xref: /linux/net/sched/sch_fq.c (revision f72aa1b276281b4e4f75261af8425bc99d903f3e)
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 	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") */
163 static 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  */
174 static void fq_flow_set_detached(struct fq_flow *f)
175 {
176 	f->age = jiffies | 1UL;
177 }
178 
179 static 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) */
185 static struct fq_flow throttled;
186 
187 static bool fq_flow_is_throttled(const struct fq_flow *f)
188 {
189 	return f->next == &throttled;
190 }
191 
192 enum new_flow {
193 	NEW_FLOW,
194 	OLD_FLOW
195 };
196 
197 static 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 
213 static 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 
220 static 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 
245 static 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 
252 static 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 
258 static 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  */
312 static 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 
337 	sk = skb->sk;
338 	if (sk && sk_fullsock(sk) && !sk_is_tcp(sk) &&
339 	    sk->sk_max_pacing_rate != ~0UL)
340 		return false;
341 
342 	if (q->flow_max_rate != ~0UL)
343 		return false;
344 
345 	return true;
346 }
347 
348 static struct fq_flow *fq_classify(struct Qdisc *sch, struct sk_buff *skb,
349 				   u64 now)
350 {
351 	struct fq_sched_data *q = qdisc_priv(sch);
352 	struct rb_node **p, *parent;
353 	struct sock *sk = skb->sk;
354 	struct rb_root *root;
355 	struct fq_flow *f;
356 
357 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
358 	 * or a listener (SYNCOOKIE mode)
359 	 * 1) request sockets are not full blown,
360 	 *    they do not contain sk_pacing_rate
361 	 * 2) They are not part of a 'flow' yet
362 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
363 	 *    especially if the listener set SO_MAX_PACING_RATE
364 	 * 4) We pretend they are orphaned
365 	 * TCP can also associate TIME_WAIT sockets with RST or ACK packets.
366 	 */
367 	if (!sk || sk_listener_or_tw(sk)) {
368 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
369 
370 		/* By forcing low order bit to 1, we make sure to not
371 		 * collide with a local flow (socket pointers are word aligned)
372 		 */
373 		sk = (struct sock *)((hash << 1) | 1UL);
374 		skb_orphan(skb);
375 	} else if (sk->sk_state == TCP_CLOSE) {
376 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
377 		/*
378 		 * Sockets in TCP_CLOSE are non connected.
379 		 * Typical use case is UDP sockets, they can send packets
380 		 * with sendto() to many different destinations.
381 		 * We probably could use a generic bit advertising
382 		 * non connected sockets, instead of sk_state == TCP_CLOSE,
383 		 * if we care enough.
384 		 */
385 		sk = (struct sock *)((hash << 1) | 1UL);
386 	}
387 
388 	if (fq_fastpath_check(sch, skb, now)) {
389 		q->internal.stat_fastpath_packets++;
390 		if (skb->sk == sk && q->rate_enable &&
391 		    READ_ONCE(sk->sk_pacing_status) != SK_PACING_FQ)
392 			smp_store_release(&sk->sk_pacing_status,
393 					  SK_PACING_FQ);
394 		return &q->internal;
395 	}
396 
397 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
398 
399 	fq_gc(q, root, sk);
400 
401 	p = &root->rb_node;
402 	parent = NULL;
403 	while (*p) {
404 		parent = *p;
405 
406 		f = rb_entry(parent, struct fq_flow, fq_node);
407 		if (f->sk == sk) {
408 			/* socket might have been reallocated, so check
409 			 * if its sk_hash is the same.
