xref: /linux/net/sched/sch_hfsc.c (revision afca12e35e711ae8f97e835a3704cc305592eac9)
1 /*
2  * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation; either version 2
7  * of the License, or (at your option) any later version.
8  *
9  * 2003-10-17 - Ported from altq
10  */
11 /*
12  * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
13  *
14  * Permission to use, copy, modify, and distribute this software and
15  * its documentation is hereby granted (including for commercial or
16  * for-profit use), provided that both the copyright notice and this
17  * permission notice appear in all copies of the software, derivative
18  * works, or modified versions, and any portions thereof.
19  *
20  * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
21  * WHICH MAY HAVE SERIOUS CONSEQUENCES.  CARNEGIE MELLON PROVIDES THIS
22  * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
23  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
25  * DISCLAIMED.  IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
28  * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
29  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
30  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
32  * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
33  * DAMAGE.
34  *
35  * Carnegie Mellon encourages (but does not require) users of this
36  * software to return any improvements or extensions that they make,
37  * and to grant Carnegie Mellon the rights to redistribute these
38  * changes without encumbrance.
39  */
40 /*
41  * H-FSC is described in Proceedings of SIGCOMM'97,
42  * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
43  * Real-Time and Priority Service"
44  * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
45  *
46  * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
47  * when a class has an upperlimit, the fit-time is computed from the
48  * upperlimit service curve.  the link-sharing scheduler does not schedule
49  * a class whose fit-time exceeds the current time.
50  */
51 
52 #include <linux/kernel.h>
53 #include <linux/module.h>
54 #include <linux/types.h>
55 #include <linux/errno.h>
56 #include <linux/compiler.h>
57 #include <linux/spinlock.h>
58 #include <linux/skbuff.h>
59 #include <linux/string.h>
60 #include <linux/slab.h>
61 #include <linux/list.h>
62 #include <linux/rbtree.h>
63 #include <linux/init.h>
64 #include <linux/rtnetlink.h>
65 #include <linux/pkt_sched.h>
66 #include <net/netlink.h>
67 #include <net/pkt_sched.h>
68 #include <net/pkt_cls.h>
69 #include <asm/div64.h>
70 
71 /*
72  * kernel internal service curve representation:
73  *   coordinates are given by 64 bit unsigned integers.
74  *   x-axis: unit is clock count.
75  *   y-axis: unit is byte.
76  *
77  *   The service curve parameters are converted to the internal
78  *   representation. The slope values are scaled to avoid overflow.
79  *   the inverse slope values as well as the y-projection of the 1st
80  *   segment are kept in order to avoid 64-bit divide operations
81  *   that are expensive on 32-bit architectures.
82  */
83 
84 struct internal_sc {
85 	u64	sm1;	/* scaled slope of the 1st segment */
86 	u64	ism1;	/* scaled inverse-slope of the 1st segment */
87 	u64	dx;	/* the x-projection of the 1st segment */
88 	u64	dy;	/* the y-projection of the 1st segment */
89 	u64	sm2;	/* scaled slope of the 2nd segment */
90 	u64	ism2;	/* scaled inverse-slope of the 2nd segment */
91 };
92 
93 /* runtime service curve */
94 struct runtime_sc {
95 	u64	x;	/* current starting position on x-axis */
96 	u64	y;	/* current starting position on y-axis */
97 	u64	sm1;	/* scaled slope of the 1st segment */
98 	u64	ism1;	/* scaled inverse-slope of the 1st segment */
99 	u64	dx;	/* the x-projection of the 1st segment */
100 	u64	dy;	/* the y-projection of the 1st segment */
101 	u64	sm2;	/* scaled slope of the 2nd segment */
102 	u64	ism2;	/* scaled inverse-slope of the 2nd segment */
103 };
104 
105 enum hfsc_class_flags {
106 	HFSC_RSC = 0x1,
107 	HFSC_FSC = 0x2,
108 	HFSC_USC = 0x4
109 };
110 
111 struct hfsc_class {
112 	struct Qdisc_class_common cl_common;
113 
114 	struct gnet_stats_basic_sync bstats;
115 	struct gnet_stats_queue qstats;
116 	struct net_rate_estimator __rcu *rate_est;
117 	struct tcf_proto __rcu *filter_list; /* filter list */
118 	struct tcf_block *block;
119 	unsigned int	level;		/* class level in hierarchy */
120 
121 	struct hfsc_sched *sched;	/* scheduler data */
122 	struct hfsc_class *cl_parent;	/* parent class */
123 	struct list_head siblings;	/* sibling classes */
124 	struct list_head children;	/* child classes */
125 	struct Qdisc	*qdisc;		/* leaf qdisc */
126 
127 	struct rb_node el_node;		/* qdisc's eligible tree member */
128 	struct rb_root vt_tree;		/* active children sorted by cl_vt */
129 	struct rb_node vt_node;		/* parent's vt_tree member */
130 	struct rb_root cf_tree;		/* active children sorted by cl_f */
131 	struct rb_node cf_node;		/* parent's cf_heap member */
132 
133 	u64	cl_total;		/* total work in bytes */
134 	u64	cl_cumul;		/* cumulative work in bytes done by
135 					   real-time criteria */
136 
137 	u64	cl_d;			/* deadline*/
138 	u64	cl_e;			/* eligible time */
139 	u64	cl_vt;			/* virtual time */
140 	u64	cl_f;			/* time when this class will fit for
141 					   link-sharing, max(myf, cfmin) */
142 	u64	cl_myf;			/* my fit-time (calculated from this
143 					   class's own upperlimit curve) */
144 	u64	cl_cfmin;		/* earliest children's fit-time (used
145 					   with cl_myf to obtain cl_f) */
146 	u64	cl_cvtmin;		/* minimal virtual time among the
147 					   children fit for link-sharing
148 					   (monotonic within a period) */
149 	u64	cl_vtadj;		/* intra-period cumulative vt
150 					   adjustment */
151 	u64	cl_cvtoff;		/* largest virtual time seen among
152 					   the children */
153 
154 	struct internal_sc cl_rsc;	/* internal real-time service curve */
155 	struct internal_sc cl_fsc;	/* internal fair service curve */
156 	struct internal_sc cl_usc;	/* internal upperlimit service curve */
157 	struct runtime_sc cl_deadline;	/* deadline curve */
158 	struct runtime_sc cl_eligible;	/* eligible curve */
159 	struct runtime_sc cl_virtual;	/* virtual curve */
160 	struct runtime_sc cl_ulimit;	/* upperlimit curve */
161 
162 	u8		cl_flags;	/* which curves are valid */
163 	u32		cl_vtperiod;	/* vt period sequence number */
164 	u32		cl_parentperiod;/* parent's vt period sequence number*/
165 	u32		cl_nactive;	/* number of active children */
166 };
167 
168 struct hfsc_sched {
169 	u16	defcls;				/* default class id */
170 	struct hfsc_class root;			/* root class */
171 	struct Qdisc_class_hash clhash;		/* class hash */
172 	struct rb_root eligible;		/* eligible tree */
173 	struct qdisc_watchdog watchdog;		/* watchdog timer */
174 };
175 
176 #define	HT_INFINITY	0xffffffffffffffffULL	/* infinite time value */
177 
178 
179 /*
180  * eligible tree holds backlogged classes being sorted by their eligible times.
