Re: Commit d5fd456c88aba4fcf77d35fe38024a8d5c814686 - "loopdev: use LOOP_CONFIG ioctl" broke loop on x86-64 w/ 32 bit userspace

["Krzysztof Olędzki" <ole@ans.pl>]

On 2021-07-27 at 18:24, Jens Axboe wrote:
> On 7/27/21 4:56 PM, Krzysztof Olędzki wrote:
>> On 2021-07-27 at 15:39, Krzysztof Olędzki wrote:
>>> On 2021-07-27 at 14:53, Krzysztof Olędzki wrote:
>>>> Hi,
>>>>
>>>> I have a number of (older) systems that are still based on 32 bit
>>>> userspace but are running a relatively modern 64 bit kernel -
>>>> 5.4-stable, where BTW - LOOP_CONFIGURE is not yet available.
>>>>
>>>> I noticed that starting with util-linux-2.37 it is no longer possible to
>>>> mount images using loop:
>>>>
>>>> # mount /usr/install/iso/systemrescue-8.04-amd64.iso /mnt/cdrom
>>>> mount: /mnt/cdrom: failed to setup loop device for
>>>> /usr/install/iso/systemrescue-8.04-amd64.iso.
>>>>
>>>> Reverting d5fd456c88aba4fcf77d35fe38024a8d5c814686 fixes the problem:
>>>>
>>>> /tmp/util-linux-2.37# ./mount
>>>> /usr/install/iso/systemrescue-8.04-amd64.iso /mnt/cdrom
>>>> mount: /mnt/cdrom: WARNING: source write-protected, mounted read-only.
>>>>
>>>> I have not tested if 32 bit kernel + 32 bit userspace is also affected,
>>>> but 64 bit kernel + 64 bit userspace works.
>>>
>>> Some debugging data:
>>>
>>> 30399: loopdev:      CXT: [0xff8d0f98]: using loop-control
>>> 30399: loopdev:      CXT: [0xff8d0f98]: loop0 name assigned
>>> 30399: loopdev:      CXT: [0xff8d0f98]: find_unused by loop-control [rc=0]
>>> 30399: libmount:     LOOP: [0x57cbbcb0]: trying to use /dev/loop0
>>> 30399: loopdev:      CXT: [0xff8d0f98]: set backing file=/usr/install/iso/systemrescue-8.04-amd64.iso
>>> 30399: loopdev:      CXT: [0xff8d0f98]: set flags=4
>>> 30399: loopdev:    SETUP: [0xff8d0f98]: device setup requested
>>> 30399: loopdev:    SETUP: [0xff8d0f98]: backing file open: OK
>>> 30399: loopdev:      CXT: [0xff8d0f98]: open /dev/loop0 [rw]: Success
>>> 30399: loopdev:    SETUP: [0xff8d0f98]: device open: OK
>>> 30399: loopdev:    SETUP: [0xff8d0f98]: LOOP_CONFIGURE failed: Inappropriate ioctl for device
>>> 30399: loopdev:    SETUP: [0xff8d0f98]: failed [rc=-25]
>>> 30399: libmount:     LOOP: [0x57cbbcb0]: failed to setup device
>>> 30399: loopdev:      CXT: [0xff8d0f98]: de-initialize
>>> 30399: loopdev:      CXT: [0xff8d0f98]: closing old open fd
>>> 30399: loopdev:     ITER: [0xff8d1168]: de-initialize
>>> 30399: libmount:      CXT: [0x57cbbcb0]: mount: preparing failed
>>> 30399: libmount:      CXT: [0x57cbbcb0]: excode: rc=32 message="failed to setup loop device for /usr/install/iso/systemrescue-8.04-amd64.iso"
>>> mount: /mnt/cdrom: failed to setup loop device for /usr/install/iso/systemrescue-8.04-amd64.iso.
>>> 30399: libmount:      CXT: [0x57cbbcb0]: <---- reset [status=0] ---->
>>>
>>> Seems like the code expects EINVAL (-22) but gets ENOTTY (-25), confirmed with strace:
>>> ioctl(4, LOOP_CONFIGURE, {fd=3, block_size=0, info={lo_offset=0, lo_number=0, lo_flags=LO_FLAGS_AUTOCLEAR, lo_file_name="/usr/install/iso/systemrescue-8.04-amd64.iso", ...}}) = -1 ENOTTY (Inappropriate ioctl for device)
>>>
>>> Indeed, changing the code from:
>>>      if (errno != EINVAL)
>>> to:
>>>      if (errno != EINVAL && errno != ENOTTY)
>>> allows it to work.
>>>
>>> Not that with 64-bit userspace, kernel returns EINVAL:
>>>
>>> ioctl(4, LOOP_CONFIGURE, {fd=3, block_size=0, info={lo_offset=0, lo_number=0, lo_flags=LO_FLAGS_AUTOCLEAR, lo_file_name="/usr/src/PACKAGES/systemrescue-8.04-amd64.iso", ...}}) = -1 EINVAL (Invalid argument)
>>
>> ... which is because lo_compat_ioctl returns -ENOIOCTLCMD for
>> unsupported cmds, while lo_ioctl returns -EINVAL via lo_simple_ioctl.
>>
>> And vfs_ioctl returns -ENOTTY for -ENOIOCTLCMD.
>>
>> Now the question is if this inconsistency is intended? :)
> 
> That's unfortunate, but probably not something that can get corrected at
> this time. The correct return value for an unknown ioctl is -ENOTTY
> (ENOIOCTLCMD isn't user visible, should get turned into that).

Yes, it does - as I said, vfs_ioctl handles this properly. However, this 
only works for .compat_ioctl (via mentioned lo_compat_ioctl which 
returns -ENOIOCTLCMD) but not for .ioctl (via lo_ioctl, which returns 
-EINVAL).


> But
> current behavior is set in stone at this point, even if it is
> technically incorrect.

Agreed. And even if this could be somehow fixed in further kernels, I 
believe we still need to fix the userspace to support and properly 
handle all the existing kernels.

So, to confirm - checking for both EINVAL and ENOTTY after 
LOOP_CONFIGURE is the proper way of taking care this?

https://git.kernel.org/pub/scm/utils/util-linux/util-linux.git/tree/lib/loopdev.c?id=d4423cce9b9001c9de7ebc6f64f6cc2bb854944c#n1362

Thanks,
  Krzysztof