3 ms·

Right, it's theoretically dangerous, or I'd say practically: "please think when using it"... though I personally couldn't find any other serious problematic exa

by akavel 8y ago

Right, it's theoretically dangerous, or I'd say practically: "please think when using it"... though I personally couldn't find any other serious problematic example other than typing something clearly dangerous in the first place. And I personally found it immensely useful in the other, non-dangerous cases. So I just took the chance and went with the idea, and then wanted to share it with you, so you might also find it useful, maybe! :) Also... umm... you know, a tool isn't really truly Unixy if you can't hurt yourself with it, right? ;P That said, I'm starting to think — should I include a note about the potential dangers in the readme?... they seemed glaringly obvious to me (at least in case of rm-like commands), but can there be readers/users, to whom they'd not be obvious?... edit: Ok, I tried to add a warning in the Readme, that's something I can do easily, and if it helps someone, worth it...

skybrian 8y ago
It seems like a whitelist of safe commands would be a good idea? The user can override with -f if they want to live dangerously. Even rm has a -f flag to override some safety measures, though the defaults aren't very safe.
- drb91 8y ago
  Even better, put it into a non-privileged user by default with no write permission.
  - akavel 8y ago
    Sounds interesting; but is there any way I could change to a different user if not running as root?... also, even if yes, how do I dynamically create a non-privileged user, or find a pre-existing one?
    - gwn7 8y ago
      I would try creating a non-privileged user "up" during installation and switch to that user under the hood before running the command via `su`. (Though it would complicate installation which is effectively non-existent since there is a single binary; and tightly couple the application with the environment. It's a trade-off anyway.) (Edit: Somebody mentioned the user "nobody" which seems to be a better alternative.) In any case I think that the whitelist approach would work better as it would be nice if the tool knew in advance which commands work properly and which ones don't. That way it could inform its users about the allowed & unallowed commands. Commands with side effects (or "impure" in the FP terminology) don't make much sense anyway to be used within this. The main value is fast iteration, in order to verify the expected output of the pipeline. For me it doesn't make much sense to use it with commands whose main purpose is to modify the filesystem instead of generating / transforming data and writing it to stdout. So the criteria for the commands that would make it into the whitelist may be "not having side effects", "writing to stdout", and optionally "reading from stdin". Fastly iterating side effects is unsafe as well, as has already been pointed out. This can be a fine data analysis tool actually, unfortunately only for command line geeks. The experience is actually not so far from analyzing data with VIM interactively by piping it to an external UNIX command and getting the result dataset back into the editor. VIM hackers will know :)
      - seiferteric 8y ago
        I actually like the user idea better. The problem with a whitelist is that even useful commands can have subtle or little known dangerous modes, like "find . -exec rm {}"
        gwn7 8y ago
        Ah. Good point. Whitelisting commands would have already been a bit painful, and now your comment shows that the parameters need to be whitelisted/blacklisted as well, which would be crazy. In a world in which shell commands respected the UNIX philosophy, "find" wouldn't have a silly option like "exec", and other commands wouldn't mix read / write / pure data transform operations in a single command. But it is what it is. So yeah, protection probably needs to be implemented in the user level, for maximum safety. Maybe an alternative and/or complementary solution would be to profile each inputted command to detect if they are attempting write operations (maybe with "strace" or something like that), and cancel the evaluation of the command in the next iterations and/or show a warning.
      - amalcon 8y ago
        One thing to watch out for is that a process running as "nobody" has permission to manipulate (e.g. kill) other processes running as that user. This sometimes includes system processes. Allowing any user to run arbitrary commands as nobody is technically a privilege escalation, and therefore should be avoided. A single-purpose user with its own group would have this problem only to a lesser degree (you'd be able to mess with other users' "up" invocations, but not any system processes).
        infogulch 8y ago
        Could you make a `[caller]_up` user for each different user?
        gwn7 8y ago
        I didn't know this (I'm not much of a sysadmin), thank you for the information!
      - 8y ago
CGamesPlay 8y ago
Nice tool! Your main use case for destroying files is “xargs rm”, which would be used when receiving a piped argument list. Could you have this command run as user “nobody” by default, and have a flag to run as a real user? Or as others have mentioned, use capabilities on the spawned processes to prevent anything other than standard input/output?
etatoby 8y ago
I'm not so versant about the state of permission / capability systems on Linux. But if there is a way for you to drop the capability to modify the filesystem before running the first pipeline, you should definitely do it by default (and provide a switch to override it if somebody knows what they are doing.)

