Merge remote-tracking branch 'origin/master' into 3029.install-test-plugins

.appveyor.yml

@@ -14,7 +14,7 @@ environment:
 install:
   - |
     %PYTHON%\python.exe -m pip install -U pip
-    %PYTHON%\python.exe -m pip install wheel tox virtualenv
+    %PYTHON%\python.exe -m pip install wheel tox==3.9.0 virtualenv
 
 # note:
 # %PYTHON% has: python.exe

src/allmydata/scripts/tahoe_stop.py

@@ -32,12 +32,12 @@ def stop(config):
         print("%s does not look like a running node directory (no twistd.pid)" % quoted_basedir, file=err)
         # we define rc=2 to mean "nothing is running, but it wasn't me who
         # stopped it"
-        return 2
+        return COULD_NOT_STOP
     elif pid == -1:
         print("%s contains an invalid PID file" % basedir, file=err)
         # we define rc=2 to mean "nothing is running, but it wasn't me who
         # stopped it"
-        return 2
+        return COULD_NOT_STOP
 
     # kill it hard (SIGKILL), delete the twistd.pid file, then wait for the
     # process itself to go away. If it hasn't gone away after 20 seconds, warn

src/allmydata/test/cli/test_backup.py

@@ -19,7 +19,6 @@ from .common import (
     parse_options,
 )
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
 
 def _unsupported(what):
     return "{} are not supported by Python on this platform.".format(what)
@@ -307,13 +306,6 @@ class Backup(GridTestMixin, CLITestMixin, StallMixin, unittest.TestCase):
 
         return d
 
-    # on our old dapper buildslave, this test takes a long time (usually
-    # 130s), so we have to bump up the default 120s timeout. The create-alias
-    # and initial backup alone take 60s, probably because of the handful of
-    # dirnodes being created (RSA key generation). The backup between check4
-    # and check4a takes 6s, as does the backup before check4b.
-    test_backup.timeout = 3000
-
     def _check_filtering(self, filtered, all, included, excluded):
         filtered = set(filtered)
         all = set(all)

src/allmydata/test/cli/test_check.py

@@ -12,8 +12,6 @@ from allmydata.scripts import debug
 from ..no_network import GridTestMixin
 from .common import CLITestMixin
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class Check(GridTestMixin, CLITestMixin, unittest.TestCase):
 
     def test_check(self):

src/allmydata/test/cli/test_cli.py

@@ -43,8 +43,6 @@ from twisted.python import usage
 
 from allmydata.util.encodingutil import listdir_unicode, get_io_encoding
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class CLI(CLITestMixin, unittest.TestCase):
     def _dump_cap(self, *args):
         config = debug.DumpCapOptions()

src/allmydata/test/cli/test_cp.py

@@ -14,8 +14,6 @@ from ..no_network import GridTestMixin
 from .common import CLITestMixin
 from ..common_util import skip_if_cannot_represent_filename
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class Cp(GridTestMixin, CLITestMixin, unittest.TestCase):
 
     def test_not_enough_args(self):

src/allmydata/test/cli/test_create_alias.py

@@ -9,8 +9,6 @@ from ..no_network import GridTestMixin
 from allmydata.util.encodingutil import quote_output, get_io_encoding
 from .common import CLITestMixin
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class CreateAlias(GridTestMixin, CLITestMixin, unittest.TestCase):
 
     def _test_webopen(self, args, expected_url):

src/allmydata/test/cli/test_list.py

@@ -8,8 +8,6 @@ from ..no_network import GridTestMixin
 from allmydata.util.encodingutil import quote_output, get_io_encoding
 from .common import CLITestMixin
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class List(GridTestMixin, CLITestMixin, unittest.TestCase):
     def test_list(self):
         self.basedir = "cli/List/list"

src/allmydata/test/cli/test_mv.py

@@ -5,8 +5,6 @@ from ..no_network import GridTestMixin
 from allmydata.scripts import tahoe_mv
 from .common import CLITestMixin
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class Mv(GridTestMixin, CLITestMixin, unittest.TestCase):
     def test_mv_behavior(self):
         self.basedir = "cli/Mv/mv_behavior"

src/allmydata/test/cli/test_put.py

@@ -11,8 +11,6 @@ from allmydata.util.encodingutil import get_io_encoding, unicode_to_argv
 from allmydata.util.fileutil import abspath_expanduser_unicode
 from .common import CLITestMixin
 
-timeout = 480 # deep_check takes 360s on Zandr's linksys box, others take > 240s
-
 class Put(GridTestMixin, CLITestMixin, unittest.TestCase):
 
     def test_unlinked_immutable_stdin(self):
@@ -471,4 +469,3 @@ class Put(GridTestMixin, CLITestMixin, unittest.TestCase):
                       self.failUnlessReallyEqual(rc_out_err[1], DATA))
 
         return d
-

src/allmydata/test/cli_node_api.py

@@ -0,0 +1,235 @@
+
+__all__ = [
+    "CLINodeAPI",
+    "Expect",
+    "on_stdout",
+    "on_stdout_and_stderr",
+    "on_different",
+    "wait_for_exit",
+]
+
+import os
+import sys
+from errno import ENOENT
+
+import attr
+
+from twisted.internet.error import (
+    ProcessDone,
+    ProcessTerminated,
+    ProcessExitedAlready,
+)
+from twisted.internet.interfaces import (
+    IProcessProtocol,
+)
+from twisted.python.filepath import (
+    FilePath,
+)
+from twisted.python.runtime import (
+    platform,
+)
+from twisted.internet.protocol import (
+    Protocol,
+    ProcessProtocol,
+)
+from twisted.internet.defer import (
+    Deferred,
+    succeed,
+)
+from twisted.internet.task import (
+    deferLater,
+)
+from ..client import (
+    _Client,
+)
+from ..scripts.tahoe_stop import (
+    COULD_NOT_STOP,
+)
+from ..util.eliotutil import (
+    inline_callbacks,
+)
+
+class Expect(Protocol):
+    def __init__(self):
+        self._expectations = []
+
+    def get_buffered_output(self):
+        return self._buffer
+
+    def expect(self, expectation):
+        if expectation in self._buffer:
+            return succeed(None)
+        d = Deferred()
+        self._expectations.append((expectation, d))
+        return d
+
+    def connectionMade(self):
+        self._buffer = b""
+
+    def dataReceived(self, data):
+        self._buffer += data
+        for i in range(len(self._expectations) - 1, -1, -1):
+            expectation, d = self._expectations[i]
+            if expectation in self._buffer:
+                del self._expectations[i]
+                d.callback(None)
+
+    def connectionLost(self, reason):
+        for ignored, d in self._expectations:
+            d.errback(reason)
+
+
+class _ProcessProtocolAdapter(ProcessProtocol):
+    def __init__(self, fds):
+        self._fds = fds
+
+    def connectionMade(self):
+        for proto in self._fds.values():
+            proto.makeConnection(self.transport)
+
+    def childDataReceived(self, childFD, data):
+        try:
+            proto = self._fds[childFD]
+        except KeyError:
+            pass
+        else:
+            proto.dataReceived(data)
+
+    def processEnded(self, reason):
+        notified = set()
+        for proto in self._fds.values():
+            if proto not in notified:
+                proto.connectionLost(reason)
+                notified.add(proto)
+
+
+def on_stdout(protocol):
+    return _ProcessProtocolAdapter({1: protocol})
+
+def on_stdout_and_stderr(protocol):
+    return _ProcessProtocolAdapter({1: protocol, 2: protocol})
+
+def on_different(fd_mapping):
+    return _ProcessProtocolAdapter(fd_mapping)
+
+@attr.s
+class CLINodeAPI(object):
+    reactor = attr.ib()
+    basedir = attr.ib(type=FilePath)
+    process = attr.ib(default=None)
+
+    @property
+    def twistd_pid_file(self):
+        return self.basedir.child(u"twistd.pid")
+
+    @property
+    def node_url_file(self):
+        return self.basedir.child(u"node.url")
+
+    @property
+    def storage_furl_file(self):
+        return self.basedir.child(u"private").child(u"storage.furl")
+
+    @property
+    def introducer_furl_file(self):
+        return self.basedir.child(u"private").child(u"introducer.furl")
+
+    @property
+    def config_file(self):
+        return self.basedir.child(u"tahoe.cfg")
+
+    @property
+    def exit_trigger_file(self):
+        return self.basedir.child(_Client.EXIT_TRIGGER_FILE)
+
+    def _execute(self, process_protocol, argv):
+        exe = sys.executable
+        argv = [
+            exe,
+            u"-m",
+            u"allmydata.scripts.runner",
+        ] + argv
+        return self.reactor.spawnProcess(
+            processProtocol=process_protocol,
+            executable=exe,
+            args=argv,
+            env=os.environ,
+        )
+
+    def run(self, protocol, extra_tahoe_args=()):
+        """
+        Start the node running.
+
+        :param IProcessProtocol protocol: This protocol will be hooked up to
+            the node process and can handle output or generate input.
+        """
+        if not IProcessProtocol.providedBy(protocol):
+            raise TypeError("run requires process protocol, got {}".format(protocol))
+        self.process = self._execute(
+            protocol,
+            list(extra_tahoe_args) + [u"run", self.basedir.asTextMode().path],
+        )
+        # Don't let the process run away forever.
+        try:
+            self.active()
+        except OSError as e:
+            if ENOENT != e.errno:
+                raise
+
+    def stop(self, protocol):
+        self._execute(
+            protocol,
+            [u"stop", self.basedir.asTextMode().path],
+        )
+
+    @inline_callbacks
+    def stop_and_wait(self):
+        if platform.isWindows():
+            # On Windows there is no PID file and no "tahoe stop".
+            if self.process is not None:
+                while True:
+                    try:
+                        self.process.signalProcess("TERM")
+                    except ProcessExitedAlready:
+                        break
+                    else:
+                        yield deferLater(self.reactor, 0.1, lambda: None)
+        else:
+            protocol, ended = wait_for_exit()
+            self.stop(protocol)
+            yield ended
+
+    def active(self):
+        # By writing this file, we get two minutes before the client will
+        # exit. This ensures that even if the 'stop' command doesn't work (and
+        # the test fails), the client should still terminate.
+        self.exit_trigger_file.touch()
+
+    def _check_cleanup_reason(self, reason):
+        # Let it fail because the process has already exited.
+        reason.trap(ProcessTerminated)
+        if reason.value.exitCode != COULD_NOT_STOP:
+            return reason
+        return None
+
+    def cleanup(self):
+        stopping = self.stop_and_wait()
+        stopping.addErrback(self._check_cleanup_reason)
+        return stopping
+
+
+class _WaitForEnd(ProcessProtocol):
+    def __init__(self, ended):
+        self._ended = ended
+
+    def processEnded(self, reason):
+        if reason.check(ProcessDone):
+            self._ended.callback(None)
+        else:
+            self._ended.errback(reason)
+
+
+def wait_for_exit():
+    ended = Deferred()
+    protocol = _WaitForEnd(ended)
+    return protocol, ended

src/allmydata/test/mutable/test_update.py

@@ -16,8 +16,6 @@ from .. import common_util as testutil
 SEGSIZE = 128*1024
 
 class Update(GridTestMixin, unittest.TestCase, testutil.ShouldFailMixin):
-    timeout = 400 # these tests are too big, 120s is not enough on slow
-                  # platforms
     def setUp(self):
         GridTestMixin.setUp(self)
         self.basedir = self.mktemp()

src/allmydata/test/test_checker.py

@@ -544,5 +544,3 @@ class TooParallel(GridTestMixin, unittest.TestCase):
             return res
         d.addBoth(_clean_up)
         return d
-
-    test_immutable.timeout = 80

src/allmydata/test/test_deepcheck.py

@@ -20,8 +20,6 @@ from .common_web import do_http
 from allmydata.test.no_network import GridTestMixin
 from .cli.common import CLITestMixin
 
-timeout = 2400 # One of these took 1046.091s on Zandr's ARM box.
-
 class MutableChecker(GridTestMixin, unittest.TestCase, ErrorMixin):
     def test_good(self):
         self.basedir = "deepcheck/MutableChecker/good"

src/allmydata/test/test_dirnode.py

@@ -63,7 +63,6 @@ one_nfd = u"one\u0304"
 
 class Dirnode(GridTestMixin, unittest.TestCase,
               testutil.ReallyEqualMixin, testutil.ShouldFailMixin, testutil.StallMixin, ErrorMixin):
-    timeout = 480 # It occasionally takes longer than 240 seconds on Francois's arm box.
 
     def _do_create_test(self, mdmf=False):
         c = self.g.clients[0]
@@ -1758,7 +1757,6 @@ class Dirnode2(testutil.ReallyEqualMixin, testutil.ShouldFailMixin, unittest.Tes
 
 
 class DeepStats(testutil.ReallyEqualMixin, unittest.TestCase):
-    timeout = 240 # It takes longer than 120 seconds on Francois's arm box.
     def test_stats(self):
         ds = dirnode.DeepStats(None)
         ds.add("count-files")
@@ -1823,7 +1821,6 @@ class UCWEingNodeMaker(NodeMaker):
 
 
 class Deleter(GridTestMixin, testutil.ReallyEqualMixin, unittest.TestCase):
-    timeout = 3600 # It takes longer than 433 seconds on Zandr's ARM box.
     def test_retry(self):
         # ticket #550, a dirnode.delete which experiences an
         # UncoordinatedWriteError will fail with an incorrect "you're

src/allmydata/test/test_download.py

@@ -194,7 +194,6 @@ class _Base(GridTestMixin, ShouldFailMixin):
         return d
 
 class DownloadTest(_Base, unittest.TestCase):
-    timeout = 2400 # It takes longer than 240 seconds on Zandr's ARM box.
     def test_download(self):
         self.basedir = self.mktemp()
         self.set_up_grid()

src/allmydata/test/test_encode.py

@@ -163,7 +163,6 @@ def make_data(length):
     return data[:length]
 
 class ValidatedExtendedURIProxy(unittest.TestCase):
-    timeout = 240 # It takes longer than 120 seconds on Francois's arm box.
     K = 4
     M = 10
     SIZE = 200
@@ -260,8 +259,6 @@ class ValidatedExtendedURIProxy(unittest.TestCase):
         return defer.DeferredList(dl)
 
 class Encode(unittest.TestCase):
-    timeout = 2400 # It takes longer than 240 seconds on Zandr's ARM box.
-
     def do_encode(self, max_segment_size, datalen, NUM_SHARES, NUM_SEGMENTS,
                   expected_block_hashes, expected_share_hashes):
         data = make_data(datalen)

src/allmydata/test/test_helper.py

@@ -114,7 +114,6 @@ def upload_data(uploader, data, convergence):
     return uploader.upload(u)
 
 class AssistedUpload(unittest.TestCase):
-    timeout = 240 # It takes longer than 120 seconds on Francois's arm box.
     def setUp(self):
         self.tub = t = Tub()
         t.setOption("expose-remote-exception-types", False)

src/allmydata/test/test_hung_server.py

@@ -19,15 +19,6 @@ mutable_plaintext = "muta" * 10000
 
 class HungServerDownloadTest(GridTestMixin, ShouldFailMixin, PollMixin,
                              unittest.TestCase):
-    # Many of these tests take around 60 seconds on François's ARM buildslave:
-    # http://tahoe-lafs.org/buildbot/builders/FranXois%20lenny-armv5tel
-    # allmydata.test.test_hung_server.HungServerDownloadTest.test_2_good_8_broken_duplicate_share_fail
-    # once ERRORed after 197 seconds on Midnight Magic's NetBSD buildslave:
-    # http://tahoe-lafs.org/buildbot/builders/MM%20netbsd4%20i386%20warp
-    # MM's buildslave varies a lot in how long it takes to run tests.
-
-    timeout = 240
-
     def _break(self, servers):
         for (id, ss) in servers:
             self.g.break_server(id)

src/allmydata/test/test_introducer.py

@@ -715,8 +715,6 @@ class SystemTest(SystemTestMixin, AsyncTestCase):
         self.basedir = "introducer/SystemTest/system_v2_server"
         os.makedirs(self.basedir)
         return self.do_system_test()
-    test_system_v2_server.timeout = 480
-    # occasionally takes longer than 350s on "draco"
 
 class FakeRemoteReference:
     def notifyOnDisconnect(self, *args, **kwargs): pass

src/allmydata/test/test_magic_folder.py

@@ -2418,7 +2418,6 @@ class RealTest(SingleMagicFolderTestMixin, AsyncTestCase):
 class RealTestAliceBob(MagicFolderAliceBobTestMixin, AsyncTestCase):
     """This is skipped unless both Twisted and the platform support inotify."""
     inject_inotify = False
-    timeout = 15
 
     def setUp(self):
         d = super(RealTestAliceBob, self).setUp()

src/allmydata/test/test_runner.py

@@ -4,26 +4,40 @@ from __future__ import (
 )
 
 import os.path, re, sys
+from os import linesep
 
 from twisted.trial import unittest
 
-from twisted.python import usage, runtime
-from twisted.internet.defer import inlineCallbacks, returnValue
-
+from twisted.internet import reactor
+from twisted.python import usage
+from twisted.internet.defer import (
+    inlineCallbacks,
+    returnValue,
+    DeferredList,
+)
+from twisted.python.filepath import FilePath
+from twisted.python.runtime import (
+    platform,
+)
 from allmydata.util import fileutil, pollmixin
 from allmydata.util.encodingutil import unicode_to_argv, unicode_to_output, \
     get_filesystem_encoding
-from allmydata.client import _Client
 from allmydata.test import common_util
 import allmydata
 from allmydata import __appname__
 from .common_util import parse_cli, run_cli
-
+from .cli_node_api import (
+    CLINodeAPI,
+    Expect,
+    on_stdout,
+    on_stdout_and_stderr,
+)
 from ._twisted_9607 import (
     getProcessOutputAndValue,
 )
-
-timeout = 240
+from ..util.eliotutil import (
+    inline_callbacks,
+)
 
 def get_root_from_file(src):
     srcdir = os.path.dirname(os.path.dirname(os.path.normcase(os.path.realpath(src))))
@@ -41,13 +55,7 @@ def get_root_from_file(src):
 srcfile = allmydata.__file__
 rootdir = get_root_from_file(srcfile)
 
-
 class RunBinTahoeMixin:
-    def skip_if_cannot_daemonize(self):
-        if runtime.platformType == "win32":
-            # twistd on windows doesn't daemonize. cygwin should work normally.
-            raise unittest.SkipTest("twistd does not fork under windows")
-
     @inlineCallbacks
     def find_import_location(self):
         res = yield self.run_bintahoe(["--version-and-path"])
@@ -364,357 +372,283 @@ class CreateNode(unittest.TestCase):
         # can't provide all three
         _test("create-stats-gatherer --hostname=foo --location=foo --port=foo D")
 
+
 class RunNode(common_util.SignalMixin, unittest.TestCase, pollmixin.PollMixin,
               RunBinTahoeMixin):
-    # exercise "tahoe start", for both introducer, client node, and
-    # key-generator, by spawning "tahoe start" as a subprocess. This doesn't
-    # get us figleaf-based line-level coverage, but it does a better job of
-    # confirming that the user can actually run "./bin/tahoe start" and
-    # expect it to work. This verifies that bin/tahoe sets up PYTHONPATH and
-    # the like correctly.
+    """
+    exercise "tahoe run" for both introducer, client node, and key-generator,
+    by spawning "tahoe run" (or "tahoe start") as a subprocess. This doesn't
+    get us line-level coverage, but it does a better job of confirming that
+    the user can actually run "./bin/tahoe run" and expect it to work. This
+    verifies that bin/tahoe sets up PYTHONPATH and the like correctly.
 
-    # This doesn't work on cygwin (it hangs forever), so we skip this test
-    # when we're on cygwin. It is likely that "tahoe start" itself doesn't
-    # work on cygwin: twisted seems unable to provide a version of
-    # spawnProcess which really works there.
+    This doesn't work on cygwin (it hangs forever), so we skip this test
+    when we're on cygwin. It is likely that "tahoe start" itself doesn't
+    work on cygwin: twisted seems unable to provide a version of
+    spawnProcess which really works there.
+    """
 
     def workdir(self, name):
         basedir = os.path.join("test_runner", "RunNode", name)
         fileutil.make_dirs(basedir)
         return basedir
 
+    @inline_callbacks
     def test_introducer(self):
-        self.skip_if_cannot_daemonize()
-
+        """
+        The introducer furl is stable across restarts.
+        """
         basedir = self.workdir("test_introducer")
         c1 = os.path.join(basedir, "c1")
-        exit_trigger_file = os.path.join(c1, _Client.EXIT_TRIGGER_FILE)
-        twistd_pid_file = os.path.join(c1, "twistd.pid")
-        introducer_furl_file = os.path.join(c1, "private", "introducer.furl")
-        node_url_file = os.path.join(c1, "node.url")
-        config_file = os.path.join(c1, "tahoe.cfg")
+        tahoe = CLINodeAPI(reactor, FilePath(c1))
+        self.addCleanup(tahoe.stop_and_wait)
 
-        d = self.run_bintahoe(["--quiet", "create-introducer", "--basedir", c1, "--hostname", "localhost"])
-        def _cb(res):
-            out, err, rc_or_sig = res
-            self.failUnlessEqual(rc_or_sig, 0)
+        out, err, rc_or_sig = yield self.run_bintahoe([
+            "--quiet",
+            "create-introducer",
+            "--basedir", c1,
+            "--hostname", "127.0.0.1",
+        ])
 
-            # This makes sure that node.url is written, which allows us to
-            # detect when the introducer restarts in _node_has_restarted below.
-            config = fileutil.read(config_file)
-            self.failUnlessIn('\nweb.port = \n', config)
-            fileutil.write(config_file, config.replace('\nweb.port = \n', '\nweb.port = 0\n'))
+        self.assertEqual(rc_or_sig, 0)
 
-            # by writing this file, we get ten seconds before the node will
-            # exit. This insures that even if the test fails (and the 'stop'
-            # command doesn't work), the client should still terminate.
-            fileutil.write(exit_trigger_file, "")
-            # now it's safe to start the node
-        d.addCallback(_cb)
+        # This makes sure that node.url is written, which allows us to
+        # detect when the introducer restarts in _node_has_restarted below.
+        config = fileutil.read(tahoe.config_file.path)
+        self.assertIn('{}web.port = {}'.format(linesep, linesep), config)
+        fileutil.write(
+            tahoe.config_file.path,
+            config.replace(
+                '{}web.port = {}'.format(linesep, linesep),
+                '{}web.port = 0{}'.format(linesep, linesep),
+            )
+        )
 
-        def _then_start_the_node(res):
-            return self.run_bintahoe(["--quiet", "start", c1])
-        d.addCallback(_then_start_the_node)
+        p = Expect()
+        tahoe.run(on_stdout(p))
+        yield p.expect("introducer running")
+        tahoe.active()
 
-        def _cb2(res):
-            out, err, rc_or_sig = res
+        yield self.poll(tahoe.introducer_furl_file.exists)
 
-            fileutil.write(exit_trigger_file, "")
-            errstr = "rc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr)
-            # self.failUnlessEqual(err, "", errstr) # See test_client_no_noise -- for now we ignore noise.
+        # read the introducer.furl file so we can check that the contents
+        # don't change on restart
+        furl = fileutil.read(tahoe.introducer_furl_file.path)
 
-            # the parent (twistd) has exited. However, twistd writes the pid
-            # from the child, not the parent, so we can't expect twistd.pid
-            # to exist quite yet.
+        tahoe.active()
 
-            # the node is running, but it might not have made it past the
-            # first reactor turn yet, and if we kill it too early, it won't
-            # remove the twistd.pid file. So wait until it does something
-            # that we know it won't do until after the first turn.
-        d.addCallback(_cb2)
+        # We don't keep track of PIDs in files on Windows.
+        if not platform.isWindows():
+            self.assertTrue(tahoe.twistd_pid_file.exists())
+        self.assertTrue(tahoe.node_url_file.exists())
 
-        def _node_has_started():
-            return os.path.exists(introducer_furl_file)
-        d.addCallback(lambda res: self.poll(_node_has_started))
+        # rm this so we can detect when the second incarnation is ready
+        tahoe.node_url_file.remove()
 
-        def _started(res):
-            # read the introducer.furl file so we can check that the contents
-            # don't change on restart
-            self.furl = fileutil.read(introducer_furl_file)
+        yield tahoe.stop_and_wait()
 
-            fileutil.write(exit_trigger_file, "")
-            self.failUnless(os.path.exists(twistd_pid_file))
-            self.failUnless(os.path.exists(node_url_file))
+        p = Expect()
+        tahoe.run(on_stdout(p))
+        yield p.expect("introducer running")
 
-            # rm this so we can detect when the second incarnation is ready
-            os.unlink(node_url_file)
-            return self.run_bintahoe(["--quiet", "restart", c1])
-        d.addCallback(_started)
+        # Again, the second incarnation of the node might not be ready yet, so
+        # poll until it is. This time introducer_furl_file already exists, so
+        # we check for the existence of node_url_file instead.
+        yield self.poll(tahoe.node_url_file.exists)
 
