xref: /openbmc/linux/net/sched/sch_fq.c (revision 8bd1369b)
1 /*
2  * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
3  *
4  *  Copyright (C) 2013-2015 Eric Dumazet <edumazet@google.com>
5  *
6  *	This program is free software; you can redistribute it and/or
7  *	modify it under the terms of the GNU General Public License
8  *	as published by the Free Software Foundation; either version
9  *	2 of the License, or (at your option) any later version.
10  *
11  *  Meant to be mostly used for locally generated traffic :
12  *  Fast classification depends on skb->sk being set before reaching us.
13  *  If not, (router workload), we use rxhash as fallback, with 32 bits wide hash.
14  *  All packets belonging to a socket are considered as a 'flow'.
15  *
16  *  Flows are dynamically allocated and stored in a hash table of RB trees
17  *  They are also part of one Round Robin 'queues' (new or old flows)
18  *
19  *  Burst avoidance (aka pacing) capability :
20  *
21  *  Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a
22  *  bunch of packets, and this packet scheduler adds delay between
23  *  packets to respect rate limitation.
24  *
25  *  enqueue() :
26  *   - lookup one RB tree (out of 1024 or more) to find the flow.
27  *     If non existent flow, create it, add it to the tree.
28  *     Add skb to the per flow list of skb (fifo).
29  *   - Use a special fifo for high prio packets
30  *
31  *  dequeue() : serves flows in Round Robin
32  *  Note : When a flow becomes empty, we do not immediately remove it from
33  *  rb trees, for performance reasons (its expected to send additional packets,
34  *  or SLAB cache will reuse socket for another flow)
35  */
36 
37 #include <linux/module.h>
38 #include <linux/types.h>
39 #include <linux/kernel.h>
40 #include <linux/jiffies.h>
41 #include <linux/string.h>
42 #include <linux/in.h>
43 #include <linux/errno.h>
44 #include <linux/init.h>
45 #include <linux/skbuff.h>
46 #include <linux/slab.h>
47 #include <linux/rbtree.h>
48 #include <linux/hash.h>
49 #include <linux/prefetch.h>
50 #include <linux/vmalloc.h>
51 #include <net/netlink.h>
52 #include <net/pkt_sched.h>
53 #include <net/sock.h>
54 #include <net/tcp_states.h>
55 #include <net/tcp.h>
56 
57 /*
58  * Per flow structure, dynamically allocated
59  */
60 struct fq_flow {
61 	struct sk_buff	*head;		/* list of skbs for this flow : first skb */
62 	union {
63 		struct sk_buff *tail;	/* last skb in the list */
64 		unsigned long  age;	/* jiffies when flow was emptied, for gc */
65 	};
66 	struct rb_node	fq_node;	/* anchor in fq_root[] trees */
67 	struct sock	*sk;
68 	int		qlen;		/* number of packets in flow queue */
69 	int		credit;
70 	u32		socket_hash;	/* sk_hash */
71 	struct fq_flow *next;		/* next pointer in RR lists, or &detached */
72 
73 	struct rb_node  rate_node;	/* anchor in q->delayed tree */
74 	u64		time_next_packet;
75 };
76 
77 struct fq_flow_head {
78 	struct fq_flow *first;
79 	struct fq_flow *last;
80 };
81 
82 struct fq_sched_data {
83 	struct fq_flow_head new_flows;
84 
85 	struct fq_flow_head old_flows;
86 
87 	struct rb_root	delayed;	/* for rate limited flows */
88 	u64		time_next_delayed_flow;
89 	unsigned long	unthrottle_latency_ns;
90 
91 	struct fq_flow	internal;	/* for non classified or high prio packets */
92 	u32		quantum;
93 	u32		initial_quantum;
94 	u32		flow_refill_delay;
95 	u32		flow_max_rate;	/* optional max rate per flow */
96 	u32		flow_plimit;	/* max packets per flow */
97 	u32		orphan_mask;	/* mask for orphaned skb */
98 	u32		low_rate_threshold;
99 	struct rb_root	*fq_root;
100 	u8		rate_enable;
101 	u8		fq_trees_log;
102 
103 	u32		flows;
104 	u32		inactive_flows;
105 	u32		throttled_flows;
106 
107 	u64		stat_gc_flows;
108 	u64		stat_internal_packets;
109 	u64		stat_tcp_retrans;
110 	u64		stat_throttled;
111 	u64		stat_flows_plimit;
112 	u64		stat_pkts_too_long;
113 	u64		stat_allocation_errors;
114 	struct qdisc_watchdog watchdog;
115 };
116 
117 /* special value to mark a detached flow (not on old/new list) */
