wireguard: add async mode for encryption packets

Originally wireguard doesn't support async mode for encryption packets.

This patch add async mode for encryption in wireguard and also adds
support chacha20-poly1305 algorithm in cryptodev for async handler.
In addition it contains new command line to activate async mode for wireguard:
  set wireguard async mode on|off

and also add new command to check active mode for wireguard:
  show wireguard mode

Type: improvement

Signed-off-by: Gabriel Oginski <gabrielx.oginski@intel.com>
Change-Id: I141d48b42ee8dbff0112b8542ab5205268089da6
(cherry picked from commit 492d7790ff26c569bee81617c662363652891140)

11 files changed, 616 insertions, 107 deletions

diff --git a/src/plugins/wireguard/wireguard.api b/src/plugins/wireguard/wireguard.api
index 1473d9ca39b..9a839deabd9 100644
--- a/src/plugins/wireguard/wireguard.api
+++ b/src/plugins/wireguard/wireguard.api
@@ -195,6 +195,17 @@ define wireguard_peers_details {
   vl_api_wireguard_peer_t peer;
 };
 
+/** \brief Wireguard Set Async mode
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param async_enable - wireguard async mode on or off
+*/
+autoreply define wg_set_async_mode {
+  u32 client_index;
+  u32 context;
+  bool async_enable;
+};
+
 /*
  * Local Variables:
  * eval: (c-set-style "gnu")
diff --git a/src/plugins/wireguard/wireguard.c b/src/plugins/wireguard/wireguard.c
index 8438cc126a6..40c2c090e9e 100644
--- a/src/plugins/wireguard/wireguard.c
+++ b/src/plugins/wireguard/wireguard.c
@@ -16,6 +16,7 @@
 #include <vnet/vnet.h>
 #include <vnet/plugin/plugin.h>
 #include <vpp/app/version.h>
+#include <vnet/crypto/crypto.h>
 
 #include <wireguard/wireguard_send.h>
 #include <wireguard/wireguard_key.h>
@@ -23,6 +24,31 @@
 #include <wireguard/wireguard.h>
 
 wg_main_t wg_main;
+wg_async_post_next_t wg_encrypt_async_next;
+
+void
+wg_set_async_mode (u32 is_enabled)
+{
+  vnet_crypto_request_async_mode (is_enabled);
+
+  if (is_enabled)
+    wg_op_mode_set_ASYNC ();
+  else
+    wg_op_mode_unset_ASYNC ();
+}
+
+static void
+wireguard_register_post_node (vlib_main_t *vm)
+{
+  wg_async_post_next_t *eit;
+
+  eit = &wg_encrypt_async_next;
+
+  eit->wg4_post_next =
+    vnet_crypto_register_post_node (vm, "wg4-output-tun-post-node");
+  eit->wg6_post_next =
+    vnet_crypto_register_post_node (vm, "wg6-output-tun-post-node");
+}
 
 static clib_error_t *
 wg_init (vlib_main_t * vm)
@@ -44,6 +70,8 @@ wg_init (vlib_main_t * vm)
 			CLIB_CACHE_LINE_BYTES);
 
   wg_timer_wheel_init ();
+  wireguard_register_post_node (vm);
+  wmp->op_mode_flags = 0;
 
   return (NULL);
 }
diff --git a/src/plugins/wireguard/wireguard.h b/src/plugins/wireguard/wireguard.h
index 4cbee1fcf7a..aaec3d47b23 100644
--- a/src/plugins/wireguard/wireguard.h
+++ b/src/plugins/wireguard/wireguard.h
@@ -30,6 +30,7 @@ typedef struct wg_per_thread_data_t_
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
   vnet_crypto_op_t *crypto_ops;
+  vnet_crypto_async_frame_t **async_frames;
   u8 data[WG_DEFAULT_DATA_SIZE];
 } wg_per_thread_data_t;
 typedef struct
@@ -50,12 +51,69 @@ typedef struct
   u8 feature_init;
 
   tw_timer_wheel_16t_2w_512sl_t timer_wheel;
+
+  /* operation mode flags (e.g. async) */
+  u8 op_mode_flags;
 } wg_main_t;
 
+typedef struct
+{
+  /* wg post node index for async crypto */
+  u32 wg4_post_next;
+  u32 wg6_post_next;
+} wg_async_post_next_t;
+
+extern wg_async_post_next_t wg_encrypt_async_next;
 extern wg_main_t wg_main;
 
+/**
+ * Wireguard operation mode
+ **/
+#define foreach_wg_op_mode_flags _ (0, ASYNC, "async")
+
+/**
+ * Helper function to set/unset and check op modes
+ **/
+typedef enum wg_op_mode_flags_t_
+{
+#define _(v, f, s) WG_OP_MODE_FLAG_##f = 1 << v,
+  foreach_wg_op_mode_flags
+#undef _
+} __clib_packed wg_op_mode_flags_t;
+
+#define _(a, v, s)                                                            \
+  always_inline int wg_op_mode_set_##v (void)                                 \
+  {                                                                           \
+    return (wg_main.op_mode_flags |= WG_OP_MODE_FLAG_##v);                    \
+  }                                                                           \
+  always_inline int wg_op_mode_unset_##v (void)                               \
+  {                                                                           \
+    return (wg_main.op_mode_flags &= ~WG_OP_MODE_FLAG_##v);                   \
+  }                                                                           \
+  always_inline int wg_op_mode_is_set_##v (void)                              \
+  {                                                                           \
+    return (wg_main.op_mode_flags & WG_OP_MODE_FLAG_##v);                     \
+  }
+foreach_wg_op_mode_flags
+#undef _
+
+  typedef struct
+{
+  u8 __pad[22];
+  u16 next_index;
+} wg_post_data_t;
+
+STATIC_ASSERT (sizeof (wg_post_data_t) <=
+		 STRUCT_SIZE_OF (vnet_buffer_opaque_t, unused),
+	       "Custom meta-data too large for vnet_buffer_opaque_t");
+
+#define wg_post_data(b)                                                       \
+  ((wg_post_data_t *) ((u8 *) ((b)->opaque) +                                 \
+		       STRUCT_OFFSET_OF (vnet_buffer_opaque_t, unused)))
+
 #define WG_START_EVENT	1
 void wg_feature_init (wg_main_t * wmp);
+void wg_set_async_mode (u32 is_enabled);
 
