xref: /openbmc/linux/net/sched/sch_fq.c (revision 3a448205)
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_plimit;	/* max packets per flow */
96 	unsigned long	flow_max_rate;	/* optional max rate per flow */
97 	u64		ce_threshold;
98 	u32		orphan_mask;	/* mask for orphaned skb */
99 	u32		low_rate_threshold;
100 	struct rb_root	*fq_root;
101 	u8		rate_enable;
102 	u8		fq_trees_log;
103 
104 	u32		flows;
105 	u32		inactive_flows;
106 	u32		throttled_flows;
107 
108 	u64		stat_gc_flows;
109 	u64		stat_internal_packets;
110 	u64		stat_throttled;
111 	u64		stat_ce_mark;
112 	u64		stat_flows_plimit;
113 	u64		stat_pkts_too_long;
114 	u64		stat_allocation_errors;
115 	struct qdisc_watchdog watchdog;
116 };
117 
118 /* special value to mark a detached flow (not on old/new list) */
119 static struct fq_flow detached, throttled;
120 
121 static void fq_flow_set_detached(struct fq_flow *f)
122 {
123 	f->next = &detached;
124 	f->age = jiffies;
125 }
126 
127 static bool fq_flow_is_detached(const struct fq_flow *f)
128 {
129 	return f->next == &detached;
130 }
131 
132 static bool fq_flow_is_throttled(const struct fq_flow *f)
133 {
134 	return f->next == &throttled;
135 }
136 
137 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
138 {
139 	if (head->first)
140 		head->last->next = flow;
141 	else
142 		head->first = flow;
143 	head->last = flow;
144 	flow->next = NULL;
145 }
146 
147 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
148 {
149 	rb_erase(&f->rate_node, &q->delayed);
150 	q->throttled_flows--;
151 	fq_flow_add_tail(&q->old_flows, f);
152 }
153 
154 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
155 {
156 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
157 
158 	while (*p) {
159 		struct fq_flow *aux;
160 
161 		parent = *p;
162 		aux = rb_entry(parent, struct fq_flow, rate_node);
163 		if (f->time_next_packet >= aux->time_next_packet)
164 			p = &parent->rb_right;
165 		else
166 			p = &parent->rb_left;
167 	}
168 	rb_link_node(&f->rate_node, parent, p);
169 	rb_insert_color(&f->rate_node, &q->delayed);
170 	q->throttled_flows++;
171 	q->stat_throttled++;
172 
173 	f->next = &throttled;
174 	if (q->time_next_delayed_flow > f->time_next_packet)
175 		q->time_next_delayed_flow = f->time_next_packet;
176 }
177 
178 
179 static struct kmem_cache *fq_flow_cachep __read_mostly;
180 
181 
182 /* limit number of collected flows per round */
183 #define FQ_GC_MAX 8
184 #define FQ_GC_AGE (3*HZ)
185 
186 static bool fq_gc_candidate(const struct fq_flow *f)
187 {
188 	return fq_flow_is_detached(f) &&
189 	       time_after(jiffies, f->age + FQ_GC_AGE);
190 }
191 
192 static void fq_gc(struct fq_sched_data *q,
193 		  struct rb_root *root,
194 		  struct sock *sk)
195 {
196 	struct fq_flow *f, *tofree[FQ_GC_MAX];
197 	struct rb_node **p, *parent;
198 	int fcnt = 0;
199 
200 	p = &root->rb_node;
201 	parent = NULL;
202 	while (*p) {
203 		parent = *p;
204 
205 		f = rb_entry(parent, struct fq_flow, fq_node);
206 		if (f->sk == sk)
207 			break;
208 
209 		if (fq_gc_candidate(f)) {
210 			tofree[fcnt++] = f;
211 			if (fcnt == FQ_GC_MAX)
212 				break;
213 		}
214 
215 		if (f->sk > sk)
216 			p = &parent->rb_right;
217 		else
218 			p = &parent->rb_left;
219 	}
220 
221 	q->flows -= fcnt;
222 	q->inactive_flows -= fcnt;
223 	q->stat_gc_flows += fcnt;
224 	while (fcnt) {
225 		struct fq_flow *f = tofree[--fcnt];
226 
227 		rb_erase(&f->fq_node, root);
228 		kmem_cache_free(fq_flow_cachep, f);
229 	}
230 }
231 
232 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
233 {
234 	struct rb_node **p, *parent;
235 	struct sock *sk = skb->sk;
236 	struct rb_root *root;
237 	struct fq_flow *f;
238 
239 	/* warning: no starvation prevention... */
240 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
241 		return &q->internal;
242 
243 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
244 	 * or a listener (SYNCOOKIE mode)
245 	 * 1) request sockets are not full blown,
