xref: /openbmc/linux/net/sched/sch_hfsc.c (revision a1e58bbd)
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 to avoid 64-bit divide operations
81  *   that are expensive on 32-bit architectures.
82  */
83 
84 struct internal_sc
85 {
86 	u64	sm1;	/* scaled slope of the 1st segment */
87 	u64	ism1;	/* scaled inverse-slope of the 1st segment */
88 	u64	dx;	/* the x-projection of the 1st segment */
89 	u64	dy;	/* the y-projection of the 1st segment */
90 	u64	sm2;	/* scaled slope of the 2nd segment */
91 	u64	ism2;	/* scaled inverse-slope of the 2nd segment */
92 };
93 
94 /* runtime service curve */
95 struct runtime_sc
96 {
97 	u64	x;	/* current starting position on x-axis */
98 	u64	y;	/* current starting position on y-axis */
99 	u64	sm1;	/* scaled slope of the 1st segment */
100 	u64	ism1;	/* scaled inverse-slope of the 1st segment */
101 	u64	dx;	/* the x-projection of the 1st segment */
102 	u64	dy;	/* the y-projection of the 1st segment */
103 	u64	sm2;	/* scaled slope of the 2nd segment */
104 	u64	ism2;	/* scaled inverse-slope of the 2nd segment */
105 };
106 
107 enum hfsc_class_flags
108 {
109 	HFSC_RSC = 0x1,
110 	HFSC_FSC = 0x2,
111 	HFSC_USC = 0x4
112 };
113 
114 struct hfsc_class
115 {
116 	u32		classid;	/* class id */
117 	unsigned int	refcnt;		/* usage count */
118 
119 	struct gnet_stats_basic bstats;
120 	struct gnet_stats_queue qstats;
121 	struct gnet_stats_rate_est rate_est;
122 	unsigned int	level;		/* class level in hierarchy */
123 	struct tcf_proto *filter_list;	/* filter list */
124 	unsigned int	filter_cnt;	/* filter count */
125 
126 	struct hfsc_sched *sched;	/* scheduler data */
127 	struct hfsc_class *cl_parent;	/* parent class */
128 	struct list_head siblings;	/* sibling classes */
129 	struct list_head children;	/* child classes */
130 	struct Qdisc	*qdisc;		/* leaf qdisc */
131 
132 	struct rb_node el_node;		/* qdisc's eligible tree member */
133 	struct rb_root vt_tree;		/* active children sorted by cl_vt */
134 	struct rb_node vt_node;		/* parent's vt_tree member */
135 	struct rb_root cf_tree;		/* active children sorted by cl_f */
136 	struct rb_node cf_node;		/* parent's cf_heap member */
137 	struct list_head hlist;		/* hash list member */
138 	struct list_head dlist;		/* drop list member */
139 
140 	u64	cl_total;		/* total work in bytes */
141 	u64	cl_cumul;		/* cumulative work in bytes done by
142 					   real-time criteria */
143 
144 	u64 	cl_d;			/* deadline*/
145 	u64 	cl_e;			/* eligible time */
146 	u64	cl_vt;			/* virtual time */
147 	u64	cl_f;			/* time when this class will fit for
148 					   link-sharing, max(myf, cfmin) */
149 	u64	cl_myf;			/* my fit-time (calculated from this
150 					   class's own upperlimit curve) */
151 	u64	cl_myfadj;		/* my fit-time adjustment (to cancel
152 					   history dependence) */
153 	u64	cl_cfmin;		/* earliest children's fit-time (used
154 					   with cl_myf to obtain cl_f) */
155 	u64	cl_cvtmin;		/* minimal virtual time among the
156 					   children fit for link-sharing
157 					   (monotonic within a period) */
158 	u64	cl_vtadj;		/* intra-period cumulative vt
159 					   adjustment */
160 	u64	cl_vtoff;		/* inter-period cumulative vt offset */
161 	u64	cl_cvtmax;		/* max child's vt in the last period */
162 	u64	cl_cvtoff;		/* cumulative cvtmax of all periods */
163 	u64	cl_pcvtoff;		/* parent's cvtoff at initialization
164 					   time */
165 
166 	struct internal_sc cl_rsc;	/* internal real-time service curve */
167 	struct internal_sc cl_fsc;	/* internal fair service curve */
168 	struct internal_sc cl_usc;	/* internal upperlimit service curve */
169 	struct runtime_sc cl_deadline;	/* deadline curve */
170 	struct runtime_sc cl_eligible;	/* eligible curve */
171 	struct runtime_sc cl_virtual;	/* virtual curve */
172 	struct runtime_sc cl_ulimit;	/* upperlimit curve */
173 
174 	unsigned long	cl_flags;	/* which curves are valid */
175 	unsigned long	cl_vtperiod;	/* vt period sequence number */
176 	unsigned long	cl_parentperiod;/* parent's vt period sequence number*/
177 	unsigned long	cl_nactive;	/* number of active children */
178 };
179 
180 #define HFSC_HSIZE	16
181 
182 struct hfsc_sched
183 {
184 	u16	defcls;				/* default class id */
185 	struct hfsc_class root;			/* root class */
186 	struct list_head clhash[HFSC_HSIZE];	/* class hash */
187 	struct rb_root eligible;		/* eligible tree */
188 	struct list_head droplist;		/* active leaf class list (for
189 						   dropping) */
190 	struct sk_buff_head requeue;		/* requeued packet */
191 	struct qdisc_watchdog watchdog;		/* watchdog timer */
192 };
193 
194 #define	HT_INFINITY	0xffffffffffffffffULL	/* infinite time value */
195 
196 
197 /*
198  * eligible tree holds backlogged classes being sorted by their eligible times.
199  * there is one eligible tree per hfsc instance.
