`Load failed: 5: Input/output error` on ./svc.sh run | Mac Big Sur Issues?
Describe the bug
Just deployed 2 new self-hosted runners on Mac Minis, I get the error Load failed: 5: Input/output error when I install and then start the runner as a service. Interestingly the service appears to start and the runner appears online on github.com
Silicon Intel
OS Version Big Sur 11.2.3
To Reproduce
- Install the latest self-hosted runner (v2.278.0 at the time of writing) and setup the runner
./config.sh --url $URL --token $TOKEN - Install the runner as a service
./svc.sh install - Run the service
./svc.sh run
Expected behavior No error is posted
Runner Version and Platform
Silicon Intel OS Version Big Sur 11.2.3 Runner Version v2.278.0
What's not working?
The runner appears to register with github and take work, currently testing to see if there is any functional issues, will update
Job Log Output
I've replaced our actual org and hostname values with $org and $hostname
starting actions.runner.$org.$hostname
Load failed: 5: Input/output error
status actions.runner.$org.$hostname:
/Users/runner/Library/LaunchAgents/actions.runner.$org.$hostname.plist
Started:
512 0 actions.runner.$org.$hostname
Runner and Worker's Diagnostic Logs
If you google load failed 5 input/output error you will find lots of posts about this error on a wide array of software being installed as a service with lots of references to Big Sur. Something has changed in relation to Big Sur and services and the plist that gets auto generated by the service install script and / or the values within the plist no longer being valid for some reason.
Example Search Results https://developer.apple.com/forums/thread/665661 https://www.reddit.com/r/MacOS/comments/kbko61/launchctl_broken/ https://www.reddit.com/r/MacOS/comments/kbko61/launchctl_broken/gpv2to1/
EDIT
It looks like launchctl load | unload are now legacy commands, the svc.sh script probably needs looking to setup / run the service in a > Big sur compliant way?
https://ss64.com/osx/launchctl.html
Subcommands from the previous implementation of launchd are generally
available, though some may be unimplemented. Unimplemented subcommands
are documented as such.
load | unload [-wF] [-S sessiontype] [-D searchpath] paths ...
Load the specified configuration files or directories of
configuration files. Jobs that are not on-demand will be started
as soon as possible. All specified jobs will be loaded before any
of them are allowed to start. Note that per-user configuration
files (LaunchAgents) must be owned by root (if they are located
in /Library/LaunchAgents) or the user loading them (if they are
located in $HOME/Library/LaunchAgents). All system-wide daemons
(LaunchDaemons) must be owned by root. Configuration files must
disallow group and world writes. These restrictions are in place
for security reasons, as allowing writability to a launchd
configuration file allows one to specify which executable will be
launched.
🤦 we need to test the runner on Big sur and verify the behavior.
@hross FYI, since we can't upgrade our MacBook to Big Sur yet... 😢
@callum-tait-pbx Hi,have you solved the problem? I met the same issue on the Mac M1
I haven't got around to trying to solve the issue personally, functionally they seem to work but it's definitely an issue that needs solving. Based on my research at the time I think it relates to the deprecation of the load command. If I get a chance next week I'll mess with the svc.sh script.
I had a similar sounding issue that does appear to be related to the launchctl update issues. From what I understand on https://babodee.wordpress.com/2016/04/09/launchctl-2-0-syntax/ using load is now replaced with bootstrap however that alone was not my issue.
The issue I found is that load and bootstrap both require the user to be in a gui session to configure them. I originally was only configuring over ssh. Running ./svc.sh start over vnc in a terminal session worked for me.
I would like to see the setup script modified to use the newer launchctl settings and configure it to run in the user/ domain instead of gui/ that way it can be configured solely over ssh.
As a workaround, our team was able to run the following commands unattended on an AWS Mac EC2 instance. We put the following within the instance's user-data. The runsvc.sh script start the agent but in the foreground but adding the & moves the process to the background. Not the most ideal but for now seems to work for us.
sudo su -- ec2-user ./svc.sh install
sudo su -- ec2-user ./runsvc.sh start &
I've opened https://github.com/actions/runner/pull/1102 with a long term solution. Perhaps if we could get some more 👍 they will consider merging it.