118 static struct fq_flow detached, throttled;
119 
120 static void fq_flow_set_detached(struct fq_flow *f)
121 {
122 	f->next = &detached;
123 	f->age = jiffies;
124 }
125 
126 static bool fq_flow_is_detached(const struct fq_flow *f)
127 {
128 	return f->next == &detached;
129 }
130 
131 static bool fq_flow_is_throttled(const struct fq_flow *f)
132 {
133 	return f->next == &throttled;
134 }
135 
136 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
137 {
138 	if (head->first)
139 		head->last->next = flow;
140 	else
141 		head->first = flow;
142 	head->last = flow;
143 	flow->next = NULL;
144 }
145 
146 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
147 {
148 	rb_erase(&f->rate_node, &q->delayed);
149 	q->throttled_flows--;
150 	fq_flow_add_tail(&q->old_flows, f);
151 }
152 
153 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
154 {
155 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
156 
157 	while (*p) {
158 		struct fq_flow *aux;
159 
160 		parent = *p;
161 		aux = rb_entry(parent, struct fq_flow, rate_node);
162 		if (f->time_next_packet >= aux->time_next_packet)
163 			p = &parent->rb_right;
164 		else
165 			p = &parent->rb_left;
166 	}
167 	rb_link_node(&f->rate_node, parent, p);
168 	rb_insert_color(&f->rate_node, &q->delayed);
169 	q->throttled_flows++;
170 	q->stat_throttled++;
171 
172 	f->next = &throttled;
173 	if (q->time_next_delayed_flow > f->time_next_packet)
174 		q->time_next_delayed_flow = f->time_next_packet;
175 }
176 
177 
178 static struct kmem_cache *fq_flow_cachep __read_mostly;
179 
180 
181 /* limit number of collected flows per round */
182 #define FQ_GC_MAX 8
183 #define FQ_GC_AGE (3*HZ)
184 
185 static bool fq_gc_candidate(const struct fq_flow *f)
186 {
187 	return fq_flow_is_detached(f) &&
188 	       time_after(jiffies, f->age + FQ_GC_AGE);
189 }
190 
191 static void fq_gc(struct fq_sched_data *q,
192 		  struct rb_root *root,
193 		  struct sock *sk)
194 {
195 	struct fq_flow *f, *tofree[FQ_GC_MAX];
196 	struct rb_node **p, *parent;
197 	int fcnt = 0;
198 
199 	p = &root->rb_node;
200 	parent = NULL;
201 	while (*p) {
202 		parent = *p;
203 
204 		f = rb_entry(parent, struct fq_flow, fq_node);
205 		if (f->sk == sk)
206 			break;
207 
208 		if (fq_gc_candidate(f)) {
209 			tofree[fcnt++] = f;
210 			if (fcnt == FQ_GC_MAX)
211 				break;
212 		}
213 
214 		if (f->sk > sk)
215 			p = &parent->rb_right;
216 		else
217 			p = &parent->rb_left;
218 	}
219 
220 	q->flows -= fcnt;
221 	q->inactive_flows -= fcnt;
222 	q->stat_gc_flows += fcnt;
223 	while (fcnt) {
224 		struct fq_flow *f = tofree[--fcnt];
225 
226 		rb_erase(&f->fq_node, root);
227 		kmem_cache_free(fq_flow_cachep, f);
228 	}
229 }
230 
231 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
232 {
233 	struct rb_node **p, *parent;
234 	struct sock *sk = skb->sk;
235 	struct rb_root *root;
236 	struct fq_flow *f;
237 
238 	/* warning: no starvation prevention... */
239 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
240 		return &q->internal;
241 
242 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
243 	 * or a listener (SYNCOOKIE mode)
244 	 * 1) request sockets are not full blown,
245 	 *    they do not contain sk_pacing_rate
246 	 * 2) They are not part of a 'flow' yet
247 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
248 	 *    especially if the listener set SO_MAX_PACING_RATE
249 	 * 4) We pretend they are orphaned
250 	 */
251 	if (!sk || sk_listener(sk)) {
252 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
253 
254 		/* By forcing low order bit to 1, we make sure to not
255 		 * collide with a local flow (socket pointers are word aligned)
256 		 */
257 		sk = (struct sock *)((hash << 1) | 1UL);
258 		skb_orphan(skb);
259 	}
260 
261 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
262 
263 	if (q->flows >= (2U << q->fq_trees_log) &&
264 	    q->inactive_flows > q->flows/2)
265 		fq_gc(q, root, sk);
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 			/* socket might have been reallocated, so check
275 			 * if its sk_hash is the same.
