Skip to content

Restrict daemon reading the pull secret from a merged secret (manual backport 278) - #307

Merged
HarshwardhanPatil07 merged 8 commits into
bootc-dev:release-0.2from
alicefr:backport-278
Nov 13, 2026
Merged

HarshwardhanPatil07 merged 8 commits into
bootc-dev:release-0.2from
alicefr:backport-278

Conversation

@alicefr

@alicefr alicefr commented Nov 12, 2026 •

Copy link
Copy Markdown
Collaborator

Manual backport of #278 . The backport required few tweaks because we don't have the bootcoperatorconfig in the release-0.2 branch, but almost stayed the same

The tag resolver looked up credentials with an exact match of the
registry host in the auths map. Entries written by kubectl create secret
docker-registry and docker login use other forms, most commonly
https://index.docker.io/v1/ for Docker Hub. Those missed, the resolver
silently fell back to anonymous access and tag resolution failed with
401, while the node's container runtime accepted the same secret.

Compare keys the way containers/image does for auth files: strip the
http(s):// scheme, drop the path when a scheme was present, and map
docker.io and registry-1.docker.io to index.docker.io. An exact match
still wins, and normalized keys are tried in sorted order so the result
is deterministic. Scoped keys without a scheme, such as quay.io/org, are
still not matched; the resolver looks up registries only.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit a118d6c)
The daemon will read pull secrets from a single Secret maintained by the
controller and mounted into the DaemonSet, with one key per source
secret. The new internal/pullsecret package holds the Secret name, the
daemon mount directory and the key function, "<namespace>_<name>", so
the controller and the daemon cannot drift.

A namespace and a Secret name together can exceed the 253-character
Secret key limit, and one invalid key would make the API server reject
the whole aggregated Secret. Key returns an error in that case, and
fetchPullSecretData reports it on the pool as SecretError.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit dd78c7e)
The pool reconciler required a .dockerconfigjson key and passed its
bytes unchanged to the tag resolver and, via the aggregated Secret, to
bootc. Both expect the {"auths": ...} wrapper. A legacy .dockercfg
Secret, which holds the bare auths map and which OpenShift generates for
service accounts, was rejected as missing a key; had it been passed
through, it would have meant silently no credentials.

Add pullsecret.AuthData, which returns the credentials in
.dockerconfigjson format. It goes by the Secret's keys rather than its
type, so an Opaque Secret works too. .dockerconfigjson is returned as
is and wins when both keys are set. .dockercfg is wrapped in
{"auths": ...} without re-encoding its entries, so fields such as
identitytoken and the credential bytes are preserved. Data that is not
a JSON object, including an empty .dockerconfigjson, is reported with
the key name but never with its content. fetchPullSecretData uses it, so
the reconciler normalizes once and the daemon keeps copying bytes
verbatim.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit fe0d577)
The pool reconciler decided which Secret events to act on by type,
accepting only kubernetes.io/dockerconfigjson. That ignored events for
the other forms of pull secret that fetchPullSecretData accepts, such as
legacy .dockercfg and Opaque Secrets carrying a .dockerconfigjson key,
so rotating them was never noticed. Matching on type is also not
meaningful for a Secret that is not referenced by any pool, and the
mapper listed every pool on each Secret event to find out.

Index BootcNodePool by spec.pullSecretRef and map a Secret event to the
pools in that index. A referenced Secret now triggers a reconcile
whatever its type or content, so data that becomes valid or invalid is
always seen, and Secrets no pool references enqueue nothing. The Secret
type is immutable, so type changes need no handling.

The field index is registered once per manager, in main and in the
envtest suite, since a manager rejects duplicate registrations.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit 3bad2ff)
The daemon will stop reading pull secrets through the API and read them
from a Secret mounted into the DaemonSet instead. The new
PullSecretAggregator maintains that Secret, bootc-pull-secrets in the
operator namespace, with one key per source secret referenced by a pool,
holding its credentials normalized to .dockerconfigjson format. It lives
in internal/pullsecret, next to the key function and the constants it
shares with the daemon, so the reconciler and the key format cannot
drift apart.

It is the Secret's only writer, so pool reconciles cannot conflict on
it. Pool, referenced source secret and bootc-pull-secrets events all map
to one request, found through the pool index, and each reconcile
recomputes the data from scratch, so stale keys disappear and a deleted
or edited Secret is repaired. References the pool reconciler reports as
SecretError are skipped. The aggregator also reconciles once at startup,
so the Secret exists whenever the controller runs; it is kept, with no
keys, when no pool references a pull secret.

The controller needs create on secrets in its namespace, and update and
patch on bootc-pull-secrets only. It now requires the operator namespace
to be known, since that is where the Secret lives.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit 359ee69)
The DaemonSet now mounts bootc-pull-secrets, maintained by the
controller, at /var/run/bootc-pull-secrets. The volume is optional, as
the daemon may start before the Secret exists, and uses no subPath, so
the kubelet keeps it up to date.

While an update is pending, before staging, syncPullSecret copies the
entry for the node's pullSecretRef to /run/ostree/auth.json instead of
fetching the Secret through the API. Without a reference or a readable
entry, it now removes the auth file instead of leaving stale credentials
behind, and bootc falls back to /etc or /usr/lib. A missing entry needs
no special handling: the kubelet may take a minute to sync a new or
rotated Secret, and a failed stage is retried with backoff and re-reads
the file.

The daemon tests now use temporary directories for the host root and
the mount, so they never touch the real /run/ostree.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit c8eda7b)
The daemon reads pull secrets from the mounted bootc-pull-secrets
Secret, so it no longer needs to get Secrets through the API. Remove the
rule from its ClusterRole, which allowed every node to read any Secret
in the cluster, and the now unused APIReader.

The kubelet still reads the mounted Secret: the node authorizer allows
a node to read Secrets used by pods scheduled on it.

make helm-test now checks that no ClusterRole, the daemon's included,
grants access to secrets.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit 3c43ec7)
The controller now creates bootc-pull-secrets in the release namespace.
Helm does not know about it, and helm uninstall keeps the namespace
created by --create-namespace, so the Secret is left behind. Document
deleting the namespace to remove it.

Assisted-by: AI
Signed-off-by: Alice Frosi <afrosi@redhat.com>
(cherry picked from commit c436dae)
@HarshwardhanPatil07
HarshwardhanPatil07 merged commit baa1c6d into bootc-dev:release-0.2 Nov 13, 2026
12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants