On Tue, Dec 14, 2021 at 11:26:47AM +0100, Magnus Karlsson wrote:
> From: Magnus Karlsson <magnus.karlsson@xxxxxxxxx>
>
> Add a test for the tx_writeable condition to the batched Tx processing
> path. This test is in the skb and non-batched code paths but not in the
> batched code path. So add it there. This test makes sure that a
> process is not woken up until there are a sufficiently large number of
> free entries in the Tx ring. Currently, any driver using the batched
> interface will be woken up even if there is only one free entry,
> impacting performance negatively.
I gave this patch a shot on ice driver with the Tx batching patch that i'm
about to send which is using the xsk_tx_peek_release_desc_batch(). I ran
the 2 core setup with no busy poll and it turned out that this change has
a negative impact on performance - it degrades by 5%.
After a short chat with Magnus he said it's due to the touch to the global
state of a ring that xsk_tx_writeable() is doing.
So maintainers, please do not apply this yet, we'll come up with a
solution.
Also, should this be sent to bpf tree (not bpf-next) ?
Thanks!
>
> Fixes: 3413f04141aa ("xsk: Change the tx writeable condition")
> Signed-off-by: Magnus Karlsson <magnus.karlsson@xxxxxxxxx>
> ---
> net/xdp/xsk.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
> index 28ef3f4465ae..3772fcaa76ed 100644
> --- a/net/xdp/xsk.c
> +++ b/net/xdp/xsk.c
> @@ -392,7 +392,8 @@ u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *
>
> xskq_cons_release_n(xs->tx, nb_pkts);
> __xskq_cons_release(xs->tx);
> - xs->sk.sk_write_space(&xs->sk);
> + if (xsk_tx_writeable(xs))
> + xs->sk.sk_write_space(&xs->sk);
>
> out:
> rcu_read_unlock();
>
> base-commit: d27a662290963a1cde26cdfdbac71a546c06e94a
> --
> 2.29.0
>