 #endif /* __included_wg_h__ */
 
diff --git a/src/plugins/wireguard/wireguard_api.c b/src/plugins/wireguard/wireguard_api.c
index 4cb40a15231..67f68d069f1 100644
--- a/src/plugins/wireguard/wireguard_api.c
+++ b/src/plugins/wireguard/wireguard_api.c
@@ -365,6 +365,18 @@ wg_api_peer_event (index_t peeri, wg_peer_flags flags)
     };
 }
 
+static void
+vl_api_wg_set_async_mode_t_handler (vl_api_wg_set_async_mode_t *mp)
+{
+  wg_main_t *wmp = &wg_main;
+  vl_api_wg_set_async_mode_reply_t *rmp;
+  int rv = 0;
+
+  wg_set_async_mode (mp->async_enable);
+
+  REPLY_MACRO (VL_API_WG_SET_ASYNC_MODE_REPLY);
+}
+
 /* set tup the API message handling tables */
 #include <wireguard/wireguard.api.c>
 static clib_error_t *
diff --git a/src/plugins/wireguard/wireguard_cli.c b/src/plugins/wireguard/wireguard_cli.c
index f821a3c313a..5e0b7243a83 100644
--- a/src/plugins/wireguard/wireguard_cli.c
+++ b/src/plugins/wireguard/wireguard_cli.c
@@ -356,6 +356,61 @@ VLIB_CLI_COMMAND (wg_show_itfs_command, static) =
   .short_help = "show wireguard",
   .function = wg_show_if_command_fn,
 };
+
+static clib_error_t *
+wg_set_async_mode_command_fn (vlib_main_t *vm, unformat_input_t *input,
+			      vlib_cli_command_t *cmd)
+{
+  unformat_input_t _line_input, *line_input = &_line_input;
+  int async_enable = 0;
+
+  if (!unformat_user (input, unformat_line_input, line_input))
+    return 0;
+
+  while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (line_input, "on"))
+	async_enable = 1;
+      else if (unformat (line_input, "off"))
+	async_enable = 0;
+      else
+	return (clib_error_return (0, "unknown input '%U'",
+				   format_unformat_error, line_input));
+    }
+
+  wg_set_async_mode (async_enable);
+
+  unformat_free (line_input);
+  return (NULL);
+}
+
+VLIB_CLI_COMMAND (wg_set_async_mode_command, static) = {
+  .path = "set wireguard async mode",
+  .short_help = "set wireguard async mode on|off",
+  .function = wg_set_async_mode_command_fn,
+};
+
+static clib_error_t *
+wg_show_mode_command_fn (vlib_main_t *vm, unformat_input_t *input,
+			 vlib_cli_command_t *cmd)
+{
+  vlib_cli_output (vm, "Wireguard mode");
+
+#define _(v, f, s)                                                            \
+  vlib_cli_output (vm, "\t%s: %s", s,                                         \
+		   (wg_op_mode_is_set_##f () ? "enabled" : "disabled"));
+  foreach_wg_op_mode_flags
+#undef _
+
+    return (NULL);
+}
+
+VLIB_CLI_COMMAND (wg_show_modemode_command, static) = {
+  .path = "show wireguard mode",
+  .short_help = "show wireguard mode",
+  .function = wg_show_mode_command_fn,
+};
+
 /* *INDENT-ON* */
 
 /*
diff --git a/src/plugins/wireguard/wireguard_noise.c b/src/plugins/wireguard/wireguard_noise.c
index c8605f117cd..6efec28fef8 100644
--- a/src/plugins/wireguard/wireguard_noise.c
+++ b/src/plugins/wireguard/wireguard_noise.c
@@ -630,58 +630,6 @@ error:
 }
 
