tcp: add CDG congestion control

CAIA Delay-Gradient (CDG) is a TCP congestion control that modifies
the TCP sender in order to [1]:

  o Use the delay gradient as a congestion signal.
  o Back off with an average probability that is independent of the RTT.
  o Coexist with flows that use loss-based congestion control, i.e.,
    flows that are unresponsive to the delay signal.
  o Tolerate packet loss unrelated to congestion. (Disabled by default.)

Its FreeBSD implementation was presented for the ICCRG in July 2012;
slides are available at http://www.ietf.org/proceedings/84/iccrg.html

Running the experiment scenarios in [1] suggests that our implementation
achieves more goodput compared with FreeBSD 10.0 senders, although it also
causes more queueing delay for a given backoff factor.

The loss tolerance heuristic is disabled by default due to safety concerns
for its use in the Internet [2, p. 45-46].

We use a variant of the Hybrid Slow start algorithm in tcp_cubic to reduce
the probability of slow start overshoot.

[1] D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
    delay gradients." In Networking 2011, pages 328-341. Springer, 2011.
[2] K.K. Jonassen. "Implementing CAIA Delay-Gradient in Linux."
    MSc thesis. Department of Informatics, University of Oslo, 2015.

Cc: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: David Hayes <davihay@ifi.uio.no>
Cc: Andreas Petlund <apetlund@simula.no>
Cc: Dave Taht <dave.taht@bufferbloat.net>
Cc: Nicolas Kuhn <nicolas.kuhn@telecom-bretagne.eu>
Signed-off-by: Kenneth Klette Jonassen <kennetkl@ifi.uio.no>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/ipv4/Kconfig

@@ -615,6 +615,22 @@ config TCP_CONG_DCTCP
 	For further details see:
 	  http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
 
+config TCP_CONG_CDG
+	tristate "CAIA Delay-Gradient (CDG)"
+	default n
+	---help---
+	CAIA Delay-Gradient (CDG) is a TCP congestion control that modifies
+	the TCP sender in order to:
+
+	  o Use the delay gradient as a congestion signal.
+	  o Back off with an average probability that is independent of the RTT.
+	  o Coexist with flows that use loss-based congestion control.
+	  o Tolerate packet loss unrelated to congestion.
+
+	For further details see:
+	  D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
+	  delay gradients." In Networking 2011. Preprint: http://goo.gl/No3vdg
+
 choice
 	prompt "Default TCP congestion control"
 	default DEFAULT_CUBIC
@@ -646,6 +662,9 @@ choice
 	config DEFAULT_DCTCP
 		bool "DCTCP" if TCP_CONG_DCTCP=y
 
+	config DEFAULT_CDG
+		bool "CDG" if TCP_CONG_CDG=y
+
 	config DEFAULT_RENO
 		bool "Reno"
 endchoice
@@ -668,6 +687,7 @@ config DEFAULT_TCP_CONG
 	default "veno" if DEFAULT_VENO
 	default "reno" if DEFAULT_RENO
 	default "dctcp" if DEFAULT_DCTCP
+	default "cdg" if DEFAULT_CDG
 	default "cubic"
 
 config TCP_MD5SIG

net/ipv4/Makefile

@@ -42,6 +42,7 @@ obj-$(CONFIG_INET_TCP_DIAG) += tcp_diag.o
 obj-$(CONFIG_INET_UDP_DIAG) += udp_diag.o
 obj-$(CONFIG_NET_TCPPROBE) += tcp_probe.o
 obj-$(CONFIG_TCP_CONG_BIC) += tcp_bic.o
+obj-$(CONFIG_TCP_CONG_CDG) += tcp_cdg.o
 obj-$(CONFIG_TCP_CONG_CUBIC) += tcp_cubic.o
 obj-$(CONFIG_TCP_CONG_DCTCP) += tcp_dctcp.o
 obj-$(CONFIG_TCP_CONG_WESTWOOD) += tcp_westwood.o

