Tags: netlink nftables netfilter
Rating:
The challenge gives us a domain easynft.hackable.software and a Tarball (files.tar.gz), which contains an easynft.pcap and a README.md
The README.md state:
One of our servers is acting strangely lately.
There are even rumours that it gives out flags if [asked in a right way](https://xkcd.com/424/).
Our forensic team didn't find any backdoors, but when we try to list firewall rules, `nft` just hangs.
The best we could get is this the netlink dump attached in easynft.pcap.
Could you help us figure out what's going on?
P.S. Obviously, the flag was redacted from the dump.
So apparently, the pcap contains traces of interactions between the nft binary and the netfilters backend.
Quickly scanning the capture with wireshark, we identify the typical keywords from a routing table (filter, input, output, forward...), several interesting keywords (flag, hack) and a fake flag dnrgs{REDACTEDREDACTEDREDACTEDREDACTED}.
The rest of the data seems binary and we cannot get much more from the capture.
We also notice that all packets start with a constant 16 bytes Linux netlink (cooked header), followed by a variable length and content Netlink message.
My tshark jutsu isn't that great, so we start by dumping the whole capture with an hexdump of each packets:
tshark -r easynft.pcap -x
And we clean it up to keep only the hex bytes on a single line
tshark -r easynft.pcap -x | grep -v "0000" | awk -F " " '{print $2}' | tr -d ' ' | perl -00 -lpe 'tr/\n//d' | grep -Ev '^\s* > easynft_netlink.dump
We now have a 23 lines file with the Netlink messages ready to parse. (see easynft_netlink.dump)
As I usually pick go as my first language choice, I've found the github.com/mdlayher/netlink library exposing a promising Message.UnmarshalBinary([]byte) error method we could feed with our packet bytes and see what happen.
We also notice when checking the source code of this method, that the first 4 bytes of the message are its length. And some of the packet hex strings contains more bytes than these first bytes indicate. This means we can have multiple netlink.Message per packet.
From here, the parsing code is quite straigtforward:
package main
import (
"bufio"
"encoding/binary"
"encoding/hex"
"fmt"
"os"
"github.com/mdlayher/netlink"
)
func main() {
f, err := os.Open("hex_netlink.dump")
if err != nil {
panic(err)
}
defer f.Close()
scanner := bufio.NewScanner(f)
var messages []netlink.Message
for scanner.Scan() {
packetHex := scanner.Text()
d, err := hex.DecodeString(packetHex)
if err != nil {
panic(err)
}
// packet may contains multiple messages, so split on size
// https://github.com/mdlayher/netlink/blob/v1.4.1/message.go#L234
for {
size := binary.LittleEndian.Uint32(d[:4])
packet := d
if len(d) > int(size) {
packet = d[:size]
}
d = d[size:]
msg := netlink.Message{}
if err := msg.UnmarshalBinary(packet); err != nil {
panic(fmt.Errorf("failed to unmarshal: %v - packet: %x", err, d))
}
messages = append(messages, msg)
if len(d) == 0 {
break
}
}
}
fmt.Printf("Parsed %d messages\n", len(messages))
for _, m := range messages {
fmt.Printf("%#v\n", m)
}
}
As an output, we get 28 decoded netlink.Message
Parsed 28 messages
netlink.Message{Header:netlink.Header{Length:0x14, Type:0xa10, Flags:0x1, Sequence:0x0, PID:0x0}, Data:[]uint8{0x0, 0x0, 0x0, 0x0}}
netlink.Message{Header:netlink.Header{Length:0x2c, Type:0xa0f, Flags:0x0, Sequence:0x0, PID:0x200a}, Data:[]uint8{0x0, 0x0, 0xbb, 0x75, 0x8, 0x0, 0x1, 0x0, 0x0, 0x0, 0xbb, 0x75, 0x8, 0x0, 0x2, 0x0, 0x0, 0x0, 0x20, 0xa, 0x8, 0x0, 0x3, 0x0, 0x6e, 0x66, 0x74, 0x0}}
netlink.Message{Header:netlink.Header{Length:0x14, Type:0xa01, Flags:0x301, Sequence:0x0, PID:0x0}, Data:[]uint8{0x0, 0x0, 0x0, 0x0}}
...truncated...
We now have more raw bytes in the Data field, which the Type field could help us identify and understand further.
