packetimpact/

README

Packetimpact

What is packetimpact?

Packetimpact is a tool for platform-independent network testing. It is heavily inspired by packetdrill. It creates two network namespaces. One is for the test bench, which operates the test. The other is for the device-under-test (DUT), which is the software being tested. The test bench communicates over the network with the DUT to check correctness of the network.

Goals

Packetimpact aims to provide:

• A multi-platform solution that can test both Linux and gVisor.
• Conciseness on par with packetdrill scripts.
• Control-flow like for loops, conditionals, and variables.
• Flexibility to specify every byte in a packet or use multiple sockets.

How to run packetimpact tests?

Run a test, e.g. fin_wait2_timeout, against Linux:

$ bazel test //test/packetimpact/tests:fin_wait2_timeout_native_test

Run the same test, but against gVisor:

$ bazel test //test/packetimpact/tests:fin_wait2_timeout_netstack_test

When to use packetimpact?

There are a few ways to write networking tests for gVisor currently:

• Go unit tests
• syscall tests
• packetdrill tests
• packetimpact tests

The right choice depends on the needs of the test.

Feature	Go unit test	syscall test	packetdrill	packetimpact
Multi-platform	no	YES	YES	YES
Concise	no	somewhat	somewhat	VERY
Control-flow	YES	YES	no	YES
Flexible	VERY	no	somewhat	VERY

Go unit tests

If the test depends on the internals of gVisor and doesn't need to run on Linux or other platforms for comparison purposes, a Go unit test can be appropriate. They can observe internals of gVisor networking. The downside is that they are not concise and not multi-platform. If you require insight on gVisor internals, this is the right choice.

Syscall tests

Syscall tests are multi-platform but cannot examine the internals of gVisor networking. They are concise. They can use control-flow structures like conditionals, for loops, and variables. However, they are limited to only what the POSIX interface provides so they are not flexible. For example, you would have difficulty writing a syscall test that intentionally sends a bad IP checksum. Or if you did write that test with raw sockets, it would be very verbose to write a test that intentionally send wrong checksums, wrong protocols, wrong sequence numbers, etc.

Packetdrill tests

Packetdrill tests are multi-platform and can run against both Linux and gVisor. They are concise and use a special packetdrill scripting language. They are more flexible than a syscall test in that they can send packets that a syscall test would have difficulty sending, like a packet with a calcuated ACK number. But they are also somewhat limimted in flexibiilty in that they can't do tests with multiple sockets. They have no control-flow ability like variables or conditionals. For example, it isn't possible to send a packet that depends on the window size of a previous packet because the packetdrill language can't express that. Nor could you branch based on whether or not the other side supports window scaling, for example.

Packetimpact tests

Packetimpact tests are similar to Packetdrill tests except that they are written in Go instead of the packetdrill scripting language. That gives them all the control-flow abilities of Go (loops, functions, variables, etc). They are multi-platform in the same way as packetdrill tests but even more flexible because Go is more expressive than the scripting language of packetdrill. However, Go is not as concise as the packetdrill language. Many design decisions below are made to mitigate that.

How it works

     Testbench                           Device-Under-Test (DUT)
    +-------------------+               +------------------------+
    |                   |   TEST NET    |                        |
    | rawsockets.go <-->| <===========> | <---+                  |
    |           ^       |               |     |                  |
    |           |       |               |     |                  |
    |           v       |               |     |                  |
    |     unittest      |               |     |                  |
    |           ^       |               |     |                  |
    |           |       |               |     |                  |
    |           v       |               |     v                  |
    |         dut.go <========gRPC========> posix server         |
    |                   |  CONTROL NET  |                        |
    +-------------------+               +------------------------+

Two network namespaces are created by the test runner, one for the testbench and the other for the device under test (DUT). The runner connects the two namespaces with a control veth pair and test veth pair. It also does some other tasks like waiting until the DUT is ready before starting the test and installing iptables rules so that RST won't be generated for TCP segments from the DUT that the kernel has no knowledge about.