246 	 *    they do not contain sk_pacing_rate
247 	 * 2) They are not part of a 'flow' yet
248 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
249 	 *    especially if the listener set SO_MAX_PACING_RATE
250 	 * 4) We pretend they are orphaned
251 	 */
252 	if (!sk || sk_listener(sk)) {
253 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
254 
255 		/* By forcing low order bit to 1, we make sure to not
256 		 * collide with a local flow (socket pointers are word aligned)
257 		 */
258 		sk = (struct sock *)((hash << 1) | 1UL);
259 		skb_orphan(skb);
260 	}
261 
262 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
263 
264 	if (q->flows >= (2U << q->fq_trees_log) &&
265 	    q->inactive_flows > q->flows/2)
266 		fq_gc(q, root, sk);
267 
268 	p = &root->rb_node;
269 	parent = NULL;
270 	while (*p) {
271 		parent = *p;
272 
273 		f = rb_entry(parent, struct fq_flow, fq_node);
274 		if (f->sk == sk) {
275 			/* socket might have been reallocated, so check
276 			 * if its sk_hash is the same.
277 			 * It not, we need to refill credit with
278 			 * initial quantum
279 			 */
280 			if (unlikely(skb->sk &&
281 				     f->socket_hash != sk->sk_hash)) {
282 				f->credit = q->initial_quantum;
283 				f->socket_hash = sk->sk_hash;
284 				if (fq_flow_is_throttled(f))
285 					fq_flow_unset_throttled(q, f);
286 				f->time_next_packet = 0ULL;
287 			}
288 			return f;
289 		}
290 		if (f->sk > sk)
291 			p = &parent->rb_right;
292 		else
293 			p = &parent->rb_left;
294 	}
295 
296 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
297 	if (unlikely(!f)) {
298 		q->stat_allocation_errors++;
299 		return &q->internal;
300 	}
301 	fq_flow_set_detached(f);
302 	f->sk = sk;
303 	if (skb->sk)
304 		f->socket_hash = sk->sk_hash;
305 	f->credit = q->initial_quantum;
306 
307 	rb_link_node(&f->fq_node, parent, p);
308 	rb_insert_color(&f->fq_node, root);
309 
310 	q->flows++;
311 	q->inactive_flows++;
312 	return f;
313 }
314 
315 
316 /* remove one skb from head of flow queue */
317 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
318 {
319 	struct sk_buff *skb = flow->head;
320 
321 	if (skb) {
322 		flow->head = skb->next;
323 		skb_mark_not_on_list(skb);
324 		flow->qlen--;
325 		qdisc_qstats_backlog_dec(sch, skb);
326 		sch->q.qlen--;
327 	}
328 	return skb;
329 }
330 
331 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
332 {
333 	struct sk_buff *head = flow->head;
334 
335 	skb->next = NULL;
336 	if (!head)
337 		flow->head = skb;
338 	else
339 		flow->tail->next = skb;
340 
341 	flow->tail = skb;
342 }
343 
344 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
345 		      struct sk_buff **to_free)
346 {
347 	struct fq_sched_data *q = qdisc_priv(sch);
348 	struct fq_flow *f;
349 
350 	if (unlikely(sch->q.qlen >= sch->limit))
351 		return qdisc_drop(skb, sch, to_free);
352 
353 	f = fq_classify(skb, q);
354 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
355 		q->stat_flows_plimit++;
356 		return qdisc_drop(skb, sch, to_free);
357 	}
358 
359 	f->qlen++;
360 	qdisc_qstats_backlog_inc(sch, skb);
361 	if (fq_flow_is_detached(f)) {
362 		struct sock *sk = skb->sk;
363 
364 		fq_flow_add_tail(&q->new_flows, f);
365 		if (time_after(jiffies, f->age + q->flow_refill_delay))
366 			f->credit = max_t(u32, f->credit, q->quantum);
367 		if (sk && q->rate_enable) {
368 			if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
369 				     SK_PACING_FQ))
370 				smp_store_release(&sk->sk_pacing_status,
371 						  SK_PACING_FQ);
372 		}
373 		q->inactive_flows--;
374 	}
375 
376 	/* Note: this overwrites f->age */
377 	flow_queue_add(f, skb);
378 
379 	if (unlikely(f == &q->internal)) {
380 		q->stat_internal_packets++;
381 	}
382 	sch->q.qlen++;
383 
384 	return NET_XMIT_SUCCESS;
385 }
386 
387 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
388 {
389 	unsigned long sample;
390 	struct rb_node *p;
391 
392 	if (q->time_next_delayed_flow > now)
393 		return;
394 
395 	/* Update unthrottle latency EWMA.
