Add opt-in no_login_cache to release the token login session on key free.

[dantecmelo]

Pull Request Type

• Bug fix
• New feature
• Code style / formatting / renaming
• Refactoring (no functional or API changes)
• Build / CI related changes
• Documentation
• Other (please describe):

Related Issue

Issue number: N/A

Current Behavior

libp11 caches the token login session for the lifetime of the process and would
normally release it at exit. However, its exit-time teardown is deliberately
skipped: the atexit handler in util_uri.c sets g_shutdown_mode, which makes
UTIL_CTX_free_libp11 skip PKCS11_release_all_slots / C_Finalize (to avoid
calling into the module during OpenSSL's teardown). As a result the login session
is never released.

On a token with a hard limit on concurrent authenticated sessions — notably the
YubiHSM 2 (max 16) — running one short-lived openssl process per file leaks one
session per process (on the YubiHSM, C_CloseSession does not free the
authenticated session; only C_Logout / C_Finalize does). After 16 files the
17th fails with CKR_SESSION_COUNT ("could not read private key").

New Behavior

Adds an opt-in no_login_cache provider parameter (env PKCS11_NO_LOGIN_CACHE)
and a matching public PKCS11_set_no_login_cache() API. When enabled, the token
is logged out and its sessions are closed as soon as a private key object is
freed — during normal runtime instead of at the skipped exit. Re-login happens
automatically on the next key load. With the parameter unset, behaviour is
byte-for-byte unchanged.

Scope of Changes

• Core: a self-contained, session-only logout helper
  pkcs11_slot_logout_session_only() (p11_slot.c), a per-context/per-slot
  no_login_cache flag, and the public PKCS11_set_no_login_cache()
  (p11_front.c, libp11.h, libp11.exports).
• Provider front-end: a no_login_cache parameter mirroring force_login
  (provider_helpers.c, util_uri.c, util.h). Because the provider's
  keymgmt_load transfers ownership of the key to OpenSSL, the wrapped key's
  legacy RSA/EC finish callback does not run during the process — so the logout
  is triggered from keymgmt_free (provider.c) via p11_keydata_release_login(),
  which reaches the slot through the wrapped key's RSA/EC ex-data
  (pkcs11_release_login_for_pkey() in p11_key.c).
• Engine front-end: the same release is triggered from the RSA/EC key finish
  callbacks (p11_rsa.c, p11_ec.c), where keys are freed via that path.

Testing

• Existing tests
• New tests added
• Manual testing

Verified on a YubiHSM 2 (max 16 sessions) with the OpenSSL 3 provider on Debian:
a 50-iteration loop of openssl dgst -sha512 -sign, one process per file,
previously failed at the 17th with CKR_SESSION_COUNT; with no_login_cache = 1
it now completes all 50. A PKCS#11 trace shows one C_Logout per key free, and
the device's active session count returns to baseline between runs.

Additional Notes

• Opt-in and backward compatible: with the parameter unset, behaviour is
  byte-for-byte unchanged.
• Safety / re-entrancy: the logout helper is self-contained — it does not call
  pkcs11_wipe_cache or pkcs11_get_session, so it cannot re-enter the key-free
  callback or nest slot->lock; it is idempotent via a lockless logged_in check.
• Key-type scope: covers RSA and ECDSA (EC). EdDSA (Ed25519/Ed448) is out of
  scope — libp11 has no per-key finish callback for it, so there is no equivalent
  hook.
• Trade-off: intended for many short-lived signing processes. A long-lived
  process that signs repeatedly benefits from the cached login, so it should
  leave the option off — hence opt-in.
• Background: the skipped exit-time teardown that makes runtime release necessary
  is the behaviour libp11 adopted to avoid an exit-time crash/deadlock (cf. Library cleanup issues with OpenSSL - libp11 - yubihsm_pkcs11.so chain #527).

License Declaration

• I hereby agree to license my contribution under the project's license.