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	unsigned int	refcnt;		/* usage count */
114
115	struct gnet_stats_basic_packed bstats;
116	struct gnet_stats_queue qstats;
117	struct gnet_stats_rate_est64 rate_est;
118	unsigned int	level;		/* class level in hierarchy */
119	struct tcf_proto __rcu *filter_list; /* filter list */
120	unsigned int	filter_cnt;	/* filter count */
121
122	struct hfsc_sched *sched;	/* scheduler data */
123	struct hfsc_class *cl_parent;	/* parent class */
124	struct list_head siblings;	/* sibling classes */
125	struct list_head children;	/* child classes */
126	struct Qdisc	*qdisc;		/* leaf qdisc */
127
128	struct rb_node el_node;		/* qdisc's eligible tree member */
129	struct rb_root vt_tree;		/* active children sorted by cl_vt */
130	struct rb_node vt_node;		/* parent's vt_tree member */
131	struct rb_root cf_tree;		/* active children sorted by cl_f */
132	struct rb_node cf_node;		/* parent's cf_heap member */
133	struct list_head dlist;		/* drop list member */
134
135	u64	cl_total;		/* total work in bytes */
136	u64	cl_cumul;		/* cumulative work in bytes done by
137					   real-time criteria */
138
139	u64	cl_d;			/* deadline*/
140	u64	cl_e;			/* eligible time */
141	u64	cl_vt;			/* virtual time */
142	u64	cl_f;			/* time when this class will fit for
143					   link-sharing, max(myf, cfmin) */
144	u64	cl_myf;			/* my fit-time (calculated from this
145					   class's own upperlimit curve) */
146	u64	cl_myfadj;		/* my fit-time adjustment (to cancel
147					   history dependence) */
148	u64	cl_cfmin;		/* earliest children's fit-time (used
149					   with cl_myf to obtain cl_f) */
150	u64	cl_cvtmin;		/* minimal virtual time among the
151					   children fit for link-sharing
152					   (monotonic within a period) */
153	u64	cl_vtadj;		/* intra-period cumulative vt
154					   adjustment */
155	u64	cl_vtoff;		/* inter-period cumulative vt offset */
156	u64	cl_cvtmax;		/* max child's vt in the last period */
157	u64	cl_cvtoff;		/* cumulative cvtmax of all periods */
158	u64	cl_pcvtoff;		/* parent's cvtoff at initialization
159					   time */
160
161	struct internal_sc cl_rsc;	/* internal real-time service curve */
162	struct internal_sc cl_fsc;	/* internal fair service curve */
163	struct internal_sc cl_usc;	/* internal upperlimit service curve */
164	struct runtime_sc cl_deadline;	/* deadline curve */
165	struct runtime_sc cl_eligible;	/* eligible curve */
166	struct runtime_sc cl_virtual;	/* virtual curve */
167	struct runtime_sc cl_ulimit;	/* upperlimit curve */
168
169	unsigned long	cl_flags;	/* which curves are valid */
170	unsigned long	cl_vtperiod;	/* vt period sequence number */
171	unsigned long	cl_parentperiod;/* parent's vt period sequence number*/
172	unsigned long	cl_nactive;	/* number of active children */
173};
174
175struct hfsc_sched {
176	u16	defcls;				/* default class id */
177	struct hfsc_class root;			/* root class */
178	struct Qdisc_class_hash clhash;		/* class hash */
179	struct rb_root eligible;		/* eligible tree */
180	struct list_head droplist;		/* active leaf class list (for
181						   dropping) */
182	struct qdisc_watchdog watchdog;		/* watchdog timer */
183};
184
185#define	HT_INFINITY	0xffffffffffffffffULL	/* infinite time value */
186
187
188/*
189 * eligible tree holds backlogged classes being sorted by their eligible times.
190 * there is one eligible tree per hfsc instance.
191 */
192
193static void
194eltree_insert(struct hfsc_class *cl)
195{
196	struct rb_node **p = &cl->sched->eligible.rb_node;
197	struct rb_node *parent = NULL;
198	struct hfsc_class *cl1;
199
200	while (*p != NULL) {
201		parent = *p;
202		cl1 = rb_entry(parent, struct hfsc_class, el_node);
203		if (cl->cl_e >= cl1->cl_e)
204			p = &parent->rb_right;
205		else
206			p = &parent->rb_left;
207	}
208	rb_link_node(&cl->el_node, parent, p);
209	rb_insert_color(&cl->el_node, &cl->sched->eligible);
210}
211
212static inline void
213eltree_remove(struct hfsc_class *cl)
214{
215	rb_erase(&cl->el_node, &cl->sched->eligible);
216}
217
218static inline void
219eltree_update(struct hfsc_class *cl)
220{
221	eltree_remove(cl);
222	eltree_insert(cl);
223}
224
225/* find the class with the minimum deadline among the eligible classes */
226static inline struct hfsc_class *
227eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
228{
229	struct hfsc_class *p, *cl = NULL;
230	struct rb_node *n;
231
232	for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
233		p = rb_entry(n, struct hfsc_class, el_node);
234		if (p->cl_e > cur_time)
235			break;
236		if (cl == NULL || p->cl_d < cl->cl_d)
237			cl = p;
238	}
239	return cl;
240}
241
242/* find the class with minimum eligible time among the eligible classes */
243static inline struct hfsc_class *
244eltree_get_minel(struct hfsc_sched *q)
245{
246	struct rb_node *n;
247
248	n = rb_first(&q->eligible);
249	if (n == NULL)
250		return NULL;
251	return rb_entry(n, struct hfsc_class, el_node);
252}
253
254/*
255 * vttree holds holds backlogged child classes being sorted by their virtual
256 * time. each intermediate class has one vttree.
