xref: /linux/net/sched/sch_fq.c (revision 05c1280a2bcfca187fe7fa90bb240602cf54af0a)
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 
53 struct fq_skb_cb {
54 	u64	time_to_send;
55 	u8	band;
56 };
57 
58 static 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  */
69 struct 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 
97 struct fq_flow_head {
98 	struct fq_flow *first;
99 	struct fq_flow *last;
100 };
101 
102 struct 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 
111 struct fq_sched_data {
112 /* Read mostly cache line */
113 
114 	u32		quantum;
115 	u32		initial_quantum;
116 	u32		flow_refill_delay;
117 	u32		flow_plimit;	/* max packets per flow */
118 	unsigned long	flow_max_rate;	/* optional max rate per flow */
119 	u64		ce_threshold;
120 	u64		horizon;	/* horizon in ns */
121 	u32		orphan_mask;	/* mask for orphaned skb */
122 	u32		low_rate_threshold;
123 	struct rb_root	*fq_root;
124 	u8		rate_enable;
125 	u8		fq_trees_log;
126 	u8		horizon_drop;
127 	u8		prio2band[FQ_PRIO2BAND_CRUMB_SIZE];
128 	u32		timer_slack; /* hrtimer slack in ns */
129 
130 /* Read/Write fields. */
131 
132 	unsigned int band_nr; /* band being serviced in fq_dequeue() */
133 
134 	struct fq_perband_flows band_flows[FQ_BANDS];
135 
136 	struct fq_flow	internal;	/* fastpath queue. */
137 	struct rb_root	delayed;	/* for rate limited flows */
138 	u64		time_next_delayed_flow;
139 	unsigned long	unthrottle_latency_ns;
140 
141 	u32		band_pkt_count[FQ_BANDS];
142 	u32		flows;
143 	u32		inactive_flows; /* Flows with no packet to send. */
144 	u32		throttled_flows;
145 
146 	u64		stat_throttled;
147 	struct qdisc_watchdog watchdog;
148 	u64		stat_gc_flows;
149 
150 /* Seldom used fields. */
151 
152 	u64		stat_band_drops[FQ_BANDS];
153 	u64		stat_ce_mark;
154 	u64		stat_horizon_drops;
155 	u64		stat_horizon_caps;
156 	u64		stat_flows_plimit;
157 	u64		stat_pkts_too_long;
158 	u64		stat_allocation_errors;
159 };
160 
161 /* return the i-th 2-bit value ("crumb") */
162 static u8 fq_prio2band(const u8 *prio2band, unsigned int prio)
163 {
164 	return (READ_ONCE(prio2band[prio / 4]) >> (2 * (prio & 0x3))) & 0x3;
165 }
166 
167 /*
168  * f->tail and f->age share the same location.
169  * We can use the low order bit to differentiate if this location points
170  * to a sk_buff or contains a jiffies value, if we force this value to be odd.
171  * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
172  */
173 static void fq_flow_set_detached(struct fq_flow *f)
174 {
175 	f->age = jiffies | 1UL;
176 }
177 
178 static bool fq_flow_is_detached(const struct fq_flow *f)
179 {
180 	return !!(f->age & 1UL);
181 }
182 
183 /* special value to mark a throttled flow (not on old/new list) */
184 static struct fq_flow throttled;
185 
186 static bool fq_flow_is_throttled(const struct fq_flow *f)
187 {
188 	return f->next == &throttled;
189 }
190 
191 enum new_flow {
192 	NEW_FLOW,
193 	OLD_FLOW
194 };
195 
196 static void fq_flow_add_tail(struct fq_sched_data *q, struct fq_flow *flow,
197 			     enum new_flow list_sel)
198 {
199 	struct fq_perband_flows *pband = &q->band_flows[flow->band];
200 	struct fq_flow_head *head = (list_sel == NEW_FLOW) ?
201 					&pband->new_flows :
202 					&pband->old_flows;
203 
204 	if (head->first)
205 		head->last->next = flow;
206 	else
207 		head->first = flow;
208 	head->last = flow;
209 	flow->next = NULL;
210 }
211 
212 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
213 {
214 	rb_erase(&f->rate_node, &q->delayed);
215 	q->throttled_flows--;
216 	fq_flow_add_tail(q, f, OLD_FLOW);
217 }
218 
219 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
220 {
221 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
222 
223 	while (*p) {
224 		struct fq_flow *aux;
225 
226 		parent = *p;
227 		aux = rb_entry(parent, struct fq_flow, rate_node);
228 		if (f->time_next_packet >= aux->time_next_packet)
229 			p = &parent->rb_right;
230 		else
231 			p = &parent->rb_left;
232 	}
233 	rb_link_node(&f->rate_node, parent, p);
234 	rb_insert_color(&f->rate_node, &q->delayed);
235 	q->throttled_flows++;
236 	q->stat_throttled++;
237 
238 	f->next = &throttled;
239 	if (q->time_next_delayed_flow > f->time_next_packet)
240 		q->time_next_delayed_flow = f->time_next_packet;
241 }
242 
243 
244 static struct kmem_cache *fq_flow_cachep __read_mostly;
245 
246 
247 /* limit number of collected flows per round */
248 #define FQ_GC_MAX 8
249 #define FQ_GC_AGE (3*HZ)
250 
251 static bool fq_gc_candidate(const struct fq_flow *f)
252 {
253 	return fq_flow_is_detached(f) &&
254 	       time_after(jiffies, f->age + FQ_GC_AGE);
255 }
256 
257 static void fq_gc(struct fq_sched_data *q,
258 		  struct rb_root *root,
259 		  struct sock *sk)
260 {
261 	struct rb_node **p, *parent;
262 	void *tofree[FQ_GC_MAX];
263 	struct fq_flow *f;
264 	int i, fcnt = 0;
265 
266 	p = &root->rb_node;
267 	parent = NULL;
268 	while (*p) {
269 		parent = *p;
270 
271 		f = rb_entry(parent, struct fq_flow, fq_node);
272 		if (f->sk == sk)
273 			break;
274 
275 		if (fq_gc_candidate(f)) {
276 			tofree[fcnt++] = f;
277 			if (fcnt == FQ_GC_MAX)
278 				break;
279 		}
280 
281 		if (f->sk > sk)
282 			p = &parent->rb_right;
283 		else
284 			p = &parent->rb_left;
285 	}
286 
287 	if (!fcnt)
288 		return;
289 
290 	for (i = fcnt; i > 0; ) {
291 		f = tofree[--i];
292 		rb_erase(&f->fq_node, root);
293 	}
294 	q->flows -= fcnt;
295 	q->inactive_flows -= fcnt;
296 	q->stat_gc_flows += fcnt;
297 
298 	kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
299 }
300 
301 /* Fast path can be used if :
302  * 1) Packet tstamp is in the past.
