• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
fq_skb_cb(struct sk_buff * skb)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 
fq_flow_set_detached(struct fq_flow * f)130 static void fq_flow_set_detached(struct fq_flow *f)
131 {
132 	f->next = &detached;
133 	f->age = jiffies;
134 }
135 
fq_flow_is_detached(const struct fq_flow * f)136 static bool fq_flow_is_detached(const struct fq_flow *f)
137 {
138 	return f->next == &detached;
139 }
140 
fq_flow_is_throttled(const struct fq_flow * f)141 static bool fq_flow_is_throttled(const struct fq_flow *f)
142 {
143 	return f->next == &throttled;
144 }
145 
fq_flow_add_tail(struct fq_flow_head * head,struct fq_flow * flow)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 
fq_flow_unset_throttled(struct fq_sched_data * q,struct fq_flow * f)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 
fq_flow_set_throttled(struct fq_sched_data * q,struct fq_flow * f)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 
fq_gc_candidate(const struct fq_flow * f)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 
fq_gc(struct fq_sched_data * q,struct rb_root * root,struct sock * sk)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 
fq_classify(struct sk_buff * skb,struct fq_sched_data * q)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 (q->rate_enable)
305 					smp_store_release(&sk->sk_pacing_status,
306 							  SK_PACING_FQ);
307 				if (fq_flow_is_throttled(f))
308 					fq_flow_unset_throttled(q, f);
309 				f->time_next_packet = 0ULL;
310 			}
311 			return f;
312 		}
313 		if (f->sk > sk)
314 			p = &parent->rb_right;
315 		else
316 			p = &parent->rb_left;
317 	}
318 
319 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
320 	if (unlikely(!f)) {
321 		q->stat_allocation_errors++;
322 		return &q->internal;
323 	}
324 	/* f->t_root is already zeroed after kmem_cache_zalloc() */
325 
326 	fq_flow_set_detached(f);
327 	f->sk = sk;
328 	if (skb->sk == sk) {
329 		f->socket_hash = sk->sk_hash;
330 		if (q->rate_enable)
331 			smp_store_release(&sk->sk_pacing_status,
332 					  SK_PACING_FQ);
333 	}
334 	f->credit = q->initial_quantum;
335 
336 	rb_link_node(&f->fq_node, parent, p);
337 	rb_insert_color(&f->fq_node, root);
338 
339 	q->flows++;
340 	q->inactive_flows++;
341 	return f;
342 }
343 
fq_peek(struct fq_flow * flow)344 static struct sk_buff *fq_peek(struct fq_flow *flow)
345 {
346 	struct sk_buff *skb = skb_rb_first(&flow->t_root);
347 	struct sk_buff *head = flow->head;
348 
349 	if (!skb)
350 		return head;
351 
352 	if (!head)
353 		return skb;
354 
355 	if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
356 		return skb;
357 	return head;
358 }
359 
fq_erase_head(struct Qdisc * sch,struct fq_flow * flow,struct sk_buff * skb)360 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
361 			  struct sk_buff *skb)
362 {
363 	if (skb == flow->head) {
364 		flow->head = skb->next;
365 	} else {
366 		rb_erase(&skb->rbnode, &flow->t_root);
367 		skb->dev = qdisc_dev(sch);
368 	}
369 }
370 
371 /* remove one skb from head of flow queue */
fq_dequeue_head(struct Qdisc * sch,struct fq_flow * flow)372 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
373 {
374 	struct sk_buff *skb = fq_peek(flow);
375 
376 	if (skb) {
377 		fq_erase_head(sch, flow, skb);
378 		skb_mark_not_on_list(skb);
379 		flow->qlen--;
380 		qdisc_qstats_backlog_dec(sch, skb);
381 		sch->q.qlen--;
382 	}
383 	return skb;
384 }
385 
flow_queue_add(struct fq_flow * flow,struct sk_buff * skb)386 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
