Rootless Containers & Unresolved issues
- 1. Rootless Containers & Unresolved Issues Akihiro Suda / NTT (@_AkihiroSuda_) May 17, 2019 1
- 2. Agenda • Introduction to Rootless Containers • How it works • Adoption status • Unresolved issues • containerd dev plan 2
- 4. Rootless Containers • Run containers, runtimes, and orchestrators as a non-root user • Don’t confuse with: – usermod -aG docker penguin – docker run --user – dockerd --userns-remap 4
- 5. Motivation of Rootless Containers • To mitigate potential vulnerability of container runtimes and orchestrator (the primary motivation) • To allow users of shared machines (e.g. HPC) to run containers without the risk of breaking other users environments – Still unsuitable for “multi-tenancy” where you can’t really trust other users • To isolate nested containers, e.g. “Docker-in-Docker” 5
- 6. Runtime vulnerabilities • Docker “Shocker” (2014) – A malicious container was allowed to access the host file system, as CAP_DAC_READ_SEARCH was effective by default • Docker CVE-2014-9357 – A malicious docker build container could run arbitrary binary on the host as the root due to an LZMA archive issue • containerd #2001 (2018) – A malicious container image could remove /tmp on the host when the image was pulled (not when actually launched!) 6
- 7. Runtime vulnerabilities • Docker “Shocker” (2014) – A malicious container was allowed to access the host file system, as CAP_DAC_READ_SEARCH was effective by default • Docker CVE-2014-9357 – A malicious docker build container could run arbitrary binary on the host as the root due to an LZMA archive issue • containerd #2001 (2018) – A malicious container image could remove /tmp on the host when the image was pulled (not when actually launched!) 7 Vulnerability of daemons, not containers per se So --userns-remap is not effective
- 8. Runtime vulnerabilities • runc #1962 (2019) – Container break-out via /proc/sys/kernel/core_pattern or /sys/kernel/uevent_helper – Hosts with the initrd rootfs (DOCKER_RAMDISK) were affected (e.g. Minikube) • runc CVE-2019-5736 – Container break-out via /proc/self/exe 8
- 9. Other vulnerabilities • Kubernetes CVE-2017-1002101, CVE-2017-1002102 – A malicious container was allowed to access the host filesystem via vulnerabilities related to volumes • Kubernetes CVE-2018-1002105 – A malicious API call could be used to gain cluster-admin (and hence the root privileges on the nodes) • Git CVE-2018-11235 (affected Kubernetes gitRepo volumes) – A malicious repo could execute an arbitrary binary as the root when it was cloned 9
- 10. Other vulnerabilities • Kubernetes CVE-2017-1002101, CVE-2017-1002102 – A malicious container was allowed to access the host filesystem via vulnerabilities related to volumes • Kubernetes CVE-2018-1002105 – A malicious API call could be used to gain cluster-admin (and hence the root privileges on the nodes) • Git CVE-2018-11235 (affected Kubernetes gitRepo volumes) – A malicious repo could execute an arbitrary binary as the root when it was cloned 10 --userns-remap might not be effective
- 11. Play-with-Docker.com vulnerability • Play-with-Docker.com: Online Docker playground, implemented using Docker-in-Docker with custom AppArmor profiles • Malicious kernel module was loadable due to AppArmor misconfiguration (revealed on Jan 14, 2019) – Not really an issue of Docker 11https://www.cyberark.com/threat-research-blog/how-i-hacked-play-with-docker-and-remotely-ran-code-on-the-host/
- 12. What Rootless Containers can • Prohibit accessing files owned by other users • Prohibit modifying firmware and kernel (→ undetectable malware) • Prohibit other privileged operations like ARP spoofing, rebooting,... 12
- 13. What Rootless Containers cannot • If a container was broke out, the attacker still might be able to – Mine cryptocurrencies – Springboard-attack to other hosts • Not effective for kernel / VM/ HW vulns – But we could use gVisor together for mitigating some of them 13
- 14. How it works 14
- 15. User Namespaces • User namespaces allow non-root users to pretend to be the root • Root-in-UserNS can have “fake” UID 0 and also create other namespaces (MountNS, NetNS..) 15
- 16. User Namespaces 16 $ id -u 1001 $ ls -ln -rw-rw---- 1 1001 1001 42 May 1 12:00 foo $ docker-rootless run -v $(pwd):/mnt -it alpine / # id -u 0 / # ls -ln /mnt -rw-rw---- 1 0 0 42 May 1 12:00 foo
- 17. User Namespaces 17 $ docker-rootless run -v /:/host -it alpine / # ls -ln /host/dev/sda brw-rw---- 1 65534 65534 8, 0 May 1 12:00 /host/dev/sda / # cat /host/dev/sda cat: can’t open ‘/host/dev/sda’: Permission denied