Hit this too.... macOS 11.6, EC2 Mac
Can confirm that we are also hitting this for our runners as well for pytorch
hit same problem, Mac M1 on Big Sur
i've been using this to fix it:
sed -i -Ee 's;(launchctl) load -w;\1 bootstrap gui/$user_id;' -e 's;(launchctl) unload;\1 bootout gui/$user_id;' -e 's;(launchctl) list;\1 print gui/$user_id;' svc.sh
i've been using this to fix it:
sed -i -Ee 's;(launchctl) load -w;\1 bootstrap gui/$user_id;' -e 's;(launchctl) unload;\1 bootout gui/$user_id;' -e 's;(launchctl) list;\1 print gui/$user_id;' svc.sh
@jordanb-afs That seems to be broken
i've been using this to fix it: sed -i -Ee 's;(launchctl) load -w;\1 bootstrap gui/$user_id;' -e 's;(launchctl) unload;\1 bootout gui/$user_id;' -e 's;(launchctl) list;\1 print gui/$user_id;' svc.sh
@jordanb-afs That seems to be broken
hmm, it's still working on my latest builds. I run ./config.sh first, then that on top. Make sure the $user_id isn't getting replaced by your shell or etc, you could also just put the actual user id there. You'll also need to setup the runners to auto-login to the gui, otherwise it may not be able to bootstrap into gui/$user_id (which is usually gui/501 in my case)
seems to be broken for me too -
m1@848070ff-3717-456a-a062-daf9e05bb634 actions-runner % sed -i -Ee 's;(launchctl) load -w;\1 bootstrap gui/501;' -e 's;(launchctl) unload;\1 bootout gui/501;' -e 's;(launchctl) list;\1 print gui/501;' svc.sh
sed: 1: "s;(launchctl) load -w;\ ...": \1 not defined in the RE
```
This is also an issue with monterey on ec2 mac metal instance
Command: ./svc.sh start
Result:
starting actions.runner.X
Load failed: 5: Input/output error
Try running `launchctl bootstrap` as root for richer errors.
status actions.runner.X:
/Users/ec2-user/Library/LaunchAgents/actions.runner.X.plist
Stopped
ec2-user@X actions-runner % launchctl list | grep action
(Didnt start)
i've been using this to fix it:
sed -i -Ee 's;(launchctl) load -w;\1 bootstrap gui/$user_id;' -e 's;(launchctl) unload;\1 bootout gui/$user_id;' -e 's;(launchctl) list;\1 print gui/$user_id;' svc.sh
Running headlessly so won't have access to gui, has anything else worked?
@jordanb-afs Do you have any additional information about making a headless mac auto-login to gui?
It would be great if there were a purely ssh solution that didn't depend on the GUI at all. In the meantime, this can be helpful: go to System Preferences -> Users and Groups -> Login Options -> Automatic Login, and set the 'runner' or 'gha' user to login automatically. Reboot. The user will login at boot time and the services will start.
@sdarwin Will try this thanks, do you know if this works for all GUI issues
A workaround that worked for me is to install the plist file in /Library/LaunchDaemons and start the service with
sudo /bin/launchctl load /Library/LaunchDaemons/actions.runner.sebsto-xcodeinstall.plist
My plist is as follow
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key>
<string>actions.runner.sebsto-xcodeinstall.ip-172-31-67-99</string>
<key>ProgramArguments</key>
<array>
<string>/Users/ec2-user/actions-runner/runsvc.sh</string>
</array>
<key>UserName</key>
<string>ec2-user</string>
<key>GroupName</key>
<string>staff</string>
<key>WorkingDirectory</key>
<string>/Users/ec2-user/actions-runner</string>
<key>RunAtLoad</key>
<true/>
<key>StandardOutPath</key>
<string>/Users/ec2-user/Library/Logs/actions.runner.sebsto-xcodeinstall.ip-172-31-67-99/stdout.log</string>
<key>StandardErrorPath</key>
<string>/Users/ec2-user/Library/Logs/actions.runner.sebsto-xcodeinstall.ip-172-31-67-99/stderr.log</string>
<key>EnvironmentVariables</key>
<dict>
<key>ACTIONS_RUNNER_SVC</key>
<string>1</string>
</dict>
<!--
<key>ProcessType</key>
<string>Interactive</string>
-->
<key>SessionCreate</key>
<true/>
</dict>
</plist>
@sebsto has the right idea. Real services start on boot, not on login so on macOS this means the service needs to be a LaunchDaemon and not a LaunchAgent
There is more problems on Sonoma, what worked on Ventura fine to start runner agent on boot, has issues with unlocking keychain keys for code signing.
With LimitLoadToSessionType no longer able to start the service completely...
