sch_fq.c

   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
   4 *
   5 *  Copyright (C) 2013-2023 Eric Dumazet <edumazet@google.com>
   6 *
   7 *  Meant to be mostly used for locally generated traffic :
   8 *  Fast classification depends on skb->sk being set before reaching us.
   9 *  If not, (router workload), we use rxhash as fallback, with 32 bits wide hash.
  10 *  All packets belonging to a socket are considered as a 'flow'.
  11 *
  12 *  Flows are dynamically allocated and stored in a hash table of RB trees
  13 *  They are also part of one Round Robin 'queues' (new or old flows)
  14 *
  15 *  Burst avoidance (aka pacing) capability :
  16 *
  17 *  Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a
  18 *  bunch of packets, and this packet scheduler adds delay between
  19 *  packets to respect rate limitation.
  20 *
  21 *  enqueue() :
  22 *   - lookup one RB tree (out of 1024 or more) to find the flow.
  23 *     If non existent flow, create it, add it to the tree.
  24 *     Add skb to the per flow list of skb (fifo).
  25 *   - Use a special fifo for high prio packets
  26 *
  27 *  dequeue() : serves flows in Round Robin
  28 *  Note : When a flow becomes empty, we do not immediately remove it from
  29 *  rb trees, for performance reasons (its expected to send additional packets,
  30 *  or SLAB cache will reuse socket for another flow)
  31 */
  32
  33#include <linux/module.h>
  34#include <linux/types.h>
  35#include <linux/kernel.h>
  36#include <linux/jiffies.h>
  37#include <linux/string.h>
  38#include <linux/in.h>
  39#include <linux/errno.h>
  40#include <linux/init.h>
  41#include <linux/skbuff.h>
  42#include <linux/slab.h>
  43#include <linux/rbtree.h>
  44#include <linux/hash.h>
  45#include <linux/prefetch.h>
  46#include <linux/vmalloc.h>
  47#include <net/netlink.h>
  48#include <net/pkt_sched.h>
  49#include <net/sock.h>
  50#include <net/tcp_states.h>
  51#include <net/tcp.h>
  52
  53struct fq_skb_cb {
  54	u64	time_to_send;
  55	u8	band;
  56};
  57
  58static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
  59{
  60	qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb));
  61	return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data;
  62}
  63
  64/*
  65 * Per flow structure, dynamically allocated.
  66 * If packets have monotically increasing time_to_send, they are placed in O(1)
  67 * in linear list (head,tail), otherwise are placed in a rbtree (t_root).
  68 */
  69struct fq_flow {
  70/* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
  71	struct rb_root	t_root;
  72	struct sk_buff	*head;		/* list of skbs for this flow : first skb */
  73	union {
  74		struct sk_buff *tail;	/* last skb in the list */
  75		unsigned long  age;	/* (jiffies | 1UL) when flow was emptied, for gc */
  76	};
  77	union {
  78		struct rb_node	fq_node;	/* anchor in fq_root[] trees */
  79		/* Following field is only used for q->internal,
  80		 * because q->internal is not hashed in fq_root[]
  81		 */
  82		u64		stat_fastpath_packets;
  83	};
  84	struct sock	*sk;
  85	u32		socket_hash;	/* sk_hash */
  86	int		qlen;		/* number of packets in flow queue */
  87
  88/* Second cache line */
  89	int		credit;
  90	int		band;
  91	struct fq_flow *next;		/* next pointer in RR lists */
  92
  93	struct rb_node  rate_node;	/* anchor in q->delayed tree */
  94	u64		time_next_packet;
  95};
  96
  97struct fq_flow_head {
  98	struct fq_flow *first;
  99	struct fq_flow *last;
 100};
 101
 102struct fq_perband_flows {
 103	struct fq_flow_head new_flows;
 104	struct fq_flow_head old_flows;
 105	int		    credit;
 106	int		    quantum; /* based on band nr : 576KB, 192KB, 64KB */
 107};
 108
 109#define FQ_PRIO2BAND_CRUMB_SIZE ((TC_PRIO_MAX + 1) >> 2)
 110
 111struct fq_sched_data {
 112/* Read mostly cache line */
 113
 114	u64		offload_horizon;
 115	u32		quantum;
 116	u32		initial_quantum;
 117	u32		flow_refill_delay;
 118	u32		flow_plimit;	/* max packets per flow */
 119	unsigned long	flow_max_rate;	/* optional max rate per flow */
 120	u64		ce_threshold;
 121	u64		horizon;	/* horizon in ns */
 122	u32		orphan_mask;	/* mask for orphaned skb */
 123	u32		low_rate_threshold;
 124	struct rb_root	*fq_root;
 125	u8		rate_enable;
 126	u8		fq_trees_log;
 127	u8		horizon_drop;
 128	u8		prio2band[FQ_PRIO2BAND_CRUMB_SIZE];
 129	u32		timer_slack; /* hrtimer slack in ns */
 130
 131/* Read/Write fields. */
 132
 133	unsigned int band_nr; /* band being serviced in fq_dequeue() */
 134
 135	struct fq_perband_flows band_flows[FQ_BANDS];
 136
 137	struct fq_flow	internal;	/* fastpath queue. */
 138	struct rb_root	delayed;	/* for rate limited flows */
 139	u64		time_next_delayed_flow;
 140	unsigned long	unthrottle_latency_ns;
 141
 142	u32		band_pkt_count[FQ_BANDS];
 143	u32		flows;
 144	u32		inactive_flows; /* Flows with no packet to send. */
 145	u32		throttled_flows;
 146
 147	u64		stat_throttled;
 148	struct qdisc_watchdog watchdog;
 149	u64		stat_gc_flows;
 150
 151/* Seldom used fields. */
 152
 153	u64		stat_band_drops[FQ_BANDS];
 154	u64		stat_ce_mark;
 155	u64		stat_horizon_drops;
 156	u64		stat_horizon_caps;
 157	u64		stat_flows_plimit;
 158	u64		stat_pkts_too_long;
 159	u64		stat_allocation_errors;
 160};
 161
 162/* return the i-th 2-bit value ("crumb") */
 163static u8 fq_prio2band(const u8 *prio2band, unsigned int prio)
 164{
 165	return (READ_ONCE(prio2band[prio / 4]) >> (2 * (prio & 0x3))) & 0x3;
 166}
 167
 168/*
 169 * f->tail and f->age share the same location.
