It is normally safe for direct reclaim to enter filesystems
even when a page is locked - as can happen if ->writepage
allocates memory with GFP_KERNEL (which xfs does).
However if a localhost NFS mount is present, then a flush-*
thread might hold a page locked and then in direct reclaim,
ask nfs to commit an inode (nfs_release_page). When nfsd
performs the fsync it might try to lock the same page, which leads to
a deadlock.
A ->writepage should not allocate much memory, or do so very often, so
it is safe to set PF_FSTRANS, and this removes the possible deadlock.
This was not detected by lockdep as it doesn't monitor the page lock.
It was found as a real deadlock in testing.
Signed-off-by: NeilBrown <neilb@xxxxxxx>
---
mm/page-writeback.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 7106cb1aca8e..572e70b9a3f7 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -1909,6 +1909,7 @@ retry:
for (i = 0; i < nr_pages; i++) {
struct page *page = pvec.pages[i];
+ unsigned int pflags;
/*
* At this point, the page may be truncated or
@@ -1960,8 +1961,10 @@ continue_unlock:
if (!clear_page_dirty_for_io(page))
goto continue_unlock;
+ current_set_flags_nested(&pflags, PF_FSTRANS);
trace_wbc_writepage(wbc, mapping->backing_dev_info);
ret = (*writepage)(page, wbc, data);
+ current_restore_flags_nested(&pflags, PF_FSTRANS);
if (unlikely(ret)) {
if (ret == AOP_WRITEPAGE_ACTIVATE) {
unlock_page(page);
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
