xref: /openbmc/linux/net/sched/sch_fq.c (revision 4eb5928d)
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 
53 struct fq_skb_cb {
54 	u64	        time_to_send;
55 };
56 
57 static 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  */
68 struct fq_flow {
69 /* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
70 	struct rb_root	t_root;
71 	struct sk_buff	*head;		/* list of skbs for this flow : first skb */
72 	union {
73 		struct sk_buff *tail;	/* last skb in the list */
74 		unsigned long  age;	/* (jiffies | 1UL) when flow was emptied, for gc */
75 	};
76 	struct rb_node	fq_node;	/* anchor in fq_root[] trees */
77 	struct sock	*sk;
78 	u32		socket_hash;	/* sk_hash */
79 	int		qlen;		/* number of packets in flow queue */
80 
81 /* Second cache line, used in fq_dequeue() */
82 	int		credit;
83 	/* 32bit hole on 64bit arches */
84 
85 	struct fq_flow *next;		/* next pointer in RR lists */
86 
87 	struct rb_node  rate_node;	/* anchor in q->delayed tree */
88 	u64		time_next_packet;
89 } ____cacheline_aligned_in_smp;
90 
91 struct fq_flow_head {
92 	struct fq_flow *first;
93 	struct fq_flow *last;
94 };
95 
96 struct fq_sched_data {
97 	struct fq_flow_head new_flows;
98 
99 	struct fq_flow_head old_flows;
100 
101 	struct rb_root	delayed;	/* for rate limited flows */
102 	u64		time_next_delayed_flow;
103 	u64		ktime_cache;	/* copy of last ktime_get_ns() */
104 	unsigned long	unthrottle_latency_ns;
105 
106 	struct fq_flow	internal;	/* for non classified or high prio packets */
107 	u32		quantum;
108 	u32		initial_quantum;
109 	u32		flow_refill_delay;
110 	u32		flow_plimit;	/* max packets per flow */
111 	unsigned long	flow_max_rate;	/* optional max rate per flow */
112 	u64		ce_threshold;
113 	u64		horizon;	/* horizon in ns */
114 	u32		orphan_mask;	/* mask for orphaned skb */
115 	u32		low_rate_threshold;
116 	struct rb_root	*fq_root;
117 	u8		rate_enable;
118 	u8		fq_trees_log;
119 	u8		horizon_drop;
120 	u32		flows;
121 	u32		inactive_flows;
122 	u32		throttled_flows;
123 
124 	u64		stat_gc_flows;
125 	u64		stat_internal_packets;
126 	u64		stat_throttled;
127 	u64		stat_ce_mark;
128 	u64		stat_horizon_drops;
129 	u64		stat_horizon_caps;
130 	u64		stat_flows_plimit;
131 	u64		stat_pkts_too_long;
132 	u64		stat_allocation_errors;
133 
134 	u32		timer_slack; /* hrtimer slack in ns */
135 	struct qdisc_watchdog watchdog;
136 };
137 
138 /*
139  * f->tail and f->age share the same location.
140  * We can use the low order bit to differentiate if this location points
141  * to a sk_buff or contains a jiffies value, if we force this value to be odd.