181  * there is one eligible tree per hfsc instance.
182  */
183 
184 static void
185 eltree_insert(struct hfsc_class *cl)
186 {
187 	struct rb_node **p = &cl->sched->eligible.rb_node;
188 	struct rb_node *parent = NULL;
189 	struct hfsc_class *cl1;
190 
191 	while (*p != NULL) {
192 		parent = *p;
193 		cl1 = rb_entry(parent, struct hfsc_class, el_node);
194 		if (cl->cl_e >= cl1->cl_e)
195 			p = &parent->rb_right;
196 		else
197 			p = &parent->rb_left;
198 	}
199 	rb_link_node(&cl->el_node, parent, p);
200 	rb_insert_color(&cl->el_node, &cl->sched->eligible);
201 }
202 
203 static inline void
204 eltree_remove(struct hfsc_class *cl)
205 {
206 	rb_erase(&cl->el_node, &cl->sched->eligible);
207 }
208 
209 static inline void
210 eltree_update(struct hfsc_class *cl)
211 {
212 	eltree_remove(cl);
213 	eltree_insert(cl);
214 }
215 
216 /* find the class with the minimum deadline among the eligible classes */
217 static inline struct hfsc_class *
218 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
219 {
220 	struct hfsc_class *p, *cl = NULL;
221 	struct rb_node *n;
222 
223 	for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
224 		p = rb_entry(n, struct hfsc_class, el_node);
225 		if (p->cl_e > cur_time)
226 			break;
227 		if (cl == NULL || p->cl_d < cl->cl_d)
228 			cl = p;
229 	}
230 	return cl;
231 }
232 
233 /* find the class with minimum eligible time among the eligible classes */
234 static inline struct hfsc_class *
235 eltree_get_minel(struct hfsc_sched *q)
236 {
237 	struct rb_node *n;
238 
239 	n = rb_first(&q->eligible);
240 	if (n == NULL)
241 		return NULL;
242 	return rb_entry(n, struct hfsc_class, el_node);
243 }
244 
245 /*
246  * vttree holds holds backlogged child classes being sorted by their virtual
247  * time. each intermediate class has one vttree.
248  */
249 static void
250 vttree_insert(struct hfsc_class *cl)
251 {
252 	struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
253 	struct rb_node *parent = NULL;
254 	struct hfsc_class *cl1;
255 
256 	while (*p != NULL) {
257 		parent = *p;
258 		cl1 = rb_entry(parent, struct hfsc_class, vt_node);
259 		if (cl->cl_vt >= cl1->cl_vt)
260 			p = &parent->rb_right;
261 		else
262 			p = &parent->rb_left;
263 	}
264 	rb_link_node(&cl->vt_node, parent, p);
265 	rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
266 }
267 
268 static inline void
269 vttree_remove(struct hfsc_class *cl)
270 {
271 	rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
272 }
273 
274 static inline void
275 vttree_update(struct hfsc_class *cl)
276 {
277 	vttree_remove(cl);
278 	vttree_insert(cl);
279 }
280 
281 static inline struct hfsc_class *
282 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
283 {
284 	struct hfsc_class *p;
285 	struct rb_node *n;
286 
287 	for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
288 		p = rb_entry(n, struct hfsc_class, vt_node);
289 		if (p->cl_f <= cur_time)
290 			return p;
291 	}
292 	return NULL;
293 }
294 
295 /*
296  * get the leaf class with the minimum vt in the hierarchy
297  */
298 static struct hfsc_class *
299 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
300 {
301 	/* if root-class's cfmin is bigger than cur_time nothing to do */
302 	if (cl->cl_cfmin > cur_time)
303 		return NULL;
304 
305 	while (cl->level > 0) {
306 		cl = vttree_firstfit(cl, cur_time);
307 		if (cl == NULL)
308 			return NULL;
309 		/*
310 		 * update parent's cl_cvtmin.
311 		 */
312 		if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
313 			cl->cl_parent->cl_cvtmin = cl->cl_vt;
314 	}
315 	return cl;
316 }
317 
318 static void
319 cftree_insert(struct hfsc_class *cl)
320 {
321 	struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
322 	struct rb_node *parent = NULL;
323 	struct hfsc_class *cl1;
324 
325 	while (*p != NULL) {
326 		parent = *p;
327 		cl1 = rb_entry(parent, struct hfsc_class, cf_node);
328 		if (cl->cl_f >= cl1->cl_f)
329 			p = &parent->rb_right;
330 		else
331 			p = &parent->rb_left;
332 	}
333 	rb_link_node(&cl->cf_node, parent, p);
334 	rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
335 }
336 
337 static inline void
338 cftree_remove(struct hfsc_class *cl)
339 {
340 	rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
341 }
342 
343 static inline void
344 cftree_update(struct hfsc_class *cl)
345 {
346 	cftree_remove(cl);
347 	cftree_insert(cl);
348 }
349 
350 /*
351  * service curve support functions
352  *
353  *  external service curve parameters
354  *	m: bps
355  *	d: us
356  *  internal service curve parameters
357  *	sm: (bytes/psched_us) << SM_SHIFT
358  *	ism: (psched_us/byte) << ISM_SHIFT
359  *	dx: psched_us
360  *
361  * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
362  *
363  * sm and ism are scaled in order to keep effective digits.
364  * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
365  * digits in decimal using the following table.
