Loading...
1/*
2 * net/sched/sch_sfq.c Stochastic Fairness Queueing discipline.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 */
11
12#include <linux/module.h>
13#include <linux/types.h>
14#include <linux/kernel.h>
15#include <linux/jiffies.h>
16#include <linux/string.h>
17#include <linux/in.h>
18#include <linux/errno.h>
19#include <linux/init.h>
20#include <linux/skbuff.h>
21#include <linux/jhash.h>
22#include <linux/slab.h>
23#include <linux/vmalloc.h>
24#include <net/netlink.h>
25#include <net/pkt_sched.h>
26#include <net/flow_keys.h>
27#include <net/red.h>
28
29
30/* Stochastic Fairness Queuing algorithm.
31 =======================================
32
33 Source:
34 Paul E. McKenney "Stochastic Fairness Queuing",
35 IEEE INFOCOMM'90 Proceedings, San Francisco, 1990.
36
37 Paul E. McKenney "Stochastic Fairness Queuing",
38 "Interworking: Research and Experience", v.2, 1991, p.113-131.
39
40
41 See also:
42 M. Shreedhar and George Varghese "Efficient Fair
43 Queuing using Deficit Round Robin", Proc. SIGCOMM 95.
44
45
46 This is not the thing that is usually called (W)FQ nowadays.
47 It does not use any timestamp mechanism, but instead
48 processes queues in round-robin order.
49
50 ADVANTAGE:
51
52 - It is very cheap. Both CPU and memory requirements are minimal.
53
54 DRAWBACKS:
55
56 - "Stochastic" -> It is not 100% fair.
57 When hash collisions occur, several flows are considered as one.
58
59 - "Round-robin" -> It introduces larger delays than virtual clock
60 based schemes, and should not be used for isolating interactive
61 traffic from non-interactive. It means, that this scheduler
62 should be used as leaf of CBQ or P3, which put interactive traffic
63 to higher priority band.
64
65 We still need true WFQ for top level CSZ, but using WFQ
66 for the best effort traffic is absolutely pointless:
67 SFQ is superior for this purpose.
68
69 IMPLEMENTATION:
70 This implementation limits :
71 - maximal queue length per flow to 127 packets.
72 - max mtu to 2^18-1;
73 - max 65408 flows,
74 - number of hash buckets to 65536.
75
76 It is easy to increase these values, but not in flight. */
77
78#define SFQ_MAX_DEPTH 127 /* max number of packets per flow */
79#define SFQ_DEFAULT_FLOWS 128
80#define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */
81#define SFQ_EMPTY_SLOT 0xffff
82#define SFQ_DEFAULT_HASH_DIVISOR 1024
83
84/* We use 16 bits to store allot, and want to handle packets up to 64K
85 * Scale allot by 8 (1<<3) so that no overflow occurs.
86 */
87#define SFQ_ALLOT_SHIFT 3
88#define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT)
89
90/* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */
91typedef u16 sfq_index;
92
93/*
94 * We dont use pointers to save space.
95 * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array
96 * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH]
97 * are 'pointers' to dep[] array
98 */
99struct sfq_head {
100 sfq_index next;
101 sfq_index prev;
102};
103
104struct sfq_slot {
105 struct sk_buff *skblist_next;
106 struct sk_buff *skblist_prev;
107 sfq_index qlen; /* number of skbs in skblist */
108 sfq_index next; /* next slot in sfq RR chain */
109 struct sfq_head dep; /* anchor in dep[] chains */
110 unsigned short hash; /* hash value (index in ht[]) */
111 short allot; /* credit for this slot */
112
113 unsigned int backlog;
114 struct red_vars vars;
115};
116
117struct sfq_sched_data {
118/* frequently used fields */
119 int limit; /* limit of total number of packets in this qdisc */
120 unsigned int divisor; /* number of slots in hash table */
121 u8 headdrop;
122 u8 maxdepth; /* limit of packets per flow */
123
124 u32 perturbation;
125 u8 cur_depth; /* depth of longest slot */
126 u8 flags;
127 unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */
128 struct tcf_proto *filter_list;
129 sfq_index *ht; /* Hash table ('divisor' slots) */
130 struct sfq_slot *slots; /* Flows table ('maxflows' entries) */
131
132 struct red_parms *red_parms;
133 struct tc_sfqred_stats stats;
134 struct sfq_slot *tail; /* current slot in round */
135
136 struct sfq_head dep[SFQ_MAX_DEPTH + 1];
137 /* Linked lists of slots, indexed by depth
138 * dep[0] : list of unused flows
139 * dep[1] : list of flows with 1 packet
140 * dep[X] : list of flows with X packets
141 */
142
143 unsigned int maxflows; /* number of flows in flows array */
144 int perturb_period;
145 unsigned int quantum; /* Allotment per round: MUST BE >= MTU */
146 struct timer_list perturb_timer;
147};
148
149/*
150 * sfq_head are either in a sfq_slot or in dep[] array
151 */
152static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val)
153{
154 if (val < SFQ_MAX_FLOWS)
155 return &q->slots[val].dep;
156 return &q->dep[val - SFQ_MAX_FLOWS];
157}
158
159/*
160 * In order to be able to quickly rehash our queue when timer changes
161 * q->perturbation, we store flow_keys in skb->cb[]
162 */
163struct sfq_skb_cb {
164 struct flow_keys keys;
165};
166
167static inline struct sfq_skb_cb *sfq_skb_cb(const struct sk_buff *skb)
168{
169 qdisc_cb_private_validate(skb, sizeof(struct sfq_skb_cb));
170 return (struct sfq_skb_cb *)qdisc_skb_cb(skb)->data;
171}
172
173static unsigned int sfq_hash(const struct sfq_sched_data *q,
174 const struct sk_buff *skb)
175{
176 const struct flow_keys *keys = &sfq_skb_cb(skb)->keys;
177 unsigned int hash;
178
179 hash = jhash_3words((__force u32)keys->dst,
180 (__force u32)keys->src ^ keys->ip_proto,
181 (__force u32)keys->ports, q->perturbation);
182 return hash & (q->divisor - 1);
183}
184
185static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
186 int *qerr)
187{
188 struct sfq_sched_data *q = qdisc_priv(sch);
189 struct tcf_result res;
190 int result;
191
192 if (TC_H_MAJ(skb->priority) == sch->handle &&
193 TC_H_MIN(skb->priority) > 0 &&
194 TC_H_MIN(skb->priority) <= q->divisor)
195 return TC_H_MIN(skb->priority);
196
197 if (!q->filter_list) {
198 skb_flow_dissect(skb, &sfq_skb_cb(skb)->keys);
199 return sfq_hash(q, skb) + 1;
200 }
201
202 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
203 result = tc_classify(skb, q->filter_list, &res);
204 if (result >= 0) {
205#ifdef CONFIG_NET_CLS_ACT
206 switch (result) {
207 case TC_ACT_STOLEN:
208 case TC_ACT_QUEUED:
209 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
210 case TC_ACT_SHOT:
211 return 0;
212 }
213#endif
214 if (TC_H_MIN(res.classid) <= q->divisor)
215 return TC_H_MIN(res.classid);
216 }
