On FreeBSD6, pthread_kill() doesn't work on the main thread before the creation
of the first thread. Create therefore a dummy thread (no-op) a startup to
initialize the pthread library.
Add also a test for this use case, test written by Charles-François Natali.
Locks are implemented using a mutex and a condition variable of the pthread
library on FreeBSD6. POSIX condition variables cannot be interrupted by signals
(see pthread_cond_wait manual page).
Mac OS X adds __CF_USER_TEXT_ENCODING variable to an empty environment. Fix
also the test on the Py_ENABLE_SHARED config varible: test that the variable is
present, don't check it's value.
Mac OS X adds __CF_USER_TEXT_ENCODING variable to an empty environment. Fix
also the test on the Py_ENABLE_SHARED config varible: test that the variable is
present, don't check it's value.
shared mode
Try also to get more informations about the Mac OS X failure: display the keys
of the environment, instead of just the number of variables.
The previous tests used time.sleep() to synchronize two processes. If the host
was too slow, the test could fail.
The new tests only use one process, but they use a subprocess to:
- have only one thread
- have a timeout on the blocking read (select cannot be used in the test,
select always fail with EINTR, the kernel doesn't restart it)
- not touch signal handling of the parent process
Andrew agreed in the issue that eliminating the module file made sense.
Wrapper has only been exposed as a function, and so there is no (easy)
way to access the wrapper module, which in any case only had the one
function in it. Since __init__ already contains a couple wrapper
functions, it seems to make sense to just move wrapper there instead of
importing it from a single function module.
Analogous to the decode_header fix, this fix makes Header.append and
make_header correctly handle the unknown-8bit charset introduced by email5.1,
when the input to them is binary strings. Previous to this fix the
make_header(decode_header(x)) == x invariant was broken in the face of the
unknown-8bit charset.
All of the other methods in mailbox that create message objects take care to
close the file descriptors they use, so it seems to make sense to have
__getitem__ do so as well.
Patch by Filip Gruszczyński.
Victor Stinner diagnosed on #12167 that some reference leaks came from
util._path_created, a set used for caching; there are two tests that
cause additions to this set, so now they clear it in tearDown, avoiding
17 refleaks. (My tests show that it’s necessary to clear the set in
only one test, clearing it in both does not stop more refleaks, but
there’s no harm in doing it.)
AbstractHTTPHandler.do_open() of urllib.request closes the HTTP connection if
its getresponse() method fails with a socket error. Patch written by Ezio
Melotti.
AbstractHTTPHandler.do_open() of urllib.request closes the HTTP connection if
its getresponse() method fails with a socket error. Patch written by Ezio
Melotti.
It is not possible to unload a module written in C, so use a subprocess to run
the tests on the module compiled by test_build_ext(). Using a subprocess, we
don't have to unload the module, save/restore sys.path, and the test can be run
more than once.
This commit fixes also an access error on rmtree() on Windows: because the
module was not really unloaded, it was not possible to remove the temporary
directory (it is not possible to remove a directory on Windows if it still
contains an open file).
- Use different Metadata objects to write and read a PKG-INFO (METADATA)
file, to make sure the tested values come from the file
- No need to restore methods on an instance after monkey-patching them:
the methods are still the same on the class
- Harmonize dedent calls
multiprocessing: Process._bootstrap() keeps a reference to the old process to
delay its finalization until after _run_after_forkers() as been executed. This
change should fix a crash on Mac OS X Tiger when a lock is released after a
fork.
Patch written by Charles-François Nataliv and Antoine Pitrou.
multiprocessing: Process._bootstrap() keeps a reference to the old process to
delay its finalization until after _run_after_forkers() as been executed. This
change should fix a crash on Mac OS X Tiger when a lock is released after a
fork.
Patch written by Charles-François Nataliv and Antoine Pitrou.
packaging.tests.support.TempdirManager: removing the current directory is not
allowed on Windows or Solaris. Store the current directory and restore it
before removing the temporary directory (which is used as the working directory
during the tests).
The two public functions in database default to sys.path if the given
*paths* argument is None; the private functions don’t have default
values for their arguments anymore, which is fine as the public
functions that call them pass their arguments down. Likewise in
install, the functions will pass down their *paths* arguments down to
database functions.
A one-line unneeded function in install was removed instead of being
changed, and the few remaining tests that used brute-force restoration
of sys.path have been cleaned up to use sys.path.remove.
Use of DeviceIoControl to obtain the symlink path via the reparse tag was
removed. The code now uses GetFinalPathNameByHandle in the case of a
symbolic link and works properly given the added test which creates a symbolic
link and calls os.stat on it from multiple locations.
Victor Stinner also noticed an issue with os.lstat following the os.stat
code path when being passed bytes. The posix_lstat function was adjusted to
properly hook up win32_lstat instead of the previous STAT macro (win32_stat).
Just try to open files from the ZIP for reading, don't extract them to avoid
UnicodeEncodeError if the filename is not encodable to the filesystem encoding
(e.g. ASCII locale encoding).
packaging.util.check_environ will define HOME and PLAT if they don’t exist; for
some reason, it does not define PLAT when running the tests from a checkout (so
no regrtest warning) but does when running from an installed Python.
Cleaning up the envvar in test_dist fixes the warning on my machine, but I
suspect that a test runner using a different order to run files or running them
in parallel may have PLAT defined in its environment because of another test.
Quite a lot of code ends up calling check_environ; maybe we should just clean
up PLAT in every test. For now I’m doing this simple fix, we’ll see if we get
bug reports.
The code does not write checksum or file length for .pyc and .pyo in the RECORD
file, in compliance with PEP 376, but the test forgot to take .pyo into
account. This was not caught because there were no .pyo in the checkout, but
after installing there are .pyo files created by compileall, and the test picks
them up.
Rather than wrapping the C _isdir function in a Python function,
just import the C _isdir function directly. Additionally, add in the
docstring which was left out.
The example C extension used the “three.fast_taunt” name, but no “three” parent
was defined in the setup.cfg. This did not cause a failure nor even print a
warning, we may want to change that.
By changing to the Windows GetFileAttributes API in nt._isdir we can figure
out if the path is a directory without opening the file via os.stat. This has
the minor benefit of speeding up os.path.isdir by at least 2x for regular
files and 10-15x improvements were seen on symbolic links (which opened the
file multiple times during os.stat). Since os.path.isdir is used in
several places on interpreter startup, we get a minor speedup in startup time.
By changing to the Windows GetFileAttributes API in nt._isdir we can figure
out if the path is a directory without opening the file via os.stat. This has
the minor benefit of speeding up os.path.isdir by at least 2x for regular
files and 10-15x improvements were seen on symbolic links (which opened the
file multiple times during os.stat). Since os.path.isdir is used in
several places on interpreter startup, we get a minor speedup in startup time.
Benjamin Peterson
Vinay Sajip
Giampaolo Rodola'
Raymond Hettinger
Mark Dickinson
Ross Lagerwall
Victor Stinner
R David Murray
Éric Araujo
Brian Curtin
Charles-François Natali
Antoine Pitrou
Łukasz Langa