276 			 * It not, we need to refill credit with
277 			 * initial quantum
278 			 */
279 			if (unlikely(skb->sk &&
280 				     f->socket_hash != sk->sk_hash)) {
281 				f->credit = q->initial_quantum;
282 				f->socket_hash = sk->sk_hash;
283 				if (fq_flow_is_throttled(f))
284 					fq_flow_unset_throttled(q, f);
285 				f->time_next_packet = 0ULL;
286 			}
287 			return f;
288 		}
289 		if (f->sk > sk)
290 			p = &parent->rb_right;
291 		else
292 			p = &parent->rb_left;
293 	}
294 
295 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
296 	if (unlikely(!f)) {
297 		q->stat_allocation_errors++;
298 		return &q->internal;
299 	}
300 	fq_flow_set_detached(f);
301 	f->sk = sk;
302 	if (skb->sk)
303 		f->socket_hash = sk->sk_hash;
304 	f->credit = q->initial_quantum;
305 
306 	rb_link_node(&f->fq_node, parent, p);
307 	rb_insert_color(&f->fq_node, root);
308 
309 	q->flows++;
310 	q->inactive_flows++;
311 	return f;
312 }
313 
314 
315 /* remove one skb from head of flow queue */
316 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
317 {
318 	struct sk_buff *skb = flow->head;
319 
320 	if (skb) {
321 		flow->head = skb->next;
322 		skb->next = NULL;
323 		flow->qlen--;
324 		qdisc_qstats_backlog_dec(sch, skb);
325 		sch->q.qlen--;
326 	}
327 	return skb;
328 }
329 
330 /* We might add in the future detection of retransmits
331  * For the time being, just return false
332  */
333 static bool skb_is_retransmit(struct sk_buff *skb)
334 {
335 	return false;
336 }
337 
338 /* add skb to flow queue
339  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
340  * We special case tcp retransmits to be transmitted before other packets.
341  * We rely on fact that TCP retransmits are unlikely, so we do not waste
342  * a separate queue or a pointer.
343  * head->  [retrans pkt 1]
344  *         [retrans pkt 2]
345  *         [ normal pkt 1]
346  *         [ normal pkt 2]
347  *         [ normal pkt 3]
348  * tail->  [ normal pkt 4]
349  */
350 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
351 {
352 	struct sk_buff *prev, *head = flow->head;
353 
354 	skb->next = NULL;
355 	if (!head) {
356 		flow->head = skb;
357 		flow->tail = skb;
358 		return;
359 	}
360 	if (likely(!skb_is_retransmit(skb))) {
361 		flow->tail->next = skb;
362 		flow->tail = skb;
363 		return;
364 	}
365 
366 	/* This skb is a tcp retransmit,
367 	 * find the last retrans packet in the queue
368 	 */
369 	prev = NULL;
370 	while (skb_is_retransmit(head)) {
371 		prev = head;
372 		head = head->next;
373 		if (!head)
374 			break;
375 	}
376 	if (!prev) { /* no rtx packet in queue, become the new head */
377 		skb->next = flow->head;
378 		flow->head = skb;
379 	} else {
380 		if (prev == flow->tail)
381 			flow->tail = skb;
382 		else
383 			skb->next = prev->next;
384 		prev->next = skb;
385 	}
386 }
387 
388 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
389 		      struct sk_buff **to_free)
390 {
391 	struct fq_sched_data *q = qdisc_priv(sch);
392 	struct fq_flow *f;
393 
394 	if (unlikely(sch->q.qlen >= sch->limit))
395 		return qdisc_drop(skb, sch, to_free);
396 
397 	f = fq_classify(skb, q);
398 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
399 		q->stat_flows_plimit++;
400 		return qdisc_drop(skb, sch, to_free);
401 	}
402 
403 	f->qlen++;
404 	if (skb_is_retransmit(skb))
405 		q->stat_tcp_retrans++;
406 	qdisc_qstats_backlog_inc(sch, skb);
407 	if (fq_flow_is_detached(f)) {
408 		struct sock *sk = skb->sk;
409 
410 		fq_flow_add_tail(&q->new_flows, f);
