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