On 2022/03/16 20:22, Jan Kara wrote:
>> @@ -1244,7 +1244,7 @@ static int loop_clr_fd(struct loop_device *lo)
>> * <dev>/do something like mkfs/losetup -d <dev> causing the losetup -d
>> * command to fail with EBUSY.
>> */
>> - if (atomic_read(&lo->lo_refcnt) > 1) {
>> + if (lo->lo_disk->part0->bd_openers > 1) {
>
> But bd_openers can be read safely only under disk->open_mutex. So for this
> to be safe against compiler playing nasty tricks with optimizations, we
> need to either make bd_openers atomic_t or use READ_ONCE / WRITE_ONCE when
> accessing it.
Use of READ_ONCE() / WRITE_ONCE() are not for avoiding races but for making
sure that memory access happens only once. It is data_race() which is needed
for tolerating and annotating races. For example, data_race(lo->lo_state) is
needed when accessing lo->lo_state without lo->lo_mutex held.
Use of atomic_t for lo->lo_disk->part0->bd_openers does not help, for currently
lo->lo_mutex is held in order to avoid races. That is, it is disk->open_mutex
which loop_clr_fd() needs to hold when accessing lo->lo_disk->part0->bd_openers.
