xref: /linux/net/sched/sch_hfsc.c (revision 85502b2214d50ba0ddf2a5fb454e4d28a160d175)
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 
cl_in_el_or_vttree(struct hfsc_class * cl)178 static bool cl_in_el_or_vttree(struct hfsc_class *cl)
179 {
180 	return ((cl->cl_flags & HFSC_FSC) && cl->cl_nactive) ||
181 		((cl->cl_flags & HFSC_RSC) && !RB_EMPTY_NODE(&cl->el_node));
182 }
183 
184 /*
185  * eligible tree holds backlogged classes being sorted by their eligible times.
186  * there is one eligible tree per hfsc instance.
187  */
188 
189 static void
eltree_insert(struct hfsc_class * cl)190 eltree_insert(struct hfsc_class *cl)
191 {
192 	struct rb_node **p = &cl->sched->eligible.rb_node;
193 	struct rb_node *parent = NULL;
194 	struct hfsc_class *cl1;
195 
196 	while (*p != NULL) {
197 		parent = *p;
198 		cl1 = rb_entry(parent, struct hfsc_class, el_node);
199 		if (cl->cl_e >= cl1->cl_e)
200 			p = &parent->rb_right;
201 		else
202 			p = &parent->rb_left;
203 	}
204 	rb_link_node(&cl->el_node, parent, p);
205 	rb_insert_color(&cl->el_node, &cl->sched->eligible);
206 }
207 
208 static inline void
eltree_remove(struct hfsc_class * cl)209 eltree_remove(struct hfsc_class *cl)
210 {
211 	if (!RB_EMPTY_NODE(&cl->el_node)) {
212 		rb_erase(&cl->el_node, &cl->sched->eligible);
213 		RB_CLEAR_NODE(&cl->el_node);
214 	}
215 }
216 
217 static inline void
eltree_update(struct hfsc_class * cl)218 eltree_update(struct hfsc_class *cl)
219 {
220 	eltree_remove(cl);
221 	eltree_insert(cl);
222 }
223 
224 /* find the class with the minimum deadline among the eligible classes */
225 static inline struct hfsc_class *
eltree_get_mindl(struct hfsc_sched * q,u64 cur_time)226 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
227 {
228 	struct hfsc_class *p, *cl = NULL;
229 	struct rb_node *n;
230 
231 	for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
232 		p = rb_entry(n, struct hfsc_class, el_node);
233 		if (p->cl_e > cur_time)
234 			break;
235 		if (cl == NULL || p->cl_d < cl->cl_d)
236 			cl = p;
237 	}
238 	return cl;
239 }
240 
241 /* find the class with minimum eligible time among the eligible classes */
242 static inline struct hfsc_class *
eltree_get_minel(struct hfsc_sched * q)243 eltree_get_minel(struct hfsc_sched *q)
244 {
245 	struct rb_node *n;
246 
247 	n = rb_first(&q->eligible);
248 	if (n == NULL)
249 		return NULL;
250 	return rb_entry(n, struct hfsc_class, el_node);
251 }
252 
253 /*
254  * vttree holds holds backlogged child classes being sorted by their virtual
255  * time. each intermediate class has one vttree.
256  */
257 static void
vttree_insert(struct hfsc_class * cl)258 vttree_insert(struct hfsc_class *cl)
259 {
260 	struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
261 	struct rb_node *parent = NULL;
262 	struct hfsc_class *cl1;
263 
264 	while (*p != NULL) {
265 		parent = *p;
266 		cl1 = rb_entry(parent, struct hfsc_class, vt_node);
267 		if (cl->cl_vt >= cl1->cl_vt)
268 			p = &parent->rb_right;
269 		else
270 			p = &parent->rb_left;
271 	}
272 	rb_link_node(&cl->vt_node, parent, p);
273 	rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
274 }
275 
276 static inline void
vttree_remove(struct hfsc_class * cl)277 vttree_remove(struct hfsc_class *cl)
278 {
279 	rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
280 }
281 
282 static inline void
vttree_update(struct hfsc_class * cl)283 vttree_update(struct hfsc_class *cl)
284 {
285 	vttree_remove(cl);
286 	vttree_insert(cl);
287 }
288 
289 static inline struct hfsc_class *
vttree_firstfit(struct hfsc_class * cl,u64 cur_time)290 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
291 {
292 	struct hfsc_class *p;
293 	struct rb_node *n;
294 
295 	for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
296 		p = rb_entry(n, struct hfsc_class, vt_node);
297 		if (p->cl_f <= cur_time)
298 			return p;
299 	}
300 	return NULL;
301 }
302 
303 /*
304  * get the leaf class with the minimum vt in the hierarchy
305  */
306 static struct hfsc_class *
vttree_get_minvt(struct hfsc_class * cl,u64 cur_time)307 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
308 {
309 	/* if root-class's cfmin is bigger than cur_time nothing to do */
310 	if (cl->cl_cfmin > cur_time)
311 		return NULL;
312 
313 	while (cl->level > 0) {
314 		cl = vttree_firstfit(cl, cur_time);
315 		if (cl == NULL)
316 			return NULL;
317 		/*
318 		 * update parent's cl_cvtmin.
319 		 */
320 		if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
321 			cl->cl_parent->cl_cvtmin = cl->cl_vt;
322 	}
323 	return cl;
324 }
325 
326 static void
cftree_insert(struct hfsc_class * cl)327 cftree_insert(struct hfsc_class *cl)
328 {
329 	struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
330 	struct rb_node *parent = NULL;
331 	struct hfsc_class *cl1;
332 
333 	while (*p != NULL) {
334 		parent = *p;
335 		cl1 = rb_entry(parent, struct hfsc_class, cf_node);
336 		if (cl->cl_f >= cl1->cl_f)
337 			p = &parent->rb_right;
338 		else
339 			p = &parent->rb_left;
340 	}
341 	rb_link_node(&cl->cf_node, parent, p);
342 	rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
343 }
344 
345 static inline void
cftree_remove(struct hfsc_class * cl)346 cftree_remove(struct hfsc_class *cl)
347 {
348 	rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
349 }
350 
351 static inline void
cftree_update(struct hfsc_class * cl)352 cftree_update(struct hfsc_class *cl)
353 {
354 	cftree_remove(cl);
355 	cftree_insert(cl);
356 }
357 
358 /*
359  * service curve support functions
360  *
361  *  external service curve parameters
362  *	m: bps
363  *	d: us
364  *  internal service curve parameters
365  *	sm: (bytes/psched_us) << SM_SHIFT
366  *	ism: (psched_us/byte) << ISM_SHIFT
367  *	dx: psched_us
368  *
369  * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
370  *
371  * sm and ism are scaled in order to keep effective digits.
