1/* 2 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler. 3 * 4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente. 5 * Copyright (c) 2012 Paolo Valente. 6 * 7 * This program is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU General Public License 9 * version 2 as published by the Free Software Foundation. 10 */ 11 12#include <linux/module.h> 13#include <linux/init.h> 14#include <linux/bitops.h> 15#include <linux/errno.h> 16#include <linux/netdevice.h> 17#include <linux/pkt_sched.h> 18#include <net/sch_generic.h> 19#include <net/pkt_sched.h> 20#include <net/pkt_cls.h> 21 22 23/* Quick Fair Queueing Plus 24 ======================== 25 26 Sources: 27 28 [1] Paolo Valente, 29 "Reducing the Execution Time of Fair-Queueing Schedulers." 30 http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf 31 32 Sources for QFQ: 33 34 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient 35 Packet Scheduling with Tight Bandwidth Distribution Guarantees." 36 37 See also: 38 http://retis.sssup.it/~fabio/linux/qfq/ 39 */ 40 41/* 42 43 QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES 44 classes. Each aggregate is timestamped with a virtual start time S 45 and a virtual finish time F, and scheduled according to its 46 timestamps. S and F are computed as a function of a system virtual 47 time function V. The classes within each aggregate are instead 48 scheduled with DRR. 49 50 To speed up operations, QFQ+ divides also aggregates into a limited 51 number of groups. Which group a class belongs to depends on the 52 ratio between the maximum packet length for the class and the weight 53 of the class. Groups have their own S and F. In the end, QFQ+ 54 schedules groups, then aggregates within groups, then classes within 55 aggregates. See [1] and [2] for a full description. 56 57 Virtual time computations. 58 59 S, F and V are all computed in fixed point arithmetic with 60 FRAC_BITS decimal bits. 61 62 QFQ_MAX_INDEX is the maximum index allowed for a group. We need 63 one bit per index. 64 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight. 65 66 The layout of the bits is as below: 67 68 [ MTU_SHIFT ][ FRAC_BITS ] 69 [ MAX_INDEX ][ MIN_SLOT_SHIFT ] 70 ^.__grp->index = 0 71 *.__grp->slot_shift 72 73 where MIN_SLOT_SHIFT is derived by difference from the others. 74 75 The max group index corresponds to Lmax/w_min, where 76 Lmax=1<<MTU_SHIFT, w_min = 1 . 77 From this, and knowing how many groups (MAX_INDEX) we want, 78 we can derive the shift corresponding to each group. 79 80 Because we often need to compute 81 F = S + len/w_i and V = V + len/wsum 82 instead of storing w_i store the value 83 inv_w = (1<<FRAC_BITS)/w_i 84 so we can do F = S + len * inv_w * wsum. 85 We use W_TOT in the formulas so we can easily move between 86 static and adaptive weight sum. 87 88 The per-scheduler-instance data contain all the data structures 89 for the scheduler: bitmaps and bucket lists. 90 91 */ 92 93/* 94 * Maximum number of consecutive slots occupied by backlogged classes 95 * inside a group. 96 */ 97#define QFQ_MAX_SLOTS 32 98 99/* 100 * Shifts used for aggregate<->group mapping. We allow class weights that are 101 * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the 102 * group with the smallest index that can support the L_i / r_i configured 103 * for the classes in the aggregate. 104 * 105 * grp->index is the index of the group; and grp->slot_shift 106 * is the shift for the corresponding (scaled) sigma_i. 107 */ 108#define QFQ_MAX_INDEX 24 109#define QFQ_MAX_WSHIFT 10 110 111#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */ 112#define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT) 113 114#define FRAC_BITS 30 /* fixed point arithmetic */ 115#define ONE_FP (1UL << FRAC_BITS) 116 117#define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */ 118#define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */ 119 120#define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */ 121 122/* 123 * Possible group states. These values are used as indexes for the bitmaps 124 * array of struct qfq_queue. 125 */ 126enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE }; 127 128struct qfq_group; 129 130struct qfq_aggregate; 131 132struct qfq_class { 133 struct Qdisc_class_common common; 134 135 unsigned int refcnt; 136 unsigned int filter_cnt; 137 138 struct gnet_stats_basic_packed bstats; 139 struct gnet_stats_queue qstats; 140 struct gnet_stats_rate_est64 rate_est; 141 struct Qdisc *qdisc; 142 struct list_head alist; /* Link for active-classes list. */ 143 struct qfq_aggregate *agg; /* Parent aggregate. */ 144 int deficit; /* DRR deficit counter. */ 145}; 146 147struct qfq_aggregate { 148 struct hlist_node next; /* Link for the slot list. */ 149 u64 S, F; /* flow timestamps (exact) */ 150 151 /* group we belong to. In principle we would need the index, 152 * which is log_2(lmax/weight), but we never reference it 153 * directly, only the group. 154 */ 155 struct qfq_group *grp; 156 157 /* these are copied from the flowset. */ 158 u32 class_weight; /* Weight of each class in this aggregate. */ 159 /* Max pkt size for the classes in this aggregate, DRR quantum. */ 160 int lmax; 161 162 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */ 163 u32 budgetmax; /* Max budget for this aggregate. */ 164 u32 initial_budget, budget; /* Initial and current budget. */ 165 166 int num_classes; /* Number of classes in this aggr. */ 167 struct list_head active; /* DRR queue of active classes. */ 168 169 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */ 170}; 171 172struct qfq_group { 173 u64 S, F; /* group timestamps (approx). */ 174 unsigned int slot_shift; /* Slot shift. */ 175 unsigned int index; /* Group index. */ 176 unsigned int front; /* Index of the front slot. */ 177 unsigned long full_slots; /* non-empty slots */ 178 179 /* Array of RR lists of active aggregates. */ 180 struct hlist_head slots[QFQ_MAX_SLOTS]; 