410 			 * It not, we need to refill credit with
411 			 * initial quantum
412 			 */
413 			if (unlikely(skb->sk == sk &&
414 				     f->socket_hash != sk->sk_hash)) {
415 				f->credit = q->initial_quantum;
416 				f->socket_hash = sk->sk_hash;
417 				if (q->rate_enable)
418 					smp_store_release(&sk->sk_pacing_status,
419 							  SK_PACING_FQ);
420 				if (fq_flow_is_throttled(f))
421 					fq_flow_unset_throttled(q, f);
422 				f->time_next_packet = 0ULL;
423 			}
424 			return f;
425 		}
426 		if (f->sk > sk)
427 			p = &parent->rb_right;
428 		else
429 			p = &parent->rb_left;
430 	}
431 
432 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
433 	if (unlikely(!f)) {
434 		q->stat_allocation_errors++;
435 		return &q->internal;
436 	}
437 	/* f->t_root is already zeroed after kmem_cache_zalloc() */
438 
439 	fq_flow_set_detached(f);
440 	f->sk = sk;
441 	if (skb->sk == sk) {
442 		f->socket_hash = sk->sk_hash;
443 		if (q->rate_enable)
444 			smp_store_release(&sk->sk_pacing_status,
445 					  SK_PACING_FQ);
446 	}
447 	f->credit = q->initial_quantum;
448 
449 	rb_link_node(&f->fq_node, parent, p);
450 	rb_insert_color(&f->fq_node, root);
451 
452 	q->flows++;
453 	q->inactive_flows++;
454 	return f;
455 }
456 
457 static struct sk_buff *fq_peek(struct fq_flow *flow)
458 {
459 	struct sk_buff *skb = skb_rb_first(&flow->t_root);
460 	struct sk_buff *head = flow->head;
461 
462 	if (!skb)
463 		return head;
464 
465 	if (!head)
466 		return skb;
467 
468 	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
469 		return skb;
470 	return head;
471 }
472 
473 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
474 			  struct sk_buff *skb)
475 {
476 	if (skb == flow->head) {
477 		flow->head = skb->next;
478 	} else {
479 		rb_erase(&skb->rbnode, &flow->t_root);
480 		skb->dev = qdisc_dev(sch);
481 	}
482 }
483 
484 /* Remove one skb from flow queue.
485  * This skb must be the return value of prior fq_peek().
486  */
487 static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
488 			   struct sk_buff *skb)
489 {
490 	fq_erase_head(sch, flow, skb);
491 	skb_mark_not_on_list(skb);
492 	qdisc_qstats_backlog_dec(sch, skb);
493 	sch->q.qlen--;
494 }
495 
496 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
497 {
498 	struct rb_node **p, *parent;
499 	struct sk_buff *head, *aux;
500 
501 	head = flow->head;
502 	if (!head ||
503 	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
504 		if (!head)
505 			flow->head = skb;
506 		else
507 			flow->tail->next = skb;
508 		flow->tail = skb;
509 		skb->next = NULL;
510 		return;
511 	}
512 
513 	p = &flow->t_root.rb_node;
514 	parent = NULL;
515 
516 	while (*p) {
517 		parent = *p;
518 		aux = rb_to_skb(parent);
519 		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
520 			p = &parent->rb_right;
521 		else
522 			p = &parent->rb_left;
523 	}
524 	rb_link_node(&skb->rbnode, parent, p);
525 	rb_insert_color(&skb->rbnode, &flow->t_root);
526 }
527 
528 static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
529 				     const struct fq_sched_data *q, u64 now)
530 {
531 	return unlikely((s64)skb->tstamp > (s64)(now + q->horizon));
532 }
533 
534 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
535 		      struct sk_buff **to_free)
536 {
537 	struct fq_sched_data *q = qdisc_priv(sch);
538 	struct fq_flow *f;
539 	u64 now;
540 	u8 band;
541 
542 	band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
543 	if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
544 		q->stat_band_drops[band]++;
545 		return qdisc_drop(skb, sch, to_free);
546 	}
547 
548 	now = ktime_get_ns();
549 	if (!skb->tstamp) {
550 		fq_skb_cb(skb)->time_to_send = now;
551 	} else {
552 		/* Check if packet timestamp is too far in the future. */
553 		if (fq_packet_beyond_horizon(skb, q, now)) {
554 			if (q->horizon_drop) {
555 					q->stat_horizon_drops++;
556 					return qdisc_drop(skb, sch, to_free);
557 			}
558 			q->stat_horizon_caps++;
559 			skb->tstamp = now + q->horizon;
560 		}
561 		fq_skb_cb(skb)->time_to_send = skb->tstamp;
562 	}
563 
564 	f = fq_classify(sch, skb, now);
565 
566 	if (f != &q->internal) {
567 		if (unlikely(f->qlen >= q->flow_plimit)) {
568 			q->stat_flows_plimit++;
569 			return qdisc_drop(skb, sch, to_free);
570 		}
571 
572 		if (fq_flow_is_detached(f)) {
573 			fq_flow_add_tail(q, f, NEW_FLOW);
574 			if (time_after(jiffies, f->age + q->flow_refill_delay))
575 				f->credit = max_t(u32, f->credit, q->quantum);
576 		}
577 
578 		f->band = band;
579 		q->band_pkt_count[band]++;
580 		fq_skb_cb(skb)->band = band;
581 		if (f->qlen == 0)
582 			q->inactive_flows--;
583 	}
584 
585 	f->qlen++;
586 	/* Note: this overwrites f->age */
587 	flow_queue_add(f, skb);
588 
589 	qdisc_qstats_backlog_inc(sch, skb);
590 	sch->q.qlen++;
591 
592 	return NET_XMIT_SUCCESS;
593 }
594 
595 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
596 {
597 	unsigned long sample;
598 	struct rb_node *p;
599 
600 	if (q->time_next_delayed_flow > now + q->offload_horizon)
601 		return;
602 
603 	/* Update unthrottle latency EWMA.
