test_memmapping.py

import os
import mmap
import sys
import platform
import gc
import pickle
import itertools
from time import sleep
import subprocess
import threading

import pytest

from joblib.test.common import with_numpy, np
from joblib.test.common import setup_autokill
from joblib.test.common import teardown_autokill
from joblib.test.common import with_multiprocessing
from joblib.test.common import with_dev_shm
from joblib.testing import raises, parametrize, skipif, xfail, param
from joblib.backports import make_memmap
from joblib.parallel import Parallel, delayed

from joblib.pool import MemmappingPool
from joblib.executor import _TestingMemmappingExecutor as TestExecutor
from joblib._memmapping_reducer import has_shareable_memory
from joblib._memmapping_reducer import ArrayMemmapForwardReducer
from joblib._memmapping_reducer import _strided_from_memmap
from joblib._memmapping_reducer import _get_temp_dir
from joblib._memmapping_reducer import _WeakArrayKeyMap
from joblib._memmapping_reducer import _get_backing_memmap
import joblib._memmapping_reducer as jmr


def setup_module():
    setup_autokill(__name__, timeout=300)


def teardown_module():
    teardown_autokill(__name__)


def check_memmap_and_send_back(array):
    assert _get_backing_memmap(array) is not None
    return array


def check_array(args):
    """Dummy helper function to be executed in subprocesses

    Check that the provided array has the expected values in the provided
    range.

    """
    data, position, expected = args
    np.testing.assert_array_equal(data[position], expected)


def inplace_double(args):
    """Dummy helper function to be executed in subprocesses


    Check that the input array has the right values in the provided range
    and perform an inplace modification to double the values in the range by
    two.

    """
    data, position, expected = args
    assert data[position] == expected
    data[position] *= 2
    np.testing.assert_array_equal(data[position], 2 * expected)


@with_numpy
@with_multiprocessing
def test_memmap_based_array_reducing(tmpdir):
    """Check that it is possible to reduce a memmap backed array"""
    assert_array_equal = np.testing.assert_array_equal
    filename = tmpdir.join('test.mmap').strpath

    # Create a file larger than what will be used by a
    buffer = np.memmap(filename, dtype=np.float64, shape=500, mode='w+')

    # Fill the original buffer with negative markers to detect over of
    # underflow in case of test failures
    buffer[:] = - 1.0 * np.arange(buffer.shape[0], dtype=buffer.dtype)
    buffer.flush()

    # Memmap a 2D fortran array on a offsetted subsection of the previous
    # buffer
    a = np.memmap(filename, dtype=np.float64, shape=(3, 5, 4),
                  mode='r+', order='F', offset=4)
    a[:] = np.arange(60).reshape(a.shape)

    # Build various views that share the buffer with the original memmap

    # b is an memmap sliced view on an memmap instance
    b = a[1:-1, 2:-1, 2:4]

    # c and d are array views
    c = np.asarray(b)
    d = c.T

    # Array reducer with auto dumping disabled
    reducer = ArrayMemmapForwardReducer(None, tmpdir.strpath, 'c', True)

    def reconstruct_array_or_memmap(x):
        cons, args = reducer(x)
        return cons(*args)

    # Reconstruct original memmap
    a_reconstructed = reconstruct_array_or_memmap(a)
    assert has_shareable_memory(a_reconstructed)
    assert isinstance(a_reconstructed, np.memmap)
    assert_array_equal(a_reconstructed, a)

    # Reconstruct strided memmap view
    b_reconstructed = reconstruct_array_or_memmap(b)
    assert has_shareable_memory(b_reconstructed)
    assert_array_equal(b_reconstructed, b)

    # Reconstruct arrays views on memmap base
    c_reconstructed = reconstruct_array_or_memmap(c)
    assert not isinstance(c_reconstructed, np.memmap)
    assert has_shareable_memory(c_reconstructed)
    assert_array_equal(c_reconstructed, c)

    d_reconstructed = reconstruct_array_or_memmap(d)
    assert not isinstance(d_reconstructed, np.memmap)
    assert has_shareable_memory(d_reconstructed)
    assert_array_equal(d_reconstructed, d)

    # Test graceful degradation on fake memmap instances with in-memory
    # buffers
    a3 = a * 3
    assert not has_shareable_memory(a3)
    a3_reconstructed = reconstruct_array_or_memmap(a3)
    assert not has_shareable_memory(a3_reconstructed)
    assert not isinstance(a3_reconstructed, np.memmap)
    assert_array_equal(a3_reconstructed, a * 3)

    # Test graceful degradation on arrays derived from fake memmap instances
    b3 = np.asarray(a3)
    assert not has_shareable_memory(b3)

    b3_reconstructed = reconstruct_array_or_memmap(b3)
    assert isinstance(b3_reconstructed, np.ndarray)
    assert not has_shareable_memory(b3_reconstructed)
    assert_array_equal(b3_reconstructed, b3)


