Linux Audio

Check our new training course

Loading...
v6.13.7
   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		flow->head = skb->next;
 484	} else {
 485		rb_erase(&skb->rbnode, &flow->t_root);
 486		skb->dev = qdisc_dev(sch);
 487	}
 488}
 489
 490/* Remove one skb from flow queue.
 491 * This skb must be the return value of prior fq_peek().
 492 */
 493static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
 494			   struct sk_buff *skb)
 495{
 496	fq_erase_head(sch, flow, skb);
 497	skb_mark_not_on_list(skb);
 498	qdisc_qstats_backlog_dec(sch, skb);
 499	sch->q.qlen--;
 
 
 
 
 
 
 500}
 501
 502static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 503{
 504	struct rb_node **p, *parent;
 505	struct sk_buff *head, *aux;
 506
 
 
 507	head = flow->head;
 508	if (!head ||
 509	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 510		if (!head)
 511			flow->head = skb;
 512		else
 513			flow->tail->next = skb;
 514		flow->tail = skb;
 515		skb->next = NULL;
 516		return;
 517	}
 518
 519	p = &flow->t_root.rb_node;
 520	parent = NULL;
 521
 522	while (*p) {
 523		parent = *p;
 524		aux = rb_to_skb(parent);
 525		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 526			p = &parent->rb_right;
 527		else
 528			p = &parent->rb_left;
 529	}
 530	rb_link_node(&skb->rbnode, parent, p);
 531	rb_insert_color(&skb->rbnode, &flow->t_root);
 532}
 533
 534static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
 535				     const struct fq_sched_data *q, u64 now)
 536{
 537	return unlikely((s64)skb->tstamp > (s64)(now + q->horizon));
 538}
 539
 540static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 541		      struct sk_buff **to_free)
 542{
 543	struct fq_sched_data *q = qdisc_priv(sch);
 544	struct fq_flow *f;
 545	u64 now;
 546	u8 band;
 547
 548	band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
 549	if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
 550		q->stat_band_drops[band]++;
 551		return qdisc_drop(skb, sch, to_free);
 552	}
 553
 554	now = ktime_get_ns();
 555	if (!skb->tstamp) {
 556		fq_skb_cb(skb)->time_to_send = now;
 557	} else {
 558		/* Check if packet timestamp is too far in the future. */
 559		if (fq_packet_beyond_horizon(skb, q, now)) {
 560			if (q->horizon_drop) {
 561					q->stat_horizon_drops++;
 562					return qdisc_drop(skb, sch, to_free);
 563			}
 564			q->stat_horizon_caps++;
 565			skb->tstamp = now + q->horizon;
 566		}
 567		fq_skb_cb(skb)->time_to_send = skb->tstamp;
 568	}
 569
 570	f = fq_classify(sch, skb, now);
 571
 572	if (f != &q->internal) {
 573		if (unlikely(f->qlen >= q->flow_plimit)) {
 574			q->stat_flows_plimit++;
 575			return qdisc_drop(skb, sch, to_free);
 576		}
 577
 578		if (fq_flow_is_detached(f)) {
 579			fq_flow_add_tail(q, f, NEW_FLOW);
 580			if (time_after(jiffies, f->age + q->flow_refill_delay))
 581				f->credit = max_t(u32, f->credit, q->quantum);
 
 
 
 
 582		}
 583
 584		f->band = band;
 585		q->band_pkt_count[band]++;
 586		fq_skb_cb(skb)->band = band;
 587		if (f->qlen == 0)
 588			q->inactive_flows--;
 589	}
 590
 591	f->qlen++;
 592	/* Note: this overwrites f->age */
 593	flow_queue_add(f, skb);
 594
 595	qdisc_qstats_backlog_inc(sch, skb);
 
 
 596	sch->q.qlen++;
 597
 598	return NET_XMIT_SUCCESS;
 599}
 600
 601static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 602{
 603	unsigned long sample;
 604	struct rb_node *p;
 605
 606	if (q->time_next_delayed_flow > now + q->offload_horizon)
 607		return;
 608
 609	/* Update unthrottle latency EWMA.
 610	 * This is cheap and can help diagnosing timer/latency problems.
 611	 */
 612	sample = (unsigned long)(now - q->time_next_delayed_flow);
 613	if ((long)sample > 0) {
 614		q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 615		q->unthrottle_latency_ns += sample >> 3;
 616	}
 617	now += q->offload_horizon;
 618
 619	q->time_next_delayed_flow = ~0ULL;
 620	while ((p = rb_first(&q->delayed)) != NULL) {
 621		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 622
 623		if (f->time_next_packet > now) {
 624			q->time_next_delayed_flow = f->time_next_packet;
 625			break;
 626		}
 627		fq_flow_unset_throttled(q, f);
 628	}
 629}
 630
 631static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
 632{
 633	if (pband->credit <= 0)
 634		return NULL;
 635
 636	if (pband->new_flows.first)
 637		return &pband->new_flows;
 638
 639	return pband->old_flows.first ? &pband->old_flows : NULL;
 640}
 641
 642static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 643{
 644	struct fq_sched_data *q = qdisc_priv(sch);
 645	struct fq_perband_flows *pband;
 646	struct fq_flow_head *head;
 647	struct sk_buff *skb;
 648	struct fq_flow *f;
 649	unsigned long rate;
 650	int retry;
 651	u32 plen;
 652	u64 now;
 653
 654	if (!sch->q.qlen)
 655		return NULL;
 656
 657	skb = fq_peek(&q->internal);
 658	if (unlikely(skb)) {
 659		q->internal.qlen--;
 660		fq_dequeue_skb(sch, &q->internal, skb);
 661		goto out;
 662	}
 663
 664	now = ktime_get_ns();
 665	fq_check_throttled(q, now);
 666	retry = 0;
 667	pband = &q->band_flows[q->band_nr];
 668begin:
 669	head = fq_pband_head_select(pband);
 670	if (!head) {
 671		while (++retry <= FQ_BANDS) {
 672			if (++q->band_nr == FQ_BANDS)
 673				q->band_nr = 0;
 674			pband = &q->band_flows[q->band_nr];
 675			pband->credit = min(pband->credit + pband->quantum,
 676					    pband->quantum);
 677			if (pband->credit > 0)
 678				goto begin;
 679			retry = 0;
 680		}
 681		if (q->time_next_delayed_flow != ~0ULL)
 682			qdisc_watchdog_schedule_range_ns(&q->watchdog,
 683							q->time_next_delayed_flow,
 684							q->timer_slack);
 685		return NULL;
 686	}
 687	f = head->first;
 688	retry = 0;
 689	if (f->credit <= 0) {
 690		f->credit += q->quantum;
 691		head->first = f->next;
 692		fq_flow_add_tail(q, f, OLD_FLOW);
 693		goto begin;
 694	}
 695
 696	skb = fq_peek(f);
 697	if (skb) {
 698		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 699					     f->time_next_packet);
 700
 701		if (now + q->offload_horizon < time_next_packet) {
 702			head->first = f->next;
 703			f->time_next_packet = time_next_packet;
 704			fq_flow_set_throttled(q, f);
 705			goto begin;
 706		}
 707		prefetch(&skb->end);
 708		if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
 709			INET_ECN_set_ce(skb);
 710			q->stat_ce_mark++;
 711		}
 712		if (--f->qlen == 0)
 713			q->inactive_flows++;
 714		q->band_pkt_count[fq_skb_cb(skb)->band]--;
 715		fq_dequeue_skb(sch, f, skb);
 716	} else {
 717		head->first = f->next;
 718		/* force a pass through old_flows to prevent starvation */
 719		if (head == &pband->new_flows) {
 720			fq_flow_add_tail(q, f, OLD_FLOW);
 721		} else {
 722			fq_flow_set_detached(f);
 
