From owner-svn-src-all@freebsd.org Tue Oct 25 05:45:49 2016 Return-Path: Delivered-To: svn-src-all@mailman.ysv.freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:1900:2254:206a::19:1]) by mailman.ysv.freebsd.org (Postfix) with ESMTP id E8D01C20CE2; Tue, 25 Oct 2016 05:45:49 +0000 (UTC) (envelope-from hiren@FreeBSD.org) Received: from repo.freebsd.org (repo.freebsd.org [IPv6:2610:1c1:1:6068::e6a:0]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (Client did not present a certificate) by mx1.freebsd.org (Postfix) with ESMTPS id BA2A2988; Tue, 25 Oct 2016 05:45:49 +0000 (UTC) (envelope-from hiren@FreeBSD.org) Received: from repo.freebsd.org ([127.0.1.37]) by repo.freebsd.org (8.15.2/8.15.2) with ESMTP id u9P5jmGZ011790; Tue, 25 Oct 2016 05:45:48 GMT (envelope-from hiren@FreeBSD.org) Received: (from hiren@localhost) by repo.freebsd.org (8.15.2/8.15.2/Submit) id u9P5jm3u011781; Tue, 25 Oct 2016 05:45:48 GMT (envelope-from hiren@FreeBSD.org) Message-Id: <201610250545.u9P5jm3u011781@repo.freebsd.org> X-Authentication-Warning: repo.freebsd.org: hiren set sender to hiren@FreeBSD.org using -f From: Hiren Panchasara Date: Tue, 25 Oct 2016 05:45:48 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r307901 - in head/sys/netinet: . cc tcp_stacks X-SVN-Group: head MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-BeenThere: svn-src-all@freebsd.org X-Mailman-Version: 2.1.23 Precedence: list List-Id: "SVN commit messages for the entire src tree \(except for " user" and " projects" \)" List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 25 Oct 2016 05:45:50 -0000 Author: hiren Date: Tue Oct 25 05:45:47 2016 New Revision: 307901 URL: https://svnweb.freebsd.org/changeset/base/307901 Log: FreeBSD tcp stack used to inform respective congestion control module about the loss event but not use or obay the recommendations i.e. values set by it in some cases. Here is an attempt to solve that confusion by following relevant RFCs/drafts. Stack only sets congestion window/slow start threshold values when there is no CC module availalbe to take that action. All CC modules are inspected and updated when needed to take appropriate action on loss. tcp_stacks/fastpath module has been updated to adapt these changes. Note: Probably, the most significant change would be to not bring congestion window down to 1MSS on a loss signaled by 3-duplicate acks and letting respective CC decide that value. In collaboration with: Matt Macy Discussed on: transport@ mailing list Reviewed by: jtl MFC after: 1 month Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D8225 Modified: head/sys/netinet/cc/cc_cdg.c head/sys/netinet/cc/cc_chd.c head/sys/netinet/cc/cc_cubic.c head/sys/netinet/cc/cc_dctcp.c head/sys/netinet/cc/cc_htcp.c head/sys/netinet/cc/cc_newreno.c head/sys/netinet/tcp_input.c head/sys/netinet/tcp_stacks/fastpath.c Modified: head/sys/netinet/cc/cc_cdg.c ============================================================================== --- head/sys/netinet/cc/cc_cdg.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/cc/cc_cdg.c Tue Oct 25 05:45:47 2016 (r307901) @@ -431,6 +431,11 @@ static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type) { struct cdg *cdg_data = ccv->cc_data; + uint32_t cwin; + u_int mss; + + cwin = CCV(ccv, snd_cwnd); + mss = CCV(ccv, t_maxseg); switch(signal_type) { case CC_CDG_DELAY: @@ -448,7 +453,7 @@ cdg_cong_signal(struct cc_var *ccv, uint */ if (IN_CONGRECOVERY(CCV(ccv, t_flags)) || cdg_data->queue_state < CDG_Q_FULL) { - CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); + CCV(ccv, snd_ssthresh) = cwin; CCV(ccv, snd_recover) = CCV(ccv, snd_max); } else { /* @@ -461,13 +466,17 @@ cdg_cong_signal(struct cc_var *ccv, uint cdg_data->shadow_w, RENO_BETA); CCV(ccv, snd_ssthresh) = max(cdg_data->shadow_w, - cdg_window_decrease(ccv, CCV(ccv, snd_cwnd), - V_cdg_beta_loss)); + cdg_window_decrease(ccv, cwin, V_cdg_beta_loss)); + CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); cdg_data->window_incr = cdg_data->rtt_count = 0; } ENTER_RECOVERY(CCV(ccv, t_flags)); break; + case CC_RTO: + CCV(ccv, snd_ssthresh) = max(2*mss, cwin/2); + CCV(ccv, snd_cwnd) = mss; + break; default: newreno_cc_algo.cong_signal(ccv, signal_type); break; Modified: head/sys/netinet/cc/cc_chd.c ============================================================================== --- head/sys/netinet/cc/cc_chd.