142  * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
143  */
144 static void fq_flow_set_detached(struct fq_flow *f)
145 {
146 	f->age = jiffies | 1UL;
147 }
148 
149 static bool fq_flow_is_detached(const struct fq_flow *f)
150 {
151 	return !!(f->age & 1UL);
152 }
153 
154 /* special value to mark a throttled flow (not on old/new list) */
155 static struct fq_flow throttled;
156 
157 static bool fq_flow_is_throttled(const struct fq_flow *f)
158 {
159 	return f->next == &throttled;
160 }
161 
162 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
163 {
164 	if (head->first)
165 		head->last->next = flow;
166 	else
167 		head->first = flow;
168 	head->last = flow;
169 	flow->next = NULL;
170 }
171 
172 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
173 {
174 	rb_erase(&f->rate_node, &q->delayed);
175 	q->throttled_flows--;
176 	fq_flow_add_tail(&q->old_flows, f);
177 }
178 
179 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
180 {
181 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
182 
183 	while (*p) {
184 		struct fq_flow *aux;
185 
186 		parent = *p;
187 		aux = rb_entry(parent, struct fq_flow, rate_node);
188 		if (f->time_next_packet >= aux->time_next_packet)
189 			p = &parent->rb_right;
190 		else
191 			p = &parent->rb_left;
192 	}
193 	rb_link_node(&f->rate_node, parent, p);
194 	rb_insert_color(&f->rate_node, &q->delayed);
195 	q->throttled_flows++;
196 	q->stat_throttled++;
197 
198 	f->next = &throttled;
199 	if (q->time_next_delayed_flow > f->time_next_packet)
200 		q->time_next_delayed_flow = f->time_next_packet;
201 }
202 
203 
204 static struct kmem_cache *fq_flow_cachep __read_mostly;
205 
206 
207 /* limit number of collected flows per round */
208 #define FQ_GC_MAX 8
209 #define FQ_GC_AGE (3*HZ)
210 
211 static bool fq_gc_candidate(const struct fq_flow *f)
212 {
213 	return fq_flow_is_detached(f) &&
214 	       time_after(jiffies, f->age + FQ_GC_AGE);
215 }
216 
217 static void fq_gc(struct fq_sched_data *q,
218 		  struct rb_root *root,
219 		  struct sock *sk)
220 {
221 	struct rb_node **p, *parent;
222 	void *tofree[FQ_GC_MAX];
223 	struct fq_flow *f;
224 	int i, fcnt = 0;
225 
226 	p = &root->rb_node;
227 	parent = NULL;
228 	while (*p) {
229 		parent = *p;
230 
231 		f = rb_entry(parent, struct fq_flow, fq_node);
232 		if (f->sk == sk)
233 			break;
234 
235 		if (fq_gc_candidate(f)) {
236 			tofree[fcnt++] = f;
237 			if (fcnt == FQ_GC_MAX)
238 				break;
239 		}
240 
241 		if (f->sk > sk)
242 			p = &parent->rb_right;
243 		else
244 			p = &parent->rb_left;
245 	}
246 
247 	if (!fcnt)
248 		return;
249 
250 	for (i = fcnt; i > 0; ) {
251 		f = tofree[--i];
252 		rb_erase(&f->fq_node, root);
253 	}
254 	q->flows -= fcnt;
255 	q->inactive_flows -= fcnt;
256 	q->stat_gc_flows += fcnt;
257 
258 	kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
259 }
260 
261 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
262 {
263 	struct rb_node **p, *parent;
264 	struct sock *sk = skb->sk;
265 	struct rb_root *root;
266 	struct fq_flow *f;
267 
268 	/* warning: no starvation prevention... */
269 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
270 		return &q->internal;
271 
272 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
273 	 * or a listener (SYNCOOKIE mode)
274 	 * 1) request sockets are not full blown,
275 	 *    they do not contain sk_pacing_rate
276 	 * 2) They are not part of a 'flow' yet
277 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
278 	 *    especially if the listener set SO_MAX_PACING_RATE
279 	 * 4) We pretend they are orphaned
280 	 */
281 	if (!sk || sk_listener(sk)) {
282 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
283 
284 		/* By forcing low order bit to 1, we make sure to not
285 		 * collide with a local flow (socket pointers are word aligned)
286 		 */
287 		sk = (struct sock *)((hash << 1) | 1UL);
288 		skb_orphan(skb);
289 	} else if (sk->sk_state == TCP_CLOSE) {
290 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
291 		/*
292 		 * Sockets in TCP_CLOSE are non connected.
293 		 * Typical use case is UDP sockets, they can send packets
294 		 * with sendto() to many different destinations.
295 		 * We probably could use a generic bit advertising
296 		 * non connected sockets, instead of sk_state == TCP_CLOSE,
297 		 * if we care enough.