 723		}
 724		goto begin;
 725	}
 
 726	plen = qdisc_pkt_len(skb);
 727	f->credit -= plen;
 728	pband->credit -= plen;
 729
 730	if (!q->rate_enable)
 731		goto out;
 732
 733	rate = q->flow_max_rate;
 734
 735	/* If EDT time was provided for this skb, we need to
 736	 * update f->time_next_packet only if this qdisc enforces
 737	 * a flow max rate.
 738	 */
 739	if (!skb->tstamp) {
 740		if (skb->sk)
 741			rate = min(READ_ONCE(skb->sk->sk_pacing_rate), rate);
 742
 743		if (rate <= q->low_rate_threshold) {
 744			f->credit = 0;
 745		} else {
 746			plen = max(plen, q->quantum);
 747			if (f->credit > 0)
 748				goto out;
 749		}
 750	}
 751	if (rate != ~0UL) {
 752		u64 len = (u64)plen * NSEC_PER_SEC;
 753
 754		if (likely(rate))
 755			len = div64_ul(len, rate);
 756		/* Since socket rate can change later,
 757		 * clamp the delay to 1 second.
 758		 * Really, providers of too big packets should be fixed !
 759		 */
 760		if (unlikely(len > NSEC_PER_SEC)) {
 761			len = NSEC_PER_SEC;
 762			q->stat_pkts_too_long++;
 763		}
 764		/* Account for schedule/timers drifts.
 765		 * f->time_next_packet was set when prior packet was sent,
 766		 * and current time (@now) can be too late by tens of us.
 767		 */
 768		if (f->time_next_packet)
 769			len -= min(len/2, now - f->time_next_packet);
 770		f->time_next_packet = now + len;
 771	}
 772out:
 773	qdisc_bstats_update(sch, skb);
 774	return skb;
 775}
 776
 777static void fq_flow_purge(struct fq_flow *flow)
 778{
 779	struct rb_node *p = rb_first(&flow->t_root);
 780
 781	while (p) {
 782		struct sk_buff *skb = rb_to_skb(p);
 783
 784		p = rb_next(p);
 785		rb_erase(&skb->rbnode, &flow->t_root);
 786		rtnl_kfree_skbs(skb, skb);
 787	}
 788	rtnl_kfree_skbs(flow->head, flow->tail);
 789	flow->head = NULL;
 790	flow->qlen = 0;
 791}
 792
 793static void fq_reset(struct Qdisc *sch)
 794{
 795	struct fq_sched_data *q = qdisc_priv(sch);
 796	struct rb_root *root;
 797	struct rb_node *p;
 798	struct fq_flow *f;
 799	unsigned int idx;
 800
 801	sch->q.qlen = 0;
 802	sch->qstats.backlog = 0;
 803
 804	fq_flow_purge(&q->internal);
 805
 806	if (!q->fq_root)
 807		return;
 808
 809	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 810		root = &q->fq_root[idx];
 811		while ((p = rb_first(root)) != NULL) {
 812			f = rb_entry(p, struct fq_flow, fq_node);
 813			rb_erase(p, root);
 814
 815			fq_flow_purge(f);
 816
 817			kmem_cache_free(fq_flow_cachep, f);
 818		}
 819	}
 820	for (idx = 0; idx < FQ_BANDS; idx++) {
 821		q->band_flows[idx].new_flows.first = NULL;
 822		q->band_flows[idx].old_flows.first = NULL;
 823	}
 824	q->delayed		= RB_ROOT;
 825	q->flows		= 0;
 826	q->inactive_flows	= 0;
 827	q->throttled_flows	= 0;
 828}
 829
 830static void fq_rehash(struct fq_sched_data *q,
 831		      struct rb_root *old_array, u32 old_log,
 832		      struct rb_root *new_array, u32 new_log)
 833{
 834	struct rb_node *op, **np, *parent;
 835	struct rb_root *oroot, *nroot;
 836	struct fq_flow *of, *nf;
 837	int fcnt = 0;
 838	u32 idx;
 839
 840	for (idx = 0; idx < (1U << old_log); idx++) {
 841		oroot = &old_array[idx];
 842		while ((op = rb_first(oroot)) != NULL) {
 843			rb_erase(op, oroot);
 844			of = rb_entry(op, struct fq_flow, fq_node);
 845			if (fq_gc_candidate(of)) {
 846				fcnt++;
 847				kmem_cache_free(fq_flow_cachep, of);
 848				continue;
 849			}
 850			nroot = &new_array[hash_ptr(of->sk, new_log)];
 851
 852			np = &nroot->rb_node;
 853			parent = NULL;
 854			while (*np) {
 855				parent = *np;
 856
 857				nf = rb_entry(parent, struct fq_flow, fq_node);
 858				BUG_ON(nf->sk == of->sk);
 859
 860				if (nf->sk > of->sk)
 861					np = &parent->rb_right;
 862				else
 863					np = &parent->rb_left;
 864			}
 865
 866			rb_link_node(&of->fq_node, parent, np);
 867			rb_insert_color(&of->fq_node, nroot);
 868		}
 869	}
 870	q->flows -= fcnt;
 871	q->inactive_flows -= fcnt;
 872	q->stat_gc_flows += fcnt;
 873}
 874
 875static void fq_free(void *addr)
 876{
 877	kvfree(addr);
 878}
 879
 880static int fq_resize(struct Qdisc *sch, u32 log)
 881{
 882	struct fq_sched_data *q = qdisc_priv(sch);
 883	struct rb_root *array;
 884	void *old_fq_root;
 885	u32 idx;
 886
 887	if (q->fq_root && log == q->fq_trees_log)
 888		return 0;
 889
 890	/* If XPS was setup, we can allocate memory on right NUMA node */
 891	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 892			      netdev_queue_numa_node_read(sch->dev_queue));
 893	if (!array)
 894		return -ENOMEM;
 895
 896	for (idx = 0; idx < (1U << log); idx++)
 897		array[idx] = RB_ROOT;
 898
 899	sch_tree_lock(sch);
 900
 901	old_fq_root = q->fq_root;
 902	if (old_fq_root)
 903		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 904
 905	q->fq_root = array;
 906	WRITE_ONCE(q->fq_trees_log, log);
 907
 908	sch_tree_unlock(sch);
 909
 910	fq_free(old_fq_root);
 911
 912	return 0;
 913}
 914
 915static const struct netlink_range_validation iq_range = {
 916	.max = INT_MAX,
 917};
 918
 919static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 920	[TCA_FQ_UNSPEC]			= { .strict_start_type = TCA_FQ_TIMER_SLACK },
 921
 922	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
 923	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
 924	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
 925	[TCA_FQ_INITIAL_QUANTUM]	= NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
 926	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
 927	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
 928	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
 929	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
 930	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
 931	[TCA_FQ_ORPHAN_MASK]		= { .type = NLA_U32 },
 932	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
 933	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
 934	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
 935	[TCA_FQ_HORIZON]		= { .type = NLA_U32 },
 936	[TCA_FQ_HORIZON_DROP]		= { .type = NLA_U8 },
 937	[TCA_FQ_PRIOMAP]		= NLA_POLICY_EXACT_LEN(sizeof(struct tc_prio_qopt)),
 938	[TCA_FQ_WEIGHTS]		= NLA_POLICY_EXACT_LEN(FQ_BANDS * sizeof(s32)),
 939	[TCA_FQ_OFFLOAD_HORIZON]	= { .type = NLA_U32 },
 940};
 941
 942/* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
 943static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
 944{
 945	const int num_elems = TC_PRIO_MAX + 1;
 946	u8 tmp[FQ_PRIO2BAND_CRUMB_SIZE];
 947	int i;
 948
 949	memset(tmp, 0, sizeof(tmp));
 950	for (i = 0; i < num_elems; i++)
 951		tmp[i / 4] |= in[i] << (2 * (i & 0x3));
 952
 953	for (i = 0; i < FQ_PRIO2BAND_CRUMB_SIZE; i++)
 954		WRITE_ONCE(out[i], tmp[i]);
 955}
 956
 957static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
 958{
 959	const int num_elems = TC_PRIO_MAX + 1;
 960	int i;
 961
 962	for (i = 0; i < num_elems; i++)
 963		out[i] = fq_prio2band(in, i);
 964}
 965
 966static int fq_load_weights(struct fq_sched_data *q,
 967			   const struct nlattr *attr,
 968			   struct netlink_ext_ack *extack)
 969{
 970	s32 *weights = nla_data(attr);
 971	int i;
 972
 973	for (i = 0; i < FQ_BANDS; i++) {
 974		if (weights[i] < FQ_MIN_WEIGHT) {
 975			NL_SET_ERR_MSG_FMT_MOD(extack, "Weight %d less that minimum allowed %d",
 976					       weights[i], FQ_MIN_WEIGHT);
 977			return -EINVAL;
 978		}
 979	}
 980	for (i = 0; i < FQ_BANDS; i++)
 981		WRITE_ONCE(q->band_flows[i].quantum, weights[i]);
 982	return 0;
 983}
 984
 985static int fq_load_priomap(struct fq_sched_data *q,
 986			   const struct nlattr *attr,
 987			   struct netlink_ext_ack *extack)
 988{
 989	const struct tc_prio_qopt *map = nla_data(attr);
 990	int i;
 991
 992	if (map->bands != FQ_BANDS) {
 993		NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
 994		return -EINVAL;
 995	}
 996	for (i = 0; i < TC_PRIO_MAX + 1; i++) {
 997		if (map->priomap[i] >= FQ_BANDS) {
 998			NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
 999					       i, map->priomap[i]);
1000			return -EINVAL;
1001		}
1002	}
1003	fq_prio2band_compress_crumb(map->priomap, q->prio2band);
1004	return 0;
1005}
1006
1007static int fq_change(struct Qdisc *sch, struct nlattr *opt,
1008		     struct netlink_ext_ack *extack)
1009{
1010	struct fq_sched_data *q = qdisc_priv(sch);
1011	struct nlattr *tb[TCA_FQ_MAX + 1];
1012	int err, drop_count = 0;
1013	unsigned drop_len = 0;
1014	u32 fq_log;
1015
 