303  * 2) FQ qlen == 0   OR
304  *   (no flow is currently eligible for transmit,
305  *    AND fast path queue has less than 8 packets)
306  * 3) No SO_MAX_PACING_RATE on the socket (if any).
307  * 4) No @maxrate attribute on this qdisc,
308  *
309  * FQ can not use generic TCQ_F_CAN_BYPASS infrastructure.
310  */
311 static bool fq_fastpath_check(const struct Qdisc *sch, struct sk_buff *skb,
312 			      u64 now)
313 {
314 	const struct fq_sched_data *q = qdisc_priv(sch);
315 	const struct sock *sk;
316 
317 	if (fq_skb_cb(skb)->time_to_send > now)
318 		return false;
319 
320 	if (sch->q.qlen != 0) {
321 		/* Even if some packets are stored in this qdisc,
322 		 * we can still enable fast path if all of them are
323 		 * scheduled in the future (ie no flows are eligible)
324 		 * or in the fast path queue.
325 		 */
326 		if (q->flows != q->inactive_flows + q->throttled_flows)
327 			return false;
328 
329 		/* Do not allow fast path queue to explode, we want Fair Queue mode
330 		 * under pressure.
331 		 */
332 		if (q->internal.qlen >= 8)
333 			return false;
334 	}
335 
336 	sk = skb->sk;
337 	if (sk && sk_fullsock(sk) && !sk_is_tcp(sk) &&
338 	    sk->sk_max_pacing_rate != ~0UL)
339 		return false;
340 
341 	if (q->flow_max_rate != ~0UL)
342 		return false;
343 
344 	return true;
345 }
346 
347 static struct fq_flow *fq_classify(struct Qdisc *sch, struct sk_buff *skb,
348 				   u64 now)
349 {
350 	struct fq_sched_data *q = qdisc_priv(sch);
351 	struct rb_node **p, *parent;
352 	struct sock *sk = skb->sk;
353 	struct rb_root *root;
354 	struct fq_flow *f;
355 
356 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
357 	 * or a listener (SYNCOOKIE mode)
358 	 * 1) request sockets are not full blown,
359 	 *    they do not contain sk_pacing_rate
360 	 * 2) They are not part of a 'flow' yet
361 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
362 	 *    especially if the listener set SO_MAX_PACING_RATE
363 	 * 4) We pretend they are orphaned
364 	 */
365 	if (!sk || sk_listener(sk)) {
366 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
367 
368 		/* By forcing low order bit to 1, we make sure to not
369 		 * collide with a local flow (socket pointers are word aligned)
370 		 */
371 		sk = (struct sock *)((hash << 1) | 1UL);
372 		skb_orphan(skb);
373 	} else if (sk->sk_state == TCP_CLOSE) {
374 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
375 		/*
376 		 * Sockets in TCP_CLOSE are non connected.
377 		 * Typical use case is UDP sockets, they can send packets
378 		 * with sendto() to many different destinations.
379 		 * We probably could use a generic bit advertising
380 		 * non connected sockets, instead of sk_state == TCP_CLOSE,
381 		 * if we care enough.
382 		 */
383 		sk = (struct sock *)((hash << 1) | 1UL);
384 	}
385 
386 	if (fq_fastpath_check(sch, skb, now)) {
387 		q->internal.stat_fastpath_packets++;
388 		if (skb->sk == sk && q->rate_enable &&
389 		    READ_ONCE(sk->sk_pacing_status) != SK_PACING_FQ)
390 			smp_store_release(&sk->sk_pacing_status,
391 					  SK_PACING_FQ);
392 		return &q->internal;
393 	}
394 
395 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
396 
397 	fq_gc(q, root, sk);
398 
399 	p = &root->rb_node;
400 	parent = NULL;
401 	while (*p) {
402 		parent = *p;
403 
404 		f = rb_entry(parent, struct fq_flow, fq_node);
405 		if (f->sk == sk) {
406 			/* socket might have been reallocated, so check
407 			 * if its sk_hash is the same.