396 	 * This is cheap and can help diagnosing timer/latency problems.
397 	 */
398 	sample = (unsigned long)(now - q->time_next_delayed_flow);
399 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
400 	q->unthrottle_latency_ns += sample >> 3;
401 
402 	q->time_next_delayed_flow = ~0ULL;
403 	while ((p = rb_first(&q->delayed)) != NULL) {
404 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
405 
406 		if (f->time_next_packet > now) {
407 			q->time_next_delayed_flow = f->time_next_packet;
408 			break;
409 		}
410 		fq_flow_unset_throttled(q, f);
411 	}
412 }
413 
414 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
415 {
416 	struct fq_sched_data *q = qdisc_priv(sch);
417 	u64 now = ktime_get_ns();
418 	struct fq_flow_head *head;
419 	struct sk_buff *skb;
420 	struct fq_flow *f;
421 	unsigned long rate;
422 	u32 plen;
423 
424 	skb = fq_dequeue_head(sch, &q->internal);
425 	if (skb)
426 		goto out;
427 	fq_check_throttled(q, now);
428 begin:
429 	head = &q->new_flows;
430 	if (!head->first) {
431 		head = &q->old_flows;
432 		if (!head->first) {
433 			if (q->time_next_delayed_flow != ~0ULL)
434 				qdisc_watchdog_schedule_ns(&q->watchdog,
435 							   q->time_next_delayed_flow);
436 			return NULL;
437 		}
438 	}
439 	f = head->first;
440 
441 	if (f->credit <= 0) {
442 		f->credit += q->quantum;
443 		head->first = f->next;
444 		fq_flow_add_tail(&q->old_flows, f);
445 		goto begin;
446 	}
447 
448 	skb = f->head;
449 	if (skb) {
450 		u64 time_next_packet = max_t(u64, ktime_to_ns(skb->tstamp),
451 					     f->time_next_packet);
452 
453 		if (now < time_next_packet) {
454 			head->first = f->next;
455 			f->time_next_packet = time_next_packet;
456 			fq_flow_set_throttled(q, f);
457 			goto begin;
458 		}
459 		if (time_next_packet &&
460 		    (s64)(now - time_next_packet - q->ce_threshold) > 0) {
461 			INET_ECN_set_ce(skb);
462 			q->stat_ce_mark++;
463 		}
464 	}
465 
466 	skb = fq_dequeue_head(sch, f);
467 	if (!skb) {
468 		head->first = f->next;
469 		/* force a pass through old_flows to prevent starvation */
470 		if ((head == &q->new_flows) && q->old_flows.first) {
471 			fq_flow_add_tail(&q->old_flows, f);
472 		} else {
473 			fq_flow_set_detached(f);
474 			q->inactive_flows++;
475 		}
476 		goto begin;
477 	}
478 	prefetch(&skb->end);
479 	plen = qdisc_pkt_len(skb);
480 	f->credit -= plen;
481 
482 	if (!q->rate_enable)
483 		goto out;
484 
485 	rate = q->flow_max_rate;
486 
487 	/* If EDT time was provided for this skb, we need to
488 	 * update f->time_next_packet only if this qdisc enforces
489 	 * a flow max rate.
