To clarify on UBSAN and enums. My understanding is:
enum A { a = 1, b =2, c = 3};
class enum B : int16_t { a = 1, b = 2, c = 3};
A a = static_cast<A>(4); // UB
B b = static_cast<B>(4); // Not UB. Declaring the holding type makes this
allowable.
On Mon, Jul 13, 2020 at 3:44 PM Micah Kornfield <emkornfi...@gmail.com>
wrote:
> Please see [1]. I ran this arrow-ipc-read-write-test with UBSAN enabled
> and it passed (this isn't my normal dev environment so please double check).
>
>
> https://github.com/emkornfield/arrow/commit/7fbd0fb95f7ea164284720428c7974b87b4b2443
>
> On Mon, Jul 13, 2020 at 3:12 PM Micah Kornfield <emkornfi...@gmail.com>
> wrote:
>
>> I think this might be more complicated, let me see if i can write a test
>> that demonstrates what I'm talking about.
>>
>> On Mon, Jul 13, 2020 at 3:10 PM Wes McKinney <wesmck...@gmail.com> wrote:
>>
>>> Here's a patch that does the check
>>>
>>>
>>> https://github.com/wesm/arrow/commit/5bfdb4255a66a4ec62b1c36ba07682fad47df9a7
>>>
>>> Here is a serialized schema that uses a V6 version
>>>
>>>
>>> https://drive.google.com/file/d/1GiWh5yKXdMaLRWU5K4cnGW2ilybF0LF_/view?usp=sharing
>>>
>>> See in action
>>> https://gist.github.com/wesm/f9621a626d56491b0bd6c8a131acf518
>>>
>>> This seems hacky to me, but maybe it's OK?
>>>
>>> On Mon, Jul 13, 2020 at 4:53 PM Wes McKinney <wesmck...@gmail.com>
>>> wrote:
>>> >
>>> > On Mon, Jul 13, 2020 at 4:43 PM Micah Kornfield <emkornfi...@gmail.com>
>>> wrote:
>>> > >>
>>> > >> We don't have any test cases that have a future metadata version. I
>>> > >> made a branch where I added V6 and wrote an IPC message, then found
>>> > >> that I was unable to determine that it was out of bounds (presumably
>>> > >> UBSAN would error, though, but we need a runtime error outside of
>>> > >> ASAN/UBSAN).
>>> > >
>>> > > To clarify I don't think UBSAN will error on the existing generated
>>> code on future versions. I believe we had issues with parquet because the
>>> enums did not have an explicit type (compare [1] to [2]) . The version
>>> check needs to be done in our code (comparing against MAX [3]).
>>> > >
>>> > > Does that align with your expectations? So we don't get this for
>>> free, but I'm not sure I understand why this is difficult?
>>> >
>>> > If the metadata version comes through as the int16_t value 5
>>> > (currently 4 == V5), how do you get to a runtime error? The generated
>>> > Flatbuffers code is doing a static_cast of 5 to the enum which is UB.
>>> > Maybe I just don't know what I'm doing. It does not appear to be
>>> > possible to obtain the raw int16_t value without doing some kind of
>>> > hacking (e.g. reinterpret_cast of Message* to flatbuffers::Table* and
>>> > using GetField<int16_t>(VT_VERSION, 0))
>>> >
>>> > I can make a binary file that uses the currently non-existent V6 so
>>> > you can try to detect it and
>>> > raise an error
>>> >
>>> >
>>> >
>>> > > [1]
>>> https://github.com/apache/arrow/blob/master/cpp/src/generated/parquet_types.h#L26
>>> > > [2]
>>> https://github.com/apache/arrow/blob/master/cpp/src/generated/Schema_generated.h#L91
>>> > > [3]
>>> https://github.com/apache/arrow/blob/master/cpp/src/generated/Schema_generated.h#L109
>>> > >
>>> > > On Mon, Jul 13, 2020 at 2:34 PM Wes McKinney <wesmck...@gmail.com>
>>> wrote:
>>> > >>
>>> > >> On Mon, Jul 13, 2020 at 4:31 PM Micah Kornfield <
>>> emkornfi...@gmail.com> wrote:
>>> > >> >
>>> > >> > >
>>> > >> > >
>>> > >> > > That static cast is currently undefined behavior.
>>> > >> >
>>> > >> > Is ubsan reporting this? When looking into the feature enum I
>>> tried to
>>> > >> > understand if that was valid. At the time I read the C++ spec* if
>>> the enum
>>> > >> > has an explicitly declared type, all values in that types range
>>> are
>>> > >> > supported.
>>> > >>
>>> > >> We don't have any test cases that have a future metadata version. I
>>> > >> made a branch where I added V6 and wrote an IPC message, then found
>>> > >> that I was unable to determine that it was out of bounds (presumably
>>> > >> UBSAN would error, though, but we need a runtime error outside of
>>> > >> ASAN/UBSAN).
>>> > >>
>>> > >> > The generated enums provide a "max" [1] value that should be
>>> comparable
>>> > >> > against.
>>> > >> > <
>>> https://github.com/apache/arrow/blob/master/cpp/src/generated/Schema_generated.h#L109
>>> >
>>> > >> >
>>> > >> >
>>> > >> > * I am not a C++ lawyer
>>> > >> >
>>> > >> > [1]
>>> > >> >
>>> https://github.com/apache/arrow/blob/master/cpp/src/generated/Schema_generated.h#L109
>>> > >> >
>>> > >> > On Mon, Jul 13, 2020 at 2:19 PM Wes McKinney <wesmck...@gmail.com>
>>> wrote:
>>> > >> >
>>> > >> > > I've discovered while working on ARROW-9399 that it is very
>>> difficult
>>> > >> > > with the Flatbuffers API in C++ to detect a MetadataVersion [1]
>>> that
>>> > >> > > is higher than the current version.
>>> > >> > >
>>> > >> > > For example, suppose that 3 or 4 years from now we move from
>>> version
>>> > >> > > V5 to version V6. The generated Flatbuffers code looks like this
>>> > >> > >
>>> > >> > > org::apache::arrow::flatbuf::MetadataVersion version() const {
>>> > >> > > return
>>> > >> > >
>>> static_cast<org::apache::arrow::flatbuf::MetadataVersion>(GetField<int16_t>(VT_VERSION,
>>> > >> > > 0));
>>> > >> > > }
>>> > >> > >
>>> > >> > > That static cast is currently undefined behavior.
>>> > >> > >
>>> > >> > > One way to deal with this would be to add placeholder future
>>> versions
>>> > >> > > (e.g. V6 and V7).
>>> > >> > >
>>> > >> > > Another backward-and-forward-compatible option would be to
>>> return the
>>> > >> > > version as a short (int16_t) rather than the enum value, which
>>> is
>>> > >> > > subject to this brittleness.
>>> > >> > >
>>> > >> > > Either way unfortunately I think adding forward compatibility
>>> checks
>>> > >> > > is out of scope for 1.0.0 and the risk is low since we don't
>>> > >> > > anticipate bumping the version anytime soon.
>>> > >> > >
>>> > >> > > Thanks,
>>> > >> > > Wes
>>> > >> > >
>>> > >> > > [1]:
>>> https://github.com/apache/arrow/blob/master/format/Schema.fbs#L22
>>> > >> > >
>>>
>>
