homepage

On compilation of .pyc/.pyo bytecode files on import Python 2.6 copies
Unix file access attributes (-rwx-) from the imported file. I'd think
it's ok except for executable (-x-) bit - bytecode files are not
directly executable. That is, for a module.py with attributes -rwxr-x---
I expect python2.6 -c 'import module' would produce module.pyc with
attributes -rw-r-----.
python compileall.py . saves compiled files without the executable bit;
it doesn't copy attributes - it just saves files and saving obeys umask.
python2.5 -c 'import module.py' doesn't copy the executable bit, it
obeys umask too. I consider this as a regression in 2.6. Please mask out
executable bits on bytecode files saving.

Would you like to provide a patch?

Python writes compiled files with the same file permissions as the
source module. In your specific example:
$ python2.6 -c 'import module'
This will produce module.pyc with the same attributes as module.py
While I am not familiar with modifying C stat structs, I can say that in
trunk/Python/import.c, somewhere after line 977 but before line 1014,
the stat struct "st" must be examined for S_IXUSR, S_IXGRP, and S_IXOTH
and have those bits removed in order to never produce a compiled module
with "x" in the stat. I would assign low priority to this issue, as it
has little impact on the behavior of Python, even though it might cause
unintended consequences to a user trying to execute the bytecode, if the
user interprets the "x" to mean that it is executable by the system.

I will try to look at import.c, though I must say I am a bad C
programmer. I have switched to Python after ten years of Pascal.
Low priority is ok.

I attach a patch to correct this little bug.
Bye ;)

Are these S_IX... constants available on every platform we support?

No, they are not supported on Visual Studio.

TO georg.brandl:
I remembered that Windows wasn't POSIX compliant, but...I thought they
used the same sys/stat.h constants.
I was wrong :P
TO loewis:
ok, I've added a new patch.
Since I've never written any code for Windows, can you check if it works
fine now?
I've added a simple #ifdef WINDOWS,...

The patch of file 14070 doesn't compile, but I get the idea. I won't
have time to test it in the next few days or weeks, though.

TO loewis: 
this new patch works fine.
I've removed the first #endif.
thank you ;)

import_patch2.patch doesn't work for me. I patched and compiled Python
2.6.2 and without installing it ran ./python -c "import test" in the
build directory. It copied executable bits from test.py to test.pyc.

hmm.. the problem is that Windows doesn't support well permissions as
all the other POSIX compliant OSs ...
I've searched for a solution on the web, and I've found a complete
answer on:
http://stackoverflow.com/questions/592448/c-how-to-set-file-permissions-cross-platform
The patch doesn't work well since it only checks for User's permissions
so it works well for that. 
Maybe using the Windows API you can change the permissions as you want.
But since I don't know them, I can't help anymore :(

I am not on Windows. I am on Linux.

The patch did not apply for me. I modified the code by hand based on
the patch file, and on Gentoo linux it worked for me. Patch that
applies cleanly to trunk attached.

Thank you David.. sorry for my errors :)
but this is my first patch :P 
I've recompiled py2.6 and it works fine on my Debian.
As you can see:
-rwxr-xr-x 1 marco marco 81822 30 set 2008 setup.py
./python -c "import setup"
-rw-r--r-- 1 marco marco 42156 23 giu 09:58 setup.pyc
However, the version 2.5 doesn't have this "bug" and I've not recompiled it.

Sorry, found the bug in my process of testing. ./python uses /usr/lib/python26.so instead of ./python26.so. Setting LD_LIBRARY_PATH=. fixes the problem and the test passes. The patch is ok.

Thank you for your report :P
Otherwise I really didn't know how solve it, thank you :P
However, on *NIX-like systems it can work well; 
but can someone try it on Windows? Since I know that only NTFS (and
versions >= XP) supports permissions, if a user have a FATfs (or Win98,
95,..) I think it raises error.

Making something executable on Windows has nothing to
do with file permissions. You can set them as much
as you like, but executability is determined by file
associations, possibly in association with PATHEXT
settings. AFAICT, the current Python installer does
this already for pyc / pyo files. (When I say "nothing
to do with file permissions", obviously the file in
question must be readable)

So we should just do #ifndef MS_WINDOWS? Do we have any non-windows
non-posix platforms? (People on #python-dev don't think so.)

I can confirm the patch works on WinXP on NTFS partition and Samba-shared network drive. I have WinXP running in an emulator (VirtualBox) and I compiled Python using MSVC90 Express.

Yes, I think #ifndef MS_WINDOWS is enough.

#ifndef version of the patch applied to trunk in r73870, with tests. 
Leaving open until I merge it.

@r.david.murray:
Does this works on Windows? Are you sure Oleg? :)
Since you've done this:
#ifndef MS_WINDOWS
 /* mode = ..*/
#endif
but on Windows the compiler "jumps" over this code, so you can get a
binding error, since it doesn't find the variable "mode"...
E.g (on my Debian):
#ifndef __GNUC__
 #define X 10
#endif
int main()
{
 printf("%d\n", X);
}
it gives error: ‘X’ undeclared (first use in this function)

I didn't test ifndef-version, I tested the full version (issue6070.patch)
on both Linux and w32. You are right, 'mode' must be defined even on w32.

Drat. I should have uploaded the revised patch first for testing. I'll
update it to the tested patch.

Committed to 2.6 in r74085, py3k in r74058 (by alexandre vassalotti),
and 3.1 in r74086.