604 	 * This is cheap and can help diagnosing timer/latency problems.
605 	 */
606 	sample = (unsigned long)(now - q->time_next_delayed_flow);
607 	if ((long)sample > 0) {
608 		q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
609 		q->unthrottle_latency_ns += sample >> 3;
610 	}
611 	now += q->offload_horizon;
612 
613 	q->time_next_delayed_flow = ~0ULL;
614 	while ((p = rb_first(&q->delayed)) != NULL) {
615 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
616 
617 		if (f->time_next_packet > now) {
618 			q->time_next_delayed_flow = f->time_next_packet;
619 			break;
620 		}
621 		fq_flow_unset_throttled(q, f);
622 	}
623 }
624 
625 static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
626 {
627 	if (pband->credit <= 0)
628 		return NULL;
629 
630 	if (pband->new_flows.first)
631 		return &pband->new_flows;
632 
633 	return pband->old_flows.first ? &pband->old_flows : NULL;
634 }
635 
636 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
637 {
638 	struct fq_sched_data *q = qdisc_priv(sch);
639 	struct fq_perband_flows *pband;
640 	struct fq_flow_head *head;
641 	struct sk_buff *skb;
642 	struct fq_flow *f;
643 	unsigned long rate;
644 	int retry;
645 	u32 plen;
646 	u64 now;
647 
648 	if (!sch->q.qlen)
649 		return NULL;
650 
651 	skb = fq_peek(&q->internal);
652 	if (unlikely(skb)) {
653 		q->internal.qlen--;
654 		fq_dequeue_skb(sch, &q->internal, skb);
655 		goto out;
656 	}
657 
658 	now = ktime_get_ns();
659 	fq_check_throttled(q, now);
660 	retry = 0;
661 	pband = &q->band_flows[q->band_nr];
662 begin:
663 	head = fq_pband_head_select(pband);
664 	if (!head) {
665 		while (++retry <= FQ_BANDS) {
666 			if (++q->band_nr == FQ_BANDS)
667 				q->band_nr = 0;
668 			pband = &q->band_flows[q->band_nr];
669 			pband->credit = min(pband->credit + pband->quantum,
670 					    pband->quantum);
671 			if (pband->credit > 0)
672 				goto begin;
673 			retry = 0;
674 		}
675 		if (q->time_next_delayed_flow != ~0ULL)
676 			qdisc_watchdog_schedule_range_ns(&q->watchdog,
677 							q->time_next_delayed_flow,
678 							q->timer_slack);
679 		return NULL;
680 	}
681 	f = head->first;
682 	retry = 0;
683 	if (f->credit <= 0) {
684 		f->credit += q->quantum;
685 		head->first = f->next;
686 		fq_flow_add_tail(q, f, OLD_FLOW);
687 		goto begin;
688 	}
689 
690 	skb = fq_peek(f);
691 	if (skb) {
692 		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
693 					     f->time_next_packet);
694 
695 		if (now + q->offload_horizon < time_next_packet) {
696 			head->first = f->next;
697 			f->time_next_packet = time_next_packet;
698 			fq_flow_set_throttled(q, f);
699 			goto begin;
700 		}
701 		prefetch(&skb->end);
702 		if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
703 			INET_ECN_set_ce(skb);
704 			q->stat_ce_mark++;
705 		}
706 		if (--f->qlen == 0)
707 			q->inactive_flows++;
708 		q->band_pkt_count[fq_skb_cb(skb)->band]--;
709 		fq_dequeue_skb(sch, f, skb);
710 	} else {
711 		head->first = f->next;
712 		/* force a pass through old_flows to prevent starvation */
713 		if (head == &pband->new_flows) {
714 			fq_flow_add_tail(q, f, OLD_FLOW);
715 		} else {
716 			fq_flow_set_detached(f);
717 		}
718 		goto begin;
719 	}
720 	plen = qdisc_pkt_len(skb);
721 	f->credit -= plen;
722 	pband->credit -= plen;
723 
724 	if (!q->rate_enable)
725 		goto out;
726 
727 	rate = q->flow_max_rate;
728 
729 	/* If EDT time was provided for this skb, we need to
730 	 * update f->time_next_packet only if this qdisc enforces
731 	 * a flow max rate.