-        def _then(res):
-            out, err, rc_or_sig = res
-            fileutil.write(exit_trigger_file, "")
-            errstr = "rc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr)
-            # self.failUnlessEqual(err, "", errstr) # See test_client_no_noise -- for now we ignore noise.
-        d.addCallback(_then)
-
-        # Again, the second incarnation of the node might not be ready yet,
-        # so poll until it is. This time introducer_furl_file already
-        # exists, so we check for the existence of node_url_file instead.
-        def _node_has_restarted():
-            return os.path.exists(node_url_file)
-        d.addCallback(lambda res: self.poll(_node_has_restarted))
-
-        def _check_same_furl(res):
-            self.failUnless(os.path.exists(introducer_furl_file))
-            self.failUnlessEqual(self.furl, fileutil.read(introducer_furl_file))
-        d.addCallback(_check_same_furl)
-
-        # Now we can kill it. TODO: On a slow machine, the node might kill
-        # itself before we get a chance to, especially if spawning the
-        # 'tahoe stop' command takes a while.
-        def _stop(res):
-            fileutil.write(exit_trigger_file, "")
-            self.failUnless(os.path.exists(twistd_pid_file))
-
-            return self.run_bintahoe(["--quiet", "stop", c1])
-        d.addCallback(_stop)
-
-        def _after_stopping(res):
-            out, err, rc_or_sig = res
-            fileutil.write(exit_trigger_file, "")
-            # the parent has exited by now
-            errstr = "rc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr)
-            # self.failUnlessEqual(err, "", errstr) # See test_client_no_noise -- for now we ignore noise.
-            # the parent was supposed to poll and wait until it sees
-            # twistd.pid go away before it exits, so twistd.pid should be
-            # gone by now.
-            self.failIf(os.path.exists(twistd_pid_file))
-        d.addCallback(_after_stopping)
-        d.addBoth(self._remove, exit_trigger_file)
-        return d
-    # This test has hit a 240-second timeout on our feisty2.5 buildslave, and a 480-second timeout
-    # on Francois's Lenny-armv5tel buildslave.
-    test_introducer.timeout = 960
-
-    def test_client_no_noise(self):
-        self.skip_if_cannot_daemonize()
-
-        basedir = self.workdir("test_client_no_noise")
-        c1 = os.path.join(basedir, "c1")
-        exit_trigger_file = os.path.join(c1, _Client.EXIT_TRIGGER_FILE)
-        twistd_pid_file = os.path.join(c1, "twistd.pid")
-        node_url_file = os.path.join(c1, "node.url")
-
-        d = self.run_bintahoe(["--quiet", "create-client", "--basedir", c1, "--webport", "0"])
-        def _cb(res):
-            out, err, rc_or_sig = res
-            errstr = "cc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            assert rc_or_sig == 0, errstr
-            self.failUnlessEqual(rc_or_sig, 0)
-
-            # By writing this file, we get two minutes before the client will exit. This ensures
-            # that even if the 'stop' command doesn't work (and the test fails), the client should
-            # still terminate.
-            fileutil.write(exit_trigger_file, "")
-            # now it's safe to start the node
-        d.addCallback(_cb)
-
-        def _start(res):
-            return self.run_bintahoe(["--quiet", "start", c1])
-        d.addCallback(_start)
-
-        def _cb2(res):
-            out, err, rc_or_sig = res
-            errstr = "cc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            fileutil.write(exit_trigger_file, "")
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr) # If you emit noise, you fail this test.
-            errlines = err.split("\n")
-            self.failIf([True for line in errlines if (line != "" and "UserWarning: Unbuilt egg for setuptools" not in line
-                                                                  and "from pkg_resources import load_entry_point" not in line)], errstr)
-            if err != "":
-                raise unittest.SkipTest("This test is known not to pass on Ubuntu Lucid; see #1235.")
-
-            # the parent (twistd) has exited. However, twistd writes the pid
-            # from the child, not the parent, so we can't expect twistd.pid
-            # to exist quite yet.
-
-            # the node is running, but it might not have made it past the
-            # first reactor turn yet, and if we kill it too early, it won't
-            # remove the twistd.pid file. So wait until it does something
-            # that we know it won't do until after the first turn.
-        d.addCallback(_cb2)
-
-        def _node_has_started():
-            return os.path.exists(node_url_file)
-        d.addCallback(lambda res: self.poll(_node_has_started))
-
-        # now we can kill it. TODO: On a slow machine, the node might kill
-        # itself before we get a chance to, especially if spawning the
-        # 'tahoe stop' command takes a while.
-        def _stop(res):
-            self.failUnless(os.path.exists(twistd_pid_file),
-                            (twistd_pid_file, os.listdir(os.path.dirname(twistd_pid_file))))
-            return self.run_bintahoe(["--quiet", "stop", c1])
-        d.addCallback(_stop)
-        d.addBoth(self._remove, exit_trigger_file)
-        return d
+        # The point of this test!  After starting the second time the
+        # introducer furl file must exist and contain the same contents as it
+        # did before.
+        self.assertTrue(tahoe.introducer_furl_file.exists())
+        self.assertEqual(furl, fileutil.read(tahoe.introducer_furl_file.path))
 
+    @inline_callbacks
     def test_client(self):
-        self.skip_if_cannot_daemonize()
+        """
+        Test many things.
 
+        0) Verify that "tahoe create-node" takes a --webport option and writes
+           the value to the configuration file.
+
+        1) Verify that "tahoe run" writes a pid file and a node url file (on POSIX).
+
+        2) Verify that the storage furl file has a stable value across a
+           "tahoe run" / "tahoe stop" / "tahoe run" sequence.
+
+        3) Verify that the pid file is removed after "tahoe stop" succeeds (on POSIX).
+        """
         basedir = self.workdir("test_client")
         c1 = os.path.join(basedir, "c1")
-        exit_trigger_file = os.path.join(c1, _Client.EXIT_TRIGGER_FILE)
-        twistd_pid_file = os.path.join(c1, "twistd.pid")
-        node_url_file = os.path.join(c1, "node.url")
-        storage_furl_file = os.path.join(c1, "private", "storage.furl")
-        config_file = os.path.join(c1, "tahoe.cfg")
 
-        d = self.run_bintahoe(["--quiet", "create-node", "--basedir", c1,
-                               "--webport", "0",
-                               "--hostname", "localhost"])
-        def _cb(res):
-            out, err, rc_or_sig = res
-            self.failUnlessEqual(rc_or_sig, 0)
+        tahoe = CLINodeAPI(reactor, FilePath(c1))
+        # Set this up right now so we don't forget later.
+        self.addCleanup(tahoe.cleanup)
 
-            # Check that the --webport option worked.
-            config = fileutil.read(config_file)
-            self.failUnlessIn('\nweb.port = 0\n', config)
+        out, err, rc_or_sig = yield self.run_bintahoe([
+            "--quiet", "create-node", "--basedir", c1,
+            "--webport", "0",
+            "--hostname", "localhost",
+        ])
+        self.failUnlessEqual(rc_or_sig, 0)
 
-            # By writing this file, we get two minutes before the client will
-            # exit. This ensures that even if the 'stop' command doesn't work
-            # (and the test fails), the client should still terminate.
-            fileutil.write(exit_trigger_file, "")
-            # now it's safe to start the node
-        d.addCallback(_cb)
+        # Check that the --webport option worked.
+        config = fileutil.read(tahoe.config_file.path)
+        self.assertIn(
+            '{}web.port = 0{}'.format(linesep, linesep),
+            config,
+        )
 
