I found an interger overflow in the standard iterator object (Objects/iterobject.c)

The it_index attribute used by the iterator is a Py_ssize_t but overflow is never checked. So after the index PY_SSIZE_T_MAX, the iterator object will ask for the index PY_SSIZE_T_MIN.

Here is an example:

    import sys

    class Seq:
        def __getitem__(self, item):
            print("[-] Asked for item at <{0}>".format(item))
            return 0

    s = Seq()
    i = iter(s)
    # Manually set `it_index` to PY_SSIZE_T_MAX without a loop
    i.__setstate__(sys.maxsize)

    next(i)
    [-] Asked for item at <9223372036854775807>
    next(i)
    [-] Asked for item at <-9223372036854775808>

I would be really interested in writing a patch but first I wanted to discuss the expected behaviour and fix.

The iterator could stop after PY_SSIZE_T_MAX but it seems strange as other iterator (like enumobject) handle values bigger than PY_SSIZE_T_MAX.

Or the same technique used in enumobject could be used: adding a field PyObject* en_longindex (a python long) which would be used for values bigger than PY_SSIZE_T_MAX

For possibly relevant background information, see bpo-21444 and the issues linked from it, and bpo-14794.

The it_index attribute used by the iterator is a Py_ssize_t but overflow is never checked.

Yes it is a bug. iter_iternext() must raises an OverflowError if it->it_index is equal to PY_SSIZE_T_MAX.

I think OverflowError is good for maintained releases, but for 3.5 Clement's idea with long index looks attractive to me. In any case an exception should be raised for negative argument in __setstate__(). Let split this issue on two parts. First fix the bug by raising exceptions, and then add long index (if anyone will need it).

I think OverflowError is good for maintained releases, but for 3.5 Clement's idea with long index looks attractive to me.

What is your use case?

What is your use case?

Something like range(). I agree that it is very unlike to encounter this
problem in real work, and we can live with implementation-related limitation
(for which OverflowError exists).

Here is a first try for a patch.

There are two points I am not sure about:

1. The message for the OverflowError: is that explicit enough ?

2. The behaviour of the iterator after the raise of OverflowError.
   With this patch every call to next(it) where it have overflowed will raise OverflowError again.
   Does this behaviour seems correct our should it raise StopIteration after the first OverflowError ?

You should not rely on undefined behaviour: check if the index is greater
or equal than PY_SSIZET_MAX - 1. __setstate__ must implement the same check.

You must always raise an error, not hide it in a second call to next (raise
StopIteration). Your unit test must check this behaviour.

check if the index is greater or equal than PY_SSIZET_MAX - 1.

Just PY_SSIZET_MAX.

Added other comments on Rietveld (look the email in the Spam folder).

Thanks for the comments.
I corrected the patch and updated the test. I also added a test that check the behavior of setstate with negative indice.

Just one question:

__setstate__ must implement the same check.

Why should __setstate__ check for PY_SSIZE_T_MAX (throwing OverflowError when unpickling) if the check will be done when calling next on the resulting iterator ?

__setstate__ should check that an index is not negative. All values from 0 to PY_SSIZE_T_MAX are acceptable.

Ah, __setstate__ already handles negative indices. Then the patch LGTM.

I would think that the PY_SSIZE_T_MAX check belongs inside the:

    if (result != NULL) {
        it->it_index++;
        return result;
    }

just before the increment which could cause the overflow. Also, PY_SSIZE_T_MAX is a valid value to pass to PySequence_GetItem(), so it shouldn't be blocked unless necessary.

This doesn't matter because next() will raise an exception if it_index == PY_SSIZE_T_MAX in any case. The code should be changed much more to allow yielding an item with index PY_SSIZE_T_MAX, use other (negative) signal value and change the behavior of __setstate__ for negative values.

Also, PY_SSIZE_T_MAX is a valid value to pass to PySequence_GetItem(), so it shouldn't be blocked unless necessary.

I agree with you, that's why my first path was checking at the next call if it->it_index had overflowed. But then it relies on undefined behaviour.

I would think that the PY_SSIZE_T_MAX check belongs inside the:

if (result != NULL) {
    it-\>it_index++;
    return result;
}

If we raise the OverflowError when it->it_index really overflow (just after the getitem PY_SSIZE_T_MAX).
Is it really necessary to do the overflow check after the GetItem ? because the value returned by PySequence_GetItem(seq, PY_SSIZE_T_MAX); will be never used.

I prefer to raise an OverflowError *before* calling
PySequence_GetItem(). The call to PySequence_GetItem() may be
expensive, and we have to drop the result if an OverflowError is
raised after the call. At the end, the behaviour is the same: an
OverflowError is raised.

The call to PySequence_GetItem() may be expensive, and we have to drop
the result if an OverflowError is raised after the call.

You do realize that this error will be very rare and therefore inconsequential.

> The call to PySequence_GetItem() may be expensive, and we have to drop
> the result if an OverflowError is raised after the call.

You do realize that this error will be very rare and therefore inconsequential.

The real question is: why would you call the iterator for a new value if it will be discarded anyway ? I think it could be very misleading to see _getitem__ being called and have an OverflowError being raised afterward.

Is it necessary to raise when it_index is PY_SSIZE_T_MAX?

An alternative is to set it_index to -1 when there would be overflow and raise an exception on the next call to next(). That way a virtual sequence with PY_SSIZE_T_MAX-1 items would still work (instead of failing unexpectedly).

That way a virtual sequence with PY_SSIZE_T_MAX-1 items would still work (instead of failing unexpectedly).

Actually with PY_SSIZE_T_MAX+1 items (indices from 0 to PY_SSIZE_T_MAX inclusive).

If Raymond insists I can write more complicated patch, but I don't think that we should complicate the code for this pretty hypotetical case. I'm for committing issue22939v2.patch.

After a few months, I am back to you on this issue.
What should be the next step of the process ?

If Raymond will not stop me, I'll commit the patch tomorrow.

New changeset d6179accca20 by Serhiy Storchaka in branch '2.7':
Fixed issue number for issue bpo-22939.
https://hg.python.org/cpython/rev/d6179accca20

New changeset 7fa2f4afcf5a by Serhiy Storchaka in branch '3.4':
Fixed issue number for issue bpo-22939.
https://hg.python.org/cpython/rev/7fa2f4afcf5a

New changeset f23533fa6afa by Serhiy Storchaka in branch 'default':
Fixed issue number for issue bpo-22939.
https://hg.python.org/cpython/rev/f23533fa6afa

Thank you for your contribution Clement.