sch_hfsc.c

   1/*
   2 * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
   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 2
   7 * of the License, or (at your option) any later version.
   8 *
   9 * 2003-10-17 - Ported from altq
  10 */
  11/*
  12 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
  13 *
  14 * Permission to use, copy, modify, and distribute this software and
  15 * its documentation is hereby granted (including for commercial or
  16 * for-profit use), provided that both the copyright notice and this
  17 * permission notice appear in all copies of the software, derivative
  18 * works, or modified versions, and any portions thereof.
  19 *
  20 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
  21 * WHICH MAY HAVE SERIOUS CONSEQUENCES.  CARNEGIE MELLON PROVIDES THIS
  22 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
  23 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  25 * DISCLAIMED.  IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
  26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
  28 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  29 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  30 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
  32 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  33 * DAMAGE.
  34 *
  35 * Carnegie Mellon encourages (but does not require) users of this
  36 * software to return any improvements or extensions that they make,
  37 * and to grant Carnegie Mellon the rights to redistribute these
  38 * changes without encumbrance.
  39 */
  40/*
  41 * H-FSC is described in Proceedings of SIGCOMM'97,
  42 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
  43 * Real-Time and Priority Service"
  44 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
  45 *
  46 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
  47 * when a class has an upperlimit, the fit-time is computed from the
  48 * upperlimit service curve.  the link-sharing scheduler does not schedule
  49 * a class whose fit-time exceeds the current time.
  50 */
  51
  52#include <linux/kernel.h>
  53#include <linux/module.h>
  54#include <linux/types.h>
  55#include <linux/errno.h>
  56#include <linux/compiler.h>
  57#include <linux/spinlock.h>
  58#include <linux/skbuff.h>
  59#include <linux/string.h>
  60#include <linux/slab.h>
  61#include <linux/list.h>
  62#include <linux/rbtree.h>
  63#include <linux/init.h>
  64#include <linux/rtnetlink.h>
  65#include <linux/pkt_sched.h>
  66#include <net/netlink.h>
  67#include <net/pkt_sched.h>
  68#include <net/pkt_cls.h>
  69#include <asm/div64.h>
  70
  71/*
  72 * kernel internal service curve representation:
  73 *   coordinates are given by 64 bit unsigned integers.
  74 *   x-axis: unit is clock count.
  75 *   y-axis: unit is byte.
  76 *
  77 *   The service curve parameters are converted to the internal
  78 *   representation. The slope values are scaled to avoid overflow.
  79 *   the inverse slope values as well as the y-projection of the 1st
  80 *   segment are kept in order to avoid 64-bit divide operations
  81 *   that are expensive on 32-bit architectures.
  82 */
  83
  84struct internal_sc {
  85	u64	sm1;	/* scaled slope of the 1st segment */
  86	u64	ism1;	/* scaled inverse-slope of the 1st segment */
  87	u64	dx;	/* the x-projection of the 1st segment */
  88	u64	dy;	/* the y-projection of the 1st segment */
  89	u64	sm2;	/* scaled slope of the 2nd segment */
  90	u64	ism2;	/* scaled inverse-slope of the 2nd segment */
  91};
  92
  93/* runtime service curve */
  94struct runtime_sc {
  95	u64	x;	/* current starting position on x-axis */
  96	u64	y;	/* current starting position on y-axis */
  97	u64	sm1;	/* scaled slope of the 1st segment */
  98	u64	ism1;	/* scaled inverse-slope of the 1st segment */
  99	u64	dx;	/* the x-projection of the 1st segment */
 100	u64	dy;	/* the y-projection of the 1st segment */
 101	u64	sm2;	/* scaled slope of the 2nd segment */
 102	u64	ism2;	/* scaled inverse-slope of the 2nd segment */
 103};
 104
 105enum hfsc_class_flags {
 106	HFSC_RSC = 0x1,
 107	HFSC_FSC = 0x2,
 108	HFSC_USC = 0x4
 109};
 110
 111struct hfsc_class {
 112	struct Qdisc_class_common cl_common;
 113
 114	struct gnet_stats_basic_sync bstats;
 115	struct gnet_stats_queue qstats;
 116	struct net_rate_estimator __rcu *rate_est;
 117	struct tcf_proto __rcu *filter_list; /* filter list */
 118	struct tcf_block *block;
 119	unsigned int	level;		/* class level in hierarchy */
 120
 121	struct hfsc_sched *sched;	/* scheduler data */
 122	struct hfsc_class *cl_parent;	/* parent class */
 123	struct list_head siblings;	/* sibling classes */
 124	struct list_head children;	/* child classes */
 125	struct Qdisc	*qdisc;		/* leaf qdisc */
 126
 127	struct rb_node el_node;		/* qdisc's eligible tree member */
 128	struct rb_root vt_tree;		/* active children sorted by cl_vt */
 129	struct rb_node vt_node;		/* parent's vt_tree member */
 130	struct rb_root cf_tree;		/* active children sorted by cl_f */
 131	struct rb_node cf_node;		/* parent's cf_heap member */
 132
 133	u64	cl_total;		/* total work in bytes */
 134	u64	cl_cumul;		/* cumulative work in bytes done by
 135					   real-time criteria */
 136
 137	u64	cl_d;			/* deadline*/
 138	u64	cl_e;			/* eligible time */
 139	u64	cl_vt;			/* virtual time */
 140	u64	cl_f;			/* time when this class will fit for
 141					   link-sharing, max(myf, cfmin) */
 142	u64	cl_myf;			/* my fit-time (calculated from this
 143					   class's own upperlimit curve) */
 144	u64	cl_cfmin;		/* earliest children's fit-time (used
 145					   with cl_myf to obtain cl_f) */
 146	u64	cl_cvtmin;		/* minimal virtual time among the
 147					   children fit for link-sharing
 148					   (monotonic within a period) */
 149	u64	cl_vtadj;		/* intra-period cumulative vt
 150					   adjustment */
 151	u64	cl_cvtoff;		/* largest virtual time seen among
 152					   the children */
 153
 154	struct internal_sc cl_rsc;	/* internal real-time service curve */
 155	struct internal_sc cl_fsc;	/* internal fair service curve */
 156	struct internal_sc cl_usc;	/* internal upperlimit service curve */
 157	struct runtime_sc cl_deadline;	/* deadline curve */
 158	struct runtime_sc cl_eligible;	/* eligible curve */
 159	struct runtime_sc cl_virtual;	/* virtual curve */
 160	struct runtime_sc cl_ulimit;	/* upperlimit curve */
 161
 162	u8		cl_flags;	/* which curves are valid */
 163	u32		cl_vtperiod;	/* vt period sequence number */
 164	u32		cl_parentperiod;/* parent's vt period sequence number*/
 165	u32		cl_nactive;	/* number of active children */
 166};
 167
 168struct hfsc_sched {
 169	u16	defcls;				/* default class id */
 170	struct hfsc_class root;			/* root class */
 171	struct Qdisc_class_hash clhash;		/* class hash */
 172	struct rb_root eligible;		/* eligible tree */
 173	struct qdisc_watchdog watchdog;		/* watchdog timer */
 174};
 175
 176#define	HT_INFINITY	0xffffffffffffffffULL	/* infinite time value */
 177
 178static bool cl_in_el_or_vttree(struct hfsc_class *cl)
 179{
 180	return ((cl->cl_flags & HFSC_FSC) && cl->cl_nactive) ||
 181		((cl->cl_flags & HFSC_RSC) && !RB_EMPTY_NODE(&cl->el_node));
 182}
 183
 184/*
 185 * eligible tree holds backlogged classes being sorted by their eligible times.
