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) 2013-2015 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
53struct fq_skb_cb {
54 u64 time_to_send;
55};
56
57static 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 */
68struct fq_flow {
69/* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
70 struct rb_root t_root;
71 struct sk_buff *head; /* list of skbs for this flow : first skb */
72 union {
73 struct sk_buff *tail; /* last skb in the list */
74 unsigned long age; /* (jiffies | 1UL) when flow was emptied, for gc */
75 };
76 struct rb_node fq_node; /* anchor in fq_root[] trees */
77 struct sock *sk;
78 u32 socket_hash; /* sk_hash */
79 int qlen; /* number of packets in flow queue */
80
81/* Second cache line, used in fq_dequeue() */
82 int credit;
83 /* 32bit hole on 64bit arches */
84
85 struct fq_flow *next; /* next pointer in RR lists */
86
87 struct rb_node rate_node; /* anchor in q->delayed tree */
88 u64 time_next_packet;
89} ____cacheline_aligned_in_smp;
90
91struct fq_flow_head {
92 struct fq_flow *first;
93 struct fq_flow *last;
94};
95
96struct fq_sched_data {
97 struct fq_flow_head new_flows;
98
99 struct fq_flow_head old_flows;
100
101 struct rb_root delayed; /* for rate limited flows */
102 u64 time_next_delayed_flow;
103 u64 ktime_cache; /* copy of last ktime_get_ns() */
104 unsigned long unthrottle_latency_ns;
105
106 struct fq_flow internal; /* for non classified or high prio packets */
107 u32 quantum;
108 u32 initial_quantum;
109 u32 flow_refill_delay;
110 u32 flow_plimit; /* max packets per flow */
111 unsigned long flow_max_rate; /* optional max rate per flow */
112 u64 ce_threshold;
113 u64 horizon; /* horizon in ns */
114 u32 orphan_mask; /* mask for orphaned skb */
115 u32 low_rate_threshold;
116 struct rb_root *fq_root;
117 u8 rate_enable;
118 u8 fq_trees_log;
119 u8 horizon_drop;
120 u32 flows;
121 u32 inactive_flows;
122 u32 throttled_flows;
123
124 u64 stat_gc_flows;
125 u64 stat_internal_packets;
126 u64 stat_throttled;
127 u64 stat_ce_mark;
128 u64 stat_horizon_drops;
129 u64 stat_horizon_caps;
130 u64 stat_flows_plimit;
131 u64 stat_pkts_too_long;
132 u64 stat_allocation_errors;
133
134 u32 timer_slack; /* hrtimer slack in ns */
135 struct qdisc_watchdog watchdog;
136};
137
138/*
139 * f->tail and f->age share the same location.
140 * We can use the low order bit to differentiate if this location points
141 * to a sk_buff or contains a jiffies value, if we force this value to be odd.
142 * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
143 */
144static void fq_flow_set_detached(struct fq_flow *f)
145{
146 f->age = jiffies | 1UL;
147}
148
149static bool fq_flow_is_detached(const struct fq_flow *f)
150{
151 return !!(f->age & 1UL);
152}
153
154/* special value to mark a throttled flow (not on old/new list) */
155static struct fq_flow throttled;
156
157static bool fq_flow_is_throttled(const struct fq_flow *f)
158{
159 return f->next == &throttled;
160}
161
162static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
163{
164 if (head->first)
165 head->last->next = flow;
166 else
167 head->first = flow;
168 head->last = flow;
169 flow->next = NULL;
170}
171
172static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
173{
174 rb_erase(&f->rate_node, &q->delayed);
175 q->throttled_flows--;
176 fq_flow_add_tail(&q->old_flows, f);
177}
178
179static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
180{
181 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
182
183 while (*p) {
184 struct fq_flow *aux;
185
186 parent = *p;
187 aux = rb_entry(parent, struct fq_flow, rate_node);
188 if (f->time_next_packet >= aux->time_next_packet)
189 p = &parent->rb_right;
190 else
191 p = &parent->rb_left;
192 }
193 rb_link_node(&f->rate_node, parent, p);
194 rb_insert_color(&f->rate_node, &q->delayed);
195 q->throttled_flows++;
196 q->stat_throttled++;
197
198 f->next = &throttled;
199 if (q->time_next_delayed_flow > f->time_next_packet)
200 q->time_next_delayed_flow = f->time_next_packet;
201}
202
203
204static struct kmem_cache *fq_flow_cachep __read_mostly;
205
206
207/* limit number of collected flows per round */
208#define FQ_GC_MAX 8
209#define FQ_GC_AGE (3*HZ)
210
211static bool fq_gc_candidate(const struct fq_flow *f)
212{
213 return fq_flow_is_detached(f) &&
214 time_after(jiffies, f->age + FQ_GC_AGE);
215}
216
217static void fq_gc(struct fq_sched_data *q,
218 struct rb_root *root,
219 struct sock *sk)
220{
221 struct rb_node **p, *parent;
222 void *tofree[FQ_GC_MAX];
223 struct fq_flow *f;
224 int i, fcnt = 0;
225
226 p = &root->rb_node;
227 parent = NULL;
228 while (*p) {
229 parent = *p;
230
231 f = rb_entry(parent, struct fq_flow, fq_node);
232 if (f->sk == sk)
233 break;
234
235 if (fq_gc_candidate(f)) {
236 tofree[fcnt++] = f;
237 if (fcnt == FQ_GC_MAX)
238 break;
239 }
240
241 if (f->sk > sk)
242 p = &parent->rb_right;
243 else
244 p = &parent->rb_left;
245 }
246
247 if (!fcnt)
248 return;
249
250 for (i = fcnt; i > 0; ) {
251 f = tofree[--i];
252 rb_erase(&f->fq_node, root);
253 }
254 q->flows -= fcnt;
255 q->inactive_flows -= fcnt;
256 q->stat_gc_flows += fcnt;
257
258 kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
259}
260
261static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
262{
263 struct rb_node **p, *parent;
264 struct sock *sk = skb->sk;
265 struct rb_root *root;
266 struct fq_flow *f;
267
268 /* warning: no starvation prevention... */
269 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
270 return &q->internal;
271
272 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
273 * or a listener (SYNCOOKIE mode)
274 * 1) request sockets are not full blown,
275 * they do not contain sk_pacing_rate
276 * 2) They are not part of a 'flow' yet
277 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
278 * especially if the listener set SO_MAX_PACING_RATE
279 * 4) We pretend they are orphaned
280 */
281 if (!sk || sk_listener(sk)) {
282 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
283
284 /* By forcing low order bit to 1, we make sure to not
285 * collide with a local flow (socket pointers are word aligned)
286 */
287 sk = (struct sock *)((hash << 1) | 1UL);
288 skb_orphan(skb);
289 } else if (sk->sk_state == TCP_CLOSE) {
290 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
291 /*
292 * Sockets in TCP_CLOSE are non connected.