-        def _start(res):
-            return self.run_bintahoe(["--quiet", "start", c1])
-        d.addCallback(_start)
+        # After this it's safe to start the node
+        tahoe.active()
 
-        def _cb2(res):
-            out, err, rc_or_sig = res
-            fileutil.write(exit_trigger_file, "")
-            errstr = "rc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr)
-            # self.failUnlessEqual(err, "", errstr) # See test_client_no_noise -- for now we ignore noise.
+        p = Expect()
+        # This will run until we stop it.
+        tahoe.run(on_stdout(p))
+        # Wait for startup to have proceeded to a reasonable point.
+        yield p.expect("client running")
+        tahoe.active()
 
-            # the parent (twistd) has exited. However, twistd writes the pid
-            # from the child, not the parent, so we can't expect twistd.pid
-            # to exist quite yet.
+        # read the storage.furl file so we can check that its contents don't
+        # change on restart
+        storage_furl = fileutil.read(tahoe.storage_furl_file.path)
 
-            # the node is running, but it might not have made it past the
-            # first reactor turn yet, and if we kill it too early, it won't
-            # remove the twistd.pid file. So wait until it does something
-            # that we know it won't do until after the first turn.
-        d.addCallback(_cb2)
+        # We don't keep track of PIDs in files on Windows.
+        if not platform.isWindows():
+            self.assertTrue(tahoe.twistd_pid_file.exists())
 
-        def _node_has_started():
-            return os.path.exists(node_url_file)
-        d.addCallback(lambda res: self.poll(_node_has_started))
+        # rm this so we can detect when the second incarnation is ready
+        tahoe.node_url_file.remove()
+        yield tahoe.stop_and_wait()
 
-        def _started(res):
-            # read the storage.furl file so we can check that its contents
-            # don't change on restart
-            self.storage_furl = fileutil.read(storage_furl_file)
+        p = Expect()
+        # We don't have to add another cleanup for this one, the one from
+        # above is still registered.
+        tahoe.run(on_stdout(p))
+        yield p.expect("client running")
+        tahoe.active()
 
-            fileutil.write(exit_trigger_file, "")
-            self.failUnless(os.path.exists(twistd_pid_file))
+        self.assertEqual(
+            storage_furl,
+            fileutil.read(tahoe.storage_furl_file.path),
+        )
 
-            # rm this so we can detect when the second incarnation is ready
-            os.unlink(node_url_file)
-            return self.run_bintahoe(["--quiet", "restart", c1])
-        d.addCallback(_started)
+        if not platform.isWindows():
+            self.assertTrue(
+                tahoe.twistd_pid_file.exists(),
+                "PID file ({}) didn't exist when we expected it to.  "
+                "These exist: {}".format(
+                    tahoe.twistd_pid_file,
+                    tahoe.twistd_pid_file.parent().listdir(),
+                ),
+            )
+        yield tahoe.stop_and_wait()
 
-        def _cb3(res):
-            out, err, rc_or_sig = res
+        if not platform.isWindows():
+            # twistd.pid should be gone by now.
+            self.assertFalse(tahoe.twistd_pid_file.exists())
 
-            fileutil.write(exit_trigger_file, "")
-            errstr = "rc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr)
-            # self.failUnlessEqual(err, "", errstr) # See test_client_no_noise -- for now we ignore noise.
-        d.addCallback(_cb3)
-
-        # again, the second incarnation of the node might not be ready yet,
-        # so poll until it is
-        d.addCallback(lambda res: self.poll(_node_has_started))
-
-        def _check_same_furl(res):
-            self.failUnlessEqual(self.storage_furl,
-                                 fileutil.read(storage_furl_file))
-        d.addCallback(_check_same_furl)
-
-        # now we can kill it. TODO: On a slow machine, the node might kill
-        # itself before we get a chance to, especially if spawning the
-        # 'tahoe stop' command takes a while.
-        def _stop(res):
-            fileutil.write(exit_trigger_file, "")
-            self.failUnless(os.path.exists(twistd_pid_file),
-                            (twistd_pid_file, os.listdir(os.path.dirname(twistd_pid_file))))
-            return self.run_bintahoe(["--quiet", "stop", c1])
-        d.addCallback(_stop)
-
-        def _cb4(res):
-            out, err, rc_or_sig = res
-
-            fileutil.write(exit_trigger_file, "")
-            # the parent has exited by now
-            errstr = "rc=%d, OUT: '%s', ERR: '%s'" % (rc_or_sig, out, err)
-            self.failUnlessEqual(rc_or_sig, 0, errstr)
-            self.failUnlessEqual(out, "", errstr)
-            # self.failUnlessEqual(err, "", errstr) # See test_client_no_noise -- for now we ignore noise.
-            # the parent was supposed to poll and wait until it sees
-            # twistd.pid go away before it exits, so twistd.pid should be
-            # gone by now.
-            self.failIf(os.path.exists(twistd_pid_file))
-        d.addCallback(_cb4)
-        d.addBoth(self._remove, exit_trigger_file)
-        return d
 
     def _remove(self, res, file):
         fileutil.remove(file)
         return res
 
-    def test_baddir(self):
-        self.skip_if_cannot_daemonize()
-        basedir = self.workdir("test_baddir")
+    def test_run_bad_directory(self):
+        """
+        If ``tahoe run`` is pointed at a non-node directory, it reports an error
+        and exits.
+        """
+        return self._bad_directory_test(
+            u"test_run_bad_directory",
+            "tahoe run",
+            lambda tahoe, p: tahoe.run(p),
+            "is not a recognizable node directory",
+        )
+
+    def test_run_bogus_directory(self):
+        """
+        If ``tahoe run`` is pointed at a non-directory, it reports an error and
+        exits.
+        """
+        return self._bad_directory_test(
+            u"test_run_bogus_directory",
+            "tahoe run",
+            lambda tahoe, p: CLINodeAPI(
+                tahoe.reactor,
+                tahoe.basedir.sibling(u"bogus"),
+            ).run(p),
+            "does not look like a directory at all"
+        )
+
+    def test_stop_bad_directory(self):
+        """
+        If ``tahoe run`` is pointed at a directory where no node is running, it
+        reports an error and exits.
+        """
+        return self._bad_directory_test(
+            u"test_stop_bad_directory",
+            "tahoe stop",
+            lambda tahoe, p: tahoe.stop(p),
+            "does not look like a running node directory",
+        )
+
+    @inline_callbacks
+    def _bad_directory_test(self, workdir, description, operation, expected_message):
+        """
+        Verify that a certain ``tahoe`` CLI operation produces a certain expected
+        message and then exits.
+
+        :param unicode workdir: A distinct path name for this test to operate
+            on.
+
+        :param unicode description: A description of the operation being
+            performed.
+
+        :param operation: A two-argument callable implementing the operation.
+            The first argument is a ``CLINodeAPI`` instance to use to perform
+            the operation.  The second argument is an ``IProcessProtocol`` to
+            which the operations output must be delivered.
+
+        :param unicode expected_message: Some text that is expected in the
+            stdout or stderr of the operation in the successful case.
+
+        :return: A ``Deferred`` that fires when the assertions have been made.
+        """
+        basedir = self.workdir(workdir)
         fileutil.make_dirs(basedir)
 
