Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * net/sched/sch_cbq.c	Class-Based Queueing discipline.
   4 *
 
 
 
 
 
   5 * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
 
   6 */
   7
   8#include <linux/module.h>
   9#include <linux/slab.h>
  10#include <linux/types.h>
  11#include <linux/kernel.h>
  12#include <linux/string.h>
  13#include <linux/errno.h>
  14#include <linux/skbuff.h>
  15#include <net/netlink.h>
  16#include <net/pkt_sched.h>
  17#include <net/pkt_cls.h>
  18
  19
  20/*	Class-Based Queueing (CBQ) algorithm.
  21	=======================================
  22
  23	Sources: [1] Sally Floyd and Van Jacobson, "Link-sharing and Resource
  24		 Management Models for Packet Networks",
  25		 IEEE/ACM Transactions on Networking, Vol.3, No.4, 1995
  26
  27		 [2] Sally Floyd, "Notes on CBQ and Guaranteed Service", 1995
  28
  29		 [3] Sally Floyd, "Notes on Class-Based Queueing: Setting
  30		 Parameters", 1996
  31
  32		 [4] Sally Floyd and Michael Speer, "Experimental Results
  33		 for Class-Based Queueing", 1998, not published.
  34
  35	-----------------------------------------------------------------------
  36
  37	Algorithm skeleton was taken from NS simulator cbq.cc.
  38	If someone wants to check this code against the LBL version,
  39	he should take into account that ONLY the skeleton was borrowed,
  40	the implementation is different. Particularly:
  41
  42	--- The WRR algorithm is different. Our version looks more
  43	reasonable (I hope) and works when quanta are allowed to be
  44	less than MTU, which is always the case when real time classes
  45	have small rates. Note, that the statement of [3] is
  46	incomplete, delay may actually be estimated even if class
  47	per-round allotment is less than MTU. Namely, if per-round
  48	allotment is W*r_i, and r_1+...+r_k = r < 1
  49
  50	delay_i <= ([MTU/(W*r_i)]*W*r + W*r + k*MTU)/B
  51
  52	In the worst case we have IntServ estimate with D = W*r+k*MTU
  53	and C = MTU*r. The proof (if correct at all) is trivial.
  54
  55
  56	--- It seems that cbq-2.0 is not very accurate. At least, I cannot
  57	interpret some places, which look like wrong translations
  58	from NS. Anyone is advised to find these differences
  59	and explain to me, why I am wrong 8).
  60
  61	--- Linux has no EOI event, so that we cannot estimate true class
  62	idle time. Workaround is to consider the next dequeue event
  63	as sign that previous packet is finished. This is wrong because of
  64	internal device queueing, but on a permanently loaded link it is true.
  65	Moreover, combined with clock integrator, this scheme looks
  66	very close to an ideal solution.  */
  67
  68struct cbq_sched_data;
  69
  70
  71struct cbq_class {
  72	struct Qdisc_class_common common;
  73	struct cbq_class	*next_alive;	/* next class with backlog in this priority band */
  74
  75/* Parameters */
  76	unsigned char		priority;	/* class priority */
  77	unsigned char		priority2;	/* priority to be used after overlimit */
  78	unsigned char		ewma_log;	/* time constant for idle time calculation */
 
 
 
 
  79
  80	u32			defmap;
  81
  82	/* Link-sharing scheduler parameters */
  83	long			maxidle;	/* Class parameters: see below. */
  84	long			offtime;
  85	long			minidle;
  86	u32			avpkt;
  87	struct qdisc_rate_table	*R_tab;
  88
 
 
 
 
  89	/* General scheduler (WRR) parameters */
  90	long			allot;
  91	long			quantum;	/* Allotment per WRR round */
  92	long			weight;		/* Relative allotment: see below */
  93
  94	struct Qdisc		*qdisc;		/* Ptr to CBQ discipline */
  95	struct cbq_class	*split;		/* Ptr to split node */
  96	struct cbq_class	*share;		/* Ptr to LS parent in the class tree */
  97	struct cbq_class	*tparent;	/* Ptr to tree parent in the class tree */
  98	struct cbq_class	*borrow;	/* NULL if class is bandwidth limited;
  99						   parent otherwise */
 100	struct cbq_class	*sibling;	/* Sibling chain */
 101	struct cbq_class	*children;	/* Pointer to children chain */
 102
 103	struct Qdisc		*q;		/* Elementary queueing discipline */
 104
 105
 106/* Variables */
 107	unsigned char		cpriority;	/* Effective priority */
 108	unsigned char		delayed;
 109	unsigned char		level;		/* level of the class in hierarchy:
 110						   0 for leaf classes, and maximal
 111						   level of children + 1 for nodes.
 112						 */
 113
 114	psched_time_t		last;		/* Last end of service */
 115	psched_time_t		undertime;
 116	long			avgidle;
 117	long			deficit;	/* Saved deficit for WRR */
 118	psched_time_t		penalized;
 119	struct gnet_stats_basic_packed bstats;
 120	struct gnet_stats_queue qstats;
 121	struct net_rate_estimator __rcu *rate_est;
 122	struct tc_cbq_xstats	xstats;
 123
 124	struct tcf_proto __rcu	*filter_list;
 125	struct tcf_block	*block;
 126
 
 127	int			filters;
 128
 129	struct cbq_class	*defaults[TC_PRIO_MAX + 1];
 130};
 131
 132struct cbq_sched_data {
 133	struct Qdisc_class_hash	clhash;			/* Hash table of all classes */
 134	int			nclasses[TC_CBQ_MAXPRIO + 1];
 135	unsigned int		quanta[TC_CBQ_MAXPRIO + 1];
 136
 137	struct cbq_class	link;
 138
 139	unsigned int		activemask;
 140	struct cbq_class	*active[TC_CBQ_MAXPRIO + 1];	/* List of all classes
 141								   with backlog */
 142
 143#ifdef CONFIG_NET_CLS_ACT
 144	struct cbq_class	*rx_class;
 145#endif
 146	struct cbq_class	*tx_class;
 147	struct cbq_class	*tx_borrowed;
 148	int			tx_len;
 149	psched_time_t		now;		/* Cached timestamp */
 
 150	unsigned int		pmask;
 151
 152	struct hrtimer		delay_timer;
 153	struct qdisc_watchdog	watchdog;	/* Watchdog timer,
 154						   started when CBQ has
 155						   backlog, but cannot
 156						   transmit just now */
 157	psched_tdiff_t		wd_expires;
 158	int			toplevel;
 159	u32			hgenerator;
 160};
 161
 162
 163#define L2T(cl, len)	qdisc_l2t((cl)->R_tab, len)
 164
 165static inline struct cbq_class *
 166cbq_class_lookup(struct cbq_sched_data *q, u32 classid)
 167{
 168	struct Qdisc_class_common *clc;
 169
 170	clc = qdisc_class_find(&q->clhash, classid);
 171	if (clc == NULL)
 172		return NULL;
 173	return container_of(clc, struct cbq_class, common);
 174}
 175
 176#ifdef CONFIG_NET_CLS_ACT
 177
 178static struct cbq_class *
 179cbq_reclassify(struct sk_buff *skb, struct cbq_class *this)
 180{
 181	struct cbq_class *cl;
 182
 183	for (cl = this->tparent; cl; cl = cl->tparent) {
 184		struct cbq_class *new = cl->defaults[TC_PRIO_BESTEFFORT];
 185
 186		if (new != NULL && new != this)
 187			return new;
 188	}
 189	return NULL;
 190}
 191
 192#endif
 193
 194/* Classify packet. The procedure is pretty complicated, but
 195 * it allows us to combine link sharing and priority scheduling
 196 * transparently.
 197 *
 198 * Namely, you can put link sharing rules (f.e. route based) at root of CBQ,
 199 * so that it resolves to split nodes. Then packets are classified
 200 * by logical priority, or a more specific classifier may be attached
 201 * to the split node.
 202 */
 203
 204static struct cbq_class *
 205cbq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
 206{
 207	struct cbq_sched_data *q = qdisc_priv(sch);
 208	struct cbq_class *head = &q->link;
 209	struct cbq_class **defmap;
 210	struct cbq_class *cl = NULL;
 211	u32 prio = skb->priority;
 212	struct tcf_proto *fl;
 213	struct tcf_result res;
 214
 215	/*
 216	 *  Step 1. If skb->priority points to one of our classes, use it.
 217	 */
 218	if (TC_H_MAJ(prio ^ sch->handle) == 0 &&
 219	    (cl = cbq_class_lookup(q, prio)) != NULL)
 220		return cl;
 221
 222	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 223	for (;;) {
 224		int result = 0;
 225		defmap = head->defaults;
 226
 227		fl = rcu_dereference_bh(head->filter_list);
 228		/*
 229		 * Step 2+n. Apply classifier.
 230		 */
 231		result = tcf_classify(skb, fl, &res, true);
 232		if (!fl || result < 0)
 233			goto fallback;
 234
 235		cl = (void *)res.class;
 236		if (!cl) {
 237			if (TC_H_MAJ(res.classid))
 238				cl = cbq_class_lookup(q, res.classid);
 239			else if ((cl = defmap[res.classid & TC_PRIO_MAX]) == NULL)
 240				cl = defmap[TC_PRIO_BESTEFFORT];
 241
 242			if (cl == NULL)
 243				goto fallback;
 244		}
 245		if (cl->level >= head->level)
 246			goto fallback;
 247#ifdef CONFIG_NET_CLS_ACT
 248		switch (result) {
 249		case TC_ACT_QUEUED:
 250		case TC_ACT_STOLEN:
 251		case TC_ACT_TRAP:
 252			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 253			/* fall through */
 254		case TC_ACT_SHOT:
 255			return NULL;
 256		case TC_ACT_RECLASSIFY:
 257			return cbq_reclassify(skb, cl);
 258		}
 259#endif
 260		if (cl->level == 0)
 261			return cl;
 262
 263		/*
 264		 * Step 3+n. If classifier selected a link sharing class,
 265		 *	   apply agency specific classifier.
 266		 *	   Repeat this procdure until we hit a leaf node.
 267		 */
 268		head = cl;
 269	}
 270
 271fallback:
 272	cl = head;
 273
 274	/*
 275	 * Step 4. No success...
 276	 */
 277	if (TC_H_MAJ(prio) == 0 &&
 278	    !(cl = head->defaults[prio & TC_PRIO_MAX]) &&
 279	    !(cl = head->defaults[TC_PRIO_BESTEFFORT]))
 280		return head;
 281
 282	return cl;
 283}
 284
 285/*
 286 * A packet has just been enqueued on the empty class.
 287 * cbq_activate_class adds it to the tail of active class list
 288 * of its priority band.
 289 */
 290
 291static inline void cbq_activate_class(struct cbq_class *cl)
 292{
 293	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 294	int prio = cl->cpriority;
 295	struct cbq_class *cl_tail;
 296
 297	cl_tail = q->active[prio];
 298	q->active[prio] = cl;
 299
 300	if (cl_tail != NULL) {
 301		cl->next_alive = cl_tail->next_alive;
 302		cl_tail->next_alive = cl;
 303	} else {
 304		cl->next_alive = cl;
 305		q->activemask |= (1<<prio);
 306	}
 307}
 308
 309/*
 310 * Unlink class from active chain.
 311 * Note that this same procedure is done directly in cbq_dequeue*
 312 * during round-robin procedure.
 313 */
 314
 315static void cbq_deactivate_class(struct cbq_class *this)
 316{
 317	struct cbq_sched_data *q = qdisc_priv(this->qdisc);
 318	int prio = this->cpriority;
 319	struct cbq_class *cl;
 320	struct cbq_class *cl_prev = q->active[prio];
 321
 322	do {
 323		cl = cl_prev->next_alive;
 324		if (cl == this) {
 325			cl_prev->next_alive = cl->next_alive;
 326			cl->next_alive = NULL;
 327
 328			if (cl == q->active[prio]) {
 329				q->active[prio] = cl_prev;
 330				if (cl == q->active[prio]) {
 331					q->active[prio] = NULL;
 332					q->activemask &= ~(1<<prio);
 333					return;
 334				}
 335			}
 336			return;
 337		}
 338	} while ((cl_prev = cl) != q->active[prio]);
 339}
 340
 341static void
 342cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl)
 343{
 344	int toplevel = q->toplevel;
 345
 346	if (toplevel > cl->level) {
 347		psched_time_t now = psched_get_time();
 
 
 
 
 
 348
 349		do {
 350			if (cl->undertime < now) {
 351				q->toplevel = cl->level;
 352				return;
 353			}
 354		} while ((cl = cl->borrow) != NULL && toplevel > cl->level);
 355	}
 356}
 357
 358static int
 359cbq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 360	    struct sk_buff **to_free)
 361{
 362	struct cbq_sched_data *q = qdisc_priv(sch);
 363	int uninitialized_var(ret);
 364	struct cbq_class *cl = cbq_classify(skb, sch, &ret);
 365
 366#ifdef CONFIG_NET_CLS_ACT
 367	q->rx_class = cl;
 368#endif
 369	if (cl == NULL) {
 370		if (ret & __NET_XMIT_BYPASS)
 371			qdisc_qstats_drop(sch);
 372		__qdisc_drop(skb, to_free);
 373		return ret;
 374	}
 375
 376	ret = qdisc_enqueue(skb, cl->q, to_free);
 
 
 
 377	if (ret == NET_XMIT_SUCCESS) {
 378		sch->q.qlen++;
 379		cbq_mark_toplevel(q, cl);
 380		if (!cl->next_alive)
 381			cbq_activate_class(cl);
 382		return ret;
 383	}
 384
 385	if (net_xmit_drop_count(ret)) {
 386		qdisc_qstats_drop(sch);
 387		cbq_mark_toplevel(q, cl);
 388		cl->qstats.drops++;
 389	}
 390	return ret;
 391}
 392
 393/* Overlimit action: penalize leaf class by adding offtime */
 394static void cbq_overlimit(struct cbq_class *cl)
 
 
 
 395{
 396	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 397	psched_tdiff_t delay = cl->undertime - q->now;
 398
 399	if (!cl->delayed) {
 400		delay += cl->offtime;
 401
 402		/*
 403		 * Class goes to sleep, so that it will have no
 404		 * chance to work avgidle. Let's forgive it 8)
 405		 *
 406		 * BTW cbq-2.0 has a crap in this
 407		 * place, apparently they forgot to shift it by cl->ewma_log.
 408		 */
 409		if (cl->avgidle < 0)
 410			delay -= (-cl->avgidle) - ((-cl->avgidle) >> cl->ewma_log);
 411		if (cl->avgidle < cl->minidle)
 412			cl->avgidle = cl->minidle;
 413		if (delay <= 0)
 414			delay = 1;
 415		cl->undertime = q->now + delay;
 416
 417		cl->xstats.overactions++;
 418		cl->delayed = 1;
 419	}
 420	if (q->wd_expires == 0 || q->wd_expires > delay)
 421		q->wd_expires = delay;
 422
 423	/* Dirty work! We must schedule wakeups based on
 424	 * real available rate, rather than leaf rate,
 425	 * which may be tiny (even zero).
 426	 */
 427	if (q->toplevel == TC_CBQ_MAXLEVEL) {
 428		struct cbq_class *b;
 429		psched_tdiff_t base_delay = q->wd_expires;
 430
 431		for (b = cl->borrow; b; b = b->borrow) {
 432			delay = b->undertime - q->now;
 433			if (delay < base_delay) {
 434				if (delay <= 0)
 435					delay = 1;
 436				base_delay = delay;
 437			}
 438		}
 439
 440		q->wd_expires = base_delay;
 441	}
 442}
 443
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 444static psched_tdiff_t cbq_undelay_prio(struct cbq_sched_data *q, int prio,
 445				       psched_time_t now)
 446{
 447	struct cbq_class *cl;
 448	struct cbq_class *cl_prev = q->active[prio];
 449	psched_time_t sched = now;
 450
 451	if (cl_prev == NULL)
 452		return 0;
 453
 454	do {
 455		cl = cl_prev->next_alive;
 456		if (now - cl->penalized > 0) {
 457			cl_prev->next_alive = cl->next_alive;
 458			cl->next_alive = NULL;
 459			cl->cpriority = cl->priority;
 460			cl->delayed = 0;
 461			cbq_activate_class(cl);
 462
 463			if (cl == q->active[prio]) {
 464				q->active[prio] = cl_prev;
 465				if (cl == q->active[prio]) {
 466					q->active[prio] = NULL;
 467					return 0;
 468				}
 469			}
 470
 471			cl = cl_prev->next_alive;
 472		} else if (sched - cl->penalized > 0)
 473			sched = cl->penalized;
 474	} while ((cl_prev = cl) != q->active[prio]);
 475
 476	return sched - now;
 477}
 478
 479static enum hrtimer_restart cbq_undelay(struct hrtimer *timer)
 480{
 481	struct cbq_sched_data *q = container_of(timer, struct cbq_sched_data,
 482						delay_timer);
 483	struct Qdisc *sch = q->watchdog.qdisc;
 484	psched_time_t now;
 485	psched_tdiff_t delay = 0;
 486	unsigned int pmask;
 487
 488	now = psched_get_time();
 489
 490	pmask = q->pmask;
 491	q->pmask = 0;
 492
 493	while (pmask) {
 494		int prio = ffz(~pmask);
 495		psched_tdiff_t tmp;
 496
 497		pmask &= ~(1<<prio);
 498
 499		tmp = cbq_undelay_prio(q, prio, now);
 500		if (tmp > 0) {
 501			q->pmask |= 1<<prio;
 502			if (tmp < delay || delay == 0)
 503				delay = tmp;
 504		}
 505	}
 506
 507	if (delay) {
 508		ktime_t time;
 509
 510		time = 0;
 511		time = ktime_add_ns(time, PSCHED_TICKS2NS(now + delay));
 512		hrtimer_start(&q->delay_timer, time, HRTIMER_MODE_ABS_PINNED);
 513	}
 514
 