408 			 * It not, we need to refill credit with
409 			 * initial quantum
410 			 */
411 			if (unlikely(skb->sk == sk &&
412 				     f->socket_hash != sk->sk_hash)) {
413 				f->credit = q->initial_quantum;
414 				f->socket_hash = sk->sk_hash;
415 				if (q->rate_enable)
416 					smp_store_release(&sk->sk_pacing_status,
417 							  SK_PACING_FQ);
418 				if (fq_flow_is_throttled(f))
419 					fq_flow_unset_throttled(q, f);
420 				f->time_next_packet = 0ULL;
421 			}
422 			return f;
423 		}
424 		if (f->sk > sk)
425 			p = &parent->rb_right;
426 		else
427 			p = &parent->rb_left;
428 	}
429 
430 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
431 	if (unlikely(!f)) {
432 		q->stat_allocation_errors++;
433 		return &q->internal;
434 	}
435 	/* f->t_root is already zeroed after kmem_cache_zalloc() */
436 
437 	fq_flow_set_detached(f);
438 	f->sk = sk;
439 	if (skb->sk == sk) {
440 		f->socket_hash = sk->sk_hash;
441 		if (q->rate_enable)
442 			smp_store_release(&sk->sk_pacing_status,
443 					  SK_PACING_FQ);
444 	}
445 	f->credit = q->initial_quantum;
446 
447 	rb_link_node(&f->fq_node, parent, p);
448 	rb_insert_color(&f->fq_node, root);
449 
450 	q->flows++;
451 	q->inactive_flows++;
452 	return f;
453 }
454 
455 static struct sk_buff *fq_peek(struct fq_flow *flow)
456 {
457 	struct sk_buff *skb = skb_rb_first(&flow->t_root);
458 	struct sk_buff *head = flow->head;
459 
460 	if (!skb)
461 		return head;
462 
463 	if (!head)
464 		return skb;
465 
466 	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
467 		return skb;
468 	return head;
469 }
470 
471 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
472 			  struct sk_buff *skb)
473 {
474 	if (skb == flow->head) {
475 		flow->head = skb->next;
476 	} else {
477 		rb_erase(&skb->rbnode, &flow->t_root);
478 		skb->dev = qdisc_dev(sch);
479 	}
480 }
481 
482 /* Remove one skb from flow queue.
483  * This skb must be the return value of prior fq_peek().
484  */
485 static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
486 			   struct sk_buff *skb)
487 {
488 	fq_erase_head(sch, flow, skb);
489 	skb_mark_not_on_list(skb);
490 	qdisc_qstats_backlog_dec(sch, skb);
491 	sch->q.qlen--;
492 }
493 
494 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
495 {
496 	struct rb_node **p, *parent;
497 	struct sk_buff *head, *aux;
498 
499 	head = flow->head;
500 	if (!head ||
501 	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
502 		if (!head)
503 			flow->head = skb;
504 		else
505 			flow->tail->next = skb;
506 		flow->tail = skb;
507 		skb->next = NULL;
508 		return;
509 	}
510 
511 	p = &flow->t_root.rb_node;
512 	parent = NULL;
513 
514 	while (*p) {
515 		parent = *p;
516 		aux = rb_to_skb(parent);
517 		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
518 			p = &parent->rb_right;
519 		else
520 			p = &parent->rb_left;
521 	}
522 	rb_link_node(&skb->rbnode, parent, p);
523 	rb_insert_color(&skb->rbnode, &flow->t_root);
524 }
525 
526 static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
527 				     const struct fq_sched_data *q, u64 now)
528 {
529 	return unlikely((s64)skb->tstamp > (s64)(now + q->horizon));
530 }
531 
532 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
533 		      struct sk_buff **to_free)
534 {
535 	struct fq_sched_data *q = qdisc_priv(sch);
536 	struct fq_flow *f;
537 	u64 now;
538 	u8 band;
539 
540 	band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
541 	if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
542 		q->stat_band_drops[band]++;
543 		return qdisc_drop(skb, sch, to_free);
544 	}
545 
546 	now = ktime_get_ns();
547 	if (!skb->tstamp) {
548 		fq_skb_cb(skb)->time_to_send = now;
549 	} else {
550 		/* Check if packet timestamp is too far in the future. */
551 		if (fq_packet_beyond_horizon(skb, q, now)) {
552 			if (q->horizon_drop) {
553 					q->stat_horizon_drops++;
554 					return qdisc_drop(skb, sch, to_free);
555 			}
556 			q->stat_horizon_caps++;
557 			skb->tstamp = now + q->horizon;
558 		}
559 		fq_skb_cb(skb)->time_to_send = skb->tstamp;
560 	}
561 
562 	f = fq_classify(sch, skb, now);
563 
564 	if (f != &q->internal) {
565 		if (unlikely(f->qlen >= q->flow_plimit)) {
566 			q->stat_flows_plimit++;
567 			return qdisc_drop(skb, sch, to_free);
568 		}
569 
570 		if (fq_flow_is_detached(f)) {
571 			fq_flow_add_tail(q, f, NEW_FLOW);
572 			if (time_after(jiffies, f->age + q->flow_refill_delay))
573 				f->credit = max_t(u32, f->credit, q->quantum);
574 		}
575 
576 		f->band = band;
577 		q->band_pkt_count[band]++;
578 		fq_skb_cb(skb)->band = band;
579 		if (f->qlen == 0)
580 			q->inactive_flows--;
581 	}
582 
583 	f->qlen++;
584 	/* Note: this overwrites f->age */
585 	flow_queue_add(f, skb);
586 
587 	qdisc_qstats_backlog_inc(sch, skb);
588 	sch->q.qlen++;
589 
590 	return NET_XMIT_SUCCESS;
591 }
592 
593 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
594 {
595 	unsigned long sample;
596 	struct rb_node *p;
597 
598 	if (q->time_next_delayed_flow > now)
599 		return;
600 
601 	/* Update unthrottle latency EWMA.
