Tlog
is a terminal I/O recording and playback package suitable for
implementing centralized user session recording.
Whereas most other similar packages write the recorded data to a file in their
own format, or upload it to a custom server, tlog
sends it to a logging
service. Both the standard syslog and the journald interfaces are supported.
The recorded data is encoded in JSON in a way which keeps it
human-readable and searchable as much as possible.
The primary purpose of logging in JSON format is to eventually deliver the recorded data to a storage service such as Elasticsearch, where it can be searched and queried, and from where it can be played back.
Tlog
contains three tools: tlog-rec
for recording terminal I/O of programs
or shells in general, tlog-rec-session
for recording I/O of whole terminal
sessions, with protection from recorded users, and tlog-play
for playing
back the recordings. You can run tlog-rec
for testing or recording specific
commands or shell sessions for yourself, or you can integrate it into another
solution. Tlog-rec-session
is intended to be a user's login shell. It puts
itself between the actual user's shell and the terminal upon user login,
logging everything that passes through. Lastly, tlog-play
can playback
recordings from Elasticsearch or from a file, made with either tlog-rec
or
tlog-rec-session
. There is no difference in log format between tlog-rec
and tlog-rec-session
.
Build dependencies are systemd, cURL, json-c, and libutempter, which development
packages are systemd-devel
, json-c-devel
, libcurl-devel
, and
libutempter-devel
on RPM-based distros, and pkg-config
, libjson-c-dev
,
libsystemd-journal-dev
/libsystemd-dev
, libcurl-*-dev
, and libutempter-dev
on Debian-based distros.
To build from Git you'll need autoconf
, automake
and libtool
packages.
For creating RPM package rpm-build
is also required.
If Systemd Journal support is not required, it can be disabled with
configure's --disable-journal
option, removing the systemd dependency as
well.
Updating the system utmp and wtmp files can be enabled with the
--enable-utempter
configure option, utilizing the libutempter library.
If you'd like to build tlog
from the Git source tree, you need to first
generate the build system files:
autoreconf -i -f
After that, or if you're building a release source tarball, you need to follow the usual configure & make approach:
./configure --prefix=/usr --sysconfdir=/etc --localstatedir=/var && make
To generate a source tarball:
./configure --prefix=/usr --sysconfdir=/etc && make dist
From a source tarball you can build an SRPM package:
rpmbuild -ts <tarball>
Or an RPM package:
rpmbuild -tb <tarball>
You can use one of the release binary RPM packages and install them with your favorite tool. The RPM package does all the necessary setup for you.
Otherwise, if you built tlog
from source, you can install it with the usual
make install
:
sudo make install
If you are recording other user sessions, and don't want them to be able to
affect the recording process, make sure you use tlog-rec-session
and its
executable is SUID/SGID to separate and dedicated user and group. It doesn't
require running as root and will be safer with a regular, dedicated user.
You will also need to create the session lock directory /var/run/tlog
and
make it writable (only) for the user(s) tlog-rec-session
runs as. On systems
where (/var)/run is a tmpfs you will also need to make sure the session lock
directory is recreated on the next boot. In that case, on systems with systemd
you'll need to create a tmpfiles.d configuration file, and on init.d systems -
a startup script, creating the directory for you.
You can test if session recording and playback work in general with a freshly
installed tlog, by recording a session into a file with tlog-rec
and then
playing it back with tlog-play
.
To record into a file, execute tlog-rec
on the command line as such:
tlog-rec --writer=file --file-path=tlog.log
Both during, and after the recording you can play the session back with
tlog-play
:
tlog-play --reader=file --file-path=tlog.log
To record into the Systemd Journal, execute tlog-rec
as such:
tlog-rec --writer=journal
Along with the regular JSON log messages, when recording to Journal, tlog copies a few JSON fields to Journal fields (unless explicitly disabled) to aid searching and extracting (parts of) particular recordings:
TLOG_USER
- the user the recording was started as (user
in JSON),TLOG_SESSION
- the audit session ID of the recording process (session
in JSON),TLOG_REC
- host-unique recording ID (rec
in JSON),TLOG_ID
- log message ID within the recording (id
in JSON).
