Re: [PATCH v2 09/46] mm: add MADV_SPLIT to enable HugeTLB HGM

[James Houghton <jthoughton@google.com>]

On Fri, Feb 17, 2023 at 5:58 PM Mina Almasry <almasrymina@google.com> wrote:
>
> On Fri, Feb 17, 2023 at 4:28 PM James Houghton <jthoughton@google.com> wrote:
> >
> > Issuing ioctl(MADV_SPLIT) on a HugeTLB address range will enable
> > HugeTLB HGM. MADV_SPLIT was chosen for the name so that this API can be
> > applied to non-HugeTLB memory in the future, if such an application is
> > to arise.
> >
> > MADV_SPLIT provides several API changes for some syscalls on HugeTLB
> > address ranges:
> > 1. UFFDIO_CONTINUE is allowed for MAP_SHARED VMAs at PAGE_SIZE
> >    alignment.
> > 2. read()ing a page fault event from a userfaultfd will yield a
> >    PAGE_SIZE-rounded address, instead of a huge-page-size-rounded
> >    address (unless UFFD_FEATURE_EXACT_ADDRESS is used).
> >
> > There is no way to disable the API changes that come with issuing
> > MADV_SPLIT. MADV_COLLAPSE can be used to collapse high-granularity page
> > table mappings that come from the extended functionality that comes with
> > using MADV_SPLIT.
> >
>
> So is a hugetlb page or VMA that has been MADV_SPLIT + MADV_COLLAPSE
> distinct from a hugetlb page or vma that has not been? I thought
> COLLAPSE would reverse the effects on SPLIT completely.

Right now, MADV_COLLAPSE does *not* completely undo the effects of an
MADV_SPLIT. The API changes that come from MADV_SPLIT aren't undone
with an MADV_COLLAPSE.

>
> > For post-copy live migration, the expected use-case is:
> > 1. mmap(MAP_SHARED, some_fd) primary mapping
> > 2. mmap(MAP_SHARED, some_fd) alias mapping
> > 3. MADV_SPLIT the primary mapping
> > 4. UFFDIO_REGISTER/etc. the primary mapping
> > 5. Copy memory contents into alias mapping and UFFDIO_CONTINUE the
> >    corresponding PAGE_SIZE sections in the primary mapping.
> >
>
> Huh, so MADV_SPLIT doesn't actually split an existing PMD mapping into
> high granularity mappings. Instead it says that future mappings may be
> high granularity? I assume they may not even be high granularity, like
> if the alias mapping faulted in a full hugetlb page (without
> UFFDIO_CONTINUE) that page would be regular mapped not high
> granularity mapped.

MADV_SPLIT just means "userspace is aware that they are able to start
mapping HugeTLB pages at high-granularity". Right now the only way to
get high-granularity mappings is with UFFDIO_CONTINUE, but there may
be other ways in the future.

As of this series, if you MADV_SPLIT a HugeTLB VMA and you aren't
using userfaultfd minor faults, it's basically a no-op. The mappings
that are created will still be huge. I could change this, but I don't
really see a reason to right now.

>
> This may be bikeshedding but I do think a clearer name is warranted.
> Maybe MADV_MAY_SPLIT or something.

I agree -- MADV_MAY_SPLIT more accurately describes the HugeTLB
functionality. I really don't mind what the MADV is called.

I think enabling the high-granularity userfaultfd bits with a
userfaultfd feature[1] worked reasonably well. There is some API
discussion in that thread[1].

[1]: https://lore.kernel.org/linux-mm/20221021163703.3218176-34-jthoughton@google.com/