(&tbl->lock);
+               return -ENOMEM;
        }
 
        mh->key.actions = (void *)mh + sizeof(*mh);
@@ -374,18 +385,23 @@ static int mlx5e_attach_mod_hdr(struct mlx5e_priv *priv,
        spin_lock_init(&mh->flows_lock);
        INIT_LIST_HEAD(&mh->flows);
        refcount_set(&mh->refcnt, 1);
+       init_completion(&mh->res_ready);
+
+       hash_add(tbl->hlist, &mh->mod_hdr_hlist, hash_key);
+       mutex_unlock(&tbl->lock);
 
        err = mlx5_modify_header_alloc(priv->mdev, namespace,
                                       mh->key.num_actions,
                                       mh->key.actions,
                                       &mh->mod_hdr_id);
-       if (err)
-               goto out_err;
-
-       hash_add(tbl->hlist, &mh->mod_hdr_hlist, hash_key);
+       if (err) {
+               mh->compl_result = err;
+               goto alloc_header_err;
+       }
+       mh->compl_result = 1;
+       complete_all(&mh->res_ready);
 
 attach_flow:
-       mutex_unlock(&tbl->lock);
        flow->mh = mh;
        spin_lock(&mh->flows_lock);
        list_add(&flow->mod_hdr, &mh->flows);
@@ -397,9 +413,10 @@ attach_flow:
 
        return 0;
 
-out_err:
-       mutex_unlock(&tbl->lock);
-       kfree(mh);
+alloc_header_err:
+       complete_all(&mh->res_ready);
+attach_header_err:
+       mlx5e_mod_hdr_put(priv, mh, namespace);
        return err;
 }