298 		 */
299 		sk = (struct sock *)((hash << 1) | 1UL);
300 	}
301 
302 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
303 
304 	if (q->flows >= (2U << q->fq_trees_log) &&
305 	    q->inactive_flows > q->flows/2)
306 		fq_gc(q, root, sk);
307 
308 	p = &root->rb_node;
309 	parent = NULL;
310 	while (*p) {
311 		parent = *p;
312 
313 		f = rb_entry(parent, struct fq_flow, fq_node);
314 		if (f->sk == sk) {
315 			/* socket might have been reallocated, so check
316 			 * if its sk_hash is the same.
317 			 * It not, we need to refill credit with
318 			 * initial quantum
319 			 */
320 			if (unlikely(skb->sk == sk &&
321 				     f->socket_hash != sk->sk_hash)) {
322 				f->credit = q->initial_quantum;
323 				f->socket_hash = sk->sk_hash;
324 				if (q->rate_enable)
325 					smp_store_release(&sk->sk_pacing_status,
326 							  SK_PACING_FQ);
327 				if (fq_flow_is_throttled(f))
328 					fq_flow_unset_throttled(q, f);
329 				f->time_next_packet = 0ULL;
330 			}
331 			return f;
332 		}
333 		if (f->sk > sk)
334 			p = &parent->rb_right;
335 		else
336 			p = &parent->rb_left;
337 	}
338 
339 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
340 	if (unlikely(!f)) {
341 		q->stat_allocation_errors++;
342 		return &q->internal;
343 	}
344 	/* f->t_root is already zeroed after kmem_cache_zalloc() */
345 
346 	fq_flow_set_detached(f);
347 	f->sk = sk;
348 	if (skb->sk == sk) {
349 		f->socket_hash = sk->sk_hash;
350 		if (q->rate_enable)
351 			smp_store_release(&sk->sk_pacing_status,
352 					  SK_PACING_FQ);
353 	}
354 	f->credit = q->initial_quantum;
355 
356 	rb_link_node(&f->fq_node, parent, p);
357 	rb_insert_color(&f->fq_node, root);
358 
359 	q->flows++;
360 	q->inactive_flows++;
361 	return f;
362 }
363 
364 static struct sk_buff *fq_peek(struct fq_flow *flow)
365 {
366 	struct sk_buff *skb = skb_rb_first(&flow->t_root);
367 	struct sk_buff *head = flow->head;
368 
369 	if (!skb)
370 		return head;
371 
372 	if (!head)
373 		return skb;
374 
375 	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
376 		return skb;
377 	return head;
378 }
379 
380 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
381 			  struct sk_buff *skb)
382 {
383 	if (skb == flow->head) {
384 		flow->head = skb->next;
385 	} else {
386 		rb_erase(&skb->rbnode, &flow->t_root);
387 		skb->dev = qdisc_dev(sch);
388 	}
389 }
390 
391 /* Remove one skb from flow queue.
392  * This skb must be the return value of prior fq_peek().
393  */
394 static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
395 			   struct sk_buff *skb)
396 {
397 	fq_erase_head(sch, flow, skb);
398 	skb_mark_not_on_list(skb);
399 	flow->qlen--;
400 	qdisc_qstats_backlog_dec(sch, skb);
401 	sch->q.qlen--;
402 }
403 
404 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
405 {
406 	struct rb_node **p, *parent;
407 	struct sk_buff *head, *aux;
408 
409 	head = flow->head;
410 	if (!head ||
411 	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
412 		if (!head)
413 			flow->head = skb;
414 		else
415 			flow->tail->next = skb;
416 		flow->tail = skb;
417 		skb->next = NULL;
418 		return;
419 	}
420 
421 	p = &flow->t_root.rb_node;
422 	parent = NULL;
423 
424 	while (*p) {
425 		parent = *p;
426 		aux = rb_to_skb(parent);
427 		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
428 			p = &parent->rb_right;
429 		else
430 			p = &parent->rb_left;
431 	}
432 	rb_link_node(&skb->rbnode, parent, p);
433 	rb_insert_color(&skb->rbnode, &flow->t_root);
434 }
435 
436 static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
437 				    const struct fq_sched_data *q)
438 {
439 	return unlikely((s64)skb->tstamp > (s64)(q->ktime_cache + q->horizon));
440 }
441 
442 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
443 		      struct sk_buff **to_free)
444 {
445 	struct fq_sched_data *q = qdisc_priv(sch);
446 	struct fq_flow *f;
447 
448 	if (unlikely(sch->q.qlen >= sch->limit))
449 		return qdisc_drop(skb, sch, to_free);
450 
451 	if (!skb->tstamp) {
452 		fq_skb_cb(skb)->time_to_send = q->ktime_cache = ktime_get_ns();
453 	} else {
454 		/* Check if packet timestamp is too far in the future.
