xref: /openbmc/linux/net/sched/sch_hfsc.c (revision 9ac8d3fb)
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 	struct Qdisc_class_common cl_common;
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 dlist;		/* drop list member */
138 
139 	u64	cl_total;		/* total work in bytes */
140 	u64	cl_cumul;		/* cumulative work in bytes done by
141 					   real-time criteria */
142 
143 	u64 	cl_d;			/* deadline*/
144 	u64 	cl_e;			/* eligible time */
145 	u64	cl_vt;			/* virtual time */
146 	u64	cl_f;			/* time when this class will fit for
147 					   link-sharing, max(myf, cfmin) */
148 	u64	cl_myf;			/* my fit-time (calculated from this
149 					   class's own upperlimit curve) */
150 	u64	cl_myfadj;		/* my fit-time adjustment (to cancel
151 					   history dependence) */
152 	u64	cl_cfmin;		/* earliest children's fit-time (used
153 					   with cl_myf to obtain cl_f) */
154 	u64	cl_cvtmin;		/* minimal virtual time among the
155 					   children fit for link-sharing
156 					   (monotonic within a period) */
157 	u64	cl_vtadj;		/* intra-period cumulative vt
158 					   adjustment */
159 	u64	cl_vtoff;		/* inter-period cumulative vt offset */
160 	u64	cl_cvtmax;		/* max child's vt in the last period */
161 	u64	cl_cvtoff;		/* cumulative cvtmax of all periods */
162 	u64	cl_pcvtoff;		/* parent's cvtoff at initialization
163 					   time */
164 
165 	struct internal_sc cl_rsc;	/* internal real-time service curve */
166 	struct internal_sc cl_fsc;	/* internal fair service curve */
167 	struct internal_sc cl_usc;	/* internal upperlimit service curve */
168 	struct runtime_sc cl_deadline;	/* deadline curve */
169 	struct runtime_sc cl_eligible;	/* eligible curve */
170 	struct runtime_sc cl_virtual;	/* virtual curve */
171 	struct runtime_sc cl_ulimit;	/* upperlimit curve */
172 
173 	unsigned long	cl_flags;	/* which curves are valid */
174 	unsigned long	cl_vtperiod;	/* vt period sequence number */
175 	unsigned long	cl_parentperiod;/* parent's vt period sequence number*/
176 	unsigned long	cl_nactive;	/* number of active children */
177 };
178 
179 struct hfsc_sched
180 {
181 	u16	defcls;				/* default class id */
182 	struct hfsc_class root;			/* root class */
183 	struct Qdisc_class_hash clhash;		/* class hash */
184 	struct rb_root eligible;		/* eligible tree */
185 	struct list_head droplist;		/* active leaf class list (for
186 						   dropping) */
187 	struct sk_buff_head requeue;		/* requeued packet */
188 	struct qdisc_watchdog watchdog;		/* watchdog timer */
189 };
190 
191 #define	HT_INFINITY	0xffffffffffffffffULL	/* infinite time value */
192 
193 
194 /*
195  * eligible tree holds backlogged classes being sorted by their eligible times.
196  * there is one eligible tree per hfsc instance.
197  */
198 
199 static void
200 eltree_insert(struct hfsc_class *cl)
201 {
202 	struct rb_node **p = &cl->sched->eligible.rb_node;
203 	struct rb_node *parent = NULL;
204 	struct hfsc_class *cl1;
205 
206 	while (*p != NULL) {
207 		parent = *p;
208 		cl1 = rb_entry(parent, struct hfsc_class, el_node);
209 		if (cl->cl_e >= cl1->cl_e)
210 			p = &parent->rb_right;
211 		else
212 			p = &parent->rb_left;
213 	}
214 	rb_link_node(&cl->el_node, parent, p);
215 	rb_insert_color(&cl->el_node, &cl->sched->eligible);
216 }
217 
218 static inline void
219 eltree_remove(struct hfsc_class *cl)
220 {
221 	rb_erase(&cl->el_node, &cl->sched->eligible);
222 }
223 
224 static inline void
225 eltree_update(struct hfsc_class *cl)
226 {
227 	eltree_remove(cl);
228 	eltree_insert(cl);
229 }
230 
231 /* find the class with the minimum deadline among the eligible classes */
232 static inline struct hfsc_class *
233 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
234 {
235 	struct hfsc_class *p, *cl = NULL;
236 	struct rb_node *n;
237 
238 	for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
239 		p = rb_entry(n, struct hfsc_class, el_node);
240 		if (p->cl_e > cur_time)
241 			break;
242 		if (cl == NULL || p->cl_d < cl->cl_d)
243 			cl = p;
244 	}
245 	return cl;
246 }
247 
248 /* find the class with minimum eligible time among the eligible classes */
249 static inline struct hfsc_class *
250 eltree_get_minel(struct hfsc_sched *q)
251 {
252 	struct rb_node *n;
253 
254 	n = rb_first(&q->eligible);
255 	if (n == NULL)
256 		return NULL;
257 	return rb_entry(n, struct hfsc_class, el_node);
258 }
259 
260 /*
261  * vttree holds holds backlogged child classes being sorted by their virtual
262  * time. each intermediate class has one vttree.
263  */
264 static void
265 vttree_insert(struct hfsc_class *cl)
266 {
267 	struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
268 	struct rb_node *parent = NULL;
269 	struct hfsc_class *cl1;
270 
271 	while (*p != NULL) {
272 		parent = *p;
273 		cl1 = rb_entry(parent, struct hfsc_class, vt_node);
274 		if (cl->cl_vt >= cl1->cl_vt)
275 			p = &parent->rb_right;
276 		else
277 			p = &parent->rb_left;
278 	}
279 	rb_link_node(&cl->vt_node, parent, p);
280 	rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
281 }
282 
283 static inline void
284 vttree_remove(struct hfsc_class *cl)
285 {
286 	rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
287 }
288 
289 static inline void
290 vttree_update(struct hfsc_class *cl)
291 {
292 	vttree_remove(cl);
293 	vttree_insert(cl);
294 }
295 
296 static inline struct hfsc_class *
297 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
298 {
299 	struct hfsc_class *p;
300 	struct rb_node *n;
301 
302 	for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
303 		p = rb_entry(n, struct hfsc_class, vt_node);
304 		if (p->cl_f <= cur_time)
305 			return p;
306 	}
307 	return NULL;
308 }
309 
310 /*
311  * get the leaf class with the minimum vt in the hierarchy
312  */
313 static struct hfsc_class *
314 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
315 {
316 	/* if root-class's cfmin is bigger than cur_time nothing to do */
317 	if (cl->cl_cfmin > cur_time)
318 		return NULL;
319 
320 	while (cl->level > 0) {
321 		cl = vttree_firstfit(cl, cur_time);
322 		if (cl == NULL)
323 			return NULL;
324 		/*
325 		 * update parent's cl_cvtmin.
