net/mlx5_core: Introduce flow steering autogrouped flow table
authorMaor Gottlieb <maorg@mellanox.com>
Mon, 11 Jan 2016 08:25:57 +0000 (10:25 +0200)
committerDavid S. Miller <davem@davemloft.net>
Mon, 11 Jan 2016 22:48:52 +0000 (17:48 -0500)
When user add rule to autogrouped flow table, we search
for flow group with the same match criteria, if we don't
find such group then we create new flow group with the
required match criteria and insert the rule to this group.

We divide the flow table into required_groups + 1,
in order to reserve a part of the flow table for rules
which don't match any existing group.

Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Moni Shoua <monis@mellanox.com>
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx5/core/fs_core.c
drivers/net/ethernet/mellanox/mlx5/core/fs_core.h
include/linux/mlx5/fs.h

index f7d62fe595f6373574a414d9b9b68f49c5d92779..7d24bbba58ba670fd75c3814ea7986bd669b6a59 100644 (file)
@@ -85,6 +85,12 @@ static struct init_tree_node {
        }
 };
 
+enum fs_i_mutex_lock_class {
+       FS_MUTEX_GRANDPARENT,
+       FS_MUTEX_PARENT,
+       FS_MUTEX_CHILD
+};
+
 static void del_rule(struct fs_node *node);
 static void del_flow_table(struct fs_node *node);
 static void del_flow_group(struct fs_node *node);
@@ -119,10 +125,11 @@ static void tree_get_node(struct fs_node *node)
        atomic_inc(&node->refcount);
 }
 
-static void nested_lock_ref_node(struct fs_node *node)
+static void nested_lock_ref_node(struct fs_node *node,
+                                enum fs_i_mutex_lock_class class)
 {
        if (node) {
-               mutex_lock_nested(&node->lock, SINGLE_DEPTH_NESTING);
+               mutex_lock_nested(&node->lock, class);
                atomic_inc(&node->refcount);
        }
 }
@@ -481,9 +488,7 @@ struct mlx5_flow_table *mlx5_create_flow_table(struct mlx5_flow_namespace *ns,
        list_add_tail(&ft->node.list, &fs_prio->node.children);
        fs_prio->num_ft++;
        unlock_ref_node(&fs_prio->node);
-
        return ft;
-
 free_ft:
        kfree(ft);
 unlock_prio:
@@ -491,8 +496,32 @@ unlock_prio:
        return ERR_PTR(err);
 }
 
-struct mlx5_flow_group *mlx5_create_flow_group(struct mlx5_flow_table *ft,
-                                              u32 *fg_in)
+struct mlx5_flow_table *mlx5_create_auto_grouped_flow_table(struct mlx5_flow_namespace *ns,
+                                                           int prio,
+                                                           int num_flow_table_entries,
+                                                           int max_num_groups)
+{
+       struct mlx5_flow_table *ft;
+
+       if (max_num_groups > num_flow_table_entries)
+               return ERR_PTR(-EINVAL);
+
+       ft = mlx5_create_flow_table(ns, prio, num_flow_table_entries);
+       if (IS_ERR(ft))
+               return ft;
+
+       ft->autogroup.active = true;
+       ft->autogroup.required_groups = max_num_groups;
+
+       return ft;
+}
+
+/* Flow table should be locked */
+static struct mlx5_flow_group *create_flow_group_common(struct mlx5_flow_table *ft,
+                                                       u32 *fg_in,
+                                                       struct list_head
+                                                       *prev_fg,
+                                                       bool is_auto_fg)
 {
        struct mlx5_flow_group *fg;
        struct mlx5_core_dev *dev = get_dev(&ft->node);
@@ -505,18 +534,33 @@ struct mlx5_flow_group *mlx5_create_flow_group(struct mlx5_flow_table *ft,
        if (IS_ERR(fg))
                return fg;
 