490 	 */
491 	if (!skb->tstamp) {
492 		if (skb->sk)
493 			rate = min(skb->sk->sk_pacing_rate, rate);
494 
495 		if (rate <= q->low_rate_threshold) {
496 			f->credit = 0;
497 		} else {
498 			plen = max(plen, q->quantum);
499 			if (f->credit > 0)
500 				goto out;
501 		}
502 	}
503 	if (rate != ~0UL) {
504 		u64 len = (u64)plen * NSEC_PER_SEC;
505 
506 		if (likely(rate))
507 			len = div64_ul(len, rate);
508 		/* Since socket rate can change later,
509 		 * clamp the delay to 1 second.
510 		 * Really, providers of too big packets should be fixed !
511 		 */
512 		if (unlikely(len > NSEC_PER_SEC)) {
513 			len = NSEC_PER_SEC;
514 			q->stat_pkts_too_long++;
515 		}
516 		/* Account for schedule/timers drifts.
517 		 * f->time_next_packet was set when prior packet was sent,
518 		 * and current time (@now) can be too late by tens of us.
519 		 */
520 		if (f->time_next_packet)
521 			len -= min(len/2, now - f->time_next_packet);
522 		f->time_next_packet = now + len;
523 	}
524 out:
525 	qdisc_bstats_update(sch, skb);
526 	return skb;
527 }
528 
529 static void fq_flow_purge(struct fq_flow *flow)
530 {
531 	rtnl_kfree_skbs(flow->head, flow->tail);
532 	flow->head = NULL;
533 	flow->qlen = 0;
534 }
535 
536 static void fq_reset(struct Qdisc *sch)
537 {
538 	struct fq_sched_data *q = qdisc_priv(sch);
539 	struct rb_root *root;
540 	struct rb_node *p;
541 	struct fq_flow *f;
542 	unsigned int idx;
543 
544 	sch->q.qlen = 0;
545 	sch->qstats.backlog = 0;
546 
547 	fq_flow_purge(&q->internal);
548 
549 	if (!q->fq_root)
550 		return;
551 
552 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
553 		root = &q->fq_root[idx];
554 		while ((p = rb_first(root)) != NULL) {
555 			f = rb_entry(p, struct fq_flow, fq_node);
556 			rb_erase(p, root);
557 
558 			fq_flow_purge(f);
559 
560 			kmem_cache_free(fq_flow_cachep, f);
561 		}
562 	}
563 	q->new_flows.first	= NULL;
564 	q->old_flows.first	= NULL;
565 	q->delayed		= RB_ROOT;
566 	q->flows		= 0;
567 	q->inactive_flows	= 0;
568 	q->throttled_flows	= 0;
569 }
570 
571 static void fq_rehash(struct fq_sched_data *q,
572 		      struct rb_root *old_array, u32 old_log,
573 		      struct rb_root *new_array, u32 new_log)
574 {
575 	struct rb_node *op, **np, *parent;
576 	struct rb_root *oroot, *nroot;
577 	struct fq_flow *of, *nf;
578 	int fcnt = 0;
579 	u32 idx;
580 
581 	for (idx = 0; idx < (1U << old_log); idx++) {
582 		oroot = &old_array[idx];
583 		while ((op = rb_first(oroot)) != NULL) {
584 			rb_erase(op, oroot);
585 			of = rb_entry(op, struct fq_flow, fq_node);
586 			if (fq_gc_candidate(of)) {
587 				fcnt++;
588 				kmem_cache_free(fq_flow_cachep, of);
589 				continue;
590 			}
591 			nroot = &new_array[hash_ptr(of->sk, new_log)];
592 
593 			np = &nroot->rb_node;
594 			parent = NULL;
595 			while (*np) {
596 				parent = *np;
597 
598 				nf = rb_entry(parent, struct fq_flow, fq_node);
599 				BUG_ON(nf->sk == of->sk);
600 
601 				if (nf->sk > of->sk)
602 					np = &parent->rb_right;
603 				else
604 					np = &parent->rb_left;
605 			}
606 
607 			rb_link_node(&of->fq_node, parent, np);
608 			rb_insert_color(&of->fq_node, nroot);
609 		}
610 	}
611 	q->flows -= fcnt;
612 	q->inactive_flows -= fcnt;
613 	q->stat_gc_flows += fcnt;
614 }
615 
616 static void fq_free(void *addr)
617 {
618 	kvfree(addr);
619 }
620 
621 static int fq_resize(struct Qdisc *sch, u32 log)
622 {
623 	struct fq_sched_data *q = qdisc_priv(sch);
624 	struct rb_root *array;
625 	void *old_fq_root;
626 	u32 idx;
627 
628 	if (q->fq_root && log == q->fq_trees_log)
629 		return 0;
630 
631 	/* If XPS was setup, we can allocate memory on right NUMA node */
632 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
633 			      netdev_queue_numa_node_read(sch->dev_queue));
634 	if (!array)
635 		return -ENOMEM;
636 
637 	for (idx = 0; idx < (1U << log); idx++)
638 		array[idx] = RB_ROOT;
639 
640 	sch_tree_lock(sch);
641 
642 	old_fq_root = q->fq_root;
643 	if (old_fq_root)
644 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
645 
646 	q->fq_root = array;
