Welcome to the Linux Foundation Forum!

Why is execve so monolitic?

mouaimouai Posts: 1
edited October 14 in Software Development

Hi, I have a tought "why stuff is'nt that way question" that just got closed on SO. I am really lost on where to find someone able to answer me. If you thinks it's not the place, please let me know. Basically I don't understand why no interface exists that separate the loading phase from the execution one. To me its just a matter of code organization and should have been done long ago. So I suspect to something be wrong in my representation.

Original Question

https://stackoverflow.com/questions/64336459/why-execve-is-so-monolitic

I have a C project where I pass most of the time executing the same small external process (Always the same but chosen by user, that run with different arguments). The external process is really short and fast, and I measured that the fork/execve call is taking an overwhelming time of the operation.

It just strike me how there is just no way to save the whole loading phase. For each execution you need the OS to go from a path to the whole process image again just to change argv, and I pay the price millions of times. This involve opening the executable file, detecting format/shebang, parsing the format, for ELF resolving dynamic libraries calls by calling ld-linux, creating kernels structures to monitor process...

My dreams execution interface would expose two functions:

1.A function that from a path create a C structure representing some sort of ready to execute process image which would use the already existing first part of execve code. This image would have an empty stack.

2.And another one that map that image where it need to be executed, taking care to set up the stack so it would match given argv and envp.

It would obviously not change anything in regard to the execve interface which would just call internally the two functions in a row.

Not to mention the fact that this would probably also facilitate the implementation of software willing to copy there current state to disk for later continuation.

Clearly there is something that I don't understand and would like someone that know well about this subject explaining what is preventing the execution to work like that? Is it some kind of security problem? Or have I a wrong representation on how process are loaded and this could not work maybe? In that case can someone point me toward where is my misconception?

Thanks for the attention.

What can I do?

Also, the original question got closed as unfocused on SO because I also asked for suggestion to diminish the overhead of execution. I think this website must probably be more open than SO in this regard. So if you have any idea, experience in this subject, I will take it.

What I planned to do (I would like your feed back on thoses as well):

1- Use posix_spawn.
2- Create a ram file copy of the exe and run posix_spawn on it unstead of disk file.
3- Use prelink on the exe beforehand.

Thanks!

Sign In or Register to comment.