classification
Title: sys._enablelegacywindowsfsencoding() don't apply to os.fsencode and os.fsdecode
Type: behavior Stage:
Components: Windows Versions: Python 3.7, Python 3.6
process
Status: open Resolution:
Dependencies: Superseder:
Assigned To: Nosy List: JGoutin, lemburg, paul.moore, steve.dower, tim.golden, vstinner, zach.ware
Priority: normal Keywords:

Created on 2017-01-11 12:28 by JGoutin, last changed 2018-02-19 23:01 by steve.dower.

Messages (16)
msg285220 - (view) Author: (JGoutin) * Date: 2017-01-11 12:28
The doc say that calling "sys._enablelegacywindowsfsencoding()" is equivalent to use "PYTHONLEGACYWINDOWSFSENCODING" environment variable.

In fact, this no apply to "os.fsencode" and "os.fsdecode".

Example with Python 3.6 64Bits on Windows 7 64 bits :


EXAMPLE CODE 1 (sys._enablelegacywindowsfsencoding()): 

import sys
import os

# Force the use of legacy encoding like versions of Python prior to 3.6.
sys._enablelegacywindowsfsencoding()

# Show actual file system encoding
encoding = sys.getfilesystemencoding()
print('File system encoding:', encoding)

# os.fsencode(filename) VS filename.encode(File system encoding)
print(os.fsencode('é'), 'é'.encode(encoding))

>>> File system encoding: mbcs
>>> b'\xc3\xa9' b'\xe9'


First is encoded with "utf-8" and not "mbcs" (The actual File system encoding)


EXAMPLE CODE 2 (PYTHONLEGACYWINDOWSFSENCODING): 

import sys
import os

# Force the use of legacy encoding like versions of Python prior to 3.6.
# "PYTHONLEGACYWINDOWSFSENCODING" environment variable set before running Python.

# Show actual file system encoding
encoding = sys.getfilesystemencoding()
print('File system encoding:', encoding)

# os.fsencode(filename) VS filename.encode(File system encoding)
print(os.fsencode('é'), 'é'.encode(encoding))

>>> File system encoding: mbcs
>>> b'\xe9' b'\xe9'


Everything encoded with "mbcs" (The actual File system encoding)


In "os.fsencode" and "os.fsdecode" encoding and errors are cached on start and never updated by "sys._enablelegacywindowsfsencoding()" after.
msg285238 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2017-01-11 16:43
If you import os first then that's acceptable and we should document it more clearly. Try calling enable before importing os.

I wouldn't be surprised if os is imported automatically, in which case we need to figure out some alternate caching mechanism that can be reset by the enable call (or demonstrate that the caching is of no benefit).
msg285290 - (view) Author: (JGoutin) * Date: 2017-01-12 07:07
import sys

# Force the use of legacy encoding like versions of Python prior to 3.6.
sys._enablelegacywindowsfsencoding()

# Show actual file system encoding
encoding = sys.getfilesystemencoding()
print('File system encoding:', encoding)

# os.fsencode(filename) VS filename.encode(File system encoding)
import os
print(os.fsencode('é'), 'é'.encode(encoding))

>>> File system encoding: mbcs
>>> b'\xc3\xa9' b'\xe9'


The result is the same.
msg285354 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2017-01-12 23:49
Then we do in fact need to make os.fsencode/fsdecode either stop caching the encoding completely, or figure out a way to reset the cache when that function is called.
msg285374 - (view) Author: Marc-Andre Lemburg (lemburg) * (Python committer) Date: 2017-01-13 10:03
Adding Victor, who implemented the fs codec.

AFAIK, it's not possible to change the encoding after interpreter initialization, since it will have been already used for many different things by the time you get to executing code.
msg285393 - (view) Author: STINNER Victor (vstinner) * (Python committer) Date: 2017-01-13 12:46
My experience with changing the Python "filesystem encoding" (sys.getfilesystemencoding()) at runtime: it doesn't work.

The filesystem encoding must be set as soon as possible and must never change later. As soon as possible: before the first call to os.fsdecode(), which is implemented in C as Py_DecodeLocale(). For example, the encoding must be set before Python imports the first module.

