On Fri 27-11-20 16:36:15, Christoph Hellwig wrote: > On Fri, Nov 27, 2020 at 01:45:37PM +0100, Jan Kara wrote: > > > 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. > > > > Well, but once device model ref to part is gone, we're going to free the > > bdev inode ref as well. Thus there's nothing which pins the bdev containing > > hd_struct? > > > > But now as I'm thinking about it you later switch the device model reference > > to just pure inode reference and use igrab() which will reliably return > > NULL if the inode is on it's way to be destroyed so probably we are safe in > > the final state. > > igrab always succeeds. But we should switch to a tryget. No. If the inode is I_FREEING or I_WILL_FREE, it will return NULL... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
