Switch to a more formally-specified `kmsg`-style format for reading the logs. - update the spec in docs/logging.md - check for bad names in pkg/memlogd with unit test Signed-off-by: David Scott <dave.scott@docker.com>
3.9 KiB
Logging
By default LinuxKit will write onboot and service logs directly to files in
/var/log
and /var/log/onboot
.
It is tricky to write the logs to a disk or a network service as no disks
or networks are available until the onboot
containers run. We work around
this by splitting the logging into 2 pieces:
memlogd
: an in-memory circular buffer which receives logs (including all the earlyonboot
logs)- a log writing
service
that starts later and can download and process the logs frommemlogd
To use this new logging system, you should add the memlogd
container to
the init
block in the LinuxKit yml. On boot memlogd
will be started
from init.d
and it will listen on a Unix domain socket:
/var/run/linuxkit-external-logging.sock
The init
/service
process will look for this socket and redirect the
stdout
and stderr
of both onboot
and services
to memlogd
.
memlogd: an in-memory circular buffer
The memlogd
daemon reads the logs from the onboot
and services
containers
and stores them together with a timestamp and the name of the originating
container in a circular buffer in memory.
The contents of the circular buffer can be read over the Unix domain socket
/var/run/memlogq.sock
The circular buffer has a fixed size (overridden by the command-line argument
-max-lines
) and when it fills up, the oldest messages will be overwritten.
To store the logs somewhere more permanent, for example a disk or a remote
network service, a service should be added to the yaml which connects to
memlogd
and streams the logs. The example program logread
in the memlogd
package demonstrates how to do this.
Message format
The format used to read logs is similar to kmsg:
<timestamp>,<log>;<body>
where <timestamp>
is an RFC3339-formatted timestamp, <log>
is the name of
the log (e.g. docker-ce.out
) and <body>
is the output. The <log>
must
not contain the character ;
.
Usage examples
/ # logread -f
2018-07-05T13:22:32Z,memlogd;memlogd started
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: waiting for carrier
2018-07-05T13:22:32Z,onboot.001-dhcpcd.err;eth0: could not detect a useable init
system
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: carrier acquired
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;DUID 00:01:00:01:22:d0:d8:18:02:50:00:00:00:02
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: IAID 00:00:00:02
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: adding address fe80::d33a:3936:
2ee4:5c8c
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: soliciting an IPv6 router
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: soliciting a DHCP lease
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: offered 192.168.65.4 from 192.1
68.65.1 `vpnkit'
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: leased 192.168.65.4 for 7200 se
conds
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: adding route to 192.168.65.0/24
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;eth0: adding default route via 192.16
8.65.1
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;exiting due to oneshot
2018-07-05T13:22:32Z,onboot.001-dhcpcd.out;dhcpcd exited
^C
Current issues and limitations:
- No docker logger plugin support yet - it could be nice to add support to memlogd, so the docker container logs would also be gathered in one place
- No syslog compatibility at the moment and
/dev/log
doesn’t exist. This socket could be created to keep syslog compatibility, e.g. by using https://github.com/mcuadros/go-syslog. Processes that require syslog should then be able to log directly to memlogd. - Kernel messages not read on startup yet (but can be captured with
logwrite dmesg
) - Currently no direct external hooks exposed - but options available that
could be added. Should also be possible to pipe output to e.g.
oklog
fromlogread
(https://github.com/oklog/oklog)