While functionally compatible with the implementation in ElementalIRCd, our approach is different,
specifically pre-calculating the bitmask at config load time. This is more efficient, and allows us
to report errors as part of the configuration phase.
- Implemented changes suggested by Jilles
- Remove some unused parameters in functions
- Remove some unused ssl procs
- 63-bit time_t support in TS deltas
- const char * vs char * cleanup
- struct alignment (void *) casts
- signed vs unsigned fixes
- bad memset() call
- Bad LT_MAIN in libratbox
- char -> unsigned char casts for isdigit/isspace/etc calls
Thanks Jilles!
This allows multiple improvements to m_sasl. With this change, the SASL
authentication gets aborted immediately when services are offline.
Additionally, we send the SASL ENCAP messages directly to the specified
SASL agent.
When receiving bans from a bursting server, if kline_delay is set to 0
(the default), rb_event_addonce will be called to schedule an event for
0 seconds in the future. While this works fine for the fallback
rb_event_run function, the epoll implementation ends up scheduling a
timerfd for the event in the past, which is then never executed.
While fixing this, I also made rb_event_add and rb_event_addonce reject
attempts to add events scheduled for 0 seconds in the future; they're
instead rewritten to run 1 second in the future.
This change prevents the log file paths from being leaked when
rehashing. Additionally, fname_killlog was added to two places where it
was previously forgotten.
Add the flags (auth{} spoof, dynamic spoof) to struct Whowas and add a
show_ip_whowas().
Normal users now see IPs of unspoofed users, and remote opers can see IPs
behind dynamic spoofs. Also, general::hide_spoof_ips is now applied when
the IP is shown, not when the client exits.
s_assert requires some higher-level functionality that shouldn't be
present in ircd_defs.h. ircd_defs.h is used by ssld, which has no notion
of logging or sending IRC messages. Additionally, some of the headers
s_assert depends on result in conflicting definitions in ssld.c.
This change also fixes the compile when using --enable-assert=soft.
This was broken by 6f7b36d5d0 in February
2013, as join failures are the only situation where a non-trivial
numeric is passed through from other code to be sent to a client. Fix it
by porting more code from ircd-ratbox 3.1.
With this comes an example module to block the killing of services.
NOTE: this will not cancel remote kills. Those are still accepted, per
the TS 6 specification.
Altering localClient->F without updating the FD hash leaves the struct
Client in the FD hash indefinitely which causes a crash later if the
struct is reused for a remote client. It also prevents error messages
from ssld showing up on IRC properly.
This will allow us to modularize message processing, e.g. having new modules to manipulate
channel and private messages in new ways.
Yes: it can be used to intercept messages, but such modules are already out in the wild for
charybdis anyway -- so this doesn't really change anything there.
If you are changing the text, then it is your responsibility to provide a pointer to a new
buffer. This buffer should be statically allocated and stored in your module's BSS segment.
We will not, and cannot, free your buffer in core, so dynamically allocated buffers will
cause a memory leak.
This will allow us to simplify m_message considerably, by moving channel mode logic out to
their own modules.
Arbitrarily prefer a forward channel to no forward channel and an
alphabetically higher forward channel to a lower one.
This is a simplistic implementation that generates one MODE message to
local clients for each ban removed (to be replaced).
For simplicity and to avoid amplification of incoming MODE messages,
regular modes may still desync the forward channel of a ban.
Add two mechanism for avoiding name-collisions in a system-wide
installation of charybdis. The ssld and bandb daemons, intended to be
directly used by ircd and not the user, install into libexec when
--enable-fhs-paths is set. For binaries which are meant to be in PATH
(bindir), such as ircd and viconf, there is now an option
--with-program-prefix=progprefix inspired by automake. If the user
specifies --with-program-prefix=charybdis, the ircd binary is named
charybdisircd when installed.
Add support for saving the pidfile to a rundir and storing the ban
database in localstatedir instead of in sysconfdir. This is, again,
conditional on --enable-fhs-paths.
Fix(?) genssl.sh to always write created SSL key/certificate/dh
parameters to the sysconfdir specified during ./configure. The
previous behavior was to assume that the user ran genssl.sh after
ensuring that his current working directory was either sysconfdir or a
sibling directory of sysconfdir.
This becomes important because of away-notify sending aways to common
channels much like nick changes (which are also paced).
Marking as unaway is not limited (but obviously only does something if the
user was away before). To allow users to fix typos in away messages, two
aways are allowed in sequence if away has not been used recently.
If a netburst adds bans or a channel TS change removes bans, +b/-b modes
are generated. Make sure these contain the forward channel, if any.