293 * Typical use case is UDP sockets, they can send packets
294 * with sendto() to many different destinations.
295 * We probably could use a generic bit advertising
296 * non connected sockets, instead of sk_state == TCP_CLOSE,
297 * if we care enough.
298 */
299 sk = (struct sock *)((hash << 1) | 1UL);
300 }
301
302 root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
303
304 if (q->flows >= (2U << q->fq_trees_log) &&
305 q->inactive_flows > q->flows/2)
306 fq_gc(q, root, sk);
307
308 p = &root->rb_node;
309 parent = NULL;
310 while (*p) {
311 parent = *p;
312
313 f = rb_entry(parent, struct fq_flow, fq_node);
314 if (f->sk == sk) {
315 /* socket might have been reallocated, so check
316 * if its sk_hash is the same.
317 * It not, we need to refill credit with
318 * initial quantum
319 */
320 if (unlikely(skb->sk == sk &&
321 f->socket_hash != sk->sk_hash)) {
322 f->credit = q->initial_quantum;
323 f->socket_hash = sk->sk_hash;
324 if (q->rate_enable)
325 smp_store_release(&sk->sk_pacing_status,
326 SK_PACING_FQ);
327 if (fq_flow_is_throttled(f))
328 fq_flow_unset_throttled(q, f);
329 f->time_next_packet = 0ULL;
330 }
331 return f;
332 }
333 if (f->sk > sk)
334 p = &parent->rb_right;
335 else
336 p = &parent->rb_left;
337 }
338
339 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
340 if (unlikely(!f)) {
341 q->stat_allocation_errors++;
342 return &q->internal;
343 }
344 /* f->t_root is already zeroed after kmem_cache_zalloc() */
345
346 fq_flow_set_detached(f);
347 f->sk = sk;
348 if (skb->sk == sk) {
349 f->socket_hash = sk->sk_hash;
350 if (q->rate_enable)
351 smp_store_release(&sk->sk_pacing_status,
352 SK_PACING_FQ);
353 }
354 f->credit = q->initial_quantum;
355
356 rb_link_node(&f->fq_node, parent, p);
357 rb_insert_color(&f->fq_node, root);
358
359 q->flows++;
360 q->inactive_flows++;
361 return f;
362}
363
364static struct sk_buff *fq_peek(struct fq_flow *flow)
365{
366 struct sk_buff *skb = skb_rb_first(&flow->t_root);
367 struct sk_buff *head = flow->head;
368
369 if (!skb)
370 return head;
371
372 if (!head)
373 return skb;
374
375 if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
376 return skb;
377 return head;
378}
379
380static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
381 struct sk_buff *skb)
382{
383 if (skb == flow->head) {
384 flow->head = skb->next;
385 } else {
386 rb_erase(&skb->rbnode, &flow->t_root);
387 skb->dev = qdisc_dev(sch);
388 }
389}
390
391/* Remove one skb from flow queue.
392 * This skb must be the return value of prior fq_peek().
393 */
394static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
395 struct sk_buff *skb)
396{
397 fq_erase_head(sch, flow, skb);
398 skb_mark_not_on_list(skb);
399 flow->qlen--;
400 qdisc_qstats_backlog_dec(sch, skb);
401 sch->q.qlen--;
402}
403
404static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
405{
406 struct rb_node **p, *parent;
407 struct sk_buff *head, *aux;
408
409 head = flow->head;
410 if (!head ||
411 fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
412 if (!head)
413 flow->head = skb;
414 else
415 flow->tail->next = skb;
416 flow->tail = skb;
417 skb->next = NULL;
418 return;
419 }
420
421 p = &flow->t_root.rb_node;
422 parent = NULL;
423
424 while (*p) {
425 parent = *p;
426 aux = rb_to_skb(parent);
427 if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
428 p = &parent->rb_right;
429 else
430 p = &parent->rb_left;
431 }
432 rb_link_node(&skb->rbnode, parent, p);
433 rb_insert_color(&skb->rbnode, &flow->t_root);
434}
435
436static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
437 const struct fq_sched_data *q)
438{
439 return unlikely((s64)skb->tstamp > (s64)(q->ktime_cache + q->horizon));
440}
441
442static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
443 struct sk_buff **to_free)
444{
445 struct fq_sched_data *q = qdisc_priv(sch);
446 struct fq_flow *f;
447
448 if (unlikely(sch->q.qlen >= sch->limit))
449 return qdisc_drop(skb, sch, to_free);
450
451 if (!skb->tstamp) {
452 fq_skb_cb(skb)->time_to_send = q->ktime_cache = ktime_get_ns();
453 } else {
454 /* Check if packet timestamp is too far in the future.
455 * Try first if our cached value, to avoid ktime_get_ns()
456 * cost in most cases.
457 */
458 if (fq_packet_beyond_horizon(skb, q)) {
459 /* Refresh our cache and check another time */
460 q->ktime_cache = ktime_get_ns();
461 if (fq_packet_beyond_horizon(skb, q)) {
462 if (q->horizon_drop) {
463 q->stat_horizon_drops++;
464 return qdisc_drop(skb, sch, to_free);
465 }
466 q->stat_horizon_caps++;
467 skb->tstamp = q->ktime_cache + q->horizon;
468 }
469 }
470 fq_skb_cb(skb)->time_to_send = skb->tstamp;
471 }
472
473 f = fq_classify(skb, q);
474 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
475 q->stat_flows_plimit++;
476 return qdisc_drop(skb, sch, to_free);
477 }
478
479 f->qlen++;
480 qdisc_qstats_backlog_inc(sch, skb);
481 if (fq_flow_is_detached(f)) {
482 fq_flow_add_tail(&q->new_flows, f);
483 if (time_after(jiffies, f->age + q->flow_refill_delay))
484 f->credit = max_t(u32, f->credit, q->quantum);
485 q->inactive_flows--;
486 }
487
488 /* Note: this overwrites f->age */
489 flow_queue_add(f, skb);
490
491 if (unlikely(f == &q->internal)) {
492 q->stat_internal_packets++;
493 }
494 sch->q.qlen++;
495
496 return NET_XMIT_SUCCESS;
497}
498
499static void fq_check_throttled(struct fq_sched_data *q, u64 now)
500{
501 unsigned long sample;
502 struct rb_node *p;
503
504 if (q->time_next_delayed_flow > now)
505 return;
506
507 /* Update unthrottle latency EWMA.
