Deferred TLS initialization is used with Mozilla NSS. The real
initialization takes place when the TLS context is needed for the first
time. If the initialization parameters were freed immediately after
tlsm_ctx_init was called, they were not available at the time of
deferred initialization which caused segmentation fault.
With this patch, initialization parameters are copied and stored until
the deferred initialization is finished. The parameters are freed
afterwards.
Red Hat Bugzilla: #783431
PEM nss is not thread safe when establishing the initial connection
using SSL_ForceHandshake. Create a new mutex - tlsm_pem_mutex - to
protect this function call.
The call to SSL_ConfigServerSessionIDCache() is not thread-safe - move it
to the init section and protect it with the init mutex.
Unfortunately automated checkers don't seem to read the documentation
for how APIs are expected to be used, and the C declaration syntax
isn't expressive enough to encode the documented usage.
The NSS_InitContext et. al, and their corresponding shutdown functions,
are not thread safe. There can only be one thread at a time calling
these functions. Protect the calls with a mutex. Create the mutex
using a PR_CallOnce to ensure that the mutex is only created once and
not used before created. Move the registration of the nss shutdown
callback to also use a PR_CallOnce. Removed the call to
SSL_ClearSessionCache() because it is always called at shutdown, and we must
not call it more than once.
If server certificate hostname does not match the server hostname,
connection is closed even if client has set TLS_REQCERT to 'allow'. This
is wrong - the documentation says, that bad certificates are being
ignored when TLS_REQCERT is set to 'allow'.
If the olcTLSVerifyClient is set to a value other than "never", the server
should request that the client send a client certificate for possible use
with client cert auth (e.g. SASL/EXTERNAL).
If set to "allow", if the client sends a cert, and there are problems with
it, the server will warn about problems, but will allow the SSL session to
proceed without a client cert.
If set to "try", if the client sends a cert, and there are problems with
it, the server will warn about those problems, and shutdown the SSL session.
If set to "demand" or "hard", the client must send a cert, and the server
will shutdown the SSL session if there are problems.
I added a new member of the tlsm context structure - tc_warn_only - if this
is set, tlsm_verify_cert will only warn about errors, and only if TRACE
level debug is set. This allows the server to warn but allow bad certs
if "allow" is set, and warn and fail if "try" is set.
If tlsm_find_and_verify_cert_key finds the cert and/or key, and it fails
to verify them, it will leave them allocated for the caller to dispose of.
There were a couple of places that were not disposing of the cert and key
upon error.
When server certificate is not required in a TLS session (e.g.
TLS_REQCERT is set to 'never'), ignore expired issuer certificate error
and do not terminate the connection.
In tlsm_auth_cert_handler, we get the peer's cert from the socket using
SSL_PeerCertificate. This value is allocated and/or cached. We must
destroy it using CERT_DestroyCertificate.
OpenLDAP built with OpenSSL allows most any value of cacertdir - directory
is a file, directory does not contain any CA certs, directory does not
exist - users expect if they specify TLS_REQCERT=never, no matter what
the TLS_CACERTDIR setting is, TLS/SSL will just work.
TLS_CACERT, on the other hand, is a hard error. Even if TLS_REQCERT=never,
if TLS_CACERT is specified and is not a valid CA cert file, TLS/SSL will
fail. This patch makes CACERT errors hard errors, and makes CACERTDIR
errors "soft" errors. The code checks CACERT first and, even though
the function will return an error, checks CACERTDIR anyway so that if the
user sets TRACE mode they will get CACERTDIR processing messages.