647 	q->fq_trees_log = log;
648 
649 	sch_tree_unlock(sch);
650 
651 	fq_free(old_fq_root);
652 
653 	return 0;
654 }
655 
656 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
657 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
658 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
659 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
660 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
661 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
662 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
663 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
664 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
665 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
666 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
667 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
668 };
669 
670 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
671 		     struct netlink_ext_ack *extack)
672 {
673 	struct fq_sched_data *q = qdisc_priv(sch);
674 	struct nlattr *tb[TCA_FQ_MAX + 1];
675 	int err, drop_count = 0;
676 	unsigned drop_len = 0;
677 	u32 fq_log;
678 
679 	if (!opt)
680 		return -EINVAL;
681 
682 	err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy, NULL);
683 	if (err < 0)
684 		return err;
685 
686 	sch_tree_lock(sch);
687 
688 	fq_log = q->fq_trees_log;
689 
690 	if (tb[TCA_FQ_BUCKETS_LOG]) {
691 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
692 
693 		if (nval >= 1 && nval <= ilog2(256*1024))
694 			fq_log = nval;
695 		else
696 			err = -EINVAL;
697 	}
698 	if (tb[TCA_FQ_PLIMIT])
699 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
700 
701 	if (tb[TCA_FQ_FLOW_PLIMIT])
702 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
703 
704 	if (tb[TCA_FQ_QUANTUM]) {
705 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
706 
707 		if (quantum > 0)
708 			q->quantum = quantum;
709 		else
710 			err = -EINVAL;
711 	}
712 
713 	if (tb[TCA_FQ_INITIAL_QUANTUM])
714 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
715 
716 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
717 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
718 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
719 
720 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
721 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
722 
723 		q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
724 	}
725 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
726 		q->low_rate_threshold =
727 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
728 
729 	if (tb[TCA_FQ_RATE_ENABLE]) {
730 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
731 
732 		if (enable <= 1)
733 			q->rate_enable = enable;
734 		else
735 			err = -EINVAL;
736 	}
737 
738 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
739 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
740 
741 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
742 	}
743 
744 	if (tb[TCA_FQ_ORPHAN_MASK])
745 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
746 
747 	if (tb[TCA_FQ_CE_THRESHOLD])
748 		q->ce_threshold = (u64)NSEC_PER_USEC *
749 				  nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
750 
751 	if (!err) {
752 		sch_tree_unlock(sch);
753 		err = fq_resize(sch, fq_log);
754 		sch_tree_lock(sch);
755 	}
756 	while (sch->q.qlen > sch->limit) {
757 		struct sk_buff *skb = fq_dequeue(sch);
758 
759 		if (!skb)
760 			break;
761 		drop_len += qdisc_pkt_len(skb);
762 		rtnl_kfree_skbs(skb, skb);
763 		drop_count++;
764 	}
765 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
766 
767 	sch_tree_unlock(sch);
768 	return err;
769 }
770 
771 static void fq_destroy(struct Qdisc *sch)
772 {
773 	struct fq_sched_data *q = qdisc_priv(sch);
774 
775 	fq_reset(sch);
776 	fq_free(q->fq_root);
777 	qdisc_watchdog_cancel(&q->watchdog);
778 }
779 
780 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
781 		   struct netlink_ext_ack *extack)
782 {
783 	struct fq_sched_data *q = qdisc_priv(sch);
784 	int err;
785 
786 	sch->limit		= 10000;
787 	q->flow_plimit		= 100;
