ofproto: Optimize datapath actions.
authorEthan Jackson <ethan@nicira.com>
Fri, 27 May 2011 17:48:58 +0000 (10:48 -0700)
committerEthan Jackson <ethan@nicira.com>
Wed, 1 Jun 2011 21:30:10 +0000 (14:30 -0700)
The translation from ofproto to datapath actions has historically
been fairly naive.  Since redundant ODP actions result in more code
execution in the fast path, this patch changes xlate_actions() to
emit fewer ODP actions for a given flow.

This change is not simply a theoretical optimization.  It actually
reduces the number of ODP actions for real flow tables which
Nicira's controllers use in practice.  Furthermore, removing
unnecessary modification actions after the last output action has
the added benefit of saving a whole skb_clone() in the fast path.

ofproto/ofproto-dpif.c

index f632ad0..55d109f 100644 (file)
@@ -184,9 +184,9 @@ struct action_xlate_ctx {
  * reason to look at them. */
 
     int recurse;                /* Recursion level, via xlate_table_action. */
-    int last_pop_priority;      /* Offset in 'odp_actions' just past most
-                                 * recent ODP_ACTION_ATTR_SET_PRIORITY. */
     uint32_t priority;          /* Current flow priority. 0 if none. */
+    struct flow base_flow;      /* Flow at the last commit. */
+    uint32_t base_priority;     /* Priority at the last commit. */
 };
 
 static void action_xlate_ctx_init(struct action_xlate_ctx *,
@@ -2709,6 +2709,71 @@ static void do_xlate_actions(const union ofp_action *in, size_t n_in,
                              struct action_xlate_ctx *ctx);
 static bool xlate_normal(struct action_xlate_ctx *);
 
+static void
+commit_odp_actions(struct action_xlate_ctx *ctx)
+{
+    const struct flow *flow = &ctx->flow;
+    struct flow *base = &ctx->base_flow;
+    struct ofpbuf *odp_actions = ctx->odp_actions;
+
+    if (base->tun_id != flow->tun_id) {
+        nl_msg_put_be64(odp_actions, ODP_ACTION_ATTR_SET_TUNNEL, flow->tun_id);
+        base->tun_id = flow->tun_id;
+    }
+
+    if (base->nw_src != flow->nw_src) {
+        nl_msg_put_be32(odp_actions, ODP_ACTION_ATTR_SET_NW_SRC, flow->nw_src);
+        base->nw_src = flow->nw_src;
+    }
+
+    if (base->nw_dst != flow->nw_dst) {
+        nl_msg_put_be32(odp_actions, ODP_ACTION_ATTR_SET_NW_DST, flow->nw_dst);
+        base->nw_dst = flow->nw_dst;
+    }
+
+    if (base->vlan_tci != flow->vlan_tci) {
+        if (!(flow->vlan_tci & htons(VLAN_CFI))) {
+            nl_msg_put_flag(odp_actions, ODP_ACTION_ATTR_STRIP_VLAN);
+        } else {
+            nl_msg_put_be16(odp_actions, ODP_ACTION_ATTR_SET_DL_TCI,
+                            flow->vlan_tci & ~htons(VLAN_CFI));
+        }
+        base->vlan_tci = flow->vlan_tci;
+    }
+
+    if (base->tp_src != flow->tp_src) {
+        nl_msg_put_be16(odp_actions, ODP_ACTION_ATTR_SET_TP_SRC, flow->tp_src);
+        base->tp_src = flow->tp_src;
+    }
+
+    if (base->tp_dst != flow->tp_dst) {
+        nl_msg_put_be16(odp_actions, ODP_ACTION_ATTR_SET_TP_DST, flow->tp_dst);
+        base->tp_dst = flow->tp_dst;
+    }
+
+    if (!eth_addr_equals(base->dl_src, flow->dl_src)) {
+        nl_msg_put_unspec(odp_actions, ODP_ACTION_ATTR_SET_DL_SRC,
+                          flow->dl_src, ETH_ADDR_LEN);
+        memcpy(base->dl_src, flow->dl_src, ETH_ADDR_LEN);
+    }
+
+    if (!eth_addr_equals(base->dl_dst, flow->dl_dst)) {
+        nl_msg_put_unspec(odp_actions, ODP_ACTION_ATTR_SET_DL_DST,
+                          flow->dl_dst, ETH_ADDR_LEN);
+        memcpy(base->dl_dst, flow->dl_dst, ETH_ADDR_LEN);
+    }
+
+    if (ctx->base_priority != ctx->priority) {
+        if (ctx->priority) {
+            nl_msg_put_u32(odp_actions, ODP_ACTION_ATTR_SET_PRIORITY,
+                           ctx->priority);
+        } else {
+            nl_msg_put_flag(odp_actions, ODP_ACTION_ATTR_POP_PRIORITY);
+        }
+        ctx->base_priority = ctx->priority;
+    }
+}
+
 static void
 add_output_action(struct action_xlate_ctx *ctx, uint16_t ofp_port)
 {
@@ -2728,6 +2793,7 @@ add_output_action(struct action_xlate_ctx *ctx, uint16_t ofp_port)
          */
     }
 