 enum noise_state_crypt
-noise_sync_remote_encrypt (vlib_main_t *vm, vnet_crypto_op_t **crypto_ops,
-			   noise_remote_t *r, uint32_t *r_idx, uint64_t *nonce,
-			   uint8_t *src, size_t srclen, uint8_t *dst, u32 bi,
-			   u8 *iv, f64 time)
-{
-  noise_keypair_t *kp;
-  enum noise_state_crypt ret = SC_FAILED;
-
-  if ((kp = r->r_current) == NULL)
-    goto error;
-
-  /* We confirm that our values are within our tolerances. We want:
-   *  - a valid keypair
-   *  - our keypair to be less than REJECT_AFTER_TIME seconds old
-   *  - our receive counter to be less than REJECT_AFTER_MESSAGES
-   *  - our send counter to be less than REJECT_AFTER_MESSAGES
-   */
-  if (!kp->kp_valid ||
-      wg_birthdate_has_expired_opt (kp->kp_birthdate, REJECT_AFTER_TIME,
-				    time) ||
-      kp->kp_ctr.c_recv >= REJECT_AFTER_MESSAGES ||
-      ((*nonce = noise_counter_send (&kp->kp_ctr)) > REJECT_AFTER_MESSAGES))
-    goto error;
-
-  /* We encrypt into the same buffer, so the caller must ensure that buf
-   * has NOISE_AUTHTAG_LEN bytes to store the MAC. The nonce and index
-   * are passed back out to the caller through the provided data pointer. */
-  *r_idx = kp->kp_remote_index;
-
-  wg_prepare_sync_op (vm, crypto_ops, src, srclen, dst, NULL, 0, *nonce,
-		      VNET_CRYPTO_OP_CHACHA20_POLY1305_ENC, kp->kp_send_index,
-		      bi, iv);
-
-  /* If our values are still within tolerances, but we are approaching
-   * the tolerances, we notify the caller with ESTALE that they should
-   * establish a new keypair. The current keypair can continue to be used
-   * until the tolerances are hit. We notify if:
-   *  - our send counter is valid and not less than REKEY_AFTER_MESSAGES
-   *  - we're the initiator and our keypair is older than
-   *    REKEY_AFTER_TIME seconds */
-  ret = SC_KEEP_KEY_FRESH;
-  if ((kp->kp_valid && *nonce >= REKEY_AFTER_MESSAGES) ||
-      (kp->kp_is_initiator && wg_birthdate_has_expired_opt (
-				kp->kp_birthdate, REKEY_AFTER_TIME, time)))
-    goto error;
-
-  ret = SC_OK;
-error:
-  return ret;
-}
-
-enum noise_state_crypt
 noise_sync_remote_decrypt (vlib_main_t *vm, vnet_crypto_op_t **crypto_ops,
 			   noise_remote_t *r, uint32_t r_idx, uint64_t nonce,
 			   uint8_t *src, size_t srclen, uint8_t *dst, u32 bi,
@@ -791,14 +739,6 @@ noise_remote_handshake_index_drop (noise_remote_t * r)
     u->u_index_drop (hs->hs_local_index);
 }
 
-static uint64_t
-noise_counter_send (noise_counter_t * ctr)
-{
-  uint64_t ret;
-  ret = ctr->c_send++;
-  return ret;
-}
-
 static void
 noise_kdf (uint8_t * a, uint8_t * b, uint8_t * c, const uint8_t * x,
 	   size_t a_len, size_t b_len, size_t c_len, size_t x_len,
diff --git a/src/plugins/wireguard/wireguard_noise.h b/src/plugins/wireguard/wireguard_noise.h
index ef1e7dcbfca..33ac189b0c0 100644
--- a/src/plugins/wireguard/wireguard_noise.h
+++ b/src/plugins/wireguard/wireguard_noise.h
@@ -136,6 +136,14 @@ noise_local_get (uint32_t locali)
   return (pool_elt_at_index (noise_local_pool, locali));
 }
 
+static_always_inline uint64_t
+noise_counter_send (noise_counter_t *ctr)
+{
+  uint64_t ret;
+  ret = ctr->c_send++;
+  return ret;
+}
+
 void noise_local_init (noise_local_t *, struct noise_upcall *);
 bool noise_local_set_private (noise_local_t *,
 			      const uint8_t[NOISE_PUBLIC_KEY_LEN]);
@@ -189,12 +197,6 @@ noise_remote_encrypt (vlib_main_t * vm, noise_remote_t *,
 		      uint8_t * src, size_t srclen, uint8_t * dst);
 
 enum noise_state_crypt
-noise_sync_remote_encrypt (vlib_main_t *vm, vnet_crypto_op_t **crypto_ops,
-			   noise_remote_t *r, uint32_t *r_idx, uint64_t *nonce,
-			   uint8_t *src, size_t srclen, uint8_t *dst, u32 bi,
-			   u8 *iv, f64 time);
-
-enum noise_state_crypt
 noise_sync_remote_decrypt (vlib_main_t *vm, vnet_crypto_op_t **crypto_ops,
 			   noise_remote_t *, uint32_t r_idx, uint64_t nonce,
 			   uint8_t *src, size_t srclen, uint8_t *dst, u32 bi,
diff --git a/src/plugins/wireguard/wireguard_output_tun.c b/src/plugins/wireguard/wireguard_output_tun.c
index 2feb0570a31..f7ae4d174c0 100644
--- a/src/plugins/wireguard/wireguard_output_tun.c
+++ b/src/plugins/wireguard/wireguard_output_tun.c
@@ -21,11 +21,12 @@
 #include <wireguard/wireguard.h>
 #include <wireguard/wireguard_send.h>
 
-#define foreach_wg_output_error                                         \
- _(NONE, "No error")							\
- _(PEER, "Peer error")                                                  \
- _(KEYPAIR, "Keypair error")                                            \
- _(TOO_BIG, "packet too big")                                           \
+#define foreach_wg_output_error                                               \
+  _ (NONE, "No error")                                                        \
+  _ (PEER, "Peer error")                                                      \
+  _ (KEYPAIR, "Keypair error")                                                \
+  _ (TOO_BIG, "packet too big")                                               \
+  _ (CRYPTO_ENGINE_ERROR, "crypto engine error (packet dropped)")
 
 typedef enum
 {
@@ -56,6 +57,12 @@ typedef struct
   u8 is_ip4;
 } wg_output_tun_trace_t;
 