DUT

The DUT namespace runs a program called the "posix_server". The posix_server is written in c++ for maximum portability. Its job is to receive directions from the test bench on what actions to take. For this, the posix_server does three steps in a loop:

1. Listen for a request from the test bench.
2. Execute a command.
3. Send the response back to the test bench.

The requests and responses are protobufs and the communication is done with gRPC. The commands run are POSIX socket commands, with the inputs and outputs converted into protobuf requests and responses. All communication is on the control network, so that the test network is unaffected by extra packets.

For example, this is the request and response pair to call socket():

message SocketRequest {
  int32 domain = 1;
  int32 type = 2;
  int32 protocol = 3;
}

message SocketResponse {
  int32 fd = 1;
  int32 errno_ = 2;
}

Alternatives considered

• We could have use JSON for communication instead. It would have been a lighter-touch than protobuf but protobuf handles all the data type and has strict typing to prevent a class of errors. The test bench could be written in other languages, too.
• Instead of mimicking the POSIX interfaces, arguments could have had a more natural form, like the bind() getting a string IP address instead of bytes in a sockaddr_t. However, conforming to the existing structures keeps more of the complexity in Go and keeps the posix_server simpler and thus more likely to compile everywhere.

Test Bench

The test bench does most of the work in a test. It is a Go program that compiles on the host and is run inside the test bench's namespace. It is a regular go unit test that imports the test bench framework. The test bench framework is based on three basic utilities:

• Commanding the DUT to run POSIX commands and return responses.
• Sending raw packets to the DUT on the test network.
• Listening for raw packets from the DUT on the test network.

DUT commands

To keep the interface to the DUT consistent and easy-to-use, each POSIX command supported by the posix_server is wrapped in functions with signatures similar to the ones in the Go unix package. This way all the details of endianess and (un)marshalling of go structs such as unix.Timeval is handled in one place. This also makes it straight-forward to convert tests that use unix. or syscall. calls to dut. calls.

For example, creating a connection to the DUT and commanding it to make a socket looks like this:

dut := testbench.NewDut(t)
fd, err := dut.SocketWithErrno(unix.AF_INET, unix.SOCK_STREAM, unix.IPPROTO_IP)
if fd < 0 {
  t.Fatalf(...)
}

Because the usual case is to fail the test when the DUT fails to create a socket, there is a concise version of each of the ...WithErrno functions that does that:

dut := testbench.NewDut(t)
fd := dut.Socket(unix.AF_INET, unix.SOCK_STREAM, unix.IPPROTO_IP)

The DUT and other structs in the code store a *testing.T so that they can provide versions of functions that call t.Fatalf(...). This helps keep tests concise.

Alternatives considered

• Instead of mimicking the unix. go interface, we could have invented a more natural one, like using float64 instead of Timeval. However, using the same function signatures that unix. has makes it easier to convert code to dut.. Also, using an existing interface ensures that we don't invent an interface that isn't extensible. For example, if we invented a function for bind() that didn't support IPv6 and later we had to add a second bind6() function.

Sending/Receiving Raw Packets

The framework wraps POSIX sockets for sending and receiving raw frames. Both send and receive are synchronous commands. SO_RCVTIMEO is used to set a timeout on the receive commands. For ease of use, these are wrapped in an Injector and a Sniffer. They have functions:

func (s *Sniffer) Recv(timeout time.Duration) []byte {...}
func (i *Injector) Send(b []byte) {...}

Alternatives considered

• gopacket pcap has raw socket support but requires cgo. cgo is not guaranteed to be portable from the host to the container and in practice, the container doesn't recognize binaries built on the host if they use cgo. Packetimpact used to be based on docker, so the library was not adopted, now we can start to consider using the library.
• Both gVisor and gopacket have the ability to read and write pcap files without cgo but that is insufficient here because we can't just replay pcap files, we need a more dynamic solution.
• The sniffer and injector can't share a socket because they need to be bound differently.
• Sniffing could have been done asynchronously with channels, obviating the need for SO_RCVTIMEO. But that would introduce asynchronous complication. SO_RCVTIMEO is well supported on the test bench.