181}; 182 183struct qfq_sched { 184 struct tcf_proto __rcu *filter_list; 185 struct Qdisc_class_hash clhash; 186 187 u64 oldV, V; /* Precise virtual times. */ 188 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */ 189 u32 num_active_agg; /* Num. of active aggregates */ 190 u32 wsum; /* weight sum */ 191 u32 iwsum; /* inverse weight sum */ 192 193 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */ 194 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */ 195 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */ 196 197 u32 max_agg_classes; /* Max number of classes per aggr. */ 198 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */ 199}; 200 201/* 202 * Possible reasons why the timestamps of an aggregate are updated 203 * enqueue: the aggregate switches from idle to active and must scheduled 204 * for service 205 * requeue: the aggregate finishes its budget, so it stops being served and 206 * must be rescheduled for service 207 */ 208enum update_reason {enqueue, requeue}; 209 210static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid) 211{ 212 struct qfq_sched *q = qdisc_priv(sch); 213 struct Qdisc_class_common *clc; 214 215 clc = qdisc_class_find(&q->clhash, classid); 216 if (clc == NULL) 217 return NULL; 218 return container_of(clc, struct qfq_class, common); 219} 220 221static void qfq_purge_queue(struct qfq_class *cl) 222{ 223 unsigned int len = cl->qdisc->q.qlen; 224 225 qdisc_reset(cl->qdisc); 226 qdisc_tree_decrease_qlen(cl->qdisc, len); 227} 228 229static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = { 230 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 }, 231 [TCA_QFQ_LMAX] = { .type = NLA_U32 }, 232}; 233 234/* 235 * Calculate a flow index, given its weight and maximum packet length. 236 * index = log_2(maxlen/weight) but we need to apply the scaling. 237 * This is used only once at flow creation. 238 */ 239static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift) 240{ 241 u64 slot_size = (u64)maxlen * inv_w; 242 unsigned long size_map; 243 int index = 0; 244 245 size_map = slot_size >> min_slot_shift; 246 if (!size_map) 247 goto out; 248 249 index = __fls(size_map) + 1; /* basically a log_2 */ 250 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1))); 251 252 if (index < 0) 253 index = 0; 254out: 255 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n", 256 (unsigned long) ONE_FP/inv_w, maxlen, index); 257 258 return index; 259} 260 261static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *); 262static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *, 263 enum update_reason); 264 265static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg, 266 u32 lmax, u32 weight) 267{ 268 INIT_LIST_HEAD(&agg->active); 269 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); 270 271 agg->lmax = lmax; 272 agg->class_weight = weight; 273} 274 275static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q, 276 u32 lmax, u32 weight) 277{ 278 struct qfq_aggregate *agg; 279 280 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next) 281 if (agg->lmax == lmax && agg->class_weight == weight) 282 return agg; 283 284 return NULL; 285} 286 287 288/* Update aggregate as a function of the new number of classes. */ 289static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg, 290 int new_num_classes) 291{ 292 u32 new_agg_weight; 293 294 if (new_num_classes == q->max_agg_classes) 295 hlist_del_init(&agg->nonfull_next); 296 297 if (agg->num_classes > new_num_classes && 298 new_num_classes == q->max_agg_classes - 1) /* agg no more full */ 299 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); 300 301 /* The next assignment may let 302 * agg->initial_budget > agg->budgetmax 303 * hold, we will take it into account in charge_actual_service(). 304 */ 305 agg->budgetmax = new_num_classes * agg->lmax; 306 new_agg_weight = agg->class_weight * new_num_classes; 307 agg->inv_w = ONE_FP/new_agg_weight; 308 309 if (agg->grp == NULL) { 310 int i = qfq_calc_index(agg->inv_w, agg->budgetmax, 311 q->min_slot_shift); 312 agg->grp = &q->groups[i]; 313 } 314 315 q->wsum += 316 (int) agg->class_weight * (new_num_classes - agg->num_classes); 317 q->iwsum = ONE_FP / q->wsum; 318 319 agg->num_classes = new_num_classes; 320} 321 322/* Add class to aggregate. */ 323static void qfq_add_to_agg(struct qfq_sched *q, 324 struct qfq_aggregate *agg, 325 struct qfq_class *cl) 326{ 327 cl->agg = agg; 328 329 qfq_update_agg(q, agg, agg->num_classes+1); 330 if (cl->qdisc->q.qlen > 0) { /* adding an active class */ 331 list_add_tail(&cl->alist, &agg->active); 332 if (list_first_entry(&agg->active, struct qfq_class, alist) == 333 cl && q->in_serv_agg != agg) /* agg was inactive */ 334 qfq_activate_agg(q, agg, enqueue); /* schedule agg */ 335 } 336} 337 338static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *); 339 340static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg) 341{ 342 if (!hlist_unhashed(&agg->nonfull_next)) 343 hlist_del_init(&agg->nonfull_next); 344 q->wsum -= agg->class_weight; 345 if (q->wsum != 0) 346 q->iwsum = ONE_FP / q->wsum; 347 348 if (q->in_serv_agg == agg) 349 q->in_serv_agg = qfq_choose_next_agg(q); 350 kfree(agg); 351} 352 353/* Deschedule class from within its parent aggregate. */ 354static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl) 355{ 356 struct qfq_aggregate *agg = cl->agg; 357 358 359 list_del(&cl->alist); /* remove from RR queue of the aggregate */ 360 if (list_empty(&agg->active)) /* agg is now inactive */ 361 qfq_deactivate_agg(q, agg); 362} 363 364/* Remove class from its parent aggregate. */ 365static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) 366{ 367 struct qfq_aggregate *agg = cl->agg; 368 369 cl->agg = NULL; 370 if (agg->num_classes == 1) { /* agg being emptied, destroy it */ 371 qfq_destroy_agg(q, agg); 372 return; 373 } 374 qfq_update_agg(q, agg, agg->num_classes-1); 375} 376 377/* Deschedule class and remove it from