1 /* 2 * Fair Queue CoDel 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 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com> 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/codel.h> 27 28 /* Fair Queue CoDel. 29 * 30 * Principles : 31 * Packets are classified (internal classifier or external) on flows. 32 * This is a Stochastic model (as we use a hash, several flows 33 * might be hashed on same slot) 34 * Each flow has a CoDel managed queue. 35 * Flows are linked onto two (Round Robin) lists, 36 * so that new flows have priority on old ones. 37 * 38 * For a given flow, packets are not reordered (CoDel uses a FIFO) 39 * head drops only. 40 * ECN capability is on by default. 41 * Low memory footprint (64 bytes per flow) 42 */ 43 44 struct fq_codel_flow { 45 struct sk_buff *head; 46 struct sk_buff *tail; 47 struct list_head flowchain; 48 int deficit; 49 u32 dropped; /* number of drops (or ECN marks) on this flow */ 50 struct codel_vars cvars; 51 }; /* please try to keep this structure <= 64 bytes */ 52 53 struct fq_codel_sched_data { 54 struct tcf_proto __rcu *filter_list; /* optional external classifier */ 55 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */ 56 u32 *backlogs; /* backlog table [flows_cnt] */ 57 u32 flows_cnt; /* number of flows */ 58 u32 perturbation; /* hash perturbation */ 59 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */ 60 struct codel_params cparams; 61 struct codel_stats cstats; 62 u32 drop_overlimit; 63 u32 new_flow_count; 64 65 struct list_head new_flows; /* list of new flows */ 66 struct list_head old_flows; /* list of old flows */ 67 }; 68 69 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q, 70 struct sk_buff *skb) 71 { 72 u32 hash = skb_get_hash_perturb(skb, q->perturbation); 73 74 return reciprocal_scale(hash, q->flows_cnt); 75 } 76 77 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch, 78 int *qerr) 79 { 80 struct fq_codel_sched_data *q = qdisc_priv(sch); 81 struct tcf_proto *filter; 82 struct tcf_result res; 83 int result; 84 85 if (TC_H_MAJ(skb->priority) == sch->handle && 86 TC_H_MIN(skb->priority) > 0 && 87 TC_H_MIN(skb->priority) <= q->flows_cnt) 88 return TC_H_MIN(skb->priority); 89 90 filter = rcu_dereference_bh(q->filter_list); 91 if (!filter) 92 return fq_codel_hash(q, skb) + 1; 93 94 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 95 result = tc_classify(skb, filter, &res, false); 96 if (result >= 0) { 97 #ifdef CONFIG_NET_CLS_ACT 98 switch (result) { 99 case TC_ACT_STOLEN: 100 case TC_ACT_QUEUED: 101 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; 102 case TC_ACT_SHOT: 103 return 0; 104 } 105 #endif 106 if (TC_H_MIN(res.classid) <= q->flows_cnt) 107 return TC_H_MIN(res.classid); 108 } 109 return 0; 110 } 111 112 /* helper functions : might be changed when/if skb use a standard list_head */ 113 114 /* remove one skb from head of slot queue */ 115 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow) 116 { 117 struct sk_buff *skb = flow->head; 118 119 flow->head = skb->next; 120 skb->next = NULL; 121 return skb; 122 } 123 124 /* add skb to flow queue (tail add) */ 125 static inline void flow_queue_add(struct fq_codel_flow *flow, 126 struct sk_buff *skb) 127 { 128 if (flow->head == NULL) 129 flow->head = skb; 130 else 131 flow->tail->next = skb; 132 flow->tail = skb; 133 skb->next = NULL; 134 } 135 136 static unsigned int fq_codel_drop(struct Qdisc *sch) 137 { 138 struct fq_codel_sched_data *q = qdisc_priv(sch); 139 struct sk_buff *skb; 140 unsigned int maxbacklog = 0, idx = 0, i, len; 141 struct fq_codel_flow *flow; 142 143 /* Queue is full! Find the fat flow and drop packet from it. 144 * This might sound expensive, but with 1024 flows, we scan 145 * 4KB of memory, and we dont need to handle a complex tree 146 * in fast path (packet queue/enqueue) with many cache misses. 