ipsec: bind an SA to a worker

the sequence number increment and the anti-replay window
checks must be atomic. Given the vector nature of VPP we
can't simply use atomic increments for sequence numbers,
since a vector on thread 1 with lower sequence numbers could
be 'overtaken' by packets on thread 2 with higher sequence
numbers.
The anti-replay logic requires a critical section, not just
atomics, and we don't want that.
So when the SA see the first packet it is bound to that worker
all subsequent packets, that arrive on a different worker,
are subject to a handoff.

Type: feature

Change-Id: Ia20a8645fb50622ea6235ab015a537f033d531a4
Signed-off-by: Neale Ranns <nranns@cisco.com>

1 files changed, 39 insertions, 9 deletions

diff --git a/src/vnet/ipsec/esp_decrypt.c b/src/vnet/ipsec/esp_decrypt.c
index cddda1f8e54..a56a7843ef5 100644
--- a/src/vnet/ipsec/esp_decrypt.c
+++ b/src/vnet/ipsec/esp_decrypt.c
@@ -27,7 +27,8 @@
 #define foreach_esp_decrypt_next                \
 _(DROP, "error-drop")                           \
 _(IP4_INPUT, "ip4-input-no-checksum")           \
-_(IP6_INPUT, "ip6-input")
+_(IP6_INPUT, "ip6-input")                       \
+_(HANDOFF, "handoff")
 
 #define _(v, s) ESP_DECRYPT_NEXT_##v,
 typedef enum
@@ -177,6 +178,21 @@ esp_decrypt_inline (vlib_main_t * vm,
 	  cpd.sa_index = current_sa_index;
 	}
 
+      if (PREDICT_FALSE (~0 == sa0->decrypt_thread_index))
+	{
+	  /* this is the first packet to use this SA, claim the SA
+	   * for this thread. this could happen simultaneously on
+	   * another thread */
+	  clib_atomic_cmp_and_swap (&sa0->decrypt_thread_index, ~0,
+				    ipsec_sa_assign_thread (thread_index));
+	}
+
+      if (PREDICT_TRUE (thread_index != sa0->decrypt_thread_index))
+	{
+	  next[0] = ESP_DECRYPT_NEXT_HANDOFF;
+	  goto next;
+	}
+
       /* store packet data for next round for easier prefetch */
       pd->sa_data = cpd.sa_data;
       pd->current_data = b[0]->current_data;
@@ -595,9 +611,10 @@ VLIB_REGISTER_NODE (esp4_decrypt_node) = {
 
   .n_next_nodes = ESP_DECRYPT_N_NEXT,
   .next_nodes = {
-#define _(s,n) [ESP_DECRYPT_NEXT_##s] = n,
-    foreach_esp_decrypt_next
-#undef _
+    [ESP_DECRYPT_NEXT_DROP] = "ip4-drop",
+    [ESP_DECRYPT_NEXT_IP4_INPUT] = "ip4-input-no-checksum",
+    [ESP_DECRYPT_NEXT_IP6_INPUT] = "ip6-input",
+    [ESP_DECRYPT_NEXT_HANDOFF] = "esp4-decrypt-handoff",
   },
 };
 
@@ -612,9 +629,10 @@ VLIB_REGISTER_NODE (esp6_decrypt_node) = {
 
   .n_next_nodes = ESP_DECRYPT_N_NEXT,
   .next_nodes = {
-#define _(s,n) [ESP_DECRYPT_NEXT_##s] = n,
-    foreach_esp_decrypt_next
-#undef _
+    [ESP_DECRYPT_NEXT_DROP] = "ip6-drop",
+    [ESP_DECRYPT_NEXT_IP4_INPUT] = "ip4-input-no-checksum",
+    [ESP_DECRYPT_NEXT_IP6_INPUT] = "ip6-input",
+    [ESP_DECRYPT_NEXT_HANDOFF]=  "esp6-decrypt-handoff",
   },
 };
 
@@ -625,7 +643,13 @@ VLIB_REGISTER_NODE (esp4_decrypt_tun_node) = {
   .type = VLIB_NODE_TYPE_INTERNAL,
   .n_errors = ARRAY_LEN(esp_decrypt_error_strings),
   .error_strings = esp_decrypt_error_strings,
-  .sibling_of = "esp4-decrypt",
+  .n_next_nodes = ESP_DECRYPT_N_NEXT,
+  .next_nodes = {
+    [ESP_DECRYPT_NEXT_DROP] = "ip4-drop",
+    [ESP_DECRYPT_NEXT_IP4_INPUT] = "ip4-input-no-checksum",
+    [ESP_DECRYPT_NEXT_IP6_INPUT] = "ip6-input",
+    [ESP_DECRYPT_NEXT_HANDOFF] = "esp4-decrypt-handoff",
+  },
 };
 
 VLIB_REGISTER_NODE (esp6_decrypt_tun_node) = {
@@ -635,7 +659,13 @@ VLIB_REGISTER_NODE (esp6_decrypt_tun_node) = {
   .type = VLIB_NODE_TYPE_INTERNAL,
   .n_errors = ARRAY_LEN(esp_decrypt_error_strings),
   .error_strings = esp_decrypt_error_strings,
-  .sibling_of = "esp6-decrypt",
+  .n_next_nodes = ESP_DECRYPT_N_NEXT,
+  .next_nodes = {
+    [ESP_DECRYPT_NEXT_DROP] = "ip6-drop",
+    [ESP_DECRYPT_NEXT_IP4_INPUT] = "ip4-input-no-checksum",
+    [ESP_DECRYPT_NEXT_IP6_INPUT] = "ip6-input",
+    [ESP_DECRYPT_NEXT_HANDOFF]=  "esp6-decrypt-handoff",
+  },
 };
 /* *INDENT-ON* */