602 	 * This is cheap and can help diagnosing timer/latency problems.
603 	 */
604 	sample = (unsigned long)(now - q->time_next_delayed_flow);
605 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
606 	q->unthrottle_latency_ns += sample >> 3;
607 
608 	q->time_next_delayed_flow = ~0ULL;
609 	while ((p = rb_first(&q->delayed)) != NULL) {
610 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
611 
612 		if (f->time_next_packet > now) {
613 			q->time_next_delayed_flow = f->time_next_packet;
614 			break;
615 		}
616 		fq_flow_unset_throttled(q, f);
617 	}
618 }
619 
620 static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
621 {
622 	if (pband->credit <= 0)
623 		return NULL;
624 
625 	if (pband->new_flows.first)
626 		return &pband->new_flows;
627 
628 	return pband->old_flows.first ? &pband->old_flows : NULL;
629 }
630 
631 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
632 {
633 	struct fq_sched_data *q = qdisc_priv(sch);
634 	struct fq_perband_flows *pband;
635 	struct fq_flow_head *head;
636 	struct sk_buff *skb;
637 	struct fq_flow *f;
638 	unsigned long rate;
639 	int retry;
640 	u32 plen;
641 	u64 now;
642 
643 	if (!sch->q.qlen)
644 		return NULL;
645 
646 	skb = fq_peek(&q->internal);
647 	if (unlikely(skb)) {
648 		q->internal.qlen--;
649 		fq_dequeue_skb(sch, &q->internal, skb);
650 		goto out;
651 	}
652 
653 	now = ktime_get_ns();
654 	fq_check_throttled(q, now);
655 	retry = 0;
656 	pband = &q->band_flows[q->band_nr];
657 begin:
658 	head = fq_pband_head_select(pband);
659 	if (!head) {
660 		while (++retry <= FQ_BANDS) {
661 			if (++q->band_nr == FQ_BANDS)
662 				q->band_nr = 0;
663 			pband = &q->band_flows[q->band_nr];
664 			pband->credit = min(pband->credit + pband->quantum,
665 					    pband->quantum);
666 			if (pband->credit > 0)
667 				goto begin;
668 			retry = 0;
669 		}
670 		if (q->time_next_delayed_flow != ~0ULL)
671 			qdisc_watchdog_schedule_range_ns(&q->watchdog,
672 							q->time_next_delayed_flow,
673 							q->timer_slack);
674 		return NULL;
675 	}
676 	f = head->first;
677 	retry = 0;
678 	if (f->credit <= 0) {
679 		f->credit += q->quantum;
680 		head->first = f->next;
681 		fq_flow_add_tail(q, f, OLD_FLOW);
682 		goto begin;
683 	}
684 
685 	skb = fq_peek(f);
686 	if (skb) {
687 		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
688 					     f->time_next_packet);
689 
690 		if (now < time_next_packet) {
691 			head->first = f->next;
692 			f->time_next_packet = time_next_packet;
693 			fq_flow_set_throttled(q, f);
694 			goto begin;
695 		}
696 		prefetch(&skb->end);
697 		if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
698 			INET_ECN_set_ce(skb);
699 			q->stat_ce_mark++;
700 		}
701 		if (--f->qlen == 0)
702 			q->inactive_flows++;
703 		q->band_pkt_count[fq_skb_cb(skb)->band]--;
704 		fq_dequeue_skb(sch, f, skb);
705 	} else {
706 		head->first = f->next;
707 		/* force a pass through old_flows to prevent starvation */
708 		if (head == &pband->new_flows) {
709 			fq_flow_add_tail(q, f, OLD_FLOW);
710 		} else {
711 			fq_flow_set_detached(f);
712 		}
713 		goto begin;
714 	}
715 	plen = qdisc_pkt_len(skb);
716 	f->credit -= plen;
717 	pband->credit -= plen;
718 
719 	if (!q->rate_enable)
720 		goto out;
721 
722 	rate = q->flow_max_rate;
723 
724 	/* If EDT time was provided for this skb, we need to
725 	 * update f->time_next_packet only if this qdisc enforces
726 	 * a flow max rate.
727 	 */
728 	if (!skb->tstamp) {
729 		if (skb->sk)
730 			rate = min(READ_ONCE(skb->sk->sk_pacing_rate), rate);
731 
732 		if (rate <= q->low_rate_threshold) {
733 			f->credit = 0;
734 		} else {
735 			plen = max(plen, q->quantum);
736 			if (f->credit > 0)
737 				goto out;
738 		}
739 	}
740 	if (rate != ~0UL) {
741 		u64 len = (u64)plen * NSEC_PER_SEC;
742 
743 		if (likely(rate))
744 			len = div64_ul(len, rate);
745 		/* Since socket rate can change later,
746 		 * clamp the delay to 1 second.
