-
Notifications
You must be signed in to change notification settings - Fork 274
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
add autonat v2 spec #538
add autonat v2 spec #538
Changes from all commits
d663611
1db8613
0ff8ac6
f2a431c
62123df
0771bab
3e57202
f6def9a
b769d79
e4efaae
f28511c
d4da279
5b8d37d
05a0de2
8b52643
dd2750c
4e6ecaa
2af3309
6b1604b
f979fac
209b215
094089b
b4a856b
1c76613
03718ef
0195203
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,7 @@ | ||
# NAT Discovery <!-- omit in toc --> | ||
> How we detect if we're behind a NAT. | ||
|
||
Specifications: | ||
- [autonat v1](autonat-v1.md) | ||
- [autonat v2](autonat-v2.md) |
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,17 @@ | ||
@startuml | ||
participant Cli | ||
participant Srv | ||
|
||
skinparam sequenceMessageAlign center | ||
skinparam defaultFontName monospaced | ||
|
||
|
||
== Amplification Attack Prevention == | ||
|
||
Cli -> Srv: [dial] DialRequest:{nonce: 0xabcd, addrs: (addr1, addr2, addr3)} | ||
Srv -> Cli: [dial] DialDataRequest:{addrIdx: 1, numBytes: 120k} | ||
Cli -> Srv: [dial] {120k bytes} | ||
Srv -> Cli: [attempt]addr2 DialAttempt:{nonce: 0xabcd} | ||
Srv -> Cli: [dial] DialResponse:{status: OK, dialStatuses:(E_TRANSPORT_NOT_SUPPORTED, OK)} | ||
|
||
@enduml |
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,299 @@ | ||
# AutonatV2: spec | ||
sukunrt marked this conversation as resolved.
Show resolved
Hide resolved
|
||
|
||
|
||
| Lifecycle Stage | Maturity | Status | Latest Revision | | ||
|-----------------|--------------------------|--------|-----------------| | ||
| 1A | Working Draft | Active | r2, 2023-04-15 | | ||
|
||
Authors: [@sukunrt] | ||
|
||
Interest Group: [@marten-seemann], [@marcopolo], [@mxinden] | ||
|
||
[@sukunrt]: https://github.com/sukunrt | ||
[@marten-seemann]: https://github.com/marten-seemann | ||
[@mxinden]: https://github.com/mxinden | ||
[@marcopolo]: https://github.com/marcopolo | ||
|
||
|
||
## Overview | ||
|
||
A priori, a node cannot know if it is behind a NAT / firewall or if it is | ||
publicly reachable. Moreover, the node may be publicly reachable on some of its | ||
addresses and not on others. Knowing reachability for its addresses is essential | ||
for the node to be well-behaved in the network: A node doesn't need to advertise | ||
its unreachable addresses to the rest of the network, preventing superfluous | ||
dials from other peers. Furthermore, in case it has no publicly reachable | ||
addresses, it might actively seek to improve its connectivity by finding a relay | ||
server, which would allow other peers to establish a relayed connection. | ||
|
||
In `autonat v2` client sends a request with a priority ordered list of addresses | ||
and a nonce. On receiving this request the server dials the first address in the | ||
list that it is capable of dialing and provides the nonce. Upon completion of | ||
the dial, the server responds to the client with the response containing the | ||
dial outcome. | ||
|
||
As the server dials _exactly_ one address from the list, `autonat v2` allows | ||
nodes to determine reachability for individual addresses. Using `autonat v2` | ||
nodes can build an address pipeline where they can test individual addresses | ||
discovered by different sources like identify, upnp mappings, circuit addresses | ||
etc for reachability. Having a priority ordered list of addresses provides the | ||
ability to verify low priority addresses. Implementations can generate low | ||
priority address guesses and add them to requests for high priority addresses as | ||
a nice to have. This is especially helpful when introducing a new transport. | ||
Initially, such a transport will not be widely supported in the network. | ||
Requests for verifying such addresses can be reused to get information about | ||
other addresses | ||
|
||
The client can verify the server did successfully dial an address of the same | ||
transport as it reported in the response by checking the local address of the | ||
connection on which the nonce was received on. | ||
|
||
Compared to `autonat v1` there are three major differences | ||
1. `autonat v1` allowed testing reachability for the node. `autonat v2` allows | ||
testing reachability for an individual address. | ||
2. `autonat v2` provides a mechanism for nodes to verify whether the peer | ||
actually successfully dialled an address. | ||
3. `autonat v2` provides a mechanism for nodes to dial an IP address different | ||
from the requesting node's observed IP address without risking amplification | ||
attacks. `autonat v1` disallowed such dials to prevent amplification attacks. | ||
|
||
|
||
## AutoNAT V2 Protocol | ||
sukunrt marked this conversation as resolved.