508 * This is cheap and can help diagnosing timer/latency problems.
509 */
510 sample = (unsigned long)(now - q->time_next_delayed_flow);
511 q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
512 q->unthrottle_latency_ns += sample >> 3;
513
514 q->time_next_delayed_flow = ~0ULL;
515 while ((p = rb_first(&q->delayed)) != NULL) {
516 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
517
518 if (f->time_next_packet > now) {
519 q->time_next_delayed_flow = f->time_next_packet;
520 break;
521 }
522 fq_flow_unset_throttled(q, f);
523 }
524}
525
526static struct sk_buff *fq_dequeue(struct Qdisc *sch)
527{
528 struct fq_sched_data *q = qdisc_priv(sch);
529 struct fq_flow_head *head;
530 struct sk_buff *skb;
531 struct fq_flow *f;
532 unsigned long rate;
533 u32 plen;
534 u64 now;
535
536 if (!sch->q.qlen)
537 return NULL;
538
539 skb = fq_peek(&q->internal);
540 if (unlikely(skb)) {
541 fq_dequeue_skb(sch, &q->internal, skb);
542 goto out;
543 }
544
545 q->ktime_cache = now = ktime_get_ns();
546 fq_check_throttled(q, now);
547begin:
548 head = &q->new_flows;
549 if (!head->first) {
550 head = &q->old_flows;
551 if (!head->first) {
552 if (q->time_next_delayed_flow != ~0ULL)
553 qdisc_watchdog_schedule_range_ns(&q->watchdog,
554 q->time_next_delayed_flow,
555 q->timer_slack);
556 return NULL;
557 }
558 }
559 f = head->first;
560
561 if (f->credit <= 0) {
562 f->credit += q->quantum;
563 head->first = f->next;
564 fq_flow_add_tail(&q->old_flows, f);
565 goto begin;
566 }
567
568 skb = fq_peek(f);
569 if (skb) {
570 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
571 f->time_next_packet);
572
573 if (now < time_next_packet) {
574 head->first = f->next;
575 f->time_next_packet = time_next_packet;
576 fq_flow_set_throttled(q, f);
577 goto begin;
578 }
579 prefetch(&skb->end);
580 if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
581 INET_ECN_set_ce(skb);
582 q->stat_ce_mark++;
583 }
584 fq_dequeue_skb(sch, f, skb);
585 } else {
586 head->first = f->next;
587 /* force a pass through old_flows to prevent starvation */
588 if ((head == &q->new_flows) && q->old_flows.first) {
589 fq_flow_add_tail(&q->old_flows, f);
590 } else {
591 fq_flow_set_detached(f);
592 q->inactive_flows++;
593 }
594 goto begin;
595 }
596 plen = qdisc_pkt_len(skb);
597 f->credit -= plen;
598
599 if (!q->rate_enable)
600 goto out;
601
602 rate = q->flow_max_rate;
603
604 /* If EDT time was provided for this skb, we need to
605 * update f->time_next_packet only if this qdisc enforces
606 * a flow max rate.
607 */
608 if (!skb->tstamp) {
609 if (skb->sk)
610 rate = min(skb->sk->sk_pacing_rate, rate);
611
612 if (rate <= q->low_rate_threshold) {
613 f->credit = 0;
614 } else {
615 plen = max(plen, q->quantum);
616 if (f->credit > 0)
617 goto out;
618 }
619 }
620 if (rate != ~0UL) {
621 u64 len = (u64)plen * NSEC_PER_SEC;
622
623 if (likely(rate))
624 len = div64_ul(len, rate);
625 /* Since socket rate can change later,
626 * clamp the delay to 1 second.
627 * Really, providers of too big packets should be fixed !
628 */
629 if (unlikely(len > NSEC_PER_SEC)) {
630 len = NSEC_PER_SEC;
631 q->stat_pkts_too_long++;
632 }
633 /* Account for schedule/timers drifts.
634 * f->time_next_packet was set when prior packet was sent,
635 * and current time (@now) can be too late by tens of us.
636 */
637 if (f->time_next_packet)
638 len -= min(len/2, now - f->time_next_packet);
639 f->time_next_packet = now + len;
640 }
641out:
642 qdisc_bstats_update(sch, skb);
643 return skb;
644}
645
646static void fq_flow_purge(struct fq_flow *flow)
647{
648 struct rb_node *p = rb_first(&flow->t_root);
649
650 while (p) {
651 struct sk_buff *skb = rb_to_skb(p);
652
653 p = rb_next(p);
654 rb_erase(&skb->rbnode, &flow->t_root);
655 rtnl_kfree_skbs(skb, skb);
656 }
657 rtnl_kfree_skbs(flow->head, flow->tail);
658 flow->head = NULL;
659 flow->qlen = 0;
660}
661
662static void fq_reset(struct Qdisc *sch)
663{
664 struct fq_sched_data *q = qdisc_priv(sch);
665 struct rb_root *root;
666 struct rb_node *p;
667 struct fq_flow *f;
668 unsigned int idx;
669
670 sch->q.qlen = 0;
671 sch->qstats.backlog = 0;
672
673 fq_flow_purge(&q->internal);
674
675 if (!q->fq_root)
676 return;
677
678 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
679 root = &q->fq_root[idx];
680 while ((p = rb_first(root)) != NULL) {
681 f = rb_entry(p, struct fq_flow, fq_node);
682 rb_erase(p, root);
683
684 fq_flow_purge(f);
685
686 kmem_cache_free(fq_flow_cachep, f);
687 }
688 }
689 q->new_flows.first = NULL;
690 q->old_flows.first = NULL;
691 q->delayed = RB_ROOT;
692 q->flows = 0;
693 q->inactive_flows = 0;
694 q->throttled_flows = 0;
695}
696
697static void fq_rehash(struct fq_sched_data *q,
698 struct rb_root *old_array, u32 old_log,
699 struct rb_root *new_array, u32 new_log)
700{
701 struct rb_node *op, **np, *parent;
702 struct rb_root *oroot, *nroot;
703 struct fq_flow *of, *nf;
704 int fcnt = 0;
705 u32 idx;
706
707 for (idx = 0; idx < (1U << old_log); idx++) {
708 oroot = &old_array[idx];
709 while ((op = rb_first(oroot)) != NULL) {
710 rb_erase(op, oroot);
711 of = rb_entry(op, struct fq_flow, fq_node);
712 if (fq_gc_candidate(of)) {
713 fcnt++;
714 kmem_cache_free(fq_flow_cachep, of);
715 continue;
716 }
717 nroot = &new_array[hash_ptr(of->sk, new_log)];
718
719 np = &nroot->rb_node;
720 parent = NULL;
721 while (*np) {
722 parent = *np;
723
724 nf = rb_entry(parent, struct fq_flow, fq_node);
725 BUG_ON(nf->sk == of->sk);
726
727 if (nf->sk > of->sk)
728 np = &parent->rb_right;
729 else
730 np = &parent->rb_left;
731 }
732
733 rb_link_node(&of->fq_node, parent, np);
734 rb_insert_color(&of->fq_node, nroot);
735 }
736 }
737 q->flows -= fcnt;
738 q->inactive_flows -= fcnt;
739 q->stat_gc_flows += fcnt;
740}
741
742static void fq_free(void *addr)
743{
744 kvfree(addr);
745}
746
747static int fq_resize(struct Qdisc *sch, u32 log)
748{
749 struct fq_sched_data *q = qdisc_priv(sch);
750 struct rb_root *array;
751 void *old_fq_root;
752 u32 idx;
753
754 if (q->fq_root && log == q->fq_trees_log)