its parent aggregate. */ 378static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) 379{ 380 if (cl->qdisc->q.qlen > 0) /* class is active */ 381 qfq_deactivate_class(q, cl); 382 383 qfq_rm_from_agg(q, cl); 384} 385 386/* Move class to a new aggregate, matching the new class weight and/or lmax */ 387static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight, 388 u32 lmax) 389{ 390 struct qfq_sched *q = qdisc_priv(sch); 391 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight); 392 393 if (new_agg == NULL) { /* create new aggregate */ 394 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC); 395 if (new_agg == NULL) 396 return -ENOBUFS; 397 qfq_init_agg(q, new_agg, lmax, weight); 398 } 399 qfq_deact_rm_from_agg(q, cl); 400 qfq_add_to_agg(q, new_agg, cl); 401 402 return 0; 403} 404 405static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, 406 struct nlattr **tca, unsigned long *arg) 407{ 408 struct qfq_sched *q = qdisc_priv(sch); 409 struct qfq_class *cl = (struct qfq_class *)*arg; 410 bool existing = false; 411 struct nlattr *tb[TCA_QFQ_MAX + 1]; 412 struct qfq_aggregate *new_agg = NULL; 413 u32 weight, lmax, inv_w; 414 int err; 415 int delta_w; 416 417 if (tca[TCA_OPTIONS] == NULL) { 418 pr_notice("qfq: no options\n"); 419 return -EINVAL; 420 } 421 422 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy); 423 if (err < 0) 424 return err; 425 426 if (tb[TCA_QFQ_WEIGHT]) { 427 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]); 428 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) { 429 pr_notice("qfq: invalid weight %u\n", weight); 430 return -EINVAL; 431 } 432 } else 433 weight = 1; 434 435 if (tb[TCA_QFQ_LMAX]) { 436 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]); 437 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) { 438 pr_notice("qfq: invalid max length %u\n", lmax); 439 return -EINVAL; 440 } 441 } else 442 lmax = psched_mtu(qdisc_dev(sch)); 443 444 inv_w = ONE_FP / weight; 445 weight = ONE_FP / inv_w; 446 447 if (cl != NULL && 448 lmax == cl->agg->lmax && 449 weight == cl->agg->class_weight) 450 return 0; /* nothing to change */ 451 452 delta_w = weight - (cl ? cl->agg->class_weight : 0); 453 454 if (q->wsum + delta_w > QFQ_MAX_WSUM) { 455 pr_notice("qfq: total weight out of range (%d + %u)\n", 456 delta_w, q->wsum); 457 return -EINVAL; 458 } 459 460 if (cl != NULL) { /* modify existing class */ 461 if (tca[TCA_RATE]) { 462 err = gen_replace_estimator(&cl->bstats, NULL, 463 &cl->rate_est, 464 qdisc_root_sleeping_lock(sch), 465 tca[TCA_RATE]); 466 if (err) 467 return err; 468 } 469 existing = true; 470 goto set_change_agg; 471 } 472 473 /* create and init new class */ 474 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL); 475 if (cl == NULL) 476 return -ENOBUFS; 477 478 cl->refcnt = 1; 479 cl->common.classid = classid; 480 cl->deficit = lmax; 481 482 cl->qdisc = qdisc_create_dflt(sch->dev_queue, 483 &pfifo_qdisc_ops, classid); 484 if (cl->qdisc == NULL) 485 cl->qdisc = &noop_qdisc; 486 487 if (tca[TCA_RATE]) { 488 err = gen_new_estimator(&cl->bstats, NULL, 489 &cl->rate_est, 490 qdisc_root_sleeping_lock(sch), 491 tca[TCA_RATE]); 492 if (err) 493 goto destroy_class; 494 } 495 496 sch_tree_lock(sch); 497 qdisc_class_hash_insert(&q->clhash, &cl->common); 498 sch_tree_unlock(sch); 499 500 qdisc_class_hash_grow(sch, &q->clhash); 501 502set_change_agg: 503 sch_tree_lock(sch); 504 new_agg = qfq_find_agg(q, lmax, weight); 505 if (new_agg == NULL) { /* create new aggregate */ 506 sch_tree_unlock(sch); 507 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL); 508 if (new_agg == NULL) { 509 err = -ENOBUFS; 510 gen_kill_estimator(&cl->bstats, &cl->rate_est); 511 goto destroy_class; 512 } 513 sch_tree_lock(sch); 514 qfq_init_agg(q, new_agg, lmax, weight); 515 } 516 if (existing) 517 qfq_deact_rm_from_agg(q, cl); 518 qfq_add_to_agg(q, new_agg, cl); 519 sch_tree_unlock(sch); 520 521 *arg = (unsigned long)cl; 522 return 0; 523 524destroy_class: 525 qdisc_destroy(cl->qdisc); 526 kfree(cl); 527 return err; 528} 529 530static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl) 531{ 532 struct qfq_sched *q = qdisc_priv(sch); 533 534 qfq_rm_from_agg(q, cl); 535 gen_kill_estimator(&cl->bstats, &cl->rate_est); 536 qdisc_destroy(cl->qdisc); 537 kfree(cl); 538} 539 540static int qfq_delete_class(struct Qdisc *sch, unsigned long arg) 541{ 542 struct qfq_sched *q = qdisc_priv(sch); 543 struct qfq_class *cl = (struct qfq_class *)arg; 544 545 if (cl->filter_cnt > 0) 546 return -EBUSY; 547 548 sch_tree_lock(sch); 549 550 qfq_purge_queue(cl); 551 qdisc_class_hash_remove(&q->clhash, &cl->common); 552 553 BUG_ON(--cl->refcnt == 0); 554 /* 555 * This shouldn't happen: we "hold" one cops->get() when called 556 * from tc_ctl_tclass; the destroy method is done from cops->put(). 557 */ 558 559 sch_tree_unlock(sch); 560 return 0; 561} 562 563static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid) 564{ 565 struct qfq_class *cl = qfq_find_class(sch, classid); 566 567 if (cl != NULL) 568 cl->refcnt++; 569 570 return (unsigned long)cl; 571} 572 573static void qfq_put_class(struct Qdisc *sch, unsigned long arg) 574{ 575 struct qfq_class *cl = (struct qfq_class *)arg; 576 577 if (--cl->refcnt == 0) 578 qfq_destroy_class(sch, cl); 579} 580 581static struct tcf_proto __rcu **qfq_tcf_chain(struct Qdisc *sch, 582 unsigned long cl) 583{ 584 struct qfq_sched *q = qdisc_priv(sch); 585 586 if (cl) 587 return NULL; 588 589 return &q->filter_list; 590} 591 592static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent, 593 u32 classid) 594{ 595 struct qfq_class *cl = qfq_find_class(sch, classid); 596 597 if (cl != NULL) 598 cl->filter_cnt++; 599 600 return (unsigned long)cl; 601} 602 603static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg) 604{ 605 struct qfq_class *cl = (struct qfq_class *)arg; 606 607 cl->filter_cnt--; 608} 609 610static int qfq_graft_class(struct Qdisc *sch, unsigned long