I'm preparing to PR a succession of privs changes with the ultimate goal
of severely limiting the scope of the binary oper/user dichotomy and
move conceptually distinct oper functions into their own privs.
Accomplishing this is a non-trivial task, and can wait, but it's
inconvenient now to have such functions enabled by the same mechanism
that grants any privs at all--so I'm moving all of them to a
transitional priv with the intention of eroding that later.
rejectcache entries can now use either a K-line aconf or a static
string as a reason. This will be sent in a 465 numeric before the usual
ERROR. In the case of K-lines, it resembles the 465 you would have been
sent without being rejected:
; nc -s 127.6.6.6 127.0.0.1 5000
:staberinde.local 465 * :You are banned from this server- Temporary
K-line 4320 min. - abc123 (2019/12/31 01.07)
ERROR :Closing Link: (*** Banned (cache))
; nc -s 127.128.0.0 127.0.0.1 5000
:staberinde.local 465 * :You are not authorised to use this server.
ERROR :Closing Link: (*** Banned (cache))
/modrestart used to be implemented as a normal command and could crash
when used remotely because it would reload m_encap, which was on the
call stack at the time. This was fixed in 41390bfe5f. However,
/modreload has exactly the same problem, so I'm giving it the
same treatment.
Incidentally: This bug was first discovered in ircd-seven, where the
`/mod*` commands themselves live in the core, so m_encap was the only way
the crash could happen (and it didn't most of the time, because m_encap
would only be moved if you got unlucky). But `/mod*` are in modules in
charybdis, so /modrestart would have unloaded the code it was in the
middle of executing. With that in mind, I'm not sure how it ever
appeared to work.
Move opername and privset storage to struct User, so it can exist for
remote opers.
On /oper and when bursting opers, send:
:foo OPER opername privset
which sets foo's opername and privset. The contents of the privset on
remote servers come from the remote server's config, so the potential
for confusion exists if these do not match.
If an oper's privset does not exist on a server that sees it, it will
complain, but create a placeholder privset. If the privset is created by
a rehash, this will be reflected properly.
/privs is udpated to take an optional argument, the server to query, and
is now local by default:
/privs [[nick_or_server] nick]
As it stands, oper hiding is rather messy and inconsistent. Add
SeesOper(target, source), which is true iff target should appear as an
oper to source. If I haven't missed something, all commands that reveal
oper status now use the same logic.
general::hide_opers_in_whois is a special case, and affects /whois only.
general::hide_opers is introduced, and has the same effect as giving
everyone oper:hidden. All commands that reveal oper status respect both.
Reloading modules sends CAP DEL followed by an immediate CAP NEW:
:staberinde.local CAP * DEL :account-tag
:staberinde.local CAP * NEW :account-tag
This isn't very nice. /modrestart is particularly bad. In order to avoid
doing this, we remember the capability set at the beginning of module
operations, compare that with the set afterwards, and report only the
differences with CAP {DEL,NEW}.
When a server disconnects the client_exit hook will only be called once
but there could be multiple servers and clients behind that server.
After any client exits, check if the agent is still present.
This only supports two addresses as the intended use is 1 IPv4 and 1 IPv6
address on a single-homed host, and the only supported configuration of
outgoing connections to other servers is to bind a single IPv4 or IPv6
address.
Build the same message but send it to the local client first,
so that the echo-message capability works. But don't do it when
sending a message to yourself.
As well as leaking a connid and leaving the connection open,
these calls to free_client() leave the client in the unknown_list
causing check_unknowns_list() to crash when either ptr->data
(ptr being the freed client_p->localClient->tnode) is NULL or
when client_p->localClient is NULL.
Flag the client as an IO error so that we don't try to send it
any data (as this is not a normal plaintext connection).
Links that are 'no-export' are not distributed to the rest of the IRC network (including local peers).
This provides a core primitive for 'anycasting' services (but the actual issue of synchronizing data in
a services package is left to the authors of the services package).
It's useful to allow authd to run in parallel with ssl negotiation,
but if the ssld connection has plaintext data ready for reading
there's a race condition between authd calling read_packet() and
ssl_process_certfp() storing the certificate fingerprint. This
scenario would be bad for a server connecting because fingerprint
verification will fail.
Allow either operation to complete first, but wait until
ssl_process_open_fd() calls the ssl open callback before calling
read_packet().
Don't use the librb callback type as we're always passing client_p.
Provide a return value so that the connect handler can exit_client()
and the accept handler can opt to use the default dead handler.
There's no need to pass information around that sslproc already has access
to, so use ServerInfo directly. Remove the extra NULL checks as these are
already performed before setting ircd_ssl_ok = true.
Fix the server connection configuration so that it can simultaneously
handle a hostname/IPv4/IPv6 for connecting and a hostname/IPv4/IPv6
for binding. Maintains backwards compatibility for matching a hostname
with a mask.
Multiple host/vhost entries can be specified and the last value for
each address family is stored. Hostnames that resolve automatically
overwrite the IP address.
Server connections can now be made to either IPv4 or IPv6 at random
as well as preferring a specific address family.
This also lays the groundwork for the netjoin batch type, but that isn't
implemented yet. I don't like how some of this is implemented but it'll
have to do for now...
Compile tested, needs more testing.
It's a bit of a hack, but better than before. Rather than rehashing
(which could get us into an endless loop), we now segregate the
configuration phase (creating entries ircd-side in case we restart authd
later) and sending phases (when configure_authd() is called). Since we
have to call configure_authd() no matter what (to send timeouts etc.)
and we have to send this data to configure authd anyway, and sending
duplicate data is bad, this is the only way I can think of for now.
