Impact
Contrary to the OCI runtime specification, runc's implementation of the linux.resources.devices list was a black-list by default. This means that users who created their own config.json objects and didn't prefix a deny-all rule ({"allow": false, "permissions": "rwm"} or equivalent) were not provided protection by the devices cgroup. This would allow malicious containers (with sufficient privileges) to create arbitrary device inodes (assuming they have CAP_MKNOD) and operate on any device inodes they may have access to (assuming they have regular Unix DAC permissions).
However, most (if not all) programs that make use of runc include this deny-all rule. This was most likely added before the specification mandated a white-list of devices, and the fact that all programs wrote their own deny-all rule obscured the existence of this bug for several years. In fact, even the specification's examples include a default deny-all rule! We therefore believe that while this is a security bug (and has been fixed as such), it was almost certainly not exploitable in the wild due to the inclusion of default deny-all rules by all known users of runc -- hence why this advisory has low severity.
Patches
This issue has been fixed in a patch that was part of a larger rework of the devices cgroup code in runc -- which lead to the discovery of this security bug. Users should upgrade to 1.0.0-rc91 as soon as it is released, or wait for your distribution to backport the relevant fixes.
Workarounds
If you are using runc directly, ensure that there is a deny-all entry at the beginning of linux.resources.devices -- such an entry would look like {"allow": false, "permissions": "rwm"} (all other fields are ignored, though type must be set to "a" or null if it is present).
Users which consume runc through another program should check whether their containers are operating under a white-list -- this can be done by reading /sys/fs/cgroup/devices/devices.list inside the container. If the file contains only the entry a *:* rwm (meaning the cgroup is in black-list mode, which likely means "allow all device access") then your containers are vulnerable to this issue.
As always, we recommend in the strongest possible terms that all of our users enable user namespaces on all of their workloads (or pressure their vendors to do so). User namespaces are one of the most significant defense-in-depth protections you can enable for containers, and have prevented many container-related vulnerabilities (both kernel 0days as well as bugs in container runtimes, such as this one).
References
For more information
If you have any questions or comments about this advisory:
cyphar Analyst
Impact
Contrary to the OCI runtime specification,
runc's implementation of thelinux.resources.deviceslist was a black-list by default. This means that users who created their ownconfig.jsonobjects and didn't prefix a deny-all rule ({"allow": false, "permissions": "rwm"}or equivalent) were not provided protection by thedevicescgroup. This would allow malicious containers (with sufficient privileges) to create arbitrary device inodes (assuming they haveCAP_MKNOD) and operate on any device inodes they may have access to (assuming they have regular Unix DAC permissions).However, most (if not all) programs that make use of
runcinclude this deny-all rule. This was most likely added before the specification mandated a white-list of devices, and the fact that all programs wrote their own deny-all rule obscured the existence of this bug for several years. In fact, even the specification's examples include a default deny-all rule! We therefore believe that while this is a security bug (and has been fixed as such), it was almost certainly not exploitable in the wild due to the inclusion of default deny-all rules by all known users ofrunc-- hence why this advisory has low severity.Patches
This issue has been fixed in a patch that was part of a larger rework of the devices cgroup code in runc -- which lead to the discovery of this security bug. Users should upgrade to 1.0.0-rc91 as soon as it is released, or wait for your distribution to backport the relevant fixes.
Workarounds
If you are using
runcdirectly, ensure that there is a deny-all entry at the beginning oflinux.resources.devices-- such an entry would look like{"allow": false, "permissions": "rwm"}(all other fields are ignored, thoughtypemust be set to"a"ornullif it is present).Users which consume
runcthrough another program should check whether their containers are operating under a white-list -- this can be done by reading/sys/fs/cgroup/devices/devices.listinside the container. If the file contains only the entrya *:* rwm(meaning the cgroup is in black-list mode, which likely means "allow all device access") then your containers are vulnerable to this issue.As always, we recommend in the strongest possible terms that all of our users enable user namespaces on all of their workloads (or pressure their vendors to do so). User namespaces are one of the most significant defense-in-depth protections you can enable for containers, and have prevented many container-related vulnerabilities (both kernel 0days as well as bugs in container runtimes, such as this one).
References
For more information
If you have any questions or comments about this advisory: