gso: do not gro on small packets

This has two benefits: fix an issue where small packets can be
padded to 64 bytes by the ethernet layer, and we included these bytes
in the tcp stream; and also if we receive a small packet, it signals
that this tcp stream is likely more latency-sensitive than throughput-
sensitive, and skipping gro buffering should be beneficial.

Type: fix
Change-Id: I165b97673d8cdce95ebf0a66c362ae9f9e6f3f34
Signed-off-by: Aloys Augustin <aloaugus@cisco.com>

2 files changed, 143 insertions, 3 deletions

diff --git a/src/vnet/gso/gro_func.h b/src/vnet/gso/gro_func.h
index b29d4a5f944..c9464bdc063 100644
--- a/src/vnet/gso/gro_func.h
+++ b/src/vnet/gso/gro_func.h
@@ -21,10 +21,14 @@
 #include <vnet/gso/hdr_offset_parser.h>
 #include <vnet/ip/ip4.h>
 #include <vnet/ip/ip6.h>
+#include <vnet/ip/ip6_inlines.h>
 #include <vnet/udp/udp_packet.h>
 #include <vnet/tcp/tcp_packet.h>
 #include <vnet/vnet.h>
 
+#define GRO_MIN_PACKET_SIZE    256
+#define GRO_PADDED_PACKET_SIZE 64
+
 static_always_inline u8
 gro_is_bad_packet (vlib_buffer_t * b, u8 flags, i16 l234_sz)
 {
@@ -160,6 +164,34 @@ gro_validate_checksum (vlib_main_t * vm, vlib_buffer_t * b0,
 }
 
 static_always_inline u32
+gro_fix_padded_packet_len (vlib_buffer_t *b0, generic_header_offset_t *gho0,
+			   ip4_header_t *ip4_0, ip6_header_t *ip6_0,
+			   u32 pkt_len0, u16 l234_sz0)
+{
+  u32 tcp_payload_len0 = 0;
+  if (gho0->gho_flags & GHO_F_IP4)
+    {
+      tcp_payload_len0 = clib_net_to_host_u16 (ip4_0->length) -
+			 ip4_header_bytes (ip4_0) - gho0->l4_hdr_sz;
+    }
+  else
+    {
+      tcp_payload_len0 =
+	clib_net_to_host_u16 (ip6_0->payload_length) - gho0->l4_hdr_sz;
+    }
+
+  ASSERT (l234_sz0 + tcp_payload_len0 <= pkt_len0);
+
+  if (PREDICT_FALSE (l234_sz0 + tcp_payload_len0 < pkt_len0))
+    {
+      /* small packet with padding at the end, remove padding */
+      b0->current_length = l234_sz0 + tcp_payload_len0;
+      pkt_len0 = b0->current_length;
+    }
+  return pkt_len0;
+}
+
+static_always_inline u32
 gro_get_packet_data (vlib_main_t *vm, vlib_buffer_t *b0,
 		     generic_header_offset_t *gho0, gro_flow_key_t *flow_key0,
 		     u8 is_l2)
@@ -222,6 +254,11 @@ gro_get_packet_data (vlib_main_t *vm, vlib_buffer_t *b0,
   if (PREDICT_FALSE (pkt_len0 >= TCP_MAX_GSO_SZ))
     return 0;
 
+  if (PREDICT_FALSE (pkt_len0 <= GRO_PADDED_PACKET_SIZE))
+    {
+      pkt_len0 =
+	gro_fix_padded_packet_len (b0, gho0, ip4_0, ip6_0, pkt_len0, l234_sz0);
+    }
   return pkt_len0;
 }
 
@@ -264,8 +301,8 @@ gro_coalesce_buffers (vlib_main_t *vm, vlib_buffer_t *b0, vlib_buffer_t *b1,
   pkt_len0 = vlib_buffer_length_in_chain (vm, b0);
   pkt_len1 = vlib_buffer_length_in_chain (vm, b1);
 
-  if (((gho0.gho_flags & GHO_F_TCP) == 0)
-      || ((gho1.gho_flags & GHO_F_TCP) == 0))
+  if (((gho0.gho_flags & GHO_F_TCP) == 0 || pkt_len0 <= GRO_MIN_PACKET_SIZE) ||
+      ((gho1.gho_flags & GHO_F_TCP) == 0 || pkt_len1 <= GRO_MIN_PACKET_SIZE))
     return 0;
 
   ip4_0 =
@@ -483,7 +520,8 @@ vnet_gro_flow_table_inline (vlib_main_t * vm, gro_flow_table_t * flow_table,
     }
 
   tcp0 = (tcp_header_t *) (vlib_buffer_get_current (b0) + gho0.l4_hdr_offset);
-  if (PREDICT_TRUE ((tcp0->flags & TCP_FLAG_PSH) == 0))
+  if (PREDICT_TRUE (((tcp0->flags & TCP_FLAG_PSH) == 0) &&
+		    (pkt_len0 > GRO_MIN_PACKET_SIZE)))
     gro_flow = gro_flow_table_find_or_add_flow (flow_table, &flow_key0);
   else
     {
diff --git a/test/test_gro.py b/test/test_gro.py
index 33215d65fa7..6c68299d7f4 100644
--- a/test/test_gro.py
+++ b/test/test_gro.py
@@ -138,5 +138,107 @@ class TestGRO(VppTestCase):
             self.assertEqual(rx[TCP].sport, 1234)
             self.assertEqual(rx[TCP].dport, 4321)
 