@with_multiprocessing
@skipif((sys.platform != "win32") or (),
        reason="PermissionError only easily triggerable on Windows")
def test_resource_tracker_retries_when_permissionerror(tmpdir):
    # Test resource_tracker retry mechanism when unlinking memmaps.  See more
    # thorough information in the ``unlink_file`` documentation of joblib.
    filename = tmpdir.join('test.mmap').strpath
    cmd = """if 1:
    import os
    import numpy as np
    import time
    from joblib.externals.loky.backend import resource_tracker
    resource_tracker.VERBOSE = 1

    # Start the resource tracker
    resource_tracker.ensure_running()
    time.sleep(1)

    # Create a file containing numpy data
    memmap = np.memmap(r"{filename}", dtype=np.float64, shape=10, mode='w+')
    memmap[:] = np.arange(10).astype(np.int8).data
    memmap.flush()
    assert os.path.exists(r"{filename}")
    del memmap

    # Create a np.memmap backed by this file
    memmap = np.memmap(r"{filename}", dtype=np.float64, shape=10, mode='w+')
    resource_tracker.register(r"{filename}", "file")

    # Ask the resource_tracker to delete the file backing the np.memmap , this
    # should raise PermissionError that the resource_tracker will log.
    resource_tracker.maybe_unlink(r"{filename}", "file")

    # Wait for the resource_tracker to process the maybe_unlink before cleaning
    # up the memmap
    time.sleep(2)
    """.format(filename=filename)
    p = subprocess.Popen([sys.executable, '-c', cmd], stderr=subprocess.PIPE,
                         stdout=subprocess.PIPE)
    p.wait()
    out, err = p.communicate()
    assert p.returncode == 0
    assert out == b''
    msg = 'tried to unlink {}, got PermissionError'.format(filename)
    assert msg in err.decode()


@with_numpy
@with_multiprocessing
def test_high_dimension_memmap_array_reducing(tmpdir):
    assert_array_equal = np.testing.assert_array_equal

    filename = tmpdir.join('test.mmap').strpath

    # Create a high dimensional memmap
    a = np.memmap(filename, dtype=np.float64, shape=(100, 15, 15, 3),
                  mode='w+')
    a[:] = np.arange(100 * 15 * 15 * 3).reshape(a.shape)

    # Create some slices/indices at various dimensions
    b = a[0:10]
    c = a[:, 5:10]
    d = a[:, :, :, 0]
    e = a[1:3:4]

    # Array reducer with auto dumping disabled
    reducer = ArrayMemmapForwardReducer(None, tmpdir.strpath, 'c', True)

    def reconstruct_array_or_memmap(x):
        cons, args = reducer(x)
        return cons(*args)

    a_reconstructed = reconstruct_array_or_memmap(a)
    assert has_shareable_memory(a_reconstructed)
    assert isinstance(a_reconstructed, np.memmap)
    assert_array_equal(a_reconstructed, a)

    b_reconstructed = reconstruct_array_or_memmap(b)
    assert has_shareable_memory(b_reconstructed)
    assert_array_equal(b_reconstructed, b)

    c_reconstructed = reconstruct_array_or_memmap(c)
    assert has_shareable_memory(c_reconstructed)
    assert_array_equal(c_reconstructed, c)

    d_reconstructed = reconstruct_array_or_memmap(d)
    assert has_shareable_memory(d_reconstructed)
    assert_array_equal(d_reconstructed, d)

    e_reconstructed = reconstruct_array_or_memmap(e)
    assert has_shareable_memory(e_reconstructed)
    assert_array_equal(e_reconstructed, e)


@with_numpy
def test__strided_from_memmap(tmpdir):
    fname = tmpdir.join('test.mmap').strpath
    size = 5 * mmap.ALLOCATIONGRANULARITY
    offset = mmap.ALLOCATIONGRANULARITY + 1
    # This line creates the mmap file that is reused later
    memmap_obj = np.memmap(fname, mode='w+', shape=size + offset)
    # filename, dtype, mode, offset, order, shape, strides, total_buffer_len
    memmap_obj = _strided_from_memmap(fname, dtype='uint8', mode='r',
                                      offset=offset, order='C', shape=size,
                                      strides=None, total_buffer_len=None,
                                      unlink_on_gc_collect=False)
    assert isinstance(memmap_obj, np.memmap)
    assert memmap_obj.offset == offset
    memmap_backed_obj = _strided_from_memmap(
        fname, dtype='uint8', mode='r', offset=offset, order='C',
        shape=(size // 2,), strides=(2,), total_buffer_len=size,
        unlink_on_gc_collect=False
    )
    assert _get_backing_memmap(memmap_backed_obj).offset == offset


@with_numpy
@with_multiprocessing
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_pool_with_memmap(factory, tmpdir):
    """Check that subprocess can access and update shared memory memmap"""
    assert_array_equal = np.testing.assert_array_equal

    # Fork the subprocess before allocating the objects to be passed
    pool_temp_folder = tmpdir.mkdir('pool').strpath
    p = factory(10, max_nbytes=2, temp_folder=pool_temp_folder)
    try:
        filename = tmpdir.join('test.mmap').strpath
        a = np.memmap(filename, dtype=np.float32, shape=(3, 5), mode='w+')
        a.fill(1.0)

        p.map(inplace_double, [(a, (i, j), 1.0)
                               for i in range(a.shape[0])
                               for j in range(a.shape[1])])

        assert_array_equal(a, 2 * np.ones(a.shape))