 186 * there is one eligible tree per hfsc instance.
 187 */
 188
 189static void
 190eltree_insert(struct hfsc_class *cl)
 191{
 192	struct rb_node **p = &cl->sched->eligible.rb_node;
 193	struct rb_node *parent = NULL;
 194	struct hfsc_class *cl1;
 195
 196	while (*p != NULL) {
 197		parent = *p;
 198		cl1 = rb_entry(parent, struct hfsc_class, el_node);
 199		if (cl->cl_e >= cl1->cl_e)
 200			p = &parent->rb_right;
 201		else
 202			p = &parent->rb_left;
 203	}
 204	rb_link_node(&cl->el_node, parent, p);
 205	rb_insert_color(&cl->el_node, &cl->sched->eligible);
 206}
 207
 208static inline void
 209eltree_remove(struct hfsc_class *cl)
 210{
 211	if (!RB_EMPTY_NODE(&cl->el_node)) {
 212		rb_erase(&cl->el_node, &cl->sched->eligible);
 213		RB_CLEAR_NODE(&cl->el_node);
 214	}
 215}
 216
 217static inline void
 218eltree_update(struct hfsc_class *cl)
 219{
 220	eltree_remove(cl);
 221	eltree_insert(cl);
 222}
 223
 224/* find the class with the minimum deadline among the eligible classes */
 225static inline struct hfsc_class *
 226eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
 227{
 228	struct hfsc_class *p, *cl = NULL;
 229	struct rb_node *n;
 230
 231	for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
 232		p = rb_entry(n, struct hfsc_class, el_node);
 233		if (p->cl_e > cur_time)
 234			break;
 235		if (cl == NULL || p->cl_d < cl->cl_d)
 236			cl = p;
 237	}
 238	return cl;
 239}
 240
 241/* find the class with minimum eligible time among the eligible classes */
 242static inline struct hfsc_class *
 243eltree_get_minel(struct hfsc_sched *q)
 244{
 245	struct rb_node *n;
 246
 247	n = rb_first(&q->eligible);
 248	if (n == NULL)
 249		return NULL;
 250	return rb_entry(n, struct hfsc_class, el_node);
 251}
 252
 253/*
 254 * vttree holds holds backlogged child classes being sorted by their virtual
 255 * time. each intermediate class has one vttree.
 256 */
 257static void
 258vttree_insert(struct hfsc_class *cl)
 259{
 260	struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
 261	struct rb_node *parent = NULL;
 262	struct hfsc_class *cl1;
 263
 264	while (*p != NULL) {
 265		parent = *p;
 266		cl1 = rb_entry(parent, struct hfsc_class, vt_node);
 267		if (cl->cl_vt >= cl1->cl_vt)
 268			p = &parent->rb_right;
 269		else
 270			p = &parent->rb_left;
 271	}
 272	rb_link_node(&cl->vt_node, parent, p);
 273	rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
 274}
 275
 276static inline void
 277vttree_remove(struct hfsc_class *cl)
 278{
 279	rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
 280}
 281
 282static inline void
 283vttree_update(struct hfsc_class *cl)
 284{
 285	vttree_remove(cl);
 286	vttree_insert(cl);
 287}
 288
 289static inline struct hfsc_class *
 290vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
 291{
 292	struct hfsc_class *p;
 293	struct rb_node *n;
 294
 295	for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
 296		p = rb_entry(n, struct hfsc_class, vt_node);
 297		if (p->cl_f <= cur_time)
 298			return p;
 299	}
 300	return NULL;
 301}
 302
 303/*
 304 * get the leaf class with the minimum vt in the hierarchy
 305 */
 306static struct hfsc_class *
 307vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
 308{
 309	/* if root-class's cfmin is bigger than cur_time nothing to do */
 310	if (cl->cl_cfmin > cur_time)
 311		return NULL;
 312
 313	while (cl->level > 0) {
 314		cl = vttree_firstfit(cl, cur_time);
 315		if (cl == NULL)
 316			return NULL;
 317		/*
 318		 * update parent's cl_cvtmin.
 319		 */
 320		if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
 321			cl->cl_parent->cl_cvtmin = cl->cl_vt;
 322	}
 323	return cl;
 324}
 325
 326static void
 327cftree_insert(struct hfsc_class *cl)
 328{
 329	struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
 330	struct rb_node *parent = NULL;
 331	struct hfsc_class *cl1;
 332
 333	while (*p != NULL) {
 334		parent = *p;
 335		cl1 = rb_entry(parent, struct hfsc_class, cf_node);
 336		if (cl->cl_f >= cl1->cl_f)
 337			p = &parent->rb_right;
 338		else
 339			p = &parent->rb_left;
 340	}
 341	rb_link_node(&cl->cf_node, parent, p);
 342	rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
 343}
 344
 345static inline void
 346cftree_remove(struct hfsc_class *cl)
 347{
 348	rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
 349}
 350
 351static inline void
 352cftree_update(struct hfsc_class *cl)
 353{
 354	cftree_remove(cl);
 355	cftree_insert(cl);
 356}
 357
 358/*
 359 * service curve support functions
 360 *
 361 *  external service curve parameters
 362 *	m: bps
 363 *	d: us
 364 *  internal service curve parameters
 365 *	sm: (bytes/psched_us) << SM_SHIFT
 366 *	ism: (psched_us/byte) << ISM_SHIFT
 367 *	dx: psched_us
 368 *
 369 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
 370 *
 371 * sm and ism are scaled in order to keep effective digits.
 372 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
 373 * digits in decimal using the following table.