 515	__netif_schedule(qdisc_root(sch));
 516	return HRTIMER_NORESTART;
 517}
 518
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 519/*
 520 * It is mission critical procedure.
 521 *
 522 * We "regenerate" toplevel cutoff, if transmitting class
 523 * has backlog and it is not regulated. It is not part of
 524 * original CBQ description, but looks more reasonable.
 525 * Probably, it is wrong. This question needs further investigation.
 526 */
 527
 528static inline void
 529cbq_update_toplevel(struct cbq_sched_data *q, struct cbq_class *cl,
 530		    struct cbq_class *borrowed)
 531{
 532	if (cl && q->toplevel >= borrowed->level) {
 533		if (cl->q->q.qlen > 1) {
 534			do {
 535				if (borrowed->undertime == PSCHED_PASTPERFECT) {
 536					q->toplevel = borrowed->level;
 537					return;
 538				}
 539			} while ((borrowed = borrowed->borrow) != NULL);
 540		}
 541#if 0
 542	/* It is not necessary now. Uncommenting it
 543	   will save CPU cycles, but decrease fairness.
 544	 */
 545		q->toplevel = TC_CBQ_MAXLEVEL;
 546#endif
 547	}
 548}
 549
 550static void
 551cbq_update(struct cbq_sched_data *q)
 552{
 553	struct cbq_class *this = q->tx_class;
 554	struct cbq_class *cl = this;
 555	int len = q->tx_len;
 556	psched_time_t now;
 557
 558	q->tx_class = NULL;
 559	/* Time integrator. We calculate EOS time
 560	 * by adding expected packet transmission time.
 561	 */
 562	now = q->now + L2T(&q->link, len);
 563
 564	for ( ; cl; cl = cl->share) {
 565		long avgidle = cl->avgidle;
 566		long idle;
 567
 568		cl->bstats.packets++;
 569		cl->bstats.bytes += len;
 570
 571		/*
 572		 * (now - last) is total time between packet right edges.
 573		 * (last_pktlen/rate) is "virtual" busy time, so that
 574		 *
 575		 *	idle = (now - last) - last_pktlen/rate
 576		 */
 577
 578		idle = now - cl->last;
 579		if ((unsigned long)idle > 128*1024*1024) {
 580			avgidle = cl->maxidle;
 581		} else {
 582			idle -= L2T(cl, len);
 583
 584		/* true_avgidle := (1-W)*true_avgidle + W*idle,
 585		 * where W=2^{-ewma_log}. But cl->avgidle is scaled:
 586		 * cl->avgidle == true_avgidle/W,
 587		 * hence:
 588		 */
 589			avgidle += idle - (avgidle>>cl->ewma_log);
 590		}
 591
 592		if (avgidle <= 0) {
 593			/* Overlimit or at-limit */
 594
 595			if (avgidle < cl->minidle)
 596				avgidle = cl->minidle;
 597
 598			cl->avgidle = avgidle;
 599
 600			/* Calculate expected time, when this class
 601			 * will be allowed to send.
 602			 * It will occur, when:
 603			 * (1-W)*true_avgidle + W*delay = 0, i.e.
 604			 * idle = (1/W - 1)*(-true_avgidle)
 605			 * or
 606			 * idle = (1 - W)*(-cl->avgidle);
 607			 */
 608			idle = (-avgidle) - ((-avgidle) >> cl->ewma_log);
 609
 610			/*
 611			 * That is not all.
 612			 * To maintain the rate allocated to the class,
 613			 * we add to undertime virtual clock,
 614			 * necessary to complete transmitted packet.
 615			 * (len/phys_bandwidth has been already passed
 616			 * to the moment of cbq_update)
 617			 */
 618
 619			idle -= L2T(&q->link, len);
 620			idle += L2T(cl, len);
 621
 622			cl->undertime = now + idle;
 623		} else {
 624			/* Underlimit */
 625
 626			cl->undertime = PSCHED_PASTPERFECT;
 627			if (avgidle > cl->maxidle)
 628				cl->avgidle = cl->maxidle;
 629			else
 630				cl->avgidle = avgidle;
 631		}
 632		if ((s64)(now - cl->last) > 0)
 633			cl->last = now;
 634	}
 635
 636	cbq_update_toplevel(q, this, q->tx_borrowed);
 637}
 638
 639static inline struct cbq_class *
 640cbq_under_limit(struct cbq_class *cl)
 641{
 642	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 643	struct cbq_class *this_cl = cl;
 644
 645	if (cl->tparent == NULL)
 646		return cl;
 647
 648	if (cl->undertime == PSCHED_PASTPERFECT || q->now >= cl->undertime) {
 649		cl->delayed = 0;
 650		return cl;
 651	}
 652
 653	do {
 654		/* It is very suspicious place. Now overlimit
 655		 * action is generated for not bounded classes
 656		 * only if link is completely congested.
 657		 * Though it is in agree with ancestor-only paradigm,
 658		 * it looks very stupid. Particularly,
 659		 * it means that this chunk of code will either
 660		 * never be called or result in strong amplification
 661		 * of burstiness. Dangerous, silly, and, however,
 662		 * no another solution exists.
 663		 */
 664		cl = cl->borrow;
 665		if (!cl) {
 666			this_cl->qstats.overlimits++;
 667			cbq_overlimit(this_cl);
 668			return NULL;
 669		}
 670		if (cl->level > q->toplevel)
 671			return NULL;
 672	} while (cl->undertime != PSCHED_PASTPERFECT && q->now < cl->undertime);
 673
 674	cl->delayed = 0;
 675	return cl;
 676}
 677
 678static inline struct sk_buff *
 679cbq_dequeue_prio(struct Qdisc *sch, int prio)
 680{
 681	struct cbq_sched_data *q = qdisc_priv(sch);
 682	struct cbq_class *cl_tail, *cl_prev, *cl;
 683	struct sk_buff *skb;
 684	int deficit;
 685
 686	cl_tail = cl_prev = q->active[prio];
 687	cl = cl_prev->next_alive;
 688
 689	do {
 690		deficit = 0;
 691
 692		/* Start round */
 693		do {
 694			struct cbq_class *borrow = cl;
 695
 696			if (cl->q->q.qlen &&
 697			    (borrow = cbq_under_limit(cl)) == NULL)
 698				goto skip_class;
 699
 700			if (cl->deficit <= 0) {
 701				/* Class exhausted its allotment per
 702				 * this round. Switch to the next one.
 703				 */
 704				deficit = 1;
 705				cl->deficit += cl->quantum;
 706				goto next_class;
 707			}
 708
 709			skb = cl->q->dequeue(cl->q);
 710
 711			/* Class did not give us any skb :-(
 712			 * It could occur even if cl->q->q.qlen != 0
 713			 * f.e. if cl->q == "tbf"
 714			 */
 715			if (skb == NULL)
 716				goto skip_class;
 717
 718			cl->deficit -= qdisc_pkt_len(skb);
 719			q->tx_class = cl;
 720			q->tx_borrowed = borrow;
 721			if (borrow != cl) {
 722#ifndef CBQ_XSTATS_BORROWS_BYTES
 723				borrow->xstats.borrows++;
 724				cl->xstats.borrows++;
 725#else
 726				borrow->xstats.borrows += qdisc_pkt_len(skb);
 727				cl->xstats.borrows += qdisc_pkt_len(skb);
 728#endif
 729			}
 730			q->tx_len = qdisc_pkt_len(skb);
 731
 732			if (cl->deficit <= 0) {
 733				q->active[prio] = cl;
 734				cl = cl->next_alive;
 735				cl->deficit += cl->quantum;
 736			}
 737			return skb;
 738
 739skip_class:
 740			if (cl->q->q.qlen == 0 || prio != cl->cpriority) {
 741				/* Class is empty or penalized.
 742				 * Unlink it from active chain.
 743				 */
 744				cl_prev->next_alive = cl->next_alive;
 745				cl->next_alive = NULL;
 746
 747				/* Did cl_tail point to it? */
 748				if (cl == cl_tail) {
 749					/* Repair it! */
 750					cl_tail = cl_prev;
 751
 752					/* Was it the last class in this band? */
 753					if (cl == cl_tail) {
 754						/* Kill the band! */
 755						q->active[prio] = NULL;
 756						q->activemask &= ~(1<<prio);
 757						if (cl->q->q.qlen)
 758							cbq_activate_class(cl);
 759						return NULL;
 760					}
 761
 762					q->active[prio] = cl_tail;
 763				}
 764				if (cl->q->q.qlen)
 765					cbq_activate_class(cl);
 766
 767				cl = cl_prev;
 768			}
 769
 770next_class:
 771			cl_prev = cl;
 772			cl = cl->next_alive;
 773		} while (cl_prev != cl_tail);
 774	} while (deficit);
 775
 776	q->active[prio] = cl_prev;
 777
 778	return NULL;
 779}
 780
 781static inline struct sk_buff *
 782cbq_dequeue_1(struct Qdisc *sch)
 783{
 784	struct cbq_sched_data *q = qdisc_priv(sch);
 785	struct sk_buff *skb;
 786	unsigned int activemask;
 787
 788	activemask = q->activemask & 0xFF;
 789	while (activemask) {
 790		int prio = ffz(~activemask);
 791		activemask &= ~(1<<prio);
 792		skb = cbq_dequeue_prio(sch, prio);
 793		if (skb)
 794			return skb;
 795	}
 796	return NULL;
 797}
 798
 799static struct sk_buff *
 800cbq_dequeue(struct Qdisc *sch)
 801{
 802	struct sk_buff *skb;
 803	struct cbq_sched_data *q = qdisc_priv(sch);
 804	psched_time_t now;
 
 805
 806	now = psched_get_time();
 
 807
 808	if (q->tx_class)
 
 
 
 
 
 
 
 
 
 
 809		cbq_update(q);
 810
 811	q->now = now;
 
 
 
 812
 813	for (;;) {
 814		q->wd_expires = 0;
 815
 816		skb = cbq_dequeue_1(sch);
 817		if (skb) {
 818			qdisc_bstats_update(sch, skb);
 819			sch->q.qlen--;
 
 820			return skb;
 821		}
 822
 823		/* All the classes are overlimit.
 824		 *
 825		 * It is possible, if:
 826		 *
 827		 * 1. Scheduler is empty.
 828		 * 2. Toplevel cutoff inhibited borrowing.
 829		 * 3. Root class is overlimit.
 830		 *
 831		 * Reset 2d and 3d conditions and retry.
 832		 *
 833		 * Note, that NS and cbq-2.0 are buggy, peeking
 834		 * an arbitrary class is appropriate for ancestor-only
 835		 * sharing, but not for toplevel algorithm.
 836		 *
 837		 * Our version is better, but slower, because it requires
 838		 * two passes, but it is unavoidable with top-level sharing.
 839		 */
 840
 841		if (q->toplevel == TC_CBQ_MAXLEVEL &&
 842		    q->link.undertime == PSCHED_PASTPERFECT)
 843			break;
 844
 845		q->toplevel = TC_CBQ_MAXLEVEL;
 846		q->link.undertime = PSCHED_PASTPERFECT;
 847	}
 848
 849	/* No packets in scheduler or nobody wants to give them to us :-(
 850	 * Sigh... start watchdog timer in the last case.
 851	 */
 852
 853	if (sch->q.qlen) {
 854		qdisc_qstats_overlimit(sch);
 855		if (q->wd_expires)
 856			qdisc_watchdog_schedule(&q->watchdog,
 857						now + q->wd_expires);
 858	}
 859	return NULL;
 860}
 861
 862/* CBQ class maintanance routines */
 863
 864static void cbq_adjust_levels(struct cbq_class *this)
 865{
 866	if (this == NULL)
 867		return;
 868
 869	do {
 870		int level = 0;
 871		struct cbq_class *cl;
 872
 873		cl = this->children;
 874		if (cl) {
 875			do {
 876				if (cl->level > level)
 877					level = cl->level;
 878			} while ((cl = cl->sibling) != this->children);
 879		}
 880		this->level = level + 1;
 881	} while ((this = this->tparent) != NULL);
 882}
 883
 884static void cbq_normalize_quanta(struct cbq_sched_data *q, int prio)
 885{
 886	struct cbq_class *cl;
 
 887	unsigned int h;
 888
 889	if (q->quanta[prio] == 0)
 890		return;
 891
 892	for (h = 0; h < q->clhash.hashsize; h++) {
 893		hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
 894			/* BUGGGG... Beware! This expression suffer of
 895			 * arithmetic overflows!
 896			 */
 897			if (cl->priority == prio) {
 898				cl->quantum = (cl->weight*cl->allot*q->nclasses[prio])/
 899					q->quanta[prio];
 900			}
 901			if (cl->quantum <= 0 ||
 902			    cl->quantum > 32*qdisc_dev(cl->qdisc)->mtu) {
 903				pr_warn("CBQ: class %08x has bad quantum==%ld, repaired.\n",
 904					cl->common.classid, cl->quantum);
 905				cl->quantum = qdisc_dev(cl->qdisc)->mtu/2 + 1;
 906			}
 907		}
 908	}
 909}
 910
 911static void cbq_sync_defmap(struct cbq_class *cl)
 912{
 913	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 914	struct cbq_class *split = cl->split;
 915	unsigned int h;
 916	int i;
 917
 918	if (split == NULL)
 919		return;
 920
 921	for (i = 0; i <= TC_PRIO_MAX; i++) {
 922		if (split->defaults[i] == cl && !(cl->defmap & (1<<i)))
 923			split->defaults[i] = NULL;
 924	}
 925
 926	for (i = 0; i <= TC_PRIO_MAX; i++) {
 927		int level = split->level;
 928
 929		if (split->defaults[i])
 930			continue;
 931
 932		for (h = 0; h < q->clhash.hashsize; h++) {
 
 933			struct cbq_class *c;
 934
 935			hlist_for_each_entry(c, &q->clhash.hash[h],
 936					     common.hnode) {
 937				if (c->split == split && c->level < level &&
 938				    c->defmap & (1<<i)) {
 939					split->defaults[i] = c;
 940					level = c->level;
 941				}
 942			}
 943		}
 944	}
 945}
 946
 947static void cbq_change_defmap(struct cbq_class *cl, u32 splitid, u32 def, u32 mask)
 948{
 949	struct cbq_class *split = NULL;
 950
 951	if (splitid == 0) {
 952		split = cl->split;
 953		if (!split)
 954			return;
 955		splitid = split->common.classid;
 956	}
 957
 958	if (split == NULL || split->common.classid != splitid) {
 959		for (split = cl->tparent; split; split = split->tparent)
 960			if (split->common.classid == splitid)
 961				break;
 962	}
 963
 964	if (split == NULL)
 965		return;
 966
 967	if (cl->split != split) {
 968		cl->defmap = 0;
 969		cbq_sync_defmap(cl);
 970		cl->split = split;
 971		cl->defmap = def & mask;
 972	} else
 973		cl->defmap = (cl->defmap & ~mask) | (def & mask);
 974
 975	cbq_sync_defmap(cl);
 976}
 977
 978static void cbq_unlink_class(struct cbq_class *this)
 979{
 980	struct cbq_class *cl, **clp;
 981	struct cbq_sched_data *q = qdisc_priv(this->qdisc);
 982
 983	qdisc_class_hash_remove(&q->clhash, &this->common);
 984
 985	if (this->tparent) {
 986		clp = &this->sibling;
 987		cl = *clp;
 988		do {
 989			if (cl == this) {
 990				*clp = cl->sibling;
 991				break;
 992			}
 993			clp = &cl->sibling;
 994		} while ((cl = *clp) != this->sibling);
 995
 996		if (this->tparent->children == this) {
 997			this->tparent->children = this->sibling;
 998			if (this->sibling == this)
 999				this->tparent->children = NULL;
1000		}
1001	} else {
1002		WARN_ON(this->sibling != this);
1003	}
1004}
1005
1006static void cbq_link_class(struct cbq_class *this)
1007{
1008	struct cbq_sched_data *q = qdisc_priv(this->qdisc);
1009	struct cbq_class *parent = this->tparent;
1010
1011	this->sibling = this;
1012	qdisc_class_hash_insert(&q->clhash, &this->common);
1013
1014	if (parent == NULL)
1015		return;
1016
1017	if (parent->children == NULL) {
1018		parent->children = this;
1019	} else {
1020		this->sibling = parent->children->sibling;
1021		parent->children->sibling = this;
1022	}
1023}
1024
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1025static void
1026cbq_reset(struct Qdisc *sch)
1027{
1028	struct cbq_sched_data *q = qdisc_priv(sch);
1029	struct cbq_class *cl;
 
