Revert "mm: don't try to NUMA-migrate COW pages that have other uses"
This reverts commit 80d47f5de5e311cbc0d01ebb6ee684e8f4c196c6.
- if the other mappings are paged out and they're not in use, they're
not interfering with NUMA balancing convergence. The only problem is
when there's more than one mapping in use, in such case it may be
impossible to converge into local memory, because local memory of
one task would still be remote memory for another task.
- if the page has additional pins that cannot be migrated, page
migration will sort it out later. All short term pins will hopefully
be gone by then. If this is meant to check for long term pins only,
then it should have checked the pincount, not the page count.
- if the page is in swapcache possibly with mapcount == 1 and
swapcount = 0 (PageAnonExclusive == true in upstream terms), there's
no reason to break NUMA balancing.
- On a side note the upstream PageAnonExclusive is also implementing
the equation "page_mapcount+page_swapcount == 1". The only
difference is in implementation terms because it provides the
result lockless atomically.
- PageAnonExclusive (if set) works safe precisely because it ignores
the page_count (see the exact math it implements in the previous
paragraph). Problem is PageAnonExclusive disappears and delivers
false negatives after fork() runs and then it's back to square one
hitting on the page_count and doing pin-breaking and unnecessary
extra cows.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
1 file changed
