[linux-2.6-block.git] / net / sched / sch_pie.c

/* Copyright (C) 2013 Cisco Systems, Inc, 2013.
 *
 * 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.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * Author: Vijay Subramanian <vijaynsu@cisco.com>
 * Author: Mythili Prabhu <mysuryan@cisco.com>
 *
 * ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
 * University of Oslo, Norway.
 *
 * References:
 * IETF draft submission: http://tools.ietf.org/html/draft-pan-aqm-pie-00
 * IEEE  Conference on High Performance Switching and Routing 2013 :
 * "PIE: A * Lightweight Control Scheme to Address the Bufferbloat Problem"
 */

#include <linux/module.h>
#include <linux/slab.h>
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/errno.h>
#include <linux/skbuff.h>
#include <net/pkt_sched.h>
#include <net/inet_ecn.h>

#define QUEUE_THRESHOLD 10000
#define DQCOUNT_INVALID -1
#define MAX_PROB  0xffffffff
#define PIE_SCALE 8

/* parameters used */
struct pie_params {
	psched_time_t target;	/* user specified target delay in pschedtime */
	u32 tupdate;		/* timer frequency (in jiffies) */
	u32 limit;		/* number of packets that can be enqueued */
	u32 alpha;		/* alpha and beta are between 0 and 32 */
	u32 beta;		/* and are used for shift relative to 1 */
	bool ecn;		/* true if ecn is enabled */
	bool bytemode;		/* to scale drop early prob based on pkt size */
};

/* variables used */
struct pie_vars {
	u32 prob;		/* probability but scaled by u32 limit. */
	psched_time_t burst_time;
	psched_time_t qdelay;
	psched_time_t qdelay_old;
	u64 dq_count;		/* measured in bytes */
	psched_time_t dq_tstamp;	/* drain rate */
	u32 avg_dq_rate;	/* bytes per pschedtime tick,scaled */
	u32 qlen_old;		/* in bytes */
};

/* statistics gathering */
struct pie_stats {
	u32 packets_in;		/* total number of packets enqueued */
	u32 dropped;		/* packets dropped due to pie_action */
	u32 overlimit;		/* dropped due to lack of space in queue */
	u32 maxq;		/* maximum queue size */
	u32 ecn_mark;		/* packets marked with ECN */
};

/* private data for the Qdisc */
struct pie_sched_data {
	struct pie_params params;
	struct pie_vars vars;
	struct pie_stats stats;
	struct timer_list adapt_timer;
};

static void pie_params_init(struct pie_params *params)
{
	params->alpha = 2;
	params->beta = 20;
	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC);	/* 30 ms */
	params->limit = 1000;	/* default of 1000 packets */
	params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC);	/* 20 ms */
	params->ecn = false;
	params->bytemode = false;
}

static void pie_vars_init(struct pie_vars *vars)
{
	vars->dq_count = DQCOUNT_INVALID;
	vars->avg_dq_rate = 0;
	/* default of 100 ms in pschedtime */
	vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
}

static bool drop_early(struct Qdisc *sch, u32 packet_size)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	u32 rnd;
	u32 local_prob = q->vars.prob;
	u32 mtu = psched_mtu(qdisc_dev(sch));

	/* If there is still burst allowance left skip random early drop */
	if (q->vars.burst_time > 0)
		return false;

	/* If current delay is less than half of target, and
	 * if drop prob is low already, disable early_drop
	 */
	if ((q->vars.qdelay < q->params.target / 2)
	    && (q->vars.prob < MAX_PROB / 5))
		return false;

	/* If we have fewer than 2 mtu-sized packets, disable drop_early,
	 * similar to min_th in RED
	 */
	if (sch->qstats.backlog < 2 * mtu)
		return false;

	/* If bytemode is turned on, use packet size to compute new
	 * probablity. Smaller packets will have lower drop prob in this case
	 */
	if (q->params.bytemode && packet_size <= mtu)
		local_prob = (local_prob / mtu) * packet_size;
	else
		local_prob = q->vars.prob;

	rnd = prandom_u32();
	if (rnd < local_prob)
		return true;

	return false;
}

static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	bool enqueue = false;

	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
		q->stats.overlimit++;
		goto out;
	}

	if (!drop_early(sch, skb->len)) {
		enqueue = true;
	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
		   INET_ECN_set_ce(skb)) {
		/* If packet is ecn capable, mark it if drop probability
		 * is lower than 10%, else drop it.
		 */
		q->stats.ecn_mark++;
		enqueue = true;
	}