200  */
201 
202 static void
203 eltree_insert(struct hfsc_class *cl)
204 {
205 	struct rb_node **p = &cl->sched->eligible.rb_node;
206 	struct rb_node *parent = NULL;
207 	struct hfsc_class *cl1;
208 
209 	while (*p != NULL) {
210 		parent = *p;
211 		cl1 = rb_entry(parent, struct hfsc_class, el_node);
212 		if (cl->cl_e >= cl1->cl_e)
213 			p = &parent->rb_right;
214 		else
215 			p = &parent->rb_left;
216 	}
217 	rb_link_node(&cl->el_node, parent, p);
218 	rb_insert_color(&cl->el_node, &cl->sched->eligible);
219 }
220 
221 static inline void
222 eltree_remove(struct hfsc_class *cl)
223 {
224 	rb_erase(&cl->el_node, &cl->sched->eligible);
225 }
226 
227 static inline void
228 eltree_update(struct hfsc_class *cl)
229 {
230 	eltree_remove(cl);
231 	eltree_insert(cl);
232 }
233 
234 /* find the class with the minimum deadline among the eligible classes */
235 static inline struct hfsc_class *
236 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
237 {
238 	struct hfsc_class *p, *cl = NULL;
239 	struct rb_node *n;
240 
241 	for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
242 		p = rb_entry(n, struct hfsc_class, el_node);
243 		if (p->cl_e > cur_time)
244 			break;
245 		if (cl == NULL || p->cl_d < cl->cl_d)
246 			cl = p;
247 	}
248 	return cl;
249 }
250 
251 /* find the class with minimum eligible time among the eligible classes */
252 static inline struct hfsc_class *
253 eltree_get_minel(struct hfsc_sched *q)
254 {
255 	struct rb_node *n;
256 
257 	n = rb_first(&q->eligible);
258 	if (n == NULL)
259 		return NULL;
260 	return rb_entry(n, struct hfsc_class, el_node);
261 }
262 
263 /*
264  * vttree holds holds backlogged child classes being sorted by their virtual
265  * time. each intermediate class has one vttree.
266  */
267 static void
268 vttree_insert(struct hfsc_class *cl)
269 {
270 	struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
271 	struct rb_node *parent = NULL;
272 	struct hfsc_class *cl1;
273 
274 	while (*p != NULL) {
275 		parent = *p;
276 		cl1 = rb_entry(parent, struct hfsc_class, vt_node);
277 		if (cl->cl_vt >= cl1->cl_vt)
278 			p = &parent->rb_right;
279 		else
280 			p = &parent->rb_left;
281 	}
282 	rb_link_node(&cl->vt_node, parent, p);
283 	rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
284 }
285 
286 static inline void
287 vttree_remove(struct hfsc_class *cl)
288 {
289 	rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
290 }
291 
292 static inline void
293 vttree_update(struct hfsc_class *cl)
294 {
295 	vttree_remove(cl);
296 	vttree_insert(cl);
297 }
298 
299 static inline struct hfsc_class *
300 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
301 {
302 	struct hfsc_class *p;
303 	struct rb_node *n;
304 
305 	for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
306 		p = rb_entry(n, struct hfsc_class, vt_node);
307 		if (p->cl_f <= cur_time)
308 			return p;
309 	}
310 	return NULL;
311 }
312 
313 /*
314  * get the leaf class with the minimum vt in the hierarchy
315  */
316 static struct hfsc_class *
317 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
318 {
319 	/* if root-class's cfmin is bigger than cur_time nothing to do */
320 	if (cl->cl_cfmin > cur_time)
321 		return NULL;
322 
323 	while (cl->level > 0) {
324 		cl = vttree_firstfit(cl, cur_time);
325 		if (cl == NULL)
326 			return NULL;
327 		/*
328 		 * update parent's cl_cvtmin.
329 		 */
330 		if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
331 			cl->cl_parent->cl_cvtmin = cl->cl_vt;
332 	}
333 	return cl;
334 }
335 
336 static void
337 cftree_insert(struct hfsc_class *cl)
338 {
339 	struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
340 	struct rb_node *parent = NULL;
341 	struct hfsc_class *cl1;
342 
343 	while (*p != NULL) {
344 		parent = *p;
345 		cl1 = rb_entry(parent, struct hfsc_class, cf_node);
346 		if (cl->cl_f >= cl1->cl_f)
347 			p = &parent->rb_right;
348 		else
349 			p = &parent->rb_left;
350 	}
351 	rb_link_node(&cl->cf_node, parent, p);
352 	rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
353 }
354 
355 static inline void
356 cftree_remove(struct hfsc_class *cl)
357 {
358 	rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
359 }
360 
361 static inline void
362 cftree_update(struct hfsc_class *cl)
363 {
364 	cftree_remove(cl);
365 	cftree_insert(cl);
366 }
367 
368 /*
369  * service curve support functions
370  *
371  *  external service curve parameters
372  *	m: bps
373  *	d: us
374  *  internal service curve parameters
375  *	sm: (bytes/psched_us) << SM_SHIFT
376  *	ism: (psched_us/byte) << ISM_SHIFT
377  *	dx: psched_us
378  *
379  * The clock source resolution with ktime is 1.024us.
380  *
381  * sm and ism are scaled in order to keep effective digits.
382  * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
383  * digits in decimal using the following table.
384  *
385  *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
386  *  ------------+-------------------------------------------------------
387  *  bytes/1.024us 12.8e-3    128e-3     1280e-3    12800e-3   128000e-3
388  *
389  *  1.024us/byte  78.125     7.8125     0.78125    0.078125   0.0078125
390  */
391 #define	SM_SHIFT	20
392 #define	ISM_SHIFT	18
393 
394 #define	SM_MASK		((1ULL << SM_SHIFT) - 1)
395 #define	ISM_MASK	((1ULL << ISM_SHIFT) - 1)
396 
397 static inline u64
398 seg_x2y(u64 x, u64 sm)
399 {
400 	u64 y;
401 
402 	/*
403 	 * compute
404 	 *	y = x * sm >> SM_SHIFT
405 	 * but divide it for the upper and lower bits to avoid overflow
406 	 */
407 	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
408 	return y;
409 }
410 
411 static inline u64
412 seg_y2x(u64 y, u64 ism)
413 {
414 	u64 x;
415 
416 	if (y == 0)
417 		x = 0;
418 	else if (ism == HT_INFINITY)
419 		x = HT_INFINITY;
420 	else {
421 		x = (y >> ISM_SHIFT) * ism
422 		    + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
423 	}
424 	return x;
425 }
426 
427 /* Convert m (bps) into sm (bytes/psched us) */
428 static u64
429 m2sm(u32 m)
430 {
431 	u64 sm;
432 
433 	sm = ((u64)m << SM_SHIFT);
434 	sm += PSCHED_TICKS_PER_SEC - 1;
435 	do_div(sm, PSCHED_TICKS_PER_SEC);
436 	return sm;
437 }
438 
439 /* convert m (bps) into ism (psched us/byte) */
440 static u64
441 m2ism(u32 m)
442 {
443 	u64 ism;
444 
445 	if (m == 0)
446 		ism = HT_INFINITY;
447 	else {
448 		ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
449 		ism += m - 1;
450 		do_div(ism, m);
451 	}
452 	return ism;
453 }
454 
455 /* convert d (us) into dx (psched us) */
456 static u64
457 d2dx(u32 d)
458 {
459 	u64 dx;
460 
461 	dx = ((u64)d * PSCHED_TICKS_PER_SEC);
462 	dx += USEC_PER_SEC - 1;
463 	do_div(dx, USEC_PER_SEC);
464 	return dx;
465 }
466 
467 /* convert sm (bytes/psched us) into m (bps) */
468 static u32
469 sm2m(u64 sm)
470 {
471 	u64 m;
472 
473 	m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
474 	return (u32)m;
475 }
476 
477 /* convert dx (psched us) into d (us) */
478 static u32
479 dx2d(u64 dx)
480 {
481 	u64 d;
482 
483 	d = dx * USEC_PER_SEC;
484 	do_div(d, PSCHED_TICKS_PER_SEC);
485 	return (u32)d;
486 }
487 
488 static void
489 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
490 {
491 	isc->sm1  = m2sm(sc->m1);
492 	isc->ism1 = m2ism(sc->m1);
493 	isc->dx   = d2dx(sc->d);
494 	isc->dy   = seg_x2y(isc->dx, isc->sm1);
495 	isc->sm2  = m2sm(sc->m2);
496 	isc->ism2 = m2ism(sc->m2);
497 }
498 
499 /*
500  * initialize the runtime service curve with the given internal
501  * service curve starting at (x, y).
