ip: reassembly: trace ip headers over worker handoffs

This change adds tracing of IP headers when doing a handoff between
worker threads. This eases debugging.

Type: feature

Change-Id: I2195b070a364cba13a658ec1cee5154fc4c3a8b0
Signed-off-by: Klement Sekera <ksekera@cisco.com>

5 files changed, 87 insertions, 27 deletions

diff --git a/src/vnet/ip/format.h b/src/vnet/ip/format.h
index 7b1740dad04..952455c975a 100644
--- a/src/vnet/ip/format.h
+++ b/src/vnet/ip/format.h
@@ -95,6 +95,7 @@ format_function_t format_ip6_address_and_length;
 format_function_t format_ip6_address_and_mask;
 unformat_function_t unformat_ip6_header;
 format_function_t format_ip6_header;
+format_function_t format_ip6_frag_hdr;
 unformat_function_t unformat_pg_ip6_header;
 
 /* Format a TCP/UDP headers. */
diff --git a/src/vnet/ip/ip6_format.c b/src/vnet/ip/ip6_format.c
index 4d1793970df..5b8df239bf2 100644
--- a/src/vnet/ip/ip6_format.c
+++ b/src/vnet/ip/ip6_format.c
@@ -417,6 +417,25 @@ format_ip46_address (u8 * s, va_list * args)
     format (s, "%U", format_ip6_address, &ip46->ip6);
 }
 
+u8 *
+format_ip6_frag_hdr (u8 * s, va_list * args)
+{
+  ip6_frag_hdr_t *h = va_arg (*args, ip6_frag_hdr_t *);
+  u32 max_header_bytes = va_arg (*args, u32);
+  u32 header_bytes;
+
+  header_bytes = sizeof (h[0]);
+  if (max_header_bytes != 0 && header_bytes > max_header_bytes)
+    return format (s, "ipv6 frag header truncated");
+
+  s =
+    format (s,
+	    "IPV6_FRAG_HDR: next_hdr: %u, rsv: %u, frag_offset_and_more: %u, id: %u",
+	    h->next_hdr, h->rsv, h->fragment_offset_and_more,
+	    clib_net_to_host_u32 (h->identification));
+  return s;
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/ip/ip6_input.c b/src/vnet/ip/ip6_input.c
index 943f6597a90..ebcfd5a9f05 100644
--- a/src/vnet/ip/ip6_input.c
+++ b/src/vnet/ip/ip6_input.c
@@ -247,25 +247,6 @@ VLIB_REGISTER_NODE (ip6_input_node) = {
 };
 /* *INDENT-ON* */
 
-static u8 *
-format_ipv6_fragmentation (u8 * s, va_list * args)
-{
-  ip6_frag_hdr_t *h = va_arg (*args, ip6_frag_hdr_t *);
-  u32 max_header_bytes = va_arg (*args, u32);
-  u32 header_bytes;
-
-  header_bytes = sizeof (h[0]);
-  if (max_header_bytes != 0 && header_bytes > max_header_bytes)
-    return format (s, "ipv6 frag header truncated");
-
-  s =
-    format (s,
-	    "fragmentation header: next_hdr: %u, rsv: %u, frag_offset_and_more: %u, id: %u",
-	    h->next_hdr, h->rsv, h->fragment_offset_and_more,
-	    clib_net_to_host_u32 (h->identification));
-  return s;
-}
-
 static clib_error_t *
 ip6_init (vlib_main_t * vm)
 {
@@ -292,7 +273,7 @@ ip6_init (vlib_main_t * vm)
       ip_protocol_info_t *info =
 	vec_elt_at_index (ip_main.protocol_infos, *u);
       ASSERT (NULL == info->format_header);
-      info->format_header = format_ipv6_fragmentation;
+      info->format_header = format_ip6_frag_hdr;
     }
   return /* no error */ 0;
 }
diff --git a/src/vnet/ip/reass/ip4_full_reass.c b/src/vnet/ip/reass/ip4_full_reass.c
index 87a677208fb..303f23337ee 100644
--- a/src/vnet/ip/reass/ip4_full_reass.c
+++ b/src/vnet/ip/reass/ip4_full_reass.c
@@ -235,6 +235,8 @@ typedef struct
   u32 fragment_first;
   u32 fragment_last;
   u32 total_data_len;
+  bool is_after_handoff;
+  ip4_header_t ip4_header;
 } ip4_full_reass_trace_t;
 