+    commit_odp_actions(ctx);
     nl_msg_put_u32(ctx->odp_actions, ODP_ACTION_ATTR_OUTPUT, odp_port);
     ctx->nf_output_iface = ofp_port;
 }
@@ -2765,20 +2831,20 @@ xlate_table_action(struct action_xlate_ctx *ctx, uint16_t in_port)
 }
 
 static void
-flood_packets(struct ofproto_dpif *ofproto,
-              uint16_t ofp_in_port, ovs_be32 mask,
-              uint16_t *nf_output_iface, struct ofpbuf *odp_actions)
+flood_packets(struct action_xlate_ctx *ctx, ovs_be32 mask)
 {
     struct ofport_dpif *ofport;
 
-    HMAP_FOR_EACH (ofport, up.hmap_node, &ofproto->up.ports) {
+    commit_odp_actions(ctx);
+    HMAP_FOR_EACH (ofport, up.hmap_node, &ctx->ofproto->up.ports) {
         uint16_t ofp_port = ofport->up.ofp_port;
-        if (ofp_port != ofp_in_port && !(ofport->up.opp.config & mask)) {
-            nl_msg_put_u32(odp_actions, ODP_ACTION_ATTR_OUTPUT,
+        if (ofp_port != ctx->flow.in_port && !(ofport->up.opp.config & mask)) {
+            nl_msg_put_u32(ctx->odp_actions, ODP_ACTION_ATTR_OUTPUT,
                            ofport->odp_port);
         }
     }
-    *nf_output_iface = NF_OUT_FLOOD;
+
+    ctx->nf_output_iface = NF_OUT_FLOOD;
 }
 
 static void
@@ -2800,14 +2866,13 @@ xlate_output_action__(struct action_xlate_ctx *ctx,
         xlate_normal(ctx);
         break;
     case OFPP_FLOOD:
-        flood_packets(ctx->ofproto, ctx->flow.in_port, htonl(OFPPC_NO_FLOOD),
-                      &ctx->nf_output_iface, ctx->odp_actions);
+        flood_packets(ctx,  htonl(OFPPC_NO_FLOOD));
         break;
     case OFPP_ALL:
-        flood_packets(ctx->ofproto, ctx->flow.in_port, htonl(0),
-                      &ctx->nf_output_iface, ctx->odp_actions);
+        flood_packets(ctx, htonl(0));
         break;
     case OFPP_CONTROLLER:
+        commit_odp_actions(ctx);
         nl_msg_put_u64(ctx->odp_actions, ODP_ACTION_ATTR_CONTROLLER, max_len);
         break;
     case OFPP_LOCAL:
@@ -2837,34 +2902,12 @@ xlate_output_action(struct action_xlate_ctx *ctx,
     xlate_output_action__(ctx, ntohs(oao->port), ntohs(oao->max_len));
 }
 
-/* If the final ODP action in 'ctx' is "pop priority", drop it, as an
- * optimization, because we're going to add another action that sets the
- * priority immediately after, or because there are no actions following the
- * pop.  */
-static void
-remove_pop_action(struct action_xlate_ctx *ctx)
-{
-    if (ctx->odp_actions->size == ctx->last_pop_priority) {
-        ctx->odp_actions->size -= NLA_ALIGN(NLA_HDRLEN);
-        ctx->last_pop_priority = -1;
-    }
-}
-
-static void
-add_pop_action(struct action_xlate_ctx *ctx)
-{
-    if (ctx->odp_actions->size != ctx->last_pop_priority) {
-        nl_msg_put_flag(ctx->odp_actions, ODP_ACTION_ATTR_POP_PRIORITY);
-        ctx->last_pop_priority = ctx->odp_actions->size;
-    }
-}
-
 static void
 xlate_enqueue_action(struct action_xlate_ctx *ctx,
                      const struct ofp_action_enqueue *oae)
 {
     uint16_t ofp_port, odp_port;
-    uint32_t priority;
+    uint32_t ctx_priority, priority;
     int error;
 
     error = dpif_queue_to_priority(ctx->ofproto->dpif, ntohl(oae->queue_id),
@@ -2883,16 +2926,10 @@ xlate_enqueue_action(struct action_xlate_ctx *ctx,
     odp_port = ofp_port_to_odp_port(ofp_port);
 