372  * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
373  * digits in decimal using the following table.
374  *
375  *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
376  *  ------------+-------------------------------------------------------
377  *  bytes/1.024us 12.8e-3    128e-3     1280e-3    12800e-3   128000e-3
378  *
379  *  1.024us/byte  78.125     7.8125     0.78125    0.078125   0.0078125
380  *
381  * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
382  */
383 #define	SM_SHIFT	(30 - PSCHED_SHIFT)
384 #define	ISM_SHIFT	(8 + PSCHED_SHIFT)
385 
386 #define	SM_MASK		((1ULL << SM_SHIFT) - 1)
387 #define	ISM_MASK	((1ULL << ISM_SHIFT) - 1)
388 
389 static inline u64
seg_x2y(u64 x,u64 sm)390 seg_x2y(u64 x, u64 sm)
391 {
392 	u64 y;
393 
394 	/*
395 	 * compute
396 	 *	y = x * sm >> SM_SHIFT
397 	 * but divide it for the upper and lower bits to avoid overflow
398 	 */
399 	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
400 	return y;
401 }
402 
403 static inline u64
seg_y2x(u64 y,u64 ism)404 seg_y2x(u64 y, u64 ism)
405 {
406 	u64 x;
407 
408 	if (y == 0)
409 		x = 0;
410 	else if (ism == HT_INFINITY)
411 		x = HT_INFINITY;
412 	else {
413 		x = (y >> ISM_SHIFT) * ism
414 		    + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
415 	}
416 	return x;
417 }
418 
419 /* Convert m (bps) into sm (bytes/psched us) */
420 static u64
m2sm(u32 m)421 m2sm(u32 m)
422 {
423 	u64 sm;
424 
425 	sm = ((u64)m << SM_SHIFT);
426 	sm += PSCHED_TICKS_PER_SEC - 1;
427 	do_div(sm, PSCHED_TICKS_PER_SEC);
428 	return sm;
429 }
430 
431 /* convert m (bps) into ism (psched us/byte) */
432 static u64
m2ism(u32 m)433 m2ism(u32 m)
434 {
435 	u64 ism;
436 
437 	if (m == 0)
438 		ism = HT_INFINITY;
439 	else {
440 		ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
441 		ism += m - 1;
442 		do_div(ism, m);
443 	}
444 	return ism;
445 }
446 
447 /* convert d (us) into dx (psched us) */
448 static u64
d2dx(u32 d)449 d2dx(u32 d)
450 {
451 	u64 dx;
452 
453 	dx = ((u64)d * PSCHED_TICKS_PER_SEC);
454 	dx += USEC_PER_SEC - 1;
455 	do_div(dx, USEC_PER_SEC);
456 	return dx;
457 }
458 
459 /* convert sm (bytes/psched us) into m (bps) */
460 static u32
sm2m(u64 sm)461 sm2m(u64 sm)
462 {
463 	u64 m;
464 
465 	m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
466 	return (u32)m;
467 }
468 
469 /* convert dx (psched us) into d (us) */
470 static u32
dx2d(u64 dx)471 dx2d(u64 dx)
472 {
473 	u64 d;
474 
475 	d = dx * USEC_PER_SEC;
476 	do_div(d, PSCHED_TICKS_PER_SEC);
477 	return (u32)d;
478 }
479 
480 static void
sc2isc(struct tc_service_curve * sc,struct internal_sc * isc)481 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
482 {
483 	isc->sm1  = m2sm(sc->m1);
484 	isc->ism1 = m2ism(sc->m1);
485 	isc->dx   = d2dx(sc->d);
486 	isc->dy   = seg_x2y(isc->dx, isc->sm1);
487 	isc->sm2  = m2sm(sc->m2);
488 	isc->ism2 = m2ism(sc->m2);
489 }
490 
491 /*
492  * initialize the runtime service curve with the given internal
493  * service curve starting at (x, y).
494  */
495 static void
rtsc_init(struct runtime_sc * rtsc,struct internal_sc * isc,u64 x,u64 y)496 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
497 {
498 	rtsc->x	   = x;
499 	rtsc->y    = y;
500 	rtsc->sm1  = isc->sm1;
501 	rtsc->ism1 = isc->ism1;
502 	rtsc->dx   = isc->dx;
503 	rtsc->dy   = isc->dy;
504 	rtsc->sm2  = isc->sm2;
505 	rtsc->ism2 = isc->ism2;
506 }
507 
508 /*
509  * calculate the y-projection of the runtime service curve by the
510  * given x-projection value
511  */
512 static u64
rtsc_y2x(struct runtime_sc * rtsc,u64 y)513 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
514 {
515 	u64 x;
516 
517 	if (y < rtsc->y)
518 		x = rtsc->x;
519 	else if (y <= rtsc->y + rtsc->dy) {
520 		/* x belongs to the 1st segment */
521 		if (rtsc->dy == 0)
522 			x = rtsc->x + rtsc->dx;
523 		else
524 			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
525 	} else {
526 		/* x belongs to the 2nd segment */
527 		x = rtsc->x + rtsc->dx
528 		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
529 	}
530 	return x;
531 }
532 
533 static u64
rtsc_x2y(struct runtime_sc * rtsc,u64 x)534 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
535 {
536 	u64 y;
537 
538 	if (x <= rtsc->x)
539 		y = rtsc->y;
540 	else if (x <= rtsc->x + rtsc->dx)
541 		/* y belongs to the 1st segment */
542 		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
543 	else
544 		/* y belongs to the 2nd segment */
545 		y = rtsc->y + rtsc->dy
546 		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
547 	return y;
548 }
549 
550 /*
551  * update the runtime service curve by taking the minimum of the current
552  * runtime service curve and the service curve starting at (x, y).
553  */
554 static void
rtsc_min(struct runtime_sc * rtsc,struct internal_sc * isc,u64 x,u64 y)555 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
556 {
557 	u64 y1, y2, dx, dy;
558 	u32 dsm;
559 
560 	if (isc->sm1 <= isc->sm2) {
561 		/* service curve is convex */
562 		y1 = rtsc_x2y(rtsc, x);
563 		if (y1 < y)
564 			/* the current rtsc is smaller */
565 			return;
566 		rtsc->x = x;
567 		rtsc->y = y;
568 		return;
569 	}
570 
571 	/*
572 	 * service curve is concave
573 	 * compute the two y values of the current rtsc
574 	 *	y1: at x
575 	 *	y2: at (x + dx)
576 	 */
577 	y1 = rtsc_x2y(rtsc, x);
578 	if (y1 <= y) {
579 		/* rtsc is below isc, no change to rtsc */
580 		return;
581 	}
582 
583 	y2 = rtsc_x2y(rtsc, x + isc->dx);
584 	if (y2 >= y + isc->dy) {
585 		/* rtsc is above isc, replace rtsc by isc */
586 		rtsc->x = x;
587 		rtsc->y = y;
588 		rtsc->dx = isc->dx;
589 		rtsc->dy = isc->dy;
590 		return;
591 	}
592 
593 	/*
594 	 * the two curves intersect
595 	 * compute the offsets (dx, dy) using the reverse
596 	 * function of seg_x2y()
597 	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
598 	 */
599 	dx = (y1 - y) << SM_SHIFT;
600 	dsm = isc->sm1 - isc->sm2;
601 	do_div(dx, dsm);
602 	/*
603 	 * check if (x, y1) belongs to the 1st segment of rtsc.