755 return 0;
756
757 /* If XPS was setup, we can allocate memory on right NUMA node */
758 array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
759 netdev_queue_numa_node_read(sch->dev_queue));
760 if (!array)
761 return -ENOMEM;
762
763 for (idx = 0; idx < (1U << log); idx++)
764 array[idx] = RB_ROOT;
765
766 sch_tree_lock(sch);
767
768 old_fq_root = q->fq_root;
769 if (old_fq_root)
770 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
771
772 q->fq_root = array;
773 q->fq_trees_log = log;
774
775 sch_tree_unlock(sch);
776
777 fq_free(old_fq_root);
778
779 return 0;
780}
781
782static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
783 [TCA_FQ_UNSPEC] = { .strict_start_type = TCA_FQ_TIMER_SLACK },
784
785 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
786 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
787 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
788 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
789 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
790 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
791 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
792 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
793 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
794 [TCA_FQ_ORPHAN_MASK] = { .type = NLA_U32 },
795 [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 },
796 [TCA_FQ_CE_THRESHOLD] = { .type = NLA_U32 },
797 [TCA_FQ_TIMER_SLACK] = { .type = NLA_U32 },
798 [TCA_FQ_HORIZON] = { .type = NLA_U32 },
799 [TCA_FQ_HORIZON_DROP] = { .type = NLA_U8 },
800};
801
802static int fq_change(struct Qdisc *sch, struct nlattr *opt,
803 struct netlink_ext_ack *extack)
804{
805 struct fq_sched_data *q = qdisc_priv(sch);
806 struct nlattr *tb[TCA_FQ_MAX + 1];
807 int err, drop_count = 0;
808 unsigned drop_len = 0;
809 u32 fq_log;
810
811 err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
812 NULL);
813 if (err < 0)
814 return err;
815
816 sch_tree_lock(sch);
817
818 fq_log = q->fq_trees_log;
819
820 if (tb[TCA_FQ_BUCKETS_LOG]) {
821 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
822
823 if (nval >= 1 && nval <= ilog2(256*1024))
824 fq_log = nval;
825 else
826 err = -EINVAL;
827 }
828 if (tb[TCA_FQ_PLIMIT])
829 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
830
831 if (tb[TCA_FQ_FLOW_PLIMIT])
832 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
833
834 if (tb[TCA_FQ_QUANTUM]) {
835 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
836
837 if (quantum > 0 && quantum <= (1 << 20)) {
838 q->quantum = quantum;
839 } else {
840 NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
841 err = -EINVAL;
842 }
843 }
844
845 if (tb[TCA_FQ_INITIAL_QUANTUM])
846 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
847
848 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
849 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
850 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
851
852 if (tb[TCA_FQ_FLOW_MAX_RATE]) {
853 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
854
855 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
856 }
857 if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
858 q->low_rate_threshold =
859 nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
860
861 if (tb[TCA_FQ_RATE_ENABLE]) {
862 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
863
864 if (enable <= 1)
865 q->rate_enable = enable;
866 else
867 err = -EINVAL;
868 }
869
870 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
871 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
872
873 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
874 }
875
876 if (tb[TCA_FQ_ORPHAN_MASK])
877 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
878
879 if (tb[TCA_FQ_CE_THRESHOLD])
880 q->ce_threshold = (u64)NSEC_PER_USEC *
881 nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
882
883 if (tb[TCA_FQ_TIMER_SLACK])
884 q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]);
885
886 if (tb[TCA_FQ_HORIZON])
887 q->horizon = (u64)NSEC_PER_USEC *
888 nla_get_u32(tb[TCA_FQ_HORIZON]);
889
890 if (tb[TCA_FQ_HORIZON_DROP])
891 q->horizon_drop = nla_get_u8(tb[TCA_FQ_HORIZON_DROP]);
892
893 if (!err) {
894
895 sch_tree_unlock(sch);
896 err = fq_resize(sch, fq_log);
897 sch_tree_lock(sch);
898 }
899 while (sch->q.qlen > sch->limit) {
900 struct sk_buff *skb = fq_dequeue(sch);
901
902 if (!skb)
903 break;
904 drop_len += qdisc_pkt_len(skb);
905 rtnl_kfree_skbs(skb, skb);
906 drop_count++;
907 }
908 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
909
910 sch_tree_unlock(sch);
911 return err;
912}
913
914static void fq_destroy(struct Qdisc *sch)
915{
916 struct fq_sched_data *q = qdisc_priv(sch);
917
918 fq_reset(sch);
919 fq_free(q->fq_root);
920 qdisc_watchdog_cancel(&q->watchdog);
921}
922
923static int fq_init(struct Qdisc *sch, struct nlattr *opt,
924 struct netlink_ext_ack *extack)
925{
926 struct fq_sched_data *q = qdisc_priv(sch);
927 int err;
928
929 sch->limit = 10000;
930 q->flow_plimit = 100;
931 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
932 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
933 q->flow_refill_delay = msecs_to_jiffies(40);
934 q->flow_max_rate = ~0UL;
935 q->time_next_delayed_flow = ~0ULL;
936 q->rate_enable = 1;
937 q->new_flows.first = NULL;
938 q->old_flows.first = NULL;
939 q->delayed = RB_ROOT;
940 q->fq_root = NULL;
941 q->fq_trees_log = ilog2(1024);
942 q->orphan_mask = 1024 - 1;
943 q->low_rate_threshold = 550000 / 8;
944
945 q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
946
947 q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
948 q->horizon_drop = 1; /* by default, drop packets beyond horizon */
949
950 /* Default ce_threshold of 4294 seconds */
951 q->ce_threshold = (u64)NSEC_PER_USEC * ~0U;
952
953 qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
954
955 if (opt)
956 err = fq_change(sch, opt, extack);
957 else
958 err = fq_resize(sch, q->fq_trees_log);
959
960 return err;
961}
962
963static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
