nfp: flower: offload merge flows
authorJohn Hurley <john.hurley@netronome.com>
Mon, 15 Apr 2019 14:56:03 +0000 (16:56 +0200)
committerDavid S. Miller <davem@davemloft.net>
Mon, 15 Apr 2019 22:45:36 +0000 (15:45 -0700)
A merge flow is formed from 2 sub flows. The match fields of the merge are
the same as the first sub flow that has formed it, with the actions being
a combination of the first and second sub flow. Therefore, a merge flow
should replace sub flow 1 when offloaded.

Offload valid merge flows by using a new 'flow mod' message type to
replace an existing offloaded rule. Track the deletion of sub flows that
are linked to a merge flow and revert offloaded merge rules if required.

Signed-off-by: John Hurley <john.hurley@netronome.com>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/netronome/nfp/flower/cmsg.h
drivers/net/ethernet/netronome/nfp/flower/main.h
drivers/net/ethernet/netronome/nfp/flower/metadata.c
drivers/net/ethernet/netronome/nfp/flower/offload.c

index 9288595..a10c29a 100644 (file)
@@ -402,6 +402,7 @@ struct nfp_flower_cmsg_hdr {
 /* Types defined for port related control messages  */
 enum nfp_flower_cmsg_type_port {
        NFP_FLOWER_CMSG_TYPE_FLOW_ADD =         0,
+       NFP_FLOWER_CMSG_TYPE_FLOW_MOD =         1,
        NFP_FLOWER_CMSG_TYPE_FLOW_DEL =         2,
        NFP_FLOWER_CMSG_TYPE_LAG_CONFIG =       4,
        NFP_FLOWER_CMSG_TYPE_PORT_REIFY =       6,
index 896f538..675f43f 100644 (file)
@@ -253,6 +253,7 @@ struct nfp_fl_payload {
        char *mask_data;
        char *action_data;
        struct list_head linked_flows;
+       bool in_hw;
 };
 
 struct nfp_fl_payload_link {
@@ -329,6 +330,8 @@ int nfp_compile_flow_metadata(struct nfp_app *app,
                              struct tc_cls_flower_offload *flow,
                              struct nfp_fl_payload *nfp_flow,
                              struct net_device *netdev);
+void __nfp_modify_flow_metadata(struct nfp_flower_priv *priv,
+                               struct nfp_fl_payload *nfp_flow);
 int nfp_modify_flow_metadata(struct nfp_app *app,
                             struct nfp_fl_payload *nfp_flow);
 
index d68307e..3d326ef 100644 (file)
@@ -276,9 +276,6 @@ nfp_check_mask_remove(struct nfp_app *app, char *mask_data, u32 mask_len,
        if (!mask_entry)
                return false;
 
-       if (meta_flags)
-               *meta_flags &= ~NFP_FL_META_FLAG_MANAGE_MASK;
-
        *mask_id = mask_entry->mask_id;
        mask_entry->ref_cnt--;
        if (!mask_entry->ref_cnt) {
@@ -367,6 +364,14 @@ err_release_stats:
        return err;
 }
 
+void __nfp_modify_flow_metadata(struct nfp_flower_priv *priv,
+                               struct nfp_fl_payload *nfp_flow)
+{
+       nfp_flow->meta.flags &= ~NFP_FL_META_FLAG_MANAGE_MASK;
+       nfp_flow->meta.flow_version = cpu_to_be64(priv->flower_version);
+       priv->flower_version++;
+}
+
 int nfp_modify_flow_metadata(struct nfp_app *app,
                             struct nfp_fl_payload *nfp_flow)
 {
@@ -375,13 +380,12 @@ int nfp_modify_flow_metadata(struct nfp_app *app,
        u8 new_mask_id = 0;
        u32 temp_ctx_id;
 
+       __nfp_modify_flow_metadata(priv, nfp_flow);
+
        nfp_check_mask_remove(app, nfp_flow->mask_data,
                              nfp_flow->meta.mask_len, &nfp_flow->meta.flags,
                              &new_mask_id);
 
