Opening a DGRAM socket from within a docker container fails (permission denied)

I'm running an application which builds and sends ICMP ECHO requests to a few different ip addresses. The application is written in Crystal. When attempting to open a socket from within the crystal docker container, Crystal raises an exception: Permission Denied.

From within the container, I have no problem running ping 8.8.8.8.

Running the application on macos, I have no problem.

Reading the https://docs.docker.com/engine/security/apparmor/ and https://docs.docker.com/engine/security/seccomp/ pages on apparmor and seccomp I was sure I'd found the solution, but the problem remains unresolved, even when running as docker run --rm --security-opt seccomp=unconfined --security-opt apparmor=unconfined socket_permission

update/edit: After digging into capabilities(7), I added the following line to my dockerfile: RUN setcap cap_net_raw+ep bin/ping trying to let the socket get opened but without change.

Thanks!

Relevant crystal socket code, full working code sample below:

  # send request
  address = Socket::IPAddress.new host, 0
  socket = IPSocket.new Socket::Family::INET, Socket::Type::DGRAM, Socket::Protocol::ICMP
  socket.send slice, to: address

Dockerfile:

FROM crystallang/crystal:0.23.1
WORKDIR /opt
COPY src/ping.cr src/
RUN mkdir bin

RUN crystal -v
RUN crystal build -o bin/ping src/ping.cr

ENTRYPOINT ["/bin/sh","-c"]
CMD ["/opt/bin/ping"]

Running the code, first native, then via docker:

#!/bin/bash
crystal run src/ping.cr
docker build -t socket_permission .
docker run --rm --security-opt seccomp=unconfined --security-opt apparmor=unconfined socket_permission

And finally, a 50 line crystal script which fails to open a socket in docker:

require "socket"

TYPE = 8_u16
IP_HEADER_SIZE_8 = 20
PACKET_LENGTH_8 = 16
PACKET_LENGTH_16 = 8
MESSAGE = " ICMP"

def ping
  sequence = 0_u16
  sender_id = 0_u16
  host = "8.8.8.8"

  # initialize packet with MESSAGE
  packet = Array(UInt16).new PACKET_LENGTH_16 do |i|
    MESSAGE[ i % MESSAGE.size ].ord.to_u16
  end

  # build out ICMP header
  packet[0] = (TYPE.to_u16 << 8)
  packet[1] = 0_u16
  packet[2] = sender_id
  packet[3] = sequence

  # calculate checksum
  checksum = 0_u32
  packet.each do |byte|
    checksum += byte
  end
  checksum += checksum >> 16
  checksum = checksum ^ 0xffff_ffff_u32
  packet[1] = checksum.to_u16

  # convert packet to 8 bit words
  slice = Bytes.new(PACKET_LENGTH_8)

  eight_bit_packet = packet.map do |word|
    [(word >> 8), (word & 0xff)]
  end.flatten.map(&.to_u8)

  eight_bit_packet.each_with_index do |chr, i|
    slice[i] = chr
  end

  # send request
  address = Socket::IPAddress.new host, 0
  socket = IPSocket.new Socket::Family::INET, Socket::Type::DGRAM, Socket::Protocol::ICMP
  socket.send slice, to: address

  # receive response
  buffer = Bytes.new(PACKET_LENGTH_8 + IP_HEADER_SIZE_8)
  count, address = socket.receive buffer
  length = buffer.size
  icmp_data = buffer[IP_HEADER_SIZE_8, length-IP_HEADER_SIZE_8]
end

ping

Solution

It turns out the answer is that Linux (and by extension docker) does not give the same permissions that macOS does for DGRAM sockets. Changing the socket declaration to socket = IPSocket.new Socket::Family::INET, Socket::Type::RAW, Socket::Protocol::ICMP allows the socket to connect under docker.

A little more still is required to run the program in a non-root context. Because raw sockets are restricted to root, the binary must also be issued the correct capability for access to a raw socket, CAP_NET_RAW. However, in docker, this isn't necessary. I was able to get the program to run outside of super-user context by running sudo setcap cap_net_raw+ep bin/ping. This is a decent primer on capabilities and the setpcap command

MacOS doesn't use the same system of permissions, so setcap is just an unrecognized command. As a result, to get the above code to compile and run successfully on macOS without super-user context, I changed the socket creation code to:

socket_type = Socket::Type::RAW

{% if flag?(:darwin) %}
  socket_type = Socket::Type::DGRAM
{% end %}

socket = IPSocket.new Socket::Family::INET, socket_type, Socket::Protocol::ICMP

Applying the CAP_NET_RAW capability for use in linux happens elsewhere in the build process if needed.

With those changes, I'm not seeing any requirement for changes to seccomp or apparmor from the default shipped with Docker in order to run the program.