In general, selecting Journal log entries for playback is done using Journal
matches and timestamp limits, with -M/--journal-match
, -S/--journal-since
,
and -U/--journal-until
options.
In practice however, playback from Journal is usually done with a single match
against the TLOG_REC
Journal field. The TLOG_REC
field contains a copy of
the rec
field from the logged JSON data, which is a host-unique ID of the
recording. For example, to playback a recording which contains this message
(output with journalctl -o verbose
and abbreviated):
Mon 2018-01-22 10:51:48.463904 EET [s=87ea0a3f655a48cd80d7f49053860806;...
_AUDIT_LOGINUID=1000
_UID=1000
_GID=1000
_AUDIT_SESSION=2
_BOOT_ID=12ca5b356065453fb50adfe57007658a
_MACHINE_ID=2d8d017e2b1144cbbdd049a8a997911b
_HOSTNAME=bard
PRIORITY=6
_TRANSPORT=journal
_SYSTEMD_OWNER_UID=1000
TLOG_REC=12ca5b356065453fb50adfe57007658a-306a-26f2910
TLOG_USER=nkondras
TLOG_SESSION=2
TLOG_ID=1
MESSAGE={"ver":"2.3","host":"bard","rec":"12ca5b356065453fb50adfe57007658a-306a-26f2910",...
SYSLOG_IDENTIFIER=tlog-rec
_PID=12394
_COMM=tlog-rec
_SOURCE_REALTIME_TIMESTAMP=1516611108463904
you can take the ID either from the TLOG_REC
field value directly, or from
the MESSAGE
field (from the JSON rec
field). You can then playback the
whole recording like this:
tlog-play -r journal -M TLOG_REC=12ca5b356065453fb50adfe57007658a-306a-26f2910
When compiled with Systemd >= 245 it is possible to read log entries from a specific
Journal namespace using parameters -N/--journal-namespace
.
By default, once tlog-play
reaches the last message a recording has so far,
it exits. However, it can be made to poll for new messages appearing with the
-f/--follow
option, which is useful for playing back ongoing recordings.
Tlog-play
accepts several command-line options affecting playback, including
-s/--speed
for setting playback speed multiplier, -g/--goto
for specifying
the location the playback should be fast-forwarded to, and -p/--paused
for
starting playback in paused state.
Several control keys are also recognized during playback:
- SPACE or
p
for pause/resume, {
and}
for halving and doubling the playback speed,.
for stepping through the recording (on pause or during playback)G
for fast-forwarding to the end of the recording (useful with--follow
), or to the specified timestamp (seetlog-play(8)
for details),- and
q
for quitting playback.
Both tlog-rec
and tlog-rec-session
can be setup to limit the rate at which
recording's messages are logged. Tlog-rec
accepts three options:
--limit-rate=NUMBER
, --limit-burst=NUMBER
, and --limit-action=STRING
,
which specify rate limit in bytes per second, burst limit in bytes, and the
limit action (pass/delay/drop), respectively. The same parameters can be
changed using limit.rate
, limit.burst
, and limit.action
configuration
parameters for both tlog-rec
and tlog-rec-session
.
The default pass
limit action lets all the messages through unhindered,
effectively disabling rate-limiting. You can throttle logging, and slow down
the user's terminal I/O using the delay
limit action. Finally, you can
simply drop the captured I/O, going above the rate and burst limits, using the
drop
action. See tlog-rec(8)
, tlog-rec.conf(5)
, and
tlog-rec-session.conf(5)
for details.
By default tlog-play
will terminate playback, if it notices out-of-order or
missing log messages. However, it is possible to make it ignore missing
messages with the --lax
option for when you need to playback a partial
or a corrupted recording.