 374 *
 375 *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
 376 *  ------------+-------------------------------------------------------
 377 *  bytes/1.024us 12.8e-3    128e-3     1280e-3    12800e-3   128000e-3
 378 *
 379 *  1.024us/byte  78.125     7.8125     0.78125    0.078125   0.0078125
 380 *
 381 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
 382 */
 383#define	SM_SHIFT	(30 - PSCHED_SHIFT)
 384#define	ISM_SHIFT	(8 + PSCHED_SHIFT)
 385
 386#define	SM_MASK		((1ULL << SM_SHIFT) - 1)
 387#define	ISM_MASK	((1ULL << ISM_SHIFT) - 1)
 388
 389static inline u64
 390seg_x2y(u64 x, u64 sm)
 391{
 392	u64 y;
 393
 394	/*
 395	 * compute
 396	 *	y = x * sm >> SM_SHIFT
 397	 * but divide it for the upper and lower bits to avoid overflow
 398	 */
 399	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
 400	return y;
 401}
 402
 403static inline u64
 404seg_y2x(u64 y, u64 ism)
 405{
 406	u64 x;
 407
 408	if (y == 0)
 409		x = 0;
 410	else if (ism == HT_INFINITY)
 411		x = HT_INFINITY;
 412	else {
 413		x = (y >> ISM_SHIFT) * ism
 414		    + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
 415	}
 416	return x;
 417}
 418
 419/* Convert m (bps) into sm (bytes/psched us) */
 420static u64
 421m2sm(u32 m)
 422{
 423	u64 sm;
 424
 425	sm = ((u64)m << SM_SHIFT);
 426	sm += PSCHED_TICKS_PER_SEC - 1;
 427	do_div(sm, PSCHED_TICKS_PER_SEC);
 428	return sm;
 429}
 430
 431/* convert m (bps) into ism (psched us/byte) */
 432static u64
 433m2ism(u32 m)
 434{
 435	u64 ism;
 436
 437	if (m == 0)
 438		ism = HT_INFINITY;
 439	else {
 440		ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
 441		ism += m - 1;
 442		do_div(ism, m);
 443	}
 444	return ism;
 445}
 446
 447/* convert d (us) into dx (psched us) */
 448static u64
 449d2dx(u32 d)
 450{
 451	u64 dx;
 452
 453	dx = ((u64)d * PSCHED_TICKS_PER_SEC);
 454	dx += USEC_PER_SEC - 1;
 455	do_div(dx, USEC_PER_SEC);
 456	return dx;
 457}
 458
 459/* convert sm (bytes/psched us) into m (bps) */
 460static u32
 461sm2m(u64 sm)
 462{
 463	u64 m;
 464
 465	m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
 466	return (u32)m;
 467}
 468
 469/* convert dx (psched us) into d (us) */
 470static u32
 471dx2d(u64 dx)
 472{
 473	u64 d;
 474
 475	d = dx * USEC_PER_SEC;
 476	do_div(d, PSCHED_TICKS_PER_SEC);
 477	return (u32)d;
 478}
 479
 480static void
 481sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
 482{
 483	isc->sm1  = m2sm(sc->m1);
 484	isc->ism1 = m2ism(sc->m1);
 485	isc->dx   = d2dx(sc->d);
 486	isc->dy   = seg_x2y(isc->dx, isc->sm1);
 487	isc->sm2  = m2sm(sc->m2);
 488	isc->ism2 = m2ism(sc->m2);
 489}
 490
 491/*
 492 * initialize the runtime service curve with the given internal
 493 * service curve starting at (x, y).
 494 */
 495static void
 496rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
 497{
 498	rtsc->x	   = x;
 499	rtsc->y    = y;
 500	rtsc->sm1  = isc->sm1;
 501	rtsc->ism1 = isc->ism1;
 502	rtsc->dx   = isc->dx;
 503	rtsc->dy   = isc->dy;
 504	rtsc->sm2  = isc->sm2;
 505	rtsc->ism2 = isc->ism2;
 506}
 507
 508/*
 509 * calculate the y-projection of the runtime service curve by the
 510 * given x-projection value
 511 */
 512static u64
 513rtsc_y2x(struct runtime_sc *rtsc, u64 y)
 514{
 515	u64 x;
 516
 517	if (y < rtsc->y)
 518		x = rtsc->x;
 519	else if (y <= rtsc->y + rtsc->dy) {
 520		/* x belongs to the 1st segment */
 521		if (rtsc->dy == 0)
 522			x = rtsc->x + rtsc->dx;
 523		else
 524			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
 525	} else {
 526		/* x belongs to the 2nd segment */
 527		x = rtsc->x + rtsc->dx
 528		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
 529	}
 530	return x;
 531}
 532
 533static u64
 534rtsc_x2y(struct runtime_sc *rtsc, u64 x)
 535{
 536	u64 y;
 537
 538	if (x <= rtsc->x)
 539		y = rtsc->y;
 540	else if (x <= rtsc->x + rtsc->dx)
 541		/* y belongs to the 1st segment */
 542		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
 543	else
 544		/* y belongs to the 2nd segment */
 545		y = rtsc->y + rtsc->dy
 546		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
 547	return y;
 548}
 549
 550/*
 551 * update the runtime service curve by taking the minimum of the current
 552 * runtime service curve and the service curve starting at (x, y).
 553 */
 554static void
 555rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
 556{
 557	u64 y1, y2, dx, dy;
 558	u64 dsm;
 559
 560	if (isc->sm1 <= isc->sm2) {
 561		/* service curve is convex */
 562		y1 = rtsc_x2y(rtsc, x);
 563		if (y1 < y)
 564			/* the current rtsc is smaller */
 565			return;
 566		rtsc->x = x;
 567		rtsc->y = y;
 568		return;
 569	}
 570
 571	/*
 572	 * service curve is concave
 573	 * compute the two y values of the current rtsc
 574	 *	y1: at x
 575	 *	y2: at (x + dx)
 576	 */
 577	y1 = rtsc_x2y(rtsc, x);
 578	if (y1 <= y) {
 579		/* rtsc is below isc, no change to rtsc */
 580		return;
 581	}
 582
 583	y2 = rtsc_x2y(rtsc, x + isc->dx);
 584	if (y2 >= y + isc->dy) {
 585		/* rtsc is above isc, replace rtsc by isc */
 586		rtsc->x = x;
 587		rtsc->y = y;
 588		rtsc->dx = isc->dx;
 589		rtsc->dy = isc->dy;
 590		return;
 591	}
 592
 593	/*
 594	 * the two curves intersect
 595	 * compute the offsets (dx, dy) using the reverse
 596	 * function of seg_x2y()
 597	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
 598	 */
 599	dx = (y1 - y) << SM_SHIFT;
 600	dsm = isc->sm1 - isc->sm2;
 601	dx = div64_u64(dx, dsm);
 602	/*
 603	 * check if (x, y1) belongs to the 1st segment of rtsc.
 604	 * if so, add the offset.
 605	 */
 606	if (rtsc->x + rtsc->dx > x)
 607		dx += rtsc->x + rtsc->dx - x;
 608	dy = seg_x2y(dx, isc->sm1);
 609
 610	rtsc->x = x;
 611	rtsc->y = y;
 612	rtsc->dx = dx;
 613	rtsc->dy = dy;
 614}
 615
 616static void
 617init_ed(struct hfsc_class *cl, unsigned int next_len)
 618{
 619	u64 cur_time = psched_get_time();
 620
 621	/* update the deadline curve */
 622	rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
 623
 624	/*
 625	 * update the eligible curve.