257 */
258static void
259vttree_insert(struct hfsc_class *cl)
260{
261	struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
262	struct rb_node *parent = NULL;
263	struct hfsc_class *cl1;
264
265	while (*p != NULL) {
266		parent = *p;
267		cl1 = rb_entry(parent, struct hfsc_class, vt_node);
268		if (cl->cl_vt >= cl1->cl_vt)
269			p = &parent->rb_right;
270		else
271			p = &parent->rb_left;
272	}
273	rb_link_node(&cl->vt_node, parent, p);
274	rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
275}
276
277static inline void
278vttree_remove(struct hfsc_class *cl)
279{
280	rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
281}
282
283static inline void
284vttree_update(struct hfsc_class *cl)
285{
286	vttree_remove(cl);
287	vttree_insert(cl);
288}
289
290static inline struct hfsc_class *
291vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
292{
293	struct hfsc_class *p;
294	struct rb_node *n;
295
296	for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
297		p = rb_entry(n, struct hfsc_class, vt_node);
298		if (p->cl_f <= cur_time)
299			return p;
300	}
301	return NULL;
302}
303
304/*
305 * get the leaf class with the minimum vt in the hierarchy
306 */
307static struct hfsc_class *
308vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
309{
310	/* if root-class's cfmin is bigger than cur_time nothing to do */
311	if (cl->cl_cfmin > cur_time)
312		return NULL;
313
314	while (cl->level > 0) {
315		cl = vttree_firstfit(cl, cur_time);
316		if (cl == NULL)
317			return NULL;
318		/*
319		 * update parent's cl_cvtmin.
320		 */
321		if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
322			cl->cl_parent->cl_cvtmin = cl->cl_vt;
323	}
324	return cl;
325}
326
327static void
328cftree_insert(struct hfsc_class *cl)
329{
330	struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
331	struct rb_node *parent = NULL;
332	struct hfsc_class *cl1;
333
334	while (*p != NULL) {
335		parent = *p;
336		cl1 = rb_entry(parent, struct hfsc_class, cf_node);
337		if (cl->cl_f >= cl1->cl_f)
338			p = &parent->rb_right;
339		else
340			p = &parent->rb_left;
341	}
342	rb_link_node(&cl->cf_node, parent, p);
343	rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
344}
345
346static inline void
347cftree_remove(struct hfsc_class *cl)
348{
349	rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
350}
351
352static inline void
353cftree_update(struct hfsc_class *cl)
354{
355	cftree_remove(cl);
356	cftree_insert(cl);
357}
358
359/*
360 * service curve support functions
361 *
362 *  external service curve parameters
363 *	m: bps
364 *	d: us
365 *  internal service curve parameters
366 *	sm: (bytes/psched_us) << SM_SHIFT
367 *	ism: (psched_us/byte) << ISM_SHIFT
368 *	dx: psched_us
369 *
370 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
371 *
372 * sm and ism are scaled in order to keep effective digits.
373 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
374 * digits in decimal using the following table.
375 *
376 *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
377 *  ------------+-------------------------------------------------------
378 *  bytes/1.024us 12.8e-3    128e-3     1280e-3    12800e-3   128000e-3
379 *
380 *  1.024us/byte  78.125     7.8125     0.78125    0.078125   0.0078125
381 *
382 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
383 */
384#define	SM_SHIFT	(30 - PSCHED_SHIFT)
385#define	ISM_SHIFT	(8 + PSCHED_SHIFT)
386
387#define	SM_MASK		((1ULL << SM_SHIFT) - 1)
388#define	ISM_MASK	((1ULL << ISM_SHIFT) - 1)
389
390static inline u64
391seg_x2y(u64 x, u64 sm)
392{
393	u64 y;
394
395	/*
396	 * compute
397	 *	y = x * sm >> SM_SHIFT
398	 * but divide it for the upper and lower bits to avoid overflow
399	 */
400	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
401	return y;
402}
403
404static inline u64
405seg_y2x(u64 y, u64 ism)
406{
407	u64 x;
408
409	if (y == 0)
410		x = 0;
411	else if (ism == HT_INFINITY)
412		x = HT_INFINITY;
413	else {
414		x = (y >> ISM_SHIFT) * ism
415		    + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
416	}
417	return x;
418}
419
420/* Convert m (bps) into sm (bytes/psched us) */
421static u64
422m2sm(u32 m)
423{
424	u64 sm;
425
426	sm = ((u64)m << SM_SHIFT);
427	sm += PSCHED_TICKS_PER_SEC - 1;
428	do_div(sm, PSCHED_TICKS_PER_SEC);
429	return sm;
430}
431
432/* convert m (bps) into ism (psched us/byte) */
433static u64
434m2ism(u32 m)
435{
436	u64 ism;
437
438	if (m == 0)
439		ism = HT_INFINITY;
440	else {
441		ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
442		ism += m - 1;
443		do_div(ism, m);
444	}
445	return ism;
446}
447
448/* convert d (us) into dx (psched us) */
449static u64
450d2dx(u32 d)
451{
452	u64 dx;
453
454	dx = ((u64)d * PSCHED_TICKS_PER_SEC);
455	dx += USEC_PER_SEC - 1;
456	do_div(dx, USEC_PER_SEC);
457	return dx;
458}
459
460/* convert sm (bytes/psched us) into m (bps) */
461static u32
462sm2m(u64 sm)
463{
464	u64 m;
465
466	m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
467	return (u32)m;
468}
469
470/* convert dx (psched us) into d (us) */
471static u32
472dx2d(u64 dx)
473{
474	u64 d;
475
476	d = dx * USEC_PER_SEC;
477	do_div(d, PSCHED_TICKS_PER_SEC);
478	return (u32)d;
479}
480
481static void
482sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
483{
484	isc->sm1  = m2sm(sc->m1);
485	isc->ism1 = m2ism(sc->m1);
486	isc->dx   = d2dx(sc->d);
487	isc->dy   = seg_x2y(isc->dx, isc->sm1);
488	isc->sm2  = m2sm(sc->m2);
489	isc->ism2 = m2ism(sc->m2);
490}
491
492/*
493 * initialize the runtime service curve with the given internal
494 * service curve starting at (x, y).