326 		 */
327 		if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
328 			cl->cl_parent->cl_cvtmin = cl->cl_vt;
329 	}
330 	return cl;
331 }
332 
333 static void
334 cftree_insert(struct hfsc_class *cl)
335 {
336 	struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
337 	struct rb_node *parent = NULL;
338 	struct hfsc_class *cl1;
339 
340 	while (*p != NULL) {
341 		parent = *p;
342 		cl1 = rb_entry(parent, struct hfsc_class, cf_node);
343 		if (cl->cl_f >= cl1->cl_f)
344 			p = &parent->rb_right;
345 		else
346 			p = &parent->rb_left;
347 	}
348 	rb_link_node(&cl->cf_node, parent, p);
349 	rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
350 }
351 
352 static inline void
353 cftree_remove(struct hfsc_class *cl)
354 {
355 	rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
356 }
357 
358 static inline void
359 cftree_update(struct hfsc_class *cl)
360 {
361 	cftree_remove(cl);
362 	cftree_insert(cl);
363 }
364 
365 /*
366  * service curve support functions
367  *
368  *  external service curve parameters
369  *	m: bps
370  *	d: us
371  *  internal service curve parameters
372  *	sm: (bytes/psched_us) << SM_SHIFT
373  *	ism: (psched_us/byte) << ISM_SHIFT
374  *	dx: psched_us
375  *
376  * The clock source resolution with ktime is 1.024us.
377  *
378  * sm and ism are scaled in order to keep effective digits.
379  * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
380  * digits in decimal using the following table.
381  *
382  *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
383  *  ------------+-------------------------------------------------------
384  *  bytes/1.024us 12.8e-3    128e-3     1280e-3    12800e-3   128000e-3
385  *
386  *  1.024us/byte  78.125     7.8125     0.78125    0.078125   0.0078125
387  */
388 #define	SM_SHIFT	20
389 #define	ISM_SHIFT	18
390 
391 #define	SM_MASK		((1ULL << SM_SHIFT) - 1)
392 #define	ISM_MASK	((1ULL << ISM_SHIFT) - 1)
393 
394 static inline u64
395 seg_x2y(u64 x, u64 sm)
396 {
397 	u64 y;
398 
399 	/*
400 	 * compute
401 	 *	y = x * sm >> SM_SHIFT
402 	 * but divide it for the upper and lower bits to avoid overflow
403 	 */
404 	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
405 	return y;
406 }
407 
408 static inline u64
409 seg_y2x(u64 y, u64 ism)
410 {
411 	u64 x;
412 
413 	if (y == 0)
414 		x = 0;
415 	else if (ism == HT_INFINITY)
416 		x = HT_INFINITY;
417 	else {
418 		x = (y >> ISM_SHIFT) * ism
419 		    + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
420 	}
421 	return x;
422 }
423 
424 /* Convert m (bps) into sm (bytes/psched us) */
425 static u64
426 m2sm(u32 m)
427 {
428 	u64 sm;
429 
430 	sm = ((u64)m << SM_SHIFT);
431 	sm += PSCHED_TICKS_PER_SEC - 1;
432 	do_div(sm, PSCHED_TICKS_PER_SEC);
433 	return sm;
434 }
435 
436 /* convert m (bps) into ism (psched us/byte) */
437 static u64
438 m2ism(u32 m)
439 {
440 	u64 ism;
441 
442 	if (m == 0)
443 		ism = HT_INFINITY;
444 	else {
445 		ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
446 		ism += m - 1;
447 		do_div(ism, m);
448 	}
449 	return ism;
450 }
451 
452 /* convert d (us) into dx (psched us) */
453 static u64
454 d2dx(u32 d)
455 {
456 	u64 dx;
457 
458 	dx = ((u64)d * PSCHED_TICKS_PER_SEC);
459 	dx += USEC_PER_SEC - 1;
460 	do_div(dx, USEC_PER_SEC);
461 	return dx;
462 }
463 
464 /* convert sm (bytes/psched us) into m (bps) */
465 static u32
466 sm2m(u64 sm)
467 {
468 	u64 m;
469 
470 	m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
471 	return (u32)m;
472 }
473 
474 /* convert dx (psched us) into d (us) */
475 static u32
476 dx2d(u64 dx)
477 {
478 	u64 d;
479 
480 	d = dx * USEC_PER_SEC;
481 	do_div(d, PSCHED_TICKS_PER_SEC);
482 	return (u32)d;
483 }
484 
485 static void
486 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
487 {
488 	isc->sm1  = m2sm(sc->m1);
489 	isc->ism1 = m2ism(sc->m1);
490 	isc->dx   = d2dx(sc->d);
491 	isc->dy   = seg_x2y(isc->dx, isc->sm1);
492 	isc->sm2  = m2sm(sc->m2);
493 	isc->ism2 = m2ism(sc->m2);
494 }
495 
496 /*
497  * initialize the runtime service curve with the given internal
498  * service curve starting at (x, y).
499  */
500 static void
501 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
502 {
503 	rtsc->x	   = x;
504 	rtsc->y    = y;
505 	rtsc->sm1  = isc->sm1;
506 	rtsc->ism1 = isc->ism1;
507 	rtsc->dx   = isc->dx;
508 	rtsc->dy   = isc->dy;
509 	rtsc->sm2  = isc->sm2;
510 	rtsc->ism2 = isc->ism2;
511 }
512 
513 /*
514  * calculate the y-projection of the runtime service curve by the
515  * given x-projection value
516  */
517 static u64
518 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
519 {
520 	u64 x;
521 
522 	if (y < rtsc->y)
523 		x = rtsc->x;
524 	else if (y <= rtsc->y + rtsc->dy) {
525 		/* x belongs to the 1st segment */
526 		if (rtsc->dy == 0)
527 			x = rtsc->x + rtsc->dx;
528 		else
529 			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
530 	} else {
531 		/* x belongs to the 2nd segment */
532 		x = rtsc->x + rtsc->dx
533 		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
534 	}
535 	return x;
536 }
537 
538 static u64
539 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
540 {
541 	u64 y;
542 
543 	if (x <= rtsc->x)
544 		y = rtsc->y;
545 	else if (x <= rtsc->x + rtsc->dx)
546 		/* y belongs to the 1st segment */
547 		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
548 	else
549 		/* y belongs to the 2nd segment */
550 		y = rtsc->y + rtsc->dy
551 		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
552 	return y;
553 }
554 
555 /*
556  * update the runtime service curve by taking the minimum of the current
557  * runtime service curve and the service curve starting at (x, y).