 extern vlib_node_registration_t ip4_full_reass_node;
@@ -274,7 +276,16 @@ format_ip4_full_reass_trace (u8 * s, va_list * args)
   u32 indent = 0;
   if (~0 != t->reass_id)
     {
-      s = format (s, "reass id: %u, op id: %u, ", t->reass_id, t->op_id);
+      if (t->is_after_handoff)
+	{
+	  s =
+	    format (s, "%U\n", format_ip4_header, &t->ip4_header,
+		    sizeof (t->ip4_header));
+	  indent = 2;
+	}
+      s =
+	format (s, "%Ureass id: %u, op id: %u, ", format_white_space, indent,
+		t->reass_id, t->op_id);
       indent = format_get_indent (s);
       s =
 	format (s,
@@ -322,7 +333,18 @@ ip4_full_reass_add_trace (vlib_main_t * vm, vlib_node_runtime_t * node,
 {
   vlib_buffer_t *b = vlib_get_buffer (vm, bi);
   vnet_buffer_opaque_t *vnb = vnet_buffer (b);
+  bool is_after_handoff = false;
+  if (vlib_buffer_get_trace_thread (b) != vm->thread_index)
+    {
+      is_after_handoff = true;
+    }
   ip4_full_reass_trace_t *t = vlib_add_trace (vm, node, b, sizeof (t[0]));
+  t->is_after_handoff = is_after_handoff;
+  if (t->is_after_handoff)
+    {
+      clib_memcpy (&t->ip4_header, vlib_buffer_get_current (b),
+		   clib_min (sizeof (t->ip4_header), b->current_length));
+    }
   if (reass)
     {
       t->reass_id = reass->id;
diff --git a/src/vnet/ip/reass/ip6_full_reass.c b/src/vnet/ip/reass/ip6_full_reass.c
index bba11e52959..aaaf56a5d59 100644
--- a/src/vnet/ip/reass/ip6_full_reass.c
+++ b/src/vnet/ip/reass/ip6_full_reass.c
@@ -215,6 +215,9 @@ typedef struct
   u32 total_data_len;
   u32 thread_id;
   u32 thread_id_to;
+  bool is_after_handoff;
+  ip6_header_t ip6_header;
+  ip6_frag_hdr_t ip6_frag_header;
 } ip6_full_reass_trace_t;
 
