On 08/09/23 20:37, Kefeng Wang wrote:
> Hi Mike
>
> On 2023/8/8 2:45, Zi Yan wrote:
> > On 7 Aug 2023, at 8:20, Kefeng Wang wrote:
> >
> > > Hi Zi Yan and Matthew and Naoya,
> > >
> > > On 2023/8/4 13:54, Kefeng Wang wrote:
> > > >
> > > >
> > > > On 2023/8/4 10:42, Zi Yan wrote:
> > > > > On 3 Aug 2023, at 21:45, Kefeng Wang wrote:
> > > > >
> > > > > > On 2023/8/3 20:30, Matthew Wilcox wrote:
> > > > > > > On Thu, Aug 03, 2023 at 03:13:21PM +0800, Kefeng Wang wrote:
> > > > > > > >
> > >
> > > ...
> > >
> > > > > >
> > > > > >
> > > > > > if (PageHuge(page)) // page must be a hugetlb page
> > > > > > if (PageHead(page)) // page must be a head page, not tail
> > > > > > isolate_hugetlb() // isolate the hugetlb page if head
> > > > > >
> > > > > > After using folio,
> > > > > >
> > > > > > if (folio_test_hugetlb(folio)) // only check folio is hugetlb or not
> > > > > >
> > > > > > I don't check the page is head or not, since the follow_page could
> > > > > > return a sub-page, so the check PageHead need be retained, right?
> > > > >
> > > > > Right. It will prevent the kernel from trying to isolate the same hugetlb page
> > > > > twice when two pages are in the same hugetlb folio. But looking at the
> > > > > code, if you try to isolate an already-isolated hugetlb folio, isolate_hugetlb()
> > > > > would return false, no error would show up. But it changes err value
> > > > > from -EACCES to -EBUSY and user will see a different page status than before.
> > > >
> > >
> > > Before e66f17ff7177 ("mm/hugetlb: take page table lock in follow_huge_pmd()")
> > > in v4.0, follow_page() will return NULL on tail page for Huagetlb page,
> > > and move_pages() will return -ENOENT errno,but after that commit,
> > > -EACCES is returned, which not match the manual,
> > >
> > > >
> > > > When check man[1], the current -EACCES is not right, -EBUSY is not
> > > > precise but more suitable for this scenario,
> > > >
> > > > -EACCES
> > > > The page is mapped by multiple processes and can be moved
> > > > only if MPOL_MF_MOVE_ALL is specified.
> > > >
> > > > -EBUSY The page is currently busy and cannot be moved. Try again
> > > > later. This occurs if a page is undergoing I/O or another
> > > > kernel subsystem is holding a reference to the page.
> > > > -ENOENT
> > > > The page is not present.
> > > >
> > > > >
> > > > > I wonder why we do not have follow_folio() and returns -ENOENT error pointer
> > > > > when addr points to a non head page. It would make this patch more folio if
> > > > > follow_folio() can be used in place of follow_page(). One caveat is that
> > > > > user will see -ENOENT instead of -EACCES after this change.
> > > > >
> > > >
> > > > -ENOENT is ok, but maybe the man need to be updated too.
> > >
> > > According to above analysis, -ENOENT is suitable when introduce the
> > > follow_folio(), but when THP migrate support is introduced by
> > > e8db67eb0ded ("mm: migrate: move_pages() supports thp migration") in
> > > v4.14, the tail page will be turned into head page and return -EBUSY,
> > >
> > > So should we unify errno(maybe use -ENOENT) about the tail page?
> > >
> > >
> > > >
> > > >
> > > >
> > > > [1] https://man7.org/linux/man-pages/man2/move_pages.2.html
> >
> > I think so. I think -EBUSY is more reasonable for tail pages. But there is
> > some subtle difference between THP and hugetlb from current code:
> >
> > For THP, compound_head() is used to get the head page for isolation, this means
> > if user specifies a tail page address in move_pages(), the whole THP can be
> > migrated.
> >
> > For hugetlb, only if user specifies the head page address of a hugetlb page,
> > the hugetlb page will be migrated. Otherwise, an error would show up.
> >
> > Cc Mike to help us clarify the expected behavior of hugetlb.
> >
> > Hi Mike, what is the expected behavior, if a user tries to use move_pages()
> > to migrate a non head page of a hugetlb page?
>
> Could you give some advise, thanks
>
Sorry, I was away for a while.
It seems unfortunate that move_pages says the passed user addresses
should be aligned to page boundaries. However, IIUC this is not checked
or enforced. Otherwise, passing a hugetlb page should return the same
error.
One thought would be that hugetlb mappings should behave the same
non-hugetlb mappings. If passed the address of a hugetlb tail page, align
the address to a hugetlb boundary and migrate the page. This changes the
existing behavior. However, it would be hard to imagine anyone depending
on this.
After taking a closer look at the add_page_for_migration(), it seems to
just ignore passed tail pages and do nothing for such passed addresses.
Correct? Or, am I missing something? Perhaps that is behavior we want/
need to preserve?
--
Mike Kravetz