217 return 0;
218}
219
220/*
221 * x : slot number [0 .. SFQ_MAX_FLOWS - 1]
222 */
223static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
224{
225 sfq_index p, n;
226 struct sfq_slot *slot = &q->slots[x];
227 int qlen = slot->qlen;
228
229 p = qlen + SFQ_MAX_FLOWS;
230 n = q->dep[qlen].next;
231
232 slot->dep.next = n;
233 slot->dep.prev = p;
234
235 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */
236 sfq_dep_head(q, n)->prev = x;
237}
238
239#define sfq_unlink(q, x, n, p) \
240 n = q->slots[x].dep.next; \
241 p = q->slots[x].dep.prev; \
242 sfq_dep_head(q, p)->next = n; \
243 sfq_dep_head(q, n)->prev = p
244
245
246static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
247{
248 sfq_index p, n;
249 int d;
250
251 sfq_unlink(q, x, n, p);
252
253 d = q->slots[x].qlen--;
254 if (n == p && q->cur_depth == d)
255 q->cur_depth--;
256 sfq_link(q, x);
257}
258
259static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
260{
261 sfq_index p, n;
262 int d;
263
264 sfq_unlink(q, x, n, p);
265
266 d = ++q->slots[x].qlen;
267 if (q->cur_depth < d)
268 q->cur_depth = d;
269 sfq_link(q, x);
270}
271
272/* helper functions : might be changed when/if skb use a standard list_head */
273
274/* remove one skb from tail of slot queue */
275static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot)
276{
277 struct sk_buff *skb = slot->skblist_prev;
278
279 slot->skblist_prev = skb->prev;
280 skb->prev->next = (struct sk_buff *)slot;
281 skb->next = skb->prev = NULL;
282 return skb;
283}
284
285/* remove one skb from head of slot queue */
286static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot)
287{
288 struct sk_buff *skb = slot->skblist_next;
289
290 slot->skblist_next = skb->next;
291 skb->next->prev = (struct sk_buff *)slot;
292 skb->next = skb->prev = NULL;
293 return skb;
294}
295
296static inline void slot_queue_init(struct sfq_slot *slot)
297{
298 memset(slot, 0, sizeof(*slot));
299 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot;
300}
301
302/* add skb to slot queue (tail add) */
303static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb)
304{
305 skb->prev = slot->skblist_prev;
306 skb->next = (struct sk_buff *)slot;
307 slot->skblist_prev->next = skb;
308 slot->skblist_prev = skb;
309}
310
311#define slot_queue_walk(slot, skb) \
312 for (skb = slot->skblist_next; \
313 skb != (struct sk_buff *)slot; \
314 skb = skb->next)
315
316static unsigned int sfq_drop(struct Qdisc *sch)
317{
318 struct sfq_sched_data *q = qdisc_priv(sch);
319 sfq_index x, d = q->cur_depth;
320 struct sk_buff *skb;
321 unsigned int len;
322 struct sfq_slot *slot;
323
324 /* Queue is full! Find the longest slot and drop tail packet from it */
325 if (d > 1) {
326 x = q->dep[d].next;
327 slot = &q->slots[x];
328drop:
329 skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot);
330 len = qdisc_pkt_len(skb);
331 slot->backlog -= len;
332 sfq_dec(q, x);
333 kfree_skb(skb);
334 sch->q.qlen--;
335 sch->qstats.drops++;
336 sch->qstats.backlog -= len;
337 return len;
338 }
339
340 if (d == 1) {
341 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
342 x = q->tail->next;
343 slot = &q->slots[x];
344 q->tail->next = slot->next;
345 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
346 goto drop;
347 }
348
349 return 0;
350}
351
352/* Is ECN parameter configured */
353static int sfq_prob_mark(const struct sfq_sched_data *q)
354{
355 return q->flags & TC_RED_ECN;
356}
357
358/* Should packets over max threshold just be marked */
359static int sfq_hard_mark(const struct sfq_sched_data *q)
360{
361 return (q->flags & (TC_RED_ECN | TC_RED_HARDDROP)) == TC_RED_ECN;
362}
363
364static int sfq_headdrop(const struct sfq_sched_data *q)
365{
366 return q->headdrop;
367}
368
369static int
370sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
371{
372 struct sfq_sched_data *q = qdisc_priv(sch);
373 unsigned int hash;
374 sfq_index x, qlen;
375 struct sfq_slot *slot;
376 int uninitialized_var(ret);
377 struct sk_buff *head;
378 int delta;
379
380 hash = sfq_classify(skb, sch, &ret);
381 if (hash == 0) {
382 if (ret & __NET_XMIT_BYPASS)
383 sch->qstats.drops++;
384 kfree_skb(skb);
385 return ret;
386 }
387 hash--;
388
389 x = q->ht[hash];
390 slot = &q->slots[x];
391 if (x == SFQ_EMPTY_SLOT) {
392 x = q->dep[0].next; /* get a free slot */
393 if (x >= SFQ_MAX_FLOWS)
394 return qdisc_drop(skb, sch);
395 q->ht[hash] = x;
396 slot = &q->slots[x];
397 slot->hash = hash;
398 slot->backlog = 0; /* should already be 0 anyway... */
399 red_set_vars(&slot->vars);
400 goto enqueue;
401 }
402 if (q->red_parms) {
403 slot->vars.qavg = red_calc_qavg_no_idle_time(q->red_parms,
404 &slot->vars,
405 slot->backlog);
406 switch (red_action(q->red_parms,
407 &slot->vars,
408 slot->vars.qavg)) {
409 case RED_DONT_MARK:
410 break;
411
412 case RED_PROB_MARK:
413 sch->qstats.overlimits++;
414 if (sfq_prob_mark(q)) {
415 /* We know we have at least one packet in queue */
416 if (sfq_headdrop(q) &&
417 INET_ECN_set_ce(slot->skblist_next)) {
418 q->stats.prob_mark_head++;
419 break;
420 }
421 if (INET_ECN_set_ce(skb)) {
422 q->stats.prob_mark++;
423 break;
424 }
425 }
426 q->stats.prob_drop++;
427 goto congestion_drop;
428
429 case RED_HARD_MARK:
430 sch->qstats.overlimits++;
431 if (sfq_hard_mark(q)) {
432 /* We know we have at least one packet in queue */
433 if (sfq_headdrop(q) &&
434 INET_ECN_set_ce(slot->skblist_next)) {
435 q->stats.forced_mark_head++;
436 break;
437 }
438 if (INET_ECN_set_ce(skb)) {
439 q->stats.forced_mark++;
440 break;
441 }
442 }
443 q->stats.forced_drop++;
444 goto congestion_drop;
445 }
446 }
447
448 if (slot->qlen >= q->maxdepth) {
449congestion_drop:
450 if (!sfq_headdrop(q))
451 return qdisc_drop(skb, sch);
452
453 /* We know we have at least one packet in queue */
454 head = slot_dequeue_head(slot);
455 delta = qdisc_pkt_len(head) - qdisc_pkt_len(skb);
456 sch->qstats.backlog -= delta;
457 slot->backlog -= delta;
458 qdisc_drop(head, sch);
459
460 slot_queue_add(slot, skb);
461 return NET_XMIT_CN;
462 }
463
464enqueue:
465 sch->qstats.backlog += qdisc_pkt_len(skb);
466 slot->backlog += qdisc_pkt_len(skb);
467 slot_queue_add(slot, skb);
468 sfq_inc(q, x);
469 if (slot->qlen == 1) { /* The flow is new */
470 if (q->tail == NULL) { /* It is the first flow */
471 slot->next = x;
472 } else {
473 slot->next = q->tail->next;
474 q->tail->next = x;
475 }
476 /* We put this flow at the end of our flow list.