	/* we can enqueue the packet */
	if (enqueue) {
		q->stats.packets_in++;
		if (qdisc_qlen(sch) > q->stats.maxq)
			q->stats.maxq = qdisc_qlen(sch);

		return qdisc_enqueue_tail(skb, sch);
	}

out:
	q->stats.dropped++;
	return qdisc_drop(skb, sch);
}

static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
	[TCA_PIE_TARGET] = {.type = NLA_U32},
	[TCA_PIE_LIMIT] = {.type = NLA_U32},
	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
	[TCA_PIE_ALPHA] = {.type = NLA_U32},
	[TCA_PIE_BETA] = {.type = NLA_U32},
	[TCA_PIE_ECN] = {.type = NLA_U32},
	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
};

static int pie_change(struct Qdisc *sch, struct nlattr *opt)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	struct nlattr *tb[TCA_PIE_MAX + 1];
	unsigned int qlen, dropped = 0;
	int err;

	if (!opt)
		return -EINVAL;

	err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy);
	if (err < 0)
		return err;

	sch_tree_lock(sch);

	/* convert from microseconds to pschedtime */
	if (tb[TCA_PIE_TARGET]) {
		/* target is in us */
		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);

		/* convert to pschedtime */
		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
	}

	/* tupdate is in jiffies */
	if (tb[TCA_PIE_TUPDATE])
		q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));

	if (tb[TCA_PIE_LIMIT]) {
		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);

		q->params.limit = limit;
		sch->limit = limit;
	}

	if (tb[TCA_PIE_ALPHA])
		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);

	if (tb[TCA_PIE_BETA])
		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);

	if (tb[TCA_PIE_ECN])
		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);

	if (tb[TCA_PIE_BYTEMODE])
		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);

	/* Drop excess packets if new limit is lower */
	qlen = sch->q.qlen;
	while (sch->q.qlen > sch->limit) {
		struct sk_buff *skb = __skb_dequeue(&sch->q);

		dropped += qdisc_pkt_len(skb);
		qdisc_qstats_backlog_dec(sch, skb);
		qdisc_drop(skb, sch);
	}
	qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);

	sch_tree_unlock(sch);
	return 0;
}

static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
{

	struct pie_sched_data *q = qdisc_priv(sch);
	int qlen = sch->qstats.backlog;	/* current queue size in bytes */

	/* If current queue is about 10 packets or more and dq_count is unset
	 * we have enough packets to calculate the drain rate. Save
	 * current time as dq_tstamp and start measurement cycle.
	 */
	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
		q->vars.dq_tstamp = psched_get_time();
		q->vars.dq_count = 0;
	}

	/* Calculate the average drain rate from this value.  If queue length
	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
	 * the dq_count to -1 as we don't have enough packets to calculate the
	 * drain rate anymore The following if block is entered only when we
	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
	 * and we calculate the drain rate for the threshold here.  dq_count is
	 * in bytes, time difference in psched_time, hence rate is in
	 * bytes/psched_time.
	 */
	if (q->vars.dq_count != DQCOUNT_INVALID) {
		q->vars.dq_count += skb->len;

		if (q->vars.dq_count >= QUEUE_THRESHOLD) {
			psched_time_t now = psched_get_time();
			u32 dtime = now - q->vars.dq_tstamp;
			u32 count = q->vars.dq_count << PIE_SCALE;

			if (dtime == 0)
				return;

			count = count / dtime;

			if (q->vars.avg_dq_rate == 0)
				q->vars.avg_dq_rate = count;
			else
				q->vars.avg_dq_rate =
				    (q->vars.avg_dq_rate -
				     (q->vars.avg_dq_rate >> 3)) + (count >> 3);

			/* If the queue has receded below the threshold, we hold
			 * on to the last drain rate calculated, else we reset
			 * dq_count to 0 to re-enter the if block when the next
			 * packet is dequeued
			 */
			if (qlen < QUEUE_THRESHOLD)
				q->vars.dq_count = DQCOUNT_INVALID;
			else {
				q->vars.dq_count = 0;
				q->vars.dq_tstamp = psched_get_time();
			}

			if (q->vars.burst_time > 0) {
				if (q->vars.burst_time > dtime)
					q->vars.burst_time -= dtime;
				else
					q->vars.burst_time = 0;
			}
		}
	}
}

static void calculate_probability(struct Qdisc *sch)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	u32 qlen = sch->qstats.backlog;	/* queue size in bytes */
	psched_time_t qdelay = 0;	/* in pschedtime */
	psched_time_t qdelay_old = q->vars.qdelay;	/* in pschedtime */
	s32 delta = 0;		/* determines the change in probability */
	u32 oldprob;
	u32 alpha, beta;
	bool update_prob = true;

	q->vars.qdelay_old = q->vars.qdelay;

	if (q->vars.avg_dq_rate > 0)
		qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
	else
		qdelay = 0;

