Nathan Lynch [off-list ref] writes:
I'm satisfied that this change does not reintroduce that problem. The
removal path still derives the node without going to the device tree,
but now performs a more efficient lookup.
Er, actually it's slightly less efficient in the remove path, as you
note in your commit message. This doesn't alter my opinion of the
change. The higher-level algorithmic inefficiencies (e.g. all the
for_each_drmem_lmb loops) in this code are more of a concern.