Show resolved
Hide resolved
|
||
|
||
![Autonat V2 Interaction](autonat-v2.svg) | ||
|
||
A client node wishing to determine reachability of its addresses sends a | ||
`DialRequest` message to a server on a stream with protocol ID | ||
`/libp2p/autonat/2/dial-request`. Each `DialRequest` is sent on a new stream. | ||
|
||
This `DialRequest` message has a list of addresses and a fixed64 `nonce`. The | ||
list is ordered in descending order of priority for verification. AutoNAT V2 is | ||
primarily for testing reachability on Public Internet. Client SHOULD NOT send any | ||
private address as defined in [RFC | ||
1918](https://datatracker.ietf.org/doc/html/rfc1918#section-3) in the list. The Server SHOULD NOT dial any private address. | ||
|
||
Upon receiving this request, the server selects an address from the list to | ||
dial. The server SHOULD use the first address it is willing to dial. The server | ||
MUST NOT dial any address other than this one. If this selected address has an | ||
IP address different from the requesting node's observed IP address, server | ||
initiates the Amplification attack prevention mechanism (see [Amplification | ||
Attack Prevention](#amplification-attack-prevention) ). On completion, the | ||
server proceeds to the next step. If the selected address has the same IP | ||
address as the client's observed IP address, server proceeds to the next step | ||
skipping Amplification Attack Prevention steps. | ||
|
||
The server dials the selected address, opens a stream with Protocol ID | ||
`/libp2p/autonat/2/dial-back` and sends a `DialBack` message with the nonce | ||
received in the request. The client on receiving this message replies with | ||
a `DialBackResponse` message with the status set to `OK`. The client MUST | ||
close this stream after sending the response. The dial back response provides | ||
the server assurance that the message was delivered so that it can close the | ||
connection. | ||
|
||
Upon completion of the dial back, the server sends a `DialResponse` message to | ||
the client node on the `/libp2p/autonat/2/dial-request` stream. The response | ||
contains `addrIdx`, the index of the address the server selected to dial and | ||
`DialStatus`, a dial status indicating the outcome of the dial back. The | ||
`DialStatus` for an address is set according to [Requirements for | ||
DialStatus](#requirements-for-dialstatus). The response also contains an | ||
appropriate `ResponseStatus` set according to [Requirements For | ||
ResponseStatus](#requirements-for-responsestatus). | ||
|
||
The client MUST check that the nonce received in the `DialBack` is the same as | ||
the nonce it sent in the `DialRequest`. If the nonce is different, it MUST | ||
discard this response. | ||
|
||
The server MUST close the stream after sending the response. The client MUST | ||
close the stream after receiving the response. | ||
|
||
|
||
### Requirements for DialStatus | ||
|
||
On receiving a `DialRequest`, the server first selects an address that it will | ||
dial. | ||
|
||
If server chooses to not dial any of the requested addresses, `ResponseStatus` | ||
is set to `E_DIAL_REFUSED`. The fields `addrIdx` and `DialStatus` are | ||
meaningless in this case. See [Requirements For | ||
ResponseStatus](#requirements-for-responsestatus). | ||
|
||
If the server selects an address for dialing, `addrIdx` is set to the | ||
index(zero-based) of the address on the list and the `DialStatus` is set | ||
according to the following consideration: | ||
|
||
If the server was unable to connect to the client on the selected address, | ||
`DialStatus` is set to `E_DIAL_ERROR`, indicating the selected address is not | ||
publicly reachable. | ||
|
||
If the server was able to connect to the client on the selected address, but an | ||
error occured while sending an nonce on the `/libp2p/autonat/2/dial-back` | ||
stream, `DialStatus` is set to `E_DIAL_BACK_ERROR`. This might happen in case of | ||
resource limited situations on client or server, or when either the client or | ||
the server is misconfigured. | ||
|
||
If the server was able to connect to the client and successfully send a nonce on | ||
the `/libp2p/autonat/2/dial-back` stream, `DialStatus` is set to `OK`. | ||
|
||
### Requirements for ResponseStatus | ||
|
||
The `ResponseStatus` sent by the server in the `DialResponse` message MUST be | ||
set according to the following requirements | ||
|
||
`E_REQUEST_REJECTED`: The server didn't serve the request because of rate | ||
limiting, resource limit reached or blacklisting. | ||
|
||
`E_DIAL_REFUSED`: The server didn't dial back any address because it was | ||
incapable of dialing or unwilling to dial any of the requested addresses. | ||
|
||
`E_INTERNAL_ERROR`: Error not classified within the above error codes occured on | ||
server preventing it from completing the request. | ||
|
||
