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/pkt_cls.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 */ 91 typedef 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 */ 99 struct sfq_head { 100 sfq_index next; 101 sfq_index prev; 102 }; 103 104 struct 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 }; 116 117 struct 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 __rcu *filter_list; 129 struct tcf_block *block; 130 sfq_index *ht; /* Hash table ('divisor' slots) */ 131 struct sfq_slot *slots; /* Flows table ('maxflows' entries) */ 132 133 struct red_parms *red_parms; 134 struct tc_sfqred_stats stats; 135 struct sfq_slot *tail; /* current slot in round */ 136 137 struct sfq_head dep[SFQ_MAX_DEPTH + 1]; 138 /* Linked lists of slots, indexed by depth 139 * dep[0] : list of unused flows 140 * dep[1] : list of flows with 1 packet 141 * dep[X] : list of flows with X packets 142 */ 143 144 unsigned int maxflows; /* number of flows in flows array */ 145 int perturb_period; 146 unsigned int quantum; /* Allotment per round: MUST BE >= MTU */ 147 struct timer_list perturb_timer; 148 struct Qdisc *sch; 149 }; 150 151 /* 152 * sfq_head are either in a sfq_slot or in dep[] array 153 */ 154 static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val) 155 { 156 if (val < SFQ_MAX_FLOWS) 157 return &q->slots[val].dep; 158 return &q->dep[val - SFQ_MAX_FLOWS]; 159 } 160 161 static unsigned int sfq_hash(const struct sfq_sched_data *q, 162 const struct sk_buff *skb) 163 { 164 return skb_get_hash_perturb(skb, q->perturbation) & (q->divisor - 1); 165 } 166 167 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch, 168 int *qerr) 169 { 170 struct sfq_sched_data *q = qdisc_priv(sch); 171 struct tcf_result res; 172 struct tcf_proto *fl; 173 int result; 174 175 if (TC_H_MAJ(skb->priority) == sch->handle && 176 TC_H_MIN(skb->priority) > 0 && 177 TC_H_MIN(skb->priority) <= q->divisor) 178 return TC_H_MIN(skb->priority); 179 180 fl = rcu_dereference_bh(q->filter_list); 181 if (!fl) 182 return sfq_hash(q, skb) + 1; 183 184 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 185 result = tcf_classify(skb, fl, &res, false); 186 if (result >= 0) { 187 #ifdef CONFIG_NET_CLS_ACT 188 switch (result) { 189 case TC_ACT_STOLEN: 190 case TC_ACT_QUEUED: 191 case TC_ACT_TRAP: 192 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; 193 /* fall through */ 194 case TC_ACT_SHOT: 195 return 0; 196 } 197 #endif 198 if (TC_H_MIN(res.classid) <= q->divisor) 199 return TC_H_MIN(res.classid); 200 } 201 return 0; 202 } 203 204 /* 205 * x : slot number [0 .. SFQ_MAX_FLOWS - 1] 206 */ 207 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x) 208 { 209 sfq_index p, n; 210 struct sfq_slot *slot = &q->slots[x]; 211 int qlen = slot->qlen; 212 213 p = qlen + SFQ_MAX_FLOWS; 214 n = q->dep[qlen].next; 215 216 slot->dep.next = n; 217 slot->dep.prev = p; 218 219 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */ 220 sfq_dep_head(q, n)->prev = x; 221 } 222 223 #define sfq_unlink(q, x, n, p) \ 224 do { \ 225 n = q->slots[x].dep.next; \ 226 p = q->slots[x].dep.prev; \ 227 sfq_dep_head(q, p)->next = n; \ 228 sfq_dep_head(q, n)->prev = p; \ 229 } while (0) 230 231 232 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x) 233 { 234 sfq_index p, n; 235 int d; 236 237 sfq_unlink(q, x, n, p); 238 239 d = q->slots[x].qlen--; 240 if (n == p && q->cur_depth == d) 241 q->cur_depth--; 242 sfq_link(q, x); 243 } 244 245 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x) 246 { 247 sfq_index p, n; 248 int d; 249 250 