495 */
496static void
497rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
498{
499	rtsc->x	   = x;
500	rtsc->y    = y;
501	rtsc->sm1  = isc->sm1;
502	rtsc->ism1 = isc->ism1;
503	rtsc->dx   = isc->dx;
504	rtsc->dy   = isc->dy;
505	rtsc->sm2  = isc->sm2;
506	rtsc->ism2 = isc->ism2;
507}
508
509/*
510 * calculate the y-projection of the runtime service curve by the
511 * given x-projection value
512 */
513static u64
514rtsc_y2x(struct runtime_sc *rtsc, u64 y)
515{
516	u64 x;
517
518	if (y < rtsc->y)
519		x = rtsc->x;
520	else if (y <= rtsc->y + rtsc->dy) {
521		/* x belongs to the 1st segment */
522		if (rtsc->dy == 0)
523			x = rtsc->x + rtsc->dx;
524		else
525			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
526	} else {
527		/* x belongs to the 2nd segment */
528		x = rtsc->x + rtsc->dx
529		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
530	}
531	return x;
532}
533
534static u64
535rtsc_x2y(struct runtime_sc *rtsc, u64 x)
536{
537	u64 y;
538
539	if (x <= rtsc->x)
540		y = rtsc->y;
541	else if (x <= rtsc->x + rtsc->dx)
542		/* y belongs to the 1st segment */
543		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
544	else
545		/* y belongs to the 2nd segment */
546		y = rtsc->y + rtsc->dy
547		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
548	return y;
549}
550
551/*
552 * update the runtime service curve by taking the minimum of the current
553 * runtime service curve and the service curve starting at (x, y).
554 */
555static void
556rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
557{
558	u64 y1, y2, dx, dy;
559	u32 dsm;
560
561	if (isc->sm1 <= isc->sm2) {
562		/* service curve is convex */
563		y1 = rtsc_x2y(rtsc, x);
564		if (y1 < y)
565			/* the current rtsc is smaller */
566			return;
567		rtsc->x = x;
568		rtsc->y = y;
569		return;
570	}
571
572	/*
573	 * service curve is concave
574	 * compute the two y values of the current rtsc
575	 *	y1: at x
576	 *	y2: at (x + dx)
577	 */
578	y1 = rtsc_x2y(rtsc, x);
579	if (y1 <= y) {
580		/* rtsc is below isc, no change to rtsc */
581		return;
582	}
583
584	y2 = rtsc_x2y(rtsc, x + isc->dx);
585	if (y2 >= y + isc->dy) {
586		/* rtsc is above isc, replace rtsc by isc */
587		rtsc->x = x;
588		rtsc->y = y;
589		rtsc->dx = isc->dx;
590		rtsc->dy = isc->dy;
591		return;
592	}
593
594	/*
595	 * the two curves intersect
596	 * compute the offsets (dx, dy) using the reverse
597	 * function of seg_x2y()
598	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
599	 */
600	dx = (y1 - y) << SM_SHIFT;
601	dsm = isc->sm1 - isc->sm2;
602	do_div(dx, dsm);
603	/*
604	 * check if (x, y1) belongs to the 1st segment of rtsc.
605	 * if so, add the offset.
606	 */
607	if (rtsc->x + rtsc->dx > x)
608		dx += rtsc->x + rtsc->dx - x;
609	dy = seg_x2y(dx, isc->sm1);
610
611	rtsc->x = x;
612	rtsc->y = y;
613	rtsc->dx = dx;
614	rtsc->dy = dy;
615}
616
617static void
618init_ed(struct hfsc_class *cl, unsigned int next_len)
619{
620	u64 cur_time = psched_get_time();
621
622	/* update the deadline curve */
623	rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
624
625	/*
626	 * update the eligible curve.
627	 * for concave, it is equal to the deadline curve.
628	 * for convex, it is a linear curve with slope m2.
629	 */
630	cl->cl_eligible = cl->cl_deadline;
631	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
632		cl->cl_eligible.dx = 0;
633		cl->cl_eligible.dy = 0;
634	}
635
636	/* compute e and d */
637	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
638	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
639
640	eltree_insert(cl);
641}
642
643static void
644update_ed(struct hfsc_class *cl, unsigned int next_len)
645{
646	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
647	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
648
649	eltree_update(cl);
650}
651
652static inline void
653update_d(struct hfsc_class *cl, unsigned int next_len)
654{
655	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
656}
657
658static inline void
659update_cfmin(struct hfsc_class *cl)
660{
661	struct rb_node *n = rb_first(&cl->cf_tree);
662	struct hfsc_class *p;
663
664	if (n == NULL) {
665		cl->cl_cfmin = 0;
666		return;
667	}
668	p = rb_entry(n, struct hfsc_class, cf_node);
669	cl->cl_cfmin = p->cl_f;
670}
671
672static void
673init_vf(struct hfsc_class *cl, unsigned int len)
674{
675	struct hfsc_class *max_cl;
676	struct rb_node *n;
677	u64 vt, f, cur_time;
678	int go_active;
679
680	cur_time = 0;
681	go_active = 1;
682	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
683		if (go_active && cl->cl_nactive++ == 0)
684			go_active = 1;
685		else
686			go_active = 0;
687
688		if (go_active) {
689			n = rb_last(&cl->cl_parent->vt_tree);
690			if (n != NULL) {
691				max_cl = rb_entry(n, struct hfsc_class, vt_node);
692				/*
693				 * set vt to the average of the min and max
694				 * classes.  if the parent's period didn't
695				 * change, don't decrease vt of the class.
696				 */
697				vt = max_cl->cl_vt;
698				if (cl->cl_parent->cl_cvtmin != 0)
699					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
700
701				if (cl->cl_parent->cl_vtperiod !=
702				    cl->cl_parentperiod || vt > cl->cl_vt)
703					cl->cl_vt = vt;
704			} else {
705				/*
706				 * first child for a new parent backlog period.
707				 * add parent's cvtmax to cvtoff to make a new
708				 * vt (vtoff + vt) larger than the vt in the
709				 * last period for all children.