 
 
1016	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
1017					  NULL);
1018	if (err < 0)
1019		return err;
1020
1021	sch_tree_lock(sch);
1022
1023	fq_log = q->fq_trees_log;
1024
1025	if (tb[TCA_FQ_BUCKETS_LOG]) {
1026		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
1027
1028		if (nval >= 1 && nval <= ilog2(256*1024))
1029			fq_log = nval;
1030		else
1031			err = -EINVAL;
1032	}
1033	if (tb[TCA_FQ_PLIMIT])
1034		WRITE_ONCE(sch->limit,
1035			   nla_get_u32(tb[TCA_FQ_PLIMIT]));
1036
1037	if (tb[TCA_FQ_FLOW_PLIMIT])
1038		WRITE_ONCE(q->flow_plimit,
1039			   nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]));
1040
1041	if (tb[TCA_FQ_QUANTUM]) {
1042		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
1043
1044		if (quantum > 0 && quantum <= (1 << 20)) {
1045			WRITE_ONCE(q->quantum, quantum);
1046		} else {
1047			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
1048			err = -EINVAL;
1049		}
1050	}
1051
1052	if (tb[TCA_FQ_INITIAL_QUANTUM])
1053		WRITE_ONCE(q->initial_quantum,
1054			   nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]));
1055
1056	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
1057		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
1058				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
1059
1060	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
1061		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
1062
1063		WRITE_ONCE(q->flow_max_rate,
1064			   (rate == ~0U) ? ~0UL : rate);
1065	}
1066	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
1067		WRITE_ONCE(q->low_rate_threshold,
1068			   nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]));
1069
1070	if (tb[TCA_FQ_RATE_ENABLE]) {
1071		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
1072
1073		if (enable <= 1)
1074			WRITE_ONCE(q->rate_enable,
1075				   enable);
1076		else
1077			err = -EINVAL;
1078	}
1079
1080	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
1081		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
1082
1083		WRITE_ONCE(q->flow_refill_delay,
1084			   usecs_to_jiffies(usecs_delay));
1085	}
1086
1087	if (!err && tb[TCA_FQ_PRIOMAP])
1088		err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
1089
1090	if (!err && tb[TCA_FQ_WEIGHTS])
1091		err = fq_load_weights(q, tb[TCA_FQ_WEIGHTS], extack);
1092
1093	if (tb[TCA_FQ_ORPHAN_MASK])
1094		WRITE_ONCE(q->orphan_mask,
1095			   nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]));
1096
1097	if (tb[TCA_FQ_CE_THRESHOLD])
1098		WRITE_ONCE(q->ce_threshold,
1099			   (u64)NSEC_PER_USEC *
1100			   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]));
1101
1102	if (tb[TCA_FQ_TIMER_SLACK])
1103		WRITE_ONCE(q->timer_slack,
1104			   nla_get_u32(tb[TCA_FQ_TIMER_SLACK]));
1105
1106	if (tb[TCA_FQ_HORIZON])
1107		WRITE_ONCE(q->horizon,
1108			   (u64)NSEC_PER_USEC *
1109			   nla_get_u32(tb[TCA_FQ_HORIZON]));
1110
1111	if (tb[TCA_FQ_HORIZON_DROP])
1112		WRITE_ONCE(q->horizon_drop,
1113			   nla_get_u8(tb[TCA_FQ_HORIZON_DROP]));
1114
1115	if (tb[TCA_FQ_OFFLOAD_HORIZON]) {
1116		u64 offload_horizon = (u64)NSEC_PER_USEC *
1117				      nla_get_u32(tb[TCA_FQ_OFFLOAD_HORIZON]);
1118
1119		if (offload_horizon <= qdisc_dev(sch)->max_pacing_offload_horizon) {
1120			WRITE_ONCE(q->offload_horizon, offload_horizon);
1121		} else {
1122			NL_SET_ERR_MSG_MOD(extack, "invalid offload_horizon");
1123			err = -EINVAL;
1124		}
1125	}
1126	if (!err) {
1127
1128		sch_tree_unlock(sch);
1129		err = fq_resize(sch, fq_log);
1130		sch_tree_lock(sch);
1131	}
1132	while (sch->q.qlen > sch->limit) {
1133		struct sk_buff *skb = fq_dequeue(sch);
1134
1135		if (!skb)
1136			break;
1137		drop_len += qdisc_pkt_len(skb);
1138		rtnl_kfree_skbs(skb, skb);
1139		drop_count++;
1140	}
1141	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
1142
1143	sch_tree_unlock(sch);
1144	return err;
1145}
1146
1147static void fq_destroy(struct Qdisc *sch)
1148{
1149	struct fq_sched_data *q = qdisc_priv(sch);
1150
1151	fq_reset(sch);
1152	fq_free(q->fq_root);
1153	qdisc_watchdog_cancel(&q->watchdog);
1154}
1155
1156static int fq_init(struct Qdisc *sch, struct nlattr *opt,
1157		   struct netlink_ext_ack *extack)
1158{
1159	struct fq_sched_data *q = qdisc_priv(sch);
1160	int i, err;
1161
1162	sch->limit		= 10000;
1163	q->flow_plimit		= 100;
1164	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
1165	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
1166	q->flow_refill_delay	= msecs_to_jiffies(40);
1167	q->flow_max_rate	= ~0UL;
1168	q->time_next_delayed_flow = ~0ULL;
1169	q->rate_enable		= 1;
1170	for (i = 0; i < FQ_BANDS; i++) {
1171		q->band_flows[i].new_flows.first = NULL;
1172		q->band_flows[i].old_flows.first = NULL;
1173	}
1174	q->band_flows[0].quantum = 9 << 16;
1175	q->band_flows[1].quantum = 3 << 16;
1176	q->band_flows[2].quantum = 1 << 16;
1177	q->delayed		= RB_ROOT;
1178	q->fq_root		= NULL;
1179	q->fq_trees_log		= ilog2(1024);
1180	q->orphan_mask		= 1024 - 1;
1181	q->low_rate_threshold	= 550000 / 8;
1182
1183	q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
1184
1185	q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
1186	q->horizon_drop = 1; /* by default, drop packets beyond horizon */
1187
1188	/* Default ce_threshold of 4294 seconds */
1189	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
1190
1191	fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
1192	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
1193
1194	if (opt)
1195		err = fq_change(sch, opt, extack);
1196	else
1197		err = fq_resize(sch, q->fq_trees_log);
1198
1199	return err;
1200}
1201
1202static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
1203{
