On Mon, Dec 4, 2023 at 10:27 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 04.12.23 17:35, Suren Baghdasaryan wrote:
> > On Mon, Dec 4, 2023 at 1:27 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>
> >> On 04/12/2023 04:09, Suren Baghdasaryan wrote:
> >>> On Sat, Dec 2, 2023 at 2:11 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>>
> >>>> On 02.12.23 09:04, Ryan Roberts wrote:
> >>>>> On 01/12/2023 20:47, David Hildenbrand wrote:
> >>>>>> On 01.12.23 10:29, Ryan Roberts wrote:
> >>>>>>> On 21/11/2023 17:16, Suren Baghdasaryan wrote:
> >>>>>>>> Add tests for new UFFDIO_MOVE ioctl which uses uffd to move source
> >>>>>>>> into destination buffer while checking the contents of both after
> >>>>>>>> the move. After the operation the content of the destination buffer
> >>>>>>>> should match the original source buffer's content while the source
> >>>>>>>> buffer should be zeroed. Separate tests are designed for PMD aligned and
> >>>>>>>> unaligned cases because they utilize different code paths in the kernel.
> >>>>>>>>
> >>>>>>>> Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> >>>>>>>> ---
> >>>>>>>> tools/testing/selftests/mm/uffd-common.c | 24 +++
> >>>>>>>> tools/testing/selftests/mm/uffd-common.h | 1 +
> >>>>>>>> tools/testing/selftests/mm/uffd-unit-tests.c | 189 +++++++++++++++++++
> >>>>>>>> 3 files changed, 214 insertions(+)
> >>>>>>>>
> >>>>>>>> diff --git a/tools/testing/selftests/mm/uffd-common.c
> >>>>>>>> b/tools/testing/selftests/mm/uffd-common.c
> >>>>>>>> index fb3bbc77fd00..b0ac0ec2356d 100644
> >>>>>>>> --- a/tools/testing/selftests/mm/uffd-common.c
> >>>>>>>> +++ b/tools/testing/selftests/mm/uffd-common.c
> >>>>>>>> @@ -631,6 +631,30 @@ int copy_page(int ufd, unsigned long offset, bool wp)
> >>>>>>>> return __copy_page(ufd, offset, false, wp);
> >>>>>>>> }
> >>>>>>>> +int move_page(int ufd, unsigned long offset, unsigned long len)
> >>>>>>>> +{
> >>>>>>>> + struct uffdio_move uffdio_move;
> >>>>>>>> +
> >>>>>>>> + if (offset + len > nr_pages * page_size)
> >>>>>>>> + err("unexpected offset %lu and length %lu\n", offset, len);
> >>>>>>>> + uffdio_move.dst = (unsigned long) area_dst + offset;
> >>>>>>>> + uffdio_move.src = (unsigned long) area_src + offset;
> >>>>>>>> + uffdio_move.len = len;
> >>>>>>>> + uffdio_move.mode = UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES;
> >>>>>>>> + uffdio_move.move = 0;
> >>>>>>>> + if (ioctl(ufd, UFFDIO_MOVE, &uffdio_move)) {
> >>>>>>>> + /* real retval in uffdio_move.move */
> >>>>>>>> + if (uffdio_move.move != -EEXIST)
> >>>>>>>> + err("UFFDIO_MOVE error: %"PRId64,
> >>>>>>>> + (int64_t)uffdio_move.move);
> >>>>>>>
> >>>>>>> Hi Suren,
> >>>>>>>
> >>>>>>> FYI this error is triggering in mm-unstable (715b67adf4c8):
> >>>>>>>
> >>>>>>> Testing move-pmd on anon... ERROR: UFFDIO_MOVE error: -16 (errno=16,
> >>>>>>> @uffd-common.c:648)
> >>>>>>>
> >>>>>>> I'm running in a VM on Apple M2 (arm64). I haven't debugged any further, but
> >>>>>>> happy to go deeper if you can direct.
> >>>>>>
> >>>>>> Does it trigger reliably? Which pagesize is that kernel using?
> >>>>>
> >>>>> Yep, although very occasionally it fails with EAGAIN. 4K kernel; see other email
> >>>>> for full config.
> >>>>>
> >>>>>>
> >>>>>> I can spot that uffd_move_pmd_test()/uffd_move_pmd_handle_fault() uses
> >>>>>> default_huge_page_size(), which reads the default hugetlb size.
> >>>>>
> >>>>> My kernel command line is explicitly seting the default huge page size to 2M.
> >>>>>
> >>>>
> >>>> Okay, so that likely won't affect it.
> >>>>
> >>>> I can only guess that it has to do with the alignment of the virtual
> >>>> area we are testing with, and that we do seem to get more odd patterns
> >>>> on arm64.
> >>>>
> >>>> uffd_move_test_common() is a bit more elaborate, but if we aligned the
> >>>> src+start area up, surely "step_count" cannot be left unmodified?
> >>>>
> >>>> So assuming we get either an unaligned source or an unaligned dst from
> >>>> mmap(), I am not convinced that we won't be moving areas that are not
> >>>> necessarily fully backed by PMDs and maybe don't even fall into the VMA
> >>>> of interest?
> >>>>
> >>>> Not sure if that could trigger the THP splitting issue, though.
> >>>>
> >>>> But I just quickly scanned that test setup, could be I am missing
> >>>> something. It might make sense to just print the mmap'ed range and the
> >>>> actual ranges we are trying to move. Maybe something "obvious" can be
> >>>> observed.
> >>>
> >>> I was able to reproduce the issue on an Android device and after
> >>> implementing David's suggestions to split the large folio and after
> >>> replacing default_huge_page_size() with read_pmd_pagesize(), the
> >>> move-pmd test started working for me. Ryan, could you please apply
> >>> attached patches (over mm-unstable) and try the test again?
> >>
> >> Yep, all fixed with those patches!
> >
> > Great! Thanks for testing and confirming. I'll post an updated
> > patchset later today and will ask Andrew to replace the current one
> > with it.
> > I'll also look into the reasons we need to split PMD on ARM64 in this
> > test. It's good that this happened and we were able to test the PMD
> > split path but I'm curious about the reason. It's possible my address
> > alignment calculations are somehow incorrect.
>
> I only skimmed the diff briefly, but likely you also want to try
> splitting in move_pages_pte(), if you encounter an already-pte-mapped THP.
Huh, good point. I might be able to move the folio splitting code into
pte-mapped case and do a retry after splitting. That should minimize
the additional code required. Will do and post a new set shortly.
Thanks!
>
> --
> Cheers,
>
> David / dhildenb
>
