I want to write software that will detect all used/created/modified/deleted files during the execution of a process (and its child processes). The process has not yet run - the user provides a command line which will later be subprocessed via bash, so we can do things before and after execution, and control the environment the command is run in.
I have thought of four methods so far that could be useful:
- Parse the command line to identify files and directories mentioned. Assume all files explicitly mentioned are used. Check directories before/after for created/deleted files. MD5 existing files before/after to see any are modified. This works on all operating systems and environments, but obviously has serious limitations (doesnt work when command is "./script.sh")
- Run the process via another process like strace (dtruss for OSX, and there are equivalent windows programs), which listens for system calls. Parse output file to find files used/modified/deleted/created. Pros that its more sensitive than MD5 method and can deal with script.sh. Cons that its very OS specific (dtruss requires root privileges even if the process being run does not - outputs of tools all different). Could also create huge log files if there are a lot of read/write operations, and will certainly slow things down.
- Integrate something similar to the above into the kernel. Obviously still OS specific, but at least now we are calling the shots, creating common output format for all OS's. Wouldn't create huge log files, and could even stop hooking syscalls to, say, read() after process has requested the first read() to the file. I think this is what the tool inotify is doing, but im not familiar with it at all, nor kernel programming!
- Run the process using the LD_PRELOAD trick (called DYLD_INSERT_LIBRARIES on OSX, not sure if it exists in Windows) which basically overwrites any call to open() by the process with our own version of open() which logs what we're opening. Same for write, read, etc. It's very simple to do, and very performant since you're essentially teaching the process to log itself. The downside is that it only works for dynamically-linked process, and i have no idea of the prevalence of dynamic/statically linked programs. I dont even know if it is possible before execution to tell if a process is dynamically or statically linked (with the intention of using this method by default, but falling back to a less-performant method if its not possible).
I need help choosing the optimal path to go down. I have already implemented the first method because it was simple and gave me a way to work on the logging backend (http://ac.gt/log) but really i need to upgrade to one of the other methods. Your advice would be invaluable :)
Take a look to the source code of "strace" (and its -f to trace children). It does basically what you are trying to do. It captures all the system calls of the process (or its childs) so you can grep for operations like "open", etc.
The following link provides some examples of implementing your own strace by using the ptrace system call:
https://blog.nelhage.com/2010/08/write-yourself-an-strace-in-70-lines-of-code/