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/ipv6.h> 21 #include <linux/skbuff.h> 22 #include <linux/jhash.h> 23 #include <linux/slab.h> 24 #include <net/ip.h> 25 #include <net/netlink.h> 26 #include <net/pkt_sched.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 maximal queue length to 128; 70 maximal mtu to 2^15-1; number of hash buckets to 1024. 71 The only goal of this restrictions was that all data 72 fit into one 4K page :-). Struct sfq_sched_data is 73 organized in anti-cache manner: all the data for a bucket 74 are scattered over different locations. This is not good, 75 but it allowed me to put it into 4K. 76 77 It is easy to increase these values, but not in flight. */ 78 79 #define SFQ_DEPTH 128 80 #define SFQ_HASH_DIVISOR 1024 81 82 /* This type should contain at least SFQ_DEPTH*2 values */ 83 typedef unsigned char sfq_index; 84 85 struct sfq_head 86 { 87 sfq_index next; 88 sfq_index prev; 89 }; 90 91 struct sfq_sched_data 92 { 93 /* Parameters */ 94 int perturb_period; 95 unsigned quantum; /* Allotment per round: MUST BE >= MTU */ 96 int limit; 97 98 /* Variables */ 99 struct tcf_proto *filter_list; 100 struct timer_list perturb_timer; 101 u32 perturbation; 102 sfq_index tail; /* Index of current slot in round */ 103 sfq_index max_depth; /* Maximal depth */ 104 105 sfq_index ht[SFQ_HASH_DIVISOR]; /* Hash table */ 106 sfq_index next[SFQ_DEPTH]; /* Active slots link */ 107 short allot[SFQ_DEPTH]; /* Current allotment per slot */ 108 unsigned short hash[SFQ_DEPTH]; /* Hash value indexed by slots */ 109 struct sk_buff_head qs[SFQ_DEPTH]; /* Slot queue */ 110 struct sfq_head dep[SFQ_DEPTH*2]; /* Linked list of slots, indexed by depth */ 111 }; 112 113 static __inline__ unsigned sfq_fold_hash(struct sfq_sched_data *q, u32 h, u32 h1) 114 { 115 return jhash_2words(h, h1, q->perturbation) & (SFQ_HASH_DIVISOR - 1); 116 } 117 118 static unsigned sfq_hash(struct sfq_sched_data *q, struct sk_buff *skb) 119 { 120 u32 h, h2; 121 122 switch (skb->protocol) { 123 case htons(ETH_P_IP): 124 { 125 const struct iphdr *iph = ip_hdr(skb); 126 h = (__force u32)iph->daddr; 127 h2 = (__force u32)iph->saddr ^ iph->protocol; 128 if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) && 129 (iph->protocol == IPPROTO_TCP || 130 iph->protocol == IPPROTO_UDP || 131 iph->protocol == IPPROTO_UDPLITE || 132 iph->protocol == IPPROTO_SCTP || 133 iph->protocol == IPPROTO_DCCP || 134 iph->protocol == IPPROTO_ESP)) 135 h2 ^= *(((u32*)iph) + iph->ihl); 136 break; 137 } 138 case htons(ETH_P_IPV6): 139 { 140 struct ipv6hdr *iph = ipv6_hdr(skb); 141 h = (__force u32)iph->daddr.s6_addr32[3]; 142 h2 = (__force u32)iph->saddr.s6_addr32[3] ^ iph->nexthdr; 143 if (iph->nexthdr == IPPROTO_TCP || 144 iph->nexthdr == IPPROTO_UDP || 145 iph->nexthdr == IPPROTO_UDPLITE || 146 iph->nexthdr == IPPROTO_SCTP || 147 iph->nexthdr == IPPROTO_DCCP || 148 iph->nexthdr == IPPROTO_ESP) 149 h2 ^= *(u32*)&iph[1]; 150 break; 151 } 152 default: 153 h = (unsigned long)skb_dst(skb) ^ (__force u32)skb->protocol; 154 h2 = (unsigned long)skb->sk; 155 } 156 157 return sfq_fold_hash(q, h, h2); 158 } 159 160 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch, 161 int *qerr) 162 { 163 struct sfq_sched_data *q = qdisc_priv(sch); 164 struct tcf_result res; 165 int result; 166 167 if (TC_H_MAJ(skb->priority) == sch->handle && 168 TC_H_MIN(skb->priority) > 0 && 