net/ipv4/tcp_cdg.c

+/*
+ * CAIA Delay-Gradient (CDG) congestion control
+ *
+ * This implementation is based on the paper:
+ *   D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
+ *   delay gradients." In IFIP Networking, pages 328-341. Springer, 2011.
+ *
+ * Scavenger traffic (Less-than-Best-Effort) should disable coexistence
+ * heuristics using parameters use_shadow=0 and use_ineff=0.
+ *
+ * Parameters window, backoff_beta, and backoff_factor are crucial for
+ * throughput and delay. Future work is needed to determine better defaults,
+ * and to provide guidelines for use in different environments/contexts.
+ *
+ * Except for window, knobs are configured via /sys/module/tcp_cdg/parameters/.
+ * Parameter window is only configurable when loading tcp_cdg as a module.
+ *
+ * Notable differences from paper/FreeBSD:
+ *   o Using Hybrid Slow start and Proportional Rate Reduction.
+ *   o Add toggle for shadow window mechanism. Suggested by David Hayes.
+ *   o Add toggle for non-congestion loss tolerance.
+ *   o Scaling parameter G is changed to a backoff factor;
+ *     conversion is given by: backoff_factor = 1000/(G * window).
+ *   o Limit shadow window to 2 * cwnd, or to cwnd when application limited.
+ *   o More accurate e^-x.
+ */
+#include <linux/kernel.h>
+#include <linux/random.h>
+#include <linux/module.h>
+#include <net/tcp.h>
+
+#define HYSTART_ACK_TRAIN	1
+#define HYSTART_DELAY		2
+
+static int window __read_mostly = 8;
+static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
+static unsigned int backoff_factor __read_mostly = 42;
+static unsigned int hystart_detect __read_mostly = 3;
+static unsigned int use_ineff __read_mostly = 5;
+static bool use_shadow __read_mostly = true;
+static bool use_tolerance __read_mostly;
+
+module_param(window, int, 0444);
+MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
+module_param(backoff_beta, uint, 0644);
+MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
+module_param(backoff_factor, uint, 0644);
+MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
+module_param(hystart_detect, uint, 0644);
+MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
+		 "(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
+module_param(use_ineff, uint, 0644);
+MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
+module_param(use_shadow, bool, 0644);
+MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
+module_param(use_tolerance, bool, 0644);
+MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
+
+struct minmax {
+	union {
+		struct {
+			s32 min;
+			s32 max;
+		};
+		u64 v64;
+	};
+};
+
+enum cdg_state {
+	CDG_UNKNOWN = 0,
+	CDG_NONFULL = 1,
+	CDG_FULL    = 2,
+	CDG_BACKOFF = 3,
+};
+
+struct cdg {
+	struct minmax rtt;
+	struct minmax rtt_prev;
+	struct minmax *gradients;
+	struct minmax gsum;
+	bool gfilled;
+	u8  tail;
+	u8  state;
+	u8  delack;
+	u32 rtt_seq;
+	u32 undo_cwnd;
+	u32 shadow_wnd;
+	u16 backoff_cnt;
+	u16 sample_cnt;
+	s32 delay_min;
+	u32 last_ack;
+	u32 round_start;
+};
+
+/**
+ * nexp_u32 - negative base-e exponential
+ * @ux: x in units of micro
+ *
+ * Returns exp(ux * -1e-6) * U32_MAX.
+ */
+static u32 __pure nexp_u32(u32 ux)
+{
+	static const u16 v[] = {
+		/* exp(-x)*65536-1 for x = 0, 0.000256, 0.000512, ... */
+		65535,
+		65518, 65501, 65468, 65401, 65267, 65001, 64470, 63422,
+		61378, 57484, 50423, 38795, 22965, 8047,  987,   14,
+	};
+	u32 msb = ux >> 8;
+	u32 res;
+	int i;
+
+	/* Cut off when ux >= 2^24 (actual result is <= 222/U32_MAX). */
+	if (msb > U16_MAX)
+		return 0;
+
+	/* Scale first eight bits linearly: */
+	res = U32_MAX - (ux & 0xff) * (U32_MAX / 1000000);
+
+	/* Obtain e^(x + y + ...) by computing e^x * e^y * ...: */
+	for (i = 1; msb; i++, msb >>= 1) {
+		u32 y = v[i & -(msb & 1)] + U32_C(1);
+
+		res = ((u64)res * y) >> 16;
+	}
+
+	return res;
