Loading...
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) 2013-2023 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
53struct fq_skb_cb {
54 u64 time_to_send;
55 u8 band;
56};
57
58static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
59{
60 qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb));
61 return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data;
62}
63
64/*
65 * Per flow structure, dynamically allocated.
66 * If packets have monotically increasing time_to_send, they are placed in O(1)
67 * in linear list (head,tail), otherwise are placed in a rbtree (t_root).
68 */
69struct fq_flow {
70/* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
71 struct rb_root t_root;
72 struct sk_buff *head; /* list of skbs for this flow : first skb */
73 union {
74 struct sk_buff *tail; /* last skb in the list */
75 unsigned long age; /* (jiffies | 1UL) when flow was emptied, for gc */
76 };
77 union {
78 struct rb_node fq_node; /* anchor in fq_root[] trees */
79 /* Following field is only used for q->internal,
80 * because q->internal is not hashed in fq_root[]
81 */
82 u64 stat_fastpath_packets;
83 };
84 struct sock *sk;
85 u32 socket_hash; /* sk_hash */
86 int qlen; /* number of packets in flow queue */
87
88/* Second cache line */
89 int credit;
90 int band;
91 struct fq_flow *next; /* next pointer in RR lists */
92
93 struct rb_node rate_node; /* anchor in q->delayed tree */
94 u64 time_next_packet;
95};
96
97struct fq_flow_head {
98 struct fq_flow *first;
99 struct fq_flow *last;
100};
101
102struct fq_perband_flows {
103 struct fq_flow_head new_flows;
104 struct fq_flow_head old_flows;
105 int credit;
106 int quantum; /* based on band nr : 576KB, 192KB, 64KB */
107};
108
109#define FQ_PRIO2BAND_CRUMB_SIZE ((TC_PRIO_MAX + 1) >> 2)
110
111struct fq_sched_data {
112/* Read mostly cache line */
113
114 u64 offload_horizon;
115 u32 quantum;
116 u32 initial_quantum;
117 u32 flow_refill_delay;
118 u32 flow_plimit; /* max packets per flow */
119 unsigned long flow_max_rate; /* optional max rate per flow */
120 u64 ce_threshold;
121 u64 horizon; /* horizon in ns */
122 u32 orphan_mask; /* mask for orphaned skb */
123 u32 low_rate_threshold;
124 struct rb_root *fq_root;
125 u8 rate_enable;
126 u8 fq_trees_log;
127 u8 horizon_drop;
128 u8 prio2band[FQ_PRIO2BAND_CRUMB_SIZE];
129 u32 timer_slack; /* hrtimer slack in ns */
130
131/* Read/Write fields. */
132
133 unsigned int band_nr; /* band being serviced in fq_dequeue() */
134
135 struct fq_perband_flows band_flows[FQ_BANDS];
136
137 struct fq_flow internal; /* fastpath queue. */
138 struct rb_root delayed; /* for rate limited flows */
139 u64 time_next_delayed_flow;
140 unsigned long unthrottle_latency_ns;
141
142 u32 band_pkt_count[FQ_BANDS];
143 u32 flows;
144 u32 inactive_flows; /* Flows with no packet to send. */
145 u32 throttled_flows;
146
147 u64 stat_throttled;
148 struct qdisc_watchdog watchdog;
149 u64 stat_gc_flows;
150
151/* Seldom used fields. */
152
153 u64 stat_band_drops[FQ_BANDS];
154 u64 stat_ce_mark;
155 u64 stat_horizon_drops;
156 u64 stat_horizon_caps;
157 u64 stat_flows_plimit;
158 u64 stat_pkts_too_long;
159 u64 stat_allocation_errors;
160};
161
162/* return the i-th 2-bit value ("crumb") */
163static u8 fq_prio2band(const u8 *prio2band, unsigned int prio)
164{
165 return (READ_ONCE(prio2band[prio / 4]) >> (2 * (prio & 0x3))) & 0x3;
166}
167
168/*
169 * f->tail and f->age share the same location.
170 * We can use the low order bit to differentiate if this location points
171 * to a sk_buff or contains a jiffies value, if we force this value to be odd.
172 * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
173 */
174static void fq_flow_set_detached(struct fq_flow *f)
175{
176 f->age = jiffies | 1UL;
177}
178
179static bool fq_flow_is_detached(const struct fq_flow *f)
180{
181 return !!(f->age & 1UL);
182}
183
184/* special value to mark a throttled flow (not on old/new list) */
185static struct fq_flow throttled;
186
187static bool fq_flow_is_throttled(const struct fq_flow *f)
188{
189 return f->next == &throttled;
190}
191
192enum new_flow {
193 NEW_FLOW,
194 OLD_FLOW
195};
196
197static void fq_flow_add_tail(struct fq_sched_data *q, struct fq_flow *flow,
198 enum new_flow list_sel)
199{
200 struct fq_perband_flows *pband = &q->band_flows[flow->band];
201 struct fq_flow_head *head = (list_sel == NEW_FLOW) ?
202 &pband->new_flows :
203 &pband->old_flows;
204
205 if (head->first)
206 head->last->next = flow;
207 else
208 head->first = flow;
209 head->last = flow;
210 flow->next = NULL;
211}
212
213static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
214{
215 rb_erase(&f->rate_node, &q->delayed);
216 q->throttled_flows--;
217 fq_flow_add_tail(q, f, OLD_FLOW);
218}
219
220static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
221{
222 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
223
224 while (*p) {
225 struct fq_flow *aux;
226
227 parent = *p;
228 aux = rb_entry(parent, struct fq_flow, rate_node);
229 if (f->time_next_packet >= aux->time_next_packet)
230 p = &parent->rb_right;
231 else
232 p = &parent->rb_left;
233 }
234 rb_link_node(&f->rate_node, parent, p);
235 rb_insert_color(&f->rate_node, &q->delayed);
236 q->throttled_flows++;
237 q->stat_throttled++;
238
239 f->next = &throttled;
240 if (q->time_next_delayed_flow > f->time_next_packet)
241 q->time_next_delayed_flow = f->time_next_packet;
242}
243
244
245static struct kmem_cache *fq_flow_cachep __read_mostly;
246
247
248/* limit number of collected flows per round */
249#define FQ_GC_MAX 8
250#define FQ_GC_AGE (3*HZ)
251
252static bool fq_gc_candidate(const struct fq_flow *f)
253{
254 return fq_flow_is_detached(f) &&
255 time_after(jiffies, f->age + FQ_GC_AGE);
256}
257
258static void fq_gc(struct fq_sched_data *q,
259 struct rb_root *root,
260 struct sock *sk)
261{
262 struct rb_node **p, *parent;
263 void *tofree[FQ_GC_MAX];
264 struct fq_flow *f;
265 int i, fcnt = 0;
266
267 p = &root->rb_node;
268 parent = NULL;
269 while (*p) {
270 parent = *p;
271
272 f = rb_entry(parent, struct fq_flow, fq_node);
273 if (f->sk == sk)
274 break;
275
276 if (fq_gc_candidate(f)) {
277 tofree[fcnt++] = f;
278 if (fcnt == FQ_GC_MAX)
279 break;
280 }
281
282 if (f->sk > sk)
283 p = &parent->rb_right;
284 else
285 p = &parent->rb_left;
286 }
287
288 if (!fcnt)
289 return;
290
291 for (i = fcnt; i > 0; ) {
292 f = tofree[--i];
293 rb_erase(&f->fq_node, root);
294 }
295 q->flows -= fcnt;
296 q->inactive_flows -= fcnt;
297 q->stat_gc_flows += fcnt;
298
299 kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
300}
301
302/* Fast path can be used if :
303 * 1) Packet tstamp is in the past, or within the pacing offload horizon.
304 * 2) FQ qlen == 0 OR
305 * (no flow is currently eligible for transmit,
306 * AND fast path queue has less than 8 packets)
307 * 3) No SO_MAX_PACING_RATE on the socket (if any).
308 * 4) No @maxrate attribute on this qdisc,
309 *
310 * FQ can not use generic TCQ_F_CAN_BYPASS infrastructure.
311 */
312static bool fq_fastpath_check(const struct Qdisc *sch, struct sk_buff *skb,
313 u64 now)
314{
315 const struct fq_sched_data *q = qdisc_priv(sch);
316 const struct sock *sk;
317
318 if (fq_skb_cb(skb)->time_to_send > now + q->offload_horizon)
319 return false;
320
321 if (sch->q.qlen != 0) {
322 /* Even if some packets are stored in this qdisc,
323 * we can still enable fast path if all of them are
324 * scheduled in the future (ie no flows are eligible)
325 * or in the fast path queue.