732 	 */
733 	if (!skb->tstamp) {
734 		if (skb->sk)
735 			rate = min(READ_ONCE(skb->sk->sk_pacing_rate), rate);
736 
737 		if (rate <= q->low_rate_threshold) {
738 			f->credit = 0;
739 		} else {
740 			plen = max(plen, q->quantum);
741 			if (f->credit > 0)
742 				goto out;
743 		}
744 	}
745 	if (rate != ~0UL) {
746 		u64 len = (u64)plen * NSEC_PER_SEC;
747 
748 		if (likely(rate))
749 			len = div64_ul(len, rate);
750 		/* Since socket rate can change later,
751 		 * clamp the delay to 1 second.
752 		 * Really, providers of too big packets should be fixed !
753 		 */
754 		if (unlikely(len > NSEC_PER_SEC)) {
755 			len = NSEC_PER_SEC;
756 			q->stat_pkts_too_long++;
757 		}
758 		/* Account for schedule/timers drifts.
759 		 * f->time_next_packet was set when prior packet was sent,
760 		 * and current time (@now) can be too late by tens of us.
761 		 */
762 		if (f->time_next_packet)
763 			len -= min(len/2, now - f->time_next_packet);
764 		f->time_next_packet = now + len;
765 	}
766 out:
767 	qdisc_bstats_update(sch, skb);
768 	return skb;
769 }
770 
771 static void fq_flow_purge(struct fq_flow *flow)
772 {
773 	struct rb_node *p = rb_first(&flow->t_root);
774 
775 	while (p) {
776 		struct sk_buff *skb = rb_to_skb(p);
777 
778 		p = rb_next(p);
779 		rb_erase(&skb->rbnode, &flow->t_root);
780 		rtnl_kfree_skbs(skb, skb);
781 	}
782 	rtnl_kfree_skbs(flow->head, flow->tail);
783 	flow->head = NULL;
784 	flow->qlen = 0;
785 }
786 
787 static void fq_reset(struct Qdisc *sch)
788 {
789 	struct fq_sched_data *q = qdisc_priv(sch);
790 	struct rb_root *root;
791 	struct rb_node *p;
792 	struct fq_flow *f;
793 	unsigned int idx;
794 
795 	sch->q.qlen = 0;
796 	sch->qstats.backlog = 0;
797 
798 	fq_flow_purge(&q->internal);
799 
800 	if (!q->fq_root)
801 		return;
802 
803 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
804 		root = &q->fq_root[idx];
805 		while ((p = rb_first(root)) != NULL) {
806 			f = rb_entry(p, struct fq_flow, fq_node);
807 			rb_erase(p, root);
808 
809 			fq_flow_purge(f);
810 
811 			kmem_cache_free(fq_flow_cachep, f);
812 		}
813 	}
814 	for (idx = 0; idx < FQ_BANDS; idx++) {
815 		q->band_flows[idx].new_flows.first = NULL;
816 		q->band_flows[idx].old_flows.first = NULL;
817 	}
818 	q->delayed		= RB_ROOT;
819 	q->flows		= 0;
820 	q->inactive_flows	= 0;
821 	q->throttled_flows	= 0;
822 }
823 
824 static void fq_rehash(struct fq_sched_data *q,
825 		      struct rb_root *old_array, u32 old_log,
826 		      struct rb_root *new_array, u32 new_log)
827 {
828 	struct rb_node *op, **np, *parent;
829 	struct rb_root *oroot, *nroot;
830 	struct fq_flow *of, *nf;
831 	int fcnt = 0;
832 	u32 idx;
833 
834 	for (idx = 0; idx < (1U << old_log); idx++) {
835 		oroot = &old_array[idx];
836 		while ((op = rb_first(oroot)) != NULL) {
837 			rb_erase(op, oroot);
838 			of = rb_entry(op, struct fq_flow, fq_node);
839 			if (fq_gc_candidate(of)) {
840 				fcnt++;
841 				kmem_cache_free(fq_flow_cachep, of);
842 				continue;
843 			}
844 			nroot = &new_array[hash_ptr(of->sk, new_log)];
845 
846 			np = &nroot->rb_node;
847 			parent = NULL;
848 			while (*np) {
849 				parent = *np;
850 
851 				nf = rb_entry(parent, struct fq_flow, fq_node);
852 				BUG_ON(nf->sk == of->sk);
853 
854 				if (nf->sk > of->sk)
855 					np = &parent->rb_right;
856 				else
857 					np = &parent->rb_left;
858 			}
859 
860 			rb_link_node(&of->fq_node, parent, np);
861 			rb_insert_color(&of->fq_node, nroot);
862 		}
863 	}
864 	q->flows -= fcnt;
865 	q->inactive_flows -= fcnt;
866 	q->stat_gc_flows += fcnt;
867 }
868 
869 static void fq_free(void *addr)
870 {
871 	kvfree(addr);
872 }
873 
874 static int fq_resize(struct Qdisc *sch, u32 log)
875 {
876 	struct fq_sched_data *q = qdisc_priv(sch);
877 	struct rb_root *array;
