On Thu, Nov 26, 2020 at 07:22:19PM +0100, Jan Kara wrote: > On Thu 26-11-20 14:04:15, Christoph Hellwig wrote: > > struct hd_struct *disk_get_part(struct gendisk *disk, int partno) > > { > > - struct hd_struct *part; > > + struct block_device *part; > > > > rcu_read_lock(); > > part = __disk_get_part(disk, partno); > > - if (part) > > - get_device(part_to_dev(part)); > > - rcu_read_unlock(); > > + if (!part) { > > + rcu_read_unlock(); > > + return NULL; > > + } > > > > - return part; > > + get_device(part_to_dev(part->bd_part)); > > + rcu_read_unlock(); > > + return part->bd_part; > > } > > This is not directly related to this particular patch but I'm wondering: > What prevents say del_gendisk() from racing with disk_get_part(), so that > delete_partition() is called just after we fetched 'part' pointer and the > last 'part' kobject ref is dropped before disk_get_part() calls > get_device()? I don't see anything preventing that and so we'd hand out > 'part' that is soon to be freed (after RCU grace period expires). At this point the hd_struct is already allocated together with the block_device, and thus only freed after the last block_device reference goes away plus the inode freeing RCU grace period. So the device model ref to part is indeed gone, but that simply does not matter any more.