747 		 * Really, providers of too big packets should be fixed !
748 		 */
749 		if (unlikely(len > NSEC_PER_SEC)) {
750 			len = NSEC_PER_SEC;
751 			q->stat_pkts_too_long++;
752 		}
753 		/* Account for schedule/timers drifts.
754 		 * f->time_next_packet was set when prior packet was sent,
755 		 * and current time (@now) can be too late by tens of us.
756 		 */
757 		if (f->time_next_packet)
758 			len -= min(len/2, now - f->time_next_packet);
759 		f->time_next_packet = now + len;
760 	}
761 out:
762 	qdisc_bstats_update(sch, skb);
763 	return skb;
764 }
765 
766 static void fq_flow_purge(struct fq_flow *flow)
767 {
768 	struct rb_node *p = rb_first(&flow->t_root);
769 
770 	while (p) {
771 		struct sk_buff *skb = rb_to_skb(p);
772 
773 		p = rb_next(p);
774 		rb_erase(&skb->rbnode, &flow->t_root);
775 		rtnl_kfree_skbs(skb, skb);
776 	}
777 	rtnl_kfree_skbs(flow->head, flow->tail);
778 	flow->head = NULL;
779 	flow->qlen = 0;
780 }
781 
782 static void fq_reset(struct Qdisc *sch)
783 {
784 	struct fq_sched_data *q = qdisc_priv(sch);
785 	struct rb_root *root;
786 	struct rb_node *p;
787 	struct fq_flow *f;
788 	unsigned int idx;
789 
790 	sch->q.qlen = 0;
791 	sch->qstats.backlog = 0;
792 
793 	fq_flow_purge(&q->internal);
794 
795 	if (!q->fq_root)
796 		return;
797 
798 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
799 		root = &q->fq_root[idx];
800 		while ((p = rb_first(root)) != NULL) {
801 			f = rb_entry(p, struct fq_flow, fq_node);
802 			rb_erase(p, root);
803 
804 			fq_flow_purge(f);
805 
806 			kmem_cache_free(fq_flow_cachep, f);
807 		}
808 	}
809 	for (idx = 0; idx < FQ_BANDS; idx++) {
810 		q->band_flows[idx].new_flows.first = NULL;
811 		q->band_flows[idx].old_flows.first = NULL;
812 	}
813 	q->delayed		= RB_ROOT;
814 	q->flows		= 0;
815 	q->inactive_flows	= 0;
816 	q->throttled_flows	= 0;
817 }
818 
819 static void fq_rehash(struct fq_sched_data *q,
820 		      struct rb_root *old_array, u32 old_log,
821 		      struct rb_root *new_array, u32 new_log)
822 {
823 	struct rb_node *op, **np, *parent;
824 	struct rb_root *oroot, *nroot;
825 	struct fq_flow *of, *nf;
826 	int fcnt = 0;
827 	u32 idx;
828 
829 	for (idx = 0; idx < (1U << old_log); idx++) {
830 		oroot = &old_array[idx];
831 		while ((op = rb_first(oroot)) != NULL) {
832 			rb_erase(op, oroot);
833 			of = rb_entry(op, struct fq_flow, fq_node);
834 			if (fq_gc_candidate(of)) {
835 				fcnt++;
836 				kmem_cache_free(fq_flow_cachep, of);
837 				continue;
838 			}
839 			nroot = &new_array[hash_ptr(of->sk, new_log)];
840 
841 			np = &nroot->rb_node;
842 			parent = NULL;
843 			while (*np) {
844 				parent = *np;
845 
846 				nf = rb_entry(parent, struct fq_flow, fq_node);
847 				BUG_ON(nf->sk == of->sk);
848 
849 				if (nf->sk > of->sk)
850 					np = &parent->rb_right;
851 				else
852 					np = &parent->rb_left;
853 			}
854 
855 			rb_link_node(&of->fq_node, parent, np);
856 			rb_insert_color(&of->fq_node, nroot);
857 		}
858 	}
859 	q->flows -= fcnt;
860 	q->inactive_flows -= fcnt;
861 	q->stat_gc_flows += fcnt;
862 }
863 
864 static void fq_free(void *addr)
865 {
866 	kvfree(addr);
867 }
868 
869 static int fq_resize(struct Qdisc *sch, u32 log)
870 {
871 	struct fq_sched_data *q = qdisc_priv(sch);
872 	struct rb_root *array;
873 	void *old_fq_root;
874 	u32 idx;
875 
876 	if (q->fq_root && log == q->fq_trees_log)
877 		return 0;
878 
879 	/* If XPS was setup, we can allocate memory on right NUMA node */
880 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
881 			      netdev_queue_numa_node_read(sch->dev_queue));
882 	if (!array)
883 		return -ENOMEM;
884 
885 	for (idx = 0; idx < (1U << log); idx++)
886 		array[idx] = RB_ROOT;
887 
888 	sch_tree_lock(sch);
889 
890 	old_fq_root = q->fq_root;
891 	if (old_fq_root)
892 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
893 
894 	q->fq_root = array;
895 	WRITE_ONCE(q->fq_trees_log, log);
896 
897 	sch_tree_unlock(sch);
898 
899 	fq_free(old_fq_root);
900 
901 	return 0;
902 }
903 
904 static const struct netlink_range_validation iq_range = {
905 	.max = INT_MAX,
906 };
907 
908 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
909 	[TCA_FQ_UNSPEC]			= { .strict_start_type = TCA_FQ_TIMER_SLACK },
910 
911 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
912 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
913 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