     /* Add ODP actions. */
-    remove_pop_action(ctx);
-    nl_msg_put_u32(ctx->odp_actions, ODP_ACTION_ATTR_SET_PRIORITY, priority);
+    ctx_priority = ctx->priority;
+    ctx->priority = priority;
     add_output_action(ctx, odp_port);
-
-    if (ctx->priority) {
-        nl_msg_put_u32(ctx->odp_actions, ODP_ACTION_ATTR_SET_PRIORITY,
-                       ctx->priority);
-    } else {
-        add_pop_action(ctx);
-    }
+    ctx->priority = ctx_priority;
 
     /* Update NetFlow output port. */
     if (ctx->nf_output_iface == NF_OUT_DROP) {
@@ -2918,20 +2955,6 @@ xlate_set_queue_action(struct action_xlate_ctx *ctx,
     }
 
     ctx->priority = priority;
-    remove_pop_action(ctx);
-    nl_msg_put_u32(ctx->odp_actions, ODP_ACTION_ATTR_SET_PRIORITY, priority);
-}
-
-static void
-xlate_set_dl_tci(struct action_xlate_ctx *ctx)
-{
-    ovs_be16 tci = ctx->flow.vlan_tci;
-    if (!(tci & htons(VLAN_CFI))) {
-        nl_msg_put_flag(ctx->odp_actions, ODP_ACTION_ATTR_STRIP_VLAN);
-    } else {
-        nl_msg_put_be16(ctx->odp_actions, ODP_ACTION_ATTR_SET_DL_TCI,
-                        tci & ~htons(VLAN_CFI));
-    }
 }
 
 struct xlate_reg_state {
@@ -2939,27 +2962,6 @@ struct xlate_reg_state {
     ovs_be64 tun_id;
 };
 
-static void
-save_reg_state(const struct action_xlate_ctx *ctx,
-               struct xlate_reg_state *state)
-{
-    state->vlan_tci = ctx->flow.vlan_tci;
-    state->tun_id = ctx->flow.tun_id;
-}
-
-static void
-update_reg_state(struct action_xlate_ctx *ctx,
-                 const struct xlate_reg_state *state)
-{
-    if (ctx->flow.vlan_tci != state->vlan_tci) {
-        xlate_set_dl_tci(ctx);
-    }
-    if (ctx->flow.tun_id != state->tun_id) {
-        nl_msg_put_be64(ctx->odp_actions,
-                        ODP_ACTION_ATTR_SET_TUNNEL, ctx->flow.tun_id);
-    }
-}
-
 static void
 xlate_autopath(struct action_xlate_ctx *ctx,
                const struct nx_action_autopath *naa)
@@ -2990,7 +2992,6 @@ xlate_nicira_action(struct action_xlate_ctx *ctx,
     const struct nx_action_multipath *nam;
     const struct nx_action_autopath *naa;
     enum nx_action_subtype subtype = ntohs(nah->subtype);
-    struct xlate_reg_state state;
     ovs_be64 tun_id;
 
     assert(nah->vendor == htonl(NX_VENDOR_ID));
@@ -3003,12 +3004,16 @@ xlate_nicira_action(struct action_xlate_ctx *ctx,
     case NXAST_SET_TUNNEL:
         nast = (const struct nx_action_set_tunnel *) nah;
         tun_id = htonll(ntohl(nast->tun_id));
-        nl_msg_put_be64(ctx->odp_actions, ODP_ACTION_ATTR_SET_TUNNEL, tun_id);
         ctx->flow.tun_id = tun_id;
         break;
 
     case NXAST_DROP_SPOOFED_ARP:
         if (ctx->flow.dl_type == htons(ETH_TYPE_ARP)) {
+            /* XXX: It's not entirely clear whether or not we need to commit
+             * here.  The safer thing to do is commit of course.  Hopefully in
+             * the near future we can rip out NXAST_DROP_SPOOFED_ARP altogether
+             * and the point will be moot. */
+            commit_odp_actions(ctx);
             nl_msg_put_flag(ctx->odp_actions,
                             ODP_ACTION_ATTR_DROP_SPOOFED_ARP);
         }
@@ -3021,21 +3026,16 @@ xlate_nicira_action(struct action_xlate_ctx *ctx,
 
     case NXAST_POP_QUEUE:
         ctx->priority = 0;
-        add_pop_action(ctx);
         break;
 
     case NXAST_REG_MOVE:
-        save_reg_state(ctx, &state);
         nxm_execute_reg_move((const struct nx_action_reg_move *) nah,
                              &ctx->flow);
-        update_reg_state(ctx, &state);
         break;
 
     case NXAST_REG_LOAD:
-        save_reg_state(ctx, &state);
         nxm_execute_reg_load((const struct nx_action_reg_load *) nah,
                              &ctx->flow);
-        update_reg_state(ctx, &state);
         break;
 
     case NXAST_NOTE:
@@ -3044,7 +3044,6 @@ xlate_nicira_action(struct action_xlate_ctx *ctx,
 
     case NXAST_SET_TUNNEL64:
         tun_id = ((const struct nx_action_set_tunnel64 *) nah)->tun_id;
-        nl_msg_put_be64(ctx->odp_actions, ODP_ACTION_ATTR_SET_TUNNEL, tun_id);
         ctx->flow.tun_id = tun_id;
         break;
 