878 	void *old_fq_root;
879 	u32 idx;
880 
881 	if (q->fq_root && log == q->fq_trees_log)
882 		return 0;
883 
884 	/* If XPS was setup, we can allocate memory on right NUMA node */
885 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
886 			      netdev_queue_numa_node_read(sch->dev_queue));
887 	if (!array)
888 		return -ENOMEM;
889 
890 	for (idx = 0; idx < (1U << log); idx++)
891 		array[idx] = RB_ROOT;
892 
893 	sch_tree_lock(sch);
894 
895 	old_fq_root = q->fq_root;
896 	if (old_fq_root)
897 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
898 
899 	q->fq_root = array;
900 	WRITE_ONCE(q->fq_trees_log, log);
901 
902 	sch_tree_unlock(sch);
903 
904 	fq_free(old_fq_root);
905 
906 	return 0;
907 }
908 
909 static const struct netlink_range_validation iq_range = {
910 	.max = INT_MAX,
911 };
912 
913 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
914 	[TCA_FQ_UNSPEC]			= { .strict_start_type = TCA_FQ_TIMER_SLACK },
915 
916 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
917 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
918 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
919 	[TCA_FQ_INITIAL_QUANTUM]	= NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
920 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
921 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
922 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
923 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
924 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
925 	[TCA_FQ_ORPHAN_MASK]		= { .type = NLA_U32 },
926 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
927 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
928 	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
929 	[TCA_FQ_HORIZON]		= { .type = NLA_U32 },
930 	[TCA_FQ_HORIZON_DROP]		= { .type = NLA_U8 },
931 	[TCA_FQ_PRIOMAP]		= NLA_POLICY_EXACT_LEN(sizeof(struct tc_prio_qopt)),
932 	[TCA_FQ_WEIGHTS]		= NLA_POLICY_EXACT_LEN(FQ_BANDS * sizeof(s32)),
933 	[TCA_FQ_OFFLOAD_HORIZON]	= { .type = NLA_U32 },
934 };
935 
936 /* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
937 static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
938 {
939 	const int num_elems = TC_PRIO_MAX + 1;
940 	u8 tmp[FQ_PRIO2BAND_CRUMB_SIZE];
941 	int i;
942 
943 	memset(tmp, 0, sizeof(tmp));
944 	for (i = 0; i < num_elems; i++)
945 		tmp[i / 4] |= in[i] << (2 * (i & 0x3));
946 
947 	for (i = 0; i < FQ_PRIO2BAND_CRUMB_SIZE; i++)
948 		WRITE_ONCE(out[i], tmp[i]);
949 }
950 
951 static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
952 {
953 	const int num_elems = TC_PRIO_MAX + 1;
954 	int i;
955 
956 	for (i = 0; i < num_elems; i++)
957 		out[i] = fq_prio2band(in, i);
958 }
959 
960 static int fq_load_weights(struct fq_sched_data *q,
961 			   const struct nlattr *attr,
962 			   struct netlink_ext_ack *extack)
963 {
964 	s32 *weights = nla_data(attr);
965 	int i;
966 
967 	for (i = 0; i < FQ_BANDS; i++) {
968 		if (weights[i] < FQ_MIN_WEIGHT) {
969 			NL_SET_ERR_MSG_FMT_MOD(extack, "Weight %d less that minimum allowed %d",
970 					       weights[i], FQ_MIN_WEIGHT);
971 			return -EINVAL;
972 		}
973 	}
974 	for (i = 0; i < FQ_BANDS; i++)
975 		WRITE_ONCE(q->band_flows[i].quantum, weights[i]);
976 	return 0;
977 }
978 
979 static int fq_load_priomap(struct fq_sched_data *q,
980 			   const struct nlattr *attr,
981 			   struct netlink_ext_ack *extack)
982 {
983 	const struct tc_prio_qopt *map = nla_data(attr);
984 	int i;
985 
986 	if (map->bands != FQ_BANDS) {
987 		NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
988 		return -EINVAL;
989 	}
990 	for (i = 0; i < TC_PRIO_MAX + 1; i++) {
991 		if (map->priomap[i] >= FQ_BANDS) {
992 			NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
993 					       i, map->priomap[i]);
