Skip to content

Commit a8842ae

Browse files
committed
docs(podman): document and track the e2e-podman rootful coverage gap
Address the rootless/rootful-scope acceptance criterion in NVIDIA#3712: document why the e2e-podman archive runs rootless-only in branch CI (unlike driver-podman, which runs both), and file a tracking issue for the one confirmed gap. - podman_host_gateway was checked directly rather than assumed: the outer container's host-gateway alias uses Podman's own native "host-gateway" value (resolved identically in both modes), and the supervisor-side resolution defaults to a hardcoded 127.0.0.1 on Linux regardless of rootful/rootless. No real coverage gap there. - podman_resource_limits is the one confirmed exception: it reads real cgroup v2 state, and rootless cgroup delegation is the harder, already-covered case. Rootful is unverified. Tracked in NVIDIA#4163, which also notes the e2e-podman tmachine playbook hardcodes rootless-only values and would need the same tmachine_container_runtime role pattern NVIDIA#3663 already proved out for the driver-podman suite. - CI.md and tests/artifacts.nix both say plainly that closing NVIDIA#4163 means running the whole archive rootful, not just the one target that needs it -- nextest archive filters select whole binaries, so there's no cheaper way to isolate just podman_resource_limits. Signed-off-by: politerealism <[email protected]>
1 parent 20cae22 commit a8842ae

2 files changed

Lines changed: 32 additions & 0 deletions

File tree

‎CI.md‎

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -132,6 +132,13 @@ naming any selected manual-only benchmarks separately, so a narrow run cannot
132132
be mistaken for full Podman coverage. See `nix run .#podman-e2e-coverage-summary`
133133
in TESTING.md.
134134

135+
The `driver-podman` testsuite runs both rootful and rootless Fedora Podman;
136+
`e2e-podman` runs rootless only for now. See the comment on `podmanE2eArchive`
137+
in `tests/artifacts.nix` for what has been checked about rootful coverage.
138+
Nextest archive filters select whole binaries, so adding a rootful leg
139+
(tracked in #4163) means running this entire archive rootful, not just one
140+
target.
141+
135142
Three opt-in labels enable the long-running E2E suites:
136143

137144
- `test:e2e` runs the Docker, rootless Podman, Kubernetes, and VM E2E suites

‎tests/artifacts.nix‎

Lines changed: 25 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -256,6 +256,31 @@ let
256256
target = muslToolchain.target;
257257
output = "artifacts/test-archives/${muslToolchain.target}/openshell-podman-tests.tar";
258258
};
259+
# Runs rootless-only in branch CI today (see .github/workflows/branch-e2e.yml's
260+
# driver-specific-integration matrix), unlike podmanDriverArchive above, which
261+
# runs both rootful and rootless. A rootful leg for this archive is tracked in
262+
# #4163. What follows records what has been checked so far, not a claim that
263+
# rootful coverage is unnecessary.
264+
#
265+
# podman_host_gateway was checked specifically (host.openshell.internal
266+
# resolution plausibly depends on the rootless pasta/slirp4netns vs. rootful
267+
# netavark/CNI network backend) and confirmed NOT privilege-sensitive: the
268+
# outer container's /etc/hosts entry uses Podman's own native "host-gateway"
269+
# alias value (container.rs), which Podman resolves identically in both
270+
# modes, and the supervisor-side resolution defaults to a hardcoded
271+
# 127.0.0.1 on Linux regardless of rootful/rootless
272+
# (PodmanComputeConfig::resolved_host_gateway_ip in config.rs). No rootful
273+
# coverage gap here.
274+
#
275+
# The one target confirmed so far to be privilege-sensitive is
276+
# podman_resource_limits, which reads real
277+
# cgroup v2 state from inside the sandbox: rootless cgroup delegation
278+
# depends on systemd-user `Delegate=` and can silently no-op if
279+
# misconfigured, making rootless the harder, higher-risk case -- already
280+
# covered here. Rootful is the lower-risk, currently-unverified path for
281+
# that one target. Tracked in #4163; since nextest archive filters select
282+
# whole binaries, closing that gap means running this entire archive
283+
# rootful too, not just the one target that needs it.
259284
podmanE2eArchive = mkTestArchive {
260285
name = "podman-e2e";
261286
workspacePath = "e2e/rust";

0 commit comments

Comments
 (0)