xref: /openbmc/linux/net/sched/sch_fq.c (revision 022dacdd)
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 
125 	u32		timer_slack; /* hrtimer slack in ns */
126 	struct qdisc_watchdog watchdog;
127 };
128 
129 /* special value to mark a detached flow (not on old/new list) */
130 static struct fq_flow detached, throttled;
131 
132 static void fq_flow_set_detached(struct fq_flow *f)
133 {
134 	f->next = &detached;
135 	f->age = jiffies;
136 }
137 
138 static bool fq_flow_is_detached(const struct fq_flow *f)
139 {
140 	return f->next == &detached;
141 }
142 
143 static bool fq_flow_is_throttled(const struct fq_flow *f)
144 {
145 	return f->next == &throttled;
146 }
147 
148 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
149 {
150 	if (head->first)
151 		head->last->next = flow;
152 	else
153 		head->first = flow;
154 	head->last = flow;
155 	flow->next = NULL;
156 }
157 
158 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
159 {
160 	rb_erase(&f->rate_node, &q->delayed);
161 	q->throttled_flows--;
162 	fq_flow_add_tail(&q->old_flows, f);
163 }
164 
165 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
166 {
167 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
168 
169 	while (*p) {
170 		struct fq_flow *aux;
171 
172 		parent = *p;
173 		aux = rb_entry(parent, struct fq_flow, rate_node);
174 		if (f->time_next_packet >= aux->time_next_packet)
175 			p = &parent->rb_right;
176 		else
177 			p = &parent->rb_left;
178 	}
179 	rb_link_node(&f->rate_node, parent, p);
180 	rb_insert_color(&f->rate_node, &q->delayed);
181 	q->throttled_flows++;
182 	q->stat_throttled++;
183 
184 	f->next = &throttled;
185 	if (q->time_next_delayed_flow > f->time_next_packet)
186 		q->time_next_delayed_flow = f->time_next_packet;
187 }
188 
189 
190 static struct kmem_cache *fq_flow_cachep __read_mostly;
191 
192 
193 /* limit number of collected flows per round */
194 #define FQ_GC_MAX 8
195 #define FQ_GC_AGE (3*HZ)
196 
197 static bool fq_gc_candidate(const struct fq_flow *f)
198 {
199 	return fq_flow_is_detached(f) &&
200 	       time_after(jiffies, f->age + FQ_GC_AGE);
201 }
202 
203 static void fq_gc(struct fq_sched_data *q,
204 		  struct rb_root *root,
205 		  struct sock *sk)
206 {
207 	struct fq_flow *f, *tofree[FQ_GC_MAX];
208 	struct rb_node **p, *parent;
209 	int fcnt = 0;
210 
211 	p = &root->rb_node;
212 	parent = NULL;
213 	while (*p) {
214 		parent = *p;
215 
216 		f = rb_entry(parent, struct fq_flow, fq_node);
217 		if (f->sk == sk)
218 			break;
219 
220 		if (fq_gc_candidate(f)) {
221 			tofree[fcnt++] = f;
222 			if (fcnt == FQ_GC_MAX)
223 				break;
224 		}
225 
226 		if (f->sk > sk)
227 			p = &parent->rb_right;
228 		else
229 			p = &parent->rb_left;
230 	}
231 
232 	q->flows -= fcnt;
233 	q->inactive_flows -= fcnt;
234 	q->stat_gc_flows += fcnt;
235 	while (fcnt) {
236 		struct fq_flow *f = tofree[--fcnt];
237 
238 		rb_erase(&f->fq_node, root);
239 		kmem_cache_free(fq_flow_cachep, f);
240 	}
241 }
242 
243 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
244 {
245 	struct rb_node **p, *parent;
246 	struct sock *sk = skb->sk;
247 	struct rb_root *root;
248 	struct fq_flow *f;
249 
250 	/* warning: no starvation prevention... */
251 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
252 		return &q->internal;
253 
254 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
255 	 * or a listener (SYNCOOKIE mode)
256 	 * 1) request sockets are not full blown,
257 	 *    they do not contain sk_pacing_rate
258 	 * 2) They are not part of a 'flow' yet
259 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
260 	 *    especially if the listener set SO_MAX_PACING_RATE
261 	 * 4) We pretend they are orphaned
262 	 */
263 	if (!sk || sk_listener(sk)) {
264 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
265 
266 		/* By forcing low order bit to 1, we make sure to not
267 		 * collide with a local flow (socket pointers are word aligned)
268 		 */
269 		sk = (struct sock *)((hash << 1) | 1UL);
270 		skb_orphan(skb);
271 	} else if (sk->sk_state == TCP_CLOSE) {
272 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
273 		/*
274 		 * Sockets in TCP_CLOSE are non connected.
