xref: /linux/net/sched/sch_fq.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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 	struct rb_root	t_root;
70 	struct sk_buff	*head;		/* list of skbs for this flow : first skb */
71 	union {
72 		struct sk_buff *tail;	/* last skb in the list */
73 		unsigned long  age;	/* jiffies when flow was emptied, for gc */
74 	};
75 	struct rb_node	fq_node;	/* anchor in fq_root[] trees */
76 	struct sock	*sk;
77 	int		qlen;		/* number of packets in flow queue */
78 	int		credit;
79 	u32		socket_hash;	/* sk_hash */
80 	struct fq_flow *next;		/* next pointer in RR lists, or &detached */
81 
82 	struct rb_node  rate_node;	/* anchor in q->delayed tree */
83 	u64		time_next_packet;
84 };
85 
86 struct fq_flow_head {
87 	struct fq_flow *first;
88 	struct fq_flow *last;
89 };
90 
91 struct fq_sched_data {
92 	struct fq_flow_head new_flows;
93 
94 	struct fq_flow_head old_flows;
95 
96 	struct rb_root	delayed;	/* for rate limited flows */
97 	u64		time_next_delayed_flow;
98 	unsigned long	unthrottle_latency_ns;
99 
100 	struct fq_flow	internal;	/* for non classified or high prio packets */
101 	u32		quantum;
102 	u32		initial_quantum;
103 	u32		flow_refill_delay;
104 	u32		flow_plimit;	/* max packets per flow */
105 	unsigned long	flow_max_rate;	/* optional max rate per flow */
106 	u64		ce_threshold;
107 	u32		orphan_mask;	/* mask for orphaned skb */
108 	u32		low_rate_threshold;
109 	struct rb_root	*fq_root;
110 	u8		rate_enable;
111 	u8		fq_trees_log;
112 
113 	u32		flows;
114 	u32		inactive_flows;
115 	u32		throttled_flows;
116 
117 	u64		stat_gc_flows;
118 	u64		stat_internal_packets;
119 	u64		stat_throttled;
120 	u64		stat_ce_mark;
121 	u64		stat_flows_plimit;
122 	u64		stat_pkts_too_long;
123 	u64		stat_allocation_errors;
124 	struct qdisc_watchdog watchdog;
125 };
126 
127 /* special value to mark a detached flow (not on old/new list) */
128 static struct fq_flow detached, throttled;
129 
130 static void fq_flow_set_detached(struct fq_flow *f)
131 {
132 	f->next = &detached;
133 	f->age = jiffies;
134 }
135 
136 static bool fq_flow_is_detached(const struct fq_flow *f)
137 {
138 	return f->next == &detached;
139 }
140 
141 static bool fq_flow_is_throttled(const struct fq_flow *f)
142 {
143 	return f->next == &throttled;
144 }
145 
146 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
147 {
148 	if (head->first)
149 		head->last->next = flow;
150 	else
151 		head->first = flow;
152 	head->last = flow;
153 	flow->next = NULL;
154 }
155 
156 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
157 {
158 	rb_erase(&f->rate_node, &q->delayed);
159 	q->throttled_flows--;
160 	fq_flow_add_tail(&q->old_flows, f);
161 }
162 
163 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
164 {
165 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
166 
167 	while (*p) {
168 		struct fq_flow *aux;
169 
170 		parent = *p;
171 		aux = rb_entry(parent, struct fq_flow, rate_node);
172 		if (f->time_next_packet >= aux->time_next_packet)
173 			p = &parent->rb_right;
174 		else
175 			p = &parent->rb_left;
176 	}
177 	rb_link_node(&f->rate_node, parent, p);
178 	rb_insert_color(&f->rate_node, &q->delayed);
179 	q->throttled_flows++;
180 	q->stat_throttled++;
181 
182 	f->next = &throttled;
