Re: [PATCH 3/8] blk: blkmap: Add basic infrastructure

[Tobias Waldekranz <tobias@waldekranz.com>]

On mån, feb 06, 2023 at 21:02, Simon Glass <sjg@chromium.org> wrote:
> Hi Tobias,
>
> On Mon, 6 Feb 2023 at 01:30, Tobias Waldekranz <tobias@waldekranz.com> wrote:
>>
>> On fre, feb 03, 2023 at 17:20, Simon Glass <sjg@chromium.org> wrote:
>> > Hi Tobias,
>> >
>> > On Fri, 3 Feb 2023 at 02:38, Tobias Waldekranz <tobias@waldekranz.com> wrote:
>> >>
>> >> On ons, feb 01, 2023 at 13:20, Simon Glass <sjg@chromium.org> wrote:
>> >> > Hi Tobias,
>> >>
>> >> Hi Simon,
>> >>
>> >> Thanks for the review!
>> >>
>> >> > On Wed, 1 Feb 2023 at 11:10, Tobias Waldekranz <tobias@waldekranz.com> wrote:
>> >> >>
>> >> >> blkmaps are loosely modeled on Linux's device mapper subsystem. The
>> >> >> basic idea is that you can create virtual block devices whose blocks
>> >> >> can be backed by a plethora of sources that are user configurable.
>> >> >>
>> >> >> This change just adds the basic infrastructure for creating and
>> >> >> removing blkmap devices. Subsequent changes will extend this to add
>> >> >> support for actual mappings.
>> >> >>
>> >> >> Signed-off-by: Tobias Waldekranz <tobias@waldekranz.com>
>> >> >> ---
>> >> >>  MAINTAINERS                      |   6 +
>> >> >>  disk/part.c                      |   1 +
>> >> >>  drivers/block/Kconfig            |  18 ++
>> >> >>  drivers/block/Makefile           |   1 +
>> >> >>  drivers/block/blk-uclass.c       |   1 +
>> >> >>  drivers/block/blkmap.c           | 275 +++++++++++++++++++++++++++++++
>> >> >>  include/blkmap.h                 |  15 ++
>> >> >>  include/dm/uclass-id.h           |   1 +
>> >> >>  include/efi_loader.h             |   4 +
>> >> >>  lib/efi_loader/efi_device_path.c |  30 ++++
>> >> >>  10 files changed, 352 insertions(+)
>> >> >>  create mode 100644 drivers/block/blkmap.c
>> >> >>  create mode 100644 include/blkmap.h
>> >> >>
>> >
>> > [..]
>> >
>> >> > This needs to be created as part of DM.  See how host_create_device()
>> >> > works. It attaches something to the uclass and then creates child
>> >> > devices from there. It also operations (struct host_ops) but you don't
>> >> > need to do that.
>> >> >
>> >> > Note that the host commands support either an label or a devnum, which
>> >> > I think is useful, so you might copy that?
>> >> >
>> >>
>> >> I took a look at the hostfs implementation. I agree that labels are much
>> >> nicer than bare integers. However, for block maps the idea is to fit in
>> >> to the existing filesystem infrastructure. Addressing block devices
>> >> using the "<interface> <dev>[:<part>]" pattern seems very well
>> >> established...
>> >
>> > You can still do that, so long as the labels are "0" and "1", etc. But
>> > it lets us move to a more flexible system in future.
>> >
>> >>
>> >> >> +{
>> >> >> +       static struct udevice *dev;
>> >> >> +       int err;
>> >> >> +
>> >> >> +       if (dev)
>> >> >> +               return dev;
>> >> >> +
>> >> >> +       err = device_bind_driver(dm_root(), "blkmap_root", "blkmap", &dev);
>> >> >> +       if (err)
>> >> >> +               return NULL;
>> >> >> +
>> >> >> +       err = device_probe(dev);
>> >> >> +       if (err) {
>> >> >> +               device_unbind(dev);
>> >> >> +               return NULL;
>> >> >> +       }
>> >> >
>> >> > Should not be needed as probing children will cause this to be probed.
>> >> >
>> >> > So this function just becomes
>> >> >
>> >> > uclass_first_device(UCLASS_BLKDEV, &
>> >> >
>> >> >> +
>> >> >> +       return dev;
>> >> >> +}
>> >> >> +
>> >> >> +int blkmap_create(int devnum)
>> >> >
>> >> > Again, please drop the use of devnum and use devices. Here you could
>> >> > use a label, perhaps?
>> >>
>> >> ...which is why I don't think a label is going to fly here. Let's say I
>> >> create a new ramdisk with a label instead, e.g.:
>> >>
>> >>     blkmap create rd
>> >>     blkmap map rd 0 0x100 mem ${loadaddr}
>> >>
>> >> How do I know which <dev> to supply to, e.g.:
>> >>
>> >>     ls blkmap <dev> /boot
>> >>
>> >> It seems like labels are a hostfs-specific feature, or am I missing
>> >> something?
>> >
>> > We have the same problem with hostfs, since we have not implemented
>> > labels in block devices. For now you must use integer labels. But we
>> > will get there.
>>
>> But there is no connection to the devnum that is allocated internally by
>> U-Boot. Here's an experiment I just ran:
>>
>> I created two squashfs images containing a single directory:
>>
>>     zero.squashfs:
>>      i_am_zero
>>
>>     one.squashfs:
>>      i_am_one
>>
>> Then I added a binding to them:
>>
>>     => host bind 1 one.squashfs
>>     => host bind 0 zero.squashfs
>>
>> When accessing them, we see that the existing filesystem utilities work
>> on the internally generated devnums, ignoring the labels:
>>
>>     => ls host 1
>>                 i_am_zero/
>>
>>     0 file(s), 1 dir(s)
>>
>>     => ls host 0
>>                 i_am_one/
>>
>>     0 file(s), 1 dir(s)
>>
>>     =>
>>
>> Doesn't it therefore make more sense to stick with the established
>> abstraction?
>
> It is pretty clear that this is a bit silly though :-)

As in "this specific example will never be used in practice"? Sure :)

My point was just that the approach to stick to integer labels is
brittle, since there is no connection between the label and the devnum
used by existing commands.

Here's an even simpler example that might actually trip up a user:

    => host bind 1 one.squashfs
    => ls host 1
    ** Bad device specification host 1 **
    Couldn't find partition host 1
    => ls host 0
                i_am_one/

    0 file(s), 1 dir(s)

    =>

> I mean, right now, it would be easier to stick with numbered devices.
> But we want to be able to support named devices (e.g. using struct
> udevice->name). A good way to be forward compatible is to support a
> label today.
>
> When we do get to it, the less code that has piled up and needs
> converting, the more likely it is to happen.

I completely understand, and agree with, the direction you want to take
this.

> Sure, you have the problem as above, but mostly people are only going
> to use one of them, so it doesn't matter.
>
> We could also have a way of obtaining a number from a label, if you
> want to go that far. But I suggest just telling people to use labels
> like "1" and "0" which should work.

Alright, well, if that is acceptable then I'll do it that way. For my
own piece of mind, I think I'll also add some way of safely doing the
reverse mapping for any label. Does the following look ok?

    blkmap get <label> dev <var>

This way you could extend it with other attributes in the future
(e.g. size).