fix(backup): gate every restore category on the permission owning its rows (#2656)

settings was gated on settings:update because a restore rewrites rows
    PUT /api/v1/settings/ owns. The same argument applies to the other three
    categories, and gating one but not the rest is the only state that is
    not defensible: a role holding Backup alone could still write spools,
    archives and K-profiles through a restore that it cannot write through
    the endpoints that own them.

    Each category now also requires that endpoint's write permission -
    inventory:update, archives:update_all and kprofiles:update. archives
    takes update_all rather than create because a restore writes rows owned
    by other users, which is exactly what update_all means.

    All missing permissions are reported in one refusal: a restore is a
    multi-select, so naming them one at a time turns picking four categories
    into four round trips.
This commit is contained in:
maziggy
2026-08-15 14:25:46 +02:00
2 changed files with 139 additions and 24 deletions
+44 -12
View File
@@ -51,6 +51,33 @@ _UNKNOWN_VISIBILITY_ERROR = (
"repo API."
)
# The permission that owns each category's rows, required on top of
# github:restore. Backup is its own permission group, so without this a role
# holding only Backup writes — via a restore — rows it cannot write through the
# endpoint that owns them.
#
# Each entry is the permission that endpoint actually gates its writes on:
#
# * SETTINGS → PUT /api/v1/settings/ (settings:update)
# * SPOOLS → POST/PATCH /api/v1/inventory/spools (inventory:update). Spool
# rows and their usage history both restore under this category.
# * ARCHIVES → archives:update_all, not archives:create. A restore writes
# rows owned by other users — that is the whole point of carrying
# created_by_id — and update_all is the permission that means "may write an
# archive that is not yours". create alone would let an operator with
# archives:create_own-shaped access seed history onto someone else.
# * KPROFILES → POST /api/v1/printers/{id}/kprofiles (kprofiles:update),
# which is what the restore ultimately calls through set_kprofiles_batch.
#
# Cloud profiles are absent because they are not a restorable category
# (RestoreCategory's docstring).
_CATEGORY_WRITE_PERMISSION = {
RestoreCategory.SETTINGS: Permission.SETTINGS_UPDATE,
RestoreCategory.SPOOLS: Permission.INVENTORY_UPDATE,
RestoreCategory.ARCHIVES: Permission.ARCHIVES_UPDATE_ALL,
RestoreCategory.KPROFILES: Permission.KPROFILES_UPDATE,
}
async def _enforce_private_repo(repo_url: str, token: str, provider: str) -> None:
"""Run a test_connection and refuse if the repo is not confirmed private.
@@ -449,25 +476,30 @@ async def restore_backup(
check exists to stop credentials leaving the instance, and this path only
reads. A config can only be saved against a private repo anyway.
The settings category needs ``settings:update`` as well — see the check
below.
Every category needs the permission that owns the rows it writes, on top of
``github:restore`` — see ``_CATEGORY_WRITE_PERMISSION`` and the check below.
"""
if RestoreCategory.SETTINGS in request.categories and current_user is not None:
# The settings category rewrites arbitrary non-auth Settings rows, which
# is what PUT /api/v1/settings/ gates on settings:update. Backup and
# Settings are separate permission groups, so a role holding only Backup
# could otherwise change settings it cannot change through the endpoint
# that owns them. This module already makes that argument — it is why
# the four protected auth keys are refused outright — so the gap was an
# inconsistency in ours, not a new policy.
if current_user is not None:
# Each category rewrites rows some other endpoint already owns, and
# Backup is its own permission group — so a role holding only Backup
# could otherwise write, through a restore, what it cannot write through
# the endpoint that owns them. This module already makes that argument;
# it is why the four protected auth keys are refused outright.
#
# current_user is None only when auth is disabled: github:restore is in
# _APIKEY_DENIED_PERMISSIONS, so an API key never gets past the
# dependency to reach this line.
if not current_user.has_all_permissions(Permission.SETTINGS_UPDATE.value):
missing = sorted(
{
permission.value
for category, permission in _CATEGORY_WRITE_PERMISSION.items()
if category in request.categories and not current_user.has_all_permissions(permission.value)
}
)
if missing:
raise HTTPException(
status_code=403,
detail=f"Missing required permissions: {Permission.SETTINGS_UPDATE.value}",
detail=f"Missing required permissions: {', '.join(missing)}",
)
result = await db.execute(select(GitHubBackupConfig).limit(1))
@@ -474,15 +474,20 @@ class TestRestoreDoesNotOpenTheMetricsEndpoint:
class TestSettingsRestoreNeedsSettingsUpdate(TestOwnershipPermissionsSetup):
"""A Backup-only role must not reach around the gate that owns settings (#2656).
"""A Backup-only role must not reach around the gate that owns the rows (#2656).
The settings category rewrites arbitrary non-auth ``Settings`` rows, which is
exactly what ``PUT /api/v1/settings/`` gates on ``settings:update``. Backup
and Settings are separate permission groups, so gating the restore endpoint
