?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
sctp: delay the authentication for the duplicated cookie-echo chunk
[linux/fpc-iii.git] / net / sched / sch_sfq.c
blobd8c2b6baaad2d3fb9af0923e9440713980bb6a9c
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/red.h>
27
28
29 /* Stochastic Fairness Queuing algorithm.
30 =======================================
31
32 Source:
33 Paul E. McKenney "Stochastic Fairness Queuing",
34 IEEE INFOCOMM'90 Proceedings, San Francisco, 1990.
35
36 Paul E. McKenney "Stochastic Fairness Queuing",
37 "Interworking: Research and Experience", v.2, 1991, p.113-131.
38
39
40 See also:
41 M. Shreedhar and George Varghese "Efficient Fair
42 Queuing using Deficit Round Robin", Proc. SIGCOMM 95.
43
44
45 This is not the thing that is usually called (W)FQ nowadays.
46 It does not use any timestamp mechanism, but instead
47 processes queues in round-robin order.
48
49 ADVANTAGE:
50
51 - It is very cheap. Both CPU and memory requirements are minimal.
52
53 DRAWBACKS:
54
55 - "Stochastic" -> It is not 100% fair.
56 When hash collisions occur, several flows are considered as one.
57
58 - "Round-robin" -> It introduces larger delays than virtual clock
59 based schemes, and should not be used for isolating interactive
60 traffic from non-interactive. It means, that this scheduler
61 should be used as leaf of CBQ or P3, which put interactive traffic
62 to higher priority band.
63
64 We still need true WFQ for top level CSZ, but using WFQ
65 for the best effort traffic is absolutely pointless:
66 SFQ is superior for this purpose.
67
68 IMPLEMENTATION:
69 This implementation limits :
70 - maximal queue length per flow to 127 packets.
71 - max mtu to 2^18-1;
72 - max 65408 flows,
73 - number of hash buckets to 65536.
74
75 It is easy to increase these values, but not in flight. */
76
77 #define SFQ_MAX_DEPTH 127 /* max number of packets per flow */
78 #define SFQ_DEFAULT_FLOWS 128
79 #define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */
80 #define SFQ_EMPTY_SLOT 0xffff
81 #define SFQ_DEFAULT_HASH_DIVISOR 1024
82
83 /* We use 16 bits to store allot, and want to handle packets up to 64K
84 * Scale allot by 8 (1<<3) so that no overflow occurs.
85 */
86 #define SFQ_ALLOT_SHIFT 3
87 #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT)
88
89 /* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */
90 typedef u16 sfq_index;
91
92 /*
93 * We dont use pointers to save space.
94 * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array
95 * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH]
96 * are 'pointers' to dep[] array
97 */
98 struct sfq_head {
99 sfq_index next;
100 sfq_index prev;
101 };
102
103 struct sfq_slot {
104 struct sk_buff *skblist_next;
105 struct sk_buff *skblist_prev;
106 sfq_index qlen; /* number of skbs in skblist */
107 sfq_index next; /* next slot in sfq RR chain */
108 struct sfq_head dep; /* anchor in dep[] chains */
109 unsigned short hash; /* hash value (index in ht[]) */
110 short allot; /* credit for this slot */
111
112 unsigned int backlog;
113 struct red_vars vars;
114 };
115
116 struct sfq_sched_data {
117 /* frequently used fields */
118 int limit; /* limit of total number of packets in this qdisc */
119 unsigned int divisor; /* number of slots in hash table */
120 u8 headdrop;
121 u8 maxdepth; /* limit of packets per flow */
122
123 u32 perturbation;
124 u8 cur_depth; /* depth of longest slot */
125 u8 flags;
126 unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */
127 struct tcf_proto __rcu *filter_list;
128 sfq_index *ht; /* Hash table ('divisor' slots) */
