mirror of
https://github.com/apple/container.git
synced 2026-10-16 02:59:21 +00:00
Michael Crosby
4294c0f37a
`ExecCommand.execInNamespaces` entered a hardcoded `CLONE_NEWCGROUP | CLONE_NEWPID | CLONE_NEWUTS | CLONE_NEWNS`, while the init process gets its namespaces from the OCI spec via `RunCommand.setupNamespaces` — which includes `CLONE_NEWIPC` whenever the spec lists an `ipc` namespace. `LinuxContainer` (and `LinuxPod`) always declare one, so the init process runs in a private IPC namespace and every exec'd process stays in the guest's root one. The effect is that an exec cannot see SysV IPC segments or POSIX message queues created by the workload — and anything the exec creates leaks into the guest's root namespace, where it outlives the container. IPC-namespaced sysctls (`kernel.shm*`, `kernel.msg*`, `kernel.sem`, `fs.mqueue.*`) also read back as the guest default, because they resolve against the reading process's IPC namespace. That last one is what fails critest's "should support safe sysctls" and "should support unsafe sysctls", which set such a sysctl and read it back over ExecSync. Add `CLONE_NEWIPC` to the mask. `setns(2)` into a namespace the caller is already in is a no-op, so the mask is now a superset of everything `setupNamespaces` can unshare rather than a set that has to be kept in sync by hand. `CLONE_NEWUSER` stays out deliberately: nothing puts a container in a user namespace today, and entering one carries constraints that deserve their own change. The integration test asserts the general invariant rather than the specific sysctl: inside an exec, `/proc/self/ns/*` must equal `/proc/1/ns/*` for `ipc`, `uts`, `mnt`, `pid`, `cgroup`, and `net`. `kernel.shm_rmid_forced` is checked alongside it as the consumer-visible symptom. Signed-off-by: michael_crosby <michael_crosby@apple.com>