183 	if (q->time_next_delayed_flow > f->time_next_packet)
184 		q->time_next_delayed_flow = f->time_next_packet;
185 }
186 
187 
188 static struct kmem_cache *fq_flow_cachep __read_mostly;
189 
190 
191 /* limit number of collected flows per round */
192 #define FQ_GC_MAX 8
193 #define FQ_GC_AGE (3*HZ)
194 
195 static bool fq_gc_candidate(const struct fq_flow *f)
196 {
197 	return fq_flow_is_detached(f) &&
198 	       time_after(jiffies, f->age + FQ_GC_AGE);
199 }
200 
201 static void fq_gc(struct fq_sched_data *q,
202 		  struct rb_root *root,
203 		  struct sock *sk)
204 {
205 	struct fq_flow *f, *tofree[FQ_GC_MAX];
206 	struct rb_node **p, *parent;
207 	int fcnt = 0;
208 
209 	p = &root->rb_node;
210 	parent = NULL;
211 	while (*p) {
212 		parent = *p;
213 
214 		f = rb_entry(parent, struct fq_flow, fq_node);
215 		if (f->sk == sk)
216 			break;
217 
218 		if (fq_gc_candidate(f)) {
219 			tofree[fcnt++] = f;
220 			if (fcnt == FQ_GC_MAX)
221 				break;
222 		}
223 
224 		if (f->sk > sk)
225 			p = &parent->rb_right;
226 		else
227 			p = &parent->rb_left;
228 	}
229 
230 	q->flows -= fcnt;
231 	q->inactive_flows -= fcnt;
232 	q->stat_gc_flows += fcnt;
233 	while (fcnt) {
234 		struct fq_flow *f = tofree[--fcnt];
235 
236 		rb_erase(&f->fq_node, root);
237 		kmem_cache_free(fq_flow_cachep, f);
238 	}
239 }
240 
241 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
242 {
243 	struct rb_node **p, *parent;
244 	struct sock *sk = skb->sk;
245 	struct rb_root *root;
246 	struct fq_flow *f;
247 
248 	/* warning: no starvation prevention... */
249 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
250 		return &q->internal;
251 
252 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
253 	 * or a listener (SYNCOOKIE mode)
254 	 * 1) request sockets are not full blown,
255 	 *    they do not contain sk_pacing_rate
256 	 * 2) They are not part of a 'flow' yet
257 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
258 	 *    especially if the listener set SO_MAX_PACING_RATE
259 	 * 4) We pretend they are orphaned
260 	 */
261 	if (!sk || sk_listener(sk)) {
262 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
263 
264 		/* By forcing low order bit to 1, we make sure to not
265 		 * collide with a local flow (socket pointers are word aligned)
266 		 */
267 		sk = (struct sock *)((hash << 1) | 1UL);
268 		skb_orphan(skb);
269 	} else if (sk->sk_state == TCP_CLOSE) {
270 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
271 		/*
272 		 * Sockets in TCP_CLOSE are non connected.
273 		 * Typical use case is UDP sockets, they can send packets
274 		 * with sendto() to many different destinations.
275 		 * We probably could use a generic bit advertising
276 		 * non connected sockets, instead of sk_state == TCP_CLOSE,
277 		 * if we care enough.
278 		 */
279 		sk = (struct sock *)((hash << 1) | 1UL);
280 	}
281 
282 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
283 
284 	if (q->flows >= (2U << q->fq_trees_log) &&
285 	    q->inactive_flows > q->flows/2)
286 		fq_gc(q, root, sk);
287 
288 	p = &root->rb_node;
289 	parent = NULL;
290 	while (*p) {
291 		parent = *p;
292 
293 		f = rb_entry(parent, struct fq_flow, fq_node);
294 		if (f->sk == sk) {
295 			/* socket might have been reallocated, so check
296 			 * if its sk_hash is the same.