964{
965 struct fq_sched_data *q = qdisc_priv(sch);
966 u64 ce_threshold = q->ce_threshold;
967 u64 horizon = q->horizon;
968 struct nlattr *opts;
969
970 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
971 if (opts == NULL)
972 goto nla_put_failure;
973
974 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
975
976 do_div(ce_threshold, NSEC_PER_USEC);
977 do_div(horizon, NSEC_PER_USEC);
978
979 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
980 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
981 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
982 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
983 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
984 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
985 min_t(unsigned long, q->flow_max_rate, ~0U)) ||
986 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
987 jiffies_to_usecs(q->flow_refill_delay)) ||
988 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
989 nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
990 q->low_rate_threshold) ||
991 nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
992 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) ||
993 nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack) ||
994 nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
995 nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop))
996 goto nla_put_failure;
997
998 return nla_nest_end(skb, opts);
999
1000nla_put_failure:
1001 return -1;
1002}
1003
1004static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1005{
1006 struct fq_sched_data *q = qdisc_priv(sch);
1007 struct tc_fq_qd_stats st;
1008
1009 sch_tree_lock(sch);
1010
1011 st.gc_flows = q->stat_gc_flows;
1012 st.highprio_packets = q->stat_internal_packets;
1013 st.tcp_retrans = 0;
1014 st.throttled = q->stat_throttled;
1015 st.flows_plimit = q->stat_flows_plimit;
1016 st.pkts_too_long = q->stat_pkts_too_long;
1017 st.allocation_errors = q->stat_allocation_errors;
1018 st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1019 ktime_get_ns();
1020 st.flows = q->flows;
1021 st.inactive_flows = q->inactive_flows;
1022 st.throttled_flows = q->throttled_flows;
1023 st.unthrottle_latency_ns = min_t(unsigned long,
1024 q->unthrottle_latency_ns, ~0U);
1025 st.ce_mark = q->stat_ce_mark;
1026 st.horizon_drops = q->stat_horizon_drops;
1027 st.horizon_caps = q->stat_horizon_caps;
1028 sch_tree_unlock(sch);
1029
1030 return gnet_stats_copy_app(d, &st, sizeof(st));
1031}
1032
1033static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1034 .id = "fq",
1035 .priv_size = sizeof(struct fq_sched_data),
1036
1037 .enqueue = fq_enqueue,
1038 .dequeue = fq_dequeue,
1039 .peek = qdisc_peek_dequeued,
1040 .init = fq_init,
1041 .reset = fq_reset,
1042 .destroy = fq_destroy,
1043 .change = fq_change,
1044 .dump = fq_dump,
1045 .dump_stats = fq_dump_stats,
1046 .owner = THIS_MODULE,
1047};
1048
1049static int __init fq_module_init(void)
1050{
1051 int ret;
1052
1053 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1054 sizeof(struct fq_flow),
1055 0, 0, NULL);
1056 if (!fq_flow_cachep)
1057 return -ENOMEM;
1058
1059 ret = register_qdisc(&fq_qdisc_ops);
1060 if (ret)
1061 kmem_cache_destroy(fq_flow_cachep);
1062 return ret;
1063}
1064
1065static void __exit fq_module_exit(void)
1066{
1067 unregister_qdisc(&fq_qdisc_ops);
1068 kmem_cache_destroy(fq_flow_cachep);
1069}
1070
1071module_init(fq_module_init)
1072module_exit(fq_module_exit)
1073MODULE_AUTHOR("Eric Dumazet");
1074MODULE_LICENSE("GPL");
1075MODULE_DESCRIPTION("Fair Queue Packet Scheduler");
1/*
2 * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
3 *
4 * Copyright (C) 2013 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 localy 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
56/*
57 * Per flow structure, dynamically allocated
58 */
59struct fq_flow {
60 struct sk_buff *head; /* list of skbs for this flow : first skb */
61 union {
62 struct sk_buff *tail; /* last skb in the list */
63 unsigned long age; /* jiffies when flow was emptied, for gc */
64 };
65 struct rb_node fq_node; /* anchor in fq_root[] trees */
66 struct sock *sk;
67 int qlen; /* number of packets in flow queue */
68 int credit;
69 u32 socket_hash; /* sk_hash */
70 struct fq_flow *next; /* next pointer in RR lists, or &detached */
71
72 struct rb_node rate_node; /* anchor in q->delayed tree */
73 u64 time_next_packet;
74};
75
76struct fq_flow_head {
77 struct fq_flow *first;
78 struct fq_flow *last;
79};
80
81struct fq_sched_data {
82 struct fq_flow_head new_flows;
83
84 struct fq_flow_head old_flows;
85
86 struct rb_root delayed; /* for rate limited flows */
87 u64 time_next_delayed_flow;
88
89 struct fq_flow internal; /* for non classified or high prio packets */
90 u32 quantum;
91 u32 initial_quantum;
92 u32 flow_refill_delay;
93 u32 flow_max_rate; /* optional max rate per flow */
94 u32 flow_plimit; /* max packets per flow */
95 struct rb_root *fq_root;
96 u8 rate_enable;
97 u8 fq_trees_log;
98
99 u32 flows;
100 u32 inactive_flows;
101 u32 throttled_flows;
102
103 u64 stat_gc_flows;
104 u64 stat_internal_packets;
105 u64 stat_tcp_retrans;
106 u64 stat_throttled;
107 u64 stat_flows_plimit;
108 u64 stat_pkts_too_long;
109 u64 stat_allocation_errors;
110 struct qdisc_watchdog watchdog;
111};
112
113/* special value to mark a detached flow (not on old/new list) */
114static struct fq_flow detached, throttled;
115
116static void fq_flow_set_detached(struct fq_flow *f)
117{
118 f->next = &detached;
119 f->age = jiffies;
120}
121
122static bool fq_flow_is_detached(const struct fq_flow *f)
123{
124 return f->next == &detached;
125}
126
127static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
128{
129 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
130
131 while (*p) {
132 struct fq_flow *aux;
133
134 parent = *p;
135 aux = container_of(parent, struct fq_flow, rate_node);
136 if (f->time_next_packet >= aux->time_next_packet)
137 p = &parent->rb_right;
138 else
139 p = &parent->rb_left;
140 }