- 18. Sub-users (and sub-groups) • Put users in your user account so you can be a user while you are being a user • Sub-users are used as non-root users in a container – USER in Dockerfile – docker run --user 18
- 19. Sub-users (and sub-groups) • If /etc/subuid contains “1001:100000:65536” • Having 65536 sub-users should be enough for most containers 19 0 1001 100000 165535 232 Host UserNS primary user sub-users start sub-users length 0 1 65536
- 20. Sub-users (and sub-groups) • Sub-users are configured via SUID binaries /usr/bin/{newuidmap, newgidmap} • SETUID binary can be dangerous; newuidmap & newgidmap had two CVEs so far: – CVE-2016-6252 (CVSS v3: 7.8): integer overflow issue – CVE-2018-7169 (CVSS v3: 5.3): supplementary GID issue 20
- 21. Sub-users (and sub-groups) • Also hard to maintain sub-users – LDAP / AD – Nesting user namespaces might need huge number of sub-users 21
- 22. Sub-users (and sub-groups) • Alternative way: Single-mapping mode • Does not require newuidmap/newgidmap • Ptrace and/or Seccomp can be used for intercepting syscalls to emulate sub-users – user.rootlesscontainers xattr can be used for chown emulation 22
- 23. Network Namespaces • An unprivileged user can create network namespaces along with user namespaces • With network namespaces, the user can – isolate abstract (pathless) UNIX sockets • important to prevent container breakout – create iptables rules – set up overlay networking with VXLAN – run tcpdump – ... 23
- 24. Network Namespaces • But an unprivileged user cannot set up veth pairs across the host and namespaces, i.e. No internet connection 24 The Internet Host UserNS + NetNS
- 25. Network Namespaces 25 • lxc-user-nic SUID binary allows unprivileged users to create veth, but we are not huge fun of SUID binaries • Our approach: use completely unprivileged usermode network (“Slirp”) with a TAP device TAP “Slirp” TAPFD send fd as a SCM_RIGHTS cmsg The Internet Host UserNS + NetNS
- 26. Network Namespaces Benchmark of several “Slirp” implementations: • slirp4netns (our own implementation based on QEMU Slirp) is the fastest because it avoids copying packets across the namespaces MTU=1500 MTU=4000 MTU=16384 MTU=65520 vde_plug 763 Mbps Unsupported Unsupported Unsupported VPNKit 514 Mbps 526 Mbps 540 Mbps Unsupported slirp4netns 1.07 Gbps 2.78 Gbps 4.55 Gbps 9.21 Gbps cf. rootful veth 52.1 Gbps 45.4 Gbps 43.6 Gbps 51.5 Gbps Benchmark: iperf3 (netns -> host), measured on Travis CI. See rootless-containers/rootlesskit#12 26
- 27. Multi-node networking • Flannel VXLAN is known to work – Encapsulates Ethernet packets in UDP packets – Provides L2 connectivity across rootless containers on different nodes • Other protocols should work as well, except ones that require access to raw Ethernet 27
- 28. Snapshotting • OverlayFS is currently unavailable in UserNS (except on Ubuntu kernel) • FUSE-OverlayFS can be used instead with kernel 4.18+ • XFS reflink can be also used to deduplicate files (but slow) 28
- 29. Cgroup • pam_cgfs can be used for delegating permissions to unprivileged users, but considered insecure by systemd folks https://github.com/containers/libpod/issues/1429 • cgroup2 provides proper support for delegation, but not adopted by OCI at the moment 29
- 30. Rootless Containers in Containers • Urge demand for building images on Kubernetes cluster • Seccomp and AppArmor needs to be disabled for the parent containers • To allow the children to mount procfs (pid-namespaced), maskedPaths and readonlyPaths for /proc/* for the parent needs to be removed (weird!) – Same applies to sysfs (net-namespaced) 30
- 31. Rootless Containers in Containers • So --privileged had been typically required anyway :( – Or at least --security-opt {seccomp,apparmor}=unconfined • Docker 19.03 supports --security-opt systempaths=unconfined for allowing procfs & sysfs mount (Kube: securityContext.procMount, but no sysMount yet) – Make sure to lock the root in the container! (passwd -l root, Alpine CVE-2019-5021 ) 31
- 33. Adoption status: runtimes 33 Docker v19.03 containerd runc Podman (≈ CRI-O) crun LXC Singularity NetNS isolation with Internet connectivity ● VPNKit ● slirp4netns ● lxc-user-nic (SUID) slirp4netns lxc-user-nic (SUID) No support Supports FUSE-OverlayFS No Yes No No Cgroup No Limited support for cgroup2 pam_cgfs No
- 34. Adoption status: runtimes::GPU • nvidia-container-runtime is known to work • Need to disable cgroup manually • Rootful nVIDIA container needs to be executed on every system startup • Probably, other devices such as FPGA should work as well (untested) 34
- 35. Adoption status: runtimes::single-mapping mode • udocker does not need subuid configuration, as it can emulate subuser with ptrace (based on PRoot) – but no persistent chown • runROOTLESS (Don’t confuse with upstream rootless runc) supports persistent chown as well, using user.rootlesscontainers xattr – the xattr value is a pair of UID and GID in protobuf encoding – the xattr convention is compatible with umoci 35