+}
+
+/* Based on the HyStart algorithm (by Ha et al.) that is implemented in
+ * tcp_cubic. Differences/experimental changes:
+ *   o Using Hayes' delayed ACK filter.
+ *   o Using a usec clock for the ACK train.
+ *   o Reset ACK train when application limited.
+ *   o Invoked at any cwnd (i.e. also when cwnd < 16).
+ *   o Invoked only when cwnd < ssthresh (i.e. not when cwnd == ssthresh).
+ */
+static void tcp_cdg_hystart_update(struct sock *sk)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	ca->delay_min = min_not_zero(ca->delay_min, ca->rtt.min);
+	if (ca->delay_min == 0)
+		return;
+
+	if (hystart_detect & HYSTART_ACK_TRAIN) {
+		u32 now_us = local_clock() / NSEC_PER_USEC;
+
+		if (ca->last_ack == 0 || !tcp_is_cwnd_limited(sk)) {
+			ca->last_ack = now_us;
+			ca->round_start = now_us;
+		} else if (before(now_us, ca->last_ack + 3000)) {
+			u32 base_owd = max(ca->delay_min / 2U, 125U);
+
+			ca->last_ack = now_us;
+			if (after(now_us, ca->round_start + base_owd)) {
+				NET_INC_STATS_BH(sock_net(sk),
+						 LINUX_MIB_TCPHYSTARTTRAINDETECT);
+				NET_ADD_STATS_BH(sock_net(sk),
+						 LINUX_MIB_TCPHYSTARTTRAINCWND,
+						 tp->snd_cwnd);
+				tp->snd_ssthresh = tp->snd_cwnd;
+				return;
+			}
+		}
+	}
+
+	if (hystart_detect & HYSTART_DELAY) {
+		if (ca->sample_cnt < 8) {
+			ca->sample_cnt++;
+		} else {
+			s32 thresh = max(ca->delay_min + ca->delay_min / 8U,
+					 125U);
+
+			if (ca->rtt.min > thresh) {
+				NET_INC_STATS_BH(sock_net(sk),
+						 LINUX_MIB_TCPHYSTARTDELAYDETECT);
+				NET_ADD_STATS_BH(sock_net(sk),
+						 LINUX_MIB_TCPHYSTARTDELAYCWND,
+						 tp->snd_cwnd);
+				tp->snd_ssthresh = tp->snd_cwnd;
+			}
+		}
+	}
+}
+
+static s32 tcp_cdg_grad(struct cdg *ca)
+{
+	s32 gmin = ca->rtt.min - ca->rtt_prev.min;
+	s32 gmax = ca->rtt.max - ca->rtt_prev.max;
+	s32 grad;
+
+	if (ca->gradients) {
+		ca->gsum.min += gmin - ca->gradients[ca->tail].min;
+		ca->gsum.max += gmax - ca->gradients[ca->tail].max;
+		ca->gradients[ca->tail].min = gmin;
+		ca->gradients[ca->tail].max = gmax;
+		ca->tail = (ca->tail + 1) & (window - 1);
+		gmin = ca->gsum.min;
+		gmax = ca->gsum.max;
+	}
+
+	/* We keep sums to ignore gradients during cwnd reductions;
+	 * the paper's smoothed gradients otherwise simplify to:
+	 * (rtt_latest - rtt_oldest) / window.
+	 *
+	 * We also drop division by window here.
+	 */
+	grad = gmin > 0 ? gmin : gmax;
+
+	/* Extrapolate missing values in gradient window: */
+	if (!ca->gfilled) {
+		if (!ca->gradients && window > 1)
+			grad *= window; /* Memory allocation failed. */
+		else if (ca->tail == 0)
+			ca->gfilled = true;
+		else
+			grad = (grad * window) / (int)ca->tail;
+	}
+
+	/* Backoff was effectual: */
+	if (gmin <= -32 || gmax <= -32)
+		ca->backoff_cnt = 0;
+
+	if (use_tolerance) {
+		/* Reduce small variations to zero: */
+		gmin = DIV_ROUND_CLOSEST(gmin, 64);
+		gmax = DIV_ROUND_CLOSEST(gmax, 64);
+
+		if (gmin > 0 && gmax <= 0)
+			ca->state = CDG_FULL;
+		else if ((gmin > 0 && gmax > 0) || gmax < 0)
+			ca->state = CDG_NONFULL;
+	}
+	return grad;
+}
+
+static bool tcp_cdg_backoff(struct sock *sk, u32 grad)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	if (prandom_u32() <= nexp_u32(grad * backoff_factor))
+		return false;
+
+	if (use_ineff) {
+		ca->backoff_cnt++;
+		if (ca->backoff_cnt > use_ineff)
+			return false;
+	}
+
+	ca->shadow_wnd = max(ca->shadow_wnd, tp->snd_cwnd);
+	ca->state = CDG_BACKOFF;
+	tcp_enter_cwr(sk);
+	return true;
+}
+
+/* Not called in CWR or Recovery state. */
+static void tcp_cdg_cong_avoid(struct sock *sk, u32 ack, u32 acked)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+	u32 prior_snd_cwnd;
+	u32 incr;
+
+	if (tp->snd_cwnd < tp->snd_ssthresh && hystart_detect)
+		tcp_cdg_hystart_update(sk);
+
+	if (after(ack, ca->rtt_seq) && ca->rtt.v64) {
+		s32 grad = 0;
+