 170 * We can use the low order bit to differentiate if this location points
 171 * to a sk_buff or contains a jiffies value, if we force this value to be odd.
 172 * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
 173 */
 174static void fq_flow_set_detached(struct fq_flow *f)
 175{
 176	f->age = jiffies | 1UL;
 177}
 178
 179static bool fq_flow_is_detached(const struct fq_flow *f)
 180{
 181	return !!(f->age & 1UL);
 182}
 183
 184/* special value to mark a throttled flow (not on old/new list) */
 185static struct fq_flow throttled;
 186
 187static bool fq_flow_is_throttled(const struct fq_flow *f)
 188{
 189	return f->next == &throttled;
 190}
 191
 192enum new_flow {
 193	NEW_FLOW,
 194	OLD_FLOW
 195};
 196
 197static void fq_flow_add_tail(struct fq_sched_data *q, struct fq_flow *flow,
 198			     enum new_flow list_sel)
 199{
 200	struct fq_perband_flows *pband = &q->band_flows[flow->band];
 201	struct fq_flow_head *head = (list_sel == NEW_FLOW) ?
 202					&pband->new_flows :
 203					&pband->old_flows;
 204
 205	if (head->first)
 206		head->last->next = flow;
 207	else
 208		head->first = flow;
 209	head->last = flow;
 210	flow->next = NULL;
 211}
 212
 213static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 214{
 215	rb_erase(&f->rate_node, &q->delayed);
 216	q->throttled_flows--;
 217	fq_flow_add_tail(q, f, OLD_FLOW);
 218}
 219
 220static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 221{
 222	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 223
 224	while (*p) {
 225		struct fq_flow *aux;
 226
 227		parent = *p;
 228		aux = rb_entry(parent, struct fq_flow, rate_node);
 229		if (f->time_next_packet >= aux->time_next_packet)
 230			p = &parent->rb_right;
 231		else
 232			p = &parent->rb_left;
 233	}
 234	rb_link_node(&f->rate_node, parent, p);
 235	rb_insert_color(&f->rate_node, &q->delayed);
 236	q->throttled_flows++;
 237	q->stat_throttled++;
 238
 239	f->next = &throttled;
 240	if (q->time_next_delayed_flow > f->time_next_packet)
 241		q->time_next_delayed_flow = f->time_next_packet;
 242}
 243
 244
 245static struct kmem_cache *fq_flow_cachep __read_mostly;
 246
 247
 248/* limit number of collected flows per round */
 249#define FQ_GC_MAX 8
 250#define FQ_GC_AGE (3*HZ)
 251
 252static bool fq_gc_candidate(const struct fq_flow *f)
 253{
 254	return fq_flow_is_detached(f) &&
 255	       time_after(jiffies, f->age + FQ_GC_AGE);
 256}
 257
 258static void fq_gc(struct fq_sched_data *q,
 259		  struct rb_root *root,
 260		  struct sock *sk)
 261{
 262	struct rb_node **p, *parent;
 263	void *tofree[FQ_GC_MAX];
 264	struct fq_flow *f;
 265	int i, fcnt = 0;
 266
 267	p = &root->rb_node;
 268	parent = NULL;
 269	while (*p) {
 270		parent = *p;
 271
 272		f = rb_entry(parent, struct fq_flow, fq_node);
 273		if (f->sk == sk)
 274			break;
 275
 276		if (fq_gc_candidate(f)) {
 277			tofree[fcnt++] = f;
 278			if (fcnt == FQ_GC_MAX)
 279				break;
 280		}
 281
 282		if (f->sk > sk)
 283			p = &parent->rb_right;
 284		else
 285			p = &parent->rb_left;
 286	}
 287
 288	if (!fcnt)
 289		return;
 290
 291	for (i = fcnt; i > 0; ) {
 292		f = tofree[--i];
 293		rb_erase(&f->fq_node, root);
 294	}
 295	q->flows -= fcnt;
 296	q->inactive_flows -= fcnt;
 297	q->stat_gc_flows += fcnt;
 298
 299	kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
 300}
 301
 302/* Fast path can be used if :
 303 * 1) Packet tstamp is in the past, or within the pacing offload horizon.
 304 * 2) FQ qlen == 0   OR
 305 *   (no flow is currently eligible for transmit,
 306 *    AND fast path queue has less than 8 packets)
 307 * 3) No SO_MAX_PACING_RATE on the socket (if any).
 308 * 4) No @maxrate attribute on this qdisc,
 309 *
 310 * FQ can not use generic TCQ_F_CAN_BYPASS infrastructure.
 311 */
 312static bool fq_fastpath_check(const struct Qdisc *sch, struct sk_buff *skb,
 313			      u64 now)
 314{
 315	const struct fq_sched_data *q = qdisc_priv(sch);
 316	const struct sock *sk;
 317
 318	if (fq_skb_cb(skb)->time_to_send > now + q->offload_horizon)
 319		return false;
 320
 321	if (sch->q.qlen != 0) {
 322		/* Even if some packets are stored in this qdisc,
 323		 * we can still enable fast path if all of them are
 324		 * scheduled in the future (ie no flows are eligible)
 325		 * or in the fast path queue.
 326		 */
 327		if (q->flows != q->inactive_flows + q->throttled_flows)
 328			return false;
 329
 330		/* Do not allow fast path queue to explode, we want Fair Queue mode
 331		 * under pressure.
 332		 */
 333		if (q->internal.qlen >= 8)
 334			return false;
 335
 336		/* Ordering invariants fall apart if some delayed flows
 337		 * are ready but we haven't serviced them, yet.