From the pcap, we saw wireshark identified the messages are using the netlink netfilter protocol. Turns out netfilter is a sub component of the whole netfilter framework, responsible of the packet routing. Luckily, another library exists extending the netlink's one, to provide us the netfilter decoding features, such as unmarshalling netlink messages into netfilter header and attributes
We can update our previous loop over the messages to add the netfilter decoding:
// add import "github.com/ti-mo/netfilter"
fmt.Printf("Parsed %d messages\n", len(messages))
for _, m := range messages {
header, attrs, err := netfilter.UnmarshalNetlink(m)
if err != nil {
panic(err)
}
fmt.Printf("%s\n", header)
for _, attr := range attrs {
fmt.Printf("\t%s\n", attr.String())
}
}
This start giving us a bit more info on what's going on:
Parsed 28 messages
<Subsystem: NFSubsysNFTables, Message Type: 16, Family: ProtoUnspec, Version: 0, ResourceID: 0>
<Subsystem: NFSubsysNFTables, Message Type: 15, Family: ProtoUnspec, Version: 0, ResourceID: 47989>
<Length 4, Type 1, Nested false, NetByteOrder false, [0 0 187 117]>
<Length 4, Type 2, Nested false, NetByteOrder false, [0 0 32 10]>
<Length 4, Type 3, Nested false, NetByteOrder false, [110 102 116 0]>
<Subsystem: NFSubsysNFTables, Message Type: 1, Family: ProtoUnspec, Version: 0, ResourceID: 0>
<Subsystem: NFSubsysNFTables, Message Type: 0, Family: ProtoIPv4, Version: 0, ResourceID: 47989>
<Length 7, Type 1, Nested false, NetByteOrder false, [102 105 108 116 101 114 0]>
<Length 4, Type 2, Nested false, NetByteOrder false, [0 0 0 0]>
<Length 4, Type 3, Nested false, NetByteOrder false, [0 0 0 5]>
<Length 8, Type 4, Nested false, NetByteOrder false, [0 0 0 0 0 0 0 150]>
<Subsystem: NFSubsysNone, Message Type: 3, Family: ProtoUnspec, Version: 0, ResourceID: 0>
...truncated...
Now, it's time to start trying to understand these message types. This has been a quite long search, which eventually ended in the linux kernel sources, on the magic nf_tables_msg_types enum. (see nf_tables.h#L101).
Using the kernel sources comments and navigating the various enum, we end up replicating them to allow looking up those meaningfull names in place of the message types using 2 global variables:
// https://elixir.bootlin.com/linux/v5.15.5/source/include/uapi/linux/netfilter/nf_tables.h#L101
var messageTypeNames = []string{
"NFT_MSG_NEWTABLE",
"NFT_MSG_GETTABLE",
"NFT_MSG_DELTABLE",
"NFT_MSG_NEWCHAIN",
"NFT_MSG_GETCHAIN",
"NFT_MSG_DELCHAIN",
"NFT_MSG_NEWRULE",
"NFT_MSG_GETRULE",
// ... truncated ...
}
var attributeTypeNames = map[string][]string{
"NFT_MSG_NEWGEN": { // https://elixir.bootlin.com/linux/v5.15.5/source/include/uapi/linux/netfilter/nf_tables.h#L1505
"NFTA_GEN_UNSPEC",
"NFTA_GEN_ID",
"NFTA_GEN_PROC_PID",
"NFTA_GEN_PROC_NAME",
},
"NFT_MSG_NEWTABLE": { // https://elixir.bootlin.com/linux/v5.15.5/source/include/uapi/linux/netfilter/nf_tables.h#L181
"NFTA_TABLE_UNSPEC",
"NFTA_TABLE_NAME",
"NFTA_TABLE_FLAGS",
"NFTA_TABLE_USE",
"NFTA_TABLE_HANDLE",
"NFTA_TABLE_PAD",
"NFTA_TABLE_USERDATA",
"NFTA_TABLE_OWNER",
"__NFTA_TABLE_MAx",
},
"NFT_MSG_NEWCHAIN": { // https://elixir.bootlin.com/linux/v5.15.5/source/include/uapi/linux/netfilter/nf_tables.h#L218
"NFTA_CHAIN_UNSPEC",
"NFTA_CHAIN_TABLE",
"NFTA_CHAIN_HANDLE",
"NFTA_CHAIN_NAME",
"NFTA_CHAIN_HOOK",
"NFTA_CHAIN_POLICY",
// ... truncated ...
