Enclose.IO: current cutting-edges and the future work

Transcript

1. Enclose.IO: current cutting-
   edges and the future work
   Minqi Pan
   

   View Slide

2. I’m Minqi Pan
   github.com/pmq20
   twitter
   @psvr
   

   View Slide

3. We made some
   modifications to …
   • Node.js Compile-time
   • Node.js Link-time / Test-time
   • Node.js Run-time
   

   View Slide

4. Compile-time
   

   View Slide

5. • Serialize project files into structs of {path,
   *source, source_len}
   • Serialize raw file contents into array of bytes
   • Bonus of Compiler hints: the compiler can tell us
   the offsets of those JS scripts
   • (TODO) better usage of resource handling of the
   compiler
   

   View Slide

6. • TODO: maybe {path, *source, source_len, atime,
   mtime, ctime, birthtime} for fs.stat()? Anybody
   uses that info?
   • TODO: better record dir, for fs.readdir
   

   View Slide

7. • skip ParseArgs
   • (cf. in Start(), ParseArgs is called before
   StartNodeInstance and CreateEnvironment)
   

   View Slide

8. • checks for memfs in process.argv[1]
   • customized entrance via automatically setting
   process.argv[1]
   • e.g. process.argv[1]="/__enclose_io_memfs__/
   node_modules/coffee-script/bin/coffee"
   • (cf. ExecuteString called on bootstrap_node.js
   when LoadEnvironment)
   

   View Slide

9. • special treatment of __enclose_io_fork__
   • special treatment of
   ENCLOSE_IO_USE_ORIGINAL_NODE
   

   View Slide

10. • adds MEMFS::stat_file
    • adds MEMFS::stat_dir
    • adds MEMFS::resolve
    

    View Slide

11. • adds MEMFS::exists?
    • adds MEMFS::read_file
    • adds MEMFS::read_dir
    

    View Slide

12. • (TODO) simulate file descriptors
    • (TODO) Libmemfs: a portable, pure C
    implementation of a Unix-style, read-only, in-
    memory filesystem.
    

    View Slide

13. Link-time / Test-time
    

    View Slide

14. • Enforces static linking
    • copies routines of openssl, libuv, … into the
    executable image
    • (TODO) copy C extensions like fsevents used by
    the project into the executable image
    • problem of GLIBC and GLIBCXX
    

    View Slide

15. passes original tests
    

    View Slide

16. add Enclose-specific tests
    

    View Slide

17. Runtime
    

    View Slide

18. • the program is loaded into memory
    • your serialized project files are initially stored
    in .text, are then “copied” by the OS into the
    data segment (.data) of the program virtual
    address space during start-up
    • (cf.) text segment, the “REAL” executable
    instructions, read-only and has a fixed size
    • (TODO) read-only data segment (.rodata)
    

    View Slide

19. – Wikipedia: Loader (computing)
    “In the case of operating systems that support
    virtual memory, the loader may not actually
    copy the contents of executable files into
    memory, but rather may simply declare to the
    virtual memory subsystem that there is a
    mapping between a region of memory allocated
    to contain the running program's code and the
    contents of the associated executable file.”
    

    View Slide

20. – Wikipedia: Loader (computing)
    “The virtual memory subsystem is then made
    aware that pages with that region of memory
    need to be filled on demand if and when
    program execution actually hits those areas of
    unfilled memory. This may mean parts of a
    program's code are not actually copied into
    memory until they are actually used, and
    unused code may never be loaded into memory
    at all.”
    

    View Slide

21. • String::NewFromUtf8 during start-up
    • need to copy memory when passed to V8
    • from data segment to the heap, what a PAIN!
    

    View Slide

22. • (TODO) to speed up startup, use zero-copy external
    string resources, V8 no longer requires that strings are
    aligned if they are one-byte strings (string data must be
    immutable and that the data must be Latin-1 and not
    UTF-8, which would require special treatment internally
    in the engine and do not allow efficient indexing)

    

    V8 3.20.9 enforces that external pointers are aligned on
    a two-byte boundary (cf. Tagged pointers, objects are
    always at least 4-byte aligned, and we can never have
    a pointer to the middle of an object in JavaScript)
    • (TODO) maybe zero-copy buffers?
    

    View Slide

23. • (TODO) to speed up startup, delay
    String::NewExternalOneByte only when
    accessed, not running literally millions of
    String::NewFromUtf8/
    String::NewExternalOneByte at start up
    

    View Slide

24. • require first finds in memfs
    • require.resolve first resolves in memfs
    

    View Slide

25. • fs.readFile, fs.readFileSync: first checks for
    memfs path
    • fs.readdir: first checks for memfs path
    

    View Slide

26. • fs.fstat, fs.lstat, fs.lstatSync, fs.statSync: first
    checks for memfs path
    • fs.watch, fs.watchFile: does not watch for memfs
    • fs.realpathSync: first checks for memfs path
    

    View Slide

27. • child_process.fork needs the original argv[1]
    semantic
    • will set __enclose_io_fork__ as argv[1] in
    child_process.fork
    

    View Slide

28. • child_process.spawn(…process.execPath…)
    needs the original node semantic
    • will set ENCLOSE_IO_USE_ORIGINAL_NODE
    

    View Slide

29. Future Work
    • see previous TODO’s
    • only include JS scripts that can be referenced
    • only include symbols that can be referenced
    • generate unoptimized machine code in compile-
    time via v8, no sources required in runtime
    

    View Slide

30. Thank you!
    https://github.com/enclose-io
    

    View Slide