1204	struct fq_sched_data *q = qdisc_priv(sch);
1205	struct tc_prio_qopt prio = {
1206		.bands = FQ_BANDS,
1207	};
1208	struct nlattr *opts;
1209	u64 offload_horizon;
1210	u64 ce_threshold;
1211	s32 weights[3];
1212	u64 horizon;
1213
1214	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1215	if (opts == NULL)
1216		goto nla_put_failure;
1217
1218	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
1219
1220	ce_threshold = READ_ONCE(q->ce_threshold);
1221	do_div(ce_threshold, NSEC_PER_USEC);
1222
1223	horizon = READ_ONCE(q->horizon);
1224	do_div(horizon, NSEC_PER_USEC);
1225
1226	offload_horizon = READ_ONCE(q->offload_horizon);
1227	do_div(offload_horizon, NSEC_PER_USEC);
1228
1229	if (nla_put_u32(skb, TCA_FQ_PLIMIT,
1230			READ_ONCE(sch->limit)) ||
1231	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT,
1232			READ_ONCE(q->flow_plimit)) ||
1233	    nla_put_u32(skb, TCA_FQ_QUANTUM,
1234			READ_ONCE(q->quantum)) ||
1235	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM,
1236			READ_ONCE(q->initial_quantum)) ||
1237	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE,
1238			READ_ONCE(q->rate_enable)) ||
1239	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
1240			min_t(unsigned long,
1241			      READ_ONCE(q->flow_max_rate), ~0U)) ||
1242	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
1243			jiffies_to_usecs(READ_ONCE(q->flow_refill_delay))) ||
1244	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK,
1245			READ_ONCE(q->orphan_mask)) ||
1246	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
1247			READ_ONCE(q->low_rate_threshold)) ||
1248	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
1249	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG,
1250			READ_ONCE(q->fq_trees_log)) ||
1251	    nla_put_u32(skb, TCA_FQ_TIMER_SLACK,
1252			READ_ONCE(q->timer_slack)) ||
1253	    nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
1254	    nla_put_u32(skb, TCA_FQ_OFFLOAD_HORIZON, (u32)offload_horizon) ||
1255	    nla_put_u8(skb, TCA_FQ_HORIZON_DROP,
1256		       READ_ONCE(q->horizon_drop)))
1257		goto nla_put_failure;
1258
1259	fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
1260	if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
1261		goto nla_put_failure;
1262
1263	weights[0] = READ_ONCE(q->band_flows[0].quantum);
1264	weights[1] = READ_ONCE(q->band_flows[1].quantum);
1265	weights[2] = READ_ONCE(q->band_flows[2].quantum);
1266	if (nla_put(skb, TCA_FQ_WEIGHTS, sizeof(weights), &weights))
1267		goto nla_put_failure;
1268
1269	return nla_nest_end(skb, opts);
1270
1271nla_put_failure:
1272	return -1;
1273}
1274
1275static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1276{
1277	struct fq_sched_data *q = qdisc_priv(sch);
1278	struct tc_fq_qd_stats st;
1279	int i;
1280
1281	st.pad = 0;
1282
1283	sch_tree_lock(sch);
1284
1285	st.gc_flows		  = q->stat_gc_flows;
1286	st.highprio_packets	  = 0;
1287	st.fastpath_packets	  = q->internal.stat_fastpath_packets;
1288	st.tcp_retrans		  = 0;
1289	st.throttled		  = q->stat_throttled;
1290	st.flows_plimit		  = q->stat_flows_plimit;
1291	st.pkts_too_long	  = q->stat_pkts_too_long;
1292	st.allocation_errors	  = q->stat_allocation_errors;
1293	st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1294				    ktime_get_ns();
1295	st.flows		  = q->flows;
1296	st.inactive_flows	  = q->inactive_flows;
1297	st.throttled_flows	  = q->throttled_flows;
1298	st.unthrottle_latency_ns  = min_t(unsigned long,
1299					  q->unthrottle_latency_ns, ~0U);
1300	st.ce_mark		  = q->stat_ce_mark;
1301	st.horizon_drops	  = q->stat_horizon_drops;
1302	st.horizon_caps		  = q->stat_horizon_caps;
1303	for (i = 0; i < FQ_BANDS; i++) {
1304		st.band_drops[i]  = q->stat_band_drops[i];
1305		st.band_pkt_count[i] = q->band_pkt_count[i];
1306	}
1307	sch_tree_unlock(sch);
1308
1309	return gnet_stats_copy_app(d, &st, sizeof(st));
1310}
1311
1312static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1313	.id		=	"fq",
1314	.priv_size	=	sizeof(struct fq_sched_data),
1315
1316	.enqueue	=	fq_enqueue,
1317	.dequeue	=	fq_dequeue,
1318	.peek		=	qdisc_peek_dequeued,
1319	.init		=	fq_init,
1320	.reset		=	fq_reset,
1321	.destroy	=	fq_destroy,
1322	.change		=	fq_change,
1323	.dump		=	fq_dump,
1324	.dump_stats	=	fq_dump_stats,
1325	.owner		=	THIS_MODULE,
1326};
1327MODULE_ALIAS_NET_SCH("fq");
1328
1329static int __init fq_module_init(void)
1330{
1331	int ret;
1332
1333	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1334					   sizeof(struct fq_flow),
1335					   0, SLAB_HWCACHE_ALIGN, NULL);
1336	if (!fq_flow_cachep)
1337		return -ENOMEM;
1338
1339	ret = register_qdisc(&fq_qdisc_ops);
1340	if (ret)
1341		kmem_cache_destroy(fq_flow_cachep);
1342	return ret;
1343}
1344
1345static void __exit fq_module_exit(void)
1346{
1347	unregister_qdisc(&fq_qdisc_ops);
1348	kmem_cache_destroy(fq_flow_cachep);
1349}
1350
1351module_init(fq_module_init)
1352module_exit(fq_module_exit)
1353MODULE_AUTHOR("Eric Dumazet");
1354MODULE_LICENSE("GPL");
1355MODULE_DESCRIPTION("Fair Queue Packet Scheduler");
v5.4
   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) 2013-2015 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};
  56
  57static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
  58{
  59	qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb));
  60	return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data;
  61}
  62
  63/*
  64 * Per flow structure, dynamically allocated.
  65 * If packets have monotically increasing time_to_send, they are placed in O(1)
  66 * in linear list (head,tail), otherwise are placed in a rbtree (t_root).
  67 */
  68struct fq_flow {
 