710				 */
711				vt = cl->cl_parent->cl_cvtmax;
712				cl->cl_parent->cl_cvtoff += vt;
713				cl->cl_parent->cl_cvtmax = 0;
714				cl->cl_parent->cl_cvtmin = 0;
715				cl->cl_vt = 0;
716			}
717
718			cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
719							cl->cl_pcvtoff;
720
721			/* update the virtual curve */
722			vt = cl->cl_vt + cl->cl_vtoff;
723			rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
724						      cl->cl_total);
725			if (cl->cl_virtual.x == vt) {
726				cl->cl_virtual.x -= cl->cl_vtoff;
727				cl->cl_vtoff = 0;
728			}
729			cl->cl_vtadj = 0;
730
731			cl->cl_vtperiod++;  /* increment vt period */
732			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
733			if (cl->cl_parent->cl_nactive == 0)
734				cl->cl_parentperiod++;
735			cl->cl_f = 0;
736
737			vttree_insert(cl);
738			cftree_insert(cl);
739
740			if (cl->cl_flags & HFSC_USC) {
741				/* class has upper limit curve */
742				if (cur_time == 0)
743					cur_time = psched_get_time();
744
745				/* update the ulimit curve */
746				rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
747					 cl->cl_total);
748				/* compute myf */
749				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
750						      cl->cl_total);
751				cl->cl_myfadj = 0;
752			}
753		}
754
755		f = max(cl->cl_myf, cl->cl_cfmin);
756		if (f != cl->cl_f) {
757			cl->cl_f = f;
758			cftree_update(cl);
759		}
760		update_cfmin(cl->cl_parent);
761	}
762}
763
764static void
765update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
766{
767	u64 f; /* , myf_bound, delta; */
768	int go_passive = 0;
769
770	if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
771		go_passive = 1;
772
773	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
774		cl->cl_total += len;
775
776		if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
777			continue;
778
779		if (go_passive && --cl->cl_nactive == 0)
780			go_passive = 1;
781		else
782			go_passive = 0;
783
784		if (go_passive) {
785			/* no more active child, going passive */
786
787			/* update cvtmax of the parent class */
788			if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
789				cl->cl_parent->cl_cvtmax = cl->cl_vt;
790
791			/* remove this class from the vt tree */
792			vttree_remove(cl);
793
794			cftree_remove(cl);
795			update_cfmin(cl->cl_parent);
796
797			continue;
798		}
799
800		/*
801		 * update vt and f
802		 */
803		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
804			    - cl->cl_vtoff + cl->cl_vtadj;
805
806		/*
807		 * if vt of the class is smaller than cvtmin,
808		 * the class was skipped in the past due to non-fit.
809		 * if so, we need to adjust vtadj.
810		 */
811		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
812			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
813			cl->cl_vt = cl->cl_parent->cl_cvtmin;
814		}
815
816		/* update the vt tree */
817		vttree_update(cl);
818
819		if (cl->cl_flags & HFSC_USC) {
820			cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
821							      cl->cl_total);
822#if 0
823			/*
824			 * This code causes classes to stay way under their
825			 * limit when multiple classes are used at gigabit
826			 * speed. needs investigation. -kaber
827			 */
828			/*
829			 * if myf lags behind by more than one clock tick
830			 * from the current time, adjust myfadj to prevent
831			 * a rate-limited class from going greedy.
832			 * in a steady state under rate-limiting, myf
833			 * fluctuates within one clock tick.
834			 */
835			myf_bound = cur_time - PSCHED_JIFFIE2US(1);
836			if (cl->cl_myf < myf_bound) {
837				delta = cur_time - cl->cl_myf;
838				cl->cl_myfadj += delta;
839				cl->cl_myf += delta;
840			}
841#endif
842		}
843
844		f = max(cl->cl_myf, cl->cl_cfmin);
845		if (f != cl->cl_f) {
846			cl->cl_f = f;
847			cftree_update(cl);
848			update_cfmin(cl->cl_parent);
849		}
850	}
851}
852
853static void
854set_active(struct hfsc_class *cl, unsigned int len)
855{
856	if (cl->cl_flags & HFSC_RSC)
857		init_ed(cl, len);
858	if (cl->cl_flags & HFSC_FSC)
859		init_vf(cl, len);
860
861	list_add_tail(&cl->dlist, &cl->sched->droplist);
862}
863
864static void
865set_passive(struct hfsc_class *cl)
866{
867	if (cl->cl_flags & HFSC_RSC)
868		eltree_remove(cl);
869
870	list_del(&cl->dlist);
871
872	/*
873	 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
874	 * needs to be called explicitly to remove a class from vttree.
875	 */
876}
877
878static unsigned int
879qdisc_peek_len(struct Qdisc *sch)
880{
881	struct sk_buff *skb;
882	unsigned int len;
883
884	skb = sch->ops->peek(sch);
885	if (skb == NULL) {
886		qdisc_warn_nonwc("qdisc_peek_len", sch);
887		return 0;
888	}
889	len = qdisc_pkt_len(skb);
890
891	return len;
892}
893
894static void
895hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
896{
897	unsigned int len = cl->qdisc->q.qlen;
898	unsigned int backlog = cl->qdisc->qstats.backlog;
899
900	qdisc_reset(cl->qdisc);
901	qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
902}
903
904static void
905hfsc_adjust_levels(struct hfsc_class *cl)
906{
907	struct hfsc_class *p;
908	unsigned int level;
909
910	do {
911		level = 0;
912		list_for_each_entry(p, &cl->children, siblings) {
913			if (p->level >= level)
914				level = p->level + 1;
915		}
916		cl->level = level;
917	} while ((cl = cl->cl_parent) != NULL);
918}
919
920static inline struct hfsc_class *
921hfsc_find_class(u32 classid, struct Qdisc *sch)
922{
923	struct hfsc_sched *q = qdisc_priv(sch);
924	struct Qdisc_class_common *clc;
925
926	clc = qdisc_class_find(&q->clhash, classid);
927	if (clc == NULL)
928		return NULL;
929	return container_of(clc, struct hfsc_class, cl_common);
930}
931
932static void
933hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
