Private hives and layers

A service or sandbox can see different registry data from an administrator looking up the same path. Before concluding that the store is inconsistent, check whether the workload uses a private hive or private layer through its effective thread credentials.

This page explains the diagnostic distinction. It does not provide a command for creating private views; reg does not manage their attachment. Use the workload's documented isolation mechanism and the technical references below.

Private hives #

A private hive is visible to threads whose tokens carry its scope identity. Private hives are considered before global hives, so a private Machine can shadow the global Machine for that thread without changing the path.

An operator reading the global path therefore has not necessarily inspected the workload's configuration. Identify which credentials and scope the workload actually uses before comparing values or planning recovery. Do not assume a global-hive backup captures a separate private hive.

Private layers #

A disabled layer can be active for a thread whose credentials name it. It then competes using the normal precedence-and-write-order rules. Disabling it globally does not remove it from those private views.

This is per thread, not necessarily per process: different impersonation tokens can give threads in one process different views. Check the relevant request or service identity rather than only the process's nominal identity.

Typical uses are session-specific overrides, test settings and sandbox configuration. Those uses do not establish who may attach a private view.

Authorisation lives in KACS #

The token mechanism governs attachment. The kernel TRM documents scope GUIDs and private-layer names entering credentials at token creation, gated by SeCreateTokenPrivilege. That is a broad token-creation gate, not per-scope ownership authorization.

With the PEI-268 kernel fix, attaching a layer above precedence zero also requires enabled SeTcbPrivilege on the token's creator. That authorization stays with the attachment, so an ordinary process can use a view deliberately delegated by a trusted service. An attachment made without that authorization cannot gain high-precedence access through a later layer promotion: registry operations using that attachment fail with EPERM while the layer is above zero. Older kernels lack this check.

Use the private-hive and private-layer TRM sections for the credential model and its documented limitations. Configured limits can also reject a registry operation when the token carries too many scopes or private layers; this is not automatically a key-permission problem.

Where to go next #

Edit this page