326 */
327 if (q->flows != q->inactive_flows + q->throttled_flows)
328 return false;
329
330 /* Do not allow fast path queue to explode, we want Fair Queue mode
331 * under pressure.
332 */
333 if (q->internal.qlen >= 8)
334 return false;
335
336 /* Ordering invariants fall apart if some delayed flows
337 * are ready but we haven't serviced them, yet.
338 */
339 if (q->time_next_delayed_flow <= now + q->offload_horizon)
340 return false;
341 }
342
343 sk = skb->sk;
344 if (sk && sk_fullsock(sk) && !sk_is_tcp(sk) &&
345 sk->sk_max_pacing_rate != ~0UL)
346 return false;
347
348 if (q->flow_max_rate != ~0UL)
349 return false;
350
351 return true;
352}
353
354static struct fq_flow *fq_classify(struct Qdisc *sch, struct sk_buff *skb,
355 u64 now)
356{
357 struct fq_sched_data *q = qdisc_priv(sch);
358 struct rb_node **p, *parent;
359 struct sock *sk = skb->sk;
360 struct rb_root *root;
361 struct fq_flow *f;
362
363 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
364 * or a listener (SYNCOOKIE mode)
365 * 1) request sockets are not full blown,
366 * they do not contain sk_pacing_rate
367 * 2) They are not part of a 'flow' yet
368 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
369 * especially if the listener set SO_MAX_PACING_RATE
370 * 4) We pretend they are orphaned
371 * TCP can also associate TIME_WAIT sockets with RST or ACK packets.
372 */
373 if (!sk || sk_listener_or_tw(sk)) {
374 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
375
376 /* By forcing low order bit to 1, we make sure to not
377 * collide with a local flow (socket pointers are word aligned)
378 */
379 sk = (struct sock *)((hash << 1) | 1UL);
380 skb_orphan(skb);
381 } else if (sk->sk_state == TCP_CLOSE) {
382 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
383 /*
384 * Sockets in TCP_CLOSE are non connected.
385 * Typical use case is UDP sockets, they can send packets
386 * with sendto() to many different destinations.
387 * We probably could use a generic bit advertising
388 * non connected sockets, instead of sk_state == TCP_CLOSE,
389 * if we care enough.
390 */
391 sk = (struct sock *)((hash << 1) | 1UL);
392 }
393
394 if (fq_fastpath_check(sch, skb, now)) {
395 q->internal.stat_fastpath_packets++;
396 if (skb->sk == sk && q->rate_enable &&
397 READ_ONCE(sk->sk_pacing_status) != SK_PACING_FQ)
398 smp_store_release(&sk->sk_pacing_status,
399 SK_PACING_FQ);
400 return &q->internal;
401 }
402
403 root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
404
405 fq_gc(q, root, sk);
406
407 p = &root->rb_node;
408 parent = NULL;
409 while (*p) {
410 parent = *p;
411
412 f = rb_entry(parent, struct fq_flow, fq_node);
413 if (f->sk == sk) {
414 /* socket might have been reallocated, so check
415 * if its sk_hash is the same.
416 * It not, we need to refill credit with
417 * initial quantum
418 */
419 if (unlikely(skb->sk == sk &&
420 f->socket_hash != sk->sk_hash)) {
421 f->credit = q->initial_quantum;
422 f->socket_hash = sk->sk_hash;
423 if (q->rate_enable)
424 smp_store_release(&sk->sk_pacing_status,
425 SK_PACING_FQ);
426 if (fq_flow_is_throttled(f))
427 fq_flow_unset_throttled(q, f);
428 f->time_next_packet = 0ULL;
429 }
430 return f;
431 }
432 if (f->sk > sk)
433 p = &parent->rb_right;
434 else
435 p = &parent->rb_left;
436 }
437
438 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
439 if (unlikely(!f)) {
440 q->stat_allocation_errors++;
441 return &q->internal;
442 }
443 /* f->t_root is already zeroed after kmem_cache_zalloc() */
444
445 fq_flow_set_detached(f);
446 f->sk = sk;
447 if (skb->sk == sk) {
448 f->socket_hash = sk->sk_hash;
449 if (q->rate_enable)
450 smp_store_release(&sk->sk_pacing_status,
451 SK_PACING_FQ);
452 }
453 f->credit = q->initial_quantum;
454
455 rb_link_node(&f->fq_node, parent, p);
456 rb_insert_color(&f->fq_node, root);
457
458 q->flows++;
459 q->inactive_flows++;
460 return f;
461}
462
463static struct sk_buff *fq_peek(struct fq_flow *flow)
464{
465 struct sk_buff *skb = skb_rb_first(&flow->t_root);
466 struct sk_buff *head = flow->head;
467
468 if (!skb)
469 return head;
470
471 if (!head)
472 return skb;
473
474 if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
475 return skb;
476 return head;
477}
478
479static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
480 struct sk_buff *skb)
481{
482 if (skb == flow->head) {
483 flow->head = skb->next;
484 } else {
485 rb_erase(&skb->rbnode, &flow->t_root);
486 skb->dev = qdisc_dev(sch);
487 }
488}
489
490/* Remove one skb from flow queue.
491 * This skb must be the return value of prior fq_peek().
492 */
493static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
494 struct sk_buff *skb)
495{
496 fq_erase_head(sch, flow, skb);
497 skb_mark_not_on_list(skb);
498 qdisc_qstats_backlog_dec(sch, skb);
499 sch->q.qlen--;
500}
501
502static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
503{
504 struct rb_node **p, *parent;
505 struct sk_buff *head, *aux;
506
507 head = flow->head;
508 if (!head ||
509 fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
510 if (!head)
511 flow->head = skb;
512 else
513 flow->tail->next = skb;
514 flow->tail = skb;
515 skb->next = NULL;
516 return;
517 }
518
519 p = &flow->t_root.rb_node;
520 parent = NULL;
521
522 while (*p) {
523 parent = *p;
524 aux = rb_to_skb(parent);
525 if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
526 p = &parent->rb_right;
527 else
528 p = &parent->rb_left;
529 }
530 rb_link_node(&skb->rbnode, parent, p);
531 rb_insert_color(&skb->rbnode, &flow->t_root);
532}
533
534static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
535 const struct fq_sched_data *q, u64 now)
536{
537 return unlikely((s64)skb->tstamp > (s64)(now + q->horizon));
538}
539
540static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
541 struct sk_buff **to_free)
542{
543 struct fq_sched_data *q = qdisc_priv(sch);
544 struct fq_flow *f;
545 u64 now;
546 u8 band;
547
548 band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
549 if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
550 q->stat_band_drops[band]++;
551 return qdisc_drop(skb, sch, to_free);
552 }
553
554 now = ktime_get_ns();
555 if (!skb->tstamp) {
556 fq_skb_cb(skb)->time_to_send = now;
557 } else {
558 /* Check if packet timestamp is too far in the future. */
559 if (fq_packet_beyond_horizon(skb, q, now)) {
560 if (q->horizon_drop) {
561 q->stat_horizon_drops++;
562 return qdisc_drop(skb, sch, to_free);
563 }
564 q->stat_horizon_caps++;
565 skb->tstamp = now + q->horizon;
566 }
567 fq_skb_cb(skb)->time_to_send = skb->tstamp;
568 }
569
570 f = fq_classify(sch, skb, now);
571
572 if (f != &q->internal) {
573 if (unlikely(f->qlen >= q->flow_plimit)) {
574 q->stat_flows_plimit++;
575 return qdisc_drop(skb, sch, to_free);
576 }
577
578 if (fq_flow_is_detached(f)) {
579 fq_flow_add_tail(q, f, NEW_FLOW);
580 if (time_after(jiffies, f->age + q->flow_refill_delay))
581 f->credit = max_t(u32, f->credit, q->quantum);
582 }
583
584 f->band = band;
585 q->band_pkt_count[band]++;
586 fq_skb_cb(skb)->band = band;
587 if (f->qlen == 0)
588 q->inactive_flows--;
589 }
590
591 f->qlen++;
592 /* Note: this overwrites f->age */
593 flow_queue_add(f, skb);
594
595 qdisc_qstats_backlog_inc(sch, skb);
596 sch->q.qlen++;
597
598 return NET_XMIT_SUCCESS;
599}
600
601static void fq_check_throttled(struct fq_sched_data *q, u64 now)
602{
603 unsigned long sample;
604 struct rb_node *p;
605
606 if (q->time_next_delayed_flow > now + q->offload_horizon)
607 return;
608
609 /* Update unthrottle latency EWMA.