411 		if (time_after(jiffies, f->age + q->flow_refill_delay))
412 			f->credit = max_t(u32, f->credit, q->quantum);
413 		if (sk && q->rate_enable) {
414 			if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
415 				     SK_PACING_FQ))
416 				smp_store_release(&sk->sk_pacing_status,
417 						  SK_PACING_FQ);
418 		}
419 		q->inactive_flows--;
420 	}
421 
422 	/* Note: this overwrites f->age */
423 	flow_queue_add(f, skb);
424 
425 	if (unlikely(f == &q->internal)) {
426 		q->stat_internal_packets++;
427 	}
428 	sch->q.qlen++;
429 
430 	return NET_XMIT_SUCCESS;
431 }
432 
433 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
434 {
435 	unsigned long sample;
436 	struct rb_node *p;
437 
438 	if (q->time_next_delayed_flow > now)
439 		return;
440 
441 	/* Update unthrottle latency EWMA.
442 	 * This is cheap and can help diagnosing timer/latency problems.
443 	 */
444 	sample = (unsigned long)(now - q->time_next_delayed_flow);
445 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
446 	q->unthrottle_latency_ns += sample >> 3;
447 
448 	q->time_next_delayed_flow = ~0ULL;
449 	while ((p = rb_first(&q->delayed)) != NULL) {
450 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
451 
452 		if (f->time_next_packet > now) {
453 			q->time_next_delayed_flow = f->time_next_packet;
454 			break;
455 		}
456 		fq_flow_unset_throttled(q, f);
457 	}
458 }
459 
460 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
461 {
462 	struct fq_sched_data *q = qdisc_priv(sch);
463 	u64 now = ktime_get_ns();
464 	struct fq_flow_head *head;
465 	struct sk_buff *skb;
466 	struct fq_flow *f;
467 	u32 rate, plen;
468 
469 	skb = fq_dequeue_head(sch, &q->internal);
470 	if (skb)
471 		goto out;
472 	fq_check_throttled(q, now);
473 begin:
474 	head = &q->new_flows;
475 	if (!head->first) {
476 		head = &q->old_flows;
477 		if (!head->first) {
478 			if (q->time_next_delayed_flow != ~0ULL)
479 				qdisc_watchdog_schedule_ns(&q->watchdog,
480 							   q->time_next_delayed_flow);
481 			return NULL;
482 		}
483 	}
484 	f = head->first;
485 
486 	if (f->credit <= 0) {
487 		f->credit += q->quantum;
488 		head->first = f->next;
489 		fq_flow_add_tail(&q->old_flows, f);
490 		goto begin;
491 	}
492 
493 	skb = f->head;
494 	if (unlikely(skb && now < f->time_next_packet &&
495 		     !skb_is_tcp_pure_ack(skb))) {
496 		head->first = f->next;
497 		fq_flow_set_throttled(q, f);
498 		goto begin;
499 	}
500 
501 	skb = fq_dequeue_head(sch, f);
502 	if (!skb) {
503 		head->first = f->next;
504 		/* force a pass through old_flows to prevent starvation */
505 		if ((head == &q->new_flows) && q->old_flows.first) {
506 			fq_flow_add_tail(&q->old_flows, f);
507 		} else {
508 			fq_flow_set_detached(f);
509 			q->inactive_flows++;
510 		}
511 		goto begin;
512 	}
513 	prefetch(&skb->end);
514 	f->credit -= qdisc_pkt_len(skb);
515 
516 	if (!q->rate_enable)
517 		goto out;
518 
519 	/* Do not pace locally generated ack packets */
520 	if (skb_is_tcp_pure_ack(skb))
521 		goto out;
522 
523 	rate = q->flow_max_rate;
524 	if (skb->sk)
525 		rate = min(skb->sk->sk_pacing_rate, rate);
526 
527 	if (rate <= q->low_rate_threshold) {
528 		f->credit = 0;
529 		plen = qdisc_pkt_len(skb);
530 	} else {
531 		plen = max(qdisc_pkt_len(skb), q->quantum);
532 		if (f->credit > 0)
533 			goto out;
534 	}
535 	if (rate != ~0U) {
536 		u64 len = (u64)plen * NSEC_PER_SEC;
537 
538 		if (likely(rate))
539 			do_div(len, rate);
540 		/* Since socket rate can change later,
541 		 * clamp the delay to 1 second.