502  */
503 static void
504 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
505 {
506 	rtsc->x	   = x;
507 	rtsc->y    = y;
508 	rtsc->sm1  = isc->sm1;
509 	rtsc->ism1 = isc->ism1;
510 	rtsc->dx   = isc->dx;
511 	rtsc->dy   = isc->dy;
512 	rtsc->sm2  = isc->sm2;
513 	rtsc->ism2 = isc->ism2;
514 }
515 
516 /*
517  * calculate the y-projection of the runtime service curve by the
518  * given x-projection value
519  */
520 static u64
521 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
522 {
523 	u64 x;
524 
525 	if (y < rtsc->y)
526 		x = rtsc->x;
527 	else if (y <= rtsc->y + rtsc->dy) {
528 		/* x belongs to the 1st segment */
529 		if (rtsc->dy == 0)
530 			x = rtsc->x + rtsc->dx;
531 		else
532 			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
533 	} else {
534 		/* x belongs to the 2nd segment */
535 		x = rtsc->x + rtsc->dx
536 		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
537 	}
538 	return x;
539 }
540 
541 static u64
542 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
543 {
544 	u64 y;
545 
546 	if (x <= rtsc->x)
547 		y = rtsc->y;
548 	else if (x <= rtsc->x + rtsc->dx)
549 		/* y belongs to the 1st segment */
550 		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
551 	else
552 		/* y belongs to the 2nd segment */
553 		y = rtsc->y + rtsc->dy
554 		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
555 	return y;
556 }
557 
558 /*
559  * update the runtime service curve by taking the minimum of the current
560  * runtime service curve and the service curve starting at (x, y).
561  */
562 static void
563 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
564 {
565 	u64 y1, y2, dx, dy;
566 	u32 dsm;
567 
568 	if (isc->sm1 <= isc->sm2) {
569 		/* service curve is convex */
570 		y1 = rtsc_x2y(rtsc, x);
571 		if (y1 < y)
572 			/* the current rtsc is smaller */
573 			return;
574 		rtsc->x = x;
575 		rtsc->y = y;
576 		return;
577 	}
578 
579 	/*
580 	 * service curve is concave
581 	 * compute the two y values of the current rtsc
582 	 *	y1: at x
583 	 *	y2: at (x + dx)
584 	 */
585 	y1 = rtsc_x2y(rtsc, x);
586 	if (y1 <= y) {
587 		/* rtsc is below isc, no change to rtsc */
588 		return;
589 	}
590 
591 	y2 = rtsc_x2y(rtsc, x + isc->dx);
592 	if (y2 >= y + isc->dy) {
593 		/* rtsc is above isc, replace rtsc by isc */
594 		rtsc->x = x;
595 		rtsc->y = y;
596 		rtsc->dx = isc->dx;
597 		rtsc->dy = isc->dy;
598 		return;
599 	}
600 
601 	/*
602 	 * the two curves intersect
603 	 * compute the offsets (dx, dy) using the reverse
604 	 * function of seg_x2y()
605 	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
606 	 */
607 	dx = (y1 - y) << SM_SHIFT;
608 	dsm = isc->sm1 - isc->sm2;
609 	do_div(dx, dsm);
610 	/*
611 	 * check if (x, y1) belongs to the 1st segment of rtsc.
612 	 * if so, add the offset.
613 	 */
614 	if (rtsc->x + rtsc->dx > x)
615 		dx += rtsc->x + rtsc->dx - x;
616 	dy = seg_x2y(dx, isc->sm1);
617 
618 	rtsc->x = x;
619 	rtsc->y = y;
620 	rtsc->dx = dx;
621 	rtsc->dy = dy;
622 	return;
623 }
624 
625 static void
626 init_ed(struct hfsc_class *cl, unsigned int next_len)
627 {
628 	u64 cur_time = psched_get_time();
629 
630 	/* update the deadline curve */
631 	rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
632 
633 	/*
634 	 * update the eligible curve.
635 	 * for concave, it is equal to the deadline curve.
636 	 * for convex, it is a linear curve with slope m2.
637 	 */
638 	cl->cl_eligible = cl->cl_deadline;
639 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
640 		cl->cl_eligible.dx = 0;
641 		cl->cl_eligible.dy = 0;
642 	}
643 
644 	/* compute e and d */
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_insert(cl);
649 }
650 
651 static void
652 update_ed(struct hfsc_class *cl, unsigned int next_len)
653 {
654 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
655 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
656 
657 	eltree_update(cl);
658 }
659 
660 static inline void
661 update_d(struct hfsc_class *cl, unsigned int next_len)
662 {
663 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
664 }
665 
666 static inline void
667 update_cfmin(struct hfsc_class *cl)
668 {
669 	struct rb_node *n = rb_first(&cl->cf_tree);
670 	struct hfsc_class *p;
671 
672 	if (n == NULL) {
673 		cl->cl_cfmin = 0;
674 		return;
675 	}
676 	p = rb_entry(n, struct hfsc_class, cf_node);
677 	cl->cl_cfmin = p->cl_f;
678 }
679 
680 static void
681 init_vf(struct hfsc_class *cl, unsigned int len)
682 {
683 	struct hfsc_class *max_cl;
684 	struct rb_node *n;
685 	u64 vt, f, cur_time;
686 	int go_active;
687 
688 	cur_time = 0;
689 	go_active = 1;
690 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
691 		if (go_active && cl->cl_nactive++ == 0)
692 			go_active = 1;
693 		else
694 			go_active = 0;
695 
696 		if (go_active) {
697 			n = rb_last(&cl->cl_parent->vt_tree);
698 			if (n != NULL) {
699 				max_cl = rb_entry(n, struct hfsc_class,vt_node);
700 				/*
701 				 * set vt to the average of the min and max
702 				 * classes.  if the parent's period didn't
703 				 * change, don't decrease vt of the class.