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/cc/cc_chd.c Tue Oct 25 05:45:47 2016 (r307901) @@ -330,10 +330,14 @@ chd_cong_signal(struct cc_var *ccv, uint struct ertt *e_t; struct chd *chd_data; int qdly; + uint32_t cwin; + u_int mss; e_t = khelp_get_osd(CCV(ccv, osd), ertt_id); chd_data = ccv->cc_data; qdly = imax(e_t->rtt, chd_data->maxrtt_in_rtt) - e_t->minrtt; + cwin = CCV(ccv, snd_cwnd); + mss = CCV(ccv, t_maxseg); switch(signal_type) { case CC_CHD_DELAY: @@ -373,6 +377,10 @@ chd_cong_signal(struct cc_var *ccv, uint } ENTER_FASTRECOVERY(CCV(ccv, t_flags)); break; + case CC_RTO: + CCV(ccv, snd_ssthresh) = max(2*mss, cwin/2); + CCV(ccv, snd_cwnd) = mss; + break; default: newreno_cc_algo.cong_signal(ccv, signal_type); Modified: head/sys/netinet/cc/cc_cubic.c ============================================================================== --- head/sys/netinet/cc/cc_cubic.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/cc/cc_cubic.c Tue Oct 25 05:45:47 2016 (r307901) @@ -225,8 +225,12 @@ static void cubic_cong_signal(struct cc_var *ccv, uint32_t type) { struct cubic *cubic_data; + uint32_t cwin; + u_int mss; cubic_data = ccv->cc_data; + cwin = CCV(ccv, snd_cwnd); + mss = CCV(ccv, t_maxseg); switch (type) { case CC_NDUPACK: @@ -235,7 +239,8 @@ cubic_cong_signal(struct cc_var *ccv, ui cubic_ssthresh_update(ccv); cubic_data->num_cong_events++; cubic_data->prev_max_cwnd = cubic_data->max_cwnd; - cubic_data->max_cwnd = CCV(ccv, snd_cwnd); + cubic_data->max_cwnd = cwin; + CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); } ENTER_RECOVERY(CCV(ccv, t_flags)); } @@ -246,7 +251,7 @@ cubic_cong_signal(struct cc_var *ccv, ui cubic_ssthresh_update(ccv); cubic_data->num_cong_events++; cubic_data->prev_max_cwnd = cubic_data->max_cwnd; - cubic_data->max_cwnd = CCV(ccv, snd_cwnd); + cubic_data->max_cwnd = cwin; cubic_data->t_last_cong = ticks; CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); ENTER_CONGRECOVERY(CCV(ccv, t_flags)); @@ -261,9 +266,13 @@ cubic_cong_signal(struct cc_var *ccv, ui * chance the first one is a false alarm and may not indicate * congestion. */ - if (CCV(ccv, t_rxtshift) >= 2) + if (CCV(ccv, t_rxtshift) >= 2) { cubic_data->num_cong_events++; cubic_data->t_last_cong = ticks; + cubic_ssthresh_update(ccv); + cubic_data->max_cwnd = cwin; + CCV(ccv, snd_cwnd) = mss; + } break; } } Modified: head/sys/netinet/cc/cc_dctcp.c ============================================================================== --- head/sys/netinet/cc/cc_dctcp.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/cc/cc_dctcp.c Tue Oct 25 05:45:47 2016 (r307901) @@ -230,10 +230,11 @@ static void dctcp_cong_signal(struct cc_var *ccv, uint32_t type) { struct dctcp *dctcp_data; - u_int win, mss; + uint32_t cwin; + u_int mss; dctcp_data = ccv->cc_data; - win = CCV(ccv, snd_cwnd); + cwin = CCV(ccv, snd_cwnd); mss = CCV(ccv, t_maxseg); switch (type) { @@ -241,16 +242,16 @@ dctcp_cong_signal(struct cc_var *ccv, ui if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { CCV(ccv, snd_ssthresh) = mss * - max(win / 2 / mss, 2); + max(cwin / 2 / mss, 2); dctcp_data->num_cong_events++; } else { /* cwnd has already updated as congestion * recovery. Reverse cwnd value using * snd_cwnd_prev and recalculate