387 {
388 	struct rb_node **p, *parent;
389 	struct sk_buff *head, *aux;
390 
391 	fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns();
392 
393 	head = flow->head;
394 	if (!head ||
395 	    fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
396 		if (!head)
397 			flow->head = skb;
398 		else
399 			flow->tail->next = skb;
400 		flow->tail = skb;
401 		skb->next = NULL;
402 		return;
403 	}
404 
405 	p = &flow->t_root.rb_node;
406 	parent = NULL;
407 
408 	while (*p) {
409 		parent = *p;
410 		aux = rb_to_skb(parent);
411 		if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
412 			p = &parent->rb_right;
413 		else
414 			p = &parent->rb_left;
415 	}
416 	rb_link_node(&skb->rbnode, parent, p);
417 	rb_insert_color(&skb->rbnode, &flow->t_root);
418 }
419 
fq_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)420 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
421 		      struct sk_buff **to_free)
422 {
423 	struct fq_sched_data *q = qdisc_priv(sch);
424 	struct fq_flow *f;
425 
426 	if (unlikely(sch->q.qlen >= sch->limit))
427 		return qdisc_drop(skb, sch, to_free);
428 
429 	f = fq_classify(skb, q);
430 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
431 		q->stat_flows_plimit++;
432 		return qdisc_drop(skb, sch, to_free);
433 	}
434 
435 	f->qlen++;
436 	qdisc_qstats_backlog_inc(sch, skb);
437 	if (fq_flow_is_detached(f)) {
438 		fq_flow_add_tail(&q->new_flows, f);
439 		if (time_after(jiffies, f->age + q->flow_refill_delay))
440 			f->credit = max_t(u32, f->credit, q->quantum);
441 		q->inactive_flows--;
442 	}
443 
444 	/* Note: this overwrites f->age */
445 	flow_queue_add(f, skb);
446 
447 	if (unlikely(f == &q->internal)) {
448 		q->stat_internal_packets++;
449 	}
450 	sch->q.qlen++;
451 
452 	return NET_XMIT_SUCCESS;
453 }
454 
fq_check_throttled(struct fq_sched_data * q,u64 now)455 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
456 {
457 	unsigned long sample;
458 	struct rb_node *p;
459 
460 	if (q->time_next_delayed_flow > now)
461 		return;
462 
463 	/* Update unthrottle latency EWMA.
464 	 * This is cheap and can help diagnosing timer/latency problems.
465 	 */
466 	sample = (unsigned long)(now - q->time_next_delayed_flow);
467 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
468 	q->unthrottle_latency_ns += sample >> 3;
469 
470 	q->time_next_delayed_flow = ~0ULL;
471 	while ((p = rb_first(&q->delayed)) != NULL) {
472 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
473 
474 		if (f->time_next_packet > now) {
475 			q->time_next_delayed_flow = f->time_next_packet;
476 			break;
477 		}
478 		fq_flow_unset_throttled(q, f);
479 	}
480 }
481 
fq_dequeue(struct Qdisc * sch)482 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
483 {
484 	struct fq_sched_data *q = qdisc_priv(sch);
485 	struct fq_flow_head *head;
486 	struct sk_buff *skb;
487 	struct fq_flow *f;
488 	unsigned long rate;
489 	u32 plen;
490 	u64 now;
491 
492 	if (!sch->q.qlen)
493 		return NULL;
494 
495 	skb = fq_dequeue_head(sch, &q->internal);
496 	if (skb)
497 		goto out;
498 
499 	now = ktime_get_ns();
500 	fq_check_throttled(q, now);
501 begin:
502 	head = &q->new_flows;
503 	if (!head->first) {
504 		head = &q->old_flows;
505 		if (!head->first) {
506 			if (q->time_next_delayed_flow != ~0ULL)
507 				qdisc_watchdog_schedule_ns(&q->watchdog,
508 							   q->time_next_delayed_flow);
509 			return NULL;
510 		}
511 	}
512 	f = head->first;
513 
514 	if (f->credit <= 0) {
515 		f->credit += q->quantum;
516 		head->first = f->next;
517 		fq_flow_add_tail(&q->old_flows, f);
