Both setgid(2) as well as setuid(2) can fail with EINVAL in addition to
EPERM, their manual pages state "EINVAL: The user/group ID specified in
uid/gid is not valid in this user namespace ".
So not only treat EPERM as an "acceptable error" and continue with
logging the error, but do the same for EINVAL.
This was triggered by the Void Linux xbps-uunshare(1) tool used for
building "XBPS source packages" and reported by luca in #ngircd. Thanks!
Periodically "ping" the service manager (every 3 seconds) and set a
status message showing connection statistics.
This enables using the systemd(8) watchdog functionality for the
"ngircd.service" unit.
Wait for the "ERROR :Closing connection" message sent by ngIRCd when
handling the QUIT command, do not wait for "Connection closed" which is
actually output by the telnet(1) command and is implementation
dependant! For example, on Haiku OS, this is not always(!) echoed (the
command seems to hang sometimes?) which results in unpredictable
failures in the test suite ...
- Try to only search for processes of the current user.
- Prefer using pgrep (in addition to pidof) when available.
- Streamline system overrides.
- Get rid of HEAD_FLAGS: all systems so far support "-1".
- Use "ps -o pid,comm" as the default, which is POSIX.1.
- Use "sort -r" to hopefully get the newest (=highest) PID, which is the
case on older systems not using randomized PIDs at least.
Although this is a nice looking solution, it is not that portable: for
example, the "openssl s_client" command of LibreSSL 3.8.2 on OpenBSD 7.4
does not support it.
So let's revert back to the "uglier" but working variant ...
This reverts commit bdb55fb4b3.
The number of possible parallel connections is limited by the process
file descriptor limit (among other things). Therefore try to upgrade the
current "soft" limit to its "hard" maximum (but limit it to 100000), and
show an information or even warning, wenn still less than the configured
"MaxConnections" setting.
Please note that ngIRCd and its linked libraries (like PAM) need file
descriptors not only for incoming and outgoing IRC connections, but for
reading files and inter-process communication, too! Therefore the actual
connection limit is _less_ than the file descriptor limit!
This introduces the new MAX_FD_LIMIT (100000) #define.
Show proper certificate information for incoming connections, too, and
not "peer did not present a certificate", regardless if the client sent
a certificate or not.
This change is for GnuTLS and similar to what was implemented in commit
for OpenSSL in "S2S-TLS/OpenSSL: Fix handling of certificate information
for incoming connections".
Without this, the S2S-TLS-Patch not even compiles with GnuTLS because
of the "new" GnuTLS certificate reload support implemented in commit
eead4a63 ("x509_cred_slot").
This includes simplifying cb_connserver_login_ssl() a bit, we do not
have to code for invalid state which was ruled out by an assert() and
therefore can get rid of the goto altogether (and don't log the same
error twice with different messages).
The verify callback in OpenSSL is called pretty early, and at that time
it is not possible yet to check which connection it belongs to, and some
connections may have relaxed requirements.
So always return success in the Verify_openssl() callback, and postpone
validation of the TLS session until starting the server handshake in
cb_connserver_login_ssl(), when we know which server this connection
belongs to and which options (like "SSLVerify") are in effect.
The code doing this was already present in cb_connserver_login_ssl(),
but this patch adds a more prominent comment to the function.
Show proper certificate information for incoming connections, too, and
not "peer did not present a certificate", regardless if the client sent
a certificate or not.
And free the client certificate structure "peer_cert" on incoming
connections as well!
Set the verification flags in the ConnSSL_SetVerifyProperties_openssl
function only, don't override them in ConnSSL_InitLibrary() afterwards.
No functional changes, now ConnSSL_SetVerifyProperties_openssl() sets
exactly the parameters which ConnSSL_InitLibrary() always overwrote ...
Setup host name verification even when the "SSLVerify" option is
disabled, because even then the peer can present a valid certificate and
validation would always(!) fail because of the missing host name
verification setup.
This patch provides code to validate the server certificate in
server links, defeating nasty man-in-the-middle attacks on server
links.
Features:
- Check whether the certificate is signed by a trusted certificate
authority (CA).
- Check the host name, including wildcard certificates and Subject
Alternative Names.
- Optionally check against a certificate revocation list (CRL).
- Implementation for both OpenSSL and GnuTLS linkage.
Left for another day:
- Parameterize the TLS parameter of an outbound connection. Currently,
it's hardcoded to disable all versions before TLSv1.1.
- Using certificate as CA-certificate. They work for GnuTLS only but
perhaps this should rather raise an error there, too.
- Optional OCSP checking.
- Checking client certificates. Code is there but this first needs some
consideration about the use cases. This could replace all other
authentication methods, for both client-server and server-server
connections.
This patch is based on a patch by Florian Westphal from 2009, which
implemented this for OpenSSL only:
From: Florian Westphal <fw@strlen.de>
Date: Mon, 18 May 2009 00:29:02 +0200
Subject: SSL/TLS: Add initial certificate support to OpenSSL backend
Commit message modified by Alex Barton.
Closes#120, "Server links using TLS/SSL need certificate validation".
Supersedes PR #8, "Options for verifying and requiring SSL client
certificates", which had (incomplete?) code for OpenSSL, no GnuTLS.
- Tweak some paragraphs and bring others more in line with texts on the
homepage ...
- Try to not duplicate information:
- Configuration is explained in doc/QuickStart.md;
- command line parameters are already better described in the
ngircd(8) manual page.
- Move all pointers to documentation to the README.md file, which is
directly shown in GitHub when browsing the repository, for example.
Correctly re-generate the "cloaked hostname" when removing the
"cloakhost" using an empty string by passing down NULL instead of the
empty string, which results in protocol violations (for example on
WHOIS).