Layer struct

A large part of packetimpact tests is creating packets to send and comparing received packets against expectations. To keep tests concise, it is useful to be able to specify just the important parts of packets that need to be set. For example, sending a packet with default values except for TCP Flags. And for packets received, it's useful to be able to compare just the necessary parts of received packets and ignore the rest.

To aid in both of those, Go structs with optional fields are created for each encapsulation type, such as IPv4, TCP, and Ethernet. This is inspired by scapy. For example, here is the struct for Ethernet:

type Ether struct {
  LayerBase
  SrcAddr *tcpip.LinkAddress
  DstAddr *tcpip.LinkAddress
  Type    *tcpip.NetworkProtocolNumber
}

Each struct has the same fields as those in the gVisor headers but with a pointer for each field that may be nil.

Alternatives considered

• Just use []byte like gVisor headers do. The drawback is that it makes the tests more verbose.
  • For example, there would be no way to call Send(myBytes) concisely and indicate if the checksum should be calculated automatically versus overridden. The only way would be to add lines to the test to calculate it before each Send, which is wordy. Or make multiple versions of Send: one that checksums IP, one that doesn't, one that checksums TCP, one that does both, etc. That would be many combinations.
  • Filtering inputs would become verbose. Either:
  • large conditionals that need to be repeated many places: h[FlagOffset] == SYN && h[LengthOffset:LengthOffset+2] == ... or
  • Many functions, one per field, like: filterByFlag(myBytes, SYN), filterByLength(myBytes, 20), filterByNextProto(myBytes, 0x8000), etc.
  • Using pointers allows us to combine Layers with reflection. So the default Layers can be overridden by a Layers with just the TCP conection's src/dst which can be overridden by one with just a test specific TCP window size.
  • It's a proven way to separate the details of a packet from the byte format as shown by scapy's success.
• Use packetgo. It's more general than parsing packets with gVisor. However:
  • packetgo doesn't have optional fields so many of the above problems still apply.
  • It would be yet another dependency.
  • It's not as well known to engineers that are already writing gVisor code.
  • It might be a good candidate for replacing the parsing of packets into Layers if all that parsing turns out to be more work than parsing by packetgo and converting that to Layer. packetgo has easier to use getters for the layers. This could be done later in a way that doesn't break tests.

Layer methods

The Layer structs provide a way to partially specify an encapsulation. They also need methods for using those partially specified encapsulation, for example to marshal them to bytes or compare them. For those, each encapsulation implements the Layer interface:

// Layer is the interface that all encapsulations must implement.
//
// A Layer is an encapsulation in a packet, such as TCP, IPv4, IPv6, etc. A
// Layer contains all the fields of the encapsulation. Each field is a pointer
// and may be nil.
type Layer interface {
    // toBytes converts the Layer into bytes. In places where the Layer's field
    // isn't nil, the value that is pointed to is used. When the field is nil, a
    // reasonable default for the Layer is used. For example, "64" for IPv4 TTL
    // and a calculated checksum for TCP or IP. Some layers require information
    // from the previous or next layers in order to compute a default, such as
    // TCP's checksum or Ethernet's type, so each Layer has a doubly-linked list
    // to the layer's neighbors.
    toBytes() ([]byte, error)

    // match checks if the current Layer matches the provided Layer. If either
    // Layer has a nil in a given field, that field is considered matching.
    // Otherwise, the values pointed to by the fields must match.
    match(Layer) bool

    // length in bytes of the current encapsulation
    length() int

    // next gets a pointer to the encapsulated Layer.
    next() Layer

    // prev gets a pointer to the Layer encapsulating this one.
    prev() Layer

    // setNext sets the pointer to the encapsulated Layer.
    setNext(Layer)

    // setPrev sets the pointer to the Layer encapsulating this one.
    setPrev(Layer)
}

The next and prev make up a link listed so that each layer can get at the information in the layer around it. This is necessary for some protocols, like TCP that needs the layer before and payload after to compute the checksum. Any sequence of Layer structs is valid so long as the parser and toBytes functions can map from type to protool number and vice-versa. When the mapping fails, an error is emitted explaining what functionality is missing. The solution is either to fix the ordering or implement the missing protocol.