297 			 * It not, we need to refill credit with
298 			 * initial quantum
299 			 */
300 			if (unlikely(skb->sk == sk &&
301 				     f->socket_hash != sk->sk_hash)) {
302 				f->credit = q->initial_quantum;
303 				f->socket_hash = sk->sk_hash;
304 				if (fq_flow_is_throttled(f))
305 					fq_flow_unset_throttled(q, f);
306 				f->time_next_packet = 0ULL;
307 			}
308 			return f;
309 		}
310 		if (f->sk > sk)
311 			p = &parent->rb_right;
312 		else
313 			p = &parent->rb_left;
314 	}
315 
316 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
317 	if (unlikely(!f)) {
318 		q->stat_allocation_errors++;
319 		return &q->internal;
320 	}
321 	/* f->t_root is already zeroed after kmem_cache_zalloc() */
322 
323 	fq_flow_set_detached(f);
324 	f->sk = sk;
325 	if (skb->sk == sk)
326 		f->socket_hash = sk->sk_hash;
327 	f->credit = q->initial_quantum;
328 
329 	rb_link_node(&f->fq_node, parent, p);
330 	rb_insert_color(&f->fq_node, root);
331 
332 	q->flows++;
333 	q->inactive_flows++;
334 	return f;
335 }
336 
337 static struct sk_buff *fq_peek(struct fq_flow *flow)
338 {
339 	struct sk_buff *skb = skb_rb_first(&flow->t_root);
340 	struct sk_buff *head = flow->head;
341 
342 	if (!skb)
343 		return head;
344 
345 	if (!head)
346 		return skb;
347 
348 	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
349 		return skb;
350 	return head;
351 }
352 
353 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
354 			  struct sk_buff *skb)
355 {
356 	if (skb == flow->head) {
357 		flow->head = skb->next;
358 	} else {
359 		rb_erase(&skb->rbnode, &flow->t_root);
360 		skb->dev = qdisc_dev(sch);
361 	}
362 }
363 
364 /* remove one skb from head of flow queue */
365 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
366 {
367 	struct sk_buff *skb = fq_peek(flow);
368 
369 	if (skb) {
370 		fq_erase_head(sch, flow, skb);
371 		skb_mark_not_on_list(skb);
372 		flow->qlen--;
373 		qdisc_qstats_backlog_dec(sch, skb);
374 		sch->q.qlen--;
375 	}
376 	return skb;
377 }
378 
379 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
380 {
381 	struct rb_node **p, *parent;
382 	struct sk_buff *head, *aux;
383 
384 	fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns();
385 
386 	head = flow->head;
387 	if (!head ||
388 	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
389 		if (!head)
390 			flow->head = skb;
391 		else
392 			flow->tail->next = skb;
393 		flow->tail = skb;
394 		skb->next = NULL;
395 		return;
396 	}
397 
398 	p = &flow->t_root.rb_node;
399 	parent = NULL;
400 
401 	while (*p) {
402 		parent = *p;
403 		aux = rb_to_skb(parent);
404 		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
405 			p = &parent->rb_right;
406 		else
407 			p = &parent->rb_left;
408 	}
409 	rb_link_node(&skb->rbnode, parent, p);
410 	rb_insert_color(&skb->rbnode, &flow->t_root);
411 }
412 
413 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
414 		      struct sk_buff **to_free)
415 {
416 	struct fq_sched_data *q = qdisc_priv(sch);
417 	struct fq_flow *f;
418 
419 	if (unlikely(sch->q.qlen >= sch->limit))
420 		return qdisc_drop(skb, sch, to_free);
421 
422 	f = fq_classify(skb, q);
423 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
424 		q->stat_flows_plimit++;
425 		return qdisc_drop(skb, sch, to_free);
426 	}
427 
428 	f->qlen++;
429 	qdisc_qstats_backlog_inc(sch, skb);
430 	if (fq_flow_is_detached(f)) {
431 		struct sock *sk = skb->sk;
432 
433 		fq_flow_add_tail(&q->new_flows, f);
434 		if (time_after(jiffies, f->age + q->flow_refill_delay))
435 			f->credit = max_t(u32, f->credit, q->quantum);
436 		if (sk && q->rate_enable) {
437 			if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
438 				     SK_PACING_FQ))
439 				smp_store_release(&sk->sk_pacing_status,
440 						  SK_PACING_FQ);
441 		}
442 		q->inactive_flows--;
443 	}
444 
445 	/* Note: this overwrites f->age */
446 	flow_queue_add(f, skb);
447 
448 	if (unlikely(f == &q->internal)) {
449 		q->stat_internal_packets++;
450 	}
451 	sch->q.qlen++;
452 
453 	return NET_XMIT_SUCCESS;
454 }
455 
456 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
457 {
458 	unsigned long sample;
459 	struct rb_node *p;
460 
461 	if (q->time_next_delayed_flow > now)
462 		return;
463 
464 	/* Update unthrottle latency EWMA.