610 * This is cheap and can help diagnosing timer/latency problems.
611 */
612 sample = (unsigned long)(now - q->time_next_delayed_flow);
613 if ((long)sample > 0) {
614 q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
615 q->unthrottle_latency_ns += sample >> 3;
616 }
617 now += q->offload_horizon;
618
619 q->time_next_delayed_flow = ~0ULL;
620 while ((p = rb_first(&q->delayed)) != NULL) {
621 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
622
623 if (f->time_next_packet > now) {
624 q->time_next_delayed_flow = f->time_next_packet;
625 break;
626 }
627 fq_flow_unset_throttled(q, f);
628 }
629}
630
631static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
632{
633 if (pband->credit <= 0)
634 return NULL;
635
636 if (pband->new_flows.first)
637 return &pband->new_flows;
638
639 return pband->old_flows.first ? &pband->old_flows : NULL;
640}
641
642static struct sk_buff *fq_dequeue(struct Qdisc *sch)
643{
644 struct fq_sched_data *q = qdisc_priv(sch);
645 struct fq_perband_flows *pband;
646 struct fq_flow_head *head;
647 struct sk_buff *skb;
648 struct fq_flow *f;
649 unsigned long rate;
650 int retry;
651 u32 plen;
652 u64 now;
653
654 if (!sch->q.qlen)
655 return NULL;
656
657 skb = fq_peek(&q->internal);
658 if (unlikely(skb)) {
659 q->internal.qlen--;
660 fq_dequeue_skb(sch, &q->internal, skb);
661 goto out;
662 }
663
664 now = ktime_get_ns();
665 fq_check_throttled(q, now);
666 retry = 0;
667 pband = &q->band_flows[q->band_nr];
668begin:
669 head = fq_pband_head_select(pband);
670 if (!head) {
671 while (++retry <= FQ_BANDS) {
672 if (++q->band_nr == FQ_BANDS)
673 q->band_nr = 0;
674 pband = &q->band_flows[q->band_nr];
675 pband->credit = min(pband->credit + pband->quantum,
676 pband->quantum);
677 if (pband->credit > 0)
678 goto begin;
679 retry = 0;
680 }
681 if (q->time_next_delayed_flow != ~0ULL)
682 qdisc_watchdog_schedule_range_ns(&q->watchdog,
683 q->time_next_delayed_flow,
684 q->timer_slack);
685 return NULL;
686 }
687 f = head->first;
688 retry = 0;
689 if (f->credit <= 0) {
690 f->credit += q->quantum;
691 head->first = f->next;
692 fq_flow_add_tail(q, f, OLD_FLOW);
693 goto begin;
694 }
695
696 skb = fq_peek(f);
697 if (skb) {
698 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
699 f->time_next_packet);
700
701 if (now + q->offload_horizon < time_next_packet) {
702 head->first = f->next;
703 f->time_next_packet = time_next_packet;
704 fq_flow_set_throttled(q, f);
705 goto begin;
706 }
707 prefetch(&skb->end);
708 if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
709 INET_ECN_set_ce(skb);
710 q->stat_ce_mark++;
711 }
712 if (--f->qlen == 0)
713 q->inactive_flows++;
714 q->band_pkt_count[fq_skb_cb(skb)->band]--;
715 fq_dequeue_skb(sch, f, skb);
716 } else {
717 head->first = f->next;
718 /* force a pass through old_flows to prevent starvation */
719 if (head == &pband->new_flows) {
720 fq_flow_add_tail(q, f, OLD_FLOW);
721 } else {
722 fq_flow_set_detached(f);
723 }
724 goto begin;
725 }
726 plen = qdisc_pkt_len(skb);
727 f->credit -= plen;
728 pband->credit -= plen;
729
730 if (!q->rate_enable)
731 goto out;
732
733 rate = q->flow_max_rate;
734
735 /* If EDT time was provided for this skb, we need to
736 * update f->time_next_packet only if this qdisc enforces
737 * a flow max rate.
738 */
739 if (!skb->tstamp) {
740 if (skb->sk)
741 rate = min(READ_ONCE(skb->sk->sk_pacing_rate), rate);
742
743 if (rate <= q->low_rate_threshold) {
744 f->credit = 0;
745 } else {
746 plen = max(plen, q->quantum);
747 if (f->credit > 0)
748 goto out;
749 }
750 }
751 if (rate != ~0UL) {
752 u64 len = (u64)plen * NSEC_PER_SEC;
753
754 if (likely(rate))
755 len = div64_ul(len, rate);
756 /* Since socket rate can change later,
757 * clamp the delay to 1 second.
758 * Really, providers of too big packets should be fixed !
759 */
760 if (unlikely(len > NSEC_PER_SEC)) {
761 len = NSEC_PER_SEC;
762 q->stat_pkts_too_long++;
763 }
764 /* Account for schedule/timers drifts.
765 * f->time_next_packet was set when prior packet was sent,
766 * and current time (@now) can be too late by tens of us.
767 */
768 if (f->time_next_packet)
769 len -= min(len/2, now - f->time_next_packet);
770 f->time_next_packet = now + len;
771 }
772out:
773 qdisc_bstats_update(sch, skb);
774 return skb;
775}
776
777static void fq_flow_purge(struct fq_flow *flow)
778{
779 struct rb_node *p = rb_first(&flow->t_root);
780
781 while (p) {
782 struct sk_buff *skb = rb_to_skb(p);
783
784 p = rb_next(p);
785 rb_erase(&skb->rbnode, &flow->t_root);
786 rtnl_kfree_skbs(skb, skb);
787 }
788 rtnl_kfree_skbs(flow->head, flow->tail);
789 flow->head = NULL;
790 flow->qlen = 0;
791}
792
793static void fq_reset(struct Qdisc *sch)
794{
795 struct fq_sched_data *q = qdisc_priv(sch);
796 struct rb_root *root;
797 struct rb_node *p;
798 struct fq_flow *f;
799 unsigned int idx;
800
801 sch->q.qlen = 0;
802 sch->qstats.backlog = 0;
803
804 fq_flow_purge(&q->internal);
805
806 if (!q->fq_root)
807 return;
808
809 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
810 root = &q->fq_root[idx];
811 while ((p = rb_first(root)) != NULL) {
812 f = rb_entry(p, struct fq_flow, fq_node);
813 rb_erase(p, root);
814
815 fq_flow_purge(f);
816
817 kmem_cache_free(fq_flow_cachep, f);
818 }
819 }
820 for (idx = 0; idx < FQ_BANDS; idx++) {
821 q->band_flows[idx].new_flows.first = NULL;
822 q->band_flows[idx].old_flows.first = NULL;
823 }
824 q->delayed = RB_ROOT;
825 q->flows = 0;
826 q->inactive_flows = 0;
827 q->throttled_flows = 0;
828}
829
830static void fq_rehash(struct fq_sched_data *q,
831 struct rb_root *old_array, u32 old_log,
832 struct rb_root *new_array, u32 new_log)
833{
834 struct rb_node *op, **np, *parent;
835 struct rb_root *oroot, *nroot;
836 struct fq_flow *of, *nf;
837 int fcnt = 0;
838 u32 idx;
839
840 for (idx = 0; idx < (1U << old_log); idx++) {
841 oroot = &old_array[idx];
842 while ((op = rb_first(oroot)) != NULL) {
843 rb_erase(op, oroot);
844 of = rb_entry(op, struct fq_flow, fq_node);
845 if (fq_gc_candidate(of)) {
846 fcnt++;
847 kmem_cache_free(fq_flow_cachep, of);
848 continue;
849 }
850 nroot = &new_array[hash_ptr(of->sk, new_log)];
851
852 np = &nroot->rb_node;
853 parent = NULL;
854 while (*np) {
855 parent = *np;
856
857 nf = rb_entry(parent, struct fq_flow, fq_node);
858 BUG_ON(nf->sk == of->sk);
859
860 if (nf->sk > of->sk)
861 np = &parent->rb_right;
862 else
863 np = &parent->rb_left;
864 }
865
866 rb_link_node(&of->fq_node, parent, np);
867 rb_insert_color(&of->fq_node, nroot);
868 }
869 }
870 q->flows -= fcnt;
871 q->inactive_flows -= fcnt;
872 q->stat_gc_flows += fcnt;
873}
874
875static void fq_free(void *addr)
876{
877 kvfree(addr);
878}
879
880static int fq_resize(struct Qdisc *sch, u32 log)
881{
882 struct fq_sched_data *q = qdisc_priv(sch);
883 struct rb_root *array;
884 void *old_fq_root;
885 u32 idx;
886
887 if (q->fq_root && log == q->fq_trees_log)
888 return 0;
889
890 /* If XPS was setup, we can allocate memory on right NUMA node */