129 struct sfq_slot *slots; /* Flows table ('maxflows' entries) */
130
131 struct red_parms *red_parms;
132 struct tc_sfqred_stats stats;
133 struct sfq_slot *tail; /* current slot in round */
134
135 struct sfq_head dep[SFQ_MAX_DEPTH + 1];
136 /* Linked lists of slots, indexed by depth
137 * dep[0] : list of unused flows
138 * dep[1] : list of flows with 1 packet
139 * dep[X] : list of flows with X packets
140 */
141
142 unsigned int maxflows; /* number of flows in flows array */
143 int perturb_period;
144 unsigned int quantum; /* Allotment per round: MUST BE >= MTU */
145 struct timer_list perturb_timer;
146 };
147
148 /*
149 * sfq_head are either in a sfq_slot or in dep[] array
150 */
151 static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val)
152 {
153 if (val < SFQ_MAX_FLOWS)
154 return &q->slots[val].dep;
155 return &q->dep[val - SFQ_MAX_FLOWS];
156 }
157
158 static unsigned int sfq_hash(const struct sfq_sched_data *q,
159 const struct sk_buff *skb)
160 {
161 return skb_get_hash_perturb(skb, q->perturbation) & (q->divisor - 1);
162 }
163
164 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
165 int *qerr)
166 {
167 struct sfq_sched_data *q = qdisc_priv(sch);
168 struct tcf_result res;
169 struct tcf_proto *fl;
170 int result;
171
172 if (TC_H_MAJ(skb->priority) == sch->handle &&
173 TC_H_MIN(skb->priority) > 0 &&
174 TC_H_MIN(skb->priority) <= q->divisor)
175 return TC_H_MIN(skb->priority);
176
177 fl = rcu_dereference_bh(q->filter_list);
178 if (!fl)
179 return sfq_hash(q, skb) + 1;
180
181 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
182 result = tc_classify(skb, fl, &res, false);
183 if (result >= 0) {
184 #ifdef CONFIG_NET_CLS_ACT
185 switch (result) {
186 case TC_ACT_STOLEN:
187 case TC_ACT_QUEUED:
188 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
189 case TC_ACT_SHOT:
190 return 0;
191 }
192 #endif
193 if (TC_H_MIN(res.classid) <= q->divisor)
194 return TC_H_MIN(res.classid);
195 }
196 return 0;
197 }
198
199 /*
200 * x : slot number [0 .. SFQ_MAX_FLOWS - 1]
201 */
202 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
203 {
204 sfq_index p, n;
205 struct sfq_slot *slot = &q->slots[x];
206 int qlen = slot->qlen;
207
208 p = qlen + SFQ_MAX_FLOWS;
209 n = q->dep[qlen].next;
210
211 slot->dep.next = n;
212 slot->dep.prev = p;
213
214 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */
215 sfq_dep_head(q, n)->prev = x;
216 }
217
218 #define sfq_unlink(q, x, n, p) \
219 do { \
220 n = q->slots[x].dep.next; \
221 p = q->slots[x].dep.prev; \
222 sfq_dep_head(q, p)->next = n; \
223 sfq_dep_head(q, n)->prev = p; \
224 } while (0)
225
226
227 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
228 {
229 sfq_index p, n;
230 int d;
231
232 sfq_unlink(q, x, n, p);
233
234 d = q->slots[x].qlen--;
235 if (n == p && q->cur_depth == d)
236 q->cur_depth--;
237 sfq_link(q, x);
238 }
239
240 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
241 {
242 sfq_index p, n;
243 int d;
244
245 sfq_unlink(q, x, n, p);
246
247 d = ++q->slots[x].qlen;
248 if (q->cur_depth < d)
249 q->cur_depth = d;
250 sfq_link(q, x);
251 }
252
253 /* helper functions : might be changed when/if skb use a standard list_head */
254
255 /* remove one skb from tail of slot queue */
256 static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot)
257 {
258 struct sk_buff *skb = slot->skblist_prev;
259
260 slot->skblist_prev = skb->prev;
261 skb->prev->next = (struct sk_buff *)slot;
262 skb->next = skb->prev = NULL;
263 return skb;
264 }
265
266 /* remove one skb from head of slot queue */