1030	int prio;
1031	unsigned int h;
1032
1033	q->activemask = 0;
1034	q->pmask = 0;
1035	q->tx_class = NULL;
1036	q->tx_borrowed = NULL;
1037	qdisc_watchdog_cancel(&q->watchdog);
1038	hrtimer_cancel(&q->delay_timer);
1039	q->toplevel = TC_CBQ_MAXLEVEL;
1040	q->now = psched_get_time();
 
1041
1042	for (prio = 0; prio <= TC_CBQ_MAXPRIO; prio++)
1043		q->active[prio] = NULL;
1044
1045	for (h = 0; h < q->clhash.hashsize; h++) {
1046		hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1047			qdisc_reset(cl->q);
1048
1049			cl->next_alive = NULL;
1050			cl->undertime = PSCHED_PASTPERFECT;
1051			cl->avgidle = cl->maxidle;
1052			cl->deficit = cl->quantum;
1053			cl->cpriority = cl->priority;
1054		}
1055	}
1056	sch->q.qlen = 0;
1057}
1058
1059
1060static int cbq_set_lss(struct cbq_class *cl, struct tc_cbq_lssopt *lss)
1061{
1062	if (lss->change & TCF_CBQ_LSS_FLAGS) {
1063		cl->share = (lss->flags & TCF_CBQ_LSS_ISOLATED) ? NULL : cl->tparent;
1064		cl->borrow = (lss->flags & TCF_CBQ_LSS_BOUNDED) ? NULL : cl->tparent;
1065	}
1066	if (lss->change & TCF_CBQ_LSS_EWMA)
1067		cl->ewma_log = lss->ewma_log;
1068	if (lss->change & TCF_CBQ_LSS_AVPKT)
1069		cl->avpkt = lss->avpkt;
1070	if (lss->change & TCF_CBQ_LSS_MINIDLE)
1071		cl->minidle = -(long)lss->minidle;
1072	if (lss->change & TCF_CBQ_LSS_MAXIDLE) {
1073		cl->maxidle = lss->maxidle;
1074		cl->avgidle = lss->maxidle;
1075	}
1076	if (lss->change & TCF_CBQ_LSS_OFFTIME)
1077		cl->offtime = lss->offtime;
1078	return 0;
1079}
1080
1081static void cbq_rmprio(struct cbq_sched_data *q, struct cbq_class *cl)
1082{
1083	q->nclasses[cl->priority]--;
1084	q->quanta[cl->priority] -= cl->weight;
1085	cbq_normalize_quanta(q, cl->priority);
1086}
1087
1088static void cbq_addprio(struct cbq_sched_data *q, struct cbq_class *cl)
1089{
1090	q->nclasses[cl->priority]++;
1091	q->quanta[cl->priority] += cl->weight;
1092	cbq_normalize_quanta(q, cl->priority);
1093}
1094
1095static int cbq_set_wrr(struct cbq_class *cl, struct tc_cbq_wrropt *wrr)
1096{
1097	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
1098
1099	if (wrr->allot)
1100		cl->allot = wrr->allot;
1101	if (wrr->weight)
1102		cl->weight = wrr->weight;
1103	if (wrr->priority) {
1104		cl->priority = wrr->priority - 1;
1105		cl->cpriority = cl->priority;
1106		if (cl->priority >= cl->priority2)
1107			cl->priority2 = TC_CBQ_MAXPRIO - 1;
1108	}
1109
1110	cbq_addprio(q, cl);
1111	return 0;
1112}
1113
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1114static int cbq_set_fopt(struct cbq_class *cl, struct tc_cbq_fopt *fopt)
1115{
1116	cbq_change_defmap(cl, fopt->split, fopt->defmap, fopt->defchange);
1117	return 0;
1118}
1119
1120static const struct nla_policy cbq_policy[TCA_CBQ_MAX + 1] = {
1121	[TCA_CBQ_LSSOPT]	= { .len = sizeof(struct tc_cbq_lssopt) },
1122	[TCA_CBQ_WRROPT]	= { .len = sizeof(struct tc_cbq_wrropt) },
1123	[TCA_CBQ_FOPT]		= { .len = sizeof(struct tc_cbq_fopt) },
1124	[TCA_CBQ_OVL_STRATEGY]	= { .len = sizeof(struct tc_cbq_ovl) },
1125	[TCA_CBQ_RATE]		= { .len = sizeof(struct tc_ratespec) },
1126	[TCA_CBQ_RTAB]		= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1127	[TCA_CBQ_POLICE]	= { .len = sizeof(struct tc_cbq_police) },
1128};
1129
1130static int cbq_opt_parse(struct nlattr *tb[TCA_CBQ_MAX + 1],
1131			 struct nlattr *opt,
1132			 struct netlink_ext_ack *extack)
1133{
1134	int err;
1135
1136	if (!opt) {
1137		NL_SET_ERR_MSG(extack, "CBQ options are required for this operation");
1138		return -EINVAL;
1139	}
1140
1141	err = nla_parse_nested_deprecated(tb, TCA_CBQ_MAX, opt,
1142					  cbq_policy, extack);
1143	if (err < 0)
1144		return err;
1145
1146	if (tb[TCA_CBQ_WRROPT]) {
1147		const struct tc_cbq_wrropt *wrr = nla_data(tb[TCA_CBQ_WRROPT]);
1148
1149		if (wrr->priority > TC_CBQ_MAXPRIO) {
1150			NL_SET_ERR_MSG(extack, "priority is bigger than TC_CBQ_MAXPRIO");
1151			err = -EINVAL;
1152		}
1153	}
1154	return err;
1155}
1156
1157static int cbq_init(struct Qdisc *sch, struct nlattr *opt,
1158		    struct netlink_ext_ack *extack)
1159{
1160	struct cbq_sched_data *q = qdisc_priv(sch);
1161	struct nlattr *tb[TCA_CBQ_MAX + 1];
1162	struct tc_ratespec *r;
1163	int err;
1164
1165	qdisc_watchdog_init(&q->watchdog, sch);
1166	hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1167	q->delay_timer.function = cbq_undelay;
1168
1169	err = cbq_opt_parse(tb, opt, extack);
1170	if (err < 0)
1171		return err;
1172
1173	if (!tb[TCA_CBQ_RTAB] || !tb[TCA_CBQ_RATE]) {
1174		NL_SET_ERR_MSG(extack, "Rate specification missing or incomplete");
1175		return -EINVAL;
1176	}
1177
1178	r = nla_data(tb[TCA_CBQ_RATE]);
1179
1180	q->link.R_tab = qdisc_get_rtab(r, tb[TCA_CBQ_RTAB], extack);
1181	if (!q->link.R_tab)
1182		return -EINVAL;
1183
1184	err = tcf_block_get(&q->link.block, &q->link.filter_list, sch, extack);
1185	if (err)
1186		goto put_rtab;
1187
1188	err = qdisc_class_hash_init(&q->clhash);
1189	if (err < 0)
1190		goto put_block;
1191
 
1192	q->link.sibling = &q->link;
1193	q->link.common.classid = sch->handle;
1194	q->link.qdisc = sch;
1195	q->link.q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1196				      sch->handle, NULL);
1197	if (!q->link.q)
1198		q->link.q = &noop_qdisc;
1199	else
1200		qdisc_hash_add(q->link.q, true);
1201
1202	q->link.priority = TC_CBQ_MAXPRIO - 1;
1203	q->link.priority2 = TC_CBQ_MAXPRIO - 1;
1204	q->link.cpriority = TC_CBQ_MAXPRIO - 1;
 
 
1205	q->link.allot = psched_mtu(qdisc_dev(sch));
1206	q->link.quantum = q->link.allot;
1207	q->link.weight = q->link.R_tab->rate.rate;
1208
1209	q->link.ewma_log = TC_CBQ_DEF_EWMA;
1210	q->link.avpkt = q->link.allot/2;
1211	q->link.minidle = -0x7FFFFFFF;
1212
 
 
 
1213	q->toplevel = TC_CBQ_MAXLEVEL;
1214	q->now = psched_get_time();
 
1215
1216	cbq_link_class(&q->link);
1217
1218	if (tb[TCA_CBQ_LSSOPT])
1219		cbq_set_lss(&q->link, nla_data(tb[TCA_CBQ_LSSOPT]));
1220
1221	cbq_addprio(q, &q->link);
1222	return 0;
1223
1224put_block:
1225	tcf_block_put(q->link.block);
1226
1227put_rtab:
1228	qdisc_put_rtab(q->link.R_tab);
1229	return err;
1230}
1231
1232static int cbq_dump_rate(struct sk_buff *skb, struct cbq_class *cl)
1233{
1234	unsigned char *b = skb_tail_pointer(skb);
1235
1236	if (nla_put(skb, TCA_CBQ_RATE, sizeof(cl->R_tab->rate), &cl->R_tab->rate))
1237		goto nla_put_failure;
1238	return skb->len;
1239
1240nla_put_failure:
1241	nlmsg_trim(skb, b);
1242	return -1;
1243}
1244
1245static int cbq_dump_lss(struct sk_buff *skb, struct cbq_class *cl)
1246{
1247	unsigned char *b = skb_tail_pointer(skb);
1248	struct tc_cbq_lssopt opt;
1249
1250	opt.flags = 0;
1251	if (cl->borrow == NULL)
1252		opt.flags |= TCF_CBQ_LSS_BOUNDED;
1253	if (cl->share == NULL)
1254		opt.flags |= TCF_CBQ_LSS_ISOLATED;
1255	opt.ewma_log = cl->ewma_log;
1256	opt.level = cl->level;
1257	opt.avpkt = cl->avpkt;
1258	opt.maxidle = cl->maxidle;
1259	opt.minidle = (u32)(-cl->minidle);
1260	opt.offtime = cl->offtime;
1261	opt.change = ~0;
1262	if (nla_put(skb, TCA_CBQ_LSSOPT, sizeof(opt), &opt))
1263		goto nla_put_failure;
1264	return skb->len;
1265
1266nla_put_failure:
1267	nlmsg_trim(skb, b);
1268	return -1;
1269}
1270
1271static int cbq_dump_wrr(struct sk_buff *skb, struct cbq_class *cl)
1272{
1273	unsigned char *b = skb_tail_pointer(skb);
1274	struct tc_cbq_wrropt opt;
1275
1276	memset(&opt, 0, sizeof(opt));
1277	opt.flags = 0;
1278	opt.allot = cl->allot;
1279	opt.priority = cl->priority + 1;
1280	opt.cpriority = cl->cpriority + 1;
1281	opt.weight = cl->weight;
1282	if (nla_put(skb, TCA_CBQ_WRROPT, sizeof(opt), &opt))
1283		goto nla_put_failure;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1284	return skb->len;
1285
1286nla_put_failure:
1287	nlmsg_trim(skb, b);
1288	return -1;
1289}
1290
1291static int cbq_dump_fopt(struct sk_buff *skb, struct cbq_class *cl)
1292{
1293	unsigned char *b = skb_tail_pointer(skb);
1294	struct tc_cbq_fopt opt;
1295
1296	if (cl->split || cl->defmap) {
1297		opt.split = cl->split ? cl->split->common.classid : 0;
1298		opt.defmap = cl->defmap;
1299		opt.defchange = ~0;
1300		if (nla_put(skb, TCA_CBQ_FOPT, sizeof(opt), &opt))
1301			goto nla_put_failure;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1302	}
1303	return skb->len;
1304
1305nla_put_failure:
1306	nlmsg_trim(skb, b);
1307	return -1;
1308}
 
1309
1310static int cbq_dump_attr(struct sk_buff *skb, struct cbq_class *cl)
1311{
1312	if (cbq_dump_lss(skb, cl) < 0 ||
1313	    cbq_dump_rate(skb, cl) < 0 ||
1314	    cbq_dump_wrr(skb, cl) < 0 ||
 
 
 
 
1315	    cbq_dump_fopt(skb, cl) < 0)
1316		return -1;
1317	return 0;
1318}
1319
1320static int cbq_dump(struct Qdisc *sch, struct sk_buff *skb)
1321{
1322	struct cbq_sched_data *q = qdisc_priv(sch);
1323	struct nlattr *nest;
1324
1325	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1326	if (nest == NULL)
1327		goto nla_put_failure;
1328	if (cbq_dump_attr(skb, &q->link) < 0)
1329		goto nla_put_failure;
1330	return nla_nest_end(skb, nest);
 
1331
1332nla_put_failure:
1333	nla_nest_cancel(skb, nest);
1334	return -1;
1335}
1336
1337static int
1338cbq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1339{
1340	struct cbq_sched_data *q = qdisc_priv(sch);
1341
1342	q->link.xstats.avgidle = q->link.avgidle;
1343	return gnet_stats_copy_app(d, &q->link.xstats, sizeof(q->link.xstats));
1344}
1345
1346static int
1347cbq_dump_class(struct Qdisc *sch, unsigned long arg,
1348	       struct sk_buff *skb, struct tcmsg *tcm)
1349{
1350	struct cbq_class *cl = (struct cbq_class *)arg;
1351	struct nlattr *nest;
1352
1353	if (cl->tparent)
1354		tcm->tcm_parent = cl->tparent->common.classid;
1355	else
1356		tcm->tcm_parent = TC_H_ROOT;
1357	tcm->tcm_handle = cl->common.classid;
1358	tcm->tcm_info = cl->q->handle;
1359
1360	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1361	if (nest == NULL)
1362		goto nla_put_failure;
1363	if (cbq_dump_attr(skb, cl) < 0)
1364		goto nla_put_failure;
1365	return nla_nest_end(skb, nest);
 
1366
1367nla_put_failure:
1368	nla_nest_cancel(skb, nest);
1369	return -1;
1370}
1371
1372static int
1373cbq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1374	struct gnet_dump *d)
1375{
1376	struct cbq_sched_data *q = qdisc_priv(sch);
1377	struct cbq_class *cl = (struct cbq_class *)arg;
1378	__u32 qlen;
1379
 
1380	cl->xstats.avgidle = cl->avgidle;
1381	cl->xstats.undertime = 0;
1382	qdisc_qstats_qlen_backlog(cl->q, &qlen, &cl->qstats.backlog);
1383
1384	if (cl->undertime != PSCHED_PASTPERFECT)
1385		cl->xstats.undertime = cl->undertime - q->now;
1386
1387	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1388				  d, NULL, &cl->bstats) < 0 ||
1389	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1390	    gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1391		return -1;
1392
1393	return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1394}
1395
1396static int cbq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1397		     struct Qdisc **old, struct netlink_ext_ack *extack)
1398{
1399	struct cbq_class *cl = (struct cbq_class *)arg;
1400
1401	if (new == NULL) {
1402		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1403					cl->common.classid, extack);
1404		if (new == NULL)
1405			return -ENOBUFS;
 
 
 
 
 
1406	}
 
 
 
 
 
 
1407
1408	*old = qdisc_replace(sch, new, &cl->q);
1409	return 0;
1410}
1411
1412static struct Qdisc *cbq_leaf(struct Qdisc *sch, unsigned long arg)
1413{
1414	struct cbq_class *cl = (struct cbq_class *)arg;
1415
1416	return cl->q;
1417}
1418
1419static void cbq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1420{
1421	struct cbq_class *cl = (struct cbq_class *)arg;
1422
1423	cbq_deactivate_class(cl);
 
1424}
1425
1426static unsigned long cbq_find(struct Qdisc *sch, u32 classid)
1427{
1428	struct cbq_sched_data *q = qdisc_priv(sch);
 
1429
1430	return (unsigned long)cbq_class_lookup(q, classid);
 
 
 