snd_ssthresh */ - win = CCV(ccv, snd_cwnd_prev); + cwin = CCV(ccv, snd_cwnd_prev); CCV(ccv, snd_ssthresh) = - max(win / 2 / mss, 2) * mss; + max(cwin / 2 / mss, 2) * mss; } ENTER_RECOVERY(CCV(ccv, t_flags)); } @@ -260,18 +261,18 @@ dctcp_cong_signal(struct cc_var *ccv, ui * Save current snd_cwnd when the host encounters both * congestion recovery and fast recovery. */ - CCV(ccv, snd_cwnd_prev) = win; + CCV(ccv, snd_cwnd_prev) = cwin; if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { if (V_dctcp_slowstart && dctcp_data->num_cong_events++ == 0) { CCV(ccv, snd_ssthresh) = - mss * max(win / 2 / mss, 2); + mss * max(cwin / 2 / mss, 2); dctcp_data->alpha = MAX_ALPHA_VALUE; dctcp_data->bytes_ecn = 0; dctcp_data->bytes_total = 0; dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); } else - CCV(ccv, snd_ssthresh) = max((win - ((win * + CCV(ccv, snd_ssthresh) = max((cwin - ((cwin * dctcp_data->alpha) >> 11)) / mss, 2) * mss; CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); ENTER_CONGRECOVERY(CCV(ccv, t_flags)); @@ -284,6 +285,8 @@ dctcp_cong_signal(struct cc_var *ccv, ui dctcp_update_alpha(ccv); dctcp_data->save_sndnxt += CCV(ccv, t_maxseg); dctcp_data->num_cong_events++; + CCV(ccv, snd_ssthresh) = max(2 * mss, cwin / 2); + CCV(ccv, snd_cwnd) = mss; } break; } Modified: head/sys/netinet/cc/cc_htcp.c ============================================================================== --- head/sys/netinet/cc/cc_htcp.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/cc/cc_htcp.c Tue Oct 25 05:45:47 2016 (r307901) @@ -271,8 +271,12 @@ static void htcp_cong_signal(struct cc_var *ccv, uint32_t type) { struct htcp *htcp_data; + uint32_t cwin; + u_int mss; htcp_data = ccv->cc_data; + cwin = CCV(ccv, snd_cwnd); + mss = CCV(ccv, t_maxseg); switch (type) { case CC_NDUPACK: @@ -287,8 +291,9 @@ htcp_cong_signal(struct cc_var *ccv, uin (htcp_data->maxrtt - htcp_data->minrtt) * 95) / 100; htcp_ssthresh_update(ccv); + CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); htcp_data->t_last_cong = ticks; - htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); + htcp_data->prev_cwnd = cwin; } ENTER_RECOVERY(CCV(ccv, t_flags)); } @@ -305,7 +310,7 @@ htcp_cong_signal(struct cc_var *ccv, uin htcp_ssthresh_update(ccv); CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); htcp_data->t_last_cong = ticks; - htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); + htcp_data->prev_cwnd = cwin; ENTER_CONGRECOVERY(CCV(ccv, t_flags)); } break; @@ -320,6 +325,8 @@ htcp_cong_signal(struct cc_var *ccv, uin */ if (CCV(ccv, t_rxtshift) >= 2) htcp_data->t_last_cong = ticks; + CCV(ccv, snd_ssthresh) = max(2 * mss, cwin / 2); + CCV(ccv, snd_cwnd) = mss; break; } } Modified: head/sys/netinet/cc/cc_newreno.c ============================================================================== --- head/sys/netinet/cc/cc_newreno.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/cc/cc_newreno.c Tue Oct 25 05:45:47 2016 (r307901) @@ -182,30 +182,39 @@ newreno_after_idle(struct cc_var *ccv) static void newreno_cong_signal(struct cc_var *ccv, uint32_t type) { - u_int win; + uint32_t cwin; + u_int mss; + + cwin = CCV(ccv, snd_cwnd); + mss = CCV(ccv, t_maxseg); /* Catch algos which mistakenly leak private signal types. */ KASSERT((type & CC_SIGPRIVMASK) == 0, ("%s: congestion signal type 0x%08x is private\n", __func__, type)); - win = max(CCV(ccv, snd_cwnd) / 2 / CCV(ccv, t_maxseg), 2) * - CCV(ccv, t_maxseg); + cwin = max(2*mss, cwin/2); switch (type) { case CC_NDUPACK: if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { - if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) - CCV(ccv, snd_ssthresh) = win; + if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { + CCV(ccv, snd_ssthresh) = cwin; + CCV(ccv, snd_cwnd) = cwin; + } ENTER_RECOVERY(CCV(ccv, t_flags)); } break; case CC_ECN: if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { - CCV(ccv, snd_ssthresh) = win; - CCV(ccv, snd_cwnd) = win; + CCV(ccv, snd_ssthresh) = cwin; + CCV(ccv, snd_cwnd) = cwin; ENTER_CONGRECOVERY(CCV(ccv, t_flags)); } break; + case CC_RTO: + CCV(ccv, snd_ssthresh) = cwin; + CCV(ccv, snd_cwnd) = mss; + break; } } Modified: head/sys/netinet/tcp_input.c ============================================================================== --- head/sys/netinet/tcp_input.