141 rb_link_node(&f->rate_node, parent, p);
142 rb_insert_color(&f->rate_node, &q->delayed);
143 q->throttled_flows++;
144 q->stat_throttled++;
145
146 f->next = &throttled;
147 if (q->time_next_delayed_flow > f->time_next_packet)
148 q->time_next_delayed_flow = f->time_next_packet;
149}
150
151
152static struct kmem_cache *fq_flow_cachep __read_mostly;
153
154static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
155{
156 if (head->first)
157 head->last->next = flow;
158 else
159 head->first = flow;
160 head->last = flow;
161 flow->next = NULL;
162}
163
164/* limit number of collected flows per round */
165#define FQ_GC_MAX 8
166#define FQ_GC_AGE (3*HZ)
167
168static bool fq_gc_candidate(const struct fq_flow *f)
169{
170 return fq_flow_is_detached(f) &&
171 time_after(jiffies, f->age + FQ_GC_AGE);
172}
173
174static void fq_gc(struct fq_sched_data *q,
175 struct rb_root *root,
176 struct sock *sk)
177{
178 struct fq_flow *f, *tofree[FQ_GC_MAX];
179 struct rb_node **p, *parent;
180 int fcnt = 0;
181
182 p = &root->rb_node;
183 parent = NULL;
184 while (*p) {
185 parent = *p;
186
187 f = container_of(parent, struct fq_flow, fq_node);
188 if (f->sk == sk)
189 break;
190
191 if (fq_gc_candidate(f)) {
192 tofree[fcnt++] = f;
193 if (fcnt == FQ_GC_MAX)
194 break;
195 }
196
197 if (f->sk > sk)
198 p = &parent->rb_right;
199 else
200 p = &parent->rb_left;
201 }
202
203 q->flows -= fcnt;
204 q->inactive_flows -= fcnt;
205 q->stat_gc_flows += fcnt;
206 while (fcnt) {
207 struct fq_flow *f = tofree[--fcnt];
208
209 rb_erase(&f->fq_node, root);
210 kmem_cache_free(fq_flow_cachep, f);
211 }
212}
213
214static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
215{
216 struct rb_node **p, *parent;
217 struct sock *sk = skb->sk;
218 struct rb_root *root;
219 struct fq_flow *f;
220
221 /* warning: no starvation prevention... */
222 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
223 return &q->internal;
224
225 if (unlikely(!sk)) {
226 /* By forcing low order bit to 1, we make sure to not
227 * collide with a local flow (socket pointers are word aligned)
228 */
229 sk = (struct sock *)(skb_get_hash(skb) | 1L);
230 }
231
232 root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
233
234 if (q->flows >= (2U << q->fq_trees_log) &&
235 q->inactive_flows > q->flows/2)
236 fq_gc(q, root, sk);
237
238 p = &root->rb_node;
239 parent = NULL;
240 while (*p) {
241 parent = *p;
242
243 f = container_of(parent, struct fq_flow, fq_node);
244 if (f->sk == sk) {
245 /* socket might have been reallocated, so check
246 * if its sk_hash is the same.
247 * It not, we need to refill credit with
248 * initial quantum
249 */
250 if (unlikely(skb->sk &&
251 f->socket_hash != sk->sk_hash)) {
252 f->credit = q->initial_quantum;
253 f->socket_hash = sk->sk_hash;
254 f->time_next_packet = 0ULL;
255 }
256 return f;
257 }
258 if (f->sk > sk)
259 p = &parent->rb_right;
260 else
261 p = &parent->rb_left;
262 }
263
264 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
265 if (unlikely(!f)) {
266 q->stat_allocation_errors++;
267 return &q->internal;
268 }
269 fq_flow_set_detached(f);
270 f->sk = sk;
271 if (skb->sk)
272 f->socket_hash = sk->sk_hash;
273 f->credit = q->initial_quantum;
274
275 rb_link_node(&f->fq_node, parent, p);
276 rb_insert_color(&f->fq_node, root);
277
278 q->flows++;
279 q->inactive_flows++;
280 return f;
281}
282
283
284/* remove one skb from head of flow queue */
285static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
286{
287 struct sk_buff *skb = flow->head;
288
289 if (skb) {
290 flow->head = skb->next;
291 skb->next = NULL;
292 flow->qlen--;
293 sch->qstats.backlog -= qdisc_pkt_len(skb);
294 sch->q.qlen--;
295 }
296 return skb;
297}
298
299/* We might add in the future detection of retransmits
300 * For the time being, just return false
301 */
302static bool skb_is_retransmit(struct sk_buff *skb)
303{
304 return false;
305}
306
307/* add skb to flow queue
308 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
309 * We special case tcp retransmits to be transmitted before other packets.
310 * We rely on fact that TCP retransmits are unlikely, so we do not waste
311 * a separate queue or a pointer.
312 * head-> [retrans pkt 1]
313 * [retrans pkt 2]
314 * [ normal pkt 1]
315 * [ normal pkt 2]
316 * [ normal pkt 3]
317 * tail-> [ normal pkt 4]
318 */
319static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
320{
321 struct sk_buff *prev, *head = flow->head;
322
323 skb->next = NULL;
324 if (!head) {
325 flow->head = skb;
326 flow->tail = skb;
327 return;
328 }
329 if (likely(!skb_is_retransmit(skb))) {
330 flow->tail->next = skb;
331 flow->tail = skb;
332 return;
333 }
334
335 /* This skb is a tcp retransmit,
336 * find the last retrans packet in the queue
337 */
338 prev = NULL;
339 while (skb_is_retransmit(head)) {
340 prev = head;
341 head = head->next;
342 if (!head)
343 break;
344 }
345 if (!prev) { /* no rtx packet in queue, become the new head */
346 skb->next = flow->head;
347 flow->head = skb;
348 } else {
349 if (prev == flow->tail)
350 flow->tail = skb;
351 else
352 skb->next = prev->next;
353 prev->next = skb;
354 }
355}
356
357static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
358{
359 struct fq_sched_data *q = qdisc_priv(sch);
360 struct fq_flow *f;
361
362 if (unlikely(sch->q.qlen >= sch->limit))
363 return qdisc_drop(skb, sch);
364
365 f = fq_classify(skb, q);
366 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
367 q->stat_flows_plimit++;
368 return qdisc_drop(skb, sch);
369 }
370
371 f->qlen++;
372 if (skb_is_retransmit(skb))
373 q->stat_tcp_retrans++;
374 sch->qstats.backlog += qdisc_pkt_len(skb);
375 if (fq_flow_is_detached(f)) {
376 fq_flow_add_tail(&q->new_flows, f);
377 if (time_after(jiffies, f->age + q->flow_refill_delay))
378 f->credit = max_t(u32, f->credit, q->quantum);
379 q->inactive_flows--;
380 qdisc_unthrottled(sch);
381 }
382
383 /* Note: this overwrites f->age */
384 flow_queue_add(f, skb);
385
386 if (unlikely(f == &q->internal)) {
387 q->stat_internal_packets++;
