Merge branch 'master' into ip-forwarding

- Merges aleksej-paschenko's with HEAD
- Adds vfs2 support for ip_forward

1 files changed, 65 insertions, 36 deletions

diff --git a/test/syscalls/linux/raw_socket_hdrincl.cc b/test/syscalls/linux/raw_socket_hdrincl.cc
index 0a27506aa..2f25aceb2 100644
--- a/test/syscalls/linux/raw_socket_hdrincl.cc
+++ b/test/syscalls/linux/raw_socket_hdrincl.cc
@@ -167,7 +167,7 @@ TEST_F(RawHDRINCL, NotReadable) {
   // nothing to be read.
   char buf[117];
   ASSERT_THAT(RetryEINTR(recv)(socket_, buf, sizeof(buf), MSG_DONTWAIT),
-              SyscallFailsWithErrno(EINVAL));
+              SyscallFailsWithErrno(EAGAIN));
 }
 
 // Test that we can connect() to a valid IP (loopback).
@@ -178,6 +178,9 @@ TEST_F(RawHDRINCL, ConnectToLoopback) {
 }
 
 TEST_F(RawHDRINCL, SendWithoutConnectSucceeds) {
+  // FIXME(gvisor.dev/issue/3159): Test currently flaky.
+  SKIP_IF(true);
+
   struct iphdr hdr = LoopbackHeader();
   ASSERT_THAT(send(socket_, &hdr, sizeof(hdr), 0),
               SyscallSucceedsWithValue(sizeof(hdr)));
@@ -273,14 +276,17 @@ TEST_F(RawHDRINCL, SendAndReceive) {
   // The network stack should have set the source address.
   EXPECT_EQ(src.sin_family, AF_INET);
   EXPECT_EQ(absl::gbswap_32(src.sin_addr.s_addr), INADDR_LOOPBACK);
-  // The packet ID should be 0, as the packet is less than 68 bytes.
-  struct iphdr iphdr = {};
-  memcpy(&iphdr, recv_buf, sizeof(iphdr));
-  EXPECT_EQ(iphdr.id, 0);
+  // The packet ID should not be 0, as the packet has DF=0.
+  struct iphdr* iphdr = reinterpret_cast<struct iphdr*>(recv_buf);
+  EXPECT_NE(iphdr->id, 0);
 }
 
-// Send and receive a packet with nonzero IP ID.
-TEST_F(RawHDRINCL, SendAndReceiveNonzeroID) {
+// Send and receive a packet where the sendto address is not the same as the
+// provided destination.
+TEST_F(RawHDRINCL, SendAndReceiveDifferentAddress) {
+  // FIXME(gvisor.dev/issue/3160): Test currently flaky.
+  SKIP_IF(true);
+
   int port = 40000;
   if (!IsRunningOnGvisor()) {
     port = static_cast<short>(ASSERT_NO_ERRNO_AND_VALUE(
@@ -292,19 +298,24 @@ TEST_F(RawHDRINCL, SendAndReceiveNonzeroID) {
   FileDescriptor udp_sock =
       ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_RAW, IPPROTO_UDP));
 
-  // Construct a packet with an IP header, UDP header, and payload. Make the
-  // payload large enough to force an IP ID to be assigned.
-  constexpr char kPayload[128] = {};
+  // Construct a packet with an IP header, UDP header, and payload.
+  constexpr char kPayload[] = "toto";
   char packet[sizeof(struct iphdr) + sizeof(struct udphdr) + sizeof(kPayload)];
   ASSERT_TRUE(
       FillPacket(packet, sizeof(packet), port, kPayload, sizeof(kPayload)));
+  // Overwrite the IP destination address with an IP we can't get to.
+  struct iphdr iphdr = {};
+  memcpy(&iphdr, packet, sizeof(iphdr));
+  iphdr.daddr = 42;
+  memcpy(packet, &iphdr, sizeof(iphdr));
 
   socklen_t addrlen = sizeof(addr_);
   ASSERT_NO_FATAL_FAILURE(sendto(socket_, &packet, sizeof(packet), 0,
                                  reinterpret_cast<struct sockaddr*>(&addr_),
                                  addrlen));
 
-  // Receive the payload.
+  // Receive the payload, since sendto should replace the bad destination with
+  // localhost.
   char recv_buf[sizeof(packet)];
   struct sockaddr_in src;
   socklen_t src_size = sizeof(src);
@@ -318,47 +329,58 @@ TEST_F(RawHDRINCL, SendAndReceiveNonzeroID) {
   // The network stack should have set the source address.
   EXPECT_EQ(src.sin_family, AF_INET);
   EXPECT_EQ(absl::gbswap_32(src.sin_addr.s_addr), INADDR_LOOPBACK);
-  // The packet ID should not be 0, as the packet was more than 68 bytes.
-  struct iphdr* iphdr = reinterpret_cast<struct iphdr*>(recv_buf);
-  EXPECT_NE(iphdr->id, 0);
+  // The packet ID should not be 0, as the packet has DF=0.
+  struct iphdr recv_iphdr = {};
+  memcpy(&recv_iphdr, recv_buf, sizeof(recv_iphdr));
+  EXPECT_NE(recv_iphdr.id, 0);
+  // The destination address should be localhost, not the bad IP we set
+  // initially.
+  EXPECT_EQ(absl::gbswap_32(recv_iphdr.daddr), INADDR_LOOPBACK);
 }
 
