fix: mount namespace propagation

[ritesh-harihar]

Fixes: #74

Summary

exec2 tasks could not see autofs-mounted filesystems (or any other host mount created after the task started) because unshare --mount-proc implicitly created a mount namespace with MS_PRIVATE propagation, severing it from the host's mount tree. Adding --propagation slave restores propagation in one direction — host mounts flow into the task, task mounts do not leak to the host.

Root Cause

exec2 creates its task process tree using unshare. The --mount-proc flag is required to give the new PID namespace a private /proc, but this forces unshare to create a new mount namespace, and without an explicit --propagation flag that namespace defaults to MS_PRIVATE for every mount point.

MS_PRIVATE does not prevent mounts that existed before the namespace was created from being visible (they are snapshotted in). It prevents mounts created on the host after the namespace exists from propagating in. This is exactly how autofs works, it creates mounts on demand, after tasks are already running.

The kernel records propagation type in /proc/self/mountinfo.

State	Root mount entry in /proc/self/mountinfo	Meaning
Before fix	... / rw,noatime - btrfs ...	✗ No peer group — MS_PRIVATE
After fix	... / rw,noatime master:35 - btrfs ...	✓ Slave of peer group 35 — receives host propagation

Options explored

Option	Issue
--propagation slave (chosen)	correct on every host, no new privileges, no config surface
--propagation unchanged	Works on some hosts, silently fails on others (it inherits whatever propagation the parent namespace has, which varies by host and container environment)
mount --make-slave in shim	Needs elevated privilege just to work around something unshare already handles natively
Config option	Exposes the broken state as a valid choice; no real use case for any value except slave

Practical example

• Before fix: task fails with No such file or directory — mount not even visible.
• After fix, without unveil: task fails with Permission denied — mount visible, Landlock blocks it.
• After fix, with unveil:

config {
  command = "/bin/cat"
  args    = ["/nfs/data/file.txt"]
  unveil  = ["r:/nfs/data"]
}

Task succeeds mount visible and Landlock allows access.

This PR fixes the prerequisite that was completely broken. The Landlock layer was already working correctly and just needs the operator to unveil the autofs paths as they would for any other filesystem path. The two layers are independent and both must be satisfied.

Testing

Details

 job "mount-test" {
  type = "batch"
  group "group" {
    task "check" {
      driver = "exec2"
      config {
        command = "/usr/bin/awk"
        args    = ["NR==1", "/proc/self/mountinfo"]
        unveil  = ["r:/proc"]
      }
    }
  }
}

Before:

riteshharihar@podman-dev:/Users/riteshharihar/Desktop/Src/nomad-driver-exec2$ nomad job run /tmp/mount-test.hcl
==> 2026-08-11T15:13:55+05:30: Monitoring evaluation "65364008"
    2026-08-11T15:13:55+05:30: Evaluation triggered by job "mount-test"
    2026-08-11T15:13:56+05:30: Allocation "ca908b94" created: node "e2feae5d", group "group"
    2026-08-11T15:13:56+05:30: Evaluation status changed: "pending" -> "complete"
==> 2026-08-11T15:13:56+05:30: Evaluation "65364008" finished with status "complete"

riteshharihar@podman-dev:/Users/riteshharihar/Desktop/Src/nomad-driver-exec2$ nomad alloc logs 
811 808 0:38 /scon/containers/01KS4F7XKHAVX614ZKDHGS499Q/rootfs / rw,noatime - btrfs /dev/vdb1 rw,nodatasum,nodatacow,ssd,discard,space_cache=v2,subvolid=333,subvol=/scon/containers/01KS4F7XKHAVX614ZKDHGS499Q

After:

riteshharihar@podman-dev:/Users/riteshharihar/Desktop/Src/nomad-driver-exec2$ nomad job run /tmp/mount-test.hcl
==> 2026-08-11T15:17:40+05:30: Monitoring evaluation "4a280b01"
    2026-08-11T15:17:40+05:30: Evaluation triggered by job "mount-test"
    2026-08-11T15:17:40+05:30: Allocation "3d37353c" created: node "7a43f8f2", group "group"
    2026-08-11T15:17:41+05:30: Allocation "3d37353c" status changed: "pending" -> "running" (Tasks are running)
    2026-08-11T15:17:41+05:30: Evaluation status changed: "pending" -> "complete"
==> 2026-08-11T15:17:41+05:30: Evaluation "4a280b01" finished with status "complete"

riteshharihar@podman-dev:/Users/riteshharihar/Desktop/Src/nomad-driver-exec2$ nomad alloc logs 3d37353c
811 808 0:38 /scon/containers/01KS4F7XKHAVX614ZKDHGS499Q/rootfs / rw,noatime master:35 - btrfs /dev/vdb1 rw,nodatasum,nodatacow,ssd,discard,space_cache=v2,subvolid=333,subvol=/scon/containers/01KS4F7XKHAVX614ZKDHGS499Q

• If a change needs to be reverted, we will roll out an update to the code within 7 days.

Changes to Security Controls

Are there any changes to security controls (access controls, encryption, logging) in this pull request? If so, explain.

[tgross]

LGTM! I don't think I realized we were implicitly creating a mount namespace with --mount-proc, although I suppose that was obvious in retrospect