147 */ 148 for (i = 0; i < q->flows_cnt; i++) { 149 if (q->backlogs[i] > maxbacklog) { 150 maxbacklog = q->backlogs[i]; 151 idx = i; 152 } 153 } 154 flow = &q->flows[idx]; 155 skb = dequeue_head(flow); 156 len = qdisc_pkt_len(skb); 157 q->backlogs[idx] -= len; 158 sch->q.qlen--; 159 qdisc_qstats_drop(sch); 160 qdisc_qstats_backlog_dec(sch, skb); 161 kfree_skb(skb); 162 flow->dropped++; 163 return idx; 164 } 165 166 static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch) 167 { 168 unsigned int prev_backlog; 169 170 prev_backlog = sch->qstats.backlog; 171 fq_codel_drop(sch); 172 return prev_backlog - sch->qstats.backlog; 173 } 174 175 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch) 176 { 177 struct fq_codel_sched_data *q = qdisc_priv(sch); 178 unsigned int idx; 179 struct fq_codel_flow *flow; 180 int uninitialized_var(ret); 181 182 idx = fq_codel_classify(skb, sch, &ret); 183 if (idx == 0) { 184 if (ret & __NET_XMIT_BYPASS) 185 qdisc_qstats_drop(sch); 186 kfree_skb(skb); 187 return ret; 188 } 189 idx--; 190 191 codel_set_enqueue_time(skb); 192 flow = &q->flows[idx]; 193 flow_queue_add(flow, skb); 194 q->backlogs[idx] += qdisc_pkt_len(skb); 195 qdisc_qstats_backlog_inc(sch, skb); 196 197 if (list_empty(&flow->flowchain)) { 198 list_add_tail(&flow->flowchain, &q->new_flows); 199 q->new_flow_count++; 200 flow->deficit = q->quantum; 201 flow->dropped = 0; 202 } 203 if (++sch->q.qlen <= sch->limit) 204 return NET_XMIT_SUCCESS; 205 206 q->drop_overlimit++; 207 /* Return Congestion Notification only if we dropped a packet 208 * from this flow. 209 */ 210 if (fq_codel_drop(sch) == idx) 211 return NET_XMIT_CN; 212 213 /* As we dropped a packet, better let upper stack know this */ 214 qdisc_tree_decrease_qlen(sch, 1); 215 return NET_XMIT_SUCCESS; 216 } 217 218 /* This is the specific function called from codel_dequeue() 219 * to dequeue a packet from queue. Note: backlog is handled in 220 * codel, we dont need to reduce it here. 221 */ 222 static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch) 223 { 224 struct fq_codel_sched_data *q = qdisc_priv(sch); 225 struct fq_codel_flow *flow; 226 struct sk_buff *skb = NULL; 227 228 flow = container_of(vars, struct fq_codel_flow, cvars); 229 if (flow->head) { 230 skb = dequeue_head(flow); 231 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb); 232 sch->q.qlen--; 233 } 234 return skb; 235 } 236 237 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch) 238 { 239 struct fq_codel_sched_data *q = qdisc_priv(sch); 240 struct sk_buff *skb; 241 struct fq_codel_flow *flow; 242 struct list_head *head; 243 u32 prev_drop_count, prev_ecn_mark; 244 245 begin: 246 head = &q->new_flows; 247 if (list_empty(head)) { 248 head = &q->old_flows; 249 if (list_empty(head)) 250 return NULL; 251 } 252 flow = list_first_entry(head, struct fq_codel_flow, flowchain); 253 254 if (flow->deficit <= 0) { 255 flow->deficit += q->quantum; 256 list_move_tail(&flow->flowchain, &q->old_flows); 257 goto begin; 258 } 259 260 prev_drop_count = q->cstats.drop_count; 261 prev_ecn_mark = q->cstats.ecn_mark; 262 263 skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats, 264 dequeue); 265 266 flow->dropped += q->cstats.drop_count - prev_drop_count; 267 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark; 268 269 if (!skb) { 270 /* force a pass through old_flows to prevent starvation */ 271 if ((head == &q->new_flows) && !list_empty(&q->old_flows)) 272 list_move_tail(&flow->flowchain, &q->old_flows); 273 else 274 list_del_init(&flow->flowchain); 275 goto begin; 276 } 277 qdisc_bstats_update(sch, skb); 278 flow->deficit -= qdisc_pkt_len(skb); 279 /* We cant call qdisc_tree_decrease_qlen() if our qlen is 0, 280 * or HTB crashes. Defer it for next round. 