On Thu 11-10-18 15:42:59, Mike Kravetz wrote:
From: Mike Kravetz mike.kravetz@oracle.com
mm: migration: fix migration of huge PMD shared pages
commit 017b1660df89f5fb4bfe66c34e35f7d2031100c7 upstream
The page migration code employs try_to_unmap() to try and unmap the source page. This is accomplished by using rmap_walk to find all vmas where the page is mapped. This search stops when page mapcount is zero. For shared PMD huge pages, the page map count is always 1 no matter the number of mappings. Shared mappings are tracked via the reference count of the PMD page. Therefore, try_to_unmap stops prematurely and does not completely unmap all mappings of the source page.
This problem can result is data corruption as writes to the original source page can happen after contents of the page are copied to the target page. Hence, data is lost.
This problem was originally seen as DB corruption of shared global areas after a huge page was soft offlined due to ECC memory errors. DB developers noticed they could reproduce the issue by (hotplug) offlining memory used to back huge pages. A simple testcase can reproduce the problem by creating a shared PMD mapping (note that this must be at least PUD_SIZE in size and PUD_SIZE aligned (1GB on x86)), and using migrate_pages() to migrate process pages between nodes while continually writing to the huge pages being migrated.
To fix, have the try_to_unmap_one routine check for huge PMD sharing by calling huge_pmd_unshare for hugetlbfs huge pages. If it is a shared mapping it will be 'unshared' which removes the page table entry and drops the reference on the PMD page. After this, flush caches and TLB.
mmu notifiers are called before locking page tables, but we can not be sure of PMD sharing until page tables are locked. Therefore, check for the possibility of PMD sharing before locking so that notifiers can prepare for the worst possible case.
Fixes: 39dde65c9940 ("shared page table for hugetlb page") Cc: stable@vger.kernel.org Signed-off-by: Mike Kravetz mike.kravetz@oracle.com
Acked-by: Michal Hocko mhocko@suse.com
with one minor nit. Please document why we are using the range invalidation (_start/_stop) only for hugetlb pages. Because this is a divergence from the upstream and we should better be explicit about that.
I've been testing without that in place and the memory corruption hasn't reproduced so far. I cannot give 100% confidence, though, due to some unrelated road blocks when testing. I might be lucky to not really have any notifier consumer for my workload as well.
Seeing an Acked-by from Jerome on this would be reassuring of course.