Commit | Line | Data |
---|---|---|
1da177e4 LT |
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 | ||
1da177e4 | 12 | #include <linux/module.h> |
1da177e4 LT |
13 | #include <linux/types.h> |
14 | #include <linux/kernel.h> | |
15 | #include <linux/jiffies.h> | |
16 | #include <linux/string.h> | |
1da177e4 LT |
17 | #include <linux/in.h> |
18 | #include <linux/errno.h> | |
1da177e4 | 19 | #include <linux/init.h> |
1da177e4 | 20 | #include <linux/ipv6.h> |
1da177e4 | 21 | #include <linux/skbuff.h> |
0ba48053 PM |
22 | #include <net/ip.h> |
23 | #include <net/netlink.h> | |
1da177e4 LT |
24 | #include <net/pkt_sched.h> |
25 | ||
26 | ||
27 | /* Stochastic Fairness Queuing algorithm. | |
28 | ======================================= | |
29 | ||
30 | Source: | |
31 | Paul E. McKenney "Stochastic Fairness Queuing", | |
32 | IEEE INFOCOMM'90 Proceedings, San Francisco, 1990. | |
33 | ||
34 | Paul E. McKenney "Stochastic Fairness Queuing", | |
35 | "Interworking: Research and Experience", v.2, 1991, p.113-131. | |
36 | ||
37 | ||
38 | See also: | |
39 | M. Shreedhar and George Varghese "Efficient Fair | |
40 | Queuing using Deficit Round Robin", Proc. SIGCOMM 95. | |
41 | ||
42 | ||
10297b99 | 43 | This is not the thing that is usually called (W)FQ nowadays. |
1da177e4 LT |
44 | It does not use any timestamp mechanism, but instead |
45 | processes queues in round-robin order. | |
46 | ||
47 | ADVANTAGE: | |
48 | ||
49 | - It is very cheap. Both CPU and memory requirements are minimal. | |
50 | ||
51 | DRAWBACKS: | |
52 | ||
10297b99 | 53 | - "Stochastic" -> It is not 100% fair. |
1da177e4 LT |
54 | When hash collisions occur, several flows are considered as one. |
55 | ||
56 | - "Round-robin" -> It introduces larger delays than virtual clock | |
57 | based schemes, and should not be used for isolating interactive | |
58 | traffic from non-interactive. It means, that this scheduler | |
59 | should be used as leaf of CBQ or P3, which put interactive traffic | |
60 | to higher priority band. | |
61 | ||
62 | We still need true WFQ for top level CSZ, but using WFQ | |
63 | for the best effort traffic is absolutely pointless: | |
64 | SFQ is superior for this purpose. | |
65 | ||
66 | IMPLEMENTATION: | |
67 | This implementation limits maximal queue length to 128; | |
68 | maximal mtu to 2^15-1; number of hash buckets to 1024. | |
69 | The only goal of this restrictions was that all data | |
70 | fit into one 4K page :-). Struct sfq_sched_data is | |
71 | organized in anti-cache manner: all the data for a bucket | |
72 | are scattered over different locations. This is not good, | |
73 | but it allowed me to put it into 4K. | |
74 | ||
75 | It is easy to increase these values, but not in flight. */ | |
76 | ||
77 | #define SFQ_DEPTH 128 | |
78 | #define SFQ_HASH_DIVISOR 1024 | |
79 | ||
80 | /* This type should contain at least SFQ_DEPTH*2 values */ | |
81 | typedef unsigned char sfq_index; | |
82 | ||
83 | struct sfq_head | |
84 | { | |
85 | sfq_index next; | |
86 | sfq_index prev; | |
87 | }; | |
88 | ||
89 | struct sfq_sched_data | |
90 | { | |
91 | /* Parameters */ | |
92 | int perturb_period; | |