604 	 * if so, add the offset.
605 	 */
606 	if (rtsc->x + rtsc->dx > x)
607 		dx += rtsc->x + rtsc->dx - x;
608 	dy = seg_x2y(dx, isc->sm1);
609 
610 	rtsc->x = x;
611 	rtsc->y = y;
612 	rtsc->dx = dx;
613 	rtsc->dy = dy;
614 }
615 
616 static void
init_ed(struct hfsc_class * cl,unsigned int next_len)617 init_ed(struct hfsc_class *cl, unsigned int next_len)
618 {
619 	u64 cur_time = psched_get_time();
620 
621 	/* update the deadline curve */
622 	rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
623 
624 	/*
625 	 * update the eligible curve.
626 	 * for concave, it is equal to the deadline curve.
627 	 * for convex, it is a linear curve with slope m2.
628 	 */
629 	cl->cl_eligible = cl->cl_deadline;
630 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
631 		cl->cl_eligible.dx = 0;
632 		cl->cl_eligible.dy = 0;
633 	}
634 
635 	/* compute e and d */
636 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
637 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
638 
639 	eltree_insert(cl);
640 }
641 
642 static void
update_ed(struct hfsc_class * cl,unsigned int next_len)643 update_ed(struct hfsc_class *cl, unsigned int next_len)
644 {
645 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
646 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
647 
648 	eltree_update(cl);
649 }
650 
651 static inline void
update_d(struct hfsc_class * cl,unsigned int next_len)652 update_d(struct hfsc_class *cl, unsigned int next_len)
653 {
654 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
655 }
656 
657 static inline void
update_cfmin(struct hfsc_class * cl)658 update_cfmin(struct hfsc_class *cl)
659 {
660 	struct rb_node *n = rb_first(&cl->cf_tree);
661 	struct hfsc_class *p;
662 
663 	if (n == NULL) {
664 		cl->cl_cfmin = 0;
665 		return;
666 	}
667 	p = rb_entry(n, struct hfsc_class, cf_node);
668 	cl->cl_cfmin = p->cl_f;
669 }
670 
671 static void
init_vf(struct hfsc_class * cl,unsigned int len)672 init_vf(struct hfsc_class *cl, unsigned int len)
673 {
674 	struct hfsc_class *max_cl;
675 	struct rb_node *n;
676 	u64 vt, f, cur_time;
677 	int go_active;
678 
679 	cur_time = 0;
680 	go_active = 1;
681 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
682 		if (go_active && cl->cl_nactive++ == 0)
683 			go_active = 1;
684 		else
685 			go_active = 0;
686 
687 		if (go_active) {
688 			n = rb_last(&cl->cl_parent->vt_tree);
689 			if (n != NULL) {
690 				max_cl = rb_entry(n, struct hfsc_class, vt_node);
691 				/*
692 				 * set vt to the average of the min and max
693 				 * classes.  if the parent's period didn't
694 				 * change, don't decrease vt of the class.
695 				 */
696 				vt = max_cl->cl_vt;
697 				if (cl->cl_parent->cl_cvtmin != 0)
698 					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
699 
700 				if (cl->cl_parent->cl_vtperiod !=
701 				    cl->cl_parentperiod || vt > cl->cl_vt)
702 					cl->cl_vt = vt;
703 			} else {
704 				/*
705 				 * first child for a new parent backlog period.
706 				 * initialize cl_vt to the highest value seen
707 				 * among the siblings. this is analogous to
708 				 * what cur_time would provide in realtime case.
709 				 */
710 				cl->cl_vt = cl->cl_parent->cl_cvtoff;
711 				cl->cl_parent->cl_cvtmin = 0;
712 			}
713 
714 			/* update the virtual curve */
715 			rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
716 			cl->cl_vtadj = 0;
717 
718 			cl->cl_vtperiod++;  /* increment vt period */
719 			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
720 			if (cl->cl_parent->cl_nactive == 0)
721 				cl->cl_parentperiod++;
722 			cl->cl_f = 0;
723 
724 			vttree_insert(cl);
725 			cftree_insert(cl);
726 
727 			if (cl->cl_flags & HFSC_USC) {
728 				/* class has upper limit curve */
729 				if (cur_time == 0)
730 					cur_time = psched_get_time();
731 
732 				/* update the ulimit curve */
733 				rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
734 					 cl->cl_total);
735 				/* compute myf */
736 				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
737 						      cl->cl_total);
738 			}
739 		}
740 
741 		f = max(cl->cl_myf, cl->cl_cfmin);
742 		if (f != cl->cl_f) {
743 			cl->cl_f = f;
744 			cftree_update(cl);
745 		}
746 		update_cfmin(cl->cl_parent);
747 	}
748 }
749 
750 static void
update_vf(struct hfsc_class * cl,unsigned int len,u64 cur_time)751 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
752 {
753 	u64 f; /* , myf_bound, delta; */
754 	int go_passive = 0;
755 
756 	if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
757 		go_passive = 1;
758 
759 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
760 		cl->cl_total += len;
761 
762 		if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
763 			continue;
764 
765 		if (go_passive && --cl->cl_nactive == 0)
766 			go_passive = 1;
767 		else
768 			go_passive = 0;
769 
770 		/* update vt */
771 		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
772 
773 		/*
774 		 * if vt of the class is smaller than cvtmin,
775 		 * the class was skipped in the past due to non-fit.
776 		 * if so, we need to adjust vtadj.