 static void
@@ -251,7 +254,19 @@ format_ip6_full_reass_trace (u8 * s, va_list * args)
   u32 indent = 0;
   if (~0 != t->reass_id)
     {
-      s = format (s, "reass id: %u, op id: %u ", t->reass_id, t->op_id);
+      if (t->is_after_handoff)
+	{
+	  s =
+	    format (s, "%U\n", format_ip6_header, &t->ip6_header,
+		    sizeof (t->ip6_header));
+	  s =
+	    format (s, "  %U\n", format_ip6_frag_hdr, &t->ip6_frag_header,
+		    sizeof (t->ip6_frag_header));
+	  indent = 2;
+	}
+      s =
+	format (s, "%Ureass id: %u, op id: %u, ", format_white_space, indent,
+		t->reass_id, t->op_id);
       indent = format_get_indent (s);
       s = format (s, "first bi: %u, data len: %u, ip/fragment[%u, %u]",
 		  t->trace_range.first_bi, t->total_data_len,
@@ -297,12 +312,33 @@ static void
 ip6_full_reass_add_trace (vlib_main_t * vm, vlib_node_runtime_t * node,
 			  ip6_full_reass_main_t * rm,
 			  ip6_full_reass_t * reass, u32 bi,
+			  ip6_frag_hdr_t * ip6_frag_header,
 			  ip6_full_reass_trace_operation_e action,
 			  u32 thread_id_to)
 {
   vlib_buffer_t *b = vlib_get_buffer (vm, bi);
   vnet_buffer_opaque_t *vnb = vnet_buffer (b);
+  bool is_after_handoff = false;
+  if (vlib_buffer_get_trace_thread (b) != vm->thread_index)
+    {
+      is_after_handoff = true;
+    }
   ip6_full_reass_trace_t *t = vlib_add_trace (vm, node, b, sizeof (t[0]));
+  t->is_after_handoff = is_after_handoff;
+  if (t->is_after_handoff)
+    {
+      clib_memcpy (&t->ip6_header, vlib_buffer_get_current (b),
+		   clib_min (sizeof (t->ip6_header), b->current_length));
+      if (ip6_frag_header)
+	{
+	  clib_memcpy (&t->ip6_frag_header, ip6_frag_header,
+		       sizeof (t->ip6_frag_header));
+	}
+      else
+	{
+	  clib_memset (&t->ip6_frag_header, 0, sizeof (t->ip6_frag_header));
+	}
+    }
   if (reass)
     {
       t->reass_id = reass->id;
@@ -434,7 +470,7 @@ ip6_full_reass_on_timeout (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (PREDICT_FALSE (b->flags & VLIB_BUFFER_IS_TRACED))
 	    {
 	      ip6_full_reass_add_trace (vm, node, rm, reass, reass->first_bi,
-					ICMP_ERROR_RT_EXCEEDED, ~0);
+					NULL, ICMP_ERROR_RT_EXCEEDED, ~0);
 	    }
 	  // fragment with offset zero received - send icmp message back
 	  if (b->flags & VLIB_BUFFER_NEXT_PRESENT)
@@ -721,7 +757,7 @@ ip6_full_reass_finalize (vlib_main_t * vm, vlib_node_runtime_t * node,
   first_b->flags &= ~VLIB_BUFFER_EXT_HDR_VALID;
   if (PREDICT_FALSE (first_b->flags & VLIB_BUFFER_IS_TRACED))
     {
-      ip6_full_reass_add_trace (vm, node, rm, reass, reass->first_bi,
+      ip6_full_reass_add_trace (vm, node, rm, reass, reass->first_bi, NULL,
 				FINALIZE, ~0);
 #if 0
       // following code does a hexdump of packet fragments to stdout ...
@@ -889,7 +925,7 @@ ip6_full_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  // overlapping fragment - not allowed by RFC 8200
 	  if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
 	    {
-	      ip6_full_reass_add_trace (vm, node, rm, reass, *bi0,
+	      ip6_full_reass_add_trace (vm, node, rm, reass, *bi0, frag_hdr,
 					RANGE_OVERLAP, ~0);
 	    }
 	  ip6_full_reass_drop_all (vm, node, rm, reass);
@@ -906,7 +942,8 @@ check_if_done_maybe:
     {
       if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
 	{
-	  ip6_full_reass_add_trace (vm, node, rm, reass, *bi0, RANGE_NEW, ~0);
+	  ip6_full_reass_add_trace (vm, node, rm, reass, *bi0, frag_hdr,
+				    RANGE_NEW, ~0);
 	}
     }
   if (~0 != reass->last_packet_octet &&
@@ -1175,7 +1212,7 @@ ip6_full_reassembly_inline (vlib_main_t * vm,
 		  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
 		    {
 		      ip6_full_reass_add_trace (vm, node, rm, NULL, bi0,
-						HANDOFF,
+						frag_hdr, HANDOFF,
 						vnet_buffer (b0)->ip.
 						reass.owner_thread_index);
 		    }