Re: [PATCH v12 13/13] dm: add power-of-2 target for zoned devices with non power-of-2 zone sizes

[Mike Snitzer <snitzer@redhat.com>]

On Fri, Sep 02 2022 at  4:55P -0400,
Mike Snitzer <snitzer@redhat.com> wrote:

> On Tue, Aug 23 2022 at  8:18P -0400,
> Pankaj Raghav <p.raghav@samsung.com> wrote:
> 
> > Only zoned devices with power-of-2(po2) number of sectors per zone(zone
> > size) were supported in linux but now non power-of-2(npo2) zone sizes
> > support has been added to the block layer.
> > 
> > Filesystems such as F2FS and btrfs have support for zoned devices with
> > po2 zone size assumption. Before adding native support for npo2 zone
> > sizes, it was suggested to create a dm target for npo2 zone size device to
> > appear as a po2 zone size target so that file systems can initially
> > work without any explicit changes by using this target.
> > 
> > The design of this target is very simple: remap the device zone size to
> > the zone capacity and change the zone size to be the nearest power of 2
> > value.
> > 
> > For e.g., a device with a zone size/capacity of 3M will have an equivalent
> > target layout as follows:
> > 
> > Device layout :-
> > zone capacity = 3M
> > zone size = 3M
> > 
> > |--------------|-------------|
> > 0             3M            6M
> > 
> > Target layout :-
> > zone capacity=3M
> > zone size = 4M
> > 
> > |--------------|---|--------------|---|
> > 0             3M  4M             7M  8M
> > 
> > The area between target's zone capacity and zone size will be emulated
> > in the target.
> > The read IOs that fall in the emulated gap area will return 0 filled
> > bio and all the other IOs in that area will result in an error.
> > If a read IO span across the emulated area boundary, then the IOs are
> > split across them. All other IO operations that span across the emulated
> > area boundary will result in an error.
> > 
> > The target can be easily created as follows:
> > dmsetup create <label> --table '0 <size_sects> po2zone /dev/nvme<id>'
> > 
> > Note that the target does not support partial mapping of the underlying
> > device.
> > 
> > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com>
> > Suggested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
> > Suggested-by: Damien Le Moal <damien.lemoal@wdc.com>
> > Suggested-by: Hannes Reinecke <hare@suse.de>
> 
> 
> This target needs more review from those who Suggested-by it.
> 
> And the header and docs needs to address:
> 
> 1) why is a partial mapping of the underlying device disallowed?
> 2) why is it assumed all IO is read-only? (talk to me and others like
>    we don't know the inherent limitations of this class of zoned hw)
> 
> On a code level:
> 1) are you certain you're properly failing all writes?
>    - are writes allowed to the "zone capacity area" but _not_
>      allowed to the "emulated zone area"? (if yes, _please document_). 
> 2) yes, you absolutely need to implement the .status target_type hook
>    (for both STATUS and TABLE).
> 3) really not loving the nested return (of DM_MAPIO_SUBMITTED or
>    DM_MAPIO_REMAPPED) from methods called from dm_po2z_map().  Would
>    prefer to not have to do a depth-first search to see where and when
>    dm_po2z_map() returns a DM_MAPIO_XXX unless there is a solid
>    justification for it.  To me it just obfuscates the DM interface a
>    bit too much. 
> 
> Otherwise, pretty clean code and nothing weird going on.
> 
> I look forward to seeing your next (final?) revision of this patchset.

Thinking further.. I'm left confused about just what the heck this
target is assuming.

E.g.: feels like its exposing a readonly end of the zone is very
bi-polar... yet no hint to upper layer it shouldn't write to that
read-only end (the "emulated zone").. but there has to be some zoned
magic assumed?  And I'm just naive?

Mike