Change the shell of the user to be recorded to tlog-rec-session
:
sudo usermod -s /usr/bin/tlog-rec-session <user>
Login as the user on a text terminal. By default the recorded terminal data
will be delivered to Journal (if tlog was built with Journal support) or to
syslog with facility "authpriv". In both cases default priority will be
"info". It will appear in Journal, if you use journald, or in
/var/log/auth.log
on Debian-based systems, and in /var/log/secure
on
Fedora and derived systems.
Customize tlog-rec-session
configuration in
/etc/tlog/tlog-rec-session.conf
as necessary (see tlog-rec-session.conf(5)
for details).
Fedora and RHEL (and some other distros) use an approach for configuring
system locale, where the login shell is responsible for reading the locale
configuration from a file (/etc/locale.conf
) itself, instead of receiving it
through the environment variables as most programs do. Since su
clears
environment when asked for imitation of a login shell (su -
, or su -l
),
the shell can only retrieve locale configuration from that file, in that case,
on these distros.
Because tlog-rec-session
is not an actual shell and cannot read
/etc/locale.conf
file itself, it will use libc routines to read the
environment, which will fall back to ANSI_X3.4-1968
(ASCII) in these cases.
Since nowadays pure ASCII is rarely used, tlog-rec-session
assumes that
locale environment was lost, assumes the actual encoding is UTF-8, and prints
a warning.
To work that around, you can implement the approach Debian and derived distros
use. I.e. use PAM's pam_env.so module to read and set the locale environment
variables before shell or tlog-rec-session
starts. To accomplish that on
Fedora or RHEL, put this into the /etc/pam.d/system-auth
file, along with
all other session
lines:
session required pam_env.so readenv=1 envfile=/etc/locale.conf
However, tlog only supports UTF-8 so far, so the above workaround only serves to silent the fallback warning.
You can create a symlink to tlog-rec-sessions
containing the shell it should
start, in its name. E.g. if you create a symlink like this:
sudo ln -s tlog-rec-session /usr/bin/tlog-rec-session-shell-bin-zsh
and execute tlog-rec-session-shell-bin-zsh
, then tlog-rec-session
will
start and execute /bin/zsh
as the shell to record. See tlog-rec-session(8)
for details.
Such symlinks can then be assigned as login shells to certain users to have a specific shell started for them, under recording.
SSSD starting with v1.16.0 allows configuring which users and/or groups should
be recorded (have tlog-rec-session
started when they login), while also
preserving the original user shell. See sssd-session-recording(5)
.
Rsyslog can be set up to deliver tlog messages to Elasticsearch. First of
all, increase the maximum message size to be 1k more than the
tlog-rec-session
payload. The default payload is 2kB, so the rsyslog
maximum message size needs to be "3k" if the defaults are used:
$MaxMessageSize 3k
The line above needs to be above any network setup in rsyslog.conf
(put it
at the top to be safe).
Then the Elasticsearch output module needs to be loaded:
$ModLoad omelasticsearch
Before sending tlog messages to Elasticsearch they need to be reformatted
and real time timestamp needs to be added, which can be done with this
rsyslog
template:
template(name="tlog" type="list") {
constant(value="{")
property(name="timegenerated"
outname="timestamp"
format="jsonf"
dateFormat="rfc3339")
constant(value=",")
property(name="msg"
regex.expression="{\\(.*\\)"
regex.submatch="1")
constant(value="\n")
}
Then, a rule routing messages originating from tlog to Elasticsearch
needs to be added. If you installed v4 or later tlog RPM package, or set up
tlog-rec-session
as SUID/SGID to a dedicated user yourself, then the rule
can use that user ID to filter genuine tlog messages securely.
If your rsyslog receives messages from journald, with the imjournal
module,
then the rule condition should be:
if $!_UID == "<TLOG_UID>" then {
# ... actions ...
}
Note that the above would only work with rsyslog v8.17.0 and later, due to an issue preventing it from parsing variable names starting with underscore.