914 	[TCA_FQ_INITIAL_QUANTUM]	= NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
915 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
916 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
917 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
918 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
919 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
920 	[TCA_FQ_ORPHAN_MASK]		= { .type = NLA_U32 },
921 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
922 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
923 	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
924 	[TCA_FQ_HORIZON]		= { .type = NLA_U32 },
925 	[TCA_FQ_HORIZON_DROP]		= { .type = NLA_U8 },
926 	[TCA_FQ_PRIOMAP]		= NLA_POLICY_EXACT_LEN(sizeof(struct tc_prio_qopt)),
927 	[TCA_FQ_WEIGHTS]		= NLA_POLICY_EXACT_LEN(FQ_BANDS * sizeof(s32)),
928 };
929 
930 /* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
931 static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
932 {
933 	const int num_elems = TC_PRIO_MAX + 1;
934 	u8 tmp[FQ_PRIO2BAND_CRUMB_SIZE];
935 	int i;
936 
937 	memset(tmp, 0, sizeof(tmp));
938 	for (i = 0; i < num_elems; i++)
939 		tmp[i / 4] |= in[i] << (2 * (i & 0x3));
940 
941 	for (i = 0; i < FQ_PRIO2BAND_CRUMB_SIZE; i++)
942 		WRITE_ONCE(out[i], tmp[i]);
943 }
944 
945 static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
946 {
947 	const int num_elems = TC_PRIO_MAX + 1;
948 	int i;
949 
950 	for (i = 0; i < num_elems; i++)
951 		out[i] = fq_prio2band(in, i);
952 }
953 
954 static int fq_load_weights(struct fq_sched_data *q,
955 			   const struct nlattr *attr,
956 			   struct netlink_ext_ack *extack)
957 {
958 	s32 *weights = nla_data(attr);
959 	int i;
960 
961 	for (i = 0; i < FQ_BANDS; i++) {
962 		if (weights[i] < FQ_MIN_WEIGHT) {
963 			NL_SET_ERR_MSG_FMT_MOD(extack, "Weight %d less that minimum allowed %d",
964 					       weights[i], FQ_MIN_WEIGHT);
965 			return -EINVAL;
966 		}
967 	}
968 	for (i = 0; i < FQ_BANDS; i++)
969 		WRITE_ONCE(q->band_flows[i].quantum, weights[i]);
970 	return 0;
971 }
972 
973 static int fq_load_priomap(struct fq_sched_data *q,
974 			   const struct nlattr *attr,
975 			   struct netlink_ext_ack *extack)
976 {
977 	const struct tc_prio_qopt *map = nla_data(attr);
978 	int i;
979 
980 	if (map->bands != FQ_BANDS) {
981 		NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
982 		return -EINVAL;
983 	}
984 	for (i = 0; i < TC_PRIO_MAX + 1; i++) {
985 		if (map->priomap[i] >= FQ_BANDS) {
986 			NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
987 					       i, map->priomap[i]);
988 			return -EINVAL;
989 		}
990 	}
991 	fq_prio2band_compress_crumb(map->priomap, q->prio2band);
992 	return 0;
993 }
994 
995 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
996 		     struct netlink_ext_ack *extack)
997 {
998 	struct fq_sched_data *q = qdisc_priv(sch);
999 	struct nlattr *tb[TCA_FQ_MAX + 1];
1000 	int err, drop_count = 0;
1001 	unsigned drop_len = 0;
1002 	u32 fq_log;
1003 
1004 	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
1005 					  NULL);
1006 	if (err < 0)
1007 		return err;
1008 
1009 	sch_tree_lock(sch);
1010 
1011 	fq_log = q->fq_trees_log;
1012 
1013 	if (tb[TCA_FQ_BUCKETS_LOG]) {
1014 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
1015 
1016 		if (nval >= 1 && nval <= ilog2(256*1024))
1017 			fq_log = nval;
1018 		else
1019 			err = -EINVAL;
1020 	}
1021 	if (tb[TCA_FQ_PLIMIT])
1022 		WRITE_ONCE(sch->limit,
1023 			   nla_get_u32(tb[TCA_FQ_PLIMIT]));
1024 
1025 	if (tb[TCA_FQ_FLOW_PLIMIT])
1026 		WRITE_ONCE(q->flow_plimit,
1027 			   nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]));
1028 
1029 	if (tb[TCA_FQ_QUANTUM]) {
1030 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
1031 
1032 		if (quantum > 0 && quantum <= (1 << 20)) {
1033 			WRITE_ONCE(q->quantum, quantum);
1034 		} else {
1035 			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
1036 			err = -EINVAL;
1037 		}
1038 	}
1039 
1040 	if (tb[TCA_FQ_INITIAL_QUANTUM])
1041 		WRITE_ONCE(q->initial_quantum,
1042 			   nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]));
1043 
1044 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
1045 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
1046 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
1047 
1048 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
1049 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
1050 
1051 		WRITE_ONCE(q->flow_max_rate,