arg, 611 struct Qdisc *new, struct Qdisc **old) 612{ 613 struct qfq_class *cl = (struct qfq_class *)arg; 614 615 if (new == NULL) { 616 new = qdisc_create_dflt(sch->dev_queue, 617 &pfifo_qdisc_ops, cl->common.classid); 618 if (new == NULL) 619 new = &noop_qdisc; 620 } 621 622 sch_tree_lock(sch); 623 qfq_purge_queue(cl); 624 *old = cl->qdisc; 625 cl->qdisc = new; 626 sch_tree_unlock(sch); 627 return 0; 628} 629 630static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg) 631{ 632 struct qfq_class *cl = (struct qfq_class *)arg; 633 634 return cl->qdisc; 635} 636 637static int qfq_dump_class(struct Qdisc *sch, unsigned long arg, 638 struct sk_buff *skb, struct tcmsg *tcm) 639{ 640 struct qfq_class *cl = (struct qfq_class *)arg; 641 struct nlattr *nest; 642 643 tcm->tcm_parent = TC_H_ROOT; 644 tcm->tcm_handle = cl->common.classid; 645 tcm->tcm_info = cl->qdisc->handle; 646 647 nest = nla_nest_start(skb, TCA_OPTIONS); 648 if (nest == NULL) 649 goto nla_put_failure; 650 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) || 651 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax)) 652 goto nla_put_failure; 653 return nla_nest_end(skb, nest); 654 655nla_put_failure: 656 nla_nest_cancel(skb, nest); 657 return -EMSGSIZE; 658} 659 660static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg, 661 struct gnet_dump *d) 662{ 663 struct qfq_class *cl = (struct qfq_class *)arg; 664 struct tc_qfq_stats xstats; 665 666 memset(&xstats, 0, sizeof(xstats)); 667 668 xstats.weight = cl->agg->class_weight; 669 xstats.lmax = cl->agg->lmax; 670 671 if (gnet_stats_copy_basic(d, NULL, &cl->bstats) < 0 || 672 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 || 673 gnet_stats_copy_queue(d, NULL, 674 &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0) 675 return -1; 676 677 return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); 678} 679 680static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) 681{ 682 struct qfq_sched *q = qdisc_priv(sch); 683 struct qfq_class *cl; 684 unsigned int i; 685 686 if (arg->stop) 687 return; 688 689 for (i = 0; i < q->clhash.hashsize; i++) { 690 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { 691 if (arg->count < arg->skip) { 692 arg->count++; 693 continue; 694 } 695 if (arg->fn(sch, (unsigned long)cl, arg) < 0) { 696 arg->stop = 1; 697 return; 698 } 699 arg->count++; 700 } 701 } 702} 703 704static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch, 705 int *qerr) 706{ 707 struct qfq_sched *q = qdisc_priv(sch); 708 struct qfq_class *cl; 709 struct tcf_result res; 710 struct tcf_proto *fl; 711 int result; 712 713 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) { 714 pr_debug("qfq_classify: found %d\n", skb->priority); 715 cl = qfq_find_class(sch, skb->priority); 716 if (cl != NULL) 717 return cl; 718 } 719 720 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 721 fl = rcu_dereference_bh(q->filter_list); 722 result = tc_classify(skb, fl, &res); 723 if (result >= 0) { 724#ifdef CONFIG_NET_CLS_ACT 725 switch (result) { 726 case TC_ACT_QUEUED: 727 case TC_ACT_STOLEN: 728 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; 729 case TC_ACT_SHOT: 730 return NULL; 731 } 732#endif 733 cl = (struct qfq_class *)res.class; 734 if (cl == NULL) 735 cl = qfq_find_class(sch, res.classid); 736 return cl; 737 } 738 739 return NULL; 740} 741 742/* Generic comparison function, handling wraparound. */ 743static inline int qfq_gt(u64 a, u64 b) 744{ 745 return (s64)(a - b) > 0; 746} 747 748/* Round a precise timestamp to its slotted value. */ 749static inline u64 qfq_round_down(u64 ts, unsigned int shift) 750{ 751 return ts & ~((1ULL << shift) - 1); 752} 753 754/* return the pointer to the group with lowest index in the bitmap */ 755static inline struct qfq_group *qfq_ffs(struct qfq_sched *q, 756 unsigned long bitmap) 757{ 758 int index = __ffs(bitmap); 759 return &q->groups[index]; 760} 761/* Calculate a mask to mimic what would be ffs_from(). */ 762static inline unsigned long mask_from(unsigned long bitmap, int from) 763{ 764 return bitmap & ~((1UL << from) - 1); 765} 766 767/* 768 * The state computation relies on ER=0, IR=1, EB=2, IB=3 769 * First compute eligibility comparing grp->S, q->V, 770 * then check if someone is blocking us and possibly add EB 771 */ 772static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp) 773{ 774 /* if S > V we are not eligible */ 775 unsigned int state = qfq_gt(grp->S, q->V); 776 unsigned long mask = mask_from(q->bitmaps[ER], grp->index); 777 struct qfq_group *next; 778 779 if (mask) { 780 next = qfq_ffs(q, mask); 781 if (qfq_gt(grp->F, next->F)) 782 state |= EB; 783 } 784 785 return state; 786} 787 788 789/* 790 * In principle 791 * q->bitmaps[dst] |= q->bitmaps[src] & mask; 792 * q->bitmaps[src] &= ~mask; 793 * but we should make sure that src != dst 794 */ 795static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask, 796 int src, int dst) 797{ 798 q->bitmaps[dst] |= q->bitmaps[src] & mask; 799 q->bitmaps[src] &= ~mask; 800} 801 802static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F) 803{ 804 unsigned long mask = mask_from(q->bitmaps[ER], index + 1); 805 struct qfq_group *next; 806 807 if (mask) { 808 next = qfq_ffs(q, mask); 809 if (!qfq_gt(next->F, old_F)) 810 return; 811 } 812 813 mask = (1UL << index) - 1; 814 qfq_move_groups(q, mask, EB, ER); 815 qfq_move_groups(q, mask, IB, IR); 816} 817 818/* 819 * perhaps 820 * 821 old_V ^= q->V; 822 old_V >>= q->min_slot_shift; 823 if (old_V) { 824 ... 