281 */ 282 if (q->cstats.drop_count && sch->q.qlen) { 283 qdisc_tree_decrease_qlen(sch, q->cstats.drop_count); 284 q->cstats.drop_count = 0; 285 } 286 return skb; 287 } 288 289 static void fq_codel_reset(struct Qdisc *sch) 290 { 291 struct fq_codel_sched_data *q = qdisc_priv(sch); 292 int i; 293 294 INIT_LIST_HEAD(&q->new_flows); 295 INIT_LIST_HEAD(&q->old_flows); 296 for (i = 0; i < q->flows_cnt; i++) { 297 struct fq_codel_flow *flow = q->flows + i; 298 299 while (flow->head) { 300 struct sk_buff *skb = dequeue_head(flow); 301 302 qdisc_qstats_backlog_dec(sch, skb); 303 kfree_skb(skb); 304 } 305 306 INIT_LIST_HEAD(&flow->flowchain); 307 codel_vars_init(&flow->cvars); 308 } 309 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32)); 310 sch->q.qlen = 0; 311 } 312 313 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = { 314 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 }, 315 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 }, 316 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 }, 317 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 }, 318 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 }, 319 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 }, 320 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 }, 321 }; 322 323 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt) 324 { 325 struct fq_codel_sched_data *q = qdisc_priv(sch); 326 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1]; 327 int err; 328 329 if (!opt) 330 return -EINVAL; 331 332 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy); 333 if (err < 0) 334 return err; 335 if (tb[TCA_FQ_CODEL_FLOWS]) { 336 if (q->flows) 337 return -EINVAL; 338 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]); 339 if (!q->flows_cnt || 340 q->flows_cnt > 65536) 341 return -EINVAL; 342 } 343 sch_tree_lock(sch); 344 345 if (tb[TCA_FQ_CODEL_TARGET]) { 346 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]); 347 348 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT; 349 } 350 351 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) { 352 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]); 353 354 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT; 355 } 356 357 if (tb[TCA_FQ_CODEL_INTERVAL]) { 358 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]); 359 360 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT; 361 } 362 363 if (tb[TCA_FQ_CODEL_LIMIT]) 364 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]); 365 366 if (tb[TCA_FQ_CODEL_ECN]) 367 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]); 368 369 if (tb[TCA_FQ_CODEL_QUANTUM]) 370 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM])); 371 372 while (sch->q.qlen > sch->limit) { 373 struct sk_buff *skb = fq_codel_dequeue(sch); 374 375 kfree_skb(skb); 376 q->cstats.drop_count++; 377 } 378 qdisc_tree_decrease_qlen(sch, q->cstats.drop_count); 379 q->cstats.drop_count = 0; 380 381 sch_tree_unlock(sch); 382 return 0; 383 } 384 385 static void *fq_codel_zalloc(size_t sz) 386 { 387 void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN); 388 389 if (!ptr) 