For each Layer there is also a parsing function. For example, this one is for Ethernet:

func ParseEther(b []byte) (Layers, error)

The parsing function converts bytes received on the wire into a Layer (actually Layers, see below) which has no nils in it. By using match(Layer) to compare against another Layer that does have nils in it, the received bytes can be partially compared. The nils behave as "don't-cares".

Alternatives considered

• Matching against []byte instead of converting to Layer first.
  • The downside is that it precludes the use of a cmp.Equal one-liner to do comparisons.
  • It creates confusion in the code to deal with both representations at different times. For example, is the checksum calculated on []byte or Layer when sending? What about when checking received packets?

Layers

type Layers []Layer

func (ls *Layers) match(other Layers) bool {...}
func (ls *Layers) toBytes() ([]byte, error) {...}

Layers is an array of Layer. It represents a stack of encapsulations, such as Layers{Ether{},IPv4{},TCP{},Payload{}}. It also has toBytes() and match(Layers), like Layer. The parse functions above actually return Layers and not Layer because they know about the headers below and sequentially call each parser on the remaining, encapsulated bytes.

All this leads to the ability to write concise packet processing. For example:

etherType := 0x8000
flags = uint8(header.TCPFlagSyn|header.TCPFlagAck)
toMatch := Layers{Ether{Type: &etherType}, IPv4{}, TCP{Flags: &flags}}
for {
  recvBytes := sniffer.Recv(time.Second)
  if recvBytes == nil {
    println("Got no packet for 1 second")
  }
  gotPacket, err := ParseEther(recvBytes)
  if err == nil && toMatch.match(gotPacket) {
    println("Got a TCP/IPv4/Eth packet with SYNACK")
  }
}

Alternatives considered

• Don't use previous and next pointers.
  • Each layer may need to be able to interrogate the layers around it, like for computing the next protocol number or total length. So some mechanism is needed for a Layer to see neighboring layers.
  • We could pass the entire array Layers to the toBytes() function. Passing an array to a method that includes in the array the function receiver itself seems wrong.

layerState

Layers represents the different headers of a packet but a connection includes more state. For example, a TCP connection needs to keep track of the next expected sequence number and also the next sequence number to send. This is stored in a layerState struct. This is the layerState for TCP:

// tcpState maintains state about a TCP connection.
type tcpState struct {
    out, in                   TCP
    localSeqNum, remoteSeqNum *seqnum.Value
    synAck                    *TCP
    portPickerFD              int
    finSent                   bool
}

The next sequence numbers for each side of the connection are stored. out and in have defaults for the TCP header, such as the expected source and destination ports for outgoing packets and incoming packets.

layerState interface

// layerState stores the state of a layer of a connection.
type layerState interface {
    // outgoing returns an outgoing layer to be sent in a frame.
    outgoing() Layer

    // incoming creates an expected Layer for comparing against a received Layer.
    // Because the expectation can depend on values in the received Layer, it is
    // an input to incoming. For example, the ACK number needs to be checked in a
    // TCP packet but only if the ACK flag is set in the received packet.
    incoming(received Layer) Layer

    // sent updates the layerState based on the Layer that was sent. The input is
    // a Layer with all prev and next pointers populated so that the entire frame
    // as it was sent is available.
    sent(sent Layer) error

    // received updates the layerState based on a Layer that is received. The
    // input is a Layer with all prev and next pointers populated so that the
    // entire frame as it was received is available.
    received(received Layer) error

    // close frees associated resources held by the LayerState.
    close() error
}

outgoing generates the default Layer for an outgoing packet. For TCP, this would be a TCP with the source and destination ports populated. Because they are static, they are stored inside the out member of tcpState. However, the sequence numbers change frequently so the outgoing sequence number is stored in the localSeqNum and put into the output of outgoing for each call.