sfq_unlink(q, x, n, p); 251 252 d = ++q->slots[x].qlen; 253 if (q->cur_depth < d) 254 q->cur_depth = d; 255 sfq_link(q, x); 256 } 257 258 /* helper functions : might be changed when/if skb use a standard list_head */ 259 260 /* remove one skb from tail of slot queue */ 261 static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot) 262 { 263 struct sk_buff *skb = slot->skblist_prev; 264 265 slot->skblist_prev = skb->prev; 266 skb->prev->next = (struct sk_buff *)slot; 267 skb->next = skb->prev = NULL; 268 return skb; 269 } 270 271 /* remove one skb from head of slot queue */ 272 static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot) 273 { 274 struct sk_buff *skb = slot->skblist_next; 275 276 slot->skblist_next = skb->next; 277 skb->next->prev = (struct sk_buff *)slot; 278 skb->next = skb->prev = NULL; 279 return skb; 280 } 281 282 static inline void slot_queue_init(struct sfq_slot *slot) 283 { 284 memset(slot, 0, sizeof(*slot)); 285 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot; 286 } 287 288 /* add skb to slot queue (tail add) */ 289 static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb) 290 { 291 skb->prev = slot->skblist_prev; 292 skb->next = (struct sk_buff *)slot; 293 slot->skblist_prev->next = skb; 294 slot->skblist_prev = skb; 295 } 296 297 static unsigned int sfq_drop(struct Qdisc *sch, struct sk_buff **to_free) 298 { 299 struct sfq_sched_data *q = qdisc_priv(sch); 300 sfq_index x, d = q->cur_depth; 301 struct sk_buff *skb; 302 unsigned int len; 303 struct sfq_slot *slot; 304 305 /* Queue is full! Find the longest slot and drop tail packet from it */ 306 if (d > 1) { 307 x = q->dep[d].next; 308 slot = &q->slots[x]; 309 drop: 310 skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot); 311 len = qdisc_pkt_len(skb); 312 slot->backlog -= len; 313 sfq_dec(q, x); 314 sch->q.qlen--; 315 qdisc_qstats_backlog_dec(sch, skb); 316 qdisc_drop(skb, sch, to_free); 317 return len; 318 } 319 320 if (d == 1) { 321 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */ 322 x = q->tail->next; 323 slot = &q->slots[x]; 324 q->tail->next = slot->next; 325 q->ht[slot->hash] = SFQ_EMPTY_SLOT; 326 goto drop; 327 } 328 329 return 0; 330 } 331 332 /* Is ECN parameter configured */ 333 static int sfq_prob_mark(const struct sfq_sched_data *q) 334 { 335 return q->flags & TC_RED_ECN; 336 } 337 338 /* Should packets over max threshold just be marked */ 339 static int sfq_hard_mark(const struct sfq_sched_data *q) 340 { 341 return (q->flags & (TC_RED_ECN | TC_RED_HARDDROP)) == TC_RED_ECN; 342 } 343 344 static int sfq_headdrop(const struct sfq_sched_data *q) 345 { 346 return q->headdrop; 347 } 348 349 static int 350 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free) 351 { 352 struct sfq_sched_data *q = qdisc_priv(sch); 353 unsigned int hash, dropped; 354 sfq_index x, qlen; 355 struct sfq_slot *slot; 356 int uninitialized_var(ret); 357 struct sk_buff *head; 358 int delta; 359 360 hash = sfq_classify(skb, sch, &ret); 361 if (hash == 0) { 362 if (ret & __NET_XMIT_BYPASS) 363 qdisc_qstats_drop(sch); 364 __qdisc_drop(skb, to_free); 365 return ret; 366 } 367 hash--; 368 369 x = q->ht[hash]; 370 slot = &q->slots[x]; 371 if (x == SFQ_EMPTY_SLOT) { 372 x = q->dep[0].next; /* get a free slot */ 373 if (x >= SFQ_MAX_FLOWS) 374 return qdisc_drop(skb, sch, to_free); 375 q->ht[hash] = x; 376 slot = &q->slots[x]; 377 slot->hash = hash; 378 slot->backlog = 0; /* should already be 0 anyway... */ 379 red_set_vars(&slot->vars); 380 goto enqueue; 381 } 382 if (q->red_parms) { 383 slot->vars.qavg = red_calc_qavg_no_idle_time(q->red_parms, 384 &slot->vars, 385 slot->backlog); 386 switch (red_action(q->red_parms, 387 &slot->vars, 388 slot->vars.qavg)) { 389 case RED_DONT_MARK: 390 break; 391 392 case RED_PROB_MARK: 393 qdisc_qstats_overlimit(sch); 394 if (sfq_prob_mark(q)) { 395 /* We know we have at least one packet in queue */ 396 if (sfq_headdrop(q) && 397 INET_ECN_set_ce(slot->skblist_next)) { 398 q->stats.prob_mark_head++; 399 break; 400 } 401 if (INET_ECN_set_ce(skb)) { 402 q->stats.prob_mark++; 403 break; 404 } 405 } 406 q->stats.prob_drop++; 407 goto congestion_drop; 408 409 case RED_HARD_MARK: 410 qdisc_qstats_overlimit(sch); 411 if (sfq_hard_mark(q)) { 412 /* We know we have at least one packet in queue */ 413 if (sfq_headdrop(q) && 414 INET_ECN_set_ce(slot->skblist_next)) { 415 q->stats.forced_mark_head++; 416 break; 417 } 418 if (INET_ECN_set_ce(skb)) { 419 q->stats.forced_mark++; 420 break; 421 } 422 } 423 q->stats.forced_drop++; 424 goto congestion_drop; 425 } 426 } 427 428 if (slot->qlen >= q->maxdepth) { 429 congestion_drop: 430 if (!sfq_headdrop(q)) 431 return qdisc_drop(skb, sch, to_free); 432 433 /* We know we have at least one packet in queue */ 434 head = slot_dequeue_head(slot); 435 delta = qdisc_pkt_len(head) - qdisc_pkt_len(skb); 436 sch->qstats.backlog -= delta; 437 slot->backlog -= delta; 438 qdisc_drop(head, sch, to_free); 439 440 slot_queue_add(slot, skb); 441 qdisc_tree_reduce_backlog(sch, 0, delta); 442 return NET_XMIT_CN; 443 } 444 445 enqueue: 446 qdisc_qstats_backlog_inc(sch, skb); 447 slot->backlog += qdisc_pkt_len(skb); 448 slot_queue_add(slot, skb); 449 sfq_inc(q, x); 450 if (slot->qlen == 1) { /* The flow is new */ 451 if (q->tail == NULL) { /* It is the first flow */ 452 slot->next = x; 453 } else { 454 slot->next = q->tail->next; 455 q->tail->next = x; 456 } 457 /* We put this flow at the end of our flow list. 458 * This might sound unfair for a new flow to wait after old ones, 459 * but we could endup servicing new flows only, and freeze old ones. 460 */ 461 q->tail = slot; 462 /* We could use a bigger initial quantum for new flows */ 463 slot->allot = q->scaled_quantum; 464 } 465 if (++sch->q.qlen <= q->limit) 466 return NET_XMIT_SUCCESS; 467 468 qlen = slot->qlen; 469 dropped = sfq_drop(sch, to_free); 470 /* Return Congestion Notification only if we dropped a packet 471 * from this flow. 472 */ 473 if (qlen != slot->qlen) { 474 qdisc_tree_reduce_backlog(sch, 0, dropped - qdisc_pkt_len(skb)); 475 return NET_XMIT_CN; 476 } 477 478 /* As we dropped a packet, better let upper stack know this */ 479 qdisc_tree_reduce_backlog(sch, 1, dropped); 480 return NET_XMIT_SUCCESS; 481 } 482 483 static struct sk_buff * 484 sfq_dequeue(struct Qdisc *sch) 485 { 486 struct sfq_sched_data *q = qdisc_priv(sch); 487 struct sk_buff *skb; 488 sfq_index a, next_a; 489 struct sfq_slot *slot; 490 491 /* No active slots */ 492 if (q->tail == NULL) 493 return NULL; 494 495 next_slot: 496 a = q->tail->next; 497 slot = &q->slots[a]; 498 if (slot->allot <= 0) { 499 q->tail = slot; 500 slot->allot += q->scaled_quantum; 501 goto next_slot; 502 } 503 skb = slot_dequeue_head(slot); 504 sfq_dec(q, a); 505 qdisc_bstats_update(sch, skb); 506 sch->q.qlen--; 507 qdisc_qstats_backlog_dec(sch, skb); 508 slot->backlog -= qdisc_pkt_len(skb); 509 /* Is the slot empty? */ 510 if (slot->qlen == 0) { 511 q->ht[slot->hash] = SFQ_EMPTY_SLOT; 512 next_a = slot->next; 513 if (a == next_a) { 514 q->tail = NULL; /* no more active slots */ 515 return skb; 516 } 517 q->tail->next = next_a; 518 } else { 519 slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb)); 520 } 521 return skb; 522 } 523 524 static void 525 sfq_reset(struct Qdisc *sch) 526 { 527 struct sk_buff *skb; 528 529 while ((skb = sfq_dequeue(sch)) != NULL) 530 rtnl_kfree_skbs(skb, skb); 531 } 532 533 /* 534 * When q->perturbation is changed, we rehash all queued skbs 535 * to avoid OOO (Out Of Order) effects. 