-        d = self.run_bintahoe(["--quiet", "start", "--basedir", basedir])
-        def _cb(res):
-            out, err, rc_or_sig = res
-            self.failUnlessEqual(rc_or_sig, 1)
-            self.failUnless("is not a recognizable node directory" in err, err)
-        d.addCallback(_cb)
+        tahoe = CLINodeAPI(reactor, FilePath(basedir))
+        # If tahoe ends up thinking it should keep running, make sure it stops
+        # promptly when the test is done.
+        self.addCleanup(tahoe.cleanup)
 
-        def _then_stop_it(res):
-            return self.run_bintahoe(["--quiet", "stop", "--basedir", basedir])
-        d.addCallback(_then_stop_it)
+        p = Expect()
+        operation(tahoe, on_stdout_and_stderr(p))
 
-        def _cb2(res):
-            out, err, rc_or_sig = res
-            self.failUnlessEqual(rc_or_sig, 2)
-            self.failUnless("does not look like a running node directory" in err)
-        d.addCallback(_cb2)
+        client_running = p.expect(b"client running")
 
-        def _then_start_in_bogus_basedir(res):
-            not_a_dir = os.path.join(basedir, "bogus")
-            return self.run_bintahoe(["--quiet", "start", "--basedir", not_a_dir])
-        d.addCallback(_then_start_in_bogus_basedir)
+        result, index = yield DeferredList([
+            p.expect(expected_message),
+            client_running,
+        ], fireOnOneCallback=True, consumeErrors=True,
+        )
 
-        def _cb3(res):
-            out, err, rc_or_sig = res
-            self.failUnlessEqual(rc_or_sig, 1)
-            self.failUnlessIn("does not look like a directory at all", err)
-        d.addCallback(_cb3)
-        return d
+        self.assertEqual(
+            index,
+            0,
+            "Expected error message from '{}', got something else: {}".format(
+                description,
+                p.get_buffered_output(),
+            ),
+        )
+
+        if not platform.isWindows():
+            # It should not be running.
+            self.assertFalse(tahoe.twistd_pid_file.exists())
+
+        # Wait for the operation to *complete*.  If we got this far it's
+        # because we got the expected message so we can expect the "tahoe ..."
+        # child process to exit very soon.  This other Deferred will fail when
+        # it eventually does but DeferredList above will consume the error.
+        # What's left is a perfect indicator that the process has exited and
+        # we won't get blamed for leaving the reactor dirty.
+        yield client_running

src/allmydata/test/test_sftp.py

@@ -35,8 +35,6 @@ from allmydata.test.no_network import GridTestMixin
 from allmydata.test.common import ShouldFailMixin
 from allmydata.test.common_util import ReallyEqualMixin
 
-timeout = 240
-
 class Handler(GridTestMixin, ShouldFailMixin, ReallyEqualMixin, unittest.TestCase):
     """This is a no-network unit test of the SFTPUserHandler and the abstractions it uses."""
 

src/allmydata/test/test_system.py

@@ -421,12 +421,6 @@ def _render_section_values(values):
 
 class SystemTestMixin(pollmixin.PollMixin, testutil.StallMixin):
 
-    # SystemTestMixin tests tend to be a lot of work, and we have a few
-    # buildslaves that are pretty slow, and every once in a while these tests
-    # run up against the default 120 second timeout. So increase the default
-    # timeout. Individual test cases can override this, of course.
-    timeout = 300
-
     def setUp(self):
         self.port_assigner = SameProcessStreamEndpointAssigner()
         self.port_assigner.setUp()
@@ -752,7 +746,6 @@ class CountingDataUploadable(upload.Data):
         return upload.Data.read(self, length)
 
 class SystemTest(SystemTestMixin, RunBinTahoeMixin, unittest.TestCase):
-    timeout = 3600 # It takes longer than 960 seconds on Zandr's ARM box.
 
     def test_connections(self):
         self.basedir = "system/SystemTest/test_connections"

src/allmydata/test/test_upload.py

@@ -27,11 +27,6 @@ MiB = 1024*1024
 def extract_uri(results):
     return results.get_uri()
 
-# Some of these took longer than 480 seconds on Zandr's arm box, but this may
-# have been due to an earlier test ERROR'ing out due to timeout, which seems
-# to screw up subsequent tests.
-timeout = 960
-
 class Uploadable(unittest.TestCase):
     def shouldEqual(self, data, expected):
         self.failUnless(isinstance(data, list))

src/allmydata/test/test_util.py

@@ -1060,7 +1060,6 @@ class Abbreviate(unittest.TestCase):
         self.failUnlessIn("fhtagn", str(e))
 
 class Limiter(unittest.TestCase):
-    timeout = 480 # This takes longer than 240 seconds on Francois's arm box.
 
     def job(self, i, foo):
         self.calls.append( (i, foo) )

src/allmydata/test/web/test_web.py

@@ -57,8 +57,6 @@ from ..status import FakeStatus
 # create a fake uploader/downloader, and a couple of fake dirnodes, then
 # create a webserver that works against them
 
