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