777 		 */
778 		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
779 			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
780 			cl->cl_vt = cl->cl_parent->cl_cvtmin;
781 		}
782 
783 		if (go_passive) {
784 			/* no more active child, going passive */
785 
786 			/* update cvtoff of the parent class */
787 			if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
788 				cl->cl_parent->cl_cvtoff = cl->cl_vt;
789 
790 			/* remove this class from the vt tree */
791 			vttree_remove(cl);
792 
793 			cftree_remove(cl);
794 			update_cfmin(cl->cl_parent);
795 
796 			continue;
797 		}
798 
799 		/* update the vt tree */
800 		vttree_update(cl);
801 
802 		/* update f */
803 		if (cl->cl_flags & HFSC_USC) {
804 			cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
805 #if 0
806 			cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
807 							      cl->cl_total);
808 			/*
809 			 * This code causes classes to stay way under their
810 			 * limit when multiple classes are used at gigabit
811 			 * speed. needs investigation. -kaber
812 			 */
813 			/*
814 			 * if myf lags behind by more than one clock tick
815 			 * from the current time, adjust myfadj to prevent
816 			 * a rate-limited class from going greedy.
817 			 * in a steady state under rate-limiting, myf
818 			 * fluctuates within one clock tick.
819 			 */
820 			myf_bound = cur_time - PSCHED_JIFFIE2US(1);
821 			if (cl->cl_myf < myf_bound) {
822 				delta = cur_time - cl->cl_myf;
823 				cl->cl_myfadj += delta;
824 				cl->cl_myf += delta;
825 			}
826 #endif
827 		}
828 
829 		f = max(cl->cl_myf, cl->cl_cfmin);
830 		if (f != cl->cl_f) {
831 			cl->cl_f = f;
832 			cftree_update(cl);
833 			update_cfmin(cl->cl_parent);
834 		}
835 	}
836 }
837 
838 static unsigned int
qdisc_peek_len(struct Qdisc * sch)839 qdisc_peek_len(struct Qdisc *sch)
840 {
841 	struct sk_buff *skb;
842 	unsigned int len;
843 
844 	skb = sch->ops->peek(sch);
845 	if (unlikely(skb == NULL)) {
846 		qdisc_warn_nonwc("qdisc_peek_len", sch);
847 		return 0;
848 	}
849 	len = qdisc_pkt_len(skb);
850 
851 	return len;
852 }
853 
854 static void
hfsc_adjust_levels(struct hfsc_class * cl)855 hfsc_adjust_levels(struct hfsc_class *cl)
856 {
857 	struct hfsc_class *p;
858 	unsigned int level;
859 
860 	do {
861 		level = 0;
862 		list_for_each_entry(p, &cl->children, siblings) {
863 			if (p->level >= level)
864 				level = p->level + 1;
865 		}
866 		cl->level = level;
867 	} while ((cl = cl->cl_parent) != NULL);
868 }
869 
870 static inline struct hfsc_class *
hfsc_find_class(u32 classid,struct Qdisc * sch)871 hfsc_find_class(u32 classid, struct Qdisc *sch)
872 {
873 	struct hfsc_sched *q = qdisc_priv(sch);
874 	struct Qdisc_class_common *clc;
875 
876 	clc = qdisc_class_find(&q->clhash, classid);
877 	if (clc == NULL)
878 		return NULL;
879 	return container_of(clc, struct hfsc_class, cl_common);
880 }
881 
882 static void
hfsc_change_rsc(struct hfsc_class * cl,struct tc_service_curve * rsc,u64 cur_time)883 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
884 		u64 cur_time)
885 {
886 	sc2isc(rsc, &cl->cl_rsc);
887 	rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
888 	cl->cl_eligible = cl->cl_deadline;
889 	if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
890 		cl->cl_eligible.dx = 0;
891 		cl->cl_eligible.dy = 0;
892 	}
893 	cl->cl_flags |= HFSC_RSC;
894 }
895 
896 static void
hfsc_change_fsc(struct hfsc_class * cl,struct tc_service_curve * fsc)897 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
898 {
899 	sc2isc(fsc, &cl->cl_fsc);
900 	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
901 	cl->cl_flags |= HFSC_FSC;
902 }
903 
904 static void
hfsc_change_usc(struct hfsc_class * cl,struct tc_service_curve * usc,u64 cur_time)905 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
906 		u64 cur_time)
907 {
908 	sc2isc(usc, &cl->cl_usc);
909 	rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
910 	cl->cl_flags |= HFSC_USC;
911 }
912 
913 static void
hfsc_upgrade_rt(struct hfsc_class * cl)914 hfsc_upgrade_rt(struct hfsc_class *cl)
915 {
916 	cl->cl_fsc = cl->cl_rsc;
917 	rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
918 	cl->cl_flags |= HFSC_FSC;
919 }
920 
921 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
922 	[TCA_HFSC_RSC]	= { .len = sizeof(struct tc_service_curve) },
923 	[TCA_HFSC_FSC]	= { .len = sizeof(struct tc_service_curve) },
924 	[TCA_HFSC_USC]	= { .len = sizeof(struct tc_service_curve) },
925 };
926 
927 static int
hfsc_change_class(struct Qdisc * sch,u32 classid,u32 parentid,struct nlattr ** tca,unsigned long * arg,struct netlink_ext_ack * extack)928 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
929 		  struct nlattr **tca, unsigned long *arg,
930 		  struct netlink_ext_ack *extack)
931 {
932 	struct hfsc_sched *q = qdisc_priv(sch);
933 	struct hfsc_class *cl = (struct hfsc_class *)*arg;
934 	struct hfsc_class *parent = NULL;
935 	struct nlattr *opt = tca[TCA_OPTIONS];
936 	struct nlattr *tb[TCA_HFSC_MAX + 1];
937 	struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
938 	u64 cur_time;
939 	int err;
940 
941 	if (opt == NULL)
942 		return -EINVAL;
943 
944 	err = nla_parse_nested_deprecated(tb, TCA_HFSC_MAX, opt, hfsc_policy,
945 					  NULL);
946 	if (err < 0)
947 		return err;
948 
949 	if (tb[TCA_HFSC_RSC]) {
950 		rsc = nla_data(tb[TCA_HFSC_RSC]);
951 		if (rsc->m1 == 0 && rsc->m2 == 0)
952 			rsc = NULL;
953 	}
954 
955 	if (tb[TCA_HFSC_FSC]) {
956 		fsc = nla_data(tb[TCA_HFSC_FSC]);
957 		if (fsc->m1 == 0 && fsc->m2 == 0)
958 			fsc = NULL;
959 	}
960 
961 	if (tb[TCA_HFSC_USC]) {
962 		usc = nla_data(tb[TCA_HFSC_USC]);
963 		if (usc->m1 == 0 && usc->m2 == 0)
964 			usc = NULL;
965 	}
966 
967 	if (cl != NULL) {
968 		int old_flags;
969 		int len = 0;
970 
971 		if (parentid) {
972 			if (cl->cl_parent &&
973 			    cl->cl_parent->cl_common.classid != parentid)
974 				return -EINVAL;
975 			if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
976 				return -EINVAL;
977 		}
978 		cur_time = psched_get_time();
979 
980 		if (tca[TCA_RATE]) {
981 			err = gen_replace_estimator(&cl->bstats, NULL,
982 						    &cl->rate_est,
983 						    NULL,
984 						    true,
985 						    tca[TCA_RATE]);
986 			if (err)
987 				return err;
988 		}
989 
990 		sch_tree_lock(sch);
991 		old_flags = cl->cl_flags;
992 
993 		if (rsc != NULL)
994 			hfsc_change_rsc(cl, rsc, cur_time);
995 		if (fsc != NULL)
996 			hfsc_change_fsc(cl, fsc);
997 		if (usc != NULL)
998 			hfsc_change_usc(cl, usc, cur_time);
999 
1000 		if (cl->qdisc->q.qlen != 0)
1001 			len = qdisc_peek_len(cl->qdisc);
1002 		/* Check queue length again since some qdisc implementations
1003 		 * (e.g., netem/codel) might empty the queue during the peek
1004 		 * operation.
