Title: is_zipfile false positives
Type: behavior Stage: patch review
Components: Library (Lib) Versions: Python 3.8, Python 3.7, Python 2.7
Status: open Resolution:
Dependencies: Superseder:
Assigned To: gregory.p.smith Nosy List: Thomas.Waldmann, alanmcintyre, gregory.p.smith, jjolly, mryan1539, serhiy.storchaka, twouters
Priority: normal Keywords: patch

Created on 2016-10-20 21:08 by Thomas.Waldmann, last changed 2019-01-30 21:23 by gregory.p.smith.

File name Uploaded Description Edit Thomas.Waldmann, 2016-10-20 21:08
isz_fail_fix.diff Thomas.Waldmann, 2016-10-20 22:26
Pull Requests
URL Status Linked Edit
PR 5053 open jjolly, 2017-12-30 18:37
Messages (9)
msg279084 - (view) Author: Thomas Waldmann (Thomas.Waldmann) Date: 2016-10-20 21:08
zipfile.is_zipfile has false positives way too easily.

I just have seen it in practive when a MoinMoin wiki site with a lot of pdf attachments crashed with 500. This was caused by a valid PDF that just happened to contain PK\005\006 somewhere in the middle - this was enough to satisfy is_zipfile() and triggered further processing as a zipfile, which then crashed with IOError (which was not catched in our code, yet).

I have looked into zipfile code: if the usual EOCD structure (with empty comment) is not at EOF, it is suspected that there might be a non-empty comment and ~64K before EOF are searched for the PK\005\006 magic. If it is somewhere there, it is assumed that the file is a zip, without any further validity check.

Attached is a failure demo that works with at least 2.7 and 3.5.
msg279088 - (view) Author: Thomas Waldmann (Thomas.Waldmann) Date: 2016-10-20 22:26
patch for py2.7

The EOCD structure is at EOF.

It either does not contain a comment (this is what the existing code checks first) or it contains a comment of the length that is specified in the structure.

The patch checks consistency specified length vs. real length (end of fixed part of structure up to EOF). If this does not match, it is likely not a zip file, but just a file that happens to have the magic 4 bytes somewhere in its last 64kB.
msg279089 - (view) Author: Thomas Waldmann (Thomas.Waldmann) Date: 2016-10-20 22:35
Note: checking the first bytes of the file (PK..) might be another option.

But this has the "problem" that a self-extracting zip starts with an executable that has different first bytes.

So whether this is an option or not depends on whether is_zipfile() should return truish for self-extracting ZIP files.
msg280341 - (view) Author: Serhiy Storchaka (serhiy.storchaka) * (Python committer) Date: 2016-11-08 19:45
The problem is that the zipfile module supports even not well-formed archives, with a data appended past a comment, and with truncated comment. There are special tests for this, and the proposed patch breaks these tests: test_comments, test_ignores_newline_at_end, test_ignores_stuff_appended_past_comments. See issue10694 and issue1622.
msg281805 - (view) Author: Thomas Waldmann (Thomas.Waldmann) Date: 2016-11-27 00:18
Well, if you have a better idea how to fix is_zipfile, go on.

I even suggested an alternative, how about that?

It is a miserable state when the is_zipfile function in the stdlib detects random crap as a zip file.
msg281817 - (view) Author: Serhiy Storchaka (serhiy.storchaka) * (Python committer) Date: 2016-11-27 10:29
No, checking the first bytes of the file is not appropriate option. zipfile should support the Python zip application format [1].

I see two options:

1. Make is_zipfile() more strict that the ZipFile constructor. The later supports ZIP files with a data past the comment or with truncated comments, but the former should reject them.

2. Make both is_zipfile() and the ZipFile constructor more robust. They should check not just the EOCD signature, but check the Zip64 end of central directory record (if exists) and the first central file header signature (if the ZIP file is not empty).

It may be that PDF files contain PK\005\006 not accidentally, but because they contain embedded ZIP files (I don't know if this is a case). In that circumstances is_zipfile() returning True is correct.

msg309245 - (view) Author: John Jolly (jjolly) * Date: 2017-12-30 19:28
Fix submitted that evaluates the ECD structure and validates the first CD entry. The fix also handles empty zipfiles.

IMO the purpose of this API is to *quickly* verify that the file is a valid zipfile. With this fix, the API only reads another 46 bytes of data (after a seek, of course). This should still qualify as "quick", especially after having potentially read 64k of data.

Perhaps a full zip validator would be appropriate in addition to is_zipfile. That would be more appropriate as a full feature rather than in this bugfix.
msg311281 - (view) Author: John Jolly (jjolly) * Date: 2018-01-30 16:33
Is there any chance that this will make it into 3.7 or is my reminder too late?
msg334595 - (view) Author: Gregory P. Smith (gregory.p.smith) * (Python committer) Date: 2019-01-30 21:23
it's a bugfix, it seems reasonable for 3.7 to me.  I agree that the previous is_zipfile check is too lenient.  I'll follow up on jjolly's PR for any specific concerns I have with the implementation.
Date User Action Args
2019-01-30 21:23:57gregory.p.smithsetassignee: serhiy.storchaka -> gregory.p.smith
messages: + msg334595
2019-01-30 21:09:30gregory.p.smithsetversions: + Python 3.8, - Python 3.5, Python 3.6
2019-01-30 20:48:14mryan1539setnosy: + mryan1539
2018-01-30 16:35:26jjollysetnosy: + gregory.p.smith
2018-01-30 16:33:55jjollysetmessages: + msg311281
2017-12-30 19:28:09jjollysetnosy: + jjolly
messages: + msg309245
2017-12-30 18:37:19jjollysetstage: patch review
pull_requests: + pull_request4934
2016-11-27 10:29:35serhiy.storchakasetmessages: + msg281817
2016-11-27 00:18:41Thomas.Waldmannsetmessages: + msg281805
2016-11-08 19:45:16serhiy.storchakasetmessages: + msg280341
2016-11-03 13:47:15serhiy.storchakasetassignee: serhiy.storchaka
2016-10-20 22:35:23Thomas.Waldmannsetmessages: + msg279089
2016-10-20 22:26:45Thomas.Waldmannsetfiles: + isz_fail_fix.diff
keywords: + patch
messages: + msg279088
2016-10-20 21:30:44serhiy.storchakasetnosy: + twouters, alanmcintyre, serhiy.storchaka

versions: + Python 3.6, Python 3.7
2016-10-20 21:08:48Thomas.Waldmanncreate