455 		 * Try first if our cached value, to avoid ktime_get_ns()
456 		 * cost in most cases.
457 		 */
458 		if (fq_packet_beyond_horizon(skb, q)) {
459 			/* Refresh our cache and check another time */
460 			q->ktime_cache = ktime_get_ns();
461 			if (fq_packet_beyond_horizon(skb, q)) {
462 				if (q->horizon_drop) {
463 					q->stat_horizon_drops++;
464 					return qdisc_drop(skb, sch, to_free);
465 				}
466 				q->stat_horizon_caps++;
467 				skb->tstamp = q->ktime_cache + q->horizon;
468 			}
469 		}
470 		fq_skb_cb(skb)->time_to_send = skb->tstamp;
471 	}
472 
473 	f = fq_classify(skb, q);
474 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
475 		q->stat_flows_plimit++;
476 		return qdisc_drop(skb, sch, to_free);
477 	}
478 
479 	f->qlen++;
480 	qdisc_qstats_backlog_inc(sch, skb);
481 	if (fq_flow_is_detached(f)) {
482 		fq_flow_add_tail(&q->new_flows, f);
483 		if (time_after(jiffies, f->age + q->flow_refill_delay))
484 			f->credit = max_t(u32, f->credit, q->quantum);
485 		q->inactive_flows--;
486 	}
487 
488 	/* Note: this overwrites f->age */
489 	flow_queue_add(f, skb);
490 
491 	if (unlikely(f == &q->internal)) {
492 		q->stat_internal_packets++;
493 	}
494 	sch->q.qlen++;
495 
496 	return NET_XMIT_SUCCESS;
497 }
498 
499 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
500 {
501 	unsigned long sample;
502 	struct rb_node *p;
503 
504 	if (q->time_next_delayed_flow > now)
505 		return;
506 
507 	/* Update unthrottle latency EWMA.
508 	 * This is cheap and can help diagnosing timer/latency problems.
509 	 */
510 	sample = (unsigned long)(now - q->time_next_delayed_flow);
511 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
512 	q->unthrottle_latency_ns += sample >> 3;
513 
514 	q->time_next_delayed_flow = ~0ULL;
515 	while ((p = rb_first(&q->delayed)) != NULL) {
516 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
517 
518 		if (f->time_next_packet > now) {
519 			q->time_next_delayed_flow = f->time_next_packet;
520 			break;
521 		}
522 		fq_flow_unset_throttled(q, f);
523 	}
524 }
525 
526 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
527 {
528 	struct fq_sched_data *q = qdisc_priv(sch);
529 	struct fq_flow_head *head;
530 	struct sk_buff *skb;
531 	struct fq_flow *f;
532 	unsigned long rate;
533 	u32 plen;
534 	u64 now;
535 
536 	if (!sch->q.qlen)
537 		return NULL;
538 
539 	skb = fq_peek(&q->internal);
540 	if (unlikely(skb)) {
541 		fq_dequeue_skb(sch, &q->internal, skb);
542 		goto out;
543 	}
544 
545 	q->ktime_cache = now = ktime_get_ns();
546 	fq_check_throttled(q, now);
547 begin:
548 	head = &q->new_flows;
549 	if (!head->first) {
550 		head = &q->old_flows;
551 		if (!head->first) {
552 			if (q->time_next_delayed_flow != ~0ULL)
553 				qdisc_watchdog_schedule_range_ns(&q->watchdog,
554 							q->time_next_delayed_flow,
555 							q->timer_slack);
556 			return NULL;
557 		}
558 	}
559 	f = head->first;
560 
561 	if (f->credit <= 0) {
562 		f->credit += q->quantum;
563 		head->first = f->next;
564 		fq_flow_add_tail(&q->old_flows, f);
565 		goto begin;
566 	}
567 
568 	skb = fq_peek(f);
569 	if (skb) {
570 		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
571 					     f->time_next_packet);
572 
573 		if (now < time_next_packet) {
574 			head->first = f->next;
575 			f->time_next_packet = time_next_packet;
576 			fq_flow_set_throttled(q, f);
577 			goto begin;
578 		}
579 		prefetch(&skb->end);
580 		if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
581 			INET_ECN_set_ce(skb);
582 			q->stat_ce_mark++;
583 		}
584 		fq_dequeue_skb(sch, f, skb);
585 	} else {
586 		head->first = f->next;
