On Tue, Nov 24, 2020 at 04:18:49PM -0500, Tejun Heo wrote:

Hello,

Please see lkml.kernel.org/r/X708BTJ5njtbC2z1@mtj.duckdns.org for a few nits
on the previous version.

Thanks, I've addressed the mapping_set_gfp mask nit and updated the
commit log.  As Jan pointed also pointed out I think we do need the
remove_inode_hash.

I might be confused but am wondering whether RCU is needed. It's currently
used to ensure that gendisk is accessible in the blkdev_get path but
wouldn't it be possible to simply pin gendisk from block_devices? The
gendisk and hd_structs hold the base refs of the block_devices and in turn
the block_devices pin the gendisk. When the gendisk gets deleted, it puts
the base refs of the block devices but stays around while the block devices
are being accessed.

Yes, that sounds sensible.  I'll take a look.

Also, would it make sense to separate out lookup_sem removal? I *think* it's
there to ensure that the same bdev doesn't get associated with old and new
gendisks at the same time but can't wrap my head around how it works
exactly. I can see that this may not be needed once the lifetimes of gendisk
and block_devices are tied together but that may warrant a bit more
explanation.

Jan added lookup_sem in commit 56c0908c855afbb to prevent a three way
race between del_gendisk and blkdev_open due to the weird bdev vs
gendisk lifetime rules.  None of that can happen with the new lookup
scheme.