on ``github:restore`` alone let a role holding only Backup change settings it
could not change through the endpoint that owns them. This module already
makes that argument — it is why the four protected auth keys are refused
outright — so the gap was an inconsistency in ours.
Each category rewrites rows some other endpoint already owns —
``PUT /api/v1/settings/`` gates on ``settings:update``, the inventory writes
on ``inventory:update``, an archive that is not yours on
``archives:update_all``, and the K-profile batch on ``kprofiles:update``.
Backup is its own permission group, so gating the restore endpoint on
``github:restore`` alone let a role holding only Backup write, through a
restore, what it could not write through the endpoint that owns them. This
module already makes that argument — it is why the four protected auth keys
are refused outright — so the gap was an inconsistency in ours.
Settings was gated first; the other three followed on review, because gating
one and not the rest is the only state that is not defensible.
"""
async def _token_for(self, async_client: AsyncClient, admin_token: str, name: str, permissions: list[str]) -> str:
@@ -530,10 +535,47 @@ class TestSettingsRestoreNeedsSettingsUpdate(TestOwnershipPermissionsSetup):
@pytest.mark.asyncio
@pytest.mark.integration
async def test_the_same_role_can_still_restore_the_other_categories(self, async_client: AsyncClient, auth_setup):
"""Control: the gate is per-category, not a blanket demotion of github:restore."""
@pytest.mark.parametrize(
("category", "permission"),
[
("spools", "inventory:update"),
("archives", "archives:update_all"),
("kprofiles", "kprofiles:update"),
],
)
async def test_backup_only_role_cannot_restore_the_other_categories(
self, async_client: AsyncClient, auth_setup, category, permission
):
"""Same argument as settings: these rows have an owning permission too."""
token = await self._token_for(
async_client, auth_setup["admin_token"], "backuponly2", ["github:backup", "github:restore"]
async_client, auth_setup["admin_token"], f"backuponly-{category}", ["github:backup", "github:restore"]
)
await _create_config(async_client, auth_setup["admin_token"])
with patch(
"backend.app.services.github_restore.github_restore_service.run_restore",
new=AsyncMock(return_value={"success": True, "message": "", "log_id": 1, "ref": "aaa1111", "results": {}}),
) as mock:
response = await async_client.post(
"/api/v1/github-backup/restore",
headers={"Authorization": f"Bearer {token}"},
json={"categories": [category]},
)
assert response.status_code == 403
assert permission in response.json()["detail"]
mock.assert_not_awaited(), "the refusal has to happen before anything is written"
@pytest.mark.asyncio
@pytest.mark.integration
async def test_every_missing_permission_is_named_at_once(self, async_client: AsyncClient, auth_setup):
"""One round trip tells the caller everything to fix, not just the first.
A restore is a multi-select, so reporting one category at a time turns
picking four into four refusals.
"""
token = await self._token_for(
async_client, auth_setup["admin_token"], "backuponly-all", ["github:backup", "github:restore"]
)
await _create_config(async_client, auth_setup["admin_token"])
@@ -544,11 +586,52 @@ class TestSettingsRestoreNeedsSettingsUpdate(TestOwnershipPermissionsSetup):
response = await async_client.post(
"/api/v1/github-backup/restore",
headers={"Authorization": f"Bearer {token}"},
json={"categories": ["spools", "archives", "kprofiles"]},
json={"categories": ["settings", "spools", "archives", "kprofiles"]},
)
assert response.status_code == 403
detail = response.json()["detail"]
for permission in ("settings:update", "inventory:update", "archives:update_all", "kprofiles:update"):
assert permission in detail
@pytest.mark.asyncio
@pytest.mark.integration
async def test_the_gate_is_per_category_not_a_blanket_demotion(self, async_client: AsyncClient, auth_setup):
"""Control: holding one category's permission is enough to restore that one."""
token = await self._token_for(
async_client,
auth_setup["admin_token"],
"backupandinventory",
["github:backup", "github:restore", "inventory:read", "inventory:update"],
)
await _create_config(async_client, auth_setup["admin_token"])
with patch(
"backend.app.services.github_restore.github_restore_service.run_restore",
new=AsyncMock(return_value={"success": True, "message": "", "log_id": 1, "ref": "aaa1111", "results": {}}),
):
response = await async_client.post(
"/api/v1/github-backup/restore",
headers={"Authorization": f"Bearer {token}"},
json={"categories": ["spools"]},
)
assert response.status_code == 200
@pytest.mark.asyncio
@pytest.mark.integration
async def test_every_restorable_category_has_an_owning_permission(self):
"""Guards the map against a category added without a gate.
A new ``RestoreCategory`` that is missing here is not a failing test
anywhere else — it simply restores under ``github:restore`` alone, which
is the hole this whole class exists to close.
"""
from backend.app.api.routes.github_backup import _CATEGORY_WRITE_PERMISSION
from backend.app.schemas.github_backup import RestoreCategory
assert set(_CATEGORY_WRITE_PERMISSION) == set(RestoreCategory)
@pytest.mark.asyncio
@pytest.mark.integration
async def test_a_role_holding_both_can_restore_settings(self, async_client: AsyncClient, auth_setup):