 626	 * for concave, it is equal to the deadline curve.
 627	 * for convex, it is a linear curve with slope m2.
 628	 */
 629	cl->cl_eligible = cl->cl_deadline;
 630	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
 631		cl->cl_eligible.dx = 0;
 632		cl->cl_eligible.dy = 0;
 633	}
 634
 635	/* compute e and d */
 636	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
 637	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 638
 639	eltree_insert(cl);
 640}
 641
 642static void
 643update_ed(struct hfsc_class *cl, unsigned int next_len)
 644{
 645	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
 646	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 647
 648	eltree_update(cl);
 649}
 650
 651static inline void
 652update_d(struct hfsc_class *cl, unsigned int next_len)
 653{
 654	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 655}
 656
 657static inline void
 658update_cfmin(struct hfsc_class *cl)
 659{
 660	struct rb_node *n = rb_first(&cl->cf_tree);
 661	struct hfsc_class *p;
 662
 663	if (n == NULL) {
 664		cl->cl_cfmin = 0;
 665		return;
 666	}
 667	p = rb_entry(n, struct hfsc_class, cf_node);
 668	cl->cl_cfmin = p->cl_f;
 669}
 670
 671static void
 672init_vf(struct hfsc_class *cl, unsigned int len)
 673{
 674	struct hfsc_class *max_cl;
 675	struct rb_node *n;
 676	u64 vt, f, cur_time;
 677	int go_active;
 678
 679	cur_time = 0;
 680	go_active = 1;
 681	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
 682		if (go_active && cl->cl_nactive++ == 0)
 683			go_active = 1;
 684		else
 685			go_active = 0;
 686
 687		if (go_active) {
 688			n = rb_last(&cl->cl_parent->vt_tree);
 689			if (n != NULL) {
 690				max_cl = rb_entry(n, struct hfsc_class, vt_node);
 691				/*
 692				 * set vt to the average of the min and max
 693				 * classes.  if the parent's period didn't
 694				 * change, don't decrease vt of the class.
 695				 */
 696				vt = max_cl->cl_vt;
 697				if (cl->cl_parent->cl_cvtmin != 0)
 698					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
 699
 700				if (cl->cl_parent->cl_vtperiod !=
 701				    cl->cl_parentperiod || vt > cl->cl_vt)
 702					cl->cl_vt = vt;
 703			} else {
 704				/*
 705				 * first child for a new parent backlog period.
 706				 * initialize cl_vt to the highest value seen
 707				 * among the siblings. this is analogous to
 708				 * what cur_time would provide in realtime case.
 709				 */
 710				cl->cl_vt = cl->cl_parent->cl_cvtoff;
 711				cl->cl_parent->cl_cvtmin = 0;
 712			}
 713
 714			/* update the virtual curve */
 715			rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
 716			cl->cl_vtadj = 0;
 717
 718			cl->cl_vtperiod++;  /* increment vt period */
 719			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
 720			if (cl->cl_parent->cl_nactive == 0)
 721				cl->cl_parentperiod++;
 722			cl->cl_f = 0;
 723
 724			vttree_insert(cl);
 725			cftree_insert(cl);
 726
 727			if (cl->cl_flags & HFSC_USC) {
 728				/* class has upper limit curve */
 729				if (cur_time == 0)
 730					cur_time = psched_get_time();
 731
 732				/* update the ulimit curve */
 733				rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
 734					 cl->cl_total);
 735				/* compute myf */
 736				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
 737						      cl->cl_total);
 738			}
 739		}
 740
 741		f = max(cl->cl_myf, cl->cl_cfmin);
 742		if (f != cl->cl_f) {
 743			cl->cl_f = f;
 744			cftree_update(cl);
 745		}
 746		update_cfmin(cl->cl_parent);
 747	}
 748}
 749
 750static void
 751update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
 752{
 753	u64 f; /* , myf_bound, delta; */
 754	int go_passive = 0;
 755
 756	if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
 757		go_passive = 1;
 758
 759	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
 760		cl->cl_total += len;
 761
 762		if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
 763			continue;
 764
 765		if (go_passive && --cl->cl_nactive == 0)
 766			go_passive = 1;
 767		else
 768			go_passive = 0;
 769
 770		/* update vt */
 771		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
 772
 773		/*
 774		 * if vt of the class is smaller than cvtmin,
 775		 * the class was skipped in the past due to non-fit.
 776		 * if so, we need to adjust vtadj.
 777		 */
 778		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
 779			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
 780			cl->cl_vt = cl->cl_parent->cl_cvtmin;
 781		}
 782
 783		if (go_passive) {
 784			/* no more active child, going passive */
 785
 786			/* update cvtoff of the parent class */
 787			if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
 788				cl->cl_parent->cl_cvtoff = cl->cl_vt;
 789
 790			/* remove this class from the vt tree */
 791			vttree_remove(cl);
 792
 793			cftree_remove(cl);
 794			update_cfmin(cl->cl_parent);
 795
 796			continue;
 797		}
 798
 799		/* update the vt tree */
 800		vttree_update(cl);
 801
 802		/* update f */
 803		if (cl->cl_flags & HFSC_USC) {
 804			cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
 805#if 0
 806			cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
 807							      cl->cl_total);
 808			/*
 809			 * This code causes classes to stay way under their
 810			 * limit when multiple classes are used at gigabit
 811			 * speed. needs investigation. -kaber
 812			 */
 813			/*
 814			 * if myf lags behind by more than one clock tick
 815			 * from the current time, adjust myfadj to prevent
 816			 * a rate-limited class from going greedy.
 817			 * in a steady state under rate-limiting, myf
 818			 * fluctuates within one clock tick.
