simplify python entrypoint

- remove redundant monitor sibling process
- use python3-login executable instead of login shell subprocess
  (same effect, but in more natural order)
- use non-blocking binary IO in tee instead of readline
  (switch to binary mode, as text wrappers don't support non-blocking mode
   see https://bugs.python.org/issue13322)

repo2docker/buildpacks/base.py

@@ -9,9 +9,6 @@ import string
 import sys
 import hashlib
 import escapism
-import xml.etree.ElementTree as ET
-
-from traitlets import Dict
 
 # Only use syntax features supported by Docker 17.09
 TEMPLATE = r"""
@@ -182,6 +179,7 @@ ENV R2D_ENTRYPOINT "{{ start_script }}"
 
 # Add entrypoint
 ENV PYTHONUNBUFFERED=1
+COPY /python3-login /usr/local/bin/python3-login
 COPY /repo2docker-entrypoint /usr/local/bin/repo2docker-entrypoint
 ENTRYPOINT ["/usr/local/bin/repo2docker-entrypoint"]
 
@@ -194,9 +192,7 @@ CMD ["jupyter", "notebook", "--ip", "0.0.0.0"]
 {% endif %}
 """
 
-ENTRYPOINT_FILE = os.path.join(
+HERE = os.path.dirname(os.path.abspath(__file__))
-    os.path.dirname(os.path.abspath(__file__)), "repo2docker-entrypoint"
-)
 
 # Also used for the group
 DEFAULT_NB_UID = 1000
@@ -583,7 +579,8 @@ class BuildPack:
             dest_path, src_path = self.generate_build_context_filename(src)
             tar.add(src_path, dest_path, filter=_filter_tar)
 
-        tar.add(ENTRYPOINT_FILE, "repo2docker-entrypoint", filter=_filter_tar)
+        for fname in ("repo2docker-entrypoint", "python3-login"):
+            tar.add(os.path.join(HERE, fname), fname, filter=_filter_tar)
 
         tar.add(".", "src/", filter=_filter_tar)
 

repo2docker/buildpacks/python3-login

@@ -0,0 +1,11 @@
+#!/bin/bash -l
+# This is an executable that launches Python in a login shell
+# to ensure that full profile setup occurs.
+# shebang on linux only allows 1 argument,
+# so we couldn't pick a login shell in one shebang line
+# for a Python script
+
+# -u means unbuffered, which one ~always wants in a container
+# otherwise output can be mysteriously missing
+
+exec python3 -u "$@"

repo2docker/buildpacks/repo2docker-entrypoint

@@ -1,4 +1,4 @@
-#!/usr/bin/env python3
+#!/usr/local/bin/python3-login
 # note: must run on Python >= 3.5, which mainly means no f-strings
 
 # goals:
@@ -6,153 +6,90 @@
 # - preserve signal handling of subprocess (kill -TERM and friends)
 # - tee output to a log file
 
-import json
+import fcntl
 import os
+import select
 import signal
 import subprocess
 import sys
-import time
-
-
-def get_login_env():
-    """Instantiate a login shell to retrieve environment variables
-
-    Serialize with Python to ensure proper escapes
-    """
-    p = subprocess.run(
-        [
-            "bash",
-            "-l",
-            "-c",
-            "python3 -c 'import os, json; print(json.dumps(dict(os.environ)))'",
-        ],
-        stdout=subprocess.PIPE,
-    )
-    if p.returncode:
-        print("Error getting login env")
-        return {}
-
-    last_line = p.stdout.splitlines()[-1]
-    try:
-        return json.loads(last_line)
-    except Exception as e:
-        print("Error getting login env: {e}".format(e=e), file=sys.stderr)
-        return {}
-
-
-def monitor_parent(parent_pid, child_pgid):
-    """Monitor parent_pid and shutdown child_pgid if parent goes away first"""
-    while True:
-        try:
-            os.kill(parent_pid, 0)
-        except ProcessLookupError:
-            # parent is gone, likely by SIGKILL
-            # send SIGKILL to child process group
-            try:
-                os.killpg(child_pgid, signal.SIGKILL)
-            except (ProcessLookupError, PermissionError):
-                # ignore if the child is already gone
-                pass
-            return
-        else:
-            time.sleep(1)
 
