Re: [RFC PATCH bpf-next 0/3] Avoid skipping sockets with socket iterators

[Jordan Rife <jrife@google.com>]

> It is a very corner case.
>
> I guess it can with GFP_ATOMIC. I just didn't think it was needed considering
> the key of the hash is addresses+ports. If we have many socks collided on the
> same addresses+ports bucket, that would be a better hashtable problem to solve
> first.

I see, thanks for explaining.

> We can also consider if the same sk batch array can be reused to store cookies
> during stop(). If the array can reuse, it would be nice but not a blocker imo.

> In term of slowness, the worst will be all the previous stored cookies cannot be
> found in the updated bucket? Not sure how often a socket is gone and how often
> there is a very large bucket (as mentioned at the hashtable's key earlier), so
> should not be an issue for iteration use case? I guess it may need some rough
> PoC code to judge if it is feasible.

I like the idea of reusing the sk batch array. Maybe create a union
batch_item containing struct sock * and __u64. I'll explore this
direction a bit and see if I can come up with a small PoC. Lots of
array scanning could be slow, but since changes to a bucket should be
rare, one optimization could be to only compare to the saved socket
cookies if the bucket has changed since it was last seen. I think
saving and checking the head, tail, and size of the bucket's linked
list should be sufficient for this?

-Jordan