tree 04febe50e11e292a211e6bba10d39a362e6f46e1
parent 7e202553cff41e7ea3b866beb331588404ad6a7a
author Al Viro <viro@zeniv.linux.org.uk> 1580574405 +0000
committer Greg Kroah-Hartman <gregkh@linuxfoundation.org> 1580907943 +0000

vfs: fix do_last() regression

commit 6404674acd596de41fd3ad5f267b4525494a891a upstream.

Brown paperbag time: fetching ->i_uid/->i_mode really should've been
done from nd->inode.  I even suggested that, but the reason for that has
slipped through the cracks and I went for dir->d_inode instead - made
for more "obvious" patch.

Analysis:

 - at the entry into do_last() and all the way to step_into(): dir (aka
   nd->path.dentry) is known not to have been freed; so's nd->inode and
   it's equal to dir->d_inode unless we are already doomed to -ECHILD.
   inode of the file to get opened is not known.

 - after step_into(): inode of the file to get opened is known; dir
   might be pointing to freed memory/be negative/etc.

 - at the call of may_create_in_sticky(): guaranteed to be out of RCU
   mode; inode of the file to get opened is known and pinned; dir might
   be garbage.

The last was the reason for the original patch.  Except that at the
do_last() entry we can be in RCU mode and it is possible that
nd->path.dentry->d_inode has already changed under us.

In that case we are going to fail with -ECHILD, but we need to be
careful; nd->inode is pointing to valid struct inode and it's the same
as nd->path.dentry->d_inode in "won't fail with -ECHILD" case, so we
should use that.

Reported-by: "Rantala, Tommi T. (Nokia - FI/Espoo)" <tommi.t.rantala@nokia.com>
Reported-by: syzbot+190005201ced78a74ad6@syzkaller.appspotmail.com
Wearing-brown-paperbag: Al Viro <viro@zeniv.linux.org.uk>
Cc: stable@kernel.org
Fixes: d0cb50185ae9 ("do_last(): fetch directory ->i_mode and ->i_uid before it's too late")
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>