incoming does the same functions for packets that arrive but instead of generating a packet to send, it generates an expect packet for filtering packets that arrive. For example, if a TCP header arrives with the wrong ports, it can be ignored as belonging to a different connection. incoming needs the received header itself as an input because the filter may depend on the input. For example, the expected sequence number depends on the flags in the TCP header.

sent and received are run for each header that is actually sent or received and used to update the internal state. incoming and outgoing should not be used for these purpose. For example, incoming is called on every packet that arrives but only packets that match ought to actually update the state. outgoing is called to created outgoing packets and those packets are always sent, so unlike incoming/received, there is one outgoing call for each sent call.

close cleans up after the layerState. For example, TCP and UDP need to keep a port reserved and then release it.

Connections

Using layerState above, we can create connections.

// Connection holds a collection of layer states for maintaining a connection
// along with sockets for sniffer and injecting packets.
type Connection struct {
    layerStates []layerState
    injector    Injector
    sniffer     Sniffer
    t           *testing.T
}

The connection stores an array of layerState in the order that the headers should be present in the frame to send. For example, Ether then IPv4 then TCP. The injector and sniffer are for writing and reading frames. A *testing.T is stored so that internal errors can be reported directly without code in the unit test.

The Connection has some useful functions:

// Close frees associated resources held by the Connection.
func (conn *Connection) Close() {...}
// CreateFrame builds a frame for the connection with layer overriding defaults
// of the innermost layer and additionalLayers added after it.
func (conn *Connection) CreateFrame(layer Layer, additionalLayers ...Layer) Layers {...}
// SendFrame sends a frame on the wire and updates the state of all layers.
func (conn *Connection) SendFrame(frame Layers) {...}
// Send a packet with reasonable defaults. Potentially override the final layer
// in the connection with the provided layer and add additionLayers.
func (conn *Connection) Send(layer Layer, additionalLayers ...Layer) {...}
// Expect a frame with the final layerStates layer matching the provided Layer
// within the timeout specified. If it doesn't arrive in time, it returns nil.
func (conn *Connection) Expect(layer Layer, timeout time.Duration) (Layer, error) {...}
// ExpectFrame expects a frame that matches the provided Layers within the
// timeout specified. If it doesn't arrive in time, it returns nil.
func (conn *Connection) ExpectFrame(layers Layers, timeout time.Duration) (Layers, error) {...}
// Drain drains the sniffer's receive buffer by receiving packets until there's
// nothing else to receive.
func (conn *Connection) Drain() {...}

CreateFrame uses the []layerState to create a frame to send. The first argument is for overriding defaults in the last header of the frame, because this is the most common need. For a TCPIPv4 connection, this would be the TCP header. Optional additionalLayers can be specified to add to the frame being created, such as a Payload for TCP.

SendFrame sends the frame to the DUT. It is combined with CreateFrame to make Send. For unittests with basic sending needs, Send can be used. If more control is needed over the frame, it can be made with CreateFrame, modified in the unit test, and then sent with SendFrame.

On the receiving side, there is Expect and ExpectFrame. Like with the sending side, there are two forms of each function, one for just the last header and one for the whole frame. The expect functions use the []layerState to create a template for the expected incoming frame. That frame is then overridden by the values in the first argument. Finally, a loop starts sniffing packets on the wire for frames. If a matching frame is found before the timeout, it is returned without error. If not, nil is returned and the error contains text of all the received frames that didn't match. Exactly one of the outputs will be non-nil, even if no frames are received at all.

