[PATCH RFC v2 1/2] block: scan partitions for hidden disks

[John Garry]

On 8/19/26 10:13, Christoph Hellwig wrote:
> On Wed, Aug 19, 2026 at 10:12:50AM +0100, John Garry wrote:
>>> It also is used for giving a handle to NVMe devices that we can't support
>>> as block devices and for which we only want a handle to issue passthrough
>>> commands.  For those attempting a partition scan is rather suboptimal.
>>
>> We could set GENHD_FL_NO_PART in that case (as well as the hidden flag).
> 
> I guess we could do that.  It still feels wrong, though :

Maybe we just don't change the GENHD_FL_HIDDEN behavior. As long as 
/proc/diskstats and /proc/partitions don't show those per-path partitions.

It seems to me that we need to have a dedicated function to clone the 
partitions, and it would be called per NS after nvme_mpath_set_live() -> 
device_add_disk(). I am not sure if that works, though. We may get 
nvme_ns_head_submit_bio() looking up the per-path partition before they 
are available. In that case I would need to keep the path disabled until 
after the partition table is available.

> 
>>> As is in general doing an extra scan per controller when we know it must
>>> give the same results.  Wouldn't it make more sense to just propagate
>>> the partition table from the ns_head to each individual path?
>>
>> So you mean that we would have something which can clone the head gendisk
>> partitions onto the per-path gendisk by itering  >part_tbl and calling
>> something like blk_add_partition(), right?
> 
> Yes.

ok