 819			 */
 820			myf_bound = cur_time - PSCHED_JIFFIE2US(1);
 821			if (cl->cl_myf < myf_bound) {
 822				delta = cur_time - cl->cl_myf;
 823				cl->cl_myfadj += delta;
 824				cl->cl_myf += delta;
 825			}
 826#endif
 827		}
 828
 829		f = max(cl->cl_myf, cl->cl_cfmin);
 830		if (f != cl->cl_f) {
 831			cl->cl_f = f;
 832			cftree_update(cl);
 833			update_cfmin(cl->cl_parent);
 834		}
 835	}
 836}
 837
 838static void
 839hfsc_adjust_levels(struct hfsc_class *cl)
 840{
 841	struct hfsc_class *p;
 842	unsigned int level;
 843
 844	do {
 845		level = 0;
 846		list_for_each_entry(p, &cl->children, siblings) {
 847			if (p->level >= level)
 848				level = p->level + 1;
 849		}
 850		cl->level = level;
 851	} while ((cl = cl->cl_parent) != NULL);
 852}
 853
 854static inline struct hfsc_class *
 855hfsc_find_class(u32 classid, struct Qdisc *sch)
 856{
 857	struct hfsc_sched *q = qdisc_priv(sch);
 858	struct Qdisc_class_common *clc;
 859
 860	clc = qdisc_class_find(&q->clhash, classid);
 861	if (clc == NULL)
 862		return NULL;
 863	return container_of(clc, struct hfsc_class, cl_common);
 864}
 865
 866static void
 867hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
 868		u64 cur_time)
 869{
 870	sc2isc(rsc, &cl->cl_rsc);
 871	rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
 872	cl->cl_eligible = cl->cl_deadline;
 873	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
 874		cl->cl_eligible.dx = 0;
 875		cl->cl_eligible.dy = 0;
 876	}
 877	cl->cl_flags |= HFSC_RSC;
 878}
 879
 880static void
 881hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
 882{
 883	sc2isc(fsc, &cl->cl_fsc);
 884	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
 885	cl->cl_flags |= HFSC_FSC;
 886}
 887
 888static void
 889hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
 890		u64 cur_time)
 891{
 892	sc2isc(usc, &cl->cl_usc);
 893	rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
 894	cl->cl_flags |= HFSC_USC;
 895}
 896
 897static void
 898hfsc_upgrade_rt(struct hfsc_class *cl)
 899{
 900	cl->cl_fsc = cl->cl_rsc;
 901	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
 902	cl->cl_flags |= HFSC_FSC;
 903}
 904
 905static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
 906	[TCA_HFSC_RSC]	= { .len = sizeof(struct tc_service_curve) },
 907	[TCA_HFSC_FSC]	= { .len = sizeof(struct tc_service_curve) },
 908	[TCA_HFSC_USC]	= { .len = sizeof(struct tc_service_curve) },
 909};
 910
 911static int
 912hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 913		  struct nlattr **tca, unsigned long *arg,
 914		  struct netlink_ext_ack *extack)
 915{
 916	struct hfsc_sched *q = qdisc_priv(sch);
 917	struct hfsc_class *cl = (struct hfsc_class *)*arg;
 918	struct hfsc_class *parent = NULL;
 919	struct nlattr *opt = tca[TCA_OPTIONS];
 920	struct nlattr *tb[TCA_HFSC_MAX + 1];
 921	struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
 922	u64 cur_time;
 923	int err;
 924
 925	if (opt == NULL)
 926		return -EINVAL;
 927
 928	err = nla_parse_nested_deprecated(tb, TCA_HFSC_MAX, opt, hfsc_policy,
 929					  NULL);
 930	if (err < 0)
 931		return err;
 932
 933	if (tb[TCA_HFSC_RSC]) {
 934		rsc = nla_data(tb[TCA_HFSC_RSC]);
 935		if (rsc->m1 == 0 && rsc->m2 == 0)
 936			rsc = NULL;
 937	}
 938
 939	if (tb[TCA_HFSC_FSC]) {
 940		fsc = nla_data(tb[TCA_HFSC_FSC]);
 941		if (fsc->m1 == 0 && fsc->m2 == 0)
 942			fsc = NULL;
 943	}
 944
 945	if (tb[TCA_HFSC_USC]) {
 946		usc = nla_data(tb[TCA_HFSC_USC]);
 947		if (usc->m1 == 0 && usc->m2 == 0)
 948			usc = NULL;
 949	}
 950
 951	if (cl != NULL) {
 952		int old_flags;
 953		int len = 0;
 954
 955		if (parentid) {
 956			if (cl->cl_parent &&
 957			    cl->cl_parent->cl_common.classid != parentid)
 958				return -EINVAL;
 959			if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
 960				return -EINVAL;
 961		}
 962		cur_time = psched_get_time();
 963
 964		if (tca[TCA_RATE]) {
 965			err = gen_replace_estimator(&cl->bstats, NULL,
 966						    &cl->rate_est,
 967						    NULL,
 968						    true,
 969						    tca[TCA_RATE]);
 970			if (err)
 971				return err;
 972		}
 973
 974		sch_tree_lock(sch);
 975		old_flags = cl->cl_flags;
 976
 977		if (rsc != NULL)
 978			hfsc_change_rsc(cl, rsc, cur_time);
 979		if (fsc != NULL)
 980			hfsc_change_fsc(cl, fsc);
 981		if (usc != NULL)
 982			hfsc_change_usc(cl, usc, cur_time);
 983
 984		if (cl->qdisc->q.qlen != 0)
 985			len = qdisc_peek_len(cl->qdisc);
 986		/* Check queue length again since some qdisc implementations
 987		 * (e.g., netem/codel) might empty the queue during the peek
 988		 * operation.
 989		 */
 990		if (cl->qdisc->q.qlen != 0) {
 991			if (cl->cl_flags & HFSC_RSC) {
 992				if (old_flags & HFSC_RSC)
 993					update_ed(cl, len);
 994				else
 995					init_ed(cl, len);
 996			}
 997
 998			if (cl->cl_flags & HFSC_FSC) {
 999				if (old_flags & HFSC_FSC)
1000					update_vf(cl, 0, cur_time);
1001				else
1002					init_vf(cl, len);
1003			}
1004		}
1005		sch_tree_unlock(sch);
1006
1007		return 0;
1008	}
1009
1010	if (parentid == TC_H_ROOT)
1011		return -EEXIST;
1012
1013	parent = &q->root;
1014	if (parentid) {
1015		parent = hfsc_find_class(parentid, sch);
1016		if (parent == NULL)
1017			return -ENOENT;
1018	}
1019
1020	if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1021		return -EINVAL;
1022	if (hfsc_find_class(classid, sch))
1023		return -EEXIST;