93 | unsigned quantum; /* Allotment per round: MUST BE >= MTU */ | |
94 | int limit; | |
95 | ||
96 | /* Variables */ | |
97 | struct timer_list perturb_timer; | |
98 | int perturbation; | |
99 | sfq_index tail; /* Index of current slot in round */ | |
100 | sfq_index max_depth; /* Maximal depth */ | |
101 | ||
102 | sfq_index ht[SFQ_HASH_DIVISOR]; /* Hash table */ | |
103 | sfq_index next[SFQ_DEPTH]; /* Active slots link */ | |
104 | short allot[SFQ_DEPTH]; /* Current allotment per slot */ | |
105 | unsigned short hash[SFQ_DEPTH]; /* Hash value indexed by slots */ | |
106 | struct sk_buff_head qs[SFQ_DEPTH]; /* Slot queue */ | |
107 | struct sfq_head dep[SFQ_DEPTH*2]; /* Linked list of slots, indexed by depth */ | |
108 | }; | |
109 | ||
110 | static __inline__ unsigned sfq_fold_hash(struct sfq_sched_data *q, u32 h, u32 h1) | |
111 | { | |
112 | int pert = q->perturbation; | |
113 | ||
114 | /* Have we any rotation primitives? If not, WHY? */ | |
115 | h ^= (h1<<pert) ^ (h1>>(0x1F - pert)); | |
116 | h ^= h>>10; | |
117 | return h & 0x3FF; | |
118 | } | |
119 | ||
120 | static unsigned sfq_hash(struct sfq_sched_data *q, struct sk_buff *skb) | |
121 | { | |
122 | u32 h, h2; | |
123 | ||
124 | switch (skb->protocol) { | |
125 | case __constant_htons(ETH_P_IP): | |
126 | { | |
eddc9ec5 | 127 | const struct iphdr *iph = ip_hdr(skb); |
1da177e4 LT |
128 | h = iph->daddr; |
129 | h2 = iph->saddr^iph->protocol; | |
130 | if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) && | |
131 | (iph->protocol == IPPROTO_TCP || | |
132 | iph->protocol == IPPROTO_UDP || | |
a8d0f952 | 133 | iph->protocol == IPPROTO_UDPLITE || |
ae82af54 PM |
134 | iph->protocol == IPPROTO_SCTP || |
135 | iph->protocol == IPPROTO_DCCP || | |
1da177e4 LT |
136 | iph->protocol == IPPROTO_ESP)) |
137 | h2 ^= *(((u32*)iph) + iph->ihl); | |
138 | break; | |
139 | } | |
140 | case __constant_htons(ETH_P_IPV6): | |
141 | { | |
0660e03f | 142 | struct ipv6hdr *iph = ipv6_hdr(skb); |
1da177e4 LT |
143 | h = iph->daddr.s6_addr32[3]; |
144 | h2 = iph->saddr.s6_addr32[3]^iph->nexthdr; | |
145 | if (iph->nexthdr == IPPROTO_TCP || | |
146 | iph->nexthdr == IPPROTO_UDP || | |
a8d0f952 | 147 | iph->nexthdr == IPPROTO_UDPLITE || |
ae82af54 PM |
148 | iph->nexthdr == IPPROTO_SCTP || |
149 | iph->nexthdr == IPPROTO_DCCP || | |
1da177e4 LT |
150 | iph->nexthdr == IPPROTO_ESP) |
151 | h2 ^= *(u32*)&iph[1]; | |
152 | break; | |
153 | } | |
154 | default: | |
155 | h = (u32)(unsigned long)skb->dst^skb->protocol; | |
156 | h2 = (u32)(unsigned long)skb->sk; | |
157 | } | |
158 | return sfq_fold_hash(q, h, h2); | |
159 | } | |
160 | ||
161 | static inline void sfq_link(struct sfq_sched_data *q, sfq_index x) | |
162 | { | |
163 | sfq_index p, n; | |
164 | int d = q->qs[x].qlen + SFQ_DEPTH; | |
165 | ||
166 | p = d; | |
167 | n = q->dep[d].next; | |
168 | q->dep[x].next = n; | |
169 | q->dep[x].prev = p; | |
170 | q->dep[p].next = q->dep[n].prev = x; | |
171 | } | |
172 | ||
173 | static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x) | |
174 | { | |
175 | sfq_index p, n; | |
176 | ||
177 | n = q->dep[x].next; | |
178 | p = q->dep[x].prev; | |
179 | q->dep[p].next = n; | |
180 | q->dep[n].prev = p; | |
181 | ||