891 array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
892 netdev_queue_numa_node_read(sch->dev_queue));
893 if (!array)
894 return -ENOMEM;
895
896 for (idx = 0; idx < (1U << log); idx++)
897 array[idx] = RB_ROOT;
898
899 sch_tree_lock(sch);
900
901 old_fq_root = q->fq_root;
902 if (old_fq_root)
903 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
904
905 q->fq_root = array;
906 WRITE_ONCE(q->fq_trees_log, log);
907
908 sch_tree_unlock(sch);
909
910 fq_free(old_fq_root);
911
912 return 0;
913}
914
915static const struct netlink_range_validation iq_range = {
916 .max = INT_MAX,
917};
918
919static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
920 [TCA_FQ_UNSPEC] = { .strict_start_type = TCA_FQ_TIMER_SLACK },
921
922 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
923 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
924 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
925 [TCA_FQ_INITIAL_QUANTUM] = NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
926 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
927 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
928 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
929 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
930 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
931 [TCA_FQ_ORPHAN_MASK] = { .type = NLA_U32 },
932 [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 },
933 [TCA_FQ_CE_THRESHOLD] = { .type = NLA_U32 },
934 [TCA_FQ_TIMER_SLACK] = { .type = NLA_U32 },
935 [TCA_FQ_HORIZON] = { .type = NLA_U32 },
936 [TCA_FQ_HORIZON_DROP] = { .type = NLA_U8 },
937 [TCA_FQ_PRIOMAP] = NLA_POLICY_EXACT_LEN(sizeof(struct tc_prio_qopt)),
938 [TCA_FQ_WEIGHTS] = NLA_POLICY_EXACT_LEN(FQ_BANDS * sizeof(s32)),
939 [TCA_FQ_OFFLOAD_HORIZON] = { .type = NLA_U32 },
940};
941
942/* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
943static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
944{
945 const int num_elems = TC_PRIO_MAX + 1;
946 u8 tmp[FQ_PRIO2BAND_CRUMB_SIZE];
947 int i;
948
949 memset(tmp, 0, sizeof(tmp));
950 for (i = 0; i < num_elems; i++)
951 tmp[i / 4] |= in[i] << (2 * (i & 0x3));
952
953 for (i = 0; i < FQ_PRIO2BAND_CRUMB_SIZE; i++)
954 WRITE_ONCE(out[i], tmp[i]);
955}
956
957static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
958{
959 const int num_elems = TC_PRIO_MAX + 1;
960 int i;
961
962 for (i = 0; i < num_elems; i++)
963 out[i] = fq_prio2band(in, i);
964}
965
966static int fq_load_weights(struct fq_sched_data *q,
967 const struct nlattr *attr,
968 struct netlink_ext_ack *extack)
969{
970 s32 *weights = nla_data(attr);
971 int i;
972
973 for (i = 0; i < FQ_BANDS; i++) {
974 if (weights[i] < FQ_MIN_WEIGHT) {
975 NL_SET_ERR_MSG_FMT_MOD(extack, "Weight %d less that minimum allowed %d",
976 weights[i], FQ_MIN_WEIGHT);
977 return -EINVAL;
978 }
979 }
980 for (i = 0; i < FQ_BANDS; i++)
981 WRITE_ONCE(q->band_flows[i].quantum, weights[i]);
982 return 0;
983}
984
985static int fq_load_priomap(struct fq_sched_data *q,
986 const struct nlattr *attr,
987 struct netlink_ext_ack *extack)
988{
989 const struct tc_prio_qopt *map = nla_data(attr);
990 int i;
991
992 if (map->bands != FQ_BANDS) {
993 NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
994 return -EINVAL;
995 }
996 for (i = 0; i < TC_PRIO_MAX + 1; i++) {
997 if (map->priomap[i] >= FQ_BANDS) {
998 NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
999 i, map->priomap[i]);
1000 return -EINVAL;
1001 }
1002 }
1003 fq_prio2band_compress_crumb(map->priomap, q->prio2band);
1004 return 0;
1005}
1006
1007static int fq_change(struct Qdisc *sch, struct nlattr *opt,
1008 struct netlink_ext_ack *extack)
1009{
1010 struct fq_sched_data *q = qdisc_priv(sch);
1011 struct nlattr *tb[TCA_FQ_MAX + 1];
1012 int err, drop_count = 0;
1013 unsigned drop_len = 0;
1014 u32 fq_log;
1015
1016 err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
1017 NULL);
1018 if (err < 0)
1019 return err;
1020
1021 sch_tree_lock(sch);
1022
1023 fq_log = q->fq_trees_log;
1024
1025 if (tb[TCA_FQ_BUCKETS_LOG]) {
1026 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
1027
1028 if (nval >= 1 && nval <= ilog2(256*1024))
1029 fq_log = nval;
1030 else
1031 err = -EINVAL;
1032 }
1033 if (tb[TCA_FQ_PLIMIT])
1034 WRITE_ONCE(sch->limit,
1035 nla_get_u32(tb[TCA_FQ_PLIMIT]));
1036
1037 if (tb[TCA_FQ_FLOW_PLIMIT])
1038 WRITE_ONCE(q->flow_plimit,
1039 nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]));
1040
1041 if (tb[TCA_FQ_QUANTUM]) {
1042 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
1043
1044 if (quantum > 0 && quantum <= (1 << 20)) {
1045 WRITE_ONCE(q->quantum, quantum);
1046 } else {
1047 NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
1048 err = -EINVAL;
1049 }
1050 }
1051
1052 if (tb[TCA_FQ_INITIAL_QUANTUM])
1053 WRITE_ONCE(q->initial_quantum,
1054 nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]));
1055
1056 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
1057 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
1058 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
1059
1060 if (tb[TCA_FQ_FLOW_MAX_RATE]) {
1061 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
1062
1063 WRITE_ONCE(q->flow_max_rate,
1064 (rate == ~0U) ? ~0UL : rate);
1065 }
1066 if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
1067 WRITE_ONCE(q->low_rate_threshold,
1068 nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]));
1069
1070 if (tb[TCA_FQ_RATE_ENABLE]) {
1071 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
1072
1073 if (enable <= 1)
1074 WRITE_ONCE(q->rate_enable,
1075 enable);
1076 else
1077 err = -EINVAL;
1078 }
1079
1080 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
1081 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
1082
1083 WRITE_ONCE(q->flow_refill_delay,
1084 usecs_to_jiffies(usecs_delay));
1085 }
1086
1087 if (!err && tb[TCA_FQ_PRIOMAP])
1088 err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
1089
1090 if (!err && tb[TCA_FQ_WEIGHTS])
1091 err = fq_load_weights(q, tb[TCA_FQ_WEIGHTS], extack);
1092
1093 if (tb[TCA_FQ_ORPHAN_MASK])
1094 WRITE_ONCE(q->orphan_mask,
1095 nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]));
1096
1097 if (tb[TCA_FQ_CE_THRESHOLD])
1098 WRITE_ONCE(q->ce_threshold,
1099 (u64)NSEC_PER_USEC *
1100 nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]));
1101
1102 if (tb[TCA_FQ_TIMER_SLACK])
1103 WRITE_ONCE(q->timer_slack,
1104 nla_get_u32(tb[TCA_FQ_TIMER_SLACK]));
1105
1106 if (tb[TCA_FQ_HORIZON])
1107 WRITE_ONCE(q->horizon,
1108 (u64)NSEC_PER_USEC *
1109 nla_get_u32(tb[TCA_FQ_HORIZON]));
1110
1111 if (tb[TCA_FQ_HORIZON_DROP])
1112 WRITE_ONCE(q->horizon_drop,
1113 nla_get_u8(tb[TCA_FQ_HORIZON_DROP]));
1114
1115 if (tb[TCA_FQ_OFFLOAD_HORIZON]) {
1116 u64 offload_horizon = (u64)NSEC_PER_USEC *
1117 nla_get_u32(tb[TCA_FQ_OFFLOAD_HORIZON]);
1118
1119 if (offload_horizon <= qdisc_dev(sch)->max_pacing_offload_horizon) {
1120 WRITE_ONCE(q->offload_horizon, offload_horizon);
1121 } else {
1122 NL_SET_ERR_MSG_MOD(extack, "invalid offload_horizon");