366  *
367  *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
368  *  ------------+-------------------------------------------------------
369  *  bytes/1.024us 12.8e-3    128e-3     1280e-3    12800e-3   128000e-3
370  *
371  *  1.024us/byte  78.125     7.8125     0.78125    0.078125   0.0078125
372  *
373  * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
374  */
375 #define	SM_SHIFT	(30 - PSCHED_SHIFT)
376 #define	ISM_SHIFT	(8 + PSCHED_SHIFT)
377 
378 #define	SM_MASK		((1ULL << SM_SHIFT) - 1)
379 #define	ISM_MASK	((1ULL << ISM_SHIFT) - 1)
380 
381 static inline u64
382 seg_x2y(u64 x, u64 sm)
383 {
384 	u64 y;
385 
386 	/*
387 	 * compute
388 	 *	y = x * sm >> SM_SHIFT
389 	 * but divide it for the upper and lower bits to avoid overflow
390 	 */
391 	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
392 	return y;
393 }
394 
395 static inline u64
396 seg_y2x(u64 y, u64 ism)
397 {
398 	u64 x;
399 
400 	if (y == 0)
401 		x = 0;
402 	else if (ism == HT_INFINITY)
403 		x = HT_INFINITY;
404 	else {
405 		x = (y >> ISM_SHIFT) * ism
406 		    + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
407 	}
408 	return x;
409 }
410 
411 /* Convert m (bps) into sm (bytes/psched us) */
412 static u64
413 m2sm(u32 m)
414 {
415 	u64 sm;
416 
417 	sm = ((u64)m << SM_SHIFT);
418 	sm += PSCHED_TICKS_PER_SEC - 1;
419 	do_div(sm, PSCHED_TICKS_PER_SEC);
420 	return sm;
421 }
422 
423 /* convert m (bps) into ism (psched us/byte) */
424 static u64
425 m2ism(u32 m)
426 {
427 	u64 ism;
428 
429 	if (m == 0)
430 		ism = HT_INFINITY;
431 	else {
432 		ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
433 		ism += m - 1;
434 		do_div(ism, m);
435 	}
436 	return ism;
437 }
438 
439 /* convert d (us) into dx (psched us) */
440 static u64
441 d2dx(u32 d)
442 {
443 	u64 dx;
444 
445 	dx = ((u64)d * PSCHED_TICKS_PER_SEC);
446 	dx += USEC_PER_SEC - 1;
447 	do_div(dx, USEC_PER_SEC);
448 	return dx;
449 }
450 
451 /* convert sm (bytes/psched us) into m (bps) */
452 static u32
453 sm2m(u64 sm)
454 {
455 	u64 m;
456 
457 	m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
458 	return (u32)m;
459 }
460 
461 /* convert dx (psched us) into d (us) */
462 static u32
463 dx2d(u64 dx)
464 {
465 	u64 d;
466 
467 	d = dx * USEC_PER_SEC;
468 	do_div(d, PSCHED_TICKS_PER_SEC);
469 	return (u32)d;
470 }
471 
472 static void
473 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
474 {
475 	isc->sm1  = m2sm(sc->m1);
476 	isc->ism1 = m2ism(sc->m1);
477 	isc->dx   = d2dx(sc->d);
478 	isc->dy   = seg_x2y(isc->dx, isc->sm1);
479 	isc->sm2  = m2sm(sc->m2);
480 	isc->ism2 = m2ism(sc->m2);
481 }
482 
483 /*
484  * initialize the runtime service curve with the given internal
485  * service curve starting at (x, y).
486  */
487 static void
488 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
489 {
490 	rtsc->x	   = x;
491 	rtsc->y    = y;
492 	rtsc->sm1  = isc->sm1;
493 	rtsc->ism1 = isc->ism1;
494 	rtsc->dx   = isc->dx;
495 	rtsc->dy   = isc->dy;
496 	rtsc->sm2  = isc->sm2;
497 	rtsc->ism2 = isc->ism2;
498 }
499 
500 /*
501  * calculate the y-projection of the runtime service curve by the
502  * given x-projection value
503  */
504 static u64
505 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
506 {
507 	u64 x;
508 
509 	if (y < rtsc->y)
510 		x = rtsc->x;
511 	else if (y <= rtsc->y + rtsc->dy) {
512 		/* x belongs to the 1st segment */
513 		if (rtsc->dy == 0)
514 			x = rtsc->x + rtsc->dx;
515 		else
516 			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
517 	} else {
518 		/* x belongs to the 2nd segment */
519 		x = rtsc->x + rtsc->dx
520 		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
521 	}
522 	return x;
523 }
524 
525 static u64
526 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
527 {
528 	u64 y;
529 
530 	if (x <= rtsc->x)
531 		y = rtsc->y;
532 	else if (x <= rtsc->x + rtsc->dx)
533 		/* y belongs to the 1st segment */
534 		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
535 	else
536 		/* y belongs to the 2nd segment */
537 		y = rtsc->y + rtsc->dy
538 		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
539 	return y;
540 }
541 
542 /*
543  * update the runtime service curve by taking the minimum of the current
544  * runtime service curve and the service curve starting at (x, y).
545  */
546 static void
547 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
548 {
549 	u64 y1, y2, dx, dy;
550 	u32 dsm;
551 
552 	if (isc->sm1 <= isc->sm2) {
553 		/* service curve is convex */
554 		y1 = rtsc_x2y(rtsc, x);
555 		if (y1 < y)
556 			/* the current rtsc is smaller */
557 			return;
558 		rtsc->x = x;
559 		rtsc->y = y;
560 		return;
561 	}
562 
563 	/*
564 	 * service curve is concave
565 	 * compute the two y values of the current rtsc
566 	 *	y1: at x
567 	 *	y2: at (x + dx)
568 	 */
569 	y1 = rtsc_x2y(rtsc, x);
570 	if (y1 <= y) {
571 		/* rtsc is below isc, no change to rtsc */
572 		return;
573 	}
574 
575 	y2 = rtsc_x2y(rtsc, x + isc->dx);
576 	if (y2 >= y + isc->dy) {
577 		/* rtsc is above isc, replace rtsc by isc */
578 		rtsc->x = x;
579 		rtsc->y = y;
580 		rtsc->dx = isc->dx;
581 		rtsc->dy = isc->dy;
582 		return;
583 	}
584 
585 	/*
586 	 * the two curves intersect
587 	 * compute the offsets (dx, dy) using the reverse
588 	 * function of seg_x2y()
589 	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
590 	 */
591 	dx = (y1 - y) << SM_SHIFT;
592 	dsm = isc->sm1 - isc->sm2;
593 	do_div(dx, dsm);
594 	/*
595 	 * check if (x, y1) belongs to the 1st segment of rtsc.
596 	 * if so, add the offset.
597 	 */
598 	if (rtsc->x + rtsc->dx > x)
599 		dx += rtsc->x + rtsc->dx - x;
600 	dy = seg_x2y(dx, isc->sm1);
601 
602 	rtsc->x = x;
603 	rtsc->y = y;
604 	rtsc->dx = dx;
605 	rtsc->dy = dy;
606 }
607 
608 static void
609 init_ed(struct hfsc_class *cl, unsigned int next_len)
610 {
611 	u64 cur_time = psched_get_time();
612 
613 	/* update the deadline curve */
614 	rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
615 
616 	/*
617 	 * update the eligible curve.
618 	 * for concave, it is equal to the deadline curve.
619 	 * for convex, it is a linear curve with slope m2.
