When a THP is present in the page cache, we can return it several times, leading to userspace seeing the same data repeatedly if doing a read() that crosses a 64-page boundary. This is probably not a security issue (since the data all comes from the same file), but it can be interpreted as a transient data corruption issue. Fortunately, it is very rare as it can only occur when CONFIG_READ_ONLY_THP_FOR_FS is enabled, and it can only happen to executables. We don't often call read() on executables.
This bug is fixed differently in v5.17 by commit 6b24ca4a1a8d ("mm: Use multi-index entries in the page cache"). That commit is unsuitable for backporting, so fix this in the clearest way. It sacrifices a little performance for clarity, but this should never be a performance path in these kernel versions.
Fixes: cbd59c48ae2b ("mm/filemap: use head pages in generic_file_buffered_read") Cc: stable@vger.kernel.org # v5.15, v5.16 Link: https://lore.kernel.org/r/df3b5d1c-a36b-2c73-3e27-99e74983de3a@suse.cz/ Analyzed-by: Adam Majer amajer@suse.com Analyzed-by: Dirk Mueller dmueller@suse.com Bisected-by: Takashi Iwai tiwai@suse.de Reported-by: Vlastimil Babka vbabka@suse.cz Signed-off-by: Matthew Wilcox (Oracle) willy@infradead.org --- mm/filemap.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
diff --git a/mm/filemap.c b/mm/filemap.c index 82a17c35eb96..1293c3409e42 100644 --- a/mm/filemap.c +++ b/mm/filemap.c @@ -2354,8 +2354,12 @@ static void filemap_get_read_batch(struct address_space *mapping, break; if (PageReadahead(head)) break; - xas.xa_index = head->index + thp_nr_pages(head) - 1; - xas.xa_offset = (xas.xa_index >> xas.xa_shift) & XA_CHUNK_MASK; + if (PageHead(head)) { + xas_set(&xas, head->index + thp_nr_pages(head)); + /* Handle wrap correctly */ + if (xas.xa_index - 1 >= max) + break; + } continue; put_page: put_page(head);