934		u64 cur_time)
935{
936	sc2isc(rsc, &cl->cl_rsc);
937	rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
938	cl->cl_eligible = cl->cl_deadline;
939	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
940		cl->cl_eligible.dx = 0;
941		cl->cl_eligible.dy = 0;
942	}
943	cl->cl_flags |= HFSC_RSC;
944}
945
946static void
947hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
948{
949	sc2isc(fsc, &cl->cl_fsc);
950	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
951	cl->cl_flags |= HFSC_FSC;
952}
953
954static void
955hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
956		u64 cur_time)
957{
958	sc2isc(usc, &cl->cl_usc);
959	rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
960	cl->cl_flags |= HFSC_USC;
961}
962
963static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
964	[TCA_HFSC_RSC]	= { .len = sizeof(struct tc_service_curve) },
965	[TCA_HFSC_FSC]	= { .len = sizeof(struct tc_service_curve) },
966	[TCA_HFSC_USC]	= { .len = sizeof(struct tc_service_curve) },
967};
968
969static int
970hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
971		  struct nlattr **tca, unsigned long *arg)
972{
973	struct hfsc_sched *q = qdisc_priv(sch);
974	struct hfsc_class *cl = (struct hfsc_class *)*arg;
975	struct hfsc_class *parent = NULL;
976	struct nlattr *opt = tca[TCA_OPTIONS];
977	struct nlattr *tb[TCA_HFSC_MAX + 1];
978	struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
979	u64 cur_time;
980	int err;
981
982	if (opt == NULL)
983		return -EINVAL;
984
985	err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
986	if (err < 0)
987		return err;
988
989	if (tb[TCA_HFSC_RSC]) {
990		rsc = nla_data(tb[TCA_HFSC_RSC]);
991		if (rsc->m1 == 0 && rsc->m2 == 0)
992			rsc = NULL;
993	}
994
995	if (tb[TCA_HFSC_FSC]) {
996		fsc = nla_data(tb[TCA_HFSC_FSC]);
997		if (fsc->m1 == 0 && fsc->m2 == 0)
998			fsc = NULL;
999	}
1000
1001	if (tb[TCA_HFSC_USC]) {
1002		usc = nla_data(tb[TCA_HFSC_USC]);
1003		if (usc->m1 == 0 && usc->m2 == 0)
1004			usc = NULL;
1005	}
1006
1007	if (cl != NULL) {
1008		if (parentid) {
1009			if (cl->cl_parent &&
1010			    cl->cl_parent->cl_common.classid != parentid)
1011				return -EINVAL;
1012			if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1013				return -EINVAL;
1014		}
1015		cur_time = psched_get_time();
1016
1017		if (tca[TCA_RATE]) {
1018			spinlock_t *lock = qdisc_root_sleeping_lock(sch);
1019
1020			err = gen_replace_estimator(&cl->bstats, NULL,
1021						    &cl->rate_est,
1022						    lock,
1023						    tca[TCA_RATE]);
1024			if (err)
1025				return err;
1026		}
1027
1028		sch_tree_lock(sch);
1029		if (rsc != NULL)
1030			hfsc_change_rsc(cl, rsc, cur_time);
1031		if (fsc != NULL)
1032			hfsc_change_fsc(cl, fsc);
1033		if (usc != NULL)
1034			hfsc_change_usc(cl, usc, cur_time);
1035
1036		if (cl->qdisc->q.qlen != 0) {
1037			if (cl->cl_flags & HFSC_RSC)
1038				update_ed(cl, qdisc_peek_len(cl->qdisc));
1039			if (cl->cl_flags & HFSC_FSC)
1040				update_vf(cl, 0, cur_time);
1041		}
1042		sch_tree_unlock(sch);
1043
1044		return 0;
1045	}
1046
1047	if (parentid == TC_H_ROOT)
1048		return -EEXIST;
1049
1050	parent = &q->root;
1051	if (parentid) {
1052		parent = hfsc_find_class(parentid, sch);
1053		if (parent == NULL)
1054			return -ENOENT;
1055	}
1056
1057	if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1058		return -EINVAL;
1059	if (hfsc_find_class(classid, sch))
1060		return -EEXIST;
1061
1062	if (rsc == NULL && fsc == NULL)
1063		return -EINVAL;
1064
1065	cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1066	if (cl == NULL)
1067		return -ENOBUFS;
1068
1069	if (tca[TCA_RATE]) {
1070		err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1071					qdisc_root_sleeping_lock(sch),
1072					tca[TCA_RATE]);
1073		if (err) {
1074			kfree(cl);
1075			return err;
1076		}
1077	}
1078
1079	if (rsc != NULL)
1080		hfsc_change_rsc(cl, rsc, 0);
1081	if (fsc != NULL)
1082		hfsc_change_fsc(cl, fsc);
1083	if (usc != NULL)
1084		hfsc_change_usc(cl, usc, 0);
1085
1086	cl->cl_common.classid = classid;
1087	cl->refcnt    = 1;
1088	cl->sched     = q;
1089	cl->cl_parent = parent;
1090	cl->qdisc = qdisc_create_dflt(sch->dev_queue,
1091				      &pfifo_qdisc_ops, classid);
1092	if (cl->qdisc == NULL)
1093		cl->qdisc = &noop_qdisc;
1094	INIT_LIST_HEAD(&cl->children);
1095	cl->vt_tree = RB_ROOT;
1096	cl->cf_tree = RB_ROOT;
1097
1098	sch_tree_lock(sch);
1099	qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1100	list_add_tail(&cl->siblings, &parent->children);
1101	if (parent->level == 0)
1102		hfsc_purge_queue(sch, parent);
1103	hfsc_adjust_levels(parent);
1104	cl->cl_pcvtoff = parent->cl_cvtoff;
1105	sch_tree_unlock(sch);
1106
1107	qdisc_class_hash_grow(sch, &q->clhash);
1108
1109	*arg = (unsigned long)cl;
1110	return 0;
1111}
1112
1113static void
1114hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1115{
1116	struct hfsc_sched *q = qdisc_priv(sch);
1117
1118	tcf_destroy_chain(&cl->filter_list);
1119	qdisc_destroy(cl->qdisc);
1120	gen_kill_estimator(&cl->bstats, &cl->rate_est);
1121	if (cl != &q->root)
1122		kfree(cl);
1123}
1124
1125static int
1126hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1127{
1128	struct hfsc_sched *q = qdisc_priv(sch);
1129	struct hfsc_class *cl = (struct hfsc_class *)arg;
1130
1131	if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1132		return -EBUSY;
1133
1134	sch_tree_lock(sch);
1135
1136	list_del(&cl->siblings);
1137	hfsc_adjust_levels(cl->cl_parent);
1138
1139	hfsc_purge_queue(sch, cl);
1140	qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1141
1142	BUG_ON(--cl->refcnt == 0);
1143	/*
1144	 * This shouldn't happen: we "hold" one cops->get() when called
1145	 * from tc_ctl_tclass; the destroy method is done from cops->put().