1005 		 */
1006 		if (cl->qdisc->q.qlen != 0) {
1007 			if (cl->cl_flags & HFSC_RSC) {
1008 				if (old_flags & HFSC_RSC)
1009 					update_ed(cl, len);
1010 				else
1011 					init_ed(cl, len);
1012 			}
1013 
1014 			if (cl->cl_flags & HFSC_FSC) {
1015 				if (old_flags & HFSC_FSC)
1016 					update_vf(cl, 0, cur_time);
1017 				else
1018 					init_vf(cl, len);
1019 			}
1020 		}
1021 		sch_tree_unlock(sch);
1022 
1023 		return 0;
1024 	}
1025 
1026 	if (parentid == TC_H_ROOT)
1027 		return -EEXIST;
1028 
1029 	parent = &q->root;
1030 	if (parentid) {
1031 		parent = hfsc_find_class(parentid, sch);
1032 		if (parent == NULL)
1033 			return -ENOENT;
1034 	}
1035 
1036 	if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1037 		return -EINVAL;
1038 	if (hfsc_find_class(classid, sch))
1039 		return -EEXIST;
1040 
1041 	if (rsc == NULL && fsc == NULL)
1042 		return -EINVAL;
1043 
1044 	cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1045 	if (cl == NULL)
1046 		return -ENOBUFS;
1047 
1048 	RB_CLEAR_NODE(&cl->el_node);
1049 
1050 	err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1051 	if (err) {
1052 		kfree(cl);
1053 		return err;
1054 	}
1055 
1056 	if (tca[TCA_RATE]) {
1057 		err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1058 					NULL, true, tca[TCA_RATE]);
1059 		if (err) {
1060 			tcf_block_put(cl->block);
1061 			kfree(cl);
1062 			return err;
1063 		}
1064 	}
1065 
1066 	if (rsc != NULL)
1067 		hfsc_change_rsc(cl, rsc, 0);
1068 	if (fsc != NULL)
1069 		hfsc_change_fsc(cl, fsc);
1070 	if (usc != NULL)
1071 		hfsc_change_usc(cl, usc, 0);
1072 
1073 	cl->cl_common.classid = classid;
1074 	cl->sched     = q;
1075 	cl->cl_parent = parent;
1076 	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1077 				      classid, NULL);
1078 	if (cl->qdisc == NULL)
1079 		cl->qdisc = &noop_qdisc;
1080 	else
1081 		qdisc_hash_add(cl->qdisc, true);
1082 	INIT_LIST_HEAD(&cl->children);
1083 	cl->vt_tree = RB_ROOT;
1084 	cl->cf_tree = RB_ROOT;
1085 
1086 	sch_tree_lock(sch);
1087 	/* Check if the inner class is a misconfigured 'rt' */
1088 	if (!(parent->cl_flags & HFSC_FSC) && parent != &q->root) {
1089 		NL_SET_ERR_MSG(extack,
1090 			       "Forced curve change on parent 'rt' to 'sc'");
1091 		hfsc_upgrade_rt(parent);
1092 	}
1093 	qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1094 	list_add_tail(&cl->siblings, &parent->children);
1095 	if (parent->level == 0)
1096 		qdisc_purge_queue(parent->qdisc);
1097 	hfsc_adjust_levels(parent);
1098 	sch_tree_unlock(sch);
1099 
1100 	qdisc_class_hash_grow(sch, &q->clhash);
1101 
1102 	*arg = (unsigned long)cl;
1103 	return 0;
1104 }
1105 
1106 static void
hfsc_destroy_class(struct Qdisc * sch,struct hfsc_class * cl)1107 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1108 {
1109 	struct hfsc_sched *q = qdisc_priv(sch);
1110 
1111 	tcf_block_put(cl->block);
1112 	qdisc_put(cl->qdisc);
1113 	gen_kill_estimator(&cl->rate_est);
1114 	if (cl != &q->root)
1115 		kfree(cl);
1116 }
1117 
1118 static int
hfsc_delete_class(struct Qdisc * sch,unsigned long arg,struct netlink_ext_ack * extack)1119 hfsc_delete_class(struct Qdisc *sch, unsigned long arg,
1120 		  struct netlink_ext_ack *extack)
1121 {
1122 	struct hfsc_sched *q = qdisc_priv(sch);
1123 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1124 
1125 	if (cl->level > 0 || qdisc_class_in_use(&cl->cl_common) ||
1126 	    cl == &q->root) {
1127 		NL_SET_ERR_MSG(extack, "HFSC class in use");
1128 		return -EBUSY;
1129 	}
1130 
1131 	sch_tree_lock(sch);
1132 
1133 	list_del(&cl->siblings);
1134 	hfsc_adjust_levels(cl->cl_parent);
1135 
1136 	qdisc_purge_queue(cl->qdisc);
1137 	qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1138 
1139 	sch_tree_unlock(sch);
1140 
1141 	hfsc_destroy_class(sch, cl);
1142 	return 0;
1143 }
1144 
1145 static struct hfsc_class *
hfsc_classify(struct sk_buff * skb,struct Qdisc * sch,int * qerr)1146 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1147 {
1148 	struct hfsc_sched *q = qdisc_priv(sch);
1149 	struct hfsc_class *head, *cl;
1150 	struct tcf_result res;
1151 	struct tcf_proto *tcf;
1152 	int result;
1153 
1154 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1155 	    (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1156 		if (cl->level == 0)
1157 			return cl;
1158 
1159 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1160 	head = &q->root;
1161 	tcf = rcu_dereference_bh(q->root.filter_list);
1162 	while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
1163 #ifdef CONFIG_NET_CLS_ACT
1164 		switch (result) {
1165 		case TC_ACT_QUEUED:
1166 		case TC_ACT_STOLEN:
1167 		case TC_ACT_TRAP:
1168 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1169 			fallthrough;
1170 		case TC_ACT_SHOT:
1171 			return NULL;
1172 		}
1173 #endif
1174 		cl = (struct hfsc_class *)res.class;
1175 		if (!cl) {
1176 			cl = hfsc_find_class(res.classid, sch);
1177 			if (!cl)
1178 				break; /* filter selected invalid classid */
1179 			if (cl->level >= head->level)
1180 				break; /* filter may only point downwards */
1181 		}
1182 
1183 		if (cl->level == 0)
1184 			return cl; /* hit leaf class */
1185 
1186 		/* apply inner filter chain */
1187 		tcf = rcu_dereference_bh(cl->filter_list);
1188 		head = cl;
1189 	}
1190 
1191 	/* classification failed, try default class */
1192 	cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle),
1193 				       READ_ONCE(q->defcls)), sch);
1194 	if (cl == NULL || cl->level > 0)
1195 		return NULL;
1196 
1197 	return cl;
1198 }
1199 
1200 static int
hfsc_graft_class(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)1201 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1202 		 struct Qdisc **old, struct netlink_ext_ack *extack)
1203 {
1204 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1205 
1206 	if (cl->level > 0)
