Basically, the SUID bit makes a program get the permissions of the owner when executed. If you set /bin/bash as SUID, suddenly every bash shell would be a root shell, kind of. Processes on Linux have a real user ID, an effective user ID, and also a saved user ID that can be used to temporarily drop privileges and gain them back again later.
So tools like sudo and doas use this mechanism to temporarily become root, then run checks to make sure you’re allowed to use sudo, then run your command. But that process is still in your user’s session and process group, and you’re still its real user ID. If anything goes wrong between sudo being root and checking permissions, that can lead to a root shell when you weren’t supposed to, and you have a root exploit. Sudo is entirely responsible for cleaning the environment before launching the child process so that it’s safe.
Run0/systemd-run acts more like an API client. The client, running as your user, asks systemd to create a process and give you its inputs and outputs, which then creates it on your behalf on a clean process tree completely separate from your user session’s process tree and group. The client never ever gets permissions, never has to check for the permissions, it’s systemd that does over D-Bus through PolKit which are both isolated and unprivileged services. So there’s no dangerous code running anywhere to exploit to gain privileges. And it makes run0 very non-special and boring in the process, it really does practically nothing. Want to make your own in Python? You can, safely and quite easily. Any app can easily integrate sudo functionnality fairly safely, and it’ll even trigger the DE’s elevated permission prompt, which is a separate process so you can grant sudo access to an app without it being able to know about your password.
Run0 takes care of interpreting what you want to do, D-Bus passes the message around, PolKit adds its stamp of approval to it, systemd takes care of spawning of the process and only the spawning of the process. Every bit does its job in isolation from the others so it’s hard to exploit.
(I’ll attempt this based on my understanding of both)
Pouring a cup of juice is something an adult needs to be involved with.
sudo is when you ask for permission to pour your own cup of juice. You ask an adult, they give you the cup and the juice, and then you’re responsible for pouring it. If the adult isn’t paying attention they may leave the fridge open for you to go back for more juice or another beverage, but otherwise you’re limited to the amount of juice the adult has given you.
run0 is when the adult just gets you a cup of juice. You tell them what you want, they go and pour the juice, and just give you the cup with the juice in it. You never enter the kitchen, so you don’t have access to the fridge, just your cup of juice.
Some people are opposed to sudo being a fairly complex program with an awkward to understand configuration language and a couple of methods that can fetch config from elsewhere. Fixing upstream sudo can’t happen because those features exist and are presumably used by some subset of people, so straight up removing them is not good, but luckily doas and sudo-rs exist as alternatives with a somewhat stripped featureset and less footguns.
Others are opposed to the concept of SUID. Underneath all the SUID stuff lies far more complexity than is obvious at first sight. There’s a pretty decent chunk of code in glibc’s libdl that will treat all kinds of environment variables differently based on whether an executable is SUID, and when that goes wrong, it’s reported as a glibc bug (last year’s glibc CVE-2023-4911 was this). And that gets all the more weird when fancy Linux features like namespaces get involved.
Removing SUID requires an entirely different implementation and the service manager is the logical place for that. That’s not just Lennart’s idea; s6, as minimal and straight to the point as it tends to be, also implements s6-sudo{,d,c}. It’s a bit more awkward to use but is a perfectly “Unix philosophy” style implementation of this very same idea.
The article talks about
sudo
anddoas
being SUID binaries and having a larger attack surface thanrun0
would. Could someone ELI5 what this means?Basically, the SUID bit makes a program get the permissions of the owner when executed. If you set
/bin/bash
as SUID, suddenly every bash shell would be a root shell, kind of. Processes on Linux have a real user ID, an effective user ID, and also a saved user ID that can be used to temporarily drop privileges and gain them back again later.So tools like
sudo
anddoas
use this mechanism to temporarily become root, then run checks to make sure you’re allowed to use sudo, then run your command. But that process is still in your user’s session and process group, and you’re still its real user ID. If anything goes wrong between sudo being root and checking permissions, that can lead to a root shell when you weren’t supposed to, and you have a root exploit. Sudo is entirely responsible for cleaning the environment before launching the child process so that it’s safe.Run0/systemd-run acts more like an API client. The client, running as your user, asks systemd to create a process and give you its inputs and outputs, which then creates it on your behalf on a clean process tree completely separate from your user session’s process tree and group. The client never ever gets permissions, never has to check for the permissions, it’s systemd that does over D-Bus through PolKit which are both isolated and unprivileged services. So there’s no dangerous code running anywhere to exploit to gain privileges. And it makes run0 very non-special and boring in the process, it really does practically nothing. Want to make your own in Python? You can, safely and quite easily. Any app can easily integrate sudo functionnality fairly safely, and it’ll even trigger the DE’s elevated permission prompt, which is a separate process so you can grant sudo access to an app without it being able to know about your password.
Run0 takes care of interpreting what you want to do, D-Bus passes the message around, PolKit adds its stamp of approval to it, systemd takes care of spawning of the process and only the spawning of the process. Every bit does its job in isolation from the others so it’s hard to exploit.
Can someone ELI3?
(I’ll attempt this based on my understanding of both)
Pouring a cup of juice is something an adult needs to be involved with.
sudo is when you ask for permission to pour your own cup of juice. You ask an adult, they give you the cup and the juice, and then you’re responsible for pouring it. If the adult isn’t paying attention they may leave the fridge open for you to go back for more juice or another beverage, but otherwise you’re limited to the amount of juice the adult has given you.
run0 is when the adult just gets you a cup of juice. You tell them what you want, they go and pour the juice, and just give you the cup with the juice in it. You never enter the kitchen, so you don’t have access to the fridge, just your cup of juice.
Okay now please ELI1
when in need, cry out for mommy!
Gagagoogoo Gagaga
Thank you, i didnt really understand what this was about ubtil now
Why not just fix sudo then?
Or just use
doas
, it’s still more secure than sudoSome people are opposed to
sudo
being a fairly complex program with an awkward to understand configuration language and a couple of methods that can fetch config from elsewhere. Fixing upstreamsudo
can’t happen because those features exist and are presumably used by some subset of people, so straight up removing them is not good, but luckilydoas
andsudo-rs
exist as alternatives with a somewhat stripped featureset and less footguns.Others are opposed to the concept of SUID. Underneath all the SUID stuff lies far more complexity than is obvious at first sight. There’s a pretty decent chunk of code in glibc’s libdl that will treat all kinds of environment variables differently based on whether an executable is SUID, and when that goes wrong, it’s reported as a glibc bug (last year’s glibc CVE-2023-4911 was this). And that gets all the more weird when fancy Linux features like namespaces get involved.
Removing SUID requires an entirely different implementation and the service manager is the logical place for that. That’s not just Lennart’s idea; s6, as minimal and straight to the point as it tends to be, also implements
s6-sudo{,d,c}
. It’s a bit more awkward to use but is a perfectly “Unix philosophy” style implementation of this very same idea.