+typedef struct
+{
+  index_t peer;
+  u32 next_index;
+} wg_output_tun_post_trace_t;
+
 u8 *
 format_ip4_udp_header (u8 * s, va_list * args)
 {
@@ -93,6 +100,47 @@ format_wg_output_tun_trace (u8 * s, va_list * args)
   return s;
 }
 
+/* post node - packet trace format function */
+static u8 *
+format_wg_output_tun_post_trace (u8 *s, va_list *args)
+{
+  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
+  CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
+
+  wg_output_tun_post_trace_t *t = va_arg (*args, wg_output_tun_post_trace_t *);
+
+  s = format (s, "peer: %d\n", t->peer);
+  s = format (s, "  wg-post: next node index %u", t->next_index);
+  return s;
+}
+
+static_always_inline void
+wg_prepare_sync_enc_op (vlib_main_t *vm, vnet_crypto_op_t **crypto_ops,
+			u8 *src, u32 src_len, u8 *dst, u8 *aad, u32 aad_len,
+			u64 nonce, vnet_crypto_key_index_t key_index, u32 bi,
+			u8 *iv)
+{
+  vnet_crypto_op_t _op, *op = &_op;
+  u8 src_[] = {};
+
+  clib_memset (iv, 0, 4);
+  clib_memcpy (iv + 4, &nonce, sizeof (nonce));
+
+  vec_add2_aligned (crypto_ops[0], op, 1, CLIB_CACHE_LINE_BYTES);
+  vnet_crypto_op_init (op, VNET_CRYPTO_OP_CHACHA20_POLY1305_ENC);
+
+  op->tag_len = NOISE_AUTHTAG_LEN;
+  op->tag = dst + src_len;
+  op->src = !src ? src_ : src;
+  op->len = src_len;
+  op->dst = dst;
+  op->key_index = key_index;
+  op->aad = aad;
+  op->aad_len = aad_len;
+  op->iv = iv;
+  op->user_data = bi;
+}
+
 static_always_inline void
 wg_output_process_ops (vlib_main_t *vm, vlib_node_runtime_t *node,
 		       vnet_crypto_op_t *ops, vlib_buffer_t *b[], u16 *nexts,
@@ -121,10 +169,148 @@ wg_output_process_ops (vlib_main_t *vm, vlib_node_runtime_t *node,
     }
 }
 
