cpython/Tools/wasm
Christian Heimes 069c96f84c
gh-90473: Skip and document more failing tests on WASI (GH-93436)
- Mark more ``umask()`` cases
- ``dup()`` is not supported
- ``/dev/null`` is not available
- document missing features
- mark more modules as not available
2022-06-03 00:44:48 +02:00
..
config.site-wasm32-emscripten bpo-40280: Add --enable-wasm-dynamic-linking (GH-32253) 2022-04-04 19:31:31 +02:00
config.site-wasm32-wasi gh-90473: Skip and document more failing tests on WASI (GH-93436) 2022-06-03 00:44:48 +02:00
python.html bpo-40280: Add limited Emscripten REPL (GH-32284) 2022-04-05 11:21:11 +02:00
python.worker.js bpo-40280: Add limited Emscripten REPL (GH-32284) 2022-04-05 11:21:11 +02:00
README.md gh-90473: Skip and document more failing tests on WASI (GH-93436) 2022-06-03 00:44:48 +02:00
Setup.local.example bpo-40280: WASM docs and smaller browser builds (GH-32412) 2022-04-10 09:29:51 +02:00
wasm_assets.py bpo-40280: WASM docs and smaller browser builds (GH-32412) 2022-04-10 09:29:51 +02:00
wasm_webserver.py bpo-40280: WASM docs and smaller browser builds (GH-32412) 2022-04-10 09:29:51 +02:00

Python WebAssembly (WASM) build

WARNING: WASM support is highly experimental! Lots of features are not working yet.

This directory contains configuration and helpers to facilitate cross compilation of CPython to WebAssembly (WASM). For now we support wasm32-emscripten builds for modern browser and for Node.js. WASI (wasm32-wasi) is work-in-progress

wasm32-emscripten build

For now the build system has two target flavors. The Emscripten/browser target (--with-emscripten-target=browser) is optimized for browsers. It comes with a reduced and preloaded stdlib without tests and threading support. The Emscripten/node target has threading enabled and can access the file system directly.

Cross compiling to the wasm32-emscripten platform needs the Emscripten SDK and a build Python interpreter. Emscripten 3.1.8 or newer are recommended. All commands below are relative to a repository checkout.

Christian Heimes maintains a container image with Emscripten SDK, Python build dependencies, WASI-SDK, wasmtime, and several additional tools.

# Fedora, RHEL, CentOS
podman run --rm -ti -v $(pwd):/python-wasm/cpython:Z quay.io/tiran/cpythonbuild:emsdk3

# other
docker run --rm -ti -v $(pwd):/python-wasm/cpython quay.io/tiran/cpythonbuild:emsdk3

Compile a build Python interpreter

mkdir -p builddir/build
pushd builddir/build
../../configure -C
make -j$(nproc)
popd

Fetch and build additional emscripten ports

embuilder build zlib bzip2

Cross compile to wasm32-emscripten for browser

mkdir -p builddir/emscripten-browser
pushd builddir/emscripten-browser

CONFIG_SITE=../../Tools/wasm/config.site-wasm32-emscripten \
  emconfigure ../../configure -C \
    --host=wasm32-unknown-emscripten \
    --build=$(../../config.guess) \
    --with-emscripten-target=browser \
    --with-build-python=$(pwd)/../build/python

emmake make -j$(nproc)
popd

Serve python.html with a local webserver and open the file in a browser.

emrun builddir/emscripten-browser/python.html

or

./Tools/wasm/wasm_webserver.py

and open http://localhost:8000/builddir/emscripten-browser/python.html . This directory structure enables the C/C++ DevTools Support (DWARF) to load C and header files with debug builds.

Cross compile to wasm32-emscripten for node

mkdir -p builddir/emscripten-node
pushd builddir/emscripten-node

CONFIG_SITE=../../Tools/wasm/config.site-wasm32-emscripten \
  emconfigure ../../configure -C \
    --host=wasm32-unknown-emscripten \
    --build=$(../../config.guess) \
    --with-emscripten-target=node \
    --with-build-python=$(pwd)/../build/python

emmake make -j$(nproc)
popd
node --experimental-wasm-threads --experimental-wasm-bulk-memory builddir/emscripten-node/python.js

wasm32-emscripten limitations and issues

Emscripten before 3.1.8 has known bugs that can cause memory corruption and resource leaks. 3.1.8 contains several fixes for bugs in date and time functions.