558  */
559 static void
560 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
561 {
562 	u64 y1, y2, dx, dy;
563 	u32 dsm;
564 
565 	if (isc->sm1 <= isc->sm2) {
566 		/* service curve is convex */
567 		y1 = rtsc_x2y(rtsc, x);
568 		if (y1 < y)
569 			/* the current rtsc is smaller */
570 			return;
571 		rtsc->x = x;
572 		rtsc->y = y;
573 		return;
574 	}
575 
576 	/*
577 	 * service curve is concave
578 	 * compute the two y values of the current rtsc
579 	 *	y1: at x
580 	 *	y2: at (x + dx)
581 	 */
582 	y1 = rtsc_x2y(rtsc, x);
583 	if (y1 <= y) {
584 		/* rtsc is below isc, no change to rtsc */
585 		return;
586 	}
587 
588 	y2 = rtsc_x2y(rtsc, x + isc->dx);
589 	if (y2 >= y + isc->dy) {
590 		/* rtsc is above isc, replace rtsc by isc */
591 		rtsc->x = x;
592 		rtsc->y = y;
593 		rtsc->dx = isc->dx;
594 		rtsc->dy = isc->dy;
595 		return;
596 	}
597 
598 	/*
599 	 * the two curves intersect
600 	 * compute the offsets (dx, dy) using the reverse
601 	 * function of seg_x2y()
602 	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
603 	 */
604 	dx = (y1 - y) << SM_SHIFT;
605 	dsm = isc->sm1 - isc->sm2;
606 	do_div(dx, dsm);
607 	/*
608 	 * check if (x, y1) belongs to the 1st segment of rtsc.
609 	 * if so, add the offset.
610 	 */
611 	if (rtsc->x + rtsc->dx > x)
612 		dx += rtsc->x + rtsc->dx - x;
613 	dy = seg_x2y(dx, isc->sm1);
614 
615 	rtsc->x = x;
616 	rtsc->y = y;
617 	rtsc->dx = dx;
618 	rtsc->dy = dy;
619 	return;
620 }
621 
622 static void
623 init_ed(struct hfsc_class *cl, unsigned int next_len)
624 {
625 	u64 cur_time = psched_get_time();
626 
627 	/* update the deadline curve */
628 	rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
629 
630 	/*
631 	 * update the eligible curve.
632 	 * for concave, it is equal to the deadline curve.
633 	 * for convex, it is a linear curve with slope m2.
634 	 */
635 	cl->cl_eligible = cl->cl_deadline;
636 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
637 		cl->cl_eligible.dx = 0;
638 		cl->cl_eligible.dy = 0;
639 	}
640 
641 	/* compute e and d */
642 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
643 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
644 
645 	eltree_insert(cl);
646 }
647 
648 static void
649 update_ed(struct hfsc_class *cl, unsigned int next_len)
650 {
651 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
652 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
653 
654 	eltree_update(cl);
655 }
656 
657 static inline void
658 update_d(struct hfsc_class *cl, unsigned int next_len)
659 {
660 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
661 }
662 
663 static inline void
664 update_cfmin(struct hfsc_class *cl)
665 {
666 	struct rb_node *n = rb_first(&cl->cf_tree);
667 	struct hfsc_class *p;
668 
669 	if (n == NULL) {
670 		cl->cl_cfmin = 0;
671 		return;
672 	}
673 	p = rb_entry(n, struct hfsc_class, cf_node);
674 	cl->cl_cfmin = p->cl_f;
675 }
676 
677 static void
678 init_vf(struct hfsc_class *cl, unsigned int len)
679 {
680 	struct hfsc_class *max_cl;
681 	struct rb_node *n;
682 	u64 vt, f, cur_time;
683 	int go_active;
684 
685 	cur_time = 0;
686 	go_active = 1;
687 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
688 		if (go_active && cl->cl_nactive++ == 0)
689 			go_active = 1;
690 		else
691 			go_active = 0;
692 
693 		if (go_active) {
694 			n = rb_last(&cl->cl_parent->vt_tree);
695 			if (n != NULL) {
696 				max_cl = rb_entry(n, struct hfsc_class,vt_node);
697 				/*
698 				 * set vt to the average of the min and max
699 				 * classes.  if the parent's period didn't
700 				 * change, don't decrease vt of the class.
701 				 */
702 				vt = max_cl->cl_vt;
703 				if (cl->cl_parent->cl_cvtmin != 0)
704 					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
705 
706 				if (cl->cl_parent->cl_vtperiod !=
707 				    cl->cl_parentperiod || vt > cl->cl_vt)
708 					cl->cl_vt = vt;
709 			} else {
710 				/*
711 				 * first child for a new parent backlog period.
712 				 * add parent's cvtmax to cvtoff to make a new
713 				 * vt (vtoff + vt) larger than the vt in the
714 				 * last period for all children.