536 * We dont use sfq_dequeue()/sfq_enqueue() because we dont want to change 537 * counters. 538 */ 539 static void sfq_rehash(struct Qdisc *sch) 540 { 541 struct sfq_sched_data *q = qdisc_priv(sch); 542 struct sk_buff *skb; 543 int i; 544 struct sfq_slot *slot; 545 struct sk_buff_head list; 546 int dropped = 0; 547 unsigned int drop_len = 0; 548 549 __skb_queue_head_init(&list); 550 551 for (i = 0; i < q->maxflows; i++) { 552 slot = &q->slots[i]; 553 if (!slot->qlen) 554 continue; 555 while (slot->qlen) { 556 skb = slot_dequeue_head(slot); 557 sfq_dec(q, i); 558 __skb_queue_tail(&list, skb); 559 } 560 slot->backlog = 0; 561 red_set_vars(&slot->vars); 562 q->ht[slot->hash] = SFQ_EMPTY_SLOT; 563 } 564 q->tail = NULL; 565 566 while ((skb = __skb_dequeue(&list)) != NULL) { 567 unsigned int hash = sfq_hash(q, skb); 568 sfq_index x = q->ht[hash]; 569 570 slot = &q->slots[x]; 571 if (x == SFQ_EMPTY_SLOT) { 572 x = q->dep[0].next; /* get a free slot */ 573 if (x >= SFQ_MAX_FLOWS) { 574 drop: 575 qdisc_qstats_backlog_dec(sch, skb); 576 drop_len += qdisc_pkt_len(skb); 577 kfree_skb(skb); 578 dropped++; 579 continue; 580 } 581 q->ht[hash] = x; 582 slot = &q->slots[x]; 583 slot->hash = hash; 584 } 585 if (slot->qlen >= q->maxdepth) 586 goto drop; 587 slot_queue_add(slot, skb); 588 if (q->red_parms) 589 slot->vars.qavg = red_calc_qavg(q->red_parms, 590 &slot->vars, 591 slot->backlog); 592 slot->backlog += qdisc_pkt_len(skb); 593 sfq_inc(q, x); 594 if (slot->qlen == 1) { /* The flow is new */ 595 if (q->tail == NULL) { /* It is the first flow */ 596 slot->next = x; 597 } else { 598 slot->next = q->tail->next; 599 q->tail->next = x; 600 } 601 q->tail = slot; 602 slot->allot = q->scaled_quantum; 603 } 604 } 605 sch->q.qlen -= dropped; 606 qdisc_tree_reduce_backlog(sch, dropped, drop_len); 607 } 608 609 static void sfq_perturbation(struct timer_list *t) 610 { 611 struct sfq_sched_data *q = from_timer(q, t, perturb_timer); 612 struct Qdisc *sch = q->sch; 613 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch)); 614 615 spin_lock(root_lock); 616 q->perturbation = prandom_u32(); 617 if (!q->filter_list && q->tail) 618 sfq_rehash(sch); 619 spin_unlock(root_lock); 620 621 if (q->perturb_period) 622 mod_timer(&q->perturb_timer, jiffies + q->perturb_period); 623 } 624 625 static int sfq_change(struct Qdisc *sch, struct nlattr *opt) 626 { 627 struct sfq_sched_data *q = qdisc_priv(sch); 628 struct tc_sfq_qopt *ctl = nla_data(opt); 629 struct tc_sfq_qopt_v1 *ctl_v1 = NULL; 630 unsigned int qlen, dropped = 0; 631 struct red_parms *p = NULL; 632 struct sk_buff *to_free = NULL; 633 struct sk_buff *tail = NULL; 634 635 if (opt->nla_len < nla_attr_size(sizeof(*ctl))) 636 return -EINVAL; 637 if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1))) 638 ctl_v1 = nla_data(opt); 639 if (ctl->divisor && 640 (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536)) 641 return -EINVAL; 642 if (ctl_v1 && !red_check_params(ctl_v1->qth_min, ctl_v1->qth_max, 643 ctl_v1->Wlog)) 644 return -EINVAL; 645 if (ctl_v1 && ctl_v1->qth_min) { 646 p = kmalloc(sizeof(*p), GFP_KERNEL); 647 if (!p) 648 return -ENOMEM; 649 } 650 sch_tree_lock(sch); 651 if (ctl->quantum) { 652 q->quantum = ctl->quantum; 653 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); 654 } 655 q->perturb_period = ctl->perturb_period * HZ; 656 if (ctl->flows) 657 q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS); 658 if (ctl->divisor) { 659 q->divisor = ctl->divisor; 660 q->maxflows = min_t(u32, q->maxflows, q->divisor); 661 } 662 if (ctl_v1) { 663 if (ctl_v1->depth) 664 q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH); 665 if (p) { 666 swap(q->red_parms, p); 667 red_set_parms(q->red_parms, 668 ctl_v1->qth_min, ctl_v1->qth_max, 669 ctl_v1->Wlog, 670 ctl_v1->Plog, ctl_v1->Scell_log, 671 NULL, 672 ctl_v1->max_P); 673 } 674 q->flags = ctl_v1->flags; 675 q->headdrop = ctl_v1->headdrop; 676 } 677 if (ctl->limit) { 678 q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows); 679 q->maxflows = min_t(u32, q->maxflows, q->limit); 680 } 681 682 qlen = sch->q.qlen; 683 while (sch->q.qlen > q->limit) { 684 dropped += sfq_drop(sch, &to_free); 685 if (!tail) 686 tail = to_free; 687 } 688 689 rtnl_kfree_skbs(to_free, tail); 690 qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped); 691 692 del_timer(&q->perturb_timer); 693 if (q->perturb_period) { 694 mod_timer(&q->perturb_timer, jiffies + q->perturb_period); 695 q->perturbation = prandom_u32(); 696 } 697 sch_tree_unlock(sch); 698 kfree(p); 699 return 0; 700 } 701 702 static void *sfq_alloc(size_t sz) 703 { 704 return kvmalloc(sz, GFP_KERNEL); 705 } 706 707 static void sfq_free(void *addr) 708 { 709 kvfree(addr); 710 } 711 712 static void sfq_destroy(struct Qdisc *sch) 713 { 714 struct sfq_sched_data *q = qdisc_priv(sch); 715 716 tcf_block_put(q->block); 717 q->perturb_period = 0; 718 del_timer_sync(&q->perturb_timer); 719 sfq_free(q->ht); 720 sfq_free(q->slots); 721 kfree(q->red_parms); 722 } 723 724 static int sfq_init(struct Qdisc *sch, struct nlattr *opt) 725 { 726 struct sfq_sched_data *q = qdisc_priv(sch); 727 int i; 728 int err; 729 730 q->sch = sch; 731 timer_setup(&q->perturb_timer, sfq_perturbation, TIMER_DEFERRABLE); 732 733 err = tcf_block_get(&q->block, &q->filter_list, sch); 734 if (err) 735 return err; 736 737 for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) { 738 q->dep[i].next = i + SFQ_MAX_FLOWS; 739 q->dep[i].prev = i + SFQ_MAX_FLOWS; 740 } 741 742 q->limit = SFQ_MAX_DEPTH; 743 q->maxdepth = SFQ_MAX_DEPTH; 744 q->cur_depth = 0; 745 q->tail = NULL; 746 q->divisor = SFQ_DEFAULT_HASH_DIVISOR; 747 q->maxflows = SFQ_DEFAULT_FLOWS; 748 q->quantum = psched_mtu(qdisc_dev(sch)); 749 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); 750 q->perturb_period = 0; 751 q->perturbation = prandom_u32(); 752 753 if (opt) { 754 int err = sfq_change(sch, opt); 755 if (err) 756 return err; 757 } 758 759 q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor); 760 q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows); 761 if (!q->ht || !q->slots) { 762 /* Note: sfq_destroy() will be called by our caller */ 763 return -ENOMEM; 764 } 765 766 for (i = 0; i < q->divisor; i++) 767 q->ht[i] = SFQ_EMPTY_SLOT; 768 769 for (i = 0; i < q->maxflows; i++) { 770 slot_queue_init(&q->slots[i]); 771 sfq_link(q, i); 772 } 773 if (q->limit >= 1) 774 sch->flags |= TCQ_F_CAN_BYPASS; 775 else 776 sch->flags &= ~TCQ_F_CAN_BYPASS; 777 return 0; 778 } 779 780 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb) 781 { 782 struct sfq_sched_data *q = qdisc_priv(sch); 783 unsigned char *b = skb_tail_pointer(skb); 784 struct tc_sfq_qopt_v1 opt; 785 struct red_parms *p = q->red_parms; 786 787 memset(&opt, 0, sizeof(opt)); 788 opt.v0.quantum = q->quantum; 