When reading a tar archive, tarfile decodes fields using "replace" error handler by default. The result is that we loose informations if there is an undecodable character.

Since the PEP 383, undecodable filenames are stored using surrogates in Python3. I think that it's a good idea to use surrogates for tar, because it's a common problem to have undecodable data in a tar archive (see the unicode section of the tarfile documentation).

lars: Do you have an opinion about this suggestion?

Yes, I will soon have ;-) Please give me a few days...

A better fix is maybe to store fields as bytes, but it would break the compatibility and unicode is pratical in Python3.

I think it is helpful to read the pax specification here:

http://www.opengroup.org/onlinepubs/009695399/utilities/pax.html

pax defines (IIUC) that all strings in a pax-compliant tar file are UTF-8 encoded. For the "invalid" option, they offer the alternatives bypass, rename, UTF-8, and write. It may be useful to provide the same options, in some form.

My patch changes test_uname_unicode() of test_tarfile for the GNU and ustar formats (but not PAX). In GNU and ustar formats, the fields can be encoded in any encoding, and may contain invalid byte sequences.

I think it is a good suggestion to use "surrogateescape" as the default, because (I hope) it produces the fewest errors and is the best choice if tarfile is used in connection with Python's filesystem calls.

- When reading tar headers, undecodable chars in filenames end up as surrogates. This way no information is lost. In principle tarfile is merely a gateway to a filesystem inside an archive, so it feels natural if it treats filenames the same as Python's filesystem calls.

- When writing tar headers, filenames with surrogate chars (e.g. from os.listdir()) will be converted back to bytes in the header (in case of gnu and ustar formats). Filenames will remain unchanged, this is exactly as one would expect.

- When writing pax headers, filenames with surrogates will raise a UnicodeError because we may only use strict utf-8 inside a pax header. This is actually no difference to the status quo.

@Martin: As I understand it, the pax "invalid"-option is supposed to deal with the case when strings from a pax header are not representable in the user's encoding. In tarfile's case we don't have this problem when reading the archive until we try to extract it.

Unfortunately, POSIX says nothing about how to store bad filenames in a pax archive. tarfile raises an error. GNU tar fails silently, it just puts the unchanged original filename into the pax header without converting it to utf-8, thus violating the standard.

Thank you for your review. I commited the patch as r80824 (I fixed the documentation, :versionadded => :versionchanged), blocked as r80825 (3.2).

--

> Unfortunately, POSIX says nothing about how to store bad filenames in
> a pax archive. tarfile raises an error. GNU tar fails silently,
> it just puts the unchanged original filename into the pax header
> without converting it to utf-8, thus violating the standard.

Right. I opened a new issue about that: #8333. I consider that it's a different problem.