@@ -3098,62 +3097,45 @@ do_xlate_actions(const union ofp_action *in, size_t n_in,
         case OFPAT_SET_VLAN_VID:
             ctx->flow.vlan_tci &= ~htons(VLAN_VID_MASK);
             ctx->flow.vlan_tci |= ia->vlan_vid.vlan_vid | htons(VLAN_CFI);
-            xlate_set_dl_tci(ctx);
             break;
 
         case OFPAT_SET_VLAN_PCP:
             ctx->flow.vlan_tci &= ~htons(VLAN_PCP_MASK);
             ctx->flow.vlan_tci |= htons(
                 (ia->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT) | VLAN_CFI);
-            xlate_set_dl_tci(ctx);
             break;
 
         case OFPAT_STRIP_VLAN:
             ctx->flow.vlan_tci = htons(0);
-            xlate_set_dl_tci(ctx);
             break;
 
         case OFPAT_SET_DL_SRC:
             oada = ((struct ofp_action_dl_addr *) ia);
-            nl_msg_put_unspec(ctx->odp_actions, ODP_ACTION_ATTR_SET_DL_SRC,
-                              oada->dl_addr, ETH_ADDR_LEN);
             memcpy(ctx->flow.dl_src, oada->dl_addr, ETH_ADDR_LEN);
             break;
 
         case OFPAT_SET_DL_DST:
             oada = ((struct ofp_action_dl_addr *) ia);
-            nl_msg_put_unspec(ctx->odp_actions, ODP_ACTION_ATTR_SET_DL_DST,
-                              oada->dl_addr, ETH_ADDR_LEN);
             memcpy(ctx->flow.dl_dst, oada->dl_addr, ETH_ADDR_LEN);
             break;
 
         case OFPAT_SET_NW_SRC:
-            nl_msg_put_be32(ctx->odp_actions, ODP_ACTION_ATTR_SET_NW_SRC,
-                            ia->nw_addr.nw_addr);
             ctx->flow.nw_src = ia->nw_addr.nw_addr;
             break;
 
         case OFPAT_SET_NW_DST:
-            nl_msg_put_be32(ctx->odp_actions, ODP_ACTION_ATTR_SET_NW_DST,
-                            ia->nw_addr.nw_addr);
             ctx->flow.nw_dst = ia->nw_addr.nw_addr;
             break;
 
         case OFPAT_SET_NW_TOS:
-            nl_msg_put_u8(ctx->odp_actions, ODP_ACTION_ATTR_SET_NW_TOS,
-                          ia->nw_tos.nw_tos);
             ctx->flow.nw_tos = ia->nw_tos.nw_tos;
             break;
 
         case OFPAT_SET_TP_SRC:
-            nl_msg_put_be16(ctx->odp_actions, ODP_ACTION_ATTR_SET_TP_SRC,
-                            ia->tp_port.tp_port);
             ctx->flow.tp_src = ia->tp_port.tp_port;
             break;
 
         case OFPAT_SET_TP_DST:
-            nl_msg_put_be16(ctx->odp_actions, ODP_ACTION_ATTR_SET_TP_DST,
-                            ia->tp_port.tp_port);
             ctx->flow.tp_dst = ia->tp_port.tp_port;
             break;
 
@@ -3194,8 +3176,9 @@ xlate_actions(struct action_xlate_ctx *ctx,
     ctx->may_set_up_flow = true;
     ctx->nf_output_iface = NF_OUT_DROP;
     ctx->recurse = 0;
-    ctx->last_pop_priority = -1;
     ctx->priority = 0;
+    ctx->base_priority = 0;
+    ctx->base_flow = ctx->flow;
 
     if (process_special(ctx->ofproto, &ctx->flow, ctx->packet)) {
         ctx->may_set_up_flow = false;
@@ -3203,8 +3186,6 @@ xlate_actions(struct action_xlate_ctx *ctx,
         do_xlate_actions(in, n_in, ctx);
     }
 
-    remove_pop_action(ctx);
-
     /* Check with in-band control to see if we're allowed to set up this
      * flow. */
     if (!connmgr_may_set_up_flow(ctx->ofproto->up.connmgr, &ctx->flow,