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/tcp_input.c Tue Oct 25 05:45:47 2016 (r307901) @@ -438,9 +438,15 @@ cc_cong_signal(struct tcpcb *tp, struct tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); - tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / - maxseg) * maxseg; - tp->snd_cwnd = maxseg; + if (CC_ALGO(tp)->cong_signal == NULL) { + /* + * RFC5681 Section 3.1 + * ssthresh = max (FlightSize / 2, 2*SMSS) eq (4) + */ + tp->snd_ssthresh = + max((tp->snd_max - tp->snd_una) / 2, 2 * maxseg); + tp->snd_cwnd = maxseg; + } break; case CC_RTO_ERR: TCPSTAT_INC(tcps_sndrexmitbad); @@ -2613,6 +2619,15 @@ tcp_do_segment(struct mbuf *m, struct tc if (awnd < tp->snd_ssthresh) { tp->snd_cwnd += maxseg; + /* + * RFC5681 Section 3.2 talks about cwnd + * inflation on additional dupacks and + * deflation on recovering from loss. + * + * We keep cwnd into check so that + * we don't have to 'deflate' it when we + * get out of recovery. + */ if (tp->snd_cwnd > tp->snd_ssthresh) tp->snd_cwnd = tp->snd_ssthresh; } @@ -2652,19 +2667,22 @@ tcp_do_segment(struct mbuf *m, struct tc TCPSTAT_INC( tcps_sack_recovery_episode); tp->sack_newdata = tp->snd_nxt; - tp->snd_cwnd = maxseg; + if (CC_ALGO(tp)->cong_signal == NULL) + tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } tp->snd_nxt = th->th_ack; - tp->snd_cwnd = maxseg; + if (CC_ALGO(tp)->cong_signal == NULL) + tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); KASSERT(tp->snd_limited <= 2, ("%s: tp->snd_limited too big", __func__)); - tp->snd_cwnd = tp->snd_ssthresh + - maxseg * - (tp->t_dupacks - tp->snd_limited); + if (CC_ALGO(tp)->cong_signal == NULL) + tp->snd_cwnd = tp->snd_ssthresh + + maxseg * + (tp->t_dupacks - tp->snd_limited); if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; goto drop; Modified: head/sys/netinet/tcp_stacks/fastpath.c ============================================================================== --- head/sys/netinet/tcp_stacks/fastpath.c Tue Oct 25 05:07:51 2016 (r307900) +++ head/sys/netinet/tcp_stacks/fastpath.c Tue Oct 25 05:45:47 2016 (r307901) @@ -1119,6 +1119,15 @@ tcp_do_slowpath(struct mbuf *m, struct t if (awnd < tp->snd_ssthresh) { tp->snd_cwnd += tp->t_maxseg; + /* + * RFC5681 Section 3.2 talks about cwnd + * inflation on additional dupacks and + * deflation on recovering from loss. + * + * We keep cwnd into check so that + * we don't have to 'deflate' it when we + * get out of recovery. + */ if (tp->snd_cwnd > tp->snd_ssthresh) tp->snd_cwnd = tp->snd_ssthresh; } @@ -1158,19 +1167,22 @@ tcp_do_slowpath(struct mbuf *m, struct t TCPSTAT_INC( tcps_sack_recovery_episode); tp->sack_newdata = tp->snd_nxt; - tp->snd_cwnd = tp->t_maxseg; + if (CC_ALGO(tp)->cong_signal == NULL) + tp->snd_cwnd = tp->t_maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } tp->snd_nxt = th->th_ack; - tp->snd_cwnd = tp->t_maxseg; + if (CC_ALGO(tp)->cong_signal == NULL) + tp->snd_cwnd = tp->t_maxseg; (void) tp->t_fb->tfb_tcp_output(tp); KASSERT(tp->snd_limited <= 2, ("%s: tp->snd_limited too big", __func__)); - tp->snd_cwnd = tp->snd_ssthresh + - tp->t_maxseg * - (tp->t_dupacks - tp->snd_limited); + if (CC_ALGO(tp)->cong_signal == NULL) + tp->snd_cwnd = tp->snd_ssthresh + + tp->t_maxseg * + (tp->t_dupacks - tp->snd_limited); if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; goto drop;