477 * This might sound unfair for a new flow to wait after old ones,
478 * but we could endup servicing new flows only, and freeze old ones.
479 */
480 q->tail = slot;
481 /* We could use a bigger initial quantum for new flows */
482 slot->allot = q->scaled_quantum;
483 }
484 if (++sch->q.qlen <= q->limit)
485 return NET_XMIT_SUCCESS;
486
487 qlen = slot->qlen;
488 sfq_drop(sch);
489 /* Return Congestion Notification only if we dropped a packet
490 * from this flow.
491 */
492 if (qlen != slot->qlen)
493 return NET_XMIT_CN;
494
495 /* As we dropped a packet, better let upper stack know this */
496 qdisc_tree_decrease_qlen(sch, 1);
497 return NET_XMIT_SUCCESS;
498}
499
500static struct sk_buff *
501sfq_dequeue(struct Qdisc *sch)
502{
503 struct sfq_sched_data *q = qdisc_priv(sch);
504 struct sk_buff *skb;
505 sfq_index a, next_a;
506 struct sfq_slot *slot;
507
508 /* No active slots */
509 if (q->tail == NULL)
510 return NULL;
511
512next_slot:
513 a = q->tail->next;
514 slot = &q->slots[a];
515 if (slot->allot <= 0) {
516 q->tail = slot;
517 slot->allot += q->scaled_quantum;
518 goto next_slot;
519 }
520 skb = slot_dequeue_head(slot);
521 sfq_dec(q, a);
522 qdisc_bstats_update(sch, skb);
523 sch->q.qlen--;
524 sch->qstats.backlog -= qdisc_pkt_len(skb);
525 slot->backlog -= qdisc_pkt_len(skb);
526 /* Is the slot empty? */
527 if (slot->qlen == 0) {
528 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
529 next_a = slot->next;
530 if (a == next_a) {
531 q->tail = NULL; /* no more active slots */
532 return skb;
533 }
534 q->tail->next = next_a;
535 } else {
536 slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb));
537 }
538 return skb;
539}
540
541static void
542sfq_reset(struct Qdisc *sch)
543{
544 struct sk_buff *skb;
545
546 while ((skb = sfq_dequeue(sch)) != NULL)
547 kfree_skb(skb);
548}
549
550/*
551 * When q->perturbation is changed, we rehash all queued skbs
552 * to avoid OOO (Out Of Order) effects.
553 * We dont use sfq_dequeue()/sfq_enqueue() because we dont want to change
554 * counters.
555 */
556static void sfq_rehash(struct Qdisc *sch)
557{
558 struct sfq_sched_data *q = qdisc_priv(sch);
559 struct sk_buff *skb;
560 int i;
561 struct sfq_slot *slot;
562 struct sk_buff_head list;
563 int dropped = 0;
564
565 __skb_queue_head_init(&list);
566
567 for (i = 0; i < q->maxflows; i++) {
568 slot = &q->slots[i];
569 if (!slot->qlen)
570 continue;
571 while (slot->qlen) {
572 skb = slot_dequeue_head(slot);
573 sfq_dec(q, i);
574 __skb_queue_tail(&list, skb);
575 }
576 slot->backlog = 0;
577 red_set_vars(&slot->vars);
578 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
579 }
580 q->tail = NULL;
581
582 while ((skb = __skb_dequeue(&list)) != NULL) {
583 unsigned int hash = sfq_hash(q, skb);
584 sfq_index x = q->ht[hash];
585
586 slot = &q->slots[x];
587 if (x == SFQ_EMPTY_SLOT) {
588 x = q->dep[0].next; /* get a free slot */
589 if (x >= SFQ_MAX_FLOWS) {
590drop: sch->qstats.backlog -= qdisc_pkt_len(skb);
591 kfree_skb(skb);
592 dropped++;
593 continue;
594 }
595 q->ht[hash] = x;
596 slot = &q->slots[x];
597 slot->hash = hash;
598 }
599 if (slot->qlen >= q->maxdepth)
600 goto drop;
601 slot_queue_add(slot, skb);
602 if (q->red_parms)
603 slot->vars.qavg = red_calc_qavg(q->red_parms,
604 &slot->vars,
605 slot->backlog);
606 slot->backlog += qdisc_pkt_len(skb);
607 sfq_inc(q, x);
608 if (slot->qlen == 1) { /* The flow is new */
609 if (q->tail == NULL) { /* It is the first flow */
610 slot->next = x;
611 } else {
612 slot->next = q->tail->next;
613 q->tail->next = x;
614 }
615 q->tail = slot;
616 slot->allot = q->scaled_quantum;
617 }
618 }
619 sch->q.qlen -= dropped;
620 qdisc_tree_decrease_qlen(sch, dropped);
621}
622
623static void sfq_perturbation(unsigned long arg)
624{
625 struct Qdisc *sch = (struct Qdisc *)arg;
626 struct sfq_sched_data *q = qdisc_priv(sch);
627 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
628
629 spin_lock(root_lock);
630 q->perturbation = net_random();
631 if (!q->filter_list && q->tail)
632 sfq_rehash(sch);
633 spin_unlock(root_lock);
634
635 if (q->perturb_period)
636 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
637}
638
639static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
640{
641 struct sfq_sched_data *q = qdisc_priv(sch);
642 struct tc_sfq_qopt *ctl = nla_data(opt);
643 struct tc_sfq_qopt_v1 *ctl_v1 = NULL;
644 unsigned int qlen;
645 struct red_parms *p = NULL;
646
647 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
648 return -EINVAL;
649 if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1)))
650 ctl_v1 = nla_data(opt);
651 if (ctl->divisor &&
652 (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536))
653 return -EINVAL;
654 if (ctl_v1 && ctl_v1->qth_min) {
655 p = kmalloc(sizeof(*p), GFP_KERNEL);
656 if (!p)
657 return -ENOMEM;
658 }
659 sch_tree_lock(sch);
660 if (ctl->quantum) {
661 q->quantum = ctl->quantum;
662 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
663 }
664 q->perturb_period = ctl->perturb_period * HZ;
665 if (ctl->flows)
666 q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS);
667 if (ctl->divisor) {
668 q->divisor = ctl->divisor;
669 q->maxflows = min_t(u32, q->maxflows, q->divisor);
670 }
671 if (ctl_v1) {
672 if (ctl_v1->depth)
673 q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH);
674 if (p) {
675 swap(q->red_parms, p);
676 red_set_parms(q->red_parms,
677 ctl_v1->qth_min, ctl_v1->qth_max,
678 ctl_v1->Wlog,
679 ctl_v1->Plog, ctl_v1->Scell_log,
680 NULL,
681 ctl_v1->max_P);
682 }
683 q->flags = ctl_v1->flags;
684 q->headdrop = ctl_v1->headdrop;
685 }
686 if (ctl->limit) {
687 q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows);
688 q->maxflows = min_t(u32, q->maxflows, q->limit);
689 }
690
691 qlen = sch->q.qlen;
692 while (sch->q.qlen > q->limit)
693 sfq_drop(sch);
694 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
695
696 del_timer(&q->perturb_timer);
697 if (q->perturb_period) {
698 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
699 q->perturbation = net_random();