+static_always_inline void
+wg_output_tun_add_to_frame (vlib_main_t *vm, vnet_crypto_async_frame_t *f,
+			    u32 key_index, u32 crypto_len,
+			    i16 crypto_start_offset, u32 buffer_index,
+			    u16 next_node, u8 *iv, u8 *tag, u8 flags)
+{
+  vnet_crypto_async_frame_elt_t *fe;
+  u16 index;
+
+  ASSERT (f->n_elts < VNET_CRYPTO_FRAME_SIZE);
+
+  index = f->n_elts;
+  fe = &f->elts[index];
+  f->n_elts++;
+  fe->key_index = key_index;
+  fe->crypto_total_length = crypto_len;
+  fe->crypto_start_offset = crypto_start_offset;
+  fe->iv = iv;
+  fe->tag = tag;
+  fe->flags = flags;
+  f->buffer_indices[index] = buffer_index;
+  f->next_node_index[index] = next_node;
+}
+
+static_always_inline enum noise_state_crypt
+wq_output_tun_process (vlib_main_t *vm, vnet_crypto_op_t **crypto_ops,
+		       noise_remote_t *r, uint32_t *r_idx, uint64_t *nonce,
+		       uint8_t *src, size_t srclen, uint8_t *dst, u32 bi,
+		       u8 *iv, f64 time)
+{
+  noise_keypair_t *kp;
+  enum noise_state_crypt ret = SC_FAILED;
+
+  if ((kp = r->r_current) == NULL)
+    goto error;
+
+  /* We confirm that our values are within our tolerances. We want:
+   *  - a valid keypair
+   *  - our keypair to be less than REJECT_AFTER_TIME seconds old
+   *  - our receive counter to be less than REJECT_AFTER_MESSAGES
+   *  - our send counter to be less than REJECT_AFTER_MESSAGES
+   */
+  if (!kp->kp_valid ||
+      wg_birthdate_has_expired_opt (kp->kp_birthdate, REJECT_AFTER_TIME,
+				    time) ||
+      kp->kp_ctr.c_recv >= REJECT_AFTER_MESSAGES ||
+      ((*nonce = noise_counter_send (&kp->kp_ctr)) > REJECT_AFTER_MESSAGES))
+    goto error;
+
+  /* We encrypt into the same buffer, so the caller must ensure that buf
+   * has NOISE_AUTHTAG_LEN bytes to store the MAC. The nonce and index
+   * are passed back out to the caller through the provided data pointer. */
+  *r_idx = kp->kp_remote_index;
+
+  wg_prepare_sync_enc_op (vm, crypto_ops, src, srclen, dst, NULL, 0, *nonce,
+			  kp->kp_send_index, bi, iv);
+
+  /* If our values are still within tolerances, but we are approaching
+   * the tolerances, we notify the caller with ESTALE that they should
+   * establish a new keypair. The current keypair can continue to be used
+   * until the tolerances are hit. We notify if:
+   *  - our send counter is valid and not less than REKEY_AFTER_MESSAGES
+   *  - we're the initiator and our keypair is older than
+   *    REKEY_AFTER_TIME seconds */
+  ret = SC_KEEP_KEY_FRESH;
+  if ((kp->kp_valid && *nonce >= REKEY_AFTER_MESSAGES) ||
+      (kp->kp_is_initiator && wg_birthdate_has_expired_opt (
+				kp->kp_birthdate, REKEY_AFTER_TIME, time)))
+    goto error;
+
+  ret = SC_OK;
+error:
+  return ret;
+}
+
+static_always_inline enum noise_state_crypt
+wg_add_to_async_frame (vlib_main_t *vm, wg_per_thread_data_t *ptd,
+		       vnet_crypto_async_frame_t *async_frame,
+		       vlib_buffer_t *b, u8 *payload, u32 payload_len, u32 bi,
+		       u16 next, u16 async_next, noise_remote_t *r,
+		       uint32_t *r_idx, uint64_t *nonce, u8 *iv, f64 time)
+{
+  wg_post_data_t *post = wg_post_data (b);
+  u8 flag = 0;
+  noise_keypair_t *kp;
+
+  post->next_index = next;
+
+  /* crypto */
+  enum noise_state_crypt ret = SC_FAILED;
+
+  if ((kp = r->r_current) == NULL)
+    goto error;
+
+  /* We confirm that our values are within our tolerances. We want:
+   *  - a valid keypair
+   *  - our keypair to be less than REJECT_AFTER_TIME seconds old
+   *  - our receive counter to be less than REJECT_AFTER_MESSAGES
+   *  - our send counter to be less than REJECT_AFTER_MESSAGES
+   */
+  if (!kp->kp_valid ||
+      wg_birthdate_has_expired_opt (kp->kp_birthdate, REJECT_AFTER_TIME,
+				    time) ||
+      kp->kp_ctr.c_recv >= REJECT_AFTER_MESSAGES ||
+      ((*nonce = noise_counter_send (&kp->kp_ctr)) > REJECT_AFTER_MESSAGES))
+    goto error;
+
+  /* We encrypt into the same buffer, so the caller must ensure that buf
+   * has NOISE_AUTHTAG_LEN bytes to store the MAC. The nonce and index
+   * are passed back out to the caller through the provided data pointer. */
+  *r_idx = kp->kp_remote_index;
+
+  clib_memset (iv, 0, 4);
+  clib_memcpy (iv + 4, nonce, sizeof (nonce));
+
+  /* this always succeeds because we know the frame is not full */
+  wg_output_tun_add_to_frame (vm, async_frame, kp->kp_send_index, payload_len,
+			      payload - b->data, bi, async_next, iv,
+			      payload + payload_len, flag);
+
+  /* If our values are still within tolerances, but we are approaching
+   * the tolerances, we notify the caller with ESTALE that they should
+   * establish a new keypair. The current keypair can continue to be used
+   * until the tolerances are hit. We notify if:
+   *  - our send counter is valid and not less than REKEY_AFTER_MESSAGES
+   *  - we're the initiator and our keypair is older than
+   *    REKEY_AFTER_TIME seconds */
+  ret = SC_KEEP_KEY_FRESH;
+  if ((kp->kp_valid && *nonce >= REKEY_AFTER_MESSAGES) ||
+      (kp->kp_is_initiator && wg_birthdate_has_expired_opt (
+				kp->kp_birthdate, REKEY_AFTER_TIME, time)))
+    goto error;
+
+  ret = SC_OK;
+error:
+  return ret;
+}
+
 /* is_ip4 - inner header flag */
 always_inline uword
 wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
-		      vlib_frame_t *frame, u8 is_ip4)
+		      vlib_frame_t *frame, u8 is_ip4, u16 async_next_node)
 {
   wg_main_t *wmp = &wg_main;
   wg_per_thread_data_t *ptd =
@@ -140,10 +326,18 @@ wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
   vlib_buffer_t *sync_bufs[VLIB_FRAME_SIZE];
   u32 thread_index = vm->thread_index;
   u16 n_sync = 0;
-  u16 drop_next = WG_OUTPUT_NEXT_ERROR;
+  const u16 drop_next = WG_OUTPUT_NEXT_ERROR;
+  const u8 is_async = wg_op_mode_is_set_ASYNC ();
+  vnet_crypto_async_frame_t *async_frame = NULL;
+  u16 n_async = 0;
+  u16 noop_nexts[VLIB_FRAME_SIZE], *noop_next = noop_nexts, n_noop = 0;
+  u16 err = !0;
+  u32 sync_bi[VLIB_FRAME_SIZE];
+  u32 noop_bi[VLIB_FRAME_SIZE];
 
   vlib_get_buffers (vm, from, bufs, n_left_from);
   vec_reset_length (ptd->crypto_ops);
+  vec_reset_length (ptd->async_frames);
 
   wg_peer_t *peer = NULL;
   u32 adj_index = 0;
@@ -169,7 +363,8 @@ wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
 			 LOAD);
 	}
 
-      next[0] = WG_OUTPUT_NEXT_ERROR;
+      noop_next[0] = WG_OUTPUT_NEXT_ERROR;
+      err = WG_OUTPUT_NEXT_ERROR;
 
       adj_index = vnet_buffer (b[0])->ip.adj_index[VLIB_TX];
 
@@ -193,9 +388,10 @@ wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
 				    wg_peer_assign_thread (thread_index));
 	}
 
-      if (PREDICT_TRUE (thread_index != peer->output_thread_index))
+      if (PREDICT_FALSE (thread_index != peer->output_thread_index))
 	{
-	  next[0] = WG_OUTPUT_NEXT_HANDOFF;
+	  noop_next[0] = WG_OUTPUT_NEXT_HANDOFF;
+	  err = WG_OUTPUT_NEXT_HANDOFF;
 	  goto next;
 	}
 
