BRT: cross-dataset block cloning fails with EXDEV on Linux 6.11

[intelfx]

System information

Type	Version/Name
Distribution Name	Arch
Distribution Version	current
Kernel Version	6.11.10
Architecture	x86_64
OpenZFS Version	2.3.0-rc3

Describe the problem you're observing

Similarly to concerns from #405 (#405 (comment)), attempting to do a cp --reflink=always between two sibling datasets fails with EXDEV. However, unlike that case, no encryption is used.

Describe how to reproduce the problem

# ccat /sys/module/zfs/parameters/*clone*
/sys/module/zfs/parameters/zfs_bclone_enabled     1
/sys/module/zfs/parameters/zfs_bclone_wait_dirty  1

# modprobe brd rd_nr=1 rd_size=$(bscalcq -k '50G')

# CFG_ZPOOL_CREATE_OPTS=(
        -o ashift=12
        -o feature@block_cloning=enabled
        -O compression=lz4
        -O acltype=posixacl -O xattr=sa -O dnodesize=auto
)
# zpool create "${CFG_ZPOOL_CREATE_OPTS[@]}" test /dev/ram0
# parallel zfs create -p ::: test/ROOT/test{1,2,3,4}
# parallel "dd if=/dev/urandom of=/test/ROOT/test{}/file{} bs=1M count=1024" ::: 1 2 3 4
# zpool sync test

# cp -a --reflink=always /test/ROOT/test2/file2 /test/ROOT/test4/file4
/usr/bin/cp: failed to clone '/test/ROOT/test4/file4' from '/test/ROOT/test2/file2': Invalid cross-device link

[amotin]

@intelfx Could you similar to quoted post run strace on the cp to make sure what syscall fails there? Just to make sure it uses copy_file_range(), since other system calls have that limitation on kernel level.

[intelfx]

@amotin

Yep, it uses FICLONE (which fails), presumably because I said --reflink=always and it is not possible to guarantee with copy_file_range() that the blocks were cloned rather than just copied in-kernel.

I'm assuming zfs uses a different struct super_block for each dataset, and consequently runs afoul of the same-sb check in the generic FICLONE implementation?

[snajpa]

AFAIK this isn't supported (due to the reason @intelfx outlies) so I'd say it's more of a Feature request than a Defect (how can one change it?)

[robn]

@intelfx yes, this is the reason.

Duplicate of #15345.