715 				 */
716 				vt = cl->cl_parent->cl_cvtmax;
717 				cl->cl_parent->cl_cvtoff += vt;
718 				cl->cl_parent->cl_cvtmax = 0;
719 				cl->cl_parent->cl_cvtmin = 0;
720 				cl->cl_vt = 0;
721 			}
722 
723 			cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
724 							cl->cl_pcvtoff;
725 
726 			/* update the virtual curve */
727 			vt = cl->cl_vt + cl->cl_vtoff;
728 			rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
729 						      cl->cl_total);
730 			if (cl->cl_virtual.x == vt) {
731 				cl->cl_virtual.x -= cl->cl_vtoff;
732 				cl->cl_vtoff = 0;
733 			}
734 			cl->cl_vtadj = 0;
735 
736 			cl->cl_vtperiod++;  /* increment vt period */
737 			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
738 			if (cl->cl_parent->cl_nactive == 0)
739 				cl->cl_parentperiod++;
740 			cl->cl_f = 0;
741 
742 			vttree_insert(cl);
743 			cftree_insert(cl);
744 
745 			if (cl->cl_flags & HFSC_USC) {
746 				/* class has upper limit curve */
747 				if (cur_time == 0)
748 					cur_time = psched_get_time();
749 
750 				/* update the ulimit curve */
751 				rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
752 					 cl->cl_total);
753 				/* compute myf */
754 				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
755 						      cl->cl_total);
756 				cl->cl_myfadj = 0;
757 			}
758 		}
759 
760 		f = max(cl->cl_myf, cl->cl_cfmin);
761 		if (f != cl->cl_f) {
762 			cl->cl_f = f;
763 			cftree_update(cl);
764 			update_cfmin(cl->cl_parent);
765 		}
766 	}
767 }
768 
769 static void
770 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
771 {
772 	u64 f; /* , myf_bound, delta; */
773 	int go_passive = 0;
774 
775 	if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
776 		go_passive = 1;
777 
778 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
779 		cl->cl_total += len;
780 
781 		if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
782 			continue;
783 
784 		if (go_passive && --cl->cl_nactive == 0)
785 			go_passive = 1;
786 		else
787 			go_passive = 0;
788 
789 		if (go_passive) {
790 			/* no more active child, going passive */
791 
792 			/* update cvtmax of the parent class */
793 			if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
794 				cl->cl_parent->cl_cvtmax = cl->cl_vt;
795 
796 			/* remove this class from the vt tree */
797 			vttree_remove(cl);
798 
799 			cftree_remove(cl);
800 			update_cfmin(cl->cl_parent);
801 
802 			continue;
803 		}
804 
805 		/*
806 		 * update vt and f
807 		 */
808 		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
809 			    - cl->cl_vtoff + cl->cl_vtadj;
810 
811 		/*
812 		 * if vt of the class is smaller than cvtmin,
813 		 * the class was skipped in the past due to non-fit.
814 		 * if so, we need to adjust vtadj.
815 		 */
816 		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
817 			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
818 			cl->cl_vt = cl->cl_parent->cl_cvtmin;
819 		}
820 
821 		/* update the vt tree */
822 		vttree_update(cl);
823 
824 		if (cl->cl_flags & HFSC_USC) {
825 			cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
826 							      cl->cl_total);
827 #if 0
828 			/*
829 			 * This code causes classes to stay way under their
830 			 * limit when multiple classes are used at gigabit
831 			 * speed. needs investigation. -kaber
832 			 */
833 			/*
834 			 * if myf lags behind by more than one clock tick
835 			 * from the current time, adjust myfadj to prevent
836 			 * a rate-limited class from going greedy.
837 			 * in a steady state under rate-limiting, myf
838 			 * fluctuates within one clock tick.
839 			 */
840 			myf_bound = cur_time - PSCHED_JIFFIE2US(1);
841 			if (cl->cl_myf < myf_bound) {
842 				delta = cur_time - cl->cl_myf;
843 				cl->cl_myfadj += delta;
844 				cl->cl_myf += delta;
845 			}
846 #endif
847 		}
848 
849 		f = max(cl->cl_myf, cl->cl_cfmin);
850 		if (f != cl->cl_f) {
851 			cl->cl_f = f;
852 			cftree_update(cl);
853 			update_cfmin(cl->cl_parent);
854 		}
855 	}
856 }
857 
858 static void
859 set_active(struct hfsc_class *cl, unsigned int len)
860 {
861 	if (cl->cl_flags & HFSC_RSC)
862 		init_ed(cl, len);
863 	if (cl->cl_flags & HFSC_FSC)
864 		init_vf(cl, len);
865 
866 	list_add_tail(&cl->dlist, &cl->sched->droplist);
867 }
868 
869 static void
870 set_passive(struct hfsc_class *cl)
871 {
872 	if (cl->cl_flags & HFSC_RSC)
873 		eltree_remove(cl);
874 
875 	list_del(&cl->dlist);
876 
877 	/*
878 	 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
879 	 * needs to be called explicitly to remove a class from vttree.
880 	 */
881 }
882 
883 /*
884  * hack to get length of first packet in queue.
885  */
886 static unsigned int
887 qdisc_peek_len(struct Qdisc *sch)
888 {
889 	struct sk_buff *skb;
890 	unsigned int len;
891 
892 	skb = sch->dequeue(sch);
893 	if (skb == NULL) {
894 		if (net_ratelimit())
895 			printk("qdisc_peek_len: non work-conserving qdisc ?\n");
896 		return 0;
897 	}
898 	len = qdisc_pkt_len(skb);
899 	if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
900 		if (net_ratelimit())
901 			printk("qdisc_peek_len: failed to requeue\n");
902 		qdisc_tree_decrease_qlen(sch, 1);
903 		return 0;
904 	}
905 	return len;
906 }
907 
908 static void
909 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
910 {
911 	unsigned int len = cl->qdisc->q.qlen;
912 
913 	qdisc_reset(cl->qdisc);
914 	qdisc_tree_decrease_qlen(cl->qdisc, len);
915 }
916 
917 static void
918 hfsc_adjust_levels(struct hfsc_class *cl)
919 {
920 	struct hfsc_class *p;
921 	unsigned int level;
922 
923 	do {
924 		level = 0;
925 		list_for_each_entry(p, &cl->children, siblings) {
926 			if (p->level >= level)
927 				level = p->level + 1;
928 		}
929 		cl->level = level;
930 	} while ((cl = cl->cl_parent) != NULL);
931 }
932 
933 static inline struct hfsc_class *
934 hfsc_find_class(u32 classid, struct Qdisc *sch)
935 {
936 	struct hfsc_sched *q = qdisc_priv(sch);
937 	struct Qdisc_class_common *clc;
938 
939 	clc = qdisc_class_find(&q->clhash, classid);
940 	if (clc == NULL)
941 		return NULL;
942 	return container_of(clc, struct hfsc_class, cl_common);
943 }
944 
945 static void
