From 35f41229b58cb8c2611207827aa4f658b82db67e Mon Sep 17 00:00:00 2001 From: Daniel Weber Date: Mon, 5 Aug 2019 14:02:30 +0200 Subject: [PATCH] mptcp: Earliest Completion First (ECF) Scheduler This scheduler works much like the default MPTCP scheduler. It always prefers the subflow with the smallest round-trip-time that is available. Signed-off-by: Daniel Weber --- net/mptcp/Kconfig | 6 + net/mptcp/Makefile | 1 + net/mptcp/mptcp_ecf.c | 384 ++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 391 insertions(+) create mode 100644 net/mptcp/mptcp_ecf.c diff --git a/net/mptcp/Kconfig b/net/mptcp/Kconfig index d22b7b47860f..dd1f859f1070 100644 --- a/net/mptcp/Kconfig +++ b/net/mptcp/Kconfig @@ -109,6 +109,12 @@ config MPTCP_REDUNDANT This scheduler sends all packets redundantly over all subflows to decreases latency and jitter on the cost of lower throughput. +config MPTCP_ECF + tristate "MPTCP ECF" + depends on (MPTCP=y) + ---help--- + This is an experimental Earliest Completion First (ECF) scheduler. + choice prompt "Default MPTCP Scheduler" default DEFAULT diff --git a/net/mptcp/Makefile b/net/mptcp/Makefile index 82a2d4d945ae..369248a2f68e 100644 --- a/net/mptcp/Makefile +++ b/net/mptcp/Makefile @@ -20,5 +20,6 @@ obj-$(CONFIG_MPTCP_NETLINK) += mptcp_netlink.o obj-$(CONFIG_MPTCP_ROUNDROBIN) += mptcp_rr.o obj-$(CONFIG_MPTCP_REDUNDANT) += mptcp_redundant.o obj-$(CONFIG_MPTCP_BLEST) += mptcp_blest.o +obj-$(CONFIG_MPTCP_ECF) += mptcp_ecf.o mptcp-$(subst m,y,$(CONFIG_IPV6)) += mptcp_ipv6.o diff --git a/net/mptcp/mptcp_ecf.c b/net/mptcp/mptcp_ecf.c new file mode 100644 index 000000000000..d61f4d2ad375 --- /dev/null +++ b/net/mptcp/mptcp_ecf.c @@ -0,0 +1,384 @@ +// SPDX-License-Identifier: GPL-2.0 +/* MPTCP ECF Scheduler + * + * Algorithm Design: + * Yeon-sup Lim + * Don Towsley + * Erich M. Nahum + * Richard J. Gibbens + * + * Initial Implementation: + * Yeon-sup Lim + * + * Additional Authors: + * Daniel Weber + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation; either version + * 2 of the License, or (at your option) any later version. + */ + +#include +#include +#include + +static unsigned int r_beta __read_mostly = 4; /* beta = 1/r_beta = 0.25 */ +module_param(r_beta, int, 0644); +MODULE_PARM_DESC(r_beta, "beta for ECF"); + +struct ecfsched_priv { + u32 last_rbuf_opti; +}; + +struct ecfsched_cb { + u32 switching_margin; /* this is "waiting" in algorithm description */ +}; + +static struct ecfsched_priv *ecfsched_get_priv(const struct tcp_sock *tp) +{ + return (struct ecfsched_priv *)&tp->mptcp->mptcp_sched[0]; +} + +static struct ecfsched_cb *ecfsched_get_cb(const struct tcp_sock *tp) +{ + return (struct ecfsched_cb *)&tp->mpcb->mptcp_sched[0]; +} + +/* This is the ECF scheduler. This function decides on which flow to send + * a given MSS. If all subflows are found to be busy or the currently best + * subflow is estimated to be slower than waiting for minsk, NULL is returned. + */ +static struct sock *ecf_get_available_subflow(struct sock *meta_sk, + struct sk_buff *skb, + bool zero_wnd_test) +{ + struct mptcp_cb *mpcb = tcp_sk(meta_sk)->mpcb; + struct sock *bestsk, *minsk = NULL; + struct tcp_sock *besttp; + struct mptcp_tcp_sock *mptcp; + struct ecfsched_cb *ecf_cb = ecfsched_get_cb(tcp_sk(meta_sk)); + u32 min_srtt = U32_MAX; + u32 sub_sndbuf = 0; + u32 sub_packets_out = 0; + + /* Answer data_fin on same subflow!!! */ + if (meta_sk->sk_shutdown & RCV_SHUTDOWN && + skb && mptcp_is_data_fin(skb)) { + mptcp_for_each_sub(mpcb, mptcp) { + bestsk = mptcp_to_sock(mptcp); + + if (tcp_sk(bestsk)->mptcp->path_index == mpcb->dfin_path_index && + mptcp_is_available(bestsk, skb, zero_wnd_test)) + return bestsk; + } + } + + /* First, find the overall best (fastest) subflow */ + mptcp_for_each_sub(mpcb, mptcp) { + bestsk = mptcp_to_sock(mptcp); + besttp = tcp_sk(bestsk); + + /* Set of states for which we are allowed to send data */ + if (!mptcp_sk_can_send(bestsk)) + continue; + + /* We do not send data on this subflow unless it is + * fully established, i.e. the 4th ack has been received. + */ + if (besttp->mptcp->pre_established) + continue; + + sub_sndbuf += bestsk->sk_wmem_queued; + sub_packets_out += besttp->packets_out; + + /* record minimal rtt */ + if (besttp->srtt_us < min_srtt) { + min_srtt = besttp->srtt_us; + minsk = bestsk; + } + } + + /* find the current best subflow according to the default scheduler */ + bestsk = get_available_subflow(meta_sk, skb, zero_wnd_test); + + /* if we decided to use a slower flow, we have the option of not using it at all */ + if (bestsk && minsk && bestsk != minsk) { + u32 mss = tcp_current_mss(bestsk); /* assuming equal MSS */ + u32 sndbuf_meta = meta_sk->sk_wmem_queued; + u32 sndbuf_minus = sub_sndbuf; + u32 sndbuf = 0; + + u32 cwnd_f = tcp_sk(minsk)->snd_cwnd; + u32 srtt_f = tcp_sk(minsk)->srtt_us >> 3; + u32 rttvar_f = tcp_sk(minsk)->rttvar_us >> 1; + + u32 cwnd_s = tcp_sk(bestsk)->snd_cwnd; + u32 srtt_s = tcp_sk(bestsk)->srtt_us >> 3; + u32 rttvar_s = tcp_sk(bestsk)->rttvar_us >> 1; + + u32 delta = max(rttvar_f, rttvar_s); + + u32 x_f; + u64 lhs, rhs; /* to avoid overflow, using u64 */ + + if (tcp_sk(meta_sk)->packets_out > sub_packets_out) + sndbuf_minus += (tcp_sk(meta_sk)->packets_out - sub_packets_out) * mss; + + if (sndbuf_meta > sndbuf_minus) + sndbuf = sndbuf_meta - sndbuf_minus; + + /* we have something to send. + * at least one time tx over fastest subflow is required + */ + x_f = sndbuf > cwnd_f * mss ? sndbuf : cwnd_f * mss; + lhs = srtt_f * (x_f + cwnd_f * mss); + rhs = cwnd_f * mss * (srtt_s + delta); + + if (r_beta * lhs < r_beta * rhs + ecf_cb->switching_margin * rhs) { + u32 x_s = sndbuf > cwnd_s * mss ? sndbuf : cwnd_s * mss; + u64 lhs_s = srtt_s * x_s; + u64 rhs_s = cwnd_s * mss * (2 * srtt_f + delta); + + if (lhs_s >= rhs_s) { + /* too slower than fastest */ + ecf_cb->switching_margin = 1; + return NULL; + } + } else { + /* use slower one */ + ecf_cb->switching_margin = 0; + } + } + + return bestsk; +} + +/* copy from mptcp_sched.c: mptcp_rcv_buf_optimization */ +static struct sk_buff *mptcp_ecf_rcv_buf_optimization(struct sock *sk, int penal) +{ + struct sock *meta_sk; + const struct tcp_sock *tp = tcp_sk(sk); + struct mptcp_tcp_sock *mptcp; + struct sk_buff *skb_head; + struct ecfsched_priv *ecf_p = ecfsched_get_priv(tp); + + meta_sk = mptcp_meta_sk(sk); + skb_head = tcp_rtx_queue_head(meta_sk); + + if (!skb_head) + return NULL; + + /* If penalization is optional (coming from mptcp_next_segment() and + * We are not send-buffer-limited we do not penalize. The retransmission + * is just an optimization to fix the idle-time due to the delay before + * we wake up the application. + */ + if (!penal && sk_stream_memory_free(meta_sk)) + goto retrans; + + /* Only penalize again after an RTT has elapsed */ + if (tcp_jiffies32 - ecf_p->last_rbuf_opti < usecs_to_jiffies(tp->srtt_us >> 3)) + goto retrans; + + /* Half the cwnd of the slow flows */ + mptcp_for_each_sub(tp->mpcb, mptcp) { + struct tcp_sock *tp_it = mptcp->tp; + + if (tp_it != tp && + TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp_it->mptcp->path_index)) { + if (tp->srtt_us < tp_it->srtt_us && inet_csk((struct sock *)tp_it)->icsk_ca_state == TCP_CA_Open) { + u32 prior_cwnd = tp_it->snd_cwnd; + + tp_it->snd_cwnd = max(tp_it->snd_cwnd >> 1U, 1U); + + /* If in slow start, do not reduce the ssthresh */ + if (prior_cwnd >= tp_it->snd_ssthresh) + tp_it->snd_ssthresh = max(tp_it->snd_ssthresh >> 1U, 2U); + + ecf_p->last_rbuf_opti = tcp_jiffies32; + } + } + } + +retrans: + + /* Segment not yet injected into this path? Take it!!! */ + if (!(TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp->mptcp->path_index))) { + bool do_retrans = false; + mptcp_for_each_sub(tp->mpcb, mptcp) { + struct tcp_sock *tp_it = mptcp->tp; + + if (tp_it != tp && + TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp_it->mptcp->path_index)) { + if (tp_it->snd_cwnd <= 4) { + do_retrans = true; + break; + } + + if (4 * tp->srtt_us >= tp_it->srtt_us) { + do_retrans = false; + break; + } else { + do_retrans = true; + } + } + } + + if (do_retrans && mptcp_is_available(sk, skb_head, false)) { + trace_mptcp_retransmit(sk, skb_head); + return skb_head; + } + } + return NULL; +} + +/* copy from mptcp_sched.c: __mptcp_next_segment */ +/* Returns the next segment to be sent from the mptcp meta-queue. + * (chooses the reinject queue if any segment is waiting in it, otherwise, + * chooses the normal write queue). + * Sets *@reinject to 1 if the returned segment comes from the + * reinject queue. Sets it to 0 if it is the regular send-head of the meta-sk, + * and sets it to -1 if it is a meta-level retransmission to optimize the + * receive-buffer. + */ +static struct sk_buff *__mptcp_ecf_next_segment(struct sock *meta_sk, int *reinject) +{ + const struct mptcp_cb *mpcb = tcp_sk(meta_sk)->mpcb; + struct sk_buff *skb = NULL; + + *reinject = 0; + + /* If we are in fallback-mode, just take from the meta-send-queue */ + if (mpcb->infinite_mapping_snd || mpcb->send_infinite_mapping) + return tcp_send_head(meta_sk); + + skb = skb_peek(&mpcb->reinject_queue); + + if (skb) { + *reinject = 1; + } else { + skb = tcp_send_head(meta_sk); + + if (!skb && meta_sk->sk_socket && + test_bit(SOCK_NOSPACE, &meta_sk->sk_socket->flags) && + sk_stream_wspace(meta_sk) < sk_stream_min_wspace(meta_sk)) { + struct sock *subsk = ecf_get_available_subflow(meta_sk, NULL, + false); + if (!subsk) + return NULL; + + skb = mptcp_ecf_rcv_buf_optimization(subsk, 0); + if (skb) + *reinject = -1; + } + } + return skb; +} + +/* copy from mptcp_sched.c: mptcp_next_segment */ +static struct sk_buff *mptcp_ecf_next_segment(struct sock *meta_sk, + int *reinject, + struct sock **subsk, + unsigned int *limit) +{ + struct sk_buff *skb = __mptcp_ecf_next_segment(meta_sk, reinject); + unsigned int mss_now; + struct tcp_sock *subtp; + u16 gso_max_segs; + u32 max_len, max_segs, window, needed; + + /* As we set it, we have to reset it as well. */ + *limit = 0; + + if (!skb) + return NULL; + + *subsk = ecf_get_available_subflow(meta_sk, skb, false); + if (!*subsk) + return NULL; + + subtp = tcp_sk(*subsk); + mss_now = tcp_current_mss(*subsk); + + if (!*reinject && unlikely(!tcp_snd_wnd_test(tcp_sk(meta_sk), skb, mss_now))) { + skb = mptcp_ecf_rcv_buf_optimization(*subsk, 1); + if (skb) + *reinject = -1; + else + return NULL; + } + + /* No splitting required, as we will only send one single segment */ + if (skb->len <= mss_now) + return skb; + + /* The following is similar to tcp_mss_split_point, but + * we do not care about nagle, because we will anyways + * use TCP_NAGLE_PUSH, which overrides this. + * + * So, we first limit according to the cwnd/gso-size and then according + * to the subflow's window. + */ + + gso_max_segs = (*subsk)->sk_gso_max_segs; + if (!gso_max_segs) /* No gso supported on the subflow's NIC */ + gso_max_segs = 1; + max_segs = min_t(unsigned int, tcp_cwnd_test(subtp, skb), gso_max_segs); + if (!max_segs) + return NULL; + + max_len = mss_now * max_segs; + window = tcp_wnd_end(subtp) - subtp->write_seq; + + needed = min(skb->len, window); + if (max_len <= skb->len) + /* Take max_win, which is actually the cwnd/gso-size */ + *limit = max_len; + else + /* Or, take the window */ + *limit = needed; + + return skb; +} + +static void ecfsched_init(struct sock *sk) +{ + struct ecfsched_priv *ecf_p = ecfsched_get_priv(tcp_sk(sk)); + struct ecfsched_cb *ecf_cb = ecfsched_get_cb(tcp_sk(mptcp_meta_sk(sk))); + + ecf_p->last_rbuf_opti = tcp_jiffies32; + ecf_cb->switching_margin = 0; +} + +struct mptcp_sched_ops mptcp_sched_ecf = { + .get_subflow = ecf_get_available_subflow, + .next_segment = mptcp_ecf_next_segment, + .init = ecfsched_init, + .name = "ecf", + .owner = THIS_MODULE, +}; + +static int __init ecf_register(void) +{ + BUILD_BUG_ON(sizeof(struct ecfsched_priv) > MPTCP_SCHED_SIZE); + BUILD_BUG_ON(sizeof(struct ecfsched_cb) > MPTCP_SCHED_DATA_SIZE); + + if (mptcp_register_scheduler(&mptcp_sched_ecf)) + return -1; + + return 0; +} + +static void ecf_unregister(void) +{ + mptcp_unregister_scheduler(&mptcp_sched_ecf); +} + +module_init(ecf_register); +module_exit(ecf_unregister); + +MODULE_AUTHOR("Yeon-sup Lim, Daniel Weber"); +MODULE_LICENSE("GPL"); +MODULE_DESCRIPTION("ECF (Earliest Completion First) scheduler for MPTCP, based on default minimum RTT scheduler"); +MODULE_VERSION("0.95");