This appears also broken in ircd-seven.
This adds a new ISUPPORT token, NICKLEN_USABLE which is strictly an informative value.
NICKLEN is always the maximum runtime NICKLEN supported by the IRCd, as other servers may
have their own usable NICKLEN settings. As NICKLEN_USABLE is strictly informative, and
NICKLEN is always the maximum possible NICKLEN, any clients which depend on NICKLEN for
memory preallocation will be unaffected by runtime changes to NICKLEN_USABLE.
The default NICKLEN is 50; the default serverinfo::nicklen in the config file is set to
30, which is the NICKLEN presently used on StaticBox.
They are now in messages, even if client_flood_message_time is not 1.
If client_flood_message_time is not 1 (by default it is), this needs a
configuration change to maintain the same behaviour.
* Deduce allow_read from the client's state (IsFloodDone) rather than
storing it in LocalUser.
* Fix the documentation (in oper /info), however strange
client_flood_burst_rate and client_flood_burst_max may seem, that is
how they currently work.
nenolod gave the thumbs-up to port ircd-seven banfowards to charybdis to spb
for a while, and people have asked about it. Might as well do it since it's a
slow weekend.
Note that as a side effect use_forward is removed from the config and
unconditionally enabled!
While what chanroles are trying to accomplish is a good idea, it is
apparently unclear this is the proper way to do it. Until we figure out
the exact way we wish to do this, it should be reverted for now.
As jilles pointed out, it is best that the chanserv access list always
remain synced with the grant list. Thus, the ability for clients to set
this is not a good idea unless services knows about the grant, but this
leads to all sorts of messy issues and likely isn't worth it.
this makes setting new roles on a user much easier as we're just setting the roles they
should be having, instead of having to try to revoke roles we don't necessarily know
about.
The theory behind this is that services sends an ENCAP * GRANT #channel
UID :+flagspec message specifying the chanroles the user has. They are
mapped into flag bits and applied to the membership of the user. They
then are restricted or permitted to what they can do based on the
permissions mask regardless of rank.
For backwards compatibility, the default permission bit (without a GRANT
statement) allows a user to to anything an existing op can do ONLY if
they are an op.
Todo: make CHANROLE_STATUS work (the ability to apply +ov to people),
which is at the moment controlled by CHANROLE_MODE.
When we broadcast a KILL message, this generates server notices on all
other servers (assuming the target user exists). Therefore, we should
also send a notice to our local opers.
Do kick_on_split_riding if services sends an SJOIN
with a lower TS and a different key. This relies on
services restoring TS (changets option in atheme) and
services not immediately parting after receiving the
KICK, which is the case in recent atheme.
For invite-only channels, still only do
kick_on_split_riding in netbursts. Services is
assumed to handle this itself (atheme does).
Any hunted parameter with wildcards is now assumed
to be a server, never a user.
Reasons:
* fewer match() calls
* do not disclose existing nicknames
* more intuitive behaviour for CONNECT
m_trace has a copy of some hunt_server logic in it
(for the RPL_TRACELINK reply), so adjust that too.
The extended-join client capability extends the JOIN message with information clients typically
query using WHO including accountname, signon TS and realname.
The reason why we do this is because some clients are dependent on receiving a numeric
for every channel join failure, even due to this limit where it can be assumed that
subsequent joins failed.
This has a separate enabling option channel::channel_target_change.
It applies to PRIVMSG, NOTICE and TOPIC by unvoiced unopped non-opers.
The same slots are used for channels and users.
The code behind this capability was never implemented, and subsequent
discussions have agreed to approach the problem differently. There seems no
reason to continue advertising a capability that does nothing.
* does not apply to NOTICE (as those may well be automated)
* mirrors +g behaviour so that no useless accept entries are added for services
* respects max_accept, if it would be exceeded the message is dropped with numeric 494
* check moved up so this is checked before floodcount/tgchange
This shouldn't provide any way for a client to get on a CALLERID list
without authorization, as if a client is +g already, a CTCP request, for
example, won't be replied to.
(resv, cmode +m, cmode +b, cmode +q, etc.).
This is only checked for local users.
For optimal compatibility, a failure for this reason still
returns ERR_CHANOPRIVSNEEDED.
Side effect: normal users cannot change topics of resv'ed
channels, even if they have ops, just like they already
cannot send messages. This only matters if resv_forcepart
is disabled, as the user would have been removed from the
channel otherwise.
If this option is yes (default), KLINE by itself sets global (propagated) bans.
If this option is no, KLINE by itself sets a local kline following cluster{},
compatible with 3.2 and older versions.
