Hi Zygo,

Would you like to test this diff?

Although I haven't find a solid reason yet, there is another report and
with the help from the reporter, it turns out that balance hangs at
relocating DATA_RELOC tree block.

After some more digging, DATA_RELOC tree doesn't need REF_COW bit at all
since we can't create snapshot for data reloc tree.

By removing the REF_COW bit, we could ensure that data reloc tree always
get cowed for relocation (just like extent tree), this would hugely
reduce the complexity for data reloc tree.

Not sure if this would help, but it passes my local balance run.

Thanks,
Qu