Enabling at least one of the ssl.verifyext modes enables a callback that
collects all the errors, which are used by mod_s2s to report better
problem descriptions.
Setting ssl.verifyext enables a callback that collects all errors from
every layer of the certificate chain. Otherwise a single string is
returned, which we did not handle before.
Various options in Prosody allow control over the behaviour of the certificate
verification process For example, some deployments choose to allow falling
back to traditional "dialback" authentication (XEP-0220), while others verify
via DANE, hard-coded fingerprints, or other custom plugins.
Implementing this flexibility requires us to override OpenSSL's default
certificate verification, to allow Prosody to verify the certificate itself,
apply custom policies and make decisions based on the outcome.
To enable our custom logic, we have to suppress OpenSSL's default behaviour of
aborting the connection with a TLS alert message. With LuaSec, this can be
achieved by using the verifyext "lsec_continue" flag.
We also need to use the lsec_ignore_purpose flag, because XMPP s2s uses server
certificates as "client" certificates (for mutual TLS verification in outgoing
s2s connections).
Commit 99d2100d2918 moved these settings out of the defaults and into mod_s2s,
because we only really need these changes for s2s, and they should be opt-in,
rather than automatically applied to all TLS services we offer.
That commit was incomplete, because it only added the flags for incoming
direct TLS connections. StartTLS connections are handled by mod_tls, which was
not applying the lsec_* flags. It previously worked because they were already
in the defaults.
This resulted in incoming s2s connections with "invalid" certificates being
aborted early by OpenSSL, even if settings such as `s2s_secure_auth = false`
or DANE were present in the config.
Outgoing s2s connections inherit verify "none" from the defaults, which means
OpenSSL will receive the cert but will not terminate the connection when it is
deemed invalid. This means we don't need lsec_continue there, and we also
don't need lsec_ignore_purpose (because the remote peer is a "server").
Wondering why we can't just use verify "none" for incoming s2s? It's because
in that mode, OpenSSL won't request a certificate from the peer for incoming
connections. Setting verify "peer" is how you ask OpenSSL to request a
certificate from the client, but also what triggers its built-in verification.
In some circumstances, particularly with 'opportunistic_writes' and
'fatal_errors' enabled in the epoll backend, the connection may be closed
halfway through the session close process (because it contains debug logging,
which in the case of the watch:log() command, will trigger a write to the
socket).
The chosen fix is to cache session.conn in a local variable (we already did
this later on, but this pulls it up to the top of the function, which is
generally more correct anyway).
This was an unintentional regression, as per-host 'ssl' options became valid
in 0.12 when SNI support was added for direct TLS ports. While we encourage
most people to use the simpler automatic certificate selection (and it seems
most do, given the overlooking of this bug), there are likely always going to
be use cases for manually-configured certificates.
The issue was introduced in commit 7e9ebdc75ce4 which inadvertently removed
the per-host option checking for SNI.
The ! commands have been broken for some time, and we're not going to
implement them right now. If we want fancier editing, we can now do that on
the client side (with readline and stuff).
Also removes mention of telnet!
This should bring some fixes and general robustness that mod_websocket had
missed out on. The duplicated code here is not at all ideal. To prevent this
happening again, we should figure out how to have the common logic in a single
place, while still being able to do the websocket-specific parts that we need.
The main known bug that this fixes is that it's possible for a session to get
into a non-destroyable state. For example, if we try to session:close() a
hibernating session, then session.conn is nil and the function will simply
return without doing anything. In the mod_c2s code we already handle this, and
just destroy the session. But if a hibernating websocket session is never
resumed or becomes non-resumable, it will become immortal!
By merging the fix from mod_c2s, the session should now be correctly
destroyed.
Although we do sometimes use single-line if blocks, I'm expanding this one to
make it easier to compare with the duplicated (but modified) code in
mod_websocket that we plan to de-duplicate one day.