169 TC_H_MIN(skb->priority) <= SFQ_HASH_DIVISOR) 170 return TC_H_MIN(skb->priority); 171 172 if (!q->filter_list) 173 return sfq_hash(q, skb) + 1; 174 175 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 176 result = tc_classify(skb, q->filter_list, &res); 177 if (result >= 0) { 178 #ifdef CONFIG_NET_CLS_ACT 179 switch (result) { 180 case TC_ACT_STOLEN: 181 case TC_ACT_QUEUED: 182 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; 183 case TC_ACT_SHOT: 184 return 0; 185 } 186 #endif 187 if (TC_H_MIN(res.classid) <= SFQ_HASH_DIVISOR) 188 return TC_H_MIN(res.classid); 189 } 190 return 0; 191 } 192 193 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x) 194 { 195 sfq_index p, n; 196 int d = q->qs[x].qlen + SFQ_DEPTH; 197 198 p = d; 199 n = q->dep[d].next; 200 q->dep[x].next = n; 201 q->dep[x].prev = p; 202 q->dep[p].next = q->dep[n].prev = x; 203 } 204 205 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x) 206 { 207 sfq_index p, n; 208 209 n = q->dep[x].next; 210 p = q->dep[x].prev; 211 q->dep[p].next = n; 212 q->dep[n].prev = p; 213 214 if (n == p && q->max_depth == q->qs[x].qlen + 1) 215 q->max_depth--; 216 217 sfq_link(q, x); 218 } 219 220 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x) 221 { 222 sfq_index p, n; 223 int d; 224 225 n = q->dep[x].next; 226 p = q->dep[x].prev; 227 q->dep[p].next = n; 228 q->dep[n].prev = p; 229 d = q->qs[x].qlen; 230 if (q->max_depth < d) 231 q->max_depth = d; 232 233 sfq_link(q, x); 234 } 235 236 static unsigned int sfq_drop(struct Qdisc *sch) 237 { 238 struct sfq_sched_data *q = qdisc_priv(sch); 239 sfq_index d = q->max_depth; 240 struct sk_buff *skb; 241 unsigned int len; 242 243 /* Queue is full! Find the longest slot and 244 drop a packet from it */ 245 246 if (d > 1) { 247 sfq_index x = q->dep[d + SFQ_DEPTH].next; 248 skb = q->qs[x].prev; 249 len = qdisc_pkt_len(skb); 250 __skb_unlink(skb, &q->qs[x]); 251 kfree_skb(skb); 252 sfq_dec(q, x); 253 sch->q.qlen--; 254 sch->qstats.drops++; 255 sch->qstats.backlog -= len; 256 return len; 257 } 258 259 if (d == 1) { 260 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */ 261 d = q->next[q->tail]; 262 q->next[q->tail] = q->next[d]; 263 q->allot[q->next[d]] += q->quantum; 264 skb = q->qs[d].prev; 265 len = qdisc_pkt_len(skb); 266 __skb_unlink(skb, &q->qs[d]); 267 kfree_skb(skb); 268 sfq_dec(q, d); 269 sch->q.qlen--; 270 q->ht[q->hash[d]] = SFQ_DEPTH; 271 sch->qstats.drops++; 272 sch->qstats.backlog -= len; 273 return len; 274 } 275 276 return 0; 277 } 278 279 static int 280 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch) 281 { 282 struct sfq_sched_data *q = qdisc_priv(sch); 283 unsigned int hash; 284 sfq_index x; 285 int uninitialized_var(ret); 286 287 hash = sfq_classify(skb, sch, &ret); 288 if (hash == 0) { 289 if (ret & __NET_XMIT_BYPASS) 290 sch->qstats.drops++; 291 kfree_skb(skb); 292 return ret; 293 } 294 hash--; 295 296 x = q->ht[hash]; 297 if (x == SFQ_DEPTH) { 298 q->ht[hash] = x = q->dep[SFQ_DEPTH].next; 299 q->hash[x] = hash; 300 } 301 302 /* If selected queue has length q->limit, this means that 303 * all another queues are empty and that we do simple tail drop, 304 * i.e. drop _this_ packet. 