518 		goto begin;
519 	}
520 
521 	skb = fq_peek(f);
522 	if (skb) {
523 		u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
524 					     f->time_next_packet);
525 
526 		if (now < time_next_packet) {
527 			head->first = f->next;
528 			f->time_next_packet = time_next_packet;
529 			fq_flow_set_throttled(q, f);
530 			goto begin;
531 		}
532 		if (time_next_packet &&
533 		    (s64)(now - time_next_packet - q->ce_threshold) > 0) {
534 			INET_ECN_set_ce(skb);
535 			q->stat_ce_mark++;
536 		}
537 	}
538 
539 	skb = fq_dequeue_head(sch, f);
540 	if (!skb) {
541 		head->first = f->next;
542 		/* force a pass through old_flows to prevent starvation */
543 		if ((head == &q->new_flows) && q->old_flows.first) {
544 			fq_flow_add_tail(&q->old_flows, f);
545 		} else {
546 			fq_flow_set_detached(f);
547 			q->inactive_flows++;
548 		}
549 		goto begin;
550 	}
551 	prefetch(&skb->end);
552 	plen = qdisc_pkt_len(skb);
553 	f->credit -= plen;
554 
555 	if (!q->rate_enable)
556 		goto out;
557 
558 	rate = q->flow_max_rate;
559 
560 	/* If EDT time was provided for this skb, we need to
561 	 * update f->time_next_packet only if this qdisc enforces
562 	 * a flow max rate.
563 	 */
564 	if (!skb->tstamp) {
565 		if (skb->sk)
566 			rate = min(skb->sk->sk_pacing_rate, rate);
567 
568 		if (rate <= q->low_rate_threshold) {
569 			f->credit = 0;
570 		} else {
571 			plen = max(plen, q->quantum);
572 			if (f->credit > 0)
573 				goto out;
574 		}
575 	}
576 	if (rate != ~0UL) {
577 		u64 len = (u64)plen * NSEC_PER_SEC;
578 
579 		if (likely(rate))
580 			len = div64_ul(len, rate);
581 		/* Since socket rate can change later,
582 		 * clamp the delay to 1 second.
583 		 * Really, providers of too big packets should be fixed !
584 		 */
585 		if (unlikely(len > NSEC_PER_SEC)) {
586 			len = NSEC_PER_SEC;
587 			q->stat_pkts_too_long++;
588 		}
589 		/* Account for schedule/timers drifts.
590 		 * f->time_next_packet was set when prior packet was sent,
591 		 * and current time (@now) can be too late by tens of us.
592 		 */
593 		if (f->time_next_packet)
594 			len -= min(len/2, now - f->time_next_packet);
595 		f->time_next_packet = now + len;
596 	}
597 out:
598 	qdisc_bstats_update(sch, skb);
599 	return skb;
600 }
601 
fq_flow_purge(struct fq_flow * flow)602 static void fq_flow_purge(struct fq_flow *flow)
603 {
604 	struct rb_node *p = rb_first(&flow->t_root);
605 
606 	while (p) {
607 		struct sk_buff *skb = rb_to_skb(p);
608 
609 		p = rb_next(p);
610 		rb_erase(&skb->rbnode, &flow->t_root);
611 		rtnl_kfree_skbs(skb, skb);
612 	}
613 	rtnl_kfree_skbs(flow->head, flow->tail);
614 	flow->head = NULL;
615 	flow->qlen = 0;
616 }
617 
fq_reset(struct Qdisc * sch)618 static void fq_reset(struct Qdisc *sch)
619 {
620 	struct fq_sched_data *q = qdisc_priv(sch);
621 	struct rb_root *root;
622 	struct rb_node *p;
623 	struct fq_flow *f;
624 	unsigned int idx;
625 
626 	sch->q.qlen = 0;
627 	sch->qstats.backlog = 0;
628 
629 	fq_flow_purge(&q->internal);
630 
631 	if (!q->fq_root)
632 		return;
633 
634 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
635 		root = &q->fq_root[idx];
636 		while ((p = rb_first(root)) != NULL) {
637 			f = rb_entry(p, struct fq_flow, fq_node);
638 			rb_erase(p, root);
639 
640 			fq_flow_purge(f);
641 
642 			kmem_cache_free(fq_flow_cachep, f);
643 		}
644 	}
645 	q->new_flows.first	= NULL;
646 	q->old_flows.first	= NULL;
647 	q->delayed		= RB_ROOT;
648 	q->flows		= 0;
649 	q->inactive_flows	= 0;
650 	q->throttled_flows	= 0;
651 }
652 
