On Wed, Sep 04, 2024 at 04:54:17PM -0700, Jakub Kicinski wrote:

On Wed, 4 Sep 2024 16:40:41 -0700 Stanislav Fomichev wrote:

quoted

quoted

I think what you are proposing seems fine; I'm just working out the
implementation details and making sure I understand before sending
another revision.  

What if instead of an extra storage index in UAPI, we make napi_id persistent?
Then we can keep using napi_id as a user-facing number for the configuration.

Having a stable napi_id would also be super useful for the epoll setup so you
don't have to match old/invalid ids to the new ones on device reset.

that'd be nice, initially I thought that we have some drivers that have
multiple instances of NAPI enabled for a single "index", but I don't
see such drivers now.

quoted

In the code, we can keep the same idea with napi_storage in netdev and
ask drivers to provide storage id, but keep that id internal.

The only complication with that is napi_hash_add/napi_hash_del that
happen in netif_napi_add_weight. So for the devices that allocate
new napi before removing the old ones (most devices?), we'd have to add
some new netif_napi_takeover(old_napi, new_napi) to remove the
old napi_id from the hash and reuse it in the new one.

So for mlx5, the flow would look like the following:

- mlx5e_safe_switch_params
  - mlx5e_open_channels
    - netif_napi_add(new_napi)
      - adds napi with 'ephemeral' napi id
  - mlx5e_switch_priv_channels
    - mlx5e_deactivate_priv_channels
      - napi_disable(old_napi)
      - netif_napi_del(old_napi) - this frees the old napi_id
  - mlx5e_activate_priv_channels
    - mlx5e_activate_channels
      - mlx5e_activate_channel
        - netif_napi_takeover(old_napi is gone, so probably take id from napi_storage?)
	  - if napi is not hashed - safe to reuse?
	- napi_enable

This is a bit ugly because we still have random napi ids during reset, but
is not super complicated implementation-wise. We can eventually improve
the above by splitting netif_napi_add_weight into two steps: allocate and
activate (to do the napi_id allocation & hashing). Thoughts?

The "takeover" would be problematic for drivers which free old NAPI
before allocating new one (bnxt?). But splitting the two steps sounds
pretty clean. We can add a helper to mark NAPI as "driver will
explicitly list/hash later", and have the driver call a new helper
which takes storage ID and lists the NAPI in the hash.

It sounds like you all are suggesting that napi_id is moved into the
napi_storage array, as well? That way NAPI IDs persist even if the
NAPI structs themselves are recreated?

I think that's interesting and I'm open to supporting that. I wrote
up an RFC that moves stuff in the direction of napi_storage and
modifies 3 drivers but:
  - is incorrect because it breaks the persistence thing we are
    talking about, and
  - it doesn't do the two step take-over thing described above to
    inherit NAPI IDs (as far as I understand ?)

I'm going to send the RFC anyway because I think it'll be easier to
pick up the discussion on code that is hopefully closer to where we
want to land.

I hope that is OK.

- Joe