Special modes like +j can be tracked easily just by adding the necessary
code to parse them to set_channel_mlock(). This will cover propagation
as well.
Such bans are not applied locally, but are propagated normally.
They can only be removed on a server that applies them.
Note that normally KLINE will not accept such bans.
This is mainly for services, differing min_wildcard and
ircd changes.
A KLINE command without the ON clause now sets a propagated
("global") ban. KLINE commands with the ON clause work as
before.
Propagated klines can only be removed with an UNKLINE command
without the ON clause, and this removes them everywhere.
In fact, they remain in a deactivated state until the latest
expiry ever used for the mask has passed.
Propagated klines are part of the netburst using a new BAN
message and capab. If such a burst has an effect, both the
server name and the original oper are shown in the server
notice.
No checks whatsoever are done on bursted klines at this time.
The system should be extended to XLINE and RESV later.
There is currently no way to list propagated klines,
but TESTLINE works normally.
The value 0 indicates the creation time is unknown (currently the case
for bandb).
Also store a creation time for xlines and resvs, but do not use it yet.
identd and SASL can cause source_p->user to be present without USER having been sent.
Without this change, that could cause a crash later on as localClient->fullcaps is not initialised.
The behaviour is the same as /msg except that where
/msg would send RPL_UMODEGMSG to the user, the /invite
is instead let through. This counts as a notification
for caller_id_wait like RPL_UMODEGMSG.
Checks are on the target user's server, which means an
error message will appear after RPL_INVITING.
This must be because the accept list is not globally
known.
Similar to /msg, inviting a user that is not in a channel
you have op or voice in requires a free target; opers always
have a free target.
Being invited adds the source as a reply target.
When a user receives a private message, notice or RPL_UMODEGMSG,
add the source to a special set of 5 target slots.
These slots are checked in the normal way when sending messages,
allowing a reply without using up a free target.
This feature will not be very useful if a user is being messaged
by many different users; to help this, messages blocked entirely
by +g or +R do not affect the targets. CTCP replies also remain
free in terms of targets.
The server protocol for this is
:<uid> ENCAP * CERTFP :<40 hex chars>
both in new user introductions and in burst.
As in oftc-hybrid, only the user themselves and opers can see the certfp.
Displaying the certfp on connect seems unnecessary to me,
the user can whois themselves if needed.
A notice will be sent to any force parted users that the channel
is temporarily/permanently unavailable on the server.
A new config option channel::resv_forcepart can be used to disable this.
from ircd-ratbox (dubkat)
Allowing this desynchronizes clients and servers that
do not know about this extension without good reason.
If you want to allow this, change it to spit out the
usual AWAY protocol.
A large group is any $$ or $# or a channel with more than
floodcount/2 local members, checked on each server separately.
Note that floodcount checks are done on the sender's server.
The special treatment is active for 15 seconds.
The change also applies to xlines sent to other servers.
Note that the backslash must be supplied in an unxline.
Incoming xlines containing '",' are rejected to avoid
inconsistency between servers.
Spaces in xline.conf files are changed to \s in memory
and recognized on unxline, but are not changed in the
file. New xlines are added to xline.conf files with \s.
xline.conf written by this commit or newer will not work
properly with charybdis older than 1124:131254925f32 (which
introduced \s support in match_esc()).
(provided channelTS != 0).
ETB is controlled by the EOPMOD capab. Fairly complicated
stuff is needed for older servers.
ETB can come from clients or servers, but at this time it
should only be sent from clients, as otherwise it is not
always possible to update older servers.
to restrict channel names to printable ascii only.
Like disable_fake_channels this only applies to joins
by local users; unlike disable_fake_channels it applies
to opers as well.
Note that in some cases (different TS delta settings,
heavy lag) it is possible only one of the servers
detects the problem and autoconnect may be left enabled.
The mechanism used for disabling is the same as
/quote set autoconn <server> 0; the A flag in /stats c
disappears and a rehash or /quote set autoconn can put
it back.
Author: Stephen Bennett <spb@exherbo.org>
Date: Sun Jan 18 16:30:27 2009 +0000
Make testline show oper reason and setter information for K:lines, where applicable
This should currently be impossible, but may be
possible with extremely long server/nick names and
additional fields.
In case of overflow, the string sent to to the user is
truncated and if it was the first overflow since the
module was loaded a notice is sent to snomask +d.
This gives a useful meaning to the cmode combo +mz-n:
messages from ops and voices go to all channel members,
messages from anyone else (on or off channel) go to ops.
With +mnz, messages from outside are not allowed at all.