275 		 * Typical use case is UDP sockets, they can send packets
276 		 * with sendto() to many different destinations.
277 		 * We probably could use a generic bit advertising
278 		 * non connected sockets, instead of sk_state == TCP_CLOSE,
279 		 * if we care enough.
280 		 */
281 		sk = (struct sock *)((hash << 1) | 1UL);
282 	}
283 
284 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
285 
286 	if (q->flows >= (2U << q->fq_trees_log) &&
287 	    q->inactive_flows > q->flows/2)
288 		fq_gc(q, root, sk);
289 
290 	p = &root->rb_node;
291 	parent = NULL;
292 	while (*p) {
293 		parent = *p;
294 
295 		f = rb_entry(parent, struct fq_flow, fq_node);
296 		if (f->sk == sk) {
297 			/* socket might have been reallocated, so check
298 			 * if its sk_hash is the same.
299 			 * It not, we need to refill credit with
300 			 * initial quantum
301 			 */
302 			if (unlikely(skb->sk == sk &&
303 				     f->socket_hash != sk->sk_hash)) {
304 				f->credit = q->initial_quantum;
305 				f->socket_hash = sk->sk_hash;
306 				if (q->rate_enable)
307 					smp_store_release(&sk->sk_pacing_status,
308 							  SK_PACING_FQ);
309 				if (fq_flow_is_throttled(f))
310 					fq_flow_unset_throttled(q, f);
311 				f->time_next_packet = 0ULL;
312 			}
313 			return f;
314 		}
315 		if (f->sk > sk)
316 			p = &parent->rb_right;
317 		else
318 			p = &parent->rb_left;
319 	}
320 
321 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
322 	if (unlikely(!f)) {
323 		q->stat_allocation_errors++;
324 		return &q->internal;
325 	}
326 	/* f->t_root is already zeroed after kmem_cache_zalloc() */
327 
328 	fq_flow_set_detached(f);
329 	f->sk = sk;
330 	if (skb->sk == sk) {
331 		f->socket_hash = sk->sk_hash;
332 		if (q->rate_enable)
333 			smp_store_release(&sk->sk_pacing_status,
334 					  SK_PACING_FQ);
335 	}
336 	f->credit = q->initial_quantum;
337 
338 	rb_link_node(&f->fq_node, parent, p);
339 	rb_insert_color(&f->fq_node, root);
340 
341 	q->flows++;
342 	q->inactive_flows++;
343 	return f;
344 }
345 
346 static struct sk_buff *fq_peek(struct fq_flow *flow)
347 {
348 	struct sk_buff *skb = skb_rb_first(&flow->t_root);
349 	struct sk_buff *head = flow->head;
350 
351 	if (!skb)
352 		return head;
353 
354 	if (!head)
355 		return skb;
356 
357 	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
358 		return skb;
359 	return head;
360 }
361 
362 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
363 			  struct sk_buff *skb)
364 {
365 	if (skb == flow->head) {
366 		flow->head = skb->next;
367 	} else {