1123 err = -EINVAL;
1124 }
1125 }
1126 if (!err) {
1127
1128 sch_tree_unlock(sch);
1129 err = fq_resize(sch, fq_log);
1130 sch_tree_lock(sch);
1131 }
1132 while (sch->q.qlen > sch->limit) {
1133 struct sk_buff *skb = fq_dequeue(sch);
1134
1135 if (!skb)
1136 break;
1137 drop_len += qdisc_pkt_len(skb);
1138 rtnl_kfree_skbs(skb, skb);
1139 drop_count++;
1140 }
1141 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
1142
1143 sch_tree_unlock(sch);
1144 return err;
1145}
1146
1147static void fq_destroy(struct Qdisc *sch)
1148{
1149 struct fq_sched_data *q = qdisc_priv(sch);
1150
1151 fq_reset(sch);
1152 fq_free(q->fq_root);
1153 qdisc_watchdog_cancel(&q->watchdog);
1154}
1155
1156static int fq_init(struct Qdisc *sch, struct nlattr *opt,
1157 struct netlink_ext_ack *extack)
1158{
1159 struct fq_sched_data *q = qdisc_priv(sch);
1160 int i, err;
1161
1162 sch->limit = 10000;
1163 q->flow_plimit = 100;
1164 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
1165 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
1166 q->flow_refill_delay = msecs_to_jiffies(40);
1167 q->flow_max_rate = ~0UL;
1168 q->time_next_delayed_flow = ~0ULL;
1169 q->rate_enable = 1;
1170 for (i = 0; i < FQ_BANDS; i++) {
1171 q->band_flows[i].new_flows.first = NULL;
1172 q->band_flows[i].old_flows.first = NULL;
1173 }
1174 q->band_flows[0].quantum = 9 << 16;
1175 q->band_flows[1].quantum = 3 << 16;
1176 q->band_flows[2].quantum = 1 << 16;
1177 q->delayed = RB_ROOT;
1178 q->fq_root = NULL;
1179 q->fq_trees_log = ilog2(1024);
1180 q->orphan_mask = 1024 - 1;
1181 q->low_rate_threshold = 550000 / 8;
1182
1183 q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
1184
1185 q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
1186 q->horizon_drop = 1; /* by default, drop packets beyond horizon */
1187
1188 /* Default ce_threshold of 4294 seconds */
1189 q->ce_threshold = (u64)NSEC_PER_USEC * ~0U;
1190
1191 fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
1192 qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
1193
1194 if (opt)
1195 err = fq_change(sch, opt, extack);
1196 else
1197 err = fq_resize(sch, q->fq_trees_log);
1198
1199 return err;
1200}
1201
1202static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
1203{
1204 struct fq_sched_data *q = qdisc_priv(sch);
1205 struct tc_prio_qopt prio = {
1206 .bands = FQ_BANDS,
1207 };
1208 struct nlattr *opts;
1209 u64 offload_horizon;
1210 u64 ce_threshold;
1211 s32 weights[3];
1212 u64 horizon;
1213
1214 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1215 if (opts == NULL)
1216 goto nla_put_failure;
1217
1218 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
1219
1220 ce_threshold = READ_ONCE(q->ce_threshold);
1221 do_div(ce_threshold, NSEC_PER_USEC);
1222
1223 horizon = READ_ONCE(q->horizon);
1224 do_div(horizon, NSEC_PER_USEC);
1225
1226 offload_horizon = READ_ONCE(q->offload_horizon);
1227 do_div(offload_horizon, NSEC_PER_USEC);
1228
1229 if (nla_put_u32(skb, TCA_FQ_PLIMIT,
1230 READ_ONCE(sch->limit)) ||
1231 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT,
1232 READ_ONCE(q->flow_plimit)) ||
1233 nla_put_u32(skb, TCA_FQ_QUANTUM,
1234 READ_ONCE(q->quantum)) ||
1235 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM,
1236 READ_ONCE(q->initial_quantum)) ||
1237 nla_put_u32(skb, TCA_FQ_RATE_ENABLE,
1238 READ_ONCE(q->rate_enable)) ||
1239 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
1240 min_t(unsigned long,
1241 READ_ONCE(q->flow_max_rate), ~0U)) ||
1242 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
1243 jiffies_to_usecs(READ_ONCE(q->flow_refill_delay))) ||
1244 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK,
1245 READ_ONCE(q->orphan_mask)) ||
1246 nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
1247 READ_ONCE(q->low_rate_threshold)) ||
1248 nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
1249 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG,
1250 READ_ONCE(q->fq_trees_log)) ||
1251 nla_put_u32(skb, TCA_FQ_TIMER_SLACK,
1252 READ_ONCE(q->timer_slack)) ||
1253 nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
1254 nla_put_u32(skb, TCA_FQ_OFFLOAD_HORIZON, (u32)offload_horizon) ||
1255 nla_put_u8(skb, TCA_FQ_HORIZON_DROP,
1256 READ_ONCE(q->horizon_drop)))
1257 goto nla_put_failure;
1258
1259 fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
1260 if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
1261 goto nla_put_failure;
1262
1263 weights[0] = READ_ONCE(q->band_flows[0].quantum);
1264 weights[1] = READ_ONCE(q->band_flows[1].quantum);
1265 weights[2] = READ_ONCE(q->band_flows[2].quantum);
1266 if (nla_put(skb, TCA_FQ_WEIGHTS, sizeof(weights), &weights))
1267 goto nla_put_failure;
1268
1269 return nla_nest_end(skb, opts);
1270
1271nla_put_failure:
1272 return -1;
1273}
1274
1275static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1276{
1277 struct fq_sched_data *q = qdisc_priv(sch);
1278 struct tc_fq_qd_stats st;
1279 int i;
1280
1281 st.pad = 0;
1282
1283 sch_tree_lock(sch);
1284
1285 st.gc_flows = q->stat_gc_flows;
1286 st.highprio_packets = 0;
1287 st.fastpath_packets = q->internal.stat_fastpath_packets;
1288 st.tcp_retrans = 0;
1289 st.throttled = q->stat_throttled;
1290 st.flows_plimit = q->stat_flows_plimit;
1291 st.pkts_too_long = q->stat_pkts_too_long;
1292 st.allocation_errors = q->stat_allocation_errors;
1293 st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1294 ktime_get_ns();
1295 st.flows = q->flows;
1296 st.inactive_flows = q->inactive_flows;
1297 st.throttled_flows = q->throttled_flows;
1298 st.unthrottle_latency_ns = min_t(unsigned long,
1299 q->unthrottle_latency_ns, ~0U);
1300 st.ce_mark = q->stat_ce_mark;
1301 st.horizon_drops = q->stat_horizon_drops;
1302 st.horizon_caps = q->stat_horizon_caps;
1303 for (i = 0; i < FQ_BANDS; i++) {
1304 st.band_drops[i] = q->stat_band_drops[i];
1305 st.band_pkt_count[i] = q->band_pkt_count[i];
1306 }
1307 sch_tree_unlock(sch);
1308
1309 return gnet_stats_copy_app(d, &st, sizeof(st));
1310}
1311
1312static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1313 .id = "fq",
1314 .priv_size = sizeof(struct fq_sched_data),
1315
1316 .enqueue = fq_enqueue,
1317 .dequeue = fq_dequeue,
1318 .peek = qdisc_peek_dequeued,
1319 .init = fq_init,
1320 .reset = fq_reset,
1321 .destroy = fq_destroy,
1322 .change = fq_change,
1323 .dump = fq_dump,
1324 .dump_stats = fq_dump_stats,
1325 .owner = THIS_MODULE,
1326};
1327MODULE_ALIAS_NET_SCH("fq");
1328
1329static int __init fq_module_init(void)
1330{
1331 int ret;
1332
1333 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1334 sizeof(struct fq_flow),
1335 0, SLAB_HWCACHE_ALIGN, NULL);
1336 if (!fq_flow_cachep)
1337 return -ENOMEM;
1338
1339 ret = register_qdisc(&fq_qdisc_ops);
1340 if (ret)
1341 kmem_cache_destroy(fq_flow_cachep);
1342 return ret;
1343}
1344
1345static void __exit fq_module_exit(void)
1346{
1347 unregister_qdisc(&fq_qdisc_ops);
1348 kmem_cache_destroy(fq_flow_cachep);
1349}
1350
1351module_init(fq_module_init)