700 }
701 sch_tree_unlock(sch);
702 kfree(p);
703 return 0;
704}
705
706static void *sfq_alloc(size_t sz)
707{
708 void *ptr = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN);
709
710 if (!ptr)
711 ptr = vmalloc(sz);
712 return ptr;
713}
714
715static void sfq_free(void *addr)
716{
717 if (addr) {
718 if (is_vmalloc_addr(addr))
719 vfree(addr);
720 else
721 kfree(addr);
722 }
723}
724
725static void sfq_destroy(struct Qdisc *sch)
726{
727 struct sfq_sched_data *q = qdisc_priv(sch);
728
729 tcf_destroy_chain(&q->filter_list);
730 q->perturb_period = 0;
731 del_timer_sync(&q->perturb_timer);
732 sfq_free(q->ht);
733 sfq_free(q->slots);
734 kfree(q->red_parms);
735}
736
737static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
738{
739 struct sfq_sched_data *q = qdisc_priv(sch);
740 int i;
741
742 q->perturb_timer.function = sfq_perturbation;
743 q->perturb_timer.data = (unsigned long)sch;
744 init_timer_deferrable(&q->perturb_timer);
745
746 for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) {
747 q->dep[i].next = i + SFQ_MAX_FLOWS;
748 q->dep[i].prev = i + SFQ_MAX_FLOWS;
749 }
750
751 q->limit = SFQ_MAX_DEPTH;
752 q->maxdepth = SFQ_MAX_DEPTH;
753 q->cur_depth = 0;
754 q->tail = NULL;
755 q->divisor = SFQ_DEFAULT_HASH_DIVISOR;
756 q->maxflows = SFQ_DEFAULT_FLOWS;
757 q->quantum = psched_mtu(qdisc_dev(sch));
758 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
759 q->perturb_period = 0;
760 q->perturbation = net_random();
761
762 if (opt) {
763 int err = sfq_change(sch, opt);
764 if (err)
765 return err;
766 }
767
768 q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor);
769 q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows);
770 if (!q->ht || !q->slots) {
771 sfq_destroy(sch);
772 return -ENOMEM;
773 }
774 for (i = 0; i < q->divisor; i++)
775 q->ht[i] = SFQ_EMPTY_SLOT;
776
777 for (i = 0; i < q->maxflows; i++) {
778 slot_queue_init(&q->slots[i]);
779 sfq_link(q, i);
780 }
781 if (q->limit >= 1)
782 sch->flags |= TCQ_F_CAN_BYPASS;
783 else
784 sch->flags &= ~TCQ_F_CAN_BYPASS;
785 return 0;
786}
787
788static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
789{
790 struct sfq_sched_data *q = qdisc_priv(sch);
791 unsigned char *b = skb_tail_pointer(skb);
792 struct tc_sfq_qopt_v1 opt;
793 struct red_parms *p = q->red_parms;
794
795 memset(&opt, 0, sizeof(opt));
796 opt.v0.quantum = q->quantum;
797 opt.v0.perturb_period = q->perturb_period / HZ;
798 opt.v0.limit = q->limit;
799 opt.v0.divisor = q->divisor;
800 opt.v0.flows = q->maxflows;
801 opt.depth = q->maxdepth;
802 opt.headdrop = q->headdrop;
803
804 if (p) {
805 opt.qth_min = p->qth_min >> p->Wlog;
806 opt.qth_max = p->qth_max >> p->Wlog;
807 opt.Wlog = p->Wlog;
808 opt.Plog = p->Plog;
809 opt.Scell_log = p->Scell_log;
810 opt.max_P = p->max_P;
811 }
812 memcpy(&opt.stats, &q->stats, sizeof(opt.stats));
813 opt.flags = q->flags;
814
815 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
816 goto nla_put_failure;
817
818 return skb->len;
819
820nla_put_failure:
821 nlmsg_trim(skb, b);
822 return -1;
823}
824
825static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg)
826{
827 return NULL;
828}
829
830static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
831{
832 return 0;
833}
834
835static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent,
836 u32 classid)
837{
838 /* we cannot bypass queue discipline anymore */
839 sch->flags &= ~TCQ_F_CAN_BYPASS;
840 return 0;
841}
842
843static void sfq_put(struct Qdisc *q, unsigned long cl)
844{
845}
846
847static struct tcf_proto **sfq_find_tcf(struct Qdisc *sch, unsigned long cl)
848{
849 struct sfq_sched_data *q = qdisc_priv(sch);
850
851 if (cl)
852 return NULL;
853 return &q->filter_list;
854}
855
856static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
857 struct sk_buff *skb, struct tcmsg *tcm)
858{
859 tcm->tcm_handle |= TC_H_MIN(cl);
860 return 0;
861}
862
863static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
864 struct gnet_dump *d)
865{
866 struct sfq_sched_data *q = qdisc_priv(sch);
867 sfq_index idx = q->ht[cl - 1];
868 struct gnet_stats_queue qs = { 0 };
869 struct tc_sfq_xstats xstats = { 0 };
870
871 if (idx != SFQ_EMPTY_SLOT) {
872 const struct sfq_slot *slot = &q->slots[idx];
873
874 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT;
875 qs.qlen = slot->qlen;
876 qs.backlog = slot->backlog;
877 }
878 if (gnet_stats_copy_queue(d, &qs) < 0)
879 return -1;
880 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
881}
882
883static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
884{
885 struct sfq_sched_data *q = qdisc_priv(sch);
886 unsigned int i;
887
888 if (arg->stop)
889 return;
890
891 for (i = 0; i < q->divisor; i++) {
892 if (q->ht[i] == SFQ_EMPTY_SLOT ||
893 arg->count < arg->skip) {
894 arg->count++;
895 continue;
896 }
897 if (arg->fn(sch, i + 1, arg) < 0) {
898 arg->stop = 1;
899 break;
900 }
901 arg->count++;
902 }
903}
904
905static const struct Qdisc_class_ops sfq_class_ops = {
906 .leaf = sfq_leaf,
907 .get = sfq_get,
908 .put = sfq_put,
909 .tcf_chain = sfq_find_tcf,
910 .bind_tcf = sfq_bind,
911 .unbind_tcf = sfq_put,
912 .dump = sfq_dump_class,
913 .dump_stats = sfq_dump_class_stats,
914 .walk = sfq_walk,
915};
916
917static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
918 .cl_ops = &sfq_class_ops,
919 .id = "sfq",
920 .priv_size = sizeof(struct sfq_sched_data),
921 .enqueue = sfq_enqueue,
922 .dequeue = sfq_dequeue,
923 .peek = qdisc_peek_dequeued,
924 .drop = sfq_drop,
925 .init = sfq_init,
926 .reset = sfq_reset,
927 .destroy = sfq_destroy,
928 .change = NULL,
929 .dump = sfq_dump,
930 .owner = THIS_MODULE,
931};
932
933static int __init sfq_module_init(void)
934{
935 return register_qdisc(&sfq_qdisc_ops);
936}
937static void __exit sfq_module_exit(void)
938{
939 unregister_qdisc(&sfq_qdisc_ops);
940}
941module_init(sfq_module_init)
942module_exit(sfq_module_exit)
943MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * net/sched/sch_sfq.c Stochastic Fairness Queueing discipline.