1207 		return -EINVAL;
1208 	if (new == NULL) {
1209 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1210 					cl->cl_common.classid, NULL);
1211 		if (new == NULL)
1212 			new = &noop_qdisc;
1213 	}
1214 
1215 	*old = qdisc_replace(sch, new, &cl->qdisc);
1216 	return 0;
1217 }
1218 
1219 static struct Qdisc *
hfsc_class_leaf(struct Qdisc * sch,unsigned long arg)1220 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1221 {
1222 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1223 
1224 	if (cl->level == 0)
1225 		return cl->qdisc;
1226 
1227 	return NULL;
1228 }
1229 
1230 static void
hfsc_qlen_notify(struct Qdisc * sch,unsigned long arg)1231 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1232 {
1233 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1234 
1235 	/* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1236 	 * needs to be called explicitly to remove a class from vttree.
1237 	 */
1238 	if (cl->cl_nactive)
1239 		update_vf(cl, 0, 0);
1240 	if (cl->cl_flags & HFSC_RSC)
1241 		eltree_remove(cl);
1242 }
1243 
1244 static unsigned long
hfsc_search_class(struct Qdisc * sch,u32 classid)1245 hfsc_search_class(struct Qdisc *sch, u32 classid)
1246 {
1247 	return (unsigned long)hfsc_find_class(classid, sch);
1248 }
1249 
1250 static unsigned long
hfsc_bind_tcf(struct Qdisc * sch,unsigned long parent,u32 classid)1251 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1252 {
1253 	struct hfsc_class *p = (struct hfsc_class *)parent;
1254 	struct hfsc_class *cl = hfsc_find_class(classid, sch);
1255 
1256 	if (cl != NULL) {
1257 		if (p != NULL && p->level <= cl->level)
1258 			return 0;
1259 		qdisc_class_get(&cl->cl_common);
1260 	}
1261 
1262 	return (unsigned long)cl;
1263 }
1264 
1265 static void
hfsc_unbind_tcf(struct Qdisc * sch,unsigned long arg)1266 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1267 {
1268 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1269 
1270 	qdisc_class_put(&cl->cl_common);
1271 }
1272 
hfsc_tcf_block(struct Qdisc * sch,unsigned long arg,struct netlink_ext_ack * extack)1273 static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1274 					struct netlink_ext_ack *extack)
1275 {
1276 	struct hfsc_sched *q = qdisc_priv(sch);
1277 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1278 
1279 	if (cl == NULL)
1280 		cl = &q->root;
1281 
1282 	return cl->block;
1283 }
1284 
1285 static int
hfsc_dump_sc(struct sk_buff * skb,int attr,struct internal_sc * sc)1286 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1287 {
1288 	struct tc_service_curve tsc;
1289 
1290 	tsc.m1 = sm2m(sc->sm1);
1291 	tsc.d  = dx2d(sc->dx);
1292 	tsc.m2 = sm2m(sc->sm2);
1293 	if (nla_put(skb, attr, sizeof(tsc), &tsc))
1294 		goto nla_put_failure;
1295 
1296 	return skb->len;
1297 
1298  nla_put_failure:
1299 	return -1;
1300 }
1301 
1302 static int
hfsc_dump_curves(struct sk_buff * skb,struct hfsc_class * cl)1303 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1304 {
1305 	if ((cl->cl_flags & HFSC_RSC) &&
1306 	    (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1307 		goto nla_put_failure;
1308 
1309 	if ((cl->cl_flags & HFSC_FSC) &&
1310 	    (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1311 		goto nla_put_failure;
1312 
1313 	if ((cl->cl_flags & HFSC_USC) &&
1314 	    (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1315 		goto nla_put_failure;
1316 
1317 	return skb->len;
1318 
1319  nla_put_failure:
1320 	return -1;
1321 }
1322 
1323 static int
hfsc_dump_class(struct Qdisc * sch,unsigned long arg,struct sk_buff * skb,struct tcmsg * tcm)1324 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1325 		struct tcmsg *tcm)
1326 {
1327 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1328 	struct nlattr *nest;
1329 
1330 	tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1331 					  TC_H_ROOT;
1332 	tcm->tcm_handle = cl->cl_common.classid;
1333 	if (cl->level == 0)
1334 		tcm->tcm_info = cl->qdisc->handle;
1335 
1336 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1337 	if (nest == NULL)
1338 		goto nla_put_failure;
1339 	if (hfsc_dump_curves(skb, cl) < 0)
1340 		goto nla_put_failure;
1341 	return nla_nest_end(skb, nest);
1342 
1343  nla_put_failure:
1344 	nla_nest_cancel(skb, nest);
1345 	return -EMSGSIZE;
1346 }
1347 
1348 static int
hfsc_dump_class_stats(struct Qdisc * sch,unsigned long arg,struct gnet_dump * d)1349 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1350 	struct gnet_dump *d)
1351 {
1352 	struct hfsc_class *cl = (struct hfsc_class *)arg;
1353 	struct tc_hfsc_stats xstats;
1354 	__u32 qlen;
1355 
1356 	qdisc_qstats_qlen_backlog(cl->qdisc, &qlen, &cl->qstats.backlog);
1357 	xstats.level   = cl->level;
1358 	xstats.period  = cl->cl_vtperiod;
1359 	xstats.work    = cl->cl_total;
1360 	xstats.rtwork  = cl->cl_cumul;
1361 
1362 	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
1363 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1364 	    gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1365 		return -1;
1366 
1367 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1368 }
1369 
1370 
1371 
1372 static void
hfsc_walk(struct Qdisc * sch,struct qdisc_walker * arg)1373 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1374 {
1375 	struct hfsc_sched *q = qdisc_priv(sch);
1376 	struct hfsc_class *cl;
1377 	unsigned int i;
1378 
1379 	if (arg->stop)
1380 		return;
1381 
1382 	for (i = 0; i < q->clhash.hashsize; i++) {
1383 		hlist_for_each_entry(cl, &q->clhash.hash[i],
1384 				     cl_common.hnode) {
1385 			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
1386 				return;
1387 		}
1388 	}
1389 }
1390 
1391 static void
hfsc_schedule_watchdog(struct Qdisc * sch)1392 hfsc_schedule_watchdog(struct Qdisc *sch)
1393 {
1394 	struct hfsc_sched *q = qdisc_priv(sch);
1395 	struct hfsc_class *cl;