	/* If qdelay is zero and qlen is not, it means qlen is very small, less
	 * than dequeue_rate, so we do not update probabilty in this round
	 */
	if (qdelay == 0 && qlen != 0)
		update_prob = false;

	/* In the algorithm, alpha and beta are between 0 and 2 with typical
	 * value for alpha as 0.125. In this implementation, we use values 0-32
	 * passed from user space to represent this. Also, alpha and beta have
	 * unit of HZ and need to be scaled before they can used to update
	 * probability. alpha/beta are updated locally below by 1) scaling them
	 * appropriately 2) scaling down by 16 to come to 0-2 range.
	 * Please see paper for details.
	 *
	 * We scale alpha and beta differently depending on whether we are in
	 * light, medium or high dropping mode.
	 */
	if (q->vars.prob < MAX_PROB / 100) {
		alpha =
		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
		beta =
		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
	} else if (q->vars.prob < MAX_PROB / 10) {
		alpha =
		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
		beta =
		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
	} else {
		alpha =
		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
		beta =
		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
	}

	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
	delta += alpha * ((qdelay - q->params.target));
	delta += beta * ((qdelay - qdelay_old));

	oldprob = q->vars.prob;

	/* to ensure we increase probability in steps of no more than 2% */
	if (delta > (s32) (MAX_PROB / (100 / 2)) &&
	    q->vars.prob >= MAX_PROB / 10)
		delta = (MAX_PROB / 100) * 2;

	/* Non-linear drop:
	 * Tune drop probability to increase quickly for high delays(>= 250ms)
	 * 250ms is derived through experiments and provides error protection
	 */

	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
		delta += MAX_PROB / (100 / 2);

	q->vars.prob += delta;

	if (delta > 0) {
		/* prevent overflow */
		if (q->vars.prob < oldprob) {
			q->vars.prob = MAX_PROB;
			/* Prevent normalization error. If probability is at
			 * maximum value already, we normalize it here, and
			 * skip the check to do a non-linear drop in the next
			 * section.
			 */
			update_prob = false;
		}
	} else {
		/* prevent underflow */
		if (q->vars.prob > oldprob)
			q->vars.prob = 0;
	}

	/* Non-linear drop in probability: Reduce drop probability quickly if
	 * delay is 0 for 2 consecutive Tupdate periods.
	 */

	if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
		q->vars.prob = (q->vars.prob * 98) / 100;

	q->vars.qdelay = qdelay;
	q->vars.qlen_old = qlen;

	/* We restart the measurement cycle if the following conditions are met
	 * 1. If the delay has been low for 2 consecutive Tupdate periods
	 * 2. Calculated drop probability is zero
	 * 3. We have atleast one estimate for the avg_dq_rate ie.,
	 *    is a non-zero value
	 */
	if ((q->vars.qdelay < q->params.target / 2) &&
	    (q->vars.qdelay_old < q->params.target / 2) &&
	    (q->vars.prob == 0) &&
	    (q->vars.avg_dq_rate > 0))
		pie_vars_init(&q->vars);
}

static void pie_timer(unsigned long arg)
{
	struct Qdisc *sch = (struct Qdisc *)arg;
	struct pie_sched_data *q = qdisc_priv(sch);
	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));

	spin_lock(root_lock);
	calculate_probability(sch);

	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
	if (q->params.tupdate)
		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
	spin_unlock(root_lock);

}

static int pie_init(struct Qdisc *sch, struct nlattr *opt)
{
	struct pie_sched_data *q = qdisc_priv(sch);

	pie_params_init(&q->params);
	pie_vars_init(&q->vars);
	sch->limit = q->params.limit;

	setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);

	if (opt) {
		int err = pie_change(sch, opt);

		if (err)
			return err;
	}

	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
	return 0;
}

static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	struct nlattr *opts;

	opts = nla_nest_start(skb, TCA_OPTIONS);
	if (opts == NULL)
		goto nla_put_failure;

