Notes from roadmapping discussion:

• fork and exec can each be implemented semi-independently
• fork may be more useful on its own than exec is, so maybe implement it first (e.g. forking-style web server)
• fork
  • need to create shared memory area and IPC for the new thread and hand it off. (similar to clone)
  • need to deal with file descriptors in shim
    • shim log
    • strace log - trickier since shadow and mprocess both write to the same descriptor. maybe can send the descriptor between shadow and the new child process over a unix socket. OK to not implement strace logging for child processes initially (but at least be careful not to keep writing to the parent's strace log)
      • possibly we could open the file twice in both processes, apparently opening in O_APPEND mode will automatically seek to the end of the file before writing (https://stackoverflow.com/questions/5284062/two-file-descriptors-to-same-file#comment5986660_5284108)
  • in shadow need to create new process and thread, and dup file descriptors
  • Need to uninit MemoryManager
    • Child regions no longer share same physical pages after a write; simplest to just treat as unmapped in Shadow and fall back to copier
  • Need to look at futexes. What happens to a futex whose physical address hasn't changed? (copy on write, but hasn't been written yet)?
  • Reset itimers (?)
  • timerfd - need to verify event fires in both parent and child (and that that's what's supposed to happen)
  • check what happens to threads (they don't run in new process; are they duplicated?)
  • pthread_atfork registers callbacks to run before fork. we may need to avoid overriding the fork libc function so that those callbacks run
• vfork
  • tricky because
    • parent is paused until child execs or exits
    • shim runs in same address space as parent
  • maybe can not implement at first. experimentally bash at least sometimes uses a fork-style clone, not vfork
• exec
  • Need to update env with handle to current process's shared memory handle
  • Shim should get reinitalized through LD_PRELOAD; may need to recognize that it was exec'd and do something different? Or might be ok as long as env has been updated.
  • Close open virtual descriptors in Shadow (if O_CLOEXEC)
  • Reset futexes
• communication between parent and child processes
  • wait4, wait, waitpid, signal on child death
• what use cases to test/prioritize
  • shell scripting (bash)
  • python subprocess management
  • site crawling with wget + python or shell wrapper script to manage a worker pool
  • tor pluggable transports
  • nginx with fork enabled
  • headless firefox (but may need substantially more functionality. e.g. sandboxing? gpu rendering? shared memory between processes :puke:)

We decided to prioritize this before project end. high cost but also high value. may help adoption, which could help future funding and further development.