825 } 826 * 827 */ 828static void qfq_make_eligible(struct qfq_sched *q) 829{ 830 unsigned long vslot = q->V >> q->min_slot_shift; 831 unsigned long old_vslot = q->oldV >> q->min_slot_shift; 832 833 if (vslot != old_vslot) { 834 unsigned long mask; 835 int last_flip_pos = fls(vslot ^ old_vslot); 836 837 if (last_flip_pos > 31) /* higher than the number of groups */ 838 mask = ~0UL; /* make all groups eligible */ 839 else 840 mask = (1UL << last_flip_pos) - 1; 841 842 qfq_move_groups(q, mask, IR, ER); 843 qfq_move_groups(q, mask, IB, EB); 844 } 845} 846 847/* 848 * The index of the slot in which the input aggregate agg is to be 849 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2' 850 * and not a '-1' because the start time of the group may be moved 851 * backward by one slot after the aggregate has been inserted, and 852 * this would cause non-empty slots to be right-shifted by one 853 * position. 854 * 855 * QFQ+ fully satisfies this bound to the slot index if the parameters 856 * of the classes are not changed dynamically, and if QFQ+ never 857 * happens to postpone the service of agg unjustly, i.e., it never 858 * happens that the aggregate becomes backlogged and eligible, or just 859 * eligible, while an aggregate with a higher approximated finish time 860 * is being served. In particular, in this case QFQ+ guarantees that 861 * the timestamps of agg are low enough that the slot index is never 862 * higher than 2. Unfortunately, QFQ+ cannot provide the same 863 * guarantee if it happens to unjustly postpone the service of agg, or 864 * if the parameters of some class are changed. 865 * 866 * As for the first event, i.e., an out-of-order service, the 867 * upper bound to the slot index guaranteed by QFQ+ grows to 868 * 2 + 869 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) * 870 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1. 871 * 872 * The following function deals with this problem by backward-shifting 873 * the timestamps of agg, if needed, so as to guarantee that the slot 874 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may 875 * cause the service of other aggregates to be postponed, yet the 876 * worst-case guarantees of these aggregates are not violated. In 877 * fact, in case of no out-of-order service, the timestamps of agg 878 * would have been even lower than they are after the backward shift, 879 * because QFQ+ would have guaranteed a maximum value equal to 2 for 880 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose 881 * service is postponed because of the backward-shift would have 882 * however waited for the service of agg before being served. 883 * 884 * The other event that may cause the slot index to be higher than 2 885 * for agg is a recent change of the parameters of some class. If the 886 * weight of a class is increased or the lmax (max_pkt_size) of the 887 * class is decreased, then a new aggregate with smaller slot size 888 * than the original parent aggregate of the class may happen to be 889 * activated. The activation of this aggregate should be properly 890 * delayed to when the service of the class has finished in the ideal 891 * system tracked by QFQ+. If the activation of the aggregate is not 892 * delayed to this reference time instant, then this aggregate may be 893 * unjustly served before other aggregates waiting for service. This 894 * may cause the above bound to the slot index to be violated for some 895 * of these unlucky aggregates. 896 * 897 * Instead of delaying the activation of the new aggregate, which is 898 * quite complex, the above-discussed capping of the slot index is 899 * used to handle also the consequences of a change of the parameters 900 * of a class. 901 */ 902static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg, 903 u64 roundedS) 904{ 905 u64 slot = (roundedS - grp->S) >> grp->slot_shift; 906 unsigned int i; /* slot index in the bucket list */ 907 908 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) { 909 u64 deltaS = roundedS - grp->S - 910 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift); 911 agg->S -= deltaS; 912 agg->F -= deltaS; 913 slot = QFQ_MAX_SLOTS - 2; 914 } 915 916 i = (grp->front + slot) % QFQ_MAX_SLOTS; 917 918 hlist_add_head(&agg->next, &grp->slots[i]); 919 __set_bit(slot, &grp->full_slots); 920} 921 922/* Maybe introduce hlist_first_entry?? */ 923static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp) 924{ 925 return hlist_entry(grp->slots[grp->front].first, 926 struct qfq_aggregate, next); 927} 928 929/* 930 * remove the entry from the slot 931 */ 932static void qfq_front_slot_remove(struct qfq_group *grp) 933{ 934 struct qfq_aggregate *agg = qfq_slot_head(grp); 935 936 BUG_ON(!agg); 937 hlist_del(&agg->next); 938 if (hlist_empty(&grp->slots[grp->front])) 939 __clear_bit(0, &grp->full_slots); 940} 941 942/* 943 * Returns the first aggregate in the first non-empty bucket of the 944 * group. As a side effect, adjusts the bucket list so the first 945 * non-empty bucket is at position 0 in full_slots. 946 */ 947static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp) 948{ 949 unsigned int i; 950 951 pr_debug("qfq slot_scan: grp %u full %#lx\n", 952 grp->index, grp->full_slots); 953 954 if (grp->full_slots == 0) 955 return NULL; 956 957 i = __ffs(grp->full_slots); /* zero based */ 958 if (i > 0) { 959 grp->front = (grp->front + i) % QFQ_MAX_SLOTS; 960 grp->full_slots >>= i; 961 } 962 963 return qfq_slot_head(grp); 964} 965 966/* 967 * adjust the bucket list. When the start time of a group decreases, 968 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to 969 * move the objects. The mask of occupied slots must be shifted 970 * because we use ffs() to find the first non-empty slot. 971 * This covers decreases in the group's start time, but what about 972 * increases of the start time ? 