        # Open a copy-on-write view on the previous data
        b = np.memmap(filename, dtype=np.float32, shape=(5, 3), mode='c')

        p.map(inplace_double, [(b, (i, j), 2.0)
                               for i in range(b.shape[0])
                               for j in range(b.shape[1])])

        # Passing memmap instances to the pool should not trigger the creation
        # of new files on the FS
        assert os.listdir(pool_temp_folder) == []

        # the original data is untouched
        assert_array_equal(a, 2 * np.ones(a.shape))
        assert_array_equal(b, 2 * np.ones(b.shape))

        # readonly maps can be read but not updated
        c = np.memmap(filename, dtype=np.float32, shape=(10,), mode='r',
                      offset=5 * 4)

        with raises(AssertionError):
            p.map(check_array, [(c, i, 3.0) for i in range(c.shape[0])])

        # depending on the version of numpy one can either get a RuntimeError
        # or a ValueError
        with raises((RuntimeError, ValueError)):
            p.map(inplace_double, [(c, i, 2.0) for i in range(c.shape[0])])
    finally:
        # Clean all filehandlers held by the pool
        p.terminate()
        del p


@with_numpy
@with_multiprocessing
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_pool_with_memmap_array_view(factory, tmpdir):
    """Check that subprocess can access and update shared memory array"""
    assert_array_equal = np.testing.assert_array_equal

    # Fork the subprocess before allocating the objects to be passed
    pool_temp_folder = tmpdir.mkdir('pool').strpath
    p = factory(10, max_nbytes=2, temp_folder=pool_temp_folder)
    try:

        filename = tmpdir.join('test.mmap').strpath
        a = np.memmap(filename, dtype=np.float32, shape=(3, 5), mode='w+')
        a.fill(1.0)

        # Create an ndarray view on the memmap instance
        a_view = np.asarray(a)
        assert not isinstance(a_view, np.memmap)
        assert has_shareable_memory(a_view)

        p.map(inplace_double, [(a_view, (i, j), 1.0)
                               for i in range(a.shape[0])
                               for j in range(a.shape[1])])

        # Both a and the a_view have been updated
        assert_array_equal(a, 2 * np.ones(a.shape))
        assert_array_equal(a_view, 2 * np.ones(a.shape))

        # Passing memmap array view to the pool should not trigger the
        # creation of new files on the FS
        assert os.listdir(pool_temp_folder) == []

    finally:
        p.terminate()
        del p


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_permission_error_windows_reference_cycle(backend):
    # Non regression test for:
    # https://github.com/joblib/joblib/issues/806
    #
    # The issue happens when trying to delete a memory mapped file that has
    # not yet been closed by one of the worker processes.
    cmd = """if 1:
        import numpy as np
        from joblib import Parallel, delayed


        data = np.random.rand(int(2e6)).reshape((int(1e6), 2))

        # Build a complex cyclic reference that is likely to delay garbage
        # collection of the memmapped array in the worker processes.
        first_list = current_list = [data]
        for i in range(10):
            current_list = [current_list]
        first_list.append(current_list)

        if __name__ == "__main__":
            results = Parallel(n_jobs=2, backend="{b}")(
                delayed(len)(current_list) for i in range(10))
            assert results == [1] * 10
    """.format(b=backend)
    p = subprocess.Popen([sys.executable, '-c', cmd], stderr=subprocess.PIPE,
                         stdout=subprocess.PIPE)
    p.wait()
    out, err = p.communicate()
    assert p.returncode == 0, out.decode() + "\n\n" + err.decode()


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_permission_error_windows_memmap_sent_to_parent(backend):
    # Second non-regression test for:
    # https://github.com/joblib/joblib/issues/806
    # previously, child process would not convert temporary memmaps to numpy
    # arrays when sending the data back to the parent process. This would lead
    # to permission errors on windows when deleting joblib's temporary folder,
    # as the memmaped files handles would still opened in the parent process.
    cmd = '''if 1:
        import os
        import time

        import numpy as np

        from joblib import Parallel, delayed
        from testutils import return_slice_of_data

        data = np.ones(int(2e6))

        if __name__ == '__main__':
            # warm-up call to launch the workers and start the resource_tracker
            _ = Parallel(n_jobs=2, verbose=5, backend='{b}')(
                delayed(id)(i) for i in range(20))

            time.sleep(0.5)

            slice_of_data = Parallel(n_jobs=2, verbose=5, backend='{b}')(
                delayed(return_slice_of_data)(data, 0, 20) for _ in range(10))
    '''.format(b=backend)

    for _ in range(3):
        env = os.environ.copy()
        env['PYTHONPATH'] = os.path.dirname(__file__)
        p = subprocess.Popen([sys.executable, '-c', cmd],
                             stderr=subprocess.PIPE,
                             stdout=subprocess.PIPE, env=env)
        p.wait()
        out, err = p.communicate()
        assert p.returncode == 0, err
        assert out == b''
        if sys.version_info[:3] not in [(3, 8, 0), (3, 8, 1)]:
            # In early versions of Python 3.8, a reference leak
            # https://github.com/cloudpipe/cloudpickle/issues/327, holds
            # references to pickled objects, generating race condition during
            # cleanup finalizers of joblib and noisy resource_tracker outputs.
            assert b'resource_tracker' not in err