  69	struct rb_root	t_root;
  70	struct sk_buff	*head;		/* list of skbs for this flow : first skb */
  71	union {
  72		struct sk_buff *tail;	/* last skb in the list */
  73		unsigned long  age;	/* jiffies when flow was emptied, for gc */
 
 
 
 
 
 
 
  74	};
  75	struct rb_node	fq_node;	/* anchor in fq_root[] trees */
  76	struct sock	*sk;
 
  77	int		qlen;		/* number of packets in flow queue */
 
 
  78	int		credit;
  79	u32		socket_hash;	/* sk_hash */
  80	struct fq_flow *next;		/* next pointer in RR lists, or &detached */
  81
  82	struct rb_node  rate_node;	/* anchor in q->delayed tree */
  83	u64		time_next_packet;
  84};
  85
  86struct fq_flow_head {
  87	struct fq_flow *first;
  88	struct fq_flow *last;
  89};
  90
  91struct fq_sched_data {
  92	struct fq_flow_head new_flows;
 
 
 
 
  93
  94	struct fq_flow_head old_flows;
  95
  96	struct rb_root	delayed;	/* for rate limited flows */
  97	u64		time_next_delayed_flow;
  98	unsigned long	unthrottle_latency_ns;
  99
 100	struct fq_flow	internal;	/* for non classified or high prio packets */
 101	u32		quantum;
 102	u32		initial_quantum;
 103	u32		flow_refill_delay;
 104	u32		flow_plimit;	/* max packets per flow */
 105	unsigned long	flow_max_rate;	/* optional max rate per flow */
 106	u64		ce_threshold;
 
 107	u32		orphan_mask;	/* mask for orphaned skb */
 108	u32		low_rate_threshold;
 109	struct rb_root	*fq_root;
 110	u8		rate_enable;
 111	u8		fq_trees_log;
 
 
 
 
 
 112
 
 
 
 
 
 
 
 
 
 
 113	u32		flows;
 114	u32		inactive_flows;
 115	u32		throttled_flows;
 116
 
 
 117	u64		stat_gc_flows;
 118	u64		stat_internal_packets;
 119	u64		stat_throttled;
 
 
 120	u64		stat_ce_mark;
 
 
 121	u64		stat_flows_plimit;
 122	u64		stat_pkts_too_long;
 123	u64		stat_allocation_errors;
 124	struct qdisc_watchdog watchdog;
 125};
 126
 127/* special value to mark a detached flow (not on old/new list) */
 128static struct fq_flow detached, throttled;
 
 
 
 129
 
 
 
 
 
 
 130static void fq_flow_set_detached(struct fq_flow *f)
 131{
 132	f->next = &detached;
 133	f->age = jiffies;
 134}
 135
 136static bool fq_flow_is_detached(const struct fq_flow *f)
 137{
 138	return f->next == &detached;
 139}
 140
 
 
 
 141static bool fq_flow_is_throttled(const struct fq_flow *f)
 142{
 143	return f->next == &throttled;
 144}
 145
 146static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 
 
 
 
 
 
 147{
 
 
 
 
 
 148	if (head->first)
 149		head->last->next = flow;
 150	else
 151		head->first = flow;
 152	head->last = flow;
 153	flow->next = NULL;
 154}
 155
 156static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 157{
 158	rb_erase(&f->rate_node, &q->delayed);
 159	q->throttled_flows--;
 160	fq_flow_add_tail(&q->old_flows, f);
 161}
 162
 163static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 164{
 165	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 166
 167	while (*p) {
 168		struct fq_flow *aux;
 169
 170		parent = *p;
 171		aux = rb_entry(parent, struct fq_flow, rate_node);
 172		if (f->time_next_packet >= aux->time_next_packet)
 173			p = &parent->rb_right;
 174		else
 175			p = &parent->rb_left;
 176	}
 177	rb_link_node(&f->rate_node, parent, p);
 178	rb_insert_color(&f->rate_node, &q->delayed);
 179	q->throttled_flows++;
 180	q->stat_throttled++;
 181
 182	f->next = &throttled;
 183	if (q->time_next_delayed_flow > f->time_next_packet)
 184		q->time_next_delayed_flow = f->time_next_packet;
 185}
 186
 187
 188static struct kmem_cache *fq_flow_cachep __read_mostly;
 189
 190
 191/* limit number of collected flows per round */
 192#define FQ_GC_MAX 8
 193#define FQ_GC_AGE (3*HZ)
 194
 195static bool fq_gc_candidate(const struct fq_flow *f)
 196{
 197	return fq_flow_is_detached(f) &&
 198	       time_after(jiffies, f->age + FQ_GC_AGE);
 199}
 200
 201static void fq_gc(struct fq_sched_data *q,
 202		  struct rb_root *root,
 203		  struct sock *sk)
 204{
 205	struct fq_flow *f, *tofree[FQ_GC_MAX];
 206	struct rb_node **p, *parent;
 207	int fcnt = 0;
 
 
 208
 209	p = &root->rb_node;
 210	parent = NULL;
 211	while (*p) {
 212		parent = *p;
 213
 214		f = rb_entry(parent, struct fq_flow, fq_node);
 215		if (f->sk == sk)
 216			break;
 217
 218		if (fq_gc_candidate(f)) {
 219			tofree[fcnt++] = f;
 220			if (fcnt == FQ_GC_MAX)
 221				break;
 222		}
 223
 224		if (f->sk > sk)
 225			p = &parent->rb_right;
 226		else
 227			p = &parent->rb_left;
 228	}
 229
 
 
 
 
 
 
 
 230	q->flows -= fcnt;
 231	q->inactive_flows -= fcnt;
 232	q->stat_gc_flows += fcnt;
 233	while (fcnt) {
 234		struct fq_flow *f = tofree[--fcnt];
 235
 236		rb_erase(&f->fq_node, root);
 237		kmem_cache_free(fq_flow_cachep, f);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 238	}
 
 
 
 
 
 
 
 
 
 
 239}
 240
 241static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 
 242{
 
 243	struct rb_node **p, *parent;
 244	struct sock *sk = skb->sk;
 245	struct rb_root *root;
 246	struct fq_flow *f;
 247
 248	/* warning: no starvation prevention... */
 249	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 250		return &q->internal;
 251
 252	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 253	 * or a listener (SYNCOOKIE mode)
 254	 * 1) request sockets are not full blown,
 255	 *    they do not contain sk_pacing_rate
 256	 * 2) They are not part of a 'flow' yet
 257	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 258	 *    especially if the listener set SO_MAX_PACING_RATE
 259	 * 4) We pretend they are orphaned
 
 260	 */
 261	if (!sk || sk_listener(sk)) {
 262		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 263
 264		/* By forcing low order bit to 1, we make sure to not
 265		 * collide with a local flow (socket pointers are word aligned)
 266		 */
 267		sk = (struct sock *)((hash << 1) | 1UL);
 268		skb_orphan(skb);
 269	} else if (sk->sk_state == TCP_CLOSE) {
 270		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 271		/*
 272		 * Sockets in TCP_CLOSE are non connected.
 273		 * Typical use case is UDP sockets, they can send packets
 274		 * with sendto() to many different destinations.
 275		 * We probably could use a generic bit advertising
 276		 * non connected sockets, instead of sk_state == TCP_CLOSE,
 277		 * if we care enough.
 278		 */
 279		sk = (struct sock *)((hash << 1) | 1UL);
 280	}
 281
 