973 * Here too we should make sure that i is less than 32 974 */ 975static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS) 976{ 977 unsigned int i = (grp->S - roundedS) >> grp->slot_shift; 978 979 grp->full_slots <<= i; 980 grp->front = (grp->front - i) % QFQ_MAX_SLOTS; 981} 982 983static void qfq_update_eligible(struct qfq_sched *q) 984{ 985 struct qfq_group *grp; 986 unsigned long ineligible; 987 988 ineligible = q->bitmaps[IR] | q->bitmaps[IB]; 989 if (ineligible) { 990 if (!q->bitmaps[ER]) { 991 grp = qfq_ffs(q, ineligible); 992 if (qfq_gt(grp->S, q->V)) 993 q->V = grp->S; 994 } 995 qfq_make_eligible(q); 996 } 997} 998 999/* Dequeue head packet of the head class in the DRR queue of the aggregate. */ 1000static void agg_dequeue(struct qfq_aggregate *agg, 1001 struct qfq_class *cl, unsigned int len) 1002{ 1003 qdisc_dequeue_peeked(cl->qdisc); 1004 1005 cl->deficit -= (int) len; 1006 1007 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */ 1008 list_del(&cl->alist); 1009 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) { 1010 cl->deficit += agg->lmax; 1011 list_move_tail(&cl->alist, &agg->active); 1012 } 1013} 1014 1015static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg, 1016 struct qfq_class **cl, 1017 unsigned int *len) 1018{ 1019 struct sk_buff *skb; 1020 1021 *cl = list_first_entry(&agg->active, struct qfq_class, alist); 1022 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc); 1023 if (skb == NULL) 1024 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n"); 1025 else 1026 *len = qdisc_pkt_len(skb); 1027 1028 return skb; 1029} 1030 1031/* Update F according to the actual service received by the aggregate. */ 1032static inline void charge_actual_service(struct qfq_aggregate *agg) 1033{ 1034 /* Compute the service received by the aggregate, taking into 1035 * account that, after decreasing the number of classes in 1036 * agg, it may happen that 1037 * agg->initial_budget - agg->budget > agg->bugdetmax 1038 */ 1039 u32 service_received = min(agg->budgetmax, 1040 agg->initial_budget - agg->budget); 1041 1042 agg->F = agg->S + (u64)service_received * agg->inv_w; 1043} 1044 1045/* Assign a reasonable start time for a new aggregate in group i. 1046 * Admissible values for \hat(F) are multiples of \sigma_i 1047 * no greater than V+\sigma_i . Larger values mean that 1048 * we had a wraparound so we consider the timestamp to be stale. 1049 * 1050 * If F is not stale and F >= V then we set S = F. 1051 * Otherwise we should assign S = V, but this may violate 1052 * the ordering in EB (see [2]). So, if we have groups in ER, 1053 * set S to the F_j of the first group j which would be blocking us. 1054 * We are guaranteed not to move S backward because 1055 * otherwise our group i would still be blocked. 1056 */ 1057static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg) 1058{ 1059 unsigned long mask; 1060 u64 limit, roundedF; 1061 int slot_shift = agg->grp->slot_shift; 1062 1063 roundedF = qfq_round_down(agg->F, slot_shift); 1064 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift); 1065 1066 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) { 1067 /* timestamp was stale */ 1068 mask = mask_from(q->bitmaps[ER], agg->grp->index); 1069 if (mask) { 1070 struct qfq_group *next = qfq_ffs(q, mask); 1071 if (qfq_gt(roundedF, next->F)) { 1072 if (qfq_gt(limit, next->F)) 1073 agg->S = next->F; 1074 else /* preserve timestamp correctness */ 1075 agg->S = limit; 1076 return; 1077 } 1078 } 1079 agg->S = q->V; 1080 } else /* timestamp is not stale */ 1081 agg->S = agg->F; 1082} 1083 1084/* Update the timestamps of agg before scheduling/rescheduling it for 1085 * service. In particular, assign to agg->F its maximum possible 1086 * value, i.e., the virtual finish time with which the aggregate 1087 * should be labeled if it used all its budget once in service. 1088 */ 1089static inline void 1090qfq_update_agg_ts(struct qfq_sched *q, 1091 struct qfq_aggregate *agg, enum update_reason reason) 1092{ 1093 if (reason != requeue) 1094 qfq_update_start(q, agg); 1095 else /* just charge agg for the service received */ 1096 agg->S = agg->F; 1097 1098 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w; 1099} 1100 1101static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg); 1102 1103static struct sk_buff *qfq_dequeue(struct Qdisc *sch) 1104{ 1105 struct qfq_sched *q = qdisc_priv(sch); 1106 struct qfq_aggregate *in_serv_agg = q->in_serv_agg; 1107 struct qfq_class *cl; 1108 struct sk_buff *skb = NULL; 1109 /* next-packet len, 0 means no more active classes in in-service agg */ 1110 unsigned int len = 0; 1111 1112 if (in_serv_agg == NULL) 1113 return NULL; 1114 1115 if (!list_empty(&in_serv_agg->active)) 1116 skb = qfq_peek_skb(in_serv_agg, &cl, &len); 1117 1118 /* 1119 * If there are no active classes in the in-service aggregate, 1120 * or if the aggregate has not enough budget to serve its next 1121 * class, then choose the next aggregate to serve. 1122 */ 1123 if (len == 0 || in_serv_agg->budget < len) { 1124 charge_actual_service(in_serv_agg); 1125 1126 /* recharge the budget of the aggregate */ 1127 in_serv_agg->initial_budget = in_serv_agg->budget = 1128 in_serv_agg->budgetmax; 1129 1130 if (!list_empty(&in_serv_agg->active)) { 1131 /* 1132 * Still active: reschedule for 1133 * service. Possible optimization: if no other 1134 * aggregate is active, then there is no point 1135 * in rescheduling this aggregate, and we can 1136 * just keep it as the in-service one. This 1137 * should be however a corner case, and to 1138 * handle it, we would need to maintain an 1139 * extra num_active_aggs field. 1140 */ 1141 qfq_update_agg_ts(q, in_serv_agg, requeue); 1142 qfq_schedule_agg(q, in_serv_agg); 1143 } else if (sch->q.qlen == 0) { /* no aggregate to serve */ 1144 q->in_serv_agg = NULL; 1145 return NULL; 1146 } 1147 1148 /* 1149 * If we get here, there are other aggregates queued: 1150 * choose the new aggregate to serve. 