mkfs.btrfs fails on s390, possible endianness error

[cmurf]

Downstream Bug 2214522 - Fedora 37: mkfs.btrfs fails on s390 with btrfs-progs 6.3.1 release

# mkfs.btrfs /dev/dasdd1
btrfs-progs v6.3.1
See https://btrfs.readthedocs.io/ for more information.

ERROR: superblock magic doesn't match
NOTE: several default settings have changed in version 5.15, please make sure
      this does not affect your deployments:
      - DUP for metadata (-m dup)
      - enabled no-holes (-O no-holes)
      - enabled free-space-tree (-R free-space-tree)

Invalid mapping for 1081344-1097728, got 17592186044416-17592190238720
Couldn't map the block 1081344
ERROR: cannot read chunk root
ERROR: open ctree failed

System: Fedora 37 Fails on both 6.3 and 6.4 kernels.

[cmurf]

We didn't build btrfs-progs 6.3 in Fedora so it could be a regression in 6.3 or 6.3.1.

[sharkcz]

I am going to bisect it, easily reproduced with make test on a s390x machine. Unless someone beats me being faster :-)

[sharkcz]

looks to me 6.3 is OK

[sharkcz]

c979ffd7872257beeae9ddc930f2be2423597e07 is the first bad commit and 164bc10dfc08d8754d23ef0d6d7281e139d6cd53 is the last good

[sharkcz]

If access to a s390x system is needed, please let me know, I can arrange that.

[adam900710]

Mind to give me the access to an s390x system? You can contact me through wqu@suse.com.

[adam900710]

Great thanks to @sharkcz for the s390x machine.

The direct cause is pretty straightforward, by somehow preprocessor didn't define __BIG_ENDIAN__, so we got incorrect keys and screws up everything.

Still looking into what's the proper macro to use to determine the endian.

[sharkcz]

You might use AC_C_BIGENDIAN if autotools are being used or rely on the compiler and __BYTE_ORDER__ being __ORDER_LITTLE_ENDIAN__ or __ORDER_BIG_ENDIAN__. There are some guides online about the right detection procedures.

[adam900710]

I found <bits/endian.h> from glibc providing all the endianness checks, thus I would go __BYTE_ORDER checks against __LITTLE_ENDIAN to send the proper fix.

The synced accessor.h is using the kernel endianness check, which doesn't work with user space headers.

[sharkcz]

There is already an endianness check in https://github.com/kdave/btrfs-progs/blob/master/configure.ac#L45, so I would probably go with it, see eg. https://github.com/chrislim2888/IP2Location-C-Library/pull/58/files for usage. Rather than introducing another dependency and method ...

[kdave]

The header is the copy from kernel so we need to use a check that's compatible on the source level, so the autoconf WORD_BIGENDIAN is not suitable here.

[kdave]

kerncompat.h or kernel-lib/bitops.h use bitops.h:#if __BYTE_ORDER == __BIG_ENDIAN

[kdave]

Qu's fix applied to devel.

[sharkcz]

The header is the copy from kernel so we need to use a check that's compatible on the source level, so the autoconf WORD_BIGENDIAN is not suitable here.

yes, it makes sense, thanks