}
Now we update our loop once again to make use of those pretty names:
fmt.Printf("Parsed %d messages\n", len(messages))
for _, m := range messages {
header, attrs, err := netfilter.UnmarshalNetlink(m)
if err != nil {
panic(err)
}
// lookup message name from its message type
fmt.Printf("%s\n", messageTypeNames[header.MessageType])
for _, attr := range attrs {
// lookup attribute name from its message type and attribute type
// or just keep default string repr if no name exists
attributeName := attr.String()
attrNames, ok := attributeTypeNames[messageTypeNames[header.MessageType]]
if ok {
attributeName = attrNames[int(attr.Type)]
}
fmt.Printf("\t%s - %q\n", attributeName, attr.Data)
}
}
and tada! we can now put some sense on all of that:
Parsed 28 messages
NFT_MSG_GETGEN
NFT_MSG_NEWGEN
NFTA_GEN_ID - "\x00\x00\xbbu"
NFTA_GEN_PROC_PID - "\x00\x00 \n"
NFTA_GEN_PROC_NAME - "nft\x00"
NFT_MSG_GETTABLE
NFT_MSG_NEWTABLE
NFTA_TABLE_NAME - "filter\x00"
NFTA_TABLE_FLAGS - "\x00\x00\x00\x00"
NFTA_TABLE_USE - "\x00\x00\x00\x05"
NFTA_TABLE_HANDLE - "\x00\x00\x00\x00\x00\x00\x00\x96"
NFT_MSG_NEWCHAIN
NFT_MSG_GETCHAIN
NFT_MSG_NEWCHAIN
NFTA_CHAIN_TABLE - "filter\x00"
NFTA_CHAIN_HANDLE - "\x00\x00\x00\x00\x00\x00\x00\x01"
NFTA_CHAIN_NAME - "input\x00"
... truncated ...
From here, we clearly see client requests (such as the one issued using the nft commands - NFT_MSG_GET...) and server response (NFT_MSG_NEW...), which then get parsed by nft to display tables, rules or whatever was requested to the terminal.
So we now start to see some human readable table, chains, rules and other pieces of a routing table:
filter {
# chains
input {}
forward {}
output {
# rule 0x5
}
hack {
# rule 0xa
}
# set
flag {
# fake flag stored here
}
}
We're now having 2 rules and a set still containing raw binary that we have to decode further the attributes. Some of these attributes clearly contains multiple sub-attributes, so we can write a simple function to decode them all:
func printAttrRecursive(data []byte, level int) {
attrs, err := netfilter.UnmarshalAttributes(data)
if err != nil {
return
}
for _, attr := range attrs {
if len(attr.Data) > 0 {
fmt.Printf("%sType: %d: %q\n", strings.Repeat("\t", level), attr.Type, attr.Data)
printAttrRecursive(attr.Data, level+1)
}
}
}
and update our previous loop on attributes:
fmt.Printf("%s\n", messageTypeNames[header.MessageType])
for _, attr := range attrs {
// lookup attribute name from its message type and attribute type
// or just keep default string repr if no name exists
attributeName := attr.String()
attrNames, ok := attributeTypeNames[messageTypeNames[header.MessageType]]
if ok {
attributeName = attrNames[int(attr.Type)]
}
switch attributeName {
case "NFTA_SET_ELEM_LIST_ELEMENTS":
fmt.Printf("\t%s\n", attributeName)
printAttrRecursive(attr.Data, 2)
case "NFTA_RULE_EXPRESSIONS":
fmt.Printf("\t%s\n", attributeName)
printAttrRecursive(attr.Data, 2)
default:
fmt.Printf("\t%s - %q\n", attributeName, attr.Data)
}
}
Now we'll start manually replacing the various types with their constant names, by looking up the parent in the kernel sources.
We end up with a nice decoded routing table, after looking up every enum values / structs for all kind of attributes such as:
filter {
# set
flag {
0: "dnrgs{REDACTEDREDACTEDREDACTEDREDACTED}"
}
# chains
input {}
forward {}
output {
immediate {
NFTA_IMMEDIATE_DREG: 0
NFTA_IMMEDIATE_DATA:
NFTA_DATA_VALUE: "\xff\xff\xff\xfd"
NFTA_DATA_VERDICT : "hack\x00"
}
}
hack {
payload {
NFTA_PAYLOAD_DREG: 1
NFTA_PAYLOAD_BASE: NFT_PAYLOAD_TRANSPORT_HEADER
NFTA_PAYLOAD_OFFSET: 0x1c
NFTA_PAYLOAD_LEN: 8
}
cmp {
NFTA_CMP_SREG: 1
NFTA_CMP_OP: NFT_CMP_EQ
NFTA_CMP_DATA:
NFTA_DATA_VALUE: dd48d0cfd3103cd4
}
immediate {
NFTA_IMMEDIATE_DREG: 0x12
NFTA_IMMEDIATE_DATA:
NFTA_DATA_VALUE: 0
}
lookup {
NFTA_LOOKUP_SET: flag
NFTA_LOOKUP_SREG: 0x12
NFTA_LOOKUP_DREG: 1
NFTA_LOOKUP_FLAGS: 0
}
payload {
NFTA_PAYLOAD_SREG: 1
NFTA_PAYLOAD_BASE: 2
NFTA_PAYLOAD_OFFSET: 0x3c
NFTA_PAYLOAD_LEN: 0x27
NFTA_PAYLOAD_CSUM_TYPE: 0
NFTA_PAYLOAD_CSUM_OFFSET: 0
NFTA_PAYLOAD_CSUM_FLAGS: 0
}
}
}
From here, we're almost done! The hack rule is now human readable, and we can see that 8 bytes are loaded from our incoming packet (offset 0x1c from the NFT_PAYLOAD_TRANSPORT_HEADER section), then compared to the dd48d0cfd3103cd4 value, and when it matches, the flag value is loaded from the set, and its 0x27 bytes are written to the response packet at the offset 0x3c. It's now time to craft our packet!