1151 */ 1152 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q); 1153 skb = qfq_peek_skb(in_serv_agg, &cl, &len); 1154 } 1155 if (!skb) 1156 return NULL; 1157 1158 sch->q.qlen--; 1159 qdisc_bstats_update(sch, skb); 1160 1161 agg_dequeue(in_serv_agg, cl, len); 1162 /* If lmax is lowered, through qfq_change_class, for a class 1163 * owning pending packets with larger size than the new value 1164 * of lmax, then the following condition may hold. 1165 */ 1166 if (unlikely(in_serv_agg->budget < len)) 1167 in_serv_agg->budget = 0; 1168 else 1169 in_serv_agg->budget -= len; 1170 1171 q->V += (u64)len * q->iwsum; 1172 pr_debug("qfq dequeue: len %u F %lld now %lld\n", 1173 len, (unsigned long long) in_serv_agg->F, 1174 (unsigned long long) q->V); 1175 1176 return skb; 1177} 1178 1179static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q) 1180{ 1181 struct qfq_group *grp; 1182 struct qfq_aggregate *agg, *new_front_agg; 1183 u64 old_F; 1184 1185 qfq_update_eligible(q); 1186 q->oldV = q->V; 1187 1188 if (!q->bitmaps[ER]) 1189 return NULL; 1190 1191 grp = qfq_ffs(q, q->bitmaps[ER]); 1192 old_F = grp->F; 1193 1194 agg = qfq_slot_head(grp); 1195 1196 /* agg starts to be served, remove it from schedule */ 1197 qfq_front_slot_remove(grp); 1198 1199 new_front_agg = qfq_slot_scan(grp); 1200 1201 if (new_front_agg == NULL) /* group is now inactive, remove from ER */ 1202 __clear_bit(grp->index, &q->bitmaps[ER]); 1203 else { 1204 u64 roundedS = qfq_round_down(new_front_agg->S, 1205 grp->slot_shift); 1206 unsigned int s; 1207 1208 if (grp->S == roundedS) 1209 return agg; 1210 grp->S = roundedS; 1211 grp->F = roundedS + (2ULL << grp->slot_shift); 1212 __clear_bit(grp->index, &q->bitmaps[ER]); 1213 s = qfq_calc_state(q, grp); 1214 __set_bit(grp->index, &q->bitmaps[s]); 1215 } 1216 1217 qfq_unblock_groups(q, grp->index, old_F); 1218 1219 return agg; 1220} 1221 1222static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch) 1223{ 1224 struct qfq_sched *q = qdisc_priv(sch); 1225 struct qfq_class *cl; 1226 struct qfq_aggregate *agg; 1227 int err = 0; 1228 1229 cl = qfq_classify(skb, sch, &err); 1230 if (cl == NULL) { 1231 if (err & __NET_XMIT_BYPASS) 1232 qdisc_qstats_drop(sch); 1233 kfree_skb(skb); 1234 return err; 1235 } 1236 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid); 1237 1238 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) { 1239 pr_debug("qfq: increasing maxpkt from %u to %u for class %u", 1240 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid); 1241 err = qfq_change_agg(sch, cl, cl->agg->class_weight, 1242 qdisc_pkt_len(skb)); 1243 if (err) 1244 return err; 1245 } 1246 1247 err = qdisc_enqueue(skb, cl->qdisc); 1248 if (unlikely(err != NET_XMIT_SUCCESS)) { 1249 pr_debug("qfq_enqueue: enqueue failed %d\n", err); 1250 if (net_xmit_drop_count(err)) { 1251 cl->qstats.drops++; 1252 qdisc_qstats_drop(sch); 1253 } 1254 return err; 1255 } 1256 1257 bstats_update(&cl->bstats, skb); 1258 ++sch->q.qlen; 1259 1260 agg = cl->agg; 1261 /* if the queue was not empty, then done here */ 1262 if (cl->qdisc->q.qlen != 1) { 1263 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) && 1264 list_first_entry(&agg->active, struct qfq_class, alist) 1265 == cl && cl->deficit < qdisc_pkt_len(skb)) 1266 list_move_tail(&cl->alist, &agg->active); 1267 1268 return err; 1269 } 1270 1271 /* schedule class for service within the aggregate */ 1272 cl->deficit = agg->lmax; 1273 list_add_tail(&cl->alist, &agg->active); 1274 1275 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl || 1276 q->in_serv_agg == agg) 1277 return err; /* non-empty or in service, nothing else to do */ 1278 1279 qfq_activate_agg(q, agg, enqueue); 1280 1281 return err; 1282} 1283 1284/* 1285 * Schedule aggregate according to its timestamps. 1286 */ 1287static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg) 1288{ 1289 struct qfq_group *grp = agg->grp; 1290 u64 roundedS; 1291 int s; 1292 1293 roundedS = qfq_round_down(agg->S, grp->slot_shift); 1294 1295 /* 1296 * Insert agg in the correct bucket. 1297 * If agg->S >= grp->S we don't need to adjust the 1298 * bucket list and simply go to the insertion phase. 1299 * Otherwise grp->S is decreasing, we must make room 1300 * in the bucket list, and also recompute the group state. 1301 * Finally, if there were no flows in this group and nobody 1302 * was in ER make sure to adjust V. 1303 */ 1304 if (grp->full_slots) { 1305 if (!qfq_gt(grp->S, agg->S)) 1306 goto skip_update; 1307 1308 /* create a slot for this agg->S */ 1309 qfq_slot_rotate(grp, roundedS); 1310 /* group was surely ineligible, remove */ 1311 __clear_bit(grp->index, &q->bitmaps[IR]); 1312 __clear_bit(grp->index, &q->bitmaps[IB]); 1313 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) && 1314 q->in_serv_agg == NULL) 1315 q->V = roundedS; 1316 1317 grp->S = roundedS; 1318 grp->F = roundedS + (2ULL << grp->slot_shift); 1319 s = qfq_calc_state(q, grp); 1320 __set_bit(grp->index, &q->bitmaps[s]); 1321 1322 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n", 1323 s, q->bitmaps[s], 1324 (unsigned long long) agg->S, 1325 (unsigned long long) agg->F, 1326 (unsigned long long) q->V); 1327 1328skip_update: 1329 qfq_slot_insert(grp, agg, roundedS); 1330} 1331 1332 1333/* Update agg ts and schedule agg for service */ 1334static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg, 1335 enum update_reason reason) 1336{ 1337 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */ 1338 1339 qfq_update_agg_ts(q, agg, reason); 1340 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */ 1341 q->in_serv_agg = agg; /* start serving this aggregate */ 1342 /* update V: to be in service, agg must be eligible */ 1343 q->oldV = q->V = agg->S; 1344 } else if (agg != q->in_serv_agg) 1345 qfq_schedule_agg(q, agg); 1346} 1347 1348static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp, 1349 struct qfq_aggregate *agg) 1350{ 1351 unsigned int i, offset; 1352 u64 roundedS; 1353 1354 roundedS = qfq_round_down(agg->S, grp->slot_shift); 1355 offset = (roundedS - grp->S) >> grp->slot_shift; 1356 1357 i = (grp->front + offset) % QFQ_MAX_SLOTS; 1358 1359 hlist_del(&agg->next); 1360 if (hlist_empty(&grp->slots[i])) 1361 __clear_bit(offset, &grp->full_slots); 1362} 1363 1364/* 1365 * Called to forcibly deschedule an aggregate. If the aggregate is 1366 * not in the front bucket, or if the latter has other aggregates in 1367 * the front bucket, we can simply remove the aggregate with no other 1368 * side effects. 