On 2019/05/15 19:21, Jan Kara wrote:
> The question is how to fix this problem. The simplest fix I can see is that
> we'd just refuse to do LOOP_SET_FD if someone has the block device
> exclusively open as there are high chances such user will be unpleasantly
> surprised by the device changing under him. OTOH this has some potential
> for userspace visible regressions. But I guess it's worth a try. Something
> like attached patch?
(1) If I understand correctly, FMODE_EXCL is set at blkdev_open() only if O_EXCL
is specified. How can we detect if O_EXCL was not used, for the reproducer
( https://syzkaller.appspot.com/text?tag=ReproC&x=135385a8a00000 ) is not
using O_EXCL ?
(2) There seems to be no serialization. What guarantees that mount_bdev()
does not start due to preempted after the check added by this patch?