542 		 * Really, providers of too big packets should be fixed !
543 		 */
544 		if (unlikely(len > NSEC_PER_SEC)) {
545 			len = NSEC_PER_SEC;
546 			q->stat_pkts_too_long++;
547 		}
548 		/* Account for schedule/timers drifts.
549 		 * f->time_next_packet was set when prior packet was sent,
550 		 * and current time (@now) can be too late by tens of us.
551 		 */
552 		if (f->time_next_packet)
553 			len -= min(len/2, now - f->time_next_packet);
554 		f->time_next_packet = now + len;
555 	}
556 out:
557 	qdisc_bstats_update(sch, skb);
558 	return skb;
559 }
560 
561 static void fq_flow_purge(struct fq_flow *flow)
562 {
563 	rtnl_kfree_skbs(flow->head, flow->tail);
564 	flow->head = NULL;
565 	flow->qlen = 0;
566 }
567 
568 static void fq_reset(struct Qdisc *sch)
569 {
570 	struct fq_sched_data *q = qdisc_priv(sch);
571 	struct rb_root *root;
572 	struct rb_node *p;
573 	struct fq_flow *f;
574 	unsigned int idx;
575 
576 	sch->q.qlen = 0;
577 	sch->qstats.backlog = 0;
578 
579 	fq_flow_purge(&q->internal);
580 
581 	if (!q->fq_root)
582 		return;
583 
584 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
585 		root = &q->fq_root[idx];
586 		while ((p = rb_first(root)) != NULL) {
587 			f = rb_entry(p, struct fq_flow, fq_node);
588 			rb_erase(p, root);
589 
590 			fq_flow_purge(f);
591 
592 			kmem_cache_free(fq_flow_cachep, f);
593 		}
594 	}
595 	q->new_flows.first	= NULL;
596 	q->old_flows.first	= NULL;
597 	q->delayed		= RB_ROOT;
598 	q->flows		= 0;
599 	q->inactive_flows	= 0;
600 	q->throttled_flows	= 0;
601 }
602 
603 static void fq_rehash(struct fq_sched_data *q,
604 		      struct rb_root *old_array, u32 old_log,
605 		      struct rb_root *new_array, u32 new_log)
606 {
607 	struct rb_node *op, **np, *parent;
608 	struct rb_root *oroot, *nroot;
609 	struct fq_flow *of, *nf;
610 	int fcnt = 0;
611 	u32 idx;
612 
613 	for (idx = 0; idx < (1U << old_log); idx++) {
614 		oroot = &old_array[idx];
615 		while ((op = rb_first(oroot)) != NULL) {
616 			rb_erase(op, oroot);
617 			of = rb_entry(op, struct fq_flow, fq_node);
618 			if (fq_gc_candidate(of)) {
619 				fcnt++;
620 				kmem_cache_free(fq_flow_cachep, of);
621 				continue;
622 			}
623 			nroot = &new_array[hash_ptr(of->sk, new_log)];
624 
625 			np = &nroot->rb_node;
626 			parent = NULL;
627 			while (*np) {
628 				parent = *np;
629 
630 				nf = rb_entry(parent, struct fq_flow, fq_node);
631 				BUG_ON(nf->sk == of->sk);
632 
633 				if (nf->sk > of->sk)
634 					np = &parent->rb_right;
635 				else
636 					np = &parent->rb_left;
637 			}
638 
639 			rb_link_node(&of->fq_node, parent, np);
640 			rb_insert_color(&of->fq_node, nroot);
641 		}
642 	}
643 	q->flows -= fcnt;
644 	q->inactive_flows -= fcnt;
645 	q->stat_gc_flows += fcnt;
646 }
647 
648 static void fq_free(void *addr)
649 {
650 	kvfree(addr);
651 }
652 
653 static int fq_resize(struct Qdisc *sch, u32 log)
654 {
655 	struct fq_sched_data *q = qdisc_priv(sch);
656 	struct rb_root *array;
657 	void *old_fq_root;
658 	u32 idx;
659 
660 	if (q->fq_root && log == q->fq_trees_log)
661 		return 0;
662 
663 	/* If XPS was setup, we can allocate memory on right NUMA node */
664 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
665 			      netdev_queue_numa_node_read(sch->dev_queue));
666 	if (!array)
667 		return -ENOMEM;
668 
669 	for (idx = 0; idx < (1U << log); idx++)
670 		array[idx] = RB_ROOT;
671 
672 	sch_tree_lock(sch);
673 
674 	old_fq_root = q->fq_root;
675 	if (old_fq_root)
676 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
677 
678 	q->fq_root = array;
679 	q->fq_trees_log = log;
680 
681 	sch_tree_unlock(sch);
682 
683 	fq_free(old_fq_root);
684 
685 	return 0;
686 }