	/* convert target from pschedtime to us */
	if (nla_put_u32(skb, TCA_PIE_TARGET,
			((u32) PSCHED_TICKS2NS(q->params.target)) /
			NSEC_PER_USEC) ||
	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
	    nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
	    nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
	    nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
	    nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
	    nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
		goto nla_put_failure;

	return nla_nest_end(skb, opts);

nla_put_failure:
	nla_nest_cancel(skb, opts);
	return -1;

}

static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	struct tc_pie_xstats st = {
		.prob		= q->vars.prob,
		.delay		= ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
				   NSEC_PER_USEC,
		/* unscale and return dq_rate in bytes per sec */
		.avg_dq_rate	= q->vars.avg_dq_rate *
				  (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
		.packets_in	= q->stats.packets_in,
		.overlimit	= q->stats.overlimit,
		.maxq		= q->stats.maxq,
		.dropped	= q->stats.dropped,
		.ecn_mark	= q->stats.ecn_mark,
	};

	return gnet_stats_copy_app(d, &st, sizeof(st));
}

static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
{
	struct sk_buff *skb;
	skb = __qdisc_dequeue_head(sch, &sch->q);

	if (!skb)
		return NULL;

	pie_process_dequeue(sch, skb);
	return skb;
}

static void pie_reset(struct Qdisc *sch)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	qdisc_reset_queue(sch);
	pie_vars_init(&q->vars);
}

static void pie_destroy(struct Qdisc *sch)
{
	struct pie_sched_data *q = qdisc_priv(sch);
	q->params.tupdate = 0;
	del_timer_sync(&q->adapt_timer);
}

static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
	.id = "pie",
	.priv_size	= sizeof(struct pie_sched_data),
	.enqueue	= pie_qdisc_enqueue,
	.dequeue	= pie_qdisc_dequeue,
	.peek		= qdisc_peek_dequeued,
	.init		= pie_init,
	.destroy	= pie_destroy,
	.reset		= pie_reset,
	.change		= pie_change,
	.dump		= pie_dump,
	.dump_stats	= pie_dump_stats,
	.owner		= THIS_MODULE,
};

static int __init pie_module_init(void)
{
	return register_qdisc(&pie_qdisc_ops);
}

static void __exit pie_module_exit(void)
{
	unregister_qdisc(&pie_qdisc_ops);
}

module_init(pie_module_init);
module_exit(pie_module_exit);

MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
MODULE_AUTHOR("Vijay Subramanian");
MODULE_AUTHOR("Mythili Prabhu");
MODULE_LICENSE("GPL");
Commit	Line	Data
d4b36210 VS	1	/* Copyright (C) 2013 Cisco Systems, Inc, 2013.
	2	*
	3	* This program is free software; you can redistribute it and/or
	4	* modify it under the terms of the GNU General Public License
	5	* as published by the Free Software Foundation; either version 2
	6	* of the License.
	7	*
	8	* This program is distributed in the hope that it will be useful,
	9	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	10	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	11	* GNU General Public License for more details.
	12	*
	13	* Author: Vijay Subramanian <vijaynsu@cisco.com>
	14	* Author: Mythili Prabhu <mysuryan@cisco.com>
	15	*
	16	* ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
	17	* University of Oslo, Norway.
219e288e VS	18	*
	19	* References:
	20	* IETF draft submission: http://tools.ietf.org/html/draft-pan-aqm-pie-00
	21	* IEEE Conference on High Performance Switching and Routing 2013 :
	22	* "PIE: A * Lightweight Control Scheme to Address the Bufferbloat Problem"
d4b36210 VS	23	*/
	24
	25	#include <linux/module.h>
	26	#include <linux/slab.h>
	27	#include <linux/types.h>
	28	#include <linux/kernel.h>
	29	#include <linux/errno.h>
	30	#include <linux/skbuff.h>
	31	#include <net/pkt_sched.h>
	32	#include <net/inet_ecn.h>
	33
	34	#define QUEUE_THRESHOLD 10000
	35	#define DQCOUNT_INVALID -1
	36	#define MAX_PROB 0xffffffff
	37	#define PIE_SCALE 8
	38
	39	/* parameters used */
	40	struct pie_params {
	41	psched_time_t target; /* user specified target delay in pschedtime */
	42	u32 tupdate; /* timer frequency (in jiffies) */
	43	u32 limit; /* number of packets that can be enqueued */
219e288e	44	u32 alpha; /* alpha and beta are between 0 and 32 */
d4b36210 VS	45	u32 beta; /* and are used for shift relative to 1 */
	46	bool ecn; /* true if ecn is enabled */
	47	bool bytemode; /* to scale drop early prob based on pkt size */
	48	};
	49
	50	/* variables used */
	51	struct pie_vars {
	52	u32 prob; /* probability but scaled by u32 limit. */
	53	psched_time_t burst_time;
	54	psched_time_t qdelay;
	55	psched_time_t qdelay_old;
	56	u64 dq_count; /* measured in bytes */
	57	psched_time_t dq_tstamp; /* drain rate */
	58	u32 avg_dq_rate; /* bytes per pschedtime tick,scaled */
	59	u32 qlen_old; /* in bytes */
	60	};
	61
	62	/* statistics gathering */
	63	struct pie_stats {
	64	u32 packets_in; /* total number of packets enqueued */
	65	u32 dropped; /* packets dropped due to pie_action */
	66	u32 overlimit; /* dropped due to lack of space in queue */
	67	u32 maxq; /* maximum queue size */
	68	u32 ecn_mark; /* packets marked with ECN */
	69	};
	70
	71	/* private data for the Qdisc */
	72	struct pie_sched_data {
	73	struct pie_params params;
	74	struct pie_vars vars;
	75	struct pie_stats stats;
	76	struct timer_list adapt_timer;
	77	};
	78
	79	static void pie_params_init(struct pie_params *params)
	80	{
	81	params->alpha = 2;
	82	params->beta = 20;
	83	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC); /* 30 ms */
	84	params->limit = 1000; /* default of 1000 packets */
	85	params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC); /* 20 ms */
	86	params->ecn = false;
	87	params->bytemode = false;
	88	}
	89
	90	static void pie_vars_init(struct pie_vars *vars)
	91	{
	92	vars->dq_count = DQCOUNT_INVALID;
	93	vars->avg_dq_rate = 0;
	94	/* default of 100 ms in pschedtime */
	95	vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
	96	}
	97
	98	static bool drop_early(struct Qdisc *sch, u32 packet_size)
	99	{
	100	struct pie_sched_data *q = qdisc_priv(sch);
	101	u32 rnd;
	102	u32 local_prob = q->vars.prob;
	103	u32 mtu = psched_mtu(qdisc_dev(sch));
	104
	105	/* If there is still burst allowance left skip random early drop */
	106	if (q->vars.burst_time > 0)
	107	return false;
	108
109	/* If current delay is less than half of target, and
110	* if drop prob is low already, disable early_drop
111	*/
112	if ((q->vars.qdelay < q->params.target / 2)
113	&& (q->vars.prob < MAX_PROB / 5))
114	return false;
115
116	/* If we have fewer than 2 mtu-sized packets, disable drop_early,
117	* similar to min_th in RED
118	*/
119	if (sch->qstats.backlog < 2 * mtu)
120	return false;
121
122	/* If bytemode is turned on, use packet size to compute new
123	* probablity. Smaller packets will have lower drop prob in this case
124	*/
125	if (q->params.bytemode && packet_size <= mtu)