587 		/* force a pass through old_flows to prevent starvation */
588 		if ((head == &q->new_flows) && q->old_flows.first) {
589 			fq_flow_add_tail(&q->old_flows, f);
590 		} else {
591 			fq_flow_set_detached(f);
592 			q->inactive_flows++;
593 		}
594 		goto begin;
595 	}
596 	plen = qdisc_pkt_len(skb);
597 	f->credit -= plen;
598 
599 	if (!q->rate_enable)
600 		goto out;
601 
602 	rate = q->flow_max_rate;
603 
604 	/* If EDT time was provided for this skb, we need to
605 	 * update f->time_next_packet only if this qdisc enforces
606 	 * a flow max rate.
607 	 */
608 	if (!skb->tstamp) {
609 		if (skb->sk)
610 			rate = min(skb->sk->sk_pacing_rate, rate);
611 
612 		if (rate <= q->low_rate_threshold) {
613 			f->credit = 0;
614 		} else {
615 			plen = max(plen, q->quantum);
616 			if (f->credit > 0)
617 				goto out;
618 		}
619 	}
620 	if (rate != ~0UL) {
621 		u64 len = (u64)plen * NSEC_PER_SEC;
622 
623 		if (likely(rate))
624 			len = div64_ul(len, rate);
625 		/* Since socket rate can change later,
626 		 * clamp the delay to 1 second.
627 		 * Really, providers of too big packets should be fixed !
628 		 */
629 		if (unlikely(len > NSEC_PER_SEC)) {
630 			len = NSEC_PER_SEC;
631 			q->stat_pkts_too_long++;
632 		}
633 		/* Account for schedule/timers drifts.
634 		 * f->time_next_packet was set when prior packet was sent,
635 		 * and current time (@now) can be too late by tens of us.
636 		 */
637 		if (f->time_next_packet)
638 			len -= min(len/2, now - f->time_next_packet);
639 		f->time_next_packet = now + len;
640 	}
641 out:
642 	qdisc_bstats_update(sch, skb);
643 	return skb;
644 }
645 
646 static void fq_flow_purge(struct fq_flow *flow)
647 {
648 	struct rb_node *p = rb_first(&flow->t_root);
649 
650 	while (p) {
651 		struct sk_buff *skb = rb_to_skb(p);
652 
653 		p = rb_next(p);
654 		rb_erase(&skb->rbnode, &flow->t_root);
655 		rtnl_kfree_skbs(skb, skb);
656 	}
657 	rtnl_kfree_skbs(flow->head, flow->tail);
658 	flow->head = NULL;
659 	flow->qlen = 0;
660 }
661 
662 static void fq_reset(struct Qdisc *sch)
663 {
664 	struct fq_sched_data *q = qdisc_priv(sch);
665 	struct rb_root *root;
666 	struct rb_node *p;
667 	struct fq_flow *f;
668 	unsigned int idx;
669 
670 	sch->q.qlen = 0;
671 	sch->qstats.backlog = 0;
672 
673 	fq_flow_purge(&q->internal);
674 
675 	if (!q->fq_root)
676 		return;
677 
678 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
679 		root = &q->fq_root[idx];
680 		while ((p = rb_first(root)) != NULL) {
681 			f = rb_entry(p, struct fq_flow, fq_node);
682 			rb_erase(p, root);
683 
684 			fq_flow_purge(f);
685 
686 			kmem_cache_free(fq_flow_cachep, f);
687 		}
688 	}
689 	q->new_flows.first	= NULL;
690 	q->old_flows.first	= NULL;
691 	q->delayed		= RB_ROOT;
692 	q->flows		= 0;
693 	q->inactive_flows	= 0;
694 	q->throttled_flows	= 0;
695 }
696 
697 static void fq_rehash(struct fq_sched_data *q,
698 		      struct rb_root *old_array, u32 old_log,
699 		      struct rb_root *new_array, u32 new_log)
700 {
701 	struct rb_node *op, **np, *parent;
702 	struct rb_root *oroot, *nroot;
703 	struct fq_flow *of, *nf;
704 	int fcnt = 0;
705 	u32 idx;
706 
707 	for (idx = 0; idx < (1U << old_log); idx++) {
708 		oroot = &old_array[idx];
709 		while ((op = rb_first(oroot)) != NULL) {
710 			rb_erase(op, oroot);
711 			of = rb_entry(op, struct fq_flow, fq_node);
712 			if (fq_gc_candidate(of)) {
713 				fcnt++;
714 				kmem_cache_free(fq_flow_cachep, of);
715 				continue;
716 			}
717 			nroot = &new_array[hash_ptr(of->sk, new_log)];
718 
719 			np = &nroot->rb_node;
720 			parent = NULL;
721 			while (*np) {