1369 * Otherwise we must propagate the event up. 1370 */ 1371static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg) 1372{ 1373 struct qfq_group *grp = agg->grp; 1374 unsigned long mask; 1375 u64 roundedS; 1376 int s; 1377 1378 if (agg == q->in_serv_agg) { 1379 charge_actual_service(agg); 1380 q->in_serv_agg = qfq_choose_next_agg(q); 1381 return; 1382 } 1383 1384 agg->F = agg->S; 1385 qfq_slot_remove(q, grp, agg); 1386 1387 if (!grp->full_slots) { 1388 __clear_bit(grp->index, &q->bitmaps[IR]); 1389 __clear_bit(grp->index, &q->bitmaps[EB]); 1390 __clear_bit(grp->index, &q->bitmaps[IB]); 1391 1392 if (test_bit(grp->index, &q->bitmaps[ER]) && 1393 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) { 1394 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1); 1395 if (mask) 1396 mask = ~((1UL << __fls(mask)) - 1); 1397 else 1398 mask = ~0UL; 1399 qfq_move_groups(q, mask, EB, ER); 1400 qfq_move_groups(q, mask, IB, IR); 1401 } 1402 __clear_bit(grp->index, &q->bitmaps[ER]); 1403 } else if (hlist_empty(&grp->slots[grp->front])) { 1404 agg = qfq_slot_scan(grp); 1405 roundedS = qfq_round_down(agg->S, grp->slot_shift); 1406 if (grp->S != roundedS) { 1407 __clear_bit(grp->index, &q->bitmaps[ER]); 1408 __clear_bit(grp->index, &q->bitmaps[IR]); 1409 __clear_bit(grp->index, &q->bitmaps[EB]); 1410 __clear_bit(grp->index, &q->bitmaps[IB]); 1411 grp->S = roundedS; 1412 grp->F = roundedS + (2ULL << grp->slot_shift); 1413 s = qfq_calc_state(q, grp); 1414 __set_bit(grp->index, &q->bitmaps[s]); 1415 } 1416 } 1417} 1418 1419static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg) 1420{ 1421 struct qfq_sched *q = qdisc_priv(sch); 1422 struct qfq_class *cl = (struct qfq_class *)arg; 1423 1424 if (cl->qdisc->q.qlen == 0) 1425 qfq_deactivate_class(q, cl); 1426} 1427 1428static unsigned int qfq_drop_from_slot(struct qfq_sched *q, 1429 struct hlist_head *slot) 1430{ 1431 struct qfq_aggregate *agg; 1432 struct qfq_class *cl; 1433 unsigned int len; 1434 1435 hlist_for_each_entry(agg, slot, next) { 1436 list_for_each_entry(cl, &agg->active, alist) { 1437 1438 if (!cl->qdisc->ops->drop) 1439 continue; 1440 1441 len = cl->qdisc->ops->drop(cl->qdisc); 1442 if (len > 0) { 1443 if (cl->qdisc->q.qlen == 0) 1444 qfq_deactivate_class(q, cl); 1445 1446 return len; 1447 } 1448 } 1449 } 1450 return 0; 1451} 1452 1453static unsigned int qfq_drop(struct Qdisc *sch) 1454{ 1455 struct qfq_sched *q = qdisc_priv(sch); 1456 struct qfq_group *grp; 1457 unsigned int i, j, len; 1458 1459 for (i = 0; i <= QFQ_MAX_INDEX; i++) { 1460 grp = &q->groups[i]; 1461 for (j = 0; j < QFQ_MAX_SLOTS; j++) { 1462 len = qfq_drop_from_slot(q, &grp->slots[j]); 1463 if (len > 0) { 1464 sch->q.qlen--; 1465 return len; 1466 } 1467 } 1468 1469 } 1470 1471 return 0; 1472} 1473 1474static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt) 1475{ 1476 struct qfq_sched *q = qdisc_priv(sch); 1477 struct qfq_group *grp; 1478 int i, j, err; 1479 u32 max_cl_shift, maxbudg_shift, max_classes; 1480 1481 err = qdisc_class_hash_init(&q->clhash); 1482 if (err < 0) 1483 return err; 1484 1485 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES) 1486 max_classes = QFQ_MAX_AGG_CLASSES; 1487 else 1488 max_classes = qdisc_dev(sch)->tx_queue_len + 1; 1489 /* max_cl_shift = floor(log_2(max_classes)) */ 1490 max_cl_shift = __fls(max_classes); 1491 q->max_agg_classes = 1<<max_cl_shift; 1492 1493 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */ 1494 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift; 1495 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX; 1496 1497 for (i = 0; i <= QFQ_MAX_INDEX; i++) { 1498 grp = &q->groups[i]; 1499 grp->index = i; 1500 grp->slot_shift = q->min_slot_shift + i; 1501 for (j = 0; j < QFQ_MAX_SLOTS; j++) 1502 INIT_HLIST_HEAD(&grp->slots[j]); 1503 } 1504 1505 INIT_HLIST_HEAD(&q->nonfull_aggs); 1506 1507 return 0; 1508} 1509 1510static void qfq_reset_qdisc(struct Qdisc *sch) 1511{ 1512 struct qfq_sched *q = qdisc_priv(sch); 1513 struct qfq_class *cl; 1514 unsigned int i; 1515 1516 for (i = 0; i < q->clhash.hashsize; i++) { 1517 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { 1518 if (cl->qdisc->q.qlen > 0) 1519 qfq_deactivate_class(q, cl); 1520 1521 qdisc_reset(cl->qdisc); 1522 } 1523 } 1524 sch->q.qlen = 0; 1525} 1526 1527static void qfq_destroy_qdisc(struct Qdisc *sch) 1528{ 1529 struct qfq_sched *q = qdisc_priv(sch); 1530 struct qfq_class *cl; 1531 struct hlist_node *next; 1532 unsigned int i; 1533 1534 tcf_destroy_chain(&q->filter_list); 1535 1536 for (i = 0; i < q->clhash.hashsize; i++) { 1537 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i], 1538 common.hnode) { 1539 qfq_destroy_class(sch, cl); 1540 } 1541 } 1542 qdisc_class_hash_destroy(&q->clhash); 1543} 1544 1545static const struct Qdisc_class_ops qfq_class_ops = { 1546 .change = qfq_change_class, 1547 .delete = qfq_delete_class, 1548 .get = qfq_get_class, 1549 .put = qfq_put_class, 1550 .tcf_chain = qfq_tcf_chain, 1551 .bind_tcf = qfq_bind_tcf, 1552 .unbind_tcf = qfq_unbind_tcf, 1553 .graft = qfq_graft_class, 1554 .leaf = qfq_class_leaf, 1555 .qlen_notify = qfq_qlen_notify, 1556 .dump = qfq_dump_class, 1557 .dump_stats = qfq_dump_class_stats, 1558 .walk = qfq_walk, 1559}; 1560 1561static struct Qdisc_ops qfq_qdisc_ops __read_mostly = { 1562 .cl_ops = &qfq_class_ops, 1563 .id = "qfq", 1564 .priv_size = sizeof(struct qfq_sched), 1565 .enqueue = qfq_enqueue, 1566 .dequeue = qfq_dequeue, 1567 .peek = qdisc_peek_dequeued, 1568 .drop = qfq_drop, 1569 .init = qfq_init_qdisc, 1570 .reset = qfq_reset_qdisc, 1571 .destroy = qfq_destroy_qdisc, 1572 .owner = THIS_MODULE, 1573}; 1574 1575static int __init qfq_init(void) 1576{ 1577 return register_qdisc(&qfq_qdisc_ops); 1578} 1579 1580static void __exit qfq_exit(void) 1581{ 1582 unregister_qdisc(&qfq_qdisc_ops); 1583} 1584 1585module_init(qfq_init); 1586module_exit(qfq_exit); 1587MODULE_LICENSE("GPL"); 1588