946 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
947 		u64 cur_time)
948 {
949 	sc2isc(rsc, &cl->cl_rsc);
950 	rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
951 	cl->cl_eligible = cl->cl_deadline;
952 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
953 		cl->cl_eligible.dx = 0;
954 		cl->cl_eligible.dy = 0;
955 	}
956 	cl->cl_flags |= HFSC_RSC;
957 }
958 
959 static void
960 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
961 {
962 	sc2isc(fsc, &cl->cl_fsc);
963 	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
964 	cl->cl_flags |= HFSC_FSC;
965 }
966 
967 static void
968 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
969 		u64 cur_time)
970 {
971 	sc2isc(usc, &cl->cl_usc);
972 	rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
973 	cl->cl_flags |= HFSC_USC;
974 }
975 
976 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
977 	[TCA_HFSC_RSC]	= { .len = sizeof(struct tc_service_curve) },
978 	[TCA_HFSC_FSC]	= { .len = sizeof(struct tc_service_curve) },
979 	[TCA_HFSC_USC]	= { .len = sizeof(struct tc_service_curve) },
980 };
981 
982 static int
983 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
984 		  struct nlattr **tca, unsigned long *arg)
985 {
986 	struct hfsc_sched *q = qdisc_priv(sch);
987 	struct hfsc_class *cl = (struct hfsc_class *)*arg;
988 	struct hfsc_class *parent = NULL;
989 	struct nlattr *opt = tca[TCA_OPTIONS];
990 	struct nlattr *tb[TCA_HFSC_MAX + 1];
991 	struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
992 	u64 cur_time;
993 	int err;
994 
995 	if (opt == NULL)
996 		return -EINVAL;
997 
998 	err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
999 	if (err < 0)
1000 		return err;
1001 
1002 	if (tb[TCA_HFSC_RSC]) {
1003 		rsc = nla_data(tb[TCA_HFSC_RSC]);
1004 		if (rsc->m1 == 0 && rsc->m2 == 0)
1005 			rsc = NULL;
1006 	}
1007 
1008 	if (tb[TCA_HFSC_FSC]) {
1009 		fsc = nla_data(tb[TCA_HFSC_FSC]);
1010 		if (fsc->m1 == 0 && fsc->m2 == 0)
1011 			fsc = NULL;
1012 	}
1013 
1014 	if (tb[TCA_HFSC_USC]) {
1015 		usc = nla_data(tb[TCA_HFSC_USC]);
1016 		if (usc->m1 == 0 && usc->m2 == 0)
1017 			usc = NULL;
1018 	}
1019 
1020 	if (cl != NULL) {
1021 		if (parentid) {
1022 			if (cl->cl_parent &&
1023 			    cl->cl_parent->cl_common.classid != parentid)
1024 				return -EINVAL;
1025 			if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1026 				return -EINVAL;
1027 		}
1028 		cur_time = psched_get_time();
1029 
1030 		sch_tree_lock(sch);
1031 		if (rsc != NULL)
1032 			hfsc_change_rsc(cl, rsc, cur_time);
1033 		if (fsc != NULL)
1034 			hfsc_change_fsc(cl, fsc);
1035 		if (usc != NULL)
1036 			hfsc_change_usc(cl, usc, cur_time);
1037 
1038 		if (cl->qdisc->q.qlen != 0) {
1039 			if (cl->cl_flags & HFSC_RSC)
1040 				update_ed(cl, qdisc_peek_len(cl->qdisc));
1041 			if (cl->cl_flags & HFSC_FSC)
1042 				update_vf(cl, 0, cur_time);
1043 		}
1044 		sch_tree_unlock(sch);
1045 
1046 		if (tca[TCA_RATE])
1047 			gen_replace_estimator(&cl->bstats, &cl->rate_est,
1048 					      qdisc_root_sleeping_lock(sch),
1049 					      tca[TCA_RATE]);
1050 		return 0;
1051 	}
1052 
1053 	if (parentid == TC_H_ROOT)
1054 		return -EEXIST;
1055 
1056 	parent = &q->root;
1057 	if (parentid) {
1058 		parent = hfsc_find_class(parentid, sch);
1059 		if (parent == NULL)
1060 			return -ENOENT;
1061 	}
1062 
1063 	if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1064 		return -EINVAL;
1065 	if (hfsc_find_class(classid, sch))
1066 		return -EEXIST;
1067 
1068 	if (rsc == NULL && fsc == NULL)
1069 		return -EINVAL;
1070 
1071 	cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1072 	if (cl == NULL)
1073 		return -ENOBUFS;
1074 
1075 	if (rsc != NULL)
1076 		hfsc_change_rsc(cl, rsc, 0);
1077 	if (fsc != NULL)
1078 		hfsc_change_fsc(cl, fsc);
1079 	if (usc != NULL)
1080 		hfsc_change_usc(cl, usc, 0);
1081 
1082 	cl->cl_common.classid = classid;
1083 	cl->refcnt    = 1;
1084 	cl->sched     = q;
1085 	cl->cl_parent = parent;
1086 	cl->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1087 				      &pfifo_qdisc_ops, classid);
1088 	if (cl->qdisc == NULL)
1089 		cl->qdisc = &noop_qdisc;
1090 	INIT_LIST_HEAD(&cl->children);
1091 	cl->vt_tree = RB_ROOT;
1092 	cl->cf_tree = RB_ROOT;
1093 
1094 	sch_tree_lock(sch);
1095 	qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1096 	list_add_tail(&cl->siblings, &parent->children);
1097 	if (parent->level == 0)
1098 		hfsc_purge_queue(sch, parent);
1099 	hfsc_adjust_levels(parent);
1100 	cl->cl_pcvtoff = parent->cl_cvtoff;
1101 	sch_tree_unlock(sch);
1102 
1103 	qdisc_class_hash_grow(sch, &q->clhash);
1104 
1105 	if (tca[TCA_RATE])
1106 		gen_new_estimator(&cl->bstats, &cl->rate_est,
1107 				  qdisc_root_sleeping_lock(sch), tca[TCA_RATE]);
1108 	*arg = (unsigned long)cl;
1109 	return 0;
1110 }
1111 
1112 static void
1113 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1114 {
1115 	struct hfsc_sched *q = qdisc_priv(sch);
1116 
1117 	tcf_destroy_chain(&cl->filter_list);
1118 	qdisc_destroy(cl->qdisc);
1119 	gen_kill_estimator(&cl->bstats, &cl->rate_est);
1120 	if (cl != &q->root)
1121 		kfree(cl);
1122 }
1123 
1124 static int
1125 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1126 {
1127 	struct hfsc_sched *q = qdisc_priv(sch);
1128 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1129 
1130 	if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1131 		return -EBUSY;
1132 
1133 	sch_tree_lock(sch);
1134 
1135 	list_del(&cl->siblings);
1136 	hfsc_adjust_levels(cl->cl_parent);
1137 
1138 	hfsc_purge_queue(sch, cl);
1139 	qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1140 
1141 	if (--cl->refcnt == 0)
1142 		hfsc_destroy_class(sch, cl);
1143 
1144 	sch_tree_unlock(sch);
1145 	return 0;
1146 }
1147 
1148 static struct hfsc_class *
1149 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1150 {
1151 	struct hfsc_sched *q = qdisc_priv(sch);
1152 	struct hfsc_class *cl;
1153 	struct tcf_result res;
1154 	struct tcf_proto *tcf;
1155 	int result;
1156 
1157 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1158 	    (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1159 		if (cl->level == 0)