1024
1025	if (rsc == NULL && fsc == NULL)
1026		return -EINVAL;
1027
1028	cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1029	if (cl == NULL)
1030		return -ENOBUFS;
1031
1032	RB_CLEAR_NODE(&cl->el_node);
1033
1034	err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1035	if (err) {
1036		kfree(cl);
1037		return err;
1038	}
1039
1040	if (tca[TCA_RATE]) {
1041		err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1042					NULL, true, tca[TCA_RATE]);
1043		if (err) {
1044			tcf_block_put(cl->block);
1045			kfree(cl);
1046			return err;
1047		}
1048	}
1049
1050	if (rsc != NULL)
1051		hfsc_change_rsc(cl, rsc, 0);
1052	if (fsc != NULL)
1053		hfsc_change_fsc(cl, fsc);
1054	if (usc != NULL)
1055		hfsc_change_usc(cl, usc, 0);
1056
1057	cl->cl_common.classid = classid;
1058	cl->sched     = q;
1059	cl->cl_parent = parent;
1060	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1061				      classid, NULL);
1062	if (cl->qdisc == NULL)
1063		cl->qdisc = &noop_qdisc;
1064	else
1065		qdisc_hash_add(cl->qdisc, true);
1066	INIT_LIST_HEAD(&cl->children);
1067	cl->vt_tree = RB_ROOT;
1068	cl->cf_tree = RB_ROOT;
1069
1070	sch_tree_lock(sch);
1071	/* Check if the inner class is a misconfigured 'rt' */
1072	if (!(parent->cl_flags & HFSC_FSC) && parent != &q->root) {
1073		NL_SET_ERR_MSG(extack,
1074			       "Forced curve change on parent 'rt' to 'sc'");
1075		hfsc_upgrade_rt(parent);
1076	}
1077	qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1078	list_add_tail(&cl->siblings, &parent->children);
1079	if (parent->level == 0)
1080		qdisc_purge_queue(parent->qdisc);
1081	hfsc_adjust_levels(parent);
1082	sch_tree_unlock(sch);
1083
1084	qdisc_class_hash_grow(sch, &q->clhash);
1085
1086	*arg = (unsigned long)cl;
1087	return 0;
1088}
1089
1090static void
1091hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1092{
1093	struct hfsc_sched *q = qdisc_priv(sch);
1094
1095	tcf_block_put(cl->block);
1096	qdisc_put(cl->qdisc);
1097	gen_kill_estimator(&cl->rate_est);
1098	if (cl != &q->root)
1099		kfree(cl);
1100}
1101
1102static int
1103hfsc_delete_class(struct Qdisc *sch, unsigned long arg,
1104		  struct netlink_ext_ack *extack)
1105{
1106	struct hfsc_sched *q = qdisc_priv(sch);
1107	struct hfsc_class *cl = (struct hfsc_class *)arg;
1108
1109	if (cl->level > 0 || qdisc_class_in_use(&cl->cl_common) ||
1110	    cl == &q->root) {
1111		NL_SET_ERR_MSG(extack, "HFSC class in use");
1112		return -EBUSY;
1113	}
1114
1115	sch_tree_lock(sch);
1116
1117	list_del(&cl->siblings);
1118	hfsc_adjust_levels(cl->cl_parent);
1119
1120	qdisc_purge_queue(cl->qdisc);
1121	qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1122
1123	sch_tree_unlock(sch);
1124
1125	hfsc_destroy_class(sch, cl);
1126	return 0;
1127}
1128
1129static struct hfsc_class *
1130hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1131{
1132	struct hfsc_sched *q = qdisc_priv(sch);
1133	struct hfsc_class *head, *cl;
1134	struct tcf_result res;
1135	struct tcf_proto *tcf;
1136	int result;
1137
1138	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1139	    (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1140		if (cl->level == 0)
1141			return cl;
1142
1143	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1144	head = &q->root;
1145	tcf = rcu_dereference_bh(q->root.filter_list);
1146	while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
1147#ifdef CONFIG_NET_CLS_ACT
1148		switch (result) {
1149		case TC_ACT_QUEUED:
1150		case TC_ACT_STOLEN:
1151		case TC_ACT_TRAP:
1152			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1153			fallthrough;
1154		case TC_ACT_SHOT:
1155			return NULL;
1156		}
1157#endif
1158		cl = (struct hfsc_class *)res.class;
1159		if (!cl) {
1160			cl = hfsc_find_class(res.classid, sch);
1161			if (!cl)
1162				break; /* filter selected invalid classid */
1163			if (cl->level >= head->level)
1164				break; /* filter may only point downwards */
1165		}
1166
1167		if (cl->level == 0)
1168			return cl; /* hit leaf class */
1169
1170		/* apply inner filter chain */
1171		tcf = rcu_dereference_bh(cl->filter_list);
1172		head = cl;
1173	}
1174
1175	/* classification failed, try default class */
1176	cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle),
1177				       READ_ONCE(q->defcls)), sch);
1178	if (cl == NULL || cl->level > 0)
1179		return NULL;
1180
1181	return cl;
1182}
1183
1184static int
1185hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1186		 struct Qdisc **old, struct netlink_ext_ack *extack)
1187{
1188	struct hfsc_class *cl = (struct hfsc_class *)arg;
1189
1190	if (cl->level > 0)
1191		return -EINVAL;
1192	if (new == NULL) {
1193		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1194					cl->cl_common.classid, NULL);
1195		if (new == NULL)
1196			new = &noop_qdisc;
1197	}
1198
1199	*old = qdisc_replace(sch, new, &cl->qdisc);
1200	return 0;
1201}
1202
1203static struct Qdisc *
1204hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1205{
1206	struct hfsc_class *cl = (struct hfsc_class *)arg;
1207
1208	if (cl->level == 0)
1209		return cl->qdisc;
1210
1211	return NULL;
1212}
1213
1214static void
1215hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1216{
1217	struct hfsc_class *cl = (struct hfsc_class *)arg;
1218
1219	/* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1220	 * needs to be called explicitly to remove a class from vttree.