4 *
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 */
7
8#include <linux/module.h>
9#include <linux/types.h>
10#include <linux/kernel.h>
11#include <linux/jiffies.h>
12#include <linux/string.h>
13#include <linux/in.h>
14#include <linux/errno.h>
15#include <linux/init.h>
16#include <linux/skbuff.h>
17#include <linux/siphash.h>
18#include <linux/slab.h>
19#include <linux/vmalloc.h>
20#include <net/netlink.h>
21#include <net/pkt_sched.h>
22#include <net/pkt_cls.h>
23#include <net/red.h>
24
25
26/* Stochastic Fairness Queuing algorithm.
27 =======================================
28
29 Source:
30 Paul E. McKenney "Stochastic Fairness Queuing",
31 IEEE INFOCOMM'90 Proceedings, San Francisco, 1990.
32
33 Paul E. McKenney "Stochastic Fairness Queuing",
34 "Interworking: Research and Experience", v.2, 1991, p.113-131.
35
36
37 See also:
38 M. Shreedhar and George Varghese "Efficient Fair
39 Queuing using Deficit Round Robin", Proc. SIGCOMM 95.
40
41
42 This is not the thing that is usually called (W)FQ nowadays.
43 It does not use any timestamp mechanism, but instead
44 processes queues in round-robin order.
45
46 ADVANTAGE:
47
48 - It is very cheap. Both CPU and memory requirements are minimal.
49
50 DRAWBACKS:
51
52 - "Stochastic" -> It is not 100% fair.
53 When hash collisions occur, several flows are considered as one.
54
55 - "Round-robin" -> It introduces larger delays than virtual clock
56 based schemes, and should not be used for isolating interactive
57 traffic from non-interactive. It means, that this scheduler
58 should be used as leaf of CBQ or P3, which put interactive traffic
59 to higher priority band.
60
61 We still need true WFQ for top level CSZ, but using WFQ
62 for the best effort traffic is absolutely pointless:
63 SFQ is superior for this purpose.
64
65 IMPLEMENTATION:
66 This implementation limits :
67 - maximal queue length per flow to 127 packets.
68 - max mtu to 2^18-1;
69 - max 65408 flows,
70 - number of hash buckets to 65536.
71
72 It is easy to increase these values, but not in flight. */
73
74#define SFQ_MAX_DEPTH 127 /* max number of packets per flow */
75#define SFQ_DEFAULT_FLOWS 128
76#define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */
77#define SFQ_EMPTY_SLOT 0xffff
78#define SFQ_DEFAULT_HASH_DIVISOR 1024
79
80/* We use 16 bits to store allot, and want to handle packets up to 64K
81 * Scale allot by 8 (1<<3) so that no overflow occurs.
82 */
83#define SFQ_ALLOT_SHIFT 3
84#define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT)
85
86/* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */
87typedef u16 sfq_index;
88
89/*
90 * We dont use pointers to save space.
91 * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array
92 * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH]
93 * are 'pointers' to dep[] array
94 */
95struct sfq_head {
96 sfq_index next;
97 sfq_index prev;
98};
99
100struct sfq_slot {
101 struct sk_buff *skblist_next;
102 struct sk_buff *skblist_prev;
103 sfq_index qlen; /* number of skbs in skblist */
104 sfq_index next; /* next slot in sfq RR chain */
105 struct sfq_head dep; /* anchor in dep[] chains */
106 unsigned short hash; /* hash value (index in ht[]) */
107 short allot; /* credit for this slot */
108
109 unsigned int backlog;
110 struct red_vars vars;
111};
112
113struct sfq_sched_data {
114/* frequently used fields */
115 int limit; /* limit of total number of packets in this qdisc */
116 unsigned int divisor; /* number of slots in hash table */
117 u8 headdrop;
118 u8 maxdepth; /* limit of packets per flow */
119
120 siphash_key_t perturbation;
121 u8 cur_depth; /* depth of longest slot */
122 u8 flags;
123 unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */
124 struct tcf_proto __rcu *filter_list;
125 struct tcf_block *block;
126 sfq_index *ht; /* Hash table ('divisor' slots) */
127 struct sfq_slot *slots; /* Flows table ('maxflows' entries) */
128
129 struct red_parms *red_parms;
130 struct tc_sfqred_stats stats;
131 struct sfq_slot *tail; /* current slot in round */
132
133 struct sfq_head dep[SFQ_MAX_DEPTH + 1];
134 /* Linked lists of slots, indexed by depth
135 * dep[0] : list of unused flows
136 * dep[1] : list of flows with 1 packet
137 * dep[X] : list of flows with X packets
138 */
139
140 unsigned int maxflows; /* number of flows in flows array */
141 int perturb_period;
142 unsigned int quantum; /* Allotment per round: MUST BE >= MTU */
143 struct timer_list perturb_timer;
144 struct Qdisc *sch;
145};
146
147/*
148 * sfq_head are either in a sfq_slot or in dep[] array
149 */
150static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val)
151{
152 if (val < SFQ_MAX_FLOWS)
153 return &q->slots[val].dep;
154 return &q->dep[val - SFQ_MAX_FLOWS];
155}
156
157static unsigned int sfq_hash(const struct sfq_sched_data *q,
158 const struct sk_buff *skb)
159{
160 return skb_get_hash_perturb(skb, &q->perturbation) & (q->divisor - 1);
161}
162
163static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
164 int *qerr)
165{
166 struct sfq_sched_data *q = qdisc_priv(sch);
167 struct tcf_result res;
168 struct tcf_proto *fl;
169 int result;
170
171 if (TC_H_MAJ(skb->priority) == sch->handle &&
172 TC_H_MIN(skb->priority) > 0 &&
173 TC_H_MIN(skb->priority) <= q->divisor)
174 return TC_H_MIN(skb->priority);
175
176 fl = rcu_dereference_bh(q->filter_list);
177 if (!fl)
178 return sfq_hash(q, skb) + 1;
179
180 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
181 result = tcf_classify(skb, fl, &res, false);
182 if (result >= 0) {
183#ifdef CONFIG_NET_CLS_ACT
184 switch (result) {
185 case TC_ACT_STOLEN:
186 case TC_ACT_QUEUED:
187 case TC_ACT_TRAP:
188 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
189 fallthrough;
190 case TC_ACT_SHOT:
191 return 0;
192 }
193#endif
194 if (TC_H_MIN(res.classid) <= q->divisor)
195 return TC_H_MIN(res.classid);
196 }
197 return 0;
198}
199
200/*
201 * x : slot number [0 .. SFQ_MAX_FLOWS - 1]
202 */
203static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
204{
205 sfq_index p, n;
206 struct sfq_slot *slot = &q->slots[x];
207 int qlen = slot->qlen;
208
209 p = qlen + SFQ_MAX_FLOWS;
210 n = q->dep[qlen].next;
211
212 slot->dep.next = n;
213 slot->dep.prev = p;
214
215 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */
216 sfq_dep_head(q, n)->prev = x;
217}
218
219#define sfq_unlink(q, x, n, p) \
220 do { \
221 n = q->slots[x].dep.next; \