126	local_prob = (local_prob / mtu) * packet_size;
127	else
128	local_prob = q->vars.prob;
129
63862b5b	130	rnd = prandom_u32();
d4b36210 VS	131	if (rnd < local_prob)
	132	return true;
	133
	134	return false;
	135	}
	136
	137	static int pie_qdisc_enqueue(struct sk_buff skb, struct Qdisc sch)
	138	{
	139	struct pie_sched_data *q = qdisc_priv(sch);
	140	bool enqueue = false;
	141
	142	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
	143	q->stats.overlimit++;
	144	goto out;
	145	}
	146
	147	if (!drop_early(sch, skb->len)) {
	148	enqueue = true;
	149	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
	150	INET_ECN_set_ce(skb)) {
	151	/* If packet is ecn capable, mark it if drop probability
	152	* is lower than 10%, else drop it.
	153	*/
	154	q->stats.ecn_mark++;
	155	enqueue = true;
	156	}
	157
	158	/* we can enqueue the packet */
	159	if (enqueue) {
	160	q->stats.packets_in++;
	161	if (qdisc_qlen(sch) > q->stats.maxq)
	162	q->stats.maxq = qdisc_qlen(sch);
	163
	164	return qdisc_enqueue_tail(skb, sch);
	165	}
	166
	167	out:
	168	q->stats.dropped++;
	169	return qdisc_drop(skb, sch);
	170	}
	171
	172	static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
	173	[TCA_PIE_TARGET] = {.type = NLA_U32},
	174	[TCA_PIE_LIMIT] = {.type = NLA_U32},
	175	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
	176	[TCA_PIE_ALPHA] = {.type = NLA_U32},
	177	[TCA_PIE_BETA] = {.type = NLA_U32},
	178	[TCA_PIE_ECN] = {.type = NLA_U32},
	179	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
	180	};
	181
	182	static int pie_change(struct Qdisc sch, struct nlattr opt)
	183	{
	184	struct pie_sched_data *q = qdisc_priv(sch);
	185	struct nlattr *tb[TCA_PIE_MAX + 1];
2ccccf5f	186	unsigned int qlen, dropped = 0;
d4b36210 VS	187	int err;
	188
	189	if (!opt)
	190	return -EINVAL;
	191
	192	err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy);
	193	if (err < 0)
	194	return err;
	195
	196	sch_tree_lock(sch);
	197
	198	/* convert from microseconds to pschedtime */
	199	if (tb[TCA_PIE_TARGET]) {
	200	/* target is in us */
	201	u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
	202
	203	/* convert to pschedtime */
	204	q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
	205	}
	206
	207	/* tupdate is in jiffies */
	208	if (tb[TCA_PIE_TUPDATE])
	209	q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
	210
	211	if (tb[TCA_PIE_LIMIT]) {
	212	u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
	213
	214	q->params.limit = limit;
	215	sch->limit = limit;
	216	}
	217
	218	if (tb[TCA_PIE_ALPHA])
	219	q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
	220
	221	if (tb[TCA_PIE_BETA])
	222	q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
	223
	224	if (tb[TCA_PIE_ECN])
	225	q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
	226
	227	if (tb[TCA_PIE_BYTEMODE])
	228	q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
	229
	230	/* Drop excess packets if new limit is lower */
	231	qlen = sch->q.qlen;
	232	while (sch->q.qlen > sch->limit) {
	233	struct sk_buff *skb = __skb_dequeue(&sch->q);
	234
2ccccf5f	235	dropped += qdisc_pkt_len(skb);
25331d6c	236	qdisc_qstats_backlog_dec(sch, skb);
d4b36210 VS	237	qdisc_drop(skb, sch);
d4b36210 VS	238	}
2ccccf5f	239	qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
d4b36210 VS	240
	241	sch_tree_unlock(sch);
	242	return 0;
	243	}
	244
	245	static void pie_process_dequeue(struct Qdisc sch, struct sk_buff skb)
	246	{
	247
	248	struct pie_sched_data *q = qdisc_priv(sch);
	249	int qlen = sch->qstats.backlog; /* current queue size in bytes */
	250
	251	/* If current queue is about 10 packets or more and dq_count is unset
	252	* we have enough packets to calculate the drain rate. Save
	253	* current time as dq_tstamp and start measurement cycle.
	254	*/
	255	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
	256	q->vars.dq_tstamp = psched_get_time();
	257	q->vars.dq_count = 0;
	258	}
	259
	260	/* Calculate the average drain rate from this value. If queue length
	261	* has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
	262	* the dq_count to -1 as we don't have enough packets to calculate the
	263	* drain rate anymore The following if block is entered only when we
	264	* have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
	265	* and we calculate the drain rate for the threshold here. dq_count is
	266	* in bytes, time difference in psched_time, hence rate is in
	267	* bytes/psched_time.
	268	*/
	269	if (q->vars.dq_count != DQCOUNT_INVALID) {
	270	q->vars.dq_count += skb->len;
	271