 
1431}
1432
1433static void cbq_destroy_class(struct Qdisc *sch, struct cbq_class *cl)
1434{
1435	struct cbq_sched_data *q = qdisc_priv(sch);
1436
1437	WARN_ON(cl->filters);
1438
1439	tcf_block_put(cl->block);
1440	qdisc_put(cl->q);
1441	qdisc_put_rtab(cl->R_tab);
1442	gen_kill_estimator(&cl->rate_est);
1443	if (cl != &q->link)
1444		kfree(cl);
1445}
1446
1447static void cbq_destroy(struct Qdisc *sch)
1448{
1449	struct cbq_sched_data *q = qdisc_priv(sch);
1450	struct hlist_node *next;
1451	struct cbq_class *cl;
1452	unsigned int h;
1453
1454#ifdef CONFIG_NET_CLS_ACT
1455	q->rx_class = NULL;
1456#endif
1457	/*
1458	 * Filters must be destroyed first because we don't destroy the
1459	 * classes from root to leafs which means that filters can still
1460	 * be bound to classes which have been destroyed already. --TGR '04
1461	 */
1462	for (h = 0; h < q->clhash.hashsize; h++) {
1463		hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1464			tcf_block_put(cl->block);
1465			cl->block = NULL;
1466		}
1467	}
1468	for (h = 0; h < q->clhash.hashsize; h++) {
1469		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[h],
1470					  common.hnode)
1471			cbq_destroy_class(sch, cl);
1472	}
1473	qdisc_class_hash_destroy(&q->clhash);
1474}
1475
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1476static int
1477cbq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, struct nlattr **tca,
1478		 unsigned long *arg, struct netlink_ext_ack *extack)
1479{
1480	int err;
1481	struct cbq_sched_data *q = qdisc_priv(sch);
1482	struct cbq_class *cl = (struct cbq_class *)*arg;
1483	struct nlattr *opt = tca[TCA_OPTIONS];
1484	struct nlattr *tb[TCA_CBQ_MAX + 1];
1485	struct cbq_class *parent;
1486	struct qdisc_rate_table *rtab = NULL;
1487
1488	err = cbq_opt_parse(tb, opt, extack);
 
 
 
1489	if (err < 0)
1490		return err;
1491
1492	if (tb[TCA_CBQ_OVL_STRATEGY] || tb[TCA_CBQ_POLICE]) {
1493		NL_SET_ERR_MSG(extack, "Neither overlimit strategy nor policing attributes can be used for changing class params");
1494		return -EOPNOTSUPP;
1495	}
1496
1497	if (cl) {
1498		/* Check parent */
1499		if (parentid) {
1500			if (cl->tparent &&
1501			    cl->tparent->common.classid != parentid) {
1502				NL_SET_ERR_MSG(extack, "Invalid parent id");
1503				return -EINVAL;
1504			}
1505			if (!cl->tparent && parentid != TC_H_ROOT) {
1506				NL_SET_ERR_MSG(extack, "Parent must be root");
1507				return -EINVAL;
1508			}
1509		}
1510
1511		if (tb[TCA_CBQ_RATE]) {
1512			rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]),
1513					      tb[TCA_CBQ_RTAB], extack);
1514			if (rtab == NULL)
1515				return -EINVAL;
1516		}
1517
1518		if (tca[TCA_RATE]) {
1519			err = gen_replace_estimator(&cl->bstats, NULL,
1520						    &cl->rate_est,
1521						    NULL,
1522						    qdisc_root_sleeping_running(sch),
1523						    tca[TCA_RATE]);
1524			if (err) {
1525				NL_SET_ERR_MSG(extack, "Failed to replace specified rate estimator");
1526				qdisc_put_rtab(rtab);
1527				return err;
1528			}
1529		}
1530
1531		/* Change class parameters */
1532		sch_tree_lock(sch);
1533
1534		if (cl->next_alive != NULL)
1535			cbq_deactivate_class(cl);
1536
1537		if (rtab) {
1538			qdisc_put_rtab(cl->R_tab);
1539			cl->R_tab = rtab;
1540		}
1541
1542		if (tb[TCA_CBQ_LSSOPT])
1543			cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1544
1545		if (tb[TCA_CBQ_WRROPT]) {
1546			cbq_rmprio(q, cl);
1547			cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1548		}
1549
 
 
 
 
 
 
 
 
1550		if (tb[TCA_CBQ_FOPT])
1551			cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1552
1553		if (cl->q->q.qlen)
1554			cbq_activate_class(cl);
1555
1556		sch_tree_unlock(sch);
1557
1558		return 0;
1559	}
1560
1561	if (parentid == TC_H_ROOT)
1562		return -EINVAL;
1563
1564	if (!tb[TCA_CBQ_WRROPT] || !tb[TCA_CBQ_RATE] || !tb[TCA_CBQ_LSSOPT]) {
1565		NL_SET_ERR_MSG(extack, "One of the following attributes MUST be specified: WRR, rate or link sharing");
1566		return -EINVAL;
1567	}
1568
1569	rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]), tb[TCA_CBQ_RTAB],
1570			      extack);
1571	if (rtab == NULL)
1572		return -EINVAL;
1573
1574	if (classid) {
1575		err = -EINVAL;
1576		if (TC_H_MAJ(classid ^ sch->handle) ||
1577		    cbq_class_lookup(q, classid)) {
1578			NL_SET_ERR_MSG(extack, "Specified class not found");
1579			goto failure;
1580		}
1581	} else {
1582		int i;
1583		classid = TC_H_MAKE(sch->handle, 0x8000);
1584
1585		for (i = 0; i < 0x8000; i++) {
1586			if (++q->hgenerator >= 0x8000)
1587				q->hgenerator = 1;
1588			if (cbq_class_lookup(q, classid|q->hgenerator) == NULL)
1589				break;
1590		}
1591		err = -ENOSR;
1592		if (i >= 0x8000) {
1593			NL_SET_ERR_MSG(extack, "Unable to generate classid");
1594			goto failure;
1595		}
1596		classid = classid|q->hgenerator;
1597	}
1598
1599	parent = &q->link;
1600	if (parentid) {
1601		parent = cbq_class_lookup(q, parentid);
1602		err = -EINVAL;
1603		if (!parent) {
1604			NL_SET_ERR_MSG(extack, "Failed to find parentid");
1605			goto failure;
1606		}
1607	}
1608
1609	err = -ENOBUFS;
1610	cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1611	if (cl == NULL)
1612		goto failure;
1613
1614	err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1615	if (err) {
1616		kfree(cl);
1617		return err;
1618	}
1619
1620	if (tca[TCA_RATE]) {
1621		err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1622					NULL,
1623					qdisc_root_sleeping_running(sch),
1624					tca[TCA_RATE]);
1625		if (err) {
1626			NL_SET_ERR_MSG(extack, "Couldn't create new estimator");
1627			tcf_block_put(cl->block);
1628			kfree(cl);
1629			goto failure;
1630		}
1631	}
1632
1633	cl->R_tab = rtab;
1634	rtab = NULL;
1635	cl->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, classid,
1636				  NULL);
1637	if (!cl->q)
1638		cl->q = &noop_qdisc;
1639	else
1640		qdisc_hash_add(cl->q, true);
1641
1642	cl->common.classid = classid;
1643	cl->tparent = parent;
1644	cl->qdisc = sch;
1645	cl->allot = parent->allot;
1646	cl->quantum = cl->allot;
1647	cl->weight = cl->R_tab->rate.rate;
1648
1649	sch_tree_lock(sch);
1650	cbq_link_class(cl);
1651	cl->borrow = cl->tparent;
1652	if (cl->tparent != &q->link)
1653		cl->share = cl->tparent;
1654	cbq_adjust_levels(parent);
1655	cl->minidle = -0x7FFFFFFF;
1656	cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1657	cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1658	if (cl->ewma_log == 0)
1659		cl->ewma_log = q->link.ewma_log;
1660	if (cl->maxidle == 0)
1661		cl->maxidle = q->link.maxidle;
1662	if (cl->avpkt == 0)
1663		cl->avpkt = q->link.avpkt;
 
 
 
 
 
 
 
1664	if (tb[TCA_CBQ_FOPT])
1665		cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1666	sch_tree_unlock(sch);
1667
1668	qdisc_class_hash_grow(sch, &q->clhash);
1669
1670	*arg = (unsigned long)cl;
1671	return 0;
1672
1673failure:
1674	qdisc_put_rtab(rtab);
1675	return err;
1676}
1677
1678static int cbq_delete(struct Qdisc *sch, unsigned long arg)
1679{
1680	struct cbq_sched_data *q = qdisc_priv(sch);
1681	struct cbq_class *cl = (struct cbq_class *)arg;
 
1682
1683	if (cl->filters || cl->children || cl == &q->link)
1684		return -EBUSY;
1685
1686	sch_tree_lock(sch);
1687
1688	qdisc_purge_queue(cl->q);
 
 
1689
1690	if (cl->next_alive)
1691		cbq_deactivate_class(cl);
1692
1693	if (q->tx_borrowed == cl)
1694		q->tx_borrowed = q->tx_class;
1695	if (q->tx_class == cl) {
1696		q->tx_class = NULL;
1697		q->tx_borrowed = NULL;
1698	}
1699#ifdef CONFIG_NET_CLS_ACT
1700	if (q->rx_class == cl)
1701		q->rx_class = NULL;
1702#endif
1703
1704	cbq_unlink_class(cl);
1705	cbq_adjust_levels(cl->tparent);
1706	cl->defmap = 0;
1707	cbq_sync_defmap(cl);
1708
1709	cbq_rmprio(q, cl);
1710	sch_tree_unlock(sch);
1711
1712	cbq_destroy_class(sch, cl);
 
 
 
 
 
1713	return 0;
1714}
1715
1716static struct tcf_block *cbq_tcf_block(struct Qdisc *sch, unsigned long arg,
1717				       struct netlink_ext_ack *extack)
1718{
1719	struct cbq_sched_data *q = qdisc_priv(sch);
1720	struct cbq_class *cl = (struct cbq_class *)arg;
1721
1722	if (cl == NULL)
1723		cl = &q->link;
1724
1725	return cl->block;
1726}
1727
1728static unsigned long cbq_bind_filter(struct Qdisc *sch, unsigned long parent,
1729				     u32 classid)
1730{
1731	struct cbq_sched_data *q = qdisc_priv(sch);
1732	struct cbq_class *p = (struct cbq_class *)parent;
1733	struct cbq_class *cl = cbq_class_lookup(q, classid);
1734
1735	if (cl) {
1736		if (p && p->level <= cl->level)
1737			return 0;
1738		cl->filters++;
1739		return (unsigned long)cl;
1740	}
1741	return 0;
1742}
1743
1744static void cbq_unbind_filter(struct Qdisc *sch, unsigned long arg)
1745{
1746	struct cbq_class *cl = (struct cbq_class *)arg;
1747
1748	cl->filters--;
1749}
1750
1751static void cbq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1752{
1753	struct cbq_sched_data *q = qdisc_priv(sch);
1754	struct cbq_class *cl;
 
1755	unsigned int h;
1756
1757	if (arg->stop)
1758		return;
1759
1760	for (h = 0; h < q->clhash.hashsize; h++) {
1761		hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1762			if (arg->count < arg->skip) {
1763				arg->count++;
1764				continue;
1765			}
1766			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1767				arg->stop = 1;
1768				return;
1769			}
1770			arg->count++;
1771		}
1772	}
1773}
1774
1775static const struct Qdisc_class_ops cbq_class_ops = {
1776	.graft		=	cbq_graft,
1777	.leaf		=	cbq_leaf,
1778	.qlen_notify	=	cbq_qlen_notify,
1779	.find		=	cbq_find,
 
1780	.change		=	cbq_change_class,
1781	.delete		=	cbq_delete,
1782	.walk		=	cbq_walk,
1783	.tcf_block	=	cbq_tcf_block,
1784	.bind_tcf	=	cbq_bind_filter,
1785	.unbind_tcf	=	cbq_unbind_filter,
1786	.dump		=	cbq_dump_class,
1787	.dump_stats	=	cbq_dump_class_stats,
1788};
1789
1790static struct Qdisc_ops cbq_qdisc_ops __read_mostly = {
1791	.next		=	NULL,
1792	.cl_ops		=	&cbq_class_ops,
1793	.id		=	"cbq",
1794	.priv_size	=	sizeof(struct cbq_sched_data),
1795	.enqueue	=	cbq_enqueue,
1796	.dequeue	=	cbq_dequeue,
1797	.peek		=	qdisc_peek_dequeued,
 
1798	.init		=	cbq_init,
1799	.reset		=	cbq_reset,
1800	.destroy	=	cbq_destroy,
1801	.change		=	NULL,
1802	.dump		=	cbq_dump,
1803	.dump_stats	=	cbq_dump_stats,
1804	.owner		=	THIS_MODULE,
1805};
1806
1807static int __init cbq_module_init(void)
1808{
1809	return register_qdisc(&cbq_qdisc_ops);
1810}
1811static void __exit cbq_module_exit(void)
1812{
1813	unregister_qdisc(&cbq_qdisc_ops);
1814}
1815module_init(cbq_module_init)
1816module_exit(cbq_module_exit)
1817MODULE_LICENSE("GPL");
v3.1
 
   1/*
   2 * net/sched/sch_cbq.c	Class-Based Queueing discipline.
   3 *
   4 *		This program is free software; you can redistribute it and/or
   5 *		modify it under the terms of the GNU General Public License
   6 *		as published by the Free Software Foundation; either version
   7 *		2 of the License, or (at your option) any later version.
   8 *
   9 * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10 *
  11 */
  12
  13#include <linux/module.h>
  14#include <linux/slab.h>
  15#include <linux/types.h>
  16#include <linux/kernel.h>
  17#include <linux/string.h>
  18#include <linux/errno.h>
  19#include <linux/skbuff.h>
  20#include <net/netlink.h>
  21#include <net/pkt_sched.h>
 
  22
  23
  24/*	Class-Based Queueing (CBQ) algorithm.
  25	=======================================
  26
  27	Sources: [1] Sally Floyd and Van Jacobson, "Link-sharing and Resource
  28		 Management Models for Packet Networks",
  29		 IEEE/ACM Transactions on Networking, Vol.3, No.4, 1995
  30
  31		 [2] Sally Floyd, "Notes on CBQ and Guaranteed Service", 1995
  32
  33		 [3] Sally Floyd, "Notes on Class-Based Queueing: Setting
  34		 Parameters", 1996
  35
  36		 [4] Sally Floyd and Michael Speer, "Experimental Results
  37		 for Class-Based Queueing", 1998, not published.
  38
  39	-----------------------------------------------------------------------
  40
  41	Algorithm skeleton was taken from NS simulator cbq.cc.
  42	If someone wants to check this code against the LBL version,
  43	he should take into account that ONLY the skeleton was borrowed,
  44	the implementation is different. Particularly:
  45
  46	--- The WRR algorithm is different. Our version looks more
  47	reasonable (I hope) and works when quanta are allowed to be
  48	less than MTU, which is always the case when real time classes
  49	have small rates. Note, that the statement of [3] is
  50	incomplete, delay may actually be estimated even if class
  51	per-round allotment is less than MTU. Namely, if per-round
  52	allotment is W*r_i, and r_1+...+r_k = r < 1
  53
  54	delay_i <= ([MTU/(W*r_i)]*W*r + W*r + k*MTU)/B
  55
  56	In the worst case we have IntServ estimate with D = W*r+k*MTU
  57	and C = MTU*r. The proof (if correct at all) is trivial.
  58
  59
  60	--- It seems that cbq-2.0 is not very accurate. At least, I cannot
  61	interpret some places, which look like wrong translations
  62	from NS. Anyone is advised to find these differences
  63	and explain to me, why I am wrong 8).
  64
  65	--- Linux has no EOI event, so that we cannot estimate true class
  66	idle time. Workaround is to consider the next dequeue event
  67	as sign that previous packet is finished. This is wrong because of
  68	internal device queueing, but on a permanently loaded link it is true.
  69	Moreover, combined with clock integrator, this scheme looks
  70	very close to an ideal solution.  */
  71
  72struct cbq_sched_data;
  73
  74
  75struct cbq_class {
  76	struct Qdisc_class_common common;
  77	struct cbq_class	*next_alive;	/* next class with backlog in this priority band */
  78
  79/* Parameters */
  80	unsigned char		priority;	/* class priority */
  81	unsigned char		priority2;	/* priority to be used after overlimit */
  82	unsigned char		ewma_log;	/* time constant for idle time calculation */
  83	unsigned char		ovl_strategy;
  84#ifdef CONFIG_NET_CLS_ACT
  85	unsigned char		police;
  86#endif
  87
  88	u32			defmap;
  89
  90	/* Link-sharing scheduler parameters */
  91	long			maxidle;	/* Class parameters: see below. */
  92	long			offtime;
  93	long			minidle;
  94	u32			avpkt;
  95	struct qdisc_rate_table	*R_tab;
  96
  97	/* Overlimit strategy parameters */
  98	void			(*overlimit)(struct cbq_class *cl);
  99	psched_tdiff_t		penalty;
 100
 101	/* General scheduler (WRR) parameters */
 102	long			allot;
 103	long			quantum;	/* Allotment per WRR round */
 104	long			weight;		/* Relative allotment: see below */
 105
 106	struct Qdisc		*qdisc;		/* Ptr to CBQ discipline */
 107	struct cbq_class	*split;		/* Ptr to split node */
 108	struct cbq_class	*share;		/* Ptr to LS parent in the class tree */
 109	struct cbq_class	*tparent;	/* Ptr to tree parent in the class tree */
 110	struct cbq_class	*borrow;	/* NULL if class is bandwidth limited;
 111						   parent otherwise */
 112	struct cbq_class	*sibling;	/* Sibling chain */
 113	struct cbq_class	*children;	/* Pointer to children chain */
 114
 115	struct Qdisc		*q;		/* Elementary queueing discipline */
 116
 117
 118/* Variables */
 119	unsigned char		cpriority;	/* Effective priority */
 120	unsigned char		delayed;
 121	unsigned char		level;		/* level of the class in hierarchy:
 122						   0 for leaf classes, and maximal
 123						   level of children + 1 for nodes.
 124						 */
 125
 126	psched_time_t		last;		/* Last end of service */
 127	psched_time_t		undertime;
 128	long			avgidle;
 129	long			deficit;	/* Saved deficit for WRR */
 130	psched_time_t		penalized;
 131	struct gnet_stats_basic_packed bstats;
 132	struct gnet_stats_queue qstats;
 133	struct gnet_stats_rate_est rate_est;
 134	struct tc_cbq_xstats	xstats;
 135
 136	struct tcf_proto	*filter_list;
 