fq_rehash(struct fq_sched_data * q,struct rb_root * old_array,u32 old_log,struct rb_root * new_array,u32 new_log)653 static void fq_rehash(struct fq_sched_data *q,
654 		      struct rb_root *old_array, u32 old_log,
655 		      struct rb_root *new_array, u32 new_log)
656 {
657 	struct rb_node *op, **np, *parent;
658 	struct rb_root *oroot, *nroot;
659 	struct fq_flow *of, *nf;
660 	int fcnt = 0;
661 	u32 idx;
662 
663 	for (idx = 0; idx < (1U << old_log); idx++) {
664 		oroot = &old_array[idx];
665 		while ((op = rb_first(oroot)) != NULL) {
666 			rb_erase(op, oroot);
667 			of = rb_entry(op, struct fq_flow, fq_node);
668 			if (fq_gc_candidate(of)) {
669 				fcnt++;
670 				kmem_cache_free(fq_flow_cachep, of);
671 				continue;
672 			}
673 			nroot = &new_array[hash_ptr(of->sk, new_log)];
674 
675 			np = &nroot->rb_node;
676 			parent = NULL;
677 			while (*np) {
678 				parent = *np;
679 
680 				nf = rb_entry(parent, struct fq_flow, fq_node);
681 				BUG_ON(nf->sk == of->sk);
682 
683 				if (nf->sk > of->sk)
684 					np = &parent->rb_right;
685 				else
686 					np = &parent->rb_left;
687 			}
688 
689 			rb_link_node(&of->fq_node, parent, np);
690 			rb_insert_color(&of->fq_node, nroot);
691 		}
692 	}
693 	q->flows -= fcnt;
694 	q->inactive_flows -= fcnt;
695 	q->stat_gc_flows += fcnt;
696 }
697 
fq_free(void * addr)698 static void fq_free(void *addr)
699 {
700 	kvfree(addr);
701 }
702 
fq_resize(struct Qdisc * sch,u32 log)703 static int fq_resize(struct Qdisc *sch, u32 log)
704 {
705 	struct fq_sched_data *q = qdisc_priv(sch);
706 	struct rb_root *array;
707 	void *old_fq_root;
708 	u32 idx;
709 
710 	if (q->fq_root && log == q->fq_trees_log)
711 		return 0;
712 
713 	/* If XPS was setup, we can allocate memory on right NUMA node */
714 	array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
715 			      netdev_queue_numa_node_read(sch->dev_queue));
716 	if (!array)
717 		return -ENOMEM;
718 
719 	for (idx = 0; idx < (1U << log); idx++)
720 		array[idx] = RB_ROOT;
721 
722 	sch_tree_lock(sch);
723 
724 	old_fq_root = q->fq_root;
725 	if (old_fq_root)
726 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
727 
728 	q->fq_root = array;
729 	q->fq_trees_log = log;
730 
731 	sch_tree_unlock(sch);
732 
733 	fq_free(old_fq_root);
734 
735 	return 0;
736 }
737 
738 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
739 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
740 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
741 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
742 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
743 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
744 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
745 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
746 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
747 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
748 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
749 	[TCA_FQ_CE_THRESHOLD]		= { .type = NLA_U32 },
750 };
751 
fq_change(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)752 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
753 		     struct netlink_ext_ack *extack)
754 {
755 	struct fq_sched_data *q = qdisc_priv(sch);
756 	struct nlattr *tb[TCA_FQ_MAX + 1];
757 	int err, drop_count = 0;
758 	unsigned drop_len = 0;
759 	u32 fq_log;
760 
761 	if (!opt)
762 		return -EINVAL;
763 
764 	err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
765 					  NULL);
766 	if (err < 0)
767 		return err;
768 
769 	sch_tree_lock(sch);
770 
771 	fq_log = q->fq_trees_log;
772 