722 				parent = *np;
723 
724 				nf = rb_entry(parent, struct fq_flow, fq_node);
725 				BUG_ON(nf->sk == of->sk);
726 
727 				if (nf->sk > of->sk)
728 					np = &parent->rb_right;
729 				else
730 					np = &parent->rb_left;
731 			}
732 
733 			rb_link_node(&of->fq_node, parent, np);
734 			rb_insert_color(&of->fq_node, nroot);
735 		}
736 	}
737 	q->flows -= fcnt;
738 	q->inactive_flows -= fcnt;
739 	q->stat_gc_flows += fcnt;
740 }
741 
742 static void fq_free(void *addr)
743 {
744 	kvfree(addr);
745 }
746 
747 static int fq_resize(struct Qdisc *sch, u32 log)
748 {
749 	struct fq_sched_data *q = qdisc_priv(sch);
750 	struct rb_root *array;
751 	void *old_fq_root;
752 	u32 idx;
753 
754 	if (q->fq_root && log == q->fq_trees_log)
755 		return 0;
756 
757 	/* If XPS was setup, we can allocate memory on right NUMA node */
758 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
759 			      netdev_queue_numa_node_read(sch->dev_queue));
760 	if (!array)
761 		return -ENOMEM;
762 
763 	for (idx = 0; idx < (1U << log); idx++)
764 		array[idx] = RB_ROOT;
765 
766 	sch_tree_lock(sch);
767 
768 	old_fq_root = q->fq_root;
769 	if (old_fq_root)
770 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
771 
772 	q->fq_root = array;
773 	q->fq_trees_log = log;
774 
775 	sch_tree_unlock(sch);
776 
777 	fq_free(old_fq_root);
778 
779 	return 0;
780 }
781 
782 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
783 	[TCA_FQ_UNSPEC]			= { .strict_start_type = TCA_FQ_TIMER_SLACK },
784 
785 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
786 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
787 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
788 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
789 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
790 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
791 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
792 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
793 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
794 	[TCA_FQ_ORPHAN_MASK]		= { .type = NLA_U32 },
795 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
796 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
797 	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
798 	[TCA_FQ_HORIZON]		= { .type = NLA_U32 },
799 	[TCA_FQ_HORIZON_DROP]		= { .type = NLA_U8 },
800 };
801 
802 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
803 		     struct netlink_ext_ack *extack)
804 {
805 	struct fq_sched_data *q = qdisc_priv(sch);
806 	struct nlattr *tb[TCA_FQ_MAX + 1];
807 	int err, drop_count = 0;
808 	unsigned drop_len = 0;
809 	u32 fq_log;
810 
811 	if (!opt)
812 		return -EINVAL;
813 
814 	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
815 					  NULL);
816 	if (err < 0)
817 		return err;
818 
819 	sch_tree_lock(sch);
820 
821 	fq_log = q->fq_trees_log;
822 
823 	if (tb[TCA_FQ_BUCKETS_LOG]) {
824 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
825 
826 		if (nval >= 1 && nval <= ilog2(256*1024))
827 			fq_log = nval;
828 		else
829 			err = -EINVAL;
830 	}
831 	if (tb[TCA_FQ_PLIMIT])
832 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
833 
834 	if (tb[TCA_FQ_FLOW_PLIMIT])
835 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
836 
837 	if (tb[TCA_FQ_QUANTUM]) {
838 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
839 
840 		if (quantum > 0 && quantum <= (1 << 20)) {
841 			q->quantum = quantum;
842 		} else {
843 			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
844 			err = -EINVAL;
845 		}
846 	}
847 
848 	if (tb[TCA_FQ_INITIAL_QUANTUM])