1396 	u64 next_time = 0;
1397 
1398 	cl = eltree_get_minel(q);
1399 	if (cl)
1400 		next_time = cl->cl_e;
1401 	if (q->root.cl_cfmin != 0) {
1402 		if (next_time == 0 || next_time > q->root.cl_cfmin)
1403 			next_time = q->root.cl_cfmin;
1404 	}
1405 	if (next_time)
1406 		qdisc_watchdog_schedule(&q->watchdog, next_time);
1407 }
1408 
1409 static int
hfsc_init_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1410 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1411 		struct netlink_ext_ack *extack)
1412 {
1413 	struct hfsc_sched *q = qdisc_priv(sch);
1414 	struct tc_hfsc_qopt *qopt;
1415 	int err;
1416 
1417 	qdisc_watchdog_init(&q->watchdog, sch);
1418 
1419 	if (!opt || nla_len(opt) < sizeof(*qopt))
1420 		return -EINVAL;
1421 	qopt = nla_data(opt);
1422 
1423 	q->defcls = qopt->defcls;
1424 	err = qdisc_class_hash_init(&q->clhash);
1425 	if (err < 0)
1426 		return err;
1427 	q->eligible = RB_ROOT;
1428 
1429 	err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1430 	if (err)
1431 		return err;
1432 
1433 	gnet_stats_basic_sync_init(&q->root.bstats);
1434 	q->root.cl_common.classid = sch->handle;
1435 	q->root.sched   = q;
1436 	q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1437 					  sch->handle, NULL);
1438 	if (q->root.qdisc == NULL)
1439 		q->root.qdisc = &noop_qdisc;
1440 	else
1441 		qdisc_hash_add(q->root.qdisc, true);
1442 	INIT_LIST_HEAD(&q->root.children);
1443 	q->root.vt_tree = RB_ROOT;
1444 	q->root.cf_tree = RB_ROOT;
1445 
1446 	qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1447 	qdisc_class_hash_grow(sch, &q->clhash);
1448 
1449 	return 0;
1450 }
1451 
1452 static int
hfsc_change_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1453 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1454 		  struct netlink_ext_ack *extack)
1455 {
1456 	struct hfsc_sched *q = qdisc_priv(sch);
1457 	struct tc_hfsc_qopt *qopt;
1458 
1459 	if (nla_len(opt) < sizeof(*qopt))
1460 		return -EINVAL;
1461 	qopt = nla_data(opt);
1462 
1463 	WRITE_ONCE(q->defcls, qopt->defcls);
1464 
1465 	return 0;
1466 }
1467 
1468 static void
hfsc_reset_class(struct hfsc_class * cl)1469 hfsc_reset_class(struct hfsc_class *cl)
1470 {
1471 	cl->cl_total        = 0;
1472 	cl->cl_cumul        = 0;
1473 	cl->cl_d            = 0;
1474 	cl->cl_e            = 0;
1475 	cl->cl_vt           = 0;
1476 	cl->cl_vtadj        = 0;
1477 	cl->cl_cvtmin       = 0;
1478 	cl->cl_cvtoff       = 0;
1479 	cl->cl_vtperiod     = 0;
1480 	cl->cl_parentperiod = 0;
1481 	cl->cl_f            = 0;
1482 	cl->cl_myf          = 0;
1483 	cl->cl_cfmin        = 0;
1484 	cl->cl_nactive      = 0;
1485 
1486 	cl->vt_tree = RB_ROOT;
1487 	cl->cf_tree = RB_ROOT;
1488 	qdisc_reset(cl->qdisc);
1489 
1490 	if (cl->cl_flags & HFSC_RSC)
1491 		rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1492 	if (cl->cl_flags & HFSC_FSC)
1493 		rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1494 	if (cl->cl_flags & HFSC_USC)
1495 		rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1496 }
1497 
1498 static void
hfsc_reset_qdisc(struct Qdisc * sch)1499 hfsc_reset_qdisc(struct Qdisc *sch)
1500 {
1501 	struct hfsc_sched *q = qdisc_priv(sch);
1502 	struct hfsc_class *cl;
1503 	unsigned int i;
1504 
1505 	for (i = 0; i < q->clhash.hashsize; i++) {
1506 		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1507 			hfsc_reset_class(cl);
1508 	}
1509 	q->eligible = RB_ROOT;
1510 	qdisc_watchdog_cancel(&q->watchdog);
1511 }
1512 
1513 static void
hfsc_destroy_qdisc(struct Qdisc * sch)1514 hfsc_destroy_qdisc(struct Qdisc *sch)
1515 {
1516 	struct hfsc_sched *q = qdisc_priv(sch);
1517 	struct hlist_node *next;
1518 	struct hfsc_class *cl;
1519 	unsigned int i;
1520 
1521 	for (i = 0; i < q->clhash.hashsize; i++) {
1522 		hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1523 			tcf_block_put(cl->block);
1524 			cl->block = NULL;
1525 		}
1526 	}
1527 	for (i = 0; i < q->clhash.hashsize; i++) {
1528 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1529 					  cl_common.hnode)
1530 			hfsc_destroy_class(sch, cl);
1531 	}
1532 	qdisc_class_hash_destroy(&q->clhash);
1533 	qdisc_watchdog_cancel(&q->watchdog);
1534 }
1535 
1536 static int
hfsc_dump_qdisc(struct Qdisc * sch,struct sk_buff * skb)1537 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1538 {
1539 	struct hfsc_sched *q = qdisc_priv(sch);
1540 	unsigned char *b = skb_tail_pointer(skb);
1541 	struct tc_hfsc_qopt qopt;
1542 
1543 	qopt.defcls = READ_ONCE(q->defcls);
1544 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1545 		goto nla_put_failure;
1546 	return skb->len;
1547 
1548  nla_put_failure:
1549 	nlmsg_trim(skb, b);
1550 	return -1;
1551 }
1552 
1553 static int
hfsc_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)1554 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1555 {
1556 	unsigned int len = qdisc_pkt_len(skb);
1557 	struct hfsc_class *cl;
1558 	int err;
1559 	bool first;
1560 
1561 	cl = hfsc_classify(skb, sch, &err);
1562 	if (cl == NULL) {
1563 		if (err & __NET_XMIT_BYPASS)
1564 			qdisc_qstats_drop(sch);
1565 		__qdisc_drop(skb, to_free);
1566 		return err;
1567 	}
1568 
1569 	first = !cl->qdisc->q.qlen;
1570 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1571 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1572 		if (net_xmit_drop_count(err)) {
1573 			cl->qstats.drops++;
1574 			qdisc_qstats_drop(sch);
1575 		}
1576 		return err;
1577 	}
1578 
1579 	sch->qstats.backlog += len;
1580 	sch->q.qlen++;
1581 
1582 	if (first && !cl_in_el_or_vttree(cl)) {
1583 		if (cl->cl_flags & HFSC_RSC)
1584 			init_ed(cl, len);
1585 		if (cl->cl_flags & HFSC_FSC)
1586 			init_vf(cl, len);
1587 		/*
1588 		 * If this is the first packet, isolate the head so an eventual
1589 		 * head drop before the first dequeue operation has no chance
1590 		 * to invalidate the deadline.
