Linux Security Summit Europe 2022

The io_uring subsystem was introduced in Linux v5.1 and provided a new way to do asynchronous I/O on Linux, improving on the existing AIO subsystem. Since then io_uring has been a source of active development, gaining the ability to delegate credentials across process boundaries in Linux v5.6. Unfortunately, all of this happened without engaging the LSM community, and as a result LSM access controls were not added to io_uring until Linux v5.16. This talk will discuss the challenges in adding LSM controls to the io_uring subsystem, thoughts on why the controls lagged the functional development, and what the LSM community might do to help reduce the changes of similar problems in the future.

Having a dynamically-sized trailing array at the end of a structure is a useful code construct in the Linux kernel. However, the lack of proper bounds-checking on such objects is concerning and worth making an effort to remediate it. This presentation is about the different strategies we, in the Kernel Self-Protection Project, have been following along the last couple of years to transform trailing zero-length and one-element arrays into modern C99 flexible-array members, and the problems we have run into in the process. We will touch on how this work is closely related and helpful to the most recent efforts to hardening key APIs like memcpy() and globally enabling options like -Warray-bounds. We will explain how we've been addressing the particularly challenging one-element array into flexible-array member transformations in UAPI. Also, we'll see how the fact that compilers like GCC have historically treated all trailing arrays, no matter what their size, as flexible-sized arrays, is problematic and causes trouble when it comes to perform proper array-bounds checking on such objects, and how the introduction of a new option like -fstrict-flex-arrays could help.

Confidential computing protects data in use by performing computation in a hardware-based Trusted Execution Environment (TEE). One use case is that a TEE virtual machine (VM) may want to transfer the workload to a device for acceleration. In order to maintain the confidentiality of the workload, the VM should perform the device attestation to ensure that the VM is communicating with an authentic device with expected status, such as secure version number and mutable firmware measurement version. This presentation will discuss the device attestation model and how to apply the model to support the device attestation in the TEE VM, including how the VM collects the device identity and evidence information and how the VM verifies the information before accepting the device for offloading the workload.

Heterogeneous SoC architectures enable a wide range of functionalities, notably for modern IoT/edge platforms. Modern SoCs contain heterogeneous CPUs (e.g., a combination of ARM and RISC-V architectures) and peripherals. As a result, the systems stack on such devices includes multiple OSs (e.g., Linux and FreeRTOS), hypervisors, or TEEs (trusted execution environments). Hence secure partitioning and sharing hardware resources within such complex system layers is challenging. Static hardware partitioning at boot time can not satisfy most use cases' security, performance, or compatibility requirements. This talk describes a hardware-assisted dynamic partitioning framework for Linux- and TEE-based heterogeneous architectures. We first summerise state-of-the-art hardware features for fine-grained privilege separation. Then we discuss how our solution modifies the Linux kernel, trusted firmware, and TEE kernel to achieve this goal while resolving various security and functionality challenges.

The Simplified Mandatory Access Control Kernel (Smack) Linux security module was introduced in 2008 and is currently used in millions of devices. Unlike SELinux and AppArmor, Smack has never been supported by a major Linux distribution. The greatest barriers to distribution support have been that only one "major" security module can be used at a time and that there has been no example of a Smack rule set to use as a basis for a distribution's policy. In this talk Casey Schaufler, the author of Smack, will describe a new effort to create a reference set of Smack rules and apply it to a major Linux distribution. The talk starts with a description of how Smack rules work and how they differ from SELinux and AppArmor policy. It moves on to cover the threats being addressed and how a "three domain" approach provides the required protection. How a distribution to target was chosen is revealed and what steps are taken to keep the scheme flexible enough to be useful elsewhere. The current state of the effort and the identified challenges are presented. Finally, there will be an invitation for additional participation in the project.

The user namespace has grown a lot since its introduction some 9 years ago. However, it is still very far from ubiquitous, even in the modern container space. In this talk, we'll be exploring the current state of things and delve into some of the exciting developments that have recently landed or will be landing very soon. This will cover the very exciting work done on the new VFS API and VFS idmap shifting, now making it very easy to setup containers without having to first mangle their root filesystem. More importantly, also allowing containers relying on shared filesystem layers to be easily run unprivileged. On the security front, we'll be covering the work to better mediate the use of the user namespace, allowing LSMs to decide who can or cannot create a user namespace. As well as the recent addition of IMA namespacing now makes it possible to have an entire system measured and checked, containers included. We'll wrap things up looking ahead for any other major blocker to the adoption of user namespace and the deprecation of much less safe container options.

We go through the current state of the Intel SGX support in the Linux kernel and userland. The topics covered include the kernel interface and its features, and available confidential computing run-times supporting SGX. Since getting into the mainline kernel late 2020, the SGX software ecosystem has started to get mature enough for production, and is the only cloud-scale confidential computing technology fully in the mainline kernel so far.

Confidential computing protects data in use by performing computation in a hardware-based Trusted Execution Environment (TEE) such as Intel SGX or Intel’s upcoming Trust Domain Extensions (Intel TDX). Today, PCIe-attached devices are outside the TEE’s trust boundary and not allowed to read/write confidential memory. This limitation requires TEE VMs to stage the data sent or received from devices in a shared memory buffer accessible to the TEE, IO device and VMM. Further, to protect the confidentiality and integrity of data in the buffer, the data must be encrypted by the sender (either the IO device or TEE VM), transferred to the shared buffer, then decrypted by the receiving entity using software-based encryption/decryption. This process results in additional latency and overhead that negatively impacts application performance. This presentation will focus on security and software changes required to support IO in trusted execution environments. The software requirements for TEE VMs to securely use a device in the Trusted Computing Base with DMA operations against confidential memory using encryption/decryption will be discussed.

Canonical uses snap application sand boxing to improve application security. While applications can be rewritten or modified to use portals and other privilege separation there are many applications that need to be run from confinement without modification. This presentation will cover the set of techniques being used and/or experimented with to improve application confinement without over burdening the user. This includes a variety of different techniques from notifying notifying userspace to allow it to provide policy updates and a more nuanced response, application and file tagging, to providing better control over environment variables and dynamic policy composition.