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_parallel_isolated_temp_folders(backend):
    # Test that consecutive Parallel call use isolated subfolders, even
    # for the loky backend that reuses its executor instance across calls.
    array = np.arange(int(1e2))
    [filename_1] = Parallel(n_jobs=2, backend=backend, max_nbytes=10)(
        delayed(getattr)(array, 'filename') for _ in range(1)
    )
    [filename_2] = Parallel(n_jobs=2, backend=backend, max_nbytes=10)(
        delayed(getattr)(array, 'filename') for _ in range(1)
    )
    assert os.path.dirname(filename_2) != os.path.dirname(filename_1)


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_managed_backend_reuse_temp_folder(backend):
    # Test that calls to a managed parallel object reuse the same memmaps.
    array = np.arange(int(1e2))
    with Parallel(n_jobs=2, backend=backend, max_nbytes=10) as p:
        [filename_1] = p(
            delayed(getattr)(array, 'filename') for _ in range(1)
        )
        [filename_2] = p(
            delayed(getattr)(array, 'filename') for _ in range(1)
        )
    assert os.path.dirname(filename_2) == os.path.dirname(filename_1)


@with_numpy
@with_multiprocessing
def test_memmapping_temp_folder_thread_safety():
    # Concurrent calls to Parallel with the loky backend will use the same
    # executor, and thus the same reducers. Make sure that those reducers use
    # different temporary folders depending on which Parallel objects called
    # them, which is necessary to limit potential race conditions during the
    # garbage collection of temporary memmaps.
    array = np.arange(int(1e2))

    temp_dirs_thread_1 = set()
    temp_dirs_thread_2 = set()

    def concurrent_get_filename(array, temp_dirs):
        with Parallel(backend='loky', n_jobs=2, max_nbytes=10) as p:
            for i in range(10):
                [filename] = p(
                    delayed(getattr)(array, 'filename') for _ in range(1)
                )
                temp_dirs.add(os.path.dirname(filename))

    t1 = threading.Thread(
        target=concurrent_get_filename, args=(array, temp_dirs_thread_1)
    )
    t2 = threading.Thread(
        target=concurrent_get_filename, args=(array, temp_dirs_thread_2)
    )

    t1.start()
    t2.start()

    t1.join()
    t2.join()

    assert len(temp_dirs_thread_1) == 1
    assert len(temp_dirs_thread_2) == 1

    assert temp_dirs_thread_1 != temp_dirs_thread_2


@with_numpy
@with_multiprocessing
def test_multithreaded_parallel_termination_resource_tracker_silent():
    # test that concurrent termination attempts of a same executor does not
    # emit any spurious error from the resource_tracker. We test various
    # situations making 0, 1 or both parallel call sending a task that will
    # make the worker (and thus the whole Parallel call) error out.
    cmd = '''if 1:
        import os
        import numpy as np
        from joblib import Parallel, delayed
        from joblib.externals.loky.backend import resource_tracker
        from concurrent.futures import ThreadPoolExecutor, wait

        resource_tracker.VERBOSE = 0

        array = np.arange(int(1e2))

        temp_dirs_thread_1 = set()
        temp_dirs_thread_2 = set()


        def raise_error(array):
            raise ValueError


        def parallel_get_filename(array, temp_dirs):
            with Parallel(backend="loky", n_jobs=2, max_nbytes=10) as p:
                for i in range(10):
                    [filename] = p(
                        delayed(getattr)(array, "filename") for _ in range(1)
                    )
                    temp_dirs.add(os.path.dirname(filename))


        def parallel_raise(array, temp_dirs):
            with Parallel(backend="loky", n_jobs=2, max_nbytes=10) as p:
                for i in range(10):
                    [filename] = p(
                        delayed(raise_error)(array) for _ in range(1)
                    )
                    temp_dirs.add(os.path.dirname(filename))


        executor = ThreadPoolExecutor(max_workers=2)

        # both function calls will use the same loky executor, but with a
        # different Parallel object.
        future_1 = executor.submit({f1}, array, temp_dirs_thread_1)
        future_2 = executor.submit({f2}, array, temp_dirs_thread_2)

        # Wait for both threads to terminate their backend
        wait([future_1, future_2])

        future_1.result()
        future_2.result()
    '''
    functions_and_returncodes = [
        ("parallel_get_filename", "parallel_get_filename", 0),
        ("parallel_get_filename", "parallel_raise", 1),
        ("parallel_raise", "parallel_raise", 1)
    ]

    for f1, f2, returncode in functions_and_returncodes:
        p = subprocess.Popen([sys.executable, '-c', cmd.format(f1=f1, f2=f2)],
                             stderr=subprocess.PIPE, stdout=subprocess.PIPE)
        p.wait()
        out, err = p.communicate()
        assert p.returncode == returncode, out.decode()
        assert b"resource_tracker" not in err, err.decode()