704 				 */
705 				vt = max_cl->cl_vt;
706 				if (cl->cl_parent->cl_cvtmin != 0)
707 					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
708 
709 				if (cl->cl_parent->cl_vtperiod !=
710 				    cl->cl_parentperiod || vt > cl->cl_vt)
711 					cl->cl_vt = vt;
712 			} else {
713 				/*
714 				 * first child for a new parent backlog period.
715 				 * add parent's cvtmax to cvtoff to make a new
716 				 * vt (vtoff + vt) larger than the vt in the
717 				 * last period for all children.
718 				 */
719 				vt = cl->cl_parent->cl_cvtmax;
720 				cl->cl_parent->cl_cvtoff += vt;
721 				cl->cl_parent->cl_cvtmax = 0;
722 				cl->cl_parent->cl_cvtmin = 0;
723 				cl->cl_vt = 0;
724 			}
725 
726 			cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
727 							cl->cl_pcvtoff;
728 
729 			/* update the virtual curve */
730 			vt = cl->cl_vt + cl->cl_vtoff;
731 			rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
732 						      cl->cl_total);
733 			if (cl->cl_virtual.x == vt) {
734 				cl->cl_virtual.x -= cl->cl_vtoff;
735 				cl->cl_vtoff = 0;
736 			}
737 			cl->cl_vtadj = 0;
738 
739 			cl->cl_vtperiod++;  /* increment vt period */
740 			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
741 			if (cl->cl_parent->cl_nactive == 0)
742 				cl->cl_parentperiod++;
743 			cl->cl_f = 0;
744 
745 			vttree_insert(cl);
746 			cftree_insert(cl);
747 
748 			if (cl->cl_flags & HFSC_USC) {
749 				/* class has upper limit curve */
750 				if (cur_time == 0)
751 					cur_time = psched_get_time();
752 
753 				/* update the ulimit curve */
754 				rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
755 					 cl->cl_total);
756 				/* compute myf */
757 				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
758 						      cl->cl_total);
759 				cl->cl_myfadj = 0;
760 			}
761 		}
762 
763 		f = max(cl->cl_myf, cl->cl_cfmin);
764 		if (f != cl->cl_f) {
765 			cl->cl_f = f;
766 			cftree_update(cl);
767 			update_cfmin(cl->cl_parent);
768 		}
769 	}
770 }
771 
772 static void
773 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
774 {
775 	u64 f; /* , myf_bound, delta; */
776 	int go_passive = 0;
777 
778 	if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
779 		go_passive = 1;
780 
781 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
782 		cl->cl_total += len;
783 
784 		if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
785 			continue;
786 
787 		if (go_passive && --cl->cl_nactive == 0)
788 			go_passive = 1;
789 		else
790 			go_passive = 0;
791 
792 		if (go_passive) {
793 			/* no more active child, going passive */
794 
795 			/* update cvtmax of the parent class */
796 			if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
797 				cl->cl_parent->cl_cvtmax = cl->cl_vt;
798 
799 			/* remove this class from the vt tree */
800 			vttree_remove(cl);
801 
802 			cftree_remove(cl);
803 			update_cfmin(cl->cl_parent);
804 
805 			continue;
806 		}
807 
808 		/*
809 		 * update vt and f
810 		 */
811 		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
812 			    - cl->cl_vtoff + cl->cl_vtadj;
813 
814 		/*
815 		 * if vt of the class is smaller than cvtmin,
816 		 * the class was skipped in the past due to non-fit.
817 		 * if so, we need to adjust vtadj.
818 		 */
819 		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
820 			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
821 			cl->cl_vt = cl->cl_parent->cl_cvtmin;
822 		}
823 
824 		/* update the vt tree */
825 		vttree_update(cl);
826 
827 		if (cl->cl_flags & HFSC_USC) {
828 			cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
829 							      cl->cl_total);
830 #if 0
831 			/*
832 			 * This code causes classes to stay way under their
833 			 * limit when multiple classes are used at gigabit
834 			 * speed. needs investigation. -kaber
835 			 */
836 			/*
837 			 * if myf lags behind by more than one clock tick
838 			 * from the current time, adjust myfadj to prevent
839 			 * a rate-limited class from going greedy.
840 			 * in a steady state under rate-limiting, myf
841 			 * fluctuates within one clock tick.
842 			 */
843 			myf_bound = cur_time - PSCHED_JIFFIE2US(1);
844 			if (cl->cl_myf < myf_bound) {
845 				delta = cur_time - cl->cl_myf;
846 				cl->cl_myfadj += delta;
847 				cl->cl_myf += delta;
848 			}
849 #endif
850 		}
851 
852 		f = max(cl->cl_myf, cl->cl_cfmin);
853 		if (f != cl->cl_f) {
854 			cl->cl_f = f;
855 			cftree_update(cl);
856 			update_cfmin(cl->cl_parent);
857 		}
858 	}
859 }
860 
861 static void
862 set_active(struct hfsc_class *cl, unsigned int len)
863 {
864 	if (cl->cl_flags & HFSC_RSC)
865 		init_ed(cl, len);
866 	if (cl->cl_flags & HFSC_FSC)
867 		init_vf(cl, len);
868 
869 	list_add_tail(&cl->dlist, &cl->sched->droplist);
870 }
871 
872 static void
873 set_passive(struct hfsc_class *cl)
874 {
875 	if (cl->cl_flags & HFSC_RSC)
876 		eltree_remove(cl);
877 
878 	list_del(&cl->dlist);
879 
880 	/*
881 	 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
882 	 * needs to be called explicitly to remove a class from vttree.
883 	 */
884 }
885 
886 /*
887  * hack to get length of first packet in queue.
