On Fri, 2020-09-11 at 15:52 +0200, Paolo Abeni wrote:
[...]
> +#define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
> + sizeof(struct tcphdr) - \
> + MAX_TCP_OPTION_SPACE - \
> + sizeof(struct ipv6hdr) - \
> + sizeof(struct frag_hdr))
> +
> +struct subflow_send_info {
> + struct sock *ssk;
> + uint64_t ratio;
> +};
> +
> static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
> u32 *sndbuf)
> {
> + struct subflow_send_info send_info[2];
> struct mptcp_subflow_context *subflow;
> - struct sock *sk = (struct sock *)msk;
> - struct sock *backup = NULL;
> - bool free;
> + int i, nr_active = 0;
> + int64_t ratio, pace;
> + struct sock *ssk;
>
> - sock_owned_by_me(sk);
> + sock_owned_by_me((struct sock *)msk);
>
> *sndbuf = 0;
> if (!mptcp_ext_cache_refill(msk))
> return NULL;
>
> - mptcp_for_each_subflow(msk, subflow) {
> - struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
> -
> - free = sk_stream_is_writeable(subflow->tcp_sock);
> - if (!free) {
> - mptcp_nospace(msk);
> + if (__mptcp_check_fallback(msk)) {
> + if (!msk->first)
> return NULL;
> + *sndbuf = msk->first->sk_sndbuf;
> + return sk_stream_memory_free(msk->first) ? msk->first : NULL;
> + }
> +
> + /* re-use last subflow, if the burst allow that */
> + if (msk->last_snd && msk->snd_burst > 0 &&
> + sk_stream_memory_free(msk->last_snd) &&
> + mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
> + mptcp_for_each_subflow(msk, subflow) {
> + ssk = mptcp_subflow_tcp_sock(subflow);
> + *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
> }
> + return msk->last_snd;
> + }
> +
> + /* pick the subflow with the lower wmem/wspace ratio */
> + for (i = 0; i < 2; ++i) {
> + send_info[i].ssk = NULL;
> + send_info[i].ratio = -1;
> + }
> + mptcp_for_each_subflow(msk, subflow) {
> + ssk = mptcp_subflow_tcp_sock(subflow);
> + if (!mptcp_subflow_active(subflow))
> + continue;
>
> + nr_active += !subflow->backup;
> *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
> - if (subflow->backup) {
> - if (!backup)
> - backup = ssk;
> + if (!sk_stream_memory_free(subflow->tcp_sock))
> + continue;
>
> + pace = READ_ONCE(ssk->sk_pacing_rate);
> + if (!pace)
> continue;
> - }
>
> - return ssk;
> + ratio = (int64_t)READ_ONCE(ssk->sk_wmem_queued) << 32 / pace;
Kbuild bot on our devel branch just noted that the above division
breaks 32 bits build.
I'll fix that in v2.
Cheers,
Paolo