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_many_parallel_calls_on_same_object(backend):
    # After #966 got merged, consecutive Parallel objects were sharing temp
    # folder, which would lead to race conditions happening during the
    # temporary resources management with the resource_tracker. This is a
    # non-regression test that makes sure that consecutive Parallel operations
    # on the same object do not error out.
    cmd = '''if 1:
        import os
        import time

        import numpy as np

        from joblib import Parallel, delayed
        from testutils import return_slice_of_data

        data = np.ones(100)

        if __name__ == '__main__':
            for i in range(5):
                slice_of_data = Parallel(
                    n_jobs=2, max_nbytes=1, backend='{b}')(
                        delayed(return_slice_of_data)(data, 0, 20)
                        for _ in range(10)
                    )
    '''.format(b=backend)
    env = os.environ.copy()
    env['PYTHONPATH'] = os.path.dirname(__file__)
    p = subprocess.Popen(
        [sys.executable, '-c', cmd],
        stderr=subprocess.PIPE,
        stdout=subprocess.PIPE,
        env=env,
    )
    p.wait()
    out, err = p.communicate()
    assert p.returncode == 0, err
    assert out == b''
    if sys.version_info[:3] not in [(3, 8, 0), (3, 8, 1)]:
        # In early versions of Python 3.8, a reference leak
        # https://github.com/cloudpipe/cloudpickle/issues/327, holds
        # references to pickled objects, generating race condition during
        # cleanup finalizers of joblib and noisy resource_tracker outputs.
        assert b'resource_tracker' not in err


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_memmap_returned_as_regular_array(backend):
    data = np.ones(int(1e3))
    # Check that child processes send temporary memmaps back as numpy arrays.
    [result] = Parallel(n_jobs=2, backend=backend, max_nbytes=100)(
        delayed(check_memmap_and_send_back)(data) for _ in range(1))
    assert _get_backing_memmap(result) is None


@with_numpy
@with_multiprocessing
@parametrize("backend", ["multiprocessing", "loky"])
def test_resource_tracker_silent_when_reference_cycles(backend):
    # There is a variety of reasons that can make joblib with loky backend
    # output noisy warnings when a reference cycle is preventing a memmap from
    # being garbage collected. Especially, joblib's main process finalizer
    # deletes the temporary folder if it was not done before, which can
    # interact badly with the resource_tracker. We don't risk leaking any
    # resources, but this will likely make joblib output a lot of low-level
    # confusing messages.
    #
    # This test makes sure that the resource_tracker is silent when a reference
    # has been collected concurrently on non-Windows platforms.
    #
    # Note that the script in ``cmd`` is the exact same script as in
    # test_permission_error_windows_reference_cycle.
    if backend == "loky" and sys.platform.startswith('win'):
        # XXX: on Windows, reference cycles can delay timely garbage collection
        # and make it impossible to properly delete the temporary folder in the
        # main process because of permission errors.
        pytest.xfail(
            "The temporary folder cannot be deleted on Windows in the "
            "presence of a reference cycle"
        )

    cmd = """if 1:
        import numpy as np
        from joblib import Parallel, delayed


        data = np.random.rand(int(2e6)).reshape((int(1e6), 2))

        # Build a complex cyclic reference that is likely to delay garbage
        # collection of the memmapped array in the worker processes.
        first_list = current_list = [data]
        for i in range(10):
            current_list = [current_list]
        first_list.append(current_list)

        if __name__ == "__main__":
            results = Parallel(n_jobs=2, backend="{b}")(
                delayed(len)(current_list) for i in range(10))
            assert results == [1] * 10
    """.format(b=backend)
    p = subprocess.Popen([sys.executable, '-c', cmd], stderr=subprocess.PIPE,
                         stdout=subprocess.PIPE)
    p.wait()
    out, err = p.communicate()
    out = out.decode()
    err = err.decode()
    assert p.returncode == 0, out + "\n\n" + err
    assert "resource_tracker" not in err, err


@with_numpy
@with_multiprocessing
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_memmapping_pool_for_large_arrays(factory, tmpdir):
    """Check that large arrays are not copied in memory"""

    # Check that the tempfolder is empty
    assert os.listdir(tmpdir.strpath) == []

    # Build an array reducers that automatically dump large array content
    # to filesystem backed memmap instances to avoid memory explosion
    p = factory(3, max_nbytes=40, temp_folder=tmpdir.strpath, verbose=2)
    try:
        # The temporary folder for the pool is not provisioned in advance
        assert os.listdir(tmpdir.strpath) == []
        assert not os.path.exists(p._temp_folder)

        small = np.ones(5, dtype=np.float32)
        assert small.nbytes == 20
        p.map(check_array, [(small, i, 1.0) for i in range(small.shape[0])])