620 	 */
621 	cl->cl_eligible = cl->cl_deadline;
622 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
623 		cl->cl_eligible.dx = 0;
624 		cl->cl_eligible.dy = 0;
625 	}
626 
627 	/* compute e and d */
628 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
629 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
630 
631 	eltree_insert(cl);
632 }
633 
634 static void
635 update_ed(struct hfsc_class *cl, unsigned int next_len)
636 {
637 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
638 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
639 
640 	eltree_update(cl);
641 }
642 
643 static inline void
644 update_d(struct hfsc_class *cl, unsigned int next_len)
645 {
646 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
647 }
648 
649 static inline void
650 update_cfmin(struct hfsc_class *cl)
651 {
652 	struct rb_node *n = rb_first(&cl->cf_tree);
653 	struct hfsc_class *p;
654 
655 	if (n == NULL) {
656 		cl->cl_cfmin = 0;
657 		return;
658 	}
659 	p = rb_entry(n, struct hfsc_class, cf_node);
660 	cl->cl_cfmin = p->cl_f;
661 }
662 
663 static void
664 init_vf(struct hfsc_class *cl, unsigned int len)
665 {
666 	struct hfsc_class *max_cl;
667 	struct rb_node *n;
668 	u64 vt, f, cur_time;
669 	int go_active;
670 
671 	cur_time = 0;
672 	go_active = 1;
673 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
674 		if (go_active && cl->cl_nactive++ == 0)
675 			go_active = 1;
676 		else
677 			go_active = 0;
678 
679 		if (go_active) {
680 			n = rb_last(&cl->cl_parent->vt_tree);
681 			if (n != NULL) {
682 				max_cl = rb_entry(n, struct hfsc_class, vt_node);
683 				/*
684 				 * set vt to the average of the min and max
685 				 * classes.  if the parent's period didn't
686 				 * change, don't decrease vt of the class.
687 				 */
688 				vt = max_cl->cl_vt;
689 				if (cl->cl_parent->cl_cvtmin != 0)
690 					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
691 
692 				if (cl->cl_parent->cl_vtperiod !=
693 				    cl->cl_parentperiod || vt > cl->cl_vt)
694 					cl->cl_vt = vt;
695 			} else {
696 				/*
697 				 * first child for a new parent backlog period.
698 				 * initialize cl_vt to the highest value seen
699 				 * among the siblings. this is analogous to
700 				 * what cur_time would provide in realtime case.
701 				 */
702 				cl->cl_vt = cl->cl_parent->cl_cvtoff;
703 				cl->cl_parent->cl_cvtmin = 0;
704 			}
705 
706 			/* update the virtual curve */
707 			rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
708 			cl->cl_vtadj = 0;
709 
710 			cl->cl_vtperiod++;  /* increment vt period */
711 			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
712 			if (cl->cl_parent->cl_nactive == 0)
713 				cl->cl_parentperiod++;
714 			cl->cl_f = 0;
715 
716 			vttree_insert(cl);
717 			cftree_insert(cl);
718 
719 			if (cl->cl_flags & HFSC_USC) {
720 				/* class has upper limit curve */
721 				if (cur_time == 0)
722 					cur_time = psched_get_time();
723 
724 				/* update the ulimit curve */
725 				rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
726 					 cl->cl_total);
727 				/* compute myf */
728 				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
729 						      cl->cl_total);
730 			}
731 		}
732 
733 		f = max(cl->cl_myf, cl->cl_cfmin);
734 		if (f != cl->cl_f) {
735 			cl->cl_f = f;
736 			cftree_update(cl);
737 		}
738 		update_cfmin(cl->cl_parent);
739 	}
740 }
741 
742 static void
743 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
744 {
745 	u64 f; /* , myf_bound, delta; */
746 	int go_passive = 0;
747 
748 	if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
749 		go_passive = 1;
750 
751 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
752 		cl->cl_total += len;
753 
754 		if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
755 			continue;
756 
757 		if (go_passive && --cl->cl_nactive == 0)
758 			go_passive = 1;
759 		else
760 			go_passive = 0;
761 
762 		/* update vt */
763 		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
764 
765 		/*
766 		 * if vt of the class is smaller than cvtmin,
767 		 * the class was skipped in the past due to non-fit.
768 		 * if so, we need to adjust vtadj.
769 		 */
770 		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
771 			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
772 			cl->cl_vt = cl->cl_parent->cl_cvtmin;
773 		}
774 
775 		if (go_passive) {
776 			/* no more active child, going passive */
777 
778 			/* update cvtoff of the parent class */
779 			if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
780 				cl->cl_parent->cl_cvtoff = cl->cl_vt;
781 
782 			/* remove this class from the vt tree */
783 			vttree_remove(cl);
784 
785 			cftree_remove(cl);
786 			update_cfmin(cl->cl_parent);
787 
788 			continue;
789 		}
790 
791 		/* update the vt tree */
792 		vttree_update(cl);
793 
794 		/* update f */
795 		if (cl->cl_flags & HFSC_USC) {
796 			cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
797 #if 0
798 			cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
799 							      cl->cl_total);
800 			/*
801 			 * This code causes classes to stay way under their
802 			 * limit when multiple classes are used at gigabit
803 			 * speed. needs investigation. -kaber
804 			 */
805 			/*
806 			 * if myf lags behind by more than one clock tick
807 			 * from the current time, adjust myfadj to prevent
808 			 * a rate-limited class from going greedy.
809 			 * in a steady state under rate-limiting, myf
810 			 * fluctuates within one clock tick.