-       nfp_flow->meta.flow_version = cpu_to_be64(priv->flower_version);
-       priv->flower_version++;
-
        /* Update flow payload with mask ids. */
        nfp_flow->unmasked_data[NFP_FL_MASK_ID_LOCATION] = new_mask_id;
 
index 1249b89..df81d86 100644 (file)
@@ -348,7 +348,7 @@ nfp_flower_calculate_key_layers(struct nfp_app *app,
                                break;
 
                        case cpu_to_be16(ETH_P_IPV6):
-                               key_layer |= NFP_FLOWER_LAYER_IPV6;
+                                       key_layer |= NFP_FLOWER_LAYER_IPV6;
                                key_size += sizeof(struct nfp_flower_ipv6);
                                break;
 
@@ -399,6 +399,7 @@ nfp_flower_allocate_new(struct nfp_fl_key_ls *key_layer)
        flow_pay->nfp_tun_ipv4_addr = 0;
        flow_pay->meta.flags = 0;
        INIT_LIST_HEAD(&flow_pay->linked_flows);
+       flow_pay->in_hw = false;
 
        return flow_pay;
 
@@ -769,6 +770,8 @@ int nfp_flower_merge_offloaded_flows(struct nfp_app *app,
                                     struct nfp_fl_payload *sub_flow1,
                                     struct nfp_fl_payload *sub_flow2)
 {
+       struct tc_cls_flower_offload merge_tc_off;
+       struct nfp_flower_priv *priv = app->priv;
        struct nfp_fl_payload *merge_flow;
        struct nfp_fl_key_ls merge_key_ls;
        int err;
@@ -810,8 +813,34 @@ int nfp_flower_merge_offloaded_flows(struct nfp_app *app,
        if (err)
                goto err_unlink_sub_flow1;
 
-       err = -EOPNOTSUPP;
+       merge_tc_off.cookie = merge_flow->tc_flower_cookie;
+       err = nfp_compile_flow_metadata(app, &merge_tc_off, merge_flow,
+                                       merge_flow->ingress_dev);
+       if (err)
+               goto err_unlink_sub_flow2;
 
+       err = rhashtable_insert_fast(&priv->flow_table, &merge_flow->fl_node,
+                                    nfp_flower_table_params);
+       if (err)
+               goto err_release_metadata;
+
+       err = nfp_flower_xmit_flow(app, merge_flow,
+                                  NFP_FLOWER_CMSG_TYPE_FLOW_MOD);
+       if (err)
+               goto err_remove_rhash;
+
+       merge_flow->in_hw = true;
+       sub_flow1->in_hw = false;
+
+       return 0;
+
+err_remove_rhash:
+       WARN_ON_ONCE(rhashtable_remove_fast(&priv->flow_table,
+                                           &merge_flow->fl_node,
+                                           nfp_flower_table_params));
+err_release_metadata:
+       nfp_modify_flow_metadata(app, merge_flow);
+err_unlink_sub_flow2:
        nfp_flower_unlink_flows(merge_flow, sub_flow2);
 err_unlink_sub_flow1:
        nfp_flower_unlink_flows(merge_flow, sub_flow1);
@@ -889,6 +918,8 @@ nfp_flower_add_offload(struct nfp_app *app, struct net_device *netdev,
        if (port)
                port->tc_offload_cnt++;
 
+       flow_pay->in_hw = true;
+
        /* Deallocate flow payload when flower rule has been destroyed. */
        kfree(key_layer);
 