182 | if (n == p && q->max_depth == q->qs[x].qlen + 1) | |
183 | q->max_depth--; | |
184 | ||
185 | sfq_link(q, x); | |
186 | } | |
187 | ||
188 | static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x) | |
189 | { | |
190 | sfq_index p, n; | |
191 | int d; | |
192 | ||
193 | n = q->dep[x].next; | |
194 | p = q->dep[x].prev; | |
195 | q->dep[p].next = n; | |
196 | q->dep[n].prev = p; | |
197 | d = q->qs[x].qlen; | |
198 | if (q->max_depth < d) | |
199 | q->max_depth = d; | |
200 | ||
201 | sfq_link(q, x); | |
202 | } | |
203 | ||
204 | static unsigned int sfq_drop(struct Qdisc *sch) | |
205 | { | |
206 | struct sfq_sched_data *q = qdisc_priv(sch); | |
207 | sfq_index d = q->max_depth; | |
208 | struct sk_buff *skb; | |
209 | unsigned int len; | |
210 | ||
211 | /* Queue is full! Find the longest slot and | |
212 | drop a packet from it */ | |
213 | ||
214 | if (d > 1) { | |
215 | sfq_index x = q->dep[d+SFQ_DEPTH].next; | |
216 | skb = q->qs[x].prev; | |
217 | len = skb->len; | |
218 | __skb_unlink(skb, &q->qs[x]); | |
219 | kfree_skb(skb); | |
220 | sfq_dec(q, x); | |
221 | sch->q.qlen--; | |
222 | sch->qstats.drops++; | |
f5539eb8 | 223 | sch->qstats.backlog -= len; |
1da177e4 LT |
224 | return len; |
225 | } | |
226 | ||
227 | if (d == 1) { | |
228 | /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */ | |
229 | d = q->next[q->tail]; | |
230 | q->next[q->tail] = q->next[d]; | |
231 | q->allot[q->next[d]] += q->quantum; | |
232 | skb = q->qs[d].prev; | |
233 | len = skb->len; | |
234 | __skb_unlink(skb, &q->qs[d]); | |
235 | kfree_skb(skb); | |
236 | sfq_dec(q, d); | |
237 | sch->q.qlen--; | |
238 | q->ht[q->hash[d]] = SFQ_DEPTH; | |
239 | sch->qstats.drops++; | |
f5539eb8 | 240 | sch->qstats.backlog -= len; |
1da177e4 LT |
241 | return len; |
242 | } | |
243 | ||
244 | return 0; | |
245 | } | |
246 | ||
247 | static int | |
248 | sfq_enqueue(struct sk_buff *skb, struct Qdisc* sch) | |
249 | { | |
250 | struct sfq_sched_data *q = qdisc_priv(sch); | |
251 | unsigned hash = sfq_hash(q, skb); | |
252 | sfq_index x; | |
253 | ||
254 | x = q->ht[hash]; | |
255 | if (x == SFQ_DEPTH) { | |
256 | q->ht[hash] = x = q->dep[SFQ_DEPTH].next; | |
257 | q->hash[x] = hash; | |
258 | } | |
f5539eb8 | 259 | sch->qstats.backlog += skb->len; |
1da177e4 LT |
260 | __skb_queue_tail(&q->qs[x], skb); |
261 | sfq_inc(q, x); | |
262 | if (q->qs[x].qlen == 1) { /* The flow is new */ | |
263 | if (q->tail == SFQ_DEPTH) { /* It is the first flow */ | |
264 | q->tail = x; | |
265 | q->next[x] = x; | |
266 | q->allot[x] = q->quantum; | |
267 | } else { | |
268 | q->next[x] = q->next[q->tail]; | |
269 | q->next[q->tail] = x; | |
270 | q->tail = x; | |
271 | } | |
272 | } | |
273 | if (++sch->q.qlen < q->limit-1) { | |
274 | sch->bstats.bytes += skb->len; | |
275 | sch->bstats.packets++; | |
276 | return 0; | |
277 | } | |
278 | ||
279 | sfq_drop(sch); | |
280 | return NET_XMIT_CN; | |
281 | } | |
282 | ||
283 | static int | |
284 | sfq_requeue(struct sk_buff *skb, struct Qdisc* sch) | |
285 | { | |
286 | struct sfq_sched_data *q = qdisc_priv(sch); | |
287 | unsigned hash = sfq_hash(q, skb); | |
288 | sfq_index x; | |
289 | ||
290 | x = q->ht[hash]; | |