        # Memory has been copied, the pool filesystem folder is unused
        assert os.listdir(tmpdir.strpath) == []

        # Try with a file larger than the memmap threshold of 40 bytes
        large = np.ones(100, dtype=np.float64)
        assert large.nbytes == 800
        p.map(check_array, [(large, i, 1.0) for i in range(large.shape[0])])

        # The data has been dumped in a temp folder for subprocess to share it
        # without per-child memory copies
        assert os.path.isdir(p._temp_folder)
        dumped_filenames = os.listdir(p._temp_folder)
        assert len(dumped_filenames) == 1

        # Check that memory mapping is not triggered for arrays with
        # dtype='object'
        objects = np.array(['abc'] * 100, dtype='object')
        results = p.map(has_shareable_memory, [objects])
        assert not results[0]

    finally:
        # check FS garbage upon pool termination
        p.terminate()
        for i in range(10):
            sleep(.1)
            if not os.path.exists(p._temp_folder):
                break
        else:  # pragma: no cover
            raise AssertionError(
                'temporary folder of {} was not deleted'.format(p)
            )
        del p


@with_numpy
@with_multiprocessing
@parametrize(
    "backend",
    [
        pytest.param(
            "multiprocessing",
            marks=pytest.mark.xfail(
                reason='https://github.com/joblib/joblib/issues/1086'
            ),
        ),
        "loky",
    ]
)
def test_child_raises_parent_exits_cleanly(backend):
    # When a task executed by a child process raises an error, the parent
    # process's backend is notified, and calls abort_everything.
    # In loky, abort_everything itself calls shutdown(kill_workers=True) which
    # sends SIGKILL to the worker, preventing it from running the finalizers
    # supposed to signal the resource_tracker when the worker is done using
    # objects relying on a shared resource (e.g np.memmaps). Because this
    # behavior is prone to :
    # - cause a resource leak
    # - make the resource tracker emit noisy resource warnings
    # we explicitly test that, when the said situation occurs:
    # - no resources are actually leaked
    # - the temporary resources are deleted as soon as possible (typically, at
    #   the end of the failing Parallel call)
    # - the resource_tracker does not emit any warnings.
    cmd = """if 1:
        import os
        from pathlib import Path
        from time import sleep

        import numpy as np
        from joblib import Parallel, delayed
        from testutils import print_filename_and_raise

        data = np.random.rand(1000)

        def get_temp_folder(parallel_obj, backend):
            if "{b}" == "loky":
                return Path(p._backend._workers._temp_folder)
            else:
                return Path(p._backend._pool._temp_folder)


        if __name__ == "__main__":
            try:
                with Parallel(n_jobs=2, backend="{b}", max_nbytes=100) as p:
                    temp_folder = get_temp_folder(p, "{b}")
                    p(delayed(print_filename_and_raise)(data)
                              for i in range(1))
            except ValueError as e:
                # the temporary folder should be deleted by the end of this
                # call but apparently on some file systems, this takes
                # some time to be visible.
                #
                # We attempt to write into the temporary folder to test for
                # its existence and we wait for a maximum of 10 seconds.
                for i in range(100):
                    try:
                        with open(temp_folder / "some_file.txt", "w") as f:
                            f.write("some content")
                    except FileNotFoundError:
                        # temp_folder has been deleted, all is fine
                        break

                    # ... else, wait a bit and try again
                    sleep(.1)
                else:
                    raise AssertionError(
                        str(temp_folder) + " was not deleted"
                    ) from e
    """.format(b=backend)
    env = os.environ.copy()
    env['PYTHONPATH'] = os.path.dirname(__file__)
    p = subprocess.Popen([sys.executable, '-c', cmd], stderr=subprocess.PIPE,
                         stdout=subprocess.PIPE, env=env)
    p.wait()
    out, err = p.communicate()
    out, err = out.decode(), err.decode()
    filename = out.split('\n')[0]
    assert p.returncode == 0, err or out
    assert err == ''  # no resource_tracker warnings.
    assert not os.path.exists(filename)


@with_numpy
@with_multiprocessing
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_memmapping_pool_for_large_arrays_disabled(factory, tmpdir):
    """Check that large arrays memmapping can be disabled"""
    # Set max_nbytes to None to disable the auto memmapping feature
    p = factory(3, max_nbytes=None, temp_folder=tmpdir.strpath)
    try:

        # Check that the tempfolder is empty
        assert os.listdir(tmpdir.strpath) == []

        # Try with a file largish than the memmap threshold of 40 bytes
        large = np.ones(100, dtype=np.float64)
        assert large.nbytes == 800
        p.map(check_array, [(large, i, 1.0) for i in range(large.shape[0])])

        # Check that the tempfolder is still empty
        assert os.listdir(tmpdir.strpath) == []

    finally:
        # Cleanup open file descriptors
        p.terminate()
        del p