388 qdisc_unthrottled(sch);
389 }
390 sch->q.qlen++;
391
392 return NET_XMIT_SUCCESS;
393}
394
395static void fq_check_throttled(struct fq_sched_data *q, u64 now)
396{
397 struct rb_node *p;
398
399 if (q->time_next_delayed_flow > now)
400 return;
401
402 q->time_next_delayed_flow = ~0ULL;
403 while ((p = rb_first(&q->delayed)) != NULL) {
404 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
405
406 if (f->time_next_packet > now) {
407 q->time_next_delayed_flow = f->time_next_packet;
408 break;
409 }
410 rb_erase(p, &q->delayed);
411 q->throttled_flows--;
412 fq_flow_add_tail(&q->old_flows, f);
413 }
414}
415
416static struct sk_buff *fq_dequeue(struct Qdisc *sch)
417{
418 struct fq_sched_data *q = qdisc_priv(sch);
419 u64 now = ktime_to_ns(ktime_get());
420 struct fq_flow_head *head;
421 struct sk_buff *skb;
422 struct fq_flow *f;
423 u32 rate;
424
425 skb = fq_dequeue_head(sch, &q->internal);
426 if (skb)
427 goto out;
428 fq_check_throttled(q, now);
429begin:
430 head = &q->new_flows;
431 if (!head->first) {
432 head = &q->old_flows;
433 if (!head->first) {
434 if (q->time_next_delayed_flow != ~0ULL)
435 qdisc_watchdog_schedule_ns(&q->watchdog,
436 q->time_next_delayed_flow);
437 return NULL;
438 }
439 }
440 f = head->first;
441
442 if (f->credit <= 0) {
443 f->credit += q->quantum;
444 head->first = f->next;
445 fq_flow_add_tail(&q->old_flows, f);
446 goto begin;
447 }
448
449 if (unlikely(f->head && now < f->time_next_packet)) {
450 head->first = f->next;
451 fq_flow_set_throttled(q, f);
452 goto begin;
453 }
454
455 skb = fq_dequeue_head(sch, f);
456 if (!skb) {
457 head->first = f->next;
458 /* force a pass through old_flows to prevent starvation */
459 if ((head == &q->new_flows) && q->old_flows.first) {
460 fq_flow_add_tail(&q->old_flows, f);
461 } else {
462 fq_flow_set_detached(f);
463 q->inactive_flows++;
464 }
465 goto begin;
466 }
467 prefetch(&skb->end);
468 f->time_next_packet = now;
469 f->credit -= qdisc_pkt_len(skb);
470
471 if (f->credit > 0 || !q->rate_enable)
472 goto out;
473
474 rate = q->flow_max_rate;
475 if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT)
476 rate = min(skb->sk->sk_pacing_rate, rate);
477
478 if (rate != ~0U) {
479 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
480 u64 len = (u64)plen * NSEC_PER_SEC;
481
482 if (likely(rate))
483 do_div(len, rate);
484 /* Since socket rate can change later,
485 * clamp the delay to 125 ms.
486 * TODO: maybe segment the too big skb, as in commit
487 * e43ac79a4bc ("sch_tbf: segment too big GSO packets")
488 */
489 if (unlikely(len > 125 * NSEC_PER_MSEC)) {
490 len = 125 * NSEC_PER_MSEC;
491 q->stat_pkts_too_long++;
492 }
493
494 f->time_next_packet = now + len;
495 }
496out:
497 qdisc_bstats_update(sch, skb);
498 qdisc_unthrottled(sch);
499 return skb;
500}
501
502static void fq_reset(struct Qdisc *sch)
503{
504 struct fq_sched_data *q = qdisc_priv(sch);
505 struct rb_root *root;
506 struct sk_buff *skb;
507 struct rb_node *p;
508 struct fq_flow *f;
509 unsigned int idx;
510
511 while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
512 kfree_skb(skb);
513
514 if (!q->fq_root)
515 return;
516
517 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
518 root = &q->fq_root[idx];
519 while ((p = rb_first(root)) != NULL) {
520 f = container_of(p, struct fq_flow, fq_node);
521 rb_erase(p, root);
522
523 while ((skb = fq_dequeue_head(sch, f)) != NULL)
524 kfree_skb(skb);
525
526 kmem_cache_free(fq_flow_cachep, f);
527 }
528 }
529 q->new_flows.first = NULL;
530 q->old_flows.first = NULL;
531 q->delayed = RB_ROOT;
532 q->flows = 0;
533 q->inactive_flows = 0;
534 q->throttled_flows = 0;
535}
536
537static void fq_rehash(struct fq_sched_data *q,
538 struct rb_root *old_array, u32 old_log,
539 struct rb_root *new_array, u32 new_log)
540{
541 struct rb_node *op, **np, *parent;
542 struct rb_root *oroot, *nroot;
543 struct fq_flow *of, *nf;
544 int fcnt = 0;
545 u32 idx;
546
547 for (idx = 0; idx < (1U << old_log); idx++) {
548 oroot = &old_array[idx];
549 while ((op = rb_first(oroot)) != NULL) {
550 rb_erase(op, oroot);
551 of = container_of(op, struct fq_flow, fq_node);
552 if (fq_gc_candidate(of)) {
553 fcnt++;
554 kmem_cache_free(fq_flow_cachep, of);
555 continue;
556 }
557 nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
558
559 np = &nroot->rb_node;
560 parent = NULL;
561 while (*np) {
562 parent = *np;
563
564 nf = container_of(parent, struct fq_flow, fq_node);
565 BUG_ON(nf->sk == of->sk);
566
567 if (nf->sk > of->sk)
568 np = &parent->rb_right;
569 else
570 np = &parent->rb_left;
571 }
572
573 rb_link_node(&of->fq_node, parent, np);
574 rb_insert_color(&of->fq_node, nroot);
575 }
576 }
577 q->flows -= fcnt;
578 q->inactive_flows -= fcnt;
579 q->stat_gc_flows += fcnt;
580}
581
582static void *fq_alloc_node(size_t sz, int node)
583{
584 void *ptr;
585
586 ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
587 if (!ptr)
588 ptr = vmalloc_node(sz, node);
589 return ptr;
590}
591
592static void fq_free(void *addr)
593{
594 if (addr && is_vmalloc_addr(addr))
595 vfree(addr);
596 else
597 kfree(addr);
598}
599
600static int fq_resize(struct Qdisc *sch, u32 log)
601{
602 struct fq_sched_data *q = qdisc_priv(sch);
603 struct rb_root *array;
604 void *old_fq_root;
605 u32 idx;
606
607 if (q->fq_root && log == q->fq_trees_log)
608 return 0;
609
610 /* If XPS was setup, we can allocate memory on right NUMA node */
611 array = fq_alloc_node(sizeof(struct rb_root) << log,
612 netdev_queue_numa_node_read(sch->dev_queue));
613 if (!array)
614 return -ENOMEM;
615
616 for (idx = 0; idx < (1U << log); idx++)
617 array[idx] = RB_ROOT;
618
619 sch_tree_lock(sch);
620
621 old_fq_root = q->fq_root;
622 if (old_fq_root)
623 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
624
625 q->fq_root = array;
626 q->fq_trees_log = log;
627
628 sch_tree_unlock(sch);
629
630 fq_free(old_fq_root);
631
632 return 0;
633}
634
635static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
636 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