1052 			   (rate == ~0U) ? ~0UL : rate);
1053 	}
1054 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
1055 		WRITE_ONCE(q->low_rate_threshold,
1056 			   nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]));
1057 
1058 	if (tb[TCA_FQ_RATE_ENABLE]) {
1059 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
1060 
1061 		if (enable <= 1)
1062 			WRITE_ONCE(q->rate_enable,
1063 				   enable);
1064 		else
1065 			err = -EINVAL;
1066 	}
1067 
1068 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
1069 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
1070 
1071 		WRITE_ONCE(q->flow_refill_delay,
1072 			   usecs_to_jiffies(usecs_delay));
1073 	}
1074 
1075 	if (!err && tb[TCA_FQ_PRIOMAP])
1076 		err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
1077 
1078 	if (!err && tb[TCA_FQ_WEIGHTS])
1079 		err = fq_load_weights(q, tb[TCA_FQ_WEIGHTS], extack);
1080 
1081 	if (tb[TCA_FQ_ORPHAN_MASK])
1082 		WRITE_ONCE(q->orphan_mask,
1083 			   nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]));
1084 
1085 	if (tb[TCA_FQ_CE_THRESHOLD])
1086 		WRITE_ONCE(q->ce_threshold,
1087 			   (u64)NSEC_PER_USEC *
1088 			   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]));
1089 
1090 	if (tb[TCA_FQ_TIMER_SLACK])
1091 		WRITE_ONCE(q->timer_slack,
1092 			   nla_get_u32(tb[TCA_FQ_TIMER_SLACK]));
1093 
1094 	if (tb[TCA_FQ_HORIZON])
1095 		WRITE_ONCE(q->horizon,
1096 			   (u64)NSEC_PER_USEC *
1097 			   nla_get_u32(tb[TCA_FQ_HORIZON]));
1098 
1099 	if (tb[TCA_FQ_HORIZON_DROP])
1100 		WRITE_ONCE(q->horizon_drop,
1101 			   nla_get_u8(tb[TCA_FQ_HORIZON_DROP]));
1102 
1103 	if (!err) {
1104 
1105 		sch_tree_unlock(sch);
1106 		err = fq_resize(sch, fq_log);
1107 		sch_tree_lock(sch);
1108 	}
1109 	while (sch->q.qlen > sch->limit) {
1110 		struct sk_buff *skb = fq_dequeue(sch);
1111 
1112 		if (!skb)
1113 			break;
1114 		drop_len += qdisc_pkt_len(skb);
1115 		rtnl_kfree_skbs(skb, skb);
1116 		drop_count++;
1117 	}
1118 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
1119 
1120 	sch_tree_unlock(sch);
1121 	return err;
1122 }
1123 
1124 static void fq_destroy(struct Qdisc *sch)
1125 {
1126 	struct fq_sched_data *q = qdisc_priv(sch);
1127 
1128 	fq_reset(sch);
1129 	fq_free(q->fq_root);
1130 	qdisc_watchdog_cancel(&q->watchdog);
1131 }
1132 
1133 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
1134 		   struct netlink_ext_ack *extack)
1135 {
1136 	struct fq_sched_data *q = qdisc_priv(sch);
1137 	int i, err;
1138 
1139 	sch->limit		= 10000;
1140 	q->flow_plimit		= 100;
1141 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
1142 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
1143 	q->flow_refill_delay	= msecs_to_jiffies(40);
1144 	q->flow_max_rate	= ~0UL;
1145 	q->time_next_delayed_flow = ~0ULL;
1146 	q->rate_enable		= 1;
1147 	for (i = 0; i < FQ_BANDS; i++) {
1148 		q->band_flows[i].new_flows.first = NULL;
1149 		q->band_flows[i].old_flows.first = NULL;
1150 	}
1151 	q->band_flows[0].quantum = 9 << 16;
1152 	q->band_flows[1].quantum = 3 << 16;
1153 	q->band_flows[2].quantum = 1 << 16;
1154 	q->delayed		= RB_ROOT;
1155 	q->fq_root		= NULL;
1156 	q->fq_trees_log		= ilog2(1024);
1157 	q->orphan_mask		= 1024 - 1;
1158 	q->low_rate_threshold	= 550000 / 8;
1159 
1160 	q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
1161 
1162 	q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
1163 	q->horizon_drop = 1; /* by default, drop packets beyond horizon */
1164 
1165 	/* Default ce_threshold of 4294 seconds */
1166 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
1167 
1168 	fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
1169 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
1170 
1171 	if (opt)
1172 		err = fq_change(sch, opt, extack);
1173 	else
1174 		err = fq_resize(sch, q->fq_trees_log);
1175 
1176 	return err;
1177 }
1178 
1179 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
1180 {
1181 	struct fq_sched_data *q = qdisc_priv(sch);
1182 	struct tc_prio_qopt prio = {
1183 		.bands = FQ_BANDS,
1184 	};
1185 	struct nlattr *opts;
1186 	u64 ce_threshold;
1187 	s32 weights[3];
1188 	u64 horizon;
1189 
1190 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1191 	if (opts == NULL)
1192 		goto nla_put_failure;
1193 
1194 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
1195 
1196 	ce_threshold = READ_ONCE(q->ce_threshold);
1197 	do_div(ce_threshold, NSEC_PER_USEC);
1198 
1199 	horizon = READ_ONCE(q->horizon);