The filesystem encoding must be set before Python decodes *any* operating system data: command line arguments, any filename or path, environment variables, etc.

Hopefully, Windows provides most operating system data as Unicode directly: command line arguments and environment variables are exposed as Unicode for example.

os.fsdecode() and os.fsencode() have an important property:
assert os.fsencode(os.fsdecode(data)) == data

On Windows, the other property is maybe more imporant:
assert os.fsdecode(os.fsencode(data)) == data

If the property becomes false, for example if the filesystem encoding is changed at runtime, you get mojibake. Example:

* os.fsdecode() decodes the filename b'h\xc3\xa9llo' from UTF-8 => 'h\xe9llo'
* sys._enablelegacywindowsfsencoding()
* os.fsencode() encodes the filename to cp1252 => you get 'h\xc3\xa9llo'
 instead of 'h\xe9llo', say hello to mojibake

--

Sorry, I didn't play with sys._enablelegacywindowsfsencoding() on Windows. I don't know if it would "work" if sys._enablelegacywindowsfsencoding() is the first instruction of an application. I expect that Python almost decodes nothing at startup on Windows, so it may work.

sys._enablelegacywindowsfsencoding() is a private method, so it shouldn't be used.

Maybe we could add a "fuse" (flag only sets to 1, cannot be reset to 0) to raise an exception if sys._enablelegacywindowsfsencoding() is called "too late", after the first call to os.fsdecode() / Py_DecodeLocale()?
msg285398 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2017-01-13 14:16
Windows doesn't use the fs encoding at all until Python code requests/provides something in bytes. Except for the caching in fsencode/fsdecode, there's no problem setting it once at the start of your program (and it can only be set once - there's no parameter and it cannot be undone).

What I'm most interested in is whether caching the encoding in fsencode/fsdecode is actually an optimization - if not, remove it, and if so make a way to reset it. I'll get around to this sooner or later but I don't want to stop someone else from doing it.
msg285421 - (view) Author: (JGoutin) * Date: 2017-01-13 17:02
Personally, I call "sys._enablelegacywindowsfsencoding()" for only one reason :
Temporary fixing issues with some third party libraries which use C code for files I/O (With filename as "mbcs" encoded bytes internally).

Theses libraries generally call "filename.encode(sys.getfilesystemencoding())" or "os.fsencode(filename)" before sending filenames from Python to C code.

Actually, I didn't see any side effect for using this function. Maybe because I call it at start before anything else.

Using the environment variable is not easy in my case.


I can look if encoding caching in fsencode is efficient (On Windows). And eventually propose a code change for this.
msg285423 - (view) Author: STINNER Victor (vstinner) * (Python committer) Date: 2017-01-13 17:13
> Temporary fixing issues with some third party libraries which use C code for files I/O (With filename as "mbcs" encoded bytes internally).
>
> Theses libraries generally call "filename.encode(sys.getfilesystemencoding())" or "os.fsencode(filename)" before sending filenames from Python to C code.

Hum, Python lacks a function to encode to/decode from the ANSI code page, something like codecs.code_page_encode() / code_page_decode() with CP_ACP. It would allow to get the same encoding in UTF-8 and legacy modes.
msg285424 - (view) Author: STINNER Victor (vstinner) * (Python committer) Date: 2017-01-13 17:14
Hum, it was long time ago since I worked on Windows. Well, Python has a "mbcs" codec which uses the ANSI code page which exists like "forever". These libraries should be patched to use "mbcs" instead of sys.getfilesystemencoding().
msg285429 - (view) Author: (JGoutin) * Date: 2017-01-13 17:31
Yes, I reported this encoding issue to some of them.

But, there is still some problems : 
- Some libraries are not updated frequently (Or not still maintained), and still use fsencode.
- Tests and CI don't see this problem if they don't have a test case for filename with accents or other uncommon characters in english.

This problem will not be easy to eliminate totally...
msg285888 - (view) Author: (JGoutin) * Date: 2017-01-20 07:14
A little encoding cache benchmark.


