Re: [PATCH 0/7] mm/thp: fix THP splitting unmap BUGs and related

[Hugh Dickins]

On Tue, 1 Jun 2021, Hugh Dickins wrote:

> Here is a batch of long-standing THP bug fixes that I had not got
> around to sending before, but prompted now by Wang Yugui's report
> https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@xxxxxxxxxxxx/
> 
> Wang Yugui has tested a rollup of these fixes applied to 5.10.39,
> and they have done no harm, but have *not* fixed that issue:
> something more is needed and I have no idea of what.
> 
> But at least these clear up related issues, and should go to stable
> (except for the last, which is just an optimization: it would be
> fine in stable, but it's not required there).
> 
> These are against 5.13-rc4: easy for others to try out, but my next
> task is to work out how to shoehorn them into mmotm and linux-next.
> 
> I would have said just before Yang Shi's related
> mm-thp-replace-debug_vm-bug-with-vm_warn-when-unmap-fails-for-split.patch
> except that (which should also go to stable) is currently placed after
> Alistair Popple's "Add support for SVM atomics in Nouveau" series,
> mm-rmap-split-try_to_munlock-from-try_to_unmap.patch etc.
> I expect I shall offer you some rediffs of Alistair's, we'll see.

And here's my attempt to rewrite history: mmotm-adjust.tar contains

before/series
before/mm-remove-special-swap-entry-functions.patch
before/mm-rmap-split-try_to_munlock-from-try_to_unmap.patch
before/mm-huge_memoryc-remove-unnecessary-tlb_remove_page_size-for-huge-zero-pmd.patch
after/series
after/mm-remove-special-swap-entry-functions.patch
after/mm-rmap-split-try_to_munlock-from-try_to_unmap.patch
after/mm-huge_memoryc-remove-unnecessary-tlb_remove_page_size-for-huge-zero-pmd.patch

"before" contains the originals from mmotm .DATE=2021-06-01-19-57:
your series file, two patches from Alistair and one from Miaohe, which
gave rejects when I inserted this 1-7/7 patchset and Yang Shi's 1-2/2
earlier.  All other patches applied cleanly, but could be refreshed.

"after" contains my replacements to those: adjusting them to how they
would have been written if this patchset and Yang Shi's went first
(and in adjusting some of the comments in Alistair's, I did not
recognize the reason given for unmap_page() using try_to_unmap()
on file THPs, so substituted my own explanation).

As diff'ing the series will show, first thing is to delete the old

mm-rmap-make-try_to_unmap-void-function-fix-fix.patch
mm-rmap-make-try_to_unmap-void-function-fix.patch
mm-rmap-make-try_to_unmap-void-function.patch
mm-thp-replace-debug_vm-bug-with-vm_warn-when-unmap-fails-for-split.patch
mm-thp-check-total_mapcount-instead-of-page_mapcount-fix-fix-fix.patch

Then insert, from the mailing list

hughd1.patch
hughd2.patch
hughd3.patch
hughd4.patch
hughd5.patch
hughd6.patch
yangshi1.patch
yangshi2.patch
hughd7.patch

You'll name those differently of course.  I have just posted a v2
of hughd3.patch, with small changes suggested by Matthew; and after
sending this, I shall reply to it with a 6.1/7 and 6.2/7 which are
rediffs of Yang Shi's related patches to fit in here: 1/7 through
6.1/7 are Cc'ed to stable, 6.2/7 and 7/7 are cleanup and optimization.

I've placed them all together early, in the "mainline-later" section;
I expect you to be more cautious and place them later - anywhere before
mm-remove-special-swap-entry-functions.patch
works (but even the two not-for-stable ones need to come before that).

Then please replace those three "before" patches by their "after"s,
unless Alistair or someone else objects.

I've probably forgotten something,
but can't remember what it is :)
Let me know if I've screwed up.

Thanks,
Hugh

> 
> 1/7 mm/thp: fix __split_huge_pmd_locked() on shmem migration entry
> 2/7 mm/thp: try_to_unmap() use TTU_SYNC for safe DEBUG_VM splitting
> 3/7 mm/thp: fix vma_address() if virtual address below file offset
> 4/7 mm/thp: fix page_address_in_vma() on file THP tails
> 5/7 mm/thp: fix page_vma_mapped_walk() if huge page mapped by ptes
> 6/7 mm/thp: unmap_mapping_page() to fix THP truncate_cleanup_page()
> 7/7 mm/thp: remap_page() is only needed on anonymous THP
> 
>  include/linux/mm.h   |    3 
>  include/linux/rmap.h |    3 
>  mm/huge_memory.c     |   47 ++++++++----
>  mm/internal.h        |   54 ++++++++++----
>  mm/memory.c          |   40 ++++++++++
>  mm/page_vma_mapped.c |  163 ++++++++++++++++++++++++++-----------------
>  mm/pgtable-generic.c |    5 -
>  mm/rmap.c            |   39 +++++++---
>  mm/truncate.c        |   43 +++++------
>  9 files changed, 266 insertions(+), 131 deletions(-)
> 
> Hugh

Attachment: mmotm-adjust.tar
Description: Unix tar archive