811 			 */
812 			myf_bound = cur_time - PSCHED_JIFFIE2US(1);
813 			if (cl->cl_myf < myf_bound) {
814 				delta = cur_time - cl->cl_myf;
815 				cl->cl_myfadj += delta;
816 				cl->cl_myf += delta;
817 			}
818 #endif
819 		}
820 
821 		f = max(cl->cl_myf, cl->cl_cfmin);
822 		if (f != cl->cl_f) {
823 			cl->cl_f = f;
824 			cftree_update(cl);
825 			update_cfmin(cl->cl_parent);
826 		}
827 	}
828 }
829 
830 static unsigned int
831 qdisc_peek_len(struct Qdisc *sch)
832 {
833 	struct sk_buff *skb;
834 	unsigned int len;
835 
836 	skb = sch->ops->peek(sch);
837 	if (unlikely(skb == NULL)) {
838 		qdisc_warn_nonwc("qdisc_peek_len", sch);
839 		return 0;
840 	}
841 	len = qdisc_pkt_len(skb);
842 
843 	return len;
844 }
845 
846 static void
847 hfsc_adjust_levels(struct hfsc_class *cl)
848 {
849 	struct hfsc_class *p;
850 	unsigned int level;
851 
852 	do {
853 		level = 0;
854 		list_for_each_entry(p, &cl->children, siblings) {
855 			if (p->level >= level)
856 				level = p->level + 1;
857 		}
858 		cl->level = level;
859 	} while ((cl = cl->cl_parent) != NULL);
860 }
861 
862 static inline struct hfsc_class *
863 hfsc_find_class(u32 classid, struct Qdisc *sch)
864 {
865 	struct hfsc_sched *q = qdisc_priv(sch);
866 	struct Qdisc_class_common *clc;
867 
868 	clc = qdisc_class_find(&q->clhash, classid);
869 	if (clc == NULL)
870 		return NULL;
871 	return container_of(clc, struct hfsc_class, cl_common);
872 }
873 
874 static void
875 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
876 		u64 cur_time)
877 {
878 	sc2isc(rsc, &cl->cl_rsc);
879 	rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
880 	cl->cl_eligible = cl->cl_deadline;
881 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
882 		cl->cl_eligible.dx = 0;
883 		cl->cl_eligible.dy = 0;
884 	}
885 	cl->cl_flags |= HFSC_RSC;
886 }
887 
888 static void
889 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
890 {
891 	sc2isc(fsc, &cl->cl_fsc);
892 	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
893 	cl->cl_flags |= HFSC_FSC;
894 }
895 
896 static void
897 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
898 		u64 cur_time)
899 {
900 	sc2isc(usc, &cl->cl_usc);
901 	rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
902 	cl->cl_flags |= HFSC_USC;
903 }
904 
905 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
906 	[TCA_HFSC_RSC]	= { .len = sizeof(struct tc_service_curve) },
907 	[TCA_HFSC_FSC]	= { .len = sizeof(struct tc_service_curve) },
908 	[TCA_HFSC_USC]	= { .len = sizeof(struct tc_service_curve) },
909 };
910 
911 static int
912 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
913 		  struct nlattr **tca, unsigned long *arg,
914 		  struct netlink_ext_ack *extack)
915 {
916 	struct hfsc_sched *q = qdisc_priv(sch);
917 	struct hfsc_class *cl = (struct hfsc_class *)*arg;
918 	struct hfsc_class *parent = NULL;
919 	struct nlattr *opt = tca[TCA_OPTIONS];
920 	struct nlattr *tb[TCA_HFSC_MAX + 1];
921 	struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
922 	u64 cur_time;
923 	int err;
924 
925 	if (opt == NULL)
926 		return -EINVAL;
927 
928 	err = nla_parse_nested_deprecated(tb, TCA_HFSC_MAX, opt, hfsc_policy,
929 					  NULL);
930 	if (err < 0)
931 		return err;
932 
933 	if (tb[TCA_HFSC_RSC]) {
934 		rsc = nla_data(tb[TCA_HFSC_RSC]);
935 		if (rsc->m1 == 0 && rsc->m2 == 0)
936 			rsc = NULL;
937 	}
938 
939 	if (tb[TCA_HFSC_FSC]) {
940 		fsc = nla_data(tb[TCA_HFSC_FSC]);
941 		if (fsc->m1 == 0 && fsc->m2 == 0)
942 			fsc = NULL;
943 	}
944 
945 	if (tb[TCA_HFSC_USC]) {
946 		usc = nla_data(tb[TCA_HFSC_USC]);
947 		if (usc->m1 == 0 && usc->m2 == 0)
948 			usc = NULL;
949 	}
950 
951 	if (cl != NULL) {
952 		int old_flags;
953 
954 		if (parentid) {
955 			if (cl->cl_parent &&
956 			    cl->cl_parent->cl_common.classid != parentid)
957 				return -EINVAL;
958 			if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
959 				return -EINVAL;
960 		}
961 		cur_time = psched_get_time();
962 
963 		if (tca[TCA_RATE]) {
964 			err = gen_replace_estimator(&cl->bstats, NULL,
965 						    &cl->rate_est,
966 						    NULL,
967 						    true,
968 						    tca[TCA_RATE]);
969 			if (err)
970 				return err;
971 		}
972 
973 		sch_tree_lock(sch);
974 		old_flags = cl->cl_flags;
975 
976 		if (rsc != NULL)
977 			hfsc_change_rsc(cl, rsc, cur_time);
978 		if (fsc != NULL)
979 			hfsc_change_fsc(cl, fsc);
980 		if (usc != NULL)
981 			hfsc_change_usc(cl, usc, cur_time);
982 
983 		if (cl->qdisc->q.qlen != 0) {
984 			int len = qdisc_peek_len(cl->qdisc);
985 
986 			if (cl->cl_flags & HFSC_RSC) {
987 				if (old_flags & HFSC_RSC)
988 					update_ed(cl, len);
989 				else
990 					init_ed(cl, len);
991 			}
992 
993 			if (cl->cl_flags & HFSC_FSC) {
994 				if (old_flags & HFSC_FSC)
995 					update_vf(cl, 0, cur_time);
996 				else
997 					init_vf(cl, len);
998 			}
999 		}
1000 		sch_tree_unlock(sch);
1001 
1002 		return 0;
1003 	}
1004 
1005 	if (parentid == TC_H_ROOT)
1006 		return -EEXIST;
1007 
1008 	parent = &q->root;
1009 	if (parentid) {
1010 		parent = hfsc_find_class(parentid, sch);
1011 		if (parent == NULL)
1012 			return -ENOENT;
1013 	}
1014 	if (!(parent->cl_flags & HFSC_FSC) && parent != &q->root) {
1015 		NL_SET_ERR_MSG(extack, "Invalid parent - parent class must have FSC");
1016 		return -EINVAL;
1017 	}
1018 
1019 	if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1020 		return -EINVAL;
1021 	if (hfsc_find_class(classid, sch))
1022 		return -EEXIST;
1023 
1024 	if (rsc == NULL && fsc == NULL)
1025 		return -EINVAL;
1026 
1027 	cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1028 	if (cl == NULL)
1029 		return -ENOBUFS;
1030 
1031 	err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1032 	if (err) {
1033 		kfree(cl);
1034 		return err;
1035 	}
1036 
1037 	if (tca[TCA_RATE]) {
1038 		err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1039 					NULL, true, tca[TCA_RATE]);