994 			return -EINVAL;
995 		}
996 	}
997 	fq_prio2band_compress_crumb(map->priomap, q->prio2band);
998 	return 0;
999 }
1000 
1001 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
1002 		     struct netlink_ext_ack *extack)
1003 {
1004 	struct fq_sched_data *q = qdisc_priv(sch);
1005 	struct nlattr *tb[TCA_FQ_MAX + 1];
1006 	int err, drop_count = 0;
1007 	unsigned drop_len = 0;
1008 	u32 fq_log;
1009 
1010 	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
1011 					  NULL);
1012 	if (err < 0)
1013 		return err;
1014 
1015 	sch_tree_lock(sch);
1016 
1017 	fq_log = q->fq_trees_log;
1018 
1019 	if (tb[TCA_FQ_BUCKETS_LOG]) {
1020 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
1021 
1022 		if (nval >= 1 && nval <= ilog2(256*1024))
1023 			fq_log = nval;
1024 		else
1025 			err = -EINVAL;
1026 	}
1027 	if (tb[TCA_FQ_PLIMIT])
1028 		WRITE_ONCE(sch->limit,
1029 			   nla_get_u32(tb[TCA_FQ_PLIMIT]));
1030 
1031 	if (tb[TCA_FQ_FLOW_PLIMIT])
1032 		WRITE_ONCE(q->flow_plimit,
1033 			   nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]));
1034 
1035 	if (tb[TCA_FQ_QUANTUM]) {
1036 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
1037 
1038 		if (quantum > 0 && quantum <= (1 << 20)) {
1039 			WRITE_ONCE(q->quantum, quantum);
1040 		} else {
1041 			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
1042 			err = -EINVAL;
1043 		}
1044 	}
1045 
1046 	if (tb[TCA_FQ_INITIAL_QUANTUM])
1047 		WRITE_ONCE(q->initial_quantum,
1048 			   nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]));
1049 
1050 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
1051 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
1052 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
1053 
1054 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
1055 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
1056 
1057 		WRITE_ONCE(q->flow_max_rate,
1058 			   (rate == ~0U) ? ~0UL : rate);
1059 	}
1060 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
1061 		WRITE_ONCE(q->low_rate_threshold,
1062 			   nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]));
1063 
1064 	if (tb[TCA_FQ_RATE_ENABLE]) {
1065 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
1066 
1067 		if (enable <= 1)
1068 			WRITE_ONCE(q->rate_enable,
1069 				   enable);
1070 		else
1071 			err = -EINVAL;
1072 	}
1073 
1074 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
1075 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
1076 
1077 		WRITE_ONCE(q->flow_refill_delay,
1078 			   usecs_to_jiffies(usecs_delay));
1079 	}
1080 
1081 	if (!err && tb[TCA_FQ_PRIOMAP])
1082 		err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
1083 
1084 	if (!err && tb[TCA_FQ_WEIGHTS])
1085 		err = fq_load_weights(q, tb[TCA_FQ_WEIGHTS], extack);
1086 
1087 	if (tb[TCA_FQ_ORPHAN_MASK])
1088 		WRITE_ONCE(q->orphan_mask,
1089 			   nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]));
1090 
1091 	if (tb[TCA_FQ_CE_THRESHOLD])
1092 		WRITE_ONCE(q->ce_threshold,
1093 			   (u64)NSEC_PER_USEC *
1094 			   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]));
1095 
1096 	if (tb[TCA_FQ_TIMER_SLACK])
1097 		WRITE_ONCE(q->timer_slack,
1098 			   nla_get_u32(tb[TCA_FQ_TIMER_SLACK]));
1099 
1100 	if (tb[TCA_FQ_HORIZON])
1101 		WRITE_ONCE(q->horizon,
1102 			   (u64)NSEC_PER_USEC *
1103 			   nla_get_u32(tb[TCA_FQ_HORIZON]));
1104 
1105 	if (tb[TCA_FQ_HORIZON_DROP])
1106 		WRITE_ONCE(q->horizon_drop,
1107 			   nla_get_u8(tb[TCA_FQ_HORIZON_DROP]));
1108 
1109 	if (tb[TCA_FQ_OFFLOAD_HORIZON]) {
1110 		u64 offload_horizon = (u64)NSEC_PER_USEC *
1111 				      nla_get_u32(tb[TCA_FQ_OFFLOAD_HORIZON]);
1112 
1113 		if (offload_horizon <= qdisc_dev(sch)->max_pacing_offload_horizon) {
1114 			WRITE_ONCE(q->offload_horizon, offload_horizon);
1115 		} else {