849 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
850 
851 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
852 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
853 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
854 
855 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
856 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
857 
858 		q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
859 	}
860 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
861 		q->low_rate_threshold =
862 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
863 
864 	if (tb[TCA_FQ_RATE_ENABLE]) {
865 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
866 
867 		if (enable <= 1)
868 			q->rate_enable = enable;
869 		else
870 			err = -EINVAL;
871 	}
872 
873 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
874 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
875 
876 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
877 	}
878 
879 	if (tb[TCA_FQ_ORPHAN_MASK])
880 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
881 
882 	if (tb[TCA_FQ_CE_THRESHOLD])
883 		q->ce_threshold = (u64)NSEC_PER_USEC *
884 				  nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
885 
886 	if (tb[TCA_FQ_TIMER_SLACK])
887 		q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]);
888 
889 	if (tb[TCA_FQ_HORIZON])
890 		q->horizon = (u64)NSEC_PER_USEC *
891 				  nla_get_u32(tb[TCA_FQ_HORIZON]);
892 
893 	if (tb[TCA_FQ_HORIZON_DROP])
894 		q->horizon_drop = nla_get_u8(tb[TCA_FQ_HORIZON_DROP]);
895 
896 	if (!err) {
897 
898 		sch_tree_unlock(sch);
899 		err = fq_resize(sch, fq_log);
900 		sch_tree_lock(sch);
901 	}
902 	while (sch->q.qlen > sch->limit) {
903 		struct sk_buff *skb = fq_dequeue(sch);
904 
905 		if (!skb)
906 			break;
907 		drop_len += qdisc_pkt_len(skb);
908 		rtnl_kfree_skbs(skb, skb);
909 		drop_count++;
910 	}
911 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
912 
913 	sch_tree_unlock(sch);
914 	return err;
915 }
916 
917 static void fq_destroy(struct Qdisc *sch)
918 {
919 	struct fq_sched_data *q = qdisc_priv(sch);
920 
921 	fq_reset(sch);
922 	fq_free(q->fq_root);
923 	qdisc_watchdog_cancel(&q->watchdog);
924 }
925 
926 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
927 		   struct netlink_ext_ack *extack)
928 {
929 	struct fq_sched_data *q = qdisc_priv(sch);
930 	int err;
931 
932 	sch->limit		= 10000;
933 	q->flow_plimit		= 100;
934 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
935 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
936 	q->flow_refill_delay	= msecs_to_jiffies(40);
937 	q->flow_max_rate	= ~0UL;
938 	q->time_next_delayed_flow = ~0ULL;
939 	q->rate_enable		= 1;
940 	q->new_flows.first	= NULL;
941 	q->old_flows.first	= NULL;
942 	q->delayed		= RB_ROOT;
943 	q->fq_root		= NULL;
944 	q->fq_trees_log		= ilog2(1024);
945 	q->orphan_mask		= 1024 - 1;
946 	q->low_rate_threshold	= 550000 / 8;
947 
948 	q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
949 
950 	q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
951 	q->horizon_drop = 1; /* by default, drop packets beyond horizon */
952 
953 	/* Default ce_threshold of 4294 seconds */
954 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
955 
956 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
957 
958 	if (opt)
959 		err = fq_change(sch, opt, extack);
960 	else
961 		err = fq_resize(sch, q->fq_trees_log);
962 
963 	return err;
964 }
965 
966 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
967 {
968 	struct fq_sched_data *q = qdisc_priv(sch);
969 	u64 ce_threshold = q->ce_threshold;
970 	u64 horizon = q->horizon;
