This issue tracker has been migrated to GitHub, and is currently read-only.
For more information, see the GitHub FAQs in the Python's Developer Guide.

Title: os.makedirs with exist_ok=True raises PermissionError on Windows 7^
Type: behavior Stage: resolved
Components: Library (Lib), Windows Versions: Python 3.6, Python 3.4, Python 3.5
Status: closed Resolution: fixed
Dependencies: Superseder:
Assigned To: Nosy List: martin.panter, paul.moore, plakhotich, python-dev, r.david.murray, steve.dower, tim.golden, zach.ware
Priority: normal Keywords: patch

Created on 2015-11-08 14:34 by plakhotich, last changed 2022-04-11 14:58 by admin. This issue is now closed.

File name Uploaded Description Edit
makedirs-exist.patch martin.panter, 2015-11-11 03:07 review
makedirs-exist.2.patch martin.panter, 2015-11-13 22:28 review
Messages (17)
msg254339 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-08 14:34
Since Windows 7 (or even Vista), Windows gives permission error(5, ERROR_ACCESS_DENIED
if you try to create a directory in a drive root with the same name as a drive itself,
even if you have administrative permissions. This behavior is not mentioned in Microsoft docs.

Here is an example session (Windows 7, admin):
d:\>IF EXIST . echo True
d:\>mkdir .
Access is denied.
d:\>mkdir dir
d:\>cd dir
d:\dir>mkdir .
A subdirectory or file . already exists.
d:\dir>cd ..
d:\>py -3
Python 3.4.3 (v3.4.3:9b73f1c3e601, Feb 24 2015, 22:43:06) [MSC v.1600 32 bit (In
tel)] on win32
>>> import os
>>> os.path.isdir('.')
>>> os.makedirs('.', exist_ok=True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Python34\lib\", line 237, in makedirs
    mkdir(name, mode)
PermissionError: [WinError 5] ...
>>> try:
...   os.mkdir('.')
... except OSError as e:
...   print(e.errno)

This means that if you want to write portable code, you still need to write like in Python 2:
    if not os.path.isdir(path):
Which makes exist_ok useless.

The actual problem is in this line (Lib/
    if not exist_ok or e.errno != errno.EEXIST or not path.isdir(name):

Due the reasons described above, makedirs shouldn't rely on e.errno, so the right code will be:
    if not (exist_ok and path.isdir(name)):

I think the issue is pretty serious to be backported.
msg254341 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-08 15:28
Probably better solution:

    if not (exist_ok and path.isdir(name) and
            e.errno in (errno.EEXIST, errno.EACCES)):
msg254350 - (view) Author: R. David Murray (r.david.murray) * (Python committer) Date: 2015-11-08 17:37
If you are trying to create a directory named '.' your code will not do anything useful, you might as well skip the call.  What's the use case?

That said, the fix looks reasonable.
msg254355 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-08 20:16
Of course in examples I create '.' by hand, but in real code such things are
mostly automatic. Assume, for example, that you have a function for downloading

def download_file(url, save_as):
  # Before saving, you must ensure that path exists:
  os.makedirs(os.path.dirname(save_as), exist_ok=True)

os.path.abspath is not a solution, because, as it was mentioned, mkdir with
adrive name gives the same result:
d:\>mkdir d:\\
Access is denied.

Anyway, skipping the calls must be a job of exist_ok=True, otherwise it has
any sense.

By the way, do not pay attention to my second message(msg254341): as I said
at first, mkdirs shouldn't use e.errno at all. Because, as said in docstring,
makedirs must first check whether a folder exists, and only then call mkdir.
But currently it works in reverse order, which, in essence, is the source
of the problem.

I don't know why it was done that way, because if you try "timeit"
"try->mkdir->except" vs "if not isdir->mkdir", you will see that the second
is much faster (x3 (!) times on Windows, x0.7 times on Linux (on ext4 partition)
on my machine).

So the best solution is the most straightforward - to replace try-except block with:
  if not path.isdir(name):
      mkdir(name, mode)
msg254356 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-08 20:35
I meant x1.3 times faster on Linux :)
msg254358 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-08 22:05
Of course, exist_ok must be taken into account:
    if not (exist_ok and path.isdir(name)):
        mkdir(name, mode)
msg254365 - (view) Author: Martin Panter (martin.panter) * (Python committer) Date: 2015-11-09 04:05
Daniel: your latest suggestions look like they introduce a race condition. What happens if another thread or process, perhaps also calling makedirs(), creates the directory just after isdir() says it doesn’t exist? Similar to Issue 1608579.

Perhaps the existing code comment needs to clarify that the exception handling is for a real race condition, not just an excuse to “be happy” :)
msg254377 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-09 10:43
Maybe the solution is to leave OSError catching after the conditional
(probably with some additional comments):

   if not (exist_ok and path.isdir(name)):
          mkdir(name, mode)
      except OSError as e:
          if not exist_ok or e.errno != errno.EEXIST or not path.isdir(name):

This should solve the problem. It also gives more or less guarantee
that errno will be EEXIST, while the current code will also raise an exception
on EROFS (read-only FS), ENOSPC (no space left) in addition to EACCES on Windows
(and possibly some other values on various systems - who knows?).
msg254416 - (view) Author: Martin Panter (martin.panter) * (Python committer) Date: 2015-11-09 22:16
Yes that looks like an improvement, though I wonder what’s wrong with your original proposal (performance maybe?). In any case, it definitely needs a comment explaining the first isdir() avoids competing failures that mask EEXIST, and the exception handling avoids the race to create the directory.