1160 			return cl;
1161 
1162 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1163 	tcf = q->root.filter_list;
1164 	while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1165 #ifdef CONFIG_NET_CLS_ACT
1166 		switch (result) {
1167 		case TC_ACT_QUEUED:
1168 		case TC_ACT_STOLEN:
1169 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1170 		case TC_ACT_SHOT:
1171 			return NULL;
1172 		}
1173 #endif
1174 		if ((cl = (struct hfsc_class *)res.class) == NULL) {
1175 			if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1176 				break; /* filter selected invalid classid */
1177 		}
1178 
1179 		if (cl->level == 0)
1180 			return cl; /* hit leaf class */
1181 
1182 		/* apply inner filter chain */
1183 		tcf = cl->filter_list;
1184 	}
1185 
1186 	/* classification failed, try default class */
1187 	cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1188 	if (cl == NULL || cl->level > 0)
1189 		return NULL;
1190 
1191 	return cl;
1192 }
1193 
1194 static int
1195 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1196 		 struct Qdisc **old)
1197 {
1198 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1199 
1200 	if (cl == NULL)
1201 		return -ENOENT;
1202 	if (cl->level > 0)
1203 		return -EINVAL;
1204 	if (new == NULL) {
1205 		new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1206 					&pfifo_qdisc_ops,
1207 					cl->cl_common.classid);
1208 		if (new == NULL)
1209 			new = &noop_qdisc;
1210 	}
1211 
1212 	sch_tree_lock(sch);
1213 	hfsc_purge_queue(sch, cl);
1214 	*old = xchg(&cl->qdisc, new);
1215 	sch_tree_unlock(sch);
1216 	return 0;
1217 }
1218 
1219 static struct Qdisc *
1220 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1221 {
1222 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1223 
1224 	if (cl != NULL && cl->level == 0)
1225 		return cl->qdisc;
1226 
1227 	return NULL;
1228 }
1229 
1230 static void
1231 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1232 {
1233 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1234 
1235 	if (cl->qdisc->q.qlen == 0) {
1236 		update_vf(cl, 0, 0);
1237 		set_passive(cl);
1238 	}
1239 }
1240 
1241 static unsigned long
1242 hfsc_get_class(struct Qdisc *sch, u32 classid)
1243 {
1244 	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1245 
1246 	if (cl != NULL)
1247 		cl->refcnt++;
1248 
1249 	return (unsigned long)cl;
1250 }
1251 
1252 static void
1253 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1254 {
1255 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1256 
1257 	if (--cl->refcnt == 0)
1258 		hfsc_destroy_class(sch, cl);
1259 }
1260 
1261 static unsigned long
1262 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1263 {
1264 	struct hfsc_class *p = (struct hfsc_class *)parent;
1265 	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1266 
1267 	if (cl != NULL) {
1268 		if (p != NULL && p->level <= cl->level)
1269 			return 0;
1270 		cl->filter_cnt++;
1271 	}
1272 
1273 	return (unsigned long)cl;
1274 }
1275 
1276 static void
1277 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1278 {
1279 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1280 
1281 	cl->filter_cnt--;
1282 }
1283 
1284 static struct tcf_proto **
1285 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1286 {
1287 	struct hfsc_sched *q = qdisc_priv(sch);
1288 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1289 
1290 	if (cl == NULL)
1291 		cl = &q->root;
1292 
1293 	return &cl->filter_list;
1294 }
1295 
1296 static int
1297 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1298 {
1299 	struct tc_service_curve tsc;
1300 
1301 	tsc.m1 = sm2m(sc->sm1);
1302 	tsc.d  = dx2d(sc->dx);
1303 	tsc.m2 = sm2m(sc->sm2);
1304 	NLA_PUT(skb, attr, sizeof(tsc), &tsc);
1305 
1306 	return skb->len;
1307 
1308  nla_put_failure:
1309 	return -1;
1310 }
1311 
1312 static inline int
1313 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1314 {
1315 	if ((cl->cl_flags & HFSC_RSC) &&
1316 	    (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1317 		goto nla_put_failure;
1318 
1319 	if ((cl->cl_flags & HFSC_FSC) &&
1320 	    (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1321 		goto nla_put_failure;
1322 
1323 	if ((cl->cl_flags & HFSC_USC) &&
1324 	    (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1325 		goto nla_put_failure;
1326 
1327 	return skb->len;
1328 
1329  nla_put_failure:
1330 	return -1;
1331 }
1332 
1333 static int
1334 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1335 		struct tcmsg *tcm)
1336 {
1337 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1338 	struct nlattr *nest;
1339 
1340 	tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1341 					  TC_H_ROOT;
1342 	tcm->tcm_handle = cl->cl_common.classid;
1343 	if (cl->level == 0)
1344 		tcm->tcm_info = cl->qdisc->handle;
1345 
1346 	nest = nla_nest_start(skb, TCA_OPTIONS);
1347 	if (nest == NULL)
1348 		goto nla_put_failure;
1349 	if (hfsc_dump_curves(skb, cl) < 0)
1350 		goto nla_put_failure;
1351 	nla_nest_end(skb, nest);
1352 	return skb->len;
1353 
1354  nla_put_failure:
1355 	nla_nest_cancel(skb, nest);
1356 	return -EMSGSIZE;
1357 }
1358 
1359 static int
1360 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1361 	struct gnet_dump *d)
1362 {
1363 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1364 	struct tc_hfsc_stats xstats;
1365 
1366 	cl->qstats.qlen = cl->qdisc->q.qlen;
1367 	xstats.level   = cl->level;
1368 	xstats.period  = cl->cl_vtperiod;
1369 	xstats.work    = cl->cl_total;
1370 	xstats.rtwork  = cl->cl_cumul;
1371 
1372 	if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1373 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1374 	    gnet_stats_copy_queue(d, &cl->qstats) < 0)
1375 		return -1;
1376 
1377 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1378 }
1379 
1380 
1381 
1382 static void
1383 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1384 {
1385 	struct hfsc_sched *q = qdisc_priv(sch);
1386 	struct hlist_node *n;
1387 	struct hfsc_class *cl;
1388 	unsigned int i;
1389 
1390 	if (arg->stop)
1391 		return;
1392 
1393 	for (i = 0; i < q->clhash.hashsize; i++) {
1394 		hlist_for_each_entry(cl, n, &q->clhash.hash[i],