1040 		if (err) {
1041 			tcf_block_put(cl->block);
1042 			kfree(cl);
1043 			return err;
1044 		}
1045 	}
1046 
1047 	if (rsc != NULL)
1048 		hfsc_change_rsc(cl, rsc, 0);
1049 	if (fsc != NULL)
1050 		hfsc_change_fsc(cl, fsc);
1051 	if (usc != NULL)
1052 		hfsc_change_usc(cl, usc, 0);
1053 
1054 	cl->cl_common.classid = classid;
1055 	cl->sched     = q;
1056 	cl->cl_parent = parent;
1057 	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1058 				      classid, NULL);
1059 	if (cl->qdisc == NULL)
1060 		cl->qdisc = &noop_qdisc;
1061 	else
1062 		qdisc_hash_add(cl->qdisc, true);
1063 	INIT_LIST_HEAD(&cl->children);
1064 	cl->vt_tree = RB_ROOT;
1065 	cl->cf_tree = RB_ROOT;
1066 
1067 	sch_tree_lock(sch);
1068 	qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1069 	list_add_tail(&cl->siblings, &parent->children);
1070 	if (parent->level == 0)
1071 		qdisc_purge_queue(parent->qdisc);
1072 	hfsc_adjust_levels(parent);
1073 	sch_tree_unlock(sch);
1074 
1075 	qdisc_class_hash_grow(sch, &q->clhash);
1076 
1077 	*arg = (unsigned long)cl;
1078 	return 0;
1079 }
1080 
1081 static void
1082 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1083 {
1084 	struct hfsc_sched *q = qdisc_priv(sch);
1085 
1086 	tcf_block_put(cl->block);
1087 	qdisc_put(cl->qdisc);
1088 	gen_kill_estimator(&cl->rate_est);
1089 	if (cl != &q->root)
1090 		kfree(cl);
1091 }
1092 
1093 static int
1094 hfsc_delete_class(struct Qdisc *sch, unsigned long arg,
1095 		  struct netlink_ext_ack *extack)
1096 {
1097 	struct hfsc_sched *q = qdisc_priv(sch);
1098 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1099 
1100 	if (cl->level > 0 || qdisc_class_in_use(&cl->cl_common) ||
1101 	    cl == &q->root) {
1102 		NL_SET_ERR_MSG(extack, "HFSC class in use");
1103 		return -EBUSY;
1104 	}
1105 
1106 	sch_tree_lock(sch);
1107 
1108 	list_del(&cl->siblings);
1109 	hfsc_adjust_levels(cl->cl_parent);
1110 
1111 	qdisc_purge_queue(cl->qdisc);
1112 	qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1113 
1114 	sch_tree_unlock(sch);
1115 
1116 	hfsc_destroy_class(sch, cl);
1117 	return 0;
1118 }
1119 
1120 static struct hfsc_class *
1121 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1122 {
1123 	struct hfsc_sched *q = qdisc_priv(sch);
1124 	struct hfsc_class *head, *cl;
1125 	struct tcf_result res;
1126 	struct tcf_proto *tcf;
1127 	int result;
1128 
1129 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1130 	    (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1131 		if (cl->level == 0)
1132 			return cl;
1133 
1134 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1135 	head = &q->root;
1136 	tcf = rcu_dereference_bh(q->root.filter_list);
1137 	while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
1138 #ifdef CONFIG_NET_CLS_ACT
1139 		switch (result) {
1140 		case TC_ACT_QUEUED:
1141 		case TC_ACT_STOLEN:
1142 		case TC_ACT_TRAP:
1143 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1144 			fallthrough;
1145 		case TC_ACT_SHOT:
1146 			return NULL;
1147 		}
1148 #endif
1149 		cl = (struct hfsc_class *)res.class;
1150 		if (!cl) {
1151 			cl = hfsc_find_class(res.classid, sch);
1152 			if (!cl)
1153 				break; /* filter selected invalid classid */
1154 			if (cl->level >= head->level)
1155 				break; /* filter may only point downwards */
1156 		}
1157 
1158 		if (cl->level == 0)
1159 			return cl; /* hit leaf class */
1160 
1161 		/* apply inner filter chain */
1162 		tcf = rcu_dereference_bh(cl->filter_list);
1163 		head = cl;
1164 	}
1165 
1166 	/* classification failed, try default class */
1167 	cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1168 	if (cl == NULL || cl->level > 0)
1169 		return NULL;
1170 
1171 	return cl;
1172 }
1173 
1174 static int
1175 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1176 		 struct Qdisc **old, struct netlink_ext_ack *extack)
1177 {
1178 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1179 
1180 	if (cl->level > 0)
1181 		return -EINVAL;
1182 	if (new == NULL) {
1183 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1184 					cl->cl_common.classid, NULL);
1185 		if (new == NULL)
1186 			new = &noop_qdisc;
1187 	}
1188 
1189 	*old = qdisc_replace(sch, new, &cl->qdisc);
1190 	return 0;
1191 }
1192 
1193 static struct Qdisc *
1194 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1195 {
1196 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1197 
1198 	if (cl->level == 0)
1199 		return cl->qdisc;
1200 
1201 	return NULL;
1202 }
1203 
1204 static void
1205 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1206 {
1207 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1208 
1209 	/* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1210 	 * needs to be called explicitly to remove a class from vttree.
1211 	 */
1212 	update_vf(cl, 0, 0);
1213 	if (cl->cl_flags & HFSC_RSC)
1214 		eltree_remove(cl);
1215 }
1216 
1217 static unsigned long
1218 hfsc_search_class(struct Qdisc *sch, u32 classid)
1219 {
1220 	return (unsigned long)hfsc_find_class(classid, sch);
1221 }
1222 
1223 static unsigned long
1224 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1225 {
1226 	struct hfsc_class *p = (struct hfsc_class *)parent;
1227 	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1228 
1229 	if (cl != NULL) {
1230 		if (p != NULL && p->level <= cl->level)
1231 			return 0;
1232 		qdisc_class_get(&cl->cl_common);
1233 	}
1234 
1235 	return (unsigned long)cl;
1236 }
1237 
1238 static void
1239 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1240 {
1241 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1242 
1243 	qdisc_class_put(&cl->cl_common);
1244 }
1245 
1246 static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1247 					struct netlink_ext_ack *extack)
1248 {
1249 	struct hfsc_sched *q = qdisc_priv(sch);
1250 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1251 
