[PATCH 6.12.y 0/4] mm/shmem, swap: overdue shmem_swapin_folio() fixes

[Hugh Dickins <hughd@google.com>]

Greg Thelen (assisted by Gemini) has observed that 6.12.69 commit
a99f9a4669a0 ("mm/shmem, swap: fix race of truncate and swap entry split")
and its followup
dfc3ab6bd648 ("mm, shmem: prevent infinite loop on truncate race")
both rely on shmem_confirm_swap() to be returning int order or -1,
whereas in the 6.12.78-rc tree it is still returning bool 0 or 1.

Quite what the effect of that is, I've not tried to work out: luckily,
it's on an unlikely path which most of us find difficult to reproduce;
but better be fixed!

And applying the Stable-dep which made that change from bool to int order
reminds me of two "Cc stable" fixes from last year (and a followup fix to
one of them), which never got applied to the 6.12.y shmem_swapin_folio()
because of intervening mods.  Aside from minimizing the rejects, there
is little point in holding that "truncate and swap entry split" fix
without the more basic shmem swap order fixes found much earlier.

My own poor testing hit none of these issues: I hope others can verify.

1/4 mm: shmem: fix potential data corruption during shmem swapin
2/4 mm: shmem: avoid unpaired folio_unlock() in shmem_swapin_folio()
3/4 mm/shmem, swap: improve cached mTHP handling and fix potential hang
4/4 mm/shmem, swap: avoid redundant Xarray lookup during swapin

 mm/shmem.c | 97 ++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 80 insertions(+), 17 deletions(-)

Thanks,
Hugh