1146	 */
1147
1148	sch_tree_unlock(sch);
1149	return 0;
1150}
1151
1152static struct hfsc_class *
1153hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1154{
1155	struct hfsc_sched *q = qdisc_priv(sch);
1156	struct hfsc_class *head, *cl;
1157	struct tcf_result res;
1158	struct tcf_proto *tcf;
1159	int result;
1160
1161	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1162	    (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1163		if (cl->level == 0)
1164			return cl;
1165
1166	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1167	head = &q->root;
1168	tcf = rcu_dereference_bh(q->root.filter_list);
1169	while (tcf && (result = tc_classify(skb, tcf, &res, false)) >= 0) {
1170#ifdef CONFIG_NET_CLS_ACT
1171		switch (result) {
1172		case TC_ACT_QUEUED:
1173		case TC_ACT_STOLEN:
1174			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1175		case TC_ACT_SHOT:
1176			return NULL;
1177		}
1178#endif
1179		cl = (struct hfsc_class *)res.class;
1180		if (!cl) {
1181			cl = hfsc_find_class(res.classid, sch);
1182			if (!cl)
1183				break; /* filter selected invalid classid */
1184			if (cl->level >= head->level)
1185				break; /* filter may only point downwards */
1186		}
1187
1188		if (cl->level == 0)
1189			return cl; /* hit leaf class */
1190
1191		/* apply inner filter chain */
1192		tcf = rcu_dereference_bh(cl->filter_list);
1193		head = cl;
1194	}
1195
1196	/* classification failed, try default class */
1197	cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1198	if (cl == NULL || cl->level > 0)
1199		return NULL;
1200
1201	return cl;
1202}
1203
1204static int
1205hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1206		 struct Qdisc **old)
1207{
1208	struct hfsc_class *cl = (struct hfsc_class *)arg;
1209
1210	if (cl->level > 0)
1211		return -EINVAL;
1212	if (new == NULL) {
1213		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1214					cl->cl_common.classid);
1215		if (new == NULL)
1216			new = &noop_qdisc;
1217	}
1218
1219	*old = qdisc_replace(sch, new, &cl->qdisc);
1220	return 0;
1221}
1222
1223static struct Qdisc *
1224hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1225{
1226	struct hfsc_class *cl = (struct hfsc_class *)arg;
1227
1228	if (cl->level == 0)
1229		return cl->qdisc;
1230
1231	return NULL;
1232}
1233
1234static void
1235hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1236{
1237	struct hfsc_class *cl = (struct hfsc_class *)arg;
1238
1239	if (cl->qdisc->q.qlen == 0) {
1240		update_vf(cl, 0, 0);
1241		set_passive(cl);
1242	}
1243}
1244
1245static unsigned long
1246hfsc_get_class(struct Qdisc *sch, u32 classid)
1247{
1248	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1249
1250	if (cl != NULL)
1251		cl->refcnt++;
1252
1253	return (unsigned long)cl;
1254}
1255
1256static void
1257hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1258{
1259	struct hfsc_class *cl = (struct hfsc_class *)arg;
1260
1261	if (--cl->refcnt == 0)
1262		hfsc_destroy_class(sch, cl);
1263}
1264
1265static unsigned long
1266hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1267{
1268	struct hfsc_class *p = (struct hfsc_class *)parent;
1269	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1270
1271	if (cl != NULL) {
1272		if (p != NULL && p->level <= cl->level)
1273			return 0;
1274		cl->filter_cnt++;
1275	}
1276
1277	return (unsigned long)cl;
1278}
1279
1280static void
1281hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1282{
1283	struct hfsc_class *cl = (struct hfsc_class *)arg;
1284
1285	cl->filter_cnt--;
1286}
1287
1288static struct tcf_proto __rcu **
1289hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1290{
1291	struct hfsc_sched *q = qdisc_priv(sch);
1292	struct hfsc_class *cl = (struct hfsc_class *)arg;
1293
1294	if (cl == NULL)
1295		cl = &q->root;
1296
1297	return &cl->filter_list;
1298}
1299
1300static int
1301hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1302{
1303	struct tc_service_curve tsc;
1304
1305	tsc.m1 = sm2m(sc->sm1);
1306	tsc.d  = dx2d(sc->dx);
1307	tsc.m2 = sm2m(sc->sm2);
1308	if (nla_put(skb, attr, sizeof(tsc), &tsc))
1309		goto nla_put_failure;
1310
1311	return skb->len;
1312
1313 nla_put_failure:
1314	return -1;
1315}
1316
1317static int
1318hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1319{
1320	if ((cl->cl_flags & HFSC_RSC) &&
1321	    (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1322		goto nla_put_failure;
1323
1324	if ((cl->cl_flags & HFSC_FSC) &&
1325	    (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1326		goto nla_put_failure;
1327
1328	if ((cl->cl_flags & HFSC_USC) &&
1329	    (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1330		goto nla_put_failure;
1331
1332	return skb->len;
1333
1334 nla_put_failure:
1335	return -1;
1336}
1337
1338static int
1339hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1340		struct tcmsg *tcm)
1341{
1342	struct hfsc_class *cl = (struct hfsc_class *)arg;
1343	struct nlattr *nest;
1344
1345	tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1346					  TC_H_ROOT;
1347	tcm->tcm_handle = cl->cl_common.classid;
1348	if (cl->level == 0)
1349		tcm->tcm_info = cl->qdisc->handle;
1350
1351	nest = nla_nest_start(skb, TCA_OPTIONS);
1352	if (nest == NULL)
1353		goto nla_put_failure;
1354	if (hfsc_dump_curves(skb, cl) < 0)
1355		goto nla_put_failure;
1356	return nla_nest_end(skb, nest);
1357
1358 nla_put_failure:
1359	nla_nest_cancel(skb, nest);
1360	return -EMSGSIZE;
1361}
1362
1363static int
1364hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1365	struct gnet_dump *d)
1366{
1367	struct hfsc_class *cl = (struct hfsc_class *)arg;
1368	struct tc_hfsc_stats xstats;
1369
1370	cl->qstats.backlog = cl->qdisc->qstats.backlog;
1371	xstats.level   = cl->level;
1372	xstats.period  = cl->cl_vtperiod;
1373	xstats.work    = cl->cl_total;
1374	xstats.rtwork  = cl->cl_cumul;
1375
1376	if (gnet_stats_copy_basic(d, NULL, &cl->bstats) < 0 ||
1377	    gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
1378	    gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->qdisc->q.qlen) < 0)
1379		return -1;
1380
1381	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1382}
1383
1384
1385
1386static void
1387hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1388{