1252 	if (cl == NULL)
1253 		cl = &q->root;
1254 
1255 	return cl->block;
1256 }
1257 
1258 static int
1259 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1260 {
1261 	struct tc_service_curve tsc;
1262 
1263 	tsc.m1 = sm2m(sc->sm1);
1264 	tsc.d  = dx2d(sc->dx);
1265 	tsc.m2 = sm2m(sc->sm2);
1266 	if (nla_put(skb, attr, sizeof(tsc), &tsc))
1267 		goto nla_put_failure;
1268 
1269 	return skb->len;
1270 
1271  nla_put_failure:
1272 	return -1;
1273 }
1274 
1275 static int
1276 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1277 {
1278 	if ((cl->cl_flags & HFSC_RSC) &&
1279 	    (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1280 		goto nla_put_failure;
1281 
1282 	if ((cl->cl_flags & HFSC_FSC) &&
1283 	    (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1284 		goto nla_put_failure;
1285 
1286 	if ((cl->cl_flags & HFSC_USC) &&
1287 	    (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1288 		goto nla_put_failure;
1289 
1290 	return skb->len;
1291 
1292  nla_put_failure:
1293 	return -1;
1294 }
1295 
1296 static int
1297 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1298 		struct tcmsg *tcm)
1299 {
1300 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1301 	struct nlattr *nest;
1302 
1303 	tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1304 					  TC_H_ROOT;
1305 	tcm->tcm_handle = cl->cl_common.classid;
1306 	if (cl->level == 0)
1307 		tcm->tcm_info = cl->qdisc->handle;
1308 
1309 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1310 	if (nest == NULL)
1311 		goto nla_put_failure;
1312 	if (hfsc_dump_curves(skb, cl) < 0)
1313 		goto nla_put_failure;
1314 	return nla_nest_end(skb, nest);
1315 
1316  nla_put_failure:
1317 	nla_nest_cancel(skb, nest);
1318 	return -EMSGSIZE;
1319 }
1320 
1321 static int
1322 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1323 	struct gnet_dump *d)
1324 {
1325 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1326 	struct tc_hfsc_stats xstats;
1327 	__u32 qlen;
1328 
1329 	qdisc_qstats_qlen_backlog(cl->qdisc, &qlen, &cl->qstats.backlog);
1330 	xstats.level   = cl->level;
1331 	xstats.period  = cl->cl_vtperiod;
1332 	xstats.work    = cl->cl_total;
1333 	xstats.rtwork  = cl->cl_cumul;
1334 
1335 	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
1336 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1337 	    gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1338 		return -1;
1339 
1340 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1341 }
1342 
1343 
1344 
1345 static void
1346 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1347 {
1348 	struct hfsc_sched *q = qdisc_priv(sch);
1349 	struct hfsc_class *cl;
1350 	unsigned int i;
1351 
1352 	if (arg->stop)
1353 		return;
1354 
1355 	for (i = 0; i < q->clhash.hashsize; i++) {
1356 		hlist_for_each_entry(cl, &q->clhash.hash[i],
1357 				     cl_common.hnode) {
1358 			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
1359 				return;
1360 		}
1361 	}
1362 }
1363 
1364 static void
1365 hfsc_schedule_watchdog(struct Qdisc *sch)
1366 {
1367 	struct hfsc_sched *q = qdisc_priv(sch);
1368 	struct hfsc_class *cl;
1369 	u64 next_time = 0;
1370 
1371 	cl = eltree_get_minel(q);
1372 	if (cl)
1373 		next_time = cl->cl_e;
1374 	if (q->root.cl_cfmin != 0) {
1375 		if (next_time == 0 || next_time > q->root.cl_cfmin)
1376 			next_time = q->root.cl_cfmin;
1377 	}
1378 	if (next_time)
1379 		qdisc_watchdog_schedule(&q->watchdog, next_time);
1380 }
1381 
1382 static int
1383 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1384 		struct netlink_ext_ack *extack)
1385 {
1386 	struct hfsc_sched *q = qdisc_priv(sch);
1387 	struct tc_hfsc_qopt *qopt;
1388 	int err;
1389 
1390 	qdisc_watchdog_init(&q->watchdog, sch);
1391 
1392 	if (!opt || nla_len(opt) < sizeof(*qopt))
1393 		return -EINVAL;
1394 	qopt = nla_data(opt);
1395 
1396 	q->defcls = qopt->defcls;
1397 	err = qdisc_class_hash_init(&q->clhash);
1398 	if (err < 0)
1399 		return err;
1400 	q->eligible = RB_ROOT;
1401 
1402 	err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1403 	if (err)
1404 		return err;
1405 
1406 	gnet_stats_basic_sync_init(&q->root.bstats);
1407 	q->root.cl_common.classid = sch->handle;
1408 	q->root.sched   = q;
1409 	q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1410 					  sch->handle, NULL);
1411 	if (q->root.qdisc == NULL)
1412 		q->root.qdisc = &noop_qdisc;
1413 	else
1414 		qdisc_hash_add(q->root.qdisc, true);
1415 	INIT_LIST_HEAD(&q->root.children);
1416 	q->root.vt_tree = RB_ROOT;
1417 	q->root.cf_tree = RB_ROOT;
1418 
1419 	qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1420 	qdisc_class_hash_grow(sch, &q->clhash);
1421 
1422 	return 0;
1423 }
1424 
1425 static int
1426 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1427 		  struct netlink_ext_ack *extack)
1428 {
1429 	struct hfsc_sched *q = qdisc_priv(sch);
1430 	struct tc_hfsc_qopt *qopt;
1431 
1432 	if (nla_len(opt) < sizeof(*qopt))
1433 		return -EINVAL;
1434 	qopt = nla_data(opt);
1435 
1436 	sch_tree_lock(sch);
1437 	q->defcls = qopt->defcls;
1438 	sch_tree_unlock(sch);
1439 
1440 	return 0;
1441 }
1442 
1443 static void
1444 hfsc_reset_class(struct hfsc_class *cl)
1445 {
1446 	cl->cl_total        = 0;
1447 	cl->cl_cumul        = 0;
1448 	cl->cl_d            = 0;
1449 	cl->cl_e            = 0;
1450 	cl->cl_vt           = 0;
1451 	cl->cl_vtadj        = 0;
1452 	cl->cl_cvtmin       = 0;
1453 	cl->cl_cvtoff       = 0;
1454 	cl->cl_vtperiod     = 0;
1455 	cl->cl_parentperiod = 0;
1456 	cl->cl_f            = 0;
1457 	cl->cl_myf          = 0;
1458 	cl->cl_cfmin        = 0;
1459 	cl->cl_nactive      = 0;
1460 
1461 	cl->vt_tree = RB_ROOT;
1462 	cl->cf_tree = RB_ROOT;
1463 	qdisc_reset(cl->qdisc);
1464 
1465 	if (cl->cl_flags & HFSC_RSC)
1466 		rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1467 	if (cl->cl_flags & HFSC_FSC)
