Re: [PATCHv14 bpf-next 3/6] xdp: add a new helper for dev map multicast support

Toke Høiland-Jørgensen wrote:

Hangbin Liu [off-list ref] writes:

quoted

Hi John,

Thanks for the reviewing.

On Sun, Jan 17, 2021 at 04:10:40PM -0800, John Fastabend wrote:

quoted

quoted

+ * 		The forwarding *map* could be either BPF_MAP_TYPE_DEVMAP or
+ * 		BPF_MAP_TYPE_DEVMAP_HASH. But the *ex_map* must be
+ * 		BPF_MAP_TYPE_DEVMAP_HASH to get better performance.

Would be good to add a note ex_map _must_ be keyed by ifindex for the
helper to work. Its the obvious way to key a hashmap, but not required
iirc.

OK, I will.

[...]

quoted

quoted

WRITE_ONCE(ri->ex_map)?

quoted

 	WRITE_ONCE(ri->map, NULL);

So we needed write_once, read_once pairs for ri->map do we also need them in
the ex_map case?

Toke said this is no need for this read/write_once as there is already one.

https://lore.kernel.org/bpf/87r1wd2bqu.fsf@toke.dk/ (local)

And then I corrected that after I figured out the real reason :)

https://lore.kernel.org/bpf/878si2h3sb.fsf@toke.dk/ (local) - Quote:

quoted

The READ_ONCE() is not needed because the ex_map field is only ever read
from or written to by the CPU owning the per-cpu pointer. Whereas the
'map' field is manipulated by remote CPUs in bpf_clear_redirect_map().
So you need neither READ_ONCE() nor WRITE_ONCE() on ex_map, just like
there are none on tgt_index and tgt_value.

-Toke

Hi Hangbin, please add a comment above that code block to remind us
why the READ_ONCE/WRITE_ONCE is not needed or add it in the commit
message so we don't lose it. It seems we've hashed it over already,
but I forgot after the holidays/break so presumably I'll forget next
time I read this code as well and commit-msg or comment will help.

Thanks,
John