267 static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot)
268 {
269 struct sk_buff *skb = slot->skblist_next;
270
271 slot->skblist_next = skb->next;
272 skb->next->prev = (struct sk_buff *)slot;
273 skb->next = skb->prev = NULL;
274 return skb;
275 }
276
277 static inline void slot_queue_init(struct sfq_slot *slot)
278 {
279 memset(slot, 0, sizeof(*slot));
280 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot;
281 }
282
283 /* add skb to slot queue (tail add) */
284 static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb)
285 {
286 skb->prev = slot->skblist_prev;
287 skb->next = (struct sk_buff *)slot;
288 slot->skblist_prev->next = skb;
289 slot->skblist_prev = skb;
290 }
291
292 static unsigned int sfq_drop(struct Qdisc *sch)
293 {
294 struct sfq_sched_data *q = qdisc_priv(sch);
295 sfq_index x, d = q->cur_depth;
296 struct sk_buff *skb;
297 unsigned int len;
298 struct sfq_slot *slot;
299
300 /* Queue is full! Find the longest slot and drop tail packet from it */
301 if (d > 1) {
302 x = q->dep[d].next;
303 slot = &q->slots[x];
304 drop:
305 skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot);
306 len = qdisc_pkt_len(skb);
307 slot->backlog -= len;
308 sfq_dec(q, x);
309 sch->q.qlen--;
310 qdisc_qstats_drop(sch);
311 qdisc_qstats_backlog_dec(sch, skb);
312 kfree_skb(skb);
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 */
329 static 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 */
335 static 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
340 static int sfq_headdrop(const struct sfq_sched_data *q)
341 {
342 return q->headdrop;
343 }
344
345 static int
346 sfq_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 uninitialized_var(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 kfree_skb(skb);
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) {
425 congestion_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
441 enqueue:
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);
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
479 static struct sk_buff *
480 sfq_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
491 next_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
520 static void
521 sfq_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 */
535 static 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) {
570 drop:
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
605 static void sfq_perturbation(unsigned long arg)
606 {
607 struct Qdisc *sch = (struct Qdisc *)arg;
608 struct sfq_sched_data *q = qdisc_priv(sch);
609 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
610
611 spin_lock(root_lock);
612 q->perturbation = prandom_u32();
613 if (!q->filter_list && q->tail)
614 sfq_rehash(sch);
615 spin_unlock(root_lock);
616
617 if (q->perturb_period)
618 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
619 }
620
621 static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
622 {
623 struct sfq_sched_data *q = qdisc_priv(sch);
624 struct tc_sfq_qopt *ctl = nla_data(opt);
625 struct tc_sfq_qopt_v1 *ctl_v1 = NULL;
626 unsigned int qlen, dropped = 0;
627 struct red_parms *p = NULL;
628
629 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
630 return -EINVAL;
631 if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1)))
632 ctl_v1 = nla_data(opt);
633 if (ctl->divisor &&
634 (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536))
635 return -EINVAL;
636 if (ctl_v1 && !red_check_params(ctl_v1->qth_min, ctl_v1->qth_max,
637 ctl_v1->Wlog))
638 return -EINVAL;
639 if (ctl_v1 && ctl_v1->qth_min) {
640 p = kmalloc(sizeof(*p), GFP_KERNEL);
641 if (!p)
642 return -ENOMEM;
