keepalived fails to restart cleanly due to the wrong systemd settings
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
keepalived (Debian) |
Fix Released
|
Unknown
|
|||
keepalived (Ubuntu) |
Fix Released
|
Undecided
|
Unassigned | ||
Xenial |
Fix Released
|
Medium
|
Christian Ehrhardt | ||
systemd (Ubuntu) |
Opinion
|
Undecided
|
Unassigned | ||
Xenial |
Opinion
|
Undecided
|
Unassigned |
Bug Description
[Impact]
* Restarts of keepalived can leave stale processes with the old configuration around.
* The systemd detection of the MainPID is suboptimal, and combined with not waiting on signals being handled it can fail on second restart killing the (still) remaining process of the first start.
* Upstream has a PIDFile statement, this has proven to avoid the issue in the MainPID guessing code of systemd.
[Test Case]
* Set up keepalived, the more complex the config is the "bigger" is the reace window, below in the description is a trivial sample config that works well.
* As a test run the loop restarting the service head-to-head while staying under the max-restart limit
$ for j in $(seq 1 20); do sleep 11s; time for i in $(seq 1 5); do sudo systemctl restart keepalived; sudo systemctl status keepalived | egrep 'Main.*exited'; done; done
Expectation: no output other than timing
Without fix: sometimes MainPIDs do no more exist, in these cases the child processes are the "old" ones from last execution with the old config.
[Regression Potential]
* Low because
* A PIDFile statement is recommended by systemd for type=forking services anyway.
* Upstream keepalived has this statement in their service file
* By the kind of change, it should have no functional impact to other parts of the service other than for the PID detection of the job by Systemd.
* Yet regression potential is never zero. There might be the unlikely case, which were considered working before due to a new config not properly being picked up. After the fix they will behave correctly and might show up as false-positives then if e.g. config was bad.
[Other Info]
* Usually a fix has to be in at least the latest Development release before SRUing it. But as I outlined below in later Releases than Xenial systemd seems to have improved making this change not-required. We haven't identified the bits for this (there is a bug task here), and they might as well be very complex. I think it is correct to fix Xenial in this regard with the simple change to the service file for now.
* To eventually match I created a Debian bug task to ask them for the inclusion of the PIDFile so it can slowly tickle back down to newer Ubuntu Releases - also there more often people run backports where the issue might occur on older systemd versions (just as it does for us on Xenial)
---
Because "PIDFile=" directive is missing in the systemd unit file, keepalived sometimes fails to kill all old processes. The old processes remain with old settings and cause unexpected behaviors. The detail of this bug is described in this ticket in upstream: https:/
The official systemd unit file is available since version 1.2.24 by this commit:
https:/
This includes "PIDFile" directive correctly:
PIDFile=
We should go the same way.
I am using Ubuntu 16.04.1, kernel 4.4.0-45-generic.
Package: keepalived
Version: 1.2.19-1
=======
How to reproduce:
I used the two instances of Ubuntu 16.04.2 on DigitalOcean:
Configurations
--------------
MASTER server's /etc/keepalived
vrrp_script chk_nothing {
script "/bin/true"
interval 2
}
vrrp_instance G1 {
interface eth1
state BACKUP
priority 100
virtual_
unicast_src_ip <primal IP>
unicast_peer {
<secondal IP>
}
track_script {
chk_nothing
}
}
BACKUP server's /etc/keepalived
vrrp_script chk_nothing {
script "/bin/true"
interval 2
}
vrrp_instance G1 {
interface eth1
state MASTER
priority 200
virtual_
unicast_src_ip <secondal IP>
unicast_peer {
<primal IP>
}
track_script {
chk_nothing
}
}
Loop based probing for the Error to exist:
-------
After the setup above start keepalived on both servers:
$ sudo systemctl start keepalived.service
Then run the following loop
$ for j in $(seq 1 20); do sleep 11s; time for i in $(seq 1 5); do sudo systemctl restart keepalived; sudo systemctl status keepalived | egrep 'Main.*exited'; done; done
Expected: no error, only time reports
Error case: Showing Main PID exited, details below
Step by Step Procedures
-------
1) Start keepalived on both servers
$ sudo systemctl start keepalived.service
2) Restart keepalived on either one
$ sudo systemctl restart keepalived.service
3) Check status and PID
$ systemctl status -n0 keepalived.service
Result
------
0) Before restart
Main PID is 3402 and the subprocesses' PIDs are 3403-3406. So far so good.
root@
● keepalived.service - Keepalive Daemon (LVS and VRRP)
Loaded: loaded (/lib/systemd/
Active: active (running) since Sat 2017-03-04 01:37:12 UTC; 14min ago
Process: 3402 ExecStart=
Main PID: 3403 (keepalived)
Tasks: 3
Memory: 1.7M
CPU: 1.900s
CGroup: /system.
├─3403 /usr/sbin/
├─3405 /usr/sbin/
└─3406 /usr/sbin/
1) First restart
Now Main PID is 3403, which was one of the previous subprocesses and is actually exited. Something is wrong. Yet, the previous processes are all exited; we are not likely to see no weird behaviors here.
root@
root@
● keepalived.service - Keepalive Daemon (LVS and VRRP)
Loaded: loaded (/lib/systemd/
Active: active (running) since Sat 2017-03-04 01:51:45 UTC; 1s ago
Process: 4782 ExecStart=
Main PID: 3403 (code=exited, status=0/SUCCESS)
Tasks: 3
Memory: 1.7M
CPU: 11ms
CGroup: /system.
├─4783 /usr/sbin/
├─4784 /usr/sbin/
└─4785 /usr/sbin/
2) Second restart
Now Main PID is 4783 and subprocesses' PIDs are 4783-4785. This is problematic as 4783 is the old process, which should have exited before new processes arose. Therefore, keepalived remains in old settings while users believe it uses the new setting.
root@
root@
● keepalived.service - Keepalive Daemon (LVS and VRRP)
Loaded: loaded (/lib/systemd/
Active: active (running) since Sat 2017-03-04 01:51:49 UTC; 1s ago
Process: 4796 ExecStart=
Main PID: 4783 (keepalived)
Tasks: 3
Memory: 1.7M
CPU: 6ms
CGroup: /system.
├─4783 /usr/sbin/
├─4784 /usr/sbin/
└─4785 /usr/sbin/
description: | updated |
description: | updated |
description: | updated |
Changed in keepalived (Ubuntu): | |
status: | Fix Released → Confirmed |
Changed in systemd (Ubuntu): | |
milestone: | none → ubuntu-17.03 |
assignee: | nobody → Dimitri John Ledkov (xnox) |
Changed in keepalived (Debian): | |
status: | Unknown → New |
Changed in systemd (Ubuntu): | |
assignee: | Dimitri John Ledkov (xnox) → nobody |
milestone: | ubuntu-17.03 → none |
Changed in keepalived (Debian): | |
status: | New → Fix Released |
This patch should fix the issue.