- 36. Adoption status: runtimes::single-mapping mode • Ptrace is slow https://github.com/rootless-containers/runrootless/issues/14 • seccomp can be used for acceleration but hard to implement correctly 36
- 37. Adoption status: image builders • BuildKit / img / Buildah supports rootless mode – Works in containers as well as on the host – Does not need --privileged but Seccomp and AppArmor needs to be disabled 37
- 38. Adoption status: image builders • Similar but different work: Kaniko & Makisu – Rootful – But no need to disable seccomp and AppArmor, because they don’t create containers for RUN instructions in Dockerfile 38
- 39. Adoption status: Kubernetes • Usernetes project provides patches for rootless Kubernetes, but not proposed to the upstream yet – Supports all major CRI runtimes: dockershim, containerd, CRI-O – Flannel VXLAN is known to work – Lack of cgroup might be huge concern • But Usernetes is already integrated into k3s! (5 less than k8s) 39 $ k3s server --rootless
- 40. You can rootlesify your own project easily! • RootlessKit does almost all things for rootlessifying your container project (or almost any rootful app) – Creates UserNS with sub-users and sub-groups – Creates MountNS with writable /etc, /run but without chroot – Creates NetNS with VPNKit/slirp4netns/lxc-user-nic – Provides REST API on UNIX socket for port forwarding management 40
- 41. You can rootlesify your own project easily! 41 $ rootlesskit --net=slirp4netns --copy-up=/etc --port-driver=builtin bash # id -u 0 # touch /etc/here-is-writable-tmpfs # ip a ... 2: tap0: <BROADCAST,MULTICAST,UP,LOWER_UP> inet 10.0.2.100/24 scope global tap0 ... # rootlessctl add-ports 0.0.0.0:8080:80/tcp
- 42. You can rootlesify your own project easily! • With RootlessKit, you just need to work on disabling cgroup stuff, sysctl stuff, and changing the data path from /var/lib to /home • Used by Docker, BuildKit, k3s 42
- 44. Kernel has vulns • UserNS tends to have priv escalation vulns – CVE 2013-1858: UserNS + CLONE_FS – CVE-2014-4014: UserNS + chmod – CVE-2015-1328: UserNS + OverlayFS (Ubuntu-only) • So rootless OverlayFS is still not merged in upstream – CVE-2018-18955: UserNS + complex ID mapping 44
- 45. Kernel has vulns • A bunch of code paths that can hang up the kernel – e.g. CVE-2018-7191 (unpublished published today): creating a tap device with illegal name – And more, see https://medium.com/@jain.sm/security-challenges-with-kubernetes-818fad4a89f2 • Unlimited resources e.g. – Pending signals – Max user process – Max FDs per user (see the same URL above) 45
- 46. Kernel has vulns • So I’ve never suggested using rootless containers for real multi-tenancy ¯_(ツ)_/¯ 46
- 47. Kernel has vulns • gVisor might be able to mitigate them but significant overhead and syscall incompatibility • UML (20 yo, still alive!) is almost compatible with real Linux but it even lacks support for SMP • linuxd: similar to UML but accelerated with host kernel patches – Still no public code https://schd.ws/hosted_files/ossna18/db/Containerize%20Linux%20Kernel.pdf 47
- 48. Cgroups • cgroup2 is not adopted in OCI • crun is trying to support cgroup2 without changing OCI spec 48
- 49. Mount • Only supports: – tmpfs – bind – procfs (PID-namespaced) – sysfs (net-namespaced) – FUSE (since kernel 4.18) – Overlay (Ubuntu only) • No support for mounting any block devices (even loopback devices) 49
- 50. Landlock • landlock: unprivileged sandbox LSM • Not merged in the upstream kernel, but promising as AppArmor-alternative 50
- 51. LDAP / Active Directory • /etc/sub{u,g}id configuration is painful for LDAP/AD • Alternatively, implementing NSS module is under discussion, but no code yet https://github.com/shadow-maint/shadow/issues/154 51
- 52. Single-mapping mode • runROOTLESS / PRoot could be accelerated with seccomp but implementation is broken • Kernel 5.0 seccomp could be used for getting rid of ptrace completely 52
- 54. containerd dev plan • Implement FUSE-OverlayFS snapshotter plugin – Probably in a separate repo – Should not be difficult • Support cgroup2 – Probably we want to wait for OCI Runtime Spec and runc to be revised – But we can also consider beginning support cgroup2 right now with crun 54
- 55. containerd dev plan • Support running containerd inside gVisor – So as to allow running rootless containers in a container without disabling seccomp & apparmor – And to mitigate potential kernel vulns – Currently MountNS is not working https://github.com/google/gvisor/issues/221 55