643 }
644 sch_tree_lock(sch);
645 if (ctl->quantum) {
646 q->quantum = ctl->quantum;
647 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
648 }
649 q->perturb_period = ctl->perturb_period * HZ;
650 if (ctl->flows)
651 q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS);
652 if (ctl->divisor) {
653 q->divisor = ctl->divisor;
654 q->maxflows = min_t(u32, q->maxflows, q->divisor);
655 }
656 if (ctl_v1) {
657 if (ctl_v1->depth)
658 q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH);
659 if (p) {
660 swap(q->red_parms, p);
661 red_set_parms(q->red_parms,
662 ctl_v1->qth_min, ctl_v1->qth_max,
663 ctl_v1->Wlog,
664 ctl_v1->Plog, ctl_v1->Scell_log,
665 NULL,
666 ctl_v1->max_P);
667 }
668 q->flags = ctl_v1->flags;
669 q->headdrop = ctl_v1->headdrop;
670 }
671 if (ctl->limit) {
672 q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows);
673 q->maxflows = min_t(u32, q->maxflows, q->limit);
674 }
675
676 qlen = sch->q.qlen;
677 while (sch->q.qlen > q->limit)
678 dropped += sfq_drop(sch);
679 qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
680
681 del_timer(&q->perturb_timer);
682 if (q->perturb_period) {
683 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
684 q->perturbation = prandom_u32();
685 }
686 sch_tree_unlock(sch);
687 kfree(p);
688 return 0;
689 }
690
691 static void *sfq_alloc(size_t sz)
692 {
693 void *ptr = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN);
694
695 if (!ptr)
696 ptr = vmalloc(sz);
697 return ptr;
698 }
699
700 static void sfq_free(void *addr)
701 {
702 kvfree(addr);
703 }
704
705 static void sfq_destroy(struct Qdisc *sch)
706 {
707 struct sfq_sched_data *q = qdisc_priv(sch);
708
709 tcf_destroy_chain(&q->filter_list);
710 q->perturb_period = 0;
711 del_timer_sync(&q->perturb_timer);
712 sfq_free(q->ht);
713 sfq_free(q->slots);
714 kfree(q->red_parms);
715 }
716
717 static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
718 {
719 struct sfq_sched_data *q = qdisc_priv(sch);
720 int i;
721
722 q->perturb_timer.function = sfq_perturbation;
723 q->perturb_timer.data = (unsigned long)sch;
724 init_timer_deferrable(&q->perturb_timer);
725
726 for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) {
727 q->dep[i].next = i + SFQ_MAX_FLOWS;
728 q->dep[i].prev = i + SFQ_MAX_FLOWS;
729 }
730
731 q->limit = SFQ_MAX_DEPTH;
732 q->maxdepth = SFQ_MAX_DEPTH;
733 q->cur_depth = 0;
734 q->tail = NULL;
735 q->divisor = SFQ_DEFAULT_HASH_DIVISOR;
736 q->maxflows = SFQ_DEFAULT_FLOWS;
737 q->quantum = psched_mtu(qdisc_dev(sch));
738 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
739 q->perturb_period = 0;
740 q->perturbation = prandom_u32();
741
742 if (opt) {
743 int err = sfq_change(sch, opt);
744 if (err)
745 return err;
746 }
747
748 q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor);
749 q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows);
750 if (!q->ht || !q->slots) {
751 /* Note: sfq_destroy() will be called by our caller */
752 return -ENOMEM;
753 }
754
755 for (i = 0; i < q->divisor; i++)
756 q->ht[i] = SFQ_EMPTY_SLOT;
757
758 for (i = 0; i < q->maxflows; i++) {
759 slot_queue_init(&q->slots[i]);
760 sfq_link(q, i);
761 }
762 if (q->limit >= 1)
763 sch->flags |= TCQ_F_CAN_BYPASS;
764 else
765 sch->flags &= ~TCQ_F_CAN_BYPASS;
766 return 0;
767 }
768
769 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
770 {
771 struct sfq_sched_data *q = qdisc_priv(sch);
772 unsigned char *b = skb_tail_pointer(skb);
773 struct tc_sfq_qopt_v1 opt;
774 struct red_parms *p = q->red_parms;
775
776 memset(&opt, 0, sizeof(opt));
777 opt.v0.quantum = q->quantum;
