Hi

The behavior of os.realpath function on Linux is different from the one 
presented by the system. 
Although the path (pointing to the linked dir) is correct
and is recognized by os.path.exists once it is resolved
by the realpath() it does not exist - at the end of the 
resolved path you will see the additional parent directory name.
you will see
..... /two/two at the end of the path.

Below is the code that shows the issue.

def linkGen():
  import os
  print os.getcwd()
  test = "./test"
  os.system( "rm -rf " + test  )  
  #os.mkdir( test )
  os.makedirs( test + "/one" )
  os.makedirs( test + "/two" )
  os.symlink( "../two", test + "/two/this_dir" )
  os.symlink( "../two/this_dir/this_dir/this_dir/this_dir/", test + 
"/one/that_dir" )
  print test + "/one/that_dir", "EXISTS?",   
  print os.path.exists( test + "/one/that_dir" )
  print os.path.realpath( test + "/one/that_dir" ), "EXISTS?", 
  print os.path.exists( os.path.realpath( test + "/one/that_dir" ) )
  os.system( "ls -l " + test + "/one/that_dir" )
  # the line above will show that system can see the directory just fine

linkGen()

Hi
I meant of course 
os.path.realpath()
Sylwester

I've provided a patch.

I've also added a new test, and it passes.

I've just found a similar bug:
http://bugs.python.org/issue1646838

However, does os.path.realpath's behavior have to be as the one
specified by the POSIX standard (
http://www.kernel.org/doc/man-pages/online/pages/man3/realpath.3.html )?
If we wanted to follow the standard, we should break the
retro-compatibility, since we should raise an exception in the case the
path passed as argument doesn't exist.

Marco,

Thanks for looking deeper into it.
I would like to check the solution ASAP.
We have really crazy dir structure which can catch a lot of
the unexpected problems with paths, links, circular links etc.

Should I expect the new version to generate the exception
as you suggested?

Best Regards,
    Sywlester Warecki

Marco Buccini wrote:
> Marco Buccini <marcusbu@gmail.com> added the comment:
>
> I've just found a similar bug:
> http://bugs.python.org/issue1646838
>
> However, does os.path.realpath's behavior have to be as the one
> specified by the POSIX standard (
> http://www.kernel.org/doc/man-pages/online/pages/man3/realpath.3.html )?
> If we wanted to follow the standard, we should break the
> retro-compatibility, since we should raise an exception in the case the
> path passed as argument doesn't exist.
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue6975>
> _______________________________________
>
>

>Thanks for looking deeper into it.
>I would like to check the solution ASAP.
>We have really crazy dir structure which can catch a lot of
>the unexpected problems with paths, links, circular links etc.

Oh well, so you can see if the patch works correctly on all paths :)
 
>Should I expect the new version to generate the exception
>as you suggested?

 
Unfortunately I'm not God, so I cannot change the "code" of the current
implementation of os.path.realpath (raising an OSError exception)
without breaking any existing code. 
So, consider using the current implementation of os.path.realpath ;)
If you experience other strange behaviors, then open a new bug report.

Bye!

swarecki or markon, are either of you interested in following up this issue?

After filing a duplicate, issue 15196, I analyzed this:

What happens:
test/one/that_dir
test/one/../two/this_dir/this_dir/this_dir/this_dir
test/two/this_dir/this_dir/this_dir/this_dir
test/two/this_dir/this_dir/this_dir/../two
test/two/this_dir/this_dir/two

What should happen:
test/one/that_dir
test/two/this_dir/this_dir/this_dir/this_dir
test/two/../two/this_dir/this_dir/this_dir
test/two/this_dir/this_dir/this_dir
test/two/../two/this_dir/this_dir
test/two/this_dir/this_dir
test/two/../two/this_dir
test/two/this_dir
test/two/../two
test/two

Tried it out and it seemed to fix the issue.
Ported the patch to python3.

Previous patches contain redundant code, are not protected from all symlink loops (link -> link/x; dir/link -> ../dir/link) and when a loop is detected the returned result differs from the current one.

The proposed patch solves this issue and some others. Added many new tests.

The test for this issue was simplified:

dir/self -> ../dir
dir/link -> self/self/self
resolve dir/link

Antoine, how about this mosquito?

If no one objects I will commit this next week.

I'd like to see a review first.  I don't have time to do it myself right now though.

OK. Perhaps I will prepare similar patches for ntpath and macpath.

I will review this first thing tomorrow.

Patch rewritten using recursion. Non-recursive version is not stack limited (and a little faster), but recursive version is perhaps more understandable. Some comments added.

I don't really understand your algorithm. Why do you need a stack? It should be simply iterative:
- if the symlink is relative, prepend the symlink target to the rest
- if the symlink is absolute, discard the current path and prepend the symlink target to the rest

Here is how pathlib does it:
https://bitbucket.org/pitrou/pathlib/src/67a2524b057f1af5b3cba26370b1353e73cdda16/pathlib.py?at=default#cl-227

> I don't really understand your algorithm. Why do you need a stack?