390 ptr = vzalloc(sz); 391 return ptr; 392 } 393 394 static void fq_codel_free(void *addr) 395 { 396 kvfree(addr); 397 } 398 399 static void fq_codel_destroy(struct Qdisc *sch) 400 { 401 struct fq_codel_sched_data *q = qdisc_priv(sch); 402 403 tcf_destroy_chain(&q->filter_list); 404 fq_codel_free(q->backlogs); 405 fq_codel_free(q->flows); 406 } 407 408 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt) 409 { 410 struct fq_codel_sched_data *q = qdisc_priv(sch); 411 int i; 412 413 sch->limit = 10*1024; 414 q->flows_cnt = 1024; 415 q->quantum = psched_mtu(qdisc_dev(sch)); 416 q->perturbation = prandom_u32(); 417 INIT_LIST_HEAD(&q->new_flows); 418 INIT_LIST_HEAD(&q->old_flows); 419 codel_params_init(&q->cparams, sch); 420 codel_stats_init(&q->cstats); 421 q->cparams.ecn = true; 422 423 if (opt) { 424 int err = fq_codel_change(sch, opt); 425 if (err) 426 return err; 427 } 428 429 if (!q->flows) { 430 q->flows = fq_codel_zalloc(q->flows_cnt * 431 sizeof(struct fq_codel_flow)); 432 if (!q->flows) 433 return -ENOMEM; 434 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32)); 435 if (!q->backlogs) { 436 fq_codel_free(q->flows); 437 return -ENOMEM; 438 } 439 for (i = 0; i < q->flows_cnt; i++) { 440 struct fq_codel_flow *flow = q->flows + i; 441 442 INIT_LIST_HEAD(&flow->flowchain); 443 codel_vars_init(&flow->cvars); 444 } 445 } 446 if (sch->limit >= 1) 447 sch->flags |= TCQ_F_CAN_BYPASS; 448 else 449 sch->flags &= ~TCQ_F_CAN_BYPASS; 450 return 0; 451 } 452 453 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb) 454 { 455 struct fq_codel_sched_data *q = qdisc_priv(sch); 456 struct nlattr *opts; 457 458 opts = nla_nest_start(skb, TCA_OPTIONS); 459 if (opts == NULL) 460 goto nla_put_failure; 461 462 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET, 463 codel_time_to_us(q->cparams.target)) || 464 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT, 465 sch->limit) || 466 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL, 467 codel_time_to_us(q->cparams.interval)) || 468 nla_put_u32(skb, TCA_FQ_CODEL_ECN, 469 q->cparams.ecn) || 470 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM, 471 q->quantum) || 472 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS, 473 q->flows_cnt)) 474 goto nla_put_failure; 475 476 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD && 477 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD, 478 codel_time_to_us(q->cparams.ce_threshold))) 479 goto nla_put_failure; 480 481 return nla_nest_end(skb, opts); 482 483 nla_put_failure: 484 return -1; 485 } 486 487 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d) 488 { 489 struct fq_codel_sched_data *q = qdisc_priv(sch); 490 struct tc_fq_codel_xstats st = { 491 .type = TCA_FQ_CODEL_XSTATS_QDISC, 492 }; 493 struct list_head *pos; 494 495 st.qdisc_stats.maxpacket = q->cstats.maxpacket; 496 st.qdisc_stats.drop_overlimit = q->drop_overlimit; 497 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark; 498 st.qdisc_stats.new_flow_count = q->new_flow_count; 499 st.qdisc_stats.ce_mark = q->cstats.ce_mark; 500 501 list_for_each(pos, &q->new_flows) 502 st.qdisc_stats.new_flows_len++; 503 504 list_for_each(pos, &q->old_flows) 505 st.qdisc_stats.old_flows_len++; 506 507 return gnet_stats_copy_app(d, &st, sizeof(st)); 508 } 509 510 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg) 511 { 512 return NULL; 513 } 514 515 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid) 516 { 517 return 0; 518 } 519 520 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent, 521 u32 classid) 522 { 523 /* we