Thanks for discussing the design Peter. I have some ideas which might
make for a nicer v2; I'll massage the code a bit and see what I can
come up with.

On Tue, Sep 21, 2021 at 5:29 PM Peter Xu [off-list ref] wrote:

On Tue, Sep 21, 2021 at 01:31:12PM -0700, Axel Rasmussen wrote:

quoted

Ah, it has to do with us asserting the list of expected ioctls. The
kernel changes the list of ioctls it reports in response to a
UFFDIO_REGISTER, depending on the particular kind of vma being
registered, **as well as what mode(s) it is being registered with**.

So for example, consider the hugetlb_shared test. When registering,
the kernel might set the UFFDIO_CONTINUE bit or not, depending on
whether we registered with the MINOR mode bit set in particular.

I can understand your point, but the "capability set" of the kernel is still
the same.  In this case we should have UFFDIO_CONTINUE capability for
hugetlb_shared test globally, as long as the kernel supports it, irrelevant of
what test case we're going to have.

Then in the test, if we don't register with MINOR mode, IMHO we should just
mask out the expected_ioctls with UFFDIO_CONTINUE because it does not make
sense to request UFFDIO_CONTINUE if we will never use it in the test.

Right, this is how it was before. I didn't love how the base set
included everything, and then each test is responsible for removing
the things it isn't testing. It seems reversed: why not just have each
test compute the set of things it *is* testing?

In other words, having a "uffd_features" global variable and having it changing
all the time during tests is odd to me, but I agree it's not a big deal. :)

100% agree with this. From my perspective this is tech debt since:

8ba6e86408 userfaultfd/selftests: reinitialize test context in each test

It used to be that we just had one global context (variables like
uffd, count_verify, area_src, area_dst, etc). But this had the problem
where some previous test can mutate the context, breaking or affecting
following tests. After 8ba6e86408, we clear and reinitialize all these
variables for each test, but they're still global. I think it would be
cleaner if these instead were in a struct, which each test initialized
and then destroyed within its own scope. If we were to do such a
refactor, I would put uffd_features in that struct too - it should be
private to a test, since it's a property we get from the uffd.

But, I wasn't sure it was worth the churn to do something like this.

--
Peter Xu