971 	struct nlattr *opts;
972 
973 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
974 	if (opts == NULL)
975 		goto nla_put_failure;
976 
977 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
978 
979 	do_div(ce_threshold, NSEC_PER_USEC);
980 	do_div(horizon, NSEC_PER_USEC);
981 
982 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
983 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
984 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
985 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
986 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
987 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
988 			min_t(unsigned long, q->flow_max_rate, ~0U)) ||
989 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
990 			jiffies_to_usecs(q->flow_refill_delay)) ||
991 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
992 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
993 			q->low_rate_threshold) ||
994 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
995 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) ||
996 	    nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack) ||
997 	    nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
998 	    nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop))
999 		goto nla_put_failure;
1000 
1001 	return nla_nest_end(skb, opts);
1002 
1003 nla_put_failure:
1004 	return -1;
1005 }
1006 
1007 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1008 {
1009 	struct fq_sched_data *q = qdisc_priv(sch);
1010 	struct tc_fq_qd_stats st;
1011 
1012 	sch_tree_lock(sch);
1013 
1014 	st.gc_flows		  = q->stat_gc_flows;
1015 	st.highprio_packets	  = q->stat_internal_packets;
1016 	st.tcp_retrans		  = 0;
1017 	st.throttled		  = q->stat_throttled;
1018 	st.flows_plimit		  = q->stat_flows_plimit;
1019 	st.pkts_too_long	  = q->stat_pkts_too_long;
1020 	st.allocation_errors	  = q->stat_allocation_errors;
1021 	st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1022 				    ktime_get_ns();
1023 	st.flows		  = q->flows;
1024 	st.inactive_flows	  = q->inactive_flows;
1025 	st.throttled_flows	  = q->throttled_flows;
1026 	st.unthrottle_latency_ns  = min_t(unsigned long,
1027 					  q->unthrottle_latency_ns, ~0U);
1028 	st.ce_mark		  = q->stat_ce_mark;
1029 	st.horizon_drops	  = q->stat_horizon_drops;
1030 	st.horizon_caps		  = q->stat_horizon_caps;
1031 	sch_tree_unlock(sch);
1032 
1033 	return gnet_stats_copy_app(d, &st, sizeof(st));
1034 }
1035 
1036 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1037 	.id		=	"fq",
1038 	.priv_size	=	sizeof(struct fq_sched_data),
1039 
1040 	.enqueue	=	fq_enqueue,
1041 	.dequeue	=	fq_dequeue,
1042 	.peek		=	qdisc_peek_dequeued,
1043 	.init		=	fq_init,
1044 	.reset		=	fq_reset,
1045 	.destroy	=	fq_destroy,
1046 	.change		=	fq_change,
1047 	.dump		=	fq_dump,
1048 	.dump_stats	=	fq_dump_stats,
1049 	.owner		=	THIS_MODULE,
1050 };
1051 
1052 static int __init fq_module_init(void)
1053 {
1054 	int ret;
1055 
1056 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1057 					   sizeof(struct fq_flow),
1058 					   0, 0, NULL);
1059 	if (!fq_flow_cachep)
1060 		return -ENOMEM;
1061 
1062 	ret = register_qdisc(&fq_qdisc_ops);
1063 	if (ret)
1064 		kmem_cache_destroy(fq_flow_cachep);
1065 	return ret;
1066 }
1067 
1068 static void __exit fq_module_exit(void)
1069 {
1070 	unregister_qdisc(&fq_qdisc_ops);
1071 	kmem_cache_destroy(fq_flow_cachep);
1072 }
1073 
1074 module_init(fq_module_init)
1075 module_exit(fq_module_exit)
1076 MODULE_AUTHOR("Eric Dumazet");
1077 MODULE_LICENSE("GPL");
1078 MODULE_DESCRIPTION("Fair Queue Packet Scheduler");
1079