Drain sniffs and discards all the frames that have yet to be received. A common way to write a test is:

conn.Drain() // Discard all outstanding frames.
conn.Send(...) // Send a frame with overrides.
// Now expect a frame with a certain header and fail if it doesn't arrive.
if _, err := conn.Expect(...); err != nil { t.Fatal(...) }

Or for a test where we want to check that no frame arrives:

if gotOne, _ := conn.Expect(...); gotOne != nil { t.Fatal(...) }

Specializing Connection

Because there are some common combinations of layerState into Connection, they are defined:

// TCPIPv4 maintains the state for all the layers in a TCP/IPv4 connection.
type TCPIPv4 Connection
// UDPIPv4 maintains the state for all the layers in a UDP/IPv4 connection.
type UDPIPv4 Connection

Each has a NewXxx function to create a new connection with reasonable defaults. They also have functions that call the underlying Connection functions but with specialization and tighter type-checking. For example:

func (conn *TCPIPv4) Send(tcp TCP, additionalLayers ...Layer) {
    (*Connection)(conn).Send(&tcp, additionalLayers...)
}
func (conn *TCPIPv4) Drain() {
    conn.sniffer.Drain()
}

They may also have some accessors to get or set the internal state of the connection:

func (conn *TCPIPv4) state() *tcpState {
    state, ok := conn.layerStates[len(conn.layerStates)-1].(*tcpState)
    if !ok {
        conn.t.Fatalf("expected final state of %v to be tcpState", conn.layerStates)
    }
    return state
}
func (conn *TCPIPv4) RemoteSeqNum() *seqnum.Value {
    return conn.state().remoteSeqNum
}
func (conn *TCPIPv4) LocalSeqNum() *seqnum.Value {
    return conn.state().localSeqNum
}

Unittests will in practice use these functions and not the functions on Connection. For example, NewTCPIPv4() and then call Send on that rather than cast is to a Connection and call Send on that cast result.

Alternatives considered

• Instead of storing outgoing and incoming, store values.
  • There would be many more things to store instead, like localMac, remoteMac, localIP, remoteIP, localPort, and remotePort.
  • Construction of a packet would be many lines to copy each of these values into a []byte. And there would be slight variations needed for each encapsulation stack, like TCPIPv6 and ARP.
  • Filtering incoming packets would be a long sequence:
  • Compare the MACs, then
  • Parse the next header, then
  • Compare the IPs, then
  • Parse the next header, then
  • Compare the TCP ports. Instead it's all just one call to cmp.Equal(...), for all sequences.
  • A TCPIPv6 connection could share most of the code. Only the type of the IP addresses are different. The types of outgoing and incoming would be remain Layers.
  • An ARP connection could share all the Ethernet parts. The IP Layer could be factored out of outgoing. After that, the IPv4 and IPv6 connections could implement one interface and a single TCP struct could have either network protocol through composition.

Putting it all together

Here's what te start of a packetimpact unit test looks like. This test creates a TCP connection with the DUT. There are added comments for explanation in this document but a real test might not include them in order to stay even more concise.

func TestMyTcpTest(t *testing.T) {
  // Prepare a DUT for communication.
  dut := testbench.NewDUT(t)

  // This does:
  //   dut.Socket()
  //   dut.Bind()
  //   dut.Getsockname() to learn the new port number
  //   dut.Listen()
  listenFD, remotePort := dut.CreateListener(unix.SOCK_STREAM, unix.IPPROTO_TCP, 1)
  defer dut.Close(listenFD) // Tell the DUT to close the socket at the end of the test.

  // Monitor a new TCP connection with sniffer, injector, sequence number tracking,
  // and reasonable outgoing and incoming packet field default IPs, MACs, and port numbers.
  conn := testbench.NewTCPIPv4(t, dut, remotePort)

  // Perform a 3-way handshake: send SYN, expect SYNACK, send ACK.
  conn.Handshake()

  // Tell the DUT to accept the new connection.
  acceptFD := dut.Accept(acceptFd)
}

Adding a new packetimpact test

• Create a go test in the tests directory
• Add a packetimpact_testbench rule in BUILD
• Add the test into the ALL_TESTS list in defs.bzl, otherwise you will see an error message complaining about a missing test.

Other notes

• The time between receiving a SYN-ACK and replying with an ACK in Handshake is about 3ms. This is much slower than the native unix response, which is about 0.3ms. Packetdrill gets closer to 0.3ms. For tests where timing is crucial, packetdrill is faster and more precise.