773 	if (tb[TCA_FQ_BUCKETS_LOG]) {
774 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
775 
776 		if (nval >= 1 && nval <= ilog2(256*1024))
777 			fq_log = nval;
778 		else
779 			err = -EINVAL;
780 	}
781 	if (tb[TCA_FQ_PLIMIT])
782 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
783 
784 	if (tb[TCA_FQ_FLOW_PLIMIT])
785 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
786 
787 	if (tb[TCA_FQ_QUANTUM]) {
788 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
789 
790 		if (quantum > 0 && quantum <= (1 << 20)) {
791 			q->quantum = quantum;
792 		} else {
793 			NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
794 			err = -EINVAL;
795 		}
796 	}
797 
798 	if (tb[TCA_FQ_INITIAL_QUANTUM])
799 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
800 
801 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
802 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
803 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
804 
805 	if (tb[TCA_FQ_FLOW_MAX_RATE]) {
806 		u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
807 
808 		q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
809 	}
810 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
811 		q->low_rate_threshold =
812 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
813 
814 	if (tb[TCA_FQ_RATE_ENABLE]) {
815 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
816 
817 		if (enable <= 1)
818 			q->rate_enable = enable;
819 		else
820 			err = -EINVAL;
821 	}
822 
823 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
824 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
825 
826 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
827 	}
828 
829 	if (tb[TCA_FQ_ORPHAN_MASK])
830 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
831 
832 	if (tb[TCA_FQ_CE_THRESHOLD])
833 		q->ce_threshold = (u64)NSEC_PER_USEC *
834 				  nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
835 
836 	if (!err) {
837 		sch_tree_unlock(sch);
838 		err = fq_resize(sch, fq_log);
839 		sch_tree_lock(sch);
840 	}
841 	while (sch->q.qlen > sch->limit) {
842 		struct sk_buff *skb = fq_dequeue(sch);
843 
844 		if (!skb)
845 			break;
846 		drop_len += qdisc_pkt_len(skb);
847 		rtnl_kfree_skbs(skb, skb);
848 		drop_count++;
849 	}
850 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
851 
852 	sch_tree_unlock(sch);
853 	return err;
854 }
855 
fq_destroy(struct Qdisc * sch)856 static void fq_destroy(struct Qdisc *sch)
857 {
858 	struct fq_sched_data *q = qdisc_priv(sch);
859 
860 	fq_reset(sch);
861 	fq_free(q->fq_root);
862 	qdisc_watchdog_cancel(&q->watchdog);
863 }
864 
fq_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)865 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
866 		   struct netlink_ext_ack *extack)
867 {
868 	struct fq_sched_data *q = qdisc_priv(sch);
869 	int err;
870 
871 	sch->limit		= 10000;
872 	q->flow_plimit		= 100;
873 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
874 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
875 	q->flow_refill_delay	= msecs_to_jiffies(40);
876 	q->flow_max_rate	= ~0UL;
877 	q->time_next_delayed_flow = ~0ULL;
878 	q->rate_enable		= 1;
879 	q->new_flows.first	= NULL;
880 	q->old_flows.first	= NULL;
881 	q->delayed		= RB_ROOT;
882 	q->fq_root		= NULL;
883 	q->fq_trees_log		= ilog2(1024);
884 	q->orphan_mask		= 1024 - 1;
885 	q->low_rate_threshold	= 550000 / 8;
886 
887 	/* Default ce_threshold of 4294 seconds */
888 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
889 
890 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
891 
892 	if (opt)
