| Message ID | alpine.LSU.2.11.2104211735430.3299@eggly.anvils (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/2] mm/filemap: fix find_lock_entries hang on 32-bit THP | expand |
On Wed, Apr 21, 2021 at 05:37:33PM -0700, Hugh Dickins wrote: > - if (!xa_is_value(page) && PageTransHuge(page)) > - xas_set(&xas, page->index + thp_nr_pages(page)); > + if (!xa_is_value(page) && PageTransHuge(page)) { > + unsigned int nr_pages = thp_nr_pages(page); > + > + /* Final THP may cross MAX_LFS_FILESIZE on 32-bit */ > + xas_set(&xas, page->index + nr_pages); > + if (xas.xa_index < nr_pages) > + break; > + } Aargh. We really need to get the multi-index support in; this works perfectly when the xas_set() hack isn't needed any more.
--- 5.12-rc8/mm/filemap.c 2021-02-26 19:42:39.812156085 -0800 +++ linux/mm/filemap.c 2021-04-20 23:20:20.509464440 -0700 @@ -1969,8 +1969,14 @@ unlock: put: put_page(page); next: - if (!xa_is_value(page) && PageTransHuge(page)) - xas_set(&xas, page->index + thp_nr_pages(page)); + if (!xa_is_value(page) && PageTransHuge(page)) { + unsigned int nr_pages = thp_nr_pages(page); + + /* Final THP may cross MAX_LFS_FILESIZE on 32-bit */ + xas_set(&xas, page->index + nr_pages); + if (xas.xa_index < nr_pages) + break; + } } rcu_read_unlock();
No problem on 64-bit, or without huge pages, but xfstests generic/308 hung uninterruptibly on 32-bit huge tmpfs. Since 4.13's 0cc3b0ec23ce ("Clarify (and fix) MAX_LFS_FILESIZE macros"), MAX_LFS_FILESIZE is only a PAGE_SIZE away from wrapping 32-bit xa_index to 0, so the new find_lock_entries() has to be extra careful when handling a THP. Fixes: 5c211ba29deb ("mm: add and use find_lock_entries") Signed-off-by: Hugh Dickins <hughd@google.com> --- mm/filemap.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)