@with_numpy
@with_multiprocessing
@with_dev_shm
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_memmapping_on_large_enough_dev_shm(factory):
    """Check that memmapping uses /dev/shm when possible"""
    orig_size = jmr.SYSTEM_SHARED_MEM_FS_MIN_SIZE
    try:
        # Make joblib believe that it can use /dev/shm even when running on a
        # CI container where the size of the /dev/shm is not very large (that
        # is at least 32 MB instead of 2 GB by default).
        jmr.SYSTEM_SHARED_MEM_FS_MIN_SIZE = int(32e6)
        p = factory(3, max_nbytes=10)
        try:
            # Check that the pool has correctly detected the presence of the
            # shared memory filesystem.
            pool_temp_folder = p._temp_folder
            folder_prefix = '/dev/shm/joblib_memmapping_folder_'
            assert pool_temp_folder.startswith(folder_prefix)
            assert os.path.exists(pool_temp_folder)

            # Try with a file larger than the memmap threshold of 10 bytes
            a = np.ones(100, dtype=np.float64)
            assert a.nbytes == 800
            p.map(id, [a] * 10)
            # a should have been memmapped to the pool temp folder: the joblib
            # pickling procedure generate one .pkl file:
            assert len(os.listdir(pool_temp_folder)) == 1

            # create a new array with content that is different from 'a' so
            # that it is mapped to a different file in the temporary folder of
            # the pool.
            b = np.ones(100, dtype=np.float64) * 2
            assert b.nbytes == 800
            p.map(id, [b] * 10)
            # A copy of both a and b are now stored in the shared memory folder
            assert len(os.listdir(pool_temp_folder)) == 2
        finally:
            # Cleanup open file descriptors
            p.terminate()
            del p

        for i in range(100):
            # The temp folder is cleaned up upon pool termination
            if not os.path.exists(pool_temp_folder):
                break
            sleep(.1)
        else:  # pragma: no cover
            raise AssertionError('temporary folder of pool was not deleted')
    finally:
        jmr.SYSTEM_SHARED_MEM_FS_MIN_SIZE = orig_size


@with_numpy
@with_multiprocessing
@with_dev_shm
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_memmapping_on_too_small_dev_shm(factory):
    orig_size = jmr.SYSTEM_SHARED_MEM_FS_MIN_SIZE
    try:
        # Make joblib believe that it cannot use /dev/shm unless there is
        # 42 exabytes of available shared memory in /dev/shm
        jmr.SYSTEM_SHARED_MEM_FS_MIN_SIZE = int(42e18)

        p = factory(3, max_nbytes=10)
        try:
            # Check that the pool has correctly detected the presence of the
            # shared memory filesystem.
            pool_temp_folder = p._temp_folder
            assert not pool_temp_folder.startswith('/dev/shm')
        finally:
            # Cleanup open file descriptors
            p.terminate()
            del p

        # The temp folder is cleaned up upon pool termination
        assert not os.path.exists(pool_temp_folder)
    finally:
        jmr.SYSTEM_SHARED_MEM_FS_MIN_SIZE = orig_size


@with_numpy
@with_multiprocessing
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_memmapping_pool_for_large_arrays_in_return(factory, tmpdir):
    """Check that large arrays are not copied in memory in return"""
    assert_array_equal = np.testing.assert_array_equal

    # Build an array reducers that automatically dump large array content
    # but check that the returned datastructure are regular arrays to avoid
    # passing a memmap array pointing to a pool controlled temp folder that
    # might be confusing to the user

    # The MemmappingPool user can always return numpy.memmap object explicitly
    # to avoid memory copy
    p = factory(3, max_nbytes=10, temp_folder=tmpdir.strpath)
    try:
        res = p.apply_async(np.ones, args=(1000,))
        large = res.get()
        assert not has_shareable_memory(large)
        assert_array_equal(large, np.ones(1000))
    finally:
        p.terminate()
        del p


def _worker_multiply(a, n_times):
    """Multiplication function to be executed by subprocess"""
    assert has_shareable_memory(a)
    return a * n_times


@with_numpy
@with_multiprocessing
@parametrize("factory", [MemmappingPool, TestExecutor.get_memmapping_executor],
             ids=["multiprocessing", "loky"])
def test_workaround_against_bad_memmap_with_copied_buffers(factory, tmpdir):
    """Check that memmaps with a bad buffer are returned as regular arrays

    Unary operations and ufuncs on memmap instances return a new memmap
    instance with an in-memory buffer (probably a numpy bug).
    """
    assert_array_equal = np.testing.assert_array_equal

    p = factory(3, max_nbytes=10, temp_folder=tmpdir.strpath)
    try:
        # Send a complex, large-ish view on a array that will be converted to
        # a memmap in the worker process
        a = np.asarray(np.arange(6000).reshape((1000, 2, 3)),
                       order='F')[:, :1, :]