by akavel 8y ago

skybrian 8y ago

It seems like a whitelist of safe commands would be a good idea? The user can override with -f if they want to live dangerously. Even rm has a -f flag to override some safety measures, though the defaults aren't very safe.

drb91 8y ago
Even better, put it into a non-privileged user by default with no write permission.
- akavel 8y ago
  Sounds interesting; but is there any way I could change to a different user if not running as root?... also, even if yes, how do I dynamically create a non-privileged user, or find a pre-existing one?
  - gwn7 8y ago
    I would try creating a non-privileged user "up" during installation and switch to that user under the hood before running the command via `su`. (Though it would complicate installation which is effectively non-existent since there is a single binary; and tightly couple the application with the environment. It's a trade-off anyway.) (Edit: Somebody mentioned the user "nobody" which seems to be a better alternative.) In any case I think that the whitelist approach would work better as it would be nice if the tool knew in advance which commands work properly and which ones don't. That way it could inform its users about the allowed & unallowed commands. Commands with side effects (or "impure" in the FP terminology) don't make much sense anyway to be used within this. The main value is fast iteration, in order to verify the expected output of the pipeline. For me it doesn't make much sense to use it with commands whose main purpose is to modify the filesystem instead of generating / transforming data and writing it to stdout. So the criteria for the commands that would make it into the whitelist may be "not having side effects", "writing to stdout", and optionally "reading from stdin". Fastly iterating side effects is unsafe as well, as has already been pointed out. This can be a fine data analysis tool actually, unfortunately only for command line geeks. The experience is actually not so far from analyzing data with VIM interactively by piping it to an external UNIX command and getting the result dataset back into the editor. VIM hackers will know :)
    - seiferteric 8y ago
      I actually like the user idea better. The problem with a whitelist is that even useful commands can have subtle or little known dangerous modes, like "find . -exec rm {}"
      - gwn7 8y ago
        Ah. Good point. Whitelisting commands would have already been a bit painful, and now your comment shows that the parameters need to be whitelisted/blacklisted as well, which would be crazy. In a world in which shell commands respected the UNIX philosophy, "find" wouldn't have a silly option like "exec", and other commands wouldn't mix read / write / pure data transform operations in a single command. But it is what it is. So yeah, protection probably needs to be implemented in the user level, for maximum safety. Maybe an alternative and/or complementary solution would be to profile each inputted command to detect if they are attempting write operations (maybe with "strace" or something like that), and cancel the evaluation of the command in the next iterations and/or show a warning.
    - amalcon 8y ago
      One thing to watch out for is that a process running as "nobody" has permission to manipulate (e.g. kill) other processes running as that user. This sometimes includes system processes. Allowing any user to run arbitrary commands as nobody is technically a privilege escalation, and therefore should be avoided. A single-purpose user with its own group would have this problem only to a lesser degree (you'd be able to mess with other users' "up" invocations, but not any system processes).
      - infogulch 8y ago
        Could you make a `[caller]_up` user for each different user?
      - gwn7 8y ago
        I didn't know this (I'm not much of a sysadmin), thank you for the information!
    - 8y ago

CGamesPlay 8y ago

Nice tool! Your main use case for destroying files is “xargs rm”, which would be used when receiving a piped argument list. Could you have this command run as user “nobody” by default, and have a flag to run as a real user? Or as others have mentioned, use capabilities on the spawned processes to prevent anything other than standard input/output?

etatoby 8y ago

I'm not so versant about the state of permission / capability systems on Linux. But if there is a way for you to drop the capability to modify the filesystem before running the first pipeline, you should definitely do it by default (and provide a switch to override it if somebody knows what they are doing.)