222 p = q->slots[x].dep.prev; \
223 sfq_dep_head(q, p)->next = n; \
224 sfq_dep_head(q, n)->prev = p; \
225 } while (0)
226
227
228static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
229{
230 sfq_index p, n;
231 int d;
232
233 sfq_unlink(q, x, n, p);
234
235 d = q->slots[x].qlen--;
236 if (n == p && q->cur_depth == d)
237 q->cur_depth--;
238 sfq_link(q, x);
239}
240
241static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
242{
243 sfq_index p, n;
244 int d;
245
246 sfq_unlink(q, x, n, p);
247
248 d = ++q->slots[x].qlen;
249 if (q->cur_depth < d)
250 q->cur_depth = d;
251 sfq_link(q, x);
252}
253
254/* helper functions : might be changed when/if skb use a standard list_head */
255
256/* remove one skb from tail of slot queue */
257static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot)
258{
259 struct sk_buff *skb = slot->skblist_prev;
260
261 slot->skblist_prev = skb->prev;
262 skb->prev->next = (struct sk_buff *)slot;
263 skb->next = skb->prev = NULL;
264 return skb;
265}
266
267/* remove one skb from head of slot queue */
268static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot)
269{
270 struct sk_buff *skb = slot->skblist_next;
271
272 slot->skblist_next = skb->next;
273 skb->next->prev = (struct sk_buff *)slot;
274 skb->next = skb->prev = NULL;
275 return skb;
276}
277
278static inline void slot_queue_init(struct sfq_slot *slot)
279{
280 memset(slot, 0, sizeof(*slot));
281 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot;
282}
283
284/* add skb to slot queue (tail add) */
285static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb)
286{
287 skb->prev = slot->skblist_prev;
288 skb->next = (struct sk_buff *)slot;
289 slot->skblist_prev->next = skb;
290 slot->skblist_prev = skb;
291}
292
293static unsigned int sfq_drop(struct Qdisc *sch, struct sk_buff **to_free)
294{
295 struct sfq_sched_data *q = qdisc_priv(sch);
296 sfq_index x, d = q->cur_depth;
297 struct sk_buff *skb;
298 unsigned int len;
299 struct sfq_slot *slot;
300
301 /* Queue is full! Find the longest slot and drop tail packet from it */
302 if (d > 1) {
303 x = q->dep[d].next;
304 slot = &q->slots[x];
305drop:
306 skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot);
307 len = qdisc_pkt_len(skb);
308 slot->backlog -= len;
309 sfq_dec(q, x);
310 sch->q.qlen--;
311 qdisc_qstats_backlog_dec(sch, skb);
312 qdisc_drop(skb, sch, to_free);
313 return len;
314 }
315
316 if (d == 1) {
317 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
318 x = q->tail->next;
319 slot = &q->slots[x];
320 q->tail->next = slot->next;
321 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
322 goto drop;
323 }
324
325 return 0;
326}
327
328/* Is ECN parameter configured */
329static int sfq_prob_mark(const struct sfq_sched_data *q)
330{
331 return q->flags & TC_RED_ECN;
332}
333
334/* Should packets over max threshold just be marked */
335static int sfq_hard_mark(const struct sfq_sched_data *q)
336{
337 return (q->flags & (TC_RED_ECN | TC_RED_HARDDROP)) == TC_RED_ECN;
338}
339
340static int sfq_headdrop(const struct sfq_sched_data *q)
341{
342 return q->headdrop;
343}
344
345static int
346sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
347{
348 struct sfq_sched_data *q = qdisc_priv(sch);
349 unsigned int hash, dropped;
350 sfq_index x, qlen;
351 struct sfq_slot *slot;
352 int ret;
353 struct sk_buff *head;
354 int delta;
355
356 hash = sfq_classify(skb, sch, &ret);
357 if (hash == 0) {
358 if (ret & __NET_XMIT_BYPASS)
359 qdisc_qstats_drop(sch);
360 __qdisc_drop(skb, to_free);
361 return ret;
362 }
363 hash--;
364
365 x = q->ht[hash];
366 slot = &q->slots[x];
367 if (x == SFQ_EMPTY_SLOT) {
368 x = q->dep[0].next; /* get a free slot */
369 if (x >= SFQ_MAX_FLOWS)
370 return qdisc_drop(skb, sch, to_free);
371 q->ht[hash] = x;
372 slot = &q->slots[x];
373 slot->hash = hash;
374 slot->backlog = 0; /* should already be 0 anyway... */
375 red_set_vars(&slot->vars);
376 goto enqueue;
377 }
378 if (q->red_parms) {
379 slot->vars.qavg = red_calc_qavg_no_idle_time(q->red_parms,
380 &slot->vars,
381 slot->backlog);
382 switch (red_action(q->red_parms,
383 &slot->vars,
384 slot->vars.qavg)) {
385 case RED_DONT_MARK:
386 break;
387
388 case RED_PROB_MARK:
389 qdisc_qstats_overlimit(sch);
390 if (sfq_prob_mark(q)) {
391 /* We know we have at least one packet in queue */
392 if (sfq_headdrop(q) &&
393 INET_ECN_set_ce(slot->skblist_next)) {
394 q->stats.prob_mark_head++;
395 break;
396 }
397 if (INET_ECN_set_ce(skb)) {
398 q->stats.prob_mark++;
399 break;
400 }
401 }
402 q->stats.prob_drop++;
403 goto congestion_drop;
404
405 case RED_HARD_MARK:
406 qdisc_qstats_overlimit(sch);
407 if (sfq_hard_mark(q)) {
408 /* We know we have at least one packet in queue */
409 if (sfq_headdrop(q) &&
410 INET_ECN_set_ce(slot->skblist_next)) {
411 q->stats.forced_mark_head++;
412 break;
413 }
414 if (INET_ECN_set_ce(skb)) {
415 q->stats.forced_mark++;
416 break;
417 }
418 }
419 q->stats.forced_drop++;
420 goto congestion_drop;
421 }
422 }
423
424 if (slot->qlen >= q->maxdepth) {
425congestion_drop:
426 if (!sfq_headdrop(q))
427 return qdisc_drop(skb, sch, to_free);
428
429 /* We know we have at least one packet in queue */
430 head = slot_dequeue_head(slot);
431 delta = qdisc_pkt_len(head) - qdisc_pkt_len(skb);
432 sch->qstats.backlog -= delta;
433 slot->backlog -= delta;
434 qdisc_drop(head, sch, to_free);
435
436 slot_queue_add(slot, skb);
437 qdisc_tree_reduce_backlog(sch, 0, delta);
438 return NET_XMIT_CN;
439 }
440
441enqueue:
442 qdisc_qstats_backlog_inc(sch, skb);
443 slot->backlog += qdisc_pkt_len(skb);
444 slot_queue_add(slot, skb);
445 sfq_inc(q, x);
446 if (slot->qlen == 1) { /* The flow is new */
447 if (q->tail == NULL) { /* It is the first flow */
448 slot->next = x;
449 } else {
450 slot->next = q->tail->next;
451 q->tail->next = x;
452 }
453 /* We put this flow at the end of our flow list.
454 * This might sound unfair for a new flow to wait after old ones,
455 * but we could endup servicing new flows only, and freeze old ones.
456 */
457 q->tail = slot;
458 /* We could use a bigger initial quantum for new flows */
459 slot->allot = q->scaled_quantum;
460 }
461 if (++sch->q.qlen <= q->limit)
462 return NET_XMIT_SUCCESS;
463
464 qlen = slot->qlen;
465 dropped = sfq_drop(sch, to_free);
466 /* Return Congestion Notification only if we dropped a packet
467 * from this flow.