-       lock_ref_node(&ft->node);
        err = mlx5_cmd_create_flow_group(dev, ft, fg_in, &fg->id);
        if (err) {
                kfree(fg);
-               unlock_ref_node(&ft->node);
                return ERR_PTR(err);
        }
+
+       if (ft->autogroup.active)
+               ft->autogroup.num_groups++;
        /* Add node to tree */
-       tree_init_node(&fg->node, 1, del_flow_group);
+       tree_init_node(&fg->node, !is_auto_fg, del_flow_group);
        tree_add_node(&fg->node, &ft->node);
        /* Add node to group list */
        list_add(&fg->node.list, ft->node.children.prev);
+
+       return fg;
+}
+
+struct mlx5_flow_group *mlx5_create_flow_group(struct mlx5_flow_table *ft,
+                                              u32 *fg_in)
+{
+       struct mlx5_flow_group *fg;
+
+       if (ft->autogroup.active)
+               return ERR_PTR(-EPERM);
+
+       lock_ref_node(&ft->node);
+       fg = create_flow_group_common(ft, fg_in, &ft->node.children, false);
        unlock_ref_node(&ft->node);
 
        return fg;
@@ -614,7 +658,63 @@ static struct fs_fte *create_fte(struct mlx5_flow_group *fg,
        return fte;
 }
 
-/* Assuming parent fg(flow table) is locked */
+static struct mlx5_flow_group *create_autogroup(struct mlx5_flow_table *ft,
+                                               u8 match_criteria_enable,
+                                               u32 *match_criteria)
+{
+       int inlen = MLX5_ST_SZ_BYTES(create_flow_group_in);
+       struct list_head *prev = &ft->node.children;
+       unsigned int candidate_index = 0;
+       struct mlx5_flow_group *fg;
+       void *match_criteria_addr;
+       unsigned int group_size = 0;
+       u32 *in;
+
+       if (!ft->autogroup.active)
+               return ERR_PTR(-ENOENT);
+
+       in = mlx5_vzalloc(inlen);
+       if (!in)
+               return ERR_PTR(-ENOMEM);
+
+       if (ft->autogroup.num_groups < ft->autogroup.required_groups)
+               /* We save place for flow groups in addition to max types */
+               group_size = ft->max_fte / (ft->autogroup.required_groups + 1);
+
+       /*  ft->max_fte == ft->autogroup.max_types */
+       if (group_size == 0)
+               group_size = 1;
+
+       /* sorted by start_index */
+       fs_for_each_fg(fg, ft) {
+               if (candidate_index + group_size > fg->start_index)
+                       candidate_index = fg->start_index + fg->max_ftes;
+               else
+                       break;
+               prev = &fg->node.list;
+       }
+
+       if (candidate_index + group_size > ft->max_fte) {
+               fg = ERR_PTR(-ENOSPC);
+               goto out;
+       }
+
+       MLX5_SET(create_flow_group_in, in, match_criteria_enable,
+                match_criteria_enable);
+       MLX5_SET(create_flow_group_in, in, start_flow_index, candidate_index);
+       MLX5_SET(create_flow_group_in, in, end_flow_index,   candidate_index +
+                group_size - 1);
+       match_criteria_addr = MLX5_ADDR_OF(create_flow_group_in,
+                                          in, match_criteria);
+       memcpy(match_criteria_addr, match_criteria,
+              MLX5_ST_SZ_BYTES(fte_match_param));
+
+       fg = create_flow_group_common(ft, in, prev, true);
+out:
+       kvfree(in);
+       return fg;
+}
+
 static struct mlx5_flow_rule *add_rule_fg(struct mlx5_flow_group *fg,
                                          u32 *match_value,
                                          u8 action,
@@ -626,9 +726,9 @@ static struct mlx5_flow_rule *add_rule_fg(struct mlx5_flow_group *fg,
        struct mlx5_flow_table *ft;
        struct list_head *prev;
 
