crosvm/seccomp
Adrian Ratiu 1a7a822858 seccomp: allow rseq syscall
Starting with v4.18, the Linux kernel provides the rseq
syscall which is a mechanism for fast userspace task
synchronization.

Starting with v2.35 glibc uses the new syscall, if it
exists, to gain some performance improvements, so we
need to update the policy files to allow it.

Even on older kernels where rseq is not supported,
glibc will still probe for its existence by expecting
an -ENOSYS response.

BUG=b:235960683
TEST=Local builds against glibc 2.35

Change-Id: I704f2fbf2b058c3a4c3269c7441c3a7324012f8a
Reviewed-on: https://chromium-review.googlesource.com/c/crosvm/crosvm/+/3763901
Commit-Queue: Manoj Gupta <manojgupta@chromium.org>
Owners-Override: Dominick Ng <dominickn@google.com>
Reviewed-by: Allen Webb <allenwebb@google.com>
Reviewed-by: Dennis Kempin <denniskempin@google.com>
Tested-by: Manoj Gupta <manojgupta@chromium.org>
2022-07-15 16:19:21 +00:00
..
aarch64 seccomp: allow rseq syscall 2022-07-15 16:19:21 +00:00
arm seccomp: allow rseq syscall 2022-07-15 16:19:21 +00:00
x86_64 seccomp: allow rseq syscall 2022-07-15 16:19:21 +00:00
README.md seccomp: define naming rules for policy files 2022-06-17 04:35:09 +00:00

Policy files for crosvm

This folder holds the seccomp policies for crosvm devices, organized by architecture.

Each crosvm device can run within its owned jailed process. A jailed process is only able to perform the system calls specified in the seccomp policy file the jail has been created with, which improves security as a rogue process cannot perform any system call it wants.

Each device can run from different contexts, which require a different set of authorized system calls. This file explains how the policy files are named in order to allow these various scenario.

Naming conventions

Since Minijail only allows for one level of policy inclusion, we need to be a little bit creative in order to minimize policy duplication.

  • common_device.policy contains a set of syscalls that are common to all devices, and is never loaded directly - only included from other policy files.
  • foo.policy contains the set of syscalls that device foo is susceptible to use, regardless of the underlying virtio transport. This policy is also never loaded directly.
  • foo_device.policy is the policy that is loaded when device foo is used as an in-VMM (i.e. regular virtio) device. It will generally simply include common_device.policy as well as foo.policy.

When using vhost-user, the virtio protocol needs to be sent over a different medium, e.g. a Unix socket. Supporting this transport requires some extra system calls after the device is jailed, and thus dedicated policies:

  • vhost_user.policy contains the set of syscalls required by the regular (i.e. socket-based) vhost-user listener. It is never loaded directly.
  • vvu.policy contains the set of syscalls required by the VFIO-based vhost-user (aka Virtio-Vhost-User) listener. It is also never loaded directly.
  • foo_device_vhost_user.policy is the policy that is loaded when device foo is used as a regular vhost-user device. It will generally include common_device.policy, vhost_user.policy and foo.policy.
  • foo_device_vvu.policy is the policy that is loaded when device foo is used as a VVU device. It will generally include common_device.policy, vvu.policy and foo.policy.