@@ -910,6 +941,75 @@ err_free_key_ls:
        return err;
 }
 
+static void
+nfp_flower_remove_merge_flow(struct nfp_app *app,
+                            struct nfp_fl_payload *del_sub_flow,
+                            struct nfp_fl_payload *merge_flow)
+{
+       struct nfp_flower_priv *priv = app->priv;
+       struct nfp_fl_payload_link *link, *temp;
+       struct nfp_fl_payload *origin;
+       bool mod = false;
+       int err;
+
+       link = list_first_entry(&merge_flow->linked_flows,
+                               struct nfp_fl_payload_link, merge_flow.list);
+       origin = link->sub_flow.flow;
+
+       /* Re-add rule the merge had overwritten if it has not been deleted. */
+       if (origin != del_sub_flow)
+               mod = true;
+
+       err = nfp_modify_flow_metadata(app, merge_flow);
+       if (err) {
+               nfp_flower_cmsg_warn(app, "Metadata fail for merge flow delete.\n");
+               goto err_free_links;
+       }
+
+       if (!mod) {
+               err = nfp_flower_xmit_flow(app, merge_flow,
+                                          NFP_FLOWER_CMSG_TYPE_FLOW_DEL);
+               if (err) {
+                       nfp_flower_cmsg_warn(app, "Failed to delete merged flow.\n");
+                       goto err_free_links;
+               }
+       } else {
+               __nfp_modify_flow_metadata(priv, origin);
+               err = nfp_flower_xmit_flow(app, origin,
+                                          NFP_FLOWER_CMSG_TYPE_FLOW_MOD);
+               if (err)
+                       nfp_flower_cmsg_warn(app, "Failed to revert merge flow.\n");
+               origin->in_hw = true;
+       }
+
+err_free_links:
+       /* Clean any links connected with the merged flow. */
+       list_for_each_entry_safe(link, temp, &merge_flow->linked_flows,
+                                merge_flow.list)
+               nfp_flower_unlink_flow(link);
+
+       kfree(merge_flow->action_data);
+       kfree(merge_flow->mask_data);
+       kfree(merge_flow->unmasked_data);
+       WARN_ON_ONCE(rhashtable_remove_fast(&priv->flow_table,
+                                           &merge_flow->fl_node,
+                                           nfp_flower_table_params));
+       kfree_rcu(merge_flow, rcu);
+}
+
+static void
+nfp_flower_del_linked_merge_flows(struct nfp_app *app,
+                                 struct nfp_fl_payload *sub_flow)
+{
+       struct nfp_fl_payload_link *link, *temp;
+
+       /* Remove any merge flow formed from the deleted sub_flow. */
+       list_for_each_entry_safe(link, temp, &sub_flow->linked_flows,
+                                sub_flow.list)
+               nfp_flower_remove_merge_flow(app, sub_flow,
+                                            link->merge_flow.flow);
+}
+
 /**
  * nfp_flower_del_offload() - Removes a flow from hardware.
  * @app:       Pointer to the APP handle
@@ -917,7 +1017,7 @@ err_free_key_ls:
  * @flow:      TC flower classifier offload structure
  *
  * Removes a flow from the repeated hash structure and clears the
- * action payload.
+ * action payload. Any flows merged from this are also deleted.
  *
  * Return: negative value on error, 0 if removed successfully.
  */
@@ -939,17 +1039,22 @@ nfp_flower_del_offload(struct nfp_app *app, struct net_device *netdev,
 
        err = nfp_modify_flow_metadata(app, nfp_flow);
        if (err)
-               goto err_free_flow;
+               goto err_free_merge_flow;
 
        if (nfp_flow->nfp_tun_ipv4_addr)
                nfp_tunnel_del_ipv4_off(app, nfp_flow->nfp_tun_ipv4_addr);
 
+       if (!nfp_flow->in_hw) {
+               err = 0;
+               goto err_free_merge_flow;
+       }
+
        err = nfp_flower_xmit_flow(app, nfp_flow,
                                   NFP_FLOWER_CMSG_TYPE_FLOW_DEL);
-       if (err)
-               goto err_free_flow;
+       /* Fall through on error. */
 
-err_free_flow:
+err_free_merge_flow:
+       nfp_flower_del_linked_merge_flows(app, nfp_flow);
        if (port)
                port->tc_offload_cnt--;
        kfree(nfp_flow->action_data);