-       lock_ref_node(&fg->node);
+       nested_lock_ref_node(&fg->node, FS_MUTEX_PARENT);
        fs_for_each_fte(fte, fg) {
-               nested_lock_ref_node(&fte->node);
+               nested_lock_ref_node(&fte->node, FS_MUTEX_CHILD);
                if (compare_match_value(&fg->mask, match_value, &fte->val) &&
                    action == fte->action && flow_tag == fte->flow_tag) {
                        rule = add_rule_fte(fte, fg, dest);
@@ -669,6 +769,33 @@ unlock_fg:
        return rule;
 }
 
+static struct mlx5_flow_rule *add_rule_to_auto_fg(struct mlx5_flow_table *ft,
+                                                 u8 match_criteria_enable,
+                                                 u32 *match_criteria,
+                                                 u32 *match_value,
+                                                 u8 action,
+                                                 u32 flow_tag,
+                                                 struct mlx5_flow_destination *dest)
+{
+       struct mlx5_flow_rule *rule;
+       struct mlx5_flow_group *g;
+
+       g = create_autogroup(ft, match_criteria_enable, match_criteria);
+       if (IS_ERR(g))
+               return (void *)g;
+
+       rule = add_rule_fg(g, match_value,
+                          action, flow_tag, dest);
+       if (IS_ERR(rule)) {
+               /* Remove assumes refcount > 0 and autogroup creates a group
+                * with a refcount = 0.
+                */
+               tree_get_node(&g->node);
+               tree_remove_node(&g->node);
+       }
+       return rule;
+}
+
 struct mlx5_flow_rule *
 mlx5_add_flow_rule(struct mlx5_flow_table *ft,
                   u8 match_criteria_enable,
@@ -679,23 +806,24 @@ mlx5_add_flow_rule(struct mlx5_flow_table *ft,
                   struct mlx5_flow_destination *dest)
 {
        struct mlx5_flow_group *g;
-       struct mlx5_flow_rule *rule = ERR_PTR(-EINVAL);
+       struct mlx5_flow_rule *rule;
 
-       tree_get_node(&ft->node);
-       lock_ref_node(&ft->node);
+       nested_lock_ref_node(&ft->node, FS_MUTEX_GRANDPARENT);
        fs_for_each_fg(g, ft)
                if (compare_match_criteria(g->mask.match_criteria_enable,
                                           match_criteria_enable,
                                           g->mask.match_criteria,
                                           match_criteria)) {
-                       unlock_ref_node(&ft->node);
                        rule = add_rule_fg(g, match_value,
                                           action, flow_tag, dest);
-                       goto put;
+                       if (!IS_ERR(rule) || PTR_ERR(rule) != -ENOSPC)
+                               goto unlock;
                }
+
+       rule = add_rule_to_auto_fg(ft, match_criteria_enable, match_criteria,
+                                  match_value, action, flow_tag, dest);
+unlock:
        unlock_ref_node(&ft->node);
-put:
-       tree_put_node(&ft->node);
        return rule;
 }
 
index 4ebb97fd554456030ed66486a78bf0aedd4e826f..0f98257c0d31cba0fdf99dc3f1f183889c614594 100644 (file)
@@ -77,6 +77,11 @@ struct mlx5_flow_table {
        unsigned int                    max_fte;
        unsigned int                    level;
        enum fs_flow_table_type         type;
+       struct {
+               bool                    active;
+               unsigned int            required_groups;
+               unsigned int            num_groups;
+       } autogroup;
 };
 
 /* Type of children is mlx5_flow_rule */
index bc7ad019afdefa5558f10cd4c5a553eeeebb562a..06ac6e8fccfaf9b5dc085c3bf126161609d4ea6f 100644 (file)
@@ -61,6 +61,12 @@ struct mlx5_flow_namespace *
 mlx5_get_flow_namespace(struct mlx5_core_dev *dev,
                        enum mlx5_flow_namespace_type type);
 
+struct mlx5_flow_table *
+mlx5_create_auto_grouped_flow_table(struct mlx5_flow_namespace *ns,
+                                   int prio,
+                                   int num_flow_table_entries,
+                                   int max_num_groups);
+
 struct mlx5_flow_table *
 mlx5_create_flow_table(struct mlx5_flow_namespace *ns,
                       int prio,