 
 
 
 
 
 
 
 
 282	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 283
 284	if (q->flows >= (2U << q->fq_trees_log) &&
 285	    q->inactive_flows > q->flows/2)
 286		fq_gc(q, root, sk);
 287
 288	p = &root->rb_node;
 289	parent = NULL;
 290	while (*p) {
 291		parent = *p;
 292
 293		f = rb_entry(parent, struct fq_flow, fq_node);
 294		if (f->sk == sk) {
 295			/* socket might have been reallocated, so check
 296			 * if its sk_hash is the same.
 297			 * It not, we need to refill credit with
 298			 * initial quantum
 299			 */
 300			if (unlikely(skb->sk == sk &&
 301				     f->socket_hash != sk->sk_hash)) {
 302				f->credit = q->initial_quantum;
 303				f->socket_hash = sk->sk_hash;
 
 
 
 304				if (fq_flow_is_throttled(f))
 305					fq_flow_unset_throttled(q, f);
 306				f->time_next_packet = 0ULL;
 307			}
 308			return f;
 309		}
 310		if (f->sk > sk)
 311			p = &parent->rb_right;
 312		else
 313			p = &parent->rb_left;
 314	}
 315
 316	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 317	if (unlikely(!f)) {
 318		q->stat_allocation_errors++;
 319		return &q->internal;
 320	}
 321	/* f->t_root is already zeroed after kmem_cache_zalloc() */
 322
 323	fq_flow_set_detached(f);
 324	f->sk = sk;
 325	if (skb->sk == sk)
 326		f->socket_hash = sk->sk_hash;
 
 
 
 
 327	f->credit = q->initial_quantum;
 328
 329	rb_link_node(&f->fq_node, parent, p);
 330	rb_insert_color(&f->fq_node, root);
 331
 332	q->flows++;
 333	q->inactive_flows++;
 334	return f;
 335}
 336
 337static struct sk_buff *fq_peek(struct fq_flow *flow)
 338{
 339	struct sk_buff *skb = skb_rb_first(&flow->t_root);
 340	struct sk_buff *head = flow->head;
 341
 342	if (!skb)
 343		return head;
 344
 345	if (!head)
 346		return skb;
 347
 348	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
 349		return skb;
 350	return head;
 351}
 352
 353static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
 354			  struct sk_buff *skb)
 355{
 356	if (skb == flow->head) {
 357		flow->head = skb->next;
 358	} else {
 359		rb_erase(&skb->rbnode, &flow->t_root);
 360		skb->dev = qdisc_dev(sch);
 361	}
 362}
 363
 364/* remove one skb from head of flow queue */
 365static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 
 
 
 366{
 367	struct sk_buff *skb = fq_peek(flow);
 368
 369	if (skb) {
 370		fq_erase_head(sch, flow, skb);
 371		skb_mark_not_on_list(skb);
 372		flow->qlen--;
 373		qdisc_qstats_backlog_dec(sch, skb);
 374		sch->q.qlen--;
 375	}
 376	return skb;
 377}
 378
 379static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 380{
 381	struct rb_node **p, *parent;
 382	struct sk_buff *head, *aux;
 383
 384	fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns();
 385
 386	head = flow->head;
 387	if (!head ||
 388	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 389		if (!head)
 390			flow->head = skb;
 391		else
 392			flow->tail->next = skb;
 393		flow->tail = skb;
 394		skb->next = NULL;
 395		return;
 396	}
 397
 398	p = &flow->t_root.rb_node;
 399	parent = NULL;
 400
 401	while (*p) {
 402		parent = *p;
 403		aux = rb_to_skb(parent);
 404		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 405			p = &parent->rb_right;
 406		else
 407			p = &parent->rb_left;
 408	}
 409	rb_link_node(&skb->rbnode, parent, p);
 410	rb_insert_color(&skb->rbnode, &flow->t_root);
 411}
 412
 
 
 
 
 
 
 413static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 414		      struct sk_buff **to_free)
 415{
 416	struct fq_sched_data *q = qdisc_priv(sch);
 417	struct fq_flow *f;
 
 
 418
 419	if (unlikely(sch->q.qlen >= sch->limit))
 
 
 420		return qdisc_drop(skb, sch, to_free);
 
 421
 422	f = fq_classify(skb, q);
 423	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 424		q->stat_flows_plimit++;
 425		return qdisc_drop(skb, sch, to_free);
 
 
 
 
 
 
 
 
 
 
 426	}
 427
 428	f->qlen++;
 429	qdisc_qstats_backlog_inc(sch, skb);
 430	if (fq_flow_is_detached(f)) {
 431		struct sock *sk = skb->sk;
 
 
 
 432
 433		fq_flow_add_tail(&q->new_flows, f);
 434		if (time_after(jiffies, f->age + q->flow_refill_delay))
 435			f->credit = max_t(u32, f->credit, q->quantum);
 436		if (sk && q->rate_enable) {
 437			if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
 438				     SK_PACING_FQ))
 439				smp_store_release(&sk->sk_pacing_status,
 440						  SK_PACING_FQ);
 441		}
 442		q->inactive_flows--;
 
 
 
 
 
 443	}
 444
 
 445	/* Note: this overwrites f->age */
 446	flow_queue_add(f, skb);
 447
 448	if (unlikely(f == &q->internal)) {
 449		q->stat_internal_packets++;
 450	}
 451	sch->q.qlen++;
 452
 453	return NET_XMIT_SUCCESS;
 454}
 455
 456static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 457{
 458	unsigned long sample;
 459	struct rb_node *p;
 460
 461	if (q->time_next_delayed_flow > now)
 462		return;
 463
 464	/* Update unthrottle latency EWMA.
 465	 * This is cheap and can help diagnosing timer/latency problems.
 466	 */
 467	sample = (unsigned long)(now - q->time_next_delayed_flow);
 468	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 469	q->unthrottle_latency_ns += sample >> 3;
 
 
 
 470
 471	q->time_next_delayed_flow = ~0ULL;
 472	while ((p = rb_first(&q->delayed)) != NULL) {
 473		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 474
 475		if (f->time_next_packet > now) {
 476			q->time_next_delayed_flow = f->time_next_packet;
 477			break;
 478		}
 479		fq_flow_unset_throttled(q, f);
 480	}
 481}
 482
 
 
 
 
 
 
 
 
 
 
 
 483static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 484{
 485	struct fq_sched_data *q = qdisc_priv(sch);
 
 486	struct fq_flow_head *head;
 487	struct sk_buff *skb;
 488	struct fq_flow *f;
 489	unsigned long rate;
 
 490	u32 plen;
 491	u64 now;
 492
 493	if (!sch->q.qlen)
 494		return NULL;
 495
 496	skb = fq_dequeue_head(sch, &q->internal);
 497	if (skb)
 
 
 498		goto out;
 
 499
 500	now = ktime_get_ns();
 501	fq_check_throttled(q, now);
 
 
 502begin:
 503	head = &q->new_flows;
 504	if (!head->first) {
 505		head = &q->old_flows;
 506		if (!head->first) {
 507			if (q->time_next_delayed_flow != ~0ULL)
 508				qdisc_watchdog_schedule_ns(&q->watchdog,
 509							   q->time_next_delayed_flow);
 510			return NULL;
 
 
 
 511		}
 
 
 
 
 