888  */
889 static unsigned int
890 qdisc_peek_len(struct Qdisc *sch)
891 {
892 	struct sk_buff *skb;
893 	unsigned int len;
894 
895 	skb = sch->dequeue(sch);
896 	if (skb == NULL) {
897 		if (net_ratelimit())
898 			printk("qdisc_peek_len: non work-conserving qdisc ?\n");
899 		return 0;
900 	}
901 	len = skb->len;
902 	if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
903 		if (net_ratelimit())
904 			printk("qdisc_peek_len: failed to requeue\n");
905 		qdisc_tree_decrease_qlen(sch, 1);
906 		return 0;
907 	}
908 	return len;
909 }
910 
911 static void
912 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
913 {
914 	unsigned int len = cl->qdisc->q.qlen;
915 
916 	qdisc_reset(cl->qdisc);
917 	qdisc_tree_decrease_qlen(cl->qdisc, len);
918 }
919 
920 static void
921 hfsc_adjust_levels(struct hfsc_class *cl)
922 {
923 	struct hfsc_class *p;
924 	unsigned int level;
925 
926 	do {
927 		level = 0;
928 		list_for_each_entry(p, &cl->children, siblings) {
929 			if (p->level >= level)
930 				level = p->level + 1;
931 		}
932 		cl->level = level;
933 	} while ((cl = cl->cl_parent) != NULL);
934 }
935 
936 static inline unsigned int
937 hfsc_hash(u32 h)
938 {
939 	h ^= h >> 8;
940 	h ^= h >> 4;
941 
942 	return h & (HFSC_HSIZE - 1);
943 }
944 
945 static inline struct hfsc_class *
946 hfsc_find_class(u32 classid, struct Qdisc *sch)
947 {
948 	struct hfsc_sched *q = qdisc_priv(sch);
949 	struct hfsc_class *cl;
950 
951 	list_for_each_entry(cl, &q->clhash[hfsc_hash(classid)], hlist) {
952 		if (cl->classid == classid)
953 			return cl;
954 	}
955 	return NULL;
956 }
957 
958 static void
959 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
960 		u64 cur_time)
961 {
962 	sc2isc(rsc, &cl->cl_rsc);
963 	rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
964 	cl->cl_eligible = cl->cl_deadline;
965 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
966 		cl->cl_eligible.dx = 0;
967 		cl->cl_eligible.dy = 0;
968 	}
969 	cl->cl_flags |= HFSC_RSC;
970 }
971 
972 static void
973 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
974 {
975 	sc2isc(fsc, &cl->cl_fsc);
976 	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
977 	cl->cl_flags |= HFSC_FSC;
978 }
979 
980 static void
981 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
982 		u64 cur_time)
983 {
984 	sc2isc(usc, &cl->cl_usc);
985 	rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
986 	cl->cl_flags |= HFSC_USC;
987 }
988 
989 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
990 	[TCA_HFSC_RSC]	= { .len = sizeof(struct tc_service_curve) },
991 	[TCA_HFSC_FSC]	= { .len = sizeof(struct tc_service_curve) },
992 	[TCA_HFSC_USC]	= { .len = sizeof(struct tc_service_curve) },
993 };
994 
995 static int
996 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
997 		  struct nlattr **tca, unsigned long *arg)
998 {
999 	struct hfsc_sched *q = qdisc_priv(sch);
1000 	struct hfsc_class *cl = (struct hfsc_class *)*arg;
1001 	struct hfsc_class *parent = NULL;
1002 	struct nlattr *opt = tca[TCA_OPTIONS];
1003 	struct nlattr *tb[TCA_HFSC_MAX + 1];
1004 	struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
1005 	u64 cur_time;
1006 	int err;
1007 
1008 	if (opt == NULL)
1009 		return -EINVAL;
1010 
1011 	err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
1012 	if (err < 0)
1013 		return err;
1014 
1015 	if (tb[TCA_HFSC_RSC]) {
1016 		rsc = nla_data(tb[TCA_HFSC_RSC]);
1017 		if (rsc->m1 == 0 && rsc->m2 == 0)
1018 			rsc = NULL;
1019 	}
1020 
1021 	if (tb[TCA_HFSC_FSC]) {
1022 		fsc = nla_data(tb[TCA_HFSC_FSC]);
1023 		if (fsc->m1 == 0 && fsc->m2 == 0)
1024 			fsc = NULL;
1025 	}
1026 
1027 	if (tb[TCA_HFSC_USC]) {
1028 		usc = nla_data(tb[TCA_HFSC_USC]);
1029 		if (usc->m1 == 0 && usc->m2 == 0)
1030 			usc = NULL;
1031 	}
1032 
1033 	if (cl != NULL) {
1034 		if (parentid) {
1035 			if (cl->cl_parent && cl->cl_parent->classid != parentid)
1036 				return -EINVAL;
1037 			if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1038 				return -EINVAL;
1039 		}
1040 		cur_time = psched_get_time();
1041 
1042 		sch_tree_lock(sch);
1043 		if (rsc != NULL)
1044 			hfsc_change_rsc(cl, rsc, cur_time);
1045 		if (fsc != NULL)
1046 			hfsc_change_fsc(cl, fsc);
1047 		if (usc != NULL)
1048 			hfsc_change_usc(cl, usc, cur_time);
1049 
1050 		if (cl->qdisc->q.qlen != 0) {
1051 			if (cl->cl_flags & HFSC_RSC)
1052 				update_ed(cl, qdisc_peek_len(cl->qdisc));
1053 			if (cl->cl_flags & HFSC_FSC)
1054 				update_vf(cl, 0, cur_time);
1055 		}
1056 		sch_tree_unlock(sch);
1057 
1058 		if (tca[TCA_RATE])
1059 			gen_replace_estimator(&cl->bstats, &cl->rate_est,
1060 					      &sch->dev->queue_lock,
1061 					      tca[TCA_RATE]);
1062 		return 0;
1063 	}
1064 
1065 	if (parentid == TC_H_ROOT)
1066 		return -EEXIST;
1067 
1068 	parent = &q->root;
1069 	if (parentid) {
1070 		parent = hfsc_find_class(parentid, sch);
1071 		if (parent == NULL)
1072 			return -ENOENT;
1073 	}
1074 
1075 	if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1076 		return -EINVAL;
1077 	if (hfsc_find_class(classid, sch))
1078 		return -EEXIST;
1079 
1080 	if (rsc == NULL && fsc == NULL)
1081 		return -EINVAL;
1082 
1083 	cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1084 	if (cl == NULL)
1085 		return -ENOBUFS;
1086 
1087 	if (rsc != NULL)
1088 		hfsc_change_rsc(cl, rsc, 0);
1089 	if (fsc != NULL)
1090 		hfsc_change_fsc(cl, fsc);
1091 	if (usc != NULL)
1092 		hfsc_change_usc(cl, usc, 0);
1093 
1094 	cl->refcnt    = 1;
1095 	cl->classid   = classid;
1096 	cl->sched     = q;
1097 	cl->cl_parent = parent;
1098 	cl->qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops, classid);
1099 	if (cl->qdisc == NULL)
1100 		cl->qdisc = &noop_qdisc;
1101 	INIT_LIST_HEAD(&cl->children);
1102 	cl->vt_tree = RB_ROOT;
1103 	cl->cf_tree = RB_ROOT;
1104 
1105 	sch_tree_lock(sch);
1106 	list_add_tail(&cl->hlist, &q->clhash[hfsc_hash(classid)]);
1107 	list_add_tail(&cl->siblings, &parent->children);
1108 	if (parent->level == 0)
1109 		hfsc_purge_queue(sch, parent);
1110 	hfsc_adjust_levels(parent);