1389	struct hfsc_sched *q = qdisc_priv(sch);
1390	struct hfsc_class *cl;
1391	unsigned int i;
1392
1393	if (arg->stop)
1394		return;
1395
1396	for (i = 0; i < q->clhash.hashsize; i++) {
1397		hlist_for_each_entry(cl, &q->clhash.hash[i],
1398				     cl_common.hnode) {
1399			if (arg->count < arg->skip) {
1400				arg->count++;
1401				continue;
1402			}
1403			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1404				arg->stop = 1;
1405				return;
1406			}
1407			arg->count++;
1408		}
1409	}
1410}
1411
1412static void
1413hfsc_schedule_watchdog(struct Qdisc *sch)
1414{
1415	struct hfsc_sched *q = qdisc_priv(sch);
1416	struct hfsc_class *cl;
1417	u64 next_time = 0;
1418
1419	cl = eltree_get_minel(q);
1420	if (cl)
1421		next_time = cl->cl_e;
1422	if (q->root.cl_cfmin != 0) {
1423		if (next_time == 0 || next_time > q->root.cl_cfmin)
1424			next_time = q->root.cl_cfmin;
1425	}
1426	WARN_ON(next_time == 0);
1427	qdisc_watchdog_schedule(&q->watchdog, next_time);
1428}
1429
1430static int
1431hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1432{
1433	struct hfsc_sched *q = qdisc_priv(sch);
1434	struct tc_hfsc_qopt *qopt;
1435	int err;
1436
1437	if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1438		return -EINVAL;
1439	qopt = nla_data(opt);
1440
1441	q->defcls = qopt->defcls;
1442	err = qdisc_class_hash_init(&q->clhash);
1443	if (err < 0)
1444		return err;
1445	q->eligible = RB_ROOT;
1446	INIT_LIST_HEAD(&q->droplist);
1447
1448	q->root.cl_common.classid = sch->handle;
1449	q->root.refcnt  = 1;
1450	q->root.sched   = q;
1451	q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1452					  sch->handle);
1453	if (q->root.qdisc == NULL)
1454		q->root.qdisc = &noop_qdisc;
1455	INIT_LIST_HEAD(&q->root.children);
1456	q->root.vt_tree = RB_ROOT;
1457	q->root.cf_tree = RB_ROOT;
1458
1459	qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1460	qdisc_class_hash_grow(sch, &q->clhash);
1461
1462	qdisc_watchdog_init(&q->watchdog, sch);
1463
1464	return 0;
1465}
1466
1467static int
1468hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1469{
1470	struct hfsc_sched *q = qdisc_priv(sch);
1471	struct tc_hfsc_qopt *qopt;
1472
1473	if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1474		return -EINVAL;
1475	qopt = nla_data(opt);
1476
1477	sch_tree_lock(sch);
1478	q->defcls = qopt->defcls;
1479	sch_tree_unlock(sch);
1480
1481	return 0;
1482}
1483
1484static void
1485hfsc_reset_class(struct hfsc_class *cl)
1486{
1487	cl->cl_total        = 0;
1488	cl->cl_cumul        = 0;
1489	cl->cl_d            = 0;
1490	cl->cl_e            = 0;
1491	cl->cl_vt           = 0;
1492	cl->cl_vtadj        = 0;
1493	cl->cl_vtoff        = 0;
1494	cl->cl_cvtmin       = 0;
1495	cl->cl_cvtmax       = 0;
1496	cl->cl_cvtoff       = 0;
1497	cl->cl_pcvtoff      = 0;
1498	cl->cl_vtperiod     = 0;
1499	cl->cl_parentperiod = 0;
1500	cl->cl_f            = 0;
1501	cl->cl_myf          = 0;
1502	cl->cl_myfadj       = 0;
1503	cl->cl_cfmin        = 0;
1504	cl->cl_nactive      = 0;
1505
1506	cl->vt_tree = RB_ROOT;
1507	cl->cf_tree = RB_ROOT;
1508	qdisc_reset(cl->qdisc);
1509
1510	if (cl->cl_flags & HFSC_RSC)
1511		rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1512	if (cl->cl_flags & HFSC_FSC)
1513		rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1514	if (cl->cl_flags & HFSC_USC)
1515		rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1516}
1517
1518static void
1519hfsc_reset_qdisc(struct Qdisc *sch)
1520{
1521	struct hfsc_sched *q = qdisc_priv(sch);
1522	struct hfsc_class *cl;
1523	unsigned int i;
1524
1525	for (i = 0; i < q->clhash.hashsize; i++) {
1526		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1527			hfsc_reset_class(cl);
1528	}
1529	q->eligible = RB_ROOT;
1530	INIT_LIST_HEAD(&q->droplist);
1531	qdisc_watchdog_cancel(&q->watchdog);
1532	sch->q.qlen = 0;
1533}
1534
1535static void
1536hfsc_destroy_qdisc(struct Qdisc *sch)
1537{
1538	struct hfsc_sched *q = qdisc_priv(sch);
1539	struct hlist_node *next;
1540	struct hfsc_class *cl;
1541	unsigned int i;
1542
1543	for (i = 0; i < q->clhash.hashsize; i++) {
1544		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1545			tcf_destroy_chain(&cl->filter_list);
1546	}
1547	for (i = 0; i < q->clhash.hashsize; i++) {
1548		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1549					  cl_common.hnode)
1550			hfsc_destroy_class(sch, cl);
1551	}
1552	qdisc_class_hash_destroy(&q->clhash);
1553	qdisc_watchdog_cancel(&q->watchdog);
1554}
1555
1556static int
1557hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1558{
1559	struct hfsc_sched *q = qdisc_priv(sch);
1560	unsigned char *b = skb_tail_pointer(skb);
1561	struct tc_hfsc_qopt qopt;
1562	struct hfsc_class *cl;
1563	unsigned int i;
1564
1565	sch->qstats.backlog = 0;
1566	for (i = 0; i < q->clhash.hashsize; i++) {
1567		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1568			sch->qstats.backlog += cl->qdisc->qstats.backlog;
1569	}
1570
1571	qopt.defcls = q->defcls;
1572	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1573		goto nla_put_failure;
1574	return skb->len;
1575
1576 nla_put_failure:
1577	nlmsg_trim(skb, b);
1578	return -1;
1579}
1580
1581static int
1582hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1583{
1584	struct hfsc_class *cl;
1585	int uninitialized_var(err);
1586
1587	cl = hfsc_classify(skb, sch, &err);
1588	if (cl == NULL) {
1589		if (err & __NET_XMIT_BYPASS)
1590			qdisc_qstats_drop(sch);
1591		kfree_skb(skb);
1592		return err;
1593	}
1594
1595	err = qdisc_enqueue(skb, cl->qdisc);
1596	if (unlikely(err != NET_XMIT_SUCCESS)) {
1597		if (net_xmit_drop_count(err)) {
1598			cl->qstats.drops++;
1599			qdisc_qstats_drop(sch);
1600		}
1601		return err;
1602	}
1603
1604	if (cl->qdisc->q.qlen == 1)
1605		set_active(cl, qdisc_pkt_len(skb));
1606
1607	sch->q.qlen++;
1608
1609	return NET_XMIT_SUCCESS;
1610}
1611
1612static struct sk_buff *
1613hfsc_dequeue(struct Qdisc *sch)
1614{
1615	struct hfsc_sched *q = qdisc_priv(sch);
1616	struct hfsc_class *cl;
1617	struct sk_buff *skb;
1618	u64 cur_time;
1619	unsigned int next_len;
1620	int realtime = 0;
1621
1622	if (sch->q.qlen == 0)
1623		return NULL;
1624
1625	cur_time = psched_get_time();
1626
1627	/*
1628	 * if there are eligible classes, use real-time criteria.