+# output chunk size to read
+CHUNK_SIZE = 1024
 
 # signals to be forwarded to the child
-SIGNALS = [
+# everything catchable, excluding SIGCHLD
-    signal.SIGHUP,
+SIGNALS = set(signal.Signals) - {signal.SIGKILL, signal.SIGSTOP, signal.SIGCHLD}
-    signal.SIGINT,
-    # signal.SIGKILL,
-    signal.SIGQUIT,
-    signal.SIGTERM,
-    signal.SIGUSR1,
-    signal.SIGUSR2,
-    signal.SIGWINCH,
-]
 
 
 def main():
 
-    # load login shell environment
-    login_env = get_login_env()
-    env = os.environ.copy()
-    env.update(login_env)
-
     # open log file to send output
     log_file = open(
         os.path.join(os.environ.get("REPO_DIR", "."), ".jupyter-server-log.txt"),
-        "a",
+        "ab",
     )
 
+    # build the command
+    # like `exec "$@"`
     command = sys.argv[1:]
+    # load entrypoint override from env
     r2d_entrypoint = os.environ.get("R2D_ENTRYPOINT")
     if r2d_entrypoint:
         command.insert(0, r2d_entrypoint)
 
+    # launch the subprocess
     child = subprocess.Popen(
         command,
         bufsize=1,
-        env=env,
-        start_new_session=True,
         stdout=subprocess.PIPE,
         stderr=subprocess.STDOUT,
-        universal_newlines=True,
     )
-    child_pgid = os.getpgid(child.pid)
-
-    # if parent is forcefully shutdown,
-    # make sure child shuts down immediately as well
-    parent_pid = os.getpid()
-
-    monitor_pid = os.fork()
-    if monitor_pid == 0:
-        # child process, sibling of 'real' command
-        # avoid receiving signals sent to parent
-        os.setpgrp()
-        # terminate child if parent goes away,
-        # e.g. in ungraceful KILL not relayed to children
-        monitor_parent(parent_pid, child_pgid)
-        return
 
     # hook up ~all signals so that every signal the parent gets,
     # the children also get
 
     def relay_signal(sig, frame):
         """Relay a signal to children"""
-        print(
+        # DEBUG: show signal
-            "Forwarding signal {sig} to {child_pgid}".format(
+        child.send_signal(sig)
-                sig=sig, child_pgid=child_pgid
-            )
-        )
-        os.killpg(child_pgid, sig)
 
-    # question: maybe use all valid_signals() except a few, e.g. SIGCHLD?
-    # rather than opt-in list
     for signum in SIGNALS:
         signal.signal(signum, relay_signal)
 
     # tee output from child to both our stdout and the log file
     def tee(chunk):
-        for f in [sys.stdout, log_file]:
+        """Tee output from child to both our stdout and the log file"""
+        for f in [sys.stdout.buffer, log_file]:
             f.write(chunk)
             f.flush()
 
-    while child.poll() is None:
+    # make stdout pipe non-blocking
-        tee(child.stdout.readline())
+    # this means child.stdout.read(nbytes)
+    # will always return immediately, even if there's nothing to read
+    flags = fcntl.fcntl(child.stdout, fcntl.F_GETFL)
+    fcntl.fcntl(child.stdout, fcntl.F_SETFL, flags | os.O_NONBLOCK)
+    poller = select.poll()
+    poller.register(child.stdout)
 
-    # flush the rest
+    # while child is running, constantly relay output
+    while child.poll() is None:
+        chunk = child.stdout.read(CHUNK_SIZE)
+        if chunk:
+            tee(chunk)
+        else:
+            # empty chunk means nothing to read
+            # wait for output on the pipe
+            # timeout is in milliseconds
+            poller.poll(1000)
+
+    # child has exited, continue relaying any remaining output
+    # At this point, read() will return an empty string when it's done
     chunk = child.stdout.read()
     while chunk:
         tee(chunk)
         chunk = child.stdout.read()
 
-    # child exited, cleanup monitor
+    # make our returncode match the child's returncode
-    try:
-        os.kill(monitor_pid, signal.SIGKILL)
-    except ProcessLookupError:
-        pass
-
-    # preserve returncode
     sys.exit(child.returncode)