If your rsyslog receives messages via syslog(1) socket by itself, with the
imuxsock module, you need to enable the module's Annotate
and ParseTrusted
options. E.g. like this:
module(load="imuxsock" SysSock.Annotate="on" SysSock.ParseTrusted="on")
And then the rule condition should be:
if $!uid == "<TLOG_UID>" then {
# ... actions ...
}
The <TLOG_UID>
above should be replaced with the UID your tlog-rec-session
runs as.
Otherwise you'll need to filter by something else. For example the program
name (ident
argument to syslog(3)), which tlog specifies as tlog
. In that
case the condition could be:
if $programname == "tlog-rec-session" then {
# ... actions ...
}
However, note that any program is able to log with that program name and thus spoof tlog messages.
Once your rule condition is established, you can add the actual action sending the messages to Elasticsearch:
action(name="tlog-elasticsearch"
type="omelasticsearch"
server="localhost"
searchIndex="tlog-rsyslog"
searchType="tlog"
bulkmode="on"
template="tlog")
The action above would send messages formatted with the tlog
template,
described above, to an Elasticsearch server running on localhost and default
port, and would put them into index tlog-rsyslog
with type tlog
, using the
bulk interface.
Add the following action if you want to also send tlog messages to a dedicated file for debugging:
action(name="tlog-file"
type="omfile"
file="/var/log/tlog.log"
fileCreateMode="0600"
template="tlog")
Further, if you don't want tlog messages delivered anywhere else you can add
the discard action (~
) after both of those:
~
If you'd like to exclude tlog messages from any other logs remember to put
its rule before any other rules in rsyslog.conf
.
Here is a complete example of a rule matching messages arriving from
tlog-rec-session
running as user with UID 123, delivered from journald. It
sends them to Elasticsearch running on localhost with default port, puts them
into tlog-rsyslog
index with type tlog
, using bulk interface, stores them
in /var/log/tlog.log
file, and then stops processing, not letting them get
anywhere else.
if $!_UID == "123" then {
action(name="tlog-elasticsearch"
type="omelasticsearch"
server="localhost"
searchIndex="tlog-rsyslog"
searchType="tlog"
bulkmode="on"
template="tlog")
action(name="tlog-file"
type="omfile"
file="/var/log/tlog.log"
fileCreateMode="0600"
template="tlog")
~
}
Once you got tlog messages to Elasticsearch, you can play them back using the
still rudimentary tlog-play
command-line tool. You will need to tell it to
use the Elasticsearch reader (es
), supply it with the Elasticsearch base
URL, and the query string, which would match the messages for your session.
The base URL should point to the _search
endpoint for your type and index.
E.g. a base URL for index tlog-rsyslog
and type tlog
on localhost would
be:
http://localhost:9200/tlog-rsyslog/tlog/_search
The query string should follow the Elasticsearch query string
syntax. E.g. to look for session #17 which happened in
the last week on host server
, you can use this query string:
host:server AND timestamp:>=now-7d AND session:17
Use --reader
(or just -r
), --es-baseurl
and --es-query
options to
specify the reader, base URL, and the query string respectively. The full
command for the above parameters could look like this:
tlog-play -r es \
--es-baseurl=http://localhost:9200/tlog-rsyslog/tlog/_search \
--es-query='host:server AND timestamp:>=now-7d AND session:17'
If you're playing back an ongoing session, adding the --follow
or -f
option will make tlog-play
wait for more messages after it plays back all
that were logged so far. Just like tail -f
will wait for more lines to be
added to a file it's outputting.
Interrupt tlog-play
(e.g. press Ctrl-C) to stop the playback at any moment.
Instead of specifying the reader and the base URL on the command line every
time, you can put them in /etc/tlog/tlog-play.conf
configuration file.
Currently tlog-play
functionality is limited. It doesn't provide a way to
rewind playback, only to fast-forward. It polls files, Journal and
Elasticsearch for new messages, instead of asking to be updated when they
appear, which impairs performance. Even though the messages contain recorded
terminal window (re)sizes, it doesn't resize its own terminal to fit the
output better. Further development will be addressing these.
Please read the Contributing Guidelines for more details on the process for submitting pull requests.