1629	 * find the class with the minimum deadline among
1630	 * the eligible classes.
1631	 */
1632	cl = eltree_get_mindl(q, cur_time);
1633	if (cl) {
1634		realtime = 1;
1635	} else {
1636		/*
1637		 * use link-sharing criteria
1638		 * get the class with the minimum vt in the hierarchy
1639		 */
1640		cl = vttree_get_minvt(&q->root, cur_time);
1641		if (cl == NULL) {
1642			qdisc_qstats_overlimit(sch);
1643			hfsc_schedule_watchdog(sch);
1644			return NULL;
1645		}
1646	}
1647
1648	skb = qdisc_dequeue_peeked(cl->qdisc);
1649	if (skb == NULL) {
1650		qdisc_warn_nonwc("HFSC", cl->qdisc);
1651		return NULL;
1652	}
1653
1654	bstats_update(&cl->bstats, skb);
1655	update_vf(cl, qdisc_pkt_len(skb), cur_time);
1656	if (realtime)
1657		cl->cl_cumul += qdisc_pkt_len(skb);
1658
1659	if (cl->qdisc->q.qlen != 0) {
1660		if (cl->cl_flags & HFSC_RSC) {
1661			/* update ed */
1662			next_len = qdisc_peek_len(cl->qdisc);
1663			if (realtime)
1664				update_ed(cl, next_len);
1665			else
1666				update_d(cl, next_len);
1667		}
1668	} else {
1669		/* the class becomes passive */
1670		set_passive(cl);
1671	}
1672
1673	qdisc_unthrottled(sch);
1674	qdisc_bstats_update(sch, skb);
1675	sch->q.qlen--;
1676
1677	return skb;
1678}
1679
1680static unsigned int
1681hfsc_drop(struct Qdisc *sch)
1682{
1683	struct hfsc_sched *q = qdisc_priv(sch);
1684	struct hfsc_class *cl;
1685	unsigned int len;
1686
1687	list_for_each_entry(cl, &q->droplist, dlist) {
1688		if (cl->qdisc->ops->drop != NULL &&
1689		    (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1690			if (cl->qdisc->q.qlen == 0) {
1691				update_vf(cl, 0, 0);
1692				set_passive(cl);
1693			} else {
1694				list_move_tail(&cl->dlist, &q->droplist);
1695			}
1696			cl->qstats.drops++;
1697			qdisc_qstats_drop(sch);
1698			sch->q.qlen--;
1699			return len;
1700		}
1701	}
1702	return 0;
1703}
1704
1705static const struct Qdisc_class_ops hfsc_class_ops = {
1706	.change		= hfsc_change_class,
1707	.delete		= hfsc_delete_class,
1708	.graft		= hfsc_graft_class,
1709	.leaf		= hfsc_class_leaf,
1710	.qlen_notify	= hfsc_qlen_notify,
1711	.get		= hfsc_get_class,
1712	.put		= hfsc_put_class,
1713	.bind_tcf	= hfsc_bind_tcf,
1714	.unbind_tcf	= hfsc_unbind_tcf,
1715	.tcf_chain	= hfsc_tcf_chain,
1716	.dump		= hfsc_dump_class,
1717	.dump_stats	= hfsc_dump_class_stats,
1718	.walk		= hfsc_walk
1719};
1720
1721static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1722	.id		= "hfsc",
1723	.init		= hfsc_init_qdisc,
1724	.change		= hfsc_change_qdisc,
1725	.reset		= hfsc_reset_qdisc,
1726	.destroy	= hfsc_destroy_qdisc,
1727	.dump		= hfsc_dump_qdisc,
1728	.enqueue	= hfsc_enqueue,
1729	.dequeue	= hfsc_dequeue,
1730	.peek		= qdisc_peek_dequeued,
1731	.drop		= hfsc_drop,
1732	.cl_ops		= &hfsc_class_ops,
1733	.priv_size	= sizeof(struct hfsc_sched),
1734	.owner		= THIS_MODULE
1735};
1736
1737static int __init
1738hfsc_init(void)
1739{
1740	return register_qdisc(&hfsc_qdisc_ops);
1741}
1742
1743static void __exit
1744hfsc_cleanup(void)
1745{
1746	unregister_qdisc(&hfsc_qdisc_ops);
1747}
1748
1749MODULE_LICENSE("GPL");
1750module_init(hfsc_init);
1751module_exit(hfsc_cleanup);
1752