368 		rb_erase(&skb->rbnode, &flow->t_root);
369 		skb->dev = qdisc_dev(sch);
370 	}
371 }
372 
373 /* remove one skb from head of flow queue */
374 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
375 {
376 	struct sk_buff *skb = fq_peek(flow);
377 
378 	if (skb) {
379 		fq_erase_head(sch, flow, skb);
380 		skb_mark_not_on_list(skb);
381 		flow->qlen--;
382 		qdisc_qstats_backlog_dec(sch, skb);
383 		sch->q.qlen--;
384 	}
385 	return skb;
386 }
387 
388 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
389 {
390 	struct rb_node **p, *parent;
391 	struct sk_buff *head, *aux;
392 
393 	fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns();
394 
395 	head = flow->head;
396 	if (!head ||
397 	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
398 		if (!head)
399 			flow->head = skb;
400 		else
401 			flow->tail->next = skb;
402 		flow->tail = skb;
403 		skb->next = NULL;
404 		return;
405 	}
406 
407 	p = &flow->t_root.rb_node;
408 	parent = NULL;
409 
410 	while (*p) {
411 		parent = *p;
412 		aux = rb_to_skb(parent);
413 		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
414 			p = &parent->rb_right;
415 		else
416 			p = &parent->rb_left;
417 	}
418 	rb_link_node(&skb->rbnode, parent, p);
419 	rb_insert_color(&skb->rbnode, &flow->t_root);
420 }
421 
422 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
423 		      struct sk_buff **to_free)
424 {
425 	struct fq_sched_data *q = qdisc_priv(sch);
426 	struct fq_flow *f;
427 
428 	if (unlikely(sch->q.qlen >= sch->limit))
429 		return qdisc_drop(skb, sch, to_free);
430 
431 	f = fq_classify(skb, q);
432 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
433 		q->stat_flows_plimit++;
434 		return qdisc_drop(skb, sch, to_free);
435 	}
436 
437 	f->qlen++;
438 	qdisc_qstats_backlog_inc(sch, skb);
439 	if (fq_flow_is_detached(f)) {
440 		fq_flow_add_tail(&q->new_flows, f);
441 		if (time_after(jiffies, f->age + q->flow_refill_delay))
442 			f->credit = max_t(u32, f->credit, q->quantum);
443 		q->inactive_flows--;
444 	}
445 
446 	/* Note: this overwrites f->age */
447 	flow_queue_add(f, skb);
448 
449 	if (unlikely(f == &q->internal)) {
450 		q->stat_internal_packets++;
451 	}
452 	sch->q.qlen++;
453 
454 	return NET_XMIT_SUCCESS;
455 }
456 
457 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
458 {
459 	unsigned long sample;
460 	struct rb_node *p;
461 
462 	if (q->time_next_delayed_flow > now)
463 		return;
464 
465 	/* Update unthrottle latency EWMA.
466 	 * This is cheap and can help diagnosing timer/latency problems.
