Commas aren't legal characters in cookie keys, yet in Python 3.5, they're allowed:

>>> bool(http.cookies._is_legal_key(','))
True

The issue lies in the use of _LegalChars constructing a regular expression.

"Some people, when confronted with a problem, think 'I know, I'll use regular expressions.' Now they have two problems."

The issue arises in this line:

_is_legal_key = re.compile('[%s]+' % _LegalChars).fullmatch

Which was added in 88e1151e8e0242 referencing issue2211.

The problem is that in a regular expression, and in a character class in particular, the '-' character has a special meaning if not the first character in the class, which is "span all characters between the leading and following characters". As a result, the pattern has the unintended effect of including the comma in the pattern:

>>> http.cookies._is_legal_key.__self__
re.compile("[abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789!#$%&'*+-.^_`|~:]+")
>>> pattern = _
>>> pattern.fullmatch(',')
<_sre.SRE_Match object; span=(0, 1), match=','>
>>> ord('+')
43
>>> ord('.')
46
>>> ''.join(map(chr, range(43,47)))
'+,-.'

That's how the comma creeped in.

This issue is the underlying cause of https://bitbucket.org/cherrypy/cherrypy/issues/1405/testcookies-fails-on-python-35 and possible other cookie-related bugs in Python.

While I jest about regular expressions, I like the implementation. It just needs to account for the extraneous comma, perhaps by escaping the dash:

_is_legal_key = re.compile('[%s]+' % _LegalChars.replace('-', '\\-')).fullmatch

Also, regression tests for keys containing invalid characters should be added as well.

Ah, how can I missed this catch?

The simplest way is just move "-" to the start or the end of character list. The most error-proof way is to use re.escape().

Hi, I am a newcomer here, I am interested to work on this issue.

We just need to use '\-' instead of '-'.

```
>>> regex = re.compile("[a-z]")
>>> bool(regex.match('b'))
True
>>> regex = re.compile("[a\-z]")
>>> bool(regex.match('b'))
False
```

I have uploaded a patch.
Let me know if this needs some tests too?

A test would be much appreciated. It's worthwhile capturing the rejection of at least one invalid character, if not several common ones.

No, _LegalChars shouldn't include "\".

@serhiy.storchaka OK, I have used re.escape instead of '\'. And I have added a test too.

Also, may I know why '\' can not be in _LegalChars, so that I can remember this for future purpose?

_LegalChars contained only characters which don't require quoting, as documented in the comment above. If _LegalChars was only used to create _is_legal_key, we would just wrote the regular expression. But it is used also in other places. In this particular case adding "\" to _LegalChars doesn't lead to visible bug (except inconsistency with the comment), but we can't be sure.

_is_legal_key() is implementation detail. It would be better to test public API.

Another option might be to do away with the regular expression (personally I like to avoid REs and code generation where practical):

def _is_legal_key(key):
    return key and set(_LegalChars).issuperset(key)

The regular expression is used for performance.

I ran regex and issuperset version on a random string. The regex one gives better performance. So, I have included the re.escape in the patch. 

>>> random_str = ''.join(random.choice(_LegalChars) for _ in range(10 ** 8))
>>> is_legal_key = re.compile('[%s]+' % re.escape(_LegalChars)).fullmatch
>>> Timer("is_legal_key(random_str)", setup="from __main__ import random_str, is_legal_key").timeit(1)
0.3168252399998437
>>> def is_legal_key(key):
...     return key and set(_LegalChars).issuperset(key)
... 
>>> Timer("is_legal_key(random_str)", setup="from __main__ import random_str, is_legal_key").timeit(1)
4.3335622880001665


Also, I have updated the patch. Can you please review it? :)

The same bug is in the _CookiePattern regular expression. For illegal characters other than a comma, the CookieError does not actually seem to be raised.

I take that back about _CookiePattern having the same bug; it uses a different input variable. But it is strange that _LegalKeyChars lists a comma, but _LegalChars omits it.

_LegalKeyChars contains "\-" whereas _LegalChars just contains "-".

On Sun, Feb 7, 2016 at 4:33 PM, Martin Panter <report@bugs.python.org>
wrote:

>
> Martin Panter added the comment:
>
> I take that back about _CookiePattern having the same bug; it uses a
> different input variable. But it is strange that _LegalKeyChars lists a
> comma, but _LegalChars omits it.
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue26302>
> _______________________________________
>

The patch looks okay to me.

The inconsistency between silently rejecting cookie “morsels” and raising an exception from load() also exists in 2.7, so maybe it is not a big deal.

Is this patch ready to merge?

Looks good to me.

LGTM. Do you want to commit the patch Jason?

New changeset 758cb13aaa2c by Anish Shah <anish.shah> in branch '3.5':
Issue #26302: Correctly identify comma as an invalid character for a cookie (correcting regression in Python 3.5).
https://hg.python.org/cpython/rev/758cb13aaa2c

New changeset 91eb7ae951a1 by Jason R. Coombs in branch 'default':
Issue #26302: merge from 3.5
https://hg.python.org/cpython/rev/91eb7ae951a1

Thanks Anish for the patch, now slated for Python 3.5.2 and Python 3.6.