687 
688 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
689 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
690 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
691 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
692 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
693 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
694 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
695 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
696 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
697 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
698 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
699 };
700 
701 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
702 		     struct netlink_ext_ack *extack)
703 {
704 	struct fq_sched_data *q = qdisc_priv(sch);
705 	struct nlattr *tb[TCA_FQ_MAX + 1];
706 	int err, drop_count = 0;
707 	unsigned drop_len = 0;
708 	u32 fq_log;
709 
710 	if (!opt)
711 		return -EINVAL;
712 
713 	err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy, NULL);
714 	if (err < 0)
715 		return err;
716 
717 	sch_tree_lock(sch);
718 
719 	fq_log = q->fq_trees_log;
720 
721 	if (tb[TCA_FQ_BUCKETS_LOG]) {
722 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
723 
724 		if (nval >= 1 && nval <= ilog2(256*1024))
725 			fq_log = nval;
726 		else
727 			err = -EINVAL;
728 	}
729 	if (tb[TCA_FQ_PLIMIT])
730 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
731 
732 	if (tb[TCA_FQ_FLOW_PLIMIT])
733 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
734 
735 	if (tb[TCA_FQ_QUANTUM]) {
736 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
737 
738 		if (quantum > 0)
739 			q->quantum = quantum;
740 		else
741 			err = -EINVAL;
742 	}
743 
744 	if (tb[TCA_FQ_INITIAL_QUANTUM])
745 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
746 
747 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
748 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
749 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
750 
751 	if (tb[TCA_FQ_FLOW_MAX_RATE])
752 		q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
753 
754 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
755 		q->low_rate_threshold =
756 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
757 
758 	if (tb[TCA_FQ_RATE_ENABLE]) {
759 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
760 
761 		if (enable <= 1)
762 			q->rate_enable = enable;
763 		else
764 			err = -EINVAL;
765 	}
766 
767 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
768 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
769 
770 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
771 	}
772 
773 	if (tb[TCA_FQ_ORPHAN_MASK])
774 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
775 
776 	if (!err) {
777 		sch_tree_unlock(sch);
778 		err = fq_resize(sch, fq_log);
779 		sch_tree_lock(sch);
780 	}
781 	while (sch->q.qlen > sch->limit) {
782 		struct sk_buff *skb = fq_dequeue(sch);
783 
784 		if (!skb)
785 			break;
786 		drop_len += qdisc_pkt_len(skb);
787 		rtnl_kfree_skbs(skb, skb);
788 		drop_count++;
789 	}
790 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
791 
792 	sch_tree_unlock(sch);
793 	return err;
794 }
795 
796 static void fq_destroy(struct Qdisc *sch)
797 {
798 	struct fq_sched_data *q = qdisc_priv(sch);
799 
800 	fq_reset(sch);
801 	fq_free(q->fq_root);
802 	qdisc_watchdog_cancel(&q->watchdog);
803 }
804 
805 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
806 		   struct netlink_ext_ack *extack)
807 {
808 	struct fq_sched_data *q = qdisc_priv(sch);
809 	int err;
810 
811 	sch->limit		= 10000;
812 	q->flow_plimit		= 100;
813 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