 137
 138	int			refcnt;
 139	int			filters;
 140
 141	struct cbq_class	*defaults[TC_PRIO_MAX + 1];
 142};
 143
 144struct cbq_sched_data {
 145	struct Qdisc_class_hash	clhash;			/* Hash table of all classes */
 146	int			nclasses[TC_CBQ_MAXPRIO + 1];
 147	unsigned int		quanta[TC_CBQ_MAXPRIO + 1];
 148
 149	struct cbq_class	link;
 150
 151	unsigned int		activemask;
 152	struct cbq_class	*active[TC_CBQ_MAXPRIO + 1];	/* List of all classes
 153								   with backlog */
 154
 155#ifdef CONFIG_NET_CLS_ACT
 156	struct cbq_class	*rx_class;
 157#endif
 158	struct cbq_class	*tx_class;
 159	struct cbq_class	*tx_borrowed;
 160	int			tx_len;
 161	psched_time_t		now;		/* Cached timestamp */
 162	psched_time_t		now_rt;		/* Cached real time */
 163	unsigned int		pmask;
 164
 165	struct hrtimer		delay_timer;
 166	struct qdisc_watchdog	watchdog;	/* Watchdog timer,
 167						   started when CBQ has
 168						   backlog, but cannot
 169						   transmit just now */
 170	psched_tdiff_t		wd_expires;
 171	int			toplevel;
 172	u32			hgenerator;
 173};
 174
 175
 176#define L2T(cl, len)	qdisc_l2t((cl)->R_tab, len)
 177
 178static inline struct cbq_class *
 179cbq_class_lookup(struct cbq_sched_data *q, u32 classid)
 180{
 181	struct Qdisc_class_common *clc;
 182
 183	clc = qdisc_class_find(&q->clhash, classid);
 184	if (clc == NULL)
 185		return NULL;
 186	return container_of(clc, struct cbq_class, common);
 187}
 188
 189#ifdef CONFIG_NET_CLS_ACT
 190
 191static struct cbq_class *
 192cbq_reclassify(struct sk_buff *skb, struct cbq_class *this)
 193{
 194	struct cbq_class *cl;
 195
 196	for (cl = this->tparent; cl; cl = cl->tparent) {
 197		struct cbq_class *new = cl->defaults[TC_PRIO_BESTEFFORT];
 198
 199		if (new != NULL && new != this)
 200			return new;
 201	}
 202	return NULL;
 203}
 204
 205#endif
 206
 207/* Classify packet. The procedure is pretty complicated, but
 208 * it allows us to combine link sharing and priority scheduling
 209 * transparently.
 210 *
 211 * Namely, you can put link sharing rules (f.e. route based) at root of CBQ,
 212 * so that it resolves to split nodes. Then packets are classified
 213 * by logical priority, or a more specific classifier may be attached
 214 * to the split node.
 215 */
 216
 217static struct cbq_class *
 218cbq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
 219{
 220	struct cbq_sched_data *q = qdisc_priv(sch);
 221	struct cbq_class *head = &q->link;
 222	struct cbq_class **defmap;
 223	struct cbq_class *cl = NULL;
 224	u32 prio = skb->priority;
 
 225	struct tcf_result res;
 226
 227	/*
 228	 *  Step 1. If skb->priority points to one of our classes, use it.
 229	 */
 230	if (TC_H_MAJ(prio ^ sch->handle) == 0 &&
 231	    (cl = cbq_class_lookup(q, prio)) != NULL)
 232		return cl;
 233
 234	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 235	for (;;) {
 236		int result = 0;
 237		defmap = head->defaults;
 238
 
 239		/*
 240		 * Step 2+n. Apply classifier.
 241		 */
 242		if (!head->filter_list ||
 243		    (result = tc_classify_compat(skb, head->filter_list, &res)) < 0)
 244			goto fallback;
 245
 246		cl = (void *)res.class;
 247		if (!cl) {
 248			if (TC_H_MAJ(res.classid))
 249				cl = cbq_class_lookup(q, res.classid);
 250			else if ((cl = defmap[res.classid & TC_PRIO_MAX]) == NULL)
 251				cl = defmap[TC_PRIO_BESTEFFORT];
 252
 253			if (cl == NULL || cl->level >= head->level)
 254				goto fallback;
 255		}
 256
 
 257#ifdef CONFIG_NET_CLS_ACT
 258		switch (result) {
 259		case TC_ACT_QUEUED:
 260		case TC_ACT_STOLEN:
 
 261			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 
 262		case TC_ACT_SHOT:
 263			return NULL;
 264		case TC_ACT_RECLASSIFY:
 265			return cbq_reclassify(skb, cl);
 266		}
 267#endif
 268		if (cl->level == 0)
 269			return cl;
 270
 271		/*
 272		 * Step 3+n. If classifier selected a link sharing class,
 273		 *	   apply agency specific classifier.
 274		 *	   Repeat this procdure until we hit a leaf node.
 275		 */
 276		head = cl;
 277	}
 278
 279fallback:
 280	cl = head;
 281
 282	/*
 283	 * Step 4. No success...
 284	 */
 285	if (TC_H_MAJ(prio) == 0 &&
 286	    !(cl = head->defaults[prio & TC_PRIO_MAX]) &&
 287	    !(cl = head->defaults[TC_PRIO_BESTEFFORT]))
 288		return head;
 289
 290	return cl;
 291}
 292
 293/*
 294 * A packet has just been enqueued on the empty class.
 295 * cbq_activate_class adds it to the tail of active class list
 296 * of its priority band.
 297 */
 298
 299static inline void cbq_activate_class(struct cbq_class *cl)
 300{
 301	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 302	int prio = cl->cpriority;
 303	struct cbq_class *cl_tail;
 304
 305	cl_tail = q->active[prio];
 306	q->active[prio] = cl;
 307
 308	if (cl_tail != NULL) {
 309		cl->next_alive = cl_tail->next_alive;
 310		cl_tail->next_alive = cl;
 311	} else {
 312		cl->next_alive = cl;
 313		q->activemask |= (1<<prio);
 314	}
 315}
 316
 317/*
 318 * Unlink class from active chain.
 319 * Note that this same procedure is done directly in cbq_dequeue*
 320 * during round-robin procedure.
 321 */
 322
 323static void cbq_deactivate_class(struct cbq_class *this)
 324{
 325	struct cbq_sched_data *q = qdisc_priv(this->qdisc);
 326	int prio = this->cpriority;
 327	struct cbq_class *cl;
 328	struct cbq_class *cl_prev = q->active[prio];
 329
 330	do {
 331		cl = cl_prev->next_alive;
 332		if (cl == this) {
 333			cl_prev->next_alive = cl->next_alive;
 334			cl->next_alive = NULL;
 335
 336			if (cl == q->active[prio]) {
 337				q->active[prio] = cl_prev;
 338				if (cl == q->active[prio]) {
 339					q->active[prio] = NULL;
 340					q->activemask &= ~(1<<prio);
 341					return;
 342				}
 343			}
 344			return;
 345		}
 346	} while ((cl_prev = cl) != q->active[prio]);
 347}
 348
 349static void
 350cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl)
 351{
 352	int toplevel = q->toplevel;
 353
 354	if (toplevel > cl->level && !(qdisc_is_throttled(cl->q))) {
 355		psched_time_t now;
 356		psched_tdiff_t incr;
 357
 358		now = psched_get_time();
 359		incr = now - q->now_rt;
 360		now = q->now + incr;
 361
 362		do {
 363			if (cl->undertime < now) {
 364				q->toplevel = cl->level;
 365				return;
 366			}
 367		} while ((cl = cl->borrow) != NULL && toplevel > cl->level);
 368	}
 369}
 370
 371static int
 372cbq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 
 373{
 374	struct cbq_sched_data *q = qdisc_priv(sch);
 375	int uninitialized_var(ret);
 376	struct cbq_class *cl = cbq_classify(skb, sch, &ret);
 377
 378#ifdef CONFIG_NET_CLS_ACT
 379	q->rx_class = cl;
 380#endif
 381	if (cl == NULL) {
 382		if (ret & __NET_XMIT_BYPASS)
 383			sch->qstats.drops++;
 384		kfree_skb(skb);
 385		return ret;
 386	}
 387
 388#ifdef CONFIG_NET_CLS_ACT
 389	cl->q->__parent = sch;
 390#endif
 391	ret = qdisc_enqueue(skb, cl->q);
 392	if (ret == NET_XMIT_SUCCESS) {
 393		sch->q.qlen++;
 394		cbq_mark_toplevel(q, cl);
 395		if (!cl->next_alive)
 396			cbq_activate_class(cl);
 397		return ret;
 398	}
 399
 400	if (net_xmit_drop_count(ret)) {
 401		sch->qstats.drops++;
 402		cbq_mark_toplevel(q, cl);
 403		cl->qstats.drops++;
 404	}
 405	return ret;
 406}
 407
 408/* Overlimit actions */
 409
 410/* TC_CBQ_OVL_CLASSIC: (default) penalize leaf class by adding offtime */
 411
 412static void cbq_ovl_classic(struct cbq_class *cl)
 413{
 414	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 415	psched_tdiff_t delay = cl->undertime - q->now;
 416
 417	if (!cl->delayed) {
 418		delay += cl->offtime;
 419
 420		/*
 421		 * Class goes to sleep, so that it will have no
 422		 * chance to work avgidle. Let's forgive it 8)
 423		 *
 424		 * BTW cbq-2.0 has a crap in this
 425		 * place, apparently they forgot to shift it by cl->ewma_log.
 426		 */
 427		if (cl->avgidle < 0)
 428			delay -= (-cl->avgidle) - ((-cl->avgidle) >> cl->ewma_log);
 429		if (cl->avgidle < cl->minidle)
 430			cl->avgidle = cl->minidle;
 431		if (delay <= 0)
 432			delay = 1;
 433		cl->undertime = q->now + delay;
 434
 435		cl->xstats.overactions++;
 436		cl->delayed = 1;
 437	}
 438	if (q->wd_expires == 0 || q->wd_expires > delay)
 439		q->wd_expires = delay;
 440
 441	/* Dirty work! We must schedule wakeups based on
 442	 * real available rate, rather than leaf rate,
 443	 * which may be tiny (even zero).
 444	 */
 445	if (q->toplevel == TC_CBQ_MAXLEVEL) {
 446		struct cbq_class *b;
 447		psched_tdiff_t base_delay = q->wd_expires;
 448
 449		for (b = cl->borrow; b; b = b->borrow) {
 450			delay = b->undertime - q->now;
 451			if (delay < base_delay) {
 452				if (delay <= 0)
 453					delay = 1;
 454				base_delay = delay;
 455			}
 456		}
 457
 458		q->wd_expires = base_delay;
 459	}
 460}
 461
 462/* TC_CBQ_OVL_RCLASSIC: penalize by offtime classes in hierarchy, when
 463 * they go overlimit
 464 */
 465
 466static void cbq_ovl_rclassic(struct cbq_class *cl)
 467{
 468	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 469	struct cbq_class *this = cl;
 470
 471	do {
 472		if (cl->level > q->toplevel) {
 473			cl = NULL;
 474			break;
 475		}
 476	} while ((cl = cl->borrow) != NULL);
 477
 478	if (cl == NULL)
 479		cl = this;
 480	cbq_ovl_classic(cl);
 481}
 482
 483/* TC_CBQ_OVL_DELAY: delay until it will go to underlimit */
 484
 485static void cbq_ovl_delay(struct cbq_class *cl)
 486{
 487	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 488	psched_tdiff_t delay = cl->undertime - q->now;
 489
 490	if (test_bit(__QDISC_STATE_DEACTIVATED,
 491		     &qdisc_root_sleeping(cl->qdisc)->state))
 492		return;
 493
 494	if (!cl->delayed) {
 495		psched_time_t sched = q->now;
 496		ktime_t expires;
 497
 498		delay += cl->offtime;
 499		if (cl->avgidle < 0)
 500			delay -= (-cl->avgidle) - ((-cl->avgidle) >> cl->ewma_log);
 501		if (cl->avgidle < cl->minidle)
 502			cl->avgidle = cl->minidle;
 503		cl->undertime = q->now + delay;
 504
 505		if (delay > 0) {
 506			sched += delay + cl->penalty;
 507			cl->penalized = sched;
 508			cl->cpriority = TC_CBQ_MAXPRIO;
 509			q->pmask |= (1<<TC_CBQ_MAXPRIO);
 510
 511			expires = ktime_set(0, 0);
 512			expires = ktime_add_ns(expires, PSCHED_TICKS2NS(sched));
 513			if (hrtimer_try_to_cancel(&q->delay_timer) &&
 514			    ktime_to_ns(ktime_sub(
 515					hrtimer_get_expires(&q->delay_timer),
 516					expires)) > 0)
 517				hrtimer_set_expires(&q->delay_timer, expires);
 518			hrtimer_restart(&q->delay_timer);
 519			cl->delayed = 1;
 520			cl->xstats.overactions++;
 521			return;
 522		}
 523		delay = 1;
 524	}
 525	if (q->wd_expires == 0 || q->wd_expires > delay)
 526		q->wd_expires = delay;
 527}
 528
 529/* TC_CBQ_OVL_LOWPRIO: penalize class by lowering its priority band */
 530
 531static void cbq_ovl_lowprio(struct cbq_class *cl)
 532{
 533	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 534
 535	cl->penalized = q->now + cl->penalty;
 536
 537	if (cl->cpriority != cl->priority2) {
 538		cl->cpriority = cl->priority2;
 539		q->pmask |= (1<<cl->cpriority);
 540		cl->xstats.overactions++;
 541	}
 542	cbq_ovl_classic(cl);
 543}
 544
 545/* TC_CBQ_OVL_DROP: penalize class by dropping */
 546
 547static void cbq_ovl_drop(struct cbq_class *cl)
 548{
 549	if (cl->q->ops->drop)
 550		if (cl->q->ops->drop(cl->q))
 551			cl->qdisc->q.qlen--;
 552	cl->xstats.overactions++;
 553	cbq_ovl_classic(cl);
 554}
 555
 556static psched_tdiff_t cbq_undelay_prio(struct cbq_sched_data *q, int prio,
 557				       psched_time_t now)
 558{
 559	struct cbq_class *cl;
 560	struct cbq_class *cl_prev = q->active[prio];
 561	psched_time_t sched = now;
 562
 563	if (cl_prev == NULL)
 564		return 0;
 565
 566	do {
 567		cl = cl_prev->next_alive;
 568		if (now - cl->penalized > 0) {
 569			cl_prev->next_alive = cl->next_alive;
 570			cl->next_alive = NULL;
 571			cl->cpriority = cl->priority;
 572			cl->delayed = 0;
 573			cbq_activate_class(cl);
 574
 575			if (cl == q->active[prio]) {
 576				q->active[prio] = cl_prev;
 577				if (cl == q->active[prio]) {
 578					q->active[prio] = NULL;
 579					return 0;
 580				}
 581			}
 582
 583			cl = cl_prev->next_alive;
 584		} else if (sched - cl->penalized > 0)
 585			sched = cl->penalized;
 586	} while ((cl_prev = cl) != q->active[prio]);
 587
 588	return sched - now;
 589}
 590
 591static enum hrtimer_restart cbq_undelay(struct hrtimer *timer)
 592{
 593	struct cbq_sched_data *q = container_of(timer, struct cbq_sched_data,
 594						delay_timer);
 595	struct Qdisc *sch = q->watchdog.qdisc;
 596	psched_time_t now;
 597	psched_tdiff_t delay = 0;
 598	unsigned int pmask;
 599
 600	now = psched_get_time();
 601
 602	pmask = q->pmask;
 603	q->pmask = 0;
 604
 605	while (pmask) {
 606		int prio = ffz(~pmask);
 607		psched_tdiff_t tmp;
 608
 609		pmask &= ~(1<<prio);
 610
 611		tmp = cbq_undelay_prio(q, prio, now);
 612		if (tmp > 0) {
 613			q->pmask |= 1<<prio;
 614			if (tmp < delay || delay == 0)
 615				delay = tmp;
 616		}
 617	}
 618
 619	if (delay) {
 620		ktime_t time;
 621
 622		time = ktime_set(0, 0);
 623		time = ktime_add_ns(time, PSCHED_TICKS2NS(now + delay));
 624		hrtimer_start(&q->delay_timer, time, HRTIMER_MODE_ABS);
 625	}
 626
 627	qdisc_unthrottled(sch);
 628	__netif_schedule(qdisc_root(sch));
 629	return HRTIMER_NORESTART;
 630}
 631
 632#ifdef CONFIG_NET_CLS_ACT
 633static int cbq_reshape_fail(struct sk_buff *skb, struct Qdisc *child)
 634{
 635	struct Qdisc *sch = child->__parent;
 636	struct cbq_sched_data *q = qdisc_priv(sch);
 637	struct cbq_class *cl = q->rx_class;
 638
 639	q->rx_class = NULL;
 640
 641	if (cl && (cl = cbq_reclassify(skb, cl)) != NULL) {
 642		int ret;
 643
 644		cbq_mark_toplevel(q, cl);
 645
 646		q->rx_class = cl;
 647		cl->q->__parent = sch;
 648
 649		ret = qdisc_enqueue(skb, cl->q);
 650		if (ret == NET_XMIT_SUCCESS) {
 651			sch->q.qlen++;
 652			if (!cl->next_alive)
 653				cbq_activate_class(cl);
 654			return 0;
 655		}
 656		if (net_xmit_drop_count(ret))
 657			sch->qstats.drops++;
 658		return 0;
 659	}
 660
 661	sch->qstats.drops++;
 662	return -1;
 663}
 664#endif
 665
 666/*
 667 * It is mission critical procedure.
 668 *
 669 * We "regenerate" toplevel cutoff, if transmitting class
 670 * has backlog and it is not regulated. It is not part of
 671 * original CBQ description, but looks more reasonable.
 672 * Probably, it is wrong. This question needs further investigation.
 673 */
 674
 675static inline void
 676cbq_update_toplevel(struct cbq_sched_data *q, struct cbq_class *cl,
 677		    struct cbq_class *borrowed)
 678{
 679	if (cl && q->toplevel >= borrowed->level) {
 680		if (cl->q->q.qlen > 1) {
 681			do {
 682				if (borrowed->undertime == PSCHED_PASTPERFECT) {
 683					q->toplevel = borrowed->level;
 684					return;
 685				}
 686			} while ((borrowed = borrowed->borrow) != NULL);
 687		}
 688#if 0
 689	/* It is not necessary now. Uncommenting it
 690	   will save CPU cycles, but decrease fairness.
 691	 */
 692		q->toplevel = TC_CBQ_MAXLEVEL;
 693#endif
 694	}
 695}
 696
 697static void
 698cbq_update(struct cbq_sched_data *q)
 699{
 700	struct cbq_class *this = q->tx_class;
 701	struct cbq_class *cl = this;
 702	int len = q->tx_len;
 