788 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
789 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
790 	q->flow_refill_delay	= msecs_to_jiffies(40);
791 	q->flow_max_rate	= ~0UL;
792 	q->time_next_delayed_flow = ~0ULL;
793 	q->rate_enable		= 1;
794 	q->new_flows.first	= NULL;
795 	q->old_flows.first	= NULL;
796 	q->delayed		= RB_ROOT;
797 	q->fq_root		= NULL;
798 	q->fq_trees_log		= ilog2(1024);
799 	q->orphan_mask		= 1024 - 1;
800 	q->low_rate_threshold	= 550000 / 8;
801 
802 	/* Default ce_threshold of 4294 seconds */
803 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
804 
805 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
806 
807 	if (opt)
808 		err = fq_change(sch, opt, extack);
809 	else
810 		err = fq_resize(sch, q->fq_trees_log);
811 
812 	return err;
813 }
814 
815 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
816 {
817 	struct fq_sched_data *q = qdisc_priv(sch);
818 	u64 ce_threshold = q->ce_threshold;
819 	struct nlattr *opts;
820 
821 	opts = nla_nest_start(skb, TCA_OPTIONS);
822 	if (opts == NULL)
823 		goto nla_put_failure;
824 
825 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
826 
827 	do_div(ce_threshold, NSEC_PER_USEC);
828 
829 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
830 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
831 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
832 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
833 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
834 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
835 			min_t(unsigned long, q->flow_max_rate, ~0U)) ||
836 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
837 			jiffies_to_usecs(q->flow_refill_delay)) ||
838 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
839 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
840 			q->low_rate_threshold) ||
841 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
842 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
843 		goto nla_put_failure;
844 
845 	return nla_nest_end(skb, opts);
846 
847 nla_put_failure:
848 	return -1;
849 }
850 
851 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
852 {
853 	struct fq_sched_data *q = qdisc_priv(sch);
854 	struct tc_fq_qd_stats st;
855 
856 	sch_tree_lock(sch);
857 
858 	st.gc_flows		  = q->stat_gc_flows;
859 	st.highprio_packets	  = q->stat_internal_packets;
860 	st.tcp_retrans		  = 0;
861 	st.throttled		  = q->stat_throttled;
862 	st.flows_plimit		  = q->stat_flows_plimit;
863 	st.pkts_too_long	  = q->stat_pkts_too_long;
864 	st.allocation_errors	  = q->stat_allocation_errors;
865 	st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
866 	st.flows		  = q->flows;
867 	st.inactive_flows	  = q->inactive_flows;
868 	st.throttled_flows	  = q->throttled_flows;
869 	st.unthrottle_latency_ns  = min_t(unsigned long,
870 					  q->unthrottle_latency_ns, ~0U);
871 	st.ce_mark		  = q->stat_ce_mark;
872 	sch_tree_unlock(sch);
873 
874 	return gnet_stats_copy_app(d, &st, sizeof(st));
875 }
876 
877 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
878 	.id		=	"fq",
879 	.priv_size	=	sizeof(struct fq_sched_data),
880 
881 	.enqueue	=	fq_enqueue,
882 	.dequeue	=	fq_dequeue,
883 	.peek		=	qdisc_peek_dequeued,
884 	.init		=	fq_init,
885 	.reset		=	fq_reset,
886 	.destroy	=	fq_destroy,
887 	.change		=	fq_change,
888 	.dump		=	fq_dump,
889 	.dump_stats	=	fq_dump_stats,
890 	.owner		=	THIS_MODULE,
891 };
892 
893 static int __init fq_module_init(void)
894 {
895 	int ret;
896 
897 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
898 					   sizeof(struct fq_flow),
899 					   0, 0, NULL);
900 	if (!fq_flow_cachep)
901 		return -ENOMEM;
902 
903 	ret = register_qdisc(&fq_qdisc_ops);
904 	if (ret)
905 		kmem_cache_destroy(fq_flow_cachep);
906 	return ret;
907 }
908 
909 static void __exit fq_module_exit(void)
910 {
911 	unregister_qdisc(&fq_qdisc_ops);
912 	kmem_cache_destroy(fq_flow_cachep);
913 }
914 
915 module_init(fq_module_init)
916 module_exit(fq_module_exit)
917 MODULE_AUTHOR("Eric Dumazet");
918 MODULE_LICENSE("GPL");
919