 338		 */
 339		if (q->time_next_delayed_flow <= now + q->offload_horizon)
 340			return false;
 341	}
 342
 343	sk = skb->sk;
 344	if (sk && sk_fullsock(sk) && !sk_is_tcp(sk) &&
 345	    sk->sk_max_pacing_rate != ~0UL)
 346		return false;
 347
 348	if (q->flow_max_rate != ~0UL)
 349		return false;
 350
 351	return true;
 352}
 353
 354static struct fq_flow *fq_classify(struct Qdisc *sch, struct sk_buff *skb,
 355				   u64 now)
 356{
 357	struct fq_sched_data *q = qdisc_priv(sch);
 358	struct rb_node **p, *parent;
 359	struct sock *sk = skb->sk;
 360	struct rb_root *root;
 361	struct fq_flow *f;
 362
 363	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 364	 * or a listener (SYNCOOKIE mode)
 365	 * 1) request sockets are not full blown,
 366	 *    they do not contain sk_pacing_rate
 367	 * 2) They are not part of a 'flow' yet
 368	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 369	 *    especially if the listener set SO_MAX_PACING_RATE
 370	 * 4) We pretend they are orphaned
 371	 * TCP can also associate TIME_WAIT sockets with RST or ACK packets.
 372	 */
 373	if (!sk || sk_listener_or_tw(sk)) {
 374		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 375
 376		/* By forcing low order bit to 1, we make sure to not
 377		 * collide with a local flow (socket pointers are word aligned)
 378		 */
 379		sk = (struct sock *)((hash << 1) | 1UL);
 380		skb_orphan(skb);
 381	} else if (sk->sk_state == TCP_CLOSE) {
 382		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 383		/*
 384		 * Sockets in TCP_CLOSE are non connected.
 385		 * Typical use case is UDP sockets, they can send packets
 386		 * with sendto() to many different destinations.
 387		 * We probably could use a generic bit advertising
 388		 * non connected sockets, instead of sk_state == TCP_CLOSE,
 389		 * if we care enough.
 390		 */
 391		sk = (struct sock *)((hash << 1) | 1UL);
 392	}
 393
 394	if (fq_fastpath_check(sch, skb, now)) {
 395		q->internal.stat_fastpath_packets++;
 396		if (skb->sk == sk && q->rate_enable &&
 397		    READ_ONCE(sk->sk_pacing_status) != SK_PACING_FQ)
 398			smp_store_release(&sk->sk_pacing_status,
 399					  SK_PACING_FQ);
 400		return &q->internal;
 401	}
 402
 403	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 404
 405	fq_gc(q, root, sk);
 406
 407	p = &root->rb_node;
 408	parent = NULL;
 409	while (*p) {
 410		parent = *p;
 411
 412		f = rb_entry(parent, struct fq_flow, fq_node);
 413		if (f->sk == sk) {
 414			/* socket might have been reallocated, so check
 415			 * if its sk_hash is the same.
 416			 * It not, we need to refill credit with
 417			 * initial quantum
 418			 */
 419			if (unlikely(skb->sk == sk &&
 420				     f->socket_hash != sk->sk_hash)) {
 421				f->credit = q->initial_quantum;
 422				f->socket_hash = sk->sk_hash;
 423				if (q->rate_enable)
 424					smp_store_release(&sk->sk_pacing_status,
 425							  SK_PACING_FQ);
 426				if (fq_flow_is_throttled(f))
 427					fq_flow_unset_throttled(q, f);
 428				f->time_next_packet = 0ULL;
 429			}
 430			return f;
 431		}
 432		if (f->sk > sk)
 433			p = &parent->rb_right;
 434		else
 435			p = &parent->rb_left;
 436	}
 437
 438	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 439	if (unlikely(!f)) {
 440		q->stat_allocation_errors++;
 441		return &q->internal;
 442	}
 443	/* f->t_root is already zeroed after kmem_cache_zalloc() */
 444
 445	fq_flow_set_detached(f);
 446	f->sk = sk;
 447	if (skb->sk == sk) {
 448		f->socket_hash = sk->sk_hash;
 449		if (q->rate_enable)
 450			smp_store_release(&sk->sk_pacing_status,
 451					  SK_PACING_FQ);
 452	}
 453	f->credit = q->initial_quantum;
 454
 455	rb_link_node(&f->fq_node, parent, p);
 456	rb_insert_color(&f->fq_node, root);
 457
 458	q->flows++;
 459	q->inactive_flows++;
 460	return f;
 461}
 462
 463static struct sk_buff *fq_peek(struct fq_flow *flow)
 464{
 465	struct sk_buff *skb = skb_rb_first(&flow->t_root);
 466	struct sk_buff *head = flow->head;
 467
 468	if (!skb)
 469		return head;
 470
 471	if (!head)
 472		return skb;
 473
 474	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
 475		return skb;
 476	return head;
 477}
 478
 479static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
 480			  struct sk_buff *skb)
 481{
 482	if (skb == flow->head) {
 483		struct sk_buff *next = skb->next;
 484
 485		prefetch(next);
 486		flow->head = next;
 487	} else {
 488		rb_erase(&skb->rbnode, &flow->t_root);
 489		skb->dev = qdisc_dev(sch);
 490	}
 491}
 492
 493/* Remove one skb from flow queue.
 494 * This skb must be the return value of prior fq_peek().