468 */
469 if (qlen != slot->qlen) {
470 qdisc_tree_reduce_backlog(sch, 0, dropped - qdisc_pkt_len(skb));
471 return NET_XMIT_CN;
472 }
473
474 /* As we dropped a packet, better let upper stack know this */
475 qdisc_tree_reduce_backlog(sch, 1, dropped);
476 return NET_XMIT_SUCCESS;
477}
478
479static struct sk_buff *
480sfq_dequeue(struct Qdisc *sch)
481{
482 struct sfq_sched_data *q = qdisc_priv(sch);
483 struct sk_buff *skb;
484 sfq_index a, next_a;
485 struct sfq_slot *slot;
486
487 /* No active slots */
488 if (q->tail == NULL)
489 return NULL;
490
491next_slot:
492 a = q->tail->next;
493 slot = &q->slots[a];
494 if (slot->allot <= 0) {
495 q->tail = slot;
496 slot->allot += q->scaled_quantum;
497 goto next_slot;
498 }
499 skb = slot_dequeue_head(slot);
500 sfq_dec(q, a);
501 qdisc_bstats_update(sch, skb);
502 sch->q.qlen--;
503 qdisc_qstats_backlog_dec(sch, skb);
504 slot->backlog -= qdisc_pkt_len(skb);
505 /* Is the slot empty? */
506 if (slot->qlen == 0) {
507 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
508 next_a = slot->next;
509 if (a == next_a) {
510 q->tail = NULL; /* no more active slots */
511 return skb;
512 }
513 q->tail->next = next_a;
514 } else {
515 slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb));
516 }
517 return skb;
518}
519
520static void
521sfq_reset(struct Qdisc *sch)
522{
523 struct sk_buff *skb;
524
525 while ((skb = sfq_dequeue(sch)) != NULL)
526 rtnl_kfree_skbs(skb, skb);
527}
528
529/*
530 * When q->perturbation is changed, we rehash all queued skbs
531 * to avoid OOO (Out Of Order) effects.
532 * We dont use sfq_dequeue()/sfq_enqueue() because we dont want to change
533 * counters.
534 */
535static void sfq_rehash(struct Qdisc *sch)
536{
537 struct sfq_sched_data *q = qdisc_priv(sch);
538 struct sk_buff *skb;
539 int i;
540 struct sfq_slot *slot;
541 struct sk_buff_head list;
542 int dropped = 0;
543 unsigned int drop_len = 0;
544
545 __skb_queue_head_init(&list);
546
547 for (i = 0; i < q->maxflows; i++) {
548 slot = &q->slots[i];
549 if (!slot->qlen)
550 continue;
551 while (slot->qlen) {
552 skb = slot_dequeue_head(slot);
553 sfq_dec(q, i);
554 __skb_queue_tail(&list, skb);
555 }
556 slot->backlog = 0;
557 red_set_vars(&slot->vars);
558 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
559 }
560 q->tail = NULL;
561
562 while ((skb = __skb_dequeue(&list)) != NULL) {
563 unsigned int hash = sfq_hash(q, skb);
564 sfq_index x = q->ht[hash];
565
566 slot = &q->slots[x];
567 if (x == SFQ_EMPTY_SLOT) {
568 x = q->dep[0].next; /* get a free slot */
569 if (x >= SFQ_MAX_FLOWS) {
570drop:
571 qdisc_qstats_backlog_dec(sch, skb);
572 drop_len += qdisc_pkt_len(skb);
573 kfree_skb(skb);
574 dropped++;
575 continue;
576 }
577 q->ht[hash] = x;
578 slot = &q->slots[x];
579 slot->hash = hash;
580 }
581 if (slot->qlen >= q->maxdepth)
582 goto drop;
583 slot_queue_add(slot, skb);
584 if (q->red_parms)
585 slot->vars.qavg = red_calc_qavg(q->red_parms,
586 &slot->vars,
587 slot->backlog);
588 slot->backlog += qdisc_pkt_len(skb);
589 sfq_inc(q, x);
590 if (slot->qlen == 1) { /* The flow is new */
591 if (q->tail == NULL) { /* It is the first flow */
592 slot->next = x;
593 } else {
594 slot->next = q->tail->next;
595 q->tail->next = x;
596 }
597 q->tail = slot;
598 slot->allot = q->scaled_quantum;
599 }
600 }
601 sch->q.qlen -= dropped;
602 qdisc_tree_reduce_backlog(sch, dropped, drop_len);
603}
604
605static void sfq_perturbation(struct timer_list *t)
606{
607 struct sfq_sched_data *q = from_timer(q, t, perturb_timer);
608 struct Qdisc *sch = q->sch;
609 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
610 siphash_key_t nkey;
611
612 get_random_bytes(&nkey, sizeof(nkey));
613 spin_lock(root_lock);
614 q->perturbation = nkey;
615 if (!q->filter_list && q->tail)
616 sfq_rehash(sch);
617 spin_unlock(root_lock);
618
619 if (q->perturb_period)
620 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
621}
622
623static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
624{
625 struct sfq_sched_data *q = qdisc_priv(sch);
626 struct tc_sfq_qopt *ctl = nla_data(opt);
627 struct tc_sfq_qopt_v1 *ctl_v1 = NULL;
628 unsigned int qlen, dropped = 0;
629 struct red_parms *p = NULL;
630 struct sk_buff *to_free = NULL;
631 struct sk_buff *tail = NULL;
632
633 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
634 return -EINVAL;
635 if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1)))
636 ctl_v1 = nla_data(opt);
637 if (ctl->divisor &&
638 (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536))
639 return -EINVAL;
640
641 /* slot->allot is a short, make sure quantum is not too big. */
642 if (ctl->quantum) {
643 unsigned int scaled = SFQ_ALLOT_SIZE(ctl->quantum);
644
645 if (scaled <= 0 || scaled > SHRT_MAX)
646 return -EINVAL;
647 }
648
649 if (ctl_v1 && !red_check_params(ctl_v1->qth_min, ctl_v1->qth_max,
650 ctl_v1->Wlog, ctl_v1->Scell_log, NULL))
651 return -EINVAL;
652 if (ctl_v1 && ctl_v1->qth_min) {
653 p = kmalloc(sizeof(*p), GFP_KERNEL);
654 if (!p)
655 return -ENOMEM;
656 }
657 sch_tree_lock(sch);
658 if (ctl->quantum) {
659 q->quantum = ctl->quantum;
660 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
661 }
662 q->perturb_period = ctl->perturb_period * HZ;
663 if (ctl->flows)
664 q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS);
665 if (ctl->divisor) {
666 q->divisor = ctl->divisor;
667 q->maxflows = min_t(u32, q->maxflows, q->divisor);
668 }
669 if (ctl_v1) {
670 if (ctl_v1->depth)
671 q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH);
672 if (p) {
673 swap(q->red_parms, p);
674 red_set_parms(q->red_parms,
675 ctl_v1->qth_min, ctl_v1->qth_max,
676 ctl_v1->Wlog,
677 ctl_v1->Plog, ctl_v1->Scell_log,
678 NULL,
679 ctl_v1->max_P);
680 }
681 q->flags = ctl_v1->flags;
682 q->headdrop = ctl_v1->headdrop;
683 }
684 if (ctl->limit) {
685 q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows);
686 q->maxflows = min_t(u32, q->maxflows, q->limit);
687 }
688
689 qlen = sch->q.qlen;
690 while (sch->q.qlen > q->limit) {
691 dropped += sfq_drop(sch, &to_free);
692 if (!tail)
693 tail = to_free;
694 }
695