467 	 */
468 	sample = (unsigned long)(now - q->time_next_delayed_flow);
469 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
470 	q->unthrottle_latency_ns += sample >> 3;
471 
472 	q->time_next_delayed_flow = ~0ULL;
473 	while ((p = rb_first(&q->delayed)) != NULL) {
474 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
475 
476 		if (f->time_next_packet > now) {
477 			q->time_next_delayed_flow = f->time_next_packet;
478 			break;
479 		}
480 		fq_flow_unset_throttled(q, f);
481 	}
482 }
483 
484 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
485 {
486 	struct fq_sched_data *q = qdisc_priv(sch);
487 	struct fq_flow_head *head;
488 	struct sk_buff *skb;
489 	struct fq_flow *f;
490 	unsigned long rate;
491 	u32 plen;
492 	u64 now;
493 
494 	if (!sch->q.qlen)
495 		return NULL;
496 
497 	skb = fq_dequeue_head(sch, &q->internal);
498 	if (skb)
499 		goto out;
500 
501 	now = ktime_get_ns();
502 	fq_check_throttled(q, now);
503 begin:
504 	head = &q->new_flows;
505 	if (!head->first) {
506 		head = &q->old_flows;
507 		if (!head->first) {
508 			if (q->time_next_delayed_flow != ~0ULL)
509 				qdisc_watchdog_schedule_range_ns(&q->watchdog,
510 							q->time_next_delayed_flow,
511 							q->timer_slack);
512 			return NULL;
513 		}
514 	}
515 	f = head->first;
516 
517 	if (f->credit <= 0) {
518 		f->credit += q->quantum;
519 		head->first = f->next;
520 		fq_flow_add_tail(&q->old_flows, f);
521 		goto begin;
522 	}
523 
524 	skb = fq_peek(f);
525 	if (skb) {
526 		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
527 					     f->time_next_packet);
528 
529 		if (now < time_next_packet) {
530 			head->first = f->next;
531 			f->time_next_packet = time_next_packet;
532 			fq_flow_set_throttled(q, f);
533 			goto begin;
534 		}
535 		if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
536 			INET_ECN_set_ce(skb);
537 			q->stat_ce_mark++;
538 		}
539 	}
540 
541 	skb = fq_dequeue_head(sch, f);
542 	if (!skb) {
543 		head->first = f->next;
544 		/* force a pass through old_flows to prevent starvation */
545 		if ((head == &q->new_flows) && q->old_flows.first) {
546 			fq_flow_add_tail(&q->old_flows, f);
547 		} else {
548 			fq_flow_set_detached(f);
549 			q->inactive_flows++;
550 		}
551 		goto begin;
552 	}
553 	prefetch(&skb->end);
554 	plen = qdisc_pkt_len(skb);
555 	f->credit -= plen;
556 
557 	if (!q->rate_enable)
558 		goto out;
559 
560 	rate = q->flow_max_rate;
561 
562 	/* If EDT time was provided for this skb, we need to
563 	 * update f->time_next_packet only if this qdisc enforces
564 	 * a flow max rate.
565 	 */
566 	if (!skb->tstamp) {
567 		if (skb->sk)
568 			rate = min(skb->sk->sk_pacing_rate, rate);
569 
570 		if (rate <= q->low_rate_threshold) {
571 			f->credit = 0;
572 		} else {
573 			plen = max(plen, q->quantum);
574 			if (f->credit > 0)
575 				goto out;
576 		}
577 	}
578 	if (rate != ~0UL) {
579 		u64 len = (u64)plen * NSEC_PER_SEC;
580 
581 		if (likely(rate))
582 			len = div64_ul(len, rate);
583 		/* Since socket rate can change later,
584 		 * clamp the delay to 1 second.
585 		 * Really, providers of too big packets should be fixed !
586 		 */
587 		if (unlikely(len > NSEC_PER_SEC)) {
588 			len = NSEC_PER_SEC;
589 			q->stat_pkts_too_long++;
590 		}
591 		/* Account for schedule/timers drifts.
592 		 * f->time_next_packet was set when prior packet was sent,
593 		 * and current time (@now) can be too late by tens of us.