        # Call a non-inplace multiply operation on the worker and memmap and
        # send it back to the parent.
        b = p.apply_async(_worker_multiply, args=(a, 3)).get()
        assert not has_shareable_memory(b)
        assert_array_equal(b, 3 * a)
    finally:
        p.terminate()
        del p


def identity(arg):
    return arg


@with_numpy
@with_multiprocessing
@parametrize(
    "factory,retry_no",
    list(itertools.product(
        [MemmappingPool, TestExecutor.get_memmapping_executor], range(3))),
    ids=['{}, {}'.format(x, y) for x, y in itertools.product(
        ["multiprocessing", "loky"], map(str, range(3)))])
def test_pool_memmap_with_big_offset(factory, retry_no, tmpdir):
    # Test that numpy memmap offset is set correctly if greater than
    # mmap.ALLOCATIONGRANULARITY, see
    # https://github.com/joblib/joblib/issues/451 and
    # https://github.com/numpy/numpy/pull/8443 for more details.
    fname = tmpdir.join('test.mmap').strpath
    size = 5 * mmap.ALLOCATIONGRANULARITY
    offset = mmap.ALLOCATIONGRANULARITY + 1
    obj = make_memmap(fname, mode='w+', shape=size, dtype='uint8',
                      offset=offset)

    p = factory(2, temp_folder=tmpdir.strpath)
    result = p.apply_async(identity, args=(obj,)).get()
    assert isinstance(result, np.memmap)
    assert result.offset == offset
    np.testing.assert_array_equal(obj, result)
    p.terminate()


def test_pool_get_temp_dir(tmpdir):
    pool_folder_name = 'test.tmpdir'
    pool_folder, shared_mem = _get_temp_dir(pool_folder_name, tmpdir.strpath)
    assert shared_mem is False
    assert pool_folder == tmpdir.join('test.tmpdir').strpath

    pool_folder, shared_mem = _get_temp_dir(pool_folder_name, temp_folder=None)
    if sys.platform.startswith('win'):
        assert shared_mem is False
    assert pool_folder.endswith(pool_folder_name)


@with_numpy
@skipif(sys.platform == 'win32', reason='This test fails with a '
        'PermissionError on Windows')
@parametrize("mmap_mode", ["r+", "w+"])
def test_numpy_arrays_use_different_memory(mmap_mode):
    def func(arr, value):
        arr[:] = value
        return arr

    arrays = [np.zeros((10, 10), dtype='float64') for i in range(10)]

    results = Parallel(mmap_mode=mmap_mode, max_nbytes=0, n_jobs=2)(
        delayed(func)(arr, i) for i, arr in enumerate(arrays))

    for i, arr in enumerate(results):
        np.testing.assert_array_equal(arr, i)


@with_numpy
def test_weak_array_key_map():

    def assert_empty_after_gc_collect(container, retries=100):
        for i in range(retries):
            if len(container) == 0:
                return
            gc.collect()
            sleep(.1)
        assert len(container) == 0

    a = np.ones(42)
    m = _WeakArrayKeyMap()
    m.set(a, 'a')
    assert m.get(a) == 'a'

    b = a
    assert m.get(b) == 'a'
    m.set(b, 'b')
    assert m.get(a) == 'b'

    del a
    gc.collect()
    assert len(m._data) == 1
    assert m.get(b) == 'b'

    del b
    assert_empty_after_gc_collect(m._data)

    c = np.ones(42)
    m.set(c, 'c')
    assert len(m._data) == 1
    assert m.get(c) == 'c'

    with raises(KeyError):
        m.get(np.ones(42))

    del c
    assert_empty_after_gc_collect(m._data)

    # Check that creating and dropping numpy arrays with potentially the same
    # object id will not cause the map to get confused.
    def get_set_get_collect(m, i):
        a = np.ones(42)
        with raises(KeyError):
            m.get(a)
        m.set(a, i)
        assert m.get(a) == i
        return id(a)

    unique_ids = set([get_set_get_collect(m, i) for i in range(1000)])
    if platform.python_implementation() == 'CPython':
        # On CPython (at least) the same id is often reused many times for the
        # temporary arrays created under the local scope of the
        # get_set_get_collect function without causing any spurious lookups /
        # insertions in the map.
        assert len(unique_ids) < 100


def test_weak_array_key_map_no_pickling():
    m = _WeakArrayKeyMap()
    with raises(pickle.PicklingError):
        pickle.dumps(m)


@with_numpy
@with_multiprocessing
def test_direct_mmap(tmpdir):
    testfile = str(tmpdir.join('arr.dat'))
    a = np.arange(10, dtype='uint8')
    a.tofile(testfile)

    def _read_array():
        with open(testfile) as fd:
            mm = mmap.mmap(fd.fileno(), 0, access=mmap.ACCESS_READ, offset=0)
        return np.ndarray((10,), dtype=np.uint8, buffer=mm, offset=0)

    def func(x):
        return x**2

    arr = _read_array()

    # this is expected to work and gives the reference
    ref = Parallel(n_jobs=2)(delayed(func)(x) for x in [a])

    # now test that it work with the mmap array
    results = Parallel(n_jobs=2)(delayed(func)(x) for x in [arr])
    np.testing.assert_array_equal(results, ref)

    # also test with a mmap array read in the subprocess
    def worker():
        return _read_array()

    results = Parallel(n_jobs=2)(delayed(worker)() for _ in range(1))
    np.testing.assert_array_equal(results[0], arr)