This is a request for adding the following patches to stable 5.10.y.
Poisoned shmem and hugetlb pages are removed from the pagecache. Subsequent access to the offset in the file results in a NEW zero filled page. Application code does not get notified of the data loss, and the only 'clue' is a message in the system log. Data loss has been experienced by real users.
This was addressed upstream. Most commits were marked for backports, but some were not. This was discussed here [1] and here [2].
Patches apply cleanly to v5.4.224 and pass tests checking for this specific data loss issue. LTP mm tests show no regressions.
All patches except 4 "mm: hwpoison: handle non-anonymous THP correctly" required a small bit of change to apply correctly: mostly for context.
linux-mm Cc'ed as it would be great to get at least an ACK from others familiar with this issue.
[1] https://lore.kernel.org/linux-mm/Y2UTUNBHVY5U9si2@monkey/ [2] https://lore.kernel.org/stable/20221114131403.GA3807058@u2004/
James Houghton (1): hugetlbfs: don't delete error page from pagecache
Yang Shi (5): mm: hwpoison: remove the unnecessary THP check mm: filemap: check if THP has hwpoisoned subpage for PMD page fault mm: hwpoison: refactor refcount check handling mm: hwpoison: handle non-anonymous THP correctly mm: shmem: don't truncate page if memory failure happens
fs/hugetlbfs/inode.c | 13 ++-- include/linux/page-flags.h | 23 ++++++ mm/huge_memory.c | 2 + mm/hugetlb.c | 4 + mm/memory-failure.c | 153 ++++++++++++++++++++++++------------- mm/memory.c | 9 +++ mm/page_alloc.c | 4 +- mm/shmem.c | 51 +++++++++++-- 8 files changed, 191 insertions(+), 68 deletions(-)