-timeout = 480 # Most of these take longer than 240 seconds on Francois's arm box.
-
 class FakeStatsProvider:
     def get_stats(self):
         stats = {'stats': {}, 'counters': {}}

src/allmydata/util/gcutil.py

@@ -0,0 +1,63 @@
+"""
+Helpers for managing garbage collection.
+
+:ivar fileDescriptorResource: A garbage-collection-informing resource tracker
+    for file descriptors.  This is used to trigger a garbage collection when
+    it may be possible to reclaim a significant number of file descriptors as
+    a result.  Register allocation and release of *bare* file descriptors with
+    this object (file objects, socket objects, etc, have their own integration
+    with the garbage collector and don't need to bother with this).
+"""
+
+__all__ = [
+    "fileDescriptorResource",
+]
+
+import gc
+
+import attr
+
+@attr.s
+class _ResourceTracker(object):
+    """
+    Keep track of some kind of resource and trigger a full garbage collection
+    when allocations outnumber releases by some amount.
+
+    :ivar int _counter: The number of allocations that have happened in excess
+        of releases since the last full collection triggered by this tracker.
+
+    :ivar int _threshold: The number of excess allocations at which point a
+        full collection will be triggered.
+    """
+    _counter = attr.ib(default=0)
+    _threshold = attr.ib(default=25)
+
+    def allocate(self):
+        """
+        Register the allocation of an instance of this resource.
+        """
+        self._counter += 1
+        if self._counter > self._threshold:
+            gc.collect()
+            # Garbage collection of this resource has done what it can do.  If
+            # nothing was collected, it doesn't make any sense to trigger
+            # another full collection the very next time the resource is
+            # allocated.  Start the counter over again.  The next collection
+            # happens when we again exceed the threshold.
+            self._counter = 0
+
+
+    def release(self):
+        """
+        Register the release of an instance of this resource.
+        """
+        if self._counter > 0:
+            # If there were any excess allocations at this point, account for
+            # there now being one fewer.  It is not helpful to allow the
+            # counter to go below zero (as naturally would if a collection is
+            # triggered and then subsequently resources are released).  In
+            # that case, we would be operating as if we had set a higher
+            # threshold and that is not desired.
+            self._counter -= 1
+
+fileDescriptorResource = _ResourceTracker()

src/allmydata/util/iputil.py

@@ -2,6 +2,10 @@
 import os, re, socket, subprocess, errno
 from sys import platform
 
+from zope.interface import implementer
+
+import attr
+
 # from Twisted
 from twisted.python.reflect import requireModule
 from twisted.internet import defer, threads, reactor
@@ -10,7 +14,14 @@ from twisted.internet.error import CannotListenError
 from twisted.python.procutils import which
 from twisted.python import log
 from twisted.internet.endpoints import AdoptedStreamServerEndpoint
-from twisted.internet.interfaces import IReactorSocket
+from twisted.internet.interfaces import (
+    IReactorSocket,
+    IStreamServerEndpoint,
+)
+
+from .gcutil import (
+    fileDescriptorResource,
+)
 
 fcntl = requireModule("fcntl")
 
@@ -268,14 +279,23 @@ def _foolscapEndpointForPortNumber(portnum):
                 s.bind(('', 0))
                 portnum = s.getsockname()[1]
                 s.listen(1)
+                # File descriptors are a relatively scarce resource.  The
+                # cleanup process for the file descriptor we're about to dup
+                # is unfortunately complicated.  In particular, it involves
+                # the Python garbage collector.  See CleanupEndpoint for
+                # details of that.  Here, we need to make sure the garbage
+                # collector actually runs frequently enough to make a
+                # difference.  Normally, the garbage collector is triggered by
+                # allocations.  It doesn't know about *file descriptor*
+                # allocation though.  So ... we'll "teach" it about those,
+                # here.
+                fileDescriptorResource.allocate()
                 fd = os.dup(s.fileno())
                 flags = fcntl.fcntl(fd, fcntl.F_GETFD)
                 flags = flags | os.O_NONBLOCK | fcntl.FD_CLOEXEC
                 fcntl.fcntl(fd, fcntl.F_SETFD, flags)
-                return (
-                    portnum,
-                    AdoptedStreamServerEndpoint(reactor, fd, socket.AF_INET),
-                )
+                endpoint = AdoptedStreamServerEndpoint(reactor, fd, socket.AF_INET)
+                return (portnum, CleanupEndpoint(endpoint, fd))
             finally:
                 s.close()
         else:
@@ -287,6 +307,39 @@ def _foolscapEndpointForPortNumber(portnum):
     return (portnum, "tcp:%d" % (portnum,))
 
 
+@implementer(IStreamServerEndpoint)
+@attr.s
+class CleanupEndpoint(object):
+    """
+    An ``IStreamServerEndpoint`` wrapper which closes a file descriptor if the
+    wrapped endpoint is never used.
+
+    :ivar IStreamServerEndpoint _wrapped: The wrapped endpoint.  The
+        ``listen`` implementation is delegated to this object.
+
+    :ivar int _fd: The file descriptor to close if ``listen`` is never called
+        by the time this object is garbage collected.
+
+    :ivar bool _listened: A flag recording whether or not ``listen`` has been
+        called.
+    """
+    _wrapped = attr.ib()
+    _fd = attr.ib()
+    _listened = attr.ib(default=False)
+
+    def listen(self, protocolFactory):
+        self._listened = True
+        return self._wrapped.listen(protocolFactory)
+
+    def __del__(self):
+        """
+        If ``listen`` was never called then close the file descriptor.
+        """
+        if not self._listened:
+            os.close(self._fd)
+            fileDescriptorResource.release()
+
+
 def listenOnUnused(tub, portnum=None):
     """
     Start listening on an unused TCP port number with the given tub.

tox.ini

@@ -19,8 +19,24 @@ passenv = TAHOE_LAFS_* PIP_* SUBUNITREPORTER_* USERPROFILE HOMEDRIVE HOMEPATH
 # available to those systems.  Installing it ahead of time (with pip) avoids
 # this problem.
 deps =
+     # Pin all of these versions for the same reason you ever want to pin
+     # anything: to prevent new releases with regressions from introducing
+     # spurious failures into CI runs for whatever development work is
+     # happening at the time.  The versions selected here are just the current
+     # versions at the time.  Bumping them to keep up with future releases is
+     # fine as long as those releases are known to actually work.
+     pip==19.1.1
+     setuptools==41.0.1
+     wheel==0.33.4
+     subunitreporter==19.3.2
+     # As an exception, we don't pin certifi because it contains CA
+     # certificates which necessarily change over time.  Pinning this is
+     # guaranteed to cause things to break eventually as old certificates
+     # expire and as new ones are used in the wild that aren't present in
+     # whatever version we pin.  Hopefully there won't be functionality
+     # regressions in new releases of this package that cause us the kind of
+     # suffering we're trying to avoid with the above pins.
      certifi
-     subunitreporter
 
 # We add usedevelop=False because testing against a true installation gives
 # more useful results.
@@ -40,6 +56,12 @@ commands =
 [testenv:coverage]
 # coverage (with --branch) takes about 65% longer to run
 commands =
+	 # As an aid to debugging, dump all of the Python packages and their
+	 # versions that are installed in the test environment.  This is
+	 # particularly useful to get from CI runs - though hopefully the
+	 # version pinning we do limits the variability of this output
+	 # somewhat.
+         pip freeze
          tahoe --version
          coverage run --branch -m twisted.trial {env:TAHOE_LAFS_TRIAL_ARGS:--rterrors --reporter=timing} {posargs:allmydata}
          coverage xml