Network stack

  • Python's socket module does not work with Emscripten's emulated POSIX sockets yet. Network modules like asyncio, urllib, selectors, etc. are not available.
  • Only AF_INET and AF_INET6 with SOCK_STREAM (TCP) or SOCK_DGRAM (UDP) are available. AF_UNIX is not supported.
  • socketpair does not work.
  • Blocking sockets are not available and non-blocking sockets don't work correctly, e.g. socket.accept crashes the runtime. gethostbyname does not resolve to a real IP address. IPv6 is not available.
  • The select module is limited. select.select() crashes the runtime due to lack of exectfd support.

processes, signals

  • Processes are not supported. System calls like fork, popen, and subprocess fail with ENOSYS or ENOSUP.
  • Signal support is limited. signal.alarm, itimer, sigaction are not available or do not work correctly. SIGTERM exits the runtime.
  • Keyboard interrupt (CTRL+C) handling is not implemented yet.
  • Resource-related functions like os.nice and most functions of the resource module are not available.

threading

  • Threading is disabled by default. The configure option --enable-wasm-pthreads adds compiler flag -pthread and linker flags -sUSE_PTHREADS -sPROXY_TO_PTHREAD.
  • pthread support requires WASM threads and SharedArrayBuffer (bulk memory). The Node.JS runtime keeps a pool of web workers around. Each web worker uses several file descriptors (eventfd, epoll, pipe).
  • It's not advised to enable threading when building for browsers or with dynamic linking support; there are performance and stability issues.

file system

  • Most user, group, and permission related function and modules are not supported or don't work as expected, e.g.pwd module, grp module, os.setgroups, os.chown, and so on. lchown and `lchmod`` are not available.
  • umask is a no-op.
  • hard links (os.link) are not supported.
  • Offset and iovec I/O functions (e.g. os.pread, os.preadv) are not available.
  • os.mknod and os.mkfifo don't work and are disabled.
  • Large file support crashes the runtime and is disabled.
  • mmap module is unstable. flush (msync) can crash the runtime.

Misc

  • Heap memory and stack size are limited. Recursion or extensive memory consumption can crash Python.
  • Most stdlib modules with a dependency on external libraries are missing, e.g. ctypes, readline, sqlite3, ssl, and more.
  • Shared extension modules are not implemented yet. All extension modules are statically linked into the main binary. The experimental configure option --enable-wasm-dynamic-linking enables dynamic extensions supports. It's currently known to crash in combination with threading.
  • glibc extensions for date and time formatting are not available.
  • locales module is affected by musl libc issues, bpo-46390.
  • Python's object allocator obmalloc is disabled by default.
  • ensurepip is not available.

wasm32-emscripten in browsers

  • The interactive shell does not handle copy 'n paste and unicode support well.
  • The bundled stdlib is limited. Network-related modules, distutils, multiprocessing, dbm, tests and similar modules are not shipped. All other modules are bundled as pre-compiled pyc files.
  • In-memory file system (MEMFS) is not persistent and limited.
  • Test modules are disabled by default. Use --enable-test-modules build test modules like _testcapi.

wasm32-emscripten in node

Node builds use NODERAWFS.

  • Node RawFS allows direct access to the host file system without need to perform FS.mount() call.

Hosting Python WASM builds

The simple REPL terminal uses SharedArrayBuffer. For security reasons browsers only provide the feature in secure environents with cross-origin isolation. The webserver must send cross-origin headers and correct MIME types for the JavaScript and WASM files. Otherwise the terminal will fail to load with an error message like Browsers disable shared array buffer.

Apache HTTP .htaccess

Place a .htaccess file in the same directory as python.wasm.

# .htaccess
Header set Cross-Origin-Opener-Policy same-origin
Header set Cross-Origin-Embedder-Policy require-corp

AddType application/javascript js
AddType application/wasm wasm

<IfModule mod_deflate.c>
    AddOutputFilterByType DEFLATE text/html application/javascript application/wasm
</IfModule>

WASI (wasm32-wasi)

WASI builds require WASI SDK 15.0+ and currently wasix for POSIX compatibility stubs.