305 */ 306 if (q->qs[x].qlen >= q->limit) 307 return qdisc_drop(skb, sch); 308 309 sch->qstats.backlog += qdisc_pkt_len(skb); 310 __skb_queue_tail(&q->qs[x], skb); 311 sfq_inc(q, x); 312 if (q->qs[x].qlen == 1) { /* The flow is new */ 313 if (q->tail == SFQ_DEPTH) { /* It is the first flow */ 314 q->tail = x; 315 q->next[x] = x; 316 q->allot[x] = q->quantum; 317 } else { 318 q->next[x] = q->next[q->tail]; 319 q->next[q->tail] = x; 320 q->tail = x; 321 } 322 } 323 if (++sch->q.qlen <= q->limit) { 324 sch->bstats.bytes += qdisc_pkt_len(skb); 325 sch->bstats.packets++; 326 return 0; 327 } 328 329 sfq_drop(sch); 330 return NET_XMIT_CN; 331 } 332 333 static struct sk_buff * 334 sfq_peek(struct Qdisc *sch) 335 { 336 struct sfq_sched_data *q = qdisc_priv(sch); 337 sfq_index a; 338 339 /* No active slots */ 340 if (q->tail == SFQ_DEPTH) 341 return NULL; 342 343 a = q->next[q->tail]; 344 return skb_peek(&q->qs[a]); 345 } 346 347 static struct sk_buff * 348 sfq_dequeue(struct Qdisc *sch) 349 { 350 struct sfq_sched_data *q = qdisc_priv(sch); 351 struct sk_buff *skb; 352 sfq_index a, old_a; 353 354 /* No active slots */ 355 if (q->tail == SFQ_DEPTH) 356 return NULL; 357 358 a = old_a = q->next[q->tail]; 359 360 /* Grab packet */ 361 skb = __skb_dequeue(&q->qs[a]); 362 sfq_dec(q, a); 363 sch->q.qlen--; 364 sch->qstats.backlog -= qdisc_pkt_len(skb); 365 366 /* Is the slot empty? */ 367 if (q->qs[a].qlen == 0) { 368 q->ht[q->hash[a]] = SFQ_DEPTH; 369 a = q->next[a]; 370 if (a == old_a) { 371 q->tail = SFQ_DEPTH; 372 return skb; 373 } 374 q->next[q->tail] = a; 375 q->allot[a] += q->quantum; 376 } else if ((q->allot[a] -= qdisc_pkt_len(skb)) <= 0) { 377 q->tail = a; 378 a = q->next[a]; 379 q->allot[a] += q->quantum; 380 } 381 return skb; 382 } 383 384 static void 385 sfq_reset(struct Qdisc *sch) 386 { 387 struct sk_buff *skb; 388 389 while ((skb = sfq_dequeue(sch)) != NULL) 390 kfree_skb(skb); 391 } 392 393 static void sfq_perturbation(unsigned long arg) 394 { 395 struct Qdisc *sch = (struct Qdisc *)arg; 396 struct sfq_sched_data *q = qdisc_priv(sch); 397 398 q->perturbation = net_random(); 399 400 if (q->perturb_period) 401 mod_timer(&q->perturb_timer, jiffies + q->perturb_period); 402 } 403 404 static int sfq_change(struct Qdisc *sch, struct nlattr *opt) 405 { 406 struct sfq_sched_data *q = qdisc_priv(sch); 407 struct tc_sfq_qopt *ctl = nla_data(opt); 408 unsigned int qlen; 409 410 if (opt->nla_len < nla_attr_size(sizeof(*ctl))) 411 return -EINVAL; 412 413 sch_tree_lock(sch); 414 q->quantum = ctl->quantum ? : psched_mtu(qdisc_dev(sch)); 415 q->perturb_period = ctl->perturb_period * HZ; 416 if (ctl->limit) 417 q->limit = min_t(u32, ctl->limit, SFQ_DEPTH - 1); 418 419 qlen = sch->q.qlen; 420 while (sch->q.qlen > q->limit) 421 sfq_drop(sch); 422 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen); 423 424 del_timer(&q->perturb_timer); 425 if (q->perturb_period) { 426 mod_timer(&q->perturb_timer, jiffies + q->perturb_period); 427 q->perturbation = net_random(); 428 } 429 sch_tree_unlock(sch); 430 return 0; 431 } 432 433 static int sfq_init(struct Qdisc *sch, struct nlattr *opt) 434 { 435 struct sfq_sched_data *q = qdisc_priv(sch); 436 int i; 437 438 q->perturb_timer.function = sfq_perturbation; 439 q->perturb_timer.data = (unsigned long)sch; 440 init_timer_deferrable(&q->perturb_timer); 441 442 for (i = 0; i < SFQ_HASH_DIVISOR; i++) 443 q->ht[i] = SFQ_DEPTH; 444 445 for (i = 0; i < SFQ_DEPTH; i++) { 446 skb_queue_head_init(&q->qs[i]); 447 q->dep[i + SFQ_DEPTH].next = i + SFQ_DEPTH; 448 q->dep[i + SFQ_DEPTH].prev = i + SFQ_DEPTH; 