@@ -66,7 +66,7 @@ static const struct rhashtable_params efx_tc_match_action_ht_params = {
static void efx_tc_free_action_set(struct efx_nic *efx,
struct efx_tc_action_set *act, bool in_hw)
{
- /* Failure paths calling this on the 'running action' set in_hw=false,
+ /* Failure paths calling this on the 'cursor' action set in_hw=false,
* because if the alloc had succeeded we'd've put it in acts.list and
* not still have it in act.
*/
@@ -407,6 +407,30 @@ static int efx_tc_flower_replace(struct efx_nic *efx,
goto release;
}
+ /**
+ * DOC: TC action translation
+ *
+ * Actions in TC are sequential and cumulative, with delivery actions
+ * potentially anywhere in the order. The EF100 MAE, however, takes
+ * an 'action set list' consisting of 'action sets', each of which is
+ * applied to the _original_ packet, and consists of a set of optional
+ * actions in a fixed order with delivery at the end.
+ * To translate between these two models, we maintain a 'cursor', @act,
+ * which describes the cumulative effect of all the packet-mutating
+ * actions encountered so far; on handling a delivery (mirred or drop)
+ * action, once the action-set has been inserted into hardware, we
+ * append @act to the action-set list (@rule->acts); if this is a pipe
+ * action (mirred mirror) we then allocate a new @act with a copy of
+ * the cursor state _before_ the delivery action, otherwise we set @act
+ * to %NULL.
+ * This ensures that every allocated action-set is either attached to
+ * @rule->acts or pointed to by @act (and never both), and that only
+ * those action-sets in @rule->acts exist in hardware. Consequently,
+ * in the failure path, @act only needs to be freed in memory, whereas
+ * for @rule->acts we remove each action-set from hardware before
+ * freeing it (efx_tc_free_action_set_list()), even if the action-set
+ * list itself is not in hardware.
+ */
flow_action_for_each(i, fa, &fr->action) {
struct efx_tc_action_set save;
u16 tci;