+        #
+        # Same test with IPv6
+        #
+        p = []
+        s = 0
+        for n in range(0, 88):
+            p.append((Ether(src=self.pg0.remote_mac, dst=self.pg0.local_mac) /
+                      IPv6(src=self.pg0.remote_ip6, dst=self.pg2.remote_ip6) /
+                      TCP(sport=1234, dport=4321, seq=s, ack=n, flags='A') /
+                      Raw(b'\xa5' * 1460)))
+            s += 1460
+        p[-1][TCP].flags = 'AP'  # push to flush second packet
+
+        rxs = self.send_and_expect(self.pg0, p, self.pg2, n_rx=2)
+
+        i = 0
+        for rx in rxs:
+            i += 1
+            self.assertEqual(rx[Ether].src, self.pg2.local_mac)
+            self.assertEqual(rx[Ether].dst, self.pg2.remote_mac)
+            self.assertEqual(rx[IPv6].src, self.pg0.remote_ip6)
+            self.assertEqual(rx[IPv6].dst, self.pg2.remote_ip6)
+            self.assertEqual(rx[IPv6].plen, 64260)  # 1460 * 44 + 20 < 65536
+            self.assertEqual(rx[TCP].sport, 1234)
+            self.assertEqual(rx[TCP].dport, 4321)
+            self.assertEqual(rx[TCP].ack, (44*i - 1))
+
+        #
+        # Send a series of 1500 bytes packets each followed by a packet with a
+        # PSH flag. Verify that GRO stops everytime a PSH flag is encountered
+        #
+        p = []
+        s = 0
+        for n in range(0, n_packets):
+            p.append((Ether(src=self.pg0.remote_mac, dst=self.pg0.local_mac) /
+                      IP(src=self.pg0.remote_ip4, dst=self.pg2.remote_ip4,
+                         flags='DF') /
+                      TCP(sport=1234, dport=4321, seq=s, ack=2*n, flags='A') /
+                      Raw(b'\xa5' * 1460)))
+            s += 1460
+            p.append((Ether(src=self.pg0.remote_mac, dst=self.pg0.local_mac) /
+                      IP(src=self.pg0.remote_ip4, dst=self.pg2.remote_ip4,
+                         flags='DF') /
+                      TCP(sport=1234, dport=4321, seq=s, ack=2*n+1,
+                          flags='AP') /
+                      Raw(b'\xa5' * 1340)))
+            s += 1340
+
+        rxs = self.send_and_expect(self.pg0, p, self.pg2, n_rx=n_packets)
+
+        i = 0
+        for rx in rxs:
+            self.assertEqual(rx[Ether].src, self.pg2.local_mac)
+            self.assertEqual(rx[Ether].dst, self.pg2.remote_mac)
+            self.assertEqual(rx[IP].src, self.pg0.remote_ip4)
+            self.assertEqual(rx[IP].dst, self.pg2.remote_ip4)
+            self.assertEqual(rx[IP].len, 40 + 1460 + 1340)
+            self.assertEqual(rx[TCP].sport, 1234)
+            self.assertEqual(rx[TCP].dport, 4321)
+            self.assertEqual(rx[TCP].ack, (2*i + 1))
+            i += 1
+
+        #
+        # Send a series of 1500 bytes packets each followed by a short packet
+        # with padding. Verify that GRO removes the padding and stops on short
+        # packets
+        #
+        p = []
+        s = 0
+        for n in range(0, n_packets):
+            i = self.pg0
+            p.append((Ether(src=i.remote_mac, dst=i.local_mac) /
+                      IP(src=i.remote_ip4, dst=self.pg2.remote_ip4,
+                         flags='DF') /
+                      TCP(sport=1234, dport=4321, seq=s, ack=2*n, flags='A') /
+                      Raw(b'\xa5' * 1459)))
+            s += 1459
+            p2 = (Ether(src=i.remote_mac, dst=i.local_mac) /
+                  IP(src=i.remote_ip4, dst=self.pg2.remote_ip4,
+                     flags='DF', len=41) /
+                  TCP(sport=1234, dport=4321, seq=s, ack=2*n+1, flags='A') /
+                  Raw(b'\xa5'))
+            # first compute csum of pkt w/o padding to work around scapy bug
+            p2 = Ether(bytes(p2))
+            p.append(p2 / Raw(b'\xa5' * 5))  # 1 byte data + 5 bytes padding
+            s += 1
+
+        rxs = self.send_and_expect(self.pg0, p, self.pg2, n_rx=n_packets)
+
+        i = 0
+        for rx in rxs:
+            self.assertEqual(rx[Ether].src, self.pg2.local_mac)
+            self.assertEqual(rx[Ether].dst, self.pg2.remote_mac)
+            self.assertEqual(rx[IP].src, self.pg0.remote_ip4)
+            self.assertEqual(rx[IP].dst, self.pg2.remote_ip4)
+            self.assertEqual(rx[IP].len, 40 + 1459 + 1)
+            self.assertEqual(rx[TCP].sport, 1234)
+            self.assertEqual(rx[TCP].dport, 4321)
+            self.assertEqual(rx[TCP].ack, (2*i + 1))
+            i += 1
+
+
 if __name__ == '__main__':
     unittest.main(testRunner=VppTestRunner)