In the end, nothing serious -- a few days ago I bumped into an Amiga thread about an interesting race condition with cpusha bc + cinva bc but now I see there's no need to worry, the kernel uses only cpusha dc and optional combination with cinva ic.
But there's a very misleading comment in arch/cpu.S:
// Ozk: I think we'll have to check this code, as on my Milan040
// cpush(addr, size) dont work as expected. Gotta check this
// later. For now, "cpusha bc" is done on every cpush() call
this is not true anymore, I guess Ozk forgot to remove it.