 495 */
 496static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
 497			   struct sk_buff *skb)
 498{
 499	fq_erase_head(sch, flow, skb);
 500	skb_mark_not_on_list(skb);
 501	qdisc_qstats_backlog_dec(sch, skb);
 502	sch->q.qlen--;
 503	qdisc_bstats_update(sch, skb);
 504}
 505
 506static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 507{
 508	struct rb_node **p, *parent;
 509	struct sk_buff *head, *aux;
 510
 511	head = flow->head;
 512	if (!head ||
 513	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 514		if (!head)
 515			flow->head = skb;
 516		else
 517			flow->tail->next = skb;
 518		flow->tail = skb;
 519		skb->next = NULL;
 520		return;
 521	}
 522
 523	p = &flow->t_root.rb_node;
 524	parent = NULL;
 525
 526	while (*p) {
 527		parent = *p;
 528		aux = rb_to_skb(parent);
 529		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 530			p = &parent->rb_right;
 531		else
 532			p = &parent->rb_left;
 533	}
 534	rb_link_node(&skb->rbnode, parent, p);
 535	rb_insert_color(&skb->rbnode, &flow->t_root);
 536}
 537
 538static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
 539				     const struct fq_sched_data *q, u64 now)
 540{
 541	return unlikely((s64)skb->tstamp > (s64)(now + q->horizon));
 542}
 543
 544#define FQDR(reason) SKB_DROP_REASON_FQ_##reason
 545
 546static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 547		      struct sk_buff **to_free)
 548{
 549	struct fq_sched_data *q = qdisc_priv(sch);
 550	struct fq_flow *f;
 551	u64 now;
 552	u8 band;
 553
 554	band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
 555	if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
 556		q->stat_band_drops[band]++;
 557		return qdisc_drop_reason(skb, sch, to_free,
 558					 FQDR(BAND_LIMIT));
 559	}
 560
 561	now = ktime_get_ns();
 562	if (!skb->tstamp) {
 563		fq_skb_cb(skb)->time_to_send = now;
 564	} else {
 565		/* Check if packet timestamp is too far in the future. */
 566		if (fq_packet_beyond_horizon(skb, q, now)) {
 567			if (q->horizon_drop) {
 568				q->stat_horizon_drops++;
 569				return qdisc_drop_reason(skb, sch, to_free,
 570							 FQDR(HORIZON_LIMIT));
 571			}
 572			q->stat_horizon_caps++;
 573			skb->tstamp = now + q->horizon;
 574		}
 575		fq_skb_cb(skb)->time_to_send = skb->tstamp;
 576	}
 577
 578	f = fq_classify(sch, skb, now);
 579
 580	if (f != &q->internal) {
 581		if (unlikely(f->qlen >= q->flow_plimit)) {
 582			q->stat_flows_plimit++;
 583			return qdisc_drop_reason(skb, sch, to_free,
 584						 FQDR(FLOW_LIMIT));
 585		}
 586
 587		if (fq_flow_is_detached(f)) {
 588			fq_flow_add_tail(q, f, NEW_FLOW);
 589			if (time_after(jiffies, f->age + q->flow_refill_delay))
 590				f->credit = max_t(u32, f->credit, q->quantum);
 591		}
 592
 593		f->band = band;
 594		q->band_pkt_count[band]++;
 595		fq_skb_cb(skb)->band = band;
 596		if (f->qlen == 0)
 597			q->inactive_flows--;
 598	}
 599
 600	f->qlen++;
 601	/* Note: this overwrites f->age */
 602	flow_queue_add(f, skb);
 603
 604	qdisc_qstats_backlog_inc(sch, skb);
 605	sch->q.qlen++;
 606
 607	return NET_XMIT_SUCCESS;
 608}
 609#undef FQDR
 610
 611static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 612{
 613	unsigned long sample;
 614	struct rb_node *p;
 615
 616	if (q->time_next_delayed_flow > now + q->offload_horizon)
 617		return;
 618
 619	/* Update unthrottle latency EWMA.
 620	 * This is cheap and can help diagnosing timer/latency problems.
 621	 */
 622	sample = (unsigned long)(now - q->time_next_delayed_flow);
 623	if ((long)sample > 0) {
 624		q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 625		q->unthrottle_latency_ns += sample >> 3;
 626	}
 627	now += q->offload_horizon;
 628
 629	q->time_next_delayed_flow = ~0ULL;
 630	while ((p = rb_first(&q->delayed)) != NULL) {
 631		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 632
 633		if (f->time_next_packet > now) {
 634			q->time_next_delayed_flow = f->time_next_packet;
 635			break;
 636		}
 637		fq_flow_unset_throttled(q, f);
 638	}
 639}
 640
 641static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
 642{
 643	if (pband->credit <= 0)
 644		return NULL;
 645
 646	if (pband->new_flows.first)
 647		return &pband->new_flows;
 648
 649	return pband->old_flows.first ? &pband->old_flows : NULL;
 650}
 651
 652static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 653{
 654	struct fq_sched_data *q = qdisc_priv(sch);
 655	struct fq_perband_flows *pband;
 656	struct fq_flow_head *head;
 657	struct sk_buff *skb;
 658	struct fq_flow *f;
 659	unsigned long rate;
 660	int retry;
 661	u32 plen;
 662	u64 now;
 663
 664	if (!sch->q.qlen)
 665		return NULL;
 666
 667	skb = fq_peek(&q->internal);
 668	if (unlikely(skb)) {
 669		q->internal.qlen--;
 670		fq_dequeue_skb(sch, &q->internal, skb);
 671		goto out;
 672	}
 673
 674	now = ktime_get_ns();
 675	fq_check_throttled(q, now);
 676	retry = 0;
 677	pband = &q->band_flows[q->band_nr];
 678begin:
 679	head = fq_pband_head_select(pband);
 680	if (!head) {
 681		while (++retry <= FQ_BANDS) {
 682			if (++q->band_nr == FQ_BANDS)
 683				q->band_nr = 0;
 684			pband = &q->band_flows[q->band_nr];
 685			pband->credit = min(pband->credit + pband->quantum,
 686					    pband->quantum);
 687			if (pband->credit > 0)
 688				goto begin;
 689			retry = 0;
 690		}
 691		if (q->time_next_delayed_flow != ~0ULL)
 692			qdisc_watchdog_schedule_range_ns(&q->watchdog,
 693							q->time_next_delayed_flow,
 694							q->timer_slack);
 695		return NULL;
 696	}
 697	f = head->first;
 698	retry = 0;
 699	if (f->credit <= 0) {
 700		f->credit += q->quantum;
 701		head->first = f->next;
 702		fq_flow_add_tail(q, f, OLD_FLOW);
 703		goto begin;
 704	}
 705
 706	skb = fq_peek(f);
 707	if (skb) {
 708		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 709					     f->time_next_packet);
 710
 711		if (now + q->offload_horizon < time_next_packet) {
 712			head->first = f->next;
 713			f->time_next_packet = time_next_packet;
 714			fq_flow_set_throttled(q, f);
 715			goto begin;
 716		}
 717		prefetch(&skb->end);
 718		fq_dequeue_skb(sch, f, skb);
 719		if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
 720			INET_ECN_set_ce(skb);
 721			q->stat_ce_mark++;
 722		}
 723		if (--f->qlen == 0)
 724			q->inactive_flows++;
 725		q->band_pkt_count[fq_skb_cb(skb)->band]--;
 726	} else {
 727		head->first = f->next;
 728		/* force a pass through old_flows to prevent starvation */
 729		if (head == &pband->new_flows) {
 730			fq_flow_add_tail(q, f, OLD_FLOW);
 731		} else {
 732			fq_flow_set_detached(f);
 733		}
 734		goto begin;
 735	}
 736	plen = qdisc_pkt_len(skb);
 737	f->credit -= plen;
 738	pband->credit -= plen;
 739
 740	if (!q->rate_enable)
 741		goto out;
 742
 743	rate = q->flow_max_rate;
 744
 745	/* If EDT time was provided for this skb, we need to
 746	 * update f->time_next_packet only if this qdisc enforces
 747	 * a flow max rate.