291 | if (x == SFQ_DEPTH) { | |
292 | q->ht[hash] = x = q->dep[SFQ_DEPTH].next; | |
293 | q->hash[x] = hash; | |
294 | } | |
f5539eb8 | 295 | sch->qstats.backlog += skb->len; |
1da177e4 LT |
296 | __skb_queue_head(&q->qs[x], skb); |
297 | sfq_inc(q, x); | |
298 | if (q->qs[x].qlen == 1) { /* The flow is new */ | |
299 | if (q->tail == SFQ_DEPTH) { /* It is the first flow */ | |
300 | q->tail = x; | |
301 | q->next[x] = x; | |
302 | q->allot[x] = q->quantum; | |
303 | } else { | |
304 | q->next[x] = q->next[q->tail]; | |
305 | q->next[q->tail] = x; | |
306 | q->tail = x; | |
307 | } | |
308 | } | |
309 | if (++sch->q.qlen < q->limit - 1) { | |
310 | sch->qstats.requeues++; | |
311 | return 0; | |
312 | } | |
313 | ||
314 | sch->qstats.drops++; | |
315 | sfq_drop(sch); | |
316 | return NET_XMIT_CN; | |
317 | } | |
318 | ||
319 | ||
320 | ||
321 | ||
322 | static struct sk_buff * | |
323 | sfq_dequeue(struct Qdisc* sch) | |
324 | { | |
325 | struct sfq_sched_data *q = qdisc_priv(sch); | |
326 | struct sk_buff *skb; | |
327 | sfq_index a, old_a; | |
328 | ||
329 | /* No active slots */ | |
330 | if (q->tail == SFQ_DEPTH) | |
331 | return NULL; | |
332 | ||
333 | a = old_a = q->next[q->tail]; | |
334 | ||
335 | /* Grab packet */ | |
336 | skb = __skb_dequeue(&q->qs[a]); | |
337 | sfq_dec(q, a); | |
338 | sch->q.qlen--; | |
f5539eb8 | 339 | sch->qstats.backlog -= skb->len; |
1da177e4 LT |
340 | |
341 | /* Is the slot empty? */ | |
342 | if (q->qs[a].qlen == 0) { | |
343 | q->ht[q->hash[a]] = SFQ_DEPTH; | |
344 | a = q->next[a]; | |
345 | if (a == old_a) { | |
346 | q->tail = SFQ_DEPTH; | |
347 | return skb; | |
348 | } | |
349 | q->next[q->tail] = a; | |
350 | q->allot[a] += q->quantum; | |
351 | } else if ((q->allot[a] -= skb->len) <= 0) { | |
352 | q->tail = a; | |
353 | a = q->next[a]; | |
354 | q->allot[a] += q->quantum; | |
355 | } | |
356 | return skb; | |
357 | } | |
358 | ||
359 | static void | |
360 | sfq_reset(struct Qdisc* sch) | |
361 | { | |
362 | struct sk_buff *skb; | |
363 | ||
364 | while ((skb = sfq_dequeue(sch)) != NULL) | |
365 | kfree_skb(skb); | |
366 | } | |
367 | ||
368 | static void sfq_perturbation(unsigned long arg) | |
369 | { | |
370 | struct Qdisc *sch = (struct Qdisc*)arg; | |
371 | struct sfq_sched_data *q = qdisc_priv(sch); | |
372 | ||
373 | q->perturbation = net_random()&0x1F; | |
374 | ||
375 | if (q->perturb_period) { | |
376 | q->perturb_timer.expires = jiffies + q->perturb_period; | |
377 | add_timer(&q->perturb_timer); | |
378 | } | |
379 | } | |
380 | ||
381 | static int sfq_change(struct Qdisc *sch, struct rtattr *opt) | |
382 | { | |
383 | struct sfq_sched_data *q = qdisc_priv(sch); | |
384 | struct tc_sfq_qopt *ctl = RTA_DATA(opt); | |
5e50da01 | 385 | unsigned int qlen; |
1da177e4 LT |
386 | |
387 | if (opt->rta_len < RTA_LENGTH(sizeof(*ctl))) | |
388 | return -EINVAL; | |
389 | ||
390 | sch_tree_lock(sch); | |
391 | q->quantum = ctl->quantum ? : psched_mtu(sch->dev); | |
392 | q->perturb_period = ctl->perturb_period*HZ; | |
393 | if (ctl->limit) | |
394 | q->limit = min_t(u32, ctl->limit, SFQ_DEPTH); | |
395 | ||
5e50da01 | 396 | qlen = sch->q.qlen; |
1da177e4 LT |