Now having the expected payload and the various offsets, we can craft a packet. Seeing no mention of any specific protocol or ports on the routing rules, we could just use any. But given the specific offets of the flag (0x3c), we need somehow to control the response size. We could try forging some TCP packet, since the port 22 is open, but the response packet is to small to contains the flag. Some bigger packets could work later during the SSH handshake, but we have some easier options: ICMP !
Then, let's go with some easy to craft ICMP echo packets, where we can send a packet of at least 0x3c bytes and get it echoed back:
package main
import (
"bytes"
"encoding/hex"
"fmt"
"log"
"math/rand"
"net"
"golang.org/x/net/icmp"
"golang.org/x/net/ipv4"
)
func main() {
targetIP := "34.159.43.116"
payload, _ := hex.DecodeString("dd48d0cfd3103cd4")
payloadOffset := 0x1c
responseOffset := 0x3c
responseSize := 0x27
// craft the expected payload given the above offets
padding := bytes.Repeat([]byte("A"), payloadOffset-len(payload))
data := append(padding, payload...)
responseFill := bytes.Repeat([]byte("B"), (responseOffset - (payloadOffset + len(payload)) + responseSize))
data = append(data, responseFill...)
// Make a new ICMP message
m := icmp.Message{
Type: ipv4.ICMPTypeEcho, Code: 0,
Body: &icmp.Echo{
ID: rand.Int(), Seq: rand.Int(),
Data: data,
},
}
packet, err := m.Marshal(nil)
if err != nil {
panic(err)
}
conn, err := net.Dial("ip4:icmp", targetIP)
if err != nil {
log.Fatalf("Dial: %s\n", err)
}
n, err := conn.Write(packet)
if err != nil {
panic(err)
}
fmt.Printf("write %d bytes\n", n)
fmt.Println(hex.Dump(packet))
}
Now fire a tcpdump in a window:
sudo tcpdump -n host 34.159.43.116 -X
and run that script:
sudo go run ./packetforge/main.go
And the flag should show up on the tcpdump window:
0x0000: 4500 0077 db92 0000 3c01 e215 229f 2b74 E..w....<...".+t
0x0010: c0a8 b222 0000 59cd fd52 164f 4141 4141 ..."..Y..R.OAAAA
0x0020: 4141 4141 4141 4141 4141 4141 4141 4141 AAAAAAAAAAAAAAAA
0x0030: dd48 d0cf d310 3cd4 4242 4242 4242 4242 .H....<.BBBBBBBB
0x0040: 4242 4242 4242 4242 4242 4242 4242 4242 BBBBBBBBBBBBBBBB
0x0050: 4472 676e 537b 6338 6439 3862 3037 6434 DrgnS{c8d98b07d4
0x0060: 6332 6634 6133 6363 6332 6663 3130 6234 c2f4a3ccc2fc10b4
0x0070: 6636 3238 3135 7d f62815}
Note: we can't just read the packet response from the socket connection in the code, as the packet get modified by the routing table, and the checksum isn't recomputed (got: 0xfd52, want: 0xc180). We could've read it using a lower level connection (such as raw socket & syscalls), but tcpdump does the trick here.
Despite having missed the flag validation by a couple of minutes, this was a pretty fun and interesting challenge. The multiple rounds of decoding, each providing some more bits of information kept me hooked. Having only a very high level of understanding of nftables and overall packet routing, diving in the sources unveiled quite a lot of the magic of it, even if there's still lots of dark spot!