 703
 704	q->tx_class = NULL;
 
 
 
 
 705
 706	for ( ; cl; cl = cl->share) {
 707		long avgidle = cl->avgidle;
 708		long idle;
 709
 710		cl->bstats.packets++;
 711		cl->bstats.bytes += len;
 712
 713		/*
 714		 * (now - last) is total time between packet right edges.
 715		 * (last_pktlen/rate) is "virtual" busy time, so that
 716		 *
 717		 *	idle = (now - last) - last_pktlen/rate
 718		 */
 719
 720		idle = q->now - cl->last;
 721		if ((unsigned long)idle > 128*1024*1024) {
 722			avgidle = cl->maxidle;
 723		} else {
 724			idle -= L2T(cl, len);
 725
 726		/* true_avgidle := (1-W)*true_avgidle + W*idle,
 727		 * where W=2^{-ewma_log}. But cl->avgidle is scaled:
 728		 * cl->avgidle == true_avgidle/W,
 729		 * hence:
 730		 */
 731			avgidle += idle - (avgidle>>cl->ewma_log);
 732		}
 733
 734		if (avgidle <= 0) {
 735			/* Overlimit or at-limit */
 736
 737			if (avgidle < cl->minidle)
 738				avgidle = cl->minidle;
 739
 740			cl->avgidle = avgidle;
 741
 742			/* Calculate expected time, when this class
 743			 * will be allowed to send.
 744			 * It will occur, when:
 745			 * (1-W)*true_avgidle + W*delay = 0, i.e.
 746			 * idle = (1/W - 1)*(-true_avgidle)
 747			 * or
 748			 * idle = (1 - W)*(-cl->avgidle);
 749			 */
 750			idle = (-avgidle) - ((-avgidle) >> cl->ewma_log);
 751
 752			/*
 753			 * That is not all.
 754			 * To maintain the rate allocated to the class,
 755			 * we add to undertime virtual clock,
 756			 * necessary to complete transmitted packet.
 757			 * (len/phys_bandwidth has been already passed
 758			 * to the moment of cbq_update)
 759			 */
 760
 761			idle -= L2T(&q->link, len);
 762			idle += L2T(cl, len);
 763
 764			cl->undertime = q->now + idle;
 765		} else {
 766			/* Underlimit */
 767
 768			cl->undertime = PSCHED_PASTPERFECT;
 769			if (avgidle > cl->maxidle)
 770				cl->avgidle = cl->maxidle;
 771			else
 772				cl->avgidle = avgidle;
 773		}
 774		cl->last = q->now;
 
 775	}
 776
 777	cbq_update_toplevel(q, this, q->tx_borrowed);
 778}
 779
 780static inline struct cbq_class *
 781cbq_under_limit(struct cbq_class *cl)
 782{
 783	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 784	struct cbq_class *this_cl = cl;
 785
 786	if (cl->tparent == NULL)
 787		return cl;
 788
 789	if (cl->undertime == PSCHED_PASTPERFECT || q->now >= cl->undertime) {
 790		cl->delayed = 0;
 791		return cl;
 792	}
 793
 794	do {
 795		/* It is very suspicious place. Now overlimit
 796		 * action is generated for not bounded classes
 797		 * only if link is completely congested.
 798		 * Though it is in agree with ancestor-only paradigm,
 799		 * it looks very stupid. Particularly,
 800		 * it means that this chunk of code will either
 801		 * never be called or result in strong amplification
 802		 * of burstiness. Dangerous, silly, and, however,
 803		 * no another solution exists.
 804		 */
 805		cl = cl->borrow;
 806		if (!cl) {
 807			this_cl->qstats.overlimits++;
 808			this_cl->overlimit(this_cl);
 809			return NULL;
 810		}
 811		if (cl->level > q->toplevel)
 812			return NULL;
 813	} while (cl->undertime != PSCHED_PASTPERFECT && q->now < cl->undertime);
 814
 815	cl->delayed = 0;
 816	return cl;
 817}
 818
 819static inline struct sk_buff *
 820cbq_dequeue_prio(struct Qdisc *sch, int prio)
 821{
 822	struct cbq_sched_data *q = qdisc_priv(sch);
 823	struct cbq_class *cl_tail, *cl_prev, *cl;
 824	struct sk_buff *skb;
 825	int deficit;
 826
 827	cl_tail = cl_prev = q->active[prio];
 828	cl = cl_prev->next_alive;
 829
 830	do {
 831		deficit = 0;
 832
 833		/* Start round */
 834		do {
 835			struct cbq_class *borrow = cl;
 836
 837			if (cl->q->q.qlen &&
 838			    (borrow = cbq_under_limit(cl)) == NULL)
 839				goto skip_class;
 840
 841			if (cl->deficit <= 0) {
 842				/* Class exhausted its allotment per
 843				 * this round. Switch to the next one.
 844				 */
 845				deficit = 1;
 846				cl->deficit += cl->quantum;
 847				goto next_class;
 848			}
 849
 850			skb = cl->q->dequeue(cl->q);
 851
 852			/* Class did not give us any skb :-(
 853			 * It could occur even if cl->q->q.qlen != 0
 854			 * f.e. if cl->q == "tbf"
 855			 */
 856			if (skb == NULL)
 857				goto skip_class;
 858
 859			cl->deficit -= qdisc_pkt_len(skb);
 860			q->tx_class = cl;
 861			q->tx_borrowed = borrow;
 862			if (borrow != cl) {
 863#ifndef CBQ_XSTATS_BORROWS_BYTES
 864				borrow->xstats.borrows++;
 865				cl->xstats.borrows++;
 866#else
 867				borrow->xstats.borrows += qdisc_pkt_len(skb);
 868				cl->xstats.borrows += qdisc_pkt_len(skb);
 869#endif
 870			}
 871			q->tx_len = qdisc_pkt_len(skb);
 872
 873			if (cl->deficit <= 0) {
 874				q->active[prio] = cl;
 875				cl = cl->next_alive;
 876				cl->deficit += cl->quantum;
 877			}
 878			return skb;
 879
 880skip_class:
 881			if (cl->q->q.qlen == 0 || prio != cl->cpriority) {
 882				/* Class is empty or penalized.
 883				 * Unlink it from active chain.
 884				 */
 885				cl_prev->next_alive = cl->next_alive;
 886				cl->next_alive = NULL;
 887
 888				/* Did cl_tail point to it? */
 889				if (cl == cl_tail) {
 890					/* Repair it! */
 891					cl_tail = cl_prev;
 892
 893					/* Was it the last class in this band? */
 894					if (cl == cl_tail) {
 895						/* Kill the band! */
 896						q->active[prio] = NULL;
 897						q->activemask &= ~(1<<prio);
 898						if (cl->q->q.qlen)
 899							cbq_activate_class(cl);
 900						return NULL;
 901					}
 902
 903					q->active[prio] = cl_tail;
 904				}
 905				if (cl->q->q.qlen)
 906					cbq_activate_class(cl);
 907
 908				cl = cl_prev;
 909			}
 910
 911next_class:
 912			cl_prev = cl;
 913			cl = cl->next_alive;
 914		} while (cl_prev != cl_tail);
 915	} while (deficit);
 916
 917	q->active[prio] = cl_prev;
 918
 919	return NULL;
 920}
 921
 922static inline struct sk_buff *
 923cbq_dequeue_1(struct Qdisc *sch)
 924{
 925	struct cbq_sched_data *q = qdisc_priv(sch);
 926	struct sk_buff *skb;
 927	unsigned int activemask;
 928
 929	activemask = q->activemask & 0xFF;
 930	while (activemask) {
 931		int prio = ffz(~activemask);
 932		activemask &= ~(1<<prio);
 933		skb = cbq_dequeue_prio(sch, prio);
 934		if (skb)
 935			return skb;
 936	}
 937	return NULL;
 938}
 939
 940static struct sk_buff *
 941cbq_dequeue(struct Qdisc *sch)
 942{
 943	struct sk_buff *skb;
 944	struct cbq_sched_data *q = qdisc_priv(sch);
 945	psched_time_t now;
 946	psched_tdiff_t incr;
 947
 948	now = psched_get_time();
 949	incr = now - q->now_rt;
 950
 951	if (q->tx_class) {
 952		psched_tdiff_t incr2;
 953		/* Time integrator. We calculate EOS time
 954		 * by adding expected packet transmission time.
 955		 * If real time is greater, we warp artificial clock,
 956		 * so that:
 957		 *
 958		 * cbq_time = max(real_time, work);
 959		 */
 960		incr2 = L2T(&q->link, q->tx_len);
 961		q->now += incr2;
 962		cbq_update(q);
 963		if ((incr -= incr2) < 0)
 964			incr = 0;
 965	}
 966	q->now += incr;
 967	q->now_rt = now;
 968
 969	for (;;) {
 970		q->wd_expires = 0;
 971
 972		skb = cbq_dequeue_1(sch);
 973		if (skb) {
 974			qdisc_bstats_update(sch, skb);
 975			sch->q.qlen--;
 976			qdisc_unthrottled(sch);
 977			return skb;
 978		}
 979
 980		/* All the classes are overlimit.
 981		 *
 982		 * It is possible, if:
 983		 *
 984		 * 1. Scheduler is empty.
 985		 * 2. Toplevel cutoff inhibited borrowing.
 986		 * 3. Root class is overlimit.
 987		 *
 988		 * Reset 2d and 3d conditions and retry.
 989		 *
 990		 * Note, that NS and cbq-2.0 are buggy, peeking
 991		 * an arbitrary class is appropriate for ancestor-only
 992		 * sharing, but not for toplevel algorithm.
 993		 *
 994		 * Our version is better, but slower, because it requires
 995		 * two passes, but it is unavoidable with top-level sharing.
 996		 */
 997
 998		if (q->toplevel == TC_CBQ_MAXLEVEL &&
 999		    q->link.undertime == PSCHED_PASTPERFECT)
1000			break;
1001
1002		q->toplevel = TC_CBQ_MAXLEVEL;
1003		q->link.undertime = PSCHED_PASTPERFECT;
1004	}
1005
1006	/* No packets in scheduler or nobody wants to give them to us :-(
1007	 * Sigh... start watchdog timer in the last case.
1008	 */
1009
1010	if (sch->q.qlen) {
1011		sch->qstats.overlimits++;
1012		if (q->wd_expires)
1013			qdisc_watchdog_schedule(&q->watchdog,
1014						now + q->wd_expires);
1015	}
1016	return NULL;
1017}
1018
1019/* CBQ class maintanance routines */
1020
1021static void cbq_adjust_levels(struct cbq_class *this)
1022{
1023	if (this == NULL)
1024		return;
1025
1026	do {
1027		int level = 0;
1028		struct cbq_class *cl;
1029
1030		cl = this->children;
1031		if (cl) {
1032			do {
1033				if (cl->level > level)
1034					level = cl->level;
1035			} while ((cl = cl->sibling) != this->children);
1036		}
1037		this->level = level + 1;
1038	} while ((this = this->tparent) != NULL);
1039}
1040
1041static void cbq_normalize_quanta(struct cbq_sched_data *q, int prio)
1042{
1043	struct cbq_class *cl;
1044	struct hlist_node *n;
1045	unsigned int h;
1046
1047	if (q->quanta[prio] == 0)
1048		return;
1049
1050	for (h = 0; h < q->clhash.hashsize; h++) {
1051		hlist_for_each_entry(cl, n, &q->clhash.hash[h], common.hnode) {
1052			/* BUGGGG... Beware! This expression suffer of
1053			 * arithmetic overflows!
1054			 */
1055			if (cl->priority == prio) {
1056				cl->quantum = (cl->weight*cl->allot*q->nclasses[prio])/
1057					q->quanta[prio];
1058			}
1059			if (cl->quantum <= 0 || cl->quantum>32*qdisc_dev(cl->qdisc)->mtu) {
1060				pr_warning("CBQ: class %08x has bad quantum==%ld, repaired.\n",
1061					   cl->common.classid, cl->quantum);
 