`OK`: the server completed the request successfully. A request is considered | ||
completed successfully when the server either completes a dial(successfully or | ||
unsuccessfully) or rejects all addresses in the request as undialable. | ||
|
||
Implementations MUST discard responses with status codes they do not understand. | ||
|
||
### Amplification Attack Prevention | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. @mxinden mention that this design is takes inspiration from a different protocol but I forgot which one. Might be worth mentioning! There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. E.g. related to the QUIC anti-amplification mechanism based on data size:
https://www.rfc-editor.org/rfc/rfc9000.html#section-8.1 I am not aware of an exact origin of the AutoNATv2 mechanism. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. QUIC is always a good source of inspiration :) The specifics here is something that we came up with ourselves, in the discussion on the issue. If you're aware of any prior art (RFCs?), please let us know. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. should I mention QUIC? |
||
|
||
![Interaction](autonat-v2-amplification-attack-prevention.svg) | ||
|
||
When a client asks a server to dial an address that is not the client's observed | ||
IP address, the server asks the client to send some non trivial amount of bytes | ||
as a cost to dial a different IP address. To make amplification attacks | ||
unattractive, servers SHOULD ask for 30k to 100k bytes. Since most handshakes | ||
cost less than 10k bytes in bandwidth, 30kB is sufficient to make attacks | ||
unattractive. | ||
|
||
On receiving a `DialRequest`, the server selects the first address it is capable | ||
of dialing. If this selected address has a IP different from the client's | ||
observed IP, the server sends a `DialDataRequest` message with the selected | ||
address's index(zero-based) and `numBytes` set to a sufficiently large value on | ||
the `/libp2p/autonat/2/dial-request` stream | ||
|
||
Upon receiving a `DialDataRequest` message, the client decides whether to accept | ||
or reject the cost of dial. If the client rejects the cost, the client resets | ||
the stream and the `DialRequest` is considered aborted. If the client accepts | ||
the cost, the client starts transferring `numBytes` bytes to the server. The | ||
client transfers these bytes wrapped in `DialDataResponse` protobufs where the | ||
`data` field in each individual protobuf is limited to 4096 bytes in length. | ||
This allows implementations to use a small buffer for reading and sending the | ||
data. Only the size of the `data` field of `DialDataResponse` protobufs is | ||
counted towards the bytes transferred. Once the server has received at least | ||
numBytes bytes, it proceeds to dial the selected address. Servers SHOULD allow | ||
the last `DialDataResponse` message received from the client to be larger than | ||
the minimum required amount. This allows clients to serialize their | ||
`DialDataResponse` message once and reuse it for all Requests. | ||
|
||
|
||
If an attacker asks a server to dial a victim node, the only benefit the | ||
attacker gets is forcing the server and the victim to do a cryptographic | ||
handshake which costs some bandwidth and compute. The attacker by itself can do | ||
a lot of handshakes with the victim without spending any compute by using the | ||
same key repeatedly. The only benefit of going via the server to do this attack | ||
is not spending bandwidth required for a handshake. So the prevention mechanism | ||
only focuses on bandwidth costs. There is a minor benefit of bypassing IP | ||
blocklists, but that's made unattractive by the fact that servers may ask 5x | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I don't think we can simply shrug this off. This is called a reflection attack and has been a huge issue for open DNS resolvers. Fixing the amplification side does go a long way, but paying a 5x bandwidth cost for a bunch of free IP addresses seems like a pretty reasonable tradeoff from an attacker's standpoint (especially because said attacker isn't paying for the bandwidth, but likely needs to compromise one machine per IP address). There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Also note: home NAT users likely don't need this feature. That is:
Being willing to dial other addresses does matter for, e.g., AWS and other special settings where there are separate ingress IP addresses. But, in that case, maybe the user should just configure their node correctly rather than relying on AutoNAT? AutoNAT specifically exists to enable home users. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. It simplifies client implementations as they don't need to worry about IPv4 peer vs IPv6 peer. Though the benefit isn't huge since most IPv4 servers won't have IPv6 connectivity so they any way cannot check the IPv6 address.