449 } 450 451 q->limit = SFQ_DEPTH - 1; 452 q->max_depth = 0; 453 q->tail = SFQ_DEPTH; 454 if (opt == NULL) { 455 q->quantum = psched_mtu(qdisc_dev(sch)); 456 q->perturb_period = 0; 457 q->perturbation = net_random(); 458 } else { 459 int err = sfq_change(sch, opt); 460 if (err) 461 return err; 462 } 463 464 for (i = 0; i < SFQ_DEPTH; i++) 465 sfq_link(q, i); 466 return 0; 467 } 468 469 static void sfq_destroy(struct Qdisc *sch) 470 { 471 struct sfq_sched_data *q = qdisc_priv(sch); 472 473 tcf_destroy_chain(&q->filter_list); 474 q->perturb_period = 0; 475 del_timer_sync(&q->perturb_timer); 476 } 477 478 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb) 479 { 480 struct sfq_sched_data *q = qdisc_priv(sch); 481 unsigned char *b = skb_tail_pointer(skb); 482 struct tc_sfq_qopt opt; 483 484 opt.quantum = q->quantum; 485 opt.perturb_period = q->perturb_period / HZ; 486 487 opt.limit = q->limit; 488 opt.divisor = SFQ_HASH_DIVISOR; 489 opt.flows = q->limit; 490 491 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt); 492 493 return skb->len; 494 495 nla_put_failure: 496 nlmsg_trim(skb, b); 497 return -1; 498 } 499 500 static unsigned long sfq_get(struct Qdisc *sch, u32 classid) 501 { 502 return 0; 503 } 504 505 static struct tcf_proto **sfq_find_tcf(struct Qdisc *sch, unsigned long cl) 506 { 507 struct sfq_sched_data *q = qdisc_priv(sch); 508 509 if (cl) 510 return NULL; 511 return &q->filter_list; 512 } 513 514 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl, 515 struct sk_buff *skb, struct tcmsg *tcm) 516 { 517 tcm->tcm_handle |= TC_H_MIN(cl); 518 return 0; 519 } 520 521 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl, 522 struct gnet_dump *d) 523 { 524 struct sfq_sched_data *q = qdisc_priv(sch); 525 sfq_index idx = q->ht[cl-1]; 526 struct gnet_stats_queue qs = { .qlen = q->qs[idx].qlen }; 527 struct tc_sfq_xstats xstats = { .allot = q->allot[idx] }; 528 529 if (gnet_stats_copy_queue(d, &qs) < 0) 530 return -1; 531 return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); 532 } 533 534 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) 535 { 536 struct sfq_sched_data *q = qdisc_priv(sch); 537 unsigned int i; 538 539 if (arg->stop) 540 return; 541 542 for (i = 0; i < SFQ_HASH_DIVISOR; i++) { 543 if (q->ht[i] == SFQ_DEPTH || 544 arg->count < arg->skip) { 545 arg->count++; 546 continue; 547 } 548 if (arg->fn(sch, i + 1, arg) < 0) { 549 arg->stop = 1; 550 break; 551 } 552 arg->count++; 553 } 554 } 555 556 static const struct Qdisc_class_ops sfq_class_ops = { 557 .get = sfq_get, 558 .tcf_chain = sfq_find_tcf, 559 .dump = sfq_dump_class, 560 .dump_stats = sfq_dump_class_stats, 561 .walk = sfq_walk, 562 }; 563 564 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = { 565 .cl_ops = &sfq_class_ops, 566 .id = "sfq", 567 .priv_size = sizeof(struct sfq_sched_data), 568 .enqueue = sfq_enqueue, 569 .dequeue = sfq_dequeue, 570 .peek = sfq_peek, 571 .drop = sfq_drop, 572 .init = sfq_init, 573 .reset = sfq_reset, 574 .destroy = sfq_destroy, 575 .change = NULL, 576 .dump = sfq_dump, 577 .owner = THIS_MODULE, 578 }; 579 580 static int __init sfq_module_init(void) 581 { 582 return register_qdisc(&sfq_qdisc_ops); 583 } 584 static void __exit sfq_module_exit(void) 585 { 586 unregister_qdisc(&sfq_qdisc_ops); 587 } 588 module_init(sfq_module_init) 589 module_exit(sfq_module_exit) 590 MODULE_LICENSE("GPL"); 591