 512	}
 513	f = head->first;
 514
 515	if (f->credit <= 0) {
 516		f->credit += q->quantum;
 517		head->first = f->next;
 518		fq_flow_add_tail(&q->old_flows, f);
 519		goto begin;
 520	}
 521
 522	skb = fq_peek(f);
 523	if (skb) {
 524		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 525					     f->time_next_packet);
 526
 527		if (now < time_next_packet) {
 528			head->first = f->next;
 529			f->time_next_packet = time_next_packet;
 530			fq_flow_set_throttled(q, f);
 531			goto begin;
 532		}
 533		if (time_next_packet &&
 534		    (s64)(now - time_next_packet - q->ce_threshold) > 0) {
 535			INET_ECN_set_ce(skb);
 536			q->stat_ce_mark++;
 537		}
 538	}
 539
 540	skb = fq_dequeue_head(sch, f);
 541	if (!skb) {
 
 542		head->first = f->next;
 543		/* force a pass through old_flows to prevent starvation */
 544		if ((head == &q->new_flows) && q->old_flows.first) {
 545			fq_flow_add_tail(&q->old_flows, f);
 546		} else {
 547			fq_flow_set_detached(f);
 548			q->inactive_flows++;
 549		}
 550		goto begin;
 551	}
 552	prefetch(&skb->end);
 553	plen = qdisc_pkt_len(skb);
 554	f->credit -= plen;
 
 555
 556	if (!q->rate_enable)
 557		goto out;
 558
 559	rate = q->flow_max_rate;
 560
 561	/* If EDT time was provided for this skb, we need to
 562	 * update f->time_next_packet only if this qdisc enforces
 563	 * a flow max rate.
 564	 */
 565	if (!skb->tstamp) {
 566		if (skb->sk)
 567			rate = min(skb->sk->sk_pacing_rate, rate);
 568
 569		if (rate <= q->low_rate_threshold) {
 570			f->credit = 0;
 571		} else {
 572			plen = max(plen, q->quantum);
 573			if (f->credit > 0)
 574				goto out;
 575		}
 576	}
 577	if (rate != ~0UL) {
 578		u64 len = (u64)plen * NSEC_PER_SEC;
 579
 580		if (likely(rate))
 581			len = div64_ul(len, rate);
 582		/* Since socket rate can change later,
 583		 * clamp the delay to 1 second.
 584		 * Really, providers of too big packets should be fixed !
 585		 */
 586		if (unlikely(len > NSEC_PER_SEC)) {
 587			len = NSEC_PER_SEC;
 588			q->stat_pkts_too_long++;
 589		}
 590		/* Account for schedule/timers drifts.
 591		 * f->time_next_packet was set when prior packet was sent,
 592		 * and current time (@now) can be too late by tens of us.
 593		 */
 594		if (f->time_next_packet)
 595			len -= min(len/2, now - f->time_next_packet);
 596		f->time_next_packet = now + len;
 597	}
 598out:
 599	qdisc_bstats_update(sch, skb);
 600	return skb;
 601}
 602
 603static void fq_flow_purge(struct fq_flow *flow)
 604{
 605	struct rb_node *p = rb_first(&flow->t_root);
 606
 607	while (p) {
 608		struct sk_buff *skb = rb_to_skb(p);
 609
 610		p = rb_next(p);
 611		rb_erase(&skb->rbnode, &flow->t_root);
 612		rtnl_kfree_skbs(skb, skb);
 613	}
 614	rtnl_kfree_skbs(flow->head, flow->tail);
 615	flow->head = NULL;
 616	flow->qlen = 0;
 617}
 618
 619static void fq_reset(struct Qdisc *sch)
 620{
 621	struct fq_sched_data *q = qdisc_priv(sch);
 622	struct rb_root *root;
 623	struct rb_node *p;
 624	struct fq_flow *f;
 625	unsigned int idx;
 626
 627	sch->q.qlen = 0;
 628	sch->qstats.backlog = 0;
 629
 630	fq_flow_purge(&q->internal);
 631
 632	if (!q->fq_root)
 633		return;
 634
 635	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 636		root = &q->fq_root[idx];
 637		while ((p = rb_first(root)) != NULL) {
 638			f = rb_entry(p, struct fq_flow, fq_node);
 639			rb_erase(p, root);
 640
 641			fq_flow_purge(f);
 642
 643			kmem_cache_free(fq_flow_cachep, f);
 644		}
 645	}
 646	q->new_flows.first	= NULL;
 647	q->old_flows.first	= NULL;
 
 
 648	q->delayed		= RB_ROOT;
 649	q->flows		= 0;
 650	q->inactive_flows	= 0;
 651	q->throttled_flows	= 0;
 652}
 653
 654static void fq_rehash(struct fq_sched_data *q,
 655		      struct rb_root *old_array, u32 old_log,
 656		      struct rb_root *new_array, u32 new_log)
 657{
 658	struct rb_node *op, **np, *parent;
 659	struct rb_root *oroot, *nroot;
 660	struct fq_flow *of, *nf;
 661	int fcnt = 0;
 662	u32 idx;
 663
 664	for (idx = 0; idx < (1U << old_log); idx++) {
 665		oroot = &old_array[idx];
 666		while ((op = rb_first(oroot)) != NULL) {
 667			rb_erase(op, oroot);
 668			of = rb_entry(op, struct fq_flow, fq_node);
 669			if (fq_gc_candidate(of)) {
 670				fcnt++;
 671				kmem_cache_free(fq_flow_cachep, of);
 672				continue;
 673			}
 674			nroot = &new_array[hash_ptr(of->sk, new_log)];
 675
 676			np = &nroot->rb_node;
 677			parent = NULL;
 678			while (*np) {
 679				parent = *np;
 680
 681				nf = rb_entry(parent, struct fq_flow, fq_node);
 682				BUG_ON(nf->sk == of->sk);
 683
 684				if (nf->sk > of->sk)
 685					np = &parent->rb_right;
 686				else
 687					np = &parent->rb_left;
 688			}
 689
 690			rb_link_node(&of->fq_node, parent, np);
 691			rb_insert_color(&of->fq_node, nroot);
 692		}
 693	}
 694	q->flows -= fcnt;
 695	q->inactive_flows -= fcnt;
 696	q->stat_gc_flows += fcnt;
 697}
 698
 699static void fq_free(void *addr)
 700{
 701	kvfree(addr);
 702}
 703
 704static int fq_resize(struct Qdisc *sch, u32 log)
 705{
 706	struct fq_sched_data *q = qdisc_priv(sch);
 707	struct rb_root *array;
 708	void *old_fq_root;
 709	u32 idx;
 710
 711	if (q->fq_root && log == q->fq_trees_log)
 712		return 0;
 713
 714	/* If XPS was setup, we can allocate memory on right NUMA node */
 715	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 716			      netdev_queue_numa_node_read(sch->dev_queue));
 717	if (!array)
 718		return -ENOMEM;
 719
 720	for (idx = 0; idx < (1U << log); idx++)
 721		array[idx] = RB_ROOT;
 722
 723	sch_tree_lock(sch);
 724
 725	old_fq_root = q->fq_root;
 726	if (old_fq_root)
 727		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 728
 729	q->fq_root = array;
 730	q->fq_trees_log = log;
 731
 732	sch_tree_unlock(sch);
 733
 734	fq_free(old_fq_root);
 735
 736	return 0;
 737}
 738
 
 
 
 
 739static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 
 
 740	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
 741	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
 742	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
 743	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
 744	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
 745	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
 746	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
 747	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
 748	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
 
 749	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
 750	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
 
 
 
 
 