594 		 */
595 		if (f->time_next_packet)
596 			len -= min(len/2, now - f->time_next_packet);
597 		f->time_next_packet = now + len;
598 	}
599 out:
600 	qdisc_bstats_update(sch, skb);
601 	return skb;
602 }
603 
604 static void fq_flow_purge(struct fq_flow *flow)
605 {
606 	struct rb_node *p = rb_first(&flow->t_root);
607 
608 	while (p) {
609 		struct sk_buff *skb = rb_to_skb(p);
610 
611 		p = rb_next(p);
612 		rb_erase(&skb->rbnode, &flow->t_root);
613 		rtnl_kfree_skbs(skb, skb);
614 	}
615 	rtnl_kfree_skbs(flow->head, flow->tail);
616 	flow->head = NULL;
617 	flow->qlen = 0;
618 }
619 
620 static void fq_reset(struct Qdisc *sch)
621 {
622 	struct fq_sched_data *q = qdisc_priv(sch);
623 	struct rb_root *root;
624 	struct rb_node *p;
625 	struct fq_flow *f;
626 	unsigned int idx;
627 
628 	sch->q.qlen = 0;
629 	sch->qstats.backlog = 0;
630 
631 	fq_flow_purge(&q->internal);
632 
633 	if (!q->fq_root)
634 		return;
635 
636 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
637 		root = &q->fq_root[idx];
638 		while ((p = rb_first(root)) != NULL) {
639 			f = rb_entry(p, struct fq_flow, fq_node);
640 			rb_erase(p, root);
641 
642 			fq_flow_purge(f);
643 
644 			kmem_cache_free(fq_flow_cachep, f);
645 		}
646 	}
647 	q->new_flows.first	= NULL;
648 	q->old_flows.first	= NULL;
649 	q->delayed		= RB_ROOT;
650 	q->flows		= 0;
651 	q->inactive_flows	= 0;
652 	q->throttled_flows	= 0;
653 }
654 
655 static void fq_rehash(struct fq_sched_data *q,
656 		      struct rb_root *old_array, u32 old_log,
657 		      struct rb_root *new_array, u32 new_log)
658 {
659 	struct rb_node *op, **np, *parent;
660 	struct rb_root *oroot, *nroot;
661 	struct fq_flow *of, *nf;
662 	int fcnt = 0;
663 	u32 idx;
664 
665 	for (idx = 0; idx < (1U << old_log); idx++) {
666 		oroot = &old_array[idx];
667 		while ((op = rb_first(oroot)) != NULL) {
668 			rb_erase(op, oroot);
669 			of = rb_entry(op, struct fq_flow, fq_node);
670 			if (fq_gc_candidate(of)) {
671 				fcnt++;
672 				kmem_cache_free(fq_flow_cachep, of);
673 				continue;
674 			}
675 			nroot = &new_array[hash_ptr(of->sk, new_log)];
676 
677 			np = &nroot->rb_node;
678 			parent = NULL;
679 			while (*np) {
680 				parent = *np;
681 
682 				nf = rb_entry(parent, struct fq_flow, fq_node);
683 				BUG_ON(nf->sk == of->sk);
684 
685 				if (nf->sk > of->sk)
686 					np = &parent->rb_right;
687 				else
688 					np = &parent->rb_left;
689 			}
690 
691 			rb_link_node(&of->fq_node, parent, np);
692 			rb_insert_color(&of->fq_node, nroot);
693 		}
694 	}
695 	q->flows -= fcnt;
696 	q->inactive_flows -= fcnt;
697 	q->stat_gc_flows += fcnt;
698 }
699 
700 static void fq_free(void *addr)
701 {
702 	kvfree(addr);
703 }
704 
705 static int fq_resize(struct Qdisc *sch, u32 log)
706 {
707 	struct fq_sched_data *q = qdisc_priv(sch);
708 	struct rb_root *array;
709 	void *old_fq_root;
710 	u32 idx;
711 
712 	if (q->fq_root && log == q->fq_trees_log)
713 		return 0;
714 
715 	/* If XPS was setup, we can allocate memory on right NUMA node */
716 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
717 			      netdev_queue_numa_node_read(sch->dev_queue));
718 	if (!array)
719 		return -ENOMEM;
720 
721 	for (idx = 0; idx < (1U << log); idx++)
722 		array[idx] = RB_ROOT;
723 
724 	sch_tree_lock(sch);
725 
726 	old_fq_root = q->fq_root;
727 	if (old_fq_root)
728 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
729 
730 	q->fq_root = array;
731 	q->fq_trees_log = log;
732 
733 	sch_tree_unlock(sch);
734 
735 	fq_free(old_fq_root);
736 
737 	return 0;
738 }
739 
740 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
741 	[TCA_FQ_UNSPEC]			= { .strict_start_type = TCA_FQ_TIMER_SLACK },
742 
743 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
744 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
745 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
746 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
747 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
748 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