-// Send and receive a packet where the sendto address is not the same as the
-// provided destination.
-TEST_F(RawHDRINCL, SendAndReceiveDifferentAddress) {
+// Send and receive a packet w/ the IP_HDRINCL option set.
+TEST_F(RawHDRINCL, SendAndReceiveIPHdrIncl) {
   int port = 40000;
   if (!IsRunningOnGvisor()) {
     port = static_cast<short>(ASSERT_NO_ERRNO_AND_VALUE(
         PortAvailable(0, AddressFamily::kIpv4, SocketType::kUdp, false)));
   }
 
-  // IPPROTO_RAW sockets are write-only. We'll have to open another socket to
-  // read what we write.
-  FileDescriptor udp_sock =
+  FileDescriptor recv_sock =
+      ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_RAW, IPPROTO_UDP));
+
+  FileDescriptor send_sock =
       ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_RAW, IPPROTO_UDP));
 
+  // Enable IP_HDRINCL option so that we can build and send w/ an IP
+  // header.
+  constexpr int kSockOptOn = 1;
+  ASSERT_THAT(setsockopt(send_sock.get(), SOL_IP, IP_HDRINCL, &kSockOptOn,
+                         sizeof(kSockOptOn)),
+              SyscallSucceeds());
+  // This is not strictly required but we do it to make sure that setting
+  // IP_HDRINCL on a non IPPROTO_RAW socket does not prevent it from receiving
+  // packets.
+  ASSERT_THAT(setsockopt(recv_sock.get(), SOL_IP, IP_HDRINCL, &kSockOptOn,
+                         sizeof(kSockOptOn)),
+              SyscallSucceeds());
+
   // Construct a packet with an IP header, UDP header, and payload.
   constexpr char kPayload[] = "toto";
   char packet[sizeof(struct iphdr) + sizeof(struct udphdr) + sizeof(kPayload)];
   ASSERT_TRUE(
       FillPacket(packet, sizeof(packet), port, kPayload, sizeof(kPayload)));
-  // Overwrite the IP destination address with an IP we can't get to.
-  struct iphdr iphdr = {};
-  memcpy(&iphdr, packet, sizeof(iphdr));
-  iphdr.daddr = 42;
-  memcpy(packet, &iphdr, sizeof(iphdr));
 
   socklen_t addrlen = sizeof(addr_);
-  ASSERT_NO_FATAL_FAILURE(sendto(socket_, &packet, sizeof(packet), 0,
+  ASSERT_NO_FATAL_FAILURE(sendto(send_sock.get(), &packet, sizeof(packet), 0,
                                  reinterpret_cast<struct sockaddr*>(&addr_),
                                  addrlen));
 
-  // Receive the payload, since sendto should replace the bad destination with
-  // localhost.
+  // Receive the payload.
   char recv_buf[sizeof(packet)];
   struct sockaddr_in src;
   socklen_t src_size = sizeof(src);
-  ASSERT_THAT(recvfrom(udp_sock.get(), recv_buf, sizeof(recv_buf), 0,
+  ASSERT_THAT(recvfrom(recv_sock.get(), recv_buf, sizeof(recv_buf), 0,
                        reinterpret_cast<struct sockaddr*>(&src), &src_size),
               SyscallSucceedsWithValue(sizeof(packet)));
   EXPECT_EQ(
@@ -368,13 +390,20 @@ TEST_F(RawHDRINCL, SendAndReceiveDifferentAddress) {
   // The network stack should have set the source address.
   EXPECT_EQ(src.sin_family, AF_INET);
   EXPECT_EQ(absl::gbswap_32(src.sin_addr.s_addr), INADDR_LOOPBACK);
-  // The packet ID should be 0, as the packet is less than 68 bytes.
-  struct iphdr recv_iphdr = {};
-  memcpy(&recv_iphdr, recv_buf, sizeof(recv_iphdr));
-  EXPECT_EQ(recv_iphdr.id, 0);
-  // The destination address should be localhost, not the bad IP we set
-  // initially.
-  EXPECT_EQ(absl::gbswap_32(recv_iphdr.daddr), INADDR_LOOPBACK);
+  struct iphdr iphdr = {};
+  memcpy(&iphdr, recv_buf, sizeof(iphdr));
+  EXPECT_NE(iphdr.id, 0);
+
+  // Also verify that the packet we just sent was not delivered to the
+  // IPPROTO_RAW socket.
+  {
+    char recv_buf[sizeof(packet)];
+    struct sockaddr_in src;
+    socklen_t src_size = sizeof(src);
+    ASSERT_THAT(recvfrom(socket_, recv_buf, sizeof(recv_buf), MSG_DONTWAIT,
+                         reinterpret_cast<struct sockaddr*>(&src), &src_size),
+                SyscallFailsWithErrno(EAGAIN));
+  }
 }
 
 }  // namespace