 748	 */
 749	if (!skb->tstamp) {
 750		if (skb->sk)
 751			rate = min(READ_ONCE(skb->sk->sk_pacing_rate), rate);
 752
 753		if (rate <= q->low_rate_threshold) {
 754			f->credit = 0;
 755		} else {
 756			plen = max(plen, q->quantum);
 757			if (f->credit > 0)
 758				goto out;
 759		}
 760	}
 761	if (rate != ~0UL) {
 762		u64 len = (u64)plen * NSEC_PER_SEC;
 763
 764		if (likely(rate))
 765			len = div64_ul(len, rate);
 766		/* Since socket rate can change later,
 767		 * clamp the delay to 1 second.
 768		 * Really, providers of too big packets should be fixed !
 769		 */
 770		if (unlikely(len > NSEC_PER_SEC)) {
 771			len = NSEC_PER_SEC;
 772			q->stat_pkts_too_long++;
 773		}
 774		/* Account for schedule/timers drifts.
 775		 * f->time_next_packet was set when prior packet was sent,
 776		 * and current time (@now) can be too late by tens of us.
 777		 */
 778		if (f->time_next_packet)
 779			len -= min(len/2, now - f->time_next_packet);
 780		f->time_next_packet = now + len;
 781	}
 782out:
 783	return skb;
 784}
 785
 786static void fq_flow_purge(struct fq_flow *flow)
 787{
 788	struct rb_node *p = rb_first(&flow->t_root);
 789
 790	while (p) {
 791		struct sk_buff *skb = rb_to_skb(p);
 792
 793		p = rb_next(p);
 794		rb_erase(&skb->rbnode, &flow->t_root);
 795		rtnl_kfree_skbs(skb, skb);
 796	}
 797	rtnl_kfree_skbs(flow->head, flow->tail);
 798	flow->head = NULL;
 799	flow->qlen = 0;
 800}
 801
 802static void fq_reset(struct Qdisc *sch)
 803{
 804	struct fq_sched_data *q = qdisc_priv(sch);
 805	struct rb_root *root;
 806	struct rb_node *p;
 807	struct fq_flow *f;
 808	unsigned int idx;
 809
 810	sch->q.qlen = 0;
 811	sch->qstats.backlog = 0;
 812
 813	fq_flow_purge(&q->internal);
 814
 815	if (!q->fq_root)
 816		return;
 817
 818	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 819		root = &q->fq_root[idx];
 820		while ((p = rb_first(root)) != NULL) {
 821			f = rb_entry(p, struct fq_flow, fq_node);
 822			rb_erase(p, root);
 823
 824			fq_flow_purge(f);
 825
 826			kmem_cache_free(fq_flow_cachep, f);
 827		}
 828	}
 829	for (idx = 0; idx < FQ_BANDS; idx++) {
 830		q->band_flows[idx].new_flows.first = NULL;
 831		q->band_flows[idx].old_flows.first = NULL;
 832		q->band_pkt_count[idx] = 0;
 833	}
 834	q->delayed		= RB_ROOT;
 835	q->flows		= 0;
 836	q->inactive_flows	= 0;
 837	q->throttled_flows	= 0;
 838}
 839
 840static void fq_rehash(struct fq_sched_data *q,
 841		      struct rb_root *old_array, u32 old_log,
 842		      struct rb_root *new_array, u32 new_log)
 843{
 844	struct rb_node *op, **np, *parent;
 845	struct rb_root *oroot, *nroot;
 846	struct fq_flow *of, *nf;
 847	int fcnt = 0;
 848	u32 idx;
 849
 850	for (idx = 0; idx < (1U << old_log); idx++) {
 851		oroot = &old_array[idx];
 852		while ((op = rb_first(oroot)) != NULL) {
 853			rb_erase(op, oroot);
 854			of = rb_entry(op, struct fq_flow, fq_node);
 855			if (fq_gc_candidate(of)) {
 856				fcnt++;
 857				kmem_cache_free(fq_flow_cachep, of);
 858				continue;
 859			}
 860			nroot = &new_array[hash_ptr(of->sk, new_log)];
 861
 862			np = &nroot->rb_node;
 863			parent = NULL;
 864			while (*np) {
 865				parent = *np;
 866
 867				nf = rb_entry(parent, struct fq_flow, fq_node);
 868				BUG_ON(nf->sk == of->sk);
 869
 870				if (nf->sk > of->sk)
 871					np = &parent->rb_right;
 872				else
 873					np = &parent->rb_left;
 874			}
 875
 876			rb_link_node(&of->fq_node, parent, np);
 877			rb_insert_color(&of->fq_node, nroot);
 878		}
 879	}
 880	q->flows -= fcnt;
 881	q->inactive_flows -= fcnt;
 882	q->stat_gc_flows += fcnt;
 883}
 884
 885static void fq_free(void *addr)
 886{
 887	kvfree(addr);
 888}
 889
 890static int fq_resize(struct Qdisc *sch, u32 log)
 891{
 892	struct fq_sched_data *q = qdisc_priv(sch);
 893	struct rb_root *array;
 894	void *old_fq_root;
 895	u32 idx;
 896
 897	if (q->fq_root && log == q->fq_trees_log)
 898		return 0;
 899
 900	/* If XPS was setup, we can allocate memory on right NUMA node */
 901	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 902			      netdev_queue_numa_node_read(sch->dev_queue));
 903	if (!array)
 904		return -ENOMEM;
 905
 906	for (idx = 0; idx < (1U << log); idx++)
 907		array[idx] = RB_ROOT;
 908
 909	sch_tree_lock(sch);
 910
 911	old_fq_root = q->fq_root;
 912	if (old_fq_root)
 913		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 914
 915	q->fq_root = array;
 916	WRITE_ONCE(q->fq_trees_log, log);
 917
 918	sch_tree_unlock(sch);
 919
 920	fq_free(old_fq_root);
 921
 922	return 0;
 923}
 924
 925static const struct netlink_range_validation iq_range = {
 926	.max = INT_MAX,
 927};
 928