	272	if (q->vars.dq_count >= QUEUE_THRESHOLD) {
	273	psched_time_t now = psched_get_time();
	274	u32 dtime = now - q->vars.dq_tstamp;
	275	u32 count = q->vars.dq_count << PIE_SCALE;
	276
	277	if (dtime == 0)
	278	return;
	279
	280	count = count / dtime;
	281
	282	if (q->vars.avg_dq_rate == 0)
	283	q->vars.avg_dq_rate = count;
	284	else
	285	q->vars.avg_dq_rate =
	286	(q->vars.avg_dq_rate -
	287	(q->vars.avg_dq_rate >> 3)) + (count >> 3);
	288
	289	/* If the queue has receded below the threshold, we hold
	290	* on to the last drain rate calculated, else we reset
	291	* dq_count to 0 to re-enter the if block when the next
	292	* packet is dequeued
	293	*/
	294	if (qlen < QUEUE_THRESHOLD)
	295	q->vars.dq_count = DQCOUNT_INVALID;
	296	else {
	297	q->vars.dq_count = 0;
	298	q->vars.dq_tstamp = psched_get_time();
	299	}
	300
	301	if (q->vars.burst_time > 0) {
	302	if (q->vars.burst_time > dtime)
	303	q->vars.burst_time -= dtime;
304	else
305	q->vars.burst_time = 0;
306	}
307	}
308	}
309	}
310
311	static void calculate_probability(struct Qdisc *sch)
312	{
313	struct pie_sched_data *q = qdisc_priv(sch);
314	u32 qlen = sch->qstats.backlog; /* queue size in bytes */
315	psched_time_t qdelay = 0; /* in pschedtime */
316	psched_time_t qdelay_old = q->vars.qdelay; /* in pschedtime */
317	s32 delta = 0; /* determines the change in probability */
318	u32 oldprob;
319	u32 alpha, beta;
320	bool update_prob = true;
321
322	q->vars.qdelay_old = q->vars.qdelay;
323
324	if (q->vars.avg_dq_rate > 0)
325	qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
326	else
327	qdelay = 0;
328
329	/* If qdelay is zero and qlen is not, it means qlen is very small, less
330	* than dequeue_rate, so we do not update probabilty in this round
331	*/
332	if (qdelay == 0 && qlen != 0)
333	update_prob = false;
334
219e288e VS	335	/* In the algorithm, alpha and beta are between 0 and 2 with typical
	336	* value for alpha as 0.125. In this implementation, we use values 0-32
	337	* passed from user space to represent this. Also, alpha and beta have
	338	* unit of HZ and need to be scaled before they can used to update
	339	* probability. alpha/beta are updated locally below by 1) scaling them
	340	* appropriately 2) scaling down by 16 to come to 0-2 range.
	341	* Please see paper for details.
	342	*
	343	* We scale alpha and beta differently depending on whether we are in
	344	* light, medium or high dropping mode.
d4b36210 VS	345	*/
	346	if (q->vars.prob < MAX_PROB / 100) {
	347	alpha =
	348	(q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
	349	beta =
	350	(q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
	351	} else if (q->vars.prob < MAX_PROB / 10) {
	352	alpha =
	353	(q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
	354	beta =
	355	(q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
	356	} else {
	357	alpha =
	358	(q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
	359	beta =
	360	(q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
	361	}
	362
	363	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
	364	delta += alpha * ((qdelay - q->params.target));
	365	delta += beta * ((qdelay - qdelay_old));
	366
	367	oldprob = q->vars.prob;
	368
	369	/* to ensure we increase probability in steps of no more than 2% */
	370	if (delta > (s32) (MAX_PROB / (100 / 2)) &&
	371	q->vars.prob >= MAX_PROB / 10)
	372	delta = (MAX_PROB / 100) * 2;
	373
	374	/* Non-linear drop:
	375	* Tune drop probability to increase quickly for high delays(>= 250ms)
	376	* 250ms is derived through experiments and provides error protection
	377	*/
	378
	379	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
	380	delta += MAX_PROB / (100 / 2);
	381
	382	q->vars.prob += delta;
	383
	384	if (delta > 0) {
	385	/* prevent overflow */
	386	if (q->vars.prob < oldprob) {
	387	q->vars.prob = MAX_PROB;
	388	/* Prevent normalization error. If probability is at
	389	* maximum value already, we normalize it here, and
	390	* skip the check to do a non-linear drop in the next
	391	* section.
	392	*/
	393	update_prob = false;
	394	}
	395	} else {
	396	/* prevent underflow */
	397	if (q->vars.prob > oldprob)
	398	q->vars.prob = 0;
	399	}
	400
	401	/* Non-linear drop in probability: Reduce drop probability quickly if
	402	* delay is 0 for 2 consecutive Tupdate periods.
	403	*/
	404
	405	if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
	406	q->vars.prob = (q->vars.prob * 98) / 100;
	407
	408	q->vars.qdelay = qdelay;
409	q->vars.qlen_old = qlen;
410