@@ -245,12 +441,35 @@ wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
 	  last_adj_index = adj_index;
 	}
 
+      /* Here we are sure that can send packet to next node */
+      next[0] = WG_OUTPUT_NEXT_INTERFACE_OUTPUT;
+
       enum noise_state_crypt state;
 
-      state = noise_sync_remote_encrypt (
-	vm, crypto_ops, &peer->remote, &message_data_wg->receiver_index,
-	&message_data_wg->counter, plain_data, plain_data_len, plain_data,
-	n_sync, iv_data, time);
+      if (is_async)
+	{
+	  /* get a frame for this op if we don't yet have one or it's full  */
+	  if (NULL == async_frame ||
+	      vnet_crypto_async_frame_is_full (async_frame))
+	    {
+	      async_frame = vnet_crypto_async_get_frame (
+		vm, VNET_CRYPTO_OP_CHACHA20_POLY1305_TAG16_AAD0_ENC);
+	      /* Save the frame to the list we'll submit at the end */
+	      vec_add1 (ptd->async_frames, async_frame);
+	    }
+	  state = wg_add_to_async_frame (
+	    vm, ptd, async_frame, b[0], plain_data, plain_data_len,
+	    from[b - bufs], next[0], async_next_node, &peer->remote,
+	    &message_data_wg->receiver_index, &message_data_wg->counter,
+	    iv_data, time);
+	}
+      else
+	{
+	  state = wq_output_tun_process (
+	    vm, crypto_ops, &peer->remote, &message_data_wg->receiver_index,
+	    &message_data_wg->counter, plain_data, plain_data_len, plain_data,
+	    n_sync, iv_data, time);
+	}
 
       if (PREDICT_FALSE (state == SC_KEEP_KEY_FRESH))
 	{
@@ -261,11 +480,11 @@ wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
 	  // TODO: Maybe wrong
 	  wg_send_handshake_from_mt (peeri, false);
 	  wg_peer_update_flags (peeri, WG_PEER_ESTABLISHED, false);
+	  noop_next[0] = WG_OUTPUT_NEXT_ERROR;
 	  goto out;
 	}
 
-      /* Here we are sure that can send packet to next node */
-      next[0] = WG_OUTPUT_NEXT_INTERFACE_OUTPUT;
+      err = WG_OUTPUT_NEXT_INTERFACE_OUTPUT;
 
       if (is_ip4_out)
 	{
@@ -304,31 +523,214 @@ wg_output_tun_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
 	}
 
     next:
-      sync_bufs[n_sync] = b[0];
-      n_sync += 1;
+      if (PREDICT_FALSE (err != WG_OUTPUT_NEXT_INTERFACE_OUTPUT))
+	{
+	  noop_bi[n_noop] = from[b - bufs];
+	  n_noop++;
+	  noop_next++;
+	  goto next_left;
+	}
+      if (!is_async)
+	{
+	  sync_bi[n_sync] = from[b - bufs];
+	  sync_bufs[n_sync] = b[0];
+	  n_sync += 1;
+	  next += 1;
+	}
+      else
+	{
+	  n_async++;
+	}
+    next_left:
       n_left_from -= 1;
-      next += 1;
       b += 1;
     }
 
-  /* wg-output-process-ops */
-  wg_output_process_ops (vm, node, ptd->crypto_ops, sync_bufs, nexts,
-			 drop_next);
+  if (n_sync)
+    {
+      /* wg-output-process-ops */
+      wg_output_process_ops (vm, node, ptd->crypto_ops, sync_bufs, nexts,
+			     drop_next);
+      vlib_buffer_enqueue_to_next (vm, node, sync_bi, nexts, n_sync);
+    }
+  if (n_async)
+    {
+      /* submit all of the open frames */
+      vnet_crypto_async_frame_t **async_frame;
+
+      vec_foreach (async_frame, ptd->async_frames)
+	{
+	  if (PREDICT_FALSE (
+		vnet_crypto_async_submit_open_frame (vm, *async_frame) < 0))
+	    {
+	      u32 n_drop = (*async_frame)->n_elts;
+	      u32 *bi = (*async_frame)->buffer_indices;
+	      u16 index = n_noop;
+	      while (n_drop--)
+		{
+		  noop_bi[index] = bi[0];
+		  vlib_buffer_t *b = vlib_get_buffer (vm, bi[0]);
+		  noop_nexts[index] = drop_next;
+		  b->error = node->errors[WG_OUTPUT_ERROR_CRYPTO_ENGINE_ERROR];
+		  bi++;
+		  index++;
+		}
+	      n_noop += (*async_frame)->n_elts;
+
+	      vnet_crypto_async_reset_frame (*async_frame);
+	      vnet_crypto_async_free_frame (vm, *async_frame);
+	    }
+	}
+    }
+  if (n_noop)
+    {
+      vlib_buffer_enqueue_to_next (vm, node, noop_bi, noop_nexts, n_noop);
+    }
 
-  vlib_buffer_enqueue_to_next (vm, node, from, nexts, n_sync);
   return frame->n_vectors;
 }
 