 929static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 930	[TCA_FQ_UNSPEC]			= { .strict_start_type = TCA_FQ_TIMER_SLACK },
 931
 932	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
 933	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
 934	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
 935	[TCA_FQ_INITIAL_QUANTUM]	= NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
 936	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
 937	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
 938	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
 939	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
 940	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
 941	[TCA_FQ_ORPHAN_MASK]		= { .type = NLA_U32 },
 942	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
 943	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
 944	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
 945	[TCA_FQ_HORIZON]		= { .type = NLA_U32 },
 946	[TCA_FQ_HORIZON_DROP]		= { .type = NLA_U8 },
 947	[TCA_FQ_PRIOMAP]		= NLA_POLICY_EXACT_LEN(sizeof(struct tc_prio_qopt)),
 948	[TCA_FQ_WEIGHTS]		= NLA_POLICY_EXACT_LEN(FQ_BANDS * sizeof(s32)),
 949	[TCA_FQ_OFFLOAD_HORIZON]	= { .type = NLA_U32 },
 950};
 951
 952/* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
 953static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
 954{
 955	const int num_elems = TC_PRIO_MAX + 1;
 956	u8 tmp[FQ_PRIO2BAND_CRUMB_SIZE];
 957	int i;
 958
 959	memset(tmp, 0, sizeof(tmp));
 960	for (i = 0; i < num_elems; i++)
 961		tmp[i / 4] |= in[i] << (2 * (i & 0x3));
 962
 963	for (i = 0; i < FQ_PRIO2BAND_CRUMB_SIZE; i++)
 964		WRITE_ONCE(out[i], tmp[i]);
 965}
 966
 967static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
 968{
 969	const int num_elems = TC_PRIO_MAX + 1;
 970	int i;
 971
 972	for (i = 0; i < num_elems; i++)
 973		out[i] = fq_prio2band(in, i);
 974}
 975
 976static int fq_load_weights(struct fq_sched_data *q,
 977			   const struct nlattr *attr,
 978			   struct netlink_ext_ack *extack)
 979{
 980	s32 *weights = nla_data(attr);
 981	int i;
 982
 983	for (i = 0; i < FQ_BANDS; i++) {
 984		if (weights[i] < FQ_MIN_WEIGHT) {
 985			NL_SET_ERR_MSG_FMT_MOD(extack, "Weight %d less that minimum allowed %d",
 986					       weights[i], FQ_MIN_WEIGHT);
 987			return -EINVAL;
 988		}
 989	}
 990	for (i = 0; i < FQ_BANDS; i++)
 991		WRITE_ONCE(q->band_flows[i].quantum, weights[i]);
 992	return 0;
 993}
 994
 995static int fq_load_priomap(struct fq_sched_data *q,
 996			   const struct nlattr *attr,
 997			   struct netlink_ext_ack *extack)
 998{
 999	const struct tc_prio_qopt *map = nla_data(attr);
1000	int i;
1001
1002	if (map->bands != FQ_BANDS) {
1003		NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
1004		return -EINVAL;
1005	}
1006	for (i = 0; i < TC_PRIO_MAX + 1; i++) {
1007		if (map->priomap[i] >= FQ_BANDS) {
1008			NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
1009					       i, map->priomap[i]);
1010			return -EINVAL;
1011		}
1012	}
1013	fq_prio2band_compress_crumb(map->priomap, q->prio2band);
1014	return 0;
1015}
1016
1017static int fq_change(struct Qdisc *sch, struct nlattr *opt,
1018		     struct netlink_ext_ack *extack)
1019{
1020	unsigned int dropped_pkts = 0, dropped_bytes = 0;
1021	struct fq_sched_data *q = qdisc_priv(sch);
1022	struct nlattr *tb[TCA_FQ_MAX + 1];
1023	u32 fq_log;
1024	int err;
1025
1026	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
1027					  NULL);
1028	if (err < 0)
1029		return err;
1030
1031	sch_tree_lock(sch);
1032
1033	fq_log = q->fq_trees_log;
1034
1035	if (tb[TCA_FQ_BUCKETS_LOG]) {
1036		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
1037
1038		if (nval >= 1 && nval <= ilog2(256*1024))
1039			fq_log = nval;
1040		else
1041			err = -EINVAL;
1042	}
1043	if (tb[TCA_FQ_PLIMIT])
1044		WRITE_ONCE(sch->limit,
1045			   nla_get_u32(tb[TCA_FQ_PLIMIT]));
1046
1047	if (tb[TCA_FQ_FLOW_PLIMIT])
1048		WRITE_ONCE(q->flow_plimit,
1049			   nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]));
1050
1051	if (tb[TCA_FQ_QUANTUM]) {
1052		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
1053
1054		if (quantum > 0 && quantum <= (1 << 20)) {
1055			WRITE_ONCE(q->quantum, quantum);
1056		} else {
1057			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
1058			err = -EINVAL;
1059		}
1060	}
1061
1062	if (tb[TCA_FQ_INITIAL_QUANTUM])
1063		WRITE_ONCE(q->initial_quantum,
1064			   nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]));
1065
1066	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
1067		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
1068				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
1069
1070	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
