Re: [PATCH net-next 1/2] net: store netdevs in an xarray

[Jakub Kicinski <kuba@kernel.org>]

On Mon, 24 Jul 2023 10:18:04 +0200 Paolo Abeni wrote:
> A possibly dumb question: why using an xarray over a plain list?

We need to drop the lock during the walk. So for a list we'd need 
to either 
 - add explicit iteration "cursor" or 
 - suffer O(n) for insertion + O(n^2) for restart?

Or there's a easier way to do it?
The cursor is not the worst option, I guess, a bit less intuitive
and harder to clean up on error, but doable?

> It looks like the idea is to additionally use xarray for device lookup
> beyond for dumping?

I was measuring it to find out if we can delete the hash table without
anyone noticing, but it's not really the motivation.

> WRT the above, have you considered instead replacing dev_name_head with
> an rhashtable? (and add the mentioned list)

The main motivation is ease of iteration for netlink dumps.

I'll admit that my understanding of rhashtable is superficial but
I don't think it's possible to do consistent dumps with it. Even 
if it supported "cursors" (which I'm not sure it does) a rhash could
rearrange buckets and mess the order up. The comment on
rhltable_walk_enter() says:

 * For a completely stable walk you should construct your own data
 * structure outside the hash table.

Given the iteration process is controlled by user space trying to
constrain a rhash also used by the fast path to get consistent dumps
seems risky.

Adding Herbert and Neil to keep me honest.