Add a compile-time option for a daemon binary #46
Conversation
The "systemd-daemon" feature will compile PARSEC to be used as a systemd, socket activated daemon. Add systemd unit files and instruction on how to install the service. Signed-off-by: Hugues de Valon <[email protected]>
| // When run as a systemd daemon, a file descriptor mapping to the Domain Socket | ||
| // should have been passed to this process. | ||
| if let Some(os_error) = err.raw_os_error() { | ||
| // On Linux, 9 is EBADF (Bad file number) |
There was a problem hiding this comment.
Is it reasonable that we'll reach this point? if not => unreachable!();
There was a problem hiding this comment.
If we reach this point, it is because the accept method failed and we check if the failure comes from the fact that it was given a bad file descriptor. We gave it 3 as a file descriptor but if the process was not launched by systemd, 3 will map to nothing and trigger this error.
There was a problem hiding this comment.
It's also possible that the user wants to run the application as a service instead of socket activation. TBH, using fd 3 directly feels a bit dirty.
According to http://0pointer.de/blog/projects/socket-activation.html, sd_listen_fds would be a better solution, but my sd-notify crate doesn't support that (yet?).
There was a problem hiding this comment.
Hi @lnicola!
Currently we only have the compile-time flag systemd-daemon to differentiate between a binary to be launched by systemd or by the command line. If compiled as a daemon and launched on the command line by the user, it will trigger this error.
I agree that a better way would be to do it at runtime, the way shown in your link below and assuming that we can also do runtime checks on other daemon-dependant parts of our application and remove the compile-time flag.
For example, what would be the effect of sd-notify if not run as a daemon?
The crate systemd has Rust wrappers around this function and other very useful functions that would allow us to do that. It is licensed under LGPL-2.1+ so we would have to check for compatibility with our Apache 2.0 license.
There was a problem hiding this comment.
Currently we only have the compile-time flag systemd-daemon to differentiate between a binary to be launched by systemd or by the command line. If compiled as a daemon and launched on the command line by the user, it will trigger this error.
That shouldn't be needed in general, IMHO.
For example, what would be the effect of sd-notify if not run as a daemon?
Do you mean sd_notify::notify? I haven't checked the implementation of sd_listen_fds, but readiness notification works like this:
- if
systemdor a compatible service manager starts the daemon, it sets an well-known environment variable (NOTIFY_SOCKET) with a Unix Domain Socket address - applications can check whether the variable is set, and can send messages and control
systemdby writing to that socket - applications should unset the variable before spawning child processes
In the case of notify, if the program is not launched by systemd, the variable won't be set, and the function will return Ok(()), so the whole thing is a no-op.
The crate systemd has Rust wrappers around this function and other very useful functions that would allow us to do that. It is licensed under LGPL-2.1+ so we would have to check for compatibility with our Apache 2.0 license.
I don't think they're compatible. I can look into adding more functionality to the sd-notify crate. If you want an alternative to sd-notify, there's https://github.com/lucab/libsystemd-rs, which is licensed under MIT/Apache 2.0. The reason I started sd-notify is lucab/libsystemd-rs#24 (i.e. having ~15 transitive dependencies, including serde).
There was a problem hiding this comment.
That shouldn't be needed in general, IMHO.
I agree now with this philosophy to make one binary that works both when run as an user or by systemd but something I forgot as well: ideally (for now) we would like parsec to work on as many OSes as possible. The compile-time flag is a way to make it work also on OSes that do not have systemd. And just having conditional compilation and target Linux only might not be enough if some Linux distributions do not have systemd?
In the case of notify, if the program is not launched by systemd, the variable won't be set, and the function will return Ok(()), so the whole thing is a no-op.
Thanks for the explanation, that is very good to know!
I can look into adding more functionality to the sd-notify crate. If you want an alternative to sd-notify, there's https://github.com/lucab/libsystemd-rs, which is licensed under MIT/Apache 2.0. The reason I started sd-notify is lucab/libsystemd-rs#24.
Even if we end up keeping the compile-time flag, it would be more robust and less dirty to call sd_listen_fds as done on the blog post. Like this, even if the flag is set the binary can be launched both ways.
I checked the docs of libsystemd-rs and it does not seem to have the function.
If you were to add a sd_listen_fds function to your crate, we would use it for sure 😋 !!
There was a problem hiding this comment.
And just having conditional compilation and target Linux only might not be enough if some Linux distributions do not have systemd?
It's totally fine to run such a binary on a non-systemd distro -- unless you're linking to the C libsystemd, which might prove problematic (and you also want to avoid it because of the licensing issue?).
If you were to add a sd_listen_fds function to your crate, we would use it for sure yum !!
No promises, but I'll take a look to see what it would imply. I didn't implement FD passing in sd-notify because it's quite complex, and it's probably using the same mechanism.
There was a problem hiding this comment.
It's totally fine to run such a binary on a non-systemd distro -- unless you're linking to the C libsystemd, which might prove problematic (and you also want to avoid it because of the licensing issue?).
Oh, just got it, I did not know you could implement the function in Rust only without any systemd dependency, as you did for notify. In that case, the only conditional compilation checks would be on the OS which is good!
No promises, but I'll take a look to see what it would imply. I didn't implement FD passing in sd-notify because it's quite complex, and it's probably using the same mechanism.
Great, please ping me if you ever do 😃 !
The "systemd-daemon" feature will compile PARSEC to be used as a
systemd, socket activated daemon.
Add systemd unit files and instruction on how to install the service.
Closes #35
A
Dockerfilecould be added later to install the daemon from a local checkout ofparsecinside a container and only expose the socket to the host.@ionut-arm FYI, the systemd unit files contain two hardcoded paths:
/home/parsec/.cargo/bin/parsecfor the binary/home/parsec/parsec.sockfor the socket