465 	 * This is cheap and can help diagnosing timer/latency problems.
466 	 */
467 	sample = (unsigned long)(now - q->time_next_delayed_flow);
468 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
469 	q->unthrottle_latency_ns += sample >> 3;
470 
471 	q->time_next_delayed_flow = ~0ULL;
472 	while ((p = rb_first(&q->delayed)) != NULL) {
473 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
474 
475 		if (f->time_next_packet > now) {
476 			q->time_next_delayed_flow = f->time_next_packet;
477 			break;
478 		}
479 		fq_flow_unset_throttled(q, f);
480 	}
481 }
482 
483 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
484 {
485 	struct fq_sched_data *q = qdisc_priv(sch);
486 	struct fq_flow_head *head;
487 	struct sk_buff *skb;
488 	struct fq_flow *f;
489 	unsigned long rate;
490 	u32 plen;
491 	u64 now;
492 
493 	if (!sch->q.qlen)
494 		return NULL;
495 
496 	skb = fq_dequeue_head(sch, &q->internal);
497 	if (skb)
498 		goto out;
499 
500 	now = ktime_get_ns();
501 	fq_check_throttled(q, now);
502 begin:
503 	head = &q->new_flows;
504 	if (!head->first) {
505 		head = &q->old_flows;
506 		if (!head->first) {
507 			if (q->time_next_delayed_flow != ~0ULL)
508 				qdisc_watchdog_schedule_ns(&q->watchdog,
509 							   q->time_next_delayed_flow);
510 			return NULL;
511 		}
512 	}
513 	f = head->first;
514 
515 	if (f->credit <= 0) {
516 		f->credit += q->quantum;
517 		head->first = f->next;
518 		fq_flow_add_tail(&q->old_flows, f);
519 		goto begin;
520 	}
521 
522 	skb = fq_peek(f);
523 	if (skb) {
524 		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
525 					     f->time_next_packet);
526 
527 		if (now < time_next_packet) {
528 			head->first = f->next;
529 			f->time_next_packet = time_next_packet;
530 			fq_flow_set_throttled(q, f);
531 			goto begin;
532 		}
533 		if (time_next_packet &&
534 		    (s64)(now - time_next_packet - q->ce_threshold) > 0) {
535 			INET_ECN_set_ce(skb);
536 			q->stat_ce_mark++;
537 		}
538 	}
539 
540 	skb = fq_dequeue_head(sch, f);
541 	if (!skb) {
542 		head->first = f->next;
543 		/* force a pass through old_flows to prevent starvation */
544 		if ((head == &q->new_flows) && q->old_flows.first) {
545 			fq_flow_add_tail(&q->old_flows, f);
546 		} else {
547 			fq_flow_set_detached(f);
548 			q->inactive_flows++;
549 		}
550 		goto begin;
551 	}
552 	prefetch(&skb->end);
553 	plen = qdisc_pkt_len(skb);
554 	f->credit -= plen;
555 
556 	if (!q->rate_enable)
557 		goto out;
558 
559 	rate = q->flow_max_rate;
560 
561 	/* If EDT time was provided for this skb, we need to
562 	 * update f->time_next_packet only if this qdisc enforces
563 	 * a flow max rate.
564 	 */
565 	if (!skb->tstamp) {
566 		if (skb->sk)
567 			rate = min(skb->sk->sk_pacing_rate, rate);
568 
569 		if (rate <= q->low_rate_threshold) {
570 			f->credit = 0;
571 		} else {
572 			plen = max(plen, q->quantum);
573 			if (f->credit > 0)
574 				goto out;
575 		}
576 	}
577 	if (rate != ~0UL) {
578 		u64 len = (u64)plen * NSEC_PER_SEC;
579 
580 		if (likely(rate))
581 			len = div64_ul(len, rate);
582 		/* Since socket rate can change later,
583 		 * clamp the delay to 1 second.
584 		 * Really, providers of too big packets should be fixed !
585 		 */
586 		if (unlikely(len > NSEC_PER_SEC)) {
587 			len = NSEC_PER_SEC;
588 			q->stat_pkts_too_long++;
589 		}
590 		/* Account for schedule/timers drifts.