+		if (ca->rtt_prev.v64)
+			grad = tcp_cdg_grad(ca);
+		ca->rtt_seq = tp->snd_nxt;
+		ca->rtt_prev = ca->rtt;
+		ca->rtt.v64 = 0;
+		ca->last_ack = 0;
+		ca->sample_cnt = 0;
+
+		if (grad > 0 && tcp_cdg_backoff(sk, grad))
+			return;
+	}
+
+	if (!tcp_is_cwnd_limited(sk)) {
+		ca->shadow_wnd = min(ca->shadow_wnd, tp->snd_cwnd);
+		return;
+	}
+
+	prior_snd_cwnd = tp->snd_cwnd;
+	tcp_reno_cong_avoid(sk, ack, acked);
+
+	incr = tp->snd_cwnd - prior_snd_cwnd;
+	ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
+}
+
+static void tcp_cdg_acked(struct sock *sk, u32 num_acked, s32 rtt_us)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	if (rtt_us <= 0)
+		return;
+
+	/* A heuristic for filtering delayed ACKs, adapted from:
+	 * D.A. Hayes. "Timing enhancements to the FreeBSD kernel to support
+	 * delay and rate based TCP mechanisms." TR 100219A. CAIA, 2010.
+	 */
+	if (tp->sacked_out == 0) {
+		if (num_acked == 1 && ca->delack) {
+			/* A delayed ACK is only used for the minimum if it is
+			 * provenly lower than an existing non-zero minimum.
+			 */
+			ca->rtt.min = min(ca->rtt.min, rtt_us);
+			ca->delack--;
+			return;
+		} else if (num_acked > 1 && ca->delack < 5) {
+			ca->delack++;
+		}
+	}
+
+	ca->rtt.min = min_not_zero(ca->rtt.min, rtt_us);
+	ca->rtt.max = max(ca->rtt.max, rtt_us);
+}
+
+static u32 tcp_cdg_ssthresh(struct sock *sk)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	ca->undo_cwnd = tp->snd_cwnd;
+
+	if (ca->state == CDG_BACKOFF)
+		return max(2U, (tp->snd_cwnd * min(1024U, backoff_beta)) >> 10);
+
+	if (ca->state == CDG_NONFULL && use_tolerance)
+		return tp->snd_cwnd;
+
+	ca->shadow_wnd = min(ca->shadow_wnd >> 1, tp->snd_cwnd);
+	if (use_shadow)
+		return max3(2U, ca->shadow_wnd, tp->snd_cwnd >> 1);
+	return max(2U, tp->snd_cwnd >> 1);
+}
+
+static u32 tcp_cdg_undo_cwnd(struct sock *sk)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+
+	return max(tcp_sk(sk)->snd_cwnd, ca->undo_cwnd);
+}
+
+static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct minmax *gradients;
+
+	switch (ev) {
+	case CA_EVENT_CWND_RESTART:
+		gradients = ca->gradients;
+		if (gradients)
+			memset(gradients, 0, window * sizeof(gradients[0]));
+		memset(ca, 0, sizeof(*ca));
+
+		ca->gradients = gradients;
+		ca->rtt_seq = tp->snd_nxt;
+		ca->shadow_wnd = tp->snd_cwnd;
+		break;
+	case CA_EVENT_COMPLETE_CWR:
+		ca->state = CDG_UNKNOWN;
+		ca->rtt_seq = tp->snd_nxt;
+		ca->rtt_prev = ca->rtt;
+		ca->rtt.v64 = 0;
+		break;
+	default:
+		break;
+	}
+}
+
+static void tcp_cdg_init(struct sock *sk)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	/* We silently fall back to window = 1 if allocation fails. */
+	if (window > 1)
+		ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
+					GFP_NOWAIT | __GFP_NOWARN);
+	ca->rtt_seq = tp->snd_nxt;
+	ca->shadow_wnd = tp->snd_cwnd;
+}
+
+static void tcp_cdg_release(struct sock *sk)
+{
+	struct cdg *ca = inet_csk_ca(sk);
+
+	kfree(ca->gradients);
+}
+
+struct tcp_congestion_ops tcp_cdg __read_mostly = {
+	.cong_avoid = tcp_cdg_cong_avoid,
+	.cwnd_event = tcp_cdg_cwnd_event,
+	.pkts_acked = tcp_cdg_acked,
+	.undo_cwnd = tcp_cdg_undo_cwnd,
+	.ssthresh = tcp_cdg_ssthresh,
+	.release = tcp_cdg_release,
+	.init = tcp_cdg_init,
+	.owner = THIS_MODULE,
+	.name = "cdg",
+};
+
+static int __init tcp_cdg_register(void)
+{
+	if (backoff_beta > 1024 || window < 1 || window > 256)
+		return -ERANGE;
+	if (!is_power_of_2(window))
+		return -EINVAL;
+
+	BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
+	tcp_register_congestion_control(&tcp_cdg);
+	return 0;
+}
+
+static void __exit tcp_cdg_unregister(void)
+{
+	tcp_unregister_congestion_control(&tcp_cdg);
+}
+
+module_init(tcp_cdg_register);
+module_exit(tcp_cdg_unregister);
+MODULE_AUTHOR("Kenneth Klette Jonassen");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("TCP CDG");