cannot bypass queue discipline anymore */ 524 sch->flags &= ~TCQ_F_CAN_BYPASS; 525 return 0; 526 } 527 528 static void fq_codel_put(struct Qdisc *q, unsigned long cl) 529 { 530 } 531 532 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch, 533 unsigned long cl) 534 { 535 struct fq_codel_sched_data *q = qdisc_priv(sch); 536 537 if (cl) 538 return NULL; 539 return &q->filter_list; 540 } 541 542 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl, 543 struct sk_buff *skb, struct tcmsg *tcm) 544 { 545 tcm->tcm_handle |= TC_H_MIN(cl); 546 return 0; 547 } 548 549 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl, 550 struct gnet_dump *d) 551 { 552 struct fq_codel_sched_data *q = qdisc_priv(sch); 553 u32 idx = cl - 1; 554 struct gnet_stats_queue qs = { 0 }; 555 struct tc_fq_codel_xstats xstats; 556 557 if (idx < q->flows_cnt) { 558 const struct fq_codel_flow *flow = &q->flows[idx]; 559 const struct sk_buff *skb = flow->head; 560 561 memset(&xstats, 0, sizeof(xstats)); 562 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS; 563 xstats.class_stats.deficit = flow->deficit; 564 xstats.class_stats.ldelay = 565 codel_time_to_us(flow->cvars.ldelay); 566 xstats.class_stats.count = flow->cvars.count; 567 xstats.class_stats.lastcount = flow->cvars.lastcount; 568 xstats.class_stats.dropping = flow->cvars.dropping; 569 if (flow->cvars.dropping) { 570 codel_tdiff_t delta = flow->cvars.drop_next - 571 codel_get_time(); 572 573 xstats.class_stats.drop_next = (delta >= 0) ? 574 codel_time_to_us(delta) : 575 -codel_time_to_us(-delta); 576 } 577 while (skb) { 578 qs.qlen++; 579 skb = skb->next; 580 } 581 qs.backlog = q->backlogs[idx]; 582 qs.drops = flow->dropped; 583 } 584 if (gnet_stats_copy_queue(d, NULL, &qs, 0) < 0) 585 return -1; 586 if (idx < q->flows_cnt) 587 return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); 588 return 0; 589 } 590 591 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg) 592 { 593 struct fq_codel_sched_data *q = qdisc_priv(sch); 594 unsigned int i; 595 596 if (arg->stop) 597 return; 598 599 for (i = 0; i < q->flows_cnt; i++) { 600 if (list_empty(&q->flows[i].flowchain) || 601 arg->count < arg->skip) { 602 arg->count++; 603 continue; 604 } 605 if (arg->fn(sch, i + 1, arg) < 0) { 606 arg->stop = 1; 607 break; 608 } 609 arg->count++; 610 } 611 } 612 613 static const struct Qdisc_class_ops fq_codel_class_ops = { 614 .leaf = fq_codel_leaf, 615 .get = fq_codel_get, 616 .put = fq_codel_put, 617 .tcf_chain = fq_codel_find_tcf, 618 .bind_tcf = fq_codel_bind, 619 .unbind_tcf = fq_codel_put, 620 .dump = fq_codel_dump_class, 621 .dump_stats = fq_codel_dump_class_stats, 622 .walk = fq_codel_walk, 623 }; 624 625 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = { 626 .cl_ops = &fq_codel_class_ops, 627 .id = "fq_codel", 628 .priv_size = sizeof(struct fq_codel_sched_data), 629 .enqueue = fq_codel_enqueue, 630 .dequeue = fq_codel_dequeue, 631 .peek = qdisc_peek_dequeued, 632 .drop = fq_codel_qdisc_drop, 633 .init = fq_codel_init, 634 .reset = fq_codel_reset, 635 .destroy = fq_codel_destroy, 636 .change = fq_codel_change, 637 .dump = fq_codel_dump, 638 .dump_stats = fq_codel_dump_stats, 639 .owner = THIS_MODULE, 640 }; 641 642 static int __init fq_codel_module_init(void) 643 { 644 return register_qdisc(&fq_codel_qdisc_ops); 645 } 646 647 static void __exit fq_codel_module_exit(void) 648 { 649 unregister_qdisc(&fq_codel_qdisc_ops); 650 } 651 652 module_init(fq_codel_module_init) 653 module_exit(fq_codel_module_exit) 654 MODULE_AUTHOR("Eric Dumazet"); 655 MODULE_LICENSE("GPL"); 656