1062				cl->quantum = qdisc_dev(cl->qdisc)->mtu/2 + 1;
1063			}
1064		}
1065	}
1066}
1067
1068static void cbq_sync_defmap(struct cbq_class *cl)
1069{
1070	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
1071	struct cbq_class *split = cl->split;
1072	unsigned int h;
1073	int i;
1074
1075	if (split == NULL)
1076		return;
1077
1078	for (i = 0; i <= TC_PRIO_MAX; i++) {
1079		if (split->defaults[i] == cl && !(cl->defmap & (1<<i)))
1080			split->defaults[i] = NULL;
1081	}
1082
1083	for (i = 0; i <= TC_PRIO_MAX; i++) {
1084		int level = split->level;
1085
1086		if (split->defaults[i])
1087			continue;
1088
1089		for (h = 0; h < q->clhash.hashsize; h++) {
1090			struct hlist_node *n;
1091			struct cbq_class *c;
1092
1093			hlist_for_each_entry(c, n, &q->clhash.hash[h],
1094					     common.hnode) {
1095				if (c->split == split && c->level < level &&
1096				    c->defmap & (1<<i)) {
1097					split->defaults[i] = c;
1098					level = c->level;
1099				}
1100			}
1101		}
1102	}
1103}
1104
1105static void cbq_change_defmap(struct cbq_class *cl, u32 splitid, u32 def, u32 mask)
1106{
1107	struct cbq_class *split = NULL;
1108
1109	if (splitid == 0) {
1110		split = cl->split;
1111		if (!split)
1112			return;
1113		splitid = split->common.classid;
1114	}
1115
1116	if (split == NULL || split->common.classid != splitid) {
1117		for (split = cl->tparent; split; split = split->tparent)
1118			if (split->common.classid == splitid)
1119				break;
1120	}
1121
1122	if (split == NULL)
1123		return;
1124
1125	if (cl->split != split) {
1126		cl->defmap = 0;
1127		cbq_sync_defmap(cl);
1128		cl->split = split;
1129		cl->defmap = def & mask;
1130	} else
1131		cl->defmap = (cl->defmap & ~mask) | (def & mask);
1132
1133	cbq_sync_defmap(cl);
1134}
1135
1136static void cbq_unlink_class(struct cbq_class *this)
1137{
1138	struct cbq_class *cl, **clp;
1139	struct cbq_sched_data *q = qdisc_priv(this->qdisc);
1140
1141	qdisc_class_hash_remove(&q->clhash, &this->common);
1142
1143	if (this->tparent) {
1144		clp = &this->sibling;
1145		cl = *clp;
1146		do {
1147			if (cl == this) {
1148				*clp = cl->sibling;
1149				break;
1150			}
1151			clp = &cl->sibling;
1152		} while ((cl = *clp) != this->sibling);
1153
1154		if (this->tparent->children == this) {
1155			this->tparent->children = this->sibling;
1156			if (this->sibling == this)
1157				this->tparent->children = NULL;
1158		}
1159	} else {
1160		WARN_ON(this->sibling != this);
1161	}
1162}
1163
1164static void cbq_link_class(struct cbq_class *this)
1165{
1166	struct cbq_sched_data *q = qdisc_priv(this->qdisc);
1167	struct cbq_class *parent = this->tparent;
1168
1169	this->sibling = this;
1170	qdisc_class_hash_insert(&q->clhash, &this->common);
1171
1172	if (parent == NULL)
1173		return;
1174
1175	if (parent->children == NULL) {
1176		parent->children = this;
1177	} else {
1178		this->sibling = parent->children->sibling;
1179		parent->children->sibling = this;
1180	}
1181}
1182
1183static unsigned int cbq_drop(struct Qdisc *sch)
1184{
1185	struct cbq_sched_data *q = qdisc_priv(sch);
1186	struct cbq_class *cl, *cl_head;
1187	int prio;
1188	unsigned int len;
1189
1190	for (prio = TC_CBQ_MAXPRIO; prio >= 0; prio--) {
1191		cl_head = q->active[prio];
1192		if (!cl_head)
1193			continue;
1194
1195		cl = cl_head;
1196		do {
1197			if (cl->q->ops->drop && (len = cl->q->ops->drop(cl->q))) {
1198				sch->q.qlen--;
1199				if (!cl->q->q.qlen)
1200					cbq_deactivate_class(cl);
1201				return len;
1202			}
1203		} while ((cl = cl->next_alive) != cl_head);
1204	}
1205	return 0;
1206}
1207
1208static void
1209cbq_reset(struct Qdisc *sch)
1210{
1211	struct cbq_sched_data *q = qdisc_priv(sch);
1212	struct cbq_class *cl;
1213	struct hlist_node *n;
1214	int prio;
1215	unsigned int h;
1216
1217	q->activemask = 0;
1218	q->pmask = 0;
1219	q->tx_class = NULL;
1220	q->tx_borrowed = NULL;
1221	qdisc_watchdog_cancel(&q->watchdog);
1222	hrtimer_cancel(&q->delay_timer);
1223	q->toplevel = TC_CBQ_MAXLEVEL;
1224	q->now = psched_get_time();
1225	q->now_rt = q->now;
1226
1227	for (prio = 0; prio <= TC_CBQ_MAXPRIO; prio++)
1228		q->active[prio] = NULL;
1229
1230	for (h = 0; h < q->clhash.hashsize; h++) {
1231		hlist_for_each_entry(cl, n, &q->clhash.hash[h], common.hnode) {
1232			qdisc_reset(cl->q);
1233
1234			cl->next_alive = NULL;
1235			cl->undertime = PSCHED_PASTPERFECT;
1236			cl->avgidle = cl->maxidle;
1237			cl->deficit = cl->quantum;
1238			cl->cpriority = cl->priority;
1239		}
1240	}
1241	sch->q.qlen = 0;
1242}
1243
1244
1245static int cbq_set_lss(struct cbq_class *cl, struct tc_cbq_lssopt *lss)
1246{
1247	if (lss->change & TCF_CBQ_LSS_FLAGS) {
1248		cl->share = (lss->flags & TCF_CBQ_LSS_ISOLATED) ? NULL : cl->tparent;
1249		cl->borrow = (lss->flags & TCF_CBQ_LSS_BOUNDED) ? NULL : cl->tparent;
1250	}
1251	if (lss->change & TCF_CBQ_LSS_EWMA)
1252		cl->ewma_log = lss->ewma_log;
1253	if (lss->change & TCF_CBQ_LSS_AVPKT)
1254		cl->avpkt = lss->avpkt;
1255	if (lss->change & TCF_CBQ_LSS_MINIDLE)
1256		cl->minidle = -(long)lss->minidle;
1257	if (lss->change & TCF_CBQ_LSS_MAXIDLE) {
1258		cl->maxidle = lss->maxidle;
1259		cl->avgidle = lss->maxidle;
1260	}
1261	if (lss->change & TCF_CBQ_LSS_OFFTIME)
1262		cl->offtime = lss->offtime;
1263	return 0;
1264}
1265
1266static void cbq_rmprio(struct cbq_sched_data *q, struct cbq_class *cl)
1267{
1268	q->nclasses[cl->priority]--;
1269	q->quanta[cl->priority] -= cl->weight;
1270	cbq_normalize_quanta(q, cl->priority);
1271}
1272
1273static void cbq_addprio(struct cbq_sched_data *q, struct cbq_class *cl)
1274{
1275	q->nclasses[cl->priority]++;
1276	q->quanta[cl->priority] += cl->weight;
1277	cbq_normalize_quanta(q, cl->priority);
1278}
1279
1280static int cbq_set_wrr(struct cbq_class *cl, struct tc_cbq_wrropt *wrr)
1281{
1282	struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
1283
1284	if (wrr->allot)
1285		cl->allot = wrr->allot;
1286	if (wrr->weight)
1287		cl->weight = wrr->weight;
1288	if (wrr->priority) {
1289		cl->priority = wrr->priority - 1;
1290		cl->cpriority = cl->priority;
1291		if (cl->priority >= cl->priority2)
1292			cl->priority2 = TC_CBQ_MAXPRIO - 1;
1293	}
1294
1295	cbq_addprio(q, cl);
1296	return 0;
1297}
1298
1299static int cbq_set_overlimit(struct cbq_class *cl, struct tc_cbq_ovl *ovl)
1300{
1301	switch (ovl->strategy) {
1302	case TC_CBQ_OVL_CLASSIC:
1303		cl->overlimit = cbq_ovl_classic;
1304		break;
1305	case TC_CBQ_OVL_DELAY:
1306		cl->overlimit = cbq_ovl_delay;
1307		break;
1308	case TC_CBQ_OVL_LOWPRIO:
1309		if (ovl->priority2 - 1 >= TC_CBQ_MAXPRIO ||
1310		    ovl->priority2 - 1 <= cl->priority)
1311			return -EINVAL;
1312		cl->priority2 = ovl->priority2 - 1;
1313		cl->overlimit = cbq_ovl_lowprio;
1314		break;
1315	case TC_CBQ_OVL_DROP:
1316		cl->overlimit = cbq_ovl_drop;
1317		break;
1318	case TC_CBQ_OVL_RCLASSIC:
1319		cl->overlimit = cbq_ovl_rclassic;
1320		break;
1321	default:
1322		return -EINVAL;
1323	}
1324	cl->penalty = ovl->penalty;
1325	return 0;
1326}
1327
1328#ifdef CONFIG_NET_CLS_ACT
1329static int cbq_set_police(struct cbq_class *cl, struct tc_cbq_police *p)
1330{
1331	cl->police = p->police;
1332
1333	if (cl->q->handle) {
1334		if (p->police == TC_POLICE_RECLASSIFY)
1335			cl->q->reshape_fail = cbq_reshape_fail;
1336		else
1337			cl->q->reshape_fail = NULL;
1338	}
1339	return 0;
1340}
1341#endif
1342
1343static int cbq_set_fopt(struct cbq_class *cl, struct tc_cbq_fopt *fopt)
1344{
1345	cbq_change_defmap(cl, fopt->split, fopt->defmap, fopt->defchange);
1346	return 0;
1347}
1348
1349static const struct nla_policy cbq_policy[TCA_CBQ_MAX + 1] = {
1350	[TCA_CBQ_LSSOPT]	= { .len = sizeof(struct tc_cbq_lssopt) },
1351	[TCA_CBQ_WRROPT]	= { .len = sizeof(struct tc_cbq_wrropt) },
1352	[TCA_CBQ_FOPT]		= { .len = sizeof(struct tc_cbq_fopt) },
1353	[TCA_CBQ_OVL_STRATEGY]	= { .len = sizeof(struct tc_cbq_ovl) },
1354	[TCA_CBQ_RATE]		= { .len = sizeof(struct tc_ratespec) },
1355	[TCA_CBQ_RTAB]		= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1356	[TCA_CBQ_POLICE]	= { .len = sizeof(struct tc_cbq_police) },
1357};
1358
1359static int cbq_init(struct Qdisc *sch, struct nlattr *opt)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1360{
1361	struct cbq_sched_data *q = qdisc_priv(sch);
1362	struct nlattr *tb[TCA_CBQ_MAX + 1];
1363	struct tc_ratespec *r;
1364	int err;
1365
1366	err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy);
 
 
 
 
1367	if (err < 0)
1368		return err;
1369
1370	if (tb[TCA_CBQ_RTAB] == NULL || tb[TCA_CBQ_RATE] == NULL)
 
1371		return -EINVAL;
 
1372
1373	r = nla_data(tb[TCA_CBQ_RATE]);
1374
1375	if ((q->link.R_tab = qdisc_get_rtab(r, tb[TCA_CBQ_RTAB])) == NULL)
 
1376		return -EINVAL;
1377
 
 
 
 
1378	err = qdisc_class_hash_init(&q->clhash);
1379	if (err < 0)
1380		goto put_rtab;
1381
1382	q->link.refcnt = 1;
1383	q->link.sibling = &q->link;
1384	q->link.common.classid = sch->handle;
1385	q->link.qdisc = sch;
1386	q->link.q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1387				      sch->handle);
1388	if (!q->link.q)
1389		q->link.q = &noop_qdisc;
 
 
1390
1391	q->link.priority = TC_CBQ_MAXPRIO - 1;
1392	q->link.priority2 = TC_CBQ_MAXPRIO - 1;
1393	q->link.cpriority = TC_CBQ_MAXPRIO - 1;
1394	q->link.ovl_strategy = TC_CBQ_OVL_CLASSIC;
1395	q->link.overlimit = cbq_ovl_classic;
1396	q->link.allot = psched_mtu(qdisc_dev(sch));
1397	q->link.quantum = q->link.allot;
1398	q->link.weight = q->link.R_tab->rate.rate;
1399
1400	q->link.ewma_log = TC_CBQ_DEF_EWMA;
1401	q->link.avpkt = q->link.allot/2;
1402	q->link.minidle = -0x7FFFFFFF;
1403
1404	qdisc_watchdog_init(&q->watchdog, sch);
1405	hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1406	q->delay_timer.function = cbq_undelay;
1407	q->toplevel = TC_CBQ_MAXLEVEL;
1408	q->now = psched_get_time();
1409	q->now_rt = q->now;
1410
1411	cbq_link_class(&q->link);
1412
1413	if (tb[TCA_CBQ_LSSOPT])
1414		cbq_set_lss(&q->link, nla_data(tb[TCA_CBQ_LSSOPT]));
1415
1416	cbq_addprio(q, &q->link);
1417	return 0;
1418
 
 
 
1419put_rtab:
1420	qdisc_put_rtab(q->link.R_tab);
1421	return err;
1422}
1423
1424static int cbq_dump_rate(struct sk_buff *skb, struct cbq_class *cl)
1425{
1426	unsigned char *b = skb_tail_pointer(skb);
1427
1428	NLA_PUT(skb, TCA_CBQ_RATE, sizeof(cl->R_tab->rate), &cl->R_tab->rate);
 
1429	return skb->len;
1430
1431nla_put_failure:
1432	nlmsg_trim(skb, b);
1433	return -1;
1434}
1435
1436static int cbq_dump_lss(struct sk_buff *skb, struct cbq_class *cl)
1437{
1438	unsigned char *b = skb_tail_pointer(skb);
1439	struct tc_cbq_lssopt opt;
1440
1441	opt.flags = 0;
1442	if (cl->borrow == NULL)
1443		opt.flags |= TCF_CBQ_LSS_BOUNDED;
1444	if (cl->share == NULL)
1445		opt.flags |= TCF_CBQ_LSS_ISOLATED;
1446	opt.ewma_log = cl->ewma_log;
1447	opt.level = cl->level;
1448	opt.avpkt = cl->avpkt;
1449	opt.maxidle = cl->maxidle;
1450	opt.minidle = (u32)(-cl->minidle);
1451	opt.offtime = cl->offtime;
1452	opt.change = ~0;
1453	NLA_PUT(skb, TCA_CBQ_LSSOPT, sizeof(opt), &opt);
 
1454	return skb->len;
1455
1456nla_put_failure:
1457	nlmsg_trim(skb, b);
1458	return -1;
1459}
1460
1461static int cbq_dump_wrr(struct sk_buff *skb, struct cbq_class *cl)
1462{
1463	unsigned char *b = skb_tail_pointer(skb);
1464	struct tc_cbq_wrropt opt;
1465
 
1466	opt.flags = 0;
1467	opt.allot = cl->allot;
1468	opt.priority = cl->priority + 1;
1469	opt.cpriority = cl->cpriority + 1;
1470	opt.weight = cl->weight;
1471	NLA_PUT(skb, TCA_CBQ_WRROPT, sizeof(opt), &opt);
1472	return skb->len;
1473
1474nla_put_failure:
1475	nlmsg_trim(skb, b);
1476	return -1;
1477}
1478
1479static int cbq_dump_ovl(struct sk_buff *skb, struct cbq_class *cl)
1480{
1481	unsigned char *b = skb_tail_pointer(skb);
1482	struct tc_cbq_ovl opt;
1483
1484	opt.strategy = cl->ovl_strategy;
1485	opt.priority2 = cl->priority2 + 1;
1486	opt.pad = 0;
1487	opt.penalty = cl->penalty;
1488	NLA_PUT(skb, TCA_CBQ_OVL_STRATEGY, sizeof(opt), &opt);
1489	return skb->len;
1490
1491nla_put_failure:
1492	nlmsg_trim(skb, b);
1493	return -1;
1494}
1495
1496static int cbq_dump_fopt(struct sk_buff *skb, struct cbq_class *cl)
1497{
1498	unsigned char *b = skb_tail_pointer(skb);
1499	struct tc_cbq_fopt opt;
1500
1501	if (cl->split || cl->defmap) {
1502		opt.split = cl->split ? cl->split->common.classid : 0;
1503		opt.defmap = cl->defmap;
1504		opt.defchange = ~0;
1505		NLA_PUT(skb, TCA_CBQ_FOPT, sizeof(opt), &opt);
1506	}
1507	return skb->len;
1508
1509nla_put_failure:
1510	nlmsg_trim(skb, b);
1511	return -1;
1512}
1513
1514#ifdef CONFIG_NET_CLS_ACT
1515static int cbq_dump_police(struct sk_buff *skb, struct cbq_class *cl)
1516{
1517	unsigned char *b = skb_tail_pointer(skb);
1518	struct tc_cbq_police opt;
1519
1520	if (cl->police) {
1521		opt.police = cl->police;
1522		opt.__res1 = 0;
1523		opt.__res2 = 0;
1524		NLA_PUT(skb, TCA_CBQ_POLICE, sizeof(opt), &opt);
1525	}
1526	return skb->len;
1527
1528nla_put_failure:
1529	nlmsg_trim(skb, b);
1530	return -1;
1531}
1532#endif
1533
1534static int cbq_dump_attr(struct sk_buff *skb, struct cbq_class *cl)
1535{
1536	if (cbq_dump_lss(skb, cl) < 0 ||
1537	    cbq_dump_rate(skb, cl) < 0 ||
1538	    cbq_dump_wrr(skb, cl) < 0 ||
1539	    cbq_dump_ovl(skb, cl) < 0 ||
1540#ifdef CONFIG_NET_CLS_ACT
1541	    cbq_dump_police(skb, cl) < 0 ||
1542#endif
1543	    cbq_dump_fopt(skb, cl) < 0)
1544		return -1;
1545	return 0;
1546}
1547
1548static int cbq_dump(struct Qdisc *sch, struct sk_buff *skb)
1549{
1550	struct cbq_sched_data *q = qdisc_priv(sch);
1551	struct nlattr *nest;
1552
1553	nest = nla_nest_start(skb, TCA_OPTIONS);
1554	if (nest == NULL)
1555		goto nla_put_failure;
1556	if (cbq_dump_attr(skb, &q->link) < 0)
1557		goto nla_put_failure;
1558	nla_nest_end(skb, nest);
1559	return skb->len;
1560
1561nla_put_failure:
1562	nla_nest_cancel(skb, nest);
1563	return -1;
1564}
1565
1566static int
1567cbq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1568{
1569	struct cbq_sched_data *q = qdisc_priv(sch);
1570
1571	q->link.xstats.avgidle = q->link.avgidle;
1572	return gnet_stats_copy_app(d, &q->link.xstats, sizeof(q->link.xstats));
1573}
1574
1575static int
1576cbq_dump_class(struct Qdisc *sch, unsigned long arg,
1577	       struct sk_buff *skb, struct tcmsg *tcm)
1578{
1579	struct cbq_class *cl = (struct cbq_class *)arg;
1580	struct nlattr *nest;
1581
1582	if (cl->tparent)
1583		tcm->tcm_parent = cl->tparent->common.classid;
1584	else
1585		tcm->tcm_parent = TC_H_ROOT;
1586	tcm->tcm_handle = cl->common.classid;
1587	tcm->tcm_info = cl->q->handle;
1588
1589	nest = nla_nest_start(skb, TCA_OPTIONS);
1590	if (nest == NULL)
1591		goto nla_put_failure;
1592	if (cbq_dump_attr(skb, cl) < 0)
1593		goto nla_put_failure;
1594	nla_nest_end(skb, nest);
1595	return skb->len;
1596
1597nla_put_failure:
1598	nla_nest_cancel(skb, nest);
1599	return -1;
1600}
1601
1602static int
1603cbq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1604	struct gnet_dump *d)
1605{
1606	struct cbq_sched_data *q = qdisc_priv(sch);
1607	struct cbq_class *cl = (struct cbq_class *)arg;
 