1111 	cl->cl_pcvtoff = parent->cl_cvtoff;
1112 	sch_tree_unlock(sch);
1113 
1114 	if (tca[TCA_RATE])
1115 		gen_new_estimator(&cl->bstats, &cl->rate_est,
1116 				  &sch->dev->queue_lock, tca[TCA_RATE]);
1117 	*arg = (unsigned long)cl;
1118 	return 0;
1119 }
1120 
1121 static void
1122 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1123 {
1124 	struct hfsc_sched *q = qdisc_priv(sch);
1125 
1126 	tcf_destroy_chain(cl->filter_list);
1127 	qdisc_destroy(cl->qdisc);
1128 	gen_kill_estimator(&cl->bstats, &cl->rate_est);
1129 	if (cl != &q->root)
1130 		kfree(cl);
1131 }
1132 
1133 static int
1134 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1135 {
1136 	struct hfsc_sched *q = qdisc_priv(sch);
1137 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1138 
1139 	if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1140 		return -EBUSY;
1141 
1142 	sch_tree_lock(sch);
1143 
1144 	list_del(&cl->siblings);
1145 	hfsc_adjust_levels(cl->cl_parent);
1146 
1147 	hfsc_purge_queue(sch, cl);
1148 	list_del(&cl->hlist);
1149 
1150 	if (--cl->refcnt == 0)
1151 		hfsc_destroy_class(sch, cl);
1152 
1153 	sch_tree_unlock(sch);
1154 	return 0;
1155 }
1156 
1157 static struct hfsc_class *
1158 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1159 {
1160 	struct hfsc_sched *q = qdisc_priv(sch);
1161 	struct hfsc_class *cl;
1162 	struct tcf_result res;
1163 	struct tcf_proto *tcf;
1164 	int result;
1165 
1166 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1167 	    (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1168 		if (cl->level == 0)
1169 			return cl;
1170 
1171 	*qerr = NET_XMIT_BYPASS;
1172 	tcf = q->root.filter_list;
1173 	while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1174 #ifdef CONFIG_NET_CLS_ACT
1175 		switch (result) {
1176 		case TC_ACT_QUEUED:
1177 		case TC_ACT_STOLEN:
1178 			*qerr = NET_XMIT_SUCCESS;
1179 		case TC_ACT_SHOT:
1180 			return NULL;
1181 		}
1182 #endif
1183 		if ((cl = (struct hfsc_class *)res.class) == NULL) {
1184 			if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1185 				break; /* filter selected invalid classid */
1186 		}
1187 
1188 		if (cl->level == 0)
1189 			return cl; /* hit leaf class */
1190 
1191 		/* apply inner filter chain */
1192 		tcf = cl->filter_list;
1193 	}
1194 
1195 	/* classification failed, try default class */
1196 	cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1197 	if (cl == NULL || cl->level > 0)
1198 		return NULL;
1199 
1200 	return cl;
1201 }
1202 
1203 static int
1204 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1205 		 struct Qdisc **old)
1206 {
1207 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1208 
1209 	if (cl == NULL)
1210 		return -ENOENT;
1211 	if (cl->level > 0)
1212 		return -EINVAL;
1213 	if (new == NULL) {
1214 		new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1215 					cl->classid);
1216 		if (new == NULL)
1217 			new = &noop_qdisc;
1218 	}
1219 
1220 	sch_tree_lock(sch);
1221 	hfsc_purge_queue(sch, cl);
1222 	*old = xchg(&cl->qdisc, new);
1223 	sch_tree_unlock(sch);
1224 	return 0;
1225 }
1226 
1227 static struct Qdisc *
1228 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1229 {
1230 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1231 
1232 	if (cl != NULL && cl->level == 0)
1233 		return cl->qdisc;
1234 
1235 	return NULL;
1236 }
1237 
1238 static void
1239 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1240 {
1241 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1242 
1243 	if (cl->qdisc->q.qlen == 0) {
1244 		update_vf(cl, 0, 0);
1245 		set_passive(cl);
1246 	}
1247 }
1248 
1249 static unsigned long
1250 hfsc_get_class(struct Qdisc *sch, u32 classid)
1251 {
1252 	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1253 
1254 	if (cl != NULL)
1255 		cl->refcnt++;
1256 
1257 	return (unsigned long)cl;
1258 }
1259 
1260 static void
1261 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1262 {
1263 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1264 
1265 	if (--cl->refcnt == 0)
1266 		hfsc_destroy_class(sch, cl);
1267 }
1268 
1269 static unsigned long
1270 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1271 {
1272 	struct hfsc_class *p = (struct hfsc_class *)parent;
1273 	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1274 
1275 	if (cl != NULL) {
1276 		if (p != NULL && p->level <= cl->level)
1277 			return 0;
1278 		cl->filter_cnt++;
1279 	}
1280 
1281 	return (unsigned long)cl;
1282 }
1283 
1284 static void
1285 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1286 {
1287 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1288 
1289 	cl->filter_cnt--;
1290 }
1291 
1292 static struct tcf_proto **
1293 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1294 {
1295 	struct hfsc_sched *q = qdisc_priv(sch);
1296 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1297 
1298 	if (cl == NULL)
1299 		cl = &q->root;
1300 
1301 	return &cl->filter_list;
1302 }
1303 
1304 static int
1305 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1306 {
1307 	struct tc_service_curve tsc;
1308 
1309 	tsc.m1 = sm2m(sc->sm1);
1310 	tsc.d  = dx2d(sc->dx);
1311 	tsc.m2 = sm2m(sc->sm2);
1312 	NLA_PUT(skb, attr, sizeof(tsc), &tsc);
1313 
1314 	return skb->len;
1315 
1316  nla_put_failure:
1317 	return -1;
1318 }
1319 
1320 static inline int
1321 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1322 {
1323 	if ((cl->cl_flags & HFSC_RSC) &&
1324 	    (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1325 		goto nla_put_failure;
1326 
1327 	if ((cl->cl_flags & HFSC_FSC) &&
1328 	    (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1329 		goto nla_put_failure;
1330 
1331 	if ((cl->cl_flags & HFSC_USC) &&
1332 	    (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1333 		goto nla_put_failure;
1334 
1335 	return skb->len;
1336 
1337  nla_put_failure:
1338 	return -1;
1339 }
1340 
1341 static int
1342 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1343 		struct tcmsg *tcm)
1344 {
1345 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1346 	struct nlattr *nest;