397 | while (sch->q.qlen >= q->limit-1) |
398 | sfq_drop(sch); | |
5e50da01 | 399 | qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen); |
1da177e4 LT |
400 | |
401 | del_timer(&q->perturb_timer); | |
402 | if (q->perturb_period) { | |
403 | q->perturb_timer.expires = jiffies + q->perturb_period; | |
404 | add_timer(&q->perturb_timer); | |
405 | } | |
406 | sch_tree_unlock(sch); | |
407 | return 0; | |
408 | } | |
409 | ||
410 | static int sfq_init(struct Qdisc *sch, struct rtattr *opt) | |
411 | { | |
412 | struct sfq_sched_data *q = qdisc_priv(sch); | |
413 | int i; | |
414 | ||
415 | init_timer(&q->perturb_timer); | |
416 | q->perturb_timer.data = (unsigned long)sch; | |
417 | q->perturb_timer.function = sfq_perturbation; | |
418 | ||
419 | for (i=0; i<SFQ_HASH_DIVISOR; i++) | |
420 | q->ht[i] = SFQ_DEPTH; | |
421 | for (i=0; i<SFQ_DEPTH; i++) { | |
422 | skb_queue_head_init(&q->qs[i]); | |
423 | q->dep[i+SFQ_DEPTH].next = i+SFQ_DEPTH; | |
424 | q->dep[i+SFQ_DEPTH].prev = i+SFQ_DEPTH; | |
425 | } | |
426 | q->limit = SFQ_DEPTH; | |
427 | q->max_depth = 0; | |
428 | q->tail = SFQ_DEPTH; | |
429 | if (opt == NULL) { | |
430 | q->quantum = psched_mtu(sch->dev); | |
431 | q->perturb_period = 0; | |
432 | } else { | |
433 | int err = sfq_change(sch, opt); | |
434 | if (err) | |
435 | return err; | |
436 | } | |
437 | for (i=0; i<SFQ_DEPTH; i++) | |
438 | sfq_link(q, i); | |
439 | return 0; | |
440 | } | |
441 | ||
442 | static void sfq_destroy(struct Qdisc *sch) | |
443 | { | |
444 | struct sfq_sched_data *q = qdisc_priv(sch); | |
445 | del_timer(&q->perturb_timer); | |
446 | } | |
447 | ||
448 | static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb) | |
449 | { | |
450 | struct sfq_sched_data *q = qdisc_priv(sch); | |
27a884dc | 451 | unsigned char *b = skb_tail_pointer(skb); |
1da177e4 LT |
452 | struct tc_sfq_qopt opt; |
453 | ||
454 | opt.quantum = q->quantum; | |
455 | opt.perturb_period = q->perturb_period/HZ; | |
456 | ||
457 | opt.limit = q->limit; | |
458 | opt.divisor = SFQ_HASH_DIVISOR; | |
459 | opt.flows = q->limit; | |
460 | ||
461 | RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt); | |
462 | ||
463 | return skb->len; | |
464 | ||
465 | rtattr_failure: | |
dc5fc579 | 466 | nlmsg_trim(skb, b); |
1da177e4 LT |
467 | return -1; |
468 | } | |
469 | ||
470 | static struct Qdisc_ops sfq_qdisc_ops = { | |
471 | .next = NULL, | |
472 | .cl_ops = NULL, | |
473 | .id = "sfq", | |
474 | .priv_size = sizeof(struct sfq_sched_data), | |
475 | .enqueue = sfq_enqueue, | |
476 | .dequeue = sfq_dequeue, | |
477 | .requeue = sfq_requeue, | |
478 | .drop = sfq_drop, | |
479 | .init = sfq_init, | |
480 | .reset = sfq_reset, | |
481 | .destroy = sfq_destroy, | |
482 | .change = NULL, | |
483 | .dump = sfq_dump, | |
484 | .owner = THIS_MODULE, | |
485 | }; | |
486 | ||
487 | static int __init sfq_module_init(void) | |
488 | { | |
489 | return register_qdisc(&sfq_qdisc_ops); | |
490 | } | |
10297b99 | 491 | static void __exit sfq_module_exit(void) |
1da177e4 LT |
492 | { |
493 | unregister_qdisc(&sfq_qdisc_ops); | |
494 | } | |
495 | module_init(sfq_module_init) | |
496 | module_exit(sfq_module_exit) | |
497 | MODULE_LICENSE("GPL"); |