1608
1609	cl->qstats.qlen = cl->q->q.qlen;
1610	cl->xstats.avgidle = cl->avgidle;
1611	cl->xstats.undertime = 0;
 
1612
1613	if (cl->undertime != PSCHED_PASTPERFECT)
1614		cl->xstats.undertime = cl->undertime - q->now;
1615
1616	if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1617	    gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
1618	    gnet_stats_copy_queue(d, &cl->qstats) < 0)
 
1619		return -1;
1620
1621	return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1622}
1623
1624static int cbq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1625		     struct Qdisc **old)
1626{
1627	struct cbq_class *cl = (struct cbq_class *)arg;
1628
1629	if (new == NULL) {
1630		new = qdisc_create_dflt(sch->dev_queue,
1631					&pfifo_qdisc_ops, cl->common.classid);
1632		if (new == NULL)
1633			return -ENOBUFS;
1634	} else {
1635#ifdef CONFIG_NET_CLS_ACT
1636		if (cl->police == TC_POLICE_RECLASSIFY)
1637			new->reshape_fail = cbq_reshape_fail;
1638#endif
1639	}
1640	sch_tree_lock(sch);
1641	*old = cl->q;
1642	cl->q = new;
1643	qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1644	qdisc_reset(*old);
1645	sch_tree_unlock(sch);
1646
 
1647	return 0;
1648}
1649
1650static struct Qdisc *cbq_leaf(struct Qdisc *sch, unsigned long arg)
1651{
1652	struct cbq_class *cl = (struct cbq_class *)arg;
1653
1654	return cl->q;
1655}
1656
1657static void cbq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1658{
1659	struct cbq_class *cl = (struct cbq_class *)arg;
1660
1661	if (cl->q->q.qlen == 0)
1662		cbq_deactivate_class(cl);
1663}
1664
1665static unsigned long cbq_get(struct Qdisc *sch, u32 classid)
1666{
1667	struct cbq_sched_data *q = qdisc_priv(sch);
1668	struct cbq_class *cl = cbq_class_lookup(q, classid);
1669
1670	if (cl) {
1671		cl->refcnt++;
1672		return (unsigned long)cl;
1673	}
1674	return 0;
1675}
1676
1677static void cbq_destroy_class(struct Qdisc *sch, struct cbq_class *cl)
1678{
1679	struct cbq_sched_data *q = qdisc_priv(sch);
1680
1681	WARN_ON(cl->filters);
1682
1683	tcf_destroy_chain(&cl->filter_list);
1684	qdisc_destroy(cl->q);
1685	qdisc_put_rtab(cl->R_tab);
1686	gen_kill_estimator(&cl->bstats, &cl->rate_est);
1687	if (cl != &q->link)
1688		kfree(cl);
1689}
1690
1691static void cbq_destroy(struct Qdisc *sch)
1692{
1693	struct cbq_sched_data *q = qdisc_priv(sch);
1694	struct hlist_node *n, *next;
1695	struct cbq_class *cl;
1696	unsigned int h;
1697
1698#ifdef CONFIG_NET_CLS_ACT
1699	q->rx_class = NULL;
1700#endif
1701	/*
1702	 * Filters must be destroyed first because we don't destroy the
1703	 * classes from root to leafs which means that filters can still
1704	 * be bound to classes which have been destroyed already. --TGR '04
1705	 */
1706	for (h = 0; h < q->clhash.hashsize; h++) {
1707		hlist_for_each_entry(cl, n, &q->clhash.hash[h], common.hnode)
1708			tcf_destroy_chain(&cl->filter_list);
 
 
1709	}
1710	for (h = 0; h < q->clhash.hashsize; h++) {
1711		hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[h],
1712					  common.hnode)
1713			cbq_destroy_class(sch, cl);
1714	}
1715	qdisc_class_hash_destroy(&q->clhash);
1716}
1717
1718static void cbq_put(struct Qdisc *sch, unsigned long arg)
1719{
1720	struct cbq_class *cl = (struct cbq_class *)arg;
1721
1722	if (--cl->refcnt == 0) {
1723#ifdef CONFIG_NET_CLS_ACT
1724		spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1725		struct cbq_sched_data *q = qdisc_priv(sch);
1726
1727		spin_lock_bh(root_lock);
1728		if (q->rx_class == cl)
1729			q->rx_class = NULL;
1730		spin_unlock_bh(root_lock);
1731#endif
1732
1733		cbq_destroy_class(sch, cl);
1734	}
1735}
1736
1737static int
1738cbq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, struct nlattr **tca,
1739		 unsigned long *arg)
1740{
1741	int err;
1742	struct cbq_sched_data *q = qdisc_priv(sch);
1743	struct cbq_class *cl = (struct cbq_class *)*arg;
1744	struct nlattr *opt = tca[TCA_OPTIONS];
1745	struct nlattr *tb[TCA_CBQ_MAX + 1];
1746	struct cbq_class *parent;
1747	struct qdisc_rate_table *rtab = NULL;
1748
1749	if (opt == NULL)
1750		return -EINVAL;
1751
1752	err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy);
1753	if (err < 0)
1754		return err;
1755
 
 
 
 
 
1756	if (cl) {
1757		/* Check parent */
1758		if (parentid) {
1759			if (cl->tparent &&
1760			    cl->tparent->common.classid != parentid)
 
1761				return -EINVAL;
1762			if (!cl->tparent && parentid != TC_H_ROOT)
 
 
1763				return -EINVAL;
 
1764		}
1765
1766		if (tb[TCA_CBQ_RATE]) {
1767			rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]),
1768					      tb[TCA_CBQ_RTAB]);
1769			if (rtab == NULL)
1770				return -EINVAL;
1771		}
1772
1773		if (tca[TCA_RATE]) {
1774			err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1775						    qdisc_root_sleeping_lock(sch),
 
 
1776						    tca[TCA_RATE]);
1777			if (err) {
1778				if (rtab)
1779					qdisc_put_rtab(rtab);
1780				return err;
1781			}
1782		}
1783
1784		/* Change class parameters */
1785		sch_tree_lock(sch);
1786
1787		if (cl->next_alive != NULL)
1788			cbq_deactivate_class(cl);
1789
1790		if (rtab) {
1791			qdisc_put_rtab(cl->R_tab);
1792			cl->R_tab = rtab;
1793		}
1794
1795		if (tb[TCA_CBQ_LSSOPT])
1796			cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1797
1798		if (tb[TCA_CBQ_WRROPT]) {
1799			cbq_rmprio(q, cl);
1800			cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1801		}
1802
1803		if (tb[TCA_CBQ_OVL_STRATEGY])
1804			cbq_set_overlimit(cl, nla_data(tb[TCA_CBQ_OVL_STRATEGY]));
1805
1806#ifdef CONFIG_NET_CLS_ACT
1807		if (tb[TCA_CBQ_POLICE])
1808			cbq_set_police(cl, nla_data(tb[TCA_CBQ_POLICE]));
1809#endif
1810
1811		if (tb[TCA_CBQ_FOPT])
1812			cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1813
1814		if (cl->q->q.qlen)
1815			cbq_activate_class(cl);
1816
1817		sch_tree_unlock(sch);
1818
1819		return 0;
1820	}
1821
1822	if (parentid == TC_H_ROOT)
1823		return -EINVAL;
1824
1825	if (tb[TCA_CBQ_WRROPT] == NULL || tb[TCA_CBQ_RATE] == NULL ||
1826	    tb[TCA_CBQ_LSSOPT] == NULL)
1827		return -EINVAL;
 
1828
1829	rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]), tb[TCA_CBQ_RTAB]);
 
1830	if (rtab == NULL)
1831		return -EINVAL;
1832
1833	if (classid) {
1834		err = -EINVAL;
1835		if (TC_H_MAJ(classid ^ sch->handle) ||
1836		    cbq_class_lookup(q, classid))
 
1837			goto failure;
 
1838	} else {
1839		int i;
1840		classid = TC_H_MAKE(sch->handle, 0x8000);
1841
1842		for (i = 0; i < 0x8000; i++) {
1843			if (++q->hgenerator >= 0x8000)
1844				q->hgenerator = 1;
1845			if (cbq_class_lookup(q, classid|q->hgenerator) == NULL)
1846				break;
1847		}
1848		err = -ENOSR;
1849		if (i >= 0x8000)
 
1850			goto failure;
 
1851		classid = classid|q->hgenerator;
1852	}
1853
1854	parent = &q->link;
1855	if (parentid) {
1856		parent = cbq_class_lookup(q, parentid);
1857		err = -EINVAL;
1858		if (parent == NULL)
 
1859			goto failure;
 
1860	}
1861
1862	err = -ENOBUFS;
1863	cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1864	if (cl == NULL)
1865		goto failure;
1866
 
 
 
 
 
 
1867	if (tca[TCA_RATE]) {
1868		err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1869					qdisc_root_sleeping_lock(sch),
 
1870					tca[TCA_RATE]);
1871		if (err) {
 
 
1872			kfree(cl);
1873			goto failure;
1874		}
1875	}
1876
1877	cl->R_tab = rtab;
1878	rtab = NULL;
1879	cl->refcnt = 1;
1880	cl->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, classid);
1881	if (!cl->q)
1882		cl->q = &noop_qdisc;
 
 
 
1883	cl->common.classid = classid;
1884	cl->tparent = parent;
1885	cl->qdisc = sch;
1886	cl->allot = parent->allot;
1887	cl->quantum = cl->allot;
1888	cl->weight = cl->R_tab->rate.rate;
1889
1890	sch_tree_lock(sch);
1891	cbq_link_class(cl);
1892	cl->borrow = cl->tparent;
1893	if (cl->tparent != &q->link)
1894		cl->share = cl->tparent;
1895	cbq_adjust_levels(parent);
1896	cl->minidle = -0x7FFFFFFF;
1897	cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1898	cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1899	if (cl->ewma_log == 0)
1900		cl->ewma_log = q->link.ewma_log;
1901	if (cl->maxidle == 0)
1902		cl->maxidle = q->link.maxidle;
1903	if (cl->avpkt == 0)
1904		cl->avpkt = q->link.avpkt;
1905	cl->overlimit = cbq_ovl_classic;
1906	if (tb[TCA_CBQ_OVL_STRATEGY])
1907		cbq_set_overlimit(cl, nla_data(tb[TCA_CBQ_OVL_STRATEGY]));
1908#ifdef CONFIG_NET_CLS_ACT
1909	if (tb[TCA_CBQ_POLICE])
1910		cbq_set_police(cl, nla_data(tb[TCA_CBQ_POLICE]));
1911#endif
1912	if (tb[TCA_CBQ_FOPT])
1913		cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1914	sch_tree_unlock(sch);
1915
1916	qdisc_class_hash_grow(sch, &q->clhash);
1917
1918	*arg = (unsigned long)cl;
1919	return 0;
1920
1921failure:
1922	qdisc_put_rtab(rtab);
1923	return err;
1924}
1925
1926static int cbq_delete(struct Qdisc *sch, unsigned long arg)
1927{
1928	struct cbq_sched_data *q = qdisc_priv(sch);
1929	struct cbq_class *cl = (struct cbq_class *)arg;
1930	unsigned int qlen;
1931
1932	if (cl->filters || cl->children || cl == &q->link)
1933		return -EBUSY;
1934
1935	sch_tree_lock(sch);
1936
1937	qlen = cl->q->q.qlen;
1938	qdisc_reset(cl->q);
1939	qdisc_tree_decrease_qlen(cl->q, qlen);
1940
1941	if (cl->next_alive)
1942		cbq_deactivate_class(cl);
1943
1944	if (q->tx_borrowed == cl)
1945		q->tx_borrowed = q->tx_class;
1946	if (q->tx_class == cl) {
1947		q->tx_class = NULL;
1948		q->tx_borrowed = NULL;
1949	}
1950#ifdef CONFIG_NET_CLS_ACT
1951	if (q->rx_class == cl)
1952		q->rx_class = NULL;
1953#endif
1954
1955	cbq_unlink_class(cl);
1956	cbq_adjust_levels(cl->tparent);
1957	cl->defmap = 0;
1958	cbq_sync_defmap(cl);
1959
1960	cbq_rmprio(q, cl);
1961	sch_tree_unlock(sch);
1962
1963	BUG_ON(--cl->refcnt == 0);
1964	/*
1965	 * This shouldn't happen: we "hold" one cops->get() when called
1966	 * from tc_ctl_tclass; the destroy method is done from cops->put().
1967	 */
1968
1969	return 0;
1970}
1971
1972static struct tcf_proto **cbq_find_tcf(struct Qdisc *sch, unsigned long arg)
 
1973{
1974	struct cbq_sched_data *q = qdisc_priv(sch);
1975	struct cbq_class *cl = (struct cbq_class *)arg;
1976
1977	if (cl == NULL)
1978		cl = &q->link;
1979
1980	return &cl->filter_list;
1981}
1982
1983static unsigned long cbq_bind_filter(struct Qdisc *sch, unsigned long parent,
1984				     u32 classid)
1985{
1986	struct cbq_sched_data *q = qdisc_priv(sch);
1987	struct cbq_class *p = (struct cbq_class *)parent;
1988	struct cbq_class *cl = cbq_class_lookup(q, classid);
1989
1990	if (cl) {
1991		if (p && p->level <= cl->level)
1992			return 0;
1993		cl->filters++;
1994		return (unsigned long)cl;
1995	}
1996	return 0;
1997}
1998
1999static void cbq_unbind_filter(struct Qdisc *sch, unsigned long arg)
2000{
2001	struct cbq_class *cl = (struct cbq_class *)arg;
2002
2003	cl->filters--;
2004}
2005
2006static void cbq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2007{
2008	struct cbq_sched_data *q = qdisc_priv(sch);
2009	struct cbq_class *cl;
2010	struct hlist_node *n;
2011	unsigned int h;
2012
2013	if (arg->stop)
2014		return;
2015
2016	for (h = 0; h < q->clhash.hashsize; h++) {
2017		hlist_for_each_entry(cl, n, &q->clhash.hash[h], common.hnode) {
2018			if (arg->count < arg->skip) {
2019				arg->count++;
2020				continue;
2021			}
2022			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
2023				arg->stop = 1;
2024				return;
2025			}
2026			arg->count++;
2027		}
2028	}
2029}
2030
2031static const struct Qdisc_class_ops cbq_class_ops = {
2032	.graft		=	cbq_graft,
2033	.leaf		=	cbq_leaf,
2034	.qlen_notify	=	cbq_qlen_notify,
2035	.get		=	cbq_get,
2036	.put		=	cbq_put,
2037	.change		=	cbq_change_class,
2038	.delete		=	cbq_delete,
2039	.walk		=	cbq_walk,
2040	.tcf_chain	=	cbq_find_tcf,
2041	.bind_tcf	=	cbq_bind_filter,
2042	.unbind_tcf	=	cbq_unbind_filter,
2043	.dump		=	cbq_dump_class,
2044	.dump_stats	=	cbq_dump_class_stats,
2045};
2046
2047static struct Qdisc_ops cbq_qdisc_ops __read_mostly = {
2048	.next		=	NULL,
2049	.cl_ops		=	&cbq_class_ops,
2050	.id		=	"cbq",
2051	.priv_size	=	sizeof(struct cbq_sched_data),
2052	.enqueue	=	cbq_enqueue,
2053	.dequeue	=	cbq_dequeue,
2054	.peek		=	qdisc_peek_dequeued,
2055	.drop		=	cbq_drop,
2056	.init		=	cbq_init,
2057	.reset		=	cbq_reset,
2058	.destroy	=	cbq_destroy,
2059	.change		=	NULL,
2060	.dump		=	cbq_dump,
2061	.dump_stats	=	cbq_dump_stats,
2062	.owner		=	THIS_MODULE,
2063};
2064
2065static int __init cbq_module_init(void)
2066{
2067	return register_qdisc(&cbq_qdisc_ops);
2068}
2069static void __exit cbq_module_exit(void)
2070{
2071	unregister_qdisc(&cbq_qdisc_ops);
2072}
2073module_init(cbq_module_init)
2074module_exit(cbq_module_exit)
2075MODULE_LICENSE("GPL");