Can you elaborate here? why isn't the attacker paying for the bandwidth. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more.
@MarcoPolo @Stebalien @umgefahren thoughts? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. The potential attack vector seems to be described correctly, however I'm not sure that rust-libp2p is affected. The whole process is allowed to take a maximum of 10 seconds: Regarding IPv4 and IPv6 I stand with @thomaseizinger's comment on that matter: umgefahren/rust-libp2p#1 (comment) So it correctly handles that case, in that we don't generate false positives. |
||
more data than the bandwidth cost of a handshake. | ||
|
||
## Implementation Suggestions | ||
|
||
For any given address, client implementations SHOULD do the following | ||
- Periodically recheck reachability status. | ||
- Query multiple servers to determine reachability. | ||
|
||
The suggested heuristic for implementations is to consider an address reachable | ||
if more than 3 servers report a successful dial and to consider an address | ||
unreachable if more than 3 servers report unsuccessful dials. Implementations | ||
are free to use different heuristics than this one | ||
|
||
Servers SHOULD NOT reuse their listening port when making a dial back. In case | ||
the client has reused their listen port when dialing out to the server, not | ||
reusing the listen port for attempts prevents accidental hole punches. Clients | ||
SHOULD only rely on the nonce and not on the peerID for verifying the dial back | ||
as the server is free to use a separate peerID for the dial backs. | ||
|
||
Servers SHOULD determine whether they have IPv6 and IPv4 connectivity. IPv4 only servers SHOULD refuse requests for dialing IPv6 addresses and IPv6 only | ||
servers SHOULD refuse requests for dialing IPv4 addresses. | ||
|
||
|
||
## RPC Messages | ||
sukunrt marked this conversation as resolved.
Show resolved
Hide resolved
|
||
|
||
All RPC messages sent over a stream are prefixed with the message length in | ||
bytes, encoded as an unsigned variable length integer as defined by the | ||
[multiformats unsigned-varint spec][uvarint-spec]. | ||
|
||
All RPC messages on stream `/libp2p/autonat/2/dial-request` are of type | ||
`Message`. A `DialRequest` message is sent as a `Message` with the `msg` field | ||
set to `DialRequest`. `DialResponse` and `DialDataRequest` are handled | ||
similarly. | ||
|
||
On stream `/libp2p/autonat/2/dial-back`, a `DialAttempt` message is sent | ||
directly | ||
|
||
```proto3 | ||
|
||
message Message { | ||
oneof msg { | ||
DialRequest dialRequest = 1; | ||
DialResponse dialResponse = 2; | ||
DialDataRequest dialDataRequest = 3; | ||
DialDataResponse dialDataResponse = 4; | ||
} | ||
} | ||
|
||
|
||
message DialRequest { | ||
repeated bytes addrs = 1; | ||
fixed64 nonce = 2; | ||
} | ||
|
||
|
||
message DialDataRequest { | ||
uint32 addrIdx = 1; | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Instead of an index, why not send the address itself? In my eyes just a small complexity reduction. Feel free to ignore. Based on intuition, the additional bytes (index < address) doesn't have a performance impact. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I think your suggestion is right. I'll resolve this while taking up the implementation. This will help us avoid sharing the slice between the thread that makes the dial request and the thread that receives the response. Similar effect can be had by sharing the idx to the requesting thread, but this seems simpler to me. I'll resolve this while implementing. |
||
uint64 numBytes = 2; | ||
} | ||
|
||
|
||
enum DialStatus { | ||
UNUSED = 0; | ||
E_DIAL_ERROR = 100; | ||
E_DIAL_BACK_ERROR = 101; | ||
OK = 200; | ||
} | ||
|
||
|
||
message DialResponse { | ||
enum ResponseStatus { | ||
E_INTERNAL_ERROR = 0; | ||
E_REQUEST_REJECTED = 100; | ||
E_DIAL_REFUSED = 101; | ||
OK = 200; | ||
} | ||
|
||
ResponseStatus status = 1; | ||
uint32 addrIdx = 2; | ||
DialStatus dialStatus = 3; | ||
} | ||
|
||
|
||
message DialDataResponse { | ||
bytes data = 1; | ||
} | ||
|
||
|
||
message DialBack { | ||
fixed64 nonce = 1; | ||
} | ||
|
||
message DialBackResponse { | ||
enum DialBackStatus { | ||
OK = 0; | ||
} | ||
|
||
DialBackStatus status = 1; | ||
} | ||
``` | ||
|
||
[uvarint-spec]: https://github.com/multiformats/unsigned-varint | ||
|
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,20 @@ | ||
@startuml | ||
participant Cli | ||
participant Srv | ||
|
||
skinparam sequenceMessageAlign center | ||
skinparam defaultFontName monospaced | ||