1347 
1348 	tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->classid : TC_H_ROOT;
1349 	tcm->tcm_handle = cl->classid;
1350 	if (cl->level == 0)
1351 		tcm->tcm_info = cl->qdisc->handle;
1352 
1353 	nest = nla_nest_start(skb, TCA_OPTIONS);
1354 	if (nest == NULL)
1355 		goto nla_put_failure;
1356 	if (hfsc_dump_curves(skb, cl) < 0)
1357 		goto nla_put_failure;
1358 	nla_nest_end(skb, nest);
1359 	return skb->len;
1360 
1361  nla_put_failure:
1362 	nla_nest_cancel(skb, nest);
1363 	return -1;
1364 }
1365 
1366 static int
1367 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1368 	struct gnet_dump *d)
1369 {
1370 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1371 	struct tc_hfsc_stats xstats;
1372 
1373 	cl->qstats.qlen = cl->qdisc->q.qlen;
1374 	xstats.level   = cl->level;
1375 	xstats.period  = cl->cl_vtperiod;
1376 	xstats.work    = cl->cl_total;
1377 	xstats.rtwork  = cl->cl_cumul;
1378 
1379 	if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1380 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1381 	    gnet_stats_copy_queue(d, &cl->qstats) < 0)
1382 		return -1;
1383 
1384 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1385 }
1386 
1387 
1388 
1389 static void
1390 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1391 {
1392 	struct hfsc_sched *q = qdisc_priv(sch);
1393 	struct hfsc_class *cl;
1394 	unsigned int i;
1395 
1396 	if (arg->stop)
1397 		return;
1398 
1399 	for (i = 0; i < HFSC_HSIZE; i++) {
1400 		list_for_each_entry(cl, &q->clhash[i], hlist) {
1401 			if (arg->count < arg->skip) {
1402 				arg->count++;
1403 				continue;
1404 			}
1405 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1406 				arg->stop = 1;
1407 				return;
1408 			}
1409 			arg->count++;
1410 		}
1411 	}
1412 }
1413 
1414 static void
1415 hfsc_schedule_watchdog(struct Qdisc *sch)
1416 {
1417 	struct hfsc_sched *q = qdisc_priv(sch);
1418 	struct hfsc_class *cl;
1419 	u64 next_time = 0;
1420 
1421 	if ((cl = eltree_get_minel(q)) != NULL)
1422 		next_time = cl->cl_e;
1423 	if (q->root.cl_cfmin != 0) {
1424 		if (next_time == 0 || next_time > q->root.cl_cfmin)
1425 			next_time = q->root.cl_cfmin;
1426 	}
1427 	WARN_ON(next_time == 0);
1428 	qdisc_watchdog_schedule(&q->watchdog, next_time);
1429 }
1430 
1431 static int
1432 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1433 {
1434 	struct hfsc_sched *q = qdisc_priv(sch);
1435 	struct tc_hfsc_qopt *qopt;
1436 	unsigned int i;
1437 
1438 	if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1439 		return -EINVAL;
1440 	qopt = nla_data(opt);
1441 
1442 	q->defcls = qopt->defcls;
1443 	for (i = 0; i < HFSC_HSIZE; i++)
1444 		INIT_LIST_HEAD(&q->clhash[i]);
1445 	q->eligible = RB_ROOT;
1446 	INIT_LIST_HEAD(&q->droplist);
1447 	skb_queue_head_init(&q->requeue);
1448 
1449 	q->root.refcnt  = 1;
1450 	q->root.classid = sch->handle;
1451 	q->root.sched   = q;
1452 	q->root.qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1453 					  sch->handle);
1454 	if (q->root.qdisc == NULL)
1455 		q->root.qdisc = &noop_qdisc;
1456 	INIT_LIST_HEAD(&q->root.children);
1457 	q->root.vt_tree = RB_ROOT;
1458 	q->root.cf_tree = RB_ROOT;
1459 
1460 	list_add(&q->root.hlist, &q->clhash[hfsc_hash(q->root.classid)]);
1461 
1462 	qdisc_watchdog_init(&q->watchdog, sch);
1463 
1464 	return 0;
1465 }
1466 
1467 static int
1468 hfsc_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 
1484 static void
1485 hfsc_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 
1518 static void
1519 hfsc_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 < HFSC_HSIZE; i++) {
1526 		list_for_each_entry(cl, &q->clhash[i], hlist)
1527 			hfsc_reset_class(cl);
1528 	}
1529 	__skb_queue_purge(&q->requeue);
1530 	q->eligible = RB_ROOT;
1531 	INIT_LIST_HEAD(&q->droplist);
1532 	qdisc_watchdog_cancel(&q->watchdog);
1533 	sch->q.qlen = 0;
1534 }
1535 
1536 static void
1537 hfsc_destroy_qdisc(struct Qdisc *sch)
1538 {
1539 	struct hfsc_sched *q = qdisc_priv(sch);
1540 	struct hfsc_class *cl, *next;
1541 	unsigned int i;
1542 
1543 	for (i = 0; i < HFSC_HSIZE; i++) {
1544 		list_for_each_entry_safe(cl, next, &q->clhash[i], hlist)
1545 			hfsc_destroy_class(sch, cl);
1546 	}
1547 	__skb_queue_purge(&q->requeue);
1548 	qdisc_watchdog_cancel(&q->watchdog);
1549 }
1550 
1551 static int
1552 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1553 {
1554 	struct hfsc_sched *q = qdisc_priv(sch);
1555 	unsigned char *b = skb_tail_pointer(skb);
1556 	struct tc_hfsc_qopt qopt;
1557 
1558 	qopt.defcls = q->defcls;
1559 	NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1560 	return skb->len;
1561 
1562  nla_put_failure:
1563 	nlmsg_trim(skb, b);
1564 	return -1;
1565 }
1566 
1567 static int
1568 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1569 {
1570 	struct hfsc_class *cl;
1571 	unsigned int len;
1572 	int err;
1573 
1574 	cl = hfsc_classify(skb, sch, &err);
1575 	if (cl == NULL) {
1576 		if (err == NET_XMIT_BYPASS)
1577 			sch->qstats.drops++;
1578 		kfree_skb(skb);
1579 		return err;
1580 	}
1581 
1582 	len = skb->len;
1583 	err = cl->qdisc->enqueue(skb, cl->qdisc);
1584 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1585 		cl->qstats.drops++;
1586 		sch->qstats.drops++;
1587 		return err;
1588 	}
1589 
1590 	if (cl->qdisc->q.qlen == 1)
1591 		set_active(cl, len);
1592 
1593 	cl->bstats.packets++;
1594 	cl->bstats.bytes += len;
1595 	sch->bstats.packets++;
1596 	sch->bstats.bytes += len;
1597 	sch->q.qlen++;
1598 
1599 	return NET_XMIT_SUCCESS;
1600 }
1601 
1602 static struct sk_buff *
1603 hfsc_dequeue(struct Qdisc *sch)
1604 {
1605 	struct hfsc_sched *q = qdisc_priv(sch);
1606 	struct hfsc_class *cl;
1607 	struct sk_buff *skb;
1608 	u64 cur_time;
1609 	unsigned int next_len;
1610 	int realtime = 0;
1611 
1612 	if (sch->q.qlen == 0)
1613 		return NULL;
1614 	if ((skb = __skb_dequeue(&q->requeue)))
1615 		goto out;
1616 
1617 	cur_time = psched_get_time();
1618 
1619 	/*
1620 	 * if there are eligible classes, use real-time criteria.