1468 		rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1469 	if (cl->cl_flags & HFSC_USC)
1470 		rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1471 }
1472 
1473 static void
1474 hfsc_reset_qdisc(struct Qdisc *sch)
1475 {
1476 	struct hfsc_sched *q = qdisc_priv(sch);
1477 	struct hfsc_class *cl;
1478 	unsigned int i;
1479 
1480 	for (i = 0; i < q->clhash.hashsize; i++) {
1481 		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1482 			hfsc_reset_class(cl);
1483 	}
1484 	q->eligible = RB_ROOT;
1485 	qdisc_watchdog_cancel(&q->watchdog);
1486 }
1487 
1488 static void
1489 hfsc_destroy_qdisc(struct Qdisc *sch)
1490 {
1491 	struct hfsc_sched *q = qdisc_priv(sch);
1492 	struct hlist_node *next;
1493 	struct hfsc_class *cl;
1494 	unsigned int i;
1495 
1496 	for (i = 0; i < q->clhash.hashsize; i++) {
1497 		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1498 			tcf_block_put(cl->block);
1499 			cl->block = NULL;
1500 		}
1501 	}
1502 	for (i = 0; i < q->clhash.hashsize; i++) {
1503 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1504 					  cl_common.hnode)
1505 			hfsc_destroy_class(sch, cl);
1506 	}
1507 	qdisc_class_hash_destroy(&q->clhash);
1508 	qdisc_watchdog_cancel(&q->watchdog);
1509 }
1510 
1511 static int
1512 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1513 {
1514 	struct hfsc_sched *q = qdisc_priv(sch);
1515 	unsigned char *b = skb_tail_pointer(skb);
1516 	struct tc_hfsc_qopt qopt;
1517 
1518 	qopt.defcls = q->defcls;
1519 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1520 		goto nla_put_failure;
1521 	return skb->len;
1522 
1523  nla_put_failure:
1524 	nlmsg_trim(skb, b);
1525 	return -1;
1526 }
1527 
1528 static int
1529 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1530 {
1531 	unsigned int len = qdisc_pkt_len(skb);
1532 	struct hfsc_class *cl;
1533 	int err;
1534 	bool first;
1535 
1536 	cl = hfsc_classify(skb, sch, &err);
1537 	if (cl == NULL) {
1538 		if (err & __NET_XMIT_BYPASS)
1539 			qdisc_qstats_drop(sch);
1540 		__qdisc_drop(skb, to_free);
1541 		return err;
1542 	}
1543 
1544 	first = !cl->qdisc->q.qlen;
1545 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1546 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1547 		if (net_xmit_drop_count(err)) {
1548 			cl->qstats.drops++;
1549 			qdisc_qstats_drop(sch);
1550 		}
1551 		return err;
1552 	}
1553 
1554 	if (first) {
1555 		if (cl->cl_flags & HFSC_RSC)
1556 			init_ed(cl, len);
1557 		if (cl->cl_flags & HFSC_FSC)
1558 			init_vf(cl, len);
1559 		/*
1560 		 * If this is the first packet, isolate the head so an eventual
1561 		 * head drop before the first dequeue operation has no chance
1562 		 * to invalidate the deadline.
1563 		 */
1564 		if (cl->cl_flags & HFSC_RSC)
1565 			cl->qdisc->ops->peek(cl->qdisc);
1566 
1567 	}
1568 
1569 	sch->qstats.backlog += len;
1570 	sch->q.qlen++;
1571 
1572 	return NET_XMIT_SUCCESS;
1573 }
1574 
1575 static struct sk_buff *
1576 hfsc_dequeue(struct Qdisc *sch)
1577 {
1578 	struct hfsc_sched *q = qdisc_priv(sch);
1579 	struct hfsc_class *cl;
1580 	struct sk_buff *skb;
1581 	u64 cur_time;
1582 	unsigned int next_len;
1583 	int realtime = 0;
1584 
1585 	if (sch->q.qlen == 0)
1586 		return NULL;
1587 
1588 	cur_time = psched_get_time();
1589 
1590 	/*
1591 	 * if there are eligible classes, use real-time criteria.
1592 	 * find the class with the minimum deadline among
1593 	 * the eligible classes.
1594 	 */
1595 	cl = eltree_get_mindl(q, cur_time);
1596 	if (cl) {
1597 		realtime = 1;
1598 	} else {
1599 		/*
1600 		 * use link-sharing criteria
1601 		 * get the class with the minimum vt in the hierarchy
1602 		 */
1603 		cl = vttree_get_minvt(&q->root, cur_time);
1604 		if (cl == NULL) {
1605 			qdisc_qstats_overlimit(sch);
1606 			hfsc_schedule_watchdog(sch);
1607 			return NULL;
1608 		}
1609 	}
1610 
1611 	skb = qdisc_dequeue_peeked(cl->qdisc);
1612 	if (skb == NULL) {
1613 		qdisc_warn_nonwc("HFSC", cl->qdisc);
1614 		return NULL;
1615 	}
1616 
1617 	bstats_update(&cl->bstats, skb);
1618 	update_vf(cl, qdisc_pkt_len(skb), cur_time);
1619 	if (realtime)
1620 		cl->cl_cumul += qdisc_pkt_len(skb);
1621 
1622 	if (cl->cl_flags & HFSC_RSC) {
1623 		if (cl->qdisc->q.qlen != 0) {
1624 			/* update ed */
1625 			next_len = qdisc_peek_len(cl->qdisc);
1626 			if (realtime)
1627 				update_ed(cl, next_len);
1628 			else
1629 				update_d(cl, next_len);
1630 		} else {
1631 			/* the class becomes passive */
1632 			eltree_remove(cl);
1633 		}
1634 	}
1635 
1636 	qdisc_bstats_update(sch, skb);
1637 	qdisc_qstats_backlog_dec(sch, skb);
1638 	sch->q.qlen--;
1639 
1640 	return skb;
1641 }
1642 
1643 static const struct Qdisc_class_ops hfsc_class_ops = {
1644 	.change		= hfsc_change_class,
1645 	.delete		= hfsc_delete_class,
1646 	.graft		= hfsc_graft_class,
1647 	.leaf		= hfsc_class_leaf,
1648 	.qlen_notify	= hfsc_qlen_notify,
1649 	.find		= hfsc_search_class,
1650 	.bind_tcf	= hfsc_bind_tcf,
1651 	.unbind_tcf	= hfsc_unbind_tcf,
1652 	.tcf_block	= hfsc_tcf_block,
1653 	.dump		= hfsc_dump_class,
1654 	.dump_stats	= hfsc_dump_class_stats,
1655 	.walk		= hfsc_walk
1656 };
1657 
1658 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1659 	.id		= "hfsc",
1660 	.init		= hfsc_init_qdisc,
1661 	.change		= hfsc_change_qdisc,
1662 	.reset		= hfsc_reset_qdisc,
1663 	.destroy	= hfsc_destroy_qdisc,
1664 	.dump		= hfsc_dump_qdisc,
1665 	.enqueue	= hfsc_enqueue,
1666 	.dequeue	= hfsc_dequeue,
1667 	.peek		= qdisc_peek_dequeued,
1668 	.cl_ops		= &hfsc_class_ops,
1669 	.priv_size	= sizeof(struct hfsc_sched),
1670 	.owner		= THIS_MODULE
1671 };
1672 
1673 static int __init
1674 hfsc_init(void)
1675 {
1676 	return register_qdisc(&hfsc_qdisc_ops);
1677 }
1678 
1679 static void __exit
1680 hfsc_cleanup(void)
1681 {
1682 	unregister_qdisc(&hfsc_qdisc_ops);
1683 }
1684 
1685 MODULE_LICENSE("GPL");
1686 module_init(hfsc_init);
1687 module_exit(hfsc_cleanup);
1688