 
 751};
 752
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 753static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 754		     struct netlink_ext_ack *extack)
 755{
 756	struct fq_sched_data *q = qdisc_priv(sch);
 757	struct nlattr *tb[TCA_FQ_MAX + 1];
 758	int err, drop_count = 0;
 759	unsigned drop_len = 0;
 760	u32 fq_log;
 761
 762	if (!opt)
 763		return -EINVAL;
 764
 765	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
 766					  NULL);
 767	if (err < 0)
 768		return err;
 769
 770	sch_tree_lock(sch);
 771
 772	fq_log = q->fq_trees_log;
 773
 774	if (tb[TCA_FQ_BUCKETS_LOG]) {
 775		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 776
 777		if (nval >= 1 && nval <= ilog2(256*1024))
 778			fq_log = nval;
 779		else
 780			err = -EINVAL;
 781	}
 782	if (tb[TCA_FQ_PLIMIT])
 783		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 
 784
 785	if (tb[TCA_FQ_FLOW_PLIMIT])
 786		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 
 787
 788	if (tb[TCA_FQ_QUANTUM]) {
 789		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 790
 791		if (quantum > 0)
 792			q->quantum = quantum;
 793		else
 
 794			err = -EINVAL;
 
 795	}
 796
 797	if (tb[TCA_FQ_INITIAL_QUANTUM])
 798		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 
 799
 800	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 801		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 802				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 803
 804	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
 805		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 806
 807		q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
 
 808	}
 809	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
 810		q->low_rate_threshold =
 811			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
 812
 813	if (tb[TCA_FQ_RATE_ENABLE]) {
 814		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 815
 816		if (enable <= 1)
 817			q->rate_enable = enable;
 
 818		else
 819			err = -EINVAL;
 820	}
 821
 822	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 823		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 824
 825		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 
 826	}
 827
 
 
 
 
 
 
 828	if (tb[TCA_FQ_ORPHAN_MASK])
 829		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 
 830
 831	if (tb[TCA_FQ_CE_THRESHOLD])
 832		q->ce_threshold = (u64)NSEC_PER_USEC *
 833				  nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 834
 
 
 
 
 
 
 
 835	if (!err) {
 
 836		sch_tree_unlock(sch);
 837		err = fq_resize(sch, fq_log);
 838		sch_tree_lock(sch);
 839	}
 840	while (sch->q.qlen > sch->limit) {
 841		struct sk_buff *skb = fq_dequeue(sch);
 842
 843		if (!skb)
 844			break;
 845		drop_len += qdisc_pkt_len(skb);
 846		rtnl_kfree_skbs(skb, skb);
 847		drop_count++;
 848	}
 849	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 850
 851	sch_tree_unlock(sch);
 852	return err;
 853}
 854
 855static void fq_destroy(struct Qdisc *sch)
 856{
 857	struct fq_sched_data *q = qdisc_priv(sch);
 858
 859	fq_reset(sch);
 860	fq_free(q->fq_root);
 861	qdisc_watchdog_cancel(&q->watchdog);
 862}
 863
 864static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 865		   struct netlink_ext_ack *extack)
 866{
 867	struct fq_sched_data *q = qdisc_priv(sch);
 868	int err;
 869
 870	sch->limit		= 10000;
 871	q->flow_plimit		= 100;
 872	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
 873	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
 874	q->flow_refill_delay	= msecs_to_jiffies(40);
 875	q->flow_max_rate	= ~0UL;
 876	q->time_next_delayed_flow = ~0ULL;
 877	q->rate_enable		= 1;
 878	q->new_flows.first	= NULL;
 879	q->old_flows.first	= NULL;
 
 
 
 
 
 880	q->delayed		= RB_ROOT;
 881	q->fq_root		= NULL;
 882	q->fq_trees_log		= ilog2(1024);
 883	q->orphan_mask		= 1024 - 1;
 884	q->low_rate_threshold	= 550000 / 8;
 885
 
 
 
 
 
 886	/* Default ce_threshold of 4294 seconds */
 887	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
 888
 
 889	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
 890
 891	if (opt)
 892		err = fq_change(sch, opt, extack);
 893	else
 894		err = fq_resize(sch, q->fq_trees_log);
 895
 896	return err;
 897}
 898
 899static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 900{
 901	struct fq_sched_data *q = qdisc_priv(sch);
 902	u64 ce_threshold = q->ce_threshold;
 
 
 903	struct nlattr *opts;
 
 
 
 
 904
 905	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
 906	if (opts == NULL)
 907		goto nla_put_failure;
 908
 909	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 910
 
 911	do_div(ce_threshold, NSEC_PER_USEC);
 912
 913	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 914	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 915	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 916	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 917	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 
 
 
 
 
 
 
 
 
 
 
 918	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
 919			min_t(unsigned long, q->flow_max_rate, ~0U)) ||
 
 920	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 921			jiffies_to_usecs(q->flow_refill_delay)) ||
 922	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 
 923	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
 924			q->low_rate_threshold) ||
 925	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
 926	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 927		goto nla_put_failure;
 928
 929	return nla_nest_end(skb, opts);
 930
 931nla_put_failure:
 932	return -1;
 933}
 934
 935static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 936{
 937	struct fq_sched_data *q = qdisc_priv(sch);
 938	struct tc_fq_qd_stats st;
 
 
 
 939
 940	sch_tree_lock(sch);
 941
 942	st.gc_flows		  = q->stat_gc_flows;
 943	st.highprio_packets	  = q->stat_internal_packets;
 
 944	st.tcp_retrans		  = 0;
 945	st.throttled		  = q->stat_throttled;
 946	st.flows_plimit		  = q->stat_flows_plimit;
 947	st.pkts_too_long	  = q->stat_pkts_too_long;
 948	st.allocation_errors	  = q->stat_allocation_errors;
 949	st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
 
 950	st.flows		  = q->flows;
 951	st.inactive_flows	  = q->inactive_flows;
 952	st.throttled_flows	  = q->throttled_flows;
 953	st.unthrottle_latency_ns  = min_t(unsigned long,
 954					  q->unthrottle_latency_ns, ~0U);
 955	st.ce_mark		  = q->stat_ce_mark;
 
 
 
 
 
 
 956	sch_tree_unlock(sch);
 957
 958	return gnet_stats_copy_app(d, &st, sizeof(st));
 959}
 960
 961static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 962	.id		=	"fq",
 963	.priv_size	=	sizeof(struct fq_sched_data),
 964
 965	.enqueue	=	fq_enqueue,
 966	.dequeue	=	fq_dequeue,
 967	.peek		=	qdisc_peek_dequeued,
 968	.init		=	fq_init,
 969	.reset		=	fq_reset,
 970	.destroy	=	fq_destroy,
 971	.change		=	fq_change,
 972	.dump		=	fq_dump,
 973	.dump_stats	=	fq_dump_stats,
 974	.owner		=	THIS_MODULE,
 975};
 
 976
 977static int __init fq_module_init(void)
 978{
 979	int ret;
 980
 981	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 982					   sizeof(struct fq_flow),
 983					   0, 0, NULL);
 984	if (!fq_flow_cachep)
 985		return -ENOMEM;
 986
 987	ret = register_qdisc(&fq_qdisc_ops);
 988	if (ret)
 989		kmem_cache_destroy(fq_flow_cachep);
 990	return ret;
 991}
 992
 993static void __exit fq_module_exit(void)
 994{
 995	unregister_qdisc(&fq_qdisc_ops);
 996	kmem_cache_destroy(fq_flow_cachep);
 997}
 998
 999module_init(fq_module_init)
1000module_exit(fq_module_exit)
1001MODULE_AUTHOR("Eric Dumazet");
1002MODULE_LICENSE("GPL");