A test case for the test suite would also be good. I understand it should be easy to do for Windows, just make a directory with an absolute path including a drive root like d:\.
msg254420 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-09 23:13
You mean msg254341? As I mentioned recently, it still will raise an exception
in case of EROFS, ENOSPC and possibly other values, which, as in the case
with Windows, can be quite unexpected depending on platform and circumstances.

Of course there is no practical sense to continue when, for example, FS is
read-only (EROFS), but I think makedirs must be predictable: it should
first check and only then try to create, as it stated in the docstring.
When exist_ok=True and directory really exists, a user doesn't expect any
exceptions from the internally used mkdir, because it simply shouldn't be
called in this case.

By the way, why 3.2 and 3.3 were removed from the list? exist_ok was introduced in 3.2.
msg254423 - (view) Author: Martin Panter (martin.panter) * (Python committer) Date: 2015-11-10 00:08
I don’t think we patch 3.2 or 3.3 any more unless it is a security concern. That is why I removed them. I understand 3.4 is due for its last non-security release in a couple weeks.

I was actually referring to your original suggestion in <>, dressed up below:

except OSError:
    # Cannot rely on checking for EEXIST, since the operating system could give priority to other errors like EACCES or EROFS
    if not (exist_ok and path.isdir(name)):

There may be practical reasons to continue if a parent directory exists on a read-only FS. Some OSes can mount writable FSes inside read-only FSes. See <> involving Cygwin, and <> involving a Linux regression.

Anyway, I think I am happy with either your last fix or the first, with an appropriate comment, and hopefully also a test case.
msg254445 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-10 13:24
Yes, it's probably a better solution. If had been more careful, I wouldn't have scribbled so much text here :) .

But is there any sense in adding Windows-specific test with EACCES since the problem with errno may affect other platforms as well (at least in theory)? It's not a Windows-only issue.
msg254471 - (view) Author: Martin Panter (martin.panter) * (Python committer) Date: 2015-11-11 03:07
It is good to add a regression test for any bug if it’s not too hard. Yes this is not a Windows-only issue, but I understand it is much simpler to produce with Windows. Otherwise you need a special file system setup and a more obscure OS.

Please review my patch. It would be good to confirm that the test fails on Windows if the fix is not applied. I have only tested this on Linux, and indirectly via Wine.
msg254629 - (view) Author: Martin Panter (martin.panter) * (Python committer) Date: 2015-11-13 22:28
New patch with simpler test case and revised NEWS entry
msg254652 - (view) Author: Daniel Plakhotich (plakhotich) Date: 2015-11-14 11:43
Looks good to me.
Without the fix, the test fails on Windows 7 as expected.
msg254949 - (view) Author: Roundup Robot (python-dev) (Python triager) Date: 2015-11-20 02:39
New changeset 05d6ddf2b7c2 by Martin Panter in branch '3.4':
Issue #25583: Avoid incorrect errors raised by os.makedirs(exist_ok=True)

New changeset 515f76bf1254 by Martin Panter in branch '3.5':
Issue #25583: Merge makedirs fix from 3.4 into 3.5

New changeset f0ad5067879b by Martin Panter in branch 'default':
Issue #25583: Merge makedirs fix from 3.5

New changeset 6ec093f78266 by Martin Panter in branch 'default':
Issue #25583: Add news to 3.6 section
msg254950 - (view) Author: Martin Panter (martin.panter) * (Python committer) Date: 2015-11-20 03:47
None of the Windows buildbots are failing my particular test. There are other failures, but they look unrelated, so I am calling this fixed.
Date User Action Args
2022-04-11 14:58:23adminsetgithub: 69769
2015-11-20 03:47:36martin.pantersetstatus: open -> closed
resolution: fixed
messages: + msg254950

stage: patch review -> resolved
2015-11-20 02:39:36python-devsetnosy: + python-dev
messages: + msg254949
2015-11-14 11:43:37plakhotichsetmessages: + msg254652
2015-11-13 22:28:29martin.pantersetfiles: + makedirs-exist.2.patch

messages: + msg254629
2015-11-11 03:07:51martin.pantersetfiles: + makedirs-exist.patch
keywords: + patch
messages: + msg254471

stage: needs patch -> patch review
2015-11-10 13:24:30plakhotichsetmessages: + msg254445
2015-11-10 00:08:15martin.pantersetmessages: + msg254423
2015-11-09 23:13:14plakhotichsetmessages: + msg254420
2015-11-09 22:16:15martin.pantersetmessages: + msg254416
2015-11-09 10:43:43plakhotichsetmessages: + msg254377
2015-11-09 04:05:48martin.pantersetversions: - Python 3.2, Python 3.3
nosy: + paul.moore, tim.golden, martin.panter, zach.ware, steve.dower

messages: + msg254365

components: + Windows
stage: needs patch
2015-11-08 22:05:05plakhotichsetmessages: + msg254358
2015-11-08 20:35:15plakhotichsetmessages: + msg254356
2015-11-08 20:16:39plakhotichsetmessages: + msg254355
2015-11-08 17:37:42r.david.murraysetnosy: + r.david.murray
messages: + msg254350
2015-11-08 15:28:19plakhotichsetmessages: + msg254341
2015-11-08 14:34:33plakhotichcreate