696 rtnl_kfree_skbs(to_free, tail);
697 qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
698
699 del_timer(&q->perturb_timer);
700 if (q->perturb_period) {
701 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
702 get_random_bytes(&q->perturbation, sizeof(q->perturbation));
703 }
704 sch_tree_unlock(sch);
705 kfree(p);
706 return 0;
707}
708
709static void *sfq_alloc(size_t sz)
710{
711 return kvmalloc(sz, GFP_KERNEL);
712}
713
714static void sfq_free(void *addr)
715{
716 kvfree(addr);
717}
718
719static void sfq_destroy(struct Qdisc *sch)
720{
721 struct sfq_sched_data *q = qdisc_priv(sch);
722
723 tcf_block_put(q->block);
724 q->perturb_period = 0;
725 del_timer_sync(&q->perturb_timer);
726 sfq_free(q->ht);
727 sfq_free(q->slots);
728 kfree(q->red_parms);
729}
730
731static int sfq_init(struct Qdisc *sch, struct nlattr *opt,
732 struct netlink_ext_ack *extack)
733{
734 struct sfq_sched_data *q = qdisc_priv(sch);
735 int i;
736 int err;
737
738 q->sch = sch;
739 timer_setup(&q->perturb_timer, sfq_perturbation, TIMER_DEFERRABLE);
740
741 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
742 if (err)
743 return err;
744
745 for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) {
746 q->dep[i].next = i + SFQ_MAX_FLOWS;
747 q->dep[i].prev = i + SFQ_MAX_FLOWS;
748 }
749
750 q->limit = SFQ_MAX_DEPTH;
751 q->maxdepth = SFQ_MAX_DEPTH;
752 q->cur_depth = 0;
753 q->tail = NULL;
754 q->divisor = SFQ_DEFAULT_HASH_DIVISOR;
755 q->maxflows = SFQ_DEFAULT_FLOWS;
756 q->quantum = psched_mtu(qdisc_dev(sch));
757 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
758 q->perturb_period = 0;
759 get_random_bytes(&q->perturbation, sizeof(q->perturbation));
760
761 if (opt) {
762 int err = sfq_change(sch, opt);
763 if (err)
764 return err;
765 }
766
767 q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor);
768 q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows);
769 if (!q->ht || !q->slots) {
770 /* Note: sfq_destroy() will be called by our caller */
771 return -ENOMEM;
772 }
773
774 for (i = 0; i < q->divisor; i++)
775 q->ht[i] = SFQ_EMPTY_SLOT;
776
777 for (i = 0; i < q->maxflows; i++) {
778 slot_queue_init(&q->slots[i]);
779 sfq_link(q, i);
780 }
781 if (q->limit >= 1)
782 sch->flags |= TCQ_F_CAN_BYPASS;
783 else
784 sch->flags &= ~TCQ_F_CAN_BYPASS;
785 return 0;
786}
787
788static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
789{
790 struct sfq_sched_data *q = qdisc_priv(sch);
791 unsigned char *b = skb_tail_pointer(skb);
792 struct tc_sfq_qopt_v1 opt;
793 struct red_parms *p = q->red_parms;
794
795 memset(&opt, 0, sizeof(opt));
796 opt.v0.quantum = q->quantum;
797 opt.v0.perturb_period = q->perturb_period / HZ;
798 opt.v0.limit = q->limit;
799 opt.v0.divisor = q->divisor;
800 opt.v0.flows = q->maxflows;
801 opt.depth = q->maxdepth;
802 opt.headdrop = q->headdrop;
803
804 if (p) {
805 opt.qth_min = p->qth_min >> p->Wlog;
806 opt.qth_max = p->qth_max >> p->Wlog;
807 opt.Wlog = p->Wlog;
808 opt.Plog = p->Plog;
809 opt.Scell_log = p->Scell_log;
810 opt.max_P = p->max_P;
811 }
812 memcpy(&opt.stats, &q->stats, sizeof(opt.stats));
813 opt.flags = q->flags;
814
815 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
816 goto nla_put_failure;
817
818 return skb->len;
819
820nla_put_failure:
821 nlmsg_trim(skb, b);
822 return -1;
823}
824
825static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg)
826{
827 return NULL;
828}
829
830static unsigned long sfq_find(struct Qdisc *sch, u32 classid)
831{
832 return 0;
833}
834
835static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent,
836 u32 classid)
837{
838 return 0;
839}
840
841static void sfq_unbind(struct Qdisc *q, unsigned long cl)
842{
843}
844
845static struct tcf_block *sfq_tcf_block(struct Qdisc *sch, unsigned long cl,
846 struct netlink_ext_ack *extack)
847{
848 struct sfq_sched_data *q = qdisc_priv(sch);
849
850 if (cl)
851 return NULL;
852 return q->block;
853}
854
855static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
856 struct sk_buff *skb, struct tcmsg *tcm)
857{
858 tcm->tcm_handle |= TC_H_MIN(cl);
859 return 0;
860}
861
862static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
863 struct gnet_dump *d)
864{
865 struct sfq_sched_data *q = qdisc_priv(sch);
866 sfq_index idx = q->ht[cl - 1];
867 struct gnet_stats_queue qs = { 0 };
868 struct tc_sfq_xstats xstats = { 0 };
869
870 if (idx != SFQ_EMPTY_SLOT) {
871 const struct sfq_slot *slot = &q->slots[idx];
872
873 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT;
874 qs.qlen = slot->qlen;
875 qs.backlog = slot->backlog;
876 }
877 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
878 return -1;
879 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
880}
881
882static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
883{
884 struct sfq_sched_data *q = qdisc_priv(sch);
885 unsigned int i;
886
887 if (arg->stop)
888 return;
889
890 for (i = 0; i < q->divisor; i++) {
891 if (q->ht[i] == SFQ_EMPTY_SLOT ||
892 arg->count < arg->skip) {
893 arg->count++;
894 continue;
895 }
896 if (arg->fn(sch, i + 1, arg) < 0) {
897 arg->stop = 1;
898 break;
899 }
900 arg->count++;
901 }
902}
903
904static const struct Qdisc_class_ops sfq_class_ops = {
905 .leaf = sfq_leaf,
906 .find = sfq_find,
907 .tcf_block = sfq_tcf_block,
908 .bind_tcf = sfq_bind,
909 .unbind_tcf = sfq_unbind,
910 .dump = sfq_dump_class,
911 .dump_stats = sfq_dump_class_stats,
912 .walk = sfq_walk,
913};
914
915static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
916 .cl_ops = &sfq_class_ops,
917 .id = "sfq",
918 .priv_size = sizeof(struct sfq_sched_data),
919 .enqueue = sfq_enqueue,
920 .dequeue = sfq_dequeue,
921 .peek = qdisc_peek_dequeued,
922 .init = sfq_init,
923 .reset = sfq_reset,
924 .destroy = sfq_destroy,
925 .change = NULL,
926 .dump = sfq_dump,
927 .owner = THIS_MODULE,
928};
929
930static int __init sfq_module_init(void)
931{
932 return register_qdisc(&sfq_qdisc_ops);
933}
934static void __exit sfq_module_exit(void)
935{
936 unregister_qdisc(&sfq_qdisc_ops);
937}
938module_init(sfq_module_init)
939module_exit(sfq_module_exit)
940MODULE_LICENSE("GPL");