1116 			NL_SET_ERR_MSG_MOD(extack, "invalid offload_horizon");
1117 			err = -EINVAL;
1118 		}
1119 	}
1120 	if (!err) {
1121 
1122 		sch_tree_unlock(sch);
1123 		err = fq_resize(sch, fq_log);
1124 		sch_tree_lock(sch);
1125 	}
1126 	while (sch->q.qlen > sch->limit) {
1127 		struct sk_buff *skb = fq_dequeue(sch);
1128 
1129 		if (!skb)
1130 			break;
1131 		drop_len += qdisc_pkt_len(skb);
1132 		rtnl_kfree_skbs(skb, skb);
1133 		drop_count++;
1134 	}
1135 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
1136 
1137 	sch_tree_unlock(sch);
1138 	return err;
1139 }
1140 
1141 static void fq_destroy(struct Qdisc *sch)
1142 {
1143 	struct fq_sched_data *q = qdisc_priv(sch);
1144 
1145 	fq_reset(sch);
1146 	fq_free(q->fq_root);
1147 	qdisc_watchdog_cancel(&q->watchdog);
1148 }
1149 
1150 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
1151 		   struct netlink_ext_ack *extack)
1152 {
1153 	struct fq_sched_data *q = qdisc_priv(sch);
1154 	int i, err;
1155 
1156 	sch->limit		= 10000;
1157 	q->flow_plimit		= 100;
1158 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
1159 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
1160 	q->flow_refill_delay	= msecs_to_jiffies(40);
1161 	q->flow_max_rate	= ~0UL;
1162 	q->time_next_delayed_flow = ~0ULL;
1163 	q->rate_enable		= 1;
1164 	for (i = 0; i < FQ_BANDS; i++) {
1165 		q->band_flows[i].new_flows.first = NULL;
1166 		q->band_flows[i].old_flows.first = NULL;
1167 	}
1168 	q->band_flows[0].quantum = 9 << 16;
1169 	q->band_flows[1].quantum = 3 << 16;
1170 	q->band_flows[2].quantum = 1 << 16;
1171 	q->delayed		= RB_ROOT;
1172 	q->fq_root		= NULL;
1173 	q->fq_trees_log		= ilog2(1024);
1174 	q->orphan_mask		= 1024 - 1;
1175 	q->low_rate_threshold	= 550000 / 8;
1176 
1177 	q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
1178 
1179 	q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
1180 	q->horizon_drop = 1; /* by default, drop packets beyond horizon */
1181 
1182 	/* Default ce_threshold of 4294 seconds */
1183 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
1184 
1185 	fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
1186 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
1187 
1188 	if (opt)
1189 		err = fq_change(sch, opt, extack);
1190 	else
1191 		err = fq_resize(sch, q->fq_trees_log);
1192 
1193 	return err;
1194 }
1195 
1196 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
1197 {
1198 	struct fq_sched_data *q = qdisc_priv(sch);
1199 	struct tc_prio_qopt prio = {
1200 		.bands = FQ_BANDS,
1201 	};
1202 	struct nlattr *opts;
1203 	u64 offload_horizon;
1204 	u64 ce_threshold;
1205 	s32 weights[3];
1206 	u64 horizon;
1207 
1208 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1209 	if (opts == NULL)
1210 		goto nla_put_failure;
1211 
1212 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
1213 
1214 	ce_threshold = READ_ONCE(q->ce_threshold);
1215 	do_div(ce_threshold, NSEC_PER_USEC);
1216 
1217 	horizon = READ_ONCE(q->horizon);
1218 	do_div(horizon, NSEC_PER_USEC);
1219 
1220 	offload_horizon = READ_ONCE(q->offload_horizon);
1221 	do_div(offload_horizon, NSEC_PER_USEC);
1222 
1223 	if (nla_put_u32(skb, TCA_FQ_PLIMIT,
1224 			READ_ONCE(sch->limit)) ||
1225 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT,
1226 			READ_ONCE(q->flow_plimit)) ||
1227 	    nla_put_u32(skb, TCA_FQ_QUANTUM,
1228 			READ_ONCE(q->quantum)) ||
1229 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM,
1230 			READ_ONCE(q->initial_quantum)) ||
1231 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE,
1232 			READ_ONCE(q->rate_enable)) ||
1233 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
1234 			min_t(unsigned long,
1235 			      READ_ONCE(q->flow_max_rate), ~0U)) ||
1236 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