1221	 */
1222	if (cl->cl_nactive)
1223		update_vf(cl, 0, 0);
1224	if (cl->cl_flags & HFSC_RSC)
1225		eltree_remove(cl);
1226}
1227
1228static unsigned long
1229hfsc_search_class(struct Qdisc *sch, u32 classid)
1230{
1231	return (unsigned long)hfsc_find_class(classid, sch);
1232}
1233
1234static unsigned long
1235hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1236{
1237	struct hfsc_class *p = (struct hfsc_class *)parent;
1238	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1239
1240	if (cl != NULL) {
1241		if (p != NULL && p->level <= cl->level)
1242			return 0;
1243		qdisc_class_get(&cl->cl_common);
1244	}
1245
1246	return (unsigned long)cl;
1247}
1248
1249static void
1250hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1251{
1252	struct hfsc_class *cl = (struct hfsc_class *)arg;
1253
1254	qdisc_class_put(&cl->cl_common);
1255}
1256
1257static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1258					struct netlink_ext_ack *extack)
1259{
1260	struct hfsc_sched *q = qdisc_priv(sch);
1261	struct hfsc_class *cl = (struct hfsc_class *)arg;
1262
1263	if (cl == NULL)
1264		cl = &q->root;
1265
1266	return cl->block;
1267}
1268
1269static int
1270hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1271{
1272	struct tc_service_curve tsc;
1273
1274	tsc.m1 = sm2m(sc->sm1);
1275	tsc.d  = dx2d(sc->dx);
1276	tsc.m2 = sm2m(sc->sm2);
1277	if (nla_put(skb, attr, sizeof(tsc), &tsc))
1278		goto nla_put_failure;
1279
1280	return skb->len;
1281
1282 nla_put_failure:
1283	return -1;
1284}
1285
1286static int
1287hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1288{
1289	if ((cl->cl_flags & HFSC_RSC) &&
1290	    (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1291		goto nla_put_failure;
1292
1293	if ((cl->cl_flags & HFSC_FSC) &&
1294	    (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1295		goto nla_put_failure;
1296
1297	if ((cl->cl_flags & HFSC_USC) &&
1298	    (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1299		goto nla_put_failure;
1300
1301	return skb->len;
1302
1303 nla_put_failure:
1304	return -1;
1305}
1306
1307static int
1308hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1309		struct tcmsg *tcm)
1310{
1311	struct hfsc_class *cl = (struct hfsc_class *)arg;
1312	struct nlattr *nest;
1313
1314	tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1315					  TC_H_ROOT;
1316	tcm->tcm_handle = cl->cl_common.classid;
1317	if (cl->level == 0)
1318		tcm->tcm_info = cl->qdisc->handle;
1319
1320	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1321	if (nest == NULL)
1322		goto nla_put_failure;
1323	if (hfsc_dump_curves(skb, cl) < 0)
1324		goto nla_put_failure;
1325	return nla_nest_end(skb, nest);
1326
1327 nla_put_failure:
1328	nla_nest_cancel(skb, nest);
1329	return -EMSGSIZE;
1330}
1331
1332static int
1333hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1334	struct gnet_dump *d)
1335{
1336	struct hfsc_class *cl = (struct hfsc_class *)arg;
1337	struct tc_hfsc_stats xstats;
1338	__u32 qlen;
1339
1340	qdisc_qstats_qlen_backlog(cl->qdisc, &qlen, &cl->qstats.backlog);
1341	xstats.level   = cl->level;
1342	xstats.period  = cl->cl_vtperiod;
1343	xstats.work    = cl->cl_total;
1344	xstats.rtwork  = cl->cl_cumul;
1345
1346	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
1347	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1348	    gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1349		return -1;
1350
1351	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1352}
1353
1354
1355
1356static void
1357hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1358{
1359	struct hfsc_sched *q = qdisc_priv(sch);
1360	struct hfsc_class *cl;
1361	unsigned int i;
1362
1363	if (arg->stop)
1364		return;
1365
1366	for (i = 0; i < q->clhash.hashsize; i++) {
1367		hlist_for_each_entry(cl, &q->clhash.hash[i],
1368				     cl_common.hnode) {
1369			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
1370				return;
1371		}
1372	}
1373}
1374
1375static void
1376hfsc_schedule_watchdog(struct Qdisc *sch)
1377{
1378	struct hfsc_sched *q = qdisc_priv(sch);
1379	struct hfsc_class *cl;
1380	u64 next_time = 0;
1381
1382	cl = eltree_get_minel(q);
1383	if (cl)
1384		next_time = cl->cl_e;
1385	if (q->root.cl_cfmin != 0) {
1386		if (next_time == 0 || next_time > q->root.cl_cfmin)
1387			next_time = q->root.cl_cfmin;
1388	}
1389	if (next_time)
1390		qdisc_watchdog_schedule(&q->watchdog, next_time);
1391}
1392
1393static int
1394hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1395		struct netlink_ext_ack *extack)
1396{
1397	struct hfsc_sched *q = qdisc_priv(sch);
1398	struct tc_hfsc_qopt *qopt;
1399	int err;
1400
1401	qdisc_watchdog_init(&q->watchdog, sch);
1402
1403	if (!opt || nla_len(opt) < sizeof(*qopt))
1404		return -EINVAL;
1405	qopt = nla_data(opt);
1406
1407	q->defcls = qopt->defcls;
1408	err = qdisc_class_hash_init(&q->clhash);
1409	if (err < 0)
1410		return err;
1411	q->eligible = RB_ROOT;
1412
1413	err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1414	if (err)
1415		return err;
1416
1417	gnet_stats_basic_sync_init(&q->root.bstats);
1418	q->root.cl_common.classid = sch->handle;
1419	q->root.sched   = q;
1420	q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1421					  sch->handle, NULL);
1422	if (q->root.qdisc == NULL)
1423		q->root.qdisc = &noop_qdisc;
1424	else
1425		qdisc_hash_add(q->root.qdisc, true);
1426	INIT_LIST_HEAD(&q->root.children);
1427	q->root.vt_tree = RB_ROOT;
1428	q->root.cf_tree = RB_ROOT;
1429
1430	qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1431	qdisc_class_hash_grow(sch, &q->clhash);
1432
1433	return 0;
1434}
1435
1436static int
1437hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1438		  struct netlink_ext_ack *extack)
1439{
1440	struct hfsc_sched *q = qdisc_priv(sch);
1441	struct tc_hfsc_qopt *qopt;
1442
1443	if (nla_len(opt) < sizeof(*qopt))
1444		return -EINVAL;
1445	qopt = nla_data(opt);
1446
1447	WRITE_ONCE(q->defcls, qopt->defcls);
1448
1449	return 0;
1450}
1451
1452static void
1453hfsc_reset_class(struct hfsc_class *cl)
1454{
1455	cl->cl_total        = 0;
1456	cl->cl_cumul        = 0;
1457	cl->cl_d            = 0;
1458	cl->cl_e            = 0;
1459	cl->cl_vt           = 0;
1460	cl->cl_vtadj        = 0;
1461	cl->cl_cvtmin       = 0;
1462	cl->cl_cvtoff       = 0;
1463	cl->cl_vtperiod     = 0;
1464	cl->cl_parentperiod = 0;
1465	cl->cl_f            = 0;