1352module_exit(fq_module_exit)
1353MODULE_AUTHOR("Eric Dumazet");
1354MODULE_LICENSE("GPL");
1355MODULE_DESCRIPTION("Fair Queue Packet Scheduler");
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 */
60struct 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
77struct fq_flow_head {
78 struct fq_flow *first;
79 struct fq_flow *last;
80};
81
82struct fq_sched_data {
83 struct fq_flow_head new_flows;
84
85 struct fq_flow_head old_flows;
86
87 struct rb_root delayed; /* for rate limited flows */
88 u64 time_next_delayed_flow;
89 unsigned long unthrottle_latency_ns;
90
91 struct fq_flow internal; /* for non classified or high prio packets */
92 u32 quantum;
93 u32 initial_quantum;
94 u32 flow_refill_delay;
95 u32 flow_max_rate; /* optional max rate per flow */
96 u32 flow_plimit; /* max packets per flow */
97 u32 orphan_mask; /* mask for orphaned skb */
98 u32 low_rate_threshold;
99 struct rb_root *fq_root;
100 u8 rate_enable;
101 u8 fq_trees_log;
102
103 u32 flows;
104 u32 inactive_flows;
105 u32 throttled_flows;
106
107 u64 stat_gc_flows;
108 u64 stat_internal_packets;
109 u64 stat_tcp_retrans;
110 u64 stat_throttled;
111 u64 stat_flows_plimit;
112 u64 stat_pkts_too_long;
113 u64 stat_allocation_errors;
114 struct qdisc_watchdog watchdog;
115};
116
117/* special value to mark a detached flow (not on old/new list) */
118static struct fq_flow detached, throttled;
119
120static void fq_flow_set_detached(struct fq_flow *f)
121{
122 f->next = &detached;
123 f->age = jiffies;
124}
125
126static bool fq_flow_is_detached(const struct fq_flow *f)
127{
128 return f->next == &detached;
129}
130
131static bool fq_flow_is_throttled(const struct fq_flow *f)
132{
133 return f->next == &throttled;
134}
135
136static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
137{
138 if (head->first)
139 head->last->next = flow;
140 else
141 head->first = flow;
142 head->last = flow;
143 flow->next = NULL;
144}
145
146static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
147{
148 rb_erase(&f->rate_node, &q->delayed);
149 q->throttled_flows--;
150 fq_flow_add_tail(&q->old_flows, f);
151}
152
153static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
154{
155 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
156
157 while (*p) {
158 struct fq_flow *aux;
159
160 parent = *p;
161 aux = rb_entry(parent, struct fq_flow, rate_node);
162 if (f->time_next_packet >= aux->time_next_packet)
163 p = &parent->rb_right;
164 else
165 p = &parent->rb_left;
166 }
167 rb_link_node(&f->rate_node, parent, p);
168 rb_insert_color(&f->rate_node, &q->delayed);
169 q->throttled_flows++;
170 q->stat_throttled++;
171
172 f->next = &throttled;
173 if (q->time_next_delayed_flow > f->time_next_packet)
174 q->time_next_delayed_flow = f->time_next_packet;
175}
176
177
178static struct kmem_cache *fq_flow_cachep __read_mostly;
179
180
181/* limit number of collected flows per round */
182#define FQ_GC_MAX 8
183#define FQ_GC_AGE (3*HZ)
184
185static bool fq_gc_candidate(const struct fq_flow *f)
186{
187 return fq_flow_is_detached(f) &&
188 time_after(jiffies, f->age + FQ_GC_AGE);
189}
190
191static void fq_gc(struct fq_sched_data *q,
192 struct rb_root *root,
193 struct sock *sk)
194{
195 struct fq_flow *f, *tofree[FQ_GC_MAX];
196 struct rb_node **p, *parent;
197 int fcnt = 0;
198
199 p = &root->rb_node;
200 parent = NULL;
201 while (*p) {
202 parent = *p;
203
204 f = rb_entry(parent, struct fq_flow, fq_node);
205 if (f->sk == sk)
206 break;
207
208 if (fq_gc_candidate(f)) {
209 tofree[fcnt++] = f;
210 if (fcnt == FQ_GC_MAX)
211 break;
212 }
213
214 if (f->sk > sk)
215 p = &parent->rb_right;
216 else
217 p = &parent->rb_left;
218 }
219
220 q->flows -= fcnt;
221 q->inactive_flows -= fcnt;
222 q->stat_gc_flows += fcnt;
223 while (fcnt) {
224 struct fq_flow *f = tofree[--fcnt];
225
226 rb_erase(&f->fq_node, root);
227 kmem_cache_free(fq_flow_cachep, f);
228 }
229}
230
231static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
232{
233 struct rb_node **p, *parent;
234 struct sock *sk = skb->sk;
235 struct rb_root *root;
236 struct fq_flow *f;
237
238 /* warning: no starvation prevention... */
239 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
240 return &q->internal;
241
242 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
243 * or a listener (SYNCOOKIE mode)
244 * 1) request sockets are not full blown,
245 * they do not contain sk_pacing_rate
246 * 2) They are not part of a 'flow' yet
247 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
248 * especially if the listener set SO_MAX_PACING_RATE
249 * 4) We pretend they are orphaned
250 */
251 if (!sk || sk_listener(sk)) {
252 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
253
254 /* By forcing low order bit to 1, we make sure to not
255 * collide with a local flow (socket pointers are word aligned)
256 */
257 sk = (struct sock *)((hash << 1) | 1UL);
258 skb_orphan(skb);
259 }
260
261 root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
262
263 if (q->flows >= (2U << q->fq_trees_log) &&
264 q->inactive_flows > q->flows/2)
265 fq_gc(q, root, sk);
266
267 p = &root->rb_node;
268 parent = NULL;
269 while (*p) {
270 parent = *p;
271
272 f = rb_entry(parent, struct fq_flow, fq_node);
273 if (f->sk == sk) {
274 /* socket might have been reallocated, so check
275 * if its sk_hash is the same.
276 * It not, we need to refill credit with
277 * initial quantum
278 */
279 if (unlikely(skb->sk &&
280 f->socket_hash != sk->sk_hash)) {
281 f->credit = q->initial_quantum;
282 f->socket_hash = sk->sk_hash;
283 if (fq_flow_is_throttled(f))
284 fq_flow_unset_throttled(q, f);
285 f->time_next_packet = 0ULL;
286 }
287 return f;
288 }
289 if (f->sk > sk)
290 p = &parent->rb_right;
291 else
292 p = &parent->rb_left;
293 }
294
295 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
296 if (unlikely(!f)) {
297 q->stat_allocation_errors++;
298 return &q->internal;
299 }
300 fq_flow_set_detached(f);
301 f->sk = sk;
302 if (skb->sk)
303 f->socket_hash = sk->sk_hash;
304 f->credit = q->initial_quantum;
305
306 rb_link_node(&f->fq_node, parent, p);
307 rb_insert_color(&f->fq_node, root);
308
309 q->flows++;
310 q->inactive_flows++;
311 return f;
312}
313
314
315/* remove one skb from head of flow queue */
316static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
317{
318 struct sk_buff *skb = flow->head;
319
320 if (skb) {
321 flow->head = skb->next;
322 skb->next = NULL;
323 flow->qlen--;
324 qdisc_qstats_backlog_dec(sch, skb);
325 sch->q.qlen--;
326 }
327 return skb;
328}
329
330/* We might add in the future detection of retransmits
331 * For the time being, just return false
332 */
333static bool skb_is_retransmit(struct sk_buff *skb)
334{
335 return false;
336}
337
338/* add skb to flow queue
339 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
340 * We special case tcp retransmits to be transmitted before other packets.
341 * We rely on fact that TCP retransmits are unlikely, so we do not waste
342 * a separate queue or a pointer.