1200 	do_div(horizon, NSEC_PER_USEC);
1201 
1202 	if (nla_put_u32(skb, TCA_FQ_PLIMIT,
1203 			READ_ONCE(sch->limit)) ||
1204 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT,
1205 			READ_ONCE(q->flow_plimit)) ||
1206 	    nla_put_u32(skb, TCA_FQ_QUANTUM,
1207 			READ_ONCE(q->quantum)) ||
1208 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM,
1209 			READ_ONCE(q->initial_quantum)) ||
1210 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE,
1211 			READ_ONCE(q->rate_enable)) ||
1212 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
1213 			min_t(unsigned long,
1214 			      READ_ONCE(q->flow_max_rate), ~0U)) ||
1215 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
1216 			jiffies_to_usecs(READ_ONCE(q->flow_refill_delay))) ||
1217 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK,
1218 			READ_ONCE(q->orphan_mask)) ||
1219 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
1220 			READ_ONCE(q->low_rate_threshold)) ||
1221 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
1222 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG,
1223 			READ_ONCE(q->fq_trees_log)) ||
1224 	    nla_put_u32(skb, TCA_FQ_TIMER_SLACK,
1225 			READ_ONCE(q->timer_slack)) ||
1226 	    nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
1227 	    nla_put_u8(skb, TCA_FQ_HORIZON_DROP,
1228 		       READ_ONCE(q->horizon_drop)))
1229 		goto nla_put_failure;
1230 
1231 	fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
1232 	if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
1233 		goto nla_put_failure;
1234 
1235 	weights[0] = READ_ONCE(q->band_flows[0].quantum);
1236 	weights[1] = READ_ONCE(q->band_flows[1].quantum);
1237 	weights[2] = READ_ONCE(q->band_flows[2].quantum);
1238 	if (nla_put(skb, TCA_FQ_WEIGHTS, sizeof(weights), &weights))
1239 		goto nla_put_failure;
1240 
1241 	return nla_nest_end(skb, opts);
1242 
1243 nla_put_failure:
1244 	return -1;
1245 }
1246 
1247 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1248 {
1249 	struct fq_sched_data *q = qdisc_priv(sch);
1250 	struct tc_fq_qd_stats st;
1251 	int i;
1252 
1253 	st.pad = 0;
1254 
1255 	sch_tree_lock(sch);
1256 
1257 	st.gc_flows		  = q->stat_gc_flows;
1258 	st.highprio_packets	  = 0;
1259 	st.fastpath_packets	  = q->internal.stat_fastpath_packets;
1260 	st.tcp_retrans		  = 0;
1261 	st.throttled		  = q->stat_throttled;
1262 	st.flows_plimit		  = q->stat_flows_plimit;
1263 	st.pkts_too_long	  = q->stat_pkts_too_long;
1264 	st.allocation_errors	  = q->stat_allocation_errors;
1265 	st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1266 				    ktime_get_ns();
1267 	st.flows		  = q->flows;
1268 	st.inactive_flows	  = q->inactive_flows;
1269 	st.throttled_flows	  = q->throttled_flows;
1270 	st.unthrottle_latency_ns  = min_t(unsigned long,
1271 					  q->unthrottle_latency_ns, ~0U);
1272 	st.ce_mark		  = q->stat_ce_mark;
1273 	st.horizon_drops	  = q->stat_horizon_drops;
1274 	st.horizon_caps		  = q->stat_horizon_caps;
1275 	for (i = 0; i < FQ_BANDS; i++) {
1276 		st.band_drops[i]  = q->stat_band_drops[i];
1277 		st.band_pkt_count[i] = q->band_pkt_count[i];
1278 	}
1279 	sch_tree_unlock(sch);
1280 
1281 	return gnet_stats_copy_app(d, &st, sizeof(st));
1282 }
1283 
1284 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1285 	.id		=	"fq",
1286 	.priv_size	=	sizeof(struct fq_sched_data),
1287 
1288 	.enqueue	=	fq_enqueue,
1289 	.dequeue	=	fq_dequeue,
1290 	.peek		=	qdisc_peek_dequeued,
1291 	.init		=	fq_init,
1292 	.reset		=	fq_reset,
1293 	.destroy	=	fq_destroy,
1294 	.change		=	fq_change,
1295 	.dump		=	fq_dump,
1296 	.dump_stats	=	fq_dump_stats,
1297 	.owner		=	THIS_MODULE,
1298 };
1299 MODULE_ALIAS_NET_SCH("fq");
1300 
1301 static int __init fq_module_init(void)
1302 {
1303 	int ret;
1304 
1305 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1306 					   sizeof(struct fq_flow),
1307 					   0, SLAB_HWCACHE_ALIGN, NULL);
1308 	if (!fq_flow_cachep)
1309 		return -ENOMEM;
1310 
1311 	ret = register_qdisc(&fq_qdisc_ops);
1312 	if (ret)
1313 		kmem_cache_destroy(fq_flow_cachep);
1314 	return ret;
1315 }
1316 
1317 static void __exit fq_module_exit(void)
1318 {
1319 	unregister_qdisc(&fq_qdisc_ops);
1320 	kmem_cache_destroy(fq_flow_cachep);
1321 }
1322 
1323 module_init(fq_module_init)
1324 module_exit(fq_module_exit)
1325 MODULE_AUTHOR("Eric Dumazet");
1326 MODULE_LICENSE("GPL");
1327 MODULE_DESCRIPTION("Fair Queue Packet Scheduler");
1328