1237 			jiffies_to_usecs(READ_ONCE(q->flow_refill_delay))) ||
1238 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK,
1239 			READ_ONCE(q->orphan_mask)) ||
1240 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
1241 			READ_ONCE(q->low_rate_threshold)) ||
1242 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
1243 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG,
1244 			READ_ONCE(q->fq_trees_log)) ||
1245 	    nla_put_u32(skb, TCA_FQ_TIMER_SLACK,
1246 			READ_ONCE(q->timer_slack)) ||
1247 	    nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
1248 	    nla_put_u32(skb, TCA_FQ_OFFLOAD_HORIZON, (u32)offload_horizon) ||
1249 	    nla_put_u8(skb, TCA_FQ_HORIZON_DROP,
1250 		       READ_ONCE(q->horizon_drop)))
1251 		goto nla_put_failure;
1252 
1253 	fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
1254 	if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
1255 		goto nla_put_failure;
1256 
1257 	weights[0] = READ_ONCE(q->band_flows[0].quantum);
1258 	weights[1] = READ_ONCE(q->band_flows[1].quantum);
1259 	weights[2] = READ_ONCE(q->band_flows[2].quantum);
1260 	if (nla_put(skb, TCA_FQ_WEIGHTS, sizeof(weights), &weights))
1261 		goto nla_put_failure;
1262 
1263 	return nla_nest_end(skb, opts);
1264 
1265 nla_put_failure:
1266 	return -1;
1267 }
1268 
1269 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1270 {
1271 	struct fq_sched_data *q = qdisc_priv(sch);
1272 	struct tc_fq_qd_stats st;
1273 	int i;
1274 
1275 	st.pad = 0;
1276 
1277 	sch_tree_lock(sch);
1278 
1279 	st.gc_flows		  = q->stat_gc_flows;
1280 	st.highprio_packets	  = 0;
1281 	st.fastpath_packets	  = q->internal.stat_fastpath_packets;
1282 	st.tcp_retrans		  = 0;
1283 	st.throttled		  = q->stat_throttled;
1284 	st.flows_plimit		  = q->stat_flows_plimit;
1285 	st.pkts_too_long	  = q->stat_pkts_too_long;
1286 	st.allocation_errors	  = q->stat_allocation_errors;
1287 	st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1288 				    ktime_get_ns();
1289 	st.flows		  = q->flows;
1290 	st.inactive_flows	  = q->inactive_flows;
1291 	st.throttled_flows	  = q->throttled_flows;
1292 	st.unthrottle_latency_ns  = min_t(unsigned long,
1293 					  q->unthrottle_latency_ns, ~0U);
1294 	st.ce_mark		  = q->stat_ce_mark;
1295 	st.horizon_drops	  = q->stat_horizon_drops;
1296 	st.horizon_caps		  = q->stat_horizon_caps;
1297 	for (i = 0; i < FQ_BANDS; i++) {
1298 		st.band_drops[i]  = q->stat_band_drops[i];
1299 		st.band_pkt_count[i] = q->band_pkt_count[i];
1300 	}
1301 	sch_tree_unlock(sch);
1302 
1303 	return gnet_stats_copy_app(d, &st, sizeof(st));
1304 }
1305 
1306 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1307 	.id		=	"fq",
1308 	.priv_size	=	sizeof(struct fq_sched_data),
1309 
1310 	.enqueue	=	fq_enqueue,
1311 	.dequeue	=	fq_dequeue,
1312 	.peek		=	qdisc_peek_dequeued,
1313 	.init		=	fq_init,
1314 	.reset		=	fq_reset,
1315 	.destroy	=	fq_destroy,
1316 	.change		=	fq_change,
1317 	.dump		=	fq_dump,
1318 	.dump_stats	=	fq_dump_stats,
1319 	.owner		=	THIS_MODULE,
1320 };
1321 MODULE_ALIAS_NET_SCH("fq");
1322 
1323 static int __init fq_module_init(void)
1324 {
1325 	int ret;
1326 
1327 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1328 					   sizeof(struct fq_flow),
1329 					   0, SLAB_HWCACHE_ALIGN, NULL);
1330 	if (!fq_flow_cachep)
1331 		return -ENOMEM;
1332 
1333 	ret = register_qdisc(&fq_qdisc_ops);
1334 	if (ret)
1335 		kmem_cache_destroy(fq_flow_cachep);
1336 	return ret;
1337 }
1338 
1339 static void __exit fq_module_exit(void)
1340 {
1341 	unregister_qdisc(&fq_qdisc_ops);
1342 	kmem_cache_destroy(fq_flow_cachep);
1343 }
1344 
1345 module_init(fq_module_init)
1346 module_exit(fq_module_exit)
1347 MODULE_AUTHOR("Eric Dumazet");
1348 MODULE_LICENSE("GPL");
1349 MODULE_DESCRIPTION("Fair Queue Packet Scheduler");
1350