+always_inline uword
+wg_output_tun_post (vlib_main_t *vm, vlib_node_runtime_t *node,
+		    vlib_frame_t *frame)
+{
+  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b = bufs;
+  u16 nexts[VLIB_FRAME_SIZE], *next = nexts;
+  u32 *from = vlib_frame_vector_args (frame);
+  u32 n_left = frame->n_vectors;
+
+  index_t peeri = ~0;
+
+  vlib_get_buffers (vm, from, b, n_left);
+
+  if (n_left >= 4)
+    {
+      vlib_prefetch_buffer_header (b[0], LOAD);
+      vlib_prefetch_buffer_header (b[1], LOAD);
+      vlib_prefetch_buffer_header (b[2], LOAD);
+      vlib_prefetch_buffer_header (b[3], LOAD);
+    }
+
+  while (n_left > 8)
+    {
+      vlib_prefetch_buffer_header (b[4], LOAD);
+      vlib_prefetch_buffer_header (b[5], LOAD);
+      vlib_prefetch_buffer_header (b[6], LOAD);
+      vlib_prefetch_buffer_header (b[7], LOAD);
+
+      next[0] = (wg_post_data (b[0]))->next_index;
+      next[1] = (wg_post_data (b[1]))->next_index;
+      next[2] = (wg_post_data (b[2]))->next_index;
+      next[3] = (wg_post_data (b[3]))->next_index;
+
+      if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE))
+	{
+	  if (b[0]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      wg_output_tun_post_trace_t *tr =
+		vlib_add_trace (vm, node, b[0], sizeof (*tr));
+	      peeri = wg_peer_get_by_adj_index (
+		vnet_buffer (b[0])->ip.adj_index[VLIB_TX]);
+	      tr->peer = peeri;
+	      tr->next_index = next[0];
+	    }
+	  if (b[1]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      wg_output_tun_post_trace_t *tr =
+		vlib_add_trace (vm, node, b[1], sizeof (*tr));
+	      peeri = wg_peer_get_by_adj_index (
+		vnet_buffer (b[1])->ip.adj_index[VLIB_TX]);
+	      tr->next_index = next[1];
+	    }
+	  if (b[2]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      wg_output_tun_post_trace_t *tr =
+		vlib_add_trace (vm, node, b[2], sizeof (*tr));
+	      peeri = wg_peer_get_by_adj_index (
+		vnet_buffer (b[2])->ip.adj_index[VLIB_TX]);
+	      tr->next_index = next[2];
+	    }
+	  if (b[3]->flags & VLIB_BUFFER_IS_TRACED)
+	    {
+	      wg_output_tun_post_trace_t *tr =
+		vlib_add_trace (vm, node, b[3], sizeof (*tr));
+	      peeri = wg_peer_get_by_adj_index (
+		vnet_buffer (b[3])->ip.adj_index[VLIB_TX]);
+	      tr->next_index = next[3];
+	    }
+	}
+
+      b += 4;
+      next += 4;
+      n_left -= 4;
+    }
+
+  while (n_left > 0)
+    {
+      next[0] = (wg_post_data (b[0]))->next_index;
+      if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE) &&
+			 (b[0]->flags & VLIB_BUFFER_IS_TRACED)))
+	{
+	  wg_output_tun_post_trace_t *tr =
+	    vlib_add_trace (vm, node, b[0], sizeof (*tr));
+	  peeri = wg_peer_get_by_adj_index (
+	    vnet_buffer (b[0])->ip.adj_index[VLIB_TX]);
+	  tr->next_index = next[0];
+	}
+
+      b += 1;
+      next += 1;
+      n_left -= 1;
+    }
+
+  vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
+  return frame->n_vectors;
+}
+
+VLIB_REGISTER_NODE (wg4_output_tun_post_node) = {
+  .name = "wg4-output-tun-post-node",
+  .vector_size = sizeof (u32),
+  .format_trace = format_wg_output_tun_post_trace,
+  .type = VLIB_NODE_TYPE_INTERNAL,
+  .sibling_of = "wg4-output-tun",
+  .n_errors = ARRAY_LEN (wg_output_error_strings),
+  .error_strings = wg_output_error_strings,
+};
+
+VLIB_REGISTER_NODE (wg6_output_tun_post_node) = {
+  .name = "wg6-output-tun-post-node",
+  .vector_size = sizeof (u32),
+  .format_trace = format_wg_output_tun_post_trace,
+  .type = VLIB_NODE_TYPE_INTERNAL,
+  .sibling_of = "wg6-output-tun",
+  .n_errors = ARRAY_LEN (wg_output_error_strings),
+  .error_strings = wg_output_error_strings,
+};
+
+VLIB_NODE_FN (wg4_output_tun_post_node)
+(vlib_main_t *vm, vlib_node_runtime_t *node, vlib_frame_t *from_frame)
+{
+  return wg_output_tun_post (vm, node, from_frame);
+}
+
+VLIB_NODE_FN (wg6_output_tun_post_node)
+(vlib_main_t *vm, vlib_node_runtime_t *node, vlib_frame_t *from_frame)
+{
+  return wg_output_tun_post (vm, node, from_frame);
+}
+
 VLIB_NODE_FN (wg4_output_tun_node)
 (vlib_main_t *vm, vlib_node_runtime_t *node, vlib_frame_t *frame)
 {
-  return wg_output_tun_inline (vm, node, frame, /* is_ip4 */ 1);
+  return wg_output_tun_inline (vm, node, frame, /* is_ip4 */ 1,
+			       wg_encrypt_async_next.wg4_post_next);
 }
 