591 		 * f->time_next_packet was set when prior packet was sent,
592 		 * and current time (@now) can be too late by tens of us.
593 		 */
594 		if (f->time_next_packet)
595 			len -= min(len/2, now - f->time_next_packet);
596 		f->time_next_packet = now + len;
597 	}
598 out:
599 	qdisc_bstats_update(sch, skb);
600 	return skb;
601 }
602 
603 static void fq_flow_purge(struct fq_flow *flow)
604 {
605 	struct rb_node *p = rb_first(&flow->t_root);
606 
607 	while (p) {
608 		struct sk_buff *skb = rb_to_skb(p);
609 
610 		p = rb_next(p);
611 		rb_erase(&skb->rbnode, &flow->t_root);
612 		rtnl_kfree_skbs(skb, skb);
613 	}
614 	rtnl_kfree_skbs(flow->head, flow->tail);
615 	flow->head = NULL;
616 	flow->qlen = 0;
617 }
618 
619 static void fq_reset(struct Qdisc *sch)
620 {
621 	struct fq_sched_data *q = qdisc_priv(sch);
622 	struct rb_root *root;
623 	struct rb_node *p;
624 	struct fq_flow *f;
625 	unsigned int idx;
626 
627 	sch->q.qlen = 0;
628 	sch->qstats.backlog = 0;
629 
630 	fq_flow_purge(&q->internal);
631 
632 	if (!q->fq_root)
633 		return;
634 
635 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
636 		root = &q->fq_root[idx];
637 		while ((p = rb_first(root)) != NULL) {
638 			f = rb_entry(p, struct fq_flow, fq_node);
639 			rb_erase(p, root);
640 
641 			fq_flow_purge(f);
642 
643 			kmem_cache_free(fq_flow_cachep, f);
644 		}
645 	}
646 	q->new_flows.first	= NULL;
647 	q->old_flows.first	= NULL;
648 	q->delayed		= RB_ROOT;
649 	q->flows		= 0;
650 	q->inactive_flows	= 0;
651 	q->throttled_flows	= 0;
652 }
653 
654 static void fq_rehash(struct fq_sched_data *q,
655 		      struct rb_root *old_array, u32 old_log,
656 		      struct rb_root *new_array, u32 new_log)
657 {
658 	struct rb_node *op, **np, *parent;
659 	struct rb_root *oroot, *nroot;
660 	struct fq_flow *of, *nf;
661 	int fcnt = 0;
662 	u32 idx;
663 
664 	for (idx = 0; idx < (1U << old_log); idx++) {
665 		oroot = &old_array[idx];
666 		while ((op = rb_first(oroot)) != NULL) {
667 			rb_erase(op, oroot);
668 			of = rb_entry(op, struct fq_flow, fq_node);
669 			if (fq_gc_candidate(of)) {
670 				fcnt++;
671 				kmem_cache_free(fq_flow_cachep, of);
672 				continue;
673 			}
674 			nroot = &new_array[hash_ptr(of->sk, new_log)];
675 
676 			np = &nroot->rb_node;
677 			parent = NULL;
678 			while (*np) {
679 				parent = *np;
680 
681 				nf = rb_entry(parent, struct fq_flow, fq_node);
682 				BUG_ON(nf->sk == of->sk);
683 
684 				if (nf->sk > of->sk)
685 					np = &parent->rb_right;
686 				else
687 					np = &parent->rb_left;
688 			}
689 
690 			rb_link_node(&of->fq_node, parent, np);
691 			rb_insert_color(&of->fq_node, nroot);
692 		}
693 	}
694 	q->flows -= fcnt;
695 	q->inactive_flows -= fcnt;
696 	q->stat_gc_flows += fcnt;
697 }
698 
699 static void fq_free(void *addr)
700 {
701 	kvfree(addr);
702 }
703 
704 static int fq_resize(struct Qdisc *sch, u32 log)
705 {
706 	struct fq_sched_data *q = qdisc_priv(sch);
707 	struct rb_root *array;
708 	void *old_fq_root;
709 	u32 idx;
710 
711 	if (q->fq_root && log == q->fq_trees_log)
712 		return 0;
713 
714 	/* If XPS was setup, we can allocate memory on right NUMA node */
715 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
716 			      netdev_queue_numa_node_read(sch->dev_queue));
717 	if (!array)
718 		return -ENOMEM;
719 
720 	for (idx = 0; idx < (1U << log); idx++)
721 		array[idx] = RB_ROOT;
722 
723 	sch_tree_lock(sch);
724 
725 	old_fq_root = q->fq_root;
726 	if (old_fq_root)
727 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
728 
729 	q->fq_root = array;
730 	q->fq_trees_log = log;
731 
732 	sch_tree_unlock(sch);
733 
734 	fq_free(old_fq_root);
735 
736 	return 0;
737 }
738 
739 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
740 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
741 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
742 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
