There is a race between xas_split() and xas_load() which can result in
the wrong page being returned, and thus data corruption.  Fortunately,
it's hard to hit (syzbot took three months to find it) and often guarded
with VM_BUG_ON().

The anatomy of this race is:

thread A			thread B
order-9 page is stored at index 0x200
				lookup of page at index 0x274
page split starts
				load of sibling entry at offset 9
stores nodes at offsets 8-15
				load of entry at offset 8

The entry at offset 8 turns out to be a node, and so we descend into it,
and load the page at index 0x234 instead of 0x274.  This is hard to fix
on the split side; we could replace the entire node that contains the
order-9 page instead of replacing the eight entries.  Fixing it on
the lookup side is easier; just disallow sibling entries that point
to nodes.  This cannot ever be a useful thing as the descent would not
know the correct offset to use within the new node.

The test suite continues to pass, but I have not added a new test for
this bug.

Reported-by: syzbot+cf4cf13056f85dec2c40@syzkaller.appspotmail.com
Tested-by: syzbot+cf4cf13056f85dec2c40@syzkaller.appspotmail.com
Fixes: 6b24ca4a ("mm: Use multi-index entries in the page cache")
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>