1621 	 * find the class with the minimum deadline among
1622 	 * the eligible classes.
1623 	 */
1624 	if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1625 		realtime = 1;
1626 	} else {
1627 		/*
1628 		 * use link-sharing criteria
1629 		 * get the class with the minimum vt in the hierarchy
1630 		 */
1631 		cl = vttree_get_minvt(&q->root, cur_time);
1632 		if (cl == NULL) {
1633 			sch->qstats.overlimits++;
1634 			hfsc_schedule_watchdog(sch);
1635 			return NULL;
1636 		}
1637 	}
1638 
1639 	skb = cl->qdisc->dequeue(cl->qdisc);
1640 	if (skb == NULL) {
1641 		if (net_ratelimit())
1642 			printk("HFSC: Non-work-conserving qdisc ?\n");
1643 		return NULL;
1644 	}
1645 
1646 	update_vf(cl, skb->len, cur_time);
1647 	if (realtime)
1648 		cl->cl_cumul += skb->len;
1649 
1650 	if (cl->qdisc->q.qlen != 0) {
1651 		if (cl->cl_flags & HFSC_RSC) {
1652 			/* update ed */
1653 			next_len = qdisc_peek_len(cl->qdisc);
1654 			if (realtime)
1655 				update_ed(cl, next_len);
1656 			else
1657 				update_d(cl, next_len);
1658 		}
1659 	} else {
1660 		/* the class becomes passive */
1661 		set_passive(cl);
1662 	}
1663 
1664  out:
1665 	sch->flags &= ~TCQ_F_THROTTLED;
1666 	sch->q.qlen--;
1667 
1668 	return skb;
1669 }
1670 
1671 static int
1672 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1673 {
1674 	struct hfsc_sched *q = qdisc_priv(sch);
1675 
1676 	__skb_queue_head(&q->requeue, skb);
1677 	sch->q.qlen++;
1678 	sch->qstats.requeues++;
1679 	return NET_XMIT_SUCCESS;
1680 }
1681 
1682 static unsigned int
1683 hfsc_drop(struct Qdisc *sch)
1684 {
1685 	struct hfsc_sched *q = qdisc_priv(sch);
1686 	struct hfsc_class *cl;
1687 	unsigned int len;
1688 
1689 	list_for_each_entry(cl, &q->droplist, dlist) {
1690 		if (cl->qdisc->ops->drop != NULL &&
1691 		    (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1692 			if (cl->qdisc->q.qlen == 0) {
1693 				update_vf(cl, 0, 0);
1694 				set_passive(cl);
1695 			} else {
1696 				list_move_tail(&cl->dlist, &q->droplist);
1697 			}
1698 			cl->qstats.drops++;
1699 			sch->qstats.drops++;
1700 			sch->q.qlen--;
1701 			return len;
1702 		}
1703 	}
1704 	return 0;
1705 }
1706 
1707 static const struct Qdisc_class_ops hfsc_class_ops = {
1708 	.change		= hfsc_change_class,
1709 	.delete		= hfsc_delete_class,
1710 	.graft		= hfsc_graft_class,
1711 	.leaf		= hfsc_class_leaf,
1712 	.qlen_notify	= hfsc_qlen_notify,
1713 	.get		= hfsc_get_class,
1714 	.put		= hfsc_put_class,
1715 	.bind_tcf	= hfsc_bind_tcf,
1716 	.unbind_tcf	= hfsc_unbind_tcf,
1717 	.tcf_chain	= hfsc_tcf_chain,
1718 	.dump		= hfsc_dump_class,
1719 	.dump_stats	= hfsc_dump_class_stats,
1720 	.walk		= hfsc_walk
1721 };
1722 
1723 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1724 	.id		= "hfsc",
1725 	.init		= hfsc_init_qdisc,
1726 	.change		= hfsc_change_qdisc,
1727 	.reset		= hfsc_reset_qdisc,
1728 	.destroy	= hfsc_destroy_qdisc,
1729 	.dump		= hfsc_dump_qdisc,
1730 	.enqueue	= hfsc_enqueue,
1731 	.dequeue	= hfsc_dequeue,
1732 	.requeue	= hfsc_requeue,
1733 	.drop		= hfsc_drop,
1734 	.cl_ops		= &hfsc_class_ops,
1735 	.priv_size	= sizeof(struct hfsc_sched),
1736 	.owner		= THIS_MODULE
1737 };
1738 
1739 static int __init
1740 hfsc_init(void)
1741 {
1742 	return register_qdisc(&hfsc_qdisc_ops);
1743 }
1744 
1745 static void __exit
1746 hfsc_cleanup(void)
1747 {
1748 	unregister_qdisc(&hfsc_qdisc_ops);
1749 }
1750 
1751 MODULE_LICENSE("GPL");
1752 module_init(hfsc_init);
1753 module_exit(hfsc_cleanup);
1754