Your second link is pointing to a loop in bpf_percpu_array_copy, but there's a similar loop in bpf_percpu_array_update which seems more relevant to the question you're asking. Guessing you meant to ask about the latter function.

Anyways, the API for updating percpu array from userspace differs a bit from the API BPF programs use, but still preserves the "updates are not synchronized" expectation.

	for_each_possible_cpu(cpu) {
		bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value + off, size);
		off += size;
	}

The passed in value buffer contains a separate value for each cpu, see tools/testing/selftests/bpf/test_maps.c in kernel source for an example use. The helper called from BPF, array_map_update_elem, updates a single cpu's value.

The semantics probably differ to keep the BPF_MAP_UPDATE_ELEM interface simple, since most bpf map types aren't percpu. Most BPF programs use percpu arrays to gather data from the program itself, or as a temporary 'heap' buffer space, and rarely need userspace to populate the percpu array except perhaps with some initial values. If you have a different usecase, though, we'd love to hear about it!