1071		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
1072
1073		WRITE_ONCE(q->flow_max_rate,
1074			   (rate == ~0U) ? ~0UL : rate);
1075	}
1076	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
1077		WRITE_ONCE(q->low_rate_threshold,
1078			   nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]));
1079
1080	if (tb[TCA_FQ_RATE_ENABLE]) {
1081		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
1082
1083		if (enable <= 1)
1084			WRITE_ONCE(q->rate_enable,
1085				   enable);
1086		else
1087			err = -EINVAL;
1088	}
1089
1090	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
1091		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
1092
1093		WRITE_ONCE(q->flow_refill_delay,
1094			   usecs_to_jiffies(usecs_delay));
1095	}
1096
1097	if (!err && tb[TCA_FQ_PRIOMAP])
1098		err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
1099
1100	if (!err && tb[TCA_FQ_WEIGHTS])
1101		err = fq_load_weights(q, tb[TCA_FQ_WEIGHTS], extack);
1102
1103	if (tb[TCA_FQ_ORPHAN_MASK])
1104		WRITE_ONCE(q->orphan_mask,
1105			   nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]));
1106
1107	if (tb[TCA_FQ_CE_THRESHOLD])
1108		WRITE_ONCE(q->ce_threshold,
1109			   (u64)NSEC_PER_USEC *
1110			   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]));
1111
1112	if (tb[TCA_FQ_TIMER_SLACK])
1113		WRITE_ONCE(q->timer_slack,
1114			   nla_get_u32(tb[TCA_FQ_TIMER_SLACK]));
1115
1116	if (tb[TCA_FQ_HORIZON])
1117		WRITE_ONCE(q->horizon,
1118			   (u64)NSEC_PER_USEC *
1119			   nla_get_u32(tb[TCA_FQ_HORIZON]));
1120
1121	if (tb[TCA_FQ_HORIZON_DROP])
1122		WRITE_ONCE(q->horizon_drop,
1123			   nla_get_u8(tb[TCA_FQ_HORIZON_DROP]));
1124
1125	if (tb[TCA_FQ_OFFLOAD_HORIZON]) {
1126		u64 offload_horizon = (u64)NSEC_PER_USEC *
1127				      nla_get_u32(tb[TCA_FQ_OFFLOAD_HORIZON]);
1128
1129		if (offload_horizon <= qdisc_dev(sch)->max_pacing_offload_horizon) {
1130			WRITE_ONCE(q->offload_horizon, offload_horizon);
1131		} else {
1132			NL_SET_ERR_MSG_MOD(extack, "invalid offload_horizon");
1133			err = -EINVAL;
1134		}
1135	}
1136	if (!err) {
1137
1138		sch_tree_unlock(sch);
1139		err = fq_resize(sch, fq_log);
1140		sch_tree_lock(sch);
1141	}
1142
1143	while (sch->q.qlen > sch->limit) {
1144		struct sk_buff *skb = qdisc_dequeue_internal(sch, false);
1145
1146		if (!skb)
1147			break;
1148
1149		dropped_pkts++;
1150		dropped_bytes += qdisc_pkt_len(skb);
1151		rtnl_kfree_skbs(skb, skb);
1152	}
1153	qdisc_tree_reduce_backlog(sch, dropped_pkts, dropped_bytes);
1154
1155	sch_tree_unlock(sch);
1156	return err;
1157}
1158
1159static void fq_destroy(struct Qdisc *sch)
1160{
1161	struct fq_sched_data *q = qdisc_priv(sch);
1162
1163	fq_reset(sch);
1164	fq_free(q->fq_root);
1165	qdisc_watchdog_cancel(&q->watchdog);
1166}
1167
1168static int fq_init(struct Qdisc *sch, struct nlattr *opt,
1169		   struct netlink_ext_ack *extack)
1170{
1171	struct fq_sched_data *q = qdisc_priv(sch);
1172	int i, err;
1173
1174	sch->limit		= 10000;
1175	q->flow_plimit		= 100;
1176	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
1177	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
1178	q->flow_refill_delay	= msecs_to_jiffies(40);
1179	q->flow_max_rate	= ~0UL;
1180	q->time_next_delayed_flow = ~0ULL;
1181	q->rate_enable		= 1;
1182	for (i = 0; i < FQ_BANDS; i++) {
1183		q->band_flows[i].new_flows.first = NULL;
1184		q->band_flows[i].old_flows.first = NULL;
1185	}
1186	q->band_flows[0].quantum = 9 << 16;
1187	q->band_flows[1].quantum = 3 << 16;
1188	q->band_flows[2].quantum = 1 << 16;
1189	q->delayed		= RB_ROOT;
1190	q->fq_root		= NULL;
1191	q->fq_trees_log		= ilog2(1024);
1192	q->orphan_mask		= 1024 - 1;
1193	q->low_rate_threshold	= 550000 / 8;
1194
1195	q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
1196
1197	q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
1198	q->horizon_drop = 1; /* by default, drop packets beyond horizon */
1199
1200	/* Default ce_threshold of 4294 seconds */
1201	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
1202
1203	fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
1204	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
1205
1206	if (opt)
1207		err = fq_change(sch, opt, extack);
1208	else
1209		err = fq_resize(sch, q->fq_trees_log);
1210
1211	return err;
1212}
1213
1214static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
1215{
1216	struct fq_sched_data *q = qdisc_priv(sch);
1217	struct tc_prio_qopt prio = {
1218		.bands = FQ_BANDS,
1219	};
1220	struct nlattr *opts;
1221	u64 offload_horizon;
1222	u64 ce_threshold;
1223	s32 weights[3];
1224	u64 horizon;
1225
1226	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1227	if (opts == NULL)
1228		goto nla_put_failure;
1229