WASI limitations and issues (WASI SDK 15.0)

A lot of Emscripten limitations also apply to WASI. Noticable restrictions are:

  • Call stack size is limited. Default recursion limit and parser stack size are smaller than in regular Python builds.
  • socket(2) cannot create new socket file descriptors. WASI programs can call read/write/accept on a file descriptor that is passed into the process.
  • socket.gethostname() and host name resolution APIs like socket.gethostbyname() are not implemented and always fail.
  • chmod(2) is not available. It's not possible to modify file permissions, yet. A future version of WASI may provide a limited set_permissions API.
  • User/group related features like os.chown(), os.getuid, etc. are stubs or fail with ENOTSUP.
  • File locking (fcntl) is not available.
  • os.pipe(), os.mkfifo(), and os.mknod() are not supported.
  • process_time does not work as expected because it's implemented using wall clock.
  • os.umask() is a stub.
  • sys.executable is empty.
  • /dev/null / os.devnull may not be available.
  • os.utime*() is buggy in WASM SDK 15.0, see utimensat() with timespec=NULL sets wrong time
  • os.symlink() fails with PermissionError when attempting to create a symlink with an absolute path with wasmtime 0.36.0. The wasmtime runtime uses openat2(2) syscall with flag RESOLVE_BENEATH to open files. The flag causes the syscall to reject symlinks with absolute paths.
  • os.curdir (aka .) seems to behave differently, which breaks some importlib tests that add . to sys.path and indirectly sys.path_importer_cache.
  • WASI runtime environments may not provide a dedicated temp directory.

Detect WebAssembly builds

Python code

import os, sys

if sys.platform == "emscripten":
    # Python on Emscripten
if sys.platform == "wasi":
    # Python on WASI

if os.name == "posix":
    # WASM platforms identify as POSIX-like.
    # Windows does not provide os.uname().
    machine = os.uname().machine
    if machine.startswith("wasm"):
        # WebAssembly (wasm32, wasm64 in the future)
>>> import os, sys
>>> os.uname()
posix.uname_result(sysname='Emscripten', nodename='emscripten', release='1.0', version='#1', machine='wasm32')
>>> os.name
'posix'
>>> sys.platform
'emscripten'
>>> sys._emscripten_info
sys._emscripten_info(
    emscripten_version=(3, 1, 8),
    runtime='Mozilla/5.0 (X11; Fedora; Linux x86_64; rv:99.0) Gecko/20100101 Firefox/99.0',
    pthreads=False,
    shared_memory=False
)
>>> sys._emscripten_info
sys._emscripten_info(emscripten_version=(3, 1, 8), runtime='Node.js v14.18.2', pthreads=True, shared_memory=True)
>>> import os, sys
>>> os.uname()
posix.uname_result(sysname='wasi', nodename='(none)', release='0.0.0', version='0.0.0', machine='wasm32')
>>> os.name
'posix'
>>> sys.platform
'wasi'

C code

Emscripten SDK and WASI SDK define several built-in macros. You can dump a full list of built-ins with emcc -dM -E - < /dev/null and /path/to/wasi-sdk/bin/clang -dM -E - < /dev/null.

#ifdef __EMSCRIPTEN__
    // Python on Emscripten
#endif
  • WebAssembly __wasm__ (also __wasm)
  • wasm32 __wasm32__ (also __wasm32)
  • wasm64 __wasm64__
  • Emscripten __EMSCRIPTEN__ (also EMSCRIPTEN)
  • Emscripten version __EMSCRIPTEN_major__, __EMSCRIPTEN_minor__, __EMSCRIPTEN_tiny__
  • WASI __wasi__

Feature detection flags:

  • __EMSCRIPTEN_PTHREADS__
  • __EMSCRIPTEN_SHARED_MEMORY__
  • __wasm_simd128__
  • __wasm_sign_ext__
  • __wasm_bulk_memory__
  • __wasm_atomics__
  • __wasm_mutable_globals__