|
||
|
||
== Dial Request Success== | ||
|
||
Cli -> Srv: [dial] DialRequest:{nonce: 0xabcd, addrs: (addr1, addr2, addr3)} | ||
Srv -> Cli: [attempt]addr2 DialAttempt:{nonce: 0xabcd} | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Not really sure how to achieve this, but it would be nice if it could be made clear that this is happening on a new connection (and a different 5-tuple), whereas the rest of the exchange happens on the same stream. |
||
Srv -> Cli: [dial] DialResponse:{status: OK, dialStatuses:(E_TRANSPORT_NOT_SUPPORTED, OK)} | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This mentions |
||
|
||
== Dial Request Failure== | ||
|
||
Cli -> Srv: [dial] DialRequest:{nonce: 0xabcd, addrs: (addr1, addr2, addr3)} | ||
Srv ->x Cli: [attempt]addr2 DialAttempt:{nonce: 0xabcd} | ||
Srv -> Cli: [dial] DialResponse:{status: OK, dialStatuses:(E_TRANSPORT_NOT_SUPPORTED, E_DIAL_ERROR)} | ||
@enduml |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think we should have a message for this too, otherwise implementing the decoder for these messages is going to be quite annoying because you need to attempt deserializing them and only if that fails, consider it bare "data".
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So the protocol is if the client accepts the DialDataRequest, it starts sending that amount of data. If the client doesn't accept the DialDataRequest, it resets the stream. You do not need to deserialise anything after you send DialDataRequest here. Just read that many bytes.
Wrapping the entire data in a protobuf will make the protobuf size very large.
Do you think adding a
DataRequestAccepted
message before sending that amount of data as raw bytes will make things cleaner?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The reason I am asking for this is because at least in
rust-libp2p
, we typically implement "codecs" which are modules that turn byte streams into typed streams (and sinks) of messages. Reading bare half-way through does not play that well with strong type systems.What do you mean it will make the protobuf large? Can't you just have a dynamically sized byte buffer in the protobuf? What is the difference?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I guess we could unwrap the stream again and read a specified number of bytes.
Thoughts @mxinden?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You'll need to bring the whole serialised protobuf into memory before decoding. This will take memory of the order of 100kb. Not prohibitive, but I'd like to avoid this if we can.
If I understand correctly, you would prefer it to be a protobuf so you can have a typed stream which says first object received from the stream is of type A and the second object received is of type B, right?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That might work, but you're now hardcoding assumptions about the wire format of the Protobuf, outside of your Protobuf library. I'd like to avoid that.
Not sure I agree with this framing, but in the end, this boils down to a matter of taste.
How about repeatedly sending a smaller message (limited by our usual Protobuf size), until we've reached the required number of bytes?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So in the current proposal because the number of bytes that you need to read in the second message is not fixed and is equal to the number you asked for in the
DialDataRequest
this prevents you from writing a generic reader, right?If that's the case we can prefix the bytes sent with a unsigned-varint just like we prefix other protobuf messages sent. This will also be consistent with the other messages exchanged everywhere which I'd argue is the source of inconsistency in the design as opposed to bytes not being sent as a TLV object.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
After doing a POC implementation, I think Marten's suggestion here is the most suited to implementation. The problem from a go perspective is buffering in the protobuf msg reader on the server side. This can happen if someone does a pipelined implementation on the client side which sends data as soon as it has sent the request. see documentation for NewDelimitedReader for another case where this happens. While this situation is unlikely, it is simpler to just use protobufs for transferring data.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah that makes sense. Would that message be a regular protobuf with a static array?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm using bytes.