Current Code:
=============

import sys

def _fscodec():
    encoding = sys.getfilesystemencoding()
    errors = sys.getfilesystemencodeerrors()

    def fsencode(filename):
        filename = fspath(filename)  # Does type-checking of `filename`.
        if isinstance(filename, str):
            return filename.encode(encoding, errors)
        else:
            return filename

    def fsdecode(filename):
        filename = fspath(filename)  # Does type-checking of `filename`.
        if isinstance(filename, bytes):
            return filename.decode(encoding, errors)
        else:
            return filename

    return fsencode, fsdecode

fsencode, fsdecode = _fscodec()
del _fscodec

---------

import os

%timeit os.fsdecode(b'\xc3\xa9')
The slowest run took 21.59 times longer than the fastest. This could mean that an intermediate result is being cached.
1000000 loops, best of 3: 449 ns per loop

%timeit os.fsencode('é')
The slowest run took 24.20 times longer than the fastest. This could mean that an intermediate result is being cached.
1000000 loops, best of 3: 412 ns per loop


Modified Code:
==============

from sys import getfilesystemencoding, getfilesystemencodeerrors

def fsencode(filename):
    filename = fspath(filename)  # Does type-checking of `filename`.
    if isinstance(filename, str):
        return filename.encode(getfilesystemencoding(),
                               getfilesystemencodeerrors())
    else:
        return filename

def fsdecode(filename):
    filename = fspath(filename)  # Does type-checking of `filename`.
    if isinstance(filename, bytes):
        return filename.decode(getfilesystemencoding(),
                               getfilesystemencodeerrors())
    else:
        return filename

---------

import os

%timeit os.fsdecode(b'\xc3\xa9')
The slowest run took 15.88 times longer than the fastest. This could mean that an intermediate result is being cached.
1000000 loops, best of 3: 541 ns per loop

%timeit os.fsencode('é')
The slowest run took 19.32 times longer than the fastest. This could mean that an intermediate result is being cached.
1000000 loops, best of 3: 502 ns per loop


Result:
=======

Cache is a 17% speed up optimization.
msg285969 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2017-01-21 22:45
Thanks for checking that.

I don't think it's worth retaining the cache on Windows in the face of the broken behaviour. Any real-world case where a lot of paths are being encoded or decoded is also likely to involve file-system access which will dwarf the encoding time. Further, passing bytes on Windows will result in another decode/encode cycle anyway, so there will be a bigger performance impact in using str (though even then, probably only when the str is already represented using 16-bit characters).

Unless somebody wants to make a case for having a more complex mechanism to reset the cache, I'll make the change to remove it (protected by an 'if sys.platform.startswith('win')' check).
msg285975 - (view) Author: STINNER Victor (vstinner) * (Python committer) Date: 2017-01-22 00:29
Can't we just update the cache when the function changes the encoding?
msg285979 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2017-01-22 01:20
How?
msg312378 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2018-02-19 23:01
I took another look at this and it's still unclear whether it's worth the performance loss.

Perhaps moving fsencode and fsdecode (almost) entirely into C would be a better approach? That shouldn't send us backwards at all, and all they really do is a typecheck and then calling a function that's already written in C.
History
Date User Action Args
2018-02-19 23:01:18steve.dowersetassignee: steve.dower ->
messages: + msg312378
2017-01-22 01:20:26steve.dowersetmessages: + msg285979
2017-01-22 00:29:07vstinnersetmessages: + msg285975
2017-01-21 22:45:55steve.dowersetassignee: steve.dower
messages: + msg285969
versions: + Python 3.7
2017-01-20 07:14:19JGoutinsetmessages: + msg285888
2017-01-13 17:31:51JGoutinsetmessages: + msg285429
2017-01-13 17:14:27vstinnersetmessages: + msg285424
2017-01-13 17:13:17vstinnersetmessages: + msg285423
2017-01-13 17:02:05JGoutinsetmessages: + msg285421
2017-01-13 14:16:40steve.dowersetmessages: + msg285398
2017-01-13 12:46:24vstinnersetmessages: + msg285393
2017-01-13 10:03:11lemburgsetnosy: + lemburg, vstinner
messages: + msg285374
2017-01-12 23:49:59steve.dowersetmessages: + msg285354
2017-01-12 07:07:47JGoutinsetmessages: + msg285290
2017-01-11 16:43:22steve.dowersetmessages: + msg285238
2017-01-11 12:28:01JGoutincreate