Re: [PATCH 1/1] virtio: disable partitions scanning for no partitions block

Re: [PATCH 1/1] virtio: disable partitions scanning for no partitions block

[Christoph Hellwig]

On Mon, May 24, 2021 at 03:29:22PM +0100, Stefan Hajnoczi wrote:
> GENHD_FL_NO_PART_SCAN is not used much in other drivers. This makes me
> wonder if the same use case is addressed through other means with SCSI,
> NVMe, etc devices. Maybe Christoph or Jens can weigh in on whether
> adding a bit to disable partition scanning for a virtio-blk fits into
> the big picture?
> 
> Is your goal to avoid accidentally detecting partitions because it's
> confusing when that happens?

I'm really confused what the use case is here.  GENHD_FL_NO_PART_SCAN
has four users:

 - the block core setting it for hidden devices, for which the concept
   of paritions doesn't make sense.  Looking back this should have never
   used GENHD_FL_NO_PART_SCAN, and instead the partition scanning code
   should just check GENHD_FL_HIDDEN as well.
 - mmc uses it for boot partitions and rpmb.  I'm not even sure how
   these can be exposed as block devices as they don't require block
   granularity access IIRC, but if the allow block layer access there
   is no reason to ever set these flags.
 - loop is a bit of a mess.  IIRC the story is that originally the
   loop device did not support partitions, then in 2008 support for
   partitions was added by partitioning the minor number space, and
   then in 2011 support for partitions without that parameter was
   added using a new flag in the loop device creation ioctl that uses
   the extended dev_t space added since.  But even that might be
   something we can handled without that flag without breaking the
   userspace ABI
 - m64card sets it for no good reason at all

In other words: in a perfect would GENHD_FL_NO_PART_SCAN would not
exist, and it certainly should not be added to a new driver, never
mind a protocol.

Re: [PATCH 1/1] virtio: disable partitions scanning for no partitions block

[Ulf Hansson]

On Mon, 24 May 2021 at 16:57, Christoph Hellwig <hch@lst.de> wrote:
>
> On Mon, May 24, 2021 at 03:29:22PM +0100, Stefan Hajnoczi wrote:
> > GENHD_FL_NO_PART_SCAN is not used much in other drivers. This makes me
> > wonder if the same use case is addressed through other means with SCSI,
> > NVMe, etc devices. Maybe Christoph or Jens can weigh in on whether
> > adding a bit to disable partition scanning for a virtio-blk fits into
> > the big picture?
> >
> > Is your goal to avoid accidentally detecting partitions because it's
> > confusing when that happens?
>
> I'm really confused what the use case is here.  GENHD_FL_NO_PART_SCAN
> has four users:
>
>  - the block core setting it for hidden devices, for which the concept
>    of paritions doesn't make sense.  Looking back this should have never
>    used GENHD_FL_NO_PART_SCAN, and instead the partition scanning code
>    should just check GENHD_FL_HIDDEN as well.
>  - mmc uses it for boot partitions and rpmb.  I'm not even sure how
>    these can be exposed as block devices as they don't require block
>    granularity access IIRC, but if the allow block layer access there
>    is no reason to ever set these flags.

For RPMB, we have converted them into char devices, thus
GENHD_FL_NO_PART_SCAN is never set for them. The code needs a cleanup
to clarify this.

When it comes to eMMC boot partitions, those can be read/written to as
any other block device. Although, it's unlikely that they need
partitions as they are usually very small, 512Kb or 2MB in that
ballpark. At least, that was the thinking behind it when we added
GENHD_FL_NO_PART_SCAN for them.

If you want to drop GENHD_FL_NO_PART_SCAN for eMMC boot partitions, I
don't think it will be an issue.

>  - loop is a bit of a mess.  IIRC the story is that originally the
>    loop device did not support partitions, then in 2008 support for
>    partitions was added by partitioning the minor number space, and
>    then in 2011 support for partitions without that parameter was
>    added using a new flag in the loop device creation ioctl that uses
>    the extended dev_t space added since.  But even that might be
>    something we can handled without that flag without breaking the
>    userspace ABI
>  - m64card sets it for no good reason at all
>
> In other words: in a perfect would GENHD_FL_NO_PART_SCAN would not
> exist, and it certainly should not be added to a new driver, never
> mind a protocol.
> _______________________________________________
> Virtualization mailing list
> Virtualization@lists.linux-foundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Kind regards
Uffe