343 * head-> [retrans pkt 1]
344 * [retrans pkt 2]
345 * [ normal pkt 1]
346 * [ normal pkt 2]
347 * [ normal pkt 3]
348 * tail-> [ normal pkt 4]
349 */
350static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
351{
352 struct sk_buff *prev, *head = flow->head;
353
354 skb->next = NULL;
355 if (!head) {
356 flow->head = skb;
357 flow->tail = skb;
358 return;
359 }
360 if (likely(!skb_is_retransmit(skb))) {
361 flow->tail->next = skb;
362 flow->tail = skb;
363 return;
364 }
365
366 /* This skb is a tcp retransmit,
367 * find the last retrans packet in the queue
368 */
369 prev = NULL;
370 while (skb_is_retransmit(head)) {
371 prev = head;
372 head = head->next;
373 if (!head)
374 break;
375 }
376 if (!prev) { /* no rtx packet in queue, become the new head */
377 skb->next = flow->head;
378 flow->head = skb;
379 } else {
380 if (prev == flow->tail)
381 flow->tail = skb;
382 else
383 skb->next = prev->next;
384 prev->next = skb;
385 }
386}
387
388static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
389 struct sk_buff **to_free)
390{
391 struct fq_sched_data *q = qdisc_priv(sch);
392 struct fq_flow *f;
393
394 if (unlikely(sch->q.qlen >= sch->limit))
395 return qdisc_drop(skb, sch, to_free);
396
397 f = fq_classify(skb, q);
398 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
399 q->stat_flows_plimit++;
400 return qdisc_drop(skb, sch, to_free);
401 }
402
403 f->qlen++;
404 if (skb_is_retransmit(skb))
405 q->stat_tcp_retrans++;
406 qdisc_qstats_backlog_inc(sch, skb);
407 if (fq_flow_is_detached(f)) {
408 struct sock *sk = skb->sk;
409
410 fq_flow_add_tail(&q->new_flows, f);
411 if (time_after(jiffies, f->age + q->flow_refill_delay))
412 f->credit = max_t(u32, f->credit, q->quantum);
413 if (sk && q->rate_enable) {
414 if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
415 SK_PACING_FQ))
416 smp_store_release(&sk->sk_pacing_status,
417 SK_PACING_FQ);
418 }
419 q->inactive_flows--;
420 }
421
422 /* Note: this overwrites f->age */
423 flow_queue_add(f, skb);
424
425 if (unlikely(f == &q->internal)) {
426 q->stat_internal_packets++;
427 }
428 sch->q.qlen++;
429
430 return NET_XMIT_SUCCESS;
431}
432
433static void fq_check_throttled(struct fq_sched_data *q, u64 now)
434{
435 unsigned long sample;
436 struct rb_node *p;
437
438 if (q->time_next_delayed_flow > now)
439 return;
440
441 /* Update unthrottle latency EWMA.
442 * This is cheap and can help diagnosing timer/latency problems.
443 */
444 sample = (unsigned long)(now - q->time_next_delayed_flow);
445 q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
446 q->unthrottle_latency_ns += sample >> 3;
447
448 q->time_next_delayed_flow = ~0ULL;
449 while ((p = rb_first(&q->delayed)) != NULL) {
450 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
451
452 if (f->time_next_packet > now) {
453 q->time_next_delayed_flow = f->time_next_packet;
454 break;
455 }
456 fq_flow_unset_throttled(q, f);
457 }
458}
459
460static struct sk_buff *fq_dequeue(struct Qdisc *sch)
461{
462 struct fq_sched_data *q = qdisc_priv(sch);
463 u64 now = ktime_get_ns();
464 struct fq_flow_head *head;
465 struct sk_buff *skb;
466 struct fq_flow *f;
467 u32 rate, plen;
468
469 skb = fq_dequeue_head(sch, &q->internal);
470 if (skb)
471 goto out;
472 fq_check_throttled(q, now);
473begin:
474 head = &q->new_flows;
475 if (!head->first) {
476 head = &q->old_flows;
477 if (!head->first) {
478 if (q->time_next_delayed_flow != ~0ULL)
479 qdisc_watchdog_schedule_ns(&q->watchdog,
480 q->time_next_delayed_flow);
481 return NULL;
482 }
483 }
484 f = head->first;
485
486 if (f->credit <= 0) {
487 f->credit += q->quantum;
488 head->first = f->next;
489 fq_flow_add_tail(&q->old_flows, f);
490 goto begin;
491 }
492
493 skb = f->head;
494 if (unlikely(skb && now < f->time_next_packet &&
495 !skb_is_tcp_pure_ack(skb))) {
496 head->first = f->next;
497 fq_flow_set_throttled(q, f);
498 goto begin;
499 }
500
501 skb = fq_dequeue_head(sch, f);
502 if (!skb) {
503 head->first = f->next;
504 /* force a pass through old_flows to prevent starvation */
505 if ((head == &q->new_flows) && q->old_flows.first) {
506 fq_flow_add_tail(&q->old_flows, f);
507 } else {
508 fq_flow_set_detached(f);
509 q->inactive_flows++;
510 }
511 goto begin;
512 }
513 prefetch(&skb->end);
514 f->credit -= qdisc_pkt_len(skb);
515
516 if (!q->rate_enable)
517 goto out;
518
519 /* Do not pace locally generated ack packets */
520 if (skb_is_tcp_pure_ack(skb))
521 goto out;
522
523 rate = q->flow_max_rate;
524 if (skb->sk)
525 rate = min(skb->sk->sk_pacing_rate, rate);
526
527 if (rate <= q->low_rate_threshold) {
528 f->credit = 0;
529 plen = qdisc_pkt_len(skb);
530 } else {
531 plen = max(qdisc_pkt_len(skb), q->quantum);
532 if (f->credit > 0)
533 goto out;
534 }
535 if (rate != ~0U) {
536 u64 len = (u64)plen * NSEC_PER_SEC;
537
538 if (likely(rate))
539 do_div(len, rate);
540 /* Since socket rate can change later,
541 * clamp the delay to 1 second.
542 * Really, providers of too big packets should be fixed !
543 */
544 if (unlikely(len > NSEC_PER_SEC)) {
545 len = NSEC_PER_SEC;
546 q->stat_pkts_too_long++;
547 }
548 /* Account for schedule/timers drifts.
549 * f->time_next_packet was set when prior packet was sent,
550 * and current time (@now) can be too late by tens of us.