743 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
744 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
745 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
746 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
747 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
748 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
749 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
750 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
751 };
752 
753 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
754 		     struct netlink_ext_ack *extack)
755 {
756 	struct fq_sched_data *q = qdisc_priv(sch);
757 	struct nlattr *tb[TCA_FQ_MAX + 1];
758 	int err, drop_count = 0;
759 	unsigned drop_len = 0;
760 	u32 fq_log;
761 
762 	if (!opt)
763 		return -EINVAL;
764 
765 	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
766 					  NULL);
767 	if (err < 0)
768 		return err;
769 
770 	sch_tree_lock(sch);
771 
772 	fq_log = q->fq_trees_log;
773 
774 	if (tb[TCA_FQ_BUCKETS_LOG]) {
775 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
776 
777 		if (nval >= 1 && nval <= ilog2(256*1024))
778 			fq_log = nval;
779 		else
780 			err = -EINVAL;
781 	}
782 	if (tb[TCA_FQ_PLIMIT])
783 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
784 
785 	if (tb[TCA_FQ_FLOW_PLIMIT])
786 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
787 
788 	if (tb[TCA_FQ_QUANTUM]) {
789 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
790 
791 		if (quantum > 0)
792 			q->quantum = quantum;
793 		else
794 			err = -EINVAL;
795 	}
796 
797 	if (tb[TCA_FQ_INITIAL_QUANTUM])
798 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
799 
800 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
801 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
802 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
803 
804 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
805 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
806 
807 		q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
808 	}
809 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
810 		q->low_rate_threshold =
811 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
812 
813 	if (tb[TCA_FQ_RATE_ENABLE]) {
814 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
815 
816 		if (enable <= 1)
817 			q->rate_enable = enable;
818 		else
819 			err = -EINVAL;
820 	}
821 
822 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
823 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
824 
825 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
826 	}
827 
828 	if (tb[TCA_FQ_ORPHAN_MASK])
829 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
830 
831 	if (tb[TCA_FQ_CE_THRESHOLD])
832 		q->ce_threshold = (u64)NSEC_PER_USEC *
833 				  nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
834 
835 	if (!err) {
836 		sch_tree_unlock(sch);
837 		err = fq_resize(sch, fq_log);
838 		sch_tree_lock(sch);
839 	}
840 	while (sch->q.qlen > sch->limit) {
841 		struct sk_buff *skb = fq_dequeue(sch);
842 
843 		if (!skb)
844 			break;
845 		drop_len += qdisc_pkt_len(skb);
846 		rtnl_kfree_skbs(skb, skb);
847 		drop_count++;
848 	}
849 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
850 
851 	sch_tree_unlock(sch);
852 	return err;
853 }
854 
855 static void fq_destroy(struct Qdisc *sch)
856 {
857 	struct fq_sched_data *q = qdisc_priv(sch);
858 
859 	fq_reset(sch);
860 	fq_free(q->fq_root);
861 	qdisc_watchdog_cancel(&q->watchdog);
862 }
863 
864 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
865 		   struct netlink_ext_ack *extack)
866 {
867 	struct fq_sched_data *q = qdisc_priv(sch);
868 	int err;
869 
870 	sch->limit		= 10000;
871 	q->flow_plimit		= 100;
872 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
873 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
874 	q->flow_refill_delay	= msecs_to_jiffies(40);
875 	q->flow_max_rate	= ~0UL;