789 opt.v0.perturb_period = q->perturb_period / HZ; 790 opt.v0.limit = q->limit; 791 opt.v0.divisor = q->divisor; 792 opt.v0.flows = q->maxflows; 793 opt.depth = q->maxdepth; 794 opt.headdrop = q->headdrop; 795 796 if (p) { 797 opt.qth_min = p->qth_min >> p->Wlog; 798 opt.qth_max = p->qth_max >> p->Wlog; 799 opt.Wlog = p->Wlog; 800 opt.Plog = p->Plog; 801 opt.Scell_log = p->Scell_log; 802 opt.max_P = p->max_P; 803 } 804 memcpy(&opt.stats, &q->stats, sizeof(opt.stats)); 805 opt.flags = q->flags; 806 807 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) 808 goto nla_put_failure; 809 810 return skb->len; 811 812 nla_put_failure: 813 nlmsg_trim(skb, b); 814 return -1; 815 } 816 817 static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg) 818 { 819 return NULL; 820 } 821 822 static unsigned long sfq_find(struct Qdisc *sch, u32 classid) 823 { 824 return 0; 825 } 826 827 static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent, 828 u32 classid) 829 { 830 /* we cannot bypass queue discipline anymore */ 831 sch->flags &= ~TCQ_F_CAN_BYPASS; 832 return 0; 833 } 834 835 static void sfq_unbind(struct Qdisc *q, unsigned long cl) 836 { 837 } 838 839 static struct tcf_block *sfq_tcf_block(struct Qdisc *sch, unsigned long cl) 840 { 841 struct sfq_sched_data *q = qdisc_priv(sch); 842 843 if (cl) 844 return NULL; 845 return q->block; 846 } 847 848 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl, 849 struct sk_buff *skb, struct tcmsg *tcm) 850 { 851 tcm->tcm_handle |= TC_H_MIN(cl); 852 return 0; 853 } 854 855 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl, 856 struct gnet_dump *d) 857 { 858 struct sfq_sched_data *q = qdisc_priv(sch); 859 sfq_index idx = q->ht[cl - 1]; 860 struct gnet_stats_queue qs = { 0 }; 861 struct tc_sfq_xstats xstats = { 0 }; 862 863 if (idx != SFQ_EMPTY_SLOT) { 864 const struct sfq_slot *slot = &q->slots[idx]; 865 866 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT; 867 qs.qlen = slot->qlen; 868 qs.backlog = slot->backlog; 869 } 870 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0) 871 return -1; 872 return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); 873 } 874 875 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) 876 { 877 struct sfq_sched_data *q = qdisc_priv(sch); 878 unsigned int i; 879 880 if (arg->stop) 881 return; 882 883 for (i = 0; i < q->divisor; i++) { 884 if (q->ht[i] == SFQ_EMPTY_SLOT || 885 arg->count < arg->skip) { 886 arg->count++; 887 continue; 888 } 889 if (arg->fn(sch, i + 1, arg) < 0) { 890 arg->stop = 1; 891 break; 892 } 893 arg->count++; 894 } 895 } 896 897 static const struct Qdisc_class_ops sfq_class_ops = { 898 .leaf = sfq_leaf, 899 .find = sfq_find, 900 .tcf_block = sfq_tcf_block, 901 .bind_tcf = sfq_bind, 902 .unbind_tcf = sfq_unbind, 903 .dump = sfq_dump_class, 904 .dump_stats = sfq_dump_class_stats, 905 .walk = sfq_walk, 906 }; 907 908 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = { 909 .cl_ops = &sfq_class_ops, 910 .id = "sfq", 911 .priv_size = sizeof(struct sfq_sched_data), 912 .enqueue = sfq_enqueue, 913 .dequeue = sfq_dequeue, 914 .peek = qdisc_peek_dequeued, 915 .init = sfq_init, 916 .reset = sfq_reset, 917 .destroy = sfq_destroy, 918 .change = NULL, 919 .dump = sfq_dump, 920 .owner = THIS_MODULE, 921 }; 922 923 static int __init sfq_module_init(void) 924 { 925 return register_qdisc(&sfq_qdisc_ops); 926 } 927 static void __exit sfq_module_exit(void) 928 { 929 unregister_qdisc(&sfq_qdisc_ops); 930 } 931 module_init(sfq_module_init) 932 module_exit(sfq_module_exit) 933 MODULE_LICENSE("GPL"); 934