411	/* We restart the measurement cycle if the following conditions are met
412	* 1. If the delay has been low for 2 consecutive Tupdate periods
413	* 2. Calculated drop probability is zero
414	* 3. We have atleast one estimate for the avg_dq_rate ie.,
415	* is a non-zero value
416	*/
417	if ((q->vars.qdelay < q->params.target / 2) &&
418	(q->vars.qdelay_old < q->params.target / 2) &&
419	(q->vars.prob == 0) &&
420	(q->vars.avg_dq_rate > 0))
421	pie_vars_init(&q->vars);
422	}
423
424	static void pie_timer(unsigned long arg)
425	{
426	struct Qdisc sch = (struct Qdisc )arg;
427	struct pie_sched_data *q = qdisc_priv(sch);
428	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
429
430	spin_lock(root_lock);
431	calculate_probability(sch);
432
433	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
434	if (q->params.tupdate)
435	mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
436	spin_unlock(root_lock);
437
438	}
439
440	static int pie_init(struct Qdisc sch, struct nlattr opt)
441	{
442	struct pie_sched_data *q = qdisc_priv(sch);
443
444	pie_params_init(&q->params);
445	pie_vars_init(&q->vars);
446	sch->limit = q->params.limit;
447
448	setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);
d4b36210 VS	449
	450	if (opt) {
	451	int err = pie_change(sch, opt);
	452
	453	if (err)
	454	return err;
	455	}
	456
d5610902	457	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
d4b36210 VS	458	return 0;
	459	}
	460
	461	static int pie_dump(struct Qdisc sch, struct sk_buff skb)
	462	{
	463	struct pie_sched_data *q = qdisc_priv(sch);
	464	struct nlattr *opts;
	465
	466	opts = nla_nest_start(skb, TCA_OPTIONS);
	467	if (opts == NULL)
	468	goto nla_put_failure;
	469
	470	/* convert target from pschedtime to us */
	471	if (nla_put_u32(skb, TCA_PIE_TARGET,
	472	((u32) PSCHED_TICKS2NS(q->params.target)) /
	473	NSEC_PER_USEC) \|\|
	474	nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) \|\|
	475	nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) \|\|
	476	nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) \|\|
	477	nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) \|\|
	478	nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) \|\|
	479	nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
	480	goto nla_put_failure;
	481
	482	return nla_nest_end(skb, opts);
	483
	484	nla_put_failure:
	485	nla_nest_cancel(skb, opts);
	486	return -1;
	487
	488	}
	489
	490	static int pie_dump_stats(struct Qdisc sch, struct gnet_dump d)
	491	{
	492	struct pie_sched_data *q = qdisc_priv(sch);
	493	struct tc_pie_xstats st = {
	494	.prob = q->vars.prob,
	495	.delay = ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
	496	NSEC_PER_USEC,
	497	/* unscale and return dq_rate in bytes per sec */
	498	.avg_dq_rate = q->vars.avg_dq_rate *
	499	(PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
	500	.packets_in = q->stats.packets_in,
	501	.overlimit = q->stats.overlimit,
	502	.maxq = q->stats.maxq,
	503	.dropped = q->stats.dropped,
	504	.ecn_mark = q->stats.ecn_mark,
	505	};
	506
	507	return gnet_stats_copy_app(d, &st, sizeof(st));
	508	}
	509
	510	static struct sk_buff pie_qdisc_dequeue(struct Qdisc sch)
	511	{
	512	struct sk_buff *skb;
	513	skb = __qdisc_dequeue_head(sch, &sch->q);
	514
	515	if (!skb)
	516	return NULL;
	517
	518	pie_process_dequeue(sch, skb);
	519	return skb;
	520	}
	521
522	static void pie_reset(struct Qdisc *sch)
523	{
524	struct pie_sched_data *q = qdisc_priv(sch);
525	qdisc_reset_queue(sch);
526	pie_vars_init(&q->vars);
527	}
528
529	static void pie_destroy(struct Qdisc *sch)
530	{
531	struct pie_sched_data *q = qdisc_priv(sch);
532	q->params.tupdate = 0;
533	del_timer_sync(&q->adapt_timer);
534	}
535
536	static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
537	.id = "pie",
538	.priv_size = sizeof(struct pie_sched_data),
539	.enqueue = pie_qdisc_enqueue,
540	.dequeue = pie_qdisc_dequeue,
541	.peek = qdisc_peek_dequeued,
542	.init = pie_init,
543	.destroy = pie_destroy,
544	.reset = pie_reset,
545	.change = pie_change,
546	.dump = pie_dump,
547	.dump_stats = pie_dump_stats,
548	.owner = THIS_MODULE,
549	};
550
551	static int __init pie_module_init(void)
552	{
553	return register_qdisc(&pie_qdisc_ops);
554	}
555
556	static void __exit pie_module_exit(void)
557	{
558	unregister_qdisc(&pie_qdisc_ops);
559	}
560
561	module_init(pie_module_init);
562	module_exit(pie_module_exit);
563
564	MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
565	MODULE_AUTHOR("Vijay Subramanian");
566	MODULE_AUTHOR("Mythili Prabhu");
567	MODULE_LICENSE("GPL");