637 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
638 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
639 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
640 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
641 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
642 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
643 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
644 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
645};
646
647static int fq_change(struct Qdisc *sch, struct nlattr *opt)
648{
649 struct fq_sched_data *q = qdisc_priv(sch);
650 struct nlattr *tb[TCA_FQ_MAX + 1];
651 int err, drop_count = 0;
652 u32 fq_log;
653
654 if (!opt)
655 return -EINVAL;
656
657 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
658 if (err < 0)
659 return err;
660
661 sch_tree_lock(sch);
662
663 fq_log = q->fq_trees_log;
664
665 if (tb[TCA_FQ_BUCKETS_LOG]) {
666 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
667
668 if (nval >= 1 && nval <= ilog2(256*1024))
669 fq_log = nval;
670 else
671 err = -EINVAL;
672 }
673 if (tb[TCA_FQ_PLIMIT])
674 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
675
676 if (tb[TCA_FQ_FLOW_PLIMIT])
677 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
678
679 if (tb[TCA_FQ_QUANTUM])
680 q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
681
682 if (tb[TCA_FQ_INITIAL_QUANTUM])
683 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
684
685 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
686 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
687 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
688
689 if (tb[TCA_FQ_FLOW_MAX_RATE])
690 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
691
692 if (tb[TCA_FQ_RATE_ENABLE]) {
693 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
694
695 if (enable <= 1)
696 q->rate_enable = enable;
697 else
698 err = -EINVAL;
699 }
700
701 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
702 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
703
704 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
705 }
706
707 if (!err) {
708 sch_tree_unlock(sch);
709 err = fq_resize(sch, fq_log);
710 sch_tree_lock(sch);
711 }
712 while (sch->q.qlen > sch->limit) {
713 struct sk_buff *skb = fq_dequeue(sch);
714
715 if (!skb)
716 break;
717 kfree_skb(skb);
718 drop_count++;
719 }
720 qdisc_tree_decrease_qlen(sch, drop_count);
721
722 sch_tree_unlock(sch);
723 return err;
724}
725
726static void fq_destroy(struct Qdisc *sch)
727{
728 struct fq_sched_data *q = qdisc_priv(sch);
729
730 fq_reset(sch);
731 fq_free(q->fq_root);
732 qdisc_watchdog_cancel(&q->watchdog);
733}
734
735static int fq_init(struct Qdisc *sch, struct nlattr *opt)
736{
737 struct fq_sched_data *q = qdisc_priv(sch);
738 int err;
739
740 sch->limit = 10000;
741 q->flow_plimit = 100;
742 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
743 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
744 q->flow_refill_delay = msecs_to_jiffies(40);
745 q->flow_max_rate = ~0U;
746 q->rate_enable = 1;
747 q->new_flows.first = NULL;
748 q->old_flows.first = NULL;
749 q->delayed = RB_ROOT;
750 q->fq_root = NULL;
751 q->fq_trees_log = ilog2(1024);
752 qdisc_watchdog_init(&q->watchdog, sch);
753
754 if (opt)
755 err = fq_change(sch, opt);
756 else
757 err = fq_resize(sch, q->fq_trees_log);
758
759 return err;
760}
761
762static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
763{
764 struct fq_sched_data *q = qdisc_priv(sch);
765 struct nlattr *opts;
766
767 opts = nla_nest_start(skb, TCA_OPTIONS);
768 if (opts == NULL)
769 goto nla_put_failure;
770
771 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
772
773 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
774 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
775 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
776 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
777 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
778 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
779 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
780 jiffies_to_usecs(q->flow_refill_delay)) ||
781 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
782 goto nla_put_failure;
783
784 return nla_nest_end(skb, opts);
785
786nla_put_failure:
787 return -1;
788}
789
790static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
791{
792 struct fq_sched_data *q = qdisc_priv(sch);
793 u64 now = ktime_to_ns(ktime_get());
794 struct tc_fq_qd_stats st = {
795 .gc_flows = q->stat_gc_flows,
796 .highprio_packets = q->stat_internal_packets,
797 .tcp_retrans = q->stat_tcp_retrans,
798 .throttled = q->stat_throttled,
799 .flows_plimit = q->stat_flows_plimit,
800 .pkts_too_long = q->stat_pkts_too_long,
801 .allocation_errors = q->stat_allocation_errors,
802 .flows = q->flows,
803 .inactive_flows = q->inactive_flows,
804 .throttled_flows = q->throttled_flows,
805 .time_next_delayed_flow = q->time_next_delayed_flow - now,
806 };
807
808 return gnet_stats_copy_app(d, &st, sizeof(st));
809}
810
811static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
812 .id = "fq",
813 .priv_size = sizeof(struct fq_sched_data),
814
815 .enqueue = fq_enqueue,
816 .dequeue = fq_dequeue,
817 .peek = qdisc_peek_dequeued,
818 .init = fq_init,
819 .reset = fq_reset,
820 .destroy = fq_destroy,
821 .change = fq_change,
822 .dump = fq_dump,
823 .dump_stats = fq_dump_stats,
824 .owner = THIS_MODULE,
825};
826
827static int __init fq_module_init(void)
828{
829 int ret;
830
831 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
832 sizeof(struct fq_flow),
833 0, 0, NULL);
834 if (!fq_flow_cachep)
835 return -ENOMEM;
836
837 ret = register_qdisc(&fq_qdisc_ops);
838 if (ret)
839 kmem_cache_destroy(fq_flow_cachep);
840 return ret;
841}
842
843static void __exit fq_module_exit(void)
844{
845 unregister_qdisc(&fq_qdisc_ops);
846 kmem_cache_destroy(fq_flow_cachep);
847}
848
849module_init(fq_module_init)
850module_exit(fq_module_exit)
851MODULE_AUTHOR("Eric Dumazet");
852MODULE_LICENSE("GPL");