On 10/24/25 12:03 AM, Jakub Kicinski wrote:

On Thu, 23 Oct 2025 18:37:09 +0000 Asbjørn Sloth Tønnesen wrote:

quoted

  > C code already does this, right? We just collect the attributes
  > completely ignoring the index.

In the userspace C code, for sub-type nest the index is preserved
in struct member idx, and elided for other sub-types.

I see it now, I guess I added it for get but forgot to obey it
for put :(

I can fix that up in v2.

Do we wanna do the same for sub-type nest in the python client?

quoted

  > So why do we need to extend the spec.

For me it's mostly the naming "indexed-array". Earlier it was
"array-nest", then we renamed it to "indexed-array" because
it doesn't always contain nests.  I think it's counter-intuitive
to elide the index by default, for something called "indexed-array".
The majority of the families using it don't care about the index.

What if we called it "ordered-array", and then always counted from 1
(for families that cares) when packing in user-space code?

  > Have you found any case where the index matters and can be
  > non-contiguous (other than the known TC kerfuffle).

IFLA_OFFLOAD_XSTATS_HW_S_INFO could be re-defined as a nest,
IFLA_OFFLOAD_XSTATS_L3_STATS is the only index atm.

Isn't that pretty much always true? If the index is significant
the whole thing could be redefined as a nest, with names provided
in the spec?

I guess it depends on the range of indexes, the aboveIFLA_OFFLOAD_XSTATS_L3_STATS is aka. 3, so the new nest policy would be
(0..2 = unused, 3 = .._L3_STATS}, which is fine, but for higher indexes
it might get a bit silly, but I haven't found any too high indexes.

NLBL_CIPSOV4_A_MLSCATLST has NLBL_CIPSOV4_A_MLSCAT which is 11, but that
corner seams very dusty, so I don't expect that to get YNL support.

Nest + multi-attr is great for the cases where we want per-attr typing,
like IFLA_PROP_LIST / IFLA_ALT_IFNAME, but a bit awkward if they all
are nests with the same policy.

quoted

IFLA_INET_CONF / IFLA_INET6_CONF is on input, but those are
also special by having different types depending on direction.

I found a bunch of other ones, using a static index, but they
can also be defined as a multi-attr wrapped in an additional
outer nest, like IFLA_VF_VLAN_LIST already is.

Multi-attr with an outer nest should at least solve your wg problem
I guess? If all the attrs have type of 0 we can make it a multi-attr.

WG don't use the index, so it's fine with incrementing from 1.

Converting it to nest + multi-attr would require extra pollution in
the UAPI, so I don't think I can get that past Jason. It would also be
hard to shove in between the elements in the existing naming.

I am just trying to get the nits that came up in the previous discussions
handled, so they are fixed when I submit the WG spec again, so that I
don't have to go back and update it shortly after, as the WG/Jason
bandwidth seams pretty limited atm.