1591 		 */
1592 		if (cl->cl_flags & HFSC_RSC)
1593 			cl->qdisc->ops->peek(cl->qdisc);
1594 
1595 	}
1596 
1597 	return NET_XMIT_SUCCESS;
1598 }
1599 
1600 static struct sk_buff *
hfsc_dequeue(struct Qdisc * sch)1601 hfsc_dequeue(struct Qdisc *sch)
1602 {
1603 	struct hfsc_sched *q = qdisc_priv(sch);
1604 	struct hfsc_class *cl;
1605 	struct sk_buff *skb;
1606 	u64 cur_time;
1607 	unsigned int next_len;
1608 	int realtime = 0;
1609 
1610 	if (sch->q.qlen == 0)
1611 		return NULL;
1612 
1613 	cur_time = psched_get_time();
1614 
1615 	/*
1616 	 * if there are eligible classes, use real-time criteria.
1617 	 * find the class with the minimum deadline among
1618 	 * the eligible classes.
1619 	 */
1620 	cl = eltree_get_mindl(q, cur_time);
1621 	if (cl) {
1622 		realtime = 1;
1623 	} else {
1624 		/*
1625 		 * use link-sharing criteria
1626 		 * get the class with the minimum vt in the hierarchy
1627 		 */
1628 		cl = vttree_get_minvt(&q->root, cur_time);
1629 		if (cl == NULL) {
1630 			qdisc_qstats_overlimit(sch);
1631 			hfsc_schedule_watchdog(sch);
1632 			return NULL;
1633 		}
1634 	}
1635 
1636 	skb = qdisc_dequeue_peeked(cl->qdisc);
1637 	if (skb == NULL) {
1638 		qdisc_warn_nonwc("HFSC", cl->qdisc);
1639 		return NULL;
1640 	}
1641 
1642 	bstats_update(&cl->bstats, skb);
1643 	update_vf(cl, qdisc_pkt_len(skb), cur_time);
1644 	if (realtime)
1645 		cl->cl_cumul += qdisc_pkt_len(skb);
1646 
1647 	if (cl->cl_flags & HFSC_RSC) {
1648 		if (cl->qdisc->q.qlen != 0) {
1649 			/* update ed */
1650 			next_len = qdisc_peek_len(cl->qdisc);
1651 			/* Check queue length again since some qdisc implementations
1652 			 * (e.g., netem/codel) might empty the queue during the peek
1653 			 * operation.
1654 			 */
1655 			if (cl->qdisc->q.qlen != 0) {
1656 				if (realtime)
1657 					update_ed(cl, next_len);
1658 				else
1659 					update_d(cl, next_len);
1660 			}
1661 		} else {
1662 			/* the class becomes passive */
1663 			eltree_remove(cl);
1664 		}
1665 	}
1666 
1667 	qdisc_bstats_update(sch, skb);
1668 	qdisc_qstats_backlog_dec(sch, skb);
1669 	sch->q.qlen--;
1670 
1671 	return skb;
1672 }
1673 
1674 static const struct Qdisc_class_ops hfsc_class_ops = {
1675 	.change		= hfsc_change_class,
1676 	.delete		= hfsc_delete_class,
1677 	.graft		= hfsc_graft_class,
1678 	.leaf		= hfsc_class_leaf,
1679 	.qlen_notify	= hfsc_qlen_notify,
1680 	.find		= hfsc_search_class,
1681 	.bind_tcf	= hfsc_bind_tcf,
1682 	.unbind_tcf	= hfsc_unbind_tcf,
1683 	.tcf_block	= hfsc_tcf_block,
1684 	.dump		= hfsc_dump_class,
1685 	.dump_stats	= hfsc_dump_class_stats,
1686 	.walk		= hfsc_walk
1687 };
1688 
1689 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1690 	.id		= "hfsc",
1691 	.init		= hfsc_init_qdisc,
1692 	.change		= hfsc_change_qdisc,
1693 	.reset		= hfsc_reset_qdisc,
1694 	.destroy	= hfsc_destroy_qdisc,
1695 	.dump		= hfsc_dump_qdisc,
1696 	.enqueue	= hfsc_enqueue,
1697 	.dequeue	= hfsc_dequeue,
1698 	.peek		= qdisc_peek_dequeued,
1699 	.cl_ops		= &hfsc_class_ops,
1700 	.priv_size	= sizeof(struct hfsc_sched),
1701 	.owner		= THIS_MODULE
1702 };
1703 MODULE_ALIAS_NET_SCH("hfsc");
1704 
1705 static int __init
hfsc_init(void)1706 hfsc_init(void)
1707 {
1708 	return register_qdisc(&hfsc_qdisc_ops);
1709 }
1710 
1711 static void __exit
hfsc_cleanup(void)1712 hfsc_cleanup(void)
1713 {
1714 	unregister_qdisc(&hfsc_qdisc_ops);
1715 }
1716 
1717 MODULE_LICENSE("GPL");
1718 MODULE_DESCRIPTION("Hierarchical Fair Service Curve scheduler");
1719 module_init(hfsc_init);
1720 module_exit(hfsc_cleanup);
1721