749 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
750 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
751 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
752 	[TCA_FQ_ORPHAN_MASK]		= { .type = NLA_U32 },
753 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
754 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
755 	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
756 };
757 
758 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
759 		     struct netlink_ext_ack *extack)
760 {
761 	struct fq_sched_data *q = qdisc_priv(sch);
762 	struct nlattr *tb[TCA_FQ_MAX + 1];
763 	int err, drop_count = 0;
764 	unsigned drop_len = 0;
765 	u32 fq_log;
766 
767 	if (!opt)
768 		return -EINVAL;
769 
770 	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
771 					  NULL);
772 	if (err < 0)
773 		return err;
774 
775 	sch_tree_lock(sch);
776 
777 	fq_log = q->fq_trees_log;
778 
779 	if (tb[TCA_FQ_BUCKETS_LOG]) {
780 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
781 
782 		if (nval >= 1 && nval <= ilog2(256*1024))
783 			fq_log = nval;
784 		else
785 			err = -EINVAL;
786 	}
787 	if (tb[TCA_FQ_PLIMIT])
788 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
789 
790 	if (tb[TCA_FQ_FLOW_PLIMIT])
791 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
792 
793 	if (tb[TCA_FQ_QUANTUM]) {
794 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
795 
796 		if (quantum > 0 && quantum <= (1 << 20)) {
797 			q->quantum = quantum;
798 		} else {
799 			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
800 			err = -EINVAL;
801 		}
802 	}
803 
804 	if (tb[TCA_FQ_INITIAL_QUANTUM])
805 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
806 
807 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
808 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
809 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
810 
811 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
812 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
813 
814 		q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
815 	}
816 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
817 		q->low_rate_threshold =
818 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
819 
820 	if (tb[TCA_FQ_RATE_ENABLE]) {
821 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
822 
823 		if (enable <= 1)
824 			q->rate_enable = enable;
825 		else
826 			err = -EINVAL;
827 	}
828 
829 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
830 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
831 
832 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
833 	}
834 
835 	if (tb[TCA_FQ_ORPHAN_MASK])
836 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
837 
838 	if (tb[TCA_FQ_CE_THRESHOLD])
839 		q->ce_threshold = (u64)NSEC_PER_USEC *
840 				  nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
841 
842 	if (tb[TCA_FQ_TIMER_SLACK])
843 		q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]);
844 
845 	if (!err) {
846 		sch_tree_unlock(sch);
847 		err = fq_resize(sch, fq_log);
848 		sch_tree_lock(sch);
849 	}
850 	while (sch->q.qlen > sch->limit) {
851 		struct sk_buff *skb = fq_dequeue(sch);
852 
853 		if (!skb)
854 			break;
855 		drop_len += qdisc_pkt_len(skb);
856 		rtnl_kfree_skbs(skb, skb);
857 		drop_count++;
858 	}
859 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
860 
861 	sch_tree_unlock(sch);
862 	return err;
863 }
864 
865 static void fq_destroy(struct Qdisc *sch)
866 {
867 	struct fq_sched_data *q = qdisc_priv(sch);
868 
869 	fq_reset(sch);
870 	fq_free(q->fq_root);
871 	qdisc_watchdog_cancel(&q->watchdog);
872 }
873 
874 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
875 		   struct netlink_ext_ack *extack)
876 {
877 	struct fq_sched_data *q = qdisc_priv(sch);
878 	int err;
879 
880 	sch->limit		= 10000;
881 	q->flow_plimit		= 100;
882 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
883 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
884 	q->flow_refill_delay	= msecs_to_jiffies(40);
885 	q->flow_max_rate	= ~0UL;
886 	q->time_next_delayed_flow = ~0ULL;