814 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
815 	q->flow_refill_delay	= msecs_to_jiffies(40);
816 	q->flow_max_rate	= ~0U;
817 	q->time_next_delayed_flow = ~0ULL;
818 	q->rate_enable		= 1;
819 	q->new_flows.first	= NULL;
820 	q->old_flows.first	= NULL;
821 	q->delayed		= RB_ROOT;
822 	q->fq_root		= NULL;
823 	q->fq_trees_log		= ilog2(1024);
824 	q->orphan_mask		= 1024 - 1;
825 	q->low_rate_threshold	= 550000 / 8;
826 	qdisc_watchdog_init(&q->watchdog, sch);
827 
828 	if (opt)
829 		err = fq_change(sch, opt, extack);
830 	else
831 		err = fq_resize(sch, q->fq_trees_log);
832 
833 	return err;
834 }
835 
836 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
837 {
838 	struct fq_sched_data *q = qdisc_priv(sch);
839 	struct nlattr *opts;
840 
841 	opts = nla_nest_start(skb, TCA_OPTIONS);
842 	if (opts == NULL)
843 		goto nla_put_failure;
844 
845 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
846 
847 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
848 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
849 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
850 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
851 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
852 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
853 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
854 			jiffies_to_usecs(q->flow_refill_delay)) ||
855 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
856 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
857 			q->low_rate_threshold) ||
858 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
859 		goto nla_put_failure;
860 
861 	return nla_nest_end(skb, opts);
862 
863 nla_put_failure:
864 	return -1;
865 }
866 
867 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
868 {
869 	struct fq_sched_data *q = qdisc_priv(sch);
870 	struct tc_fq_qd_stats st;
871 
872 	sch_tree_lock(sch);
873 
874 	st.gc_flows		  = q->stat_gc_flows;
875 	st.highprio_packets	  = q->stat_internal_packets;
876 	st.tcp_retrans		  = q->stat_tcp_retrans;
877 	st.throttled		  = q->stat_throttled;
878 	st.flows_plimit		  = q->stat_flows_plimit;
879 	st.pkts_too_long	  = q->stat_pkts_too_long;
880 	st.allocation_errors	  = q->stat_allocation_errors;
881 	st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
882 	st.flows		  = q->flows;
883 	st.inactive_flows	  = q->inactive_flows;
884 	st.throttled_flows	  = q->throttled_flows;
885 	st.unthrottle_latency_ns  = min_t(unsigned long,
886 					  q->unthrottle_latency_ns, ~0U);
887 	sch_tree_unlock(sch);
888 
889 	return gnet_stats_copy_app(d, &st, sizeof(st));
890 }
891 
892 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
893 	.id		=	"fq",
894 	.priv_size	=	sizeof(struct fq_sched_data),
895 
896 	.enqueue	=	fq_enqueue,
897 	.dequeue	=	fq_dequeue,
898 	.peek		=	qdisc_peek_dequeued,
899 	.init		=	fq_init,
900 	.reset		=	fq_reset,
901 	.destroy	=	fq_destroy,
902 	.change		=	fq_change,
903 	.dump		=	fq_dump,
904 	.dump_stats	=	fq_dump_stats,
905 	.owner		=	THIS_MODULE,
906 };
907 
908 static int __init fq_module_init(void)
909 {
910 	int ret;
911 
912 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
913 					   sizeof(struct fq_flow),
914 					   0, 0, NULL);
915 	if (!fq_flow_cachep)
916 		return -ENOMEM;
917 
918 	ret = register_qdisc(&fq_qdisc_ops);
919 	if (ret)
920 		kmem_cache_destroy(fq_flow_cachep);
921 	return ret;
922 }
923 
924 static void __exit fq_module_exit(void)
925 {
926 	unregister_qdisc(&fq_qdisc_ops);
927 	kmem_cache_destroy(fq_flow_cachep);
928 }
929 
930 module_init(fq_module_init)
931 module_exit(fq_module_exit)
932 MODULE_AUTHOR("Eric Dumazet");
933 MODULE_LICENSE("GPL");
934