1395 				     cl_common.hnode) {
1396 			if (arg->count < arg->skip) {
1397 				arg->count++;
1398 				continue;
1399 			}
1400 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1401 				arg->stop = 1;
1402 				return;
1403 			}
1404 			arg->count++;
1405 		}
1406 	}
1407 }
1408 
1409 static void
1410 hfsc_schedule_watchdog(struct Qdisc *sch)
1411 {
1412 	struct hfsc_sched *q = qdisc_priv(sch);
1413 	struct hfsc_class *cl;
1414 	u64 next_time = 0;
1415 
1416 	if ((cl = eltree_get_minel(q)) != NULL)
1417 		next_time = cl->cl_e;
1418 	if (q->root.cl_cfmin != 0) {
1419 		if (next_time == 0 || next_time > q->root.cl_cfmin)
1420 			next_time = q->root.cl_cfmin;
1421 	}
1422 	WARN_ON(next_time == 0);
1423 	qdisc_watchdog_schedule(&q->watchdog, next_time);
1424 }
1425 
1426 static int
1427 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1428 {
1429 	struct hfsc_sched *q = qdisc_priv(sch);
1430 	struct tc_hfsc_qopt *qopt;
1431 	int err;
1432 
1433 	if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1434 		return -EINVAL;
1435 	qopt = nla_data(opt);
1436 
1437 	q->defcls = qopt->defcls;
1438 	err = qdisc_class_hash_init(&q->clhash);
1439 	if (err < 0)
1440 		return err;
1441 	q->eligible = RB_ROOT;
1442 	INIT_LIST_HEAD(&q->droplist);
1443 	skb_queue_head_init(&q->requeue);
1444 
1445 	q->root.cl_common.classid = sch->handle;
1446 	q->root.refcnt  = 1;
1447 	q->root.sched   = q;
1448 	q->root.qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1449 					  &pfifo_qdisc_ops,
1450 					  sch->handle);
1451 	if (q->root.qdisc == NULL)
1452 		q->root.qdisc = &noop_qdisc;
1453 	INIT_LIST_HEAD(&q->root.children);
1454 	q->root.vt_tree = RB_ROOT;
1455 	q->root.cf_tree = RB_ROOT;
1456 
1457 	qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1458 	qdisc_class_hash_grow(sch, &q->clhash);
1459 
1460 	qdisc_watchdog_init(&q->watchdog, sch);
1461 
1462 	return 0;
1463 }
1464 
1465 static int
1466 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1467 {
1468 	struct hfsc_sched *q = qdisc_priv(sch);
1469 	struct tc_hfsc_qopt *qopt;
1470 
1471 	if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1472 		return -EINVAL;
1473 	qopt = nla_data(opt);
1474 
1475 	sch_tree_lock(sch);
1476 	q->defcls = qopt->defcls;
1477 	sch_tree_unlock(sch);
1478 
1479 	return 0;
1480 }
1481 
1482 static void
1483 hfsc_reset_class(struct hfsc_class *cl)
1484 {
1485 	cl->cl_total        = 0;
1486 	cl->cl_cumul        = 0;
1487 	cl->cl_d            = 0;
1488 	cl->cl_e            = 0;
1489 	cl->cl_vt           = 0;
1490 	cl->cl_vtadj        = 0;
1491 	cl->cl_vtoff        = 0;
1492 	cl->cl_cvtmin       = 0;
1493 	cl->cl_cvtmax       = 0;
1494 	cl->cl_cvtoff       = 0;
1495 	cl->cl_pcvtoff      = 0;
1496 	cl->cl_vtperiod     = 0;
1497 	cl->cl_parentperiod = 0;
1498 	cl->cl_f            = 0;
1499 	cl->cl_myf          = 0;
1500 	cl->cl_myfadj       = 0;
1501 	cl->cl_cfmin        = 0;
1502 	cl->cl_nactive      = 0;
1503 
1504 	cl->vt_tree = RB_ROOT;
1505 	cl->cf_tree = RB_ROOT;
1506 	qdisc_reset(cl->qdisc);
1507 
1508 	if (cl->cl_flags & HFSC_RSC)
1509 		rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1510 	if (cl->cl_flags & HFSC_FSC)
1511 		rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1512 	if (cl->cl_flags & HFSC_USC)
1513 		rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1514 }
1515 
1516 static void
1517 hfsc_reset_qdisc(struct Qdisc *sch)
1518 {
1519 	struct hfsc_sched *q = qdisc_priv(sch);
1520 	struct hfsc_class *cl;
1521 	struct hlist_node *n;
1522 	unsigned int i;
1523 
1524 	for (i = 0; i < q->clhash.hashsize; i++) {
1525 		hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1526 			hfsc_reset_class(cl);
1527 	}
1528 	__skb_queue_purge(&q->requeue);
1529 	q->eligible = RB_ROOT;
1530 	INIT_LIST_HEAD(&q->droplist);
1531 	qdisc_watchdog_cancel(&q->watchdog);
1532 	sch->q.qlen = 0;
1533 }
1534 
1535 static void
1536 hfsc_destroy_qdisc(struct Qdisc *sch)
1537 {
1538 	struct hfsc_sched *q = qdisc_priv(sch);
1539 	struct hlist_node *n, *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, n, &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, n, next, &q->clhash.hash[i],
1549 					  cl_common.hnode)
1550 			hfsc_destroy_class(sch, cl);
1551 	}
1552 	qdisc_class_hash_destroy(&q->clhash);
1553 	__skb_queue_purge(&q->requeue);
1554 	qdisc_watchdog_cancel(&q->watchdog);
1555 }
1556 
1557 static int
1558 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1559 {
1560 	struct hfsc_sched *q = qdisc_priv(sch);
1561 	unsigned char *b = skb_tail_pointer(skb);
1562 	struct tc_hfsc_qopt qopt;
1563 
1564 	qopt.defcls = q->defcls;
1565 	NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1566 	return skb->len;
1567 
1568  nla_put_failure:
1569 	nlmsg_trim(skb, b);
1570 	return -1;
1571 }
1572 
1573 static int
1574 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1575 {
1576 	struct hfsc_class *cl;
1577 	int err;
1578 
1579 	cl = hfsc_classify(skb, sch, &err);
1580 	if (cl == NULL) {
1581 		if (err & __NET_XMIT_BYPASS)
1582 			sch->qstats.drops++;
1583 		kfree_skb(skb);
1584 		return err;
1585 	}
1586 
1587 	err = qdisc_enqueue(skb, cl->qdisc);
1588 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1589 		if (net_xmit_drop_count(err)) {
1590 			cl->qstats.drops++;
1591 			sch->qstats.drops++;
1592 		}
1593 		return err;
1594 	}
1595 
1596 	if (cl->qdisc->q.qlen == 1)
1597 		set_active(cl, qdisc_pkt_len(skb));
1598 
1599 	cl->bstats.packets++;
1600 	cl->bstats.bytes += qdisc_pkt_len(skb);
1601 	sch->bstats.packets++;
1602 	sch->bstats.bytes += qdisc_pkt_len(skb);
1603 	sch->q.qlen++;
1604 
1605 	return NET_XMIT_SUCCESS;
1606 }
1607 
1608 static struct sk_buff *
1609 hfsc_dequeue(struct Qdisc *sch)
1610 {
1611 	struct hfsc_sched *q = qdisc_priv(sch);
1612 	struct hfsc_class *cl;
1613 	struct sk_buff *skb;
1614 	u64 cur_time;
1615 	unsigned int next_len;
1616 	int realtime = 0;
1617 
1618 	if (sch->q.qlen == 0)
1619 		return NULL;
1620 	if ((skb = __skb_dequeue(&q->requeue)))
1621 		goto out;
1622 
1623 	cur_time = psched_get_time();
1624 
1625 	/*
1626 	 * if there are eligible classes, use real-time criteria.