778 opt.v0.perturb_period = q->perturb_period / HZ;
779 opt.v0.limit = q->limit;
780 opt.v0.divisor = q->divisor;
781 opt.v0.flows = q->maxflows;
782 opt.depth = q->maxdepth;
783 opt.headdrop = q->headdrop;
784
785 if (p) {
786 opt.qth_min = p->qth_min >> p->Wlog;
787 opt.qth_max = p->qth_max >> p->Wlog;
788 opt.Wlog = p->Wlog;
789 opt.Plog = p->Plog;
790 opt.Scell_log = p->Scell_log;
791 opt.max_P = p->max_P;
792 }
793 memcpy(&opt.stats, &q->stats, sizeof(opt.stats));
794 opt.flags = q->flags;
795
796 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
797 goto nla_put_failure;
798
799 return skb->len;
800
801 nla_put_failure:
802 nlmsg_trim(skb, b);
803 return -1;
804 }
805
806 static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg)
807 {
808 return NULL;
809 }
810
811 static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
812 {
813 return 0;
814 }
815
816 static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent,
817 u32 classid)
818 {
819 /* we cannot bypass queue discipline anymore */
820 sch->flags &= ~TCQ_F_CAN_BYPASS;
821 return 0;
822 }
823
824 static void sfq_put(struct Qdisc *q, unsigned long cl)
825 {
826 }
827
828 static struct tcf_proto __rcu **sfq_find_tcf(struct Qdisc *sch,
829 unsigned long cl)
830 {
831 struct sfq_sched_data *q = qdisc_priv(sch);
832
833 if (cl)
834 return NULL;
835 return &q->filter_list;
836 }
837
838 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
839 struct sk_buff *skb, struct tcmsg *tcm)
840 {
841 tcm->tcm_handle |= TC_H_MIN(cl);
842 return 0;
843 }
844
845 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
846 struct gnet_dump *d)
847 {
848 struct sfq_sched_data *q = qdisc_priv(sch);
849 sfq_index idx = q->ht[cl - 1];
850 struct gnet_stats_queue qs = { 0 };
851 struct tc_sfq_xstats xstats = { 0 };
852
853 if (idx != SFQ_EMPTY_SLOT) {
854 const struct sfq_slot *slot = &q->slots[idx];
855
856 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT;
857 qs.qlen = slot->qlen;
858 qs.backlog = slot->backlog;
859 }
860 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
861 return -1;
862 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
863 }
864
865 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
866 {
867 struct sfq_sched_data *q = qdisc_priv(sch);
868 unsigned int i;
869
870 if (arg->stop)
871 return;
872
873 for (i = 0; i < q->divisor; i++) {
874 if (q->ht[i] == SFQ_EMPTY_SLOT ||
875 arg->count < arg->skip) {
876 arg->count++;
877 continue;
878 }
879 if (arg->fn(sch, i + 1, arg) < 0) {
880 arg->stop = 1;
881 break;
882 }
883 arg->count++;
884 }
885 }
886
887 static const struct Qdisc_class_ops sfq_class_ops = {
888 .leaf = sfq_leaf,
889 .get = sfq_get,
890 .put = sfq_put,
891 .tcf_chain = sfq_find_tcf,
892 .bind_tcf = sfq_bind,
893 .unbind_tcf = sfq_put,
894 .dump = sfq_dump_class,
895 .dump_stats = sfq_dump_class_stats,
896 .walk = sfq_walk,
897 };
898
899 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
900 .cl_ops = &sfq_class_ops,
901 .id = "sfq",
902 .priv_size = sizeof(struct sfq_sched_data),
903 .enqueue = sfq_enqueue,
904 .dequeue = sfq_dequeue,
905 .peek = qdisc_peek_dequeued,
906 .init = sfq_init,
907 .reset = sfq_reset,
908 .destroy = sfq_destroy,
909 .change = NULL,
910 .dump = sfq_dump,
911 .owner = THIS_MODULE,
912 };
913
914 static int __init sfq_module_init(void)
915 {
916 return register_qdisc(&sfq_qdisc_ops);
917 }
918 static void __exit sfq_module_exit(void)
919 {
920 unregister_qdisc(&sfq_qdisc_ops);
921 }
922 module_init(sfq_module_init)
923 module_exit(sfq_module_exit)
924 MODULE_LICENSE("GPL");
Linux with added FPC-III support