A dialog on python-dev suggests that dict.setdefault() was intended to be atomic and that a number of experienced developers have deployed code relying on its atomicity:  http://mail.python.org/pipermail/python-3000/2007-July/008693.html

The actual implementation shows a second call to PyDict_Setitem() which can call PyObject_Hash() which can call arbitrary Python code.   http://hg.python.org/cpython/file/2.7/Objects/dictobject.c#l1937

This fix is straight-forward, use the results of the initial lookup to insert the default object.  This will make the operation atomic and it will make it faster.

+1.

Atomic and faster... +1.

+1 for atomic and more robust

maniram: We try to keep bug reports focused on one thing, and we don’t generally collect votes (we prefer that people vote with patches, tests and messages with content—for example, saying exactly what “more robust” should be).  Here I think Antoine and Jesús said +1 to agree that this should be changed even in stable versions.

Raymond: I think I could make the patch, but I’m not sure about testing the new behavior.

Eric, overload "__hash__()" and check that is only called once, while now would be called twice.

I have written a patch and a test, but since it's changing C code, I am far from being sure if it's achieve the expected behavior in the right way. There are also tests and running whole test suite didn't bring any errors.

Also: I'll be happy to work further on this patch, if I get some comments and advice.

If _PyDict_SetItemUsingHash is module-private, it should be declared "static". Also, better if it follows the usual naming of static functions inside that C file (i.e. "dict_some_lowercase_name").

Done.

The patch looks ok to me. I'll let Raymond make the final call.

Thanks for the idea Jesús, even though I didn’t get the change to use it :)

Filip: Is there a reasons for using a nonlocal count rather than e.g. self.count?  Otherwise the test looks good.

> Filip: Is there a reasons for using a nonlocal count rather than e.g.
> self.count?  Otherwise the test looks good.

Eric, please, could we stop such pointless nitpicking about one's
stylistic preferences? "nonlocal" is as good as any other solution here,
and reviewing is not supposed to be a pedantry contest. There are tons
of serious issues to be solved where your energy would be better spent,
IM(NSH)O.

I haven't given it much thought, when I was making the choice of using nonlocal rather than self.count. I was rather excited to see, if the change will work as I wanted it to. If you believe it would be better to use an attribute, I'll be happy to change it.

Well, if this is going to be applied to 2.7, "nonlocal" is invalid syntax there. I guess that being able to use the same test in 2.7 and 3.2/3.3 would be nice. Style, but justified.

> Well, if this is going to be applied to 2.7, "nonlocal" is invalid
> syntax there.

Ah, good point.

I'll send a patch, when I get home from work.

Let's just start with a working 2.7 patch and go from there.

I have tried to port patch to python2.7, but apparently I must be doing something wrong, because while most tests pass, several fail (including tests for unittest itself).

I would like to continue working on this test, but incoming week is pretty intensive for me. Is it ok if I returned to working on this during the weekend?

> I would like to continue working on this test, but incoming week is
> pretty intensive for me. Is it ok if I returned to working on this
> during the weekend?

Of course. There's no rush.
(I'm really not sure it should go in 2.7 and 3.2; this looks more like a feature than a bug, since atomicity has never been documented)

Thanks, I'll look at it over the next few days.

Patch for 2.7.

Bump! There was no activity here for two weeks. Is my patch for 2.7 ok or should I do something more about it?

Looking at the patch again, I think this isn't enough.
setdefault() will still call the lookup routine twice which, in the general case (i.e. lookdict() not lookdict_unicode()), can call arbitrary Python code through e.g. __eq__ methods.

I understand you are talking about this call:

mp->ma_lookup(mp, key, hash);

I haven't noticed that earlier. I'll try to provide a better fix (for 2.7 first, after we agree it's good enough, I will provide one for 3.3).

Do you have any idea, how I might this part? I mean how to check, that this is called only once?

> Do you have any idea, how I might this part? I mean how to check, that
> this is called only once?

Checking that __eq__ is called only once should be a good proxy.

Thanks, I'll try that. Like before I will probably find time next weekend.

No more double lookup.

> No more double lookup.

Your patch doesn't check hashed2.eq_count.
Since the dict specification doesn't say on which instance __eq__ will
be called when doing a lookup, the patch should either check
``hashed1.eq_count + hashed2.eq_count``, or make eq_count a class
attribute.

A nit: be careful not to use tabs in C files.

Fixed both issues.

New changeset 5c52e7c6d868 by Antoine Pitrou in branch '2.7':
Issue #13521: dict.setdefault() now does only one lookup for the given key, making it "atomic" for many purposes.
http://hg.python.org/cpython/rev/5c52e7c6d868

New changeset 90572ccda12c by Antoine Pitrou in branch '3.2':
Issue #13521: dict.setdefault() now does only one lookup for the given key, making it "atomic" for many purposes.
http://hg.python.org/cpython/rev/90572ccda12c

New changeset 3dfa98cf2e26 by Antoine Pitrou in branch 'default':
Issue #13521: dict.setdefault() now does only one lookup for the given key, making it "atomic" for many purposes.
http://hg.python.org/cpython/rev/3dfa98cf2e26

Thank you Filip! I've now committed the patch.