Re: [PATCH net-next v2] rocker: move netevent neigh update to processes context

[Simon Horman <horms@verge.net.au>]

Hi Scott,

> Simon, just focusing on this rocker->neigh_tbl_next_index++ issue at
> the moment, I think the change below would make _rocker_neigh_add()
> safe to call without needing to put neigh_update into process context.
> It works because entry is stored between PREPARE and COMMIT, so once
> entry->index is assigned in PREPARE (under neigh_tbl_lock), it'll be
> re-used in COMMIT, even if the NONE thread also claims it's
> entry->index (also under neigh_tbl_lock).
> 
> I know there are other issues you and Toshiaki Makita have pointed
> out, but let's see if we can peel the onion one layer at a time.
> 
> --- a/drivers/net/ethernet/rocker/rocker.c
> +++ b/drivers/net/ethernet/rocker/rocker.c
> @@ -2904,10 +2904,10 @@ static void _rocker_neigh_add(struct rocker *rocker,
>                               enum switchdev_trans trans,
>                               struct rocker_neigh_tbl_entry *entry)
>  {
> -       entry->index = rocker->neigh_tbl_next_index;
> +       if (trans != SWITCHDEV_TRANS_COMMIT)
> +               entry->index = rocker->neigh_tbl_next_index++;
>         if (trans == SWITCHDEV_TRANS_PREPARE)
>                 return;
> -       rocker->neigh_tbl_next_index++;
>         entry->ref_count++;
>         hash_add(rocker->neigh_tbl, &entry->entry,
>                  be32_to_cpu(entry->ip_addr));

I agree that should work.