1466	cl->cl_myf          = 0;
1467	cl->cl_cfmin        = 0;
1468	cl->cl_nactive      = 0;
1469
1470	cl->vt_tree = RB_ROOT;
1471	cl->cf_tree = RB_ROOT;
1472	qdisc_reset(cl->qdisc);
1473
1474	if (cl->cl_flags & HFSC_RSC)
1475		rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1476	if (cl->cl_flags & HFSC_FSC)
1477		rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1478	if (cl->cl_flags & HFSC_USC)
1479		rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1480}
1481
1482static void
1483hfsc_reset_qdisc(struct Qdisc *sch)
1484{
1485	struct hfsc_sched *q = qdisc_priv(sch);
1486	struct hfsc_class *cl;
1487	unsigned int i;
1488
1489	for (i = 0; i < q->clhash.hashsize; i++) {
1490		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1491			hfsc_reset_class(cl);
1492	}
1493	q->eligible = RB_ROOT;
1494	qdisc_watchdog_cancel(&q->watchdog);
1495}
1496
1497static void
1498hfsc_destroy_qdisc(struct Qdisc *sch)
1499{
1500	struct hfsc_sched *q = qdisc_priv(sch);
1501	struct hlist_node *next;
1502	struct hfsc_class *cl;
1503	unsigned int i;
1504
1505	for (i = 0; i < q->clhash.hashsize; i++) {
1506		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1507			tcf_block_put(cl->block);
1508			cl->block = NULL;
1509		}
1510	}
1511	for (i = 0; i < q->clhash.hashsize; i++) {
1512		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1513					  cl_common.hnode)
1514			hfsc_destroy_class(sch, cl);
1515	}
1516	qdisc_class_hash_destroy(&q->clhash);
1517	qdisc_watchdog_cancel(&q->watchdog);
1518}
1519
1520static int
1521hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1522{
1523	struct hfsc_sched *q = qdisc_priv(sch);
1524	unsigned char *b = skb_tail_pointer(skb);
1525	struct tc_hfsc_qopt qopt;
1526
1527	qopt.defcls = READ_ONCE(q->defcls);
1528	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1529		goto nla_put_failure;
1530	return skb->len;
1531
1532 nla_put_failure:
1533	nlmsg_trim(skb, b);
1534	return -1;
1535}
1536
1537static int
1538hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1539{
1540	unsigned int len = qdisc_pkt_len(skb);
1541	struct hfsc_class *cl;
1542	int err;
1543	bool first;
1544
1545	cl = hfsc_classify(skb, sch, &err);
1546	if (cl == NULL) {
1547		if (err & __NET_XMIT_BYPASS)
1548			qdisc_qstats_drop(sch);
1549		__qdisc_drop(skb, to_free);
1550		return err;
1551	}
1552
1553	first = !cl->qdisc->q.qlen;
1554	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1555	if (unlikely(err != NET_XMIT_SUCCESS)) {
1556		if (net_xmit_drop_count(err)) {
1557			cl->qstats.drops++;
1558			qdisc_qstats_drop(sch);
1559		}
1560		return err;
1561	}
1562
1563	sch->qstats.backlog += len;
1564	sch->q.qlen++;
1565
1566	if (first && !cl_in_el_or_vttree(cl)) {
1567		if (cl->cl_flags & HFSC_RSC)
1568			init_ed(cl, len);
1569		if (cl->cl_flags & HFSC_FSC)
1570			init_vf(cl, len);
1571		/*
1572		 * If this is the first packet, isolate the head so an eventual
1573		 * head drop before the first dequeue operation has no chance
1574		 * to invalidate the deadline.
1575		 */
1576		if (cl->cl_flags & HFSC_RSC)
1577			cl->qdisc->ops->peek(cl->qdisc);
1578
1579	}
1580
1581	return NET_XMIT_SUCCESS;
1582}
1583
1584static struct sk_buff *
1585hfsc_dequeue(struct Qdisc *sch)
1586{
1587	struct hfsc_sched *q = qdisc_priv(sch);
1588	struct hfsc_class *cl;
1589	struct sk_buff *skb;
1590	u64 cur_time;
1591	unsigned int next_len;
1592	int realtime = 0;
1593
1594	if (sch->q.qlen == 0)
1595		return NULL;
1596
1597	cur_time = psched_get_time();
1598
1599	/*
1600	 * if there are eligible classes, use real-time criteria.
1601	 * find the class with the minimum deadline among
1602	 * the eligible classes.
1603	 */
1604	cl = eltree_get_mindl(q, cur_time);
1605	if (cl) {
1606		realtime = 1;
1607	} else {
1608		/*
1609		 * use link-sharing criteria
1610		 * get the class with the minimum vt in the hierarchy
1611		 */
1612		cl = vttree_get_minvt(&q->root, cur_time);
1613		if (cl == NULL) {
1614			qdisc_qstats_overlimit(sch);
1615			hfsc_schedule_watchdog(sch);
1616			return NULL;
1617		}
1618	}
1619
1620	skb = qdisc_dequeue_peeked(cl->qdisc);
1621	if (skb == NULL) {
1622		qdisc_warn_nonwc("HFSC", cl->qdisc);
1623		return NULL;
1624	}
1625
1626	bstats_update(&cl->bstats, skb);
1627	update_vf(cl, qdisc_pkt_len(skb), cur_time);
1628	if (realtime)
1629		cl->cl_cumul += qdisc_pkt_len(skb);
1630
1631	if (cl->cl_flags & HFSC_RSC) {
1632		if (cl->qdisc->q.qlen != 0) {
1633			/* update ed */
1634			next_len = qdisc_peek_len(cl->qdisc);
1635			/* Check queue length again since some qdisc implementations
1636			 * (e.g., netem/codel) might empty the queue during the peek
1637			 * operation.
1638			 */
1639			if (cl->qdisc->q.qlen != 0) {
1640				if (realtime)
1641					update_ed(cl, next_len);
1642				else
1643					update_d(cl, next_len);
1644			}
1645		} else {
1646			/* the class becomes passive */
1647			eltree_remove(cl);
1648		}
1649	}
1650
1651	qdisc_bstats_update(sch, skb);
1652	qdisc_qstats_backlog_dec(sch, skb);
1653	sch->q.qlen--;
1654
1655	return skb;
1656}
1657
1658static const struct Qdisc_class_ops hfsc_class_ops = {
1659	.change		= hfsc_change_class,
1660	.delete		= hfsc_delete_class,
1661	.graft		= hfsc_graft_class,
1662	.leaf		= hfsc_class_leaf,
1663	.qlen_notify	= hfsc_qlen_notify,
1664	.find		= hfsc_search_class,
1665	.bind_tcf	= hfsc_bind_tcf,
1666	.unbind_tcf	= hfsc_unbind_tcf,
1667	.tcf_block	= hfsc_tcf_block,
1668	.dump		= hfsc_dump_class,
1669	.dump_stats	= hfsc_dump_class_stats,
1670	.walk		= hfsc_walk
1671};
1672
1673static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1674	.id		= "hfsc",
1675	.init		= hfsc_init_qdisc,
1676	.change		= hfsc_change_qdisc,
1677	.reset		= hfsc_reset_qdisc,
1678	.destroy	= hfsc_destroy_qdisc,
1679	.dump		= hfsc_dump_qdisc,
1680	.enqueue	= hfsc_enqueue,
1681	.dequeue	= hfsc_dequeue,
1682	.peek		= qdisc_peek_dequeued,
1683	.cl_ops		= &hfsc_class_ops,
1684	.priv_size	= sizeof(struct hfsc_sched),
1685	.owner		= THIS_MODULE
1686};
1687MODULE_ALIAS_NET_SCH("hfsc");
1688
1689static int __init
1690hfsc_init(void)
1691{
1692	return register_qdisc(&hfsc_qdisc_ops);
1693}
1694
1695static void __exit
1696hfsc_cleanup(void)
1697{
1698	unregister_qdisc(&hfsc_qdisc_ops);
1699}
1700
1701MODULE_LICENSE("GPL");
1702MODULE_DESCRIPTION("Hierarchical Fair Service Curve scheduler");
1703module_init(hfsc_init);
1704module_exit(hfsc_cleanup);