 VLIB_NODE_FN (wg6_output_tun_node)
 (vlib_main_t *vm, vlib_node_runtime_t *node, vlib_frame_t *frame)
 {
-  return wg_output_tun_inline (vm, node, frame, /* is_ip4 */ 0);
+  return wg_output_tun_inline (vm, node, frame, /* is_ip4 */ 0,
+			       wg_encrypt_async_next.wg6_post_next);
 }
 
 /* *INDENT-OFF* */
diff --git a/src/plugins/wireguard/wireguard_timer.c b/src/plugins/wireguard/wireguard_timer.c
index d2ae7ebf93e..b95801122fc 100644
--- a/src/plugins/wireguard/wireguard_timer.c
+++ b/src/plugins/wireguard/wireguard_timer.c
@@ -73,7 +73,7 @@ start_timer_thread_fn (void *arg)
   return 0;
 }
 
-static void
+static_always_inline void
 start_timer_from_mt (u32 peer_idx, u32 timer_id, u32 interval_ticks)
 {
   wg_timers_args a = {
@@ -204,8 +204,8 @@ wg_expired_zero_key_material (vlib_main_t * vm, wg_peer_t * peer)
     }
 }
 
-void
-wg_timers_any_authenticated_packet_traversal (wg_peer_t * peer)
+inline void
+wg_timers_any_authenticated_packet_traversal (wg_peer_t *peer)
 {
   if (peer->persistent_keepalive_interval)
     {
@@ -221,7 +221,7 @@ wg_timers_any_authenticated_packet_sent (wg_peer_t * peer)
   peer->last_sent_packet = vlib_time_now (vlib_get_main ());
 }
 
-void
+inline void
 wg_timers_any_authenticated_packet_sent_opt (wg_peer_t *peer, f64 time)
 {
   peer->last_sent_packet = time;
@@ -259,7 +259,7 @@ wg_timers_data_sent (wg_peer_t * peer)
 		       peer->new_handshake_interval_tick);
 }
 
-void
+inline void
 wg_timers_data_sent_opt (wg_peer_t *peer, f64 time)
 {
   peer->new_handshake_interval_tick =
@@ -299,7 +299,7 @@ wg_timers_any_authenticated_packet_received (wg_peer_t * peer)
   peer->last_received_packet = vlib_time_now (vlib_get_main ());
 }
 
-void
+inline void
 wg_timers_any_authenticated_packet_received_opt (wg_peer_t *peer, f64 time)
 {
   peer->last_received_packet = time;
diff --git a/src/plugins/wireguard/wireguard_timer.h b/src/plugins/wireguard/wireguard_timer.h
index cc8e123f3a2..9d5c071c86e 100644
--- a/src/plugins/wireguard/wireguard_timer.h
+++ b/src/plugins/wireguard/wireguard_timer.h
@@ -61,7 +61,7 @@ wg_birthdate_has_expired (f64 birthday_seconds, f64 expiration_seconds)
   return (birthday_seconds + expiration_seconds) < now_seconds;
 }
 
-static inline bool
+static_always_inline bool
 wg_birthdate_has_expired_opt (f64 birthday_seconds, f64 expiration_seconds,
 			      f64 time)
 {
diff --git a/src/vnet/crypto/crypto.h b/src/vnet/crypto/crypto.h
index 73b7f70ec46..eb381187f29 100644
--- a/src/vnet/crypto/crypto.h
+++ b/src/vnet/crypto/crypto.h
@@ -82,15 +82,16 @@ typedef enum
 /** async crypto **/
 
 /* CRYPTO_ID, PRETTY_NAME, KEY_LENGTH_IN_BYTES, TAG_LEN, AAD_LEN */
-#define foreach_crypto_aead_async_alg \
-  _(AES_128_GCM, "aes-128-gcm-aad8", 16, 16, 8) \
-  _(AES_128_GCM, "aes-128-gcm-aad12", 16, 16, 12) \
-  _(AES_192_GCM, "aes-192-gcm-aad8", 24, 16, 8) \
-  _(AES_192_GCM, "aes-192-gcm-aad12", 24, 16, 12) \
-  _(AES_256_GCM, "aes-256-gcm-aad8", 32, 16, 8) \
-  _(AES_256_GCM, "aes-256-gcm-aad12", 32, 16, 12) \
-  _(CHACHA20_POLY1305, "chacha20-poly1305-aad8", 32, 16, 8) \
-  _(CHACHA20_POLY1305, "chacha20-poly1305-aad12", 32, 16, 12)
+#define foreach_crypto_aead_async_alg                                         \
+  _ (AES_128_GCM, "aes-128-gcm-aad8", 16, 16, 8)                              \
+  _ (AES_128_GCM, "aes-128-gcm-aad12", 16, 16, 12)                            \
+  _ (AES_192_GCM, "aes-192-gcm-aad8", 24, 16, 8)                              \
+  _ (AES_192_GCM, "aes-192-gcm-aad12", 24, 16, 12)                            \
+  _ (AES_256_GCM, "aes-256-gcm-aad8", 32, 16, 8)                              \
+  _ (AES_256_GCM, "aes-256-gcm-aad12", 32, 16, 12)                            \
+  _ (CHACHA20_POLY1305, "chacha20-poly1305-aad8", 32, 16, 8)                  \
+  _ (CHACHA20_POLY1305, "chacha20-poly1305-aad12", 32, 16, 12)                \
+  _ (CHACHA20_POLY1305, "chacha20-poly1305", 32, 16, 0)
 
 /* CRYPTO_ID, INTEG_ID, PRETTY_NAME, KEY_LENGTH_IN_BYTES, DIGEST_LEN */
 #define foreach_crypto_link_async_alg                                         \