893 		err = fq_change(sch, opt, extack);
894 	else
895 		err = fq_resize(sch, q->fq_trees_log);
896 
897 	return err;
898 }
899 
fq_dump(struct Qdisc * sch,struct sk_buff * skb)900 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
901 {
902 	struct fq_sched_data *q = qdisc_priv(sch);
903 	u64 ce_threshold = q->ce_threshold;
904 	struct nlattr *opts;
905 
906 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
907 	if (opts == NULL)
908 		goto nla_put_failure;
909 
910 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
911 
912 	do_div(ce_threshold, NSEC_PER_USEC);
913 
914 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
915 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
916 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
917 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
918 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
919 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
920 			min_t(unsigned long, q->flow_max_rate, ~0U)) ||
921 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
922 			jiffies_to_usecs(q->flow_refill_delay)) ||
923 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
924 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
925 			q->low_rate_threshold) ||
926 	    nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
927 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
928 		goto nla_put_failure;
929 
930 	return nla_nest_end(skb, opts);
931 
932 nla_put_failure:
933 	return -1;
934 }
935 
fq_dump_stats(struct Qdisc * sch,struct gnet_dump * d)936 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
937 {
938 	struct fq_sched_data *q = qdisc_priv(sch);
939 	struct tc_fq_qd_stats st;
940 
941 	sch_tree_lock(sch);
942 
943 	st.gc_flows		  = q->stat_gc_flows;
944 	st.highprio_packets	  = q->stat_internal_packets;
945 	st.tcp_retrans		  = 0;
946 	st.throttled		  = q->stat_throttled;
947 	st.flows_plimit		  = q->stat_flows_plimit;
948 	st.pkts_too_long	  = q->stat_pkts_too_long;
949 	st.allocation_errors	  = q->stat_allocation_errors;
950 	st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
951 	st.flows		  = q->flows;
952 	st.inactive_flows	  = q->inactive_flows;
953 	st.throttled_flows	  = q->throttled_flows;
954 	st.unthrottle_latency_ns  = min_t(unsigned long,
955 					  q->unthrottle_latency_ns, ~0U);
956 	st.ce_mark		  = q->stat_ce_mark;
957 	sch_tree_unlock(sch);
958 
959 	return gnet_stats_copy_app(d, &st, sizeof(st));
960 }
961 
962 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
963 	.id		=	"fq",
964 	.priv_size	=	sizeof(struct fq_sched_data),
965 
966 	.enqueue	=	fq_enqueue,
967 	.dequeue	=	fq_dequeue,
968 	.peek		=	qdisc_peek_dequeued,
969 	.init		=	fq_init,
970 	.reset		=	fq_reset,
971 	.destroy	=	fq_destroy,
972 	.change		=	fq_change,
973 	.dump		=	fq_dump,
974 	.dump_stats	=	fq_dump_stats,
975 	.owner		=	THIS_MODULE,
976 };
977 
fq_module_init(void)978 static int __init fq_module_init(void)
979 {
980 	int ret;
981 
982 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
983 					   sizeof(struct fq_flow),
984 					   0, 0, NULL);
985 	if (!fq_flow_cachep)
986 		return -ENOMEM;
987 
988 	ret = register_qdisc(&fq_qdisc_ops);
989 	if (ret)
990 		kmem_cache_destroy(fq_flow_cachep);
991 	return ret;
992 }
993 
fq_module_exit(void)994 static void __exit fq_module_exit(void)
995 {
996 	unregister_qdisc(&fq_qdisc_ops);
997 	kmem_cache_destroy(fq_flow_cachep);
998 }
999 
1000 module_init(fq_module_init)
1001 module_exit(fq_module_exit)
1002 MODULE_AUTHOR("Eric Dumazet");
1003 MODULE_LICENSE("GPL");
1004