551 */
552 if (f->time_next_packet)
553 len -= min(len/2, now - f->time_next_packet);
554 f->time_next_packet = now + len;
555 }
556out:
557 qdisc_bstats_update(sch, skb);
558 return skb;
559}
560
561static void fq_flow_purge(struct fq_flow *flow)
562{
563 rtnl_kfree_skbs(flow->head, flow->tail);
564 flow->head = NULL;
565 flow->qlen = 0;
566}
567
568static void fq_reset(struct Qdisc *sch)
569{
570 struct fq_sched_data *q = qdisc_priv(sch);
571 struct rb_root *root;
572 struct rb_node *p;
573 struct fq_flow *f;
574 unsigned int idx;
575
576 sch->q.qlen = 0;
577 sch->qstats.backlog = 0;
578
579 fq_flow_purge(&q->internal);
580
581 if (!q->fq_root)
582 return;
583
584 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
585 root = &q->fq_root[idx];
586 while ((p = rb_first(root)) != NULL) {
587 f = rb_entry(p, struct fq_flow, fq_node);
588 rb_erase(p, root);
589
590 fq_flow_purge(f);
591
592 kmem_cache_free(fq_flow_cachep, f);
593 }
594 }
595 q->new_flows.first = NULL;
596 q->old_flows.first = NULL;
597 q->delayed = RB_ROOT;
598 q->flows = 0;
599 q->inactive_flows = 0;
600 q->throttled_flows = 0;
601}
602
603static void fq_rehash(struct fq_sched_data *q,
604 struct rb_root *old_array, u32 old_log,
605 struct rb_root *new_array, u32 new_log)
606{
607 struct rb_node *op, **np, *parent;
608 struct rb_root *oroot, *nroot;
609 struct fq_flow *of, *nf;
610 int fcnt = 0;
611 u32 idx;
612
613 for (idx = 0; idx < (1U << old_log); idx++) {
614 oroot = &old_array[idx];
615 while ((op = rb_first(oroot)) != NULL) {
616 rb_erase(op, oroot);
617 of = rb_entry(op, struct fq_flow, fq_node);
618 if (fq_gc_candidate(of)) {
619 fcnt++;
620 kmem_cache_free(fq_flow_cachep, of);
621 continue;
622 }
623 nroot = &new_array[hash_ptr(of->sk, new_log)];
624
625 np = &nroot->rb_node;
626 parent = NULL;
627 while (*np) {
628 parent = *np;
629
630 nf = rb_entry(parent, struct fq_flow, fq_node);
631 BUG_ON(nf->sk == of->sk);
632
633 if (nf->sk > of->sk)
634 np = &parent->rb_right;
635 else
636 np = &parent->rb_left;
637 }
638
639 rb_link_node(&of->fq_node, parent, np);
640 rb_insert_color(&of->fq_node, nroot);
641 }
642 }
643 q->flows -= fcnt;
644 q->inactive_flows -= fcnt;
645 q->stat_gc_flows += fcnt;
646}
647
648static void fq_free(void *addr)
649{
650 kvfree(addr);
651}
652
653static int fq_resize(struct Qdisc *sch, u32 log)
654{
655 struct fq_sched_data *q = qdisc_priv(sch);
656 struct rb_root *array;
657 void *old_fq_root;
658 u32 idx;
659
660 if (q->fq_root && log == q->fq_trees_log)
661 return 0;
662
663 /* If XPS was setup, we can allocate memory on right NUMA node */
664 array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
665 netdev_queue_numa_node_read(sch->dev_queue));
666 if (!array)
667 return -ENOMEM;
668
669 for (idx = 0; idx < (1U << log); idx++)
670 array[idx] = RB_ROOT;
671
672 sch_tree_lock(sch);
673
674 old_fq_root = q->fq_root;
675 if (old_fq_root)
676 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
677
678 q->fq_root = array;
679 q->fq_trees_log = log;
680
681 sch_tree_unlock(sch);
682
683 fq_free(old_fq_root);
684
685 return 0;
686}
687
688static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
689 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
690 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
691 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
692 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
693 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
694 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
695 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
696 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
697 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
698 [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 },
699};
700
701static int fq_change(struct Qdisc *sch, struct nlattr *opt,
702 struct netlink_ext_ack *extack)
703{
704 struct fq_sched_data *q = qdisc_priv(sch);
705 struct nlattr *tb[TCA_FQ_MAX + 1];
706 int err, drop_count = 0;
707 unsigned drop_len = 0;
708 u32 fq_log;
709
710 if (!opt)
711 return -EINVAL;
712
713 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy, NULL);
714 if (err < 0)
715 return err;
716
717 sch_tree_lock(sch);
718
719 fq_log = q->fq_trees_log;
720
721 if (tb[TCA_FQ_BUCKETS_LOG]) {
722 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
723
724 if (nval >= 1 && nval <= ilog2(256*1024))
725 fq_log = nval;
726 else
727 err = -EINVAL;
728 }
729 if (tb[TCA_FQ_PLIMIT])
730 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
731
732 if (tb[TCA_FQ_FLOW_PLIMIT])
733 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
734
735 if (tb[TCA_FQ_QUANTUM]) {
736 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
737
738 if (quantum > 0)
739 q->quantum = quantum;
740 else
741 err = -EINVAL;
742 }
743
744 if (tb[TCA_FQ_INITIAL_QUANTUM])
745 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
746
747 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
748 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
749 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
750
751 if (tb[TCA_FQ_FLOW_MAX_RATE])
752 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
753
754 if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
755 q->low_rate_threshold =
756 nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
757
758 if (tb[TCA_FQ_RATE_ENABLE]) {
759 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
760
761 if (enable <= 1)
762 q->rate_enable = enable;
763 else
764 err = -EINVAL;
765 }
766
767 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
768 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
769
770 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
771 }
772
773 if (tb[TCA_FQ_ORPHAN_MASK])
774 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
775
776 if (!err) {
777 sch_tree_unlock(sch);
778 err = fq_resize(sch, fq_log);
779 sch_tree_lock(sch);
780 }
781 while (sch->q.qlen > sch->limit) {
782 struct sk_buff *skb = fq_dequeue(sch);
783
784 if (!skb)
785 break;
786 drop_len += qdisc_pkt_len(skb);
787 rtnl_kfree_skbs(skb, skb);
788 drop_count++;
789 }
790 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
791
792 sch_tree_unlock(sch);
793 return err;
794}
795
796static void fq_destroy(struct Qdisc *sch)
797{
798 struct fq_sched_data *q = qdisc_priv(sch);
799
800 fq_reset(sch);
801 fq_free(q->fq_root);
802 qdisc_watchdog_cancel(&q->watchdog);
803}
804
805static int fq_init(struct Qdisc *sch, struct nlattr *opt,
806 struct netlink_ext_ack *extack)
807{
808 struct fq_sched_data *q = qdisc_priv(sch);
809 int err;
810
811 sch->limit = 10000;
812 q->flow_plimit = 100;
813 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
814 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
815 q->flow_refill_delay = msecs_to_jiffies(40);
816 q->flow_max_rate = ~0U;
817 q->time_next_delayed_flow = ~0ULL;
818 q->rate_enable = 1;
819 q->new_flows.first = NULL;
820 q->old_flows.first = NULL;
821 q->delayed = RB_ROOT;
822 q->fq_root = NULL;
823 q->fq_trees_log = ilog2(1024);
824 q->orphan_mask = 1024 - 1;
825 q->low_rate_threshold = 550000 / 8;
826 qdisc_watchdog_init(&q->watchdog, sch);
827
828 if (opt)
829 err = fq_change(sch, opt, extack);
830 else
831 err = fq_resize(sch, q->fq_trees_log);
832
833 return err;
834}
835
836static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
837{
838 struct fq_sched_data *q = qdisc_priv(sch);
839 struct nlattr *opts;
840
841 opts = nla_nest_start(skb, TCA_OPTIONS);
842 if (opts == NULL)
843 goto nla_put_failure;
844
845 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
846
847 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
848 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
849 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
850 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
851 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
852 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
853 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
854 jiffies_to_usecs(q->flow_refill_delay)) ||
855 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
856 nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
857 q->low_rate_threshold) ||
858 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
859 goto nla_put_failure;
860
861 return nla_nest_end(skb, opts);
862
863nla_put_failure:
864 return -1;
865}
866
867static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
868{
869 struct fq_sched_data *q = qdisc_priv(sch);
870 struct tc_fq_qd_stats st;
871
872 sch_tree_lock(sch);
873
874 st.gc_flows = q->stat_gc_flows;
875 st.highprio_packets = q->stat_internal_packets;
876 st.tcp_retrans = q->stat_tcp_retrans;
877 st.throttled = q->stat_throttled;
878 st.flows_plimit = q->stat_flows_plimit;
879 st.pkts_too_long = q->stat_pkts_too_long;
880 st.allocation_errors = q->stat_allocation_errors;
881 st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
882 st.flows = q->flows;
883 st.inactive_flows = q->inactive_flows;
884 st.throttled_flows = q->throttled_flows;
885 st.unthrottle_latency_ns = min_t(unsigned long,
886 q->unthrottle_latency_ns, ~0U);
887 sch_tree_unlock(sch);
888
889 return gnet_stats_copy_app(d, &st, sizeof(st));
890}
891
892static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
893 .id = "fq",
894 .priv_size = sizeof(struct fq_sched_data),
895
896 .enqueue = fq_enqueue,
897 .dequeue = fq_dequeue,
898 .peek = qdisc_peek_dequeued,
899 .init = fq_init,
900 .reset = fq_reset,
901 .destroy = fq_destroy,
902 .change = fq_change,
903 .dump = fq_dump,
904 .dump_stats = fq_dump_stats,
905 .owner = THIS_MODULE,
906};
907
908static int __init fq_module_init(void)
909{
910 int ret;
911
912 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
913 sizeof(struct fq_flow),
914 0, 0, NULL);
915 if (!fq_flow_cachep)
916 return -ENOMEM;
917
918 ret = register_qdisc(&fq_qdisc_ops);
919 if (ret)
920 kmem_cache_destroy(fq_flow_cachep);
921 return ret;
922}
923
924static void __exit fq_module_exit(void)
925{
926 unregister_qdisc(&fq_qdisc_ops);
927 kmem_cache_destroy(fq_flow_cachep);
928}
929
930module_init(fq_module_init)
931module_exit(fq_module_exit)
932MODULE_AUTHOR("Eric Dumazet");
933MODULE_LICENSE("GPL");