887 	q->rate_enable		= 1;
888 	q->new_flows.first	= NULL;
889 	q->old_flows.first	= NULL;
890 	q->delayed		= RB_ROOT;
891 	q->fq_root		= NULL;
892 	q->fq_trees_log		= ilog2(1024);
893 	q->orphan_mask		= 1024 - 1;
894 	q->low_rate_threshold	= 550000 / 8;
895 
896 	q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
897 
898 	/* Default ce_threshold of 4294 seconds */
899 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
900 
901 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
902 
903 	if (opt)
904 		err = fq_change(sch, opt, extack);
905 	else
906 		err = fq_resize(sch, q->fq_trees_log);
907 
908 	return err;
909 }
910 
911 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
912 {
913 	struct fq_sched_data *q = qdisc_priv(sch);
914 	u64 ce_threshold = q->ce_threshold;
915 	struct nlattr *opts;
916 
917 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
918 	if (opts == NULL)
919 		goto nla_put_failure;
920 
921 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
922 
923 	do_div(ce_threshold, NSEC_PER_USEC);
924 
925 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
926 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
927 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
928 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
929 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
930 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
931 			min_t(unsigned long, q->flow_max_rate, ~0U)) ||
932 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
933 			jiffies_to_usecs(q->flow_refill_delay)) ||
934 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
935 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
936 			q->low_rate_threshold) ||
937 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
938 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) ||
939 	    nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack))
940 		goto nla_put_failure;
941 
942 	return nla_nest_end(skb, opts);
943 
944 nla_put_failure:
945 	return -1;
946 }
947 
948 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
949 {
950 	struct fq_sched_data *q = qdisc_priv(sch);
951 	struct tc_fq_qd_stats st;
952 
953 	sch_tree_lock(sch);
954 
955 	st.gc_flows		  = q->stat_gc_flows;
956 	st.highprio_packets	  = q->stat_internal_packets;
957 	st.tcp_retrans		  = 0;
958 	st.throttled		  = q->stat_throttled;
959 	st.flows_plimit		  = q->stat_flows_plimit;
960 	st.pkts_too_long	  = q->stat_pkts_too_long;
961 	st.allocation_errors	  = q->stat_allocation_errors;
962 	st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
963 				    ktime_get_ns();
964 	st.flows		  = q->flows;
965 	st.inactive_flows	  = q->inactive_flows;
966 	st.throttled_flows	  = q->throttled_flows;
967 	st.unthrottle_latency_ns  = min_t(unsigned long,
968 					  q->unthrottle_latency_ns, ~0U);
969 	st.ce_mark		  = q->stat_ce_mark;
970 	sch_tree_unlock(sch);
971 
972 	return gnet_stats_copy_app(d, &st, sizeof(st));
973 }
974 
975 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
976 	.id		=	"fq",
977 	.priv_size	=	sizeof(struct fq_sched_data),
978 
979 	.enqueue	=	fq_enqueue,
980 	.dequeue	=	fq_dequeue,
981 	.peek		=	qdisc_peek_dequeued,
982 	.init		=	fq_init,
983 	.reset		=	fq_reset,
984 	.destroy	=	fq_destroy,
985 	.change		=	fq_change,
986 	.dump		=	fq_dump,
987 	.dump_stats	=	fq_dump_stats,
988 	.owner		=	THIS_MODULE,
989 };
990 
991 static int __init fq_module_init(void)
992 {
993 	int ret;
994 
995 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
996 					   sizeof(struct fq_flow),
997 					   0, 0, NULL);
998 	if (!fq_flow_cachep)
999 		return -ENOMEM;
1000 
1001 	ret = register_qdisc(&fq_qdisc_ops);
1002 	if (ret)
1003 		kmem_cache_destroy(fq_flow_cachep);
1004 	return ret;
1005 }
1006 
1007 static void __exit fq_module_exit(void)
1008 {
1009 	unregister_qdisc(&fq_qdisc_ops);
1010 	kmem_cache_destroy(fq_flow_cachep);
1011 }
1012 
1013 module_init(fq_module_init)
1014 module_exit(fq_module_exit)
1015 MODULE_AUTHOR("Eric Dumazet");
1016 MODULE_LICENSE("GPL");
1017