1627 	 * find the class with the minimum deadline among
1628 	 * the eligible classes.
1629 	 */
1630 	if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1631 		realtime = 1;
1632 	} else {
1633 		/*
1634 		 * use link-sharing criteria
1635 		 * get the class with the minimum vt in the hierarchy
1636 		 */
1637 		cl = vttree_get_minvt(&q->root, cur_time);
1638 		if (cl == NULL) {
1639 			sch->qstats.overlimits++;
1640 			hfsc_schedule_watchdog(sch);
1641 			return NULL;
1642 		}
1643 	}
1644 
1645 	skb = cl->qdisc->dequeue(cl->qdisc);
1646 	if (skb == NULL) {
1647 		if (net_ratelimit())
1648 			printk("HFSC: Non-work-conserving qdisc ?\n");
1649 		return NULL;
1650 	}
1651 
1652 	update_vf(cl, qdisc_pkt_len(skb), cur_time);
1653 	if (realtime)
1654 		cl->cl_cumul += qdisc_pkt_len(skb);
1655 
1656 	if (cl->qdisc->q.qlen != 0) {
1657 		if (cl->cl_flags & HFSC_RSC) {
1658 			/* update ed */
1659 			next_len = qdisc_peek_len(cl->qdisc);
1660 			if (realtime)
1661 				update_ed(cl, next_len);
1662 			else
1663 				update_d(cl, next_len);
1664 		}
1665 	} else {
1666 		/* the class becomes passive */
1667 		set_passive(cl);
1668 	}
1669 
1670  out:
1671 	sch->flags &= ~TCQ_F_THROTTLED;
1672 	sch->q.qlen--;
1673 
1674 	return skb;
1675 }
1676 
1677 static int
1678 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1679 {
1680 	struct hfsc_sched *q = qdisc_priv(sch);
1681 
1682 	__skb_queue_head(&q->requeue, skb);
1683 	sch->q.qlen++;
1684 	sch->qstats.requeues++;
1685 	return NET_XMIT_SUCCESS;
1686 }
1687 
1688 static unsigned int
1689 hfsc_drop(struct Qdisc *sch)
1690 {
1691 	struct hfsc_sched *q = qdisc_priv(sch);
1692 	struct hfsc_class *cl;
1693 	unsigned int len;
1694 
1695 	list_for_each_entry(cl, &q->droplist, dlist) {
1696 		if (cl->qdisc->ops->drop != NULL &&
1697 		    (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1698 			if (cl->qdisc->q.qlen == 0) {
1699 				update_vf(cl, 0, 0);
1700 				set_passive(cl);
1701 			} else {
1702 				list_move_tail(&cl->dlist, &q->droplist);
1703 			}
1704 			cl->qstats.drops++;
1705 			sch->qstats.drops++;
1706 			sch->q.qlen--;
1707 			return len;
1708 		}
1709 	}
1710 	return 0;
1711 }
1712 
1713 static const struct Qdisc_class_ops hfsc_class_ops = {
1714 	.change		= hfsc_change_class,
1715 	.delete		= hfsc_delete_class,
1716 	.graft		= hfsc_graft_class,
1717 	.leaf		= hfsc_class_leaf,
1718 	.qlen_notify	= hfsc_qlen_notify,
1719 	.get		= hfsc_get_class,
1720 	.put		= hfsc_put_class,
1721 	.bind_tcf	= hfsc_bind_tcf,
1722 	.unbind_tcf	= hfsc_unbind_tcf,
1723 	.tcf_chain	= hfsc_tcf_chain,
1724 	.dump		= hfsc_dump_class,
1725 	.dump_stats	= hfsc_dump_class_stats,
1726 	.walk		= hfsc_walk
1727 };
1728 
1729 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1730 	.id		= "hfsc",
1731 	.init		= hfsc_init_qdisc,
1732 	.change		= hfsc_change_qdisc,
1733 	.reset		= hfsc_reset_qdisc,
1734 	.destroy	= hfsc_destroy_qdisc,
1735 	.dump		= hfsc_dump_qdisc,
1736 	.enqueue	= hfsc_enqueue,
1737 	.dequeue	= hfsc_dequeue,
1738 	.requeue	= hfsc_requeue,
1739 	.drop		= hfsc_drop,
1740 	.cl_ops		= &hfsc_class_ops,
1741 	.priv_size	= sizeof(struct hfsc_sched),
1742 	.owner		= THIS_MODULE
1743 };
1744 
1745 static int __init
1746 hfsc_init(void)
1747 {
1748 	return register_qdisc(&hfsc_qdisc_ops);
1749 }
1750 
1751 static void __exit
1752 hfsc_cleanup(void)
1753 {
1754 	unregister_qdisc(&hfsc_qdisc_ops);
1755 }
1756 
1757 MODULE_LICENSE("GPL");
1758 module_init(hfsc_init);
1759 module_exit(hfsc_cleanup);
1760