I also found we cannot properly access Keychain while keeping unattended (no-GUI login) because code-signing during Xcode build even fails on our machines on non-Sonoma (Ventura 13.6) in certain configuration.
Of course we tried LaunchDaemons w/ ProcessType commented out like https://github.com/actions/runner/issues/1056#issuecomment-1237426462 but still failed.
We ended up in associating GitHub Actions self-hosted Runner process w/ GUI sessions like this way:
- Uninstall a launchctl-based service.
- After each reboot, login to a GUI user session using an account you set up a Runner.
- Start a GUI-bound
tmuxsession via default terminal.app or iTerm, navigate to a directory where you installed a Runner. - Inside tmux session, run
./run.shthen detach from a console (check tmux manuals for how) so you can later check logs by attaching to a tmux session via SSH or GUI session.
This is kinda last-resort-nasty-hack and not a preferred solution because it needs a GUI session manual operation on every reboot but at least it works on our env for dealing with Keychain.
@muojp this is unecessary. I have GitHub agent + LaunchDaemons configuration up and running and serving dozens of builds per day. This runs on a totally unattended Amazon EC2 Mac instance on which I never connected with a GUI.
Here is the launch daemon plist file I use
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key>
<string>actions.runner.sebsto-xcodeinstall.xcodeinstall</string>
<key>ProgramArguments</key>
<array>
<string>/Users/ec2-user/actions-runner-xcodeinstall/run.sh</string>
</array>
<key>UserName</key>
<string>ec2-user</string>
<key>WorkingDirectory</key>
<string>/Users/ec2-user/actions-runner-xcodeinstall</string>
<key>RunAtLoad</key>
<true/>
<key>StandardOutPath</key>
<string>/Users/ec2-user/Library/Logs/actions.runner.sebsto-xcodeinstall.xcodeinstall/stdout.log</string>
<key>StandardErrorPath</key>
<string>/Users/ec2-user/Library/Logs/actions.runner.sebsto-xcodeinstall.xcodeinstall/stderr.log</string>
<key>EnvironmentVariables</key>
<dict>
<key>ACTIONS_RUNNER_SVC</key>
<string>1</string>
</dict>
<!--
<key>ProcessType</key>
<string>Interactive</string>
-->
<key>SessionCreate</key>
<true/>
</dict>
</plist>
And this is how I prepare the keychain
https://github.com/sebsto/amplify-ios-getting-started/blob/main/code/ci_actions/01_keychain.sh
Hey @sebsto , is ProcessType meant to be commented out? Or are you just directing attention to it and we need to remove <!--?
@NorseGaud Sorry for the lack of accuracy here. It is an actual copy of the file on my system, I left it commented out. I should have delete it for clarity.
Still getting the same: "Load failed: 5: Input/output error" error on Sonoma.
Even tried the LaunchDaemon solution and it doesn't change anything.
Still running into this issue.
Currently when you run svc.sh on macOS, the output is something like this:
actions-runner gha$ ./svc.sh
Usage:
./svc.sh [install, start, stop, status, uninstall]
actions-runner gha$
That should be modified.
either
-
installanduninstallwill install LaunchDaemons instead of user agents.
or
- Add
install-daemonanduninstall-daemonoptions to the list of choices.
Prompt the user for sudo access if needed. How are LaunchDaemons implemented? Just as sebsto explained in this issue. Put the service file in /Library/LaunchDaemons/ and run sudo /bin/launchctl load /Library/LaunchDaemons/actions.runner....
@sebsto solution works for me, but make sure to set the user/group and permissions on the plist file:
sudo chown root:wheel /Library/LaunchDaemons/actions.runner.sebsto-xcodeinstall.plist
sudo chmod 644 /Library/LaunchDaemons/actions.runner.sebsto-xcodeinstall.plist
Without setting root:wheel I still got Load failed: 5: Input/output error.
Running on macOS 15.1 Sequoia and it works without GUI login on SSH.
Runners also start automatically on reboot.
Would be great to see an official fix for this, for example adding the option to install-daemon to svc.h as @sdarwin described. It seems ideal that runners should start on boot via LaunchDaemons.
Additionally, it would be nice to have runners work on a shared machine without requiring a manual login (or setting that one particular user to auto-login, how would that work with multiple runners?)
One question for people using the "hot fix" by @sebsto: I run some GUI fuzz testing in CI, can I specify Aqua as the LimitLoadToSessionType or am I out of luck for some reason when using LaunchDaemons?