Before resolving the symlink we mark the path as unresolved symlink for detecting infinite symlink loops. Before resolving the symlink we mark the path as resolved symlink (and cache the resolved value for speed) for allowing finite symlink recursion.

Here is a test which pathlib fails.

Vis the discussion of x[:0] in the review.  This kind of construct is only unfamiliar because it is new in Python3, and there are not *that* many places that you want to (and can) deal with both bytes and strings using the same code.  But when you can, it is the more-or-less obvious way to write the code.

In this particular case you could also write type(x)().  There are pluses and minuses to both forms, but if any of the rest of the code uses slicing to work around the fact that byte strings index into ints, then I would definitely prefer the slice form.

New changeset 6ec6dbf787f4 by Serhiy Storchaka in branch '2.7':
Issue #6975: os.path.realpath() now correctly resolves multiple nested symlinks on POSIX platforms.
http://hg.python.org/cpython/rev/6ec6dbf787f4

New changeset c5f4fa02fc86 by Serhiy Storchaka in branch '3.2':
Issue #6975: os.path.realpath() now correctly resolves multiple nested symlinks on POSIX platforms.
http://hg.python.org/cpython/rev/c5f4fa02fc86

New changeset bfe9526606e2 by Serhiy Storchaka in branch '3.3':
Issue #6975: os.path.realpath() now correctly resolves multiple nested symlinks on POSIX platforms.
http://hg.python.org/cpython/rev/bfe9526606e2

New changeset f42cabe6ccb5 by Serhiy Storchaka in branch 'default':
Issue #6975: os.path.realpath() now correctly resolves multiple nested symlinks on POSIX platforms.
http://hg.python.org/cpython/rev/f42cabe6ccb5

The above revisions have broken handling of arguments with >=2 "..".

Before these revisions:

$ cd /usr/bin
$ python3.2 -c 'import os; print(os.path.realpath(".."))'
/usr
$ python3.2 -c 'import os; print(os.path.realpath("../.."))'
/
$ python3.2 -c 'import os; print(os.path.realpath("../../.."))'
/
$ python3.2 -c 'import os; print(os.path.realpath("../../../.."))'

After these revisions:

$ cd /usr/bin
$ python3.2 -c 'import os; print(os.path.realpath(".."))'
/usr
$ python3.2 -c 'import os; print(os.path.realpath("../.."))'
/usr/bin
$ python3.2 -c 'import os; print(os.path.realpath("../../.."))'
/usr
$ python3.2 -c 'import os; print(os.path.realpath("../../../.."))'
/usr/bin

The actual output of last command in "Before these revisions:" is: /

Thank you for the report. I'm surprised that the tests are not caught it.

New changeset 50ed06b3d419 by Serhiy Storchaka in branch '2.7':
Fix posixpath.realpath() for multiple pardirs (fixes issue #6975).
http://hg.python.org/cpython/rev/50ed06b3d419

New changeset cb3fbadb65aa by Serhiy Storchaka in branch '3.2':
Fix posixpath.realpath() for multiple pardirs (fixes issue #6975).
http://hg.python.org/cpython/rev/cb3fbadb65aa

New changeset aad7e68eff0a by Serhiy Storchaka in branch '3.3':
Fix posixpath.realpath() for multiple pardirs (fixes issue #6975).
http://hg.python.org/cpython/rev/aad7e68eff0a

New changeset f99ff3b01fab by Serhiy Storchaka in branch 'default':
Fix posixpath.realpath() for multiple pardirs (fixes issue #6975).
http://hg.python.org/cpython/rev/f99ff3b01fab

New changeset 3c5517c4fa5d by Serhiy Storchaka in branch '2.7':
Disable posixpath.realpath() tests on Windows (fix for issue #6975).
http://hg.python.org/cpython/rev/3c5517c4fa5d

New changeset 0bbf7cdea551 by Serhiy Storchaka in branch '3.2':
Disable posixpath.realpath() tests on Windows (fix for issue #6975).
http://hg.python.org/cpython/rev/0bbf7cdea551

New changeset 79ea59b394bf by Serhiy Storchaka in branch '3.3':
Disable posixpath.realpath() tests on Windows (fix for issue #6975).
http://hg.python.org/cpython/rev/79ea59b394bf

New changeset aa77f7eb2bf1 by Serhiy Storchaka in branch 'default':
Disable posixpath.realpath() tests on Windows (fix for issue #6975).
http://hg.python.org/cpython/rev/aa77f7eb2bf1

You could probably test '.\\.' and '..\\..' etc. in these tests on Windows.

I have no access to Windows and can't design Windows tests.

Use os.path.sep and os.path.sep.encode() instead of hardcoding "/" and b"/".

> Use os.path.sep and os.path.sep.encode() instead of hardcoding "/" and
> b"/".

Some separators will be '\\' (if they are derived from OS functions, i.e. 
getcwd), and some will be '/' (if they are generated by posixpath). I do not 
have the ability to research where there are any. Feel free to fix these tests 
for Windows.