1230	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
1231
1232	ce_threshold = READ_ONCE(q->ce_threshold);
1233	do_div(ce_threshold, NSEC_PER_USEC);
1234
1235	horizon = READ_ONCE(q->horizon);
1236	do_div(horizon, NSEC_PER_USEC);
1237
1238	offload_horizon = READ_ONCE(q->offload_horizon);
1239	do_div(offload_horizon, NSEC_PER_USEC);
1240
1241	if (nla_put_u32(skb, TCA_FQ_PLIMIT,
1242			READ_ONCE(sch->limit)) ||
1243	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT,
1244			READ_ONCE(q->flow_plimit)) ||
1245	    nla_put_u32(skb, TCA_FQ_QUANTUM,
1246			READ_ONCE(q->quantum)) ||
1247	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM,
1248			READ_ONCE(q->initial_quantum)) ||
1249	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE,
1250			READ_ONCE(q->rate_enable)) ||
1251	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
1252			min_t(unsigned long,
1253			      READ_ONCE(q->flow_max_rate), ~0U)) ||
1254	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
1255			jiffies_to_usecs(READ_ONCE(q->flow_refill_delay))) ||
1256	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK,
1257			READ_ONCE(q->orphan_mask)) ||
1258	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
1259			READ_ONCE(q->low_rate_threshold)) ||
1260	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
1261	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG,
1262			READ_ONCE(q->fq_trees_log)) ||
1263	    nla_put_u32(skb, TCA_FQ_TIMER_SLACK,
1264			READ_ONCE(q->timer_slack)) ||
1265	    nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
1266	    nla_put_u32(skb, TCA_FQ_OFFLOAD_HORIZON, (u32)offload_horizon) ||
1267	    nla_put_u8(skb, TCA_FQ_HORIZON_DROP,
1268		       READ_ONCE(q->horizon_drop)))
1269		goto nla_put_failure;
1270
1271	fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
1272	if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
1273		goto nla_put_failure;
1274
1275	weights[0] = READ_ONCE(q->band_flows[0].quantum);
1276	weights[1] = READ_ONCE(q->band_flows[1].quantum);
1277	weights[2] = READ_ONCE(q->band_flows[2].quantum);
1278	if (nla_put(skb, TCA_FQ_WEIGHTS, sizeof(weights), &weights))
1279		goto nla_put_failure;
1280
1281	return nla_nest_end(skb, opts);
1282
1283nla_put_failure:
1284	return -1;
1285}
1286
1287static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1288{
1289	struct fq_sched_data *q = qdisc_priv(sch);
1290	struct tc_fq_qd_stats st;
1291	int i;
1292
1293	st.pad = 0;
1294
1295	sch_tree_lock(sch);
1296
1297	st.gc_flows		  = q->stat_gc_flows;
1298	st.highprio_packets	  = 0;
1299	st.fastpath_packets	  = q->internal.stat_fastpath_packets;
1300	st.tcp_retrans		  = 0;
1301	st.throttled		  = q->stat_throttled;
1302	st.flows_plimit		  = q->stat_flows_plimit;
1303	st.pkts_too_long	  = q->stat_pkts_too_long;
1304	st.allocation_errors	  = q->stat_allocation_errors;
1305	st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1306				    ktime_get_ns();
1307	st.flows		  = q->flows;
1308	st.inactive_flows	  = q->inactive_flows;
1309	st.throttled_flows	  = q->throttled_flows;
1310	st.unthrottle_latency_ns  = min_t(unsigned long,
1311					  q->unthrottle_latency_ns, ~0U);
1312	st.ce_mark		  = q->stat_ce_mark;
1313	st.horizon_drops	  = q->stat_horizon_drops;
1314	st.horizon_caps		  = q->stat_horizon_caps;
1315	for (i = 0; i < FQ_BANDS; i++) {
1316		st.band_drops[i]  = q->stat_band_drops[i];
1317		st.band_pkt_count[i] = q->band_pkt_count[i];
1318	}
1319	sch_tree_unlock(sch);
1320
1321	return gnet_stats_copy_app(d, &st, sizeof(st));
1322}
1323
1324static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1325	.id		=	"fq",
1326	.priv_size	=	sizeof(struct fq_sched_data),
1327
1328	.enqueue	=	fq_enqueue,
1329	.dequeue	=	fq_dequeue,
1330	.peek		=	qdisc_peek_dequeued,
1331	.init		=	fq_init,
1332	.reset		=	fq_reset,
1333	.destroy	=	fq_destroy,
1334	.change		=	fq_change,
1335	.dump		=	fq_dump,
1336	.dump_stats	=	fq_dump_stats,
1337	.owner		=	THIS_MODULE,
1338};
1339MODULE_ALIAS_NET_SCH("fq");
1340
1341static int __init fq_module_init(void)
1342{
1343	int ret;
1344
1345	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1346					   sizeof(struct fq_flow),
1347					   0, SLAB_HWCACHE_ALIGN, NULL);
1348	if (!fq_flow_cachep)
1349		return -ENOMEM;
1350
1351	ret = register_qdisc(&fq_qdisc_ops);
1352	if (ret)
1353		kmem_cache_destroy(fq_flow_cachep);
1354	return ret;
1355}
1356
1357static void __exit fq_module_exit(void)
1358{
1359	unregister_qdisc(&fq_qdisc_ops);
1360	kmem_cache_destroy(fq_flow_cachep);
1361}
1362
1363module_init(fq_module_init)
1364module_exit(fq_module_exit)
1365MODULE_AUTHOR("Eric Dumazet");
1366MODULE_LICENSE("GPL");
1367MODULE_DESCRIPTION("Fair Queue Packet Scheduler");