876 	q->time_next_delayed_flow = ~0ULL;
877 	q->rate_enable		= 1;
878 	q->new_flows.first	= NULL;
879 	q->old_flows.first	= NULL;
880 	q->delayed		= RB_ROOT;
881 	q->fq_root		= NULL;
882 	q->fq_trees_log		= ilog2(1024);
883 	q->orphan_mask		= 1024 - 1;
884 	q->low_rate_threshold	= 550000 / 8;
885 
886 	/* Default ce_threshold of 4294 seconds */
887 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
888 
889 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
890 
891 	if (opt)
892 		err = fq_change(sch, opt, extack);
893 	else
894 		err = fq_resize(sch, q->fq_trees_log);
895 
896 	return err;
897 }
898 
899 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
900 {
901 	struct fq_sched_data *q = qdisc_priv(sch);
902 	u64 ce_threshold = q->ce_threshold;
903 	struct nlattr *opts;
904 
905 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
906 	if (opts == NULL)
907 		goto nla_put_failure;
908 
909 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
910 
911 	do_div(ce_threshold, NSEC_PER_USEC);
912 
913 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
914 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
915 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
916 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
917 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
918 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
919 			min_t(unsigned long, q->flow_max_rate, ~0U)) ||
920 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
921 			jiffies_to_usecs(q->flow_refill_delay)) ||
922 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
923 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
924 			q->low_rate_threshold) ||
925 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
926 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
927 		goto nla_put_failure;
928 
929 	return nla_nest_end(skb, opts);
930 
931 nla_put_failure:
932 	return -1;
933 }
934 
935 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
936 {
937 	struct fq_sched_data *q = qdisc_priv(sch);
938 	struct tc_fq_qd_stats st;
939 
940 	sch_tree_lock(sch);
941 
942 	st.gc_flows		  = q->stat_gc_flows;
943 	st.highprio_packets	  = q->stat_internal_packets;
944 	st.tcp_retrans		  = 0;
945 	st.throttled		  = q->stat_throttled;
946 	st.flows_plimit		  = q->stat_flows_plimit;
947 	st.pkts_too_long	  = q->stat_pkts_too_long;
948 	st.allocation_errors	  = q->stat_allocation_errors;
949 	st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
950 	st.flows		  = q->flows;
951 	st.inactive_flows	  = q->inactive_flows;
952 	st.throttled_flows	  = q->throttled_flows;
953 	st.unthrottle_latency_ns  = min_t(unsigned long,
954 					  q->unthrottle_latency_ns, ~0U);
955 	st.ce_mark		  = q->stat_ce_mark;
956 	sch_tree_unlock(sch);
957 
958 	return gnet_stats_copy_app(d, &st, sizeof(st));
959 }
960 
961 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
962 	.id		=	"fq",
963 	.priv_size	=	sizeof(struct fq_sched_data),
964 
965 	.enqueue	=	fq_enqueue,
966 	.dequeue	=	fq_dequeue,
967 	.peek		=	qdisc_peek_dequeued,
968 	.init		=	fq_init,
969 	.reset		=	fq_reset,
970 	.destroy	=	fq_destroy,
971 	.change		=	fq_change,
972 	.dump		=	fq_dump,
973 	.dump_stats	=	fq_dump_stats,
974 	.owner		=	THIS_MODULE,
975 };
976 
977 static int __init fq_module_init(void)
978 {
979 	int ret;
980 
981 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
982 					   sizeof(struct fq_flow),
983 					   0, 0, NULL);
984 	if (!fq_flow_cachep)
985 		return -ENOMEM;
986 
987 	ret = register_qdisc(&fq_qdisc_ops);
988 	if (ret)
989 		kmem_cache_destroy(fq_flow_cachep);
990 	return ret;
991 }
992 
993 static void __exit fq_module_exit(void)
994 {
995 	unregister_qdisc(&fq_qdisc_ops);
996 	kmem_cache_destroy